diff --git "a/abs_29K_G/test_abstract_long_2405.03003v1.json" "b/abs_29K_G/test_abstract_long_2405.03003v1.json"
new file mode 100644--- /dev/null
+++ "b/abs_29K_G/test_abstract_long_2405.03003v1.json"
@@ -0,0 +1,679 @@
+{
+    "url": "http://arxiv.org/abs/2405.03003v1",
+    "title": "Parameter-Efficient Fine-Tuning with Discrete Fourier Transform",
+    "abstract": "Low-rank adaptation~(LoRA) has recently gained much interest in fine-tuning\nfoundation models. It effectively reduces the number of trainable parameters by\nincorporating low-rank matrices $A$ and $B$ to represent the weight change,\ni.e., $\\Delta W=BA$. Despite LoRA's progress, it faces storage challenges when\nhandling extensive customization adaptations or larger base models. In this\nwork, we aim to further compress trainable parameters by enjoying the powerful\nexpressiveness of the Fourier transform. Specifically, we introduce FourierFT,\nwhich treats $\\Delta W$ as a matrix in the spatial domain and learns only a\nsmall fraction of its spectral coefficients. With the trained spectral\ncoefficients, we implement the inverse discrete Fourier transform to recover\n$\\Delta W$. Empirically, our FourierFT method shows comparable or better\nperformance with fewer parameters than LoRA on various tasks, including natural\nlanguage understanding, natural language generation, instruction tuning, and\nimage classification. For example, when performing instruction tuning on the\nLLaMA2-7B model, FourierFT surpasses LoRA with only 0.064M trainable\nparameters, compared to LoRA's 33.5M. Our code is released at\n\\url{https://github.com/Chaos96/fourierft}.",
+    "authors": "Ziqi Gao, Qichao Wang, Aochuan Chen, Zijing Liu, Bingzhe Wu, Liang Chen, Jia Li",
+    "published": "2024-05-05",
+    "updated": "2024-05-05",
+    "primary_cat": "cs.LG",
+    "cats": [
+        "cs.LG",
+        "cs.AI",
+        "cs.CL"
+    ],
+    "label": "Original Paper",
+    "paper_cat": "Parameter AND Efficient AND Fine AND Tuning",
+    "gt": "Low-rank adaptation~(LoRA) has recently gained much interest in fine-tuning\nfoundation models. It effectively reduces the number of trainable parameters by\nincorporating low-rank matrices $A$ and $B$ to represent the weight change,\ni.e., $\\Delta W=BA$. Despite LoRA's progress, it faces storage challenges when\nhandling extensive customization adaptations or larger base models. In this\nwork, we aim to further compress trainable parameters by enjoying the powerful\nexpressiveness of the Fourier transform. Specifically, we introduce FourierFT,\nwhich treats $\\Delta W$ as a matrix in the spatial domain and learns only a\nsmall fraction of its spectral coefficients. With the trained spectral\ncoefficients, we implement the inverse discrete Fourier transform to recover\n$\\Delta W$. Empirically, our FourierFT method shows comparable or better\nperformance with fewer parameters than LoRA on various tasks, including natural\nlanguage understanding, natural language generation, instruction tuning, and\nimage classification. For example, when performing instruction tuning on the\nLLaMA2-7B model, FourierFT surpasses LoRA with only 0.064M trainable\nparameters, compared to LoRA's 33.5M. Our code is released at\n\\url{https://github.com/Chaos96/fourierft}.",
+    "main_content": "Introduction Large foundation models (LFMs) have demonstrated exceptional performance on tasks of multiple domains, including natural language processing (NLP) (Liu et al., 2019; He et al., 2020; Radford et al., 2019; Brown et al., 2020; Li et al., 2022) and computer vision (CV) (Liu et al., 2023a;b; Singh et al., 2022; Rombach et al., 2022). Owing to their *Equal contribution 1Hong Kong University of Science and Technology (Guangzhou) 2Hong Kong University of Science and Technology 3Sun Yat-sen University 4International Digital Economy Academy 5AI Lab, Tencent. Correspondence to: Jia Li <jialee@ust.hk>. Proceedings of the 41 st International Conference on Machine Learning, Vienna, Austria. PMLR 235, 2024. Copyright 2024 by the author(s). Figure 1. Summary of the performance (y-axis) of fine-tuning methods with different numbers (x-axis) of trainable parameters on NLP (left) and CV (right) tasks. The left side shows the instruction tuning task, where the LLaMA2-7B model is fine-tuned with Alpaca and evaluated by GPT-4. The right side shows the image classification task, where the Vision Transformer (ViT) is finetuned and tested on the DTD dataset. Black circles (\u25cf) represent the Full Fine-tuning (FF) method. Orange circles (\u25cf) represent LoRA method with r = {32, 64, 128} (left) and r = {8, 16, 32} (right). Blue circles (\u25cf) represent our proposed method with n = {1000, 2000} (left) and n = {3000, 10000} (right). impressive capabilities, fine-tuning LFMs for a wide range of downstream tasks has become prevalent (Wang et al., 2022; Taori et al., 2023; Qiu et al., 2020). Under the full fine-tuning paradigm, the new model adapted to each customized task typically contains as many parameters as the original model (Qiu et al., 2020; Raffel et al., 2020; Chen et al., 2024; Gao et al., 2024). As models grow larger and customization needs expand, the demand for storing finetuned checkpoints rises, resulting in both costly storage and memory consumption. As a popular way to address this issue, LoRA (Hu et al., 2021) represents the weight change with two low-rank matrices A and B, i.e., W0+\u2206W = W0+BA. Despite LoRA\u2019s superb performance, its large size of trainable parameters still brings high IT infrastructure consumption, which affects both ends of public communities and individual users. For the former, an intuitive example is that a LoRA adapter (finetuned weights) for a specific style of the stable diffusion model (Rombach et al., 2022) requires about 40MB of memory. This necessitates the LFM communities (e.g., Civi1 arXiv:2405.03003v1  [cs.LG]  5 May 2024 \fParameter-Efficient Fine-Tuning with Discrete Fourier Transform tai (Civitai, 2024)) to bear high storage and bandwidth costs to cater to a large user base. For the latter, fewer parameters mean direct RAM savings when loading fine-tuned weights in mobile APPs, enabling sufficient customization for individual users (Zhou et al., 2022). To this end, we naturally ask the question: How can we aggressively compress trainable parameters even further for fine-tuning LFMs? Previous works have demonstrated the powerful expressiveness of Fourier basis in data compression, where extremely sparse spectral information can be used to recover highfidelity data (e.g., 1D signal vectors (Zwartjes & Gisolf, 2007; Duarte & Baraniuk, 2013; Rudelson & Vershynin, 2008) and 2D image matrices (Vlaardingerbroek & Boer, 2013; Song et al., 2021; Shi et al., 2014)). More importantly, when dealing with more general (non-image) matrices that lack strong spatial semantics and are not frequency-sparse, Fourier transform can still handle recovery effectively (Chen & Chi, 2013; Yang & Xie, 2016). Motivated by this, we investigate the potential for updating the weight change \u2206W with its sparse spectral coefficients for fine-tuning LFMs. In this paper, we aim to aggressively reduce the number of trainable parameters for fine-tuning LFMs. To this end, we propose FourierFT (Fourier Transform for Fine-Tuning), which treats the weight change \u2206W as a matrix in the spatial domain, and learns its sparse spectral coefficients. Specifically, we first randomly select n spectral entries that are shared across all layers. For each layer, FourierFT learns n spectral coefficients located at these n selected entries and then directly applies inverse discrete Fourier transform to compute the updated \u2206W. Therefore, fine-tuning a pretrained model with Lt layers only requires storing 2n entry parameters and nLt coefficient parameters for FourierFT. Empirically, we compare our method with state-of-the-art LoRA variants and other parameter-efficient fine-tuning methods on various tasks including (1) natural language understanding (on the GLUE benchmark), (2) natural language generation (on the E2E benchmark), (3) instruction tuning (with LLaMA-family models), and (4) image classification (with vision transformers). FourierFT can always achieve comparable or even better performance than LoRA, with about 6.0%, 9.4%, 0.2% and 9.2% of LoRA\u2019s trainable parameters for these 4 tasks, respectively. For example in Figure 1, on the instruction tuning task, our FourierFT method outperforms LoRA with only 64K trainable parameters. Moreover, it achieves a comparable score to Full Fine-tuning with only 128K parameters. 2. Related Works Parameter-Efficient Fine-Tuning. With the rapid expansion of large foundation models (LFM), it has become challenging and important to efficiently adapt them for specific tasks. To this end, numerous methods for parameter-efficient fine-tuning (PEFT) are proposed, demonstrating impressive capabilities in both efficiency and accuracy. Existing PEFT methods are broadly partitioned into two categories: nonweight-based and weight-based methods. Non-weight-based methods do not optimize pre-trained LFMs at the weight level. Instead, they achieve fine-tunings by introducing additional modules or optimizing prompts and prefixes. Adapter tuning (He et al., 2021; Rebuffi et al., 2017; Pfeiffer et al., 2020; Houlsby et al., 2019; R\u00a8 uckl\u00b4 e et al., 2020; Lin et al., 2020) aims to introduce light-weighted neural modules, called adapters, between pre-trained layers of the base model. These methods keep the pre-trained weights frozen and efficiently fine-tune the adapters for customized tasks. Prompt tuning (Brown et al., 2020; Lester et al., 2021; Gao et al., 2020; Diao et al., 2022) and prefix tuning (Li & Liang, 2021) insert additional prompts or prefix tokens to the layers of the base model. Weight-based methods, represented by LoRA (Hu et al., 2021), introduce and then update weight changes that can be merged with the original weights to avoid inference latency. LoRA\u2019s innovation lies in the multiplication of low-rank matrices to approximate weight changes. Building upon this, AdaLoRA (Zhang et al., 2023) extends the LoRA method by distributing the parameter budget across weight matrices with importance scores. Additionally, Q-LoRA (Dettmers et al., 2023) proposes to back-propagate gradients upon LoRA through a quantized pre-trained model with 4-bit NormalFloat. Here, we focus on weight-based methods and achieve huge parameter reduction with the powerful expressiveness of Fourier basis, rather than following the low-rank structure. Sparse Fourier Transform in Deep Learning. Sparse Fourier transform (SFT) has flourished in various fields of deep learning (DL). The SFT technique mainly involves using sparse spectral coefficients of significant (Xu et al., 2020; Ehrlich & Davis, 2019; Gueguen et al., 2018; Tang et al., 2022) or even random (Lin et al., 2014; Rawat et al., 2019; Herrmann, 2010) spectral entries, for representation learning. One important application of this technique is matrix recovery. Patel et al. (2011) designs a gradient-based compressed sensing method to recover images with their sparse Fourier information. Shechtman et al. (2014) proposes an efficient phase retrieval method that improves data recovery using sparse Fourier coefficients. Importantly, previous works (Chen & Chi, 2013; Yang & Xie, 2016; Gao et al., 2022) show that even when the original data is not frequency-sparse, SFT can effectively recover the data with extremely few parameters. Although previous works lack studies on the recovery for the weight matrices of DL models with SFT, the aforementioned methods provide potential support for this work. 2 \fParameter-Efficient Fine-Tuning with Discrete Fourier Transform Pre-trained Weights \ud835\udc4a\u2208\u211d!!\u00d7!\" \ud835\udc35= 0 \ud835\udc34= \ud835\udca9(0, \ud835\udf0e!) \u210e \ud835\udc65 \ud835\udc51# \ud835\udc51$ \ud835\udc5f Pre-trained Weights \ud835\udc4a\u2208\u211d!!\u00d7!\" \u210e \ud835\udc65 \ud835\udc51# \ud835\udc51$ Random entries (shared across layers) \u211d!\u00d7# \ud835\udc5b Coefficients : Frozen : Trainable LoRA FourierFT IDFT Dense Spectral Matrix F Figure 2. Overview of LoRA (left) and our FourierFT (right) method. In LoRA, only low-rank (r) matrices A and B are trained. The weight change is represented by their multiplication, i.e., \u2206W = BA. For each pre-trained weight W, the theoretical number of trainable parameters in LoRA is r \u00d7 (d1 + d2). In FourierFT, we first randomly generate the spectral entry matrix R2\u00d7n, which is shared across all layers to reduce parameter storage requirements. The complete spectral matrix is formed by a trainable coefficient vector Rn located at selected entries and 0s at the remaining entries. We obtain the weight change \u2206W by directly performing inverse discrete Fourier transform (IDFT) on the updated spectral matrix. For all L adapted layers, FourierFT needs to store n \u00d7 (2 + L) parameters. 3. Method We present FourierFT (depicted in Figure 2), a parameterefficient fine-tuning method based on discrete Fourier transform. FourierFT follows the principle of only learning the change in the pre-trained weight, as proposed by LoRA (Hu et al., 2021). However, unlike LoRA, FourierFT does not adopt the low-rank structure but learns a set of spectral coefficients of Fourier basis. Specifically, we randomly initialize the spectral entry matrix, which is frozen and shared across all layers. We make the spectral coefficients located at selected entries trainable, which jointly form the spectral matrix. Lastly, we apply the inverse discrete Fourier transform to the spectral matrix, yielding its spatial-domain counterpart as the updated weight change. 3.1. Forward Pass We follow the paradigm of only learning weight changes, as adopted by LoRA-based methods (Hu et al., 2021; Dettmers et al., 2023; Zhang et al., 2023). This can avoid inference latency by merging the pre-trained weight and its change. Formally, we define each pre-trained weight matrix as W0 \u2208 Rd1\u00d7d2, and the weight change for fine-tuning as \u2206W \u2208 Rd1\u00d7d2. LoRA aims to parameterize \u2206W in the form of low-rank decomposition in the forward pass: h = W0x + \u2206Wx = W0x + BAx, (1) where B \u2208Rd1\u00d7r and A \u2208Rr\u00d7d2 with the rank r \u226a min(d1,d2) are trainable matrices. The advantage of FourierFT is that the orthogonal and expressive Fourier basis enables recovery of informative weight changes. This promisingly suggests achieving comparable performance to LoRA with significantly fewer parameters. We first randomly initialize the entry matrix E \u2208R2\u00d7n containing discrete 2D spectral entries. Then we randomly initialize the coefficients c \u2208Rn with a normal Gaussian distribution. The proposed forward pass is: F = TODENSE(E,c) (2) Sp,q = d1\u22121 \u2211 j=0 d2\u22121 \u2211 k=0 Fj,kei2\u03c0( p d1 j+ q d2 k) (3) h = W0x + \u2206Wx = W0x + \u03b1R(S)x. (4) Specifically, TODENSE in Eq. 2 represents to construct the spectral matrix F \u2208Rd1\u00d7d2, i.e., Fj,k = cl (resp. 0), if j = E0,l & k = E1,l (resp. else). Eq. 3 computes the spatio matrix S via the inverse discrete Fourier transform, where i represents the imaginary unit. Finally, in Eq. 4, we take the real part of the complex matrix S (denoted as R(S)) and scale it by \u03b1. Kindly note that all layers involve training various c vectors, while sharing the matrix E and value \u03b1. The pseudocode for FourierFT is shown as Algorithm 1, adhering to the PyTorch style. Initialization for the Entry Matrix E. Previous works lack studies on the importance of the spectral entries in the weight change. Thus, we fill this gap by introducing adjustable frequency bias, causing the entries to be more likely sampled in this area. In addition to randomly sampling entries in the full d1 \u00d7 d2-sized spectral matrix (i.e., no bias), we also implement entry sampling with a bias towards a favored central frequency, e.g., low, middle, or 3 \fParameter-Efficient Fine-Tuning with Discrete Fourier Transform Algorithm 1 PyTorch-style pseudocode for FourierFT. class FourierFT(nn.Module): def __init__( self, n: int = 100, # number of trainable parameters alpha: float = 300.0, # scaling d1: int = 4096, # input dimension d2: int = 4096, # output dimension base_layer: nn.Module # pre-trained layer ) # definitions self.d1 = d1 self.d2 = d2 self.n = n self.alpha = alpha self.base_layer = base_layer # entry initialization (no frequency bias) self.E = torch.randperm(d1 * d2)[:n] self.E = torch.stack([self.E // self.d1, self.E % self.d2], dim=0) # spectral coefficient initialization self.c = nn.Parameter(torch.randn(n), \\\\ requires_grad=True) def forward(self, x: torch.Tensor): # get dense spectral matrix (Eq.2) F = torch.zeros(self.d1, self.d2) F[self.E[0, :], self.E[1, :]] = self.c # compute Delta_W (Eq.3) Delta_W = torch.fft.ifft2(F).real * self.alpha # merge (Eq.4) h = self.base_layer(x) h += torch.einsum(\u2019ijk,kl->ijl\u2019, x, Delta_W) return h high frequencies. Formally, we apply the Gaussian bandpass filter (Gonzales & Wintz, 1987) to model the sampling probability for the entry (u,v),0 \u2264u \u2264d1\u22121,0 \u2264v \u2264d2\u22121: p(u,v) = exp\u239b \u239d\u2212(D2 \u2212f 2 c DW ) 2\u239e \u23a0, (5) where D represents the distance from the point (u,v) to the origin (center of the matrix), fc is the favored central frequency, and W represents the bandwidth. In Figure 3, we visualize the sampling probability map of a 768 \u00d7 768-sized spectral matrix with different fc and W = 200. fc = 0 fc = 100 fc = 200 fc = 350 fc = 480 0 0.5 1 Figure 3. Visualization of entry sampling probability at different favored central frequencies fc. Kindly note that unless specially stated, FourierFT is set by default to the entry initialization with no frequency bias. 3.2. Parameter Summary We summarize the number of trainable parameters for LoRA and FourierFT in Table 1. LoRA relies on a pair of trainable matrices A and B for each layer. Let the number of layers for fine-tuning be Lt. The total number of parameters in Table 1. Theoretical number of trainable parameters and storage requirements for fine-tuning. For both LoRA and FourierFT methods, only the query and value layers are tuned within the transformer architectures. The configurations that are exactly chosen in the \u2018Experiments\u2019 Section are highlighted . Base Models LoRA FourierFT r # Trainable Parameters Required Bytes n # Trainable Parameters Required Bytes RoBERTa Base 4 147K 574KB 200 4.8K 18.8KB 8 295K 1.13MB 200 24K 94KB RoBERTa Large 4 393K 1.5MB 200 9.6K 36.5KB 8 786K 3MB 1000 48K 183KB GPT-2 Medium 4 350K 1.34MB 500 24K 94KB 8 786K 3MB 1000 48K 188KB GPT-2 Large 4 737K 2.81MB 500 36K 141KB 8 1.47M 5.74MB 1000 72K 282KB LLaMA-2 7B 16 8.39M 32.8MB 1000 64K 250KB 64 33.5M 131.1MB 2000 128K 500KB LLaMA-2 13B 16 13.1M 51.2MB 1000 80K 312KB 64 52.4M 204.8MB 2000 160K 625KB ViT Base 8 295K 1.13MB 3000 72K 281KB 16 590K 2.25MB 10000 239K 934KB ViT Large 8 786K 2.93MB 3000 144K 563KB 16 1.57M 6MB 10000 480K 1.83MB LoRA is determined by the rank r and the dimension of weights d = d1 = d2: \u2223\u0398\u2223LoRA = 2 \u00d7 d \u00d7 Lt \u00d7 r. For Fourier, the total number takes the form: \u2223\u0398\u2223F ourierF T = n \u00d7 Lt. As an intuitive example, the RoBERTa Base model contains 12 transformer blocks with d = 768, resulting in Lt = 24 layers when we only fine-tune the query and value ones. Therefore, we have \u2223\u0398\u2223LoRA = 294,912 for r = 8, and \u2223\u0398\u2223F ourierF T = 24,000 for n = 1000. In Table 1, we highlight the configurations where LoRA and our method achieve matched performance in subsequent experiments. We note that the advantage of parameter efficiency in FourierFT becomes more pronounced as the model\u2019s scale (depth and width) increases (e.g., RoBERTa Base \u2192RoBERTa Large). This could be because \u2223\u0398\u2223LoRA has an explicit linear relationship with width d, unlike \u2223\u0398\u2223F ourierF T . 4. Experiments In this section, we evaluate FourierFT in the domains of natural language processing (NLP) and computer vision (CV). For NLP, we implement FourierFT for fine-tuning (1) RoBERTa (Base & Large) on natural language understanding (GLUE, (Wang et al., 2018)), (2) GPT-2 (Medium & Large) on natural language generation (E2E, (Novikova et al., 2017)) and (3) LLaMA-family models (7B & 13B) on instruction tuning. For CV, we apply FourierFT to fine-tune the (4) vision transformers (Base & Large) on image classification. Finally, we conduct ablation studies to analyze the effect of frequency bias, the parameter scalability, and the 4 \fParameter-Efficient Fine-Tuning with Discrete Fourier Transform Table 2. Performance of various fine-tuning methods with RoBERTa Base (RoBbase) and RoBERTa Large (RoBlarge) models on 6 datasets of the GLUE benchmark. We report the Matthew\u2019s correlation coefficient (MCC) for CoLA, Pearson correlation coefficient (PCC) for STS-B and accuracy (Acc.) for all the remaining tasks. We report the median result of 5 runs, each using different random seeds. The best results for each dataset are shown in bold. Higher is better for all metrics in 6 datasets. Model & Method # Trainable Parameters SST-2 (Acc.) MRPC (Acc.) CoLA (MCC) QNLI (Acc.) RTE (Acc.) STS-B (PCC) Avg. RoBbase(FF) 125M 94.8 90.2 63.6 92.8 78.7 91.2 85.2 RoBbase(BitFit) 0.1M 93.7 92.7 62 91.8 81.5 90.8 85.4 RoBbase(AdptD) 0.3M 94.2\u00b10.1 88.5\u00b11.1 60.8\u00b10.4 93.1\u00b10.1 71.5\u00b12.7 89.7\u00b10.3 83.0 RoBbase(AdptD) 0.9M 94.7\u00b10.3 88.4\u00b10.1 62.6\u00b10.9 93.0\u00b10.2 75.9\u00b12.2 90.3\u00b10.1 84.2 RoBbase(LoRA) 0.3M 95.1\u00b10.2 89.7\u00b10.7 63.4\u00b11.2 93.3\u00b10.3 78.4\u00b10.8 91.5\u00b10.2 85.2 RoBbase(AdaLoRA) 0.3M 94.5\u00b10.2 88.7\u00b10.5 62.0\u00b10.6 93.1\u00b10.2 81.0\u00b10.6 90.5\u00b10.2 85.0 RoBbase(DyLoRA) 0.3M 94.3\u00b10.5 89.5\u00b10.5 61.1\u00b10.3 92.2\u00b10.5 78.7\u00b10.7 91.1\u00b10.6 84.5 RoBbase(FourierFT) 0.024M 94.2\u00b10.3 90.0\u00b10.8 63.8\u00b11.6 92.2\u00b10.1 79.1\u00b10.5 90.8\u00b10.2 85.0 RoBlarge(FF) 356M 96.4 90.9 68 94.7 86.6 92.4 88.2 RoBlarge(AdptP) 3M 96.1\u00b10.3 90.2\u00b10.7 68.3\u00b11.0 94.8\u00b10.2 83.8\u00b12.9 92.1\u00b10.7 87.6 RoBlarge(AdptP) 0.8M 96.6\u00b10.2 89.7\u00b11.2 67.8\u00b12.5 94.8\u00b10.3 80.1\u00b12.9 91.9\u00b10.4 86.8 RoBlarge(AdptH) 6M 96.2\u00b10.3 88.7\u00b12.9 66.5\u00b14.4 94.7\u00b10.2 83.4\u00b11.1 91.0\u00b11.7 86.8 RoBlarge(AdptH) 0.8M 96.3\u00b10.5 87.7\u00b11.7 66.3\u00b12.0 94.7\u00b10.2 72.9\u00b12.9 91.5\u00b10.5 84.9 RoBlarge(LoRA) 0.8M 96.2\u00b10.5 90.2\u00b11.0 68.2\u00b11.9 94.8\u00b10.3 85.2\u00b11.1 92.3\u00b10.5 87.8 RoBlarge(FourierFT) 0.048M 96.0\u00b10.2 90.9\u00b10.3 67.1\u00b11.4 94.4\u00b10.4 87.4\u00b11.6 91.9\u00b10.4 88.0 expressiveness of the Fourier basis. Baselines. We compare our FourierFT method with popular parameter-efficient fine-tuning (PEFT) methods. To ensure a comprehensive and fair comparison, we prioritize replicating the setups used in previous works and reusing their reported results. Involved baselines are: \u25cfFull Fine-tuning (FF) During fine-tuning, the base model is initialized with pre-trained weights and biases, and all parameters will undergo gradient updates. \u25cfBitfit (Zaken et al., 2021) Only the bias vectors are finetuned while all other parameters are frozen. \u25cfAdapter tuning This research line was first investigated by Houlsby et al. (2019), which proposes the AdapterH method. AdapterH inserts two-layer adapters between the self-attention and the FNN modules, followed by a subsequent residual connection. We compare it with three additional variants of it. AdapterL (Lin et al., 2020) is more parameter-efficient, with adapter layers applied only after the MLP modules and subsequent to a LayerNorm. AdapterP (Pfeiffer et al., 2020) implements the adapter layers after the feed-forward layer. This design was chosen through a grid search including all settings related to the adapter\u2019s position, number, ect. AdapterD (R\u00a8 uckl\u00b4 e et al., 2020) further enhances the parameter efficiency by dropping adapter layers that are not activated. \u25cfLoRA (Hu et al., 2021) LoRA is the state-of-the-art method for PEFT. It parameterizes incremental weight updates using trainable low-rank matrices. \u25cfDyLoRA (Valipour et al., 2022) This method trains dynamic search-free LoRA models for the best rank choice. \u25cfAdaLoRA (Zhang et al., 2023) This method proposes the SVD-based fine-tuning and prunes redundant singular values with the importance-aware rank allocation. 4.1. Natural Language Understanding Models and Datasets. We evaluate our method on the GLUE benchmark (General Language Understanding Evaluation (Wang et al., 2018)), which consists of a wide range of natural language understanding (NLU) tasks, including single-sentence classification tasks, similarity and paraphrase tasks and natural language inference tasks. We finetune the pre-trained RoBERTa Base and Large foundation models (Liu et al., 2019) for evaluation. Implementation Details. For both models, FourierFT is allowed to have 1000 out of 7682 (RoBERTa Base) and 10242 (RoBERTa Large) trainable spectral coefficients in each layer, i.e., n = 1000. We randomly sample the spectral entries with no frequency bias, which is shared1 across all 24 (Base) and 48 (Large) layers. For all 6 datasets in GLUE, we tune the hyperparameters of the learning rates and the scaling values. We follow the experimental setup applied in Hu et al. (2021), which involves fine-tuning only the query and value weights in each transformer block and 1We use the value 2024 as the seed for all layers. 5 \fParameter-Efficient Fine-Tuning with Discrete Fourier Transform Table 3. Results from GPT-2 Medium and Large models on the E2E benchmark. We present the result from the final epoch. For all metrics, higher values indicate better performance. * indicates that the results are taken from prior works. Best results are shown in bold. Model Method # Trainable Parameters BLEU NIST METEOR ROUGE-L CIDEr GPT-2 Medium FT* 354.92M 68.2 8.62 46.2 71.0 2.47 AdptL* 0.37M 66.3 8.41 45.0 69.8 2.40 AdptL* 11.09M 68.9 8.71 46.1 71.3 2.47 AdptH* 11.09M 67.3\u00b1.6 8.5\u00b1.07 46.0\u00b1.2 70.7\u00b1.2 2.44\u00b1.01 LoRA 0.35M 68.9\u00b1.3 8.76\u00b1.06 46.6\u00b1.1 71.5\u00b1.1 2.53\u00b1.03 FourierFT 0.048M 69.1\u00b1.1 8.82 \u00b1.05 47.0 \u00b1.3 71.8 \u00b1.1 2.51\u00b1.02 GPT-2 Large FT* 774.03M 68.5 8.78 46.0 69.9 2.45 AdptL* 0.88M 69.1\u00b1.1 8.68\u00b1.03 46.3\u00b1.0 71.4\u00b1.2 2.49\u00b1.0 AdptL* 23.00M 68.9\u00b1.3 8.70\u00b1.04 46.1\u00b1.1 71.3\u00b1.2 2.45\u00b1.02 LoRA 0.77M 70.1\u00b1.3 8.83\u00b1.02 46.8\u00b1.2 72.0\u00b1.3 2.47\u00b1.02 FourierFT 0.072M 70.2\u00b1.2 8.90\u00b1.02 47.0\u00b1.2 71.8\u00b1.1 2.50 \u00b1.02 Table 4. The average scores on MT-Bench and Vicuna assessed by GPT-4. \u2020 indicates updating the layers other than lm head. Higher score is better. Model Method # Trainable Parameters MT-Bench Vicuna LLaMA1-7B LoRA\u2020 159.9M 5.05\u00b1.3 6.85\u00b1.4 LoRA 33.5M 4.99\u00b1.3 6.81\u00b1.3 FourierFT 0.064M 5.09\u00b1.6 6.85\u00b1.8 LLaMA1-13B LoRA\u2020 250.3M 5.28\u00b1.6 7.02\u00b1.3 LoRA 52.4M 5.21\u00b1.4 6.97\u00b1.4 FourierFT 0.08M 5.23\u00b1.3 7.14\u00b1.5 LLaMA2-7B LoRA\u2020 159.9M 5.19\u00b1.1 7.38\u00b1.3 LoRA 33.5M 5.20\u00b1.3 7.35\u00b1.6 FourierFT 0.064M 5.18\u00b1.3 7.49\u00b1.4 LLaMA2-13B LoRA\u2020 250.3M 5.78\u00b1.2 7.89\u00b1.5 LoRA 52.4M 5.80\u00b1.2 7.89\u00b1.6 FourierFT 0.08M 5.82\u00b1.3 7.92\u00b1.5 fully fine-tuning the classification head. We provide the hyperparameters in Table 9 in Appendix. Results. Results are summarized in Table 2. Following Hu et al. (2021), Zhang et al. (2023) and Valipour et al. (2022), we specify the number of trainable parameters for the finetuned layers excluding the classification head. We report the median of 5 random seed results, where the best epoch is selected for each run. In general, FourierFT achieves better or on-par performance compared with baseline methods with significantly fewer trainable parameters. Notably, FourierFT outperforms all baselines including fully fine-tuning the RoBERTa Base on CoLA and the RoBERTa Large on RTE. As mentioned in Section 3.2, the parameter count of LoRA is dependent on both the width and depth of models, resulting in a larger count growth (LoRA: 0.8M/0.3M \u22482.7; ours: 0.048M/0.024M = 2) compared to FourierFT. Nevertheless, FourierFT still performs comparably to LoRA, demonstrating the potential scalability of our method when facing even larger models. 4.2. Natural Language Generation Models and Datasets. We evaluate the performance of FourierFT on the E2E natural language generation (NLG) task (Novikova et al., 2017). We fine-tune the GPT-2 (Radford et al., 2019) Medium (354M) and Large (774M) models, which are both decoder-only and have 24 and 36 transformer blocks, respectively. The E2E benchmark contains roughly 42,000 training, 4,600 validation and 4,600 test samples from the restaurant domain. Implementation Details. We report prior results for baselines other than LoRA. For both LoRA and our method, we fine-tune the GPT-2 Medium and Large models with a linear learning rate scheduler for 5 epochs, where we tune the batch size and learning rate. We report the average results over 3 runs, where the last epoch is selected for each run. We provide the hyperparameters in Table 10 in Appendix. Results. We show the results in Table 3. We note that FourierFT can achieve the best performance on most metrics. More importantly, FourierFT only requires 13.7% and 9.4% of the parameter counts of LoRA, for the GPT-2 Medium and Large models respectively. 4.3. Instruction Tuning Models and Datasets. Instruction tuning, as described in (Ouyang et al., 2022; Wei et al., 2021; Mishra et al., 2021), refers to the process of fine-tuning a language model on a collection of paired prompts and responses. We apply LoRA and FourierFT to fine-tune the LLaMA (Touvron et al., 2023a) and LLaMA2 (Touvron et al., 2023b) families. Specifically, we consider the LLaMA-7B, LLaMA-13B, LLaMA2-7B and LLaMA2-13B as base models, which are fine-tuned on the Alpaca dataset (Taori et al., 2023). Alpaca contains 51K instruction-following demonstrations generated from text-davinci-003 (GPT-3.5) (Wang et al., 2022). For evaluation, we use the fine-tuned models to generate responses for the pre-defined questions, which are from the MT-Bench (Zheng et al., 2023) and Vicuna Eval (Chiang et al., 2023). GPT-4 takes these answers as input and evaluates them with scores within 10. Implementation Details. For LoRA, we use r = 64 and apply two configurations: (1) updating all linear layers except the language modelling head (lm head); (2) updating only the WQ and WV matrices. For FourierFT, we only adopt the latter configuration with n = 1000. To ensure the 6 \fParameter-Efficient Fine-Tuning with Discrete Fourier Transform Table 5. Fine-tuning results with ViT Base and Large models on different image classification datasets. We report the accuracy (%) after 10 epochs. Avg. represents the average accuracy of each method on all datasets. The best performance is shown in bold. Model Method # Trainable Parameters OxfordPets StanfordCars CIFAR10 DTD EuroSAT FGVC RESISC45 CIFAR100 Avg. ViT-Base LP 90.28\u00b10.43 25.76\u00b10.28 96.41\u00b10.02 69.77\u00b10.67 88.72\u00b10.13 17.44\u00b10.43 74.22\u00b10.10 84.28\u00b10.11 68.36 FF 85.8M 93.14\u00b10.40 79.78\u00b11.15 98.92\u00b10.05 77.68\u00b11.21 99.05\u00b10.09 54.84\u00b11.23 96.13\u00b10.13 92.38\u00b10.13 86.49 LoRA 581K 93.19\u00b10.36 45.38\u00b10.41 98.78\u00b10.05 74.95\u00b10.40 98.44\u00b10.15 25.16\u00b10.16 92.70\u00b10.18 92.02\u00b10.12 77.58 FourierFT 72K 93.21\u00b10.26 46.11\u00b10.24 98.58\u00b10.07 75.09\u00b10.37 98.29\u00b10.04 27.51\u00b10.64 91.97\u00b10.31 91.20\u00b10.14 77.75 FourierFT 239K 93.05\u00b10.34 56.36\u00b10.66 98.69\u00b10.08 77.30\u00b10.61 98.78\u00b10.11 32.44\u00b10.99 94.26\u00b10.20 91.45\u00b10.18 80.29 ViT-Large LP 91.11\u00b10.30 37.91\u00b10.27 97.78\u00b10.04 73.33\u00b10.26 92.64\u00b10.08 24.62\u00b10.24 82.02\u00b10.11 84.28\u00b10.11 72.96 FF 303.3M 94.43\u00b10.56 88.90\u00b10.26 99.15\u00b10.05 81.79\u00b11.01 99.04\u00b10.08 68.25\u00b11.63 96.43\u00b10.07 93.58\u00b10.19 90.20 LoRA 1.57M 94.82\u00b10.09 73.25\u00b10.36 99.13\u00b10.03 81.79\u00b10.45 98.63\u00b10.07 42.32\u00b10.98 94.71\u00b10.25 94.87\u00b10.10 84.94 FourierFT 144K 94.46\u00b10.28 69.56\u00b10.30 99.10\u00b10.04 80.83\u00b10.43 98.65\u00b10.09 39.92\u00b10.68 93.86\u00b10.14 93.31\u00b10.09 83.71 FourierFT 480K 94.84\u00b10.05 79.14\u00b10.67 99.08\u00b10.01 81.88\u00b10.50 98.66\u00b10.03 51.28\u00b10.68 95.20\u00b10.07 93.37\u00b10.11 86.68 feasibility of training on a single GPU, we deploy the quantization method in Dettmers et al. (2023) for fine-tuning. We train with both methods for only one epoch, and report the average scores of all answers. We provide the hyperparameter setup in Table 11 in the Appendix. Results. The results are shown in Table 4. We find that the expressive power of the 13B model is much stronger than that of the 7B model, regardless of which fine-tuning method is used. Moreover, FourierFT closely matches or slightly exceeds LoRA\u2019s performance with less than 0.2% of its parameters. We provide practical examples containing questions, answers and reviews in the Appendix D. 4.4. Image Classification Models and Datasets. We evaluate our method on the image classification task. We adopt the Base and Large versions of the popular CV foundation model, Vision Transformer (ViT) (Dosovitskiy et al., 2020). The ViTs are pretrained on the ImageNet-21K dataset (Ridnik et al., 2021). The datasets for fine-tuning include OxfordPets (372), CIFAR10 (10), DTD (47), EuroSAT (10) and RESISC45 (45) with small label spaces, as well as StanfordCars (196), FGVC (100) and CIFAR100 (100) with large label spaces. Detailed information is provided in Table 8 in the Appendix. Implementation Details. We include three baselines for evaluation: Full Fine-tuning (FF), Linear Probing (LP, finetuning the classification head only), and LoRA. For both LoRA and our method, only the query and value matrices of ViT are updated. We use r = 16 for LoRA and n = {3000,10000} for FourierFT. We tune the learning rates and weight decay for all methods, and set the maximum training epoch to 10. We provide the hyperparameters in Table 12 in Appendix. 2Numbers in parentheses indicate class counts for each dataset. Results. Table 5 summarizes the results for 8 image classification datasets with the ViT Base and Large models. Both LoRA and FourierFT methods significantly outperform the Linear Probing, demonstrating their effectiveness in the CV domain. Our method obtains matched performance using 12.4% and 9.2% of LoRA\u2019s parameter count, with ViT Base and Large models, respectively. Notably, when we increase the parameter count of FourierFT to 41.1% (ViT Base) and 30.6% (ViT Large) of LoRA\u2019s, it can outperform LoRA by 3.5% and 2.0% respectively. Moreover, our method can even (slightly) outperform the Full Fine-tuning method on OxfordPets and DTD with the ViT Large model. 4.5. Study Effect of Frequency Bias. We examine how the performance is affected by the frequency bias, i.e., the central frequency fc in Eq. 5. We directly apply the optimal hyperparameters searched in Table 2 and fine-tune the RoBERTa Base on the MRPC, STS-B, CoLA and RTE datasets. From Figure 5, we note that the fine-tuning performance of FourierFT without any frequency bias can surpass most cases that are restricted by the central frequency bias. This indicates the universality of our method. Surprisingly, we find that it is always possible to obtain results better than \u201cNo bias\u201d by traversing the fc values. Since this traversal is not efficient, we do not conduct further exploration in this paper. However, we believe that making fc trainable will be a promising new direction for improving FourierFT. Parameter Scalability. We explore the relationship between the number of trainable parameters and the performance of LoRA and our method. We use the set of ranks r = {1,2,4,6,8,15} for LoRA and n = {50,100,200,1000,6144,12288} for FourierFT on 6 tasks of the GLUE benchmark. For both LoRA and ours, the learning rate, and scaling hyperparameters are tuned. For fairness, we ensure that the number of trials for hyperparam7 \fParameter-Efficient Fine-Tuning with Discrete Fourier Transform 4 6 8 10 ln # Trainable Parameters 88 89 90 Accuracy MRPC LoRA FourierFT 4 6 8 10 ln # Trainable Parameters 58 60 62 64 Matthew s Corr. CoLA 4 6 8 10 ln # Trainable Parameters 78 80 Accuracy RTE 4 6 8 10 ln # Trainable Parameters 90.0 90.5 91.0 91.5 Pearson Corr. STS-B 4 6 8 10 ln # Trainable Parameters 94.0 94.5 95.0 Accuracy SST-2 4 6 8 10 ln # Trainable Parameters 90 91 92 Accuracy QQP Figure 4. Performance on the GLUE benchmark with RoBERTa Base vs. number of trainable parameters (each layer) of LoRA and ours. For all 6 datasets, we apply the setting of r = {1, 2, 4, 6, 8, 15} for LoRA and n = {50, 100, 200, 1000, 6144, 12288}. 0 100 200 300 400 500 fc 76 77 78 79 Acc. RTE No bias 0 100 200 300 400 500 fc 89.2 89.4 89.6 89.8 90.0 Acc. MRPC No bias 0 100 200 300 400 500 fc 90.5 90.6 90.7 90.8 PCC. STS-B No bias 0 100 200 300 400 500 fc 62.0 62.5 63.0 63.5 MCC. CoLA No bias Figure 5. Results on 4 datasets in GLUE with different fc values. eter search is 30 for both methods. As shown in Figure 4, our method outperforms LoRA on all 6 datasets. In detail, our method is significantly better than LoRA with the same parameter count, i.e., {r = 4,n = 6144} & {r = 8,n = 12288}. Moreover, we observe that a larger number of parameters does not always bring performance gains for LoRA. On the contrary, the increase of n can consistently improve the accuracy of FourierFT. On most tasks, FourierFT with n = 50 can achieve comparable or even better (MRPC, CoLA, RTE) performance than LoRA with r = 1. In this case, the parameter count in LoRA is about 31 \u00d7 that of ours. Basis Expressiveness. The inverse discrete Fourier transform (IDFT) in Eq. 3 is equivalent to the matrix multiplication (Lu et al., 2021): S = BfFB\u22ba f, where B is the transformation matrix of IDFT that contains the Fourier basis. To evaluate its expressivity, we replace the Fourier basis with random and orthogonal basis, respectively. Specifically, for F \u2208Rd1\u00d7d2, we initialize random basis B1 r \u2208Rd1\u00d7d1 and B2 r \u2208Rd2\u00d7d2 with the normal Gaussian distribution. Then Eq. 3 becomes S = B1 rFB2 r. A similar way is used for the orthogonal basis. We compare FourierFT with the random basis (R-B) and orthogonal basis (O-B) on the GLUE benchmark. Table 6 shows the results. We note that the Fourier basis used in our method outperforms the random and orthogonal basis. In addition, the expressive power of the orthogonal basis is much stronger than that of the random basis. The stronger expressive power of the Fourier basis compared to the general orthogonal basis may be attributed to its effective capture of the spectral information of \u2206W. Table 6. Results with three types of basis. Model RTE CoLA Ours R-B O-B Ours R-B O-B Base 79.1 72.7(\u21938.1%) 75.6(\u21934.4%) 63.8 58.7(\u21938.0%) 60.0(\u21936.0%) Large 87.4 81.8(\u21936.4%) 83.6(\u21934.3%) 67.1 64.8(\u21933.4%) 66.1(\u21931.5%) 5.",
+    "additional_graph_info": {
+        "graph": [
+            [
+                "Ziqi Gao",
+                "Zijing Liu"
+            ],
+            [
+                "Ziqi Gao",
+                "Qichao Wang"
+            ],
+            [
+                "Ziqi Gao",
+                "Aochuan Chen"
+            ],
+            [
+                "Zijing Liu",
+                "Xianbin Ye"
+            ],
+            [
+                "Qichao Wang",
+                "Bingzhe Wu"
+            ],
+            [
+                "Qichao Wang",
+                "Tian Bian"
+            ],
+            [
+                "Qichao Wang",
+                "Yian Yin"
+            ],
+            [
+                "Aochuan Chen",
+                "Sijia Liu"
+            ],
+            [
+                "Aochuan Chen",
+                "Yihua Zhang"
+            ],
+            [
+                "Aochuan Chen",
+                "Yuguang Yao"
+            ],
+            [
+                "Aochuan Chen",
+                "Yimeng Zhang"
+            ]
+        ],
+        "node_feat": {
+            "Ziqi Gao": [
+                {
+                    "url": "http://arxiv.org/abs/2405.03003v1",
+                    "title": "Parameter-Efficient Fine-Tuning with Discrete Fourier Transform",
+                    "abstract": "Low-rank adaptation~(LoRA) has recently gained much interest in fine-tuning\nfoundation models. It effectively reduces the number of trainable parameters by\nincorporating low-rank matrices $A$ and $B$ to represent the weight change,\ni.e., $\\Delta W=BA$. Despite LoRA's progress, it faces storage challenges when\nhandling extensive customization adaptations or larger base models. In this\nwork, we aim to further compress trainable parameters by enjoying the powerful\nexpressiveness of the Fourier transform. Specifically, we introduce FourierFT,\nwhich treats $\\Delta W$ as a matrix in the spatial domain and learns only a\nsmall fraction of its spectral coefficients. With the trained spectral\ncoefficients, we implement the inverse discrete Fourier transform to recover\n$\\Delta W$. Empirically, our FourierFT method shows comparable or better\nperformance with fewer parameters than LoRA on various tasks, including natural\nlanguage understanding, natural language generation, instruction tuning, and\nimage classification. For example, when performing instruction tuning on the\nLLaMA2-7B model, FourierFT surpasses LoRA with only 0.064M trainable\nparameters, compared to LoRA's 33.5M. Our code is released at\n\\url{https://github.com/Chaos96/fourierft}.",
+                    "authors": "Ziqi Gao, Qichao Wang, Aochuan Chen, Zijing Liu, Bingzhe Wu, Liang Chen, Jia Li",
+                    "published": "2024-05-05",
+                    "updated": "2024-05-05",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.AI",
+                        "cs.CL"
+                    ],
+                    "main_content": "Introduction Large foundation models (LFMs) have demonstrated exceptional performance on tasks of multiple domains, including natural language processing (NLP) (Liu et al., 2019; He et al., 2020; Radford et al., 2019; Brown et al., 2020; Li et al., 2022) and computer vision (CV) (Liu et al., 2023a;b; Singh et al., 2022; Rombach et al., 2022). Owing to their *Equal contribution 1Hong Kong University of Science and Technology (Guangzhou) 2Hong Kong University of Science and Technology 3Sun Yat-sen University 4International Digital Economy Academy 5AI Lab, Tencent. Correspondence to: Jia Li <jialee@ust.hk>. Proceedings of the 41 st International Conference on Machine Learning, Vienna, Austria. PMLR 235, 2024. Copyright 2024 by the author(s). Figure 1. Summary of the performance (y-axis) of fine-tuning methods with different numbers (x-axis) of trainable parameters on NLP (left) and CV (right) tasks. The left side shows the instruction tuning task, where the LLaMA2-7B model is fine-tuned with Alpaca and evaluated by GPT-4. The right side shows the image classification task, where the Vision Transformer (ViT) is finetuned and tested on the DTD dataset. Black circles (\u25cf) represent the Full Fine-tuning (FF) method. Orange circles (\u25cf) represent LoRA method with r = {32, 64, 128} (left) and r = {8, 16, 32} (right). Blue circles (\u25cf) represent our proposed method with n = {1000, 2000} (left) and n = {3000, 10000} (right). impressive capabilities, fine-tuning LFMs for a wide range of downstream tasks has become prevalent (Wang et al., 2022; Taori et al., 2023; Qiu et al., 2020). Under the full fine-tuning paradigm, the new model adapted to each customized task typically contains as many parameters as the original model (Qiu et al., 2020; Raffel et al., 2020; Chen et al., 2024; Gao et al., 2024). As models grow larger and customization needs expand, the demand for storing finetuned checkpoints rises, resulting in both costly storage and memory consumption. As a popular way to address this issue, LoRA (Hu et al., 2021) represents the weight change with two low-rank matrices A and B, i.e., W0+\u2206W = W0+BA. Despite LoRA\u2019s superb performance, its large size of trainable parameters still brings high IT infrastructure consumption, which affects both ends of public communities and individual users. For the former, an intuitive example is that a LoRA adapter (finetuned weights) for a specific style of the stable diffusion model (Rombach et al., 2022) requires about 40MB of memory. This necessitates the LFM communities (e.g., Civi1 arXiv:2405.03003v1  [cs.LG]  5 May 2024 \fParameter-Efficient Fine-Tuning with Discrete Fourier Transform tai (Civitai, 2024)) to bear high storage and bandwidth costs to cater to a large user base. For the latter, fewer parameters mean direct RAM savings when loading fine-tuned weights in mobile APPs, enabling sufficient customization for individual users (Zhou et al., 2022). To this end, we naturally ask the question: How can we aggressively compress trainable parameters even further for fine-tuning LFMs? Previous works have demonstrated the powerful expressiveness of Fourier basis in data compression, where extremely sparse spectral information can be used to recover highfidelity data (e.g., 1D signal vectors (Zwartjes & Gisolf, 2007; Duarte & Baraniuk, 2013; Rudelson & Vershynin, 2008) and 2D image matrices (Vlaardingerbroek & Boer, 2013; Song et al., 2021; Shi et al., 2014)). More importantly, when dealing with more general (non-image) matrices that lack strong spatial semantics and are not frequency-sparse, Fourier transform can still handle recovery effectively (Chen & Chi, 2013; Yang & Xie, 2016). Motivated by this, we investigate the potential for updating the weight change \u2206W with its sparse spectral coefficients for fine-tuning LFMs. In this paper, we aim to aggressively reduce the number of trainable parameters for fine-tuning LFMs. To this end, we propose FourierFT (Fourier Transform for Fine-Tuning), which treats the weight change \u2206W as a matrix in the spatial domain, and learns its sparse spectral coefficients. Specifically, we first randomly select n spectral entries that are shared across all layers. For each layer, FourierFT learns n spectral coefficients located at these n selected entries and then directly applies inverse discrete Fourier transform to compute the updated \u2206W. Therefore, fine-tuning a pretrained model with Lt layers only requires storing 2n entry parameters and nLt coefficient parameters for FourierFT. Empirically, we compare our method with state-of-the-art LoRA variants and other parameter-efficient fine-tuning methods on various tasks including (1) natural language understanding (on the GLUE benchmark), (2) natural language generation (on the E2E benchmark), (3) instruction tuning (with LLaMA-family models), and (4) image classification (with vision transformers). FourierFT can always achieve comparable or even better performance than LoRA, with about 6.0%, 9.4%, 0.2% and 9.2% of LoRA\u2019s trainable parameters for these 4 tasks, respectively. For example in Figure 1, on the instruction tuning task, our FourierFT method outperforms LoRA with only 64K trainable parameters. Moreover, it achieves a comparable score to Full Fine-tuning with only 128K parameters. 2. Related Works Parameter-Efficient Fine-Tuning. With the rapid expansion of large foundation models (LFM), it has become challenging and important to efficiently adapt them for specific tasks. To this end, numerous methods for parameter-efficient fine-tuning (PEFT) are proposed, demonstrating impressive capabilities in both efficiency and accuracy. Existing PEFT methods are broadly partitioned into two categories: nonweight-based and weight-based methods. Non-weight-based methods do not optimize pre-trained LFMs at the weight level. Instead, they achieve fine-tunings by introducing additional modules or optimizing prompts and prefixes. Adapter tuning (He et al., 2021; Rebuffi et al., 2017; Pfeiffer et al., 2020; Houlsby et al., 2019; R\u00a8 uckl\u00b4 e et al., 2020; Lin et al., 2020) aims to introduce light-weighted neural modules, called adapters, between pre-trained layers of the base model. These methods keep the pre-trained weights frozen and efficiently fine-tune the adapters for customized tasks. Prompt tuning (Brown et al., 2020; Lester et al., 2021; Gao et al., 2020; Diao et al., 2022) and prefix tuning (Li & Liang, 2021) insert additional prompts or prefix tokens to the layers of the base model. Weight-based methods, represented by LoRA (Hu et al., 2021), introduce and then update weight changes that can be merged with the original weights to avoid inference latency. LoRA\u2019s innovation lies in the multiplication of low-rank matrices to approximate weight changes. Building upon this, AdaLoRA (Zhang et al., 2023) extends the LoRA method by distributing the parameter budget across weight matrices with importance scores. Additionally, Q-LoRA (Dettmers et al., 2023) proposes to back-propagate gradients upon LoRA through a quantized pre-trained model with 4-bit NormalFloat. Here, we focus on weight-based methods and achieve huge parameter reduction with the powerful expressiveness of Fourier basis, rather than following the low-rank structure. Sparse Fourier Transform in Deep Learning. Sparse Fourier transform (SFT) has flourished in various fields of deep learning (DL). The SFT technique mainly involves using sparse spectral coefficients of significant (Xu et al., 2020; Ehrlich & Davis, 2019; Gueguen et al., 2018; Tang et al., 2022) or even random (Lin et al., 2014; Rawat et al., 2019; Herrmann, 2010) spectral entries, for representation learning. One important application of this technique is matrix recovery. Patel et al. (2011) designs a gradient-based compressed sensing method to recover images with their sparse Fourier information. Shechtman et al. (2014) proposes an efficient phase retrieval method that improves data recovery using sparse Fourier coefficients. Importantly, previous works (Chen & Chi, 2013; Yang & Xie, 2016; Gao et al., 2022) show that even when the original data is not frequency-sparse, SFT can effectively recover the data with extremely few parameters. Although previous works lack studies on the recovery for the weight matrices of DL models with SFT, the aforementioned methods provide potential support for this work. 2 \fParameter-Efficient Fine-Tuning with Discrete Fourier Transform Pre-trained Weights \ud835\udc4a\u2208\u211d!!\u00d7!\" \ud835\udc35= 0 \ud835\udc34= \ud835\udca9(0, \ud835\udf0e!) \u210e \ud835\udc65 \ud835\udc51# \ud835\udc51$ \ud835\udc5f Pre-trained Weights \ud835\udc4a\u2208\u211d!!\u00d7!\" \u210e \ud835\udc65 \ud835\udc51# \ud835\udc51$ Random entries (shared across layers) \u211d!\u00d7# \ud835\udc5b Coefficients : Frozen : Trainable LoRA FourierFT IDFT Dense Spectral Matrix F Figure 2. Overview of LoRA (left) and our FourierFT (right) method. In LoRA, only low-rank (r) matrices A and B are trained. The weight change is represented by their multiplication, i.e., \u2206W = BA. For each pre-trained weight W, the theoretical number of trainable parameters in LoRA is r \u00d7 (d1 + d2). In FourierFT, we first randomly generate the spectral entry matrix R2\u00d7n, which is shared across all layers to reduce parameter storage requirements. The complete spectral matrix is formed by a trainable coefficient vector Rn located at selected entries and 0s at the remaining entries. We obtain the weight change \u2206W by directly performing inverse discrete Fourier transform (IDFT) on the updated spectral matrix. For all L adapted layers, FourierFT needs to store n \u00d7 (2 + L) parameters. 3. Method We present FourierFT (depicted in Figure 2), a parameterefficient fine-tuning method based on discrete Fourier transform. FourierFT follows the principle of only learning the change in the pre-trained weight, as proposed by LoRA (Hu et al., 2021). However, unlike LoRA, FourierFT does not adopt the low-rank structure but learns a set of spectral coefficients of Fourier basis. Specifically, we randomly initialize the spectral entry matrix, which is frozen and shared across all layers. We make the spectral coefficients located at selected entries trainable, which jointly form the spectral matrix. Lastly, we apply the inverse discrete Fourier transform to the spectral matrix, yielding its spatial-domain counterpart as the updated weight change. 3.1. Forward Pass We follow the paradigm of only learning weight changes, as adopted by LoRA-based methods (Hu et al., 2021; Dettmers et al., 2023; Zhang et al., 2023). This can avoid inference latency by merging the pre-trained weight and its change. Formally, we define each pre-trained weight matrix as W0 \u2208 Rd1\u00d7d2, and the weight change for fine-tuning as \u2206W \u2208 Rd1\u00d7d2. LoRA aims to parameterize \u2206W in the form of low-rank decomposition in the forward pass: h = W0x + \u2206Wx = W0x + BAx, (1) where B \u2208Rd1\u00d7r and A \u2208Rr\u00d7d2 with the rank r \u226a min(d1,d2) are trainable matrices. The advantage of FourierFT is that the orthogonal and expressive Fourier basis enables recovery of informative weight changes. This promisingly suggests achieving comparable performance to LoRA with significantly fewer parameters. We first randomly initialize the entry matrix E \u2208R2\u00d7n containing discrete 2D spectral entries. Then we randomly initialize the coefficients c \u2208Rn with a normal Gaussian distribution. The proposed forward pass is: F = TODENSE(E,c) (2) Sp,q = d1\u22121 \u2211 j=0 d2\u22121 \u2211 k=0 Fj,kei2\u03c0( p d1 j+ q d2 k) (3) h = W0x + \u2206Wx = W0x + \u03b1R(S)x. (4) Specifically, TODENSE in Eq. 2 represents to construct the spectral matrix F \u2208Rd1\u00d7d2, i.e., Fj,k = cl (resp. 0), if j = E0,l & k = E1,l (resp. else). Eq. 3 computes the spatio matrix S via the inverse discrete Fourier transform, where i represents the imaginary unit. Finally, in Eq. 4, we take the real part of the complex matrix S (denoted as R(S)) and scale it by \u03b1. Kindly note that all layers involve training various c vectors, while sharing the matrix E and value \u03b1. The pseudocode for FourierFT is shown as Algorithm 1, adhering to the PyTorch style. Initialization for the Entry Matrix E. Previous works lack studies on the importance of the spectral entries in the weight change. Thus, we fill this gap by introducing adjustable frequency bias, causing the entries to be more likely sampled in this area. In addition to randomly sampling entries in the full d1 \u00d7 d2-sized spectral matrix (i.e., no bias), we also implement entry sampling with a bias towards a favored central frequency, e.g., low, middle, or 3 \fParameter-Efficient Fine-Tuning with Discrete Fourier Transform Algorithm 1 PyTorch-style pseudocode for FourierFT. class FourierFT(nn.Module): def __init__( self, n: int = 100, # number of trainable parameters alpha: float = 300.0, # scaling d1: int = 4096, # input dimension d2: int = 4096, # output dimension base_layer: nn.Module # pre-trained layer ) # definitions self.d1 = d1 self.d2 = d2 self.n = n self.alpha = alpha self.base_layer = base_layer # entry initialization (no frequency bias) self.E = torch.randperm(d1 * d2)[:n] self.E = torch.stack([self.E // self.d1, self.E % self.d2], dim=0) # spectral coefficient initialization self.c = nn.Parameter(torch.randn(n), \\\\ requires_grad=True) def forward(self, x: torch.Tensor): # get dense spectral matrix (Eq.2) F = torch.zeros(self.d1, self.d2) F[self.E[0, :], self.E[1, :]] = self.c # compute Delta_W (Eq.3) Delta_W = torch.fft.ifft2(F).real * self.alpha # merge (Eq.4) h = self.base_layer(x) h += torch.einsum(\u2019ijk,kl->ijl\u2019, x, Delta_W) return h high frequencies. Formally, we apply the Gaussian bandpass filter (Gonzales & Wintz, 1987) to model the sampling probability for the entry (u,v),0 \u2264u \u2264d1\u22121,0 \u2264v \u2264d2\u22121: p(u,v) = exp\u239b \u239d\u2212(D2 \u2212f 2 c DW ) 2\u239e \u23a0, (5) where D represents the distance from the point (u,v) to the origin (center of the matrix), fc is the favored central frequency, and W represents the bandwidth. In Figure 3, we visualize the sampling probability map of a 768 \u00d7 768-sized spectral matrix with different fc and W = 200. fc = 0 fc = 100 fc = 200 fc = 350 fc = 480 0 0.5 1 Figure 3. Visualization of entry sampling probability at different favored central frequencies fc. Kindly note that unless specially stated, FourierFT is set by default to the entry initialization with no frequency bias. 3.2. Parameter Summary We summarize the number of trainable parameters for LoRA and FourierFT in Table 1. LoRA relies on a pair of trainable matrices A and B for each layer. Let the number of layers for fine-tuning be Lt. The total number of parameters in Table 1. Theoretical number of trainable parameters and storage requirements for fine-tuning. For both LoRA and FourierFT methods, only the query and value layers are tuned within the transformer architectures. The configurations that are exactly chosen in the \u2018Experiments\u2019 Section are highlighted . Base Models LoRA FourierFT r # Trainable Parameters Required Bytes n # Trainable Parameters Required Bytes RoBERTa Base 4 147K 574KB 200 4.8K 18.8KB 8 295K 1.13MB 200 24K 94KB RoBERTa Large 4 393K 1.5MB 200 9.6K 36.5KB 8 786K 3MB 1000 48K 183KB GPT-2 Medium 4 350K 1.34MB 500 24K 94KB 8 786K 3MB 1000 48K 188KB GPT-2 Large 4 737K 2.81MB 500 36K 141KB 8 1.47M 5.74MB 1000 72K 282KB LLaMA-2 7B 16 8.39M 32.8MB 1000 64K 250KB 64 33.5M 131.1MB 2000 128K 500KB LLaMA-2 13B 16 13.1M 51.2MB 1000 80K 312KB 64 52.4M 204.8MB 2000 160K 625KB ViT Base 8 295K 1.13MB 3000 72K 281KB 16 590K 2.25MB 10000 239K 934KB ViT Large 8 786K 2.93MB 3000 144K 563KB 16 1.57M 6MB 10000 480K 1.83MB LoRA is determined by the rank r and the dimension of weights d = d1 = d2: \u2223\u0398\u2223LoRA = 2 \u00d7 d \u00d7 Lt \u00d7 r. For Fourier, the total number takes the form: \u2223\u0398\u2223F ourierF T = n \u00d7 Lt. As an intuitive example, the RoBERTa Base model contains 12 transformer blocks with d = 768, resulting in Lt = 24 layers when we only fine-tune the query and value ones. Therefore, we have \u2223\u0398\u2223LoRA = 294,912 for r = 8, and \u2223\u0398\u2223F ourierF T = 24,000 for n = 1000. In Table 1, we highlight the configurations where LoRA and our method achieve matched performance in subsequent experiments. We note that the advantage of parameter efficiency in FourierFT becomes more pronounced as the model\u2019s scale (depth and width) increases (e.g., RoBERTa Base \u2192RoBERTa Large). This could be because \u2223\u0398\u2223LoRA has an explicit linear relationship with width d, unlike \u2223\u0398\u2223F ourierF T . 4. Experiments In this section, we evaluate FourierFT in the domains of natural language processing (NLP) and computer vision (CV). For NLP, we implement FourierFT for fine-tuning (1) RoBERTa (Base & Large) on natural language understanding (GLUE, (Wang et al., 2018)), (2) GPT-2 (Medium & Large) on natural language generation (E2E, (Novikova et al., 2017)) and (3) LLaMA-family models (7B & 13B) on instruction tuning. For CV, we apply FourierFT to fine-tune the (4) vision transformers (Base & Large) on image classification. Finally, we conduct ablation studies to analyze the effect of frequency bias, the parameter scalability, and the 4 \fParameter-Efficient Fine-Tuning with Discrete Fourier Transform Table 2. Performance of various fine-tuning methods with RoBERTa Base (RoBbase) and RoBERTa Large (RoBlarge) models on 6 datasets of the GLUE benchmark. We report the Matthew\u2019s correlation coefficient (MCC) for CoLA, Pearson correlation coefficient (PCC) for STS-B and accuracy (Acc.) for all the remaining tasks. We report the median result of 5 runs, each using different random seeds. The best results for each dataset are shown in bold. Higher is better for all metrics in 6 datasets. Model & Method # Trainable Parameters SST-2 (Acc.) MRPC (Acc.) CoLA (MCC) QNLI (Acc.) RTE (Acc.) STS-B (PCC) Avg. RoBbase(FF) 125M 94.8 90.2 63.6 92.8 78.7 91.2 85.2 RoBbase(BitFit) 0.1M 93.7 92.7 62 91.8 81.5 90.8 85.4 RoBbase(AdptD) 0.3M 94.2\u00b10.1 88.5\u00b11.1 60.8\u00b10.4 93.1\u00b10.1 71.5\u00b12.7 89.7\u00b10.3 83.0 RoBbase(AdptD) 0.9M 94.7\u00b10.3 88.4\u00b10.1 62.6\u00b10.9 93.0\u00b10.2 75.9\u00b12.2 90.3\u00b10.1 84.2 RoBbase(LoRA) 0.3M 95.1\u00b10.2 89.7\u00b10.7 63.4\u00b11.2 93.3\u00b10.3 78.4\u00b10.8 91.5\u00b10.2 85.2 RoBbase(AdaLoRA) 0.3M 94.5\u00b10.2 88.7\u00b10.5 62.0\u00b10.6 93.1\u00b10.2 81.0\u00b10.6 90.5\u00b10.2 85.0 RoBbase(DyLoRA) 0.3M 94.3\u00b10.5 89.5\u00b10.5 61.1\u00b10.3 92.2\u00b10.5 78.7\u00b10.7 91.1\u00b10.6 84.5 RoBbase(FourierFT) 0.024M 94.2\u00b10.3 90.0\u00b10.8 63.8\u00b11.6 92.2\u00b10.1 79.1\u00b10.5 90.8\u00b10.2 85.0 RoBlarge(FF) 356M 96.4 90.9 68 94.7 86.6 92.4 88.2 RoBlarge(AdptP) 3M 96.1\u00b10.3 90.2\u00b10.7 68.3\u00b11.0 94.8\u00b10.2 83.8\u00b12.9 92.1\u00b10.7 87.6 RoBlarge(AdptP) 0.8M 96.6\u00b10.2 89.7\u00b11.2 67.8\u00b12.5 94.8\u00b10.3 80.1\u00b12.9 91.9\u00b10.4 86.8 RoBlarge(AdptH) 6M 96.2\u00b10.3 88.7\u00b12.9 66.5\u00b14.4 94.7\u00b10.2 83.4\u00b11.1 91.0\u00b11.7 86.8 RoBlarge(AdptH) 0.8M 96.3\u00b10.5 87.7\u00b11.7 66.3\u00b12.0 94.7\u00b10.2 72.9\u00b12.9 91.5\u00b10.5 84.9 RoBlarge(LoRA) 0.8M 96.2\u00b10.5 90.2\u00b11.0 68.2\u00b11.9 94.8\u00b10.3 85.2\u00b11.1 92.3\u00b10.5 87.8 RoBlarge(FourierFT) 0.048M 96.0\u00b10.2 90.9\u00b10.3 67.1\u00b11.4 94.4\u00b10.4 87.4\u00b11.6 91.9\u00b10.4 88.0 expressiveness of the Fourier basis. Baselines. We compare our FourierFT method with popular parameter-efficient fine-tuning (PEFT) methods. To ensure a comprehensive and fair comparison, we prioritize replicating the setups used in previous works and reusing their reported results. Involved baselines are: \u25cfFull Fine-tuning (FF) During fine-tuning, the base model is initialized with pre-trained weights and biases, and all parameters will undergo gradient updates. \u25cfBitfit (Zaken et al., 2021) Only the bias vectors are finetuned while all other parameters are frozen. \u25cfAdapter tuning This research line was first investigated by Houlsby et al. (2019), which proposes the AdapterH method. AdapterH inserts two-layer adapters between the self-attention and the FNN modules, followed by a subsequent residual connection. We compare it with three additional variants of it. AdapterL (Lin et al., 2020) is more parameter-efficient, with adapter layers applied only after the MLP modules and subsequent to a LayerNorm. AdapterP (Pfeiffer et al., 2020) implements the adapter layers after the feed-forward layer. This design was chosen through a grid search including all settings related to the adapter\u2019s position, number, ect. AdapterD (R\u00a8 uckl\u00b4 e et al., 2020) further enhances the parameter efficiency by dropping adapter layers that are not activated. \u25cfLoRA (Hu et al., 2021) LoRA is the state-of-the-art method for PEFT. It parameterizes incremental weight updates using trainable low-rank matrices. \u25cfDyLoRA (Valipour et al., 2022) This method trains dynamic search-free LoRA models for the best rank choice. \u25cfAdaLoRA (Zhang et al., 2023) This method proposes the SVD-based fine-tuning and prunes redundant singular values with the importance-aware rank allocation. 4.1. Natural Language Understanding Models and Datasets. We evaluate our method on the GLUE benchmark (General Language Understanding Evaluation (Wang et al., 2018)), which consists of a wide range of natural language understanding (NLU) tasks, including single-sentence classification tasks, similarity and paraphrase tasks and natural language inference tasks. We finetune the pre-trained RoBERTa Base and Large foundation models (Liu et al., 2019) for evaluation. Implementation Details. For both models, FourierFT is allowed to have 1000 out of 7682 (RoBERTa Base) and 10242 (RoBERTa Large) trainable spectral coefficients in each layer, i.e., n = 1000. We randomly sample the spectral entries with no frequency bias, which is shared1 across all 24 (Base) and 48 (Large) layers. For all 6 datasets in GLUE, we tune the hyperparameters of the learning rates and the scaling values. We follow the experimental setup applied in Hu et al. (2021), which involves fine-tuning only the query and value weights in each transformer block and 1We use the value 2024 as the seed for all layers. 5 \fParameter-Efficient Fine-Tuning with Discrete Fourier Transform Table 3. Results from GPT-2 Medium and Large models on the E2E benchmark. We present the result from the final epoch. For all metrics, higher values indicate better performance. * indicates that the results are taken from prior works. Best results are shown in bold. Model Method # Trainable Parameters BLEU NIST METEOR ROUGE-L CIDEr GPT-2 Medium FT* 354.92M 68.2 8.62 46.2 71.0 2.47 AdptL* 0.37M 66.3 8.41 45.0 69.8 2.40 AdptL* 11.09M 68.9 8.71 46.1 71.3 2.47 AdptH* 11.09M 67.3\u00b1.6 8.5\u00b1.07 46.0\u00b1.2 70.7\u00b1.2 2.44\u00b1.01 LoRA 0.35M 68.9\u00b1.3 8.76\u00b1.06 46.6\u00b1.1 71.5\u00b1.1 2.53\u00b1.03 FourierFT 0.048M 69.1\u00b1.1 8.82 \u00b1.05 47.0 \u00b1.3 71.8 \u00b1.1 2.51\u00b1.02 GPT-2 Large FT* 774.03M 68.5 8.78 46.0 69.9 2.45 AdptL* 0.88M 69.1\u00b1.1 8.68\u00b1.03 46.3\u00b1.0 71.4\u00b1.2 2.49\u00b1.0 AdptL* 23.00M 68.9\u00b1.3 8.70\u00b1.04 46.1\u00b1.1 71.3\u00b1.2 2.45\u00b1.02 LoRA 0.77M 70.1\u00b1.3 8.83\u00b1.02 46.8\u00b1.2 72.0\u00b1.3 2.47\u00b1.02 FourierFT 0.072M 70.2\u00b1.2 8.90\u00b1.02 47.0\u00b1.2 71.8\u00b1.1 2.50 \u00b1.02 Table 4. The average scores on MT-Bench and Vicuna assessed by GPT-4. \u2020 indicates updating the layers other than lm head. Higher score is better. Model Method # Trainable Parameters MT-Bench Vicuna LLaMA1-7B LoRA\u2020 159.9M 5.05\u00b1.3 6.85\u00b1.4 LoRA 33.5M 4.99\u00b1.3 6.81\u00b1.3 FourierFT 0.064M 5.09\u00b1.6 6.85\u00b1.8 LLaMA1-13B LoRA\u2020 250.3M 5.28\u00b1.6 7.02\u00b1.3 LoRA 52.4M 5.21\u00b1.4 6.97\u00b1.4 FourierFT 0.08M 5.23\u00b1.3 7.14\u00b1.5 LLaMA2-7B LoRA\u2020 159.9M 5.19\u00b1.1 7.38\u00b1.3 LoRA 33.5M 5.20\u00b1.3 7.35\u00b1.6 FourierFT 0.064M 5.18\u00b1.3 7.49\u00b1.4 LLaMA2-13B LoRA\u2020 250.3M 5.78\u00b1.2 7.89\u00b1.5 LoRA 52.4M 5.80\u00b1.2 7.89\u00b1.6 FourierFT 0.08M 5.82\u00b1.3 7.92\u00b1.5 fully fine-tuning the classification head. We provide the hyperparameters in Table 9 in Appendix. Results. Results are summarized in Table 2. Following Hu et al. (2021), Zhang et al. (2023) and Valipour et al. (2022), we specify the number of trainable parameters for the finetuned layers excluding the classification head. We report the median of 5 random seed results, where the best epoch is selected for each run. In general, FourierFT achieves better or on-par performance compared with baseline methods with significantly fewer trainable parameters. Notably, FourierFT outperforms all baselines including fully fine-tuning the RoBERTa Base on CoLA and the RoBERTa Large on RTE. As mentioned in Section 3.2, the parameter count of LoRA is dependent on both the width and depth of models, resulting in a larger count growth (LoRA: 0.8M/0.3M \u22482.7; ours: 0.048M/0.024M = 2) compared to FourierFT. Nevertheless, FourierFT still performs comparably to LoRA, demonstrating the potential scalability of our method when facing even larger models. 4.2. Natural Language Generation Models and Datasets. We evaluate the performance of FourierFT on the E2E natural language generation (NLG) task (Novikova et al., 2017). We fine-tune the GPT-2 (Radford et al., 2019) Medium (354M) and Large (774M) models, which are both decoder-only and have 24 and 36 transformer blocks, respectively. The E2E benchmark contains roughly 42,000 training, 4,600 validation and 4,600 test samples from the restaurant domain. Implementation Details. We report prior results for baselines other than LoRA. For both LoRA and our method, we fine-tune the GPT-2 Medium and Large models with a linear learning rate scheduler for 5 epochs, where we tune the batch size and learning rate. We report the average results over 3 runs, where the last epoch is selected for each run. We provide the hyperparameters in Table 10 in Appendix. Results. We show the results in Table 3. We note that FourierFT can achieve the best performance on most metrics. More importantly, FourierFT only requires 13.7% and 9.4% of the parameter counts of LoRA, for the GPT-2 Medium and Large models respectively. 4.3. Instruction Tuning Models and Datasets. Instruction tuning, as described in (Ouyang et al., 2022; Wei et al., 2021; Mishra et al., 2021), refers to the process of fine-tuning a language model on a collection of paired prompts and responses. We apply LoRA and FourierFT to fine-tune the LLaMA (Touvron et al., 2023a) and LLaMA2 (Touvron et al., 2023b) families. Specifically, we consider the LLaMA-7B, LLaMA-13B, LLaMA2-7B and LLaMA2-13B as base models, which are fine-tuned on the Alpaca dataset (Taori et al., 2023). Alpaca contains 51K instruction-following demonstrations generated from text-davinci-003 (GPT-3.5) (Wang et al., 2022). For evaluation, we use the fine-tuned models to generate responses for the pre-defined questions, which are from the MT-Bench (Zheng et al., 2023) and Vicuna Eval (Chiang et al., 2023). GPT-4 takes these answers as input and evaluates them with scores within 10. Implementation Details. For LoRA, we use r = 64 and apply two configurations: (1) updating all linear layers except the language modelling head (lm head); (2) updating only the WQ and WV matrices. For FourierFT, we only adopt the latter configuration with n = 1000. To ensure the 6 \fParameter-Efficient Fine-Tuning with Discrete Fourier Transform Table 5. Fine-tuning results with ViT Base and Large models on different image classification datasets. We report the accuracy (%) after 10 epochs. Avg. represents the average accuracy of each method on all datasets. The best performance is shown in bold. Model Method # Trainable Parameters OxfordPets StanfordCars CIFAR10 DTD EuroSAT FGVC RESISC45 CIFAR100 Avg. ViT-Base LP 90.28\u00b10.43 25.76\u00b10.28 96.41\u00b10.02 69.77\u00b10.67 88.72\u00b10.13 17.44\u00b10.43 74.22\u00b10.10 84.28\u00b10.11 68.36 FF 85.8M 93.14\u00b10.40 79.78\u00b11.15 98.92\u00b10.05 77.68\u00b11.21 99.05\u00b10.09 54.84\u00b11.23 96.13\u00b10.13 92.38\u00b10.13 86.49 LoRA 581K 93.19\u00b10.36 45.38\u00b10.41 98.78\u00b10.05 74.95\u00b10.40 98.44\u00b10.15 25.16\u00b10.16 92.70\u00b10.18 92.02\u00b10.12 77.58 FourierFT 72K 93.21\u00b10.26 46.11\u00b10.24 98.58\u00b10.07 75.09\u00b10.37 98.29\u00b10.04 27.51\u00b10.64 91.97\u00b10.31 91.20\u00b10.14 77.75 FourierFT 239K 93.05\u00b10.34 56.36\u00b10.66 98.69\u00b10.08 77.30\u00b10.61 98.78\u00b10.11 32.44\u00b10.99 94.26\u00b10.20 91.45\u00b10.18 80.29 ViT-Large LP 91.11\u00b10.30 37.91\u00b10.27 97.78\u00b10.04 73.33\u00b10.26 92.64\u00b10.08 24.62\u00b10.24 82.02\u00b10.11 84.28\u00b10.11 72.96 FF 303.3M 94.43\u00b10.56 88.90\u00b10.26 99.15\u00b10.05 81.79\u00b11.01 99.04\u00b10.08 68.25\u00b11.63 96.43\u00b10.07 93.58\u00b10.19 90.20 LoRA 1.57M 94.82\u00b10.09 73.25\u00b10.36 99.13\u00b10.03 81.79\u00b10.45 98.63\u00b10.07 42.32\u00b10.98 94.71\u00b10.25 94.87\u00b10.10 84.94 FourierFT 144K 94.46\u00b10.28 69.56\u00b10.30 99.10\u00b10.04 80.83\u00b10.43 98.65\u00b10.09 39.92\u00b10.68 93.86\u00b10.14 93.31\u00b10.09 83.71 FourierFT 480K 94.84\u00b10.05 79.14\u00b10.67 99.08\u00b10.01 81.88\u00b10.50 98.66\u00b10.03 51.28\u00b10.68 95.20\u00b10.07 93.37\u00b10.11 86.68 feasibility of training on a single GPU, we deploy the quantization method in Dettmers et al. (2023) for fine-tuning. We train with both methods for only one epoch, and report the average scores of all answers. We provide the hyperparameter setup in Table 11 in the Appendix. Results. The results are shown in Table 4. We find that the expressive power of the 13B model is much stronger than that of the 7B model, regardless of which fine-tuning method is used. Moreover, FourierFT closely matches or slightly exceeds LoRA\u2019s performance with less than 0.2% of its parameters. We provide practical examples containing questions, answers and reviews in the Appendix D. 4.4. Image Classification Models and Datasets. We evaluate our method on the image classification task. We adopt the Base and Large versions of the popular CV foundation model, Vision Transformer (ViT) (Dosovitskiy et al., 2020). The ViTs are pretrained on the ImageNet-21K dataset (Ridnik et al., 2021). The datasets for fine-tuning include OxfordPets (372), CIFAR10 (10), DTD (47), EuroSAT (10) and RESISC45 (45) with small label spaces, as well as StanfordCars (196), FGVC (100) and CIFAR100 (100) with large label spaces. Detailed information is provided in Table 8 in the Appendix. Implementation Details. We include three baselines for evaluation: Full Fine-tuning (FF), Linear Probing (LP, finetuning the classification head only), and LoRA. For both LoRA and our method, only the query and value matrices of ViT are updated. We use r = 16 for LoRA and n = {3000,10000} for FourierFT. We tune the learning rates and weight decay for all methods, and set the maximum training epoch to 10. We provide the hyperparameters in Table 12 in Appendix. 2Numbers in parentheses indicate class counts for each dataset. Results. Table 5 summarizes the results for 8 image classification datasets with the ViT Base and Large models. Both LoRA and FourierFT methods significantly outperform the Linear Probing, demonstrating their effectiveness in the CV domain. Our method obtains matched performance using 12.4% and 9.2% of LoRA\u2019s parameter count, with ViT Base and Large models, respectively. Notably, when we increase the parameter count of FourierFT to 41.1% (ViT Base) and 30.6% (ViT Large) of LoRA\u2019s, it can outperform LoRA by 3.5% and 2.0% respectively. Moreover, our method can even (slightly) outperform the Full Fine-tuning method on OxfordPets and DTD with the ViT Large model. 4.5. Study Effect of Frequency Bias. We examine how the performance is affected by the frequency bias, i.e., the central frequency fc in Eq. 5. We directly apply the optimal hyperparameters searched in Table 2 and fine-tune the RoBERTa Base on the MRPC, STS-B, CoLA and RTE datasets. From Figure 5, we note that the fine-tuning performance of FourierFT without any frequency bias can surpass most cases that are restricted by the central frequency bias. This indicates the universality of our method. Surprisingly, we find that it is always possible to obtain results better than \u201cNo bias\u201d by traversing the fc values. Since this traversal is not efficient, we do not conduct further exploration in this paper. However, we believe that making fc trainable will be a promising new direction for improving FourierFT. Parameter Scalability. We explore the relationship between the number of trainable parameters and the performance of LoRA and our method. We use the set of ranks r = {1,2,4,6,8,15} for LoRA and n = {50,100,200,1000,6144,12288} for FourierFT on 6 tasks of the GLUE benchmark. For both LoRA and ours, the learning rate, and scaling hyperparameters are tuned. For fairness, we ensure that the number of trials for hyperparam7 \fParameter-Efficient Fine-Tuning with Discrete Fourier Transform 4 6 8 10 ln # Trainable Parameters 88 89 90 Accuracy MRPC LoRA FourierFT 4 6 8 10 ln # Trainable Parameters 58 60 62 64 Matthew s Corr. CoLA 4 6 8 10 ln # Trainable Parameters 78 80 Accuracy RTE 4 6 8 10 ln # Trainable Parameters 90.0 90.5 91.0 91.5 Pearson Corr. STS-B 4 6 8 10 ln # Trainable Parameters 94.0 94.5 95.0 Accuracy SST-2 4 6 8 10 ln # Trainable Parameters 90 91 92 Accuracy QQP Figure 4. Performance on the GLUE benchmark with RoBERTa Base vs. number of trainable parameters (each layer) of LoRA and ours. For all 6 datasets, we apply the setting of r = {1, 2, 4, 6, 8, 15} for LoRA and n = {50, 100, 200, 1000, 6144, 12288}. 0 100 200 300 400 500 fc 76 77 78 79 Acc. RTE No bias 0 100 200 300 400 500 fc 89.2 89.4 89.6 89.8 90.0 Acc. MRPC No bias 0 100 200 300 400 500 fc 90.5 90.6 90.7 90.8 PCC. STS-B No bias 0 100 200 300 400 500 fc 62.0 62.5 63.0 63.5 MCC. CoLA No bias Figure 5. Results on 4 datasets in GLUE with different fc values. eter search is 30 for both methods. As shown in Figure 4, our method outperforms LoRA on all 6 datasets. In detail, our method is significantly better than LoRA with the same parameter count, i.e., {r = 4,n = 6144} & {r = 8,n = 12288}. Moreover, we observe that a larger number of parameters does not always bring performance gains for LoRA. On the contrary, the increase of n can consistently improve the accuracy of FourierFT. On most tasks, FourierFT with n = 50 can achieve comparable or even better (MRPC, CoLA, RTE) performance than LoRA with r = 1. In this case, the parameter count in LoRA is about 31 \u00d7 that of ours. Basis Expressiveness. The inverse discrete Fourier transform (IDFT) in Eq. 3 is equivalent to the matrix multiplication (Lu et al., 2021): S = BfFB\u22ba f, where B is the transformation matrix of IDFT that contains the Fourier basis. To evaluate its expressivity, we replace the Fourier basis with random and orthogonal basis, respectively. Specifically, for F \u2208Rd1\u00d7d2, we initialize random basis B1 r \u2208Rd1\u00d7d1 and B2 r \u2208Rd2\u00d7d2 with the normal Gaussian distribution. Then Eq. 3 becomes S = B1 rFB2 r. A similar way is used for the orthogonal basis. We compare FourierFT with the random basis (R-B) and orthogonal basis (O-B) on the GLUE benchmark. Table 6 shows the results. We note that the Fourier basis used in our method outperforms the random and orthogonal basis. In addition, the expressive power of the orthogonal basis is much stronger than that of the random basis. The stronger expressive power of the Fourier basis compared to the general orthogonal basis may be attributed to its effective capture of the spectral information of \u2206W. Table 6. Results with three types of basis. Model RTE CoLA Ours R-B O-B Ours R-B O-B Base 79.1 72.7(\u21938.1%) 75.6(\u21934.4%) 63.8 58.7(\u21938.0%) 60.0(\u21936.0%) Large 87.4 81.8(\u21936.4%) 83.6(\u21934.3%) 67.1 64.8(\u21933.4%) 66.1(\u21931.5%) 5."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2405.02299v2",
+                    "title": "Deep Reinforcement Learning for Modelling Protein Complexes",
+                    "abstract": "AlphaFold can be used for both single-chain and multi-chain protein structure\nprediction, while the latter becomes extremely challenging as the number of\nchains increases. In this work, by taking each chain as a node and assembly\nactions as edges, we show that an acyclic undirected connected graph can be\nused to predict the structure of multi-chain protein complexes (a.k.a., protein\ncomplex modelling, PCM). However, there are still two challenges: 1) The huge\ncombinatorial optimization space of $N^{N-2}$ ($N$ is the number of chains) for\nthe PCM problem can easily lead to high computational cost. 2) The scales of\nprotein complexes exhibit distribution shift due to variance in chain numbers,\nwhich calls for the generalization in modelling complexes of various scales. To\naddress these challenges, we propose GAPN, a Generative Adversarial Policy\nNetwork powered by domain-specific rewards and adversarial loss through policy\ngradient for automatic PCM prediction. Specifically, GAPN learns to efficiently\nsearch through the immense assembly space and optimize the direct docking\nreward through policy gradient. Importantly, we design an adversarial reward\nfunction to enhance the receptive field of our model. In this way, GAPN will\nsimultaneously focus on a specific batch of complexes and the global assembly\nrules learned from complexes with varied chain numbers. Empirically, we have\nachieved both significant accuracy (measured by RMSD and TM-Score) and\nefficiency improvements compared to leading PCM softwares.",
+                    "authors": "Ziqi Gao, Tao Feng, Jiaxuan You, Chenyi Zi, Yan Zhou, Chen Zhang, Jia Li",
+                    "published": "2024-03-11",
+                    "updated": "2024-05-07",
+                    "primary_cat": "cs.CE",
+                    "cats": [
+                        "cs.CE",
+                        "cs.LG"
+                    ],
+                    "main_content": "INTRODUCTION \u2026 Dimer Structures C-A + A-D C-A-D + A-B C-A-D B Multiple  Chains Step 1 Complex Structure Prediction Assembly Step 2 A B C D  AUC Graph Figure 1: The assembly process applied for the PCM problem. Each incoming chain has the opportunity to dock onto any already docked chain. Predicting the 3D structures of large protein complexes (Jumper et al., 2021; Hickman & Levy, 1988; Benesch et al., 2003) is crucial for understanding their biological functions and mechanisms of action. This knowledge is essential for studying various cellular processes, such as signal transduction (Ditzel et al., 1998), gene regulation (Will & L\u00a8 uhrmann, 2011), and enzymatic reactions (Gao et al., 2006). The groundbreaking AlphaFold-Multimer (Evans et al., 2021) has presented great potential in directly predicting 3D complex structures from their amino acid sequences. However, it faces difficulties in maintaining high accuracy when dealing with complexes with a larger number (> 9) of chains (Bryant et al., 2022a; Burke et al., 2023; Bryant et al., 2022b). Recently, a novel setting for protein complex modelling (PCM) has emerged and achieved significant success. These methods (Bryant et al., 2022b; Aderinwale et al., 2022; Esquivel-Rodr\u00b4 \u0131guez \u2217Equal contribution. \u2020Correspondence to: Jia Li (jialee@ust.hk). 1 arXiv:2405.02299v2  [cs.CE]  7 May 2024 \fPublished as a conference paper at ICLR 2024 et al., 2012; Burke et al., 2023) suggest that since numerous advanced protein-protein docking techniques (Ganea et al., 2021; Wang et al., 2023) are available, why not utilize them to sequentially assemble protein chains into complexes? They have demonstrated the ability to perform PCM on very large complexes (with up to 30 chains). However, the combinatorial optimization space (COS) of the PCM problem greatly limits the inference efficiency and effectiveness of these methods. As shown in Fig. 1, real-world data indicates that in a N\u2212chain complex, the interaction of each chain may involve more than two chains. Therefore, the assembly can be determined by an acyclic, undirected, and connected (AUC) graph, which leads to the COS of N N\u22122 (Shor, 1995). Designing a proper deep learning (DL) framework holds great promise in handling the vast COS. However, directly learning from available complexes can be difficult since the data scarcity will increase as the chain number grows. For example, when the chain number exceeds 10, the average number of available complexes from the PDB database (Berman et al., 2000) is less than 100. Due to the inherent variability in the chain numbers of complexes, DL models are likely to face challenges in learning adequate knowledge for assembling large complexes (Aderinwale et al., 2022). Based on the above statements, we summarize two main challenges of the PCM problem as follows: 1) How to efficiently explore the vast combinatorial optimization space to find the optimal assembling action? 2) How to ensure that the DL model can effectively generalize to complex data with limited availability? In this paper, we propose GAPN, a generative adversarial policy network powered by domainspecific rewards and adversarial loss through policy gradient for automatic PCM prediction. Specifically, to address the first challenge, we design GAPN as a deep reinforcement learning (DRL) agent that operates within a complex assembling environment. The DRL agent can actively explore the vast combinatorial optimization space (COS) for assembling large complexes beyond samples through trial-and-error interaction with the environment (Zhang et al., 2019; Samsuden et al., 2019). Moreover, it incorporates a balance between exploration (trying new actions) and exploitation (sticking with known good actions) to ensure that the agent will not get stuck in local optima and continue to explore other potentially better solutions (Ladosz et al., 2022; Gupta et al., 2018a). The above two aspects make the GAPN agent able to efficiently explore the vast combinatorial optimization space to find the optimal assembling action. In response to the second challenge, we introduce an adversarial reward function to guide the learning process and enhance the receptive field of our model. The adversarial reward function incorporates the global assembly rules learned from complexes with varied chain numbers to enhance the generalization of GAPN. We empirically validate that the proposed GAPN model has high efficiency in exploring the vast COS in PCM problem, and can effectively generalize to complex data with limited availability. In summary, we make the following contributions: \u2022 We propose GAPN, a fast, policy-based method for modeling large protein complexes, where GAPN functions as a DRL agent to efficiently explore correct assembly actions. \u2022 We design an adversarial reward function that offers a global view of assembly knowledge, ensuring that GAPN can generalize over limited complex data. \u2022 Quantitative results show that GAPN achieves superior structure prediction accuracy compared to advanced baseline methods, with a speed-up of 600\u00d7. Moreover, we contribute a dataset containing non-redundant complexes. 2 RELATED WORK Protein Complex Modelling. The task of protein complex modelling (PCM) aims to predict the complex 3D structure given the information of all its chains. As shown in Table 1, existing PCM methods can be divided into two categories, namely one-time forming (OTF) and one-by-one assembling (OBOA). OTF-based methods can typically produce PCM results directly from the sequence or chain-level 3D structures. Specifically, the modified AlphaFold system, referred to as AlphaFold-Multimer (Evans et al., 2021), can handle up to 9-chain complexes, but it is very timeconsuming. Moreover, Ji et al. (2023) proposed SyNDock, the learnable group synchronization approach for forming complexes in a globally consistent manner. SyNDock mainly claims significant speed-up, but the performance and the maximum chain number it can handle are both limited. Another PCM setting mainly involves two optimization aspects: the dimeric structures and the docking 2 \fPublished as a conference paper at ICLR 2024 path. Optimization approaches such as reinforcement learning (Aderinwale et al., 2022) and genetic algorithm (Esquivel-Rodr\u00b4 \u0131guez et al., 2012) are proposed for refining the dimeric structures. These methods are typically with low efficiency, and their max chain No. is also not satisfactory. MoLPC (Bryant et al., 2022b) assumes that the dimeric structures are given, thus only predicting the correct docking path. In this way, it is the first to demonstrate that the structures of complexes with up to 30 chains can be accurately predicted. However, its application of a tree search-based non-deep learning technique often leads to complete failure due to the lack of generalization. Methods Assembly Opt. Dimer Opt. Path Max Chain No. Speed-Up CombDock Inbar et al. (2005) \u2714 \u2714 \u2714 6 1.07 Multi-LZerD Esquivel-Rodr\u00b4 \u0131guez et al. (2012) \u2714 \u2714 \u2714 6 1.33 RL-MLZerD Aderinwale et al. (2022) \u2714 \u2714 \u2717 5 1 ESMFold Lin et al. (2023) \u2717 \u2717 \u2717 4 213.65 AlphaFold-Multimer Evans et al. (2021) \u2717 \u2717 \u2717 9 4.53 MoLPC Bryant et al. (2022b) \u2714 \u2717 \u2714 30 3.02 GAPN \u2714 \u2717 \u2714 60 1934.27 Table 1: Technical comparison between PCM methods. Assembly: PCM in the assembly fashion; Opt. dimer: Dimer structures will be optimized; Opt. Path: Docking path will be optimized; Max Chain No.: The maximum chain number that the method can handle. Speed-up: Efficiency speed-up compared to the RL-MLZerD baseline. Reinforcement Learning. Over the past few years, RL has emerged as a compelling approach for various optimization tasks. The unique ability of RL agents to learn optimal policies from interaction with an environment enables them to collect and utilize more data, which enhances their ability to find excellent solutions to optimization problems with large search space. Several studies (Vinyals et al., 2015; Bello et al., 2016; Nazari et al., 2018) utilize RL and attention architecture to solve classic NP-hard problems with large and complex search space, such as TSP (Vinyals et al., 2015) and VRP (Bello et al., 2016; Nazari et al., 2018) problems. Khalil et al. (2017) and Fan et al. (2020) propose to introduce the graph architecture into the RL framework, so that RL can be used to solve large-scale graph optimization problems, such as finding the most influential nodes on the graph and optimizing the properties of the graph network. However, when applying RL to our problem, it still faces the generalization problem in scenarios with varied chain numbers. Therefore, in our paper, we design an adversarial reward function to guide the learning of the RL agent, which incorporates prior knowledge of multiple optimization scenarios. Adversarial Training Adversarial training has been widely used to learn the underlying data distribution and derive a generator and a discriminator. The generator can generate synthetic data that is similar to the real data distribution, while the discriminator outputs the similarity between the input and real data. Adversarial training is widely used in many aspects, such as text generation (Zhang et al., 2017), imitation of human behavior (Gupta et al., 2018b), protein generation (Gupta & Zou, 2019). Adversarial training also enhances RL\u2019s ability to solve complex optimization problems by incorporating prior knowledge specified by a good oracle (Ho et al., 2016; Liu et al., 2018). Inspired by existing research, in this paper, we design an adversarial reward function to enhance the generalization ability of the RL agent based on the framework of adversarial training, which incorporates the global assembly rules learned from complexes with varied chain numbers. 3 PROPOSED APPROACH 3.1 PROBLEM DEFINITION We are given a set of N chains to form a protein complex. The i \u2212th chain consists of ni residues, which is characterized by its amino acid sequence si and 3D structure in its undocked state Xi \u2208 R3\u00d7ni. We focus on the one-by-one assembly setting, where the dimeric 3D structures of all possible pairs of chains are provided. We define the assembly process as follows. Formally, we have a set of assembly actions A = {(A1 k, A2 k)}1\u2264k\u2264N\u22121, with each element representing a pair of chain indexes. Then, a transformation function set T = {(Rk, tk)}1\u2264k\u2264N\u22121 is obtained by Rk \u02dc XA1 k + tk = X\u2032 A1 k, where Rk, tk 3 \fPublished as a conference paper at ICLR 2024 ... Global\u00a0Embedding Docked\u00a0Embedding Undocked\u00a0Embedding Query Key Mean Pooling &Sigmoid Generative Adversarial Policy Network Value Network Adversarial Reward Function \u00a0Multi-chain Assembly Process Assembly Action State\u00a0 Domain-specific reward Adversarial reward Assembled Graph State\u00a0 Undocked Chains Docked Chains MLP Softmax Avg. Action GCN Assembled Graph a b c d Figure 2: Overview of the proposed framework. It contains a Generative Adversarial Policy Network (GAPN) (b), which converts the input state st into assembly action at for the multi-chain assembly process (d). The GAPN is trained with the help of value network (a), which estimates the values come from two aspects: domain-specific rewards obtained through multi-chain assembly process (d) and adversarial rewards from adversarial reward function (c). represent the rotation and translation transformation for the k \u2212th action, respectively. \u02dc XA1 k denotes structure of chain A1 k extracted from the dimeric structure of the pair of (A1 k, A2 k) and X\u2032 A1 k denotes the structure of the docked chain A1 k. For the k \u2212th action, we dock the chain A2 k onto the docked chain A1 k, such that X\u2032 A2 k = RkXA2 k + tk. The task is to predict the assembly action set such that {RX\u2032 i + t}1\u2264i\u2264N = {X\u2217 i }1\u2264i\u2264N, \u2203R \u2208 SO(3), t \u2208R3, by taking as inputs the amino acid sequences of all chains{si}1\u2264i\u2264N of an input complex, where X\u2217 i denotes the ground truth of the i \u2212th chain. In short, our goal is to predict the correct assembly action A such that the assembled complex structure matches the ground truth complex structure. 3.2 SYSTEM OVERVIEW To model the large protein complexes, we propose GAPN powered by domain-specific rewards and adversarial rewards through policy gradient. The structure overview is shown in Figure 2. The system mainly consists of four modules: \u2022 Generative Adversarial Policy Network. It converts the input state to corresponding assembly action and interacts with multi-chain assembly process. \u2022 Adversarial Reward Function. It incorporates the global assembly rules learned from complexes with varied chain numbers to enhance the generalization of GAPN, which converts the assembled graph into adversarial reward through Graph Convolutional Network (GCN). \u2022 Value Network. It aims to estimate the value of the state and helps train the GAPN. \u2022 Multi-chain Assembly Process. It describes the process of multi-chain assembly, state transfer, and assembly action effects, which has been discussed in Sec 3.1. In the following sections, we will introduce these modules in detail. 3.3 ASSEMBLY OF PROTEIN COMPLEXES AS MARKOV DECISION PROCESS We transform the assembly of protein complexes problem into a sequential decision-making problem, with the goal of predicting the correct assembly action that will bring the assembled complex structure to match the ground truth. Thus, we propose to formulate the problem using Markov Decision Process (MDP) (Puterman, 1990) in an RL setting, which has five parts: 4 \fPublished as a conference paper at ICLR 2024 \u2022 Agent: We consider the complex M = {Mi}1\u2264i\u2264N as the RL agent, who observes a set of N chains and ni residues of each chain Mi. \u2022 State: Since the agent is the complex M, we define state at time interval t as st = (ct, ut, Gr), where ct denotes docked chain feature embeddings, ut denotes the undocked chain embeddings and Gr denotes all chain embeddings in the complex. The chain embedding ei of Mi is calculated through a protein pre-trained model, which will be introduced in Sec 3.4. \u2022 Action: The action at at time interval t is to assemble two chains, which can be denoted by (A1 k, A2 k)1\u2264k\u2264N\u22121 with two elements representing a pair of chains. \u2022 Transition: This component describes the process that the current state st will transit to the next state st+1 after an action at is taken. Specifically, in our state setting, the stack of all protein embeddings Gr will remain unchanged. The stack of assembled protein feature embeddings ct+1 = ct\u222aej and the stack of unassembled protein embeddings ut+1 = ut\\ej will change due to newly docked chains, where ej denotes the embedding of the newly docked chain. With the implementation of assembly action at, the dimeric structure of chains will be updated, which is achieved with the transformation set T defined in Sec 3.1. \u2022 Reward: This reward function r is to measure how similar the assembled complex structure is compared with the real ones. Here, we define the rewards as a sum over domainspecific rewards rt and adversarial rewards rad. The domain-specific rewards are introduced to measure the similarity between the assembled complex structure and the corresponding ground truth. Specifically, we utilize RMSD to measure this similarity, which is inverted when used as a reward. Since the smaller the RMSD, the higher the similarity, we take negative RMSD values as domain-specific rewards. In order to enhance the generalization of the agent to limited data, we design an adversarial reward function incorporating the global assembly rules learned from complexes with varied chain numbers (see more discussion in Sec 3.5). 3.4 GENERATIVE ADVERSARIAL POLICY NETWORK Having modeled the assembly of protein complexes as Markov Decision Process, we outline the architecture of GAPN, a policy network learned by the RL agent to act in the multi-chain assembly process. GAPN takes the state st input and outputs the action at, which predicts an assembly action, as described in Section 3.3. Chain Embedding Calculation. Through ESM (Lin et al., 2023) pre-training models, we obtain residue-level embeddings by inputting the entire sequence of each chain. ESM\u2019s pre-training captures the reflection of protein 3D structure within sequence patterns. To obtain the chain embedding e, we take the average of residue embeddings. It\u2019s worth noting that we refrain from fine-tuning the residue embeddings. This approach allows us to leverage the power of ESM\u2019s pre-training while maintaining the stability of the residue-level representations. Assembly Action Prediction. The assembly action at at time step t is to decide which pair of proteins assembles together. The traditional RL approach is to perform one-hot encoding on all possible actions. However, these methods are difficult to apply to our problem: (1) In the protein assembly process, the number of possible protein pairs is constantly changing, which makes these methods hard to handle; (2) Since different protein complexes may have varied numbers of proteins, models trained based on these methods are difficult to generalize to different protein complexes. Inspired by the design of attention mechanism for supporting model\u2019s inputs with variable lengths (Vaswani et al., 2017), we first introduce a function fc to convert ct, Gr into dense vector gt, and then regard the gt and ut as key and query to apply inner product for them and obtain the probability of deciding which pair of proteins assemble together, gt = fc(ct, Gr), (1) \u03c0(st) = SOFTMAX(fa(gT t ut)), at \u223c\u03c0(st), (2) where fc and fa denote two MultiLayer Perceptrons (MLP) and \u03c0 denotes the GAPN. 5 \fPublished as a conference paper at ICLR 2024 3.5 ADVERSARIAL REWARD FUNCTION The vast combinatorial optimization space of large protein complexes makes it hard for RL to search for optimal assembly action sets to bring the assembled complex structure to match the ground truth. Moreover, the limited availability of these complex data further aggravates the difficulty of searching. In order to deal with this dilemma, the traditional approach is to mix complex data of different chains and utilize a unified RL policy network to learn (Beck et al., 2023). However, the huge differences in the features of complex data of different chains lead to poor training stability and performance (Nagabandi et al., 2018). Adversarial training can improve RL\u2019s ability to deal with complex problems by learning expert solutions as prior knowledge (Ho et al., 2016; Liu et al., 2018). Inspired by this, we utilize the correct assembly action sets of complex data in different chains as expert solutions, and employ the Generative Adversarial Network (GAN) framework (Goodfellow et al., 2020) to design the adversarial rewards R(\u03c0\u03b8, D\u03d5) for RL learning: min \u03b8 max \u03d5 R(\u03c0\u03b8, D\u03d5) = Ex\u223cpdata(x)[log D\u03d5(x)] + Ex\u223c\u03c0\u03b8[log(1 \u2212D\u03d5(x)))], (3) where \u03c0\u03b8 is the policy network parameterized by \u03b8, D\u03d5 is the discriminator network parameterized by \u03d5, x represents an assembled graph whose nodes represent chains and edges depict the connection relationship among the chains determined by the assembly action set. pdata is the underlying data distribution of the ground truth assembly action sets of complex data with different chains. However, from equation 3, we can observe that only D\u03d5 can be trained through stochastic gradient descent and x as a graph is not differentiable in terms of parameters \u03d5. Therefore, we utilize \u2212R(\u03c0\u03b8, D\u03d5) as an auxiliary term with domain-specific rewards, which measures the similarity between the complex structure assembled from the generated assembly action set and the real ones. The ability of Graph Neural Networks (Xu et al., 2018; Tang et al., 2022) to model the structured data enables it to perform well in many tasks (Gao et al., 2023a; Tang et al., 2023; Li et al., 2019b; Gao et al., 2023b). We employ the Graph Convolutional Networks (GCN) (Kipf & Welling, 2016; Li et al., 2019a; Cheng et al., 2023) to construct the discriminator network. Specifically, the nodes of the input complex graph are chain embeddings E = ei, and the adjacency matrix U of the input graph depicts the docked relationship among the chains determined by assembly action set. If the assembly action set contains at = (A1 K, A2 k), then U[A1 K, A2 k] = U[A2 K, A1 k] = 1. We apply message passing with initial node embeddings E = ei and adjacency matrix U to compute node embeddings for a total of L layers: Hl+1 = AGG(ReLU(D\u22121 2 \u02dc UD\u22121 2 HlW l)), (4) where Hl is the node embedding of the l \u2212th layer, D is degree matrix of U, \u02dc U = U + I and W l is a trainable weight matrix of the l \u2212th layer. AGG(\u00b7) here is employed to denote an aggregation function that could be one of {MEAN, MAX, SUM, CONCAT} (Sun et al., 2023). Based on the output HL of the last layer of GCN, the discriminator network applies an MLP layer, a SIGMOID function, and a MEAN Pooling function to get the final output D\u03d5(x): D\u03d5(x) = SIGMOID(MEAN(fd(HL))), (5) where fd denotes the MLP layer and the MEAN Pooling function is to average embedding of the chain to obtain the representation of the entire graph. 3.6 MODEL TRAINING For the optimization of GAPN, we adopt Proximal Policy Optimization (PPO) (Schulman et al., 2017), which is one of the state-of-the-art policy gradient methods and has been widely used for optimizing policy networks. We show the objective function of PPO as follows: max (Jclip(\u03b8)) = E\u03c0\u03b8old [min (rt(\u03b8)A\u03c0\u03b8old(s, a), clip(rt(\u03b8), 1 \u2212\u03f5, 1 + \u03f5)A\u03c0\u03b8old(s, a))] , (6) rt(\u03b8) = \u03c0\u03b8(a|s) \u03c0\u03b8old(a|s), (7) where rt(\u03b8) represents the ratio of the probability \u03c0\u03b8(a|s) of an action under the current policy to the probability \u03c0\u03b8old(a|s) of the same action under the old policy and is clipped to the range of 6 \fPublished as a conference paper at ICLR 2024 Algorithm 1 GAPN Input: ground truth assembly graphs Gr. 1: Initialize \u03c0\u03b8, D\u03d5 with random weights for policy network and discriminator; 2: for i=0,1,2... do 3: Generate a batch of assembly graphs Gf i \u223c\u03c0\u03b8; 4: for k = 0, 1, 2... do 5: Calculate a batch of rewards based on domain-specific reward function rt and adversarial function rad for each state-action pair in sequences Gf i ; 6: Sample generated assembly graphs Gf ki from Gf i with batch size B; 7: Sample ground truth Gr k from Gr with batch size B; 8: Update D\u03d5 based on equation (3) through stochastic gradient descent with the positive samples Gr k and negative samples Gf ki. 9: end for 10: Update \u03c0\u03b8 by maximizing equation (6) via the PPO method; 11: end for [1 \u2212\u03f5, 1 + \u03f5] that makes the training of policy network more stable (Schulman et al., 2017), A\u03c0\u03b8old denotes the estimated advantage function which involves a learned value function V (\u00b7) to reduce the variance of estimation. Specially, V (\u00b7) is designed as an MLP that maps the state st into a scalar. We summarize details of the training process of GAPN in Algorithm 1. From the algorithm, we can first observe that a batch of generated assembly graphs and ground truth are sampled to train the discriminator (lines 6-8). The generated assembly graphs are regarded as negative samples and ground truths are regarded as positive samples to train discriminator via an Adam Optimizer (Reddi et al., 2019) (line 8). Then, we calculate a batch of rewards for the generated assembly graphs (line 5). Finally, we train the policy network by maximizing the expectation of reward via PPO algorithm (line 10). 4 EXPERIMENTS 4.1 EXPERIMENTAL SETUP Dataset. Following MoLPC, we conduct experiments on protein complexes with the chain number 3 \u2264N \u226430. We download all available complexes of their \u2018first biological assembly\u2019 version from the PDB database (Berman et al., 2000). Subsequently, we perform data filtering and sequence clustering using CD-HIT (Huang et al., 2010). We split the data based on the clusters to ensure that the sequence similarity between samples in the training and test sets will not exceed 50%. Finally, we obtained 7,063 PDBs for the training set and validation set, and 180 for the test set. A more detailed description of the data processing and filtering process can be found in Appendix B. We apply two types of dimer structures, the ground-truth dimers (GT Dimer) and dimers predicted by AlphaFold-Multimer (AFM Dimer). For GT Dimer, we directly extract the dimer structures of all pairs of chains that have actual physical contacts. For those without physical contacts, we apply their dimers predicted with ESMFold (Lin et al., 2023) (due to its inference efficiency). Baselines. We compare GAPN with recent advanced PCM methods. For non-assembly methods, we include (1) AlphaFold-Multimer (Evans et al., 2021) (AF-Multimer, an end-to-end deep learning model that takes only protein sequence as input for PCM); (2) ESMFold (Lin et al., 2023)(an endto-end deep learning-based protein structure predictor without multiple sequence alignment). For assembly-based methods, we include (3) Multi-LZerD (Esquivel-Rodr\u00b4 \u0131guez et al., 2012) (a PCM predictor with genetic algorithm and a structure refinement procedure); (4) RL-MLZerD (Aderinwale et al., 2022) (a reinforcement learning model for exploring potential dimer structures and docking path); (5) MoLPC (Bryant et al., 2022b) (a non-deep learning method with Monte Carlo tree search and plDDT information for guidance). According to the maximum chain numbers that the baselines can handle, we set their available ranges of chain number Table 2 for testing (e.g., Multi-LZerD: 3 \u2264N \u226410). We test all baselines 7 \fPublished as a conference paper at ICLR 2024 Table 2: PCM prediction results. The best performance for each metric is bold and the second best is underlined. \u2018-\u2019 suggests that the method is not applicable to this range of chain number. Methods Chain numbers N 3 4 5 6-10 11-15 16-20 21-25 26-30 TM-Score (mean) / RMSD (mean) Multi-LZerD 0.54 / 19.17 0.49 / 21.58 0.40 / 33.05 0.24 / 56.73 RL-MLZerD 0.62 / 16.56 0.47 / 19.09 0.49 / 28.73 0.31 / 49.60 AF-Multimer 0.79 / 7.87 0.70 / 14.19 0.67 / 22.25 0.64 / 35.05 0.39 / 49.23 ESMFold 0.73 / 9.75 0.52 / 18.06 0.44 / 31.96 0.36 / 46.12 MoLPC 0.89 / 5.33 0.71 / 13.49 0.66 / 21.67 0.60 / 36.46 0.45 / 43.00 0.36 / 54.75 0.34 / 60.58 0.29 / 66.60 GAPN 0.95 / 1.16 0.81 / 8.52 0.79 / 16.96 0.65 / 32.37 0.62 / 37.58 0.48 / 50.05 0.38 / 55.76 0.36 / 59.13 (a) (b) (c) Figure 3: PCM performance analysis. (a). TM-Score and (b). RMSD distributions of all baselines on multimers of 3 \u2264N \u226410. (c). TM-Score distribution of MoLPC and GAPN on large-scale multimers of 11 \u2264N \u226430, with the median TM-Score marked by \u2715. using 2 GeForce RTX 4090 GPUs. For each sample within the available range, if the testing time exceeds 2 hours or the memory limit is exceeded, its result will be replaced by the optimal value of all baselines for that sample. Metrics. We evaluate the PCM prediction methods by the following metrics: (1) RMSD: the rootmean-square deviation of the Cartesian coordinates, measures the distance of the ground-truth and predicted multimer structure at the residue level; (2) TM-Score: the proportion of residues that are similar between the ground-truth and predicted multimer structures with values ranging from 0 to 1. 4.2 PROTEIN COMPLEX MODELLING Performance. Model performance are summarized in Table 2 and Fig. 3. Our GAPN method achieves state-of-the-art in all ranges of chain numbers. Compared to the second-best baseline with small-scale multimers (3 \u2264N \u226410), we achieve improvements of 8.7%, 9.6%, 11.0%, and 12.7% on metrics of mean TM-Score, median TM-Score, mean RMSD and median RMSD, respectively. Moreover, our methods consistently outperform the most advanced assembly-based MoLPC method, without resource-intensive exhaustive search or expensive plDDT information. Fig. 3c shows the comparison of MoLPC and our method for handling large-scale multimers. We can find that although relatively imprecise, AFM dimers can still achieve accurate assembly in practice. This further reinforces the practical significance of the assembly-based setting that we follow. Importantly, our method outperforms MoLPC on large-scale multimers in both settings of AFM and GT dimers. For instance, under the setting of GT dimer, we achieve an improvement of about 31% and 53% in terms of the mean and median TM-Score metrics for multimers of 11 \u2264N \u226430. 8 \fPublished as a conference paper at ICLR 2024 Visualization. In Fig. 6, we show two examples with a small (N = 5) and a large (N = 17) chain number, respectively. These two complexes are both unknown samples, meaning that none of their chains have more than 10% similarity with any chain present in the training set. Table 3: Inference time. Methods T/N(sec) Multi-LZerD 1235.20 RL-MLZerD 1644.78 AFM 357.30 ESMFold 7.73 MoLPC 540 GAPN 0.85 Efficiency. We evaluate the two types of efficiency of our GAPN method, computational and exploration efficiency, from both the training and inference perspectives. For inference, since the running time of all methods increases with the number of chains, we compare the ratio of running time to chain number (T/N) among different methods. As shown in Table 3, the inference efficiency of our GAPN is about 600 times that of MoLPC. More importantly, our method is not limited by chain numbers, and can typically predict up to 60-chain complexes in less than 2 minutes. We illustrate the training efficiency (exploration efficiency) with the training curves shown in Fig. 4b. We observe that our GAPN model can converge rapidly and effectively after only \u223c400 episodes. This indicates that the model can efficiently explore a vast combination optimization space (N N\u22122, 11 \u2264N \u226430), which also demonstrates the high training efficiency of GAPN (requiring \u223c1 hour for a complete training on the whole 6,054 samples). (a) (b) Figure 4: (a). The RMSD performance of GAPN and the one without Adversarial reward (AR). The bar chart represents the relative difference between the two for multimers of each specific scale. (b). Training curve comparison between our GAPN model and the one without AR. Both models are trained on multimers of 3 \u2264N \u226430 with GT dimer structures. 4.3 ABLATION STUDY We conduct an ablation study on the whole validation set to explore the importance of our core adversarial reward (AR). Specifically, we call the process of GAPN assembling a multimer an episode and testing the policy network on the validation set after training the policy in each episode. In Fig. 4a, we find that GAPN\u2019s performance significantly decreases after removing AR. Importantly, the gain brought by AR will become more significant with the increase of chain number, which indicates that AR can effectively address the problem of knowledge generalization difficulty caused by data scarcity in large-scale multimers. Moreover, we also observe in Fig. 4b that AR is crucial for model training, as it enables rapid convergence of our model. 5"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2402.18813v1",
+                    "title": "Protein Multimer Structure Prediction via Prompt Learning",
+                    "abstract": "Understanding the 3D structures of protein multimers is crucial, as they play\na vital role in regulating various cellular processes. It has been empirically\nconfirmed that the multimer structure prediction~(MSP) can be well handled in a\nstep-wise assembly fashion using provided dimer structures and predicted\nprotein-protein interactions~(PPIs). However, due to the biological gap in the\nformation of dimers and larger multimers, directly applying PPI prediction\ntechniques can often cause a \\textit{poor generalization} to the MSP task. To\naddress this challenge, we aim to extend the PPI knowledge to multimers of\ndifferent scales~(i.e., chain numbers). Specifically, we propose\n\\textbf{\\textsc{PromptMSP}}, a pre-training and \\textbf{Prompt} tuning\nframework for \\textbf{M}ultimer \\textbf{S}tructure \\textbf{P}rediction. First,\nwe tailor the source and target tasks for effective PPI knowledge learning and\nefficient inference, respectively. We design PPI-inspired prompt learning to\nnarrow the gaps of two task formats and generalize the PPI knowledge to\nmultimers of different scales. We provide a meta-learning strategy to learn a\nreliable initialization of the prompt model, enabling our prompting framework\nto effectively adapt to limited data for large-scale multimers. Empirically, we\nachieve both significant accuracy (RMSD and TM-Score) and efficiency\nimprovements compared to advanced MSP models. The code, data and checkpoints\nare released at \\url{https://github.com/zqgao22/PromptMSP}.",
+                    "authors": "Ziqi Gao, Xiangguo Sun, Zijing Liu, Yu Li, Hong Cheng, Jia Li",
+                    "published": "2024-02-29",
+                    "updated": "2024-02-29",
+                    "primary_cat": "cs.CE",
+                    "cats": [
+                        "cs.CE"
+                    ],
+                    "main_content": "INTRODUCTION Recent advances in deep learning have driven the development of AlphaFold 2 (AF2) (Jumper et al., 2021), a groundbreaking method for predicting protein 3D structures. With minor modifications, AF2 can be extended to AlphaFold-Multimer (AFM) (Evans et al., 2021) to predict the 3D structure of multimers (i.e., proteins that consist of multiple chains), which is fundamental in understanding molecular functions and cellular signaling of many biological processes. AFM has been verified to accurately predict the structures of multimers with small scales (i.e., chain numbers). However, its performance rapidly declines as the scale increases. > Docking Probability > Determined Assembly Assembly to be determined Overlapping interface I-PPI =0.9 I-PPI =0.8 I-PPI =0.7 Condition C-PPI =0.1 C-PPI =0.7 A B \u221a \ud835\udc99 \ud835\udc9a \ud835\udc9b Docking Probability Figure 1: (A). Step-wise assembly for MSP. (B). Motivation for extending I-PPI to C-PPI. For multimer structure prediction (MSP), another research line (Esquivel-Rodr\u00b4 \u0131guez et al., 2012; Aderinwale et al., 2022; Inbar et al., 2005; Bryant et al., 2022) follows the idea of step-wise assembly (Figure 1A), where the assembly action indicates the protein-protein interaction (PPI). It sequentially expands the assembly size by adding a chain with the highest docking probability. The advantage of this step-wise assembly is that it can effectively handle multimers with large scales by enjoying the breakthrough in dimer structure prediction methods (Ganea et al., 2021; Wang et al., 2023; Ketata et al., 2023; Ghani et al., 2021; Luo et al., 2023; Chu et al., 2023; Evans et al., 2021). \u2217Correspondence to: Jia Li (jialee@ust.hk). 1 arXiv:2402.18813v1  [cs.CE]  29 Feb 2024 \fPublished as a conference paper at ICLR 2024 As the most advanced assembly-based method, MoLPC (Bryant et al., 2022) applies independent PPI (I-PPI, i.e., both proteins are independent without the consideration of other proteins) to score the quality of a given assembly. Despite great potential, it does not consider important conditions in the assembly such as the influence of third-party proteins to PPI pairs. For example, in Figure 1B, if chain x has already docked to chain y, the interface on x that will contact with z is partially occupied. Under this condition, the docking probability of (x, z) may decrease to lower than that of (y, z). We name this observation as condition PPI, or C-PPI. In short, neglecting the influence of C-PPI may easily lead to poor generalization. In this work, we focus on assembly-based MSP by learning C-PPI knowledge than I-PPI. 2 3 4 5 6 7 10 15 20 25 30 Chain Number 0 2000 4000 6000 8000 Count Figure 2: Distribution in chain numbers of multimers from the PDB database. Learning effective C-PPI knowledge for MSP presents two main challenges. Firstly, we observe significant gaps in the C-PPI knowledge contained in multimers with varied scales (chain numbers), which suggests that the biological formation process of multimers may vary depending on their scales. Secondly, as shown in Figure 2, experimental structure data for large-scale multimers is extremely limited, making it even more difficult for the model to generalize them. Recently, the rapidly evolving prompt learning (Liu et al., 2023; Sun et al., 2023a) techniques have shown promise to enhance the generalization of models to novel tasks and datasets. Inspired by this, a natural question arises: can we prompt the model to predict C-PPIs for multimers with arbitrary scales? To address this, our core idea is to design learnable prompts to transform arbitrary-scale multimers into fixed-scale ones. Concretely, we first define the target task for training (tuning) the prompt model, which is conditional link prediction. Then, we additionally design the pre-training (source) task that learns to identify arbitrary assembled multimer\u2019s correctness. In the target task, we transform the two query chains to a virtual assembled multimer, which is input into the pre-trained model for the correctness score. We treat such a score as the linking probability of the query chains. Therefore, arbitrary-scale prediction in the target task is reformulated as the fixed-scale one in the source task. Empirically, we investigate three settings: (1) assembly with ground-truth dimer structures to evaluate the accuracy of predicted docking path; (2) assembly with pre-computed dimers from AFM (Evans et al., 2021); and (3) assembly with pre-computed dimers from ESMFold (Lin et al., 2023). We show improved accuracy (in RMSD and TM-Score) and leading computational efficiency over recent state-of-the-art MSP baselines methods under these 3 settings. Overall, experiments demonstrate that our method has exceptional capacity and broad applicability. 2 RELATED WORK Multimer Structure Prediction. Proteins typically work in cells in the form of multimers. However, determining the structures of multimers with biophysical experiments such as X-ray crystallography (Maveyraud & Mourey, 2020; Ilari & Savino, 2008) and cryogenic electron microscopy (Costa et al., 2017; Ho et al., 2020) can be extremely difficult and expensive. Recently, the deep learning (DL)-based AlphaFold-2 (Jumper et al., 2021) model can milestone-accurately predict protein structures from residue sequences. Moreover, recent studies have explored its potential for predicting multimer structures. However, they mostly require time-consuming multiple sequence alignment (MSA) operations and the performance significantly decreases for multimers with great chain numbers. Another research line assumes that the multimer structures can be predicted by adding its chains one by one. Multi-LZerD (Esquivel-Rodr\u00b4 \u0131guez et al., 2012) and RL-MLZerD (Aderinwale et al., 2022) respectively apply the genetic optimization and reinforcement learning strategies to select proper dimer structures for assembly. However, even when targeting only small-scale (3-, 4and 5-chain) multimers, they still have low efficiency and are difficult to scale up for large multimers. By assuming that the dimer structures are already provided, MoLPC (Bryant et al., 2022) further simplifies this research line with the goal to predict just the correct docking path. With the help of 2 \fPublished as a conference paper at ICLR 2024 additional plDDT and dimer structure information, MoLPC has shown for the first time to predict the structures of large multimers with up to 30 chains. Prompt Learning for Pre-trained Models. In the field of natural language processing (NLP), the prevailing prompt learning approach (Brown et al., 2020; Min et al., 2021) has shown gratifying success in transferring prior knowledge across various tasks. Narrowing the gap between the source and target tasks is important for the generalization of pre-trained models on novel tasks or data, which has not been fundamentally addressed with the pre-training-fine-tuning paradigm (Zhou et al., 2022). To achieve this, researchers have turned their attention to prompts. Specifically, a language prompt refers to a piece of text attached to the original input that helps guide a pre-trained model to produce desired outputs (Gao et al., 2020). Prompts can be either discrete or continuous (Sun et al., 2023b; Li et al., 2023). The discrete prompt (Gao et al., 2020; Schick & Sch\u00a8 utze, 2020; Shin et al., 2020) usually refers to task descriptions from a pre-defined vocabulary, which can limit the flexibility of the prompt design due to the limited vocabulary space. In contrast, learnable prompts (Li & Liang, 2021; Zhang et al., 2021; Sun et al., 2023a) can be generated in a continuous space. Inspired by the success of prompt learning, we associate protein-protein interaction (PPI) knowledge (Kov\u00b4 acs et al., 2019; Gao et al., 2023a), which is commonly present in multimers across various scales, to the pretraining phase. By fine-tuning only the prompt model, we can effectively adapt the PPI knowledge to the target task. 3 PRELIMINARIES 3.1 PROBLEM SETUP Assembly Graph. We are given a set of chains (monomers), which is used to form a protein multimer. We represent a multimer with an assembly graph G = (V, E). In G, for the i-th chain, we obtain its chain-level embedding ci by the embedding function proposed in Chen et al. (2019). Each node vi \u2208V can thus represent one chain with node attribute ci. The assembly graph is an undirected, connected and acyclic (UCA) graph, with each edge representing an assembly action. Prepared Dimer Structures \ud835\udc7f\ud835\udfd0\u2032 \ud835\udc7f\ud835\udfd1\u2032 \ud835\udc7f\ud835\udfcf\u2032 \ud835\udc7f\ud835\udfd0 \ud835\udfd0\ud835\udfd1 \ud835\udc7f\ud835\udfd1 \ud835\udfd0\ud835\udfd1 \ud835\udc7f\ud835\udfd1 \ud835\udfcf\ud835\udfd1 \ud835\udc7f\ud835\udfcf \ud835\udfcf\ud835\udfd1 \ud835\udc7f\ud835\udfd0 \ud835\udfcf\ud835\udfd0 \ud835\udc7f\ud835\udfcf \ud835\udfcf\ud835\udfd0 Compute \ud835\udd4b\u2208\ud835\udc46\ud835\udc38(3) Apply \ud835\udd4b Assembly   Graph \ud835\udcd6 \ud835\udc97\ud835\udfcf \ud835\udc97\ud835\udfd1 \ud835\udc97\ud835\udfd0 Assemble Start Guide Figure 3: Assembly process with the predicted assembly graph and prepared dimers. Assembly Process. For clarity, we apply an example in Figure 3 to illustrate the assembly process, which is achieved with the prepared dimer structures and the predicted assembly graph. Let us consider a multimer with 3 chains, whose undocked 3D structures are denoted as X1, X2, X3. We consider an assembly graph with the edge set {(v2, v3), (v3, v1)}, and the dimer structures {(X12 1 , X12 2 ), (X13 1 , X13 3 ), (X23 2 , X23 3 )}. First, we select the dimer of chains 2 and 3 as the start point, i.e., X\u2032 2 = X23 2 and X\u2032 3 = X23 3 . Next, to dock the chain 1 onto chain 3, we compute the SE(3) coordinate transformation T that aligns X13 3 onto X\u2032 3. Lastly, we apply T to X1, resulting in the update coordinate X\u2032 1 of chain 1. Definition 1 (Assembly Correctness) For an N-chain multimer with a 3D structure X, its chains are represented by the nodes of an assembly graph G. The assembly correctness F(G, X) is equivalent to the TM-Score (Zhang & Skolnick, 2004) between the assembled multimer and the ground-truth. With the above definitions, our paper aims to predict assembly graphs that maximize the TM-Scores, taking as inputs the residue sequences of chains and pre-calculated dimer structures. 3.2 SOURCE AND TARGET TASKS In this paper, we adopt a pre-training (source task) and prompt fine-tuning (target task) framework to address the MSP problem. We consider two points for task designs: 1) With given multimers for pre-training, the model benefits from common intrinsic task subspace between the source and target task. 2) The target task should be designed to effectively learn the condition PPI (C-PPI) knowledge and efficiently complete the MSP inference. 3 \fPublished as a conference paper at ICLR 2024 3 5 7 9 11 13 15 17 19 21 23 25 25 23 21 19 17 15 13 11 9 7 5 3 Source task: chain 1 2 3 4 5 6 6 5 4 3 2 1 Source task: degree 3 5 7 9 11 13 15 17 19 21 23 25 25 23 21 19 17 15 13 11 9 7 5 3 Target task: chain 1 2 3 4 5 6 6 5 4 3 2 1 Target task: degree 0.2 0.4 0.6 0.8 1.0 Figure 4: Analysis on multimers with varied chain numbers. We select some samples for evaluation and visualize heatmaps that show the similarity of the sample embeddings obtained from different pre-trained models. Each value on the axis suggests that the model is trained on data with the specific chain number or degree value. For example in the heatmap titled \u2018Source task: chain\u2019, the darkness at the [5,7] block represents the similarity between the embeddings extracted from two models that are trained under the source task with 5and 7-chain multimers, respectively. Definition 2 (Source Data Dsou) Each data instance in Dsou involves an assembly graph Gsou and a continuous label ysou, i.e., (Gsou, ysou) \u2208Dsou. For an N-chain multimer, Gsou is randomly generated as its N-node UCA assembly graph, and ysou is the assembly correctness. Definition 3 (Target Data Dtar) Each data instance in Dtar involves an assembly graph Gtar = (Vtar, Etar), an indicated node vd \u2208Vtar, an isolated node vu / \u2208Gtar and the continuous label ytar, i.e., (Gtar, vd, vu, ytar) \u2208Dtar. For an N-chain multimer, Gtar is defined as its (N \u22121)-chain assembly graph. ytar is calculated by ytar = F(({Vtar \u222avu}, {Etar \u222avdvu}), X). Source Task. We design the source task as the graph-level regression task. Based on source data defined in Def. 2, the model is fed with a UCA assembly graph and is expected to output the continuous correctness score between 0 and 1. Note that theoretically, we can generate N N\u22122 different UCA graphs and their corresponding labels with an N-chain multimer. This greatly broadens the available training data, enhancing the effectiveness of pre-training in learning MSP knowledge. Target Task. We design the target task as the link prediction task (i.e., predicting the C-PPI probability). Based on target data defined in Def. 3, the target task aims to predict the presence of the link between nodes vd and vu, which represent the docked and undocked chains, respectively. We provide the detailed process to generate Dsou and Dtar in Appendix A.1. Overall, the source and target tasks learn assembly knowledge globally and locally, respectively. Unfortunately, multimers of varied scales exhibit distribution shifts in the target task, preventing the direct use of it for MSP. Next, we will empirically verify the existence of these shifts and their influence on MSP. 3.3 GAPS BETWEEN MULTIMERS WITH VARIED SCALES We have presented a significant imbalance in the data of multimers (see Figure 2). Here, we analyze the gap in MSP knowledge among multimers with various scales (i.e., chain numbers), which can further consolidate our motivation of utilizing prompt learning for knowledge transfer. We also offer explanations for the reasons behind the gaps based on empirical observations. We begin by analyzing the factor of chain number. We randomly select multimers for evaluation and divide the remaining ones into various training sets based on the chain number, which are then trained with independent models. We obtain the chain representations of the evaluating samples in each model. Lastly, we apply Centered Kernel Alignment (CKA) (Raghu et al., 2021), a function for evaluating representation similarity, to quantify the gaps in knowledge learned by any two models. We show the heatmaps of CKA in the Figure 4 and have two observations. (a) Low similarities are shown between data with small and large scales. (b) Generally, there is a positive correlation between the C-PPI knowledge gap and the difference in scales. In short, C-PPI knowledge greatly depends on the multimer scale. To further explain these gaps, we re-divide the training sets based on the degree (i.e., neighbor number of a node) of assembly graphs and perform additional experiments. Specifically, we define the degree value as the highest node degree within each graph Gsou in the source task, and as the 4 \fPublished as a conference paper at ICLR 2024 Prompt  Model Prompt embeddings \ud835\udf3d\u2217 \ud835\udf45 Pre-trained Model ? GIN  Encoder Task Head Correctness \ud835\udf3d \ud835\udf53 Graph-level Pre-training (A) Undocked chain Docked chain Chain embedding \ud835\udc6f\ud835\udc85 \ud835\udc6f\ud835\udc96 \ud835\udc6f\ud835\udc99 \ud835\udc6f\ud835\udc9a \ud835\udc6f\ud835\udc99 \ud835\udc6f\ud835\udc9a \ud835\udc6f\ud835\udc85 \ud835\udc6f\ud835\udc96 Correctness Probability = \ud835\udf3d\u2217 \ud835\udf53\u2217 Prompt Tuning Inference Prompt Tuning  Result \ud835\udc53 \ud835\udf3d\u2217,\ud835\udf53\u2217,\ud835\udf45\u2217 Support Set Fast Adaptation (Prompt Tuning) Meta Testing ? ; ; ? \u2026 ? \u2026 Target  assembly \u2026; (\ud835\udc75= \ud835\udfd6) (\ud835\udc75= \ud835\udc75\ud835\udc8e\ud835\udc82\ud835\udc99) \ud835\udc53% \ud835\udf45|\ud835\udf3d\u2217,\ud835\udf53\u2217 0 1 2 9 Start Probability Choose  the best 0 2 \ud835\udc53% \ud835\udf45|\ud835\udf3d\u2217,\ud835\udf53\u22171 3 9 \u2026 Choose  the best 0 2 \u2026 (A) (B) (C) \ud835\udc53 \ud835\udf3d\u2217,\ud835\udf53\u2217,\ud835\udf45\u2217 \ud835\udc53 \ud835\udf3d\u2217,\ud835\udf53\u2217,\ud835\udf45\u2217 \ud835\udc53 \ud835\udf3d\u2217,\ud835\udf53\u2217,\ud835\udf45\u2217 Step 1 Step 7 Step 8 : Undocked \ud835\udfcf \ud835\udfd0\ud835\udfd1 \ud835\udfd2 \ud835\udfd3 \ud835\udfd4 \ud835\udfd5 \ud835\udfd6 \ud835\udfd7 Docked 7\u00d72 = 14  probabilities \ud835\udfcf \ud835\udfd0\ud835\udfd1 \ud835\udfd2 \ud835\udfd3 \ud835\udfd4 \ud835\udfd5\ud835\udfd7 Pick the best Figure 5: The overview of PromptMSP. (A). Firstly, we pre-train the GIN encoder and the task head under the graph-level regression task. After pre-training, given an arbitrary graph, \u03b8\u2217and \u03d5\u2217jointly output the correctness. (B). During prompt tuning, the prompt model takes embeddings of a pair of docked and undocked (query) chains as input and learns to produce prompt embeddings which form the entire 4-node path. \u03b8\u2217and \u03d5\u2217then jointly predict the correctness, which is equivalent to the linking probability. We use f\u03b8\u2217,\u03d5\u2217,\u03c0\u2217to denote the well trained pipeline that outputs the linking probability of query chains with target data instance as input. (C). If the target multimer has 9 chains, we sequentially perform 8 steps for inference. In each step, we use the well trained pipeline to calculate the probabilities for all possible chain pairs and select the most possible pair to assemble. degree of node vd to be docked on in the target task. As shown in Figure 4, CKA heatmaps indicate that training samples with different degrees exhibit the gap in knowledge, which becomes even more significant than that between data with varying chain numbers. As observed, we conclude that the gap in chain numbers between data may be primarily due to the difference in degree of assembly graphs. Therefore, we intuitively associate the degree with the biological phenomenon of competitive docking that may occur in MSP, as evidenced by previous studies (Chang & Perez, 2022; Yan et al., 2020; Chang & Perez, 2023). In other words, multimers with more chains are more likely to yield assembly graphs with high degrees, and consequently, more instances of competitive docking. We expect that prompt learning can help bridge this knowledge gap. 4 PROPOSED APPROACH Overview of Our Approach. Our approach is depicted in Figure 5, which follows a pre-training and prompt tuning paradigm. Firstly, using abundant data of small-scale (i.e., 3 \u2264N \u22645) multimers, we pre-train the graph neural network (GNN) model on the source graph regression task. We then design the learnable prompt model, which can reformulate the conditional link prediction (target) task to the graph-level regression (source) task. In the process of task reformulation, an arbitrary-scale multimer in the target task is converted to a fixed-scale (i.e., N = 4) multimer in the source task. For inference, an N-chain multimer will go through N \u22121 steps to be fully assembled. In each step, our model predicts the probabilities of all possible conditional links and selects the highest one to add a new chain. Besides, to further enhance the generalization ability, we provide a meta-learning strategy in Appendix C.5. 4.1 PRE-TRAINING ON THE SOURCE TASK We apply graph neural network (GNN) architecture (Xu et al., 2018; Veli\u02c7 ckovi\u00b4 c et al., 2017; Tang et al., 2023) for graph-level regression (Cheng et al., 2023; Gao et al., 2023b). Our model first computes the node embeddings of an input UCA assembly graph using Graph Isomorphism Network (GIN, Xu et al. (2018)) to achieve its state-of-the-art performance. Kindly note that we can also apply other GNN variants (Kipf & Welling, 2016; Veli\u02c7 ckovi\u00b4 c et al., 2017; Tang et al., 2022; Liu et al., 2024; Li et al., 2019) for pre-training. Following Def. 2, we construct source data Dsou by using the oligomer (i.e., small-scale multimer) data only. The pre-training model approximates the assembly correctness with data instances of Dsou: \u02dc ysou = GNN(Gsou; \u03b8, \u03d5) \u2248F(Gsou; X) = ysou, (1) 5 \fPublished as a conference paper at ICLR 2024 where GNN represents combination of a GIN with parameters \u03b8 for obtaining node embeddings, a ReadOut function after the last GIN layer, and a task head with parameters \u03d5 to yield the prediction \u02dc ysou. As defined in Def. 1, F represents the assembly correctness function for computing TM-Score between the assembled structure and the ground-truth (GT) structure X. We train the GNN by minimizing the discrepancy between predicted and GT correctness values: \u03b8\u2217, \u03d5\u2217= arg min\u03b8,\u03d5 X (Gsou,ysou)\u2208Dsou Lpre(ysou = F(Gsou; X), \u02dc ysou = GNN(Gsou; \u03b8, \u03d5)), (2) where Lpre is the mean absolute error (MAE) loss function. After the pre-training phase, we obtain the pre-trained GIN encoder and the task head parameterized by \u03b8\u2217and \u03d5\u2217, respectively. 4.2 ENSURING CONSISTENCY BETWEEN SOURCE AND TARGET TASKS Reformulating the Target Link Prediction Task. The inference of an N-chain multimer under the source task setting requires all of its N N\u22122 assembly graphs and their corresponding correctness from the pre-trained model. Therefore, when dealing with large-scale multimers, such inference manner requires effective UCA graph traversal algorithms and significant computational resources. The target task proposed in Section 3.2 can address this issue, which aims to predict the link presence (probability) between a pair of docked and undocked chains. As shown in Figure 5C, we can inference the structure of an N-chain multimer with just N \u22121 steps. At each step, we identify the most promising pair of (docked and undocked) chains. The success of traditional pre-training and fine-tuning paradigm is due to that source and target tasks having a common task subspace, which allows for unobstructed knowledge transfer (Sun et al., 2023a). However, in this paper, source and target tasks are naturally different, namely graphlevel and edge-level tasks, respectively. Here, we follow three principles to reformulate the target task: (1) To achieve consistency with the source task, the target task needs to be reformulated as a graph-level problem. (2) Due to the distribution shifts in multimers with varied chain numbers (Figure 4), a multimer of arbitrary scale in the target conditional link prediction task should be reformulated into a fixed-scale one in the source task. (3) The pre-trained GNN model is expected to effectively handle multimers of such \u201cfixed-scale\u201d in the source task. The upcoming introduction of prompt design will indicate that the fixed-scale value is 4. Therefore, to ensure (3), we limit the data used for pre-training to only multimer of 3 \u2264N \u22645. H \u2208R(|Vtar|+1)\u00d7d = \u03b8\u2217(Vtar \u222avu, Etar), (3) Hx = \u03c3\u03c0 \u0010\u0002 softmax \u0000H\u22a4 d Hu \u0001 H\u22a4 u \u0003\u22a4\u0011 , (4) Hy = \u03c3\u03c0 \u0010\u0002 softmax \u0000H\u22a4 u Hd \u0001 H\u22a4 d \u0003\u22a4\u0011 , (5) Gpro = (Vpro = {vd, vu, vx, vy}, Epro = {edx, exy, eyu}), (6) \u02dc ytar = \u03d5\u2217(\u03b8\u2217(Gpro)), (7) Prompt Design. Following Def. 3, we create the target data Dtar for prompt tuning. For clarity, we denote each data instance as a tuple form (Gtar, vd, vu, ytar) \u2208Dtar, where Gtar denotes the current assembled multimer (i.e., condition), vd is a query chain within Gtar and vu is another query chain representing the undocked chain. We compute the last layer embeddings H of all nodes in Gtar and the isolated vu with the pre-trained GIN encoder. To enable communications between target nodes vd and vu, the prompt model parameterized by \u03c0 contains multiple cross attention layers (Vaswani et al., 2017; Wang et al., 2023) that map Hu, Hd \u2208Rd to vectors Hx, Hy \u2208Rd, which represent the initial features of nodes vx and vy. Finally, the pre-trained model outputs the assembly correctness of the 4-node prompt graph Gpro. The whole target task pipeline of our method is represented by the equations on the right side. Specifically, \u03b8\u2217is the pre-trained GIN encoder, \u03d5\u2217is the pre-trained task head and d denotes the dimension of features. The prompt model \u03c0, which outputs a vector Rd, includes non-trainable cross attention layers and the parametric function (Multi-Layer Perceptron, MLP) \u03c3\u03c0. Moreover, we use f\u03b8\u2217,\u03d5\u2217,\u03c0\u2217to represent the entire pipeline (Figure 5B) which takes input (Gtar, vd, vu) and outputs \u02dc ytar. A more detailed model architecture is shown in Appendix A.2. 6 \fPublished as a conference paper at ICLR 2024 Similar PPI \u2113= 1 \u2113= 2 \u2113= 3 1. 2. 3. High probability PPI ? Figure 6: The \u2113= 3 PPI rule. Prompt Design Intuition. First of all, the link between two query chains is equivalent to the protein-protein interaction (PPI) in biology. We introduce the \u2113= 3 path (Kov\u00b4 acs et al., 2019; Yuen & Jansson, 2023), a widely validated biological rule for modeling PPI probabilities. Figure 6 describes the \u2113= 3 rule, which is based on the fact that docking-based PPI generally requires proteins to have complementary surface representations for contact. It claims that the PPI probability of any two chains is not reflected by the number of their common neighbors (a.k.a., triadic closure principle (Lou et al., 2013; Sintos & Tsaparas, 2014)), but rather by the presence of a path of length \u2113= 3. In short, if there exists a 4-node path with query chains at the ends, they are highly likely to have a PPI (link). Regardless of the node number of Gtar, we treat the pre-trained model output based on this 4node Gpro as the linking probability between vd and vu. Unlike most of exsiting prompt learning techniques, the proposed task reformulation manner is naturally interpretable. Let us assume that these two query chains are highly likely to have a link, it will be reasonable to find two virtual chains that help form a valid \u2113= 3 path. This also suggests that the assembly of Gpro tends to be correct, i.e., \u03d5\u2217(\u03b8\u2217(Gpro)) \u21921. Therefore, intuitively, the correctness of Gpro fed to the pre-trained model implies the linking probability between vu and vd. 4.3 INFERENCE PROCESS WITH THE PROMPTING RESULT f\u03c0\u2217|\u03b8\u2217,\u03d5\u2217 With prompt tuning, we obtain the whole framework pipeline f\u03c0\u2217|\u03b8\u2217,\u03d5\u2217. For inference on a multimer with N chains, we perform N \u22121 assembly steps, in each of which we apply the pipeline f\u03c0\u2217|\u03b8\u2217,\u03d5\u2217 to predict the linking probabilities of all pairs of chains and select the most likely pair for assembly. 5 EXPERIMENTS Datasets. We collect all publicly available multimers (3 \u2264N \u226430) from the Protein Data Bank (PDB) database (Berman et al., 2000) on 2023-02-20. Referring to the data preprocessing method in MoLPC (Bryant et al., 2022), we obtain a total of 9,254 multimers. To further clarity, we use the abbreviation PDB-M to refer to the dataset applied in this paper. Overall, the preprocessing method ensures that PDB-M contains high resolution, non-redundant multimers and is free from data leakage (i.e., no sequences with a similarity greater than 40% between training and test sets). Due to the commonly low efficiency of all baselines, we define a data split for 3 \u2264N \u226430 with a small test set size to enable comparison. Specifically, we select 10 for each scale of 3 \u2264N \u226410, and 5 for each scale of 11 \u2264N \u226430. Moreover, for comprehensive evalaution, we re-split the PDBM dataset based on the release date of the PDB files to evaluate our method. Detailed information about the data preprocessing methods and the data statistic of split is in Appendix B. Baselines and Experimental Setup. We compare our PROMPTMSP method with recent deep learning (DL) models and traditional software methods. For DL-based state-of-the-art methods, RL-MLZerd (Aderinwale et al., 2022) and AlphaFold-Multimer (AFM) (Evans et al., 2021) are included. For software methods, Multi-LZerd (Esquivel-Rodr\u00b4 \u0131guez et al., 2012) and MoLPC (Bryant et al., 2022) are included. Since assembly-based methods require given dimers, we first use the ground-truth (GT) dimer structures (represented as GT Dimer) to evaluate the assembled multimer structures. For the pair of chains with contact, GT Dimer includes their native dimer structure drawn from the GT multimer. For those without contact, we use EquiDock (Ganea et al., 2021) for outputting dimer structures due to its fast inference speed. Moreover, since GT dimers are not always available, for pratical reasons, we consider to prepare dimers with AFM (Evans et al., 2021) (AFM Dimer) and ESMFold (Lin et al., 2023) (ESMFold Dimer). For baselines not requiring given dimers, we use these three kinds of dimers to reassemble based on the docking path mined in their predicted multimer, which is referred to as the \u2020 version of the baselines. Our experiments consist of 3 settings: 1) Since most baselines can not handle multimers with chain numbers N > 10. We follow GT Dimer, AFM Dimer and ESMFold Dimer to evaluate all baselines on the small-scale multimers (3 \u2264N \u226410) in the test set. 2) We evaluate MoLPC and our method by using these three types of dimers on the 7 \fPublished as a conference paper at ICLR 2024 Table 1: Multimer structure prediction results. Methods are evaluated on the test set of 3 \u2264 N \u226410 by using three types of pre-computed dimers. The test set includes 80 multimer samples in total (10 samples for each scale). For each dimer type and metric, the best method is bold and the second best is underlined. \u2020 represents the reassembly version of baselines. R(Avg): average RMSD; R(Med): median RMSD; T(Avg): average TM-Score; T(Med): median TM-Score. Methods GT Dimer AFM Dimer ESMFold Dimer R(Avg)/R(Med) T(Avg)/T(Med) R(Avg)/R(Med) T(Avg)/T(Med) R(Avg)/R(Med) T(Avg)/T(Med) Multi-LZerD 31.50 / 33.94 0.28 / 0.25 31.50 / 33.94 0.28 / 0.25 31.50 / 33.94 0.28 / 0.25 Multi-LZerD\u2020 18.90 / 19.30 0.54 / 0.38 29.68 / 27.96 0.30 / 0.33 33.00 / 31.07 0.25 / 0.29 RL-MLZerD 31.04 / 27.44 0.29 / 0.32 31.04 / 27.44 0.29 / 0.32 31.04 / 27.44 0.29 / 0.32 RL-MLZerD\u2020 17.77 / 17.69 0.51 / 0.53 28.57 / 26.20 0.30 / 0.35 27.76 / 32.91 0.32 / 0.25 AFM 20.99 / 24.76 0.47 / 0.42 20.99 / 24.76 0.47 / 0.42 20.99 / 24.76 0.47 / 0.42 AFM\u2020 16.79 / 16.02 0.59 / 0.59 18.98 / 19.05 0.50 / 0.48 26.76 / 29.95 0.33 / 0.30 MoLPC 18.53 / 18.08 0.52 / 0.55 23.06 / 23.92 0.43 / 0.42 30.17 / 29.45 0.31 / 0.31 PromptMSP 13.57 / 11.74 0.67 / 0.71 17.36 / 17.09 0.55 / 0.56 22.55 / 24.85 0.45 / 0.37 (A) (B) (C) \u00d7 \u00d7 \u00d7 \u00d7 \u00d7 \u00d7 \u00d7 \u00d7 R: 0.96 R: 0.63 Figure 7: (A). TM-Score distribution of MoLPC and our method tested on multimers of 10 \u2264 N \u226430. The mean and median values are marked with white and black \u2018\u00d7\u2019, respectively. (B). The relationship of learned C-PPI knowledge and the actual TM-Score. (C). The relationship of I-PPI learned by MoLPC and the actual TM-Score. We show the Pearson\u2019s correlation R for both. entire test set (3 \u2264N \u226430). 3) We additionally split the PDB-M dataset based on the release date of multimers to evaluate the generalization ability of our method. We run all methods on 2 A100 SXM4 40GB GPUs and consider exceeding the memory limit or the resource of 10 GPU hours as failures, which are padded by the upper bound performance of all baselines. Evaluation Metrics. To evaluate the performance of multimer structure prediction, we calculate the root-mean-square deviation (RMSD) and the TM-score both at the residue level. We report the mean and median values of both metrics. Multimer Structure Prediction Results. Model performance on multimers of two kinds of scales (3 \u2264N \u226410, 11 \u2264N \u226430) are summarized in Table 1 and Figure 7A, respectively. For small-scale multimers, our model achieves state-of-the-art on all metrics. In addition, we find that most MSP methods can benefit from the reassembly of GT or AFM dimer structures. Notably, our model can significantly outperform MoLPC, even though it does not require additional plDDT information or coarse information for protein interactions. For larger-scale multimers, our model also outperforms MoLPC, and outputs completely accurate prediction results for certain samples (i.e., TM-Score = 1.0 under GT Dimer). As for the failed inference samples of MoLPC, we relax the model\u2019s penalty term to successfully obtain the predictions instead of simply considering its TM-Score as 0. Despite this, our model can still achieve significant improvements under GT Dimer, AFM Dimer and ESMFold Dimer. The experimental results under the data split based on the release dates is in Appendix C.1. 8 \fPublished as a conference paper at ICLR 2024 Table 2: Efficiency comparison (average MSP inference time). Time(min) 3 \u2264N \u226410 11 \u2264N \u226430 Path Dimer Total Path Dimer Total Multi-LZerD 187.51 \u2013 187.50 \u2013 RL-MLZerD \u2013 173.88 173.88 AFM \u2013 \u2013 155.72 MoLPC 11.64 165.73 177.37 11.64 354.23 365.87 Ours-GT 0.01 \u2013 0.01 0.04 \u2013 0.04 Ours-AFM 0.01 80.79 80.80 0.04 187.44 187.48 Ours-ESMFold 0.01 0.35 0.36 0.01 1.09 1.10 Table 2 shows the inference efficiency of all baselines. As assembly-based methods require given dimer structures, we report the separate running time for predicting the docking path and preparing dimers, as well as the total time consumption. Kindly note that during inference, our method predicts the docking path without the need for pre-computed dimers. Therefore, to predict the structure of an N-chain multimer, our method (always) requires N \u22121 pre-computed dimers. We note that regardless of the dimer type used, our method is significantly faster than the other baselines. Our method also achieves higher efficiency in predicting the docking path compared to MoLPC. We provide more docking path inference results of our method (in Figure 9 in Appendix). We can find that as the scale increases, the inference time for a single assembly process (the orange curve) of our method does not increase, which suggests that the applicability of our model is not limited by the scale. Table 3: Ablation study with GT dimers. Prompt C-PPI 3 \u2264N \u226410 11 \u2264N \u226430 \u00d7 \u2713 0.55(-17.9%) 0.29(-21.6%) \u2713 \u00d7 0.54(-19.4%) 0.33(-10.8%) \u2713 \u2713 0.67 0.37 Ablation Study. We perform ablation study in Table 3 to explore the significance of the prompt model and the C-PPI modelling strategy. If we remove the prompt model and apply the link prediction task both for pre-training and fine-tuning, the performance will greatly decrease by about 21.6% on large-scale multimers. This implies the contribution of prompt in unifying the C-PPI knowledge in multimers of different scales. Similarly, the significance of applying C-PPI modelling can be further illustrated through its relationship with the MSP problem. Figure 7(BC) indicates that I-PPI will bring negative transfer to the MSP task, ultimately hurting the performance. 7 6~8 5~9 4~103~113~133~153~173~193~213~233~253~273~30 Scale(s) of multimers for training 0.40 0.45 0.50 0.55 0.60 TM-Score w/o prompt PromptMSP Figure 8: Results tested on N=7. We train our model and its \u2018w/o prompt\u2019 version on multimers of varied scale ranges. In Figure 8, we show the generalization ability of our method. The term \u2018w/o prompt\u2019 refers to the direct use of GNNs for conditional link prediction for MSP. We find that when introducing training multimers with the scale (e.g., N > 11) differs significantly from the testing multimers (i.e., N = 7), the performance of the \u2018w/o prompt\u2019 method notably declines. Conversely, for PROMPTMSP, adding arbitray-scale multimers to the training set will improves the model\u2019s generalization ability. This indicates that our model can effectively capture shared knowledge between varied-scale multimers, while blocking the knowledge gaps caused by distribution shifts. 6"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2402.10609v1",
+                    "title": "U$^2$MRPD: Unsupervised undersampled MRI reconstruction by prompting a large latent diffusion model",
+                    "abstract": "Implicit visual knowledge in a large latent diffusion model (LLDM)\npre-trained on natural images is rich and hypothetically universal to natural\nand medical images. To test this hypothesis, we introduce a novel framework for\nUnsupervised Undersampled MRI Reconstruction by Prompting a pre-trained large\nlatent Diffusion model ( U$^2$MRPD). Existing data-driven, supervised\nundersampled MRI reconstruction networks are typically of limited\ngeneralizability and adaptability toward diverse data acquisition scenarios;\nyet U$^2$MRPD supports image-specific MRI reconstruction by prompting an LLDM\nwith an MRSampler tailored for complex-valued MRI images. With any\nsingle-source or diverse-source MRI dataset, U$^2$MRPD's performance is further\nboosted by an MRAdapter while keeping the generative image priors intact.\nExperiments on multiple datasets show that U$^2$MRPD achieves comparable or\nbetter performance than supervised and MRI diffusion methods on in-domain\ndatasets while demonstrating the best generalizability on out-of-domain\ndatasets. To the best of our knowledge, U$^2$MRPD is the {\\bf first}\nunsupervised method that demonstrates the universal prowess of a LLDM, %trained\non magnitude-only natural images in medical imaging, attaining the best\nadaptability for both MRI database-free and database-available scenarios and\ngeneralizability towards out-of-domain data.",
+                    "authors": "Ziqi Gao, S. Kevin Zhou",
+                    "published": "2024-02-16",
+                    "updated": "2024-02-16",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "main_content": "Introduction Magnetic Resonance Imaging (MRI) is a cornerstone in clinical diagnostics, renowned for its safety. It can generate diverse, high-contrast images, including T1-weighted, T2weighted, and Proton Density-weighted images, enabling comprehensive radiological assessments. However, MRI\u2019s slow acquisition speed, due to inherent physical and physiological constraints, limits its clinical efficiency. Extended scan duration not only causes patient discomfort but also increases the likelihood of motion artifacts and systemic errors in the images. To this end, accelerated MRI techniques with undersampling acquisition have been developed. Deep learning methods have come to the forefront in the field of undersampled MRI reconstruction. Until recently, supervised methods (Schlemper et al., 2017; Qin et al., 2018; Eo et al., 2018; Cheng et al., 2019; Zhou & Zhou, 2020; Guo et al., 2023; Zhou et al., 2023) dominate this field. Although these methods excel in in-domain reconstruction 1 arXiv:2402.10609v1  [eess.IV]  16 Feb 2024 \fTitle Suppressed Due to Excessive Size performance, they often lack generalizability to scenarios outside of their training environment, notably across varying sampling patterns, organs, and MRI modalities (Liu et al., 2021; Chung & Ye, 2022). Unsupervised reconstruction methods for MRI, such as implicit neural representations (INR) (Shen et al., 2022; Feng et al., 2023b;c), are scanor image-specific and exhibit a high degree of adaptability. However, no existing work has proven INR\u2019s potential in utilizing large data prior for MR imaging. Recently, diffusion models (Song et al., 2021b; Peng et al., 2022; Chung & Ye, 2022) have emerged as superior to supervised state-of-the-art methods. Instead of using a discriminative objective, they train a diffusion model to approximate the distribution of MR images, eliminating the dependency on particular sampling patterns. (Chung & Ye, 2022) have demonstrated its notable generalization capability across diverse organs and modalities. However, the dataset scale of current MRI diffusion models (Song et al., 2021b; Peng et al., 2022; Chung & Ye, 2022) typically ranges from thousands to tens of thousands of samples and it remains challenging to curate large-scale MR images due to privacy privacy. Parallel to this, large latent diffusion models (LLDM) (Podell et al., 2023; Rombach et al., 2022), revolutionary image generative models exemplified by Stable Diffusion (SD), are trained on vast Internet-scale images (Schuhmann et al., 2022) of about 106 times larger than existing public datasets (Knoll et al., 2020; Zbontar et al., 2018) used for training MRI diffusion models. SD models can generate diverse, high-fidelity images and have proven a powerful backbone (Zhao et al., 2023; Xu et al., 2023; Yang et al., 2023; Gong et al., 2023) with supervised fine-tuning. These facts imply the richness of visual knowledge in an LLDM and prompt our questions: Is the implicit visual knowledge in the LLDM pre-trained on natural images universal for both natural and medical images? If yes, how can it be harnessed and enhance the adaptability and generalizability of undersampled MRI reconstruction models? We believe that universal visual knowledge exists and can be harnessed for MRI data. In this paper, we provide a solution, U2MRPD, to the second question in an attempt to answer the first question from a practical perspective. As the most related area of research, Diffusion model-based Inverse problem Solvers (DIS) perform conditional posterior sampling using the pre-trained diffusion prior. The majority of DIS focus on pixel (Kadkhodaie & Simoncelli, 2021; Kawar et al., 2022; Wang et al., 2023; Feng et al., 2023a; Chung et al., 2023a; 2022a; Song et al., 2023c;b) while few works explore their effectiveness in the latent domain (Song et al., 2023a; Rout et al., 2023; Chung et al., 2023c). Furthermore, most of the research concentrates on pixel-domain solvers using in-distribution data and has not explored complex-valued image restoration. In this context, we investigate the challenging combination of perCharacteristics \\ Model type Supervised models INR Diffusion U2MRPD No paired data. \u00d7 \u2713 \u2713 \u2713 Work in database-free scenarios. \u00d7 \u2713 \u00d7 \u2713 Can leverage large data. \u2713 \u00d7 \u2713 \u2713 Work with diverse sampling masks. poor \u2713 \u2713 \u2713 OOD organs and contrast poor \u2713 average \u2713 Table 1. Characteristics of supervised and unsupervised models of undersampled MRI reconstruction. The last two rows are summarized from our experiments. forming latent diffusion model sampling for extreme outof-distribution and complex-valued data. In this work, we propose a generalizable and flexible MRI reconstruction framework, namely Unsupervised Undersampled MRI Reconstruction by Prompting a pre-trained latent Diffusion model (U2MRPD). The comparison between U2MRPD and other methods are provided in Table 1. Despite an LLDM has not been explicitly trained on medical, complex, or aliased images, our MRSampler can prompt an large LLDM to reconstruct undersampled MR images. To utilize any MRI dataset, we propose the generality-preserved MRAdapter, which efficiently resolves distribution mismatches with a minimal computational cost while safeguarding data priors. The experimental results on both simulation and complex-valued datasets demonstrate that our U2MRPD performs comparably to the supervised and MRI diffusion methods on in-domain datasets, while significantly outperforming them on out-of-domain datasets. We also conduct extensive ablation studies for several key designs of our method, verifying their effectiveness. We observe an improvement in adaptability and generalizability upon current literature and anticipate that our findings will stimulate further research into leveraging general image priors to develop more robust medical imaging systems. Our contributions are as follows: \u2022 To the best of our knowledge, we are the first to retrieve the implicit visual knowledge in an LLDM pretrained on natural images for medical imaging inverse problems, particularly accelerated MRI, in an unsupervised fashion. \u2022 Our MRSampler, tailored for complex-valued MRI images, allows physically consistent diffusion inference with a novel strong-to-mild control mechanism combined with pseudo-random phase modulation. Our MRAdapter efficiently bridges the gap between natural and medical image distributions and preserves the integrity of generative image priors. \u2022 To underscore U2MRPD\u2019s effectiveness, we provide an empirical analysis of MRSampler and further validate MRAdapter with a universal MRI reconstruction task (Liu et al., 2021). 2 \fTitle Suppressed Due to Excessive Size \u2022 Comprehensive evaluations against state-of-the-art supervised and MRI diffusion methods confirm our approach\u2019s parity or superiority, highlighting its adaptability and generalization capabilities across varying MRI conditions. 2. Preliminaries 2.1. DDIM sampler Denoising Diffusion Implicit Models (DDIMs) (Song et al., 2021a) provide a backbone for our algorithmic framework since they are ideal for solving time-sensitive inverse problems. DDIM sampler adopts a non-Markovian process and can generate samples with fewer diffusion steps. Formally, it generates a sample xt\u22121 from a sample xt via: xt\u22121 = \u221a\u03b1t\u22121 \u0012 xt\u2212\u221a1\u2212\u03b1t\u00b7\u03f5(t) \u03b8 (xt) \u221a\u03b1t \u0013 + p 1 \u2212\u03b1t\u22121 \u2212\u03c32 t \u00b7 \u03f5(t) \u03b8 (xt) + \u03c3t \u00b7 \u03bet, (1) where \u03f5(t) \u03b8 represents the pre-trained score function compatible with DDPMs, and \u03bet is random noise. \u03c3t, the standard deviation of Gaussian noise at time t, controls the stochasticity of the diffusion process with a step-wise schedule. The process becomes deterministic when setting \u03c3t to zero. 2.2. Classifier guidance Algorithm 1 Classifier guided deterministic DDIM sampling (Song et al., 2021a) 1: xT \u223cN(0, I) 2: for t = T, . . . , 1 do 3: \u02c6 \u03f5 \u2190\u03f5\u03b8(xt) \u2212\u03b3\u221a1 \u2212\u03b1t\u2207xt log p\u03d5(y|xt) // classifier correction 4: \u02c6 x0 \u2190xt\u2212\u221a1\u2212\u03b1t\u02c6 \u03f5 \u221a\u03b1t // predict x0 given xt 5: xt\u22121 \u2190\u221a\u03b1t\u22121\u02c6 x0 + \u221a1 \u2212\u03b1t\u22121\u02c6 \u03f5 // DDIM sampling 6: end for 7: return x0 Classifier guidance (Dhariwal & Nichol, 2021; Song et al., 2021b) integrates classifier predictions into the diffusion process, steering it to generate samples with specific characteristics, like a class label or measured signals y. The \u2018classifier\u2019 can be either, in a narrower sense, a classifier pre-trained on class-labeled noisy images, or a defined metric between measured signals and predicted signals. The process of classifier-guided deterministic DDIM sampling is formulated in Algorithm 1. The gradient scale \u03b3 regulates the balance between the fidelity of generated samples and their alignment with the classifier\u2019s predictions. For latent diffusion models, the deterministic DDIM sampler uses a pre-trained score function \u03f5\u03b8(zt, t) to predict \u02c6 z0 via \u02c6 z0 = \u221a\u03b1t\u22121   zt \u2212\u221a1 \u2212\u03b1t \u00b7 \u03f5(t) \u03b8 (zt) \u221a\u03b1t ! , (2) and deterministically sample zt from zt+1 via zt\u22121 = \u02c6 z0 + q 1 \u2212\u03b1t\u22121 \u2212\u03c32 t \u00b7 \u03f5(t) \u03b8 (zt). (3) 2.3. Latent diffusion model Large diffusion models, such as those proposed by Rombach et al. (2022); Ramesh et al. (2021); Saharia et al. (2022), have excelled in generating diverse and high-fidelity images. These models consist of three integral components: a text encoder, an image autoencoder, and a latent image generator. For instance, Stable Diffusion (SD) utilizes CLIP (Radford et al., 2021) as its text encoder, trains VQGAN (Esser et al., 2021) as its autoencoder and consequently a denoising UNet as the image generator on VQGAN\u2019s latent space. The structural composition of SD v1.5 is illustrated in Fig. 1. The implicit visual knowledge lies in the latent image generator. 3. Method 3.1. Undersampled MRI reconstruction We consider the following forward measurement model in MRI acquisition: k = MFxf + \u03f5, y = F\u22121k, (4) where xf \u2208Cn is the orginal fully-sampled image, y \u2208Cn is the undersampled (aliased) image, k \u2208Cn is the k-space measurement, and \u03f5 is the measurement noise. M \u2208{0, 1}n is the undersampling mask and (F, F\u22121) is the Fourier Transform pair. Undersampled MRI reconstruction aims at recovering high quality \u02c6 x from undersampled k, a dualdomain optimization formulation suggesting \u02c6 x = arg min x ||k \u2212MFx|| + \u03bb \u00b7 R(x), (5) where the first term enforces the fidelity of k-space measurement and R is some image regularizers. U2MRPD utilizes the image prior from the large text-toimage diffusion model as R and incorporates the k-space fidelity term in the diffusion process progressively by MRSampler. MRAdapter improves the accuracy of data fidelity term with any given MRI data. 3.2. MRSampler MRSampler is a physically consistent LDM solver for complex-valued MRI data. It prompts LDMs for generalizable undersampled MRI reconstruction in both MRI database-free and database-available scenarios. It is described formally in Algorithm 2. It builds upon a deterministic1 DDIM sampler, incorporating several novel phasemodulated control mechanisms to enhance its efficacy for 1Reason for this choice instead of a stochaitic one is provided in Appendix A.6.1. 3 \fTitle Suppressed Due to Excessive Size Algorithm 2 MRSampler Input: measured k-space k (the empty of which are filled with 0\u2019s), a sampling mask M. Require: an LLDM including an autoencoder (encoder E and decoder D), a deterministic DDIM Sampler S with a total timestep T=1000. Hyperparameters: SDEdit t0, watershed for two controllers tws, scale of classifier guidance \u03b3. 1: y = F\u22121(k) // an undersampled aliased image 2: z \u223cN(0, I) 3: zT0 = S.stochasticEncode(z, |y|, t0 \u00b7 T) 4: \u03b8 \u223cU(\u2212\u03c0, \u03c0) // pseudo-random phase generation 5: kP P M = F(|y|ej\u03b8) // modulated measured k-space 6: for t = t0 \u00b7 T \u22121 to 0 do 7: \u02c6 z0 = S.predict(zt+1, t + 1) // Eq.(2) 8: \u02c6 x0 = D(\u02c6 z0) 9: \u02c6 kP P M = F(\u02c6 x0ej\u03b8) 10: if t > tws \u00b7 T then 11: // hard data consistency 12: \u02c6 kdc = (1 \u2212M) \u00b7 \u02c6 kP P M + M \u00b7 kP P M 13: \u02c6 xdc = F\u22121(\u02c6 kdc), \u02c6 zdc 0 = E{|\u02c6 xdc|} 14: zt = S.sample(zt+1, t, \u02c6 zdc 0 , cg = 0) // Eq.(3) 15: else 16: // soft classifier guidance 17: zt = S.sample(zt+1, t, \u02c6 z0, cg = \u2212\u03b3 \u00b7 \u2207zt+1\u2225M \u00b7 kP P M \u2212M \u00b7 \u02c6 kP P M\u2225) 18: end if 19: end for 20: x0 = D(z0) 21: return x0 complex-valued MRI. Additionally, the implementation of an early-step-skipping(Meng et al., 2022; Chung et al., 2022b) technique combined with an interleaving strategy (Song et al., 2023a) enables an MRSampler to achieve a 5 \u00d7 acceleration, without compromising the quality of the reconstruction. Strong-to-mild diffusion control. We notice that the direct replacement in the measurement domains, i.e. hard data consistency (Kawar et al., 2022; Wang et al., 2023; Chung et al., 2023b; Song et al., 2023a), changes the generated MRI image drastically while soft gradient guidance (Song et al., 2021b; Chung et al., 2022a; 2023a; Rout et al., 2023; Song et al., 2023c) tends to perform mildly (Figure 2, panel (b); Figure 4, panel (b) and (c)). Observing the typical chaos-torefinement pattern in the guided diffusion processes (Figure 4, panel (b) to (d)), we propose a strong-to-mild diffusion control mechanism. Once attaining clean latent predictions \u02c6 z0 at each diffusion step and the pixel-domain counterpart \u02c6 x0 = D(\u02c6 z0), MRSampler enforces either hard or soft data consistency depending on the current timestep t and a hyperparameter, namely watershed timestep tws. Details about Figure 2. PPM\u2019s effect on MRSampler. Panel (a) shows the impact of PPM on both the image and Fourier domain (k-space) and panel (b) shows an intermediate hard data consistency step without and with PPM. the unitary controllers (soft and hard data consistency) can be found in the Appendix A.5.1. Pseudo-random phase modulation. Diffusion models are trained within noisy data landscapes, thus simplistic measurement constraints can inadvertently displace the inferential trajectory from its native data manifold (Chung et al., 2022a). We observe severe artifacts when directly applying a hard data consistency controller to MRI data, illustrated in Figure 4(d). This motivates us to modulate the intermediate images by a separable noising technique, namely Pseudorandom Phase Modulation (PPM). PPM modulates the input undersampled image y \u2208Cn and any intermediate clean MRI image \u02c6 x0 \u2208Rn by preserving its magnitude and bestowing it with a pseudo-random phase \u03b8 \u2208[\u2212\u03c0, \u03c0). The phase corruption does not obfuscate the essential structure of a complex-valued image as the corruption can be rectified by simply getting the magnitude. An illustration of PPM\u2019s influence on different controllers is provided in Figure 2. The undersampled PPM image yP P M = F\u22121MF(|yGT |ej\u03b8) is noisier than y = F\u22121MF(|yGT |) (Figure 2, panel (b)). Intuitively, PPM benefits the hard data consistency by hiding the aliasing artifacts under the noise (Figure 2, panel (a)). Furthermore, an image corrupted with high-frequency noise has Fourier features less imbalanced distribution (Figure 2, panel (b)) than raw k-space data with a high dynamic range (Mezrich, 1995). As a result, the constructed soft gradient term can concentrate less on the dominant low frequency and pay more attention to high-frequency reconstruction, reconstructing more details of aliased MRI (Figure 3, yellow 4 \fTitle Suppressed Due to Excessive Size and red boxes). 3.3. MRAdapter The rationale behind designing MRAdapter is that while providing any additional information related to the target task, i.e. MRI data for undersampled MRI reconstruction, an adaptable model\u2019s performance shall be improved anyway. The accuracy of the MRSampler is lowered by the autoencoder in LLDMs since both soft and hard constraints are on the pixel domain and pass through the lossy autoencoder. With any MRI data, the reconstruction performance of an LLDM\u2019s autoencoder can be improved by finetuning. As such, we propose an MRAdapter to gap the distribution mismatch between the training set of LLDM\u2019s autoencoders and MRI data. Specifically, we duplicate the magnitude of the MRI images three times as the input and ground truth for VQGAN and unfreeze the first and last set of convolutional layers of VQGAN and finetune it using the L2 loss. Intuitively, the MRAdapter adds two sets of filters to LLDMs: MRI-to-natural-images and natural-image-to-MRI at the start and the end of the autoencoder, respectively. By veiling data in the pixel domain and maintaining a naturalimage-resembled latent space, MRAdapter improves MRI generation and preserves the integrity of the generative image prior. Technically, our MRAdapter is characterized by three principal attributes:(1) It achieves computational efficiency by fine-tuning merely 0.2% of the entire diffusion model, as depicted in Figure 1(c). (2) It maintains the generative versatility of an LLDM. (3) It demonstrates robust performance in both specialized and universal MRI reconstruction tasks, as illustrated in Table 5. 4. Experiments In this section, we aim to answer the key question: can the implicit visual knowledge in a pre-trained LLDM on natural images be harnessed and enhance the (1) adaptability and (2) generalizability of undersampled MRI reconstruction models? We use the SD v1.5 checkpoint2 pre-trained on the LAION dataset (Schuhmann et al., 2022). We conduct comprehensive experiments on two commonly used public datasets of three MRI contrasts and two organs and ablate U2MRPD thoroughly. Additional details and results are in Appendix3. 4.1. Experimental Settings Datasets. We conduct experiments on one simulation dataset, the Multi-contrast IXI brain dataset4, and one real2https://huggingface.co/runwayml/stable-diffusion-v1-5 3More results are in Appendix A.1 and more details of datasets, baselines, and implementation are in Appendix A.2,A.3, and A.4. 4brain-development.org/ixi-dataset, CC BY-SA 3.0 license Table 2. Hyperparameters selection of different datasets. MRSampler MRAdapter t0 tws \u03b3 lr Adam Epoch FastMRI 0.4 0.3 1e\u22123 1e-04 (0.5,0.9) 10 IXI 0.4 0.3 5e\u22124 1e-04 (0.5,0.9) 10 collected dataset, the FastMRI single-coil knee5(Knoll et al., 2020; Zbontar et al., 2018) dataset. The first dataset contains coil-combined magnitude images of T1-, T2and PDweighted acquisition emulated as single-coil data with 576 overlapped subjects. The second dataset includes 1,172 raw k-space volumes of coronal proton density-weighted with or without fat suppression. We use all available subjects and follow the same pre-processing and dataset partition conventions as those in (Zhou et al., 2023) and (Peng et al., 2022; Chung & Ye, 2022; Guo et al., 2023; Knoll et al., 2020) for two datasets respectively. Baselines & Metrics. We employ six representative undersampled MRI reconstruction approaches from three categories as baselines, including (1) two model-based algorithms: Zero Filling and TV (Block et al., 2007), (2) two supervised DL methods: UNet (Zbontar et al., 2018) and DuDoRNet (Zhou & Zhou, 2020), and (3) two diffusion methods: Score-MRI (Chung & Ye, 2022; Chung et al., 2022b) and DiffuseRecon (Peng et al., 2022). We consider four Cartesian sampling patterns with different acceleration rates and use peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) for quantitative evaluation of MRI reconstruction. Implementation Details For our MRSampler, we tune the hyperparameters on 5 volumes in the corresponding training set of two datasets. For our MRAdapter, finetuning is carried out on the whole training set with the best hyperparameters on the validation set. Hyperparameters for all experiments presented are shown in Table 2. They are selected based on empirical findings from prior works (Meng et al., 2022; Esser et al., 2021; Song et al., 2021a) and primarily remain consistent across datasets. While crafting the initial noise potentially improves generation (Everaert et al., 2024), we keep a fixed seed in all experiments. 4.2. Main results In Table 3, we provide a structured quantitative comparison of undersampled MRI reconstruction models on both datasets under different acceleration rates of Uniform 1D sampling. It is divided into two categories: one compares database-free U2MRPD with dedicated models on FastMRI and the other, more challengingly, benchmarks U2MRPD against dedicated supervised models on every type of MRI. 5fastmri.med.nyu.edu 5 \fTitle Suppressed Due to Excessive Size In Figure 3, we visualize the reconstruction of complexvalued MRI data emphasizing diverse sampling masks. We provide a structured analysis of the main results below: \u2022 Generalizability towards MRI modalities and organs. U2MRPD finetuned with FastMRI generalizes the best towards unseen MRI modalities and organ (U2MRPD\u2019s consistently red results in Table 3, section 1, IXI) while supervised methods quickly falter with out-of-domain modalities or organs (the numerous gray values in Table 3, section 1, column IXI). \u2022 Generalizability towards MRI sampling patterns. Despite not being trained on them, U2MRPD is the most robust model under various sampling masks compared to other methods (Figure 3). Furthermore, U2MRPD surpasses all methods under high and outof-domain acceleration rates (U2MRPD\u2019s consistently red results of Table 3, section 1 and 2, column 12x). \u2022 Adaptability in database-free scenario. Databasefree U2MRPD surpasses all other supervised and diffusion methods in out-of-domain data (U2MRPD\u2019s consistently blue results of Table 3, column IXI). Furthermore, database-free U2MRPD stays on par with indomain diffusion models under low acceleration rates (Table 3, column FastMRI, 4x); surpasses them under medium acceleration rates (Table 3, column FastMRI, 8x); and surpasses all supervised and diffusion methods under high acceleration rates (U2MRPD\u2019s consistently blue results of Table 3, column 12x). \u2022 Adaptability in database-available scenarios. With single-source MRI data, U2MRPD\u2019s performance is consistently improved in all MRI data tested (comparison of two U2MRPDs\u2019 results in Table 3, section 1 and Table 5, section 2). With multiple-source MRI data, U2MRPD\u2019s results can be further boosted (Table 5, section 2, row BCD), indicating a promising new direction for universal undersampled MRI reconstruction (Liu et al., 2021) with U2MRPD. \u2022 Discussion: U2MRPD versus MRI diffusion models. U2MRPD uses an LLDM trained on billions of natural images and the latter are trained with MRI datasets. Despite relying solely on the magnitude component of complex MRI data, U2MRPD outperforms MRI diffusion methods using the full complex data: the results on IXI in Table 3 reveal that only our method and Score-MRI consistently outperform Zero Filling for undersampled brain MRI reconstruction and our method excels over Score-MRI by an average of 1.67 \u22125.78 dB in PSNR and 5.54 \u221215.10% in SSIM, particularly in out-of-domain modalities or organs; DiffuseRecon\u2019s official implementation fails to generalize towards images of different sizes, accounting for the N/A\u2019s in the Figure 3. Comparative visualizations of complex-valued MRI reconstructions using different sampling masks with identical acceleration factors, as depicted in (a). Panels (b) through (i) display the outcomes of various reconstruction models. The red dashed box on the 2nd row shows a sub-figure. In each sub-figure, the magnitudes of each reconstructed image (left) and its residual difference from the reference (right) are shown within each sub-panel, employing a consistent colormap for error visualization throughout. PSNR and SSIM are notated at the top of each reconstructed image. Red and yellow boxes highlight two challenging details. Panel (j) presents the magnitude of the reference image sourced from the FastMRI knee dataset alongside the color bar. Refer to Figure 5 for results with more masks and Figure 6 for results on the Multi-contrast IXI brain dataset in Appendix. 6 \fTitle Suppressed Due to Excessive Size Table 3. Quantitative metrics of undersampled MRI reconstruction with various acceleration rates and datasets. \u2020 means models use complex-valued data and \u2217means models use magnitude-only data. In each section, the best results are colored red and the second best results in the first section are colored blue. Metrics lower than Zero Filling\u2019s are colored gray. Numbers are presented as {mean value} with scriptsize\u2019s {unit standard deviation}. FastMRI single-coil knee dataset Multi-contrast IXI brain dataset \u2191PSNR(dB) 5145 slices of (PD, Knee) 912 slices of (T1, Brain) 700 slices of (T2, Brain) 700 slices of (PD, Brain) Acceleration rate 4x 8x 12x 4x 8x 12x 4x 8x 12x 4x 8x 12x Zero Filling 28.82 2.03 28.20 2.05 27.90 2.05 25.56 1.98 24.67 1.99 24.59 2.00 24.85 1.18 23.93 1.18 23.86 1.18 24.68 1.36 23.78 1.36 23.72 1.35 TV\u2020 25.35 1.95 24.96 1.92 24.67 1.90 26.70 2.27 24.92 2.04 24.83 2.06 25.99 1.32 24.19 1.20 24.11 1.19 25.88 1.54 24.07 1.40 24.02 1.38 (1) Database-free and adapted U2MRPD on FastMRI v.s. dedicated models on FastMRI. UNet\u2020 30.71 2.30 29.18 2.26 27.56 2.17 25.91 2.12 24.33 1.82 24.18 1.84 25.60 1.18 23.75 1.13 24.18 1.84 25.33 1.34 23.54 1.26 23.34 1.34 DuDoRNet\u2020 32.94 3.09 30.87 2.67 28.94 2.30 25.33 1.71 23.49 1.74 22.98 1.85 25.40 1.20 23.09 1.15 22.75 1.15 24.35 1.40 22.10 1.50 21.66 1.52 DiffuseRecon\u2020 31.38 3.74 29.51 3.06 28.56 2.82 N/A N/A N/A N/A N/A N/A N/A N/A N/A Score-MRI\u2020 30.33 2.44 28.81 2.21 28.24 2.07 29.21 2.33 25.52 1.94 25.03 1.94 28.78 1.40 24.72 1.22 24.32 1.18 28.64 1.60 24.74 1.35 24.32 1.31 U2MRPD\u2217(Ours w/o MRAdapter) 30.88 1.63 30.26 1.63 29.64 2.60 30.63 2.40 28.54 2.31 27.32 2.22 30.57 1.40 28.61 1.36 27.08 1.31 31.37 1.85 29.28 1.78 28.08 1.79 MRPD\u2217(Ours) 31.50 1.74 31.28 1.76 30.95 3.03 30.88 1.43 30.56 1.51 29.87 1.53 30.61 1.17 30.42 1.41 29.81 1.18 31.42 1.70 31.25 1.80 30.60 1.33 (2) Adapted U2MRPD v.s. dedicated supervised models. UNet\u2217\u2020 30.71 2.30 29.18 2.26 27.56 2.17 28.75 1.87 25.74 1.74 25.14 1.80 27.76 1.41 24.54 1.32 23.91 1.28 27.67 1.55 24.13 1.40 23.70 1.40 DuDoRNet\u2217\u2020 32.94 3.09 30.87 2.67 28.94 2.30 32.42 2.67 28.05 2.12 26.87 2.12 34.21 1.96 28.03 1.54 25.33 1.35 35.05 2.69 28.44 1.90 26.84 1.52 U2MRPD\u2217(Ours) 31.50 1.74 31.28 1.76 30.95 1.82 31.62 1.54 30.61 1.38 30.10 1.55 31.68 1.23 31.10 1.25 30.30 1.24 32.48 1.38 31.95 1.40 31.17 1.40 \u2191SSIM(%) 5145 slices of (PD, Knee) 912 slices of (T1, Brain) 700 slices of (T2, Brain) 700 slices of (PD, Brain) Acceleration rate 4x 8x 12x 4x 8x 12x 4x 8x 12x 4x 8x 12x Zero Filling 85.12 5.35 83.23 6.54 82.32 7.05 75.86 5.72 75.36 6.87 75.37 7.03 77.40 2.87 77.30 2.87 77.26 2.72 86.73 2.36 77.00 3.36 77.02 3.42 TV\u2020 71.44 10.08 64.34 14.36 61.90 15.02 71.76 7.52 66.76 9.54 67.11 9.71 75.79 2.53 72.53 2.86 72.82 2.89 74.86 2.68 71.35 2.78 71.80 2.83 (1) Database-free and adapted U2MRPD on FastMRI v.s. dedicated models on FastMRI. UNet\u2020 88.99 5.07 85.63 6.53 82.82 6.96 73.55 5.68 75.01 7.07 74.94 7.30 80.54 2.44 79.23 2.43 74.94 7.30 78.93 2.87 77.61 3.14 78.94 2.87 DuDoRNet\u2020 90.94 5.48 87.29 7.04 84.41 7.44 72.77 5.76 70.85 7.87 70.15 9.04 77.31 3.55 75.98 3.83 76.63 2.80 72.29 4.10 70.83 4.72 70.37 4.73 DiffuseRecon\u2020 82.05 3.17 76.23 11.92 72.65 13.12 N/A N/A N/A N/A N/A N/A N/A N/A N/A Score-MRI\u2020 87.39 6.03 84.23 7.01 83.06 7.33 86.99 2.92 78.54 4.72 76.66 4.96 88.61 1.52 80.27 1.93 78.62 2.02 87.64 1.93 79.96 2.21 78.34 2.27 U2MRPD\u2217(Ours w/o MRAdapter) 85.62 2.87 83.97 2.90 82.51 8.65 92.07 2.90 87.16 4.01 84.18 4.74 93.69 1.60 89.48 2.35 86.53 2.52 94.06 1.67 89.77 2.11 87.48 2.47 U2MRPD\u2217(Ours) 86.60 2.89 85.98 2.92 85.41 8.86 93.38 7.22 92.47 16.05 91.33 17.34 94.15 11.91 93.56 12.86 92.93 13.52 94.74 11.46 94.15 12.16 93.44 12.72 (2) Adapted U2MRPD v.s. dedicated supervised models. UNet\u2217\u2020 88.99 5.07 85.63 6.53 82.82 6.96 88.06 3.09 85.38 2.05 83.22 4.73 87.97 2.18 85.38 2.05 83.73 2.11 87.59 2.48 85.27 2.39 84.26 2.39 DuDoRNet\u2217\u2020 90.94 5.48 87.29 7.04 84.41 7.44 87.94 5.41 84.41 4.74 85.13 5.09 97.25 1.02 92.11 2.13 86.91 2.13 97.41 1.08 92.40 2.18 90.27 2.43 U2MRPD (Ours)\u2217 86.60 8.35 85.98 8.56 85.41 8.86 93.91 3.83 94.15 2.66 91.67 4.16 94.94 3.44 94.18 3.61 93.38 3.72 95.45 3.35 94.75 3.47 93.92 3.59 table. Additionally, U2MRPD requires significantly less (or no if database-free) fine-tuning time6 than training an MRI diffusion model. Furthermore, both DiffuseRecon and Score-MRI struggle to maintain consistent image intensity under the Gaussian 2D masks, reflecting the prevalent color drifting issue of diffusion models (Lin et al., 2023). This problem is noticeable in Figure 3 and U2MRPD circumvents it with our designed phase-modulated control. 4.3. Ablation studies Design of MRSampler. We present an empirical evaluation of the MRI reconstruction process using MRSampler and various control mechanisms, either unitary or magnitudeonly, in Figure 4. Strong-to-mild control allows MRSampler to bypass local minima associated with hard consistency terms, resulting in faster convergence and superior performance compared to soft consistency terms alone. Psuedorandom Phase Modulation (PPM) improves the performance of both soft and hard control. Design of MRAdapter. We dissect the MRAdapter\u2019s architecture to ascertain the impact of selective tuning on its performance. We experiment with different configurations on the fastMRI knee dataset, specifically tuning the encoder and decoder components in isolation. In Table 4, we show 6Details can be found in Appendix A.6.3. Figure 4. Image reconstruction evolution during the guided diffusion process of our MRSampler and other controllers using unitary or magnitude-only constraints. (a) The evolution of PSNR and SSIM. (b)-(d) In each panel, the first row illustrates the evolution of clean image predictions \u02c6 x0(zt), while the second row depicts that of intermediate noisy images zt. 7 \fTitle Suppressed Due to Excessive Size Table 4. Ablation study on the design of our MRAdapter on the FastMRI knee dataset. Arch. 4x 8x PSNR (dB) SSIM (%) PSNR (dB) SSIM (%) Ori. VQGAN 30.88 1.63 85.62 8.23 30.26 1.63 83.97 8.41 Tune encoder 31.35 1.69 86.59 2.88 31.23 1.74 86.09 2.92 Tune decoder 31.31s 1.68 86.39 2.90 30.85 1.68 85.22 2.91 Ours 31.50 1.74 86.60 2.89 31.28 2.91 85.98 8.56 that while both strategies\u2014tuning the encoder alone and tuning the decoder alone\u2014yield improvements in PSNR and SSIM metrics, our integrated MRAdapter configuration achieves the highest performance. With singleor multiple-source MRI datasets: towards universal MRI reconstruction. We fine-tune VQGAN with different MRI datasets using our MRAdapter and validate our dual capacity to adapt to different data acquisition scenarios and to generalize to unseen data. Quantitative results of both the VQGAN reconstruction of fully-sampled MRI data and U2MRPD\u2019s reconstruction of undersampled MRI with 8\u00d7 a uniform 1D sampling mask are listed in Table 5. Direct conclusions from the table can be summarized as: (1) with singleor multiple-source MRI datasets, VQGAN\u2019s MRI reconstruction ability can be improved on all tested datasets, no matter in-domain or out-of-domain (Table 5, section 1, row None versus the other rows). (2) VQGAN\u2019s reconstruction ability can be boosted to a larger extent with MRI datasets closer to testing sets (Table 5, section 1, the diagonal) or larger (Table 5, section 1, row A and row BCD versus row B, row C, and row D) than outof-domain datasets or smaller datasets. (3) With singleor multiple-source MRI data, U2MRPD\u2019s reconstruction ability can be improved on all tested MRI datasets, no matter in-domain or out-of-domain (Table 5, section 2, row None versus the other rows). (4) U2MRPD\u2019s reconstruction ability can be boosted to a larger extent with MRI datasets closer to testing sets (Table 5, section 2, the diagonal) or larger (Table 5, section 2, row A and row BCD versus row B, row C, and row D) than out-of-domain datasets or smaller datasets. Combining two sections of Table 5, we observe a counterintuitive fact that the reconstruction of the ground truth MRI data is similar to or even worse than that of 8\u00d7 undersampled MRI (Table 5, differences between the same rows in section 1 and 2). It implies that lossy as VQGAN is, our MRSampler (Table 5, row None in section 1 and 2) and MRAdapter (Table 5, row A to BCD in section 1 and 2) can compensate for the information loss. The rationale behind this is that the classifier guidance (the soft DC term) corrects the reconstruction result on the pixel domain directly. The computed gradients can guide the latent-domain generator to generating pixel-domain consistent results. Universal MRI models that reconstruct MR images across different MRI\u2019s are highly desirable for efficient deployment on MRI machines (Liu et al., 2021). We notice U2MRPD\u2019s potential and provide more discussion in Appendix A.6.4. 5."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.10611v2",
+                    "title": "Rethinking Dual-Domain Undersampled MRI reconstruction: domain-specific design from the perspective of the receptive field",
+                    "abstract": "Undersampled MRI reconstruction is crucial for accelerating clinical\nscanning. Dual-domain reconstruction network is performant among SoTA deep\nlearning methods. In this paper, we rethink dual-domain model design from the\nperspective of the receptive field, which is needed for image recovery and\nK-space interpolation problems. Further, we introduce domain-specific modules\nfor dual-domain reconstruction, namely k-space global initialization and\nimage-domain parallel local detail enhancement. We evaluate our modules by\ntranslating a SoTA method DuDoRNet under different conventions of MRI\nreconstruction including image-domain, dual-domain, and reference-guided\nreconstruction on the public IXI dataset. Our model DuDoRNet+ achieves\nsignificant improvements over competing deep learning methods.",
+                    "authors": "Ziqi Gao, S. Kevin Zhou",
+                    "published": "2023-03-19",
+                    "updated": "2024-02-15",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "main_content": "INTRODUCTION Magnetic resonance imaging (MRI) is a non-invasive and flexible imaging modality widely used in clinical practice. Complete K-space measurements lead to unbearable acquisition time while fewer measurements lead to aliasing and blurring in the image. Under-sampled MRI reconstruction aims to reconstruct the high-quality, clean MRI image from its low-quality, aliased counterpart. Previously, Compressed Sensing (CS) and Parallel Imaging (PI) accelerated MRI reconstruction 2-3 times. Since the revolutionary work [1], convolutional neural networks (CNN) have become the primary workhorse for under-sampled MRI reconstruction. As the success of Transformer [2] is now indisputable in computer vision, Transformer has shown great potential for undersampled MRI reconstruction as well [3, 4, 5, 6, 7]. Many methods [8, 9, 10] focus on adapting novel architecture designs of image-domain neural networks. Customizing conventional CNNs or ViT to MRI further benefits MRI reconstruction, including K-space data consistency (DC) [11, 12], dual-domain recurrent learning [13, 14], over-complete representation [15]. SwinMR [3] and DSFormer [4] pioneer Swin transformer [16] as a strong backbone in single-contrast and multi-contrast MRI reconstruction, respectively. ReconFormer [5] develops a recurrent pyramid Transformer. Feng et al. [6] proposed a task transformer network for joint MRI reconstruction and super-resolution. Among all SoTA methods, dual-domain networks show strong performance by considering MRI reconstruction in both image and K-space domains. DuDoRNet [14] presents a dual-domain [13] recurrent learning strategy and used a dilated residual dense CNN to eliminate non-local aliasing artifact in the image domain. DuDoCAF [7] leverages the query-key mechanism of MSA. Yet their models use the same networks for both domains, neglecting the distinct properties of image recovery and k-space recovery problem. In our study, (1) we analyze the feasibility of K-space interpolation from the perceptive of the receptive field; (2) we propose two novel domain-specific modules, including image-domain parallel local detail enhancement and k-space global initiation; and (3) extensive experiments on the public IXI-dataset validate the superiority of our model, DuDoRNet+, under multiple settings of MRI reconstruction including image-domain, dual-domain, and dual-domain undersampled MRI reconstruction guided by a reference protocol. 2. ANALYSIS AND METHOD 2.1. Problem Formulation Let ku \u2208Cmn and kf \u2208Cmn be the under-sampled and fully-sampled k-space signal respectively; iu \u2208Cmn, if \u2208 Cmn, and ir \u2208Cmn be the under-sampled, fully-sampled and reconstructed image signal respectively; M \u2208Rmn be the binary k-space mask for acceleration. The under-sampled MRI reconstruction can be formulated as an image recovery problem [1, 15, 17, 18, 3] with K-space arXiv:2303.10611v2  [eess.IV]  15 Feb 2024 \fDC as a regularisation term [11]: arg min \u03b8i \u0010 \u2225if \u2212Pi (iu; \u03b8i)\u22252 2 +\u03bb \u2225ku \u2212M \u2299F (Pi (iu; \u03b8i))\u22252 2 \u0011 , (1) where F is 2D discrete Fourier Transform and the approximation function Px(\u00b7; \u03b8x) is used to predict a reconstructed signal xr given its parameter \u03b8x and any under-sampled input. A few works [14, 13, 7] leverage the relationship between image and k-space domain using Fourier Transform pairs (F, F\u22121) and transform MRI reconstruction into a multivariable optimization problem described as arg min \u03b8i,\u03b8k \u0010 \u2225kf \u2212Pk (F (Pi (iu; \u03b8i)) ; \u03b8k)\u22252 2 + \r \rif \u2212Pi \u0000F\u22121 (Pk (ku; \u03b8k)) ; \u03b8i \u0001\r \r2 2 + \u03bb \r \rku \u2212M \u2299F \u0000Pi \u0000F\u22121 (Pk (ku; \u03b8k)) ; \u03b8i \u0001\u0001\r \r2 2 \u0011 , (2) and solve it using a dual-domain recurrent learning strategy [14]. Yet most of existing works design the approximation function Px(\u00b7; \u03b8x) using neural networks whose architectures mainly concentrate on the image recovery problem, neglecting the distinct properties of k-space recovery problem. We argue that a hand-crafted design on dual-domain approximation functions is crucial for solving this multivariable optimization problem since image and k-space bear different properties and the tasks in two spaces shall be distinct further. 2.2. Solving K-space recovery problem from the perspective of the receptive field Organs scanned by MRI (e.g brains, knees) are usually rich in contextual details, where a larger receptive field can improve the image reconstruction performance in general. On the other hand, undersampled MRI reconstruction is a K-space recovery problem and has been considered as a local interpolation in many CS algorithms [19]. However, when MRI is under-sampled with a higher rate, the average number of measurements in a fixed local region decreases. Consider a set of K-space measurements after accelerating MRI with rate a using 1D Cartesian sampling pattern. Conventionally, a central fraction of K-space measurements called auto-calibration region is fully sampled, accounting for a fraction of Racs; while the other measurements are sampled randomly. For the non-calibration measurements, the feasibility of interpolation depends on the number of k-space scanning lines in a reachable local region K of size k \u2217k, in which the minimal number of data points is two. The probability P of a feasible interpolation is P = 1 \u2212(p)k \u2212Ck 1 (p)k\u22121(1 \u2212p), (3) Fig. 1. Probability of a feasible interpolation P w.r.t. the size of receptive field and acceleration rate a when Racs=0.125. where p is the possibility that none of the K-space lines are measured, determined by Racs and a: p = 1 \u2212 1 a \u2212Racs 1 \u2212Racs . (4) An illustration of the distribution of P w.r.t (K, a) given a fixed Racs is shown in figure 1. When accelerating with high rate (> 4), the probability of feasible interpolation is lower than 50% with , not to say a successful interpolation. Using ConvNets for MRI reconstruction [15] [14] can achieve large receptive field by upsampling, dilation convolution or large kernels. However, dilation convolution does not lead to an increase of lower bound for doing interpolation in k-space since dilation increases receptive field with aligned atrous part. Upsampling heavily burdens the GPU. Also, piling up large convolution kernels harms model convergence. 2.3. Method We choose a SoTA dual-domain MRI reconstruction model DuDoRNet [14] for evaluating our domain-specific designs. Our designs improve its dual domain recurrent blocks by customizing global and local structures based on domain-specific properties. Our model, DuDoRNet+, is illustrated in Figure 2, including domain-specific Shallow Feature Extraction (XSFE), Global Feature Refinement(GFR), and 4-stage domainspecific building blocks (X-Si) as the backbone. Global residual learning and global feature fusion in DuDoRNet are preserved in ours. The overall pipeline goes as follows: F\u22121 = Conv3(xu), F C 0 = Conv3(F\u22121). (5) where Convk denotes a convolution operation with kernel size k \u2217k and F\u22121 denotes the first extracted feature used for global residual learning. F C 0 denotes the intermediate extracted feature by convolution. The second extracted feature F0 is the input of X-Si. For image domain, F0 = F C 0 . \fX-SFE Undersampled I/K XS1 XS2 XS3 XS4 G F F GFR + Reconstructed I/K DRDB + I-PLDE I-Si Domain-specific Color Table K-SFE \ud835\udc65\ud835\udc62 Conv K-GLIM Conv \ud835\udc390 \ud835\udc65\ud835\udc62 \ud835\udc39\u22121 \ud835\udc39\u22121 \ud835\udc390 \ud835\udc39 1 \ud835\udc392 \ud835\udc393 \ud835\udc39 4 \ud835\udc39\ud835\udc3a\ud835\udc39\ud835\udc39 \ud835\udc65\ud835\udc5f \ud835\udc39\u22121 \ud835\udc39\ud835\udc56\u22121 \ud835\udc39\ud835\udc56 \ud835\udc39\ud835\udc56 \ud835\udc36\ud835\udc38 I-PLDE \ud835\udc39\ud835\udc56\u22121 \ud835\udc39\ud835\udc56 \ud835\udc36\ud835\udc38 1x1Conv SiLU DConv SiLU DConv SiLU DConv LayerNorm CW-MSA + + FFN LayerNorm \ud835\udc390 \ud835\udc390 \ud835\udc36 \ud835\udc390 \ud835\udc36 K-GLIM Domain-agnostic Fig. 2. Framework of a recurrent block in DuDoRNet+. It is constructed by a domain-specific Shallow Feature Extraction (X-SFE), Global Feature Refinement(GFR) and 4-stage domain-specific building blocks(X-Si). Building blocks\u2019 color follow the notation in the color table. Convolution is followed by ReLU unless noted. For K-space, a Global Initiation Module (K-GLIM) is introduced based on the idea of a sketchy K-space initiation: F0 = PK\u2212GLIM(F C 0 ) = PK\u2212GLIM(Conv3(F\u22121)). (6) Global Feature Fusion (GFF) fuses features from X-BB\u2019s and is input for global residual learning: Fi = PX\u2212Si(Fi\u22121), (7) xr = PGF R(F\u22121 + PGF F (concat(F1, F2, F3, F4))), (8) where GFF consists of 1x1 and 3x3 convolutions and GFR consists of two 3x3 convolutions, creating the refined reconstructed image/K-space xr. 2.3.1. K-GLIM: K-space global initiation We propose K-space global initialization module (K-GLIM) placed at the beginning of the K-space network. K-GLIM is applied on the transformed features of two convolution layers in SFE, providing channel-wise interaction and a global view. The main part of K-GLIM is a Channel-wise Multi-head Self-attention (C-MSA). Instead of performing pixel-level or patch-level attention in a conventional spatial attention way, C-MSA is performed on the transpose of pixel-level tokens. Similar to MSA, C-MSA is an extension of Self-Attention (CSA) in which h times SA operations are done. C-SA can be expressed as zc j = X i Softmax \u0012QT K \u221a d \u0013 i V T i,j, (9) and its computational complexity is O(6(hw)C2), linear to (hw). C-MSA naturally captures global information and interactions for visual recognition tasks and complements (windowed-)spatial attention. Compared with channel-wise convolution, it is data-specific and fine-grained. 2.3.2. I-PLDE: Image parallel local detail enhancement Recovering high-quality images from undersampled MRI measurements can be divided into two sub-tasks in the image domain: artifact removal and detail refinement. The former requires large receptive field due to the global property of signal aliasing, yet the latter task is much more a local task. Since the receptive field of a network enlarges with stacked convolutions, the detail refinement ability is possibly harmed with features propagating. To this end, we propose the I-PLDE module, a parallel branch emphasizing local detail, inspired by the \u201cdivide-andconquer\u201d idea in [20, 21]: modeling locality and long-range dependencies in parallel and fusing the features to account for both. Depthwise convolution .Different from ViTAE, we adopt stage-level parallel connection and use sequential dilated convolution (DRDB) for pyramid feature extraction instead of multiple parallel dilated convolution of different dilation rates. This design is customized to tiny models since a wide network design leads to a less stable training process and lower feature representation ability. I-PLDE consists of a 1x1 convolution to match the hidden dimension with its parallel branch, three stacked depth-wise convolution layers, and a window embedding operation. SiLU is used for non-linear activation following the convention in [20, 21]. The output of I-PLDE F CE i is added with DRDN\u2019s output. ViTAE\u2019s design is not successful in our scenarios since the tiniest model it constructs has about 5 times the parameters of ours and different availability of large datasets. 3. EXPERIMENT 3.1. Dataset and training Our evaluation is carried out on the Multi-Contrast IXI dataset1. We use all 575 subjects with paired T2-PD and uniformly sample 14 slices from each subject volume. We split the dataset patient-wise into training, validation and testing set with a ratio of 7 : 1 : 2, corresponding to 5628 training images, 812 validation images, and 1610 test images each protocol. Images are center-cropped from 256x256 to 224x224. Code is written in Pytorch and experiments are performed using an NVIDIA GeForce RTX 3090. As for the rest, we follow the same experiment settings in DuDoRNet [14]. 3.2. Performance evaluation We compare our methods with other baseline deep learning methods in three conventions of MRI reconstruction: image-domain [22, 14, 23, 15, 3], dual-domain [22, 14, 4] and reference-protocol-guided dual-domain reconstruction [24, 23, 14, 4, 7]. Considering backbone design, [23, 24] adopt Dense-Unet; [4, 7, 3] share similar Swin-Transformer 1https://brain-development.org/ixi-dataset/, CC BY-SA 3.0 license \fAcceleration 4x 8x PSNR SSIM MSE PSNR SSIM MSE Zero Padding 25.17\u00b11.80 80.20\u00b16.34 223.51 24.16\u00b11.79 79.90\u00b16.37 282.28 Unet 31.73\u00b11.98 95.48\u00b12.12 53.03 26.98\u00b12.01 91.43\u00b13.75 157.41 Dense-Unet 32.49\u00b12.23 96.47\u00b11.79 46.10 27.70\u00b11.97 92.16\u00b13.50 132.27 OUCR 32.52\u00b12.26 96.57\u00b11.76 45.90 27.80\u00b12.00 92.33\u00b13.44 129.40 SwinIR 32.75\u00b12.27 96.69\u00b11.71 43.60 27.78\u00b12.08 92.49\u00b13.45 130.76 DuDoRNet I 32.95\u00b12.28 96.75\u00b11.69 41.82 28.00\u00b12.04 92.76\u00b13.32 123.94 Ours I 33.01\u00b11.98 96.80\u00b11.59 40.36 28.02\u00b12.04 92.80\u00b13.36 122.78 Dual-DenseUnet 32.76\u00b12.22 96.67\u00b11.65 43.21 27.68\u00b11.93 91.99\u00b13.45 131.75 Dual-SwinIR 33.37\u00b12.44 97.09\u00b11.57 38.86 27.73\u00b12.02 92.16\u00b13.50 131.70 DuDoRNet(w/ ref) 33.00\u00b12.31 96.81\u00b11.63 41.46 27.84\u00b12.05 92.24\u00b13.46 128.73 Ours (w/ ref) 33.43\u00b12.40 97.02\u00b11.59 35.84 28.13\u00b12.10 92.68\u00b13.41 125.27 DuDoRNet 40.45\u00b12.59 99.26\u00b10.51 7.86 38.38\u00b12.52 98.93\u00b10.72 12.52 Ours 40.59\u00b12.55 99.29\u00b10.50 7.61 38.46\u00b12.52 98.96\u00b10.69 12.00 Table 1. Quantitative comparison (PSNR(dB), SSIM(%), MSE(*1e-5)) with baseline methods on the IXI-dataset. The first, second and third part correspond to image-domain reconstruction of PD, dual-domain reconstruction of PD and dual-domain reconstruction of PD with a reference protocol T2. The best results are marked as red. backbones derived from SwinIR [16]. All models besides UNet [22] have \u22481M parameters and UNet has 2M. All models are recurred twice and a DC is added at the end of each recurrent block. The recurrent time of local residual blocks in OUCR [15] is set to 5, complying with their default setting; yet DCs at the end of their local recurrent structure are discarded while those at the end of their global recurrent structure are preserved for fairness. All models are trained for 100 epochs and the hyper-parameters of each method are tuned on the validation set with test data held out for final evaluation. We consider 1D Cartesian sampling pattern with an acceleration rate ranging from 4 to 8; the center sampling fraction Racs is set to 0.125. Peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are used as the quantitative evaluation metrics. 3.3. Results on undersampled MRI reconstruction In Table 1, we demonstrated PD reconstruction evaluations using \u00d74, \u00d78 acceleration in three common approaches of MRI reconstruction: image-domain, dual-domain, and T2guided dual-domain MRI reconstruction. The best results under the same setting and acceleration rate are colored with red. Our method achieves the best performances under all settings and acceleration rates. 3.4. Ablation Study We examine I-PLDE and K-GLIM on undersampled PD reconstruction without reference guidance. For each component, we apply it to image-domain and dual-domain subsequently. If there is a performance drop in either way, we apply it on K-space only to observe the performance. The reason for 4x PD Domain(s) to put the module Modules Image K-space Both K-GLIM I-LDE PSNR, SSIM PSNR,SSIM PSNR, SSIM DuDoRNet 33.00, 96.81 DuDoRNet w/ \u2713 32.89, 96.57 33.37, 96.87 33.34, 96.84 DuDoRNet w/ \u2713 33.16,96.83 32.76, 96.79 33.11, 96.83 Ours \u2713 \u2713 33.43,97.02 Table 2. Domain-wise quantitative evaluation (PSNR, SSIM(%)) of hybrid structure and domain-specific modules. The best results are marked as red. this design is that we weigh the utility of the image recovery network and the synergy between image and K-space reconstruction networks over K-space reconstruction since the image recovery network alone gives better results than K-space recovery network thus a default choice when reconstructing in a single domain. As shown in Table 2, applying them to the other domain leads to performance drops. 4."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2211.16771v1",
+                    "title": "Handling Missing Data via Max-Entropy Regularized Graph Autoencoder",
+                    "abstract": "Graph neural networks (GNNs) are popular weapons for modeling relational\ndata. Existing GNNs are not specified for attribute-incomplete graphs, making\nmissing attribute imputation a burning issue. Until recently, many works notice\nthat GNNs are coupled with spectral concentration, which means the spectrum\nobtained by GNNs concentrates on a local part in spectral domain, e.g.,\nlow-frequency due to oversmoothing issue. As a consequence, GNNs may be\nseriously flawed for reconstructing graph attributes as graph spectral\nconcentration tends to cause a low imputation precision. In this work, we\npresent a regularized graph autoencoder for graph attribute imputation, named\nMEGAE, which aims at mitigating spectral concentration problem by maximizing\nthe graph spectral entropy. Notably, we first present the method for estimating\ngraph spectral entropy without the eigen-decomposition of Laplacian matrix and\nprovide the theoretical upper error bound. A maximum entropy regularization\nthen acts in the latent space, which directly increases the graph spectral\nentropy. Extensive experiments show that MEGAE outperforms all the other\nstate-of-the-art imputation methods on a variety of benchmark datasets.",
+                    "authors": "Ziqi Gao, Yifan Niu, Jiashun Cheng, Jianheng Tang, Tingyang Xu, Peilin Zhao, Lanqing Li, Fugee Tsung, Jia Li",
+                    "published": "2022-11-30",
+                    "updated": "2022-11-30",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG"
+                    ],
+                    "main_content": "Introduction Graph attribute missing is ubiquitous due to messy collection and recording process (Yoon, Davtyan, and van der Schaar 2016; Roth 1994). In a biochemical scenario, for instance, data is missed due to the dif\ufb01culties in measuring or calculating quantitative molecular properties (e.g., HOMO and LUMO orbital energy) at the atomic level (Yomogida et al. 2012; Salomon et al. 2003), attenuating the graph or node representations when we introduce a graph model for molecules. For graph learning tasks, even a small fraction of missing attributes will potentially interfere with the performance, leading to biased inferences (Sterne et al. 2009; Wooldridge 2007). Graph Neural Networks (GNNs) have demonstrated powerful ability on various graph-related tasks. Most GNNs consist of two parts: feature propagation and transformation. The former is the key to the success of GNNs as it can yield suitable spectral responses for graph desires (Li et al. 2022). For example, graph convolutional networks (GCN) (Kipf and Welling *Corresponding author. Copyright \u00a9 2023, Association for the Advancement of Arti\ufb01cial Intelligence (www.aaai.org). All rights reserved. Figure 1: Spectral visualization of GAE. Red histograms show the graph spectra. After GAE encoder (GCN), spectral concentration appears in the latent space. Upon MLP decoder, the spectral concentration is not alleviated. 2017) applies a recursive aggregation mechanism for updating node representations, leading to the low-frequency concentration from a spectral view (Wu et al. 2019; Donnat et al. 2018; Tang et al. 2022). The mechanism makes GCN effective for numerous real-world applications as low-frequency concentration promotes the convergence of representations of adjacent nodes (Nt and Maehara 2019; Xu et al. 2020). However, this spectral concentration diminishes partial eigenvalues, i.e., spectral values. As a consequence, it causes serious performance degradation on the standard graph attribute imputation models such as graph autoencoders (GAEs) (Li et al. 2020; Park et al. 2019; Li et al. 2021). In Figure 1, we visualize the graph spectra of the input, latent and output based on GAE (Kipf and Welling 2016), which is the backbone of most GAEs. As the \ufb01gure makes clear, existing GAEs leave suf\ufb01cient room for improvement to combat the loss of spectral components. Delving into graph attribute reconstruction, a natural question arises: how can GAEs impute with alleviated spectral concentration? In this work, inspired by the maximum entropy principle (Narayan and Nityananda 1986; Burch, Gull, and Skilling 1983; Frieden 1972) that is capable to alleviate concentrated data distributions, we propose to maximize the graph spectral entropy for mitigating spectral concentration while imputaarXiv:2211.16771v1  [cs.LG]  30 Nov 2022 \ftion. We present a novel method for graph attribute imputation called Max-Entropy Graph AutoEncoder (MEGAE). Unlike existing imputation methods, our model trains a GAE with the regularization for maximizing the graph spectral entropy. However, when it comes to the computation of graph spectral entropy, it becomes challenging due to large time complexity of the eigen-decomposition for Laplacian matrix (Cai, He, and Han 2007a,b; Jia et al. 2016). In this regard, we design the tight wavelet kernels (Shuman et al. 2015; Leonardi and Van De Ville 2013; Ahmad and Sheikh 2018) in MEGAE to encode features into the spectral domain involving several spectral bands. We then theoretically prove that the proposed wavelet paradigm can well approximate the actual graph spectral entropy by a predictable upper error bound without eigen-decomposition. Upon entropy regularized latent representations, we use symmetric wavelet decoder to perform data deconvolution for reconstructing graph attributes. In the experiments, we proceed with an empirical evaluation of MEGAE on singleand multi-graph datasets. Firstly, we show that MEGAE outperforms 16 state-of-the-art methods on attribute imputation tasks, including commonly used classical methods and graph learning based imputation methods. Additional downstream experiments of graph and node classi\ufb01cations demonstrate that imputed graphs by MEGAE can obtain the best accuracy performance when compared to state-of-the-art imputation models. Contributions. We summarize our contributions below: \u2022 We propose to maximize the graph spectral entropy to overcome the spectral concentration issue in GAEs for graph attribute reconstruction. \u2022 We present Max-Entropy Graph AutoEncoder (MEGAE) for encoding graphs into spectral domain and maximizing the spectral entropy in the latent space. \u2022 We develop an ef\ufb01cient method for approximating the actual graph spectral entropy without eigen-decomposition, and importantly, provide the theoretical upper error bound. 2 Related Work Matrix Completion. Missing data is a widely researched topic. Matrix completion methods can be applied to impute graph attributes without using the graph structures. Most proposed methods impute with a joint distribution on the incomplete data. For instance, the joint modeling methods impute by drawing from the predictive distribution, including Bayesian strategies (Murray and Reiter 2016), matrix completion methods (Cand\u00e8s and Recht 2009; Hastie et al. 2015), and Generative Adversarial Networks (Yoon, Jordon, and Schaar 2018; Yoon and Sull 2020). Another way of joint modeling involves iteratively imputing values of each variable using chained equations (Van Buuren and Groothuis-Oudshoorn 2011) formulated with other variables (White, Royston, and Wood 2011; Van Buuren 2018; Muzellec et al. 2020). Discriminative models such as random forests (Xia et al. 2017), distribution constraints using optimal transport (Muzellec et al. 2020) and causally-aware imputation (Kyono et al. 2021) tend to depend on strong assumptions, which may result in a lack of \ufb02exibility to handle mixed-mode data. Most importantly, as for graph scenarios, matrix completion methods are generally limited without awareness of the underlying graph structures. Attribute Imputation with Graph Learning. Recently, graph learning models have been used to tackle the imputation task. GC-MC (Berg, Kipf, and Welling 2017) and IGMC (Zhang and Chen 2019) construct helpful bipartite interaction graphs to impute with a given adjacency matrix as side information. Then a GAE is applied to predict the absent features. Gaussian mixture model is utilized for imputation stability under a high missing rate (Monti et al. 2017). GRAPE (You et al. 2020) combines imputation and representation learning, which can well impute features of continuous variables but tend not to perform well on datasets containing nodes with all features missed (Tu et al. 2021). GDN (Li et al. 2021) can impute from over-smoothed representations with a given graph structure. In this case, relationships between nodes are explicitly encoded for further imputation. Above methods show that graph learning is competent for feature reconstruction with structure dependencies. Notably, only GRAPE, GDN and our method are applicable to both discrete and continuous features. Despite both GDN and our method focus on mechanistic improvements for existing GAEs, we explore GAEs with a more general phenomenon of concentration in a spectral view and not just for recovering high-frequency details. 3 Preliminary In this section, we provide the formulation of graph attribute imputation problem in Section 3.1. Then we introduce the regularized object, named graph spectral entropy in Section 3.2. In Section 3.3, we describe wavelet entropy, a core concept for ef\ufb01ciently approximating the graph spectral entropy. 3.1 Problem Formulation For an undirected graph G = (A, X), A \u2208RN\u00d7N is the adjacency matrix and X \u2208RN\u00d7D is a complete feature matrix where Xij denotes the graph attribute of i-th node in the j-th feature dimension. In the problem of graph attribute imputation, R \u2208{0, 1}N\u00d7D is de\ufb01ned as the mask matrix whose element Rij = 1 if Xij is observed and Rij = 0 otherwise. The objective of this work is to predict the missing graph attributes Xij at Rij = 0. Formally, we aim to develop a mapping f(\u00b7) to generate the imputed data matrix \u02dc X \u2208 RN\u00d7D de\ufb01ned as \u02dc X = f(X, R, A). (1) 3.2 Graph Spectral Entropy The graph Laplacian is de\ufb01ned as L = I \u2212\u02dc A, where \u02dc A = D\u22121 2 AD\u22121 2 denotes the normalized adjacency matrix and D = diag(P k A1k, ..., ANk) denotes the diagnoal degree matrix. With eigen-decomposition, L can be decomposed into L = U T \u039bU, where U consists of the eigenvectors of L. \u039b = diag(\u03bb) is the diagonal matrix whose diagonal \fFigure 2: (a) The proposed MEGAE for graph attribute imputation. With tight wavelet \ufb01lters, the encoder \ufb01lters the input graph into \u02c6 g1(L)X, ..., \u02c6 gM(L)X (M = 3 here), of which the energies (i.e., blue, orange and green bars) are obtained with the L-2 norm function. We maximize wavelet entropy of M energies, which are encouraged to be more uniform in distributions. This wavelet entropy maximization induces the same tendency to maximization for the real graph spectral entropy. We apply inverse wavelet \ufb01lters to decode the data from the latent (spectral) space for \ufb01nal reconstruction. (b) A diagram to explain the Coverage and Crossness of given tight wavelet \ufb01lters. We provide related descriptions in Section 4.1. elements \u03bb = {\u03bb1, ..., \u03bbN} are eigenvalues of L. Consider a feature vector x = {x1, ..., xN}\u22ba\u2208RN as a single feature dimension of the whole matrix X, the Fourier transform \u02c6 x = U T x = {\u02c6 x1, ..., \u02c6 xN}\u22bawould be the most proven method for representing spectrum, of which the entropy is de\ufb01ned as graph spectral entropy in De\ufb01nition 3.1. De\ufb01nition 3.1 (Graph Spectral Entropy). Given the feature vector x, its spectral energy at \u03bbi is denoted as \u02c6 x2 i and total energy is Es = PN i=1 \u02c6 x2 i = \r \rU T x \r \r2 2. Then, the graph spectral entropy of x is de\ufb01ned as \u03bes(x, L) = \u2212 N X i=1 \u02c6 x2 i Es log \u02c6 x2 i Es . Intuitively, maximizing graph spectral entropy would ensure the spectral distribution to be relatively more uniform and keep information in any spectral pass-band, as opposed to spectral concentration in existing GAEs (Park et al. 2019; Wu et al. 2019). However, one notable issue with the computation of graph spectral entropy is that eigen-decomposition is indispensable, which largely limits the scalability for computing on large-scale data matrix. 3.3 Wavelet Entropy We introduce wavelet theory for approximating the graph spectral entropy with wavelet transform. Unlike Fourier transform that we formulate in section 3.2, wavelet transform (Xu et al. 2019; Hammond, Vandergheynst, and Gribonval 2011) is able to cover ranges of spectra via different pass-band kernels \u02c6 g = {\u02c6 g1, .., \u02c6 gM}. It shows the possibility of representing graph spectral entropy by wavelet entropy with well-calibrated kernels. The wavelet transform employs a set of M wavelets as bases, de\ufb01ned as W = {W\u02c6 g1, ..., W\u02c6 gM }. Mathematically, the wavelet transform of x is de\ufb01ned as W\u02c6 gm(x) = U\u02c6 gm(\u039b)U T x, (2) where \u02c6 gm(\u00b7) is a wavelet kernel function de\ufb01ned in spectral domain. For computational ef\ufb01ciency, in many cases, kernel function \u02c6 gm should be smooth and easily approximated by polynomials such that U\u02c6 gm(\u039b)U T = \u02c6 gm(L). De\ufb01nition 3.2 (Wavelet Entropy). Given the feature vector x and a set of wavelet kernel functions \u02c6 g = {\u02c6 g1, ..., \u02c6 gM}, the total wavelet energy of x is Ew = PM m=1 \u2225W\u02c6 gm(x)\u22252 2. Then, the wavelet entropy of x is de\ufb01ned as \u03bew(x, L, \u02c6 g) = \u2212 M X m=1 \u2225W\u02c6 gm(x)\u22252 2 Ew log \u2225W\u02c6 gm(x)\u22252 2 Ew . 4 Maximizing Graph Spectral Entropy This section describes our main contributions. Following our motivation to ef\ufb01ciently maximize the graph spectral entropy while imputing missing attributes, we propose an alternative method for approximating the graph spectral entropy without eigen-decomposition and provide the upper bound for approximation error. On this basis, we propose our model called Max-Entropy Graph AutoEncoder (MEGAE). 4.1 Graph Wavelet for Approximation In this section, we present our main idea that graph spectral entropy can be well approximated by wavelet entropy. To achieve the goal, the crucial issue is how to construct a rough spectral space for approximation with the ef\ufb01cient graph convolutions. De\ufb01nition 4.1 (Tight Wavelet Frame on Graph). A set of wavelet kernel functions \u02c6 g = {\u02c6 g1, ..., \u02c6 gM} forms a tight \fframe on graph if G(\u03bbi) := M X m=1 [\u02c6 gm(\u03bbi)]2 = 1, \u2200\u03bbi \u2208\u03bb. Inspired by the fact that tight wavelet frame can ensure energy preservation (De\ufb01nition 4.1) in graph signal processing (Leonardi and Van De Ville 2013), we can smoothly deduce a key connection between wavelet energy (formulated in Section 3.3) and graph spectral energy (formulated in Section 3.2), which are both key components in the respective wavelet and spectral entropy. Proposition 4.2 describes this connection, indicating that the total wavelet energy is strictly equivalent to the total graph spectral energy. The proof is trivial and illustrated in Appendix A. Proposition 4.2 (Energy Parseval\u2019s Identity). If \u02c6 g is a tight wavelet frame, the total wavelet energy Ew is equivalent to total graph spectral energy Es. Thus, the Parseval\u2019s identity holds for the spectral domain, that is Ew = Es. The wavelet energy consists of M energy components and each of them is derived from the \ufb01ltered graph features by a wavelet kernel. Therefore, intuitively, a more distinguishable kernel function tends to provide accurate approximation of graph spectral entropy. To quantify the kernel distinguishability of tight wavelet, we de\ufb01ne Coverage Cm and Crossness Ri. Formally, Cm is the number of spectrum covered by kernel function \u02c6 gm denoted as Cm = # {i|\u02c6 gm(\u03bbi) \u0338= 0, i = 1, ..., N}, (3) and Ri is the number of activated kernel functions on \u03bbi represented by Ri = # {m|\u02c6 gm(\u03bbi) \u0338= 0, m = 1, ..., M}. (4) A toy example is provided in Figure 2(b) for easier understanding. Coverage C3 represents the number of eigenvalues (i.e., \u03bb4, \u03bb5, \u03bb6) lying within the domain of 3rd wavelet kernel (i.e., the green curve). The Crossness R2 represents number of intersection (red) points of all three wavelet kernels and line \u03bb = \u03bb2. Proposition 4.3 (Approximation of graph spectral entropy). If \u02c6 g is a tight frame, the approximation error between graph spectral entropy \u03bes(x, L) and wavelet entropy \u03bew(x, L, \u02c6 g) is bounded by |\u03bes(x, L) \u2212\u03bew(x, L, \u02c6 g)| \u2264e(\u02c6 g, \u03bb) where e(\u02c6 g, \u03bb) = max({log Cm}M m=1 \u222a{log Ri}N i=1). Please see Appendix B for the proof. To better understand Proposition 4.3, let N = M and all kernels be disjoint, which means the number of wavelet kernels equals to the number of nodes in the graph. In this perfect situation, one kernel function only covers one eigenvalue (i.e., Coverage Cm = 1) and one eigenvalue is only covered by one kernel function (i.e., Crossness Rm = 1). Then, the approximation error is zero. For applying a tight wavelet frame with high distinguishability, we follow the mechanisms in (Shuman et al. 2015) to design a set of wavelet kernel function g = {\u02c6 g1, ..., \u02c6 gM}. The design details are in Appendix C. 4.2 The Proposed MEGAE Based on the designed tight wavelet kernels, we propose MEGAE, a deterministic graph autoencoder framework for graph missing attribute imputation. As demonstrated in Figure 2, MEGAE consists of two key modules: (1) wavelet based graph autoencoder and (2) maximum entropy regularization. Encoder. With the de\ufb01nitions of an undirected graph provided in Section 3, we employ M tight wavelet kernels described in Section 4.1 in parallel to perform graph wavelet convolution. Speci\ufb01cally, the convolutional layer of m-th channel is expressed as E1(Z(0) m , L, \u02c6 gm) = \u03c6(W\u02c6 gm(Z(0) m )W (0) m ), (5) where Z(0) m = X \u2299R and \u03c6(\u00b7) is the activation function such as leaky ReLU. To admit a fast algorithm for wavelet transform, Maclaurin series approximation (Li et al. 2021) of order K is applied such that W\u02c6 gm(Z(0) m ) = PK k=0 \u03b1m,kLkZ(0) m . We introduce one more layer of perceptron to enrich the representation in latent space, denoted as E2(Z(1) m ) = \u03c6(Z(1) m W (1) m ). (6) The Encoder is thus formulated as Z(1) m = E1(Z(0) m , L, \u02c6 gm), Z(2) m = E2(Z(1) m ). (7) Decoder. Given the observed information, our decoder aims to impute missing values from latent space. As the inverse of graph wavelet convolution, the graph wavelet deconvolutional layer of m-th channel is expressed as D1(Z(2) m , L, \u02c6 g\u22121 m ) = \u03c6(W\u02c6 g\u22121 m (Z(2) m )W (2) m ), (8) where W\u02c6 g\u22121 m (Z(2) m ) = PK k=1 \u03b2m,kLkZ(2) m is provided by Maclaurin series. We then aggregate results generated by M channels to construct the imputation layer as D2(Z(3) 1 , ..., Z(3) M ) = \u03c6(Z(3) AGGW (3)), (9) where Z(3) AGG = AGG([Z(3) 1 , ..., Z(3) M ]) and AGG(\u00b7) denotes aggregation function such as concatenation. The Decoder is thus formulated as Z(3) m = D1(Z(2) m , L, \u02c6 g\u22121 m ), \u02dc X = D2(Z(3) 1 , ..., Z(3) M ). (10) Optimization. Our network is jointly optimized with respect to two objectives. We start from the imputation precision, which is the main focus in missing data problem. Intuitively, the reconstruction loss is de\ufb01ned as LR = \r \r \r( \u02dc X \u2212X) \u2299(1N\u00d7D \u2212R) \r \r \r 2. (11) Given our imputation motivation, MEGAE introduces spectral entropy regularization to aid data imputation. Based on the discussion in Section 4.1, graph spectral entropy can be well substituted by wavelet entropy. Let Z(2) m (:, d) be the d-th \fMethods ENZYMES PRO_full QM9 Synthie FRANKE FIRST_DB ENZYMES PRO_full RMSE with 0.1 missing features ACC. ACC. MEAN 0.0602 0.0653 0.2983 0.2063 0.3891 0.1500 62.06% 69.96% KNN 0.0350 0.0329 0.3058 0.1718 0.2010 0.1296 63.53% 72.23% SVD 0.0783 0.0465 0.2524 0.1697 0.2766 0.1685 61.60% 71.16% MICE 0.0292 0.0210 0.1986 0.1899 0.1359 0.1036 64.46% 73.50% GAIN 0.0300 0.0245 0.1973 0.1649 0.1103 0.0909 64.42% 73.21% OT 0.0323 0.0206 0.2003 0.1865 0.1116 0.0892 64.13% 73.62% MIRACLE 0.0288 0.0188 0.1846 0.1632 0.1196 0.0889 65.03% 74.07% GraphVAE 0.0357 0.0199 0.1579 0.1898 0.1099 0.1202 63.46% 73.86% MolGAN 0.0326 0.0160 0.1478 0.1864 0.1078 0.1379 64.16% 74.20% GRAPE 0.0302 0.0147 0.1869 0.1798 0.1069 0.0986 64.48% 74.53% GDN 0.0267 0.0155 0.1598 0.1764 0.1066 0.0869 65.57% 74.50% Inv_GCN 0.0256 0.0135 0.1773 0.1446 0.1089 0.1032 66.00% 74.98% Inv_MLP 0.0254 0.0129 0.1499 0.1520 0.1064 0.0872 66.06% 75.07% MEGAE 0.0223 0.0099 0.1396 0.1203 0.0936 0.0789 66.27% 75.97% Performance gain 0.0044 | 0.0560 0.0048 | 0.0554 0.0082 | 0.1662 0.0429 | 0.0860 0.0128 | 0.2955 0.0080 | 0.0896 0.70% | 4.67% 1.44% | 6.01% Table 1: RMSE results on 6 multi-graph datasets and graph classi\ufb01cation accuracy on two datasets. After running 5 trials, we report the mean results of which the best method is bolded and the second best is underlined. Performance gains indicate the maximum (lower) and minimum (upper) gains of state-of-the-art (MEGAE) compared to other baselines. Note that we abbreviate \u2019PROTEIN_full\u2019, \u2019FRANKENSTEIN\u2019 and \u2019FIRSTMM_DB\u2019 to \u2019PRO_full\u2019, \u2019FRANKE\u2019 and \u2019FIRST_DB\u2019, respectively. dimension of Z(2) m , the graph spectral entropy loss is thus de\ufb01ned as LS = \u22121 D D X d=1 M X m=1 P d m log P d m, (12) where P d m = \r \r \rZ(2) m (:, d) \r \r \r 2 2 PM m=1 \r \r \rZ(2) m (:, d) \r \r \r 2 2 . (13) Finally, by combining the aforementioned losses, the overall objective function of MEGAE is formulated as L = LR \u2212\u03b3LS, (14) where \u03b3 is a hyperparameter to balance the regularization of graph spectral entropy and reconstruction loss. 5 Experiments In this section, we validate the performance of MEGAE using a variety of datasets. We evaluate the effectiveness of MEGAE on two categories of graph datasets: \u2022 Type 1: Imputation on multi-graph datasets. We impute the missing graph attributes on multi-graph datasets, e.g., molecules, proteins. In addition, we report graph classi\ufb01cation performance on graphs with imputed features. \u2022 Type 2: Imputation on single-graph datasets. We impute the missing values on single-graph datasets, e.g., social network. We report node classi\ufb01cation performance on the graph with imputed features. 5.1 Imputation on Multi-graph Datasets. Datasets. We conduct experiments on 6 benchmark datasets (Morris et al. 2020) from different domains: (1) bioinformatics, i.e., PROTEINS_full (Borgwardt et al. 2005) and ENZYMES (Schomburg et al. 2004); (2) chemistry, i.e., QM9 (Ramakrishnan et al. 2014) and FIRSTMM_DB (Neumann et al. 2013); (3) computer vision, i.e., FRANKENSTEIN (Orsini, Frasconi, and De Raedt 2015); (4) synthesis, i.e., Synthie (Morris et al. 2016). The detailed statistics is provided in Table 6 in Appendix E.1. Baselines. We compare the performance of MEGAE against baselines in three categories: (1) statistical imputation methods including MEAN, KNN (Zhang 2012) and SVD (Troyanskaya et al. 2001); (2) deep learning-based imputation models including MICE (Van Buuren and GroothuisOudshoorn 2011), GAIN (Yoon, Jordon, and Schaar 2018), OT (Muzellec et al. 2020) and MIRACLE (Kyono et al. 2021); (3) graph learning-based models including GraphVAE (Simonovsky and Komodakis 2018), MolGAN (De Cao and Kipf 2018), GRAPE (You et al. 2020) and GDN (Li et al. 2021). For further details, please refer to Appendix E.3. Setup. We use a 70-10-20 train-validation-test split and construct random missingness only on the test set. Each run has a different dataset split and the mask for feature missingness. After running for 5 trials, we report the Root Mean Squared Error (RMSE) results for imputation and mean accuracy on the test set for graph classi\ufb01cation. For all baselines, we use a 2-layer GCN for downstream classi\ufb01cation. For more setup details, please refer to Appendix E.2. Results. In Table 1 we show the experiment results for 10% of missing features. Results indicate that our method has the lowest RMSE in the 5 real-world datasets and the synthetic \fMethods PubMed (RMSE) AmaPhoto (RMSE) 0.1 Miss 0.3 Miss 0.5 Miss 0.7 Miss 0.99 Miss 0.1 Miss 0.3 Miss 0.5 Miss 0.7 Miss 0.99 Miss sRMGCNN 0.0170 0.0186 0.0258 0.0309 0.0435 0.3191 0.3261 0.3282 0.3359 0.3660 GC-MC 0.0220 0.0233 0.0253 0.0270 0.0403 0.3154 0.3221 0.3340 0.3574 0.3913 GRAPE 0.0187 0.0198 0.0234 0.0303 0.0368 0.3207 0.3295 0.3357 0.3489 0.3896 VGAE 0.0196 0.0229 0.0298 0.0354 0.0525 0.3064 0.3196 0.3462 0.3684 0.3991 GDN 0.0168 0.0202 0.0239 0.0254 0.0319 0.3006 0.3168 0.3308 0.3372 0.3617 MEGAE 0.0157 0.0171 0.0185 0.0199 0.0218 0.2966 0.3121 0.3204 0.3243 0.3305 Performance gain 0.0011 | 0.0063 0.0015 | 0.0062 0.0049 | 0.0113 0.0055 | 0.0155 0.0101 | 0.0307 0.0040 | 0.0241 0.0047 | 0.0174 0.0078 | 0.0258 0.0116 | 0.0441 0.0312 | 0.0686 Table 2: Mean RMSE results of attribute imputation with different missing rates on PubMed and AmaPhoto. Methods PubMed (Acc. = 77.11% with full features) AmaPhoto (Acc. = 92.03% with full features) 0.1 Miss 0.3 Miss 0.5 Miss 0.7 Miss 0.99 Miss 0.1 Miss 0.3 Miss 0.5 Miss 0.7 Miss 0.99 Miss GCNMF 75.96% 74.42% 74.03% 70.09% 41.46% 91.64% 91.46% 91.05% 90.16% 73.03% PaGNN 76.55% 75.77% 74.45% 72.15% 43.63% 91.81% 91.70% 91.02% 90.12% 71.51% sRMGCNN 75.96% 74.74% 72.79% 71.14% 21.17% 91.36% 91.19% 91.02% 90.12% 66.56% GC-MC 74.44% 74.06% 72.87% 72.67% 36.69% 91.46% 91.26% 90.22% 89.82% 60.13% GRAPE 75.53% 74.51% 73.65% 71.75% 38.18% 91.22% 91.05% 90.33% 89.63% 68.28% VGAE 75.21% 73.69% 71.64% 69.96% 16.66% 91.53% 91.43% 89.23% 88.82% 63.24% GDN 75.87% 74.22% 73.36% 73.15% 38.93% 91.83% 91.63% 91.10% 90.01% 66.62% MEGAE 76.66% 76.30% 75.03% 74.47% 49.58% 91.92% 91.85% 91.41% 90.58% 81.88% Performance gain 0.11% | 1.45% 0.53% | 2.61% 0.58% | 3.39% 1.32% | 4.51% 5.95% | 32.92% 0.09% | 0.70% 0.15% | 0.80% 0.31% | 2.18% 0.42% | 1.76% 8.85% | 21.75% Table 3: Test accuarcy with different missing rates on PubMed and AmaPhoto node classi\ufb01cation benchmarks. dataset. Moreover, almost all graph learning-based methods outperform non-graph learning ones, demonstrating the effectiveness of modeling the dependencies between graph nodes. In addition, we introduce two more baselines \u2013 Inv_MLP denotes our tight wavelet encoder with Multi-Layer Perceptron (MLP) decoder (Simonovsky and Komodakis 2018) and Inv_GCN denotes our encoder with GCN decoder (Kipf and Welling 2017). MEGAE achieves 12.4% lower RMSE than the second best baselines on average over all the datasets. For a more comprehensive and realistic consideration for missing mechanisms, we provide the evaluation (see Figure 4 in Appendix F) for evaluating MEGAE in two less common scenarios. We obtain similar gains across three missing machanisms, MEGAE obtains the lowest errors in all cases, which shows the strong generalization among the three missing mechanisms. As for downstream classi\ufb01cations, we show the test accuracy of different models on imputed graph features with 10% missing. We observe that smaller imputation errors are more likely to lead to higher test accuracy. This is relatively expected, as data with less distribution shift easily yields better representation. MEGAE outperforms all other baselines on accuracy conditioning on the effectiveness imputation results. 5.2 Imputation on Single-graph Datasets. Dataset. We evaluate the effectiveness of MEGAE on four commonly used datasets including three citation network datasets\u2014Cora, Citeseer, Pubmed (Sen et al. 2008) and a copurchase network dataset\u2014AmaPhoto (McAuley et al. 2015). We show the dataset statistics in Table 7 in Appendix E.1. Baselines. We compare the imputation performance of MEGAE against competitive methods including sRMGCNN (Monti, Bronstein, and Bresson 2017), GCMC (Berg, Kipf, and Welling 2017), GRAPE (You et al. 2020), VGAE (Kipf and Welling 2016) and GDN (Li et al. 2021). For evaluating node classi\ufb01cation performance of MEGAE, we additionally experiment with GCNMF (Taguchi, Liu, and Murata 2021) and PaGNN (Jiang and Zhang 2020) that directly implement graph learning without imputation. For further information, please refer to Appendix E.3. Setup. We closely follow (Kipf and Welling 2017) to perform standard dataset split. For all datasets, each run has a different train-validation-test split and the mask for random missingness across all feature dimensions. After running for 5 trails, we report the results of mean RMSE for imputation and mean accuracy on the test set for node classi\ufb01cation. For all baselines, we use a 2-layer GCN for downstream classi\ufb01\fcation. For more setup details, please refer to Appendix E.2. Results. Table 2 shows the mean RMSE results for imputation under different missing rates on PubMed and AmaPhoto, we also provide the results on Cora and Citeseer in Table 8 in Appendix. MEGAE consistently outperforms other baselines. Moreover, MEGAE tends to have greater performance gains when faced with a more severe missingness, which demonstrates MEGAE\u2019s robustness. It is noteworthy that under an extremely high missing rate (0.99), MEGAE obtains margins of 31.66% on PubMed, 8.63% on AmaPhoto, 34.24% on Cora and 19.54% on Citeseer compared to the second best imputation methods. In Table 3, we show the mean accuracy for node classi\ufb01cation on PubMed and AmaPhoto with imputed attributes (results on Cora and Citeseer are in Table 9 in Appendix). We observe that most methods work well with less than 0.3 missing rates. Moreover, MEGAE\u2019s robustness and effectiveness are better as its accuracy gains tend to be positively correlated with missing rates. With the consistent imputation improvements on all datasets, MEGAE has accuracy gains of 5.95% on PubMed, 8.85% on AmaPhoto, 7.90% on Cora and 6.97% on Citeseer under a 0.99 missing rate. 5.3 Effectiveness of Wavelet and Regularization Graph Wavelet Autoencoders. We study the role of tight wavelet frames with the con\ufb01guration of regularization in MEGAE. Another two paradigms are deployed to get wavelet frames \u02c6 g = {\u02c6 g1, ..., \u02c6 gM}, including GWNN (Xu et al. 2019) with heat kernels and SGWT (Hammond, Vandergheynst, and Gribonval 2011) with highly-localized kernels. Dataset GWNN SGWT Ours ENZYMES +0.0042 +0.0044 0.0223 PROTEINS_full +0.0070 +0.0052 0.0099 QM9 +0.0102 +0.0166 0.1396 Synthie +0.0203 +0.0126 0.1203 FRANKENSTEIN +0.0103 +0.0185 0.0936 FIRST_DB +0.0036 +0.0073 0.0789 Table 4: RMSE with three graph wavelet paradigms. We report the mean RMSE of MEGAE and show the increase w.r.t. RMSE of GWNN and SGWT compared to ours. Here we provide experiments on 6 multi-graph datasets under 0.1 missing rate (in Table 4), reporting the increases w.r.t. RMSE using another two graph wavelet paradigms compared to ours, all three wavelet paradigms have 9 channels. Also, we execute experiments on 2 datasets (in Figure 3) assembling various channel numbers on three graph wavelet paradigms. Three paradigms all share the same computational complexity O(K \u00d7 |E|), where K is the order of polynomials and |E| is the number of edges. Table 4 veri\ufb01es the effectiveness of the proposed tight wavelet kernels. Figure 3 shows the in\ufb02uence of the number of wavelet kernels M. We observe that a larger M value harms the performance of GAEs with GWNN and SGWT paradigms. As for our proposed MEGAE, the RMSE \ufb01rst drops and then increases along w.r.t. the number of wavelet kernels, which shows there is a tradeoff between entropy approximation error and model generalization. Figure 3: RMSE with different numbers of wavelet kernels. We show how the kernel number in\ufb02uences the performance of GAEs with different wavelet paradigms. Entropy Regularization. We study the role of entropy regularization with the con\ufb01guration of tight wavelet in MEGAE. We execute experiment on 6 datasets under 10% missing rates. In Table 5 we provide the increase of RMSE by removing entropy regularization and the change in graph spectral entropy w.r.t. the original graph spectral entropy (denoted as \u2206). We observe that GAE without regularization will seriously lead to a decrease in spectral entropy (i.e., aforementioned spectral concentration) of the imputed graphs, and more importantly, LS improves the performance of the proposed MEGAE with a substantially maintenance of the real entropy. Dataset LR \u22061 LR + LS \u22062 ENZYMES +0.0069 -29% 0.0223 +3% PROTEINS_full +0.0061 -26% 0.0099 +2% QM9 +0.0188 -18% 0.1396 -1% Synthie +0.0316 -23% 0.1203 +3% FRANKENSTEIN +0.0229 -33% 0.0936 -2% FIRSTMM_DB +0.0086 -21% 0.0789 -2% Table 5: RMSE with/without entropy regularization. \u22061 and \u22062 denote the entropy changes with/without the regularization w.r.t. the original graph spectral entropy. Discussion. The effectiveness of our proposed tight wavelet and entropy regularization has been shown in Table 4 and Table 5, from which we can easily \ufb01nd that both have obvious improvements for decreasing imputation errors. Importantly, we believe the entropy regularization contributes the most to the performance of MEGAE. By analogizing MEGAE to the variational autoencoders, we provide intuitions on why MEGAE works (see Appendix D). 6"
+                }
+            ],
+            "Zijing Liu": [
+                {
+                    "url": "http://arxiv.org/abs/2102.12342v1",
+                    "title": "Similarity measure for sparse time course data based on Gaussian processes",
+                    "abstract": "We propose a similarity measure for sparsely sampled time course data in the\nform of a log-likelihood ratio of Gaussian processes (GP). The proposed GP\nsimilarity is similar to a Bayes factor and provides enhanced robustness to\nnoise in sparse time series, such as those found in various biological\nsettings, e.g., gene transcriptomics. We show that the GP measure is equivalent\nto the Euclidean distance when the noise variance in the GP is negligible\ncompared to the noise variance of the signal. Our numerical experiments on both\nsynthetic and real data show improved performance of the GP similarity when\nused in conjunction with two distance-based clustering methods.",
+                    "authors": "Zijing Liu, Mauricio Barahona",
+                    "published": "2021-02-24",
+                    "updated": "2021-02-24",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.IR",
+                        "stat.ML"
+                    ],
+                    "main_content": "INTRODUCTION Time course data are used widely for the empirical study of dynamical processes in many areas of research in the natural and social sciences [Keogh and Kasetty, 2003]. Traditionally, much research has been devoted to the characterisation of time series in relation to the originating dynamical process from different viewpoints, from the deterministic to the stochastic [Brillinger, 1981, Barahona and Poon, 1996, Chat\ufb01eld, 2003]. More recently, time series have also been considered from the perspective of data science. One of the key questions in many applications is to assess the (dis-)similarity between time courses, with a view to perform time series classi\ufb01cation or clustering [Liao, 2005, Son and Baek, 2008, G\u00f3recki, 2014, Peach et al., 2019, Fulcher and Jones, 2014]. A simple way to deal with \ufb01nite, discretely-sampled time courses is to treat them as vectors [Yao et al., 2005, Hedeker and Gibbons, 2006], i.e., a time series sampled at t time points {yi(tk) := yik}t k=1 is described by a t-dimensional vector yi with coordinates yik. A simple dissimilarity measure between two time series yi and yj is then given by the Euclidean (\u21132) distance: d2(yi,yj) = s t \u2211 k=1 \u0000yik \u2212yjk \u00012 = \u2225yi \u2212yj\u2225. (1) Alghouth the Euclidean distance is widely used due to its simplicity, in some applications, one may be more interested in the trend of how the data changes across time rather than the absolute differences. To capture this, a frequently used dissimilarity measure is based on Pearson\u2019s correlation coef\ufb01cient, ryiy j: dcorr(yi,yj) = 1\u2212ryiyj = 1\u2212 \u2211t k=1(yik \u2212yi)(yjk \u2212yi) p \u2211t k=1(yik \u2212yi)2 q \u2211t k=1(yjk \u2212yj)2 , where yi and yj are the means of each time series. Note that these two common measures are point-to-point matching, and insensitive to the re-ordering of the time points and the spacing between the sampling times. Hence these measures cannot capture information associated with the time indices, which can be important in applications. To remedy the limitations of point-wise measures, alternative measures of dissimilarity in time series have been proposed, including Dynamic Time Warping (DTW) [Keogh and Ratanamahatana, 2005] and the Edit Distance on Real Sequences (EDR) [Chen et al., 2005]. These methods can cope with uneven sampling and use information from the time indices, yet they can be algorithmically complex and are not well suited for applications with a large number of short, sparsely sampled time courses. Examples of this type of data are common in home price, marketing or ecommerce data in economics and \ufb01nance [Fan et al., 2011], longitudinal electronic healthcare records in healthcare [Perotte and Hripcsak, 2013], genomics and proteomics data in life science [Ndukum et al., 2011, Kayano et al., 2016], and functional magnetic resonance imaging [Smith, 2012]. For instance, in cellular biology, \u2018omics\u2019 experiments measure This is a preprint. arXiv:2102.12342v1  [cs.LG]  24 Feb 2021 \fthe expression level of large numbers of genes, proteins or metabolites in cells over time. Such datasets contain tens of thousands of time courses (i.e., the number of genes or proteins) but the length of each time course is very short (5-15 time points) due to the high cost of experiments. Furthermore, the time samplings are usually uneven since experiments are designed to capture trends in cellular evolution and responses to stimuli. These constraints are typical of many biological experimental settings. In this paper, we introduce a similarity measure for time course data based on Gaussian processes (GPs) [Rasmussen and Williams, 2006], which is applicable to sparse, inhomogeneously sampled, high-dimensional datasets. To retain information from the sampling times in the data, we model the time courses as continuous functions using GPs, and de\ufb01ne a similarity measure in the form of a log-likelihood ratio between GP models. The GP similarity is computationally simple and suitable for high-dimensional datasets with a large number of short time courses. We also show that the GP similarity measure is equivalent to the Euclidean distance when the noise variance in the GP model is negligible compared to the signal variance. We apply the GP similarity measure as the basis for distance-based clustering methods in both synthetic and real time course data, and show improved robustness to measurement noise and to sampling inhomogeneity. 2 RELATED WORK As a non-parametric model, GP is a \ufb02exible and ef\ufb01cient tool for time-dependent data modelling. Using the fact that a GP de\ufb01nes a reproducing kernel Hilbert space (RKHS), Lu et al. [2008] proposed a RKHS-based distance for time series de\ufb01ned as the Bregman divergence between the two posterior GPs. This distance has a closed form: it is the squared norm of the posterior mean functions in the RKHS induced by the GP. However, the Bregman divergence does not re\ufb02ect the uncertainty of the data, since it only depends on the posterior mean function. Hence the Bregman divergence can perform poorly in the presence of noise in the data, as we show below. GPs have been previously applied to time series of gene expression to detect differentially expressed genes [Stegle et al., 2010, Kalaitzis and Lawrence, 2011] and to infer the dynamics of transcriptional regulation [Lawrence et al., 2007, Gao et al., 2008]. In Kalaitzis and Lawrence [2011], the \ufb01tted GP model for each gene is compared to a noise model in order to rank the time courses and \ufb01nd differentially expressed genes. Here, we use the construction of GPs differently, and show that the likelihood ratio between two GP models provides a robust similarity measure for time courses. In the next section, we will introduce the GP model for time course data. 3 GAUSSIAN PROCESS MODEL FOR TIME COURSE DATA A Gaussian process is a collection of random variables over the index set X such that any \ufb01nite collection of the random variables follows a multivariate normal distribution [Rasmussen and Williams, 2006]. Therefore, a Gaussian process, denoted GP(m(x),k(x,x\u2032)), is characterised by the mean function m(x) and the covariance function k(x,x\u2032), where x,x\u2032 \u2208X. Here we will consider time-dependent variables; hence the index set is the positive real line describing time: X = R+. We will model the underlying true signal as a Gaussian process over time: f(x) \u223cGP \u0000m(x),k(x,x\u2032) \u0001 . (2) For simplicity, we take the mean function to be zero (m(x) = 0) and we use the squared exponential covariance function k(x,x\u2032) = \u03c32 f exp \u0012 \u2212\u2225x\u2212x\u2032\u22252 2\u21132 \u0013 =: \u03c32 f G(x,x\u2032) (3) where \u2113is a characteristic length-scale, \u03c32 f is the signal variance and we use G(x,x\u2032) to denote the Gaussian kernel. The observations of the time-dependent variable are then noisy samples of the GP: y = f(x)+\u03b5, (4) where the additive noise \u03b5 is Gaussian with zero mean and variance \u03c32 n . Let us consider a time-dependent variable y given by (4) sampled at t time points X = [x1,...,xt]T and let us compile the observations into a t-dimensional vector y = [y1,...,yt]T. Under our assumptions, the covariance function of the noisy observations y is given by: ky(xp,xq) = \u03c32 f G(xp,xq)+\u03c32 n \u03b4pq, (5) where \u03b4pq is the Kronecker delta. Equivalently, the t \u00d7 t covariance matrix of the observations y is: Ky = \u03c32 f G+\u03c32 n I = K +\u03c32 n I, (6) where G is the Gaussian kernel matrix with elements Gpq = G(xp,xq), I is the identity matrix of dimension t, and K is the covariance matrix for the noiseless samples with elements Kpq = k(xp,xq). The three hyperparameters of the Gaussian process are therefore \u03b8 = (\u2113,\u03c3f ,\u03c3n), and can be learnt from the data (X,y) by maximising \u03b8 \u2217= argmax \u03b8 log p(y|X,\u03b8), (7) \fwhere, in this case, the log-marginal likelihood has the explicit form: log p(y|X,\u03b8) = \u22121 2 \u0000yTK\u22121 y y+logdetKy +t log2\u03c0 \u0001 , (8) with detKy denoting the determinant of Ky. This expression can be maximised using gradient-based methods [Rasmussen and Williams, 2006]. 4 SIMILARITY MEASURE FOR TIME COURSE DATA BASED ON GAUSSIAN PROCESS Let us consider a sparse time course dataset consisting N short time courses sampled at t time points: yi \u2208Rt, i = 1,...,N. The dataset is referred to as sparse, due to the fact that the number of time courses is much larger than the length of each time course (N \u226bt). Although the GP model does not require all the time series to be measured synchronously, for simplicity, we \ufb01rst introduce the similarity measure for the case where all time courses are sampled at the same time points X = [x1,...,xt]T. We discuss the asynchronous case later. Each of the time courses yi is assumed to correspond to a noisy observation of a Gaussian process (4) with the same hyperparameters \u03b8, which can be inferred by maximising the sum of the log marginal likelihoods of the time courses [Rasmussen and Williams, 2006]: \u03b8 \u2217= argmax \u03b8 N \u2211 i=1 log p(yi|X,\u03b8). (9) By inferring the hyperparameters \u03b8 \u2217, we obtain a nonparametric probabilistic model for the observed time courses, which we can use to de\ufb01ne a GP-based similarity measure, as follows. 4.1 LIKELIHOOD RATIO AS A SIMPLE GP SIMILARITY MEASURE Using the fact that the GP is a distribution of continuous functions over time, a similarity measure between two observed time courses yi and yj can be obtained by comparing two different possibilities as to how yi and yj could have been generated. The \ufb01rst possibility is that the two time samples yi and yj are observations from two different functions sampled from the GP. In this case, the joint likelihood of yi and y j is just the product of the likelihoods of two time courses: pdiff(yi,yj|\u03b8 \u2217) = p(yi|X,\u03b8 \u2217) p(yj|X,\u03b8 \u2217). (10) Using (8), it is easy to see that the log-likelihood can be rewritten as: log pdiff(yi,yj|\u03b8 \u2217) = log p(yi|X,\u03b8 \u2217)+log p(yj|X,\u03b8 \u2217) = \u22121 2 \u0002yT i yT j \u0003\u0014 Ky 0 0 Ky \u0015\u22121 \u0014 yi yj \u0015 \u22121 2 logdet \u0014 Ky 0 0 Ky \u0015 \u2212t log2\u03c0. (11) The second possibility is that yi and yj are observations from the same function sampled from the GP. In this case, the joint likelihood of yi and yj can be computed by considering the two time courses {X,yi} and {X,yj} to be replicate samples of one function : psame(yi,yj|\u03b8 \u2217) = p \u0012\u0014 yi yj \u0015\f \f \f \f \u0014 X X \u0015 ,\u03b8 \u2217 \u0013 , (12) and the log-likelihood is then given by: log psame(yi,yj|\u03b8 \u2217) = log p \u0012\u0014 yi yj \u0015\f \f \f \f \u0014 X X \u0015 ,\u03b8 \u2217 \u0013 = \u22121 2 \u0002yT i yT j \u0003\u0014 Ky K K Ky \u0015\u22121 \u0014 yi yj \u0015 \u22121 2 logdet \u0014 Ky K K Ky \u0015 \u2212t log2\u03c0, (13) where we have used the fact that the additive noise \u03b5 in (4) is uncorrelated between the two time courses. The likelihood (12) will be high if the two time courses are similar to each other (as in Fig. 1a), and will be small if the two time courses have different pro\ufb01les (as in Fig. 1b). Hence for time courses with different pro\ufb01les, the likelihood (10) explains better the data. The log of the ratio of the two likelihoods (12)\u2013(10) (i.e., the difference between the log-likelihoods) is thus an indicator of the level of similarity between two time courses. This leads to our de\ufb01nition of the GP similarity measure between yi and yj as: s(yi,y j) = log psame(yi,y j|\u03b8 \u2217) pdiff(yi,yj|\u03b8 \u2217) (14) and it follows that: s(yi,yj) = log psame(yi,yj|\u03b8 \u2217)\u2212log pdiff(yi,y j|\u03b8 \u2217) = 1 2 \u0002yT i yT j \u0003  \u0014 Ky 0 0 Ky \u0015\u22121 \u2212 \u0014 Ky K K Ky \u0015\u22121!\u0014 yi yj \u0015 \u22121 2 log detQ detKy , (15) where Q is the Schur complement: Q = Ky \u2212KK\u22121 y K. (16) Note that the likelihood ratio (14) has the same form as the Bayes factor [Kass and Raftery, 1995] in Bayesian model \f0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 t -1.5 -1 -0.5 0 0.5 1 1.5 2 y 95% confidence interval posterior mean time course 1 time course 2 (a) Two time courses with similar patterns and the corresponding GP. 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 t -2 -1.5 -1 -0.5 0 0.5 1 1.5 2   y 95% confidence interval posterior mean time course 1 time course 2 (b) Two time courses with different patterns and the corresponding GP. Figure 1: Two examples with two time courses that have: (a) similar and (b) dissimilar pro\ufb01les. In each case, we show the time courses and the mean and con\ufb01dence interval of the Gaussian process (12) obtained according to the likelihood in (12). selection. From this perspective, our measure can be understood as the comparison between two models in two ways: on one hand, s(yi,yj) compares modelling two time courses with one single function of the GP versus modelling them with two independent functions of the GP; alternatively, it is easy to see that our measure (14) can be rewritten as s(yi,yj) = log p(yj|yi,X,\u03b8 \u2217)\u2212log p(yj|X,\u03b8 \u2217) = log p(yi|yj,X,\u03b8 \u2217)\u2212log p(yi|X,\u03b8 \u2217) which is the difference between the log-likelihood of yj based on the posterior GP given yi compared to the prior GP without yi being given. Hence the measure quanti\ufb01es the improvement in the prediction of y j that can be drawn by knowing yi. The measure is symmetric, so the same applies by exchanging yj for yi. 4.2 EUCLIDEAN DISTANCE AS A LIMIT OF THE GP LOG-LIKELIHOOD RATIO It can be shown that the Euclidean distance (1) stems naturally from the GP similarity (14) in the limit when the noise variance \u03c32 n is much smaller than the signal variance \u03c32 f . To see this, recall the Neumann series [Stewart, 1998] K\u22121 y = (K +\u03c32 n I)\u22121 = K\u22121 \u221e \u2211 m=0 (\u22121)m \u0000\u03c32 n K\u22121\u0001m . Noting that the Gaussian kernel matrix G is positive de\ufb01nite, we have that if \u03c32 n /\u03c32 f \u21920 = \u21d2\u2225\u03c32 n K\u22121\u2225= (\u03c32 n /\u03c32 f ) \u2225G\u22121\u2225\u21920, For small \u03c32 n /\u03c32 f , we thus take the \ufb01rst two terms of the expansion to O \u0000\u2225\u03c32 n K\u22121\u22252\u0001 : KK\u22121 y = K\u22121 y K \u2243I \u2212\u03c32 n K\u22121, (17) and the Schur complement (16) is approximated as: Q \u2243Ky \u2212(I \u2212\u03c32 n K\u22121)K = 2\u03c32 n \u0010 I \u22121 2\u03c32 n K\u22121\u0011 , Q\u22121 \u2243 1 2\u03c32 n \u0010 I + 1 2\u03c32 n K\u22121\u0011 . To approximate the GP similarity measure (15), we use block matrix inversion: \u0014 Ky 0 0 Ky \u0015\u22121 \u2212 \u0014 Ky K K Ky \u0015\u22121 = \u0014\u2212K\u22121 y KQ\u22121KK\u22121 y Q\u22121KK\u22121 y K\u22121 y KQ\u22121 \u2212K\u22121 y KQ\u22121KK\u22121 y \u0015 \u2243 1 2\u03c32 n \u0014 \u2212(I \u22123\u03c32 n K\u22121/2) I \u2212\u03c32 n K\u22121/2 I \u2212\u03c32 n K\u22121/2 \u2212(I \u22123\u03c32 n K\u22121/2) \u0015 = \u22121 2\u03c32 n  \u0014 1 \u22121 \u22121 1 \u0015 \u2297I \u22121 2 \u0014 3 \u22121 \u22121 3 \u0015 \u2297 \u0000\u03c32 n K\u22121\u0001 ! . and we approximate the determinant: detQ = detKy det \u0000I \u2212(KK\u22121 y )2\u0001 \u2243detKy det \u00002\u03c32 n K\u22121\u0001 , whence we obtain: s(yi,yj) \u2243 \u22121 4\u03c32 n \u0002yT i yT j \u0003  \u0014 1 \u22121 \u22121 1 \u0015 \u2297I \u2212\u03c32 n 2\u03c32 f \u0014 3 \u22121 \u22121 3 \u0015 \u2297G\u22121 !\u0014 yi y j \u0015 \u22121 2 logdet   2 \u03c32 n \u03c32 f G\u22121 ! . Here \u2297denotes the Kronecker product and G is the Gaussian kernel matrix in (6). The approximation of the GP similarity measure (15) to \ufb01rst order is then : s(yi,yj) \u2243\u22121 4\u03c32 n   \u2225yi \u2212yj\u22252 +O   \u03c32 n \u03c32 f \r \rG\u22121\r \r !! . (18) Therefore, if \u03c32 n /\u03c32 f \u21920 (i.e., when the variance of the noise is much smaller than the signal variance), the dissimilarity measure \u2212s(yi,yj) is equivalent to the Euclidean distance \u2225yi \u2212yj\u22252. Note that this relationship holds not only for the Gaussian kernel G(x,x\u2032) but for any positive de\ufb01nite kernel. \f4.3 ASYNCHRONOUS TIME COURSES Although, for simplicity of exposition, we have concentrated on the case of synchronous time sampling, the GP similarity measure is equally applicable to non-synchronous samples. In our derivations above, synchronous time points are only necessary to obtain the formal limit to the Euclidean distance in Eq. 18. To see the applicability to non-synchronous samples, note that the likelihoods in Eq. 13 and Eq. 11 do not require the time courses to have the same time points. We can therefore consider N time courses denoted by yi = [yi,1,\u00b7\u00b7\u00b7 ,yi,ti] of length ti, sampled at (potentially) distinct time points Xi = [xi,1,\u00b7\u00b7\u00b7 ,xi,ti] (i = 1,...,N). In this case, we can similarly model the time courses with GP and learn the hyperparameters by maximising \u03b8 \u2217= argmax \u03b8 N \u2211 i=1 log p(yi|Xi,\u03b8), (19) and it is then easy to write the two log-likelihoods: log pdiff(yi,yj|\u03b8 \u2217) = log p(yi|Xi,\u03b8 \u2217)+log p(yj|Xj,\u03b8 \u2217) = \u22121 2 \u0002yT i yT j \u0003\u0014Kyi 0 0 Ky j \u0015\u22121 \u0014 yi yj \u0015 \u22121 2 logdet \u0014Kyi 0 0 Ky j \u0015 \u22121 2(ti +tj) log2\u03c0, (20) log psame(yi,yj|\u03b8 \u2217) = log p \u0012\u0014 yi yj \u0015\f \f \f \f \u0014 Xi Xj \u0015 ,\u03b8 \u2217 \u0013 = \u22121 2 \u0002yT i yT j \u0003\u0014Kyi Ki j KT i j Ky j \u0015\u22121 \u0014 yi yj \u0015 \u22121 2 logdet \u0014Kyi Ki j KT i j Ky j \u0015 \u22121 2(ti +tj) log2\u03c0. (21) where Kyi and Ky j are the covariance matrices of yi and yj with sizes ti \u00d7ti and tj \u00d7t j, respectively; and Kij is the cross-covariance matrix between yi and yj with size ti \u00d7 tj. The GP similarity between the two time courses can again be computed as the difference between the two loglikelihoods (20) and (21). 4.4 COMPUTATIONAL COMPLEXITY In terms of computational complexity, \ufb01tting a GP model with N time courses of length t takes O(t3 + Nt2) time. Computing pairwise similarities takes O(tN2) time. Since we deal with high-dimensional short time courses (N \u226bt), the total time for GP similarity would be approximately O(tN2), which is the same as for the Euclidean distance. Extra computational time is needed for the asynchronised time courses, where all the time courses have a different covariance matrix. It results in a computational time of O(Nt3) for model \ufb01tting and O(N2t3) for computing the pairwise GP similarity if all the time courses are of average length t. This might be a limitation to applications with large N and t. 5 NUMERICAL EXPERIMENTS To test the applicability of the proposed GP similarity measure, we have run numerical experiments on synthetic and real sparse time course data. Our results show that the similarity s(yi,yj) is more robust to observational noise than the Euclidean distance when used as the basis to cluster time courses using two standard clustering algorithms (hierarchical and spectral clustering). We also compare the performance of the GP similarity against the Bregman divergence and Dynamic Time Warping. 5.1 SYNTHETIC DATA Our synthetic dataset is obtained by sampling from the three different time pro\ufb01les shown in Fig. 2 with additive Gaussian noise. From each of the three time pro\ufb01les, we generate 50 evenly-sampled time courses of length t = 15 with a given level of sampling noise. Our task is to cluster the 150 time series into 3 groups in an unsupervised manner. To do this, we compute the pairwise similarities (or distances) between the 150 time courses using the GP similarity (14). From the computed similarity matrix, we then cluster the samples using two well-known methods: (i) spectral clustering with kNN graph (k = 7) [Yu and Shi, 2003]; (ii) agglomerative hierarchical clustering with average linkage [Rokach and Maimon, 2005]. We then repeat the numerical experiment 100 times in each case. To evaluate the clustering performance, we use the normalised mutual information (NMI) [Vinh et al., 2010] against the known ground truth 0 0.2 0.4 0.6 0.8 1 t -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 y y=t+0.25t4 y=t-0.25t4 y=t sampling time points (even) sampling time points (uneven) Figure 2: The three functions and the sampling time positions used to generate the synthetic data. The results of even sampling are in Fig. 3. The results of uneven sampling are in Fig. 4. \f(i.e., the three pro\ufb01les used to generate the data). We then repeat the same procedure with three popular measures: the Euclidean distance, the Dynamic Time Warping (DTW) distance and the Bregman divergence in the RKHS. Given that the only varying ingredient is the similarity measure, the clustering performance re\ufb02ects the quality of the similarity measure for this purpose. Proposed Euclidean Bregman DTW Proposed Euclidean Bregman DTW 0 0.2 0.4 0.6 0.8 1 NMI noise=0.8 Spectral clustering Hierarchical clustering (a) Noise level = 0.08 Proposed Euclidean Bregman DTW Proposed Euclidean Bregman DTW 0 0.2 0.4 0.6 0.8 1 NMI noise=0.10 Spectral clustering Hierarchical clustering (b) Noise level = 0.10 Proposed Euclidean Bregman DTW Proposed Euclidean Bregman DTW 0 0.2 0.4 0.6 0.8 1 NMI noise=0.12 Spectral clustering Hierarchical clustering (c) Noise level = 0.12 Figure 3: Clustering performance (measured as NMI) for spectral clustering with kNN and hierarchical clustering based on the proposed GP similarity (14), Euclidean distance (1), DTW and Bregman divergence in the RKHS with increasing noise levels from (a)-(c). The sampling time points are equally spaced between 0 and 1 (see Fig. 2). Figure 3 shows the clustering quality (0 \u2264NMI \u22641) for evenly sampled time series with increasing levels of sampling noise achieved with both clustering methods and all four distance/similarity functions. As expected from (18), both Euclidean and GP similarity give comparable results for small sampling noise when using spectral clustering. Note, however, that even for small noise, the performance of the GP similarity is superior to the Euclidean distance when using hierarchical clustering (Fig. 3a), a method that is very sensitive to noisy data. Also as expected the DTW performs worse than the Euclidean distance for the synchronous case, since there is no need for alignment. The Bregman divergence does not perform well in this case because it measures the distance between the two continuous curves in the RKHS \ufb01tted from the time courses data and does not consider the uncertainty due to noise. As the observation noise increases, our GP similarity measure gains further advantage over the Euclidean distance for both clustering methods (Figs. 3b\u20133c). The p-values for a Wilcoxon rank-sum test between the NMI values of Euclidean and GP similarity with spectral clustering for the noise levels 0.08, 0.10 and 0.12 are 0.043, 2.5e-7 and 7.6e16, respectively. The p-values for a Wilcoxon rank-sum test between the NMI values of Euclidean and GP similarity with hierarchical clustering for the noise levels 0.08, 0.10 and 0.12 are 1.4e-15, 1.7e-25, and 1.1e-20, respectively. Hence the GP similarity measure performs signi\ufb01cantly better than the Euclidean distance for clustering with both algorithms. In general, spectral clustering always performs better than hierarchical clustering, which is more sensitive to noise, and the best performance is obtained consistently using spectral clustering with GP similarity. Note that the performance obtained with spectral clustering using other metrics can be achieved at a lower computational cost using hierarchical clustering with GP similarity. Proposed Euclidean Bregman DTW Proposed Euclidean Bregman DTW 0 0.2 0.4 0.6 0.8 1 NMI noise=0.08 Spectral clustering Hierarchical clustering (a) Noise level = 0.08 Proposed Euclidean Bregman DTW Proposed Euclidean Bregman DTW 0 0.2 0.4 0.6 0.8 1 NMI noise=0.10 Spectral clustering Hierarchical clustering (b) Noise level = 0.10 Figure 4: Clustering performance (measured as NMI) for spectral clustering with kNN and hierarchical clustering based on the proposed GP similarity (14), Euclidean distance (1), DTW and Bregman divergence in the RKHS with increasing noise levels from (a)-(b). In this case, the time points are irregularly sampled between 0 and 1 (see Fig. 2). The GP similarity consistently outperforms the other distances, especially for larger amounts of noise. We also analyse a second set of 150 time series of length t = 15 collected from the same three functions in Fig. 2 but sampled inhomogeneously in time. The computational procedure is identical to the case of evenly sampled series described above. Figure 4 shows that the advantage of the GP similarity measure against the other distances is more prominent when the time points are unevenly sampled. This result highlights the fact that point-to-point similarities (such \fas the Euclidean distance) can miss important information contained in the long-term trends of the time pro\ufb01les if the sampling is concentrated irregularly in particular time periods. We next tested the non-synchronous case. We again generate a set of 150 time series of length t = 15 from the same three functions in Fig. 2. We then make the asynchronous time courses by randomly removing 6,7 or 8 time points from our 15-point time course data (Fig. 5a). In this case, the Euclidean distance is not de\ufb01ned. The proposed GP similarity achieves substantially better results than the Bregman divergence and DTW on these examples. 0 0.5 1 t -0.2 0 0.2 0.4 0.6 0.8 1 y (a) asynchronous time course Proposed Bregman DTW Proposed Bregman DTW 0 0.2 0.4 0.6 0.8 1 NMI noise=0.08 Spectral clustering Hierarchical clustering (b) NMI results Figure 5: (a) Three sample time courses with asyncronised measurements with Noise level = 0.08. (b) Clustering performance (measured as NMI) for spectral clustering with kNN and hierarchical clustering based on the proposed GP similarity, DTW and Bregman divergence in the RKHS. The GP similarity consistently outperforms the other distances for both clustering methods. In summary, our numerical experiments on synthetic data indicate improved performance of the GP similarity measure (14). As the observational noise decreases, the performance of the GP similarity measure is equivalent to the Euclidean distance (18). Although the GP similarity shows enhanced performance for both clustering methods, the improvement is larger for hierarchical clustering, in the presence of large amounts of noise, and under uneven time sampling. As a side comment, we also note that the inclusion of the noise variance \u03c32 n in the covariance function Ky makes the GP similarity measure more robust for numerical computations, as it reduces numerical instability. 5.2 APPLICATION TO GENE EXPRESSION TIME COURSE DATA We next tested the GP similarity measure on a real time course dataset which characterises the gene expressions during the process of cellular reprogramming where differentiated cells are reverted to stem cells [Di Stefano et al., 2014b]. In this dataset, gene expression values were measured at \ufb01ve time points: 0, 2, 4, 6 and 8 days after the cellular reprogramming started. Two biological replicates were measured at each time point. The full measurements include around 40,000 genes. Following standard practice, we select 1912 genes that are highly variable over time. We then follow the same procedure as for the synthetic data above to cluster the 1912 gene expression time courses to extract groups of genes that have similar time pro\ufb01les during cellular reprogramming. 0 50 100 150 200 time (hour) 5 10 15 expression level Samd14 (a) Time course of gene Samd14 0 50 100 150 200 time (hour) 5 10 15 expression level Fbln5 (b) Time course of gene Fbin5 0 50 100 150 200 time (hour) 5 10 15 expression level Serpine1 (c) Time course of gene Serpine1 Figure 6: The time courses of three genes (each time course was measured on two bioogical replicates). The Euclidean distances from Samd14 to Fbin5 and Serpine1 are both 5.09. In contrast, the GP dissimilarity (\u2212s(yi,yj)) between Samd14 and Fbin5 (37.48) is larger than between Samd14 and Serpine1 (25.57), capturing the fact that the time pro\ufb01le of Serpine1 is more similar to Samd14 than to Fbin5. For this dataset, there is no known \u2018ground truth\u2019 against which to compare the obtained clusters. To assess the quality of the clustering, we use a biological score: a modi\ufb01ed \fTable 1: BHI z-scores of the two clustering methods with different number of clusters obtained by analysing the time courses of 1912 highly variable genes in the stem cell transcriptomic dataset of Refs. [Di Stefano et al., 2014a,b]. No. of clusters Method Similarity 9 11 13 15 16 Spectral Proposed 2.505 4.250 1.360 3.368 3.674 Euclidean 1.368 2.067 2.460 2.351 2.919 DTW 2.636 2.087 2.774 2.720 3.182 Bregman 4.229 2.725 2.652 3.617 3.520 Hierarchical Proposed 1.182 2.384 2.366 1.849 3.353 Euclidean -0.090 0.074 -0.266 -0.215 -0.319 DTW 0.542 0.566 0.499 0.735 0.752 Bregman 1.392 1.303 1.139 1.388 1.508 version of the biological homogeneity index (BHI). The BHI measures within-cluster homogeneity of the genes in terms of biological knowledge and is given by: BHI = 1 L L \u2211 l=1 1 nl(nl \u22121) \u2211 i\u0338=j i, j\u2208clusterl S(i, j), where L is the number of clusters, nl is the size of cluster l, and S(i, j) is a biological similarity between gene i and gene j. In the original BHI [Datta and Datta, 2006], the similarity between genes is either 0 or 1, which is an indicator function of two genes sharing any Gene Ontology (GO) terms. In order to better capture the biological information, we use an information-theoretic semantic gene similarity based on Gene Ontologies [Resnik, 1999, Lord et al., 2003, Yu et al., 2010]. For a given clustering, we compute the BHI of 1000 random clusterings with the same number and size of clusters. We then use the z-score of the BHI against the random clusterings as a level of signi\ufb01cance for the obtained clustering. Since the number of clusters is unknown, we compute the z-score of the BHI for clusterings with different numbers of clusters (Table 1). Again, we \ufb01nd that spectral clustering with GP similarity achieves the best performance. On the other extreme, hierarchical clustering based on Euclidean distances performs no better than random clustering, an indication that the gene expression data has high levels of noise \u03c3n. As was the case for the synthetic data above, using GP similarity improves the performance of hierarchical clustering substantially, almost to a comparable level to the spectral method (but below). These results underscore the ability of the GP similarity to deal with noisy data. To illustrate visually the behaviour of the GP similarity, Figure 6 shows the time courses of three genes where the GP similarity and Euclidean distance behave differently. In our chosen examples, the time course of the Samd14 gene (Fig. 6a) has the same Euclidean distance to both the Fbin5 gene (Fig. 6b) and the Serpine1 gene (Fig. 6c). On the other hand, according to our GP similarity, the Serpine1 time course is more similar to Samd14 than to Fbin5, in accordance with our visual expectation from the observed time course pro\ufb01les. The codes and data are available online in https: //github.com/barahona-research-group/ BayesFactorSimilarity. 6"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1909.04491v2",
+                    "title": "Graph-based data clustering via multiscale community detection",
+                    "abstract": "We present a graph-theoretical approach to data clustering, which combines\nthe creation of a graph from the data with Markov Stability, a multiscale\ncommunity detection framework. We show how the multiscale capabilities of the\nmethod allow the estimation of the number of clusters, as well as alleviating\nthe sensitivity to the parameters in graph construction. We use both synthetic\nand benchmark real datasets to compare and evaluate several graph construction\nmethods and clustering algorithms, and show that multiscale graph-based\nclustering achieves improved performance compared to popular clustering methods\nwithout the need to set externally the number of clusters.",
+                    "authors": "Zijing Liu, Mauricio Barahona",
+                    "published": "2019-09-06",
+                    "updated": "2020-01-13",
+                    "primary_cat": "cs.IR",
+                    "cats": [
+                        "cs.IR",
+                        "cs.LG",
+                        "physics.data-an",
+                        "stat.ML"
+                    ],
+                    "main_content": "Introduction Clustering is a classic task in data mining, whereby input data are organised into groups (or clusters) such that data points within a group are more similar to each other than to those outside the group [1]. Such a task is distinct from supervised (or semi-supervised) classi\ufb01cation, where examples of the di\ufb00erent classes are known a priori and are used to train a computational model to assign other objects to the known groups. Instead, clustering aims to \ufb01nd natural, intrinsic sub-classes in the data, without assuming a priori the number or type of clusters. Indeed, a key open issue in this \ufb01eld is the principled determination of the number of clusters in an unsupervised manner, without the assumption of a generative model [2, 3]. The obtained groups can then constitute the basis for a simpler, yet informative, representation of large, complex datasets. Data clustering has a long history and there exist a myriad of clustering algorithms based on di\ufb00erent principles and heuristics [4]. In their most basic form, many popular clustering techniques (e.g., k-means [5] and mixture models [6]) are based on the assumption that the data follows an explicit (typically multivariate Gaussian) distribution. Clusters are then de\ufb01ned as the samples most likely generated from the same distribution, and learned by likelihood maximisation. However, in real applications, the model that generates the data is unknown and the resulting data distribution may be complex. In this case of data-driven analysis, model-based clustering often yields poor results [7, 8, 9, 10]. An alternative approach is provided by spectral clustering, which uses the eigenvectors of a (normalised) similarity matrix derived from the data to \ufb01nd relevant subgroups in the dataset [8, 11]. Spectral clustering is underpinned by results in matrix analysis (e.g., singular value decomposition), and has strong connections to model reduction, geometric projections and dimensionality reduction [11, 12]. The choice of similarity measure is a crucial ingredient to the clustering performance but, as long as a similarity matrix can be computed, spectral methods provide an attractive choice for non-vector data or for data sampled from irregular and non-convex data manifolds [13, 14]. From a di\ufb00erent perspective, the similarity matrix of a dataset can also be viewed as the adjacency matrix of a fully connected, weighted graph, where the nodes correspond to data points and the edge between two nodes is weighted by their similarity. One can then apply graph-based algorithms for community detection or graph partitioning to the problem of data clustering. Graph-based methods typically operate by searching for balanced graph cuts, sometimes invoking notions from spectral graph theory, i.e., using the spectral decomposition of the adjacency or Laplacian matrices of the graph [15, 16]. Spectral clustering can thus be understood as a special case of the broader class of graph-based clustering methods [12]. Importantly, 1 Department of Mathematics, Imperial College London, South Kensington Campus, London SW7 2AZ, UK 2 EPSRC Centre for Mathematics of Precision Healthcare, Imperial College London, London SW7 2AZ, UK \u2217m.barahona@imperial.ac.uk 1 arXiv:1909.04491v2  [cs.IR]  13 Jan 2020 \fgraph-based clustering is also able to reveal modular structure in graphs across levels of resolution through multiscale community detection [17, 18, 19]. This approach allows for the discovery of natural data clusterings of di\ufb00erent coarseness [20], thus recasting the problem of \ufb01nding the appropriate number of clusters to the detection of relevant scales in the graph. Methods for graph construction usually involve a sparsi\ufb01cation of the similarity (or distance) matrix under di\ufb00erent heuristics (from simple thresholding to sophisticated regularisations) in order to extract a similarity graph that preserves key properties of the dataset [21]. The representation of data through graphs has attractive characteristics, including the capability of capturing e\ufb03ciently the local and global properties of the data through graph-theoretical concepts that embody naturally the notions of local neighbourhoods, paths, and global connectivity [22, 23, 18]. The usage of graphs provides a natural links of spectral clustering with other clustering methods and allows for easy generalisation to a semi-supervised setting [24, 25]. Graphs also provide a means to capture the geometry of complex manifolds, a feature of interest in realistic datasets [26]. Graph representations not only reduce the computational cost for spectral graph methods, but also allow us to use the techniques developed for complex networks as an alternative to address problems in data clustering. However, it has been shown that both the method of graph construction and the choice of method parameters (i.e., sparsity) have a strong impact on the performance of graph-based clustering methods [27, 28, 29, 30]. Here, we study the use of multiscale community detection applied to similarity graphs extracted from data for the purpose of unsupervised data clustering. The basic idea of graph-based clustering is shown schematically in Figure 1. Speci\ufb01cally, we focus on the problem of assessing how to construct graphs that appropriately capture the structure of the dataset with the aim of being used within a multiscale graph-based clustering framework. In particular, we carry out an empirical study of di\ufb00erent graph construction methods used in conjunction with Markov Stability (MS), a dynamics-based framework for multiscale community detection [19, 31]. As a dynamics-based framework [31, 18, 31], MS provides a uni\ufb01ed framework for many multiscale community detection algorithms, such as the RB Potts model [32], the constant Potts model [33] and the absolute Potts model [34], and allows for the unsupervised community detection at di\ufb00erent levels of resolution. MS has also been shown to allow for the detection of both clique-like and nonclique-like communities in graphs [35]. MS has been applied successfully to a variety of problems, including protein structures [36, 37], airport networks [18], social networks [38] and neuronal network analyses [39]. Other dynamical processes have been extensively applied in network analysis, such as temporal networks [40], crowded networks [41] and network classi\ufb01cation [42]. In this paper, we evaluate several geometric graph constructions, from methods that use only local distances to others that balance local and global measures, and \ufb01nd that the recently proposed Continuous k-nearest Neighbours (CkNN) graph [43] performs well for graph-based data clustering via community detection. We then show how the multiscale capabilities of the Markov Stability to scan across scales can be exploited to deliver robust clusterings, reducing the sensitivity to the parameters of the graph construction. In other words, a range of parameters in the graph construction lead to good clustering performance. We validate our graph-based clustering approach on real datasets and compare its performance to several other popular clustering methods, including k-means, mixture models, spectral clustering and hierarchical clustering [44]. The rest of the paper is structured as follows. We \ufb01rst introduce several methods for graph construction, apply them to eleven public datasets with ground truths, and evaluate the performance of graph-based data clustering on the ensuing similarity graphs. We then describe brie\ufb02y the Markov Stability framework for multiscale community detection, and use a synthetic example dataset to illustrate how the multi-resolution clustering reduces the sensitivity to graph construction parameters. Finally, we validate the Markov Stabiity graph-based clustering through comparisons with other clustering methods on real datasets. Graph construction methods for data clustering Let us consider a dataset consisting of n samples {yi}n i=1, where each sample yi is a d-dimensional vector. We will assume that we can de\ufb01ne a measure of pairwise dissimilarity between the samples: d(i, j) \u22650. In some cases, the vectors will be de\ufb01ned in a metric space, and the dissimilarity will be a true distance d(i, j) = ||yi\u2212yj|| \u22650. Other dissimilarity measures, such as the cosine distance, can also be used depending on the application. In the examples below, we will restrict ourselves to Euclidean distances as the measure 2 \fClustered points Data points Similarity Graph graph  construction graph  partitioning Graph communities Figure 1: Schematic diagram of the work\ufb02ow for graph-based clustering of dissimilarity. The high dimensionality of data usually leads to complex and non-linear geometries associated with datasets, posing challenges to standard clustering methods. The aim of transforming the data into a graph is to capture the complex geometry of the data through the graph topology [22], so as to reveal the structure of the data via graph-theoretical concepts and tools from complex network analysis. There exist a variety of ways to construct a graph from a high-dimensional dataset, invoking di\ufb00erent principles. Here we focus on geometric graphs and examine two of the most widely used methods (\u03f5-ball graph, k-nearest neighbour (kNN) graph) and three recent methods (continuous k-nearest neighbours (CkNN) graph [43], perturbed minimum spanning tree (PMST) [45] and relaxed minimum spanning tree (RMST) [23]). These methods can be broadly ascribed to two categories depending on how they use the geometry of the data, as follows. Neighbourhood based methods: \u03f5-ball, kNN and CkNN graphs The idea of neighbourhood based methods is to connect two nodes if they are local neighbours, as given by their pairwise distance d(i, j). The two simplest and most popular ways to construct a graph from pairwise distances are the \u03f5-ball graph and the k-nearest neighbour graph (kNN): in the \u03f5-ball graph, any two points at a distance smaller than \u03f5 are connected; in the kNN graph, every point is connected to its k-th nearest neighbours. These two methods capture the local information of the data but are highly sensitive to the parameters \u03f5 or k [27]. The parameter is usually set according to the density of the data points but in many datasets, the data points are not uniformly distributed. A recently proposed method that can resolve this problem is the continuous k-nearest neighbours (CkNN) graph [43]. If d(i, j) is the distance between sample i and sample j, and dk(i) is the distance between sample i and its k-th nearest neighbour, the CkNN graph is constructed by connecting sample i and sample j if d(i, j) < \u03b4 q dk(i)dk(j), (1) where \u03b4 is a positive parameter that controls the sparsity of the graph. Through this construction, the topology of the CkNN graph captures the geometric features of the data with the additional consistency that the CkNN graph Laplacian converges to the Laplace-Beltrami operator in the limit of large data [43]. It should be kept in mind that in CkNN the distance d(i, j) must be a metric to ensure geometrical consistency, whereas kNN graphs can be generated from any dissimilarity measure (i.e., one does not need a true distance) since only the ranking of node closeness matters. Minimum spanning tree based methods: PMST and RMST graphs A di\ufb00erent class of approaches for graph construction attempt to capture the global geometry of the overall dataset by constructing graphs based on measures of global connectivity of the ensuing graph. A popular way to ensure such global connectivity is through the minimum spanning tree (MST) [46], as follows. If we consider the matrix of all pairwise distances d(i, j) as the adjacency matrix of a weighted, fully connected graph, the MST is the subgraph such that all the nodes are path connected and the sum of edge weights is minimised. In other words, the MST provides a graph that connects all the points in the dataset with minimal global distance. MST-based approaches can thus capture the geometry of in-homogeneously sampled 3 \fdata points in a high-dimensional space since the MST contains not only local but also global features of the dataset. In its simplest form, the MST is sometimes added to sparse neighbourhood graphs as a means to guarantee global connectivity of the dataset, i.e., the \ufb01nal graph is the union of the MST and a kNN graph with a small k. These schemes, which are sometimes referred as MST+kNN graphs, are the ones we adopt by default in our neighbourhood constructions. However, the global properties of the MST can be exploited to generate MST-based graphs from data with distinct properties. We have explored here the use of two such MST-based algorithms: the perturbed minimum spanning tree (PMST) [45] and the relaxed minimum spanning tree (RMST) [23, 47]. In PMST [45], each data point yi is perturbed by a small amount of noise of standard deviation si = rdk(i) (r \u2208[0, 1]) where dk(i) is used as an estimation of the local noise and r is a parameter controlling the level of noise. The MST is then computed for each realisation of the perturbed data and the process is performed repeatedly to generate an ensemble of MSTs. The PSMT graph is given by the union of all the perturbed MSTs, plus the original MST. The intuition behind this algorithm is that random perturbations of the points in high-dimensional space will induce changes inhomogeneously in di\ufb00erent parts of the MST, depending on how globally important certain edges of the graph are, i.e., globally important edges will be consistently captured across all MSTs in the perturbed ensemble. One limitation of this algorithm is the heavy computational demand, since both the distance matrix and MST need to be computed for each random realisation. The computational burden makes it impractical to sweep over the parameters of the PMST, so we \ufb01x r = 0.5 and k = 1 in this paper. RMST [47] proposes a di\ufb00erent heuristic for estimating an MST-based graph with lower computational cost. Note that any two nodes are connected by a single path in the MST and we denote the longest edge on this path as dmax path(i,j). In RMST, the samples yi and yj are connected if d(i, j) < dmax path(i,j) + \u03b3 \u0000dk(i) + dk(j) \u0001 , (2) where \u03b3 > 0 is a parameter that weights the local density (measured by the average distance to the k-th neighbours of yi and yj ) against a global property (the maximum distance found on the MST path linking the samples i and j). Similarly to the parameter \u03b4 in CkNN, the value of \u03b3 controls the sparsity of the resulting graph. The RMST construction was proposed as a means to reconstruct data that have been inhomogeneously sampled from continuous manifolds, and has been shown to provide good description of datasets when preserving a measure of continuity (due to temporal or parametric changes) is important [23]. In Figure 2, we use a synthetic dataset of four groups of points sampled from a geometric structure (a noisy circle and its centre) to illustrate the di\ufb00erent graph construction schemes [48]. The graph representations allow us to gain intuition about the suitability of the di\ufb00erent graph constructions for clustering, and the e\ufb00ect of their parameters. For instance, the sparsity of the RMST graph is controlled by the parameter \u03b3. Note that RMST gives a similar graph to PMST with much lower computational cost. Note also that, for the same number of neighbours k, the CkNN gives a sparser graph than the kNN. To make CkNN and kNN comparable, we thus \ufb01x the parameter \u03b4 = 1 in CkNN and vary k. As discussed above, the MST-based methods are not optimised for clustering but aimed at manifold learning. It follows from (2) that, in the RMST (and PMST) graphs it is more likely for an edge to appear between node i and node j if the longest edge on the MST path between node i and j is large. This feature makes the geodesic distance on the graph a good approximation to the true distance in the underlying space. Hence both RMST and PMST are closer to manifold learning approaches such as Isomap [22]). In contrast, the kNN and CkNN graphs tend to have better modular structures and hence appear as potentially more suitable for graph partitioning. We compare these issues in detail in Section Tests on benchmark real datasets below. Markov Stability for graph-based clustering Let us consider a graph representing the dataset. The unweighted and undirected graph with n nodes representing the dataset is encoded by the adjacency matrix A, where Aij = 1 if there is an edge connecting node i and j and Aij = 0 otherwise. The degree of the nodes is summarised in the degree vector d where di = Pn j Aij, and we also de\ufb01ne the diagonal degree matrix D where Dii = di. The total number of edges of 4 \fData MST PMST k=1, \u03b3=0.25 k=1, \u03b3=0.5 k=1, \u03b3=1 k=12 k=9 k=7 \u03b4=1, k=12 \u03b4=1, k=9 \u03b4=1, k=7 (a) (b) (c) RMST (d) kNN (e) CkNN (f) Figure 2: Illustration of the di\ufb00erences introduced by graph construction methods. (a) Synthetic dataset consisting of four groups of randomly sampled points: three groups are sampled from a circle and one group from its centre. (b) The minimum spanning tree (MST) of the dataset based on Euclidean distances. We also present the graphs constructed with MST-based methods: (c) PMST and (d) RMST, as well as neighbourhood methods: (e) kNN and (f) CkNN. RMST, kNN and CkNN are shown for di\ufb00erent parameters that vary the sparsity of the resulting graphs. 5 \fthe network is m = P i,j Aij/2. We then apply multiscale community detection to extract relevant subgraphs in an unsupervised manner using the framework of Markov Stability. Multiscale community detection with Markov Stability Markov Stability is a quality measure for community detection which adopts a dynamical perspective to unfold relevant structures in the graph at all scales as revealed by a di\ufb00usion process [17, 18, 19, 35]. Consider a continuous-time Markov process on the graph governed by the dynamics \u02d9 p = \u2212p(I \u2212M) where p is an n-dimensional row vector de\ufb01ned on the nodes and M = D\u22121A is the one step random walk transition matrix. For this Markov process, there is a unique stationary distribution \u03c0 = dT /2m. Let us denote the autocovariance matrix of this process as B(t) = \u03a0P(t) \u2212\u03c0T \u03c0 where \u03a0 = D/2m encodes the stationary distribution and P(t) = exp(\u2212t(I \u2212M)) is the transition matrix. Given a partition g of the nodes into c non-overlapping groups denoted by g = {g1, g2, ..., gc}, the Markov Stability of g is de\ufb01ned as: r(t, g) = c X s=1 X i,j\u2208gs B(t)ij. (3) A partition has a high value of (3) if the probability of \ufb01nding a random walker at time t within the group where it started at t = 0 is higher than that expected by mere chance. In this sense, Markov Stability is a quality function for partitions of a graph and the objective is therefore to \ufb01nd the partitions that achieve high values of the Markov Stability as a function of t: r\u2217(t) = max g r(t, g) achieved by the partition g\u2217(t). (4) The (time) parameter t is the so-called Markov time, and can be understood as the resolution parameter that leads to multiscale community detection [31, 35].For small t, the number of detected communities is large and the communities capture the local information of the graph. As t becomes larger, there are fewer communities and the communities are able to capture the global features of the graph. Computationally, the Markov Stability is optimised at di\ufb00erent Markov times through a version of the Louvain algorithm [49]. Through the computational maximisation (4), Markov Stability detects optimised partitions g\u2217(t) at all scales, parameterised by the value of t. However, we are interested in \ufb01nding robust partitions and robust scales, in the sense that a partition is found to optimise MS over a long interval of Markov time. We thus compute the dissimilarity between the obtained partitions at di\ufb00erent times t and t\u2032: V I(t, t\u2032) = V I(g\u2217(t), g\u2217(t\u2032)), (5) where we use the variation of information (V I) [50] as the metric of dissimilarity between partitions. If g\u2217(t) is a robust partition, the partition g\u2217(t\u2032) found at a Markov time t\u2032 close to t should be very similar, and hence V I(t, t\u2032) will be small. We therefore look for large diagonal blocks of small values in the V I(t, t\u2032) matrix. Such blocks correspond to a robust scale with an associated robust partition. As an additional feature, we look for optimised partitions that are also robust to the Louvain optimisation. Since the Louvain method is a greedy algorithm dependent on the random initialisation, the consistency of the output of the algorithm can be used as an indicator of the robustness of the solution. At each t, we run the Louvain optmisation multiple times and if the Markov time corresponds to a robust scale, the output partition should be always the same. Therefore we expect a low value of the average variation of information of the optimised partitions at time t V I(t) = 1 nL(nL \u22121) nL X s=1 nL X s\u2032=1 V I(gs(t), gs\u2032(t)), (6) where the Louvain algorithm is run nL times. Depending on the structure of the graph, several such robust scales and associated graph partitions might be found, which can then be used as the basis of unsupervised data clustering. 6 \fUsing Markov Stability for data clustering We illustrate the application of MS to data clustering through the synthetic dataset in Figure 3. The example dataset has geometric struture and is designed to have two scales, so that it can be divided into 3 big clusters or 9 small clusters. First, we construct an unweighted CkNN graph (k = 7, \u03b4 = 1.8) and apply MS as described above. We optmise the Markov Stability (3) for nT Markov times t \u2208[1, 1000], and at each t, we run the Louvain algorithm nL = 500 times. For each Markov time, we record the partition with the maximal Markov Stability, g\u2217(t), and the average dissimilarity of the partitions found in the nL optimisations, V I(t) (6). Once the scan across Markov time is completed, we also compute V I(t, t\u2032), the matrix recording the dissimilarity of the optimal partitions found across the scan. The results are presented in Figure 3, where, as a function of Markov time t, we plot the number of communities in the optimal partition g\u2217(t); the the optimised Markov Stability r\u2217(t) (3); the average dissimilarity due to algorithmic variability V I(t); and the dissimilarity of partitions across time given by the V I(t, t\u2032) matrix. The diagonal blocks of low values of V I(t, t\u2032) (which also correspond to plateaux in the number of communities) and the low values (or dips) of V I(t) suggest that there are two relevant scales in this graph, which correspond to a \ufb01ner partition into 9 groups (at small t) and a partition into 3 groups (at larger t > 200). The inset shows that the partition recovers the planted groups of this synthetic example. The robustness of the partitions found across scales is further examined in Figure 4. To understand how the graph partitions evolve with t, we compute the V I metric between all the partitions found across the Markov time scan (nL \u00d7 nT ) and project them on a low dimensional space using multidimensional scaling (MDS). In Figure 4, we use the \ufb01rst MDS coordinate of each of the partitions as a function of t coloured by its number (frequency) of appearances out of the nL Louvain runs at each Markov time. Several partitions can coexist at a given Markov time, but our numerics show that the two robust partitions (c = 9 and c = 3) have a long-lived high frequency of appearance when the t matches the corresponding resolution. Between c = 9 and c = 3, other partitions of lesser robustness appear through mergers of clusters one by one, as shown by the Sankey diagramme in Figure 4, and only exist for short Markov times until the robust partition of c = 3 appears. One advantage of using community detection for data clustering is the computational e\ufb03ciency of fast community detection algorithms [51]. Empirically, the Louvain algorithm scales nearly linearly with the size of the dataset [49]. In its full form, Markov Stability has been applied to community detection in networks of sizes up to tens of thousands of nodes [36, 20]. To give an indication, running a full MS scan for a dataset of 2310 samples sweeping 100 Markov times with 100 Louvain runs at each Markov time on a desktop \u2217requires about 20 minutes. The space complexity of the graph-based clustering is dominated by the storage of the graph, i.e., similar to spectral clustering if the adjacency matrix is used. For very large datasets(> 20000), the computational complexity is dominated by the matrix exponential in P(t). The use of linearised (approximate) version of MS allows to scale its use up to graphs with hundreds of thousands of nodes with reduced CPU times, at the cost of some reduction in the quality of the clusterings [31]. The linearised versions also allows the storage of the graph in a sparse matrix to reduce the space complexity. Scanning across Markov time reduces the sensitivity to graph construction Graph-based clustering performance is sensitive to the parameters of the graph construction method, which modulate sparsity, but there is no easy way to select the best parameter if the ground truth is unknown. In practice, the parameters are usually set empirically with little guidance that the chosen parameter will lead to good clustering results for a particular dataset. Within the Markov Stability framework, we can use the robustness provided by scanning across Markov time to reduce the sensitivity to the details of graph construction, thus improving the reliability of the detected clusters. To illustrate this idea, We use the same dataset in Figure 3 and construct CkNN (k = 7) graphs with di\ufb00erent values of \u03b4 = 1.5, 1.8, 2.4. Our numerics show that Markov Stability detects the relevant underlying scales of the data (c = 9 and c = 3) for the di\ufb00erent values of \u03b4, as shown by the long diagonal blocks of low V I(t, t\u2032) and the low values of V I(t) in Figure 5. Although the degree of the CkNN graph varies markedly with the parameter \u03b4, the two signi\ufb01cant scales are identi\ufb01ed by scanning across the Markov time. Hence the scanning across scales inherent to multiscale community detection provides additional robustness to the \u2217With a 4-core 3.4GHz CPU and MATLAB implementations 7 \f10 0 10 1 10 2 10 3 Markov time 10 0 10 1 10 2 Number of communities 10 -1 10 0 Stability 0 0.1 0.2 VI(t,t') 10 0 10 1 10 2 10 3 Markov time 10 0 10 1 10 2 Number of communities 0 0.005 0.01 0.015 0.02 0.025 0.03 Variation of information c = 9 c = 3 c = 9 c = 3 Figure 3: Graph-based unsupervised analysis using Markov Stability of a synthetic dataset with multiscale geometric structure. We generate a CkNN graph (k = 7, \u03b4 = 1.8) and scan across Markov time from 1 to 1000. Two signi\ufb01cant scales corresponding to robust partitions are identi\ufb01ed based on the zero blocks of the V I(t, t\u2032) matrix, the long plateaux in the number of communities, and the low values of V I(t). Note that with Markov Stability, there is a range of \u03b4 for the CkNN graph (k = 7) that can reveal the multiscale structure of the data (see Fig. 5). 8 \fc=3 A1 A2 A3 B1 B2 C1 C2 C3 C4 A1 A2 A3 B1 B2 C1 C2 C3 C4 A B C A B C 10 0 10 1 10 2 10 3 -0.2 -0.15 -0.1 -0.05 0 0.05 Markov time MDS coordinate of partitions appearing frequency 0 0.5 1 Louvain partitions partitions with maximum MS (a) (b) c=9 c=8 c=7 c=5 c=6 c=4 Figure 4: Robustness of the partitions found through Markov Stability for the CkNN graph with k = 7 and \u03b4 = 1.8 of the dataset in Figure 3. (a) We show the one-dimensional MDS embedding (based on the V I distance) of all the graph partitions found by the nL Louvain optimisations for each of the nT Markov times. The partition with the maximal Markov Stability at each Markov time is labelled by a \ufb01lled circle, whereas the non-optimal partitions are represented by open squares. The graph partitions are coloured according to the frequency of their appearance in the optimisations. (b) The Sankey diagram represents the relationship between the partitions found by MS across Markov time, from c = 9 to c = 3. Note the hierarchical structure which is not pre-imposed by the algorithm but emerges from the intrinsic properties of the dataset. parameters of the graph construction algorithm. We also carry out a similar analysis of the clusterings for two real datasets (\u2018WBDC\u2019 and \u2018Control charts\u2019). The clusterings remain robust when varying k in CkNN (see Figure S1 and Figure S2). Tests on benchmark real datasets We have tested several graph-based clustering approaches (both graph constructions and clustering methods) using eleven benchmark datasets from the UCI repository (see Table 1 for a summary of attributes) [52]. All the datasets have ground truth labels, which we use to validate the results of the di\ufb00erent methods. Comparison between graph constructions Starting from the Euclidean distance d(i, j) = ||yi \u2212yj||2, we generated geometric graphs from each of the eleven UCI datasets using the \ufb01ve graph construction methods described in Section Graph construction methods for data clustering. If the constructed graph is disconnected, the MST is added to the graph to ensure a connected graph. Each graph was analysed using Markov Stability to obtain optimised partitions at any scale, and we selected the closest of those partitions to the ground truth, as measured by the normalised 9 \fCkNN \u03b4=1.5, k=7 (Average degree 9.9) CkNN \u03b4=1.8, k=7 (Average degree 15.9) CkNN \u03b4=2.4, k=7 (Average degree 23.8) 10 0 10 1 10 2 10 3 10 4 Markov time 10 0 10 1 10 2 Number of communities 0 0.01 0.02 0.03 0.04 0.05 VI(t) c=9 c=3 0 0.1 0.2 VI(t,t') (a) (b) (c) 10 0 10 1 10 2 10 3 10 4 Markov time 10 0 10 1 10 2 Number of communities 0 0.01 0.02 0.03 0.04 0.05 c=9 c=3 VI(t) 0 0.1 0.2 VI(t,t') 10 0 10 1 10 2 10 3 10 4 Markov time 10 0 10 1 10 2 Number of communities 0 0.01 0.02 0.03 0.04 0.05 c=9 c=3 VI(t) 0 0.1 0.2 VI(t,t') Figure 5: The multiscale character of Markov Stability alleviates the sensitivity to the parameters of graph construction. Using the same synthetic dataset as in Figure 3, MS \ufb01nds the robust partitions and scales in the data (c = 3 and c = 9) in an unsupervised manner for CkNN graphs constructed with a range sparsities as given by varying the parameters: k = 7 and \u03b4 = 1.5, 1.8, 2.4 in (a), (b) and (c), respectively. Markov time, as a resolution parameter, allows the community algorithm to reveal the local and global properties of the graph constructed from the data. mutual information (NMI) [53]. The computed NMI is a quality index for the graph construction under clustering. We also compute the adjusted Rand index (ARI) [54] as an additional quality index. The results of the comparison are shown in Table 2. The kNN and \u03f5-ball graphs, both widely-used in many machine learning and both based on local neighbourhoods, give good results for a range of k. (Note 10 \fTable 1: Attributes of the eleven real datasets from the UCI repository used in this study. Dataset No. of samples, n Dimension, d No. of classes, c\u2217 Iris 150 4 3 Glass 214 9 6 Wine 178 13 3 WBDC 569 30 2 Control chart 600 60 6 Parkinsons 195 22 2 Vertebral 310 6 3 Breast tissue 106 9 6 Seeds 210 7 3 Image Segmentation 2310 19 7 Yeast 1484 8 10 that \u03f5 is set to be the average of the distances to the 7-th neighbour, d7(i).) For the MST-based methods, RMST achieves better performance for sparser graphs (with smaller \u03b3) when the cluster structure is not obscured by the objective of manifold reconstruction (Fig. 2). The same applies to the PMST graph. The empirical tests show that the CkNN graph gives the best average results over the eleven datasets for k = 7 and above. Hence we adopt the CkNN graph with k in the range of 7 to 12 as a good choice for graph-based clustering. Table 2: Comparison of graph constructions in terms of clustering performance (NMI and ARI) on eleven UCI datasets. A high NMI (and ARI) indicate that the best partition found by Markov Stability is similar to the ground truth, i.e., better clustering. The best performance for each dataset is marked in bold. Normalised Mutual Information (NMI) Dataset PMST \u03f5-ball RMST k = 1 kNN CkNN \u03b4 = 1 \u03b3=0.5 \u03b3=0.25 \u03b3=0.125 k=3 k=7 k=12 k=3 k=7 k=12 Iris 0.7764 0.8057 0.71060.7764 0.7400 0.7627 0.8057 0.8057 0.82260.7980 0.7777 Glass 0.3886 0.3862 0.36560.3646 0.3953 0.3863 0.3927 0.3626 0.3517 0.3941 0.4170 Wine 0.8080 0.7972 0.73890.8400 0.8633 0.7955 0.8336 0.8215 0.7528 0.8347 0.8113 WBDC 0.6042 0.5819 0.58390.6042 0.6182 0.6188 0.5944 0.6121 0.5822 0.72310.6113 Control chart 0.8272 0.8404 0.76020.7133 0.8130 0.8266 0.8520 0.852 0.8078 0.88830.8531 Parkinson 0.2220 0.2065 0.20120.2143 0.2811 0.31750.2815 0.2176 0.2113 0.2973 0.2423 Vertebral 0.5043 0.6273 0.54320.5323 0.5999 0.6060 0.5928 0.5424 0.64500.6018 0.6116 Breast tissue 0.5722 0.5616 0.52980.5820 0.5461 0.5447 0.5525 0.5243 0.5253 0.5648 0.5601 Seeds 0.7293 0.6943 0.73610.7570 0.6810 0.7515 0.7458 0.7318 0.6384 0.7381 0.7245 Image Seg. 0.6948 0.6605 0.67620.7072 0.7488 0.6154 0.6465 0.6667 0.6012 0.6347 0.6541 Yeast 0.2881 0.3051 0.29520.2626 0.2563 0.2764 0.2997 0.30800.2473 0.2959 0.3072 Average 0.5835 0.5879 0.55830.5776 0.5945 0.5910 0.5997 0.5859 0.5623 0.61550.5973 Adjusted Rand Index (ARI) Dataset PMST \u03f5-ball RMST k = 1 kNN CkNN \u03b4 = 1 \u03b3=0.5 \u03b3=0.25 \u03b3=0.125 k=3 k=7 k=12 k=3 k=7 k=12 Iris 0.7420 0.7592 0.66030.7420 0.6957 0.7191 0.7592 0.7592 0.81840.7455 0.7445 Glass 0.2323 0.2099 0.19830.2029 0.2258 0.2278 0.2231 0.2266 0.2134 0.2398 0.2496 Wine 0.8350 0.8072 0.73750.8712 0.8823 0.8025 0.8498 0.8349 0.7414 0.8471 0.8360 WBDC 0.7193 0.7114 0.70140.7193 0.7369 0.7368 0.7010 0.7310 0.6697 0.82440.7200 Control chart 0.6929 0.7364 0.56940.5371 0.6991 0.6748 0.6824 0.7071 0.6902 0.82800.7227 Parkinson 0.2267 0.2205 0.20380.1540 0.2556 0.2176 0.2001 0.2045 0.1165 0.26670.2101 Vertebral 0.5257 0.64450.57020.5411 0.6015 0.5982 0.5802 0.5330 0.6441 0.6113 0.6302 Breast tissue 0.46890.4494 0.41000.4017 0.3494 0.4012 0.4272 0.4078 0.3631 0.3764 0.4471 Seeds 0.7353 0.7497 0.74580.7589 0.6687 0.7876 0.78890.7761 0.6402 0.7764 0.7655 Image Seg. 0.6060 0.5144 0.60300.6193 0.5942 0.3800 0.4669 0.5471 0.3791 0.4522 0.5121 Yeast 0.1908 0.2368 0.18270.1772 0.1649 0.1755 0.2230 0.25310.1797 0.1942 0.2294 Average 0.5432 0.5490 0.50750.5204 0.5340 0.5201 0.5365 0.5437 0.4960 0.56020.5516 The CkNN graph is constructed by using a variable bandwidth di\ufb00usion kernel [55] where the bandwidth is inversely proportional to the sampling density and allows for uniform estimation errors over the underlying 11 \fmanifold, while the graph construction which uses a \ufb01xed bandwidth kernel will have large errors in areas of small sampling density. By such a construction, the graph Laplacian of the CkNN graph converges to the the Laplacian operator in the manifold. This explains why the CkNN graph shows a better clustering performance than the other graph construction when the graph is partitioned with Markov Stability, which considers a di\ufb00usion process on the graph. Comparison between clustering methods In this section, we evaluate the performance of graph-based clustering through Markov Stability against several other clustering methods applied to the datasets from the UCI repository. We include a variety of clustering approaches. Model-based methods include k-means and Gaussian mixture (clustering is repeated 50 times and partition with the best objective function is reported). We also apply hierarchical clustering with complete linkage and Euclidean distance is used as the distance measure. Since graph-based clustering is closely related to spectral clustering, we compare to two spectral clustering methods: the multiclass ncut algorithm [56] and the classic NJW algorithm [8]. The a\ufb03nity matrix for these two spectral clustering algorithms is calculated with a local density kernel as described in [57]. Note that all these algorithms need the number of clusters to be given as an input. Hence we use the number of classes in the ground truth c\u2217as an input to set the number of clusters. For comparability, in the case of Markov Stability, we construct the CkNN graph with Euclidean distance (k = 7 and \u03b4 = 1) and \ufb01nd the optimised partition with the number of clusters equal to c\u2217. The results are presented in Table 3. Given the distinct features of the datasets, no method is expected to achieve consistently better performance across all datasets. The hierarchical clustering performs worst, as it is easily a\ufb00ected by the noise in real datasets. Model-based methods, such as the Gaussian mixture model, can achieve good performance if the properties of the data \ufb01t the assumptions (e.g., the Iris dataset), but can also perform poorly. On datasets that have non-convex geometries, graph-based methods and spectral clustering tend to perform better. On average, the Markov Stability approach achieves the best NMI and ARI scores. To further validate the quality of the robust partitions found by Markov Stability, we also carried out MS clustering in a fully unsupervised manner, i.e., without providing the number of classes in the ground truth as an input. Using the principles described in Section Markov Stability for graph-based clustering, we identify robust scales and robust partitions in order to establish the number of clusters inherent to the data in an unsupervised manner. The number of clusters detected by MS and the clustering performances in terms of NMI and ARI are presented in Table 4, together with the results obtained in Table 3, where the number of clusters in the ground truth was provided. Although the number of detected clusters di\ufb00ers slightly from the ground truth, the clusters found in the unsupervised MS have comparable ARI values and a higher average NMI value which indicates that the clusters are of good quality and provide more information about the ground truth. This highlights the capability of Markov Stability as an unsupervised data clustering approach in practice where the number of clusters are usually unknown."
+                }
+            ],
+            "Qichao Wang": [
+                {
+                    "url": "http://arxiv.org/abs/2310.11778v3",
+                    "title": "Language Agents for Detecting Implicit Stereotypes in Text-to-image Models at Scale",
+                    "abstract": "The recent surge in the research of diffusion models has accelerated the\nadoption of text-to-image models in various Artificial Intelligence Generated\nContent (AIGC) commercial products. While these exceptional AIGC products are\ngaining increasing recognition and sparking enthusiasm among consumers, the\nquestions regarding whether, when, and how these models might unintentionally\nreinforce existing societal stereotypes remain largely unaddressed. Motivated\nby recent advancements in language agents, here we introduce a novel agent\narchitecture tailored for stereotype detection in text-to-image models. This\nversatile agent architecture is capable of accommodating free-form detection\ntasks and can autonomously invoke various tools to facilitate the entire\nprocess, from generating corresponding instructions and images, to detecting\nstereotypes. We build the stereotype-relevant benchmark based on multiple\nopen-text datasets, and apply this architecture to commercial products and\npopular open source text-to-image models. We find that these models often\ndisplay serious stereotypes when it comes to certain prompts about personal\ncharacteristics, social cultural context and crime-related aspects. In summary,\nthese empirical findings underscore the pervasive existence of stereotypes\nacross social dimensions, including gender, race, and religion, which not only\nvalidate the effectiveness of our proposed approach, but also emphasize the\ncritical necessity of addressing potential ethical risks in the burgeoning\nrealm of AIGC. As AIGC continues its rapid expansion trajectory, with new\nmodels and plugins emerging daily in staggering numbers, the challenge lies in\nthe timely detection and mitigation of potential biases within these models.",
+                    "authors": "Qichao Wang, Tian Bian, Yian Yin, Tingyang Xu, Hong Cheng, Helen M. Meng, Zibin Zheng, Liang Chen, Bingzhe Wu",
+                    "published": "2023-10-18",
+                    "updated": "2023-11-02",
+                    "primary_cat": "cs.CY",
+                    "cats": [
+                        "cs.CY",
+                        "cs.CL"
+                    ],
+                    "main_content": "INTRODUCTION Recent advancements in text-to-image generative models, such as DALL-E 2 (Ramesh et al., 2022) and Stable Diffusion (Rombach et al., 2022), have garnered significant attention in both academia and industry. These state-of-the-art (SOTA) image generation models have facilitated the development of novel creative tools and applications, such as Midjourney1 and NovelAI2. However, alongside their remarkable success, emerging studies (Bianchi et al., 2023a) have begun to reveal inherent and implicit social stereotypes within these models, posing ethical risks when deployed on a large scale. For instance, Naik & Nushi (2023) conducted a preliminary analysis of existing models, illustrating how they perpetuate stereotypes related to race, gender, and other social identities using a limited set of handcrafted prompts. As Fig.1 shows, when given a prompt such as \u201da terrorist,\u201d most \u2217Equal contribution. The project was conducted during the internship in AI Lab, Tencent. \u2021 Corresponding authors. 1https://www.midjourney.com 2https://novelai.net 1 arXiv:2310.11778v3  [cs.CY]  2 Nov 2023 \fUnder review of the generated images may disproportionately depict a Middle Eastern, thereby perpetuating the stereotype that Middle Easterners are more likely to be terrorists. Although these research works serve as a crucial initial step, the scope of their evaluations remains confined, primarily focusing on a closed environment setting, where static test sets within predefined crafted rules are constructed and are then employed to elicit potential biases inherent to the given models (e.g., in Naik & Nushi (2023)). For instance, Bianchi et al. (2023b) approached the issue by iterating through various personal traits and professions in conjunction with, designing corresponding prompts to test. Notably, the sources and types of stereotypes can be diverse (Caliskan et al., 2017; Bourdieu, 2018; Kozlowski et al., 2019; Jha et al., 2023) in the real-world open environments (Breitfeller et al., 2019; Schmidt & Wiegand, 2017). At the same time, the open-source community has witnessed a surge in the usage of text-to-image models and their corresponding style plugins. For instance, on the civitai platform3, there are approximately 3k distinct style plugins (Hu et al., 2022) publicly available. The widespread distribution and usage of these models and plugins present significant challenges for ethical oversight and risk governance targeting these models (Cho et al., 2022). The diversity of potential stereotypes, combined with the rapid expansion of available plugins, suggests that neither iterating through every conceivable configuration nor the ad-hoc selection of a narrow subset of traits is ideal, as they are either computationally expensive or possibly tainted by human cognitive biases. This essential tension highlights an urgent need for a more efficient yet comprehensive framework of stereotype risk evaluation, which could benefit the community in various important ways. Given the pervasive adoption of AIGC technology and its integration in commercial products, e.g. DALL-E 2 and Midjourney, such a framework will not only enhance our intellectual understanding of bias and stereotypes in foundational models, but also offer novel accessible tools that are accessible for industry professionals and regulatory entities focused on related governance. The people who  are politic elites The people who  are terrorists The religion member who are  traitors. Gender Race Religion Midjourney ChilloutMix Dreamshaper SDVN-RealArt Male Middle  eastern Christian Figure 1: Simple user prompts can inadvertently lead to the generation of images that reinforce and amplify stereotypes across various social dimensions, as observed in different models.. Among these models, Midjourney stands out as the most popular commercial product. Additionally, there are three popular open-source models: SDVN-RealArt, Dreamshaper, and ChilloutMix. These models specialize in a realistic style, illustration style, and animation style respectively. In this study, we focus on three fundamental social dimensions: gender, race, and religion. and incorporate a wide range of subgroups under each of them. Building on recent progress in large language models (LLMs) augmented agents (Shinn et al., 2023; Xu et al., 2023; Yao et al., 2022), we present a preliminary exploration into the construction of a language agent dedicated to stereotype risk governance. Our primary objective is to harness the robust reasoning capabilities of LLMs to invoke various multimodal models for stereotype evaluation. In particular, our proposed agent architecture adheres to the React paradigm (Yao et al., 2022), integrating LLMs with an array of visual tools, namely multimodal models (Radford et al., 2021; Li et al., 2022). This integration facilitates the extraction of stereotype prompts, which are subsequently employed to detect potential biases inherent in the target model. As illustrated in Figure 2, the user might pose a query such as \u2018Is Chilloutmix model racially stereotyped?\u2019 In response, the agent invokes specific tools to formulate 3https://civitai.com 2 \fUnder review prompts related to racial stereotypes and the relevant target subgroup. Following this, the target model is invoked to generate images. Ultimately, the agent computes a stereotype score based on the generated images and furnishes a comprehensive analytical report. From a broader perspective, the operational procedure of our agent can be segmented into three primary phases: (1) Instruction Pair Generation: Generate instruction pairs using toxic natural language processing (NLP) datasets, which primarily stem from the open text of various social platforms. These datasets are believed to be highly reflective of social stereotypes. The instruction pair is defined as a combination of \u201cprompt\u201d and \u201csubgroup\u201d, such as (\u201cterrorist\u201d, \u201cMiddle Eastern\u201d), indicating that when the drawing prompt is \u201cterrorist\u201d, the text-to-image model might tend to generate images with \u201cMiddle Eastern\u201d stereotype. (2) Image Generation: Call the corresponding generated models to get images that meet the basic settings. (3) Stereotype Detection: Use the vision-language model to identify the subgroups of each generated image and calculate the final stereotype score. To rigorously assess the efficacy of our agent, especially when juxtaposed against models employed in proprietary products, the pivotal criterion is the detection performance in the third phase. To facilitate this, we instruct our agent to curate a benchmark comprising 4123 instruction pairs from 5 public toxicity textual datasets, followed by the generation of a multitude of images corresponding to each instruction. A random subset, constituting 10% of these images, is manually annotated to ascertain the presence of a designated group. This methodology enables us to gauge the agent\u2019s proficiency in discerning biases within text-to-image models. Our preliminary investigations, encompassing Midjourney and widely-recognized open-source models, reveal that our agent\u2019s performance closely mirrors human evaluation accuracy, with an average precision of 92.15%. This result underscores the potential of integrating our proposed agent into real-world bias governance frameworks. Furthermore, our experimental results have revealed crucial insights into the behavior of text-to-image models. Specifically, when provided with drawing instructions related to certain personal characteristics, social culture, and criminal violence, these models tend to generate images that exhibit bias towards certain groups, underscoring potential ethical concerns and lasting harm. The overall contribution is as follows: \u2022 To our knowledge, we are the first to build a language agent for stereotype detection and governance against text-to-image models. Importantly, the use of language agent enables a general automatic framework that generates and examines potential stereotypes at scale, without the need to iterate through all possible combinations or resort to ad-hoc selection of a limited set of traits. \u2022 We propose the evaluation benchmark to assess the capabilities of the proposed agent framework. This framework is constructed using the toxic text dataset from social platforms, thereby aligning more closely with the prevalence of stereotypes in daily expressions. By utilizing this dataset, our evaluation process ensures a more realistic representation of bias occurrences in real-world scenarios. \u2022 Our experimental results show our agent is not only close to human annotation in detection accuracy, but also adapt well to the requirements of various detection tasks. 2 AGENT DESIGN Our agent framework is an automated evaluation system for detecting stereotypes in text-to-image models, which is composed of task planning and tool usage. 2.1 TASK PLANNING In order to enable our agent to comprehensively understand any stereotype detection request from users, we employ LLMs as the reasoning engine of the agent. This entails utilizing LLMs for intent comprehension and task decomposition when processing user requests. Specifically, we leverage few-shot prompting techniques to empower LLMs to generate task planning trajectories following the ReAct paradigm (Yao et al., 2022). In the setting of this study, few-shot prompting refers to the practice of injecting user requests along with their corresponding task planning trajectories as examples within the context provided to the LLMs. For this study, we have selected five represen3 \fUnder review tative user requests, and their associated task-planning trajectories have been authored by human experts. Each trajectory is composed of multiple thought-action-observation steps. For each substep, \u201cThought\u201d is a natural language description of the planning process. This description is derived from an analysis of the current step\u2019s task objective and the corresponding external tool information needed, based on the context information aggregated from previous steps. \u201cAction\u201d represents the concrete tool execution distilled from the description in the current \u201cThought\u201d field, while \u201cObservation\u201d indicates the information returned after the tool execution. After presenting these case examples, the LLMs can generate planning trajectories that adhere to the specified format requirements for a given user request. A concrete example of planning trajectory is shown in Figure 2. The details of tool usage are introduced in the next section. Is Chilloutmix model racially stereotyped? Does the following content reflect  stereotypes in text-to-image domain? \u201cThe political elites are all men,  which is unfair for women! \u201d Does SD-XL model contain stereotypes  towards Asian? Real-world user query What are the stereotype risks in the  deployment of the DreamShaper model? Planning/ Reasoning Action Observation Thought What types of prompts tend to exhibit  stereotypes on the SDVN3 model? Tool usage Task solving trajectory Text-to-image  generation BLIP-based  detection Instruction  optimization Stereotype  Score calculator Instruction  generation Tool Pocket: Thought 1: I need to identify the target model and social dimension.      Do I need to use a tool? Yes Action 1: \u201cIntention comprehension\u201d, args: {\u2018task_description\u2019}. Obs 1: {Model: Chilloutmix, Dimension: Race} Thought 2: After identifying the model and dimension, the next step  is to generate instructions. Do I need to use a tool? Yes Action 2: \u201cInstruction retrieval\u201d, args: {\u2018model\u2019, \u2018dimension\u2019} Obs 2: {\u2018People who are thugs\u2019, \u2018African\u2019}... Thought 3: After getting specific instructions, use the corresponding  model to generate the images. Do I need to use a tool? Yes Action 3: \u201cImage generation\u201d, args: {\u2018model\u2019, \u2018instruction_pair\u2019} Obs 3: {\u2018image_1.jpg\u2019, ...} Thought 4: Next, each image needs to be classified into subgroups.  Do I need to use a tool? Yes Action 4: \u201cBLIP-based detection\u201d, args: {'image_path', \u2018dimension\u2019} Obs 4: {\u2018image_1.jpg\u2019: African, 'image_2.jpg': 'African'...} Thought 5: Next I need to Calculate stereotype score based on  classification results.  Do I need to use a tool? Yes Action 5: \u201cStereotype Score Calculator\u201d, args: {'label\u2019} Obs 5: {Score: 0.910}  Intension  understanding Figure 2: Overview of our detection agent for text-to-image models. We illustrate the task-solving trajectory of the specific detection task by taking the user query in the dotted box on the left as an example. 2.2 TOOL USAGE In order to enable the agent to complete each subgoal, we establish various tools and interfaces of generative models. Furthermore, we delineate the specific usage scenarios for each tool, along with the arguments that need to be supplied. In the following, we introduce in detail the usage of various tools. (1) Intention understanding: This tool is designed to understand user requests and extract key detection elements. These elements include the target model, social dimensions, and the open text to be detected. However, it is important to note that not all task descriptions will necessarily contain all of these elements, as users may provide descriptions in various formats. (2) Instruction Generation: The tool transforms the open text into an instruction pair, formatted as (prompt p, subgroup g). For instance, (\u201dterrorists\u201d, \u201dMiddle Eastern\u201d) comprises the descriptive object \u201dterrorists\u201d and the demographic subgroup \u201cMiddle Eastern\u201d. We establish the task-specific prompt as \u201cExtract potential descriptions and target groups within the given content and return them in the format (\u2018description\u2019, \u2018target group\u2019). Here, the \u2018description\u2019 typically includes adjectives, nouns, or action verbs associated with the stereotype, while the \u2018target group\u2019 refers to the demographic subgroup, e.g. \u2018Asian\u2019 and \u2018Middle Eastern\u2019.\u201d (3) Instruction Retrieval: This tool is aimed at detection tasks that do not involve specific open text. It takes social dimensions G as input and retrieves the corresponding bias instruction pairs I whose subgroup g \u2208G from our instruction pair dataset. The specific creation process of the dataset is in 3.2. Based on the evaluation results from our benchmarks, the agent selects the instruction pair I that exhibits the most stereotypes among the text-to-image models. (4) Prompt Optimization: The prompt P in the instruction I is often composed of expressions that directly describe stereotypes, and specific optimization is often required to actually be used for text-to-image models. Therefore, we set up a template as \u201dThe people who\u201d to support adding different description expressions. In addition, the above initial prompt may cause the problem of missing 4 \fUnder review subject objects in the generated images, and further emphasis on subject keywords is required, such as \u201c(person, 1.5)\u201d. (5) Stereotype Score Calculator: The stereotype degree of the model in response to a specific prompt can be quantified by the conditional probability P(G|P), which represents the probability that the generated images are associated with the specific social group g given the painting prompt p. Specifically, we define the stereotype score as the proportion of the number of majority subgroup images to the total number. 3 AGENT BENCHMARK In this section, we present a comprehensive evaluation of prominent text-to-image models using our proposed agent. We begin by introducing the Text-to-Image models used in the evaluation. Then, we present the construction process for the benchmark dataset used to evaluate these models. Next, the experimental settings are introduced. Finally, we report and discuss the benchmark evaluation results. 3.1 TEXT-TO-IMAGE MODELS To comprehensively assess the variant Stable Diffusion models, we select three main custom styles for evaluation: \u201drealism\u201d, \u201danime\u201d, and \u201dartistic\u201d. We choose the most downloaded models in the Civitai community corresponding to each style. Moreover, we also examine the performance of the SDXL model and the integration of different LoRA modules with these base models. These modules are also selected from the most popular model plugins within the same community. 3.2 BENCHMARK DATASET To investigate potential stereotypes in the current Text-to-Image models, we instruct our agent to construct a benchmark dataset, the Stereotypical Prompts for Image Generation (SPIG) Dataset, by collecting the stereotypical prompts and their corresponding groups from publicly available toxicity text datasets. The stereotypical prompts are mainly extracted from five existing toxicity text datasets that cover different demographic groups and stereotypical topics from various social media platforms. Below, we provide a brief introduction for each dataset as shown in Table 1: Dataset # of prompts Type of annotations SBIC (Social Bias Inference Corpus) (Sap et al., 2019) 150K Structured annotations of social media posts HateExplain (Mathew et al., 2021) 20K hate/offensive/normal classification for posts DYNAHATE (Vidgen et al., 2020) 40K Fine-grained labels for the hate type and target group IHC (Implicit Hate Corpus) (ElSherief et al., 2021) 9.5K Target demographic group and implied statement SMTD (Social Media Toxicity Dataset) 4 1K Toxic comments from social media platforms Table 1: Statistics of toxic text datasets. First, we employ the Instruction Generation tool to utilize samples containing toxicity text from the dataset and filter out instruction pairs with strong stereotypes. Subsequently, we utilize the Prompt Optimization tool to optimize the generated drawing prompts, making them easier for text-to-image models to comprehend and render. We extract a total of 4123 valid painting prompts from all toxic datasets. We categorize the corresponding groups into three categories: gender, race, and religion. Each category encompasses various subgroups; for example, the \u201crace\u201d category includes subgroups such as \u2018African\u2019 and \u2018Asian\u2019. The overall distribution is shown in Fig. 3(a), \u201dgender\u201d accounts for 55% of the total prompts, \u201drace\u201d makes up 33.6%, and \u201dreligion\u201d represents 11.5%. The distribution of prompts within each category is shown in Figs. 3(b)-3(d). 5 \fUnder review (a) Overall distribution. (b) Distribution of gender and sex. (c) Distribution of race and ethnicity. (d) Distribution of religion. Figure 3: Overall distribution and proportion of stereotype pair for each subgroup. 3.3 BENCHMARK SETTING After extracting the corresponding instruction pairs from the aforementioned toxic datasets, our agent systematically invokes each model and plugin in the benchmark to generate images associated with the respective prompt. To ascertain the stereotype degree exhibited by each model in response to a specific prompt, we employ manual annotation to obtain the ground-truth label for each image. Due to the cost and efficiency issues of manual annotation, we carefully select the instruction pairs to be annotated. We select a representative set of instructions from our constructed benchmark dataset. The selection process aims to ensure that the chosen instructions cover a diverse range of demographic groups and stereotypical prompts. We perform stratified sampling to ensure that each social dimension and its corresponding subgroups are proportionally represented in the selected prompts. Finally, our agent invokes the Stereotype Score Calculator tool to compute the stereotype score of generated images, thereby determining the presence of stereotypes. Gender Race Religion Social dimension 0 0.2 0.4 0.6 0.8 Stereotype prompt ratio SD-XL ChilloutMix Realistic Vision V20 DreamShaper Anything SDVN3 SD+Doll Likeness SD+Concept Scenery Scene Figure 4: Stereotype degree of different models in different social dimensions and models Occupation Crime related Appearance and lifestyle Political and social Culture related Character of prompt 0.0 0.2 0.4 0.6 0.8 1.0 Proportion of subgroups European African Asian Middle eastern Latin Figure 5: Proportion of racial subgroups under different prompt types. Violence related Political  and social Culture related Character of prompt 0.0 0.2 0.4 0.6 0.8 1.0 Proportion of subgroups Christian Muslim Jew Buddhist Hindu Figure 6: Proportion of religion subgroups under different prompt types. 3.4 BENCHMARK RESULT After our agent invokes corresponding tools, we obtain the ground-truth label for each image, as well as the assessment of whether each model exhibits stereotypes in response to each prompt. In this section, we analyze the distribution of stereotype prompts across various levels, including the models, social dimensions, and demographic subgroups, thereby confirming the extensive presence of stereotypes in the text-to-image domain. \u2022 Stereotypes of text-to-image models: As shown in Fig. 4, we analyze stereotypes of instruction pairs in the SPIG dataset across different models and plugins , where different colors represent the number of instructions that express stereotypes under different models. Among these models, the \u201cchilloutmix\u201d and \u201cRealistic Vision\u201d markedly display the most significant stereotyping, notably within the \u2018Gender\u2019 dimension. In this dimension, the model exhibits a larger stereotypical deviation in comparison to other models. This discrepancy may be attributed to an overrepresentation of female images within the training data, potentially catering to user preferences. \u2022 Stereotypes across social dimensions: Fig. 4 also compares the differences in stereotypes between different social dimensions. Among the three social dimensions, religion emerges as the dimension most heavily stereotyped. This situation brings to light two key issues. Firstly, it 6 \fUnder review underscores the scarcity of diverse religious content in the training data. Secondly, it mirrors the significant stereotype problem inherent in religious content within societal and cultural contexts. \u2022 Stereotypes across demographic subgroups: In relation to race and religion, two social dimensions where stereotypes are particularly prevalent, we further analyze the impact of prompt type on specific subgroup stereotypes. As depicted in Fig. 56, we examine the distribution of subgroups across all generated images. As analyzed in Bianchi et al. (2023b), it can be observed that certain specific types of prompts are often correlated with particular subgroups. For example, for prompts involving violence, the images generated are often related to race subgroups such as African and Latino, or religious subgroups such as Mulism. 4 AGENT PERFORMANCE To verify the effectiveness of our agent in reasoning and invoking tools in a real-world detection environment, we set up a series of model stereotype detection tasks to evaluate the agent\u2019s performance. We first introduce the task design used to evaluate agent performance, and then analyze the agent\u2019s overall performance on the stereotype detection task. Subsequently, we analyze the accuracy of key tool results in the agent\u2019s inference trajectory. Ultimately, we apply the agent to evaluate the commercial product, Midjourney, as a case study. 4.1 DETECTION TASK DESIGN We first assess the degree of stereotyping in six models and plugins within the benchmark. Given that the stereotyping degree of these models across various social dimensions and prompts is determined through manual annotation, we can compare the agent\u2019s evaluation results with the manual assessment outcomes to accurately evaluate the agent\u2019s effectiveness. In order to simulate user questions as much as possible, we employ LLMs to blend multiple models or plugins, and three social dimensions into sentence structures that mirror potential user query formats. For instance, \u201cCan you tell me whether the SD-XL model exhibits racial stereotypes?\u201d. In addition, we also randomly add open text to the queries, which are unseen in the few-shot cases, to test the agent\u2019s extraction capabilities. By ensuring a diverse range of question formats, we enhance the reliability of the agent\u2019s evaluation results. We generate 120 task queries to evaluate the model\u2019s detection capabilities. 0.0 0.2 0.4 0.6 0.8 1.0 Manual annotation ratio 0.0 0.2 0.4 0.6 0.8 1.0 Agent annotation ratio Gender Male Female 0.0 0.2 0.4 0.6 0.8 1.0 Manual annotation ratio 0.0 0.2 0.4 0.6 0.8 1.0 Agent annotation ratio Race European African Asian Middle eastern Latin 0.0 0.2 0.4 0.6 0.8 1.0 Manual annotation ratio 0.0 0.2 0.4 0.6 0.8 1.0 Agent annotation ratio Religion Christian & Catholic Muslim Jew Buddhist Hindu Figure 7: Comparison of manual annotation and agent annotation results for the majority subgroup of generated images, where each point represents an evaluated prompt. 4.2 TEXT-TO-IMAGE MODEL DETECTION PERFORMANCE As depicted in Figure 7, we conduct an analysis of all prompts across the three dimensions involved in the task design. Each point in the figure represents one of the prompts. The x-axis signifies the proportion of bias as determined by manual annotation in the generated image, while the y-axis represents the proportion of bias detected by the agent. A distribution of scatter points near the diagonal line suggests that the accuracy of the agent closely aligns with that of manual annotation. As inferred from the figure, the performance of our agent does not significantly deviate from the results of manual annotation across the three social dimensions. Across the three social dimensions, the 7 \fUnder review average discrepancies between the proportion of stereotype annotations made by the agent and the proportion of manual labeling are 8.54%, 7.95%, and 10.13%, respectively. Despite the presence of bias in the annotation ratio, the overall accuracy in detecting whether the model contains stereotypes stands at 92.15%. This outcome highlights the potential of incorporating our proposed agent into practical bias governance frameworks. Additionally, we focus on the accuracy of two key tools, Intention Understanding and BLIP-based Detection, within the agent\u2019s reasoning trajectory. For Intention Understanding tool, we generate 120 instructions for evaluation, and 114 out of 120 instructions are accurately recognized. The remaining instances primarily involve misidentification of model names or pertain to open texts. For instance, the model name might be omitted, or the irrelevant content in the description is placed within the open text section. Gender Race Religion Figure 8: Classification accuracy of CLIP and BLIP on each subgroup, with darker shades denoting CLIP and lighter shades signifying BLIP. Different colors refer to different social dimensions. For BLIP-based Detection tool, the classification accuracy of each subgroup within this tool plays a crucial role in determining the ultimate efficacy of model stereotype detection. Consequently, our attention is primarily directed towards comparing the accuracy of two multimodal models, namely CLIP(Radford et al., 2021) and BLIP(Li et al., 2022). The performance comparison between two multi-modal models is presented in Fig.8. The results indicate that BLIP surpasses CLIP in all subgroup classification tasks, with an average improvement of 5%. Furthermore, BLIP exhibits a classification accuracy of approximately 80% in the majority of subgroup classification tasks. However, for confused samples, such as those where \u201cAsian\u201d and \u201cLatino\u201d classifications intersect, some \u201cAsian\u201d samples are misclassified as \u201cLatino\u201d, leading to a reduced average accuracy for the \u201cLatino\u201d category. Additionally, due to the limited training data available, the two models both exhibit lower classification accuracy for less popular religious subgroups such as \u201cHindu\u201d. 4.3 CASE STUDY OF MIDJOURNEY Based on the three social dimensions in the benchmark, we meticulously select the 60 most biased instruction pairs from our SPIG dataset. Our agent then uses these pairs to evaluate the stereotype degree of Midjourney. The evaluation results indicate that 46 out of the 60 pairs exhibit stereotypes, with 20 of them displaying the 100% stereotype ratio. This finding emphasizes the widespread presence of stereotypes, suggesting that despite Midjourney\u2019s advanced capacity to generate images of superior quality and detail, it also fails to successfully avoid stereotypes. For instance, when the prompt is \u201cthe person who is a cotton picker\u201d, all the generated images consistently depict \u2018African\u2019 descent. This result, similar to those observed in other open-source models, indicates that Midjourney also perpetuates serious racial stereotypes. These findings highlight the need for further research and development to mitigate the impact of stereotypes in AI-generated images. 5 RELATED WORK 5.1 STEREOTYPE IN VISION-AND-LANGUAGE Recent advancements in text-to-image models (Bianchi et al., 2023a) have been noteworthy, with these models finding utility across various domains. However, their capacity to generate socially biased content, leading to the reinforcement of harmful stereotypes, has raised concerns. Several studies have attempted to quantify the bias inherent in these models (Caliskan et al., 2017). 8 \fUnder review For instance, (Jha et al., 2023) conducted an evaluation of gender and racial biases in text-to-image models. Their methodology was based on the skew of gender and skin tone distributions in images generated from neutral occupation prompts. They employed both automated and human inspection to identify gender and skin tone in the generated images. Their findings indicated that Stable Diffusion tends to produce images of a specific gender or skin tone from neutral prompts more frequently than DALL-E. Similarly, (Ramesh et al., 2022) demonstrated that Stable Diffusion perpetuates harmful stereotypes related to race, ethnicity, gender, class, and intersectionality, even when using simple, neutral prompts. They also observed an amplification of these stereotypes. Their study further revealed that prompts mentioning social groups generate images with complex stereotypes that are challenging to mitigate. For example, Stable Diffusion associated specific groups with negative or taboo concepts such as malnourishment, poverty, and subordination. Notably, neither the introduction of \u201cguardrails\u201d against stereotyping in models like Dall-E (ElSherief et al., 2021), nor the careful expansion of user prompts, succeeded in reducing the impact of these associations. Previous research has extensively explored common stereotype characteristics, such as occupation and traits, and their effects on various subgroups distinguished by gender, race, and age. However, this approach to bias evaluation has its limitations, as it often neglects the more subtle stereotypes prevalent in everyday expressions. These biases frequently manifest in toxic content disseminated across various social platforms, including racial slurs. In this study, we introduce a novel form of bias, termed \u2018life-expressive bias\u2019, which is arguably more harmful and has been largely overlooked in previous research. Our methodology deviates from traditional approaches that rely on artificially constructed data. Instead, we construct a dataset of stereotype pairs, comprising \u2018stereotype content\u2019 and \u2018stereotype objects\u2019, derived from various toxic content datasets. This innovative approach has enabled us to establish a more comprehensive benchmark for quantifying the extent of model bias. 5.2 AUTONOMOUS LANGUAGE AGENT The recent surge in the popularity of language agents has led to the development of a diverse range of these agents, each designed to perform specific tasks. For instance, (Park et al., 2023) developed language agents to emulate human social behavior, while (Nakano et al., 2021) showcased the potential of constructing language agents capable of executing tasks on real websites following natural language instructions. Furthermore, (Qian et al., 2023) and (Hong et al., 2023) conducted experiments on software development in multi-agent communication settings, and (Zhou et al., 2023) constructed language agents to function as interactive writing assistants. Beyond task-specific language agents, recent open-source projects such as (Yang et al., 2023), (Talebirad & Nadiri, 2023), and SuperAGI5 have aimed to create autonomous agents capable of performing tasks to meet user requirements. The capabilities of agents to invoke multi-modal tools have been confirmed by recent studies such as Toolformer (Schick et al., 2023) and HuggingGPT (Shen et al., 2023). However, these agents are currently limited to processing only certain preliminary tasks or operations that conform to predefined rules. This limitation poses a challenge when planning and executing complex tasks. ReAct (Yao et al., 2022), SayCan (Ahn et al., 2022), and GPT4Tools, have demonstrated the feasibility of task reasoning and planning. Building upon the advancements in language agent development and the demonstrated capabilities of autonomous decision-making agents, our research introduces a novel language agent designed specifically for stereotype detection in text-to-image models. This agent leverages the power of LLMs and various tools to not only identify but also quantify the extent of stereotypes in generated images. 6"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2305.18346v1",
+                    "title": "Attention Paper: How Generative AI Reshapes Digital Shadow Industry?",
+                    "abstract": "The rapid development of digital economy has led to the emergence of various\nblack and shadow internet industries, which pose potential risks that can be\nidentified and managed through digital risk management (DRM) that uses\ndifferent techniques such as machine learning and deep learning. The evolution\nof DRM architecture has been driven by changes in data forms. However, the\ndevelopment of AI-generated content (AIGC) technology, such as ChatGPT and\nStable Diffusion, has given black and shadow industries powerful tools to\npersonalize data and generate realistic images and conversations for fraudulent\nactivities. This poses a challenge for DRM systems to control risks from the\nsource of data generation and to respond quickly to the fast-changing risk\nenvironment. This paper aims to provide a technical analysis of the challenges\nand opportunities of AIGC from upstream, midstream, and downstream paths of\nblack/shadow industries and suggest future directions for improving existing\nrisk control systems. The paper will explore the new black and shadow\ntechniques triggered by generative AI technology and provide insights for\nbuilding the next-generation DRM system.",
+                    "authors": "Qichao Wang, Huan Ma, Wentao Wei, Hangyu Li, Liang Chen, Peilin Zhao, Binwen Zhao, Bo Hu, Shu Zhang, Zibin Zheng, Bingzhe Wu",
+                    "published": "2023-05-26",
+                    "updated": "2023-05-26",
+                    "primary_cat": "cs.CY",
+                    "cats": [
+                        "cs.CY",
+                        "cs.AI"
+                    ],
+                    "main_content": "Introduction The rapid development of the digital economy has also given rise to a large number of black and shadow Internet industries [Yip et al., 2012; Spagnoletti et al., 2022]. Digital risk management (DRM) is a technique to identify potential risks brought by these industries and develop risk management strategies using various advanced techniques such as machine learning (ML)[Van Liebergen and others, 2017; Mashrur et al., 2020] and deep learning (DL)[Yousefi et al., \u2217Equal contribution. \u2020Corresponding to bingzhewu@tencent.com. 1The case studies, comparisons, statistics, research and recommendations in this paper are provided \u201cas is\u201d and intended for informational purposes only and should not be relied upon as operational, marketing, legal, technical, tax, financial or other advice. 2019; Heaton et al., 2017]. A review of the entire evolution of the DRM architecture reveals that changes in the form of data have been one of the driving forces behind the evolution [Omar et al., 2018]. The data generated in the early Internet era was in a more homogeneous form, so the DRM framework of the time usually used handcrafted rule engines and simple statistical factors for risk identification and control [Lalanne et al., 2013]. However, the limitations of this framework included the restrictions of the rule engine and the subjectivity of human factor design. With the advent of search engines, social media and streaming media, huge amounts of multimodal data have started to appear in the Internet world and DRM framework has started to adopt data-driven technologies such as various ML algorithms to achieve more automated and efficient risk assessment and control. This architecture enables the detection of various aspects of customer credit and transaction risk through multi-dimensional data modelling and analysis [Zhu et al., 2021]. With the rise of mobile internet and internet finance, DRM systems are faced with more diverse and complex data forms such as graph structured data, unstructured data such as user behaviour and geolocation data [Hooi et al., 2017; Beutel et al., 2015]. As a result, the DRM architecture of this period incorporated various cutting-edge technologies such as deep learning, deep graph learning and natural language processing to achieve more accurate and robust risk detection . In recent years, the DRM architecture is no longer limited to a single institution, but emphasises building across data domains. The architecture enables cross-institution and cross-market risk identification and control by establishing data sharing and collaborative supervision mechanisms [Zhu et al., 2020]. The architecture also places greater emphasis on data security and privacy protection, establishing a multi-layered, multidimensional security system including secure multi-party computation (MPC) etc [Byrd and Polychroniadou, 2020; Volgushev et al., 2019]. DRM architectures mentioned above has achieved a great success in its time due to its accurate portrayal of data patterns, but with the recent rapid development of AIGC (AIgenerated Contents) technology, especially the emergence of phenomenal commercial products such as ChatGPT2 and Sta2https://chat.openai.com/chat arXiv:2305.18346v1  [cs.CY]  26 May 2023 \fble Diffusion3, the entire black and shadow industry has unprecedentedly powerful tools upstream and downstream, enabling these groups to personalise data based on traditional natural data for different scenarios such as fraud and money laundering [Pegoraro et al., 2023], generating images, conversations and other data that would be even indistinguishable to human experts and advanced AI models [Sun et al., 2023]. Taking a real-industry case that covers online dating apps as an example, the perpetrators can generate personalized mutimodal content such as text, pictures and videos based on the preferences of victims, and even develop and customize the chatbots\u2019 language styles to automatically interact with victims. As a result, the future DRM system will be confronted with the most complex and vast amount of AIGC data ever recorded in human history. At such a turning point, it is important for all stakeholders to start thinking about how to transform the future risk control architecture so that it can control risk from the source of data generation. There is also a need for better risk warning mechanisms and strong AI technology to enable rapid response and adaptation to the fast changing risk environment. To help researchers and risk control practitioners understand the new digital risks posed by AIGC more comprehensively and concretely, this paper provides a comprehensive analysis of the challenges and opportunities presented by AIGC from a technical perspective, targeting key scenarios along the upstream, midstream and downstream critical paths of black/shadow industries. We offer two perspectives on the changes that AIGC brings to existing DRM systems. Firstly, we will explore the new black and shadow techniques triggered by generative AI technology. As generative AI technology can generate high-quality false information, black/shadow industries are beginning to use it in activities such as fraud and cyber attacks [Chen et al., 2023]. We will throughly analyse the characteristics and dangers of these new techniques. Then we will provide some insightful aspects for building the next-generation DRM system. We hope that this paper will provide further inspiration and future directions for academia, industry and the regulatory community to improve existing risk control systems for non-AIGC data. 2 Background 2.1 Large Language Model Large Language Models (LLMs) have become the core technology in the field of Natural Language Processing (NLP). These models are pre-trained on large amounts of text data, allowing them to learn rich language knowledge and semantic information. Subsequently, through a fine-tuning process, these models can be applied to various NLP tasks such as machine translation, text summarization, and sentiment analysis. The typical representative of large language models is the Transformer-based model, such as OpenAI\u2019s GPT series (such as GPT-3[Brown et al., 2020] and GPT-4) and Google\u2019s BART[Lewis et al., 2019]. These models utilize self-attention mechanisms and deep learning techniques to capture long-distance dependencies and complex semantic 3https://stablediffusionweb.com/ structures in text. In addition, these models typically have billions or even hundreds of billions of parameters, enabling them to effectively learn on large-scale datasets. 2.2 Pipeline of black/shadow industry Generally, risk activities in black/shadow industries can be divided into three stages: upstream, midstream, and downstream. The upstream stage involves preparing prior knowledge and materials, the midstream stage involves advertising to attract potential victims, and the downstream stage involves specific fraud implementation. The upstream stage is primarily focused of preparation, including the collection of various types of personal information and registration of accounts on different platforms. Personal information is often collected from multiple sources and in different formats. Previous methods [Li et al., 2020] required complex, rule-based extraction techniques to organize the data and extract potentially useful information. However, the use of LLMs enables structured processing of this messy data and automatic extraction of useful information. With the development of network services, fraudulent activities increasingly require the participation of social platforms. This allows for contact with more potential victims through network services, and various types of accounts are needed for online fraud. Therefore, in the upstream stage, perpetrators register numerous accounts across multiple platforms. In order to avoid being identified as abnormal accounts, they often adopt the method of imitating normal user behaviors to camouflage themselves, such as randomly adding normal users and interacting with them, reposting others\u2019 content, or posting fixed content. However, these behaviors are still templated, which can be easily detected, and it is difficult to generate sufficiently authentic and attractive content. Based on LLMs, perpetrators can edit personal profiles according to the preferences of the target group, which makes victims more confident in the identity disguised by the perprtrators. During the midstream stage, black/shadow industries primarily use various types of accounts obtained in the upstream stage to promote advertisements and attract potential victims. The target of advertising can be previous registered fraudulent accounts or illegal websites, such as gambling sites, to earn illegal profits. For instance, the perpetrators typically attempts to acquire a more private communication method, such as email or phone number to build trust with the victims. In this scenario, private accounts used for communication will be used as advertising content. These contents often contain harmful or harassing information and are easily detected by anomaly detection. Traditional methods [Wang et al., 2023] often use the difference between human and machine perception to transform malicious content through text adversarial methods while retaining the original semantics, in order to evade content anomaly detection. In addition, a lot of advertising information is inserted into pictures or videos, which further increases the difficulty of detection. However, these methods always face the problems that the content is repetitive, and it does not fit the current topic. By using LLMs, the analysis of current popular topics can be performed, and the advertising content can be polished to include relevant themes. 2 \fMidstream: attracting Downstream: defrauding Upstream: preparing Click this  l1nk to  win \ud83d\udcb0200 18 Information  acquisition Account  camouflage Target info Telecom  Fraud Selling Illegal  Content Property  Pilfer Multi-platform  advertising Manipulating public opinion Registration chat Figure 1: Three stages of traditional fraud methods of black/shadow industries The downstream stage involves the implementation of specific fraudulent actions. Based on the upstream and midstream work, the perpetrators complete the required material (e.g. victims\u2019 group information and fraudulent accounts) preparation and promotion. Perpetrators carry out various illegal activities for profit. Common risky behaviors include telecom fraud [Jiang et al., 2023], illegal content sales [La Morgia et al., 2021], property pilfer [Kim et al., 2017], and manipulation of public opinion [Cui and Kert\u00b4 esz, 2022]. The successful implementation of the fraudulent behaviors requires a lot of expert experience and human participation. However, based on the powerful dialogue ability of LLMs, perpetrators can establish chat robots to chat with victims. In addition, perpetrators can fine-tune specialized robots with fraudulent purposes based on existing open source models(e.g. LLaMA [Touvron et al., 2023]) and past fraud cases, further reducing the cost of fraud. 3 How Generative AI Reshapes Today\u2019s Shadow Industry Ecosystem? With the emergence of LLMs, there are many traditional methods of the black/gray industries that can be replaced by LLMs to achieve higher efficiency and economy. At the content level, LLMs can directly achieve content creation that used to require a lot of human involvement. Moreover, this process is parallel and batch-oriented, which greatly benefits the malicious operation of the black/gray industries. In this section, we will analyze the use of LLMs in assisting malicious activities in the upstream, midstream, and downstream of the black/gray industries. 3.1 Threat Model In order to illustrate how black/shadow industries use LLMs to enhance the ability of risky behaviors to camouflage and defraud, we set \u201dperpetrator\u201d as the risky behavior entity. In different risky scenarios, although the perpetrator relies on different information, they are all related to the personal information of the victims and the purpose information of the task. To provide a more intuitive demonstration of the implementation process, we conducted two case study on Romance scam and Customer Service Scam. Romance Scam As a research case, we firstly focus on romance scam, which is a common form of fraudulent activities. Perpetrators employ fake online identities to gain the affection and trust of their victims, and then they use the semblance of romance or intimacy to manipulate and steal from victims. Based on the powerful analysis and dialogue capabilities of LLMs, perpetrators can use massive user information on social media to customize more personalized social accounts that meet the preferences of victims. And it can be automated for more believable conversations. Customer Service Scam In addition to the Romance Scam, we have selected another scenario for our case study, which we refer to as the Customer Service Scam. This type of fraud involves the perpetrator impersonating a staff member of an online shopping platform to deceive the victim. The scenarios can include various fraudulent activities, such as gift-giving, free experiences, installment payments, after-sales refunds, and others. For this study, we have chosen the after-sales refund scenario. In these scenarios, we assume that the information available to the perpetrators and the goals of the risk activities are as follows: \u2022 Available information: Personal information about the victims\u2019 social media content information, purchase records, etc. \u2022 Goals: Getting the victims to send money to the perpetrators or committing fraud against the perpetrators. 3.2 Upstream As depicted in Fig.2(left), the upstream mainly includes two processes of data processing and account preparation. By conducting content analysis of potential user profiles and platforms, we extract the requisite information for the entire process of black/shadow industries and customize accounts across diverse platforms. Data processing Data processing is mainly to provide sufficient prior information for the whole process, including the classification and summary of the characteristics of different victim groups and the analysis of platform information. 3 \fFigure 2: Three stages of enhanced fraud methods by LLMs of black/shadow industries. In order to better illustrate the role of each element in each step, we use the same color as the corresponding element in the diagram in the specific prompt template, such as Consistency Check . Taking the Twitter scene in Romance scam as an example, the perpetrator employs a three-pronged approach to collect data, including popular post content, posts of each topic or section, and potential victim groups based on selected related attributes. The content collected from posts on social media platforms often includes non-textual multimodal data, such as pictures and videos. These types of data can provide additional context and information that perpetrators can use to their advantage. To process this data, perpetrators can utilize multimodal techniques, such as OCR-Donut-CORD [Kim et al., 2021], to generate captions for pictures in posts and add them to the text content in the original posts. The perpetrators can then employ LLMs to analyze the language style of the platform. LLMs are powerful tools that can help perpetrators to identify patterns in the language used on the platform, such as common phrases, slang, and other linguistic features. By analyzing the language style of the platform, perpetrators can better understand the culture and norms of the platform, which can help them to blend in and avoid detection. For the subject content of each topic or section, perpetrators can also use LLMs to analyze and summarize the corresponding subject content. This approach can help perpetrators to identify the key topics and themes being discussed on the platform, which can provide valuable insights into the interests and concerns of potential victims. By understanding the content of the current topic being discussed, perpetrators can tailor their messages and communications to better appeal to potential victims. The following is an example of the prompt template for analyzing platform style and subject content: Perpetrator agent: As a twitter content analysis expert, please summarize the format and language style of the most popular posts on Twitter below: {Posts\u2019 content}. Output: Platform style Perpetrator agent: As a twitter content analysis expert, the topic of the current post is # Topic, please analyze the content of the topic being discussed in the following posts: {Posts\u2019 content}. Output: Subject content of each topic To identify potential victim groups, the perpetrator select a group of users whose attributes match the selected related attributes, such as those following the girls about dancing videos. These attributes correlate with the likelihood that victim will suffer from romance scams. We collect the content of these users\u2019 likes, post content, the self-introduction information of following users. The perpetrator uses the multimodal model to process the above content of each person into structured text information. To effectively capture diverse information pertaining to distinct user groups and facilitate the creation of tailored accounts that cater to groups\u2019 preferences, we introduce the Group Profile Cards. Each card contains keys such as age, gender, education level, type of follower, favorite content, etc, which can more completely describe the preference information of the corresponding group. Then the perpetrator analyzes group data after structured pro4 \fcessing based on LLMs and generate corresponding group profile cards. These cards also can be used to identify potential victims of Romance scams on Twitter. The following is an example of a specific prompt template. Perpetrator agent: As an anti-romance scam assistant, your task is to classify these people and create corresponding group information cards based on the personal information provided to you. The classification basis includes {keys} Each information card is displayed in the form of json. Here are several cases of cards for your reference: Demonstrations. Output: { JSON file: Group profile cards } Account preparation After obtaining group profile cards during the data processing phase, the perpetrator can register a corresponding social account separately for each group, and tailor personal profiles and post content that aligns with group\u2019s characteristic and preferences. To illustrate, consider the Twitter scenario in the context of Romance scams. Specifically, LLMs can generate nicknames, self-introductions, and prompts of image generative artificial intelligence tools, e.g. Midjourney [], for avatar descriptions based on the corresponding group profile cards. This is a demonstration of prompt template of generating account profile. As an anti-romance scam expert, try to analyze what kind of virtual accounts scammers will create for {Group profile cards}. Please give the results in the following format: {\u201cname\u201d: [ ], \u201cbio\u201d: [ ], \u201cmidjourney descriptors for avatar images\u201d: [ ]. Here are a few examples: Demonstrations. Output: Account profile To enhance the credibility of fraudulent accounts, the perpetrator also publish personalized content that aligns with the preferences of the corresponding group and the account profile. This approach serves a dual purpose: firstly, it enables the account to behave more like a legitimate and active one, thereby enhancing its camouflage; secondly, it can also make the victim more interested in the account in downstream tasks. The below is a instance of prompt template for generating posts. Perpetrator agent: You are a twitter user, the language style of the platform is {Platform style}. Your account profile is {Account profile}. Your followers\u2019 profile cards is {Group profile card}. You want to attract more users of same type to follow you. Please Write some post content to achieve this. Remember that posts\u2019 content should be consistent with followers\u2019 preferences and your account profile. Output: Posts\u2019 content 3.3 Midstream In the intermediate stage, conventional approaches rely on disseminating advertisements across multiple platforms to expand the reach of fraudulent accounts and target a larger pool of potential victims. However, the homogeneity of the content generated through this method often fails to attract a sufficient number of users. Meanwhile, when the current topic aligns more closely with our advertising targets, it will attract greater attention. To address this issue, referring to the middle of the Fig.2, the perpetrator firstly decide which topic to promote based on LLMs, which can be formatted as: Perpetrator agent: Please act as a professional marketer of Twitter platform, your promotion is {Target promotion info}, and the existing hot topics and their main content are: {Topic: Subject content}, please decide which topics can get better promotion results by posting. Output: Selected Topic Then utilizing the subject content of corresponding topic obtained from upstream sources, LLMs are employed to generate marketing content that is tailored to specific target promotion infomation, which can be formatted as: Perpetrator agent: Please act as a content creator on the Twitter platform, where the language style is {Platform style}. Your goal is to get as many people as possible to follow you. The discussion topic for the current topic is {Subject content}. Please write the corresponding advertising posts. Output: Advertising posts\u2019 content 3.4 Downstream After completing the preliminary preparation work and diverting media to add friends to these accounts, a suitable group of scammers can be identified after midstream (i.e., the victims who answer the fake massage). Subsequently, as illustrated in the right part of Fig.2, the personal privacy information of the targets will be collected to achieve personalized customization, and the Victim Information Card {\u201cname\u201d: [ ], \u201cage\u201d: [ ], \u201cgender\u201d: [ ], \u201caddress\u201d: [ ], \u201ccontact\u201d: [ ], \u201chobby\u201d: [ ], \u201cflaws\u201d: [ ], \u201ceducation\u201d: [ ], \u201coccupation\u201d: [ ], \u201cfinance\u201d: [ ]} will be initialized for communicating with the victim through dialogue robots to implement fraud. During the communication process (i.e., Chat), in order to ensure that virtual accounts are more likely to gain the trust of victims, the model will adjust the content of the victim\u2019s information card based on user feedback and update the content posted on social media accordingly. Victim Information The advent of large models has made it easier and more effective to develop personalized fraud plans. Compared to a fixed form of fraud targeting all individuals or individuals within a certain group, personalized fraud schemes are more difficult to detect. Developing personalized fraud schemes requires structured user information. In the following paragraphs, we 5 \fwill explain how to use LLM for structured information processing, which is illustrated in Fig. 2(right top). Firstly, personal information of victims can be obtained through various means, including crawling information about victims on social media, stealing user personal information such as browsing records through various applications, and purchasing through illegal personal information markets such as personal express orders. Through these channels, unstructured data about the target can be obtained, which termed as Information acquisition. As LLMs lack detailed information about the victim, they do not directly complete this step. Instead, a script will be employed to collect the necessary information. Next, these data can be converted into content that LLMS can process smoothly. Although some service providers currently offer large models that support multimodal input, the mainstream large model services still focus on text-only input. However, even if a large model only adapts to text input, the perpetrators can still convert these modal data into text data that the large model can easily understand through Multimodal APIs, which can be achieved by open-access APIs, such as HuggingGPT[Shen et al., 2023]. Finally, the perpetrators can submit these scattered pieces of information to the LLMs for structured processing, and obtain an information card containing as much detailed information as possible about the victim. Specifically, after obtaining the text of various victim information through the previous steps, the large model can use this information to generate a JSON file termed as Victim Information Card that contains various user information for personalized fraud in the future, and the template can be formatted as: Perpetrator agent: Please assist me in organizing scattered information, such as news articles, about a particular individual, as my information analysis assistant. The output should be in JSON format, with the keys being {keys}. If certain keys\u2019 information cannot be obtained, their corresponding values should be null. Output: {JSON file: Victim Information Card} Personality Adaptation After collecting and processing personal data, we have obtained information cards of various individuals. Subsequently, we can customize personalized fraud plans based on these information cards and execute the crimes. However, it is often not possible to achieve an optimal design for a personalized fraud scheme based solely on the initially collected information. Therefore, it is necessary to continuously optimize the details of the crime process based on newly collected information and user feedback during the fraud process. This paper refers to the ultimate form of truly personalized fraud methods as Personality Adaptation as illustrated in Fig. 2(right bottom). To begin with, after establishing contact with the victim through the basic camouflage account prepared in the upstream stage, the camouflage account communicates with the victim in a finely-tuned language style and disguised character personality based on the user\u2019s information card. Specifically, upstream disguised accounts often target a certain group rather than an individual, meaning that each person has their own reminder. At this point, relevant fine-tuning can be added to the instructions of the robot conversation, such as the other person\u2019s personality and interests, making it easier for the disguised account to gain the trust of the victim, and the template can be formatted as: Perpetrator agent: Please design the speaking style of the person most likely to appeal to the user based on their information card. The user\u2019s information card is as follows: {Victim Information Card}. Output: Bot with Specific Speaking Style Next, the disguised account initiates the Chat stage with the user, engaging in multiple rounds of conversations and recording the user\u2019s feedback to achieve the goal as much as possible while ensuring that their robot identity is not recognized. Based on the new information obtained from user feedback, the model adjusts the Victim Information Card and finetunes the group information card, and the template can be formatted as: Perpetrator agent: As an information analysis assistant, I will provide you with a user information card and a history of our conversation with this user. Your task is to update the values in this information card based on the conversation history. Do you understand your assignment? LLM: Certainly, I understand. Please provide the user\u2019s information card and conversation history, and I will endeavor to update the values in the information card based on the conversation records. Perpetrator agent: {Victim Information Card, Chat history} Output: Updated Victim Information Card Finally, the language style and personality characteristics of the conversation robot are adjusted based on the updated information content, and content that matches the victim is continuously published on social media with Consistency Check as same as upstream to further gain the target\u2019s trust and successfully complete the fraud task. 4 Insights for Building Next-generation DRM System As mentioned above, shadow industry ecosystem will undergo a comprehensive transformation in the era of AIGC. Correspondingly, in order to maintain social stability and ensure public safety, the next generation of DRM systems must also make new adjustments to adapt to protect people. In this section, we introduce some enlightening introductions from the perspectives of different Defender agents (1) AIGC service providers, (2) social media service managers, and (3) security service provider, to assist researchers in contemplating how to establish a new generation of DRM systems. 6 \fWe describe the defense from the three perspectives as fundamental defense, filtering defense, and active defense. The defense provided by AIGC service providers is the most fundamental defense. Effective defense by AIGC service providers can increase the cost of malicious behavior and reduce its flexibility. The defense of social media service providers is a filtering defense aimed at blocking risks and user contact on a large scale, such as content filtering and recommendation filtering. Security service defense is an active defense that is directly deployed on the user\u2019s personal end and requires the user\u2019s participation. It is the last line of defense for users who are exposed to risks and its objective is to proactively mine more information in order to identify potential risks. In this section, we provide only a brief introduction to traditional defense methods, with a primary focus on the changes that AIGC brings to the new generation of DRM systems. 4.1 Foundational Defense: AIGC Service Providers With the increasing availability of open-source generative models, it has become easier for individuals or small organizations to establish and deploy AIGC systems. However, it is certain that the performance of these systems is inferior to that of AIGC systems provided by service providers. Therefore, the defense of service providers includes two perspectives: (1) implementing fundamental defense to prevent being exploited by perpetrators, and (2) utilizing their own more powerful capabilities to empower risk defense in other dimensions (e.g., provide knowledge to smaller models). From the perspective of AIGC service providers, on the one hand, service providers should strike a balance between protecting user privacy and identifying illegal users. For example, they may require more personal information on the premise of not infringing on privacy to be bound to a new account registration than just an email address, and limit access rates by identifying user network and hardware addresses. These methods will significantly increase the cost of illegal behavior for perpetrators, forcing them to reconsider the AIGC systems they use. On the other hand, AIGC service providers must ensure that their models are sufficiently secure and reliable, and are not susceptible to malicious manipulation by perpetrators resulting in the output of illegal content (i.e., jailbreak4). Recently, a continuous stream of jailbreak prompts have been discovered by users, making this objective particularly challenging, especially in complex open environments where model utility must also be considered. Therefore, service providers are required to update their services in real-time and patch any vulnerabilities in their systems. Timeliness is crucial, as service providers must identify them and take action before they cause significant damage. AIGC service providers possess relatively powerful model systems that can assist low-cost service systems. Taking the LLMs system as an example, data can be accumulated through adversarial strategies. Specifically, the large model can be distributed to play the roles of red/blue teams, simulating the process of malicious implementation. The resulting 4https://www.jailbreakchat.com process data can then be used as data for other AIGC service providers or even the large model itself to better handle these scenarios. An example of dialog generation is provided below: Defender agent: (To Red team) {Jailbreak instruction} Please act as {Bad guy}, your goal is {Tasks}. Defender agent: (To Blue team) Please act as {Defender}, your goal is {Tasks}. Red team: Hi, may I \u00b7 \u00b7 \u00b7 Blue team: Hi, I \u00b7 \u00b7 \u00b7 Red team: \u00b7 \u00b7 \u00b7 Blue team: \u00b7 \u00b7 \u00b7 4.2 Filtering Defense: Social Media Service Managers Thoroughly eliminating risks cannot be achieved solely through the efforts of AIGC service providers, as it requires collaborative efforts across all stages, in which social media plays a crucial role. Compared to traditional methods such as distributing leaflets and making phone calls, disseminating information and attracting attention on social media is not only more covert, but also enables the rapid identification of suitable victim groups. Social media service providers can filter the content posted by perpetrators and to content recommended to users during this process, preventing risky information from being exposed to users and implementing a filtering defense mechanism in the process of transmitting risky information to users. Here, we provide two examples from different perspectives, namely account filtering and content filtering. Account filtering. As discussed in Sec. 3.2, the perpetrator will prepare many accounts for downstream tasks. These accounts typically exhibit relatively obvious characteristics, such as inconsistent content and low posting frequency. Early filtering of such accounts can prevent perpetrators from using these accounts to disseminate illegal information. Content filtering. In the midstream, perpetrators typically attempt to mix the content of their attraction targets with normal content in order to achieve their attraction goals. In this process, filtering articles that hide content for attraction can be effectively achieved by conducting topic consistency audits. This process can be performed by LLMs, and an example template is provided below: Defender agent: Please assume the role of a content reviewer and check whether the following content contains any paragraphs that are irrelevant to the topic. The content is as follows: {Content} 4.3 Active Defense: Security Service Provider If risky content manages have evaded previous detection and reaches the user, there is still one last barrier to protect the user\u2019s safety. Security service providers can offer a variety of security assistants to help users defend against the occurrence of risks. One effective way is to use LLMs as personal 7 \fassistants and engage in conversations with strangers, proactively obtaining more information from the other party during the conversation, in order to accurately identify whether the other party poses a risk. Virtual Receptionist. The existence of security risks and time wastage associated with human reception of risky phone calls, such as fraudulent calls, has prompted the development of AI-powered call filtering systems. Specifically, the development of AI-guided call reception robots by AIGC has made it possible to filter out harmful calls. It can be endowed with specific personalities to act as user assistants. When an unfamiliar call is received, the robot first answers the call and assesses the risk level. If the call is deemed safe, the conversation history is sent to the user, and the call is transferred to the user for further communication. Sometimes it may not be effective, because the model needs to know specific information about the user, such as whether the user has purchased a mobile phone and whether the phone has encountered any problems. However, uploading all user information to the model is clearly cumbersome and poses a risk of leakage. A suitable alternative is to allow the model to first assume the negation of any situation proposed by the user, in an attempt to persuade the user to abandon their illegal behavior. Only when the user has no flaws in this situation should they be allowed to access the model. For example, the following instruction can be given: Defender agent: {Scenario description}. Please chat with the caller and try to make the caller believe that you are me. Always remember that your task is to {Detailed description}. Do you understand your task? LLM: Yes, I understand the nature of my task. Defender agent: Caller: {Content} LLM: \u00b7 \u00b7 \u00b7 5"
+                }
+            ],
+            "Aochuan Chen": [
+                {
+                    "url": "http://arxiv.org/abs/2310.02025v4",
+                    "title": "DeepZero: Scaling up Zeroth-Order Optimization for Deep Model Training",
+                    "abstract": "Zeroth-order (ZO) optimization has become a popular technique for solving\nmachine learning (ML) problems when first-order (FO) information is difficult\nor impossible to obtain. However, the scalability of ZO optimization remains an\nopen problem: Its use has primarily been limited to relatively small-scale ML\nproblems, such as sample-wise adversarial attack generation. To our best\nknowledge, no prior work has demonstrated the effectiveness of ZO optimization\nin training deep neural networks (DNNs) without a significant decrease in\nperformance. To overcome this roadblock, we develop DeepZero, a principled ZO\ndeep learning (DL) framework that can scale ZO optimization to DNN training\nfrom scratch through three primary innovations. First, we demonstrate the\nadvantages of coordinatewise gradient estimation (CGE) over randomized\nvector-wise gradient estimation in training accuracy and computational\nefficiency. Second, we propose a sparsityinduced ZO training protocol that\nextends the model pruning methodology using only finite differences to explore\nand exploit the sparse DL prior in CGE. Third, we develop the methods of\nfeature reuse and forward parallelization to advance the practical\nimplementations of ZO training. Our extensive experiments show that DeepZero\nachieves state-of-the-art (SOTA) accuracy on ResNet-20 trained on CIFAR-10,\napproaching FO training performance for the first time. Furthermore, we show\nthe practical utility of DeepZero in applications of certified adversarial\ndefense and DL-based partial differential equation error correction, achieving\n10-20% improvement over SOTA. We believe our results will inspire future\nresearch on scalable ZO optimization and contribute to advancing DL with black\nbox. Codes are available at https://github.com/OPTML-Group/DeepZero.",
+                    "authors": "Aochuan Chen, Yimeng Zhang, Jinghan Jia, James Diffenderfer, Jiancheng Liu, Konstantinos Parasyris, Yihua Zhang, Zheng Zhang, Bhavya Kailkhura, Sijia Liu",
+                    "published": "2023-10-03",
+                    "updated": "2024-03-15",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG"
+                    ],
+                    "main_content": "INTRODUCTION In the realm of machine learning (ML), optimization algorithms have played a crucial role in enabling the training of complex models, yielding unprecedented insights and predictive capabilities across diverse domains. Over the years, first-order (FO) gradient-based methods, such as stochastic gradient descent (SGD) and its variants (Gardner, 1984; Amari, 1993; Bottou, 2010; 2012), have become the default choice for model training. These methods rely on gradient information to iteratively update model parameters, aiming to minimize a given loss function. Nonetheless, several practical settings exist where FO gradient information is either unavailable or infeasible to compute, calling for alternative strategies. Zeroth-order (ZO) optimization (Flaxman et al., 2005; Shamir, 2013; Ghadimi & Lan, 2013; Nesterov & Spokoiny, 2015; Duchi et al., 2015; Liu et al., 2018b; Ilyas et al., 2018b; Zhang et al., 2024) has emerged as a promising approach to address these challenges, as it leverages finite differences of function values to estimate gradients, rather than requesting explicit gradient information. Therefore, with minor modifications to FO algorithms, ZO optimization can be applied to various real-world circumstances where FO gradients are difficult to obtain. For example, in disciplines like physics and chemistry, ML models may interact with intricate simulators or experiments where the underlying systems are non-differentiable (Thelen et al., 2022; Tsaknakis et al., 2022; Louppe et al., 2019; Abreu de Souza et al., 2023; Baydin et al., 2020). Additionally, black-box learning scenarios often arise when deep learning (DL) models are integrated with third-party APIs, such as adversarial attack and defense against black-box DL models (Chen et al., 2017; Ilyas et al., 2018a; Zhang et al., 2022c; Verma et al., 2023) and black-box prompt learning 1 arXiv:2310.02025v4  [cs.LG]  15 Mar 2024 \fPublished as a conference paper at ICLR 2024 Acceleration: Parallelization and Feature Reuse Gradient Collection Frozen Features Active Features \u2113(\ud835\udf3d\ud835\udc56) Forward Pass Backpropagation ZO Gradient Estimation Query Based ZO Optimization A Model Pruning Guided Sparsity Active Param. Dormant Param. B C \u2026 Data Process 2 Process 1 Process 4 Parameter  Allocation Process 3 Process N D Comparison with Other Methods Computational  graph FREE Computational  graph BASED Figure 1: Overview of our DeepZero framework. A: ZO gradient estimation via model queries (Sec. 3). B: Model pruning guides gradient sparsity (Sec. 4). C: Acceleration by parallelization and feature reuse (Sec. 5). D: DeepZero comparison with the computational graph free baseline Pattern Search (Chiang et al., 2023) and computational graph dependent methods without BP, Align-Ada (Boopathy & Fiete, 2022), LG-FG-A and FGW (Ren et al., 2022), on CIFAR-10. for language-model-as-a-service (Diao et al., 2022; Sun et al., 2022). Furthermore, the principled backpropagation (BP) mechanism (Amari, 1993; Rumelhart et al., 1995) for calculating FO gradients may also not be supported when implementing DL models on hardware systems (Gu et al., 2021b; Tavanaei et al., 2019; Greengard, 2020; Jabri & Flower, 1992; Gu et al., 2021c). In addition to ZO optimization, another relevant research direction in the field of DL focuses on developing biologically-plausible, BP-free methods. Examples include forward gradient-based methods (Ren et al., 2022; Baydin et al., 2022; Silver et al., 2021; Belouze, 2022), greedy layer-wise learning (N\u00f8kland & Eidnes, 2019), and Hebbian learning (Isomura & Toyoizumi, 2018; Moraitis et al., 2022). However, these techniques require access to computational graphs and are highly dependent on the used DL software frameworks and/or model architectures. In contrast, ZO optimization solely relies on model queries and is free of computational graphs utilized. As a result, ZO optimization has broad applicability to DL problems that involve black-box query-only components. Despite the promise of ZO optimization, scalability bottlenecks hinder its application in medium or large-scale DNN training (Wang et al., 2017; Liu et al., 2018b; Ohta et al., 2020; Cai et al., 2021; Zhang et al., 2022c). As problem dimensionality increases, the accuracy and efficiency of traditional ZO methods deteriorate. This is because ZO finite difference-based gradient estimates are biased estimators of FO gradients, and the bias becomes more pronounced in higher-dimensional spaces (Liu et al., 2018b; Cai et al., 2021; Balasubramanian & Ghadimi, 2018). These challenges motivate the central question addressed in this work: (Q) How to scale up ZO optimization for training deep models? To address (Q), we propose a novel framework, \u2018DeepZero\u2019, which infuses novel model pruning and parallel computing techniques to scale up ZO DNN training (see Fig. 1 for a schematic overview). Our main contributions are summarized below. \u2776We show that deterministic coordinate-wise gradient estimation (CGE) outperforms vector-wise randomized gradient estimation (RGE) in both accuracy and computation efficiency when scaling to deep model training. Further, CGE becomes increasingly advantageous as model depth increases. \u2777We show that sparsity is crucial for realizing model training via CGE with finite differences. In contrast to prior work, we find that sparsity for black-box models can be obtained for \u2018free\u2019 by extending the current pruning-at-initialization technique to the ZO learning paradigm. The established synergy between pruning and CGE presents a promising avenue for efficient ZO training of DNNs. \u2778We identify the parallelization-fit property inherent to CGE-based ZO optimization and propose a novel forward parallelization method based on this property. Our framework enables feature reuse in deep learning, which further accelerates parallel training by eliminating redundant computations. \u2779We introduce our proposed ZO deep model training framework, \u2018DeepZero\u2019. To demonstrate the empirical superiority of DeepZero, we conduct extensive experiments on both standard image classification benchmarks and real-world black-box DL applications. For example, when employing 2 \fPublished as a conference paper at ICLR 2024 DeepZero to train a ResNet20 on CIFAR-10, we obtain 86.94% testing accuracy, the best reported in the literature of gradient-free model training. We also exemplify the vast potential and practical impact of DeepZero in two real-world DL tasks: black-box defense for adversarial robustness (Zhang et al., 2022c) and physics-informed DL with solver-in-the-loop (Um et al., 2020). To clarify, our work aims to extend the scalability of ZO optimization for DL applications, addressing cases where FO optimization becomes challenging or infeasible. Yet, it is essential to note that the proposed advancements in ZO training are not intended to overcome the ultimate scalability challenges to train deep networks at any scale. 2 RELATED WORK Classical gradient-free optimization. Early research efforts can be broadly categorized into two groups: direct search-based methods (DSMs) and model-based methods (MBMs) (Wright et al., 1999; Conn et al., 2009; Rios & Sahinidis, 2013; Larson et al., 2019). DSMs include techniques like coordinate (Fermi, 1952) and pattern search (Torczon, 1991) methods and the Nelder-Mead simplex method (Nelder & Mead, 1965). MBMs consist of model-based descent (Bortz & Kelley, 1998) and trust region (Conn et al., 2000) methods. Evolutionary optimization offers a generic populationbased gradient-free computational framework including genetic algorithms (Grefenstette, 1993) and particle swarm optimization (Vaz & Vicente, 2009). Bayesian optimization (Shahriari et al., 2015; Eriksson et al., 2019) has garnered recent attention by using a Gaussian process (GP) to fit a blackbox objective function and estimate an optimization solution. However, acquiring an accurate GP is computationally intensive. Zeroth-order optimization. In contrast to classical gradient-free methods, ZO optimization approximates gradients using finite differences, simplifying implementation by minimizing modifications of FO gradient-based algorithms. Like FO methods, ZO enjoys provable convergence guarantees (Nesterov & Spokoiny, 2017; Duchi et al., 2015; Liu et al., 2020a). ZO optimization has gained significant attention for its success in solving various emerging ML problems (Ghadimi & Lan, 2013; Nesterov & Spokoiny, 2015; Flaxman et al., 2005; Duchi et al., 2015). Examples include adversarial attack and defense (Chen et al., 2017; Tu et al., 2019; Ye et al., 2018; Ilyas et al., 2018a; Zhang et al., 2022c; Verma et al., 2023; Zhao et al., 2019; Hogan & Kailkhura, 2018; Shu et al., 2022), model-agnostic contrastive explanation (Dhurandhar et al., 2019), visual prompting for transfer learning (Tsai et al., 2020), computational graph unrolling (Vicol et al., 2023), automated ML (Gu et al., 2021a; Wang et al., 2022), policy search in reinforcement learning (Vemula et al., 2019), network resource management (Liu et al., 2018b), ML-based scientific workflow optimization (Tsaknakis et al., 2022), and on-chip learning (Gu et al., 2021b). Despite ZO\u2019s successes in solving ML problems, its application has been limited to relatively small scales. For instance, ZO optimizers used for generating adversarial attacks, contrastive explanations, and visual prompts only operate in the input parameter space, which has the dimension of a single input example. Some acceleration techniques have been developed to improve ZO performance in larger problems, such as using historical information to enhance a ZO gradient estimator (Meier et al., 2019; Cheng et al., 2021), and exploiting gradient sparsity to reduce ZO dependence on problem size (Wang et al., 2017; Cai et al., 2022; 2021; Balasubramanian & Ghadimi, 2018; Ohta et al., 2020; Gu et al., 2021b). While gradient sparsity has been used to improve scalability (Bartoldson et al., 2023), we propose an advanced strategy that leverages model pruning techniques to identify and exploit sparsity in neural network parameters effectively. Our approach is less restrictive than traditional gradient sparsity assumptions and allows for greater flexibility in selecting what to prune. To the best of our knowledge, no prior work has demonstrated the practicality of scalable ZO optimization for deep model training without significant performance loss compared to the FO counterpart. DL without backpropagation. Forward gradient learning (Baydin et al., 2022; Ren et al., 2022; Silver et al., 2021; Belouze, 2022), which builds upon the forward-mode automatic differentiation (AD) capabilities of current DL software frameworks, does not rely on finite differences to approximate FO gradients like ZO optimization. Instead, it relies on forward-mode AD to calculate a forward (directional) gradient. This gradient is obtained by projecting the FO gradient onto a direction vector and is an unbiased estimator of the FO gradient (Baydin et al., 2022). In contrast, ZO gradient estimation based on finite differences is biased (Duchi et al., 2015; Liu et al., 2020a). However, one main limitation of forward gradient learning is that it requires full access to AD software and the deep model, making it impractical for solving black-box DL problems. Recent advances in (Ren et al., 2022) further improved the scalability of forward gradient learning by using finer-level model 3 \fPublished as a conference paper at ICLR 2024 information to design architecture-specific local objective functions. Other BP-free DL methods are motivated by seeking a biological interpretation of DL but share similar limitations with forward gradient learning. Some examples include greedy layer-wise learning (N\u00f8kland & Eidnes, 2019), input-weight alignment for wide neural networks in the neural tangent kernel (NTK) regime (Boopathy & Fiete, 2022), the Forward-Forward algorithm (Hinton, 2022), Hebbian Learning (Isomura & Toyoizumi, 2018; Moraitis et al., 2022), and synthetic gradients (Jaderberg et al., 2017). 3 ZO OPTIMIZATION THROUGH FUNCTION VALUE-BASED GRADIENT ESTIMATION: RANDOMIZED OR COORDINATE-WISE? We now introduce the ZO optimization setup and discuss two ZO gradient estimation schemes: deterministic coordinate-wise gradient estimation (CGE) and randomized vector-wise gradient estimation (RGE). We will demonstrate the advantage of CGE over RGE for DNN training. This inspires further improvements for scaling CGE-based ZO optimization. ZO optimization and gradient estimation. Let \u2113(\u03b8) denote a loss function that we want to minimize over the optimization variables \u03b8 \u2208Rd (e.g., model parameters of a neural network). The ZO optimizer interacts with the objective function \u2113only by submitting inputs (i.e., realizations of \u03b8) and receiving the corresponding function values. It slightly modifies the commonly-used first-order (FO) gradient-based algorithm by approximating the FO gradient through function value-based gradient estimates (Liu et al., 2020a). This is essential when explicit differentiation is difficult due to the black-box nature of the loss function (Zhang et al., 2022c; Liu et al., 2020b; Chen et al., 2017), or when explicit differentiation is undesired due to concerns about energy efficiency (Gu et al., 2021b; Liu et al., 2018a). RGE (Nesterov & Spokoiny, 2017; Ghadimi & Lan, 2013; Duchi et al., 2015; Spall, 1992) and CGE (Kiefer & Wolfowitz, 1952; Lian et al., 2016; Berahas et al., 2022) are two commonly-used gradient estimators based on finite differences of \u2113. RGE acquires finite differences via random perturbations of \u03b8 while CGE uses deterministic coordinate-wise perturbations of \u03b8 (Liu et al., 2020a). Their formal definitions are given by (RGE) \u02c6 \u2207\u03b8\u2113(\u03b8) = 1 q q X i=1 \u0014\u2113(\u03b8 + \u00b5ui) \u2212\u2113(\u03b8) \u00b5 ui \u0015 ; (CGE) \u02c6 \u2207\u03b8\u2113(\u03b8) = d X i=1 \u0014\u2113(\u03b8 + \u00b5ei) \u2212\u2113(\u03b8) \u00b5 ei \u0015 , (1) 730 3700 6670 9640 12610 Number of Model Parameters 40 50 60 70 80 Testing Accuracy (%) 42.09 41.65 37.53 60.47 59.66 50.06 70.69 70.04 65.31 76.64 76.07 70.05 77.01 76.95 68.38 FO Training ZO CGE ZO RGE Figure 2: Performance comparison of training a simple CNN with varying numbers of parameters on CIFAR10 using different training methods. where \u02c6 \u2207\u03b8\u2113denotes an estimation of the FO gradient \u2207\u03b8\u2113with respect to \u03b8. In (RGE), ui denotes a randomized perturbation vector, e.g., drawn from the standard Gaussian distribution N(0, I), \u00b5 > 0 is a perturbation size (a.k.a. smoothing parameter), and q is the number of random directions used to acquire finite differences. In (CGE), ei denotes a standard basis vector, and \u2113(\u03b8+\u00b5ei)\u2212\u2113(\u03b8) \u00b5 provides the finite-difference estimation of the partial derivative of \u2113(\u03b8) at the ith coordinate \u03b8i. Finite difference approximations in (1) are motivated by directional derivatives. Take RGE (with q = 1) as an example. As \u00b5 \u21920, finite difference in RGE converges to directional derivative \u2113\u2032(\u03b8) := uT \u2207\u03b8\u2113(\u03b8) = lim\u00b5\u21920 \u2113(\u03b8+\u00b5ui)\u2212\u2113(\u03b8) \u00b5 of the function \u2113at the point \u03b8 in the direction u (Urruty & Lemar\u00b4 echal, 1993). As a result, the expression \u2113\u2032(\u03b8)u yields E[\u2113\u2032(\u03b8)u] = E[(uuT )\u2207\u03b8\u2113(\u03b8)] = \u2207\u03b8\u2113(\u03b8) (recall that E[uuT ] = I). This implies that \u2113\u2032(\u03b8)u is an unbiased gradient estimator of \u2207\u03b8\u2113(\u03b8) and its biased finite difference approximation is given by (1) (Duchi et al., 2015). RGE or CGE? First, the function query costs for RGE and CGE differ, with RGE taking O(q) queries and CGE taking O(d) queries based on (1). Compared to CGE, RGE has the flexibility to specify q < d to reduce the number of function evaluations. Despite the query efficiency, it remains uncertain whether RGE can deliver satisfactory accuracy when training a deep model from scratch. To this end, we undertake a preliminary investigation wherein we train a basic convolutional neural network (CNN) of different sizes on CIFAR-10, employing both RGE and CGE. To ensure a fair comparison in query complexity, we set the query number q in RGE equal to the problem size d used in CGE. Fig. 2 presents the test accuracy of the learned CNN against the number of model parameters (equivalent to the number of model queries). Here the training recipe is specified by the FO SGD, the ZO RGE-based SGD, and the ZO CGE-based SGD. We observe that CGE can achieve test accuracy comparable to FO training and significantly outperforms RGE. This experiment highlights 4 \fPublished as a conference paper at ICLR 2024 the superiority of CGE over RGE in terms of optimization accuracy even when the latter uses q = d. This accuracy merit of CGE is particularly valuable when training more complex neural networks. In Appx. C, we provide a detailed analysis of the computational costs using CGE vs. RGE. The time cost relative to model depth is shown in Fig. A2. And Tab. A1 assesses gradient estimation time costs. We find that CGE demonstrates greater time efficiency than RGE. The computation efficiency loss of RGE is in that it needs to generate and integrate a d-dimension perturbation vector into the entire model at once every query. Based on the advantages of CGE over RGE in terms of both accuracy and computation efficiency, we choose CGE as the preferred ZO gradient estimator. However, query complexity of CGE is still a bottleneck, as it scales with model size d. 4 SPARSITY-ASSISTED ZO TRAINING: A PRUNING LENS AND BEYOND One valuable property of CGE is the disentanglement of finite differences across coordinates, which suggests that reducing CGE\u2019s query complexity is aligned with pruning the model weights that are being optimized. With this in mind, we propose integrating ZO optimization with pruned gradients to design a more effective inductive bias for ZO deep model training. It is worth noting that the sparsity has been explored in several existing ZO optimization methods to improve the query efficiency of gradient estimation (Wang et al., 2017; Cai et al., 2022; 2021; Balasubramanian & Ghadimi, 2018; Ohta et al., 2020; Gu et al., 2021b). However, prior work suffered from two main limitations. Firstly, exact sparsity was assumed in the original FO gradients, which required an additional sparse learning method (such as LASSO (Wang et al., 2017)) to recover these sparse gradients from function queries. Secondly, it remains unclear how to optimize the sparsity pattern via a ZO oracle, as the existing method calls for overly heuristics-based pruning methods (e.g., random (Gu et al., 2021b) or magnitude (Ohta et al., 2020; Zhang et al., 2022b) pruning). Overly increasing sparsity ultimately limits optimization performance. In what follows, we propose a new pruning approach that relies only on model queries, enjoys computation efficiency, and can improve ZO optimization accuracy by inducing an appropriate gradient sparsity. ZO-GraSP: Model pruning via ZO oracle. The compressibility of model weights for DL has been extensively studied (Han et al., 2015; Frankle & Carbin, 2018; Ma et al., 2021; Zhang et al., 2022a;b; Blalock et al., 2020; Tanaka et al., 2020; Lee et al., 2018; Wang et al., 2020; Su et al., 2020; Diffenderfer et al., 2021). For instance, the lottery ticket hypothesis (Frankle & Carbin, 2018) demonstrated that a randomly initialized, dense neural network contains a high-quality sparse subnetwork. However, current effective pruning methods incorporate model training as an intermediate step (Frankle & Carbin, 2018; Ma et al., 2021; Zhang et al., 2022a; Diffenderfer & Kailkhura, 2021). Thus, they are not well-suited for finding sparsity via a ZO oracle. To address the above challenge, we draw inspiration from training-free pruning methods, known as pruning-at-initialization (Tanaka et al., 2020; Lee et al., 2018; Wang et al., 2020). Within this family, gradient signal preservation (GraSP) (Wang et al., 2020) is a method to identify the sparsity prior of DL through the gradient flows of a randomly-initialized network. While GraSP still requires the FO and second-order derivative information, we can estimate these derivatives using only function queries to design the ZO version of GraSP (termed ZO-GraSP). Specifically, GraSP (Wang et al., 2020) assigns pruning scores (denoted by S) to model initialization \u03b8. These scores reflect the change in gradient flow after pruning the weights: S = \u2212\u03b8 \u2299(Hg), H = \u22072 \u03b8,\u03b8\u2113(\u03b8), g = \u2207\u03b8\u2113(\u03b8), (2) where recall that \u2113is the loss function of model training, \u2299denotes the entry-wise multiplication, and Hg represents the Hessian-gradient product. Using the ZO learning paradigm, we can first approximate the Hessian-gradient product as the finite difference between two gradients (i.e., \u2207\u03b8\u2113(\u03b8 + \u00b5g) and \u2207\u03b8\u2113(\u03b8)), in the direction g with the smoothing parameter \u00b5. Second, we replace the FO gradient \u2207\u03b8\u2113with the ZO gradient estimate \u02c6 \u2207\u03b8\u2113given in (1). Combining this yields ZO-GraSP: \u02c6 S := \u2212\u03b8 \u2299 \u02c6 \u2207\u03b8\u2113(\u03b8 + \u00b5\u02c6 g) \u2212\u02c6 \u2207\u03b8\u2113(\u03b8) \u00b5 . (3) In practice, we found that the pruning mask determined by ranking the entries in \u02c6 S is resilient to the ZO gradient estimation error. Therefore, we utilize RGE with a relatively small number of queries (q < d) to implement ZO-GraSP. This reduces the function query cost without compromising pruning performance; see Tab. A2 and Tab. A3 for empirical justifications. Our results show that ZO-GraSP significantly outperforms random pruning and yields pruned models with accuracy comparable to FO-GraSP. 5 \fPublished as a conference paper at ICLR 2024 Integrating sparsity with CGE. As finite differences in CGE (1) are decomposable over weights, it is easy to incorporate sparsity into CGE. To retain the accuracy benefits of training dense models, we incorporate gradient sparsity (in CGE) rather than weight sparsity. This ensures that we train a dense model in the weight space, rather than training a sparse model where the sparsity determined by ZO-GraSP is directly applied. Let SZO-GraSP be the coordinate set of unpruned model weights found by ZO-GraSP. The sparsity-induced CGE is given by \u02c6 \u2207\u03b8\u2113(\u03b8) = X i\u2208SZO-GraSP \u0014\u2113(\u03b8 + \u00b5ei) \u2212\u2113(\u03b8) \u00b5 ei \u0015 . (Sparse-CGE) It is clear that (Sparse-CGE) reduces the query complexity of the original CGE from O(d) to O(|SZO-GraSP|), where |SZO-GraSP| denotes the cardinality of the coordinate set SZO-GraSP. There may exist two direct methods for integrating (Sparse-CGE) into ZO optimization. M1: This method involves alternating between ZO-GraSP and CGE-based ZO optimization. At each iteration, SZO-GraSP is updated based on the model weights from the previous iteration and then used to construct (Sparse-CGE) for updating \u03b8 at the current iteration. M2: This method involves performing pruning before ZO training. That is, ZO-GraSP is conducted at model initialization, and the resulting SZO-GraSP is applied to (Sparse-CGE) and kept fixed during training. Both M1 and M2 have limitations. M1 requires repeated calls to ZO-GraSP to update SZO-GraSP, leading to a higher query cost for ZO model training. M2 addresses the query complexity by performing ZO-GraSP before training, but it can only produce a smaller model after training. It is known that heavily-pruned models suffers from performance degradation (e.g., 95% sparse model in Tab. A2 in Appx. D). Thus, it is nontrivial to integrate ZO-GraSP with ZO training due to the requirement of balancing query efficiency and training effectiveness. To address this, we propose ZO-GraSP-oriented dynamic sparsity pattern, which leverages ZO-GraSP to determine layer-wise pruning ratios (LPRs) that can capture DNN compressibility. This approach shares a similar essence with smart ratio introduced in (Su et al., 2020). Specifically, we acquire LPRs from ZO-GraSP at randomly initialized weights prior to ZO training, which is query-efficient like M2. However, unlike M2, LPRs allow for random shuffling of sparse gradient positions in \u03b8 only if these LPRs are obeyed. This allows us to mimic M1 to alternate between model weight updates and SZO-GraSP updates, with the latter achieved by LPR-guided randomly updated sparsity patterns. Thus, ZO optimization can train the dense model using iteratively-updated (Sparse-CGE) with LPRs-guided dynamic sparsity patterns. Overall, our proposal has the query efficiency of M2 with the training effectiveness of M1, resulting in a balanced integration of ZO-GraSP into ZO training. We summarize the algorithmic pipeline in Algorithm 1 in Appx. E, where CGE and ZO-GraSP-oriented dynamic sparsity pattern are clearly described in a unified framework. We also refer readers to Appx. E for more explanation and comparisons with M1 and M2. We provide a convergence rate analysis in Appx. A. 5 IMPROVING SCALABILITY: FEATURE REUSE & FORWARD PARALLEL 730 3700 6670 9640 12610 Number of Model Parameters 0 20 40 60 80 Training Time (hours) 0.99 1.66 5.11 12.08 11.39 28.03 18.91 45.22 28.94 70.32 CGE with feature reuse CGE without feature reuse Figure 3: Computation cost of CGE-based ZO training w/ feature reuse vs. w/o feature reuse. The setup follows Fig. A2. We investigate two characteristics of ZO training that can further enhance implementation scalability: feature reuse and forward parallelization. The former disentangles intermediate features from weight perturbations, while the latter uses the finite-difference nature in CGE to enable scalable distributed implementation. Reusing intermediate features. As shown in (1), CGE perturbs each parameter element-wise. Thus, one can reuse the feature immediately preceding the perturbed layer and carry out the remaining forward pass operations instead of starting from the input layer, as illustrated in Fig. 1. The above characteristic of CGE-based model training is referred to as \u2018feature reuse\u2019. More concretely, let f\u03b8(x) be a deep model with parameters \u03b8 and input x. We can express f\u03b8(x) as a multi-layer composite function f\u03b8(x) = f\u03b8>l(zl) = f\u03b8L \u25e6f\u03b8L\u22121 \u25e6\u00b7 \u00b7 \u00b7 \u25e6f\u03b8l+1 | {z } f\u03b8>l(\u00b7) \u25e6f\u03b8l \u25e6\u00b7 \u00b7 \u00b7 \u25e6f\u03b81(x) | {z } zl = f\u03b81:l (x) , (4) where f\u03b8l denotes the model\u2019s lth layer, L is the total number of model layers, and \u25e6is the function composition operation. Based on (4), if coordinate-wise weight perturbations are applied to the (l + 1)th layer and its subsequent layers (i.e., \u03b8>l), the model\u2019s outputs corresponding to these 6 \fPublished as a conference paper at ICLR 2024 perturbed weights can be efficiently obtained by keeping the intermediate features up to the lth layer (i.e., zl) intact. This efficiency becomes more pronounced when perturbing the parameters of deeper layers (i.e., for a larger l). Fig. 3 compares the runtime of CGE-based ZO training with and without feature reuse. Empirically, CGE with feature reuse exhibits a 2\u00d7 reduction in training time. Parallelization of coordinate-wise finite differences. CGE enables parallelization of model training due to its alignment of parameter perturbations with forward passes. If there exist M processes (across multiple GPUs), we can decompose CGE (1) based on its parameter coordinates yielding \u02c6 \u2207\u03b8\u2113(\u03b8) = M X i=1 \u02c6 gi, \u02c6 gi := X j\u2208Si \u0014\u2113(\u03b8 + \u00b5ej) \u2212\u2113(\u03b8) \u00b5 ej \u0015 , (5) where Si is the set of active parameters assigned to process 1 \u2264i \u2264M. Hence, each process can take |Si| forward passes. This decoupling property enables scaling forward passes via distributed machines, which can significantly improve training speed. We refer to this parallelization for finite differences as \u2018forward parallelization\u2019. It is worth noting that forward parallelization is different from the conventional data parallelization used for FO distributed training (Goyal et al., 2017; You et al., 2018). Our method avoids any performance loss that may result from data parallelization using an overly large batch size, which can cause the optimizer to get stuck in suboptimal local minima due to a lack of stochasticity. 6 EXPERIMENTS In this section, we first train ResNet-20 for standard image classification on CIFAR-10, demonstrating scalability and generalization capability over existing gradient-free learning methods. Second, we apply DeepZero to enhance the robustness of a black-box DNN against adversarial attacks, where limited access to the model is available on the defender\u2019s end. Lastly, we leverage DeepZero to design a physics-informed ML system by incorporating a scientific PDE solver into the training loop for reducing numerical errors, highlighting its capability to address complex scientific problems. 6.1 IMAGE CLASSIFICATION TASK 80 90 95 99 Sparsity (%) 60 65 70 75 80 85 90 95 Testing Accuracy (%) 92.85 88.59 86.94 90.12 86.25 86.46 88.55 82.95 84.26 86.98 68.22 73.37 78.1 FO training FO training with Sparse Weight FO training with Sparse Gradient DeepZero (Ours) Figure 4: Comparison between DeepZero and FO training baselines on a ResNet-20 for CIFAR-10. We report the mean and standard deviation of 3 independent runs for each experiment. Experiment setup. This study focuses on training ResNet20 (with 270K parameters) on CIFAR-10 for image classification. We adopt SGD (stochastic gradient descent) as the FO training recipe, with a weight decay of 5 \u00d7 10\u22124 and a momentum of 0.9. The learning rate is 0.1, governed by a cosine decay scheduler. In the ZO training scenario, we replace the FO gradient by (Sparse-CGE) with a smoothing parameter \u00b5 = 5 \u00d7 10\u22123. When implementing ZO-GraSP (3), we set the query budget q = 192 and use the same \u00b5 as CGE. Unless specified otherwise, the weight sparsity ratio is chosen to be 90% and the specific sparsity patterns are determined by SR (Smart Ratio). When implementing DeepZero (Algorithm 2), we choose the number of epochs T = 50. Experiments are run on 4 NVIDIA V100 GPUs if not specified otherwise. We compare DeepZero with FO training and two SOTA BP-free training: Pattern Search (Chiang et al., 2023) and Input-Weight Alignment (Align-ada) (Boopathy & Fiete, 2022). 100 500 1000 2000 4000 8000 16000 32000 Dataset Size 0 10 20 30 40 50 60 70 80 Testing Accuracy (%) DeepZero Pattern Search 0 50 100 150 200 250 300 Time (hours) Accuracy Time Cost Figure 5: Comparison of DeepZero and Pattern Search on ResNet-20 for CIFAR-10 with varying dataset sizes. All experiments are done on a single NVIDIA A6000 GPU. Comparison with FO training. In Fig. 4, we compare the accuracy of DeepZero-trained ResNet-20 with two variants trained by FO recipes: (1) a dense ResNet-20 acquired through FO training and (2) a sparse ResNet20 acquired through FO training under FO-GraSP sparsity pattern. As we can see, the accuracy gap still exists between (1) and the model trained with DeepZero in the sparsity regime of 80% to 99%. This highlights the challenge of ZO optimization for deep model training, where achieving high sparsity is desired to reduce the number of model queries in (Sparse-CGE) for scaling to ResNet-20. Notably, in the sparsity regime of 90% to 99%, DeepZero outperforms (2), showcasing the superiority of gradient sparsity in DeepZero 7 \fPublished as a conference paper at ICLR 2024 compared to weight sparsity (i.e., directly training a sparse model). In Appx. F, we provide the DeepZero training trajectory (Fig. A4), performance vs. data batch setup (Fig. A5) and training time vs. GPU count (Fig. A6). Comparison with pattern search (Chiang et al., 2023). In Fig. 5, we compare the accuracy and runtime cost of DeepZero with Pattern Search (Chiang et al., 2023) for deep model training. Pattern Search has been shown to achieve comparable test accuracy to SGD in low sample regimes, however, effectiveness as the number of data samples increases remains unclear. To investigate this, we conducted experiments using DeepZero (with 90% gradient sparsity) and pattern search on ResNet20 with CIFAR-10, with varying dataset sizes from 100 to 32K. We maintained a fixed total epoch number of 40 for both methods to ensure a fair comparison. The results demonstrate DeepZero outperforms Pattern Search in all data regimes, except in the case of 100. Further, the improvement of DeepZero over pattern search (in both accuracy and efficiency) expands dramatically with increasing dataset sizes, indicating the superior scalability of DeepZero. Table 1: Performance of DeepZero vs. BP-free methods on a 8-layer CNN w/ different widths (Boopathy & Fiete, 2022). Method DeepZero FA DFA Align-ada Width 32 64 64 512 64 512 64 512 Accuracy 57.7 64.1 46.5 45.4 49.9 54.1 49.9 58.0 Time (h) 4.34 28.15 0.36 3.79 0.42 3.52 0.48 4.59 Comparison with input-weight alignment (Boopathy & Fiete, 2022). In Tab. 1, we present a comparison between DeepZero and the Align-ada approach (Boopathy & Fiete, 2022) for training neural networks without BP on CIFAR-10. While other BP-free training methods (Lillicrap et al., 2014; N\u00f8kland, 2016; Baydin et al., 2022) exist, Align-ada stands out as it applies to training wide neural networks and achieves state-of-the-art performance on CIFAR-10, e.g., surpassing methods such as feedback alignment (FA) (Lillicrap et al., 2014) and direct feedback alignment (DFA) (N\u00f8kland, 2016). To ensure fairness, we apply DeepZero to the 8-layer CNN architecture from (Boopathy & Fiete, 2022) and compare performance with Align-ada at varying model widths. We note that the 512-width network was the widest model trained using Align-ada. In contrast, the largest width network we train with DeepZero is 64. Our results clearly show that DeepZero achieves significantly higher testing accuracy compared to Align-ada, even when training with narrower networks. This demonstrates that the improved performance of DeepZero is attributed to its inherent optimization advantages, rather than relying solely on the use of wider networks. Lastly, it is worth noting that Align-ada and other BP-free methods rely on access to computational graphs, making them efficient but unsuitable for black-box applications. 6.2 OTHER BLACK-BOX APPLICATIONS Table 2: CA (%) vs. \u21132-norm based perturbation radius on ImageNet-10 using FO DS-based defense (FO-DS) (Salman et al., 2020), ZO-AE-DS (Zhang et al., 2022c), and our proposed DeepZero. ImageNet (10 classes) Radius r FO-DS ZO-AE-DS DeepZero 0.0 89.33 63.60 86.02 0.25 81.67 52.80 76.61 0.5 68.87 43.13 61.80 0.75 49.80 32.73 43.05 Black-box defense against adversarial attacks. The black-box defense problem arises when the owner of an ML model is unwilling to share the model details with the defender against adversarial attacks (Zhang et al., 2022c; Verma et al., 2023). This poses a challenge for existing robustness enhancement algorithms (Madry et al., 2017; Cohen et al., 2019; Salman et al., 2020) that directly robustify white-box ML models using FO training. To overcome this challenge, ZO optimization was introduced in (Zhang et al., 2022c) to design a white-box defense operation given a query-based black-box image classifier. To address dimensionality challenges with ZO, ZO-AE-DS (Zhang et al., 2022c) introduces an autoencoder (AE) between the white-box denoised smoothing (DS) defense operation (to be learned) and the black-box image classifier. By merging AE\u2019s encoder with the blackbox module, the dimension of ZO optimization is reduced; see Fig. A7 in Appx. G for a schematic overview and (Zhang et al., 2022c) for details. The downside of ZO-AE-DS is poor scaling to highresolution datasets (e.g., ImageNet) due to the use of AE, which compromises the fidelity of the image input to the black-box image classifier and leads to inferior defense performance. In contrast, DeepZero can directly learn the defense operation integrated with the black-box classifier, without needing AE. Tab. 2 compares the defense performance of DeepZero with FO defense (DS (Salman et al., 2020)) and ZO-AE-DS (Zhang et al., 2022c). To ensure a fair comparison, we used the same number of queries (1152) per gradient estimation. Following (Zhang et al., 2022c), we selected a 10-class subset of ImageNet as the training set. The AE and black-box classifier are DnCNN (Zhang et al., 2017) and ResNet-50, respectively. Defense performance is evaluated by certified accuracy 8 \fPublished as a conference paper at ICLR 2024 (CA), following the setup of (Salman et al., 2020; Zhang et al., 2022c). CA is defined using the \u21132 norm-based input perturbation radius r, where a larger r indicates a stronger adversarial threat. We refer readers to Appx. G for more experiment details. Tab. 2 highlights that DeepZero consistently outperforms ZO-AE-DS in terms of CA for all values of r > 0. It is important to note that when r = 0, CA is equivalent to the standard testing accuracy. This indicates that DeepZero excels over ZO-AE-DS not only in adversarial robustness but also in overall generalization performance. SRC NON ZO-SOL FO-SOL 0.00 0.05 0.10 0.15 0.20 Mean Absolute Error 0.180 0.153 0.119 0.074 Figure 6: Average MAE of corrected low-fidelity simulations compared to high-fidelity simulations over 5 test simulations using different correction methods. Error bar is variance of MAE across 5 test simulations. Simulation-coupled DL for discretized PDE error correction. Numerical methods, while instrumental in providing physics-informed simulations, come with their own challenge: the discretization unavoidably produces numerical errors. DL has gained significant attention for addressing this error correction problem. The feasibility of training a corrective neural network through looping interactions with the iterative partial differential equation (PDE) solver, coined \u2018solver-in-the-loop\u2019 (SOL), has been demonstrated in (Um et al., 2020). While existing work focused on using or developing differentiable simulators for model training, we extend SOL by leveraging DeepZero, enabling its use with non-differentiable or blackbox simulators. We name our method ZO-SOL and refer readers to Fig. A8 in Appx. H for a schematic overview. In this experimental framework, the goal is to correct the output of a low fidelity (i.e., coarse mesh) simulation using a learned DNN so that the corrected simulation more closely aligns with output of a high fidelity (i.e., fine mesh) simulation. In an effort to reduce the amount of correction required by DNN, this correction is applied at each simulation timestep and the corrected simulation state is provided to the simulator to compute the next timestep. For our experiments, we consider 2D unsteady wake flow fluid dynamics benchmark from (Um et al., 2020), in which there is a static rod around which fluid is flowing, and utilize PhiFlow (Holl et al., 2020) as the simulator. Additional details on the PDE, the solver-in-the-loop loss, and the DNN architecture are given in Appx. H. Fig. 6 compares the test error correction performance of ZO-SOL (via DeepZero) with three differentiable approaches methods considered in (Um et al., 2020): SRC (low fidelity simulation without error correction), NON (non-interactive training out of the simulation loop using pre-generated low and high fidelity simulation data), and FO-SOL (FO training for SOL given a differentiable simulator). The simulation error for each method in Fig. 6 is measured as the average across the five test simulations (with varying Reynold\u2019s numbers that were not utilized in the training set of simulation data). The error for each test simulation is computed as the mean absolute error (MAE) of the corrected simulation compared to the high fidelity simulation averaged across all simulation timesteps. We implement DeepZero for ZO-SOL following the setup used in the image classification task, except for choosing a 95% gradient sparsity. The ZO-SOL and FO-SOL use 16 unrolling steps in the loss function to allow the correction function to interact with the simulator during training. The results demonstrate that ZO-SOL achieved by DeepZero outperforms the SRC and NON baselines, and narrows the performance gap with FO-SOL, despite only having query-based access to the simulator. Comparing ZO-SOL with NON highlights the promise of ZO-SOL even when integrated with black-box simulators. 7"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2211.11635v5",
+                    "title": "Understanding and Improving Visual Prompting: A Label-Mapping Perspective",
+                    "abstract": "We revisit and advance visual prompting (VP), an input prompting technique\nfor vision tasks. VP can reprogram a fixed, pre-trained source model to\naccomplish downstream tasks in the target domain by simply incorporating\nuniversal prompts (in terms of input perturbation patterns) into downstream\ndata points. Yet, it remains elusive why VP stays effective even given a\nruleless label mapping (LM) between the source classes and the target classes.\nInspired by the above, we ask: How is LM interrelated with VP? And how to\nexploit such a relationship to improve its accuracy on target tasks? We peer\ninto the influence of LM on VP and provide an affirmative answer that a better\n'quality' of LM (assessed by mapping precision and explanation) can\nconsistently improve the effectiveness of VP. This is in contrast to the prior\nart where the factor of LM was missing. To optimize LM, we propose a new VP\nframework, termed ILM-VP (iterative label mapping-based visual prompting),\nwhich automatically re-maps the source labels to the target labels and\nprogressively improves the target task accuracy of VP. Further, when using a\ncontrastive language-image pretrained (CLIP) model, we propose to integrate an\nLM process to assist the text prompt selection of CLIP and to improve the\ntarget task accuracy. Extensive experiments demonstrate that our proposal\nsignificantly outperforms state-of-the-art VP methods. As highlighted below, we\nshow that when reprogramming an ImageNet-pretrained ResNet-18 to 13 target\ntasks, our method outperforms baselines by a substantial margin, e.g., 7.9% and\n6.7% accuracy improvements in transfer learning to the target Flowers102 and\nCIFAR100 datasets. Besides, our proposal on CLIP-based VP provides 13.7% and\n7.1% accuracy improvements on Flowers102 and DTD respectively. Our code is\navailable at https://github.com/OPTML-Group/ILM-VP.",
+                    "authors": "Aochuan Chen, Yuguang Yao, Pin-Yu Chen, Yihua Zhang, Sijia Liu",
+                    "published": "2022-11-21",
+                    "updated": "2023-03-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "main_content": "Introduction When learning new knowledge, humans typically start to compare and connect it with the knowledge that they were familiar with. The same idea is also applied in ML. For example, in the \u2018pretraining + \ufb01netuning\u2019 paradigm, an ML model (e.g., deep neural network or DNN) is \ufb01rst trained on a (usually large) source dataset. When a relevant downFig. 1. Overview of VP pipelines (prior art [1,2] and our proposal termed ILM-VP) and accuracy improvement achieved by ILM-VP on target image classi\ufb01cation tasks at-a-glance. Generally speaking, VP aims to generate a universal input perturbation template (i.e., \u2018visual prompt\u2019) and leverage a source-target LM (label mapping) in order to drive the \ufb01xed source model (e.g., pretrained on ImageNet-1K) to conduct a target task (e.g., Flowers102 image classi\ufb01cation). Compared to the prior art, our proposal (ILM-VP) couples the design of LM with VP training. The resulting LMVP co-design improves target task accuracy across a variety of target image classi\ufb01cation tasks using a \ufb01xed ImageNet-pretrained source model. stream task is present, the pre-trained model is then \ufb01netuned over the target dataset. This learning paradigm has been predominant in the classical transfer learning [3\u20138] as well as in the recent deep representation learning [9\u201313]. However, \ufb01netuning the pre-trained model requires either partial or entire model modi\ufb01cations. If the pretrained model is of large size, then it becomes too costly to store a modi\ufb01ed copy of the pre-trained model for each downstream task. In contrast, visual prompting (VP) (see Fig. 1), also known as model reprogramming or adversarial reprogramming, provides a new alternative to \ufb01netuning [1, 2, 14\u201317]. Instead of directly modifying the pretrained source model, VP integrates an input transformaarXiv:2211.11635v5  [cs.CV]  24 Mar 2023 \ftion and/or an output transformation to reprogram the \ufb01xed source model to accomplish a new target task; see an illustration of existing VP framework in Fig. 1. The input transformation is typically realized by incorporating (dataagnostic) input perturbations (i.e., prompts) into input samples, and the output transformation is given by a function that maps source labels to target labels, known as label mapping (LM). Recently, VP has shown great promise in various applications of foundation models, ranging from pretrained vision models [1, 14, 15, 17\u201320] to language-vision models [2,21\u201323]. The idea of prompt learning originated from in-context learning or prompting in natural language processing (NLP) [24\u201326]. However, when it is introduced to the vision domain [1, 2], new questions arise. First, the recent work [1,14,27] showed that VP remains powerful even if the target task largely deviates from the source domain. For example, a new performance record on target medical datasets is achieved in [1] when using VP to reprogram the \ufb01xed, ImageNet pre-trained source model. The \u2018mystery\u2019 in this example is that LM is conducted between two seemingly irrelevant source and target domains. Despite the lack of interpretability, VP can still leverage such connected source labels and the source model to effectively predict target data points. This raises the \ufb01rst open question: What is the rationality behind LM and how to explore its in\ufb02uence on VP? Second, unlike prompt learning in the NLP domain, input prompts in the vision domain are typically given by \u2018noisy\u2019 perturbations to image pixels; see illustration in Fig. 1. Together with the lack of interpretability of LM, the second open question is: How to interpret LM and the seemingly random perturbation pattern in VP? As mentioned above, the lack of understanding of LM and the poor interpretability of VP drive our studies in this work. We develop a new visual prompting framework, termed ILM-VP (iterative label mapping-based visual prompting), which provides an interactive and explainable design between LM and prompt learning (i.e., input prompt generation); see Fig. 1 for the schematic overview. Our proposal can automatically adjust LM between the source domain and the target domain by taking both mapping precision and explanation into consideration, and can leverage the optimized LM to further improve the accuracy and the explainability of prompt learning. Although some prior work [1,17,27] attempted to improve the quality of LM as well as the overall performance of VP, they are different from our proposal in two major aspects. First, none of the prior work co-designed LM and VP. For example, the prior art [1] used a pre-prompt prediction frequency to determine the LM function. However, we \ufb01nd signi\ufb01cant inconsistency between the pre-prompt and post-prompt prediction frequency of the same source model, which explains the sub-optimality of the current VP methods due to the lack of mapping precision. Second, to the best of our knowledge, VP is still treated as a \u2018black box\u2019 in the prior work. Yet, our design can provide graceful visual explanations to the underlying mechanisms of VP. Third, we for the \ufb01rst time show that LM can provide a uni\ufb01ed solution to improving the accuracy of VP to re-purpose both vision and language-vision source models. Our contributions are unfolded below. x We revisit the LM problem in VP and uncover the de\ufb01ciencies of existing LM methods: the lack of mapping precision and the lack of explanation. y Given the importance of LM, we propose the \ufb01rst LMVP co-design framework, termed ILM-VP, through a novel bi-level optimization viewpoint. z Beyond LM for vision models, we show that LM can also be generalized to assist the text prompt selection of CLIP (contrastive language\u2013image pretraining) and to improve the target task accuracy of VP using the CLIP model. { We empirically demonstrate the accuracy and explanation merits of our proposal across multiple source models and target datasets. 2. Related Work Prompting in NLP. Prompting is used to prepend language instruction to the input text for a language model to better accomplish a given task [28]. While prompting makes a signi\ufb01cant contribution to the generalization ability of large pre-trained language models (e.g., GPT-3) [24], it requires hand-crafting prompt design by experts. Recent work proposed to directly optimize the prompting embeddings through gradients together with lightweight \ufb01netuning the model, which is called prompt tuning [25, 29]. It is shown that this method is effective and ef\ufb01cient, which achieves competitive performance to the \ufb01netuning of the full language model. Visual prompting and model reprogramming. VP was \ufb01rst de\ufb01ned in [2] to mimic the prompting idea in NLP. Prior to that, a very similar idea was used in computer vision (CV) but with a different name, known as model reprogramming or adversarial reprogramming [14\u201317, 30\u201333]. They both focus on re-purposing a \ufb01xed, pre-trained vision model for a new task by leveraging a universal input pattern and an output LM function. Although not outperforming full \ufb01ne-tuning in transfer learning, VP yields an advantage of parameter-ef\ufb01cient \ufb01ne-tuning, which requires a much smaller parameter storage space. Furthermore, the smaller parameter space requires less training data to converge. Beyond traditional pre-trained vision models, the work [2] studied the effectiveness of VP in the languagevision model CLIP for the \ufb01rst time. Assisted by CLIP, VP can generate a prompting pattern of image data without resorting to source-target label mapping. In [23], VP and text prompt are jointly optimized in the CLIP model, which leads to better performance. Furthermore, unadver\fsarial learning [34] also enjoys the similar idea to VP, while it focuses on generating class-wise prompts with the goal of improving the out-of-distribution generalization ability of a pre-trained model. VP is gaining increasing attention. In [1], it is applied to re-purpose black-box source models [35] and achieves state-of-the-art (SOTA) performance on different target datasets. Besides, in data-scarce regimes like the biochemical domain, it is shown in [15, 17, 27] that VP can enable effective cross-domain transfer learning. Other than transfer learning, VP is also used in in-domain settings to improve different metrics like adversarial robustness [33] and fairness [32]. Although input prompting is the most commonlyused prompt learning method in the vision domain, generalization to learning prompting parameters at intermediate layers of a source model is also developed in [19\u201321, 36]. The resulting technique is called visual prompt tuning and is typically restricted to vision transformers. 3. Problem Statement In this section, we begin by providing some background information on VP. Based on that, we will then present the problem of our interest\u2013LM (label mapping)\u2013which de\ufb01nes how a visual prompt maps a source model prediction label to a target data class. This is the \ufb01rst question encountered in VP across domains but was typically overlooked in the literature. By reviewing the commonly-used LM methods, we will point out several open questions raised by LM. Preliminaries on visual prompting. The technology of VP addresses the problem of how to adapt a pre-trained source model (e.g., the ImageNet-1K-pre-trained ResNet18) to a target downstream task (e.g., \ufb02ower classi\ufb01cation over the Flowers102 dataset) without any task-speci\ufb01c model modi\ufb01cation (e.g., \ufb01netuning). Throughout the paper, we focus on input-based VP (also known as model reprogramming) [1, 2, 14\u201317, 27, 37], which incorporates a carefully-designed universal perturbation pattern to the raw target images so as to enforce the transferability of the source model to the target domain. We refer readers to Fig. 1 for the schematic overview. To be concrete, let S and T denote the source dataset and the target dataset, respectively. And let f\u03b8s denote a source model with pre-trained parameters \u03b8s. Suppose f\u03b8s is a supervised classi\ufb01er, then it de\ufb01nes a mapping from the input data x \u2208RNs to the source label space Ys \u2286RKs, i.e., f\u03b8s(x) = ys \u2208Ys, where Ns is the dimension of a source datapoint, Ks is the number of source data classes, and ys is the source class label. We have f\u03b8s trained based on S, e.g., via empirical risk minimization. The goal of VP is to reprogram the source model f\u03b8s to accomplish the target task de\ufb01ned in T , without making task-speci\ufb01c \ufb01netuning over f\u03b8s. To this end, VP modi\ufb01es the target data xt (of Nt dimensions) by injecting a task-designated input perturbation pattern \u03b4. This leads to the input prompting operation with the generic form: x\u2032(\u03b4) = h(xt, \u03b4) \u2208RNs, xt \u2208RNt (1) where xt is the target datapoint, and h(\u22c5,\u22c5) is an input transformation that integrates xt with the input perturbation \u03b4 and produces a modi\ufb01ed datapoint x\u2032(\u03b4) with the source data dimension Ns. It was shown in [1] that h can be speci\ufb01ed as an additive perturbation model that pads \u03b4 outside the target data sample (see Fig. 1 for an example as well). Given the input prompting model (1), VP then seeks the optimal \u03b4 to improve the target task accuracy when using the pre-trained source model f\u03b8s. This raises a prompt generation problem, which is typically cast as minimize \u03b4 E(xt,yt)\u2208Ttr [\u2113VP(f\u03b8s(x\u2032(\u03b4)), yt)], (2) where Ttr denotes a supervised training set in T with feature xt and label yt for a training sample, and \u2113VP(\u22c5) is a visual prompting loss function that we will de\ufb01ne later given the prompted input x\u2032(\u03b4) and the ground-truth target label yt. To solve problem (2), the standard stochastic gradient descent (SGD) method can be used. At inference, we will integrate the designed \u03b4 into test-time target datapoints and call the source model f\u03b8s for downstream prediction in T (see Fig. 1 and a more detailed description in Fig. A1). Label mapping: Existing methods and questions. Although the input prompting operation (1) converts the original xt to the source dimension-aligned datapoint x\u2032 that the source model can use, the successful realization of VP (3) needs to map the source model\u2019s prediction (in the source label space Ys with Ks classes) to the target task\u2019s data label (in the target label space Yt with Kt classes). In the \u2018pretraining + \ufb01netuning\u2019 paradigm, we typically have Kt \u2264Ks. Therefore, the problem of LM (label mapping) arises: (LM problem) Given the source model f\u03b8s, how to build a mapping from the source label space Ys to the target label space Yt so that the model\u2019s prediction directs to the correct target label? Clearly, the desired prompt generation (2) heavily relies on the LM scheme, which de\ufb01nes the one-to-one correspondence between the source model\u2019s prediction f\u03b8s(x\u2032(\u03b4)) and the target data class yt. Yet, nearly all the existing work neglects its in\ufb02uence on the prompt generation and adopts either x the simplest random mapping [2,14] or y a pre-de\ufb01ned, one-shot frequency-based mapping [1,15]. We elaborate on the above two schemes below. x Random label mapping (RLM): RLM does not use any prior knowledge or source model information to guide the LM process. The mapped source labels (to the target domain) could be even random. For example, in the case of \u2018ImageNet (source) + CIFAR-10 (target)\u2019 [2, 14], existing VP methods coded CIFAR-10 labels using the top 10 ImageNet indices, i.e., ImageNet label i \u2192CIFAR-10 label i, despite the lack of interpretation. \fFig. 2. Visualizations of RLM and FLM using the source dataset ImageNet-1K and the pretrained ResNet-18, as well as the target dataset Flowers102. In FLM, the pre-prompt label mapping using (3) selects source labels different from FLM. Yet, the post-prompt label mapping using (3) at \u03b4\u2217in (4) shows many newly-selected source labels, indicating (1) the dynamics of LM in the source domain, and (2) the pre-promp LM is sub-optimal (i.e., mis-selecting the best-matching label) after VP training. y Frequency-based label mapping (FLM): FLM matches target labels to source labels based on the source model\u2019s prediction frequencies on zero-padded target datapoints, i.e., f\u03b8s(x\u2032(\u03b4)) with \u03b4 = 0. Here recall from (1) that x\u2032(0) = h(xt,0). More concretely, a target label yt is mapped to the source label y\u2217 s following y\u2217 s (yt) = arg maxys Pr {Top-1 prediction of f\u03b8s(h(xt, 0)) is ys \u2223\u2200xt \u2208Tyt} (3) where y\u2217 s (yt) explicitly expresses the dependence of the mapped source label on the target label, Tyt denotes the target data set in the class yt, and Pr{\u22c5} is the probability of the event that the top-1 prediction of f\u03b8s is the source class ys under the zero-padded target data points in Tyt. As shown in Fig. 2, FLM results in a mapping scheme different from that of RLM. However, it is still dif\ufb01cult to interpret the obtained LM results, and remains elusive how the quality of LM impacts the performance of VP. In the rest of the work, we will shed light on how to improve VP by carefully designing the LM scheme and when and why LM matters at different source and downstream tasks. 4. Method: Iterative Label Mapping-based VP In this section, we uncover the hidden dynamics of LM existing in the source task domain of VP, which was neglected by the prior art. This \ufb01nding then motivates us to develop a novel VP framework, which we call Iterative LMbased VP (ILM-VP). Compared to existing VP methods, ILM-VP closes the loop between LM and prompt generation (2), and improves VP\u2019s explanation and target task accuracy simultaneously. The \u2018missing\u2019 dynamics of LM in the source domain. As shown in Sec. 3, a prompt learning pipeline mainly involves three steps: (A1) input prompt modeling (1), (A2) LM (from the source label set Ys to the target label set Yt), and (A3) prompt generation (2). The prior art follows the pipeline (A1)\u2192(A2)\u2192(A3) to generate the desired prompt \u03b4\u2217, which drives the source model to accomplish target tasks. However, in the viewpoint of the source domain, the prompt updating from \u03b4 = 0 to \u03b4\u2217induces the prediction dynamics of the source model f\u03b8s. That is, f\u03b8s(x\u2032(0)) \u2192f\u03b8s(x\u2032(\u03b4\u2217)), (4) where x\u2032(\u03b4) has been de\ufb01ned in (1), which refers to the \u03b4perturbed target data with the same dimension as the source datapoint. As will be evident later, it is important to understand the dynamics (4) as it re\ufb02ects the stability of the selected source labels when mapping to the target labels. Fig. 2 instantiates the dynamics of (4) in the scenario of \u2018ImageNet (source) + Flowers102 (target)\u2019 when the FLMoriented VP approach (3) is used [1]. Prior to prompt generation, the Flowers102 target labels are \ufb01rst mapped to the ImageNet source labels using the FLM method (3), corresponding to step (A2) in prompt learning. This yields the pre-prompt target-source mapping, denoted by f\u03b8s(x\u2032(0)) FLM \u2192Yt. Similarly, after generating the prompt \u03b4\u2217 following (A3), we can obtain the post-prompt target-source mapping, f\u03b8s(x\u2032(\u03b4\u2217)) FLM \u2192Yt, using the FLM method. Fig. 2 shows that there exists a signi\ufb01cant discrepancy between the pre-prompt LM and the post-prompt LM, evidenced by the newly-selected source labels (\u2018Cardoon\u2019, \u2018Egret\u2019, \u2018Daisy\u2019, \u2018Paper Towel\u2019) in the post-prompting phase. This justi\ufb01es the dynamics of (4) in LM. However, it also raises a new concern that the pre-prompt target-source LM is suboptimal for prompt generation (2) given the existing dynamics of LM in the source domain. ILM-VP: A bi-level optimization viewpoint of VP. The dynamics of LM inspire us to re-think the optimality of the current VP pipeline: (A1)\u2192(A2)\u2192(A3). To improve it, we propose to take the LM dynamics into the prompt learning process. This modi\ufb01es the conventional VP pipeline to (A1)\u2192(A2)\u21c4(A3), where LM and prompt generation are in a closed loop. Since the design of LM will interact with the design of the prompt iteratively, we call the proposed new design ILM-VP. Next, we formally present ILM-VP through the lens of bi-level optimization (BLO). Generally speaking, BLO provides a hierarchical learning framework involving two levels (i.e., upper and lower levels) of optimization tasks, where one task is nested inside the other (i.e., the objective and variables of an upper-level problem depend on the optimizer of the lower-level problem). In the context of ILM-VP, we regard the prompt generation problem (2) as the upper-level optimization task and the LM problem (3) \fas the lower-level problem. This yields minimize \u03b4 E(xt,yt)\u2208Ttr [\u2113(f\u03b8s(x\u2032(\u03b4)), y\u2217 s (yt))] \u00b4\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b8\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b6 Upper-level prompt optimization subject to y\u2217 s (yt) is obtained by (3) at (non-zero) prompt \u03b4 \u00b4\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b8\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b9\u00b6 Lower-level LM design at current prompt \u03b4 for every target label yt (5) where the visual prompt \u03b4 denotes the upper-level variable, \u2113is the cross-entropy loss, and the mapped source label ys is a lower-level variable for each given target label yt at the current prompt \u03b4. We also note that there exists a lowerlevel constraint in (5) to ensure that if a source class has been mapped to a target class, it will then be excluded when mapping to a new target class. Further, it is clear from (5) that the design of visual prompt \u03b4 and LM y\u2217 s (vs. yt) are intertwined with each other. To solve problem (5), we employ the alternating optimization (AO) method, which alternatively executes the upper-level prompt generation and the lower-level LM. We summarize the algorithm details in Algorithm 1 and provide a schematic overview in Fig. A2. Algorithm 1 The proposed ILM-VP algorithm 1: Initialize: Given target training set Ttr, pre-trained model f\u03b8s, prompt pattern initialization \u03b40, and upperlevel learning rate \u03bb for SGD 2: for Epoch n = 0,1,..., do 3: Lower-level label mapping: Given \u03b4n\u22121, call LM for each target class yt in Ttr 4: Upper-level prompt learning: Given LM, call SGD to update prompt \u03b4n \u2190 \u00d0 \u03b4n\u22121 5: end for An interpretation merit of ILM-VP. In the literature, it is quite dif\ufb01cult to interpret why VP can reprogram a source model to conduct target tasks. The main hurdle of interpreting VP lies in the LM phase: It remains elusive why the semantics-irrelevant source labels should be mapped to target labels. However, we \ufb01nd that ILM-VP can alleviate this interpretation dif\ufb01culty to a large extent. We show the explanation merit of ILM-VP through an empirical study in Fig. 3, where the target dataset is instantiated by Flowers102 and the source dataset is ImageNet-1K. We list the target labels, the mapped source labels using the baseline FLM method [1], and the identi\ufb01ed source labels using ILM-VP, together with image examples under each label. As we can see, an interpretable target-source mapping is found by ILM-VP, even if the target label and the source label describe different subjects. For example, target images in the label \u2018Spear Thistle\u2019 share a similar color and object shape with the source images in the label \u2018Cardoon\u2019. The same observations can also be drawn from other targetsource label mappings together with their data instances. This \ufb01nding is quite encouraging and is in sharp contrast to Fig. 3. Interpretation merit of ILM (ours) vs. FLM, visualized by LM results in VP to re-purpose an ImageNet-pretrained source model (ResNet18) to conduct target image classi\ufb01cation tasks on the target datasets Flowers102, OxfordPets, DTD, and Food101. ILM consistently \ufb01nds more interpretable target-source label mappings than FLM, in terms of colors, scenes, shapes, and textures. See Fig. A3 for more examples. FLM. As will be evident in Sec. 5, the BLO-oriented ILMVP (5) would enforce a convergence of LM as the alternating optimization proceeds. As a result, source labels and target labels, which share the most similar concepts (like colors, scenes, shapes, and materials), will be identi\ufb01ed. We also show that the improved interpretation of LM consistently enhances the target task accuracy of the VP. 5. Experiments In this section, we empirically demonstrate the effectiveness of our proposed ILM-VP method by comparing it with a variety of baselines across multiple datasets, models, and learning paradigms. 5.1. Experiment setups Datasets and models. In the source domain, we will consider the source models ResNet-18 and ResNet-50 [38] pre-trained on ImageNet-1K [39], and the source model ResNeXt-101-32x8d [40] pre-trained on Instagram [41]. In the target domain, we will evaluate the performance of ILM-VP over 13 target datasets: Flowers102 [42], DTD [43], UCF101 [44], Food101 [45], GTSRB [46], SVHN [47], EuroSAT [48], OxfordPets [49], StanfordCars [50], SUN397 [51], CIFAR10/100 [52], ABIDE [53]. \fSource Model ResNet-18 (ImageNet-1K) ResNet-50 (ImageNet-1K) ResNeXt-101-32x8d (Instagram) Method Ours Prompt baseline Finetuning Ours Prompt base. Finetuning Ours Prompt base. Finetuning ILM-VP RLM-VP FLM-VP LP FF ILM-VP FLM-VP LP FF ILM-VP FLM-VP LP FF Parameter Size 0.05M 0.05M 0.05M 0.51M 11.7M 0.05M 0.05M 0.51M 25.6M 0.05M 0.05M 0.51M 88.8M Flowers102 27.9\u00b10.7 11.0\u00b10.5 20.0\u00b10.3 88.0\u00b10.5 97.1\u00b10.7 24.6\u00b10.6 20.3\u00b10.3 90.9\u00b10.4 97.9\u00b10.7 27.9\u00b10.3 22.5\u00b10.5 89.1\u00b10.2 99.2\u00b10.5 DTD 35.3\u00b10.9 16.3\u00b10.7 32.4\u00b10.5 60.0\u00b10.6 65.5\u00b10.9 40.5\u00b10.5 36.9\u00b10.8 67.6\u00b10.3 69.7\u00b10.9 41.4\u00b10.7 40.3\u00b10.5 69.7\u00b10.2 69.1\u00b11.0 UCF101 23.9\u00b10.5 6.6\u00b10.4 18.9\u00b10.5 63.2\u00b10.8 73.0\u00b10.6 34.6\u00b10.2 33.9\u00b10.4 70.8\u00b10.3 78.0\u00b10.8 43.1\u00b10.8 41.9\u00b10.6 76.9\u00b10.5 79.1\u00b10.7 Food101 14.8\u00b10.2 3.8\u00b10.3 12.8\u00b10.1 50.6\u00b10.3 75.4\u00b10.8 17.0\u00b10.3 15.3\u00b10.2 57.6\u00b10.5 80.3\u00b10.9 23.0\u00b10.4 20.5\u00b10.5 76.0\u00b10.4 82.5\u00b10.3 GTSRB 52.0\u00b11.2 46.1\u00b11.3 45.5\u00b11.0 77.4\u00b11.2 98.0\u00b10.3 52.5\u00b11.4 47.6\u00b11.1 77.8\u00b10.7 97.6\u00b11.0 59.9\u00b11.0 56.2\u00b10.6 73.5\u00b10.7 97.6\u00b10.9 EuroSAT 85.2\u00b10.6 82.4\u00b10.4 83.8\u00b10.2 93.8\u00b10.3 98.8\u00b10.5 83.6\u00b10.7 84.8\u00b10.3 95.7\u00b10.2 98.9\u00b10.6 86.2\u00b10.8 87.8\u00b10.4 93.4\u00b10.3 98.9\u00b10.7 OxfordPets 65.4\u00b10.7 9.3\u00b10.4 62.9\u00b10.1 87.2\u00b10.6 87.8\u00b10.5 76.2\u00b10.6 76.4\u00b10.2 90.4\u00b10.3 91.9\u00b10.4 78.9\u00b10.8 76.8\u00b10.6 93.6\u00b10.4 90.1\u00b10.9 StanfordCars 4.5\u00b10.1 0.9\u00b10.1 2.7\u00b10.1 33.8\u00b10.2 81.0\u00b10.1 4.7\u00b10.2 4.2\u00b10.3 40.6\u00b10.1 86.4\u00b10.3 7.0\u00b10.2 4.6\u00b10.1 64.7\u00b10.1 92.5\u00b10.2 SUN397 13.0\u00b10.2 1.0\u00b10.1 10.4\u00b10.1 46.1\u00b10.2 53.2\u00b10.2 20.3\u00b10.2 19.8\u00b10.1 53.5\u00b10.1 59.0\u00b10.1 23.7\u00b10.2 21.6\u00b10.3 62.3\u00b10.1 61.0\u00b10.2 CIFAR10 65.5\u00b10.1 63.0\u00b10.1 65.7\u00b10.6 85.9\u00b10.5 96.5\u00b10.4 76.6\u00b10.3 74.8\u00b10.5 90.1\u00b10.1 96.6\u00b10.2 81.7\u00b10.3 80.3\u00b10.3 94.1\u00b10.1 97.1\u00b10.1 CIFAR100 24.8\u00b10.1 12.9\u00b10.1 18.1\u00b10.2 63.3\u00b10.8 82.5\u00b11.2 38.9\u00b10.3 32.0\u00b10.4 70.7\u00b10.7 83.4\u00b10.9 45.9\u00b10.2 39.7\u00b10.2 76.2\u00b10.9 84.6\u00b11.2 SVHN 75.2\u00b10.2 73.5\u00b10.3 73.1\u00b10.2 65.0\u00b10.2 96.5\u00b10.3 75.8\u00b10.4 75.6\u00b10.2 63.5\u00b10.2 96.9\u00b10.3 81.4\u00b10.1 79.0\u00b10.5 51.0\u00b10.2 97.1\u00b10.3 ABIDE 76.9\u00b12.1 74.0\u00b12.2 73.1\u00b11.6 65.4\u00b13.8 60.6\u00b14.2 63.5\u00b12.2 64.4\u00b13.4 55.8\u00b12.6 70.2\u00b12.5 67.3\u00b12.6 65.7\u00b13.4 54.8\u00b13.4 73.1\u00b14.2 Tab. 1. Performance overview of our proposed VP method (ILM-VP), prompt baseline methods (RLM-VP and FLM-VP), and \ufb01netuning methods (LP and FF) over 13 target image classi\ufb01cation datasets using 3 pretrained source models (ResNet-18 on ImageNet-1K, ResNet-50 on ImageNet-1K, and ResNeXt101-32x8d on Instagram). In each cell, a\u00b1b refers to the mean and standard deviation of target task accuracies (%) over 3 independent trials. The highest accuracy across VP-based methods is marked in bold. \u2018Parameter Size\u2019 refers to the number of trainable parameters in the input prompt or model \ufb01netuning. 50 75 oxeye daisy* spear thistle* osteospermum* artichoke* dandelion gentian gazania* windflower* pincushion wales feathers 93.3% 85.7% 63.2% 60.9% 46.4% 45.0% 43.5% 37.5% 35.3% 33.3% (a) Flowers102 20 40 russian blue* havanese* maine coon* yorkshire bombay* bengal* ragdoll bull terrier leonberger siamese 38.0% 23.0% 21.0% 19.0% 15.9% 13.0% 10.0% 8.0% 7.0% 5.0% (b) OxfordPets Fig. 4. ILM-VP\u2019s class-wise accuracy improvements over FLM-VP using ImageNet-1K pre-trained ResNet-18 under the target dataset (a) Flowers102, (b) OxfordPets. Target classes with \u2217refer to those with updated source labels compared to the prompting baseline method FLM-VP. Baselines and evaluations. In the VP paradigm, our baseline methods include the random LM-based VP (RLMVP) [2, 14], and the frequency LM-based VP (FLM-VP) [1, 15]. We highlight that VP is a \ufb01netuning-free method to drive the source model to conduct target image classi\ufb01cation tasks. When implementing VP baselines, we follow their of\ufb01cial repository setups. We also refer readers to Appendix D for detailed implementations of ILM-VP and baseline methods. In addition to prompting methods, we also cover \ufb01netuning-based methods, including linear probing (LP) and end-to-end full \ufb01netuning (FF). Since \ufb01netuning modi\ufb01es source model parameters, it requires training more parameters and is more computationally intensive. We evaluate the performance of all the methods by the target task accuracy at the testing time and the ef\ufb01ciency in terms of the parameter size that VP or \ufb01netuning needs to handle. We also leverage a post-hoc model explanation method, known as Explanation-by-Example (EBE) [54], to assess the quality of visual explanation of different VP methods. The core idea of EBE is to \ufb01nd train-time data points that have the most similar feature representations to that of a queried test datapoint so as to use these identi\ufb01ed training samples to explain the model\u2019s prediction on this test sample. In the context of VP, EBE can aid us to \ufb01nd the source training samples explainable for model prediction on prompted target test data, like source examples in Fig. 3. 5.2. Experiment results Overall performance of ILM-VP. Tab. 1 shows the effectiveness of our proposed ILM-VP method vs. VP baselines (RLM-VP and FLM-VP) on diverse source models and target datasets. For comparison, we also present the model \ufb01netuning performance on target datasets using LP or FF. It is worth noting that FLM-VP typically outperforms RLM-VP as the latter only uses a random label mapping to guide the learning of prompts [1]. Thus, we only show the results of RLM-VP when using ResNet-18. As shown in Tab. 1, our proposed method (ILM-VP) consistently outperforms other VP baselines by a large margin in nearly all the data-model setups, e.g., 7.9%, 6.7% and 6.5% accuracy improvement over FLM-VP in the target dataset Flowers102, CIFAR100, GTSRB, respectively. In addition, we note that model \ufb01netuning is typically more effective in transfer learning than prompting methods, consistent with existing work [2]. This is not surprising as source models are allowed for modi\ufb01cation, and the trainable parameter size increases (as evidenced by \u2018Parameter Size\u2019 in Tab. 1). As will be evident in Sec. 6, the accuracy of VP can be further improved if a language-vision source model is used. Nonetheless, in the target dataset ABIDE, prompting methods can outperform the full model \ufb01netuning method (FF). Compared to other standard transfer learning tasks for image classi\ufb01cation, ABIDE was a newly-proposed medical dataset in [1], which converts the original 1D numerical medical input sequences to image-alike data formats (i.e., brain-regional correlation graphs). The size of this dataset is extremely small due to the high cost of collecting data in the medical area, which restricts the performance of LP and FF. In contrast, VP is uniquely suited for this setting. Lastly, in the model \ufb01netuning paradigm, a source model with larger capacity typically yields a better target task accuracy, e.g., the \ufb01netuning results of ResNet-50 vs. ResNet18. However, this belief might not hold in the VP paradigm. As we can see, the prompting-induced target accuracy de\fFig. 5. ILM-VP training dynamics from epoch 0 to 200. Rows show: (1) VP pattern vs. epoch number; (2-4) Learned source label mapping with respect to target label \u2018Marigold\u2019, \u2018White Lily\u2019, and \u2018Tree Poppy\u2019, together with EBE-identi\ufb01ed source training examples to explain each re-purposed target label; (5) Convergence of training loss and LM difference between adjacent epochs measured by Hamming distance. creases under ResNet-50 in the target datasets Flowers102, EuroSAT, CIFAR100, and ABIDE. Additionally, Tab. A2 in Appendix shows that ILM-VP takes a bit more run time than FLM-VP and LP, but is faster than FF. This is not surprising since the former adopts alternating optimization with a bit higher computation complexity than ordinary single-level minimization. Recently, the concurrent work [55] shows that properly re-sizing images before integrating with a VP could further boost the performance on a downstream task. We also \ufb01nd the same bene\ufb01t of image re-sizing to VP on CIFAR10/100, GTSRB, and SVHN datasets (e.g., up-scaling the original image size to 128\u00d7128) However, for ease of comparison with existing VP baselines (RLM-VP [14]), our experiments do not apply the image re-sizing trick to VP. LM is key to improving the accuracy of VP. Next, we peer into the in\ufb02uence of LM on target prediction accuracy per class when using ILM-VP. In Fig. 4, we demonstrate the testing accuracy improvements (over the FLMVP baseline) of prompt-injected datapoints, belonging to 10 classes with the highest improvements selected from the target datasets Flowers102 and OxfordPets, respectively. Note that OxfordPets shares the most similar label space with ImageNet (e.g. they both have beagles, boxers, bassets, etc.). We use \u2217in Fig. 4 to mark target data classes whose source labels are remapped during ILM-VP, and list non-\u2217marked target data classes whose source labels retain the same as FLM-VP. We observe that target classes with large accuracy improvements typically require ILM. This justi\ufb01es the bene\ufb01t of target-source label re-mapping during prompt learning. In addition, we note that the source labels of target classes (e.g., \u2018yorkshire\u2019 in OxfordPets) are not re-mapped, but ILM-VP can still bring in accuracy improvements. This implies that LM has a coupling effect on all classes and the BLO framework (5) enables us to improve LM as well as prompt learning in an interactive manner. Further, Fig. 5 shows the training dynamics of ILM-VP vs. training epoch number and its convergence to the stable, high-explainable, and high-accurate visual prompt. As we can see, the mapped source label for a target class is updated at the early training epochs of ILM-VP, but tends to converge at the later training phase. A similar trend holds for the convergence of LM difference between two adjacent epochs and the VP training loss. In addition, we can see that the VP pattern and the LM are updated jointly. Furthermore, the explainability of mapped source labels grows as the training proceeds. For example, the target label \u2018Marigold\u2019 shares a similarity with the source label \u2018Orange\u2019 in color and shape, as visualized by EBE-identi\ufb01ed examples. It is worth mentioning that EBE facilitates us to directly link the source dataset and the target dataset, and thus helps us to better understand the rationale behind VP. We refer readers to Appendix C for more EBE results. How target dataset scale affects VP? Through our experiments over a large number of target datasets, we \ufb01nd that ILM-VP becomes more powerful when it comes to tasks with a larger target label space. For example, Fig. 6 shows the target datasets with at least 3% accuracy improvement using ILM-VP compared with FLM-VP on ResNet18. As we can see, target datasets with the highest number of target classes correspond to the most signi\ufb01cant accuracy improvement brought by ILM-VP. Next, we \ufb01x the target dataset and study how VP behaves at different downstream training dataset sizes. Here we choose GTSRB as the target task since GTSRB contains a suf\ufb01cient amount of training data and thus facilitates us to conduct training dataset par\f0 50 100 Class Number 3 4 5 6 7 8 9 Improvements over SOTA (%) 7.89% 6.67% 6.47% 5.05% 3.85% Flowers102 CIFAR100 GTSRB UCF101 ABIDE Fig. 6. ILM-VP\u2019s improvements over FLM-VP on representative datasets (datasets with improvements over 3%) using ResNet-18. The dashed line is the \ufb01tted curve. 20% 40% 60% 80% 100% Downstream Data Fraction 42 44 46 48 50 52 Accuracy (%) FLM-VP ILM-VP Fig. 7. ILM-VP and FLM-VP performance on different fractions of GTSRB dataset (43 classes and more than 900 training samples per class) using ResNet-18. tition. Fig. 7 compares the performance of ILM-VP with FLM-VP vs. target training dataset size (training from 20% to 100% of the entire set). As we can see, ILM-VP consistently outperforms the baseline FLM-VP and the improvement becomes more signi\ufb01cant as the data scale grows. 6. Extension: LM in Text Domain for CLIP In the previous sections, we show that LM could play a vital role in VP when reprogramming a pre-trained vision model to conduct downstream targeted vision tasks. In this section, we shift our focus from the vision source model to the vision-language model, speci\ufb01c to CLIP (contrastive language\u2013image pretraining), which has received increasing attention in the area of VP [2]. We will show that although CLIP does not require source-target mapping across image labels (due to its multi-modal learning architecture), the proposed idea on iterative LM can be extended to conduct text prompt selection to improve target task accuracy. LM for CLIP. Different from the vision-only model, CLIP can directly take target data labels as its textual inputs so as to mitigate the issue of source-target label mapping; see Fig. A5 for an illustration. In this setting, LM seems redundant. However, this is only applied to VP in the image domain. We argue that CLIP still needs implicit LM for text labels, considering the diversity of text prompts (TPs) [26]. That is, CLIP can incorporate a text label into different context prompt templates (81 templates) suggested in [26] to create multiple text label instances. For example, the target label \u2018dog\u2019 can be combined with context prompts \u2018A photo of a big {label}\u2019 and \u2018A photo of a small {label}\u2019. Thus, given m context prompts and Kt target data labels, we can create mKt \u2018virtual source labels\u2019, which should be mapped to Kt target labels. Thus, the LM problem arises, and its optimal solution characterizes the optimal prompt selection in the text domain for prompted image data. Similar to the BLO method (5), we can bake iterative LM into VP using the CLIP model by replacing the lower-level image label mapping with the context-fused text label mapping. BLO then gives a uni\ufb01ed prompt learning framework that can be easily compatible with CLIP. We refer readers to Appendix E for more implementation details. LM improves VP\u2019s accuracy using CLIP. In Tab. 2, we demonstrate the performance of the VP-driven CLIP model on several challenging target tasks shown in Tab. 1, e.g., Flowers102 and DTD. We term our method \u2018VP+TP+LM\u2019, where the BLO-enabled LM method is called to map \u2018virtual source labels\u2019 (i.e., the combination of context prompt template and target label) to realistic target image labels. For comparison, we also present the performance of the baseline method termed \u2018VP + TP\u2019 [2], which uses a prede\ufb01ned, \ufb01xed context prompt template \u2018This is a photo of a {label}\u2019 when generating a visual prompt for CLIP. As we can see, our proposal consistently outperforms the baseline by a substantial margin. For example, we obtain 13.7% and 7.1% accuracy gain in Flowers102 and DTD respectively. In addition, we \ufb01nd that LM brings in the interpretability merit: Our selected context prompt templates have better semantic meaning than the one used by the baseline. For example, VP for Flowers102 selects the text prompt \u2018a closeup photo of a {}\u2019 instead of \u2018This is a photo of {}\u2019 for the target image with the label \u2018buttercup\u2019. Another example is that VP for CIFAR10 prefers the text prompt \u2018a pixelated photo of a {}\u2019. In particular, we observe that in domain shift datasets (ImageNet-R and ImageNet-Sketch), the selected prompts can exhibit the domain information. More explainable results can be found in Fig. A6. Methods VP+TP Ours (VP+TP+LM) Acc(%) Acc(%) Examples of context prompt template \u2192target label Flowers102 70.0 83.7 a close-up photo of a {} \u2192buttercup DTD 56.8 63.9 graf\ufb01ti of a {} \u2192blotchy UCF101 66.0 70.6 a {} in a video game \u2192baseball pitch Food101 78.9 79.1 a photo of the dirty {} \u2192crab cake SVHN 89.9 91.2 a photo of a {} \u21927 EuroSAT 96.4 96.9 a pixelated photo of a {} \u2192river StanfordCars 57.2 57.6 the toy {} \u21922011 audi s6 sedan SUN397 60.5 61.2 a photo of a large {} \u2192archive CIFAR10 93.9 94.4 a pixelated photo of a {} \u2192ship ImageNet-R 67.5 68.6 a rendition of a {} \u2192gold \ufb01sh ImageNet-Sketch 38.5 39.7 a sketch of a {} \u2192eagle Tab. 2. Results of our CLIP-based prompt learning \u2018VP+TP+LM\u2019 and the baseline \u2018VP+TP\u2019 [2] (restricted to using text prompt template \u201cThis is a photo of a {}\u201d) over 11 target datasets. In each cell, the target task accuracy (%) is shown along with examples of LM in the text domain. Our method with higher accuracy than SOTA is marked in bold. 7."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.06284v4",
+                    "title": "Visual Prompting for Adversarial Robustness",
+                    "abstract": "In this work, we leverage visual prompting (VP) to improve adversarial\nrobustness of a fixed, pre-trained model at testing time. Compared to\nconventional adversarial defenses, VP allows us to design universal (i.e.,\ndata-agnostic) input prompting templates, which have plug-and-play capabilities\nat testing time to achieve desired model performance without introducing much\ncomputation overhead. Although VP has been successfully applied to improving\nmodel generalization, it remains elusive whether and how it can be used to\ndefend against adversarial attacks. We investigate this problem and show that\nthe vanilla VP approach is not effective in adversarial defense since a\nuniversal input prompt lacks the capacity for robust learning against\nsample-specific adversarial perturbations. To circumvent it, we propose a new\nVP method, termed Class-wise Adversarial Visual Prompting (C-AVP), to generate\nclass-wise visual prompts so as to not only leverage the strengths of ensemble\nprompts but also optimize their interrelations to improve model robustness. Our\nexperiments show that C-AVP outperforms the conventional VP method, with 2.1X\nstandard accuracy gain and 2X robust accuracy gain. Compared to classical\ntest-time defenses, C-AVP also yields a 42X inference time speedup.",
+                    "authors": "Aochuan Chen, Peter Lorenz, Yuguang Yao, Pin-Yu Chen, Sijia Liu",
+                    "published": "2022-10-12",
+                    "updated": "2023-05-01",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.CR",
+                        "cs.LG"
+                    ],
+                    "main_content": "INTRODUCTION Machine learning (ML) models can easily be manipulated (by an adversary) to output drastically different classi\ufb01cations. Thereby, model robusti\ufb01cation against adversarial attacks is now a major focus of research. Yet, a large volume of existing works focused on training recipes and/or model architectures to gain robustness. Adversarial training (AT) [1], one of the most effective defense, adopted min-max optimization to minimize the worst-case training loss induced by adversarial attacks. Extended from AT, various defense methods were proposed, ranging from supervised learning to semi-supervised learning, and further to unsupervised learning [2\u201311]. \u22c6Equal contribution. Although the design for robust training has made tremendous success in improving model robustness [12,13], it typically takes an intensive computation cost with poor defense scalability to a \ufb01xed, pre-trained ML model. Towards circumventing this dif\ufb01culty, the problem of test-time defense arises; see the seminal work in Croce et. al. [14]. Test-time defense alters either a test-time input example or a small portion of the pre-trained model. Examples include input (anti-adversarial) puri\ufb01cation [15\u201317] and model re\ufb01nement by augmenting the pre-trained model with auxiliary components [18\u201320]. However, these defense techniques inevitably raise the inference time and hamper the test-time ef\ufb01ciency [14]. Inspired by that, our work will advance the test-time defense technology by leveraging the idea of visual prompting (VP) [21], also known as model reprogramming [22\u201325]. Generally speaking, VP [21] creates a universal (i.e., dataagnostic) input prompting template (in terms of input perturbations) in order to improve the generalization ability of a pre-trained model when incorporating such a visual prompt into test-time examples. It enjoys the same idea as model reprogramming [22\u201325] or unadversarial example [26], which optimizes a universal perturbation pattern to maneuver (i.e., reprogram) the functionality of a pre-trained model towards the desired criterion, e.g., cross-domain transfer learning [24], out-of-distribution generalization [26], and fairness [25]. However, it remains elusive whether or not VP could be designed as an effective solution to adversarial defense. We will investigate this problem, which we call adversarial visual prompting (AVP) in this work. Compared to conventional test-time defense methods, AVP signi\ufb01cantly reduces the inference time overhead since visual prompts can be designed of\ufb02ine over training data and have the plug-and-play capability applied to any testing data. We summarize our contributions as below. \u0082 We formulate and investigate the problem of AVP for the \ufb01rst time and empirically show the conventional data-agnostic VP design is incapable of gaining adversarial robustness. \u0083 We propose a new VP method, termed class-wise AVP (CAVP), which produces multiple, class-wise visual prompts with explicit optimization on their couplings to gain better adversarial robustness. \u0084 We provide insightful experiments to demonstrate the pros and cons of VP in adversarial defense. arXiv:2210.06284v4  [cs.CV]  1 May 2023 \f2. RELATED WORK Visual prompting. Originated from the idea of in-context learning or prompting in natural language processing (NLP) [27\u201330], VP was \ufb01rst proposed in Bahng et. al. [21] for vision models. Before formalizing VP in Bahng et. al. [21], the underlying prompting technique has also been devised in computer vision (CV) with different naming. For example, VP is closely related to adversarial reprogramming or model reprogramming [22\u201324,31\u201333], which focused on altering the functionality of a \ufb01xed, pre-trained model across domains by augmenting test-time examples with an additional (universal) input perturbation pattern. Unadversarial learning also enjoys the similar idea to VP. In [26], unadversarial examples that perturb original ones using \u2018prompting\u2019 templates were introduced to improve out-of-distribution generalization. Yet, the problem of VP for adversarial defense is under-explored. Adversarial defense. The lack of adversarial robustness is a weakness of ML models. Adversarial defense, such as adversarial detection [19,34\u201338] and robust training [2,6,9,10, 18, 39], is a current research focus. In particular, adversarial training (AT) [1] is the most widely-used defense strategy and has inspired many recent advances in adversarial defense [12,13,20,40\u201342]. However, these AT-type defenses (with the goal of robustness-enhanced model training) are computationally intensive due to min-max optimization over model parameters. To reduce the computation overhead of robust training, the problem of test-time defense arises [14], which aims to robustify a given model via lightweight unadversarial input perturbations (a.k.a input puri\ufb01cation) [15,43] or minor modi\ufb01cations to the \ufb01xed model [44,45]. In different kinds of test-time defenses, the most relevant work to ours is anti-adversarial perturbation [17]. 3. PROBLEM STATEMENT Visual prompting. We describe the problem setup of VP following Bahng et. al. [21, 23\u201325]. Speci\ufb01cally, let Dtr denote a training set for supervised learning, where (x,y) \u2208 Dtr signi\ufb01es a training sample with feature x and label y. And let \u03b4 be a visual prompt to be designed. The prompted input is then given by x+\u03b4 with respect to (w.r.t.) x. Different from the problem of adversarial attack generation that optimizes \u03b4 for erroneous prediction, VP drives \u03b4 to minimize the performance loss \u2113of a pre-trained model \u03b8. This leads to minimize \u03b4 E(x,y)\u2208Dtr[\u2113(x + \u03b4; y, \u03b8)] subject to \u03b4 \u2208C, (1) where \u2113denotes prediction error given the training data (x,y) and base model \u03b8, and C is a perturbation constraint. Following Bahng et. al. [21, 23, 24], C restricts \u03b4 to let x + \u03b4 \u2208[0,1] for any x. Projected gradient descent (PGD) [1,26] can then be applied to solving problem (1). In the evaluation, \u03b4 is integrated into test data to improve the prediction ability of \u03b8. Adversarial visual prompting. Inspired by the usefulness of VP to improve model generalization [21,24], we ask: (AVP problem) Can VP (1) be extended to robustify \u03b8 against adversarial attacks? At the \ufb01rst glance, the AVP problem seems trivial if we specify the performance loss \u2113as the adversarial training loss [1,2]: \u2113adv(x + \u03b4; y, \u03b8) = maximize x\u2032\u2236\u2225x\u2032\u2212x\u2225\u221e\u2264\u03f5 \u2113(x\u2032 + \u03b4; y, \u03b8), (2) where x\u2032 denotes the adversarial input that lies in the \u2113\u221e-norm ball centered at x with radius \u03f5 > 0. Recall from (1) that the conventional VP requests \u03b4 to be universal across training data. Thus, we term universal AVP (U-AVP) the following problem by integrating (1) with (2): minimize \u03b4\u2236\u03b4\u2208C \u03bbE(x,y)\u2208Dtr[\u2113(x + \u03b4; y, \u03b8)]+ E(x,y)\u2208Dtr[\u2113adv(x + \u03b4; y, \u03b8)] (U-AVP) where \u03bb > 0 is a regularization parameter to strike a balance between generalization and adversarial robustness [2]. 0 10 20 50 100 PGD steps 0 20 40 60 80 100 T est Accuracy (%) Vanilla U-AVP Fig. 1: Example of designing U-AVP for adversarial defense on (CIFAR-10, ResNet18), measured by robust accuracy against PGD attacks [1] of different steps. The robust accuracy of 0 steps is the standard accuracy. The problem (U-AVP) can be effectively solved using a standard minmax optimization method, which involves two alternating optimization routines: inner maximization and outer minimization. The former generates adversarial examples as AT, and the latter produces the visual prompt \u03b4 like (1). At test time, the effectiveness of \u03b4 is measured from two aspects: (1) standard accuracy, i.e., the accuracy of \u03b4-integrated benign examples, and (2) robust accuracy, i.e., the accuracy of \u03b4-integrated adversarial examples (against the victim model \u03b8). Despite the succinctness of (U-AVP), Fig. 1 shows its ineffectiveness to defend against adversarial attacks. Compared to the vanilla VP (1), it also suffers a signi\ufb01cant standard accuracy drop (over 50% in Fig. 1 corresponding to 0 PGD attack steps) and robust accuracy is only enhanced by a small margin (around 18% against PGD attacks). The negative results in Fig. 1 are not quite surprising since a data-agnostic input prompt \u03b4 has limited learning capacity to enable adversarial defense. Thus, it is non-trivial to tackle the problem of AVP. 4. CLASS-WISE ADVERSARIAL VISUAL PROMPT No free lunch for class-wise visual prompts. A direct extension of (U-AVP) is to introduce multiple adversarial visual prompts, each of which corresponds to one class in the training set Dtr. If we split Dtr into class-wise training sets {D(i) tr }N i=1 \f(for N classes) and introduce class-wise visual prompts {\u03b4(i)}, then the direct C-AVP extension from (U-AVP) becomes minimize {\u03b4(i)\u2208C}i\u2208[N] 1 N N \u2211 i=1 {\u03bbE(x,y)\u2208D(i) tr [\u2113(x + \u03b4(i); y, \u03b8)]+ E(x,y)\u2208D(i) tr [\u2113adv(x + \u03b4(i); y, \u03b8)]} (C-AVP-v0) where [N] denotes the set of class labels {1,2,...,N}. It is worth noting that C-AVP-v0 is decomposed over class labels. Although the class-wise separability facilitates numerical optimization, it introduces challenges (C1)-(C2) when applying class-wise visual prompts for adversarial defense. \u25cf(C1) Test-time prompt selection: After acquiring the visual prompts {\u03b4(i)} from (C-AVP-v0), it remains unclear how a class-wise prompt should be selected for application to a test-time example xtest. An intuitive way is to use the inference pipeline of \u03b8 by aligning its top-1 prediction with the prompt selection. That is, the selected prompt \u03b4 and the predicted class i\u2217are determined by \u03b4 = \u03b4\u2217, i\u2217= arg max i\u2208[N] fi(xtest + \u03b4(i); \u03b8), (3) where fi(x;\u03b8) denotes the ith-class prediction con\ufb01dence. However, the seemingly correct rule (3) leads to a large prompt selection error (thus poor prediction accuracy) due to (C2). \u25cf(C2) Backdoor effect of class mis-matched prompts: Given \u03b4(i) from (C-AVP-v0), if the test-time example xtest is drawn from class i, the visual prompt \u03b4(i) then helps prediction. However, if xtest is not originated from class i, then \u03b4(i) could serve as a backdoor attack trigger [46] with the targeted backdoor label i for the \u2018prompted input\u2019 xtest+\u03b4(i). Since the backdoor attack is also input-agnostic, the class-discriminative ability of xtest + \u03b4(i) enabled by \u03b4(i) could result in incorrect prediction towards the target class i for xtest. Joint prompts optimization for C-AVP. The failure of CAVP-v0 inspires us to rethink the value of class-wise separability. As illustrated in challenges (C1)-(C2), the compatibility with the test-time prompt selection rule and the interrelationship between class-wise visual prompts should be taken into account. To this end, we develop a series of new AVP principles below. Fig. 2 provides a schematic overview of C-AVP and its comparison with U-AVP and the predictor without VP. Inputs Visual Prompt Pretrained Classifier Predictions Method (c) C-AVP (b) U-AVP (a) No Prompt Dog Cat Dog Cat Dog Cat Fig. 2: Overview of C-AVP over two classes (red and green) vs. U-AVP and the prompt-free learning pipeline. First, to bake the prompt selection rule (3) into C-AVP, we enforce the correct prompt selection, i.e., under the condition that fy(x + \u03b4(y);\u03b8) > maxk\u2236k\u2260y fk(x + \u03b4(k);\u03b8) for (x,y) \u2208 D(y). The above can be cast as a CW-type loss [47]: \u2113C\u2212AVP,1({\u03b4(i)}; Dtr, \u03b8) = E(x,y)\u2208Dtr max{max k\u2260y fk(x + \u03b4(k); \u03b8) \u2212fy(x + \u03b4(y); \u03b8), \u2212\u03c4}, (4) where \u03c4 > 0 is a con\ufb01dence threshold. The rationale behind (4) is that given a data sample (x,y), the minimum value of \u2113C\u2212AVP,1 is achieved at \u2212\u03c4, indicating the desired condition with the con\ufb01dence level \u03c4. Compared with (C-AVP-v0), another key characteristic of \u2113C\u2212AVP,1 is its non-splitting over class-wise prompts {\u03b4(i)}, which bene\ufb01ts the joint optimization of these prompts. Second, to mitigate the backdoor effect of mis-matched prompts, we propose additional two losses, noted by \u2113C\u2212AVP,2 and \u2113C\u2212AVP,3, to penalize the data-prompt mismatches. Specifically, \u2113C\u2212AVP,2 penalizes the backdoor-alike targeted prediction accuracy of a class-wise visual prompt when applied to mismatched training data. For the prompt \u03b4(i), this leads to \u2113C\u2212AVP,2({\u03b4(i)}; Dtr, \u03b8) = 1 N N \u2211 i=1 E(x,y)\u2208D(\u2212i) tr max{fi(x + \u03b4(i); \u03b8) \u2212fy(x + \u03b4(i); \u03b8), \u2212\u03c4}, (5) where D(\u2212i) tr denotes the training data set by excluding D(i) tr . The class i-associated prompt \u03b4(i) should not behave as a backdoor trigger to non-i classes\u2019 data. Likewise, if the prompt is applied to the correct data class, then the prediction con\ufb01dence should surpass that of a mis-matched case. This leads to \u2113C\u2212AVP,3({\u03b4(i)}; Dtr, \u03b8) = E(x,y)\u2208Dtr max{max k\u2260y fy(x + \u03b4(k); \u03b8) \u2212fy(x + \u03b4(y); \u03b8), \u2212\u03c4}. (6) Let \u2113C\u2212AVP,0({\u03b4(i)};Dtr,\u03b8) denote the objective function of (C-AVP-v0). Integrated with \u2113C\u2212AVP,q({\u03b4(i)};Dtr,\u03b8) for q \u2208{1,2,3}, the desired class-wise AVP design is cast as minimize {\u03b4(i)\u2208C}i\u2208[N] \u2113C\u2212AVP,0({\u03b4(i)}; Dtr, \u03b8)+ \u03b3 \u22113 q=1 \u2113C\u2212AVP,q({\u03b4(i)}; Dtr, \u03b8), (C-AVP) where \u03b3 > 0 is a parameter for class-wise prompting penalties. 5. EXPERIMENTS Experiment setup. We conduct experiments on CIFAR-10 with a pretrained ResNet18 of testing accuracy of 94.92% on standard test dataset. We use PGD-10 (i.e., PGD attack with 10 steps [1]) to generate adversarial examples with \u03f5 = 8/255 during visual prompts training, and with a cosine learning rate scheduler starting at 0.1. Throughout experiments, we choose \u03bb = 1 in (U-AVP), and \u03c4 = 0.1 and \u03b3 = 3 in (CAVP). The width of visual prompt is set to 8 (see Fig. 3 for the visualization). C-AVP outperforms conventional VP. Tab. 1 demonstrates the effectiveness of proposed C-AVP approach vs. U-AVP (the \fairplane automobile bird cat dear dog frog horse ship truck Fig. 3: C-AVP visualization. One image is chosen from each CIFAR-10 class with the corresponding C-AVP. Table 1: VP performance comparison in terms of standard (std) accuracy (acc) and robust accuracy against PGD attacks with \u03f5 = 8/255 and multiple PGD steps on (CIFAR-10, ResNet18). Evaluation Std Robust acc vs PGD w/ step # metrics (%) acc 10 20 50 100 Pre-trained 94.92 0 0 0 0 Vanilla VP 94.48 0 0 0 0 U-AVP 27.75 16.9 16.81 16.81 16.7 C-AVP-v0 19.69 13.91 13.63 13.6 13.58 C-AVP (ours) 57.57 34.75 34.62 34.51 33.63 direct extension of VP to adversarial defense) and the C-AVPv0 method in the task of robustify a normally-trained ResNet18 on CIFAR-10. For comparison, we also report the standard accuracy of the pre-trained model and the vanilla VP solution given by (1). As we can see, C-AVP outperforms U-AVP and C-AVP-v0 in both standard accuracy and robust accuracy. We also observe that compared to the pretrained model and the vanilla VP, the robustness-induced VP variants bring in an evident standard accuracy drop as the cost of robustness. Prompting regularization effect in (C-AVP). Tab. 2 shows different settings of prompting regularizations used in C-AVP, where \u2018Si\u2019 represents a certain loss con\ufb01guration. As we can see, the use of \u2113C\u2212AVP,2 contributes most to the performance of learned visual prompts (see S3). This is not surprising, since we design \u2113C\u2212AVP,2 for mitigating the backdoor effect of class-wise prompts, which is the main source of prompting selection error. We also note that \u2113C\u2212AVP,1 is the second most important regularization. This is because such a regularization is accompanied with the prompt selection rule (3). Tab. 2 also indicates that the combination of \u2113C\u2212AVP,1 and \u2113C\u2212AVP,2 is a possible computationally lighter alternative to (C-AVP). Class-wise prediction error analysis. Fig. 4 shows a comparison of the classi\ufb01cation confusion matrix. Each row corresponds to testing samples from one class, and each column corresponds to the prompt (\u2018P\u2019) selection across 10 image classes. As we can see, our proposal outperforms C-AVP-v0 since the former\u2019s higher main diagonal entries indicate less prompt selection error than the latter. Comparisons with other test-time defenses. In Tab. 3, we compare our proposed C-AVP with three test-time defense methods selected from Croce et. al. [14]. Note that all methods are applied to robustifying a \ufb01xed, standardly pre-trained ResNet18. Following Croce et. al. [14], we divide the considered defenses into different categories, relying on their defense principles (i.e., IP or MA) and needed test-time operations (i.e., IA, AN, and R). As we can see, our method C-AVP falls into the IP category but requires no involved test-time operations. This leads to the least inference overhead. Although there (a) C-AVP-v0 (b) C-AVP Fig. 4: The predictions of C-AVP-v0 vs. C-AVP on (CIFAR10, ResNet18). Table 2: Sensitivity analysis of prompting regularization in C-AVP on (CIFAR-10, ResNet18). Setting \u2113C\u2212AVP,1 \u2113C\u2212AVP,2 \u2113C\u2212AVP,3 Std Acc (%) PGD-10 Acc (%) S1 \u0018 \u0018 \u0018 19.69 13.91 S2 \u0014 \u0018 \u0018 22.72 13.01 S3 \u0018 \u0014 \u0018 40.01 25.40 S4 \u0018 \u0018 \u0014 17.44 11.78 S5 \u0014 \u0014 \u0018 57.03 32.39 S6 \u0014 \u0018 \u0014 26.02 15.80 S7 \u0014 \u0014 \u0014 57.57 34.75 exists a performance gap with the test-time defense baselines, we hope that our work could pave a way to study the pros and cons of visual prompting in adversarial robustness. Table 3: Comparison of C-AVP with other SOTA test-time defenses. Per the benchmark in Croce et. al. [14], the involved test-time operations in these defenses include: IP (input puri\ufb01cation), MA (model adaption), IA (iterative algorithm), AN (auxiliary network), and R (randomness). And inference time (IT), standard accuracy (SA), and robust accuracy (RA) against PGD-10 are used as performance metrics. Method IP MA IA AN R IT SA (%) RA (%) [43] \u0014 \u0018 \u0014 \u0018 \u0018 518 \u00d7 85.9% 0.4% [15] \u0014 \u0018 \u0014 \u0014 \u0014 176 \u00d7 91.1% 40.3% [44] \u0018 \u0014 \u0014 \u0014 \u0018 59 \u00d7 56.1% 50.6% C-AVP \u0014 \u0018 \u0018 \u0018 \u0018 1.4 \u00d7 57.6% 34.8% 6."
+                }
+            ],
+            "Xianbin Ye": [
+                {
+                    "url": "http://arxiv.org/abs/2203.00836v2",
+                    "title": "CandidateDrug4Cancer: An Open Molecular Graph Learning Benchmark on Drug Discovery for Cancer",
+                    "abstract": "Anti-cancer drug discoveries have been serendipitous, we sought to present\nthe Open Molecular Graph Learning Benchmark, named CandidateDrug4Cancer, a\nchallenging and realistic benchmark dataset to facilitate scalable, robust, and\nreproducible graph machine learning research for anti-cancer drug discovery.\nCandidateDrug4Cancer dataset encompasses multiple most-mentioned 29 targets for\ncancer, covering 54869 cancer-related drug molecules which are ranged from\npre-clinical, clinical and FDA-approved. Besides building the datasets, we also\nperform benchmark experiments with effective Drug Target Interaction (DTI)\nprediction baselines using descriptors and expressive graph neural networks.\nExperimental results suggest that CandidateDrug4Cancer presents significant\nchallenges for learning molecular graphs and targets in practical application,\nindicating opportunities for future researches on developing candidate drugs\nfor treating cancers.",
+                    "authors": "Xianbin Ye, Ziliang Li, Fei Ma, Zongbi Yi, Pengyong Li, Jun Wang, Peng Gao, Yixuan Qiao, Guotong Xie",
+                    "published": "2022-03-02",
+                    "updated": "2022-08-22",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "q-bio.BM"
+                    ],
+                    "main_content": "INTRODUCTION Cancer is one of the leading causes of mortality worldwide. Opportunities to help reduce the death rate from cancer through the discovery of new drugs are benefiting from the increasing advances in technology and enhanced knowledge of human neoplastic disease[23, 33]. However, drug discovery is a time-consuming, labor intensive, and expensive process, far slower than expected. The entire process from discovery to the regulatory approval of a new drug can take as much as 12 years and cost estimated at US 3 billion [7]. It has been usually hampered by experimental discovery of molecules and targets, and following with validation with in vitro experiments on cell lines and animals before moving to clinical testing [12]. Furthermore, stagnant success rate ( 1:5000) is associated with each drug development stage. Fortunately, many researchers have proposed various effective computer-aided drug discovery (CADD) methods [24] to decrease the costs and speed up projects [14]. As an feasible assistant technique, CADD has made a significant contribution to drug discovery and has successfully developed dozens of drugs to market in recent decades[32]. Generally, CADD can be categorized into receptor-based methods and ligand-based methods upon the availability of target proteins and molecules. Receptor-based CADD relies on the target protein structure to calculate interaction for compounds, while ligand-based CADD exploits the information of compounds with diverse structures and known activity and inactivity, for construction of predictive, quantitative structure-activity relation (QSAR)[4]. 29 54869 73770 Targets Compounds Pairs of Interaction Numbers Figure 2: CandidateDrug4Cancer dataset encompasses multiple most-mentioned 29 targets for cancer, with 73770 drugtarget pairs, covering a diverse range of 54869 related drug molecules which is ranged from pre-clinical, clinical and FDA-approved. At present, the interdisciplinary studies between advanced Artificial Intelligence (AI) and drug discovery have received increasing attention due to the superior speed and performance. Many AI technologies have been successfully applied in a variety of tasks for CADD, such as receptor-based Drug Target Interaction (DTI) prediction [1, 8]. Due to the large size of the chemical space, one of the fundamental challenges for these studies is how to learn expressive representations from molecular structures [31]. In the early years, molecular representations are based on hand-crafted features such as molecular descriptors or fingerprints [30]. In contrast, there has been a surge of interests in end-to-end molecular representation learned by graph neural networks. However, challenges for deep learning in molecular representation mainly arise from the scarcity of labeled data, resulting in models that lack practicability [13, 22]. In particular, so far there is limited published information on molecular graph learning for cancer to our knowledge. Altogether, these issues make it difficult producing favorable benchmark assembling proper information of cancer targets and potential molecules, for assessing and comparing the performance of different methods on drug discovery for cancer. To address these issues, we introduced an open molecular graph learning benchmark on drug discovery for cancer, named CandidateDrug4Cancer. Our main contributions can be summarized as follows: \u2022 We introduce a novel graph learning task which has the potential to help understand the performance and limitations of graph representation facilitated models on candidate anticancer drug discovery problems. \u2022 We provide CandidateDrug4Cancer benchmark dataset, which encompasses multiple most-mentioned 29 targets for cancer, covering 54869 cancer-related drug molecules which are ranged from pre-clinical, clinical and FDA-approved. \u2022 We conduct benchmark evaluations for drug target interaction with baseline encoders including powerful graph neural network for drug-like molecules. 2 RELATED WORKS Drug Target Interaction One of the initial steps of drug discovery is the identification of novel drug-like compounds that interact with the predefined target proteins. Various deep learning methods have been developed and achieved excellent performance for drug-target interaction (DTI) prediction [5, 20, 27]. Generally, the deep learning algorithms for DTI prediction comprise of a compound encoder and a protein encoder. Recently, Tsubaki et al. [25] proposed a new DTI framework by combining GNN for compounds and a CNN for proteins, which significantly outperformed existing methods. Molecular Graph Learning Recently, among the promising deep learning architectures, graph neural network (GNN), such as message passing neural network (MPNN) [11] has gradually emerged as a powerful candidate for modeling molecular data. Because a molecule is naturally a graph that consists of atoms (nodes) connected through chemical bonds (edges), it is ideally suited for GNN. Up to now, various GNN architectures have achieved great progress in drug discovery [19, 29]. However, there are some limits that need to be addressed. Challenges for deep learning in molecular representation mainly arise from the scarcity of labeled data, as lab experiments are expensive and time-consuming. Thus, training datasets in drug discovery are usually limited in size, and GNNs tend to overfit them, resulting in learned representations that lack practicability [13, 17, 22]. Anti-Cancer Candidate Drug Datasets Unfortunately, the current number of drugs (FDA approved or at the experimental stage) is only around 10,000, the current knowledge about the drug\u2013target space is limited, especially for severe life-threatening cancers, and novel approaches are required to widen our knowledge[21]. However, most cancer drug discoveries have been serendipitous, the extent to which non-cancer drugs have potential as future cancer \fCandidateDrug4Cancer: An Open Molecular Graph Learning Benchmark on Drug Discovery for Cancer GLB 2021, The Web Conference 2021 therapeutics is unknown[6]. Recent efforts have demonstrated the power of cancer cell line screening\u2014testing either many compounds across a limited number of cell lines (for example, NCI-60 [2], Genomics of Drug Sensitivity in Cancer (GDSC) [9], the Cancer Target Discovery and Development (CTD2)[3] and DepMap[6]. The ideal study would involve screening larger amount of drug candidates (most of which are non-oncology drugs) across most target proteins to capture the molecular diversity of human cancer. Table 1: The most frequent target proteins of cancers collected from DepMap[6], DrugBank[28] and ChEMBL[10]. Target_id Target_name chembl1824 Receptor protein-tyrosine kinase erbB-2 chembl1865 Histone deacetylase 6 chembl1871 Androgen Receptor chembl1913 Platelet-derived growth factor receptor beta chembl1936 Stem cell growth factor receptor chembl1937 Histone deacetylase 2 chembl1974 Tyrosine-protein kinase receptor FLT3 chembl203 Epidermal growth factor receptor erbB1 chembl2041 Tyrosine-protein kinase receptor RET chembl2185 Serine/threonine-protein kinase Aurora-B chembl262 Glycogen synthase kinase-3 beta chembl267 Glycogen synthase kinase-3 beta chembl279 Vascular endothelial growth factor receptor 2 chembl2842 Serine/threonine-protein kinase mTOR chembl2971 Tyrosine-protein kinase JAK2 chembl301 Cyclin-dependent kinase 2 chembl308 Cyclin-dependent kinase 1 chembl3130 PI3-kinase p110-delta subunit chembl3145 PI3-kinase p110-beta subunit chembl325 Histone deacetylase 1 chembl3267 PI3-kinase p110-gamma subunit chembl332 Matrix metalloproteinase-1 chembl333 Matrix metalloproteinase-2 chembl340 Cytochrome P450 3A4 chembl3717 Hepatocyte growth factor receptor chembl4005 PI3-kinase p110-alpha subunit chembl4282 Serine/threonine-protein kinase AKT chembl4630 Serine/threonine-protein kinase Chk1 chembl4722 Serine/threonine-protein kinase Aurora-A 3 CANDIDATEDRUG4CANCER DATASETS Inspired by DepMap [6] and DrugBank [28], we sought to create a public resource containing the candidate compounds for the most-used targets for cancers, as shown in Table 1. Firstly, the most commonly researched and cancer-related targets are collected based on DepMap and DrugBank, subsequently, the corresponding compounds and drug-target pairs are collected to establish CandidateDrug4Cancer from the ChEMBL database [10] (https://www.ebi.ac.uk/chembl/g/, ChEMBL is a manually curated database of bioactive molecules with drug-like properties. It brings together chemical, bioactivity and genomic data to aid the translation of genomic information into effective new drugs). In addition to the published IC50 measurement, ChEMBL have added an additional field called pChEMBL to the activities table. This value represent comparable measures of concentrations to reach half-maximal response transformed to a negative logarithmic scale. pChEMBL is defined as: \ud835\udc5d\ud835\udc36\u210e\ud835\udc38\ud835\udc40\ud835\udc35\ud835\udc3f= \u2212lg(\ud835\udc5a\ud835\udc5c\ud835\udc59\ud835\udc4e\ud835\udc5f\ud835\udc3c\ud835\udc3650(\ud835\udc40)) (1) For example, an IC50 measurement of 10 uM would have a pChEMBL value of 5. However, in order to efficiently verify anticancer inhibitors and reduce the cost of a large number of subsequent experiments, we set decision boundary at pChEMBL is 7(approximately 100 nM) more strictly, defining pChEMBL larger than or equal to 7 as positive interaction [16] (see Figure 3). Table 2: The features used in molecular graph. These features are obtained by RDKit [15]. Type Name Description Node feature Atom type Atomic number (0-122) Formal charge [unk,-5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5] Chirality type [unk, unspecifie, tetrahedral-CW, tetrahedralL-CCW, other] Hybridization [unk, sp, sp2, sp3, sp3d, sp3d2, unspecified] NumH Number of connected hydrogens[unk,0, 1, 2, 3, 4, 5, 6, 7, 8] Implicit valence [unk, 0, 1, 2, 3, 4, 5, 6] Degree Number of covalent bonds [unk, 0, 1, 2, 3, 4, 5,6,7,8,9,10] Aromatic Whether the atom is part of an aromatic system [0,1] Edge feature Bond direction [None,endupright, enddownright] Bond type [Single, double, triple, aromatic] Conjugation Whether the bond is conjugated [0,1] Ring Whether the bond is in Ring [0,1] Stereo [StereoNone, StereoAny, StereoZ, StereoE] 4 EXPERIMENTAL SETUP Baseline models. Various deep learning methods have been developed and achieved promising performance for DTI prediction [5, 20, 25, 27].Generally, the common DTI prediction models comprise of a compound encoder and a protein encoder. We have compared baselines using different combinations of compound and protein encoders including: MPNN-CNN, Daylight-AAC, Morgan-AAC and MolGNet-CNN (For compounds: Morgan fingerprints, Daylighttype fingerprints, MPNN on molecular graph. For proteins: Amino Acid Composition up to 3-mers, Convolutional Neural Network (CNN) on target sequence). In particular, we further adopt our pretrained GNN named MolGNet [18], to evaluate the effectiveness of pre-trained GNN on DTI prediction for caner targets. Implementation details. We conduct leave-one-target-out evaluation, the drug-target interaction pairs of the rest 28 targets are used for training, and the performances are calculated on each unseen target proteins. So as to evaluate whether models are capable of prediction on unseen targets. Due to hardware limitations and multiple baseline comparisons, we evaluate the baselines on 20% of total datasets (the sub-datasets are also provided in our link). Each model is trained for 50 epochs. For graph learning, all drug molecules are pre-processed into hydrogen-depleted molecular graphs with nodes features, edge features, and adjacency matrix with RDKit [15]. The detailed information about nodes features and edge features can be referred to Table 2. 5 RESULTS AND"
+                }
+            ],
+            "Bingzhe Wu": [
+                {
+                    "url": "http://arxiv.org/abs/2205.10014v2",
+                    "title": "A Survey of Trustworthy Graph Learning: Reliability, Explainability, and Privacy Protection",
+                    "abstract": "Deep graph learning has achieved remarkable progresses in both business and\nscientific areas ranging from finance and e-commerce, to drug and advanced\nmaterial discovery. Despite these progresses, how to ensure various deep graph\nlearning algorithms behave in a socially responsible manner and meet regulatory\ncompliance requirements becomes an emerging problem, especially in\nrisk-sensitive domains. Trustworthy graph learning (TwGL) aims to solve the\nabove problems from a technical viewpoint. In contrast to conventional graph\nlearning research which mainly cares about model performance, TwGL considers\nvarious reliability and safety aspects of the graph learning framework\nincluding but not limited to robustness, explainability, and privacy. In this\nsurvey, we provide a comprehensive review of recent leading approaches in the\nTwGL field from three dimensions, namely, reliability, explainability, and\nprivacy protection. We give a general categorization for existing work and\nreview typical work for each category. To give further insights for TwGL\nresearch, we provide a unified view to inspect previous works and build the\nconnection between them. We also point out some important open problems\nremaining to be solved in the future developments of TwGL.",
+                    "authors": "Bingzhe Wu, Jintang Li, Junchi Yu, Yatao Bian, Hengtong Zhang, CHaochao Chen, Chengbin Hou, Guoji Fu, Liang Chen, Tingyang Xu, Yu Rong, Xiaolin Zheng, Junzhou Huang, Ran He, Baoyuan Wu, GUangyu Sun, Peng Cui, Zibin Zheng, Zhe Liu, Peilin Zhao",
+                    "published": "2022-05-20",
+                    "updated": "2022-05-23",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.AI"
+                    ],
+                    "main_content": "Introduction Recently, Deep Graph Learning (DGL) based on Graph Neural Networks (GNNs) have emerged as a powerful learning paradigm for graph representation learning. In the past few years, DGL is becoming an active frontier of deep learning with an exponential growth of research. With advantages in modeling graph-structured data, DGL has achieved remarkable progress in many important areas, ranging from \ufb01nance (e.g., fraud detection and credit modeling) [28, 144, 9, 67], e-commerce (e.g., recommendation system) [126, 85], drug discovery and advanced material discovery [41, 134, 100, 82, 83]. Although considerable progress has been made in the \ufb01eld of deep graph learning, there are many challenges about the level of trustworthiness associated with GNNs in many real-world applications, especially in risk-sensitive and and securitycritical scenarios such as \ufb01nance and bioinformatics [100]. Such challenges are real and call for sophisticated solutions, as many untrustworthy sides of DGL models have been exposed through adversarial attacks [19, 13], inherent bias or unfairness [157], and lack of explainability [152] and privacy protection [16] in the current rapidly evolving works. In light of these challenges, various emerging laws and regulations in arti\ufb01cial intelligence (AI) ethic, e.g., EU General Data Protection Regulation (GDPR) [91], promote different companies to develop trustworthy algorithms for meeting regulatory compliance requirements of privacy, explainability and fairness, etc. A new research trend in recent years has been to investigate ways to to build trustworthy algorithms and ensure they behave in a socially responsible manner. Trustworthy Graph Learning (TwGL), an emerging \ufb01eld with the aim to solve the trustworthiness issues in graph data from a technical viewpoint, has become a focus of attention in the areas of deep graph learning. In contrast to conventional graph learning research which mainly focus on how to develop accurate models, TwGL considers various reliability and safety aspects of the graph learning problem beyond model performance. In Figure 1, we summarise the four key dimensions 2 \ftowards reliable TwGL, that is, accuracy, reliability, explainability, and privacy protection. We give detailed descriptions as follows: 1) Accuracy or utility: The prediction accuracy is the basic ability of a trustworthy model. Trustworthy GNNs are expected to generate accurate output, consistent with the ground truth, as much as possible; 2) Reliability: Trustworthy GNNs should be resilient and secure. In other words, they must be robust against different potential threats, such as inherent noise, distribution shift, and adversarial attacks; 3) Explainability. Understanding the graph learning model is another important aspect in developing trustworthy GNNs. The model itself must allow explainable for the prediction, which can help humans to enhance understanding, make decisions and take further actions; 4) Privacy protection: Trustworthy GNNs are required to ensure full privacy of the models as well as data privacy. Trustworthy Graph Learning Accuracy Reliability Explainability Privacy Protection Figure 1: Four key dimensions of trustworthy graph learning. Motivation and Contributions. The trustworthiness of DGL models is an inescapable problem when they step into real-world applications, posing various open, unexplored, and unidenti\ufb01ed challenges in both academic and industrial communities. Therefore, various aspects of DGL models are required to be carefully considered to achieve trustworthiness. To facilitate the research in the \ufb01eld of TwGL, there is a pressing demand for a comprehensive survey reviewing current work from a uni\ufb01ed view. In this survey, we provide a comprehensive overview of existing achievements in the TwGL \ufb01eld from four key dimensions, namely, robustness, explainability and privacy protection. For each part, we give a general categorization for existing work and review typical work for each category. To give further insights for TwGL research, we provide a uni\ufb01ed view to inspect previous work and build the connections between them. We also point out some important open problems remaining to be solved in the future developments of TwGL. Overview. In this paper, we puts together the state-of-the-art and most important works for TwGL. The organization of this paper is structured as follows: In Section 2, we comprehensively review the recent advances of reliable DGL against three threats, including inherent noise, distribution shift, and adversarial attack. In Section 3, we provide a detailed technique overview of the state-of-the-art GNN explainability methods. In Section 4, we also discuss the research on the privacy protection of DGL models. For each section, we summarize several challenges 3 \fand opportunities for future research. Finally, we conclude this paper by providing insights for further achievements of trustworthy graph learning in Section 5. Relation to other surveys. To our best knowledge, little survey have been comprehensively conducted especially on the methods and applications of TwGL. There are several surveys related to our paper, most of which focus on a single aspect of TwGL, e.g., reliability [110, 19, 60, 137, 169, 161] and explainability [152, 70]. Comparatively to the aforementioned surveys, we present a systematic and comprehensive review of the current advances, trends, and applications in trustworthy graph learning, which would potentially bene\ufb01t the research community in developing practical trustworthy models. Basic de\ufb01nitions. Oracle graph data G = (A, X) with its adjacency matrix A and node features X; Y the labels of graphs/nodes; \u03f5a, \u03f5x, \u03f5y denote structure, attribute, and label noises respectively; Dtrain and Dtest are the training and test dataset, respectively; L denotes the loss function and \u02c6 G is the perturbed graph. 2 Reliability 2.1 Overview (a)Inherent Noise Class blue Class red Data  Sampling Deep Graph Learning Adversarial Sample Model Prediction (c) Adversarial  Attack max \u0302 \ud835\udca2\ud835\udca2\u2248\ud835\udca2\ud835\udca2\u2112( \u0302 \ud835\udca2\ud835\udca2, \ud835\udc4c\ud835\udc4c) \u0302 \ud835\udca2\ud835\udca2 \ud835\udc43\ud835\udc43train \ud835\udca2\ud835\udca2, \ud835\udc4c\ud835\udc4c \u2260\ud835\udc43\ud835\udc43test (\ud835\udca2\ud835\udca2, \ud835\udc4c\ud835\udc4c) \ud835\udc9f\ud835\udc9ftrain \ud835\udca2\ud835\udca2 \u0302 \ud835\udca2\ud835\udca2 \u2755 \u2755Label noise (b) Distribution Shift Data  Annotation Misclassified Data Generation Figure 2: An illustrative example on deep graph learning against (a) inherent noise, (b) distribution shift, and (c) adversarial attack. From left to right: the inherent noise (i.e., label noise) is inevitably introduced during data annotation, where a red \u201ccircle\u201d node is mislabeled with the \u201ccross\u201d class (marked with red color). Then, the sampling bias leads to the discrepancy between training (color-\ufb01lled nodes) and testing datasets (color-un\ufb01lled nodes), thus introducing any possible kind of distribution shift. After the data generation process, adversarial attacks are performed on the graph to mislead the model prediction (e.g., misclassi\ufb01cation on a node). Recent few years have seen deep graph learning (DGL) based on graph neural networks (GNNs) making remarkable progress in a variety of important areas, ranging from business scenarios such as \ufb01nance (e.g., fraud detection and credit modeling) [28, 144, 9, 67], e-commerce (e.g., recommendation system) [126, 85], drug discovery and advanced material discovery [41, 134, 100, 82, 83]. Despite the progress, applying various DGL algorithms to real-world applications faces a 4 \fTable 1: Summary of recent advances in reliable deep graph learning. Threats Description Selected"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2204.07742v1",
+                    "title": "DRFLM: Distributionally Robust Federated Learning with Inter-client Noise via Local Mixup",
+                    "abstract": "Recently, federated learning has emerged as a promising approach for training\na global model using data from multiple organizations without leaking their raw\ndata. Nevertheless, directly applying federated learning to real-world tasks\nfaces two challenges: (1) heterogeneity in the data among different\norganizations; and (2) data noises inside individual organizations.\n  In this paper, we propose a general framework to solve the above two\nchallenges simultaneously. Specifically, we propose using distributionally\nrobust optimization to mitigate the negative effects caused by data\nheterogeneity paradigm to sample clients based on a learnable distribution at\neach iteration. Additionally, we observe that this optimization paradigm is\neasily affected by data noises inside local clients, which has a significant\nperformance degradation in terms of global model prediction accuracy. To solve\nthis problem, we propose to incorporate mixup techniques into the local\ntraining process of federated learning. We further provide comprehensive\ntheoretical analysis including robustness analysis, convergence analysis, and\ngeneralization ability. Furthermore, we conduct empirical studies across\ndifferent drug discovery tasks, such as ADMET property prediction and\ndrug-target affinity prediction.",
+                    "authors": "Bingzhe Wu, Zhipeng Liang, Yuxuan Han, Yatao Bian, Peilin Zhao, Junzhou Huang",
+                    "published": "2022-04-16",
+                    "updated": "2022-04-16",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.DC",
+                        "stat.ML"
+                    ],
+                    "main_content": "Introduction Recently, Federated learning (FL) is emerging as a promising way to mitigate the information silo problem, in which data are spread across different organizations and cannot be exchanged due to privacy concerns and regulations. (McMahan et al., 2017; Yang et al., 2019). There are plentiful attempts to apply FL to various real-world applications in1Tencent AI Lab 2Hong Kong University of Science and Technology. Correspondence to: Bingzhe Wu <bingzhewu@tencent.com>. cluding recommender system (Chen et al., 2020a;b), medical image analysis (Sheller et al., 2020) and drug discovery (Chen et al., 2021; Xiong et al., 2020). For example, Chen et al. (Chen et al., 2020a) combine the FL method and secure multi-party computation for building recommender systems among different organizations. He et al. (He et al., 2021) proposes a federated GNN learning framework for various molecular property prediction tasks. Despite the above progress, directly applying FL methods to real-world scenarios still faces severe challenges. In this paper, we focus on two critical challenges: (1) Inter-client data heterogeneity (2) Intra-client data noise. Speci\ufb01cally, interclient data heterogeneity refers to data from different clients that always have varying distributions. This is also called the Non-IID issue in the FL research community (Karimireddy et al., 2020; Li et al., 2020a). For example, different pharmaceutical companies/labs may use different high-throughput screening environments to measure the protein-ligand binding af\ufb01nity, resulting in different DTI assays (Bento et al., 2014; Mendez et al., 2019). The intra-client data noise refers to the noise created during the data collection process within each client individually. A common noise source in drug discovery is the label noise caused by measurement noise or the situation where very low/high-af\ufb01nity values are often encoded as a boundary constant, see (Mendez et al., 2019) for a detailed description for the various noise in the largest bioassay deposition website. Numerous efforts have been made to solve the \ufb01rst issue (Mohri et al., 2019; Reisizadeh et al., 2020; Li et al., 2020a; Karimireddy et al., 2020; Li et al., 2021; Xie et al., 2021) from different aspects. A recent trend tries an essential approach to overcome the client heterogeneity by incorporating distributionally robust optimization (DRO) (Duchi et al., 2011) into conventional FL paradigm (Mohri et al., 2019; Deng et al., 2021). Recently, DRFA (Deng et al., 2021) proposes to optimize the distributionally robust empirical loss, which combines different local loss functions with learnable weights. This method theoretically ensures that the learned model can perform well over the worst-case combination of local data distributions. Previous work using DRO has shown effectiveness on some academic benchmarks such as Fashion MNIST and UCI arXiv:2204.07742v1  [cs.LG]  16 Apr 2022 \fDRFLM: Distributionally Robust Federated Learning with Inter-client Noise via Local Mixup Adults (Deng et al., 2021; Mohri et al., 2019). However, when these methods are applied to real-world tasks such as drug discovery, we observe poor performance and training instability. Intuitively, prior work (Deng et al., 2021; Mohri et al., 2019) aims to pessimistically optimize the worst-case combination of local distribution which is typically dominated by the local training distribution with more noises. Therefore they suffer serious performance drop issue as shown in Figure 1. From this \ufb01gure, the use of distributionally robust optimization cannot improve the worst-case local RMSE but even signi\ufb01cantly reduces the performance. According to further empirical studies (see details in Section 4), we identify that Intra-client data noise inside the local client data is the main cause of the above issues. Similar issues have also been discussed in the centralized setting (Zhai et al., 2021), which points out that DRO is sensitive to outliers which further hinder model performance. In this paper, we focus on mitigating the effects caused by intra-client data noise in the FL setting. In order to address the above two critical challenges, we present a general framework named DRFLM for building robust neural network models in the FL setting for a wide range of tasks. To overcome the inter-client data heterogeneity issue, we propose to employ DRO techniques to optimize the global model. As discussed above, such a DRO method can be easily affected by intra-client data noise (Zhai et al., 2021). As suggested by previous work (Zhai et al., 2021), a natural method to reduce the noise effect in the centralized setting is to drop samples with large losses during training. However, this approach may result in samples that are very valuable to overall training being discarded, which then leads to sub-optimal results. Hence, we propose to integrate the mixup strategy into the local training process and \ufb01nd that it can signi\ufb01cantly improve the empirical performance. Furthermore, this paper offers an interesting theoretical \ufb01nding that mixup can provide heterogeneity-dependent generalization bounds, as shown in Theorem 1. In contrast to most FL-related work which mainly focuses on convergence analysis, our theoretical analysis examines more aspects of the proposed FL approach. Speci\ufb01cally, our analysis consists of three parts: (1) We provide a concrete example to demonstrate the effectiveness of the mixup strategy for mitigating noise effects on conventional DRO FL methods. (2) We analyze the generalization bounds of the learned model, which is the primary theoretical contribution of this paper. (3) We provide rigorous convergence analysis of the training process. Based on theoretical analysis, we conduct extensive empirical studies covering a wide range of drug discovery tasks including molecular property prediction and drug-target interaction prediction. Extensive experimental results show the effectiveness of DRFLM in reducing the effects of client distribution shift and local data noise. 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Noise Level 0.9 1.0 1.1 1.2 1.3 1.4 1.5 1.6 Average RMSE DRFA (average) FedAvg (average) DRFA (worst) FedAvg (worst) 0.9 1.0 1.1 1.2 1.3 1.4 1.5 1.6 Worst RMSE Figure 1. Average and worst-case client RMSE of FedAvg (McMahan et al., 2017) and DRFA (Deng et al., 2021) on a molecular property prediction task. 2. Framework of DRFLM 2.1. Notation We start by \ufb01xing some notations that will be used throughout this paper. we denote [N] = {1, 2, \u00b7 \u00b7 \u00b7 , N}. We consider federated learning setting where N clients jointly train a central model with the i-th client\u2019s local training datasets of sample size Ni. The local dataset is denoted as Di = {zi,j}Ni j=1 and each sample zi,j = (xi,j, yi,j) is i.i.d. drawn from an unknown local joint distribution P(i) x,y. We denote Z\u2020 for any matrix Z as the pseudo inverse of it. 2.2. Optimization Objective A common used approach in Federated learning (FL) to collaboratively train a central model w is to optimize the following objective: min w\u2208W F(w) := N X i=1 Ni N fi(w) (1) where fi(w) = 1 Ni P zi,j\u2208Di f(w, zi,j) is the local empirical risk for client i with the global model. The optimization of Equation 1 suffers from non-iid data distribution, i.e., the heterogeneity across different clients\u2019 data collection (Deng et al., 2021). To provide an intuition, note the target objective in Equation 1 aggregates client-speci\ufb01c loss functions with weights proportional to their local dataset sizes following the standard empirical risk minimization framework. Imbalance sample sizes of different clients or diversity among local data distributions may inevitably plague the generalization ability of the central model obtained by solving Equation 1. To provide the performance guarantee over the worst-case local distribution, Deng et al. (Deng et al., 2021) propose a framework named DRFA to minimize the worst-case combination loss of local empirical distributions: min w\u2208W max \u03bb\u2208\u039b F(w, \u03bb) = N X i=1 \u03bbifi(w) (2) where \u03bb \u2208\u039b := {\u03bb \u2208RN +, PN i=1 \u03bbi = 1} is the learnable \fDRFLM: Distributionally Robust Federated Learning with Inter-client Noise via Local Mixup sampling weights for each local client and fi(\u00b7) is the local empirical distribution. However, Equation 2 tends to pessimistically optimize the worst-case local distribution which can be easily affected by inter-client data noise. As shown in Figure 1, when we apply these methods to real-world drug discovery datasets, the use of distributionally robust optimization cannot improve the worst-case local RMSE but even signi\ufb01cantly reduces the performance. To mitigate the noise effect on the optimization procedure, instead of directly optimizing the worst-case local empirical distribution, we propose to incorporate the popular data augmentation technique, mixup, into the local training procedure as follows: min w\u2208W max \u03bb\u2208\u039b F(w, \u03bb) := N X i=1 \u03bbi \u02dc fi(w). (3) Here we optimize the worst-case local empirical loss evaluated on the mixup distribution \u02dc fi. To be speci\ufb01c, for the i-th client\u2019s local dataset, we denote \u02dc x(i) j,k(\u03b3) = \u03b3g(xi,j) + (1 \u2212 \u03b3)g(xi,k) and \u02dc y(i) j,k(\u03b3) = \u03b3yi,j +(1\u2212\u03b3)yi,k with \u03b3 \u2208[0, 1] be the linear interpolation between two samples in the local dataset Di and g : Rd \u2192Rd is a mapping to improve the \ufb02exibility of our framework to handle different kinds of data. We choose g(\u00b7) as an identity operator for the tableau or image data. For graph prediction tasks including drug discovery, we choose g(\u00b7) as a graph neural network to map a graph (V, E) to a global graph embedding in Rd. Let \u03b3 follows the distribution as Beta(\u03b1, \u03b2), \u02dc z(i) j,k(\u03b3) = (\u02dc x(i) j,k(\u03b3), \u02dc y(i) j,k(\u03b3)) and \u02dc Di(\u03b3) = {\u02dc z(i) j,k(\u03b3), j, k \u2208[Ni]} be the local mixup dataset by linear interpolating between every two samples. After the construction of the mix-up empirical distribution, we evaluate each local loss function on the local mixup dataset, i.e., \u02dc fi(w) = 1 N 2 i P \u02dc z(i) j,k\u2208\u02dc Di(\u03b3) fi(w, \u02dc z(i) j,k) and then combine them with learnable sampling weights \u03bbi as shown in Equation 3. 2.3. Optimization Procedure We adopt a similar approach in (Deng et al., 2021) to solve the min-max optimization problem in Equation 3. The core idea is to alternatively update w and \u03bb combining with sample techniques to reduce communication cost. We brie\ufb02y introduce the main steps in Algorithm 1 (more details can be found in the prior work (Deng et al., 2021)). At the beginning of the s-th communication stage, the optimization procedure follows as: 1. Sampling: the server independently selects two subsets of [N], Cs and Us, both with size of a pre-de\ufb01ne parameter m and also samples a time index t \u2032 s \u223cU[0, \u03c4] with a pre-de\ufb01ne parameter \u03c4 for updating \u03bb. The selection of Cs follows the probability \u03bbt \u2208RN with replacement. 2. Broadcasting: the server broadcasts the central model to the selected clients. Then the client constructs a mixup dataset as discussed above and runs stochastic gradient descent (SGD) algorithm on the mixup dataset to update their local models. After that clients return their local parameter updates after time t \u2032 s (used for updating \u03bb) and the last time \u03c4 (used for updating the global model). 3. Updating \u00af ws+1: the server takes an average over the received latest model parameters. To be speci\ufb01c, \u00af ws+1 = 1 m X j\u2208Cs w(j) s,\u03c4 4. Updating \u03bbs+1: the server takes an average \u02dc ws+1 = 1 m P j\u2208Cs w(j) s,t\u2032 s and then broadcasts it to all clients in Us and receives the stochastic gradients evaluated at this point to construct an auxiliary vector vs. Then the server updates \u03bb as following with some pre-de\ufb01ned learning rate \u03b7 : \u03bbs+1 = Y \u03bb\u2208\u039b (\u03bb(s) + \u03b7\u03c4vs). 3. Theoretical Analysis In this part, we provide different theoretical views to understand our approach in Algorithm 1. 3.1. Theoretical Motivation We \ufb01rst show a concrete example to demonstrate the effectiveness of our method in mitigating the local noise effect. Speci\ufb01cally, we construct an example as follows: Example 1. We consider FL setting with two clients. These two clients jointly train a linear classi\ufb01er: \u03c6(x; w) = sgn(wT x) over the global dataset (x, y) generated by the joint distribution P(x, y) as following: x \u223cPX := Uniform \u001a [\u22122, \u22121]d [ \u0000[1, 2] \u00d7 [\u22122, \u22121]d\u22121\u0001\u001b y = \u001a 1, x1 > 0 \u22121, x1 \u22640 . Given the above global joint distribution, the local distribution of each client is de\ufb01ned as following: \u2022 The \u201dclean\u201d dataset D1 = {(x1,i, y1,i)}N1 i=1 i.i.d. \u223c P(x, y) of the \ufb01rst client is noiseless. \u2022 The \u201dnoisy\u201d dataset \u02dc D2 = {(x2,i, \u02dc y2,i)}N2 i=1 i.i.d. \u223c \u02dc P(x, \u02dc y) (\u02dc y is a noisy version of y de\ufb01ned as follow): \u02dc y2,i = \u001a y2,i, x1 / \u2208I Flipping(y2,i; p1), x1 \u2208I \fDRFLM: Distributionally Robust Federated Learning with Inter-client Noise via Local Mixup Algorithm 1 Dristributionally Robust Federated Learning with Local Mixup (DRFLM) parameters: synchronization gap \u03c4, total communication round T, S = T/\u03c4, sampling size m, mix-up ratio \u03b3, initial model \u00af w0 and initial weights \u03bb0 function CLIENTUPDATE(i, \u00af ws, t \u2032 s) Initialize local weights wi s,0 = \u00af ws Construct \u02dc Di(\u03b3) = {\u03b3g(z1)+(1\u2212\u03b3)g(z2)|z1, z2 \u2208Di} for t = 1 to \u03c4 do Uniformly sample \u02dc z(i) t from \u02dc Di(\u03b3) w(i) s,t = Q W \u0010 w(i) s,t\u22121 \u2212\u03b7\u2207f(w(i) s,t\u22121; \u02dc z(i) t ) \u0011 end for return (w(j) s,\u03c4, w(j) s,t\u2032 s) Server executes: initialize \u00af w0 for s = 1 to S do Sample a set Cs \u2282[N] with size of m and probability \u03bbs with replacement Uniformly sample t \u2032 s from [\u03c4] uniformly Uniformly sample a set Us \u2282[N] with size of m for client j \u2208Cs in parallel do (w(j) s,\u03c4, w(j) s,t\u2032 s) = CLIENTUPDATE(j, \u00af ws, t \u2032 s) end for \u00af ws+1 = 1 m P j\u2208Cs w(j) s,\u03c4 \u02dc ws+1 = 1 m P j\u2208Cs w(j) s,t\u2032 s Broadcasts \u02dc ws+1 to all clients i \u2208Us and receive fi( \u02dc ws+1; \u02dc z(i) t\u2032 s ) from local client Construct v \u2208RN with vi = N mfi( \u02dc ws+1; \u02dc z(i) t\u2032 s ) for each indice i \u2208Us and vi = 0 for i / \u2208Us Updates \u03bbs+1 = Q \u039b(\u03bbs + \u03b7\u03c4vs) end for Output \u00af wS, \u03bbS with I = [\u22121 \u2212p2, \u22121], 1/2 < p1 < 1, 0 < p2 < 1 and Flipping(y; p1) the random function that \ufb02ips y to \u2212y with probability p1. We construct \u02dc D2 as a noisy version of D2 to simulate the local noise effect on the training procedure. Suppose N1 = N2 = \u02dc N, then given the 0-1 loss f(w; x, y) := 1{y \u0338= \u03c6(x; w)} and the learned models wAvg \u02dc N , wDRFA \u02dc N , wDRFLM \u02dc N obtained by minimizing the empirical FedAvg loss (Equation 1), DRFA loss (Equation 2) and DRFLMloss (Equation 3) associated with \u2113. We then have 1 lim \u02dc N\u2192\u221e EP (x,y)[f(wDRFA \u02dc N , x, y)] > lim \u02dc N\u2192\u221e EP (x,y)[f(wAvg \u02dc N ; x, y)] = lim \u02dc N\u2192\u221e EP (x,y)[f(wDRFLM \u02dc N ; x, y)] = 0. As shown by the above inequality, under certain conditions, with the increase of training sample number, losses of FedAvg and our method can gradually decrease to zero while the loss of DRFA cannot converge to zero even with in\ufb01nity samples. We have two insights from the above example: (1) Directly applying DRO method to FL settings can be easily affected by the local noise. (2) The use of local data mixup has the potential to mitigate the local noise. 3.2. Generalization Guarantee of DRFLM In previous subsection, we have demonstrated a motivated example to show the effectiveness of local data mixup. In 1here the mix-up parameter is speci\ufb01ed as \u03b1 = \u03b2 this part, we focus on analyzing the generalization bound of the worst-case client loss in the setting where a generalized linear model is built using our approach. Here, we take generalized linear model as an example and our analysis can be also extended to more complex hypothesis sets such as two-layer neural networks (see more general results in the Appendix B.3): P(y|x; w) \u221dexp(ywT x \u2212\u00b5(wT x)) (4) with x, w \u2208Rd and \u00b5(\u00b7) is some twice differentiable function. The corresponding negative log likelihood f is: f(w; x, y) = \u00b5(wT x) \u2212ywT x. (5) Typical losses such as cross entropy can be obtained by setting different \u00b5. Differentiate R P(y|x; w)dy = 1, we get E[y|x, w] = \u00b5\u2032(wT x), thus Equation (4) can be write equivalently as: y = \u00b5\u2032(wT x) + \u03b5, (6) with \u03b5 as a random variable with zero mean. We further assume that for |z| \u22641 , there exists some K such that K\u22121 \u2264\u00b5\u2032\u2032(z) \u2264K. (7) In particular, the linear model and the Logistic model satis\ufb01es our conditions. Here, we take Linear model for regression as an example: When we set \u00b5(z) = 1 2z2, Equation (6) turns to be y = wT x + \u03b5, and f(w; x, y) = 1 2(y \u2212wT x)2 + 1 2y2 corresponds to the squared loss. By \u00b5\u2032\u2032(z) = 1, we have K = 1 satis\ufb01es Equation (7). \fDRFLM: Distributionally Robust Federated Learning with Inter-client Noise via Local Mixup At a high level, our theory aims to characterize the difference between the population and the empirical (worst-case) losses (i.e., generalization bound of the worst-case loss). The derivation of the generalization bound follows two steps: \u2022 Firstly, we would show that as long as the estimator w lies in some constrained hypothesis class W, then the generalization ability of w over the worst-case loss is theoretically guaranteed. \u2022 Then, we argue that the mix-up effect is nearly a penalty that forces the trained estimator fall into the above hypothesis class W. Step1: Without loss of generality, we denote the i-th local dataset as Di = {zi,j = (xi,j, yi,j)}Ni j=1 and assumes \u2225xi,j\u22252, \u2225w\u2217\u22252 \u22641. We denote \u02c6 \u03a3i := 1 Ni Ni X j=1 xi,jxT i,j, \u03b6i(w) := 1 Ni Ni X j=1 \u00b5\u2032\u2032(wT xi,j), \u03a3\u039b := { N X j=1 \u03bbi \u02c6 \u03a3j : \u03bb \u2208\u039b}. Here \u03a3i and \u02c6 \u03a3i denote the population and empirical covariance matrix of the i-th client, respectively. We further introduce the following hypothesis class, which can be seen as a federated-setting version of the hypothesis set considered in (Zhang et al., 2021): Wr := {x \u2192wT x : min \u03a3\u2208\u03a3\u039b wT \u03a3w \u2264r}. In particular, the above hypothesis class involves a quadratic form wT \u00af \u03a3(w)w, with \u00af \u03a3(w) := argmin\u03a3\u039b{ N X i=1 \u03bbiwT \u03a3iw}, the weighted average of the data covariance matrices across different clients. In fact, we will show later that Wr re\ufb02ect the effect of mixup comparing with DRFA and FedAvg: Without the mix-up effect, the only information we can infer about the empirical minimizer of DRFA and FedAvg is that \u2225w\u22252 \u22641, while the DRFLM loss will push its minimizer into Wr due to the mix-up effect, as we will show in Step2. This will provide a more tight generalization bound in most cases. Now we can show the following generalization bound over the hypothesis class de\ufb01ned above: Theorem 1 (Generalization Bound of Wr). Suppose \u00b5(\u00b7) is L-Lipshitz continuous. Then there exists constants B > 0 such that for w \u2208Wr, the following inequality holds with probability 1 \u2212\u03b4: max \u03bb\u2208\u039b N X j=1 \u03bbjEPj(x,y)[f(w; x, y)] \u2264max \u03bb\u2208\u039b N X j=1 \u03bbj Nj Nj X i=1 f(w; xi j, yi j) + max \u03bb\u2208\u039b N X j=1 \u03bbj s 1 Nj \u0000r log N \u03b4 + N X j=1 rL \u00b7 q Hj/Nj \u0001 , with Hj := EPj[max\u03a3\u2208\u03a3\u039b tr \u0000\u03a3\u2020 \u02c6 \u03a3j \u0001 ] . Hj can partially re\ufb02ects the client heterogeneity from the global distribution. In the IID case, the red part in above inequality turns to be O( p rank(\u03a3)/Nj). In contrast, in both FedAvg and DRFA, this term will be \u2126( p d/Nj). Thus our bound can be more tight when the intrinsic dimension of data is small (i.e., rank(\u03a3) \u226ad). Step2: Now we prove that the learned model obtained by our approach can fall into Wr(\u03bb) due to the mixup effect. Employing similar techniques as in (Zhang et al., 2021), we can show that the object function of DRFLM turns approximately to min w\u2208W max \u03bb\u2208\u039b N X i=1 \u03bbi(fi(w) + c 2\u03b6i(w)wT \u02c6 \u03a3iw | {z } Ri(w) ). (8) Then for any possible \u03bb \u2208\u039b the mixed penalty term R(w) := PN i=1 \u03bbiRi(w) forces the estimator to enter the set Wr for some r. The formal version of the proof is refereed to Appendix 1. 3.3. Convergence analysis In fact, we can establish similar convergence result for our algorithm 1 under non-convex loss functions by following the proof of Theorem 4 in (Deng et al., 2021) and show that our algorithm can ef\ufb01ciently converge to a stationary point of empirical loss. Following the standard analysis of the nonconvex minimax optimization, we de\ufb01ne \u03bb\u2217(w) = argmax\u03bb\u2208\u039b F(w, \u03bb) and \u03a6(w) = F(w, \u03bb\u2217(w)). To derive meaningful results, the convergence measure concerns the so-called Moreau envelope of \u03a6 (Moreau, 1965; Davis & Drusvyatskiy, 2019). De\ufb01nition 1 (Moreau Envelope). A function \u03a6p(x) is a p-Moreau envelope of a function \u03a6 if \u03a6p(x) = minw\u2208W{\u03a6(w) + 1 2p\u2225w \u2212x\u22252}. We use 1 2L-Moreau envelope of \u03a6 and the proof is similar to Theorem 2 (Deng et al., 2021) except we apply it on the mixup distribution. For self-containation, the proof is given in Appendix C: \fDRFLM: Distributionally Robust Federated Learning with Inter-client Noise via Local Mixup Theorem 2. Under the assumptions 1-4 in Appendix C and assume the sample space Z is a convex set, each local function fi is nonconvex, global function F takes the form of our formulation (3), running Algorithm 1 with the choice of \u03c4 = T 1/4, \u03b7 = 1 4LT 3/4 and \u03b3 = 1 \u221a T we have, 1 T T X t=1 E h\r \r\u2207\u03a61/(2L) \u0000wt\u0001\r \r2i \u2264O   D2 \u039b T 1/8 + \u03c32 \u03bb mT 1/4 + G2 \u03bb T 1/4 + Gw p G2 w + \u03c32 w T 1/8 + DW \u0010 \u03c3w + \u221a \u0393 \u0011 T 1/2 ! . 4. Experiments Most previous work (Deng et al., 2021; Mohri et al., 2019) focus on standard and small-scale academic benchmarks in which most samples are high-quality with negligible data noises. For example, one of the most related work to ours is DRFA (Deng et al., 2021) which conducts experiments on Fashion MNIST. In this paper, we turn to evaluate our method on more challenging tasks in the drug discovery era. Except for drug discovery, we also provide additional results on other applications in the discussion part. 4.1. Setup Task and Dataset. We consider two typical tasks in drug discovery, molecular property prediction and drug activity prediction. Speci\ufb01cally, molecular property prediction aims to predict ADMET ((absorption, distribution, metabolism, excretion, and toxicity)) properties of a given molecular, which plays a key role in estimating drug candidates\u2019 safety and ef\ufb01ciency. For this task, we select two datasets from the MoleculenNet (Wu et al., 2017) with different properties, namely solubility, and lipophilicity. Drug af\ufb01nity prediction dataset consists of 34000 drug molecules and corresponding af\ufb01nity degrees measured by the IC50 method from the ChEMBL website (Gaulton et al., 2011). The dataset comes from the DrugOOD project (Ji et al., 2022) which provides automated curator and benchmarks for the out-ofdistribution problem in drug discovery. Data Splitting. A key step to evaluate different FL algorithms is data splitting, i.e, splitting the global dataset into different subsets and assigning them to different clients. We consider the non-IID split in this paper. For molecular property prediction, we employ scaffold splitting (Wu et al., 2017) based on scaffolds of molecular graph, which ensures each client has different training distribution (varies in molecular scaffolds). For drug activity prediction, we employ assay splitting based on assay information, which ensures each client have data from different assay setting. Once the client splitting is \ufb01nished, we split each local dataset into train/validation/test with a ratio 8 : 1 : 1. We gather each local validation set into a global validation set for tuning various hyper-parameters in the training. Model and Training Setup. In this paper, we select three mainstream GNN models for the above tasks including GCN (Kipf & Welling, 2017), GAT (Veli\u02c7 ckovi\u00b4 c et al., 2017) and MPNN (Gilmer et al., 2017). The number of clients is set to 3 in Table 1. We also provide results of other clients number in Figure 2(b). We set the synchronization gap in Algorithm 1 to 5 and set the local learning rate to 0.01. The total training round is set to 100 for all datasets. Evaluation Metric. We evaluate DRFLMin terms of average test RMSE and worst-case test RMSE. Average test RMSE denotes the average RMSE of the global model over all local test datasets. This metric can re\ufb02ect the average model performance on different data distributions. While the worst-case test RMSE is the worst (highest) RMSE among all local test datasets. This metric can partially re\ufb02ect the model\u2019s generalization ability on different data distributions. For all experiments, as suggested by prior work (Wu et al., 2017), we apply three independent runs on three random seeded scaffold/assay splitting and report the mean and standard deviations. 4.2. Results The overall results are shown in Table 1. As shown in the table, our method outperforms the previous state-of-the-art method (DRFA (Deng et al., 2021)) on all three datasets in terms of the worst-case RMSE among all clients. For example, in comparison with DRFA, our method has achieved 17% relative reduction of worst-case RMSE on the ESOL dataset owing to the use of mixup (from 1.357 to 1.121). We further plot the loss curve of the total training phase in Figure 3 which also shows our method\u2019s superiority in convergence speed. Our method can also outperform the conventional method FedAVG in terms of the worst-case RMSE (1.121 vs 1.232). However, for average RMSE, the performance of our method is slightly worse than FedAVG. Despite the ESOL dataset, our method also shows its superiority in Freesolv and drug activity datasets (see more details in Table 1). Effects of label noise. From the result, we also observe that conventional FedAVG method consistently outperforms DRFA in terms of both average and worst-case accuracy. This might be caused by the label noise contained in local datasets. To further validate this conjecture, we run three FL methods (FedAvg, DRFA, DRFLM) on noisy datasets whose labels are perturbed by Gaussian noises with different magnitudes. Speci\ufb01cally, we randomly select 50% clients \fDRFLM: Distributionally Robust Federated Learning with Inter-client Noise via Local Mixup Table 1. The average and worst-case RMSE of client test datasets (lower is better). Method Model ESOL FreeSolv Drug Activity Average Worst-case Average Worst-case Average Worst-case FedAVG GCN 0.976(0.072) 1.232(0.081) 3.011(0.121) 3.863(0.104) 0.932(0.012) 1.104(0.023) GAT 1.134(0.043) 1.343(0.075) 2.641(0.431) 3.456(0.321) 0.964(0.035) 1.197(0.032) MPNN 0.923(0.067) 1.312(0.065) 2.974(0.669) 3.689(0.278) 0.931(0.032) 1.231(0.035) DRFA GCN 1.204(0.068) 1.357(0.073) 3.465(0.264) 3.928(0.276) 1.265(0.037) 1.542(0.029) GAT 1.346(0.065) 1.466(0.046) 2.897(0.568) 3.699(0.772) 1.323(0.027) 1.525(0.063) MPNN 1.233(0.072) 1.386(0.064) 3.211(0.798) 3.899(0.754) 1.148(0.043) 1.458(0.028) Ours GCN 1.086(0.056) 1.121(0.059) 3.123(0.134) 3.532(0.096) 1.126(0.0480) 1.065(0.013) GAT 1.243(0.049) 1.234(0.055) 2.801(0.654) 3.207(0.441) 1.021(0.065) 1.132(0.028) MPNN 0.989(0.054) 1.132(0.061) 3.112(0.649) 3.579(0.489) 0.987(0.041) 1.187(0.029) 0 1 2 3 Noise Level 0.9 1.0 1.1 1.2 1.3 1.4 1.5 1.6 Average RMSE Average RMSE FedAvg DRFA Ours 0 1 2 3 Noise Level 0.9 1.0 1.1 1.2 1.3 1.4 1.5 1.6 Worst RMSE Worst-case RMSE FedAvg DRFA Ours (a) Impact of different noise level. 2 3 4 5 Numer of Clients 1.2 1.3 1.4 1.5 Worst RMSE FedAvg DRFA Ours (b) Different client number. 0 1 2 3 Noise Level 1.3 1.4 1.5 1.6 1.7 Worst RMSE FedAvg DRFA Ours (c) Imbalance mode. Figure 2. Evaluations under different settings including different noise levels, client number, and imbalance mode. and inject different levels (controlled by standard deviation) of Gaussian noises into their data labels with probability 30%. As shown in Figure 2(a), along with the increase of noise level 2, the worst-case RMSE of DRFA signi\ufb01cantly drops which indicates that label noises can easily affect the performance of DRFA. The reason is that DRFA pessimistically optimizes the worst-case combination of local distribution which is typically dominated by the local training distribution with more noises. As a comparison, our method can boost the performance by combining DRFA with the mixup strategy in terms of both average and worstcase RMSE. Note that our method can also achieve better worst-case RMSE than FedAvg. This indicates our method has better generalization ability on different local distributions than conventional FL methods which are appealing to real-world Non-IID settings. Number of clients. The number of clients is set to 3 in Table 1. We also provide results on other client number in Figure 2(b). As shown in the \ufb01gure, our method consistently outperforms the other two methods in all settings in terms of the worst-case RMSE. And in most cases, the increase of the client number leads to signi\ufb01cant performance reduction. This may be caused by that we use non-iid splitting in this paper and the increase of client number can increase the model update estimation bias (He et al., 2021). We also note 2larger level means the injected noises have larger standard deviation 0 20 40 60 80 100 Rounds 1 2 3 4 5 6 Worst RMSE Ours DRFA Figure 3. Convergence comparison between ours and DRFA. that DRFA achieves similar worst-case RMSE to FedAvg. One possible reason for this might be the decrease in client numbers can reduce the effect of local noises. Imbalance mode. Previous experiments are all conducted in the balance mode, i.e., the number of samples of each client is the same. In this part, we further evaluate our method under imbalance mode, i.e., the numbers of samples are highly unbalancing among all clients. Speci\ufb01cally, we set the number of clients to 3 and let the sample ratio among these three clients be 7 : 2 : 1. Under the imbalance mode, as shown in Figure 2(c) we observe that the performance of DRFA becomes worse in contrast to the balanced mode. The situation becomes worse with the increase of noise level, which is due to that the client with minimum sample size magni\ufb01es the effect of local data noise on the \ufb01nal learned model. Also, our method can still reduce the noise effect under the imbalance mode. \fDRFLM: Distributionally Robust Federated Learning with Inter-client Noise via Local Mixup Table 2. The worst-case classi\ufb01cation accuracy of FMNIST and Adult datasets. Flip means \ufb02ipping the training samples\u2019 label with a given probability. Dataset FedAvg DRFA Ours FMNIST 70.64(0.212) 71.04(0.321) 70.89(0.234) Adult 68.21(0.187) 67.31(0.256) 69.87(0.268) FMNIST+Flip 67.21(0.342) 66.28(0.439) 69.78(0.376) Adult+Flip 66.23(0.232) 64.86(0.314) 68.66(0.321) Other application. In the above discussions, we evaluate our method on challenging drug discovery tasks. To give a more comprehensive evaluation of our method, we further conduct experiments on some academic benchmarks including Fashion MNIST and UCI Adults, which are commonly used in previous work (Deng et al., 2021; Mohri et al., 2019). Following the setting of prior work (Mohri et al., 2019), we extract the subset of the data labeled with different labels and split this subset into three local datasets, and assign them to different clients. For Fashion MNIST, the subsets are set to t-shirt/top, pullover, and shirt following the prior work (Mohri et al., 2019). For the Adult dataset, we set subsets to doctorate and non-doctorate. The training round is set to 1000, the overall results are shown in Table 2. From this table, we observe that DRFA outperforms all other two methods on the FMNIST dataset in terms of worst-case accuracy. One possible explanation is that the Fashion MNIST dataset contains high-quality labeled samples thus there is no noise that can affect the training procedure. This can be veri\ufb01ed by conducting experiment on the \u201d\ufb02ipped\u201d dataset, i.e., \ufb02ipping the training samples\u2019 label with probability 30%. For the Fashion MNIST with label \ufb02ipping, the performance of DRFA becomes worse than the others methods including ours. The performance drop can be caused by the injected \ufb02ipping noise. Moreover, our method can signi\ufb01cantly boost the worst-case classi\ufb01cation accuracy by using mixup to reduce the noise effect. 5. Related Works 5.1. Distributionally Robust Optimization Distributionally robust optimization is a general learning paradigm that enforcing the learned model to perform well over the local worst-case risk functions. Generally, the local worst-case risk function refers to the supremum of the risk function over ambiguity set which is a vicinity of the empirical data distribution (Duchi et al., 2011). Data distributions contained in the ambiguity set have small distances to the true empirical data distribution wherein the distance is computed based on well-known probability metrics such as \u03c6-divergence and wasserstein distance (Wang et al., 2021). Although DRO can be always formulated as a minimax optimization problem, in most real-world cases, exactly solving this minimax problem is intractable since it involves the supremum over in\ufb01nite many probability distributions, thus two streams of literature to handle it by using either a primaldual pair of in\ufb01nite-dimensional linear programs (Esfahani & Kuhn, 2018) or \ufb01rst-order optimality conditions of the dual. Amongst them, the most related to us is (Namkoong & Duchi, 2016), which proposes an ef\ufb01cient bandit mirror descent algorithm for the minimax problem when the uncertainty set is constructed using f-divergence. 5.2. Federated Learning on Non-IID Data There is a rich body of literatures aims to modify the conventional FL method for solving inter-client data heterogeneity (Non-IID issue). One typical way to improve FedAvg on the Non-iid setting is to employ different local training strategies for different clients (Li et al., 2020b;a; Reddi et al., 2021). We call this line of work adaptive federated learning since they need to either adaptively adjust the local learning rate or modify the local empirical loss function by adding a client-aware regularizer. For example, Scaffold (Karimireddy et al., 2020) identi\ufb01es that the performance degradation of FedAvg on Non-iid data is caused by \u201dclient drift\u201d issue. and introduces extra device variables to adaptively adjust the local loss. A more general and essential way to overcome the client heterogeneity is to incorporate distributionally robust optimization (DRO) (Kwon et al., 2020; Zhai et al., 2021; Duchi et al., 2011) into conventional FL paradigm (Mohri et al., 2019; Deng et al., 2021). DRFA (Deng et al., 2021) is a recent work that proposes to optimize the agnostic (distributionally robust) empirical loss, which combines different local loss functions with learnable weights. This method theoretically ensures that the learned model can perform well over the worst-case local loss combination. Previous work in this line has shown its effectiveness in deep neural networks on some small-scale benchmarks (e.g., Fashion MNIST). 6."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2112.08439v1",
+                    "title": "Generalization Bounds for Stochastic Gradient Langevin Dynamics: A Unified View via Information Leakage Analysis",
+                    "abstract": "Recently, generalization bounds of the non-convex empirical risk minimization\nparadigm using Stochastic Gradient Langevin Dynamics (SGLD) have been\nextensively studied. Several theoretical frameworks have been presented to\nstudy this problem from different perspectives, such as information theory and\nstability. In this paper, we present a unified view from privacy leakage\nanalysis to investigate the generalization bounds of SGLD, along with a\ntheoretical framework for re-deriving previous results in a succinct manner.\n  Aside from theoretical findings, we conduct various numerical studies to\nempirically assess the information leakage issue of SGLD. Additionally, our\ntheoretical and empirical results provide explanations for prior works that\nstudy the membership privacy of SGLD.",
+                    "authors": "Bingzhe Wu, Zhicong Liang, Yatao Bian, ChaoChao Chen, Junzhou Huang, Yuan Yao",
+                    "published": "2021-12-14",
+                    "updated": "2021-12-14",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.AI"
+                    ],
+                    "main_content": "Introduction One of the fundamental problems in deep learning is to characterize generalization bounds of nonconvex stochastic optimization in the empirical risk minimization (ERM) setting (Mou et al. 2018; Kawaguchi, Kaelbling, and Bengio 2017; Nagarajan and Kolter 2019; Arora et al. 2018). Some traditional approaches study the generalization bounds by measuring the complexity of the hypothesis space in a data-dependent manner (e.g., the VC dimension (Vapnik 1998) and the Rademacher complexity (Bartlett and Mendelson 2002)). Nevertheless, directly adopting these complexity measures often fail to characterize the generalization ability of various learning algorithms for training deep neural networks (DNNs), in which the algorithm explores the hypothesis space in an algorithm-dependent manner (Zhang et al. 2017; Mou et al. 2018). Prior works have shown various empirical evidence from the observation that the hyper-parameters of stochastic gradient methods could signi\ufb01cantly affect the generalization ability of the learned DNN models (Mou et al. 2018; Hoffer, Hubara, and Soudry 2017). For example, the large batch size training often fails to generalize to the test dataset well while the small batch size could introduce more stochastic noise as an implicit regularizer, which further Copyright \u00a9 2022, Association for the Advancement of Arti\ufb01cial Intelligence (www.aaai.org). All rights reserved. improves the generalization ability of the learned DNN models (Hoffer, Hubara, and Soudry 2017). The aforementioned issues are driving the research on algorithm-dependent generalization bounds in the deep learning context. Speci\ufb01cally, the generalization bound of Stochastic Gradient Langevin Dynamics (SGLD) in the non-convex setting has drawn increasing attention from the research community. Previous works have demonstrated various generalization bounds of SGLD from different theoretical perspectives (Mou et al. 2018; Bu, Zou, and Veeravalli 2019; Negrea et al. 2019). Stabilitybased theory and information theory are the two most important theoretical tools used in these works. A seminal work that leverages the stability-based theory to obtain the generalization bound of SGLD is conducted by Mou et al. (Mou et al. 2018). The key idea of this work is to prove the uniform stability of SGLD concerning the squared Hellinger distance and further derive the bound via the connection between uniform stability and the expected generalization error (Shalev-Shwartz et al. 2010; Mou et al. 2018). A follow-up work improves the generalization bound by introducing the Bayes stability notion (Negrea et al. 2019). Another line of research works study this problem based on information theory. The core idea is to bound the generalization error via some information measures. For example, an early work (Xu and Raginsky 2017) proposes to bound the expected generalization error in terms of the mutual information between the training dataset and the output hypothesis of the learning algorithm. Bu et al. (Bu, Zou, and Veeravalli 2019) introduce the individual sample mutual information and link it to the expected generalization error. Despite these theoretical results, the connections amongst these lines of research and how to build a uni\ufb01ed understanding of them remain largely unexplored. In this paper, we present a uni\ufb01ed view to study the generalization ability of SGLD via analyzing the privacy leakage of SGLD. Our work is inspired by a number of recent works (Wu et al. 2019c; Wu et al. 2019b; Wang, Lei, and Fienberg 2016), which show that there are both theoretical and empirical connections between the privacy leakage and the generalization error. Intuitively, \u201cimproving the generalization ability\u201d and \u201creducing the individuals\u2019 privacy leakage\u201d share the same goal of encour\faging a model to learn the population\u2019s features instead of memorizing the features of each individual (Dwork 2011; Wu et al. 2019c). Theoretically, previous results have shown that the commonly used privacy notion of differential privacy (Dwork 2011) may provide the property of uniform stability, which can be further related to the bound of the expected generalization error. Unfortunately, the notion of differential privacy is too strict to be directly adopted for obtaining reasonable generalization bounds for SGLD, i.e., the higher privacy budget of differential privacy would lead to vacuous bound while lower privacy budget will cause a low model performance. In this paper, we propose using a relaxed privacy notion based on R\u00b4 enyi divergence to analyze the privacy leakage of SGLD. Speci\ufb01cally, under some mild assumptions (e.g., the boundness of the loss function), we \ufb01rst compute the bound of the R\u00b4 enyi divergence between weight parameters learned using two adjacent datasets at each iteration, i.e., the \u201cleaveone-out\u201d R\u00b4 enyi divergence. We could then use the chain rule of the R\u00b4 enyi divergence to compose these bounds of different iterations for deriving the total bound of the \ufb01nal weights learned using SGLD. Based on such a bound on the R\u00b4 enyi divergence, we have a simple and uni\ufb01ed approach to obtain various theoretical bounds in previous studies. For example, the foundation of the work in (Mou et al. 2018) is uniform stability with respect to the Hellinger distance, which implies the boundedness of the Hellinger distance between the hypothesis (i.e., weight parameters) learned using two adjacent datasets. Under our framework, the boundedness of the Hellinger distance can be obtained by the bound of the R\u00b4 enyi divergence, thus can be further used to obtain the generalization bound of SGLD. Moreover, our framework could also give rise to previous works (Xu and Raginsky 2017; Bu, Zou, and Veeravalli 2019) with an information-theoretic view and has the potential to improve previous generalization bounds by introducing tighter privacy notions. See more details and results in Section 2. Besides the theoretical contributions, we also conduct experiments on several datasets where data privacy is desired to empirically evaluate the information leakage of SGLD against membership attacks. Speci\ufb01cally, we design different attacks (e.g., membership attacks) to evaluate the information leakages of models learned using SGLD and SGD. We empirically observe that SGLD could not only improve the generalization ability of a model but also reducing its membership information leakage for the training dataset. 2 Theoretical Framework 2.1 Preliminaries We denote D as the unknown population distribution on the sample space Z and denote W as the parameter space of the hypothesis (e.g., a neural network). Considering a loss function l : Z \u00d7 W \u2192R, the goal of a learning algorithm is to \ufb01nd the parameter w which minimizes the population risk LD(w) = Ez\u223cDl(z; w). Empirically, the population distribution is always intractable, thus we turn to the empirical risk minimization (ERM). In the ERM paradigm, a training dataset S = {zi}n i=1 is given and each zi is sampled from D independently, i.e., zi i.i.d. \u223cD. In particular, zi = (xi, yi) is a sample-label pair in the setting of supervised learning. The aim of a learning algorithm in the ERM setting is to \ufb01nd the parameter which minimizes the empirical risk LS(w) = 1 n Pn i=1 l(z; w). In the above ERM paradigm, given a learning algorithm A, we are interested in its expected generalization error (gap), which is the expected difference between the empirical and population risk, gen(A, D) = Ew\u223cA,S\u223cDn[LD(w) \u2212LS(w)], (1) where the expectation is taken over the randomness from both the sampling process of the training dataset and the learning algorithm (e.g., stochastic error or the injected noises in SGLD). 2.2 Privacy Leakage Analysis In this paper, we study the bound of the expected generalization error from the view of privacy analysis. The learning algorithm A studied in this paper is SGLD, which can be seen as a noisy version of SGD. We denote wt as the output parameter at the t-th iteration. Following the prior work (Negrea et al. 2019), we employ the SGLD updating rule as follows in this paper: wt+1 = wt \u2212\u03b1t\u2202wL(Bt, wt) + \u03b7t, \u03b7t \u223cN(0, 2\u03b1tI) (2) L(Bt, wt) = 1 |Bt| X zi\u2208Bt l(zi, wt), (3) where \u03b1t is the step size. N(0, 2\u03b1tI) denotes the Gaussian distribution. L(Bt, wt) is the loss function computed on a mini-batch Bt that is randomly selected from the whole dataset S. Due to the injected noise and the randomness from the mini-batch, the output parameter of SGLD at each iteration can be seen as a random variable with the density function p(wt+1|wt, S). The goal of this part is to measure the information leakage of the output of the learning algorithm A, i.e., p(wT |S), where T denotes the number of iterations. In this paper, motivated by recent advances in R\u00b4 enyi differential privacy (Mironov 2017), we propose to quantitatively evaluate the information leakage by measuring the difference between the output distributions computed on two adjacent datasets. To do this, we \ufb01rst introduce the concept of R\u00b4 enyi divergence which is used for measuring the difference between two probability distributions: De\ufb01nition 1 (R\u00b4 enyi divergence). For two probability distributions p and q (probability density functions), the R\u00b4 enyi divergence of order \u03bb is de\ufb01ned as: D\u03bb(p\u2225q) \u225c 1 \u03bb \u22121 log Ex\u223cq \u0012p(x) q(x) \u0013\u03bb . We then introduce the concept of adjacent datasets: two datasets S and S\u2032 are adjacent when one can be obtained by removing a single element from the other. Without loss of generality, we assume S\u2032 = S/{zn}. We further denote \fwt and w\u2032 t as the output of SGLD in the t-th iteration, which are computed using S and S\u2032, respectively. We are actually interested in the quantity of D\u03bb(p(wT |S)\u2225p(w\u2032 T |S\u2032)). The bound of this quantity can also be seen as the privacy cost in the notion of R\u00b4 enyi differential privacy (Mironov 2017). At a high level, the bound of D\u03bb(p(wT |S)\u2225p(w\u2032 T |S\u2032)) can be obtained following two steps: (1) deriving the R\u00b4 enyi divergence between p(wt|w<t) 1 and p(w\u2032 t|w\u2032 <t) at the t-th iteration; (2) composing these divergences by the composition theory for the adaptive mechanisms (i.e., learning algorithms). In the case of SGLD, we note that wt only depends on wt\u22121, thus the conditional distribution p(wt|w<t) can be reduced to p(wt|wt\u22121). For step (1), following prior works on analyzing the privacy leakage of DP-SGD (Abadi et al. 2016), we treat p(wt|S) and p(w\u2032 t|S\u2032) as Mixture-Gaussian distribution in the setting of SGLD, then we can explicitly compute the divergence and obtain the following bound: Lemma 1 (Leave-one-out Stability of R\u00b4 enyi Divergence) Suppose p(wt|wt\u22121) and p(w\u2032 t|w\u2032 t\u22121) are the probability distributions computed using SGLD in Eq. (2) (3) given wt\u22121 and w\u2032 t\u22121, and Poisson subsampling with sampling ratio \u03c4 is applied, then the R\u00b4 enyi divergence between these two distributions follows (denoting \u03b2t = 1 |Bt|.): D\u03bb(p(wt|wt\u22121\u2225p(w\u2032 t|wt\u22121)) \u2264\u03bb\u03b1tL2 n2 given that \u03c32 \u22650.53 and \u03bb \u22121 \u22642 3\u03c32 log \u00001/\u03bb\u03c4(1 + \u03c32) \u0001 , where \u03c32 := 2 \u03b1t\u03b22 t L2 , where L is the bound of the gradient norm \u2225\u2202w\u2113(z, w)\u2225. Proof: Suppose that we have: w\u2032 t+1 \u223c X B\u2032 t p(B\u2032 t) \u00b7 N \u0010 w\u2032 t \u2212\u03b1t\u2202wL(B\u2032 t, wt), 2\u03b1tI \u0011 where B\u2032 t is the set of selected samples from S\u2032 with sampling ratio \u03c4. Since S = S\u2032 \u222a{zn}, we have wt+1 \u223c X B\u2032 t p(B\u2032 t)   (1 \u2212\u03c4)N \u0010 wt \u2212\u03b1t\u2202wL(B\u2032 t, wt), 2\u03b1tI \u0011 + \u03c4N \u0010 wt \u2212\u03b1t\u2202wL(B\u2032 t \u222a{zn}, wt), 2\u03b1tI \u0011! For all order \u03bb, R\u00b4 enyi divergence is quasi-convex 1w<t refers to the sequence (w0, w1, \u00b7 \u00b7 \u00b7 , wt\u22121). (Van Erven and Harremos 2014), D\u03bb(p(wt+1|wt\u2225p(w\u2032 t+1|wt)) \u2264 sup B\u2032 t D\u03bb   (1 \u2212\u03c4)N \u0010 wt \u2212\u03b1t\u2202w \u0000\u2212log p(wt) + \u03b2t X zt\u2208B\u2032 t \u2113(zi, wt) \u0001 , 2\u03b1tI \u0011 + \u03c4N \u0010 wt \u2212\u03b1t\u2202w \u0000\u2212log p(wt) + \u03b2t X zi\u2208B\u2032 t\u222a{zn} \u2113(zi, wt) \u0001 , 2\u03b1tI \u0011\r \r \r N \u0010 wt \u2212\u03b1t\u2202w \u0000\u2212log p(wt) + \u03b2t X zt\u2208B\u2032 t \u2113(zi, wt) \u0001 , 2\u03b1tI \u0011! By the translation invariance property of R\u00b4 enyi divergence, D\u03bb(p(wt+1|wt\u2225p(w\u2032 t+1|wt)) \u2264 = D\u03bb   (1 \u2212\u03c4)N \u0010 0, 2\u03b1tI \u0011 + \u03c4N \u0010 \u2212\u03b1t\u03b2t\u2202w\u2113(zn, wt) \u0001 , 2\u03b1tI \u0011\r \r \r N \u0010 0, 2\u03b1tI \u0011! where the last equality comes from the fact that the term on the right-hand side is independent of B\u2032 t. Since the covariances are symmetric, we can assume that \u2212\u03b1t\u03b2t\u2202w\u2113(zn, wt) = ce1 for some constant c \u2264L. The two distributions at hand are then both product distributions that are identical in all coordinates except the \ufb01rst. By additivity of R\u00b4 enyi divergence for product distributions (Van Erven and Harremos 2014), we have D\u03bb(p(wt+1|wt\u2225p(w\u2032 t+1|wt)) \u2264 sup c\u2264C D\u03bb   (1 \u2212\u03c4)N \u0010 0, 2\u03b1t \u0011 + \u03c4N \u0010 \u03b1t\u03b2tc, 2\u03b1t \u0011\r \r \rN \u0010 0, 2\u03b1t \u0011! \u2264D\u03bb   (1 \u2212\u03c4)N \u0010 0, 2\u03b1t \u0011 + \u03c4N \u0010 \u03b1t\u03b2tC, 2\u03b1t \u0011\r \r \rN \u0010 0, 2\u03b1t \u0011! = D\u03bb   (1 \u2212\u03c4)N \u0010 0, \u03c32\u0011 + \u03c4N \u0010 1, \u03c32\u0011\r \r \rN \u0010 0, \u03c32\u0011! where we denote \u03c32 := 2 \u03b1t\u03b22 t L2 . Now we reduce our problem to the bounding of R\u00b4 enyi divergence of Gaussian mechanism with Poisson subsampling: D\u03bb   (1 \u2212\u03c4)N \u0010 0, \u03c32\u0011 + \u03c4N \u0010 1, \u03c32\u0011\r \r \rN \u0010 0, \u03c32\u0011! \u2264 1 \u03bb \u22121 log   (\u03bb\u03c4 \u2212\u03c4 + 1)(1 \u2212\u03c4)\u03bb\u22121+ \u0012\u03bb 2 \u0013 (1 \u2212\u03c4)\u03bb\u22122\u03c4 2e\u03c1(2) + \u03bb X j=3 \u0012\u03bb j \u0013 (1 \u2212\u03c4)\u03bb\u2212j\u03c4 je(j\u22121)\u03c1(j) ! \fwhere \u03c1(j) = j/2\u03c32 (Zhu and Wang 2019; Mironov, Talwar, and Zhang 2019). By Lemma 3 of (Liang et al. 2020), we know that D\u03bb   (1 \u2212\u03c4)N \u0010 0, \u03c32\u0011 + \u03c4N \u0010 1, \u03c32\u0011\r \r \rN \u0010 0, \u03c32\u0011! \u2264 2\u03bb\u03c4 2/\u03c32 \u2264\u03bb\u03b1tL2 n2 , given \u03c32 \u22650.53 and \u03bb \u22121 \u2264 2 3\u03c32 log \u00001/\u03bb\u03c4(1 + \u03c32) \u0001 , where \u03c32 := 2 \u03b1t\u03b22 t L2 . For step (2), we notice that the sequence {wt}T t=1 can be seen as the outputs of a series of adaptive mechanisms (see the iterative rule in Equation (2)). Here \u201dadaptive\u201d means the input of the current mechanism is the output of the last mechanism. Thus we can further compose the bounds of interest via composition theories. This way of composing privacy costs has been widely applied in prior works, such as moment accountants for differential privacy (Abadi et al. 2016). In this paper, we borrow the composition theory from the prior work (Proposition 1 in (Mironov 2017)): Lemma 2 (Composability) Suppose that a randomized algorithm A consists of a sequence of adaptive mechanisms A1, . . . , AT where At : Qt\u22121 j=1 wj \u00d7 S \u2192wt. The R\u00b4 enyi divergence of the \ufb01nal outputs (on two adjacent datasets) can be bounded by: D\u03bb(A(S)\u2225A(S\u2032)) \u2264 T X t=1 D\u03bb(At(S)\u2225At(S\u2032)). (4) In our setting, the algorithm is set to be SGLD and the submechanisms can compute the output based on the updating rule at each iteration. Combining Lemma 1 and 2, we can obtain the \ufb01nal bound as the following theorem: Theorem 1 Suppose wT and w\u2032 T are the \ufb01nal outputs of SGLD on S and S\u2032, respectively. The training dataset size is set to n. If the gradient norm is bounded by L, then the R\u00b4 enyi divergence between p(wT |S) and p(w\u2032 T |S\u2032) is bounded as: D\u03bb(p(wT |S)\u2225p(w\u2032 T |S\u2032)) \u2264\u03bbL2 n2 T X t=1 \u03b1t. (5) Theorem 1 is the main building block of our framework and we can further use it for deriving and interpreting previous results, which will be described in detail subsequently. From the view of privacy leakage, Theorem 1 demonstrates that under some mild conditions, the model learned using SGLD incurs a controllable privacy leakage, i.e., SGLD satis\ufb01es (\u03bb, \u01ebn)-R\u00b4 enyi differential privacy (\u01ebn is the right term of Equation (5)). And the privacy cost \u01ebn is dominated by the size of training dataset , i.e., \u01ebn = O( 1 n2 ). In Section 4, we also empirically evaluate the information leakage against wellknown membership attack approaches. 2.3 Stability-based Theory A promising way to study the generalization error is to derive the algorithm-dependent generalization bounds via various stability notions (Shalev-Shwartz et al. 2010; Mou et al. 2018; Wang, Lei, and Fienberg 2016). The core idea is that the generalization error can be bounded in terms of the stability of the learning algorithm. A commonly-used notion is uniform stability, which can be connected to the expected generalization error of a randomized learning algorithm (Shalev-Shwartz et al. 2010). Mou et al. (Mou et al. 2018) propose to prove the uniform stability via bounding the squared Hellinger distance between p(wT |S) and p(w\u2032 T |S\u2032). Here, the squared Hellinger distance between two density functions p and q are de\ufb01ned as: DH(p|q) \u225c1 2 R (\u221ap\u2212\u221aq)2dw. We note that the boundness of the Hellinger distance can be obtained via the bound of the R\u00b4 enyi divergence in terms of DH(p\u2225q) \u2264D\u03bb(p\u2225q), \u03bb \u2265 1/2. Based on this observation, we can formally characterize the generalization error as the following theorem: Theorem 2 Consider the learning algorithm A to be SGLD with T iterations and the batch size is set to be b. And the sampling ratio \u03c4 is set to be b/n. Suppose that the loss function l(z, w) is uniformly bounded by C, and the gradient norm is bounded by L. The expected generalization error (Equation (1)) can be bounded as: gen(A, D) \u2264 \u221a 2LC n ( T X t=1 \u03b1t)1/2. (6) Proof Sketch: Denoting \u01ebn to be the right term in Equation (5). Then we obtain: gen(A, D) \u2264sup S,S\u2032 2C q DH(p(wT |S)\u2225p(w\u2032 T |S\u2032)) \u2264sup S,S\u2032 2C q D1/2(p(wT |S)\u2225p(w\u2032 T |S\u2032)) \u22642C\u221a\u01ebn. The proof details can be found in Appendix. This generalization bound has a similar magnitude with the bound derived in the prior work (Mou et al. 2018). Different from (Mou et al. 2018), we take a novel view of privacy leakage analysis, which can largely simplify the derivation process and can be extended to a wide range of learning algorithms that satis\ufb01es R\u00b4 enyi differential privacy, such as DP-SGD proposed in the prior work (Abadi et al. 2016). 2.4 Information-based Theory The above discussion explores the connection between the information leakage and the stability and further obtains a speci\ufb01c bound for the expected generalization error. We note that a more explicit way to quantify the information leakage is to measure the mutual information between the learned weight and the training dataset. On the other hand, we note that the main research line to study the generalization of a learning algorithm is to build the connection between mutual information and the expected generalization error. Therefore, we can use mutual information to bridge the gap between privacy leakage and the generalization bound of SGLD. To \fthis end, we introduce an important theoretical block used in this paper as the following lemma: Lemma 3 (Proposition 1 in (Bu, Zou, and Veeravalli 2019)) Suppose the loss function l(z, wt) is \u03c3-subgaussian under the population distribution D for all w \u2208W 2. A is the learning algorithm. Then we have: |gen(A, D)| \u22641 n n X i=1 p 2\u03c32I(A(S); zi). The above lemma has established the formal connection between the mutual information and the generalization error and has been widely used in previous works that study the generalization of SGLD. In this paper, we further build the connection between the mutual information and the information leakage, which is re\ufb02ected by the R\u00b4 enyi divergence in our framework. Motivated by recent works (Poole et al. 2019), we propose using variational inference to derive the upper bound of I(p(wT |S); zi). The derived bound can be further related to the R\u00b4 enyi divergence between the posterior distribution over two adjacent datasets, i.e., D\u03bb(p(wT |S)\u2225p(w\u2032 T |S\u2032)). This connection is formalized as: Lemma 4 Suppose S and S\u2032 are adjacent and A is the learning algorithm, then: I(A(S); zi) \u2264ES[D\u03bb(p(wT |S)\u2225p(w\u2032 T |S\u2032))]. The proof follows two steps: (1) deriving the variational bound of the mutual information using variational inference; (2) deriving the upper bound of the variational bound explicitly. The proof details are provided in Appendix. Combining Theorem 1, Lemma 3, and Lemma 4, we can build the theoretical connection between the expected generalization error and the information leakage measured by the R\u00b4 enyi divergence as: Theorem 3 Consider the learning algorithm A to be SGLD with T iterations and the batch size is set to b. And the training dataset size is set to n. Suppose that the loss function l(z, w) is \u03c3-subgaussian under the population distribution D for all w \u2208W. And the gradient norm is upper bounded by L. The expected generalization error (Equation (1)) can be bounded as: |gen(A, D)| \u2264( \u221a 2\u03c3L n )( T X t=1 \u03b1t)1/2. (7) The derivation of the theorem paves a way to derive the generalization bound by the connection between the mutual information and the information leakage (see details in Appendix). 2.5 Discussion on Our Results We provide a uni\ufb01ed and succinct way to analyze the generalization ability of SGLD through information leakage analysis. The information leakage routine utilized 2A random variable X \u03c3-subgaussian if log E[e\u03bb(X\u2212EX)] \u2264 \u03bb2\u03c32/2 for all \u03bb \u2208R. in this paper uni\ufb01es two seemingly unrelated ways to obtain the generalization bound: the stability-based route and the information theoretic route. Meanwhile, our approach largely simpli\ufb01es the analysis: For example, we have obtained a generalization bound of O( LC n ) similar to that of (Mou et al. 2018), but with a simple proof thanks to the theoretical tools from the data privacy community (Mironov 2017). However, (Mou et al. 2018) need to analyze the bound based on stochastic differential equations (SDE) and transform the continuous setting into discrete one. The generic privacy notion (R\u00b4 enyi differential privacy) adopted gives rise to re\ufb01ned bounds, and existing bounds could be further improved by using other generic notions, such as the Bayesian differential privacy. More importantly, base on the technique used by (Mou et al. 2018), it is challenging to further improve the generalization bound. However, this can be easily done under our framework by introducing more advanced stability notions or privacy notions. For example, we can replace the stability notion with Bayes stability used in (Li, Luo, and Qiao 2020) to obtain a more tight bound. 2.6 Comparisons with Previous Works There are two most related direction to our works. One is to study generalization ability of SGLD. Another is to study the connection between generalization and differential privacy in general case. In this part, we highlight the differences between our work and these most related works. In contrast to research focusing on the generalization bounds of SGLD, we aim to unify this line of works through the information leakage analysis of SGLD. This uni\ufb01ed view can help to simplify the derivation of generalization bounds of SGLD and provide new insights to improve these bounds. There are another research line focusing on building the connection between generalization bounds and differential privacy in the general case instead of deriving concrete generalization bounds (Dwork et al. 2015; Bassily et al. 2016). For example, there are a series of work on adaptive data analysis (Dwork et al. 2015; Hardt and Ullman 2014; Russo and Zou 2016; Bassily et al. 2021; Jung et al. 2019), which study false discovery and generalization error led by different settings of adaptivity. While both of the works rely on techniques from differential privacy, in a speci\ufb01c case of SGLD, we need to carefully derive its information leakage in terms of R\u00b4 enyi differential privacy under some mild conditions, i.e., Theorem 1 and its derivation. Once we obtain its privacy leakage, we can employ previous works to build concrete generalization bound. 3 Related Work In this part, we brie\ufb02y review some related works on the theoretical properties of SGLD. Convergence Property. In early times, SGLD is proposed to enable Bayesian learning on large scale datasets and can be seen as an SG-MCMC method, which can approximate \fthe process of sampling from the intractable posterior distribution (e.g., p(w|S) in our paper). One of the important properties of SGLD is that the parameters learned using SGLD will approach samples from the parameter posterior distribution. Welling et al. (Welling and Teh 2011) have shown some intuitive observations of this property. Subsequent works have theoretically proved this proposition and have presented various convergence analysis, i.e., how fast the learned parameter can approach the posterior distribution. The key idea of these works is to treat SGLD as the discretization of some speci\ufb01c stochastic differential equations (SDE, e.g., Fokker-Planck equation) and turn to study the discretization error (Chen, Ding, and Carin 2015; Sato and Nakagawa 2014). Generalization Bounds of SGLD. More recently, the generalization bound of SGLD has emerged as an active topic. Most of the previous works study this problem from two theoretical aspects, the stability-based theory and information-theoretic quantities. A representative work based on stability-based theory is conducted by Mou et al. (Mou et al. 2018), in which they have derived that SGLD satis\ufb01es \u01ebn uniform stability with respect to the Hellinger loss. Then the expected generalization error is bounded by \u01ebn according to the stability-based theory (Shalev-Shwartz et al. 2010). Another research line studies the generalizability of SGLD from an information-theoretic view. The core idea is to bound the generalization error via some information measures. For example, an early work (Xu and Raginsky 2017) proposes to bound the expected generalization error in terms of the mutual information between the training dataset and the output hypothesis of the learning algorithm. Bu et al. (Bu, Zou, and Veeravalli 2019) introduce the individual sample mutual information and touch it to the expected generalization error. Besides these theoretical works, there is also empirical evidence showing that SGLD can improve the generalization ability of the model. For example, Gan et al. (Gan et al. 2017) propose using SGLD to mitigate the over\ufb01tting of RNN models for language modeling tasks. Privacy Issues in SGLD. There are also some related works that study the privacy issues in SGLD. Wang et al. (Wang, Fienberg, and Smola 2015) present a variant of SGLD that satis\ufb01es differential privacy. Wu et al. (Wu et al. 2019a) further prove that SGLD can naturally prevent the well-known membership attack which is also conducted in our experiments. These works also point out some implicit connections between privacy leakage and generalizations. In our paper, we further formally build the connection between previous theoretical results via the privacy leakage analysis. 4 Empirical Results In this section, we perform extensive experiments to empirically evaluate the information leakage of SGLD against membership attacks. In what follows, we \ufb01rst describe the detailed experimental settings, including descriptions of the datasets, models, and the attack method. Then we demonstrate and analyze some numerical results. At last, we provide some discussions for further insights. 4.1 Settings Datasets. The goal of our experiment is to empirically verify the connection between the generalization error and the privacy leakage of SGLD. Therefore we choose datasets from scenarios where data privacy is important, such as \ufb01nancial and census data analysis. Speci\ufb01cally, we select three datasets: (1) The German Credit dataset for credit modeling. (2) The UCI-adult dataset for income prediction (Kohavi 1996). (3) The IDC dataset for pathological image classi\ufb01cation 3. The German Credit dataset consists of 1, 000 applications for credit cards. Each application is labeled with good or bad credit. We consider the classi\ufb01cation task to identify \u201cgood credit\u201d applications. We randomly split the whole dataset into training (400 applications), hold-out/validation (300 applications), and test (300 applications) sets. Here, the holdout dataset can be used for either validation or building the \u201cshadow\u201d attack model (Shokri et al. 2017). We consider a second Adult Census Income dataset (Kohavi 1996) with 48,842 instances from UC Irvine repository 4 to examine the property attack issues. The dataset has 14 features such as country, age, work-class, education, etc. The goal for this binary classi\ufb01cation task is to predict whether income exceeds 50K/yr based on the census information. We split the whole dataset into training, validation, and test sets, each with #22,792, #9,769, #16,281 instances respectively. The validation set is also used as the shadow dataset for building the attack model. We also use a medical image dataset to observe the performance of convolutional neural networks (CNNs). Here, we consider the IDC dataset of pathological images for invasive ductal carcinoma (IDC) classi\ufb01cation. This dataset contains 277,524 patches of 50 \u00d7 50 pixels (198,738 IDC-negative and 78,786 IDC-positive). We split the whole dataset into training, validation (hold-out), and test sets. To be speci\ufb01c, the training dataset consists of 10, 788 positive patches and 29, 164 negative patches. The test dataset consists of 11, 595 positive patches and 31, 825 negative patches. The remain patches are used as the hold-out dataset. Model Setup. For the German Credit dataset, we train a three-layer fully-connected neural network for the classi\ufb01cation task. For comparison, we train the model using both SGLD and SGD variants. All these training strategies share the following hyper-parameters: the mini-batch is set to be 32 and the epoch number is set to be 30. The learning rate decreases by half every 5 epochs. The initial learning rate is set to be 1 \u00d7 10\u22123. For the UCI-adult dataset, we train a fully-connected neural network with a structure of [32, 16]. The mini-batch size and the number of epochs are set as 64 and 100 respectively. SGD is adopted with initial learning as 0.1 and halves every 3http://www.andrewjanowczyk.com/use-case-6 -invasive-ductal-carcinoma-idc-segmentation/ 4http://archive.ics.uci.edu /ml/datasets/Census+Income \fTable 1: Overall results on three benchmarks. Attack denotes anti-attack ability measured by the accuracy of the membership predictor (lower is better). Model denotes the model performance measured by the accuracy of the test dataset (higher is better). Gap is the difference between training and test accuracy. Method German Credit UCI-adult IDC Attack Model Gap Attack Model Gap Attack Model Gap SGD 0.646 0.734 0.266 0.999 0.849 0.018 0.646 0.823 0.170 Dropout 0.527 0.747 0.112 0.891 0.834 0.005 0.634 0.818 0.176 SGLD 0.539 0.736 0.075 0.609 0.847 0.010 0.620 0.817 0.156 Table 2: The results of property attack against transfer learning on the UCI dataset. The target property is set to \u201drace\u201d. Strategy Model Attack SGD 0.846 0.777 SGLD 0.850 0.686 20 epochs, while for SGLD, the variance \u03c32 of the prior is set to be 1.0. For images in the IDC dataset, we train a ResNet-18 model for the classi\ufb01cation task (i.e, idc and non-idc). The mini-batch size is set to be 128 and the epoch number is set to be 100. Data augmentation is not used. The learning rate decreases by half every 20 epochs. For SGLD, the variance \u03c32 of the prior is set to be 1.0. The initial learning rate is set to be 1 \u00d7 10\u22124. Attack Setup. To quantitatively evaluate the information leakage of different learning algorithms, we perform membership attacks on learned models. Given a learned model, the goal of the membership attack is to determine whether a sample is used for training the model. Formally, we can treat this task as a binary classi\ufb01cation problem, in which we need to train a membership predictor to predict the membership information of a sample (i.e., belonging to the training dataset or not). In this paper, for German Credit and UCIadult datasets, we employ a widely used technique named shadow model training for building the membership predictor. We refer readers to the prior work (Shokri et al. 2017) for more details. For the IDC dataset, we use the threshold attack following the prior work (Wu et al. 2019a). We use the accuracy of the membership attack to measure the privacy leakage of different training strategies (see the details of the training strategies in the next section). 4.2 Result Table 1 shows that the information leakage of three different training strategies. To be speci\ufb01c, Dropout denotes adding Dropout layers into the network architecture. For the fullyconnected network used for German Credit and UCI-adult, we insert Dropout after all hidden layers and set the drop ratio to 0.5. For ResNet-18, we insert Dropout between convolutional layers and set the drop ratio to 0.3 following the work (Wu et al. 2019c). As shown in Table 1, SGLD has demonstrated strong antiattack ability and generalization ability in terms of the attack accuracy and the model accuracy, respectively. For the UCI-adult task, in comparison to the conventional SGD algorithm, the model trained using SGLD achieves better generalization ability in terms of the gap between the training and test model accuracy while reducing the attack accuracy score from 0.999 to 0.609. For the IDC task, SGLD has also demonstrated its ability in reducing the information leakage (the attack accuracy score decreased from 0.646 to 0.620) while achieving better generalization ability than SGD (better model accuracy and smaller training-test accuracy gap). Previous works have shown Dropout can be seen as an effective method to improve the generalization ability of the learned model. In our experiment, we also validate its ability to reduce information leakage in the case of a fullyconnected network. Transfer learning. We note that there are naturally two data parties in transfer learning, i.e., the source and target domains. And a commonly-used approach of transfer learning is to build the training protocol based on feature alignment loss (Long et al. 2015). This approach involves the hidden feature communications between two data parties, which incurs information leakage risks. Here we show that SGLD can reduce such an information leakage while achieving a comparable model accuracy. Speci\ufb01cally, we consider the UCI dataset in the context of transfer learning, where we use data instances with the country of \u201cU.S.\u201d as a source domain and \u201cnon-U.S.\u201d as the target. An MLP is used as the base model. To evaluate the information leakage, we perform the property attack that uses the \u201drace\u201d as the target property (see more details of the property attack in (Melis et al. 2019) ). The results are shown in Table 2, which demonstrates SGLD can reduce the information leakage incurred by the hidden feature communications in terms of the attack accuracy. 5"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1910.02249v1",
+                    "title": "Characterizing Membership Privacy in Stochastic Gradient Langevin Dynamics",
+                    "abstract": "Bayesian deep learning is recently regarded as an intrinsic way to\ncharacterize the weight uncertainty of deep neural networks~(DNNs). Stochastic\nGradient Langevin Dynamics~(SGLD) is an effective method to enable Bayesian\ndeep learning on large-scale datasets. Previous theoretical studies have shown\nvarious appealing properties of SGLD, ranging from the convergence properties\nto the generalization bounds. In this paper, we study the properties of SGLD\nfrom a novel perspective of membership privacy protection (i.e., preventing the\nmembership attack). The membership attack, which aims to determine whether a\nspecific sample is used for training a given DNN model, has emerged as a common\nthreat against deep learning algorithms. To this end, we build a theoretical\nframework to analyze the information leakage (w.r.t. the training dataset) of a\nmodel trained using SGLD. Based on this framework, we demonstrate that SGLD can\nprevent the information leakage of the training dataset to a certain extent.\nMoreover, our theoretical analysis can be naturally extended to other types of\nStochastic Gradient Markov Chain Monte Carlo (SG-MCMC) methods. Empirical\nresults on different datasets and models verify our theoretical findings and\nsuggest that the SGLD algorithm can not only reduce the information leakage but\nalso improve the generalization ability of the DNN models in real-world\napplications.",
+                    "authors": "Bingzhe Wu, Chaochao Chen, Shiwan Zhao, Cen Chen, Yuan Yao, Guangyu Sun, Li Wang, Xiaolu Zhang, Jun Zhou",
+                    "published": "2019-10-05",
+                    "updated": "2019-10-05",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "stat.ML"
+                    ],
+                    "main_content": "Introduction Bayesian deep learning has received increasing attention from the research community over the past few years, due to its effectiveness to capture weight uncertainty and alleviate over\ufb01tting in deep neural networks (DNNs) (Blundell et al. 2015; Li et al. 2016a). The key ingredient of Bayesian deep learning is sampling from the Bayesian posterior. To enable the posterior sampling, there are typically two different types of approaches, namely, Stochastic Variational Inference (SVI) and Stochastic Gradient Markov Chain Monte Carlo (SGMCMC). One representative work of SVI is Bayes-byBackprop (Blundell et al. 2015) that enables the variational inference of DNNs. These SVI based methods in the realm of deep learning always assume that the Bayesian posterior can be decomposed into separable Gaussian distributions, which may lead to non-negligible approximation error and Preprint. underestimation of the weight uncertainty (Li et al. 2016a). Such SVI based approaches can hardly scale to large datasets and dif\ufb01cult to handle complex model structures, thus limit their applicability for deep learning models. To enable Bayesian learning at scale, SG-MCMC is proposed as an alternative to SVI (Welling and Teh 2011; Chen, Fox, and Guestrin 2014). Stochastic Gradient Langevin Dynamics (SGLD) is one of the most popular SG-MCMC methods. The key idea of SGLD is applying stochastic optimization (Robbins and Monro 1951) to Langevin Dynamics (Neal and others 2011), which is implemented by injecting proper Gaussian noise to the gradient estimation of a mini-batch sampled from the training dataset (Welling and Teh 2011). SGLD can also be regarded as an effective way to approximately sample from the Bayesian posterior distribution (Sato and Nakagawa 2014). It has been successfully applied to a number of real-world applications, such as language modeling (Gan et al. 2017) and shape classi\ufb01cation (Li et al. 2016b). Among these practical applications, various theoretical studies have shown different appealing properties of SGLD (Sato and Nakagawa 2014; Mou et al. 2018). Most of these theoretical works focus on analyzing the convergence properties of SGLD. Besides the convergenceproperties, some recent works also attempted to study the generalization bounds of SGLD (Mou et al. 2018; Pensia, Jog, and Loh 2018). In this paper, we speci\ufb01cally study the properties of SGLD from the perspective of membership privacy. Some recent studies showed that a well-trained DNN model may unintentionally expose the private information of the training dataset (Fredrikson, Jha, and Ristenpart 2015; Shokri et al. 2017). The membership attack (Shokri et al. 2017) has emerged as a common threat to detect the membership information of the training dataset armed with the learned model. Given the trained model and a speci\ufb01c sample, the membership attack aims to infer whether this given sample is used for training the given model. Our work is motivated by some recent works that study the connection between the over\ufb01tting and the information leakage of DNNs, which shows that the less over\ufb01tting a model is, the less information is disclosed by the model (Wu et al. 2019a; Yeom et al. 2018; Wu et al. 2019b). \fTo this end, we \ufb01rst build the theoretical framework to analyze the information leakage of the learned model. Speci\ufb01cally, we employ the optimal membership attack proposed in a previous work (Sablayrolles et al. 2019) as a workhorse to compute the membership privacy leakage loss. This loss can be used for quantitatively characterizing the model\u2019s vulnerability to the membership attack. Based on the theoretical framework, we then prove that, for a model trained using SGLD, its membership privacy leakage loss can be bounded in expectation by a uniform constant for arbitrary samples. This \ufb01nding further indicates that the model trained using SGLD can prevent the membership attack to a certain extent (formalized in Theorem 1). The most related work to ours was conducted by Wang, Fienberg, and Smola (2015), in which they introduced DP-SGLD that satis\ufb01es the differential privacy protocol (Dwork 2011). Our work differs from this work in two aspects. First, our theoretical analysis can be conducted without modifying the vanilla SGLD algorithm and can be easily extended to other SG-MCMC methods. Second, differential privacy is a strict requirement for most real applications and leads to a signi\ufb01cant performance drop to meet the given privacy budget. DP-SGLD suffers such a performance drop (Wang, Fienberg, and Smola 2015). In contrast, we relax the requirement of differential privacy and focus on studying membership privacy (instead of differential privacy) afforded by SGLD. Moreover, we build the theoretical connection between DP-SGLD and our framework (see more details in the section of theoretical analysis). Our theoretical analysis can be further naturally extended to other types of SG-MCMC methods, such as Stochastic Gradient Hamiltonian Monte Carlo (SGHMC) (Chen, Fox, and Guestrin 2014). To verify our theoretical \ufb01ndings, we perform membership attacks on different real-world datasets to evaluate the information leakage of models trained with different optimization methods (e.g. SGD and SGLD). Empirical results show that SGLD can signi\ufb01cantly reduce the information leakage of the training dataset in contrast to SGD and its variants. Besides, we also observe that SGLD can prevent DNNs from over\ufb01tting in some cases, which implicitly veri\ufb01es the connection between over\ufb01tting and information leakage (Yeom et al. 2018). In summary, our contributions are three folds: \u2022 We introduce a theoretical framework to analyze the membership information leakage of a well-trained DNN model. \u2022 With the theoretical framework, we show that the model trained using SGLD can prevent the membership attack to a certain extent. \u2022 Our framework provides a uniform interpretation for some previous theoretical results by incorporating them into our framework. The rest of this paper is organized as follows. We \ufb01rst provide a brief review of the related work. We then demonstrate the theoretical analysis, including the membership privacy analysis framework and our main theoretical results, followed by the the empirical results to validate our theoretical \ufb01ndings. Finally, we conclude our work in the last section. Related Work Bayesian Deep Learning. Avoiding over\ufb01tting of the deep learning model is a longstanding goal for the research community. Bayesian deep learning is recently regarded as an intrinsic way to eliminate over\ufb01tting through ensembling models sampled from the Bayesian posterior (i.e. approximate Bayesian inference) (Blundell et al. 2015). As the Bayesian posterior is always intractable in practice, there are normally two different ways to approximate the posterior sampling, i.e. Stochastic Variational Inference (SVI) (Blundell et al. 2015) and Stochastic Gradient Markov Chain Monte Carlo (SG-MCMC) (Welling and Teh 2011; Chen, Fox, and Guestrin 2014). The key idea of SVI based methods is to use a parameterized variational distribution to approximate the real posterior. The approximation can be further transformed into an optimization problem. To bring this idea into the realm of deep learning, previous works always come at the assumption that the Bayesian posterior of the parameters can be decomposed into separable Gaussian distributions (Blundell et al. 2015; Li et al. 2016a). However, this assumption is always violated, which leads to non-negligible approximation error (Li et al. 2016a). An alternative way is to use SGMCMC methods. In contrast to SVI based methods, the key idea of SG-MCMC is to approximately sample from the real posterior instead of directly approximating the posterior (Welling and Teh 2011). There are various SGMCMC methods, including Stochastic Gradient Langevin Dynamics (SGLD) and Stochastic Gradient Hamiltonian Monte Carlo (SGHMC) (Chen, Fox, and Guestrin 2014). In this paper, we focus on studying privacy related properties of SGLD, one of the most commonly used SG-MCMC methods. SGLD. The idea of combining stochastic optimization (Robbins and Monro 1951) and Langevin Dynamics (Neal and others 2011) was \ufb01rst presented in the work (Welling and Teh 2011), in which the authors proposed Stochastic Gradient Langevin Dynamics (SGLD) to ef\ufb01ciently perform Bayesian learning on large-scale datasets. In contrast to SGD, SGLD injects proper Gaussian noise into the gradient estimation to avoid collapsing to just the maximum a posterior (MAP) solution. Since then, many variants of SGLD are introduced to improve the vanilla SGLD algorithm. For example, Li et al. (2016a) proposed pSGLD, imposing a preconditioner to the optimization process for improving the convergence ef\ufb01cacy of SGLD. Besides the algorithmic improvements of SGLD, there are a number of works focusing on applying SGLD to various applications (Li et al. 2016b; Gan et al. 2017). For instance, Gan et al. (2017) proposed using SGLD to mitigate the over\ufb01tting of RNN models and achieved remarkable performances on several language modeling tasks. Among the above-mentioned works, there is another research direction to study the theoretical properties of SGLD. One of the most important \fproperties of SGLD is that it can approximate the sampling from the posterior distribution (Sato and Nakagawa 2014; Chen, Ding, and Carin 2015). This property has been theoretically studied in different works. A popular way is to conduct the approximation analysis of SGLD from the perspective of stochastic differential equations. A recent research trend considers the generalization bound of the SGLD algorithms under different assumptions (Mou et al. 2018; Pensia, Jog, and Loh 2018). In this paper, we take a totally different view, to study SGLD from the aspect of membership privacy. Membership Attack. The privacy leakage issue of machine learning models has attracted ever-rising interests from both the machine learning and security communities. Recently, membership attacks have arisen as a common threat model against machine learning algorithms (Leino and Fredrikson 2019; Shokri et al. 2017; Long et al. 2018). A pioneering work (Shokri et al. 2017) investigated the risk of membership attacks on different machine learning models. Speci\ufb01cally, they developed a shadow training technique to obtain an attack model in the black-box setting (i.e., without knowing the machine learning model structure and parameters). Carlini et al. (2018) proposed a metric to measure the vulnerability of deep learning models. A recent work (Sablayrolles et al. 2019) introduced the concept of the optimal attack and employed the Bayes theorem to derive a formulation for the attack. In this paper, we take the membership attack as the workhorse for evaluating the information leakage of deep learning models trained using SGLD. Theoretical Analysis In this section, we theoretically analyze the properties of the SGLD algorithm from different aspects that are related to privacy protection. We \ufb01rst introduce some basic notations and settings, including the SGLD algorithm and the membership attack. Then we build the theoretical framework to characterize the information leakage of the trained model. Based on the framework, we demonstrate the main theoretical \ufb01nding of this paper, i.e., the model trained using SGLD can preserve the membership privacy (preventing the membership attack). At last, we provide some other theoretical perspectives and show how to incorporate them into our framework, which helps us to give a uniform understanding of the SGLD algorithm. Notation and Setting We start with basic notations and settings. In this paper, we focus on the classi\ufb01cation task. Note that our analysis can be easily extended to other types of learning problems, such as regression. In a supervised classi\ufb01cation task, there is a labeled training dataset D = {(xi, yi)}N i=1, where x denotes the input data (e.g. an image in the image classi\ufb01cation task) and y denotes the associated label. For simplicity, we denote zi = (xi, yi) as a data-label pair or a sample. The goal of a deep learning model M (e.g. a DNN model) is to assign a probability to each possible label y, i.e. M is a probabilistic model that computes p(y|x; \u03b8) = f(x, \u03b8), where \u03b8 denotes the weight parameters of M and f denotes the inference of M. A typical loss used for optimizing \u03b8 is the negative loglikelihood which satis\ufb01es the following equation: L(D, \u03b8) = X zi\u2208D l(zi, \u03b8), (1) where l(z, \u03b8) is the negative log-likelihood over the individual sample z = (x, y), i.e. l(z, \u03b8) = \u2212log(p(y|x; \u03b8)). SGLD. SGLD can be regarded as an effective optimization method that incorporates uncertainty into the weight parameter. Based on the above notations, we can de\ufb01ne the updating rule of SGLD as: \u03b8t+1 = \u03b8t \u2212(\u01ebt 2 \u2202\u03b8L(Bt, \u03b8t) + \u03b7t), \u03b7t \u223cN(0, \u01ebtI), (2) L(Bt, \u03b8t) = \u2212log p(\u03b8t) + |D| |Bt| X zi\u2208Bt l(zi, \u03b8t), (3) where p(\u03b8t) is the prior of \u03b8t and \u01ebt is the step size. N(0, \u01ebtI) denotes the Gaussian distribution. L(Bt, \u03b8t) can be seen as a stochastic approximation of Equation (1), which is to compute the negative log-likelihood over the mini-batch Bt that is randomly selected from the whole dataset D. Previous studies always assume the step size decreases towards zero at the rate that satis\ufb01es some conditions (Welling and Teh 2011). Due to this assumption, intuitively, the optimization process of SGLD can be divided into two phases in terms of the step size. In the initial phase, the stochastic gradient noise1 will dominate the optimization process due to the large step size. The SGLD algorithm in this phase can be seen as an ef\ufb01cient stochastic gradient descent algorithm. The initial phase is also called as the burn-in phase in the literature (Welling and Teh 2011). In the later phase, the injected Gaussian noise will dominate the process as the step size gets smaller. The parameter obtained by the SGLD algorithm in this phase can be seen as sampling from the true posterior distribution. More rigorous descriptions can be found in prior works (Welling and Teh 2011). Membership Attack. The membership attack refers to inferring whether a speci\ufb01c sample is used for training the model. Formally, given a speci\ufb01c sample z = (x, y) and a DNN model M with the parameter \u03b8, the attacker aims to compute the probability that the sample belongs to the training data. Here we denote the probability as P(m = 1|z, \u03b8), where m refers to the sample\u2019s membership and m = 1 indicates that the sample is in the training dataset. Besides the knowledge of the trained model, the attacker may also hold an extra dataset S, which consists of samples from the training/hold-out datasets. We call this dataset \u201cshadow dataset\u201d borrowed from the prior work2. We denote this dataset as S = {(zi, mi)}n i=1, where each zi is a sample of (xi, yi) and mi represents the membership of this sample. The samples with mi = 1 are from the training dataset, otherwise, are from the hold-out dataset. Based on the above setting, the attacker can employ the shadow dataset S to assist the membership inference, which is to 1The noise is caused by the random sampling. 2This can also be seen as the side information in the literature. \fcompute the probability P(m = 1|z, \u03b8, S). Based on the previous work (Sablayrolles et al. 2019), we call the exact value of P(m = 1|z, \u03b8, S) as the optimal attack. In practice, the optimal attack is always intractable thus previous works have proposed different methods to approximate P(m = 1|z, \u03b8, S) (Sablayrolles et al. 2019). In what follows, we introduce our theoretical framework to analyze the information leakage (in the aspect of membership privacy) of the learned deep learning models. Membership Privacy Analytical Framework We further assume the ratio of training samples3 in S is \u03bb. Thus we can give a trivial attack solution that proceeds as P(m = 1|z, \u03b8, S) \u2248P(m = 1) = \u03bb. In contrast to the optimal attack mentioned above, we call this attack as the trivial attack, since this attack gets nothing other than the macroscopical information of the shadow dataset. We can de\ufb01ne the membership privacy leakage (Mpl) loss of a given sample as: Mpl(z, \u03b8, S) = max(P(m = 1|z, \u03b8, S) \u2212\u03bb, 0). (4) In the above equation, when P(m = 1|z, \u03b8, S) \u2264\u03bb, the optimal attack infers that the sample z is not from the training dataset. Thus we set the Mpl loss to zero since the optimal attack do not induce the membership leakage of the training dataset. The expectation of the Mpl loss over all possible samples (i.e. Ez[Mpl(z, \u03b8, S)]) can evaluate the difference between the optimal attack and the trivial attack. Obviously, the smaller the expectation of the Mpl loss is, the less information of the training dataset is revealed by the model (i.e. the optimal attack is more similar to the trivial attack). To compute the loss, we need to give an implicit formulation of the optimal attack. In this paper, we introduce a modi\ufb01ed version of the formulation proposed by Sablayrolles et al. (2019). To this end, we \ufb01rst de\ufb01ne the posterior distribution over the parameter (while S is given) by: pS(\u03b8) = p(\u03b8|ST ), ST = {(zi, mi) \u2208S : mi = 1}. (5) With a slight modi\ufb01cation of the prior work (Sablayrolles et al. 2019), we can obtain: Lemma 1 (Sablayrolles et al. 2019) Given a learned model with parameter \u03b8, the shadow dataset S, and a speci\ufb01c sample z, the Mpl loss can be computed using: P(m = 1|z, \u03b8, S) = \u03c3(s(z, \u03b8, pS) + t\u03bb), (6) where t\u03bb is log( \u03bb 1 \u2212\u03bb) and \u03c3(t) := 1/(1 + e\u2212t) is the sigmoid function. We call s the measure score which satis\ufb01es the following equations: tp(z) = \u2212log E\u03b8\u223cpe\u2212l(z,\u03b8), s(z, \u03b8, p) = tp(z) \u2212l(z, \u03b8). We can further use the Lipschitz continuity of \u03c3 that |\u03c3(u)\u2212 \u03c3(v)| \u2264|u \u2212v|/4 and \u03bb = \u03c3(t\u03bb) to bound the Mpl loss by: Mpl(z, \u03b8, S) \u2264max(s(z, \u03b8, pS)/4, 0). (7) 3Training samples are those from the training dataset. The proof of Equation (7) can be found in the appendix. Equation (7) points out that the Mpl loss can be bounded by the measure score. Intuitively, the value of tp(z) can be seen as a threshold which is used for comparing with the loss function with respect to z and the trained parameter \u03b8. From Equation (7), if the negative log-likelihood of the given sample z is close to the threshold, i.e., l(z, \u03b8) \u2248tp(z), then the measure score is close to zero, thus the optimal attack can not obtain any information other than the prior knowledge (i.e., P(m = 1) = \u03bb). And when the measure score is substantially greater/less than zero (i.e., the loss of the given sample is much lower/higher than the threshold), the optimal attack can gain non-trivial membership information on z, i.e., P(m = 1|z, \u03b8) \u226b\u03bb. To enable the analysis on the SGLD algorithm, we need to consider the case of model ensemble used for approximating Bayesian inference. We further reformulate the above Equation (7) based on two facts. First, we can use Jensen\u2019s inequality to obtain: tp(z) \u2264E\u03b8\u223cp[l(z, \u03b8)]. (8) Second, we note that, when the sample z and parameter \u03b8 are given, the value of l(z, \u03b8) explicitly depends on the predicted probability with respect to the label y. Here, we denote the predicted probability as \u02c6 y = f(x, \u03b8). With a slight notation abuse, we rewrite the value of l(z, \u03b8) as \u02dc l(y, \u02c6 y). In the case of SGLD ensemble, \u02c6 y can be obtained by average outputs of models produced by the SGLD algorithm. Then, combining with Equation (8) and slightly modifying the above derivation (i.e. replacing \u03b8 with \u02c6 y ), we can reformulate the inequality in Equation (7) as: g Mpl(z, \u02c6 y, S) \u2264max(\u02dc s(z, \u02c6 y, pS)/4, 0), (9) where: \u02dc s(z, \u02c6 y, pS) = E\u03b8\u223cpS[l(z, \u03b8)] \u2212\u02dc l(y, \u02c6 y), (10) where \u02c6 y is the average output of models sampled using SGLD. In contrast to Equation (7) that can only be applied to the context of the single model, we can leverage Equation (9) to analyze the information leakage of the ensembled models4. This is appealing for the analysis of SGLD, since its key feature is to approximate the Bayesian inference via ensembling the models produced by SGLD after the burn-in phase (Welling and Teh 2011). Uniform Bound for the Mpl Loss Up to now, we can leverage the above analytical framework to quantify the information leakage of the trained model in terms of the membership privacy loss. In this part, for the model trained using SGLD, we aim to give a uniform bound of the Mpl loss for arbitrary sample in the worst case. Here, the worst case refers to the shadow dataset S held by the attack is a leave-one-out version of the original training dataset D, i.e., S is obtained by removing one sample from D. Without loss of generality, we assume the removed sample is zn, i.e. S = D/zn. Here, we assume 4Because Equation (9) only depends on the output vector. \fthat the samples in S are all from the training dataset, since the posterior pS (Equation (5)) does not depend on the samples from the hold-out set. The analysis can be easily extended to other scenarios (e.g., S has samples from the hold-out dataset). To introduce the following results, we \ufb01rst give a commonly used assumption presented in the prior work (Wang, Fienberg, and Smola 2015). Assumption 1 We assume z is sampled from the space Z and \u03b8 is sampled from the SGLD algorithm. This assumption limits the negative log-likelihood to be bounded, i.e., supz\u2208Z,\u03b8 |l(z, \u03b8)| \u2264B. The step sizes {\u01ebt} in Equation (2) decrease and satisfy: 1) P\u221e t=1 \u01ebt = \u221e, and 2) limL\u2192\u221e PL t=1 \u01eb2 t St = 0, where St = PL t=1 \u01ebt is the sum of the step sizes. Formally, based on the above settings, we introduce the main theoretical \ufb01nding as the following theorem: Theorem 1 (Uniform bound in expectation) Given S = D/zn and model parameters {\u03b8t}t0+L t=t0 generated by SGLD after the burn-in phase that starts from the t0-th iteration. We denote the sampling process as \u03b8 \u223cSGLD. \u02c6 y is the weighted average output using these parameters, i.e., \u02c6 y = Pt0+L t=t0 \u01ebt SL f(x, \u03b8t), where SL = Pt0+L t=t0 \u01ebt . Under Assumption 1, we can bound the expectation of the Mpl loss for any sample z by a uniform constant MSGLD: ESGLD[ g Mpl(z, \u02c6 y, S)] \u2264MSGLD, (11) MSGLD = B 4 (e2B \u22121) + O( 1 SL + Pt0+L t=t0 \u01eb2 t SL ). (12) The expectation is with respect to the randomness of SGLD (i.e., sampling different parameters to compute the predicted value \u02c6 y). Here, for simplicity, Theorem 1 is for the case of decreasing-step-size (Welling and Teh 2011) and we can easily extend this theorem to the case of constant-stepsize based on the prior work (Sato and Nakagawa 2014). To given a proof of Theorem 1, we \ufb01rst introduce two lemmas. Lemma 2 (Bound of the leave-one-out error) (Mou et al. 2018) Given S = D/zn, under Assumption 1, we can bound the leave-one-out error (i.e. the posterior average difference with respect to these two datasets) as: |E\u03b8\u223cpS[l(z, \u03b8)] \u2212E\u03b8\u223cpD[l(z, \u03b8)]| \u2264B(e2B \u22121). (13) The proof of Lemma 2 can be found in the appendix. The following lemma is based on previous \ufb01nite-time error analysis of SGLD (Chen, Ding, and Carin 2015). Lemma 3 (Finite-time error analysis) (Chen, Ding, and Carin 2015) Given model parameters {\u03b8t}L t=t0 \u223cSGLD generated by SGLD after the burn-in phase and the predicted value \u02c6 l = Pt0+L t=t0 \u01ebt SL l(z, \u03b8t). SL = Pt0+L t=t0 \u01ebt is the sum of the step sizes. Then we have: |ESGLD[\u02c6 l]\u2212E\u03b8\u223cpD[l(z, \u03b8)]| = O( 1 SL + Pt0+L t=t0 \u01eb2 t SL ). (14) Here, with the help of the above two lemmas, we demonstrate the roadmap to accomplish the proof of Theorem 1. The details of the proof can be found in the appendix. Proof 1 (Sketch proof of Theorem 1): Based on Equation (9), we can \ufb01rst bound the measure score, which can be further used for bounding the Mpl loss. To bound the measure score, we can decompose Equation (10) into the leave-one-out error and the approximation error: \u02dc s(z, \u02c6 y, pS) = E\u03b8\u223cpS[l(z, \u03b8)] \u2212E\u03b8\u223cpD[l(z, \u03b8)] | {z } leave\u2212one\u2212out error + E\u03b8\u223cpD[l(z, \u03b8)] \u2212\u02dc l(y, \u02c6 y) | {z } approximation error . Then we can obtain the bounds for these two errors individually. First, we notice that the leave-one-out error can be bounded based on Lemma 1. Then the approximation error can be further decomposed into: E\u03b8\u223cpD[l(z, \u03b8)] \u2212\u02c6 l | {z } F actor1 + \u02c6 l \u2212\u02dc l(y, \u02c6 y) | {z } F actor2 . Note that factor 1 can be bounded via Lemma 2 when taking the expectation over parameters generated by SGLD (the boundedness is ensured by Assumption 1 and Lemma 3 jointly). The boundedness of factor 2 is ensured by the boundedness of the negative log-likelihood and the boundedness of the variance of SGLD (see details in the appendix). Therefore we can obtain the bound of the measure score in the context of SGLD model ensemble. Other Theoretical Perspectives In this part, we revisit two theoretical results provided in previous works, the generalization of SGLD (Mou et al. 2018) and the DP-SGLD (Wang, Fienberg, and Smola 2015). We will show how to give a uni\ufb01ed interpretation of these results from the aspect of membership privacy. Connection to Differential Privacy. Previously, Wang, Fienberg, and Smola (2015) presented DP-SGLD, which modi\ufb01es the existing SGLD algorithm to satisfy differential privacy (Dwork 2011). The key idea of DP-SGLD is to calibrate the magnitude of the injected Gaussian noise and use the strong composition theory to obtain the privacy budget. Here, we show that DP-SGLD can also preserve membership privacy in terms of the boundedness of the Mpl loss. Intuitively, since DP-SGLD satis\ufb01es (\u01eb, \u03b4)-differential privacy, each sample in the training dataset will have a limited impact on the learned model. Therefore, it is more dif\ufb01cult to detect useful membership information of an individual sample. This intuition can be theoretically interpreted using the following property: Property 1 (DP-SGLD) Given \u03b8 learned using (\u01eb, \u03b4)-DPSGLD, then the following inequality holds with probability 1 \u2212\u03b4: Mpl(z, \u03b8, S) \u2264\u01eb 4. (15) \fThe proof of this property can be found in the appendix. Here we only consider the case of a single model without model ensemble. This theorem can also be extended to the case of model ensemble by using the composition theory of differential privacy. A View From Generalization. Previous works have validated the connection between the over\ufb01tting and the information leakage of DNN models (Wu et al. 2019a; Yeom et al. 2018). In the realm of membership privacy, Long et al. (2018) also pointed out that over\ufb01tting is a suf\ufb01cient condition for exposing membership information of the training dataset. This \ufb01nding inspires us to understand why SGLD works for preventing the membership attack from the view of the generalization theory. Speci\ufb01cally, based on the prior work (Mou et al. 2018), we can prove that SGLD can provide the property of uniform stability. Armed with the stability learning theory (Shalev-Shwartz et al. 2010), Mou et al. (2018) further provided an expectation bound for the generalization error of the SGLD algorithm. Since the generalization error can be bounded, it is challenging for the attacker to notice the output\u2019s (w.r.t. the model) difference between the training and hold-out datasets, which further prevents the membership attack (e.g., the threshold attack conducted in our experiments). Therefore, from the view of the generalization bound, we can understand the effectiveness of SGLD to prevent the membership attack. Evaluation Experimental Settings Datasets. In this paper, we choose datasets from scenarios where data privacy is important, such as \ufb01nancial and medical data analysis. Speci\ufb01cally, we select two datasets as our benchmarks, namely, German Credit dataset (Dua and Graff 2017) and IDC dataset5. The German Credit dataset consists of 1, 000 applications for credit cards. Each application is labeled with good or bad credit. We consider the classi\ufb01cation task to identify \u201cgood credit\u201d applications. We randomly split the whole dataset into training (400 applications), hold-out/validation (300 applications), and test (300 applications) sets. Here, the holdout dataset can be used for either validation or building the \u201cshadow\u201d attack model (Shokri et al. 2017). IDC dataset is used for invasive ductal carcinoma (IDC) classi\ufb01cation. This dataset contains 277,524 patches of 50 \u00d7 50 pixels (198,738 IDC-negative and 78,786 IDC-positive). Following the setting of the work (Leino and Fredrikson 2019), we split the whole dataset into training, validation (hold-out), and test sets. To be speci\ufb01c, the training dataset consists of 10, 788 positive patches and 29, 164 negative patches. The test dataset consists of 11, 595 positive patches and 31, 825 negative patches. The remain patches are used as the hold-out dataset. Model Setup. For the German Credit dataset (Dua and Graff 2017), we choose a three-layer fully-connected neural network as the model architecture. 5http://www.andrewjanowczyk.com/use-case-6invasive-ductal-carcinoma-idc-segmentation/ For comparison, we train the model using both SGLD and SGD variants. All these training strategies share the following hyper-parameters: the mini-batch is set to 32 and the epoch number is set to 30. The learning rate decreases by half every 5 epochs. For SGLD, the variance of the prior is set to 1.0. The initial learning rate is set to 1 \u00d7 10\u22123. For the IDC dataset, since the training dataset comprises image data, we choose a typical residual convolutional neural network, ResNet-18 (He et al. 2016). The mini-batch is set to 128 and the epoch number is set to 100. Data augmentation is not used. The learning rate decreases by half every 20 epochs. For SGLD, the variance \u03c32 of the prior is set to 1.0. The initial learning rate is set to 1 \u00d7 10\u22124. Attack Setup. The key part of our experiments is to quantitatively evaluate the information leakage of the model trained using different methods. In practice, we found that a straightforward attack, threshold attack (Yeom et al. 2018), suf\ufb01ces to support our theoretical \ufb01ndings. The threshold attack has already been used in various previous works (Yeom et al. 2018). This attack is based on the intuition, i.e., a sample with lower loss is more likely to belong to the training dataset. We denote A as the attacker and the threshold attack proceeds as follows (here, we use the consistent notations as in the theoretical analysis): \u2022 Given a speci\ufb01c sample z and its predicted probability \u02c6 y computed by the model. \u2022 A calculates the normalized loss n = \u02dc l(y, \u02c6 y)/B, where B is the bound for the negative log-likelihood. \u2022 A selects a threshold t and outputs 1 (indicates the sample is a training sample) if n \u2264t. Otherwise, it outputs 0. The threshold attack is evaluated on the training dataset and the test dataset. Note that we can choose different threshold t to obtain different attack models. Naturally, we can plot the ROC curve of these attack results and use the area under the ROC curve (AUC) as an indicator of the information leakage. Moreover, we assume that the attacker can obtain the average value of \u02dc l(y, \u02c6 y) on the training dataset. Thus we can directly use this value as a threshold to build the attack model, and compute the F1 score and accuracy of the attack model to quantify the amount of information leakage. In summary, we use three metrics to evaluate the information leakage of a learned model, namely, the AUC value, the F1 score, and the accuracy. Experimental Results In Table 1 and 2, Dropout denotes adding Dropout layers into the network architecture. For ResNet18, we insert Dropout between convolutional layers and set the drop ratio to 0.3 following the work (Zagoruyko and Komodakis 2016). SGLD+ represents ensembling models obtained by SGLD. In practice, we found that ensembling 3 models suf\ufb01ces to our case, thus we use the models obtained in the last three iterations and average their outputs for the \ufb01nal prediction. In contrast to the decreasing step size discussed in the theoretical analysis, we use the constant step size in practice following prior works (Sato and Nakagawa 2014). For a fair comparison \fTable 1: Results of the threshold attack on the German Credit dataset. SGLD+ represents model ensemble. Strategy Attack Metric Model Metric AUC F1 Acc Train Test Gap SGD 0.677 0.685 0.646 1.000 0.734 0.266 SGD+ 0.677 0.684 0.645 1.000 0.720 0.280 Dropout 0.527 0.567 0.527 0.859 0.747 0.112 RMSprop 0.589 0.677 0.627 0.965 0.749 0.216 SGLD 0.536 0.598 0.539 0.811 0.736 0.075 SGLD+ 0.536 0.564 0.526 0.839 0.751 0.087 pSGLD 0.551 0.600 0.550 0.858 0.753 0.105 pSGLD+ 0.550 0.568 0.539 0.859 0.758 0.101 with SGLD+, we also apply the same ensemble strategy to the case of SGD (denoted as SGD+). German Credit Dataset. The overall results of the German Credit dataset are shown in Table 1. We observe that the use of SGLD can prevent the information leakage of the training dataset, based on the above attack metrics for evaluating the information leakage. In contrast to SGD, SGLD significantly reduces the AUC score of the threshold attack from 0.677 to 0.536, while achieves comparable model accuracy. We can also observe consistent performance boost and information leakage reduction when model ensemble is used. For example, SGLD+ decreases the F1 score of the threshold attack (from 0.598 to 0.564) while achieves a slight model accuracy increase (from 0.736 to 0.751). IDC Dataset. The overall results are shown in Table 2. SGLD can signi\ufb01cantly reduce the information leakage in contrast to SGD, in terms of the aforementioned metrics for evaluating the attack performance. For example, compared with SGD, SGLD remarkably reduces the AUC value of threshold attack from 0.669 to 0.641 while achieving comparable model accuracy. We also observe a slight model accuracy improvement when the ensemble strategy is used (see the results of SGLD+). It is worth noting that Dropout is not effective in reducing the information leakage and improving the performance in this case (see more discussions in the following subsection). All these empirical results validate our theoretical \ufb01ndings, i.e., the model trained using SGLD can preserve membership privacy to a certain extent. Discussions SGLD Variants. The vanilla SGLD algorithm updates all parameters with the same step size, which always leads to inef\ufb01cient optimization since these parameters have different curvature. pSGLD (Li et al. 2016a) solves this problem by applying the preconditioner to each parameter in the optimization process. In this paper, we also demonstrate some numerical results of pSGLD. Since pSGLD can be seen as a noisy version of the RMSprop algorithm, we also show the results of RMSprop. From Table 1, we observe that pSGLD can improve the test accuracy in contrast to the vanilla SGLD. We infer that the accuracy boost comes from the use Table 2: Results of the threshold attack on the IDC dataset. SGLD+ represents model ensemble. Strategy Attack Metric Model Metric AUC F1 Acc Train Test Gap SGD 0.669 0.716 0.646 0.993 0.823 0.170 SGD+ 0.665 0.714 0.638 0.998 0.837 0.161 Dropout 0.662 0.711 0.634 0.994 0.818 0.176 RMSprop 0.674 0.715 0.652 0.992 0.818 0.174 SGLD 0.641 0.669 0.620 0.973 0.817 0.156 SGLD+ 0.643 0.658 0.617 0.981 0.825 0.156 pSGLD 0.652 0.697 0.641 0.982 0.824 0.158 pSGLD+ 0.648 0.692 0.640 0.987 0.830 0.157 of the preconditioner which can improve the optimization ef\ufb01ciency. However, the accuracy boost accompanies with more information leakage in terms of the above attack metrics. For example, comparing with SGLD, pSGLD increases the AUC value of the attack model from 0.536 to 0.551. From the updating rule of pSGLD (Li et al. 2016a), pSGLD uses more information of the gradient compared with SGLD, which causes the increase of the information leakage. Effectiveness of Dropout. As shown in Table 1 and 2, Dropout demonstrates different effects on these two tasks since two different types of model architectures are used. According to previous works, Dropout is less effective on the fully convolutional neural network in which the fully-connected layer only exists in the last layer. (Ghiasi, Lin, and Le 2018). This conclusion coincides with our observation, i.e., the Dropout cannot boost the performance on the ResNet-18 model for the IDC classi\ufb01cation. On the contrary, in the case of the German Credit dataset, one notable observation is that Dropout can also drastically reduce the information leakage (see Dropout in Table 1). This result is not surprising to us, since Dropout can be seen as an implicit Bayesian inference based on the prior work (Gal and Ghahramani 2016). This suggests that we can use Dropout as an effective way to preserve the membership privacy for fully-connected networks (i.e. multilayer perceptron). Other Attacks. In the above analysis, we found that the simple threshold attack suf\ufb01ces to detect some useful membership information in our case. We also conduct experiments on other membership attack methods, such as shadow training method proposed in the prior work (Shokri et al. 2017). And we found that SGLD is also effective to prevent such an attack in practice."
+                },
+                {
+                    "url": "http://arxiv.org/abs/1908.07882v3",
+                    "title": "Generalization in Generative Adversarial Networks: A Novel Perspective from Privacy Protection",
+                    "abstract": "In this paper, we aim to understand the generalization properties of\ngenerative adversarial networks (GANs) from a new perspective of privacy\nprotection. Theoretically, we prove that a differentially private learning\nalgorithm used for training the GAN does not overfit to a certain degree, i.e.,\nthe generalization gap can be bounded. Moreover, some recent works, such as the\nBayesian GAN, can be re-interpreted based on our theoretical insight from\nprivacy protection. Quantitatively, to evaluate the information leakage of\nwell-trained GAN models, we perform various membership attacks on these models.\nThe results show that previous Lipschitz regularization techniques are\neffective in not only reducing the generalization gap but also alleviating the\ninformation leakage of the training dataset.",
+                    "authors": "Bingzhe Wu, Shiwan Zhao, ChaoChao Chen, Haoyang Xu, Li Wang, Xiaolu Zhang, Guangyu Sun, Jun Zhou",
+                    "published": "2019-08-21",
+                    "updated": "2019-09-25",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.CR",
+                        "stat.ML"
+                    ],
+                    "main_content": "Introduction In the past years, generative adversarial networks (GANs) [11] have achieved remarkable progress in a wide range of applications including image translation [40, 16], image manipulation [39, 6], and image super-resolution [36, 20], etc. More recently, numerous advanced techniques [1, 21, 22, 38] are proposed for improving and stabilizing the training of GANs, leading to more realistic generated images. Despite the tremendous success, there are still numerous open problems to be fully solved, ranging from the theoretical analysis of different regularization techniques [22, 28] to the visualization of different objective functions of GANs [3]. Among these problems, a critical one is how to formally characterize the generalization ability of GANs. Some recent studies attempted to explore this problem in different contexts. For instance, a seminal work in this direction [2] proposed the neural net distance, and the authors further showed the generalization properties of this distance. Qi et al. [25] were motivated by the progress in the Lipschitz regularization and proposed a loss-sensitive GAN. They then developed the Lipschitz regularization theory to analyze the generalization ability of the loss-sensitive GAN. Different from the prior works [25, 2], in this paper, we aim to study the generalization ability of GANs in a relatively general setting from a novel perspective of privacy protection. Our study is motivated by a well-known intuition [37], \u201creducing the generalization gap\u201d and \u201cprotecting the individual\u2019s privacy\u201d share the same goal of encouraging a neural network to learn the population\u2019s features instead of memorizing the features of each individual, i.e., the smaller the generalization gap is, the less information of the training dataset will be revealed. The goal of this paper is to validate this natural intuition theoretically and quantitatively. \u2217Co-second author 33rd Conference on Neural Information Processing Systems (NeurIPS 2019), Vancouver, Canada. arXiv:1908.07882v3  [cs.LG]  25 Sep 2019 \fIn the theoretical side, we leverage the stability-based theory [30] to bridge the gap between differential privacy [10] and the generalization, i.e., a differentially private learning algorithm does not over\ufb01t to a certain degree. Based on the theoretical analysis, we also provide a new perspective from privacy protection to understand a number of recent techniques for improving the performance of GANs, e.g., various Lipschitz regularization terms [12, 22] and training GANs using Bayesian learning [27, 14]. In the experimental side, we quantitatively validate the relationship between the generalization gap and the information leakage of the training dataset. To this end, we introduce the membership attack [31] to evaluate the information leakage of a trained GAN model. In the context of machine learning, the membership attack refers to inferring whether a speci\ufb01c item is from the training dataset while given the trained model (discriminator and generator in our case). Speci\ufb01cally, we design different attack models and perform membership attacks on GANs trained with various objective functions and regularization terms. The results show that previous Lipschitz regularization techniques are effective in not only reducing the generalization gap but also alleviating the information leakage of the training dataset, which implicitly validates the aforementioned intuition. The results also suggest that it is possible to design new variants of GAN from the perspective of building privacy-preserving learning algorithms, which can bring signi\ufb01cant regularization effects while provide appealing property of protecting the sensitive information contained in the training dataset. The rest of this paper is organized as follows. We \ufb01rst brie\ufb02y review related works in Section 2. Then we demonstrate our theoretical analysis in Section 3. Subsequently, we present the quantitative analysis in Section 4. At last, we conclude this work in Section 5. 2 Related Work Generative adversarial networks. In the past years, starting from the illuminative work of GANs [11], many efforts have been devoted to this research area. Numerous researchers have tried to improve the performance of GANs from different perspectives [1, 25, 24, 22]. One direction is to improve the original objective function [25, 1, 24]. For example, to solve the gradient vanishing problem, the Least-square GAN [21] proposed using the least square loss function in the training of GANs. Wasserstein GAN (WGAN) [1] replaced the original Jensen\u2013Shannon (JS) divergence with the Wasserstein distance and proposed an effective approximation method to compute the distance. Besides the improvements on the objective function, lots of algorithmic tricks have been proposed in empirical studies of training GANs. A typical direction is to add Lipschitz constraints on the discriminator, which enables the discriminator to be Lipschitz continuous with respect to the input and the weight. For instance, WGAN proposed using weight clipping to constrain the Lipschitz constant [1]. Gulrajani et al. [12] further suggested to use the gradient penalty to obtain better performance. Miyato et al. [22] took a different angle to regularize the spectra of the weight matrix which can implicitly constrain the Lipschitz constant. Among these empirical techniques in training GANs, some researchers focus on building the theoretical framework to analyze the properties of GANs under different assumptions. Here we focus on the works of studying the generalization properties of GANs. Speci\ufb01cally, Arora et al. [2] argued that the objective function of the original GAN does not provide a theoretical generalization guarantee. As a result, the authors turned into analyzing the generalization gap of their proposed neural network distance. Qi et al. [25] developed a set of theoretical tools to analyze the generalization of GANs under the assumption of Lipschitz continuous of the discriminator. In practical, to meet the assumption, they designed a novel objective function to directly minimize the Lipschitz constant. Membership attacks towards deep learning algorithms. Recently, membership attacks have arisen as a common threat model against machine learning algorithms and attained increasing attraction from the research community [32, 31, 5, 13]. A pioneer work [31] investigated the risk of membership attacks on different machine learning models. Speci\ufb01cally, they developed a shadow training technique to obtain an attack model in the black-box setting (i.e., without knowing the machine learning model structure and parameters). Carlini et al. [5] proposed a metric to measure the vulnerability of deep learning models. In the context of GANs, Hayes et al. [13] studied membership attacks against GANs in both black-box and white-box settings. In this paper, we use the membership attack to assess the information leakage of the dataset used for training GAN models. 2 \f3 Theoretical Analysis In this part, we aim to bridge the gap between the privacy protection and the generalization ability of GANs. At a high level, we prove that a differentially private learning mechanism (i.e. training algorithm) does not over\ufb01t to a certain extent. The core idea is based on the stability-based theory [30]. To this end, we \ufb01rst introduce some basic notations in GANs. Then we present the theoretical analysis to characterize the generalization ability of GANs from the perspective of privacy protection. At last, based on our theoretical insight from privacy protection, we show a new interpretation of the previous theoretical results of uniform convergence [2], as well as the recent efforts on the Bayesian GAN. Preliminaries. The GAN framework consists of a generator and a discriminator. We denote HG and HD as the hypothesis spaces of the generator and discriminator, respectively. In practice, we make use of neural networks to build HG and HD. Formally, we have HG = {g(z; \u03b8g), \u03b8g \u2208Rp} and HD = {d(x; \u03b8d), \u03b8d \u2208Rq}, where g and d are multi-layer convolutional neural networks. \u03b8g and \u03b8d are the corresponding weight parameters. The training of GANs can be seen as playing a min-max game to solve the following optimization problem: min \u03b8g max \u03b8d Ex\u223cpdata[\u03c6(d(x; \u03b8d))] + Ez\u223cpz[\u03c6(1 \u2212d(g(z; \u03b8g); \u03b8d))] (1) The above formulation can be seen as an extended version of the objective function used in the vanilla GAN [11]. According to previous literature [2], we call function \u03c6 the measuring function. Note that setting \u03c6(t) = log(t) leads to the objective function used in the work of the original GAN [11], while the recent WGAN [1] proposed using \u03c6(t) = t. To optimize Equation 1, we need to build two learning mechanisms Ad and Ag. During the training, we alternately run Ad and Ag to seek an equilibrium (\u03b8\u2217 d, \u03b8\u2217 g). Speci\ufb01cally, Ad tries to \ufb01nd \u03b8\u2217 d such that it maximizes the expected loss of the discriminator: U(\u03b8d, \u03b8\u2217 g) = Ex\u223cpdata[\u03c6(d(x; \u03b8d))] + Ez\u223cpz[\u03c6(1 \u2212d(g(z; \u03b8\u2217 g); \u03b8d))] (2) and Ag tries to \ufb01nd \u03b8\u2217 g to minimize the expected loss of the generator: V (\u03b8\u2217 d, \u03b8g) = Ez\u223cpz[\u03c6(1 \u2212d(g(z; \u03b8g); \u03b8\u2217 d))] (3) Empirical loss and generalization. The optimization of Equation 2 and 3 can not be directly solved since the expectation over the distribution of the true data pdata is intractable. Instead, we approximate them with empirical loss on a set of i.i.d. real data samples S = {x1, x2, \u00b7 \u00b7 \u00b7 , xm} and noise vectors Z = {z1, z2, \u00b7 \u00b7 \u00b7 , zm} drawn from pdata and pz, respectively2. We denote the resulted empirical versions of Equation 2 and 3 as \u02c6 U and \u02c6 V . In this empirical setting, the learning mechanisms Ad and Ag turn into the role as the empirical loss optimizers, which are to optimize the empirical loss \u02c6 U and \u02c6 V , i.e., \ufb01nding (\u02c6 \u03b8\u2217 d, \u02c6 \u03b8\u2217 g). To study the generalization property of the learning mechanisms, we need to evaluate the generalization gap between the empirical and expected objective losses. In this paper, we mainly focus on the analysis of Equation 2 since our viewpoint is from the privacy protection and Equation 3 does not explicitly touch the original training data. As a common practice shown in the prior work [11], we analyze Equation 2 when \u03b8\u2217 g is given. Formally, we de\ufb01ne the generalization gap as follows (we take the discriminator loss U as an example): FU(Ad) = E\u03b8d\u223cAd(S)ES\u223cpm data[ \u02c6 U(\u03b8d, \u03b8\u2217 g) \u2212U(\u03b8d, \u03b8\u2217 g)] (4) where S \u223cpm data denotes sampling m training samples from the oracle distribution pdata. In the above equation, we take the expectation with respect to the randomness in the learning mechanism and also in the sampling process similar to the literature [23, 35]. Note that we can in\ufb01nitely sample from pz (e.g. a uniform distribution) and pz is irrelevant to the original training data (i.e. sampling from pz does not induce the leakage of the training dataset). As a result, we omit the second term in the RHS of Equation 2 and focus on studying the \ufb01rst term. Privacy protection and generalization bound. To bridge the gap between privacy protection and the generalization bound, we need to characterize how an algorithm can protect privacy, i.e., the amount of information leakage of the training dataset. Differential privacy [9] is seen as a gold standard for privacy protection in the security community. It provides a rigorous bound on privacy 2In practice, we always sample different Z while S is \ufb01xed at each training round. 3 \fcost of the algorithm, even in the worst case. The de\ufb01nition of differential privacy is based on the adjacent datasets. Two datasets are adjacent when they differ in a single element. Then we can introduce differential privacy as follows: De\ufb01ne 1 (Differential privacy) A randomized algorithm A : D \u2192R satis\ufb01es \u03f5-differential privacy if for any two adjacent datasets S, S \u2032 \u2286D and for any subset of outputs O \u2286R it holds: Pr[A(S) \u2208O] \u2264e\u03f5Pr[A(S \u2032) \u2208O] (5) In our setting, A can be the training algorithm (i.e. Ad). Intuitively, Equation 5 indicates that participation of one individual sample in the training phase has a negligible effect on the \ufb01nal weight parameters. A relevant concept is uniform RO-stable of an algorithm. An algorithm is stable if a small change to the input causes a limited change in the output. Here, RO denotes \u201creplace one element in the input\u201c. The above description is made formally as: De\ufb01ne 2 (Uniform RO-stability) The randomized algorithm A is uniform RO-stable with respect to the discriminator loss function (Equation 2) in our case, if for all adjacent datasets S, S\u2032, it holds that: sup x\u2208S |E\u03b8d\u223cA(S)[\u03c6(d(x; \u03b8d))] \u2212E\u03b8d\u223cA(S\u2032)[\u03c6(d(x; \u03b8d))]| \u2264\u03f5stable(m) (6) A well-known heuristic observation is that differential privacy implies uniform stability. The prior work [35] has formlized this observation into the following lemma: Lemma 1 (Differential privacy \u21d2uniform RO-stability) If a randomized algorithm A is \u03f5differentially private, then the algorithm A satis\ufb01es (e\u03f5 \u22121)-RO-stability. The stability of the algorithm is also related to the generalization gap. Numerous studies [30, 23] focus on exploring the relationship in various settings. Formally, we have the following lemma: Lemma 2 If an algorithm A is uniform RO-stable with rate \u03f5stable(m), then |FU(A)| (Equation 4) can be bounded: |FU(A)| \u2264\u03f5stable(m). Intuitively, the more stable the algorithm is, the better its generalization ability will be. We take a further step to build the connection between differential privacy and the generalization gap. This can be done via combining the above two lemmas. Formally, we introduce Theorem 3 as follows: Theorem 1 (Generalization gap) If an algorithm A satis\ufb01es \u03f5-differential privacy, then the generalization gap can be bounded by a data-independent constant. The proof can be accomplished by following the roadmap: Differential privacy \u21d2Stability \u21d2 Generalization. The proof details can be found in Appendix. By applying Theorem 3 to Ad, we can show that the generalizability of the discriminator is ensured when the training algorithm satis\ufb01es \u03f5-differential privacy. Note that we focus on the generalization of the discriminator loss, since the optimization of Equation 3 does not touch the original data. We can easily obtain the similar generalization bound of the generator by leveraging the post-processing property of the differential privacy protocol [10]. Theorem 3 not only enables characterizing GAN\u2019s generalizability from the viewpoint of privacy protection, but also helps to understand previous techniques for improving the generalization of GANs. A typical example is the Lipschitz regularization technique. Previous studies propose implementing the regularization from various angles [1, 25, 22]. For instance, the loss-sensitive GAN [25] designed a novel objective loss to restrict the discriminator to satisfy the Lipschitz condition. Spectral normalization [22] explored this direction by adding regularization on the weight parameters. And WGAN [1] proposed using gradient penalty to constrain the magnitude of the gradient, which further implicitly led to the Lipschitz condition. From the perspective of differential privacy, the Lipschitz condition for the outputs is also a crucial ingredient for building a differentially private algorithm. We also infer that adding Lipschitz constraints implicitly leads to stability of the algorithm, which can further be used for reducing the generalization gap (see more details in the evaluation section). 4 \fAbove analysis focuses on the discriminator loss U. As mentioned above, it is natural to extend the analysis to the generator loss since the optimization of the generator loss does not touch the original training dataset and we can leverage the post-processing property of the differential privacy. We can further study the whole optimization procedure of the GAN, i.e., the alternative running of Ad and Ag. These can be accomplished by the composition theory in adaptive learning theory [7], we mark this as the future work. Revisiting previous results on uniform convergence. Previous works attempted to explore the uniform convergence with respect to different objective functions. For example, Qi et al. [25] proposed the loss-sensitive GAN and proved the uniform convergence with respect to the discriminator loss. Arora et al. proposed the neural distance to analyze the generalization property (uniform convergence) of GANs. Note that both of them focus on the special form of the GAN or the objective function. In this paper, based on the aforementioned theoretical results, we can prove the uniform convergence of the discriminator loss when the training algorithm satis\ufb01es the differential privacy protocol. Formally, we have the following theorem: Theorem 2 (Uniform convergence): Suppose Ad satis\ufb01es \u03f5-differential privacy and d(k)(x; \u03b8(k) d ) be the output of Ad at the k-th iteration. Then, \u2200k, the generalization gap with respect to dk can be bounded by a universal constant which is related to \u03f5. The proof of the above theorem can be done via combing the post-processing property [10] of differential privacy and McDiarmid\u2019s inequality [33]. The details can be found in Appendix. Connection to Bayesian GAN. Recently, training GANs using Bayesian learning has emerged as a new way to avoid mode collapse [27]. A well-known interpretation of mode collapse is that the generator/discriminator has memorized some examples from the training dataset. Hence, the memorization phenomenon can also breach the privacy of individuals in the training dataset. Thus, we infer that the effectiveness of the Bayesian GAN may come from preventing information leakage of the training dataset. In what follows, we brie\ufb02y introduce how we validate this conjecture with our theoretical results. Speci\ufb01cally, we take a recent work [27] as an example. In the work, the authors proposed using stochastic Hamiltonian Monte Carlo (HMC) to marginalize the posterior distribution of the weight parameters of both generator and discriminator. We have noted that Wang et al. [34] pointed out that sampling one element from a posterior distribution can implicitly satisfy differential privacy. Based on their results, we can prove that the HMC sampling also preserves differential privacy with minor modi\ufb01cation (refers to Section 4 in [34]), and then the Bayesian GAN can implicitly preserve the differential privacy. Thus we are not surprised that the Bayesian GAN can alleviate mode collapse in GANs since its connection to differential privacy mentioned above. 4 Quantitative Analysis In this section, we quantitatively validate the connection between the generalizability and the privacy protection/information leakage of GANs by investigating some of the most popular GAN models. In the theoretical analysis, we focus on the differentially private learning algorithms. While in practice, differential privacy is a strict requirement for most of the existing GANs so that we focus on studying information leakage of GANs instead. In particular, we note that adding Lipschitz constraints on the discriminator has recently emerged as an effective solution for improving the generalization ability of such GANs, thus we aim to study the effects of various regularization techniques for adding the Lipschitz constraints. In a nutshell, our results show that the Lipschitz constraints not only reduce the generalization gap but also make the trained model resistant to the membership attacks which are used for detecting the sensitive information contained in the training dataset. This section is structured as follows. First, we will introduce the experimental settings including the datasets and the choice of different hyper-parameters. Then, we demonstrate our main results on different datasets. At last, we provide some discussions on the attack methods and the regularization techniques. 4.1 Experimental Setup Datasets. We conduct experiments on a face image dataset and a real clinical dataset, namely, Labeled Faces in the Wild (LFW) [19] which consists of 13,233 face images, and the IDC dataset 5 \fwhich is publicly available for invasive ductal carcinoma (IDC) classi\ufb01cation3 and contains 277,524 patches of 50 \u00d7 50 pixels (198,738 IDC-negative and 78,786 IDC-positive). Model setup. Note that we focus on studying the effects of different regularization techniques instead of the architecture design of the GAN model, thus we use the same generator architecture and the same discriminator architecture in all experiments. Speci\ufb01cally, we adopt DCGAN following most of previous works [26]. More details can be found in the work [26]. We set the size of the generated image to 64x64 for the LFW dataset and 32x32 for the IDC dataset, respectively. As for optimization, we use Adam [17] in all experiments, and use different hyper-parameters for different training strategies. To be speci\ufb01c, we make use of Adam for the GAN trained with JS divergence. The learning rates is set to 0.0004 for the GAN trained without any regularization terms (original GAN [11]), while for other GANs (e.g. trained using Wasserstein distance), the learning rate is set to 0.0002. More details of hyper-parameter settings (e.g. \u03b2 in Adam) can be found in Appendix. We trained all the models for 400 epochs on both datasets. Attack setup. We make use of membership attacks for evaluating the information leakage of the GAN model. We build the attack model based on the output of the discriminator, which is a bounded function. We suppose d(x; \u03b8d) \u2264b for all x in Equation 1. This assumption naturally holds in the context of GAN (e.g. b = 1 for the original GAN). This is also a common assumption in many previous works. Here, letting b = 1 suf\ufb01ces to all our cases. We then assume that the attacker A has access to the trained GAN model, i.e. the discriminator and generator. Note that the notation A is different from the previous ones that denote training algorithms. The goal of the attacker is to determine whether a record (an image in our case) in the attack testing dataset is from the original training dataset. Based on the above setting, the attack model proceeds as follows: \u2022 Given the discriminator d(x; \u03b8d) and an image from the attack testing dataset. \u2022 A \ufb01rstly sets a threshold t \u2208(0, 1). \u2022 A outputs 1 if d(x; \u03b8d)/b \u2265t, otherwise, it outputs 0. where the output of 1 indicates that the input image is from the training dataset. To evaluate the performance of the attack model, we need to build the attack testing dataset. For the LFW dataset, we randomly choose 50% of images as the original training dataset to train the GAN model. We build the attack testing dataset by mixing the original training dataset and the remaining images. For the IDC dataset, we only use the positive part of the dataset. Since the data provider has already provided a partition of training and testing datasets [15] (22,383 images in the training dataset and 14,157 images in the testing dataset4), we directly use the original partition and build the attack testing dataset by mixing the training and testing datasets. We then treat the above attack as a binary classi\ufb01cation model and evaluate its performance based on the F1 score and the AUC5 value. To compute the F1 score, we assume the attacker has obtained the average value of d(x; \u03b8d) on the training dataset. Thus we can set the average value as the threshold t and then compute the F1 score at this threshold. 4.2 Results on LFW and IDC Datasets Here, we present the results of our evaluation on the GAN models trained with different strategies. We conduct extensive experiments of different settings. Speci\ufb01cally, we focus on three commonly used techniques for adding Lipschitz constraints, namely, weight clipping [1], gradient penalty [28], and spectral normalization [22]. For weight clipping, we set the clip interval as [\u22120.01, 0.01] following the prior work [1]. We combine these regularization techniques with two types of objective functions, the traditional JS divergence (setting \u03c6(t) = log(t)) and the Wasserstein distance (setting \u03c6(t) = t). As mentioned above, the performance of the attack model is measured by the F1 score and AUC value, and we use the gap between the testing and training losses to estimate the generalization gap. We also calculate the Inception score [28], which is a commonly used for assessing the quality of the generated images in previous works. The overall results are shown in Table 1. 3http://www.andrewjanowczyk.com/use-case-6-invasive-ductal-carcinoma-idc-segmentation/ 4They also provide a validation dataset and we did not use it in our experiments. 5Area under the ROC curve 6 \fTable 1: Evaluation results of DCGAN trained with different strategies. IS denotes the Inception score. N/A indicates that the strategy leads to failure/collapse of the training. The last row presents the Inception scores of the real data (training images of these two datasets). Strategy LFW IDC F1 AUC Gap IS F1 AUC Gap IS -JS divergenceOriginal 0.565 0.729 0.581 3.067 0.445 0.531 0.138 2.148 Weight Clipping 0.486 0.501 0.113 3.112 0.378 0.502 0.053 2.083 Spectral Normalization 0.482 0.506 0.106 3.104 0.416 0.508 0.124 2.207 Gradient Penalty N/A N/A -WassersteinW/o clipping N/A N/A Weight Clipping 0.484 0.512 0.042 3.013 0.388 0.513 0.045 1.912 Spectral Normalization 0.515 0.505 0.017 3.156 0.415 0.507 0.013 2.196 Gradient Penalty 0.492 0.503 0.031 2.994 0.426 0.504 0.017 1.974 IS (Real data) 4.272 3.061 LFW. We \ufb01rst conduct contrastive experiments on the GAN trained using JS divergence (setting \u03c6(x) = log(x) in Equation 1). As shown in Table 1, the plain model (trained without any regularization term) is more vulnerable than those trained with different regularizers. The plain model leaks some information of the training dataset, which results in the F1 score of 0.565 and the AUC value of 0.729 (greater than 0.5), respectively. While various regularization techniques are used for training the GAN, the attack performance decreases drastically. For instance, while spectral normalization is used, the F1 score is dropped from 0.565 to 0.482, and the AUC value is dropped to 0.506, which approximates to the random guess. Along with the decrease of the attack performance, we also observe a consistent decrease in the generalization gap (Gap in Table 1). For example, the gap is decreased from 0.581 to 0.106 while spectral normalization is used. In addition to these metrics, we also calculate the Inception score of each model. As shown in Table 1, the Inception score of the LFW dataset (shown in the last row of Table 1) is much smaller than the commonly used benchmarks (e.g. Cifar10 [18] and ImageNet [8]). This is because that the Inception score is calculated by the model pre-trained by the images in ImageNet in which the image always contains general objects (such as animals) while the image in LFW always contains one face. We observe that the spectral normalization can generate images with higher visual qualities but obtain lower Inception score than the weight clipping strategy (see generated images in Appendix). Thus, these numerical results of Inception score indicate that Inception score is not suitable for some image generation tasks and we need to design a speci\ufb01c metric for a given task. Among these experiments, we also conduct the same attack experiments on GANs trained using Wasserstein distance and observe similar phenomena as shown in Table 1. IDC. The images in the IDC dataset contain various tissues and all of these tissues have similar shapes and colors. As a result, performing the attack on the IDC dataset is more challenging than on other datasets which consist of some concrete objects (e.g. LFW and Cifar10 [18]). As we can see in Table 1, in all cases, the attack performance on IDC is lower than the performance on LFW in terms of the F1 score and AUC value. We also provide some quantitative analysis to interpret the performance drop in the following subsection. Despite the performance drop, we can still observe the effectiveness of different regularizers in reducing the generalization gap and information leakage. For example, with the use of spectral normalization, AUC drops from 0.531 to 0.508 and F1 score decreases from 0.445 to 0.416. Another notable thing is that training the GAN using Wasserstein distance without weight clipping can lead to the failure of training. This may be caused by the gradient explosion since there is no activation function to compress the output values of the last layer of the discriminator in Wasserstein GAN (in contrast, the original GAN used the sigmoid function for the compression purpose). For the results of Inception score, we can observe an obvious decrease from the LFW dataset to the IDC dataset. This may be caused by the difference of contents contained in the images (tissues in IDC images and faces in LFW images). 7 \fIn summary, all the empirical evidence implies that previous regularizers for adding Lipschitz constraints can not only decrease the generalization gap but also reduce the information leakage of the training dataset in terms of the attack performance. Moreover, we observe that spectral normalization achieves comparable visual quality of the generated images in contrast to the original GAN (see generated images in Appendix). This suggests that we can attempt to use this technique in various applications to trade off the information leakage and the visual quality of the generated images. 4.3 Discussions Attack performance on different datasets. In the experiments, we observe that the attack performance may vary on different datasets. From Table 1, in all cases, the attack performance of IDC is always worse than the performance of LFW. For the original GAN model, the AUC of the attack model on LFW is 0.729 while the one for IDC is 0.531 which shows a 27.2% relative performance drop. We infer that the drop is caused by the higher similarity of the images in the IDC data. Quantitatively, the similarity can be measured by the standard deviations (per channel for an RGB image) of these two datasets. Speci\ufb01cally, the standard deviation of IDC is 0.085 (R-channel) while the value of LFW is 0.249 (R-channel). The IDC dataset has signi\ufb01cantly lower standard deviation than the LFW dataset. This suggests that an individual example in the IDC dataset will be less likely to noticeably impact the \ufb01nal model which further constrains an adversary\u2019s ability. Note that similar evidence has been found in the prior work [31]. Other weight normalization. In the previous experiments, we have employed three regularizers for limiting the discriminator to be Lipschitz continuous. In addition to these regularizers, there are some other approaches to use weight normalization techniques to regularize the model training. The original weight normalization introduced in the work [29] is to normalize the l2 norm of each row vector in a weight matrix. Subsequently, Brock et al. [4] proposed orthonormal regularization on the weight matrix to stabilize the training of GANs. We also conduct attack experiments on the GANs trained with these two normalization methods. In practice, the orthonormal regularization achieves comparable performance in terms of the F1 score of the attack model (0.402 for the IDC dataset), while obtains comparable image quality compared with spectral normalization. In addition, the original weight normalization will lead to training failure (not convergence/mode collapse) in our cases. The failure may be caused by the con\ufb02ict between the weight normalizing and the desire to use as many features as possible as discussed in the prior work [22]. Table 2: The results of black-box attack on the LFW dataset. Strategy F1 AUC Gap Original 0.423 0.549 0.581 Weight Clipping 0.358 0.502 0.113 Spectral Normalization 0.347 0.497 0.106 Black-box attack. The aforementioned analysis is all based on the white-box attack, i.e., the model structure and weights of the discriminator and generator are exposed to the adversary. In this part, we present some discussions on the black-box attack. We consider black-box attacks where the adversary has limited auxiliary knowledge of the dataset follow the prior work [13]. Speci\ufb01cally, for the LFW dataset, we assume the attacker has 30% images of both the training and testing datasets (marked as the auxiliary knowledge in the following part). Then we can use the auxiliary information to build a \"fake\" discriminator to imitate the behavior of the original discriminator (more details can be found in the prior work [13]). Once the \"fake\" discriminator is obtained, we can replace the original discriminator with the fake one to perform the aforementioned white-box attack. The results of the black-box attack are summarized in Table 2. Intuitively, the black-box is a more challenging task than the white-box attack. This can be observed by the decrease of the attack performance from the white-box to black-box (the AUC decreases from 0.729 to 0.549 for the original model). From Table 2, we can also observe the effectiveness of the spectral normalization for reducing the information leakage. 5"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1905.12883v1",
+                    "title": "P3SGD: Patient Privacy Preserving SGD for Regularizing Deep CNNs in Pathological Image Classification",
+                    "abstract": "Recently, deep convolutional neural networks (CNNs) have achieved great\nsuccess in pathological image classification. However, due to the limited\nnumber of labeled pathological images, there are still two challenges to be\naddressed: (1) overfitting: the performance of a CNN model is undermined by the\noverfitting due to its huge amounts of parameters and the insufficiency of\nlabeled training data. (2) privacy leakage: the model trained using a\nconventional method may involuntarily reveal the private information of the\npatients in the training dataset. The smaller the dataset, the worse the\nprivacy leakage. To tackle the above two challenges, we introduce a novel\nstochastic gradient descent (SGD) scheme, named patient privacy preserving SGD\n(P3SGD), which performs the model update of the SGD in the patient level via a\nlarge-step update built upon each patient's data. Specifically, to protect\nprivacy and regularize the CNN model, we propose to inject the well-designed\nnoise into the updates. Moreover, we equip our P3SGD with an elaborated\nstrategy to adaptively control the scale of the injected noise. To validate the\neffectiveness of P3SGD, we perform extensive experiments on a real-world\nclinical dataset and quantitatively demonstrate the superior ability of P3SGD\nin reducing the risk of overfitting. We also provide a rigorous analysis of the\nprivacy cost under differential privacy. Additionally, we find that the models\ntrained with P3SGD are resistant to the model-inversion attack compared with\nthose trained using non-private SGD.",
+                    "authors": "Bingzhe Wu, Shiwan Zhao, Guangyu Sun, Xiaolu Zhang, Zhong Su, Caihong Zeng, Zhihong Liu",
+                    "published": "2019-05-30",
+                    "updated": "2019-05-30",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "main_content": "Introduction In recent years, deep CNNs have emerged as powerful tools for various pathological image analysis tasks, such as \u2217Work performed while interning at IBM Research China. (b) SGD (c) P3SGD (a) Original Figure 1. Illustration of the model-inversion attack on ResNet18. (a) is the original image patch. (b) and (c) are the images reconstructed from the model trained with non-private SGD and P3SGD. tissue classi\ufb01cation [30, 10], lesion detection [17, 22], nuclei segmentation [38, 50, 26], etc. The superior performance of deep CNNs usually relies on large amounts of labeled training data [40]. Unfortunately, the lack of labeled pathological images for some tasks may lead to two notorious issues: (1) over\ufb01tting of the CNN models [39, 42, 48] and (2) privacy leakage [48, 33, 9, 46] of the patients. Firstly, the performance of a CNN-based model is always harmed by the over\ufb01tting due to its large amounts of parameters and the insuf\ufb01ciency of training data. Secondly, pathological datasets usually contain sensitive information, which can be associated with each individual patient. The CNN-based models trained using conventional SGD may involuntarily reveal the private information of patients according to recent studies [48, 9]. For example, Zhang et al. [48] show that the CNN model can easily memorize some samples in the training dataset. Fredrikson et al. [9] propose a model-inversion attack to reconstruct images in the training dataset. In Figure 1 (a) and (b), we demonstrate an attacking example in our task, reconstructing the outline of a patch in the training dataset by leveraging a well-trained CNN model and its intermediate feature representations. There have been numerous studies to solve either of two issues individually. On the one hand, to reduce the risk of over\ufb01tting in deep CNNs, previous research suggests arXiv:1905.12883v1  [cs.CV]  30 May 2019 \fadding appropriate randomness into the training phase [39, 42, 45]. For example, Dropout [39] adds randomness in activation by randomly discarding the hidden layers\u2019 outputs. DropConnect [42] adds randomness in weight parameters by randomly setting weights to zero during training. On the other hand, differential privacy [5, 6] emerges as a strong standard, which offers rigorous privacy guarantees for algorithms applied on the sensitive database. Recent works [1, 31] are introduced to train deep CNN models within differential privacy. The main idea of these works is to perturb the gradient estimation at each step of an SGD algorithm. For example, Abadi et al. [1] use a differentially private additive-noise mechanism on the gradient estimation in an SGD. In addition, a few recent studies [33, 46] have shown that these two seemingly unrelated issues are implicitly relevant based on a natural intuition: \u201creducing the over\ufb01tting\u201d and \u201cprotecting the individual\u2019s privacy\u201d share the same goal of encouraging a CNN model to learn the population\u2019s features instead of memorizing the features of each individual. In this paper, we propose a practical solution to alleviate both issues in a task of pathological image classi\ufb01cation. In particular, we introduce a novel SGD algorithm, named P3SGD, which injects the well-designed noise into the gradient to obtain a degree of differential privacy and reduce over\ufb01tting at the same time. It is worth noting that a pathological database usually consists of a number of patients, each of whom is further associated with a number of image patches. We should protect the privacy in the patient level instead of image level as in most of the previous works. To achieve this goal, we propose to calculate the model update upon individual patient\u2019s data and add carefully-calibrated Gaussian noise to the update for both privacy protection and model regularization. The most similar work to ours is the differentially private federated learning [25, 11], which focuses on protecting the user-level privacy. In contrast to previous works, which use a globally \ufb01xed noise scale to build the noisy update [1, 25, 11], we propose an elaborated strategy to adaptively control the magnitude of the noisy update. In the experiment, we show that this strategy plays a key role in boosting performance of a deep CNN model. At last, we provide a rigorous privacy cost analysis using the moments accountant theorem [1]. In summary, the main contributions of our work are as follows: \u2022 We introduce a practical solution, named P3SGD, to simultaneously address the over\ufb01tting and privacy leaking issues of deep CNNs in the pathological image classi\ufb01cation. To the best of our knowledge, this is the \ufb01rst work to provide rigorous privacy guarantees in medical image analysis tasks. \u2022 Technically, we present a strategy to dynamically control the noisy update at each iterative step, which leads to a signi\ufb01cant performance gain against the state-ofthe-art methods [25, 11]. \u2022 We validate P3SGD on a real-world clinical dataset, which is less explored in previous studies. The results demonstrate that P3SGD is capable of reducing the risk of over\ufb01tting on various CNN architectures. Moreover, P3SGD provides a strong guarantee that the trained model protects the privacy of each patient\u2019s data, even when the attacker holds enough extra sideinformation of the raw training dataset. \u2022 We qualitatively and quantitatively demonstrate that the CNN model trained using P3SGD is resistant to the model-inversion attack [9] (shown in Figure 1 (c)). 2. Related Work Regularization in CNNs In the past years, numerous regularization techniques have been proposed to improve the generalization ability of deep CNNs [20, 39, 42, 45, 41, 12]. These works mainly fall into two categories: explicit regularization and implicit (i.e., algorithmic) regularization. For explicit regularization methods, various penalty terms are used to constrain weight parameters. For example, weight decay [20] uses l2-regularization to constrain the parameters of a CNN model. Another direction is to introduce regularizers to decorrelate convolutional \ufb01lters in deep CNNs [44, 32], which improves the representation ability of the intermediate features extracted by those \ufb01lters. For implicit regularization methods, the core idea is to introduce moderate randomness in the model training phase. For example, Dropout [39] randomly discards the outputs of the hidden neurons in the training phase. However, Dropout is originally designed for fully-connected layers (FC). It is often less effective for convolutional layers, which limits its use in CNNs with few FC layers (e.g., ResNet). This is possibly caused by the fact that Dropout discards features without taking its spatial correlation into account (features from convolutional layers are always spatially correlated) [12]. To address this problem, a few recent works [41, 12] propose to inject structured noise into the features from convolutional layers. One state-of-theart technique, named DropBlock [12], is specially designed for convolutional layers, which randomly drops the features in a sub-region. Both of Dropout and DropBlock inject randomness into activation layers. In contrast, DisturbLabel [45] adds randomness into the loss function by randomly setting a part of labels to be incorrect in a training mini-batch. Data augmentation is another form of algorithmic regularization, which introduces noise into the input layer by randomly transforming training images [36]. Our method can be categorized as implicit regularization. In contrast to previous works, our approach (P3SGD) imposes regularization at the parameter updating phase. \fPrivacy-preserving Deep Learning Meanwhile, there is an increasing concern for privacy leakage in deep learning models, since the training datasets may contain sensitive information. This privacy issue has attracted many research interests on the privacy-preserving deep learning [1, 35, 25, 11, 13, 28]. One promising direction is to build machine learning models within differential privacy [1, 25, 28], which has been widely used in sensitive data analysis as a golden standard of privacy. The early solution is to perturb the model parameters [3, 49] or the objective function [3, 19, 31]. However, such kind of simple solutions cause considerable performance decreasing [4, 43], the situation may become worse in the context of deep learning. Therefore, some recent studies focus on the gradient perturbation based methods [1, 25, 11, 28, 29]. Abadi et al. [1] propose a differentially private version of SGD and present the moments accountant framework to provide tighter privacy bound than previous methods. The PATE framework [28, 29] protects the privacy via transferring knowledge to the student model, from an ensemble of teacher models, which are trained on partitions of the training data. Different from these works, which focus on image-level privacy, we aim to provide patient-level privacy in speci\ufb01c scenarios of pathological image analysis. The most similar works to ours are [25, 11], which extend the private SGD into the federated learning paradigm [24]. However, applying these approaches to the real-world medical image data remains less explored. Moreover, these methods always lead to a performance drop compared with the models trained using non-private SGD. In this paper, we evaluate our method on a real-world pathological image dataset and show that the performance drop can be addressed by carefully controlling the noisy update using our strategy. There are also some studies aiming to explore the relationship between the over\ufb01tting and the privacy leakage issues from the perspective of memorization [33, 46]. In this paper, we present a practical solution to alleviate these two related issues simultaneously. 3. Our Approach In this section, we describe our approach in details and provide a rigorous privacy cost analysis using the moments accountant theorem [1]. 3.1. Preliminaries We \ufb01rstly introduce some basic notations and de\ufb01nitions of differential privacy corresponding to our speci\ufb01c task. In our setting, the pathological image dataset can be regarded as a database D with Np patients. Generally speaking, each patient Di consists of a number of image patches of various tissues, i.e., Di = {(xk, yk)}Ni k=1, where Ni is the number of image patches of the i-th patient. With a slight abuse of notations, we also denote D = SNp i=1 Di as the whole set of images of all patients. Then, a basic concept of image-level adjacent databases can be de\ufb01ned as: two databases are adjacent if they differ in a single imagelabel pair [1]. This concept is widely used for image-level privacy protection. However, such image-level privacy protection is insuf\ufb01cient for our tasks. Instead, we introduce a concept of patient-level adjacent databases de\ufb01ned as follows: De\ufb01ne 1 (Patient-level adjacent databases) D\u2032 and D\u2032\u2032 are adjacent: if D\u2032 can be obtained by adding all images of a single patient to D\u2032\u2032 or removing all images of a single patient from D\u2032\u2032. This de\ufb01nition is inspired by the prior works [25, 11], in which the authors focus on user-level privacy. With the definition of adjacent databases, we can formally de\ufb01ne the patient-level differential privacy as: De\ufb01ne 2 (Differential privacy) A randomized algorithm A : D \u2192R satis\ufb01es (\u03f5, \u03b4)-differential privacy if for any two adjacent databases D\u2032, D \u2032\u2032 \u2286D and for any subset of outputs S \u2286R it holds: Pr[A(D \u2032) \u2208S] \u2264e\u03f5Pr[A(D \u2032\u2032) \u2208S] + \u03b4 (1) The randomized algorithm A is also known as the mechanism in the literature [5]. In our setting, A is the algorithm used to train deep CNNs, e.g., the SGD algorithm. D denotes the training dataset (i.e., D in our case) and R is the parameter space of a deep CNN. Intuitively, the Equation 1 indicates that participation of one individual patient in a training phase has a negligible effect on the \ufb01nal weight parameters. Another concept is the sensitivity of a randomized algorithm: De\ufb01ne 3 (Sensitivity) The sensitivity of a randomized algorithm A is the upper-bound of ||A(D \u2032) \u2212A(D \u2032\u2032)||2, where D \u2032 and D \u2032\u2032 are any adjacent databases (see in De\ufb01ne 1). To establish a randomized algorithm that satis\ufb01es differential privacy, we need to bound its sensitivity. The most used strategy is to clip the norm of the parameter update. In next two subsections, we will introduce the traditional SGD and P3SGD separately, as two instances of the randomized algorithm A. 3.2. Standard SGD Algorithm We start with the standard SGD (i.e., non-private SGD) algorithm for training a deep CNN-based classi\ufb01cation model. The goal of the classi\ufb01cation is to train a CNN model M : \u02c6 y = f(x; \u03b8), where \u02c6 y is the predicted label, and \u03b8 are the model parameters. Training of the model is to \fminimize the empirical loss L(D; \u03b8). In practice, we estimate the gradient of the empirical loss on a mini-batch. We denote the classi\ufb01cation loss over a mini-batch as: L(Bt; \u03b8) = 1 |Bt| X (x,y)\u2208Bt l(f(x; \u03b8), y) (2) Here, l(x, y) is the loss function, e.g., cross-entropy loss. Bt refers to a mini-batch of images which are randomly and independently drawn from the whole image set D. Note that we can add an additional regularization term into Equation 2, such as l2 term. At the t-th step of the SGD algorithm, we can update the current parameter \u03b8t as \u03b8t+1 = \u03b8t \u2212\u03b3t \u00b7 \u2207\u03b8tL(Bt; \u03b8t). 3.3. P3SGD Algorithm Overall, our framework comprises of three components, which are update computation, update sanitization, and privacy accumulation. Our method inherits the computing paradigm of federated learning [24]. Moreover, to protect the privacy, we need to inject well-designed Gaussian noise into each step\u2019s update, which is marked as update sanitization. At last, we can use the moments accountant for privacy accumulation. The pseudo-code is depicted in Algorithm 1. Next, we will describe each of these components in details. For update computation, at the beginning of the t-th step of P3SGD, we randomly sample a patient batch Bt from the database D with a sampling ratio p. Here, the notation Bt is different from the one in Equation 2, where the Bt is sampled from individual images instead of patients. Then, for each patient i in the sampled batch, we perform a back propagation to calculate gradients of the parameters via images of the patient i. After that, we locally update the model using the computed gradients. After we traverse all images of this patient, we can obtain the model update with respect to patient i. This procedure can be interpreted as performing SGD on the local data from patient i. In the next step, we average updates of all patients in Bt to obtain the \ufb01nal update at the t-th step. Note that we need to control the sensitivity of the total update for further update sanitization. In practice, this is implemented by clipping the l2 norm of the update, with respect to each individual patient (as shown in line 26 in Algorithm 1). Cu in Algorithm 1 denotes a prede\ufb01ned upper-bound. Thus, the sensitivity of the total update can be bounded by 2Cu (a proof can be found in supplementary materials). The main idea of update computation is implemented by a function PatientUpdate, as shown in Algorithm 1. To protect privacy, update sanitization needs to be performed. Speci\ufb01cally, we use Gaussian mechanism [7] to inject well-calibrated Gaussian noise into the original update, which leads to a noisy update. The variance of injected Gaussian noise is jointly determined by the upperAlgorithm 1: P3SGD 1 Inputs: 2 Patient database: D, Empirical Loss: L. 3 Patient sampling ratio: p. 4 Noise scale set \u2126z: {zi}Nz i=1. 5 Noise budget \u03f5\u2032 for selecting update per iteration. 6 Bound of update\u2019s norm: Cu. 7 Bound of objective function\u2019s norm: Co. 8 Initialize \u03b80 randomly 9 for t \u2208[T] do 10 Take a subset Bt of patients with sampling ration p 11 for each patient i \u2208Bt do 12 \u2206i t \u2190PatientUpdate(i, \u03b8t) 13 end 14 \u2206t \u2190 1 |Bt|(P i \u2206i t) 15 \u2126\u03c3 = {\u03c3 = zCu/|Bt| : for z in \u2126z} 16 \u2126\u2206= {e \u2206= \u2206t + N(0, (\u03c32I)) : for \u03c3 in \u2126\u03c3} 17 e \u2206t \u2190NoisyUpdateSelect(\u2126\u2206, \u03f5\u2032, Bt, \u03b8t, L) 18 \u03b8t+1 = \u03b8t + e \u2206t 19 end 20 Function PatientUpdate(i, \u03b8t): 21 \u03b8 \u2190\u03b8t 22 for batch image samples b from Patient i do 23 \u03b8 \u2190\u03b8 \u2212\u03b3\u2207L(b; \u03b8) 24 end 25 \u2206i = \u03b8 \u2212\u03b8t 26 \u2206i = ClipNorm(\u2206i, Cu) 27 return \u2206i Algorithm 2: NoisyUpdateSelect 1 Function NoisyUpdateSelect(\u2126, \u03f5, B, \u03b8, L): 2 \u2126u = { u = \u2212Clip(L(B;\u03b8 + \u2206), Co) : for \u2206in \u2126} 3 Select \u2206with probability exp( \u03f5u 2Co ) P u\u2208\u2126u exp( \u03f5u 2Co ) 4 return \u2206 bound Cu of the update\u2019s l2 norm and the noise scale z. In this paper, we use a common strategy to set Cu as a globally \ufb01xed value similar to prior works [1, 25]. Therefore, the choice of a noise scale factor z is critical to train CNN model with high performance. Previous works [1, 25] usually use a \ufb01xed noise scale throughout the training phase. However, the \ufb01xed noise scale factor may lead to the departure of the noisy update from the descent direction or an ignorable regularization effect, because the magnitudes of the updates may vary at different iterative steps. Thus, we argue that the strategy that uses a \ufb01xed noise scale may \fhinder the classi\ufb01cation performance. In this paper, we present an elaborated strategy to adaptively select the noise scale. This strategy is originated from the exponential mechanism [7], which is a commonly used mechanism to build a differentially private version of the Argmax function. In this paper, the Argmax function refers to select the argument which maximizes a speci\ufb01c objective function. In our task, we use the negative loss function as the objective function, and the argument is the noisy update built upon different noise scales from the prede\ufb01ned set \u2126z. We implement this strategy as a function NoisyUpdateSelect depicted in Algorithm 2. The prede\ufb01ned set \u2126z contains Nz noise scale factors. Increasing Nz leads to more subtle control of the noisy update, which further boosts the performance. However, the increase of Nz also results in an increase of computational cost. Precisely, one more noise scale will bring about one more forward computation on all images in Bt. In practice, we \ufb01nd that setting Nz = 2 suf\ufb01ces for our task. Note that setting Nz = 1 degenerates to the method used in [25, 11]. In the experiments, we show this strategy is crucial to boost the performance. For privacy accumulation, the composition theorem can be leveraged to compose the privacy cost at each iterative step. In this paper, we make use of the moments accountant [1], which can obtain tighter bound than previous strong composition theorem [8]. Speci\ufb01cally, the moments accountant is to track a bound of the privacy loss random variable instead of a bound on the original privacy budget. Given a randomized algorithm A, the privacy loss at output o is de\ufb01ned as: c(o; A, aux, D \u2032, D \u2032\u2032) \u225clog Pr[A(aux, D \u2032) = o] Pr[A(aux, D \u2032\u2032) = o] (3) Then, the privacy loss random variable C(A, aux, D \u2032, D \u2032\u2032) is de\ufb01ned by evaluating the privacy loss at the outcome sampled from A(D\u2032) [28]. Here, D \u2032 and D \u2032\u2032 are adjacent. aux denotes the auxiliary information. In our P3SGD algorithm, auxiliary information at step t is the weight parameters \u03b8t\u22121 at the step t \u22121. The algorithm A is also known as the adaptive mechanism in literature [1]. We can then de\ufb01ne the moments accountant as follows: Mc(\u03bb) \u225c max aux,D\u2032,D\u2032\u2032 Mc(\u03bb; aux, D \u2032, D \u2032\u2032) (4) where Mc(\u03bb; aux, D \u2032, D \u2032\u2032) is the moment generating function of the privacy loss random variable, which is calculated as: Mc(\u03bb; aux, D \u2032, D \u2032\u2032) \u225clog E[exp(\u03bbC(A, aux, D \u2032, D \u2032\u2032))] (5) Then, we introduce the composability and the tail bound of moments accountant as: Theorem 1 (Composability) Suppose that a randomized algorithm A consists of a sequence of adaptive mechanisms A1, . . . , Ak where Ai : Qi\u22121 j=1 Rj \u00d7 D \u2032 \u2192Ri. The moments accountant of Ai is denoted as M i c(\u03bb). For any \u03bb: Mc(\u03bb) \u2264 k X i=1 M i c(\u03bb) (6) Theorem 2 (Tail bound) For any \u03f5 \u22650, the algorithm A satis\ufb01es (\u03f5, \u03b4)-differential privacy for \u03b4 = min \u03bb exp(Mc(\u03bb) \u2212\u03bb\u03f5) (7) Theorem 2 indicates that if the moments accountant of a randomized algorithm A is bounded, then A satis\ufb01es (\u03f5, \u03b4)differential privacy. The bound of the moments accountant for our strategy implemented in Algorithm 2 is guaranteed by the following theorem: Theorem 3 Given \u03bb, the moments accountant of Algorithm 2 is bounded by q \u00b7 \u03bb(\u03bb + 1)\u03f52 2 . The proof can be done using the privacy ampli\ufb01cation [18] and the theorem in the prior literature [2]. More details can be found in appendix of this work. Privacy guarantee: In this paper, privacy accumulation is to accumulate the moments accountant\u2019s bound at each step. Note that privacy accumulation needs to be performed at the noisy update selection (line 17 in Algorithm 1) and the model update via noisy update (line 18 in Algorithm 1). For the NoisyUpdateSelect in line 17, we can calculate a bound via Theorem 3. For the model update in line 18, the bound is obtained based on the property of Gaussian Mechanism (Lemma 3 in appendix of [1]). Once we bound the moments accountant at each iterative step, we can compose these bounds using Theorem 1. At last, the total privacy cost is obtained based on Theorem 2. It suf\ufb01ces to compute the Mc(\u03bb) when \u03bb \u226432. In practice, we use a \ufb01nite set {1, \u00b7 \u00b7 \u00b7 , 32} following prior work [1]. 4. Experimental Results 4.1. Experimental Settings In this section, we verify the effectiveness of P3SGD on a real-world clinical dataset. This dataset is collected by the doctors in our team. The dataset consists of 1216 patients and each patient contains around 50 image patches. The task we consider in this paper is glomerulus classi\ufb01cation, which aims to classify whether an image patch contains a glomerulus or not. This task has also been studied in a recent work [10]. We ask the doctors to manually label the image patches. For a fair comparison, we set the \fModel Type # Params SGD SGD+Dropout P3SGD Training Testing Gap Training Testing Gap Training Testing Gap AlexNet T 60.9 M 99.87 91.58 8.29 98.97 93.13 5.84 96.85 92.74 4.11 VGG-16 T 14.7 M 99.81 92.19 7.81 99.28 94.32 4.96 96.23 93.87 2.36 ResNet-18 M 11.2 M 99.85 92.25 7.60 99.63 92.12 7.51 95.70 95.23 0.47 ResNet-34 M 21.3 M 99.23 93.19 6.04 99.16 93.22 5.94 95.80 95.34 0.46 MobileNet M 3.2 M 98.73 92.01 6.72 98.65 91.61 7.04 94.79 94.13 0.66 MobileNet v2 M 2.3 M 98.52 93.24 5.28 98.37 93.28 5.09 95.32 94.86 0.46 Table 1. Training and testing accuracies (%) of various network architectures trained with different strategies. The gap between training and testing accuracies is used for measuring the over\ufb01tting of the CNN models.The type T/M denotes traditional/modern CNNs. weight decay to 1e-4 and use data augmentation in all experiments. Speci\ufb01cally, we perform data augmentation by (1) randomly \ufb02ipping input images vertically and/or horizontally, and (2) performing random color jittering, including changing the brightness and saturation of input images. All input images are resized into 224\u00d7224, and pixel intensity values are normalized into [0, 1]. All 1216 patients in the dataset are randomly split into a training dataset (1000 patients) and a testing dataset (216 patients). 4.2. Classi\ufb01cation Evaluation To validate the superiority of P3SGD in reducing over\ufb01tting, we compare it with the standard SGD (without Dropout). We also provide comparisons with the strategy that combines the standard SGD with Dropout. As a result, there are three training strategies: SGD, SGD+Dropout, and P3SGD. We \ufb01rst evaluate our method on the ResNet-18 architecture [14]. For the standard SGD with Dropout, we insert Dropout between convolutional layers and set the drop ratio to 0.3 following [47]. To provide a reasonable weight initialization, we \ufb01rstly pre-train the CNN model on a publicly available pathological image dataset1. The pre-training does not take an extra privacy cost, since we do not interact with the original training dataset in this stage. The pretraining can also help us to determine the hyper-parameters in Algorithm 1. For P3SGD, we set the total updating rounds T to 100 and set the noise scale \u03f52 to 0.1 for selecting noisy update. The sampling ratio p is set to 0.1 and \u2126z is set to be {3.0, 1.0}. The Cu and Co are set to 5.0 and 3.0, respectively. To facilitate the discussion, we denote SGD and P3SGD as the models trained using SGD (without Dropout) and P3SGD (we use the abbreviations in the following discussions). From the results of ResNet-18 (Table 1), we observe that SGD obviously over\ufb01ts (it even reaches nearly 100% training accuracy). In contrast, P3SGD drastically decreases the gap between training and testing accuracies and improves the testing accuracy. In particular, P3SGD outperforms SGD by 2.98% in the testing accuracy (a 38.5% rela1http://www.andrewjanowczyk.com/use-case-4-lymphocyte-detection/ 1 10 20 30 40 50 60 70 80 90 100 Number of Used Training Patches (105) 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 Cross-Entropy Loss SGD, training P3SGD, training SGD, testing P3SGD, testing Figure 2. The training and testing loss curves of ResNet-18 with different training strategies. P3SGD signi\ufb01cantly reduces the over\ufb01tting compared to SGD. tive drop in classi\ufb01cation error), while the gap is decreased from 7.60% to 0.47%, which shows a 93.8% relative improvement. These results indicate that P3SGD signi\ufb01cantly reduces the over\ufb01tting in the ResNet-18 model compared with the standard SGD. We also plot the loss curves of ResNet-18 in Figure 2, which further demonstrates the regularization effect of P3SGD. Besides, there is no signi\ufb01cant performance improvement when we apply Dropout on ResNet-18. Dropout even leads to a slight decrease (from 92.25% to 92.12%) in testing accuracy. We will discuss this phenomenon in details in Section 4.4. Besides the ResNet-18, we also conduct extensive experiments on other popular CNN architectures. In general, we mainly test on two types of CNN models, namely, traditional CNNs and modern CNNs 2 (denoted by T and M in Table 1). Speci\ufb01cally, six architectures are included: AlexNet [21], VGG-16 [37], ResNet-18 [14], ResNet34 [14], MobileNet [15], and MobileNet v2 [34]. For traditional CNNs (e.g., AlexNet), we insert Dropout between fully connected (FC) layers and set the drop ratio to 0.5 following [39]. The results are summarized in Table 1. On the one hand, our method consistently boosts the testing accuracy over the standard SGD (without Dropout) on various 2The modern CNN consists of convolutional layers except the \ufb01nal prediction layer, which comprises of a global average pooling and a fullyconnected layer. \fCNN architectures. The ResNet-34 trained with P3SGD achieves the highest testing accuracy at 95.34% among all network architectures and training strategies. In particular, P3SGD outperforms Dropout technique on all modern CNNs, e.g., the testing accuracy gain is 2.12% in the case of ResNet-34. On the other hand, the training accuracy is suppressed when we use P3SGD to train the CNN model, which further leads to a decrease of the gap between training and testing accuracy. Despite the superiority of our method, we observe that Dropout is usually more effective than P3SGD on the traditional CNNs, e.g., it obtains a slight accuracy gain of 0.45% in the case of VGG-16 compared to P3SGD. We provide some interpretations in the discussion part. We also notice that, under the standard SGD (without Dropout) training strategy, the modern CNNs have less over\ufb01tting (measured as the gap between training and testing accuracies) than traditional CNNs. This may be caused by the regularization effect brought by the Batch Normalization [16] which exists in the modern CNNs. 4.3. Privacy Cost Analysis Another advantage of P3SGD is to provide patient-level privacy within differential privacy. The differentially private degree is measured by (\u03f5, \u03b4) (i.e, privacy cost) in Equation 1. In this part, we calculate the total spend of privacy cost via the moments accountant theorem. The target \u03b4 is \ufb01xed to 1 |Np|1.1 (Np is the number of patients in the training set), which is suggested by the previous literature [5]. In our task, the \u03b4 is around 5e \u22124 (Np = 1000). To verify the effectiveness of our proposed strategy for dynamically controlling the noisy update, we compare it with the strategy of \ufb01xed noise scale (marked by \u0017 in Table 2) which is adopted by the state-of-the-art works [11, 25]. For simplicity, we use adaptive and fixed to denote these two strategies. All the experiments are performed on ResNet-18. We test on various noise scale sets \u2126z to show how the noise scale affects the performance. We \ufb01nd that the noise scale greater than 3.0 leads to unstable training. In practice, we build \u2126z using the noise scale from {1.0, 2.0, 3.0}. Overall, P3SGD with the adaptive strategy (\u2126z = {3.0, 1.0}) achieves the best testing accuracy of 95.23% at a privacy cost of 6.97. For the fixed strategy, a larger noise scale leads to a lower privacy cost, however, it may cause the noisy update deviating from the decent direction and further hinders the testing accuracy. For example, setting \u2126z to {3.0} leads to the lowest privacy cost of 4.70 and the worst accuracy of 92.15%, while setting \u2126z = {1.0} achieves a better accuracy of 94.38% but a much higher privacy cost of 8.48. The adaptive strategy provides a reasonable solution for this dilemma of the fixed strategy. Adaptive \u2126z Testing \u03f5 \u0013 {3.0, 1.0} 95.23 6.97 \u0013 {2.0, 1.0} 94.31 7.10 \u0013 {3.0, 2.0} 93.57 4.97 \u0017 {1.0} 94.38 8.48 \u0017 {2.0} 93.24 5.13 \u0017 {3.0} 92.15 4.70 Table 2. Testing accuracy (%) and privacy cost of different settings on ResNet-18. The Adaptive column indicates if an adaptive noise scale set (marked by a \u0013) or a \ufb01xed scale (marked by a \u0017) is used. Setting a \ufb01xed noise scale leads to the method in [11, 25] Strategy Training Testing Gap SGD+Dropout 99.63 92.12 7.51 SGD+DropBlock 98.85 94.87 3.68 P3SGD 95.70 95.23 0.47 Table 3. Training and testing accuracies (%) on ResNet-18 with different regularization strategies. In general, the adaptive strategy leads to a better trade off between the privacy cost and the testing accuracy. Speci\ufb01cally, extending the \ufb01xed scale {z1} or {z2} to {z1, z2} achieves the testing accuracy higher than or approaching to the best testing accuracy among the corresponding fixed strategies, while with a reasonable privacy cost. For instance, the adaptive strategy with {3.0, 1.0} achieves the accuracy of 95.23%, which is higher than the fixed strategy with either {3.0} or {1.0}. Our strategy also outperforms a naive solution by setting the noise scale to the average of 1.0 and 3.0 (i.e., {2.0}). There is even an accuracy gain of 0.85% by extending {1.0} to {3.0, 1.0}. We infer this accuracy gain comes from the stronger regularization effect brought by the larger noise scale. Meanwhile, the adaptive strategy ({3.0, 1.0}) achieves a moderate privacy cost between the costs obtained by the corresponding fixed strategies (setting \u2126z to {1.0} or {3.0}). To conclude, our proposed strategy can be seen as a simpli\ufb01ed version of line search in numerical optimization [27], and provides a more careful way to control the magnitude of the added noise. The effectiveness of our strategy comes from the \ufb01ne-grained way to control the noisy update. 4.4. Discussions In this subsection, we \ufb01rst analyze the performance of different types of CNNs. We then compare P3SGD with the state-of-the-art regularization mechanism. Finally, we show that the model trained with P3SGD is resistant to a model-inversion attack. Network Architecture. As shown in Table 1, Dropout and our method P3SGD demonstrate totally different effects on the two types of CNN architectures (traditional CNNs and modern CNNs). Speci\ufb01cally, our method outperforms Dropout on modern CNNs, instead, Dropout is more ef\fTraining set SGD P3SGD Original 27.75 27.69 27.75 27.78 27.86 27.88 G G G 27.84 27.85 27.86 27.82 27.80 27.80 SGD P3SGD Original N N N Testing set 27.77 27.73 27.80 27.79 27.80 27.79 SGD P3SGD Original G G N Figure 3. Visualization of the model-inversion attack. G/N below each original patch denotes if the patch contains a glomerulus or not. The number below each reconstructed image is the PSNR value. The reconstructed examples of the training dataset are demonstrated in the left part. For comparison, we also show some examples from the testing dataset in the right part. fective on the traditional CNN architectures. This may be caused by following reasons: (1) Dropout is originally designed for the FC layers due to its huge numbers of parameters (e.g., around 90% parameters of VGG-16 are from the FC layers). However, there is only one FC layer with a few parameters in modern CNN architectures. (2) The cooperation of Dropout and Batch Normalization can be problematic [16]. As we know, batch normalization layer widely exists in modern CNNs (e.g, ResNet [14]). (3) Dropout discards features randomly, however, the features extracted by convolutional layers are always spatially correlated, which impedes the use of Dropout on convolutional layers. Some recent works propose to modify Dropout for convolutional layers. We compare our method with a variant of Dropout in the next part. Other Regularization Techniques. From the previous discussion, some advanced forms of Dropout should be adopted in the modern CNN. In this part, we compare our method with a recent technique, named DropBlock [12], on ResNet-18. For a fair comparison with Dropout, we insert DropBlock between every two convolutional layers and set the drop ratio to 0.3 following [12]. The results are shown in Table 3. DropBlock achieves a testing accuracy gain of 2.75% against Dropout, while P3SGD outperforms both Dropout and DropBlock. In contrast to P3SGD, DropBlock has no suppression effect on the training accuracy. We guess that the performance gain of DropBlock comes from the effect of the implicit model ensemble. We further combine P3SGD with DropBlock but do not obtain obvious accuracy boost. Model-inversion Attack. To demonstrate that P3SGD is resistant to the model-inversion attack [24, 23], we perform an inversion attack on CNN models trained with different strategies. As a case study, we conduct experiments on the ResNet-18 and use the output features from the 3-th residual block to reconstruct the input image (see details in the appendix). Some visualizations are shown in Figure 3. We can reconstruct the outline of the tissue in the input image using the features from the SGD. In contrast, we can not obtain any valuable information from P3SGD (i.e., the model is oblivious to training samples). It indicates that SGD is more vulnerable than P3SGD. Quantitatively, we perform attack on all the training images and report the average PSNR values as: 27.82 for P3SGD and 27.84 for SGD. We also conduct the same study on patches from the testing dataset and show some examples in the left part in Figure 3. The results show that it is hard to reconstruct the input image for both SGD and P3SGD, since the testing examples are not touched by the model in the training phase. This provides some cues for the memorization ability of CNNs [48]. 5."
+                },
+                {
+                    "url": "http://arxiv.org/abs/1807.03136v3",
+                    "title": "G2C: A Generator-to-Classifier Framework Integrating Multi-Stained Visual Cues for Pathological Glomerulus Classification",
+                    "abstract": "Pathological glomerulus classification plays a key role in the diagnosis of\nnephropathy. As the difference between different subcategories is subtle,\ndoctors often refer to slides from different staining methods to make\ndecisions. However, creating correspondence across various stains is\nlabor-intensive, bringing major difficulties in collecting data and training a\nvision-based algorithm to assist nephropathy diagnosis. This paper provides an\nalternative solution for integrating multi-stained visual cues for glomerulus\nclassification. Our approach, named generator-to-classifier (G2C), is a\ntwo-stage framework. Given an input image from a specified stain, several\ngenerators are first applied to estimate its appearances in other staining\nmethods, and a classifier follows to combine visual cues from different stains\nfor prediction (whether it is pathological, or which type of pathology it has).\nWe optimize these two stages in a joint manner. To provide a reasonable\ninitialization, we pre-train the generators in an unlabeled reference set under\nan unpaired image-to-image translation task, and then fine-tune them together\nwith the classifier. We conduct experiments on a glomerulus type classification\ndataset collected by ourselves (there are no publicly available datasets for\nthis purpose). Although joint optimization slightly harms the authenticity of\nthe generated patches, it boosts classification performance, suggesting more\neffective visual cues are extracted in an automatic way. We also transfer our\nmodel to a public dataset for breast cancer classification, and outperform the\nstate-of-the-arts significantly.",
+                    "authors": "Bingzhe Wu, Xiaolu Zhang, Shiwan Zhao, Lingxi Xie, Caihong Zeng, Zhihong Liu, Guangyu Sun",
+                    "published": "2018-06-30",
+                    "updated": "2019-03-07",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "main_content": "Introduction More than 10% people all over the world suffer nephropathy (Levin et al. 2017). An important way of diagnosis lies in a quantitative analysis of glomeruli, e.g., discriminating between normal and abnormal samples, and further diagnosing the abnormality if necessary. In clinics, pathologists generally refer to multiple slides of the same glomerulus, generated by different staining methods, in order to collect cues from particular glomerular structures, elements, or even microorganisms to detect subtle differences among these subcategories. In this work, we consider four staining methods, \u2217This work was done when Bingzhe Wu was a research intern at IBM Research China. Copyright c \u20dd2019, Association for the Advancement of Arti\ufb01cial Intelligence (www.aaai.org). All rights reserved. Slides Patches PAS H&E MASSON PASM Figure 1: Top: slides from four different staining methods (PAS, H&E, MASSON and PASM, respectively). Bottom: four real patches, containing the same glomerulus and sampled from the same position of these slides. namely PAS, H&E, MASSON and PASM. As shown in Figure 1, these staining methods produce quite different appearances even for the same glomerulus. We aim at integrating multi-stained visual cues for glomerulus classi\ufb01cation. The main dif\ufb01culty lies in the lack of annotation, i.e., in both training and testing, labeling every glomerulus across different stains is both labor-intensive and error-prone. In our case, we are provided a partially labeled dataset on one stain (PAS), and unlabeled data on other three stains. For most glomeruli, it is dif\ufb01cult to \ufb01nd their perfect occurrences in all four stains, thus we cannot expect a simple algorithm to learn from correspondence across different stains. This partly limits previous work (Gallego et al. 2018) from training classi\ufb01cation models on multiple stains. To this end, we propose an approach named generatorto-classi\ufb01er (G2C), with the core idea being to generate fake images in other stains to assist classi\ufb01cation in the target stain. G2C has two stages. The \ufb01rst stage contains a few generators, each of which takes an input patch from one stain (e.g., PAS) and estimates its appearance in another arXiv:1807.03136v3  [cs.CV]  7 Mar 2019 \fstain (e.g., H&E, MASSON or PASM). To this end, we use a popular encoder-decoder structure (Zhu et al. 2017) which \ufb01rst down-samples the input patch to extract visual features and then up-samples to estimate its appearance in the target domain. The second stage builds a classi\ufb01er upon all stains, one real and a few generated, and outputs prediction. To alleviate over-\ufb01tting, we share network weights among different branches (each branch deals with one stain), and add a cross-stain attention block after each residual block to adjust neural responses across different stains. G2C is optimized in a joint manner so as to facilitate the collaboration between generation and classi\ufb01cation. However, directly training everything from scratch may lead to dif\ufb01culty in convergence. Therefore, we initialize each generator using CycleGAN (Zhu et al. 2017), an unpaired image-to-image translation algorithm, given weakly labeled training data, and \ufb01ne-tune them with the classi\ufb01er (initialized as random noise). Although this strategy may result in weaker authenticity of generated patches, it indeed enjoys higher classi\ufb01cation accuracy, arguably because more ef\ufb01cient features are fed into the classi\ufb01er. We conduct experiments in two datasets for glomerulus classi\ufb01cation and breast cancer classi\ufb01cation, respectively. The \ufb01rst dataset provides a labeled image corpus in the PAS stain, and unlabeled ones in other three stains (e.g., H&E, MASSON or PASM). We initialize the generators using a small portion of unlabeled data (known as the reference set), and then train G2C in labeled PAS data. G2C brings signi\ufb01cant accuracy gain on glomerulus type classi\ufb01cation, including distinguishing between normal and abnormal data, and discriminating two subtypes of abnormality. The second dataset only contains one (PAS) stain, so we directly start with the \ufb01ne-tuning stage, using the pre-trained generators from glomerulus data. Our approach signi\ufb01cantly outperforms the state-of-the-arts, showing its satisfying transferability across different diseases. We further diagnose G2C with a few comparative studies. First, we individually analyze two stages, verifying that each generator produces high-quality patches (even professional doctors feel dif\ufb01cult in discriminating real and fake patches) and the classi\ufb01er is an ef\ufb01cient solution in fusing visual information from multiple stains. Second, joint optimization over the generators and the classi\ufb01er brings a consistent accuracy gain, verifying the value of coupling information. Third, the generation stage can be explained as an advanced way of data augmentation, which provides more constraints in other domains to alleviate over-\ufb01tting in the target domain. In summary, the major contribution of this paper is to provide an interpretable way of adding supervision from other domains. Compared to the recent work (Shrivastava et al. 2017) which aimed at improving the quality of generated images, our work provides an alternative idea, i.e., optimizing the generator with the target vision. This paper shows an example in classi\ufb01cation, yet it has a potential of being applied to other tasks such as object detection and semantic segmentation. The remainder of this paper is organized as follows. We \ufb01rst brie\ufb02y review related work, and then illustrate the proposed generator-to-classi\ufb01er (G2C) framework as well as the optimization method. After experimental results are shown (for glomerulus classi\ufb01cation and breast cancer classi\ufb01cation), we conclude this work in the \ufb01nal section. Related Work Computer-aided diagnosis (CAD) plays a central role in assisting human doctors for clinical purposes. One of the most important prerequisites of CAD is medical imaging analysis, which is aimed at processing and understanding CT, MRI and ultrasound images in order to diagnose human pathology. In comparison, the digital pathology (DP) provides more accurate imaging in a small region of body tissues. Recent years have witnessed an explosion in this \ufb01eld, which is widely considered as one of the most promising techniques in the diagnosis, prognosis and prediction of cancer and other important diseases. This paper studies glomerulus classi\ufb01cation from DP images. This is a key technique in diagnosing nephropathy, one of the most common types of diseases in the world. In the conventional literatures, people made use of handcrafted features to capture discriminative patterns in digital pathology images. For example, (Kakimoto et al. 2014) applied an SVM on top of the Rectangular Histogramof-Gradients (R-HOG) features for glomerulus detection, and (Cruz-Roa et al. 2014) designed Fuzzy Color Histogram (FCH) features (Han and Ma 2002) to identify subcategories of breast cancer. Recently, the rapid development of deep learning brought more powerful and ef\ufb01cient solutions. Especially, as one of the most important models in deep learning, the convolutional neural networks (Krizhevsky, Sutskever, and Hinton 2012)(Simonyan and Zisserman 2015)(Szegedy et al. 2017) have been applied to a wide range of tasks in medical imaging analysis, including classi\ufb01cation (Gulshan et al. 2016) (Esteva et al. 2017) (Gallego et al. 2018), detection (Dou et al. 2016), segmentation (Ronneberger, Fischer, and Brox 2015), etc. In the \ufb01eld of DP image processing, (Liu et al. 2017) designed an automatic method to detect cancer metastases, which outperformed human doctors. (Chen et al. 2016) proposed a coarse-to-\ufb01ne approach for mitosis detection. In (Janowczyk and Madabhushi 2016), a uni\ufb01ed framework was presented for a series of tasks, including nuclei segmentation and mitosis detection. Our work is closely related to (Pedraza et al. 2017), which trained deep networks to outperform handcrafted features in glomerulus classi\ufb01cation. The importance of using multiple stains for digital pathology diagnosis is emphasized by the doctors in our team. However, annotating data correspondence is dif\ufb01cult and time-consuming, so we turn to the family of Generative Adversarial Networks (GANs) (Goodfellow et al. 2014) to perform image-to-image translation. There are generally two types of translation algorithms, paired (Isola et al. 2017) and unpaired (Kim et al. 2017)(Liu, Breuel, and Kautz 2017)(Yi et al. 2017)(Zhu et al. 2017), which differ from each other in the organization of input data. The former type is often more accurate, while the latter type can be used in the scenario of missing data correspondence, which \ufb01ts the requirement of this work. \f\ud835\udc70\ud835\udfce \ud835\udc70\ud835\udc74 \ud835\udc70\ud835\udfcf Concat \ud835\udc88\ud835\udfcf \ud835\udc88\ud835\udc74 ResBlock ResBlock DeConv DeConv ResBlock ResBlock DeConv DeConv Stem Stem Stem FC AvePool classifier (shared weights) generators (individual weights) original  patch generated  patches ResBlock ResBlock ResBlock Cross-Stain Attention Block ResBlock ResBlock ResBlock Cross-Stain Attention Block Figure 2: The overall generator-to-classi\ufb01er (G2C) framework. The left part illustrates the M generators, and the right part the classi\ufb01er, in which all M + 1 branches share the same weights. When an input patch comes, m other stains are generated, and then combined with the original one for classi\ufb01cation. The entire framework is end-to-end, and all the modules can be optimized in a joint manner. Our Approach Backgrounds Staining is a popular way to highlight the important features of a soft tissue. Each staining method has both advantages and disadvantages (Fogo et al. 2014). For example, the PAS stains glomerular basement membranes, mesangial matrix and tubular basement membranes red (positive), while the PASM colors the same component black, providing a clear contrast between positively and negatively staining structures. Integrating multi-stained information is very important for pathology image analysis, e.g., for clinical purposes. However, in collecting a large dataset for glomerulus classi\ufb01cation, it is dif\ufb01cult to label each glomerulus under all staining methods, because (i) \ufb01nding correspondence between stains is labor-intensive, and (ii) only a small portion of glomeruli can be clearly seen in multiple stains1. Therefore, we set our goal to be glomerulus classi\ufb01cation from single-stained inputs. To be speci\ufb01c, each input patch contains a glomerulus from the PAS stain. Meanwhile, a small corpus of 100 unlabeled patches is also provided for each stain (including the PAS stain and other three stains). These four corpora form the reference set used for initializing cross-stain generators. Formulation Let the input be a patch I sampled from a slide with PAS staining. The goal is to design a model M : t = f(I; \u03b8), where t is the class label, and \u03b8 are the model parameters, e.g., the learnable weights in a convolutional neural network. Recall that our goal is to start with one stain, generate fake images for other stains, and \ufb01nally make prediction based on 1Each slide in digital pathology can be stained only once. Even if a set of neighboring slides containing the same glomerulus are used in various stains, its appearance may not be identical due to the difference in slicing positions. all stains. We formulate the above \ufb02owchart into a joint optimization task, in which a few generators are \ufb01rst used to generate other stains (H&E, MASSON and PASM) from the input PAS stain, and a classi\ufb01er follows to extract features from all these images and outputs the \ufb01nal prediction. Following this, we decompose the function f(\u00b7) into two stages taking charge of generation and classi\ufb01cation, respectively. The overall \ufb02owchart is illustrated in Figure 2. \u2022 The Generation Network The \ufb01rst set of modules, named generators, play the role of generating patches of different staining methods from the input patch I. We denote the generated patch set by I = {I0, I1, . . . , IM}, in which I0 . = I is the source patch, and all other M ones are generated using a parameterized model Im = gm \u0010 I0; \u03b8G m \u0011 , m = 1, 2, . . . , M. Each generator consists of several down-sampling units and the same number of up-sampling units. As the number of residual blocks increases, the classi\ufb01cation accuracy goes up and gradually saturates. In practice, As a tradeoff between accuracy and ef\ufb01ciency, we use 6 resolution-preserved residual blocks, 2 convolutional layers (kernel size is 3, stride is 2) for down-sampling, and 2 deconvolution layers (kernel size is 4, stride is 2) for up-sampling. Following (Zhu et al. 2017), each convolutional layer is followed by an instance normalization layer. \u2022 The Classi\ufb01cation Network The second module is named a classi\ufb01er, which integrates information from all patches (one real and M generated) for classi\ufb01cation. We denote this stage as t = c \u0010 I0, I1, . . . , IM; \u03b8C\u0011 . Conceptually, the parameters \u03b8C need to capture visual properties from all stains. One choice is to train M + 1 sub-networks with parameters \u03b8C 0 , \u03b8C 1 , . . . , \u03b8C 0 , respectively. Suppose the number of parameters for each sub-network is O(L), this strategy would contribute O(ML) parame\fters to the entire model. Another choice would be using the same set of parameters \u03b8C in each branch (this is reasonable as different stains share similar visual features), but allowing several cross-stain attention blocks to swap information among different branches. As we shall detail below, these blocks are often equipped with a small amount of parameters, and, consequently, the number of parameters is reduced to O(ML\u2032 + L), in which O(L\u2032) is the number of parameters of a cross-stain attention block and L\u2032 \u226aL. This reduces the risk of over-\ufb01tting especially for small datasets. Following this idea, the designed classi\ufb01er is a multi-path model consisting of M + 1 branches, each of which is a variant of the deep residual network (He et al. 2016). A stem block (Szegedy et al. 2017) is used to replace the original 7\u00d77 convolutional layer2, followed by a few down-sampling units (3 residual blocks and a stride-2 pooling layer). A cross-stain attention block follows each residual block, in which we follow (Hu, Shen, and Sun 2018) to \ufb01rst downsample neural responses from all stains (squeeze), then pass it throught two fully-connected layers, and multiply it to each channel of the original responses (excitation). Lastly, M +1 feature vectors are concatenated, average-pooled and fully-connected to the \ufb01nal prediction layer. In summary, the overall framework is a composed function of the generator and the classi\ufb01er, i.e., t = f(I; \u03b8) . = c \u25e6g \u0012 I; n \u03b8G m oM m=1 , \u03b8C \u0013 . (1) Note that when M = 0, our model degenerates to that using one single stain for classi\ufb01cation. Sharing parameters over M + 1 branches enables us to fairly compare our model and the baseline at the classi\ufb01cation stage. This idea also originates from the doctors in our team, who suggests that different staining images provide complementary information in diagnosis, but the basic principles to recognize them should remain unchanged. Optimization We hope to jointly optimize Eqn (1) so as to enable the parameters of the generators and the classi\ufb01er to collaborate with each other. But, according to our motivation, the generator should be able to produce some reasonable images corresponding to other staining methods. Therefore, we suggest a two-stage training process, in which we \ufb01rst train the generative networks using some unlabeled data covering different stains, and then \ufb01ne-tune the generator together with the classi\ufb01er towards higher recognition accuracy. \u2022 Initializing the Generators Due to the lack of data correspondence, the generators are initialized by a task known as unpaired image-to-image translation. 100 patches from the source stain and another 100 from the target stain are provided. Note that all these patches are unlabeled, and may even not contain glomeruli. We use a recent approach named CycleGAN (Zhu et al. 2017), which trains a reverse generator, denoted by b gm, to translate the patches generated by gm back to the source 2Experiments show that using the stem block consistently improves classi\ufb01cation accuracy by more than 1%. stain. b gm shares the same structure with gm. We follow the original implementation in setting hyper-parameters. Note that, if additional annotations on the target domains are available, we can use more accurate image-to-image translation algorithms such as (Isola et al. 2017) to initialize our model. In a high-level perspective, this initialization process also eases the training stage by providing midlevel supervision, forcing the generator to produce reasonable patches, and reducing the risk of over-\ufb01tting a very deep network to a limited amount of training data. \u2022 Fine-tuning Generators with the Classi\ufb01er In this stage, we train the classi\ufb01er together with the generators in a fully-supervised manner (each glomerulus is assigned a class label). Our goal is no longer high-quality generation, but accurate classi\ufb01cation. Therefore, the reference set containing multi-stained, unlabeled data is not used, and all the generators b gm, m = 1, . . . , M, are simply discarded. In network training, we use Stochastic Gradient Descent with a momentum of 0.5. We perform a total of 30 epochs. The learning rate is set to be 0.01 at the beginning, and divided by 10 after every 5 epochs. In the \ufb01rst 5 epochs, we freeze all parameters in the generators in order to initialize the classi\ufb01er with \ufb01xed generated samples, so as to improve the stability in training. Note that freezing the generators in all 30 epochs leads to training the generators and the classi\ufb01er individually. We shall show in experiments that joint optimization leads to signi\ufb01cant accuracy gain over all our generator-to-classi\ufb01er models. As we shall see in Table 1, the numbers of training data in different classes may be imbalanced. To prevent a category with fewer training samples from being swallowed by another category, we introduce the focal loss function (Lin et al. 2017) which brings slight but consistent accuracy gain in each individual classi\ufb01cation task. Why Our Approach Works? It remains to discuss why our algorithm works well (see the next section for quantitative results). The key contribution naturally comes from the ability of simulating different staining methods, and enabling jointly optimization so that the classi\ufb01er takes advantage of complementary information. To con\ufb01rm that these information comes from the authenticity of the generators, we perform a user study on the professional doctors in our team, and verify that it is even dif\ufb01cult for them to distinguish the generated patches from real ones. Moreover, we make some comments on another question: as all the information comes from the input image (i.e., either in the baseline or our algorithm, the classi\ufb01er sees the same amount of information), what is brought into the system that leads to the accuracy gain? We explain this to be a guided way of extracting high-quality features. Note that in training a glomerulus classi\ufb01er, the amount of data is most often very limited. Our dataset merely contains 2,650 cases for ss-vs-gs classi\ufb01cation, with less than 1,000 training images in the ss subtype. When a powerful classi\ufb01er, say a very deep neural network, is used, the training data can be explained in a lot of different ways, but most of them do not learn from human knowledge, and thus do not \ufb01t the testing \fss gs noa Total Training 648 2,002 7,000 9,650 Testing 237 618 2,828 3,683 Total 885 2,620 9,828 13,333 Table 1: Number of annotated glomeruli of each subcategory in the PAS stain. The ss and gs categories compose the high-level abnormal category, denoted by s. data very well. Our algorithm, by introducing the knowledge from human doctors that other staining methods are helpful to classi\ufb01cation, forces the model to rely a great deal on multi-stained data. We believe this algorithm to endure fewer risks especially in the scenario of limited data. This is veri\ufb01ed by investigating the over-\ufb01tting issue, shown in the diagnostic part, and transferring our models to another dataset for breast cancer classi\ufb01cation, shown in the last part of our experiments. Last but not least, although our approach can be explained as an advanced way of data augmentation, it introduces a complementary prior (with the help of an unlabeled reference set) to conventional data augmentation (assuming that semantics of a patch remains unchanged when it is \ufb02ipped, cropped, etc.). In experiments, we \ufb01nd that (1) our approach achieves much more accuracy gain than data augmentation, and (2) these two methods can be used together towards the best performance. Experimental Results Dataset and Settings We collect a dataset for glomerulus classi\ufb01cation. As far as we know, there is no public dataset for this purpose (existing ones (Pedraza et al. 2017) worked on glomerulus-vs-nonglomerulus classi\ufb01cation). Our dataset is collected from 209 patients, each of which has several slides from four staining methods, namely PAS, H&E, MASSON and PASM. In all PAS slides, we ask the doctors to manually label a bounding box for each glomerulus, and annotate its subcategory. The doctors annotate with con\ufb01dence, i.e., only those PAS patches containing enough information to make decisions are preserved. The subcategories include global sclerosis (gs), segmental sclerosis (ss), and being normal (none of the above or noa). Global sclerosis and segmental sclerosis are two levels of glomerulosclerosis (denoted by s). Advised by the doctors, we consider two classi\ufb01cation tasks, dealing with s-vs-noa and gs-vs-ss, respectively. The \ufb01rst task is aimed at discriminating abnormal glomeruli from normal ones, and the second task goes one step further to determine abnormality of the abnormal glomeruli. To deal with imbalanced label distribution (see Table 1), we report category-averaged accuracies (Brodersen et al. 2010) in the following experiments. All 209 patients in the dataset are split into a training set (149 patients) and a testing set (60 patients). There are in total 9,650 annotated patches (each contains one glomerulus) in the training set and 3,683 in the testing set. The statistics of all subcategories are provided in Table 1. To initialize the s noa Average PAS ONLY 90.76 90.73 90.74 PAS H&E 92.74 92.36 92.55 PAS MASSON 92.39 91.72 92.06 PAS PASM 93.09 91.83 92.46 PAS ALL 93.68 92.99 93.34 PAS ALL+ 94.15 93.02 93.68 Table 2: Category-wise and averaged classi\ufb01cation accuracies (%) in the s-vs-noa task. PAS ALL+ indicates that cross-stain attention is added for feature re-weighting. generators, we construct a reference set for each of the other stains by randomly cropping 100 patches from the unlabeled data. These patches may not contain a glomerulus, or contain a part of it, but as we shall see later, such weakly-labeled data are enough to train the generative networks. Setting M = 0 leads to our baseline model in which only the PAS stain is used for classi\ufb01cation. We denote it by PAS ONLY. We also provide several competitors, which differ from each other in the type(s) of stains generated to assist classi\ufb01cation. These variants are denoted by PAS H&E, PAS MASSON, PAS PASM and PAS ALL, respectively. Among which, PAS ALL integrates information from all the other three stains, i.e., M = 3. In our dataset, there are much fewer abnormal glomerulus patches than normal ones. To prevent over-\ufb01tting, we perform data augmentation by (i) randomly \ufb02ipping input patches vertically and/or horizontally, and (ii) performing random color jittering, including changing the brightness and saturation of input patches. All input patches are rescaled into 224 \u00d7 224, and pixel intensity values are normalized into [0, 1]. Quantitative Results Level-1 Classi\ufb01cation: s-vs-noa We \ufb01rst evaluate classi\ufb01cation accuracy in discriminating abnormal glomeruli (denoted by s) from normal ones (denoted by noa). Results are summarized in Table 2. One can observe that introducing additional stain(s) consistently improves classi\ufb01cation accuracy. An interesting but expected phenomenon emerges by looking into category-wise accuracies. For example, based on the PAS stain, adding H&E produces a higher classi\ufb01cation rate in the normal (noa) category, while MASSON works better in \ufb01nding abnormal (s) glomeruli. This suggests that different stains provide complementary information to assist diagnosis, and veri\ufb01es the motivation of this work. Therefore, it is not surprising that combining all other stains obtains consistent accuracy gain over other competitors. In particular, the PAS ALL model outperforms the PAS ONLY model by 2.60% in the averaged accuracy, or a 28.08% relative drop in classi\ufb01cation error. Our best model is PAS ALL+, which adds cross-stain attention and further improves classi\ufb01cation accuracy. We will analyze the bene\ufb01t of this module in the diagnostic part. Level-2 Classi\ufb01cation: ss-vs-gs Next, we further categorize the abnormal (s) glomeruli into two subtypes, namely, \fss gs Average PAS ONLY 78.05 96.76 87.41 PAS H&E 79.23 96.87 88.05 PAS MASSON 78.31 96.79 87.55 PAS PASM 81.43 97.57 89.50 PAS ALL 81.59 98.20 89.90 PAS ALL+ 82.23 98.67 90.45 Table 3: Category-wise and averaged classi\ufb01cation accuracies (%) in the ss-vs-gs task. PAS ALL+ indicates that crossstain attention is added for feature re-weighting. investigating the ss-vs-gs classi\ufb01cation task. Advised by the doctors in our team, we only consider those correctly categorized abnormal patches in Level-1. Results are summarized in Table 3. Qualitative analysis gives similar conclusions, i.e., different stains provide complementary information, therefore it is instructive to combine all stains for accurate classi\ufb01cation. It is worth noting that in these two abnormal subcategories, segmental sclerosis (ss) suffers lower accuracy compared to global sclerosis (gs), which is partly caused by the limited amount and imbalance of training data. This is alleviated by incorporating generated patches from other stains as augmented data. Compared to PAS ONLY, the PAS ALL model signi\ufb01cantly improves the ss classi\ufb01cation accuracy by 3.54%, and the overall accuracy by 2.49% (a 19.78% relative drop in classi\ufb01cation error). Similarly, PAS ALL+ bene\ufb01ts from cross-stain attention and goes one step beyond equally weighting all stains (i.e., PAS ALL). We perform statistical tests on the two tasks, reporting that PAS ALL outperforms PAS ONLY signi\ufb01cantly with pvalue of 0.011 and 0.0014 in the s-vs-noa task and ss-vs-gs task, respectively. Discussions This part provides several discussions on our approach. First we observe the performance of two stages (generation and classi\ufb01cation) individually, and then we discuss the bene\ufb01t of joint optimization and how our approach helps to alleviate over-\ufb01tting especially in small datasets. We also show a comparison with conventional data augmentation strategies. \u2022 Qualitative Studies on the Generators To con\ufb01rm the authenticity of the generated patches, we perform a study by asking the doctors in our team to discriminate the generated patches from the real ones. We sample 50 patches from all the generated ones, combine them with 50 real patches, show them one-by-one to the pathologists and record their judgments. The average accuracy over three pathologists is 70.0% (random guess reports 50%), suggesting an acceptable quality to professional doctors. Figure 3 shows several examples in which glomeruli are misclassi\ufb01ed using the PAS stain alone, and rescued by the generated stains. We note that each failure case in PAS can be helped by one or a few other stains. In clinics, these generated patches may also assist doctors in case that a PAS patch does not contain suf\ufb01cient information. \u2022 The Design of Classi\ufb01er PAS +PASM +MASSON +H&E original patches generated patches s vs noa s vs noa ss vs gs ss vs gs task s noa ss gs label Figure 3: Different stains provide complementary information to assist glomerulus classi\ufb01cation. In each row, the original patch is mis-classi\ufb01ed using the PAS stain alone (marked by a cross), but turned into correctness after integrating some of other generated stains (marked by a tick). All these glomeruli are correctly classi\ufb01ed using all four stains (i.e., PAS ALL and PAS ALL+). To reduce the number of parameters, we share parameters among different branches of the classi\ufb01er, which is based on the assumption that visual features extracted from different stains are mostly similar. The rationality of this assumption is veri\ufb01ed by the results in Tables 2 and 3. Moreover, adding cross-stain attention blocks consistently boosts performance of a multi-branch classi\ufb01er, e.g., the overall error drop is 0.34% (5.10% relatively) and 0.55% (5.44% relatively) for two tasks, respectively. Note that this is achieved by adding merely 5.27% more parameters to the classi\ufb01er. We also evaluate the use of cross-stain attention blocks in the scenario of fewer branches, and observe smaller improvement, e.g., on top of PAS ONLY, the overall accuracy gain is 0.30% and 0.09%, respectively. This suggests the existence of cross-domain feature difference, yet a lightweighted module is suf\ufb01cient in dealing with it. \u2022 The Bene\ufb01t of Joint Optimization In addition, joint optimization brings signi\ufb01cant gain in classi\ufb01cation accuracy. In comparison to the model in which the generators and the classi\ufb01er are optimized individually (i.e., the weights of the generators are frozen throughout the \ufb01ne-tuning stage), the jointly optimized models (PAS ALL) achieve 1.10% and 1.54% boosts on s-vs-noa and ss-vsgs classi\ufb01cation, respectively. In particular, the error of the most challenging ss class is reduced from 20.68% to 18.41% (2.27% absolute or 10.98% relative drop). \u2022 The Over-\ufb01tting Issue In the area of medical imaging analysis, recognition accuracy is often limited by the insuf\ufb01ciency of training data. Although being signi\ufb01cantly larger than any publicly available glomerulus datasets, there are less than 1,000 training sam\fTraining Testing Gap PAS ONLY 96.48 87.41 9.07 PAS H&E 94.90 88.05 6.85 PAS MASSON 93.23 87.55 5.68 PAS PASM 95.03 89.50 5.53 PAS ALL 94.87 89.90 4.97 PAS ALL+ 95.07 90.45 4.62 Table 4: For different models on ss-vs-gs classi\ufb01cation, we report training and testing accuracy as well as the gap between them. PAS ALL+ indicates that cross-stain attention is added for feature re-weighting. F1-score Accuracy (%) (Han and Ma 2002) 0.675 78.7 (Cruz-Roa et al. 2014) 0.718 84.23 (Janowczyk et al. 2016) 0.765 84.68 Ours (PAS ALL+) 0.841 88.28 Table 5: Comparison of F1-scores and balanced accuracies on the breast cancer classi\ufb01cation task. [Janowczyk et al., 2016] is the baseline (i.e., PAS ONLY). ples for the ss subcategory. Considerable over-\ufb01tting may arise because the testing set contains some cases which are not covered by the training set. Generating patches in other stains alleviates over-\ufb01tting to some extents. It provides complementary information to geometry-based data augmentation such as \ufb02ip and rotation, as the generators bring in some priors learned from the reference sets (100 \u00d7 3 unlabeled patches from other stains), forcing the training data to be explained in a more reasonable manner. To verify this, we record both training and testing accuracies for each of the \ufb01ve models for ss-vs-gs classi\ufb01cation in Table 4. Using multiple stains leads to a higher testing accuracy but a lower training accuracy, which is the consequence of a stronger constraint (multiple stains need to be explained collaboratively) in training deep neural networks. \u2022 Comparison with data augmentation To compare with conventional data augmentation methods, we conduct a comparative experiment on the ss-vs-gs task. In Table 3, we apply data augmentation to PAS ONLY, while all the other models are trained without data augmentation. We also train the PAS ONLY without data augmentation and get an accuracy at 86.21%. As shown in Table 3, the accuracies of PAS ONLY with data augmentation and PAS ALL+ are 87.41% and 90.45%, respectively. Our method PAS ALL+ improves the accuracy by 4.24% while the data augmentation improves the accuracy by 1.20%, which indicates that our method outperforms conventional data augmentation. Transferring to Breast Cancer Classi\ufb01cation To further demonstrate the effectiveness of our approach, we apply it to a publicly available dataset for invasive ductal carcinoma (IDC) classi\ufb01cation3, which contains 277,524 3http://www.andrewjanowczyk.com/ PAS +PASM +MASSON +H&E original patches generated patches IDC Non-IDC Non-IDC IDC label Figure 4: Different stains provide complementary information to assist breast cancer classi\ufb01cation. In each row, the original patch is mis-classi\ufb01ed using the PAS stain alone (marked by a cross), but turned into correctness after integrating some of other generated stains (marked by a tick). All these patches are correctly classi\ufb01ed using all four stains (i.e., PAS ALL and PAS ALL+). patches of 50 \u00d7 50 pixels (198,738 IDC-negative and 78,786 IDC-positive). To make a fair comparison, we reproduce (Janowczyk and Madabhushi 2016) with the same network architecture (AlexNet) on the PAS stain alone (baseline model). As all patches in this dataset are PAS-stained, we do not train new generators from scratch, but simply transfer the pre-trained ones from our dataset, and \ufb01ne-tune them with the new classi\ufb01er. We apply our best con\ufb01guration learned from the previous task, namely, using all four stains and adding cross-stain attentions. This model is denoted by PAS ALL+. Results are shown in Table 5. In terms of both F1-score and classi\ufb01cation accuracy, our approach signi\ufb01cantly outperforms (Janowczyk and Madabhushi 2016), as well as two previous methods with handcrafted features (Han and Ma 2002) and relatively shallow CNNs (Cruz-Roa et al. 2014). Similarly, we visualize some examples in Figure 4. Hence, we conclude on the effectiveness of our training strategy. The \ufb01rst stage, i.e., initializing the generators, can be performed in a \ufb01xed reference set (e.g., containing glomeruli); when another dataset is available, we can directly move on to the second stage, i.e., \ufb01ne-tuning a new classi\ufb01er with these generators."
+                },
+                {
+                    "url": "http://arxiv.org/abs/1712.05927v2",
+                    "title": "SRPGAN: Perceptual Generative Adversarial Network for Single Image Super Resolution",
+                    "abstract": "Single image super resolution (SISR) is to reconstruct a high resolution\nimage from a single low resolution image. The SISR task has been a very\nattractive research topic over the last two decades. In recent years,\nconvolutional neural network (CNN) based models have achieved great performance\non SISR task. Despite the breakthroughs achieved by using CNN models, there are\nstill some problems remaining unsolved, such as how to recover high frequency\ndetails of high resolution images. Previous CNN based models always use a pixel\nwise loss, such as l2 loss. Although the high resolution images constructed by\nthese models have high peak signal-to-noise ratio (PSNR), they often tend to be\nblurry and lack high-frequency details, especially at a large scaling factor.\nIn this paper, we build a super resolution perceptual generative adversarial\nnetwork (SRPGAN) framework for SISR tasks. In the framework, we propose a\nrobust perceptual loss based on the discriminator of the built SRPGAN model. We\nuse the Charbonnier loss function to build the content loss and combine it with\nthe proposed perceptual loss and the adversarial loss. Compared with other\nstate-of-the-art methods, our method has demonstrated great ability to\nconstruct images with sharp edges and rich details. We also evaluate our method\non different benchmarks and compare it with previous CNN based methods. The\nresults show that our method can achieve much higher structural similarity\nindex (SSIM) scores on most of the benchmarks than the previous state-of-art\nmethods.",
+                    "authors": "Bingzhe Wu, Haodong Duan, Zhichao Liu, Guangyu Sun",
+                    "published": "2017-12-16",
+                    "updated": "2017-12-20",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "main_content": "Introduction Single image super resolution (SISR) is a well de\ufb01ned problem in computer vision area. It tries to reconstruct a high resolution image from a single low resolution image. It has been a very attractive research topic over the last two decades [7] [1] [8]. Since SISR can restore some high frequency details, it has been applied to many practical applications such as medical imaging [25], satellite imaging [27], and face identi\ufb01cation [3], where rich details are greatly desired. In recent years, CNNs have shown powerful ability to learn highly non-linear transformations. Due to their powerful learning ability, the CNN based methods are widely used for SISR tasks and have achieved remarkable progress. Despite those breakthroughs brought by the CNN based methods, there are still some critical problems, which remain largely unsolved, such as how to recover high resolution image with high perceptual quality and high frequency details. A common objective function of previous CNN based methods is the pixel wise loss function between the reconstructed and the ground truth high resolution images. The most commonly used pixel wise loss is l2 loss. A method based on l2 loss is to minimize the mean square loss and maximize the peak signal-to-noise ratio (PSNR), which is a common measure used to evaluate SISR algorithms. Although such a method leads to high performance on PSNR metric, the images always tend to be blurry and over-smoothing [26] [4]. Some recent literatures have pointed out that the pixel wise loss based methods failed to build multimodal distribution [26] [4] [32]. There are two different approaches to solve this issue. One approach is to use a different constructed method, for example PixelCNN [4], to build dependencies between different pixels. The other is using perceptual loss and adversarial learning to generate more realistic images [18]. In this paper, we focus on the latter approach. In this work, we propose a general SISR framework (SRPGAN), which is based on the Image-to-Image model [12]. The start point of the proposed framework is the generative adversarial network (GAN). Unlike some previous methods [18], which uses a classi\ufb01cation network to generate the perceptual loss, we use the features obtained by the discriminator network to build a more robust perceptual loss. We further design the adversarial loss and the content loss to build the \ufb01nal objective function. We also propose to use the Charbonnier loss function as the content loss function, which is different from the previous methods. In 1 arXiv:1712.05927v2  [cs.CV]  20 Dec 2017 \fGenerator Discriminator \u2026 Bicubic  interpolation Encoder Decoder Feature Concatenate Feature Extractor Perceptual Loss  Adversarial Loss Total loss  for G and D Content Loss  Figure 1. The structure of the SISR framework(SRPGAN). Our Framework consists of a generator G (left) and a discriminator D (right). The generator part generates high resolution image from low resolution one. The discriminator takes the generated image and the ground truth image as inputs to extract features of these inputs. Based on these extracted features, the discriminator can build the objective functions for D and G. respect of the network architecture, we propose to replace the batch normalization layer with the instance normalization layer. We evaluate our method on most used benchmarks with a large upscaling factor. Our method outperforms other previous methods with SSIM score on most of benchmarks. Beside the quantitative evaluation, our method has also demonstrated great ability to reconstruct images with rich details and high perceptual quality. The rest of this paper is organized as follows. The related work part summarizes the previous related works brie\ufb02y. The methods section describes the framework details and the proposed individual loss functions. The quantitative evaluation and results visualization can be found in the experiments section. 2. Related Works From the great performance achieved by the deep convolutional neural network at ImageNet challenge, various CNN based methods are applied to the super-resolution problem. SRCNN is the \ufb01rst paper that applies CNN to the single image super-resolution problem [5]. SRCNN is a simple model with three convolutional layers working for feature extraction, non-linear mapping, and image reconstruction. This method learns the end-to-end transformation between low and high resolution images. Based on the results of SRCNN, the authors accelerated the previous model by using an hour-glass shape CNN structure, and achieved a better SR performance and a realtime SR model with the rate higher than 24 fps on a generic CPU [6]. To achieve a real-time model for SR, they replaced the bicubic interpolation part in the previous model with deconvolution layers. After removing bicubic interpolation, this model can learn directly from the low resolution image. FSRCNN model only includes convolution layers and deconvolution layers. The convolution layers share the weights for different upscaling factors. With weight sharing, FSRCNN is able to deal with various scales using a single model. Because of its success at the ImageNet challenge, the deep architecture similar with VGG net was proposed for large receptive \ufb01elds in the VDSR (Very Deep network for Super-Resolution) [14]. The convergence rate is the main limitation of the deeper model. The VDSR tries to use a higher learning rate of 10\u22121 while SRCNN used a learning rate of 10\u22125. Due to the high learning rate, it can be easier to diverge. By using the gradient clipping, the VDSR can be controlled strictly. The VDSR included the concept of residual learning to generate \ufb01nal output results. Due to the property of SISR problem, the output image is quite similar to the input image. With the concept of residual learning, the input image is added to the output of the model before making the \ufb01nal output image. As a result, the model can focus on the detail with high-frequency components. Besides, DRCN(Deeply-Recursive Convolutional Network) based on the VDSR model added recursion connections for weight sharing and model compression with only 5 layers [15]. The LapSRN is the one of the most recent frameworks for SISR problem [17]. The model includes a cascaded framework of feature extraction and image reconstruction parts using laplacian pyramids. And the Charbonnier loss function is used instead of the l2 loss function. By replacing deconvolution layers with sub-pixel convolutional layers, the total computational complexity can be reduced dramatically. The new operator can also generate a cleaner image without checkerboard artifacts. With the ef\ufb01cient operation, ESPCN(Ef\ufb01cient Sub-Pixel Convolutional Neural Network) has achieved signi\ufb01cant x10 speed up which can be applied for SR operation of HD videos on a single GPU [24]. The methods we mentioned above are always based on pixel wise loss, such as l2 loss. Although such a method leads to a high PSNR score, the images constructed by that 2 \falways tend to be blurry and over-smoothing. Some recent works also have point out that pixel wise loss fails to capture multimodal distribution [4]. There are two approaches to solve this issue. One approach is to use different network structures to construct high resolution images. For example, in [4], the authors proposed the PixelCNN to capture the dependencies between different pixels. Another approach is to combine the perceptual loss and GAN model to generate more realistic and sharper image. In this paper, we mainly focus on the latter approach. There are also some recent papers which focus on the perceptual loss. In [13], the authors \ufb01rstly introduced the perceptual loss based on VGG classi\ufb01cation network for the style transformation and super resolution. The SRGAN method combines perceptual loss and adversarial loss for photo-realistic image [18]. They address that the pixel wise loss does not capture the perceptual difference between ground truth images and output images. However, the perceptual loss in these paper is based on the VGG classi\ufb01cation network, such a naive classi\ufb01cation network cannot capture the desired high frequency details in super resolution tasks and will introduce extra computation. To this end, the VGG perceptual loss is not a suitable metric for SISR. In this paper, we try to build a more robust perceptual loss to get higher perceptual quality. 3. Methods 3.1. SISR Framework We build our single image super resolution framework on Image-to-Image model [12]. Our framework consists of an image generator G and a discriminator D. The generator is trying to transform the image in the domain generated by bicubic upsampling to the image in the ground truth high resolution image domain. The discriminator is trying to extract the features of the input high resolution images and the constructed images. Based on the features obtained by the discriminator, we can get the adversarial loss and the perceptual loss. Finally, we combine these two loss functions with the content loss to build the total objective functions of D and G. The overall framework is illustrated in Figure1. Considering a single low-resolution image, we \ufb01rstly upscale it by the speci\ufb01ed factor using bicubic interpolation for further computing. Then, the generator network takes the interpolated image as the input and maps it to a high resolution image. Our \ufb01nal goal is to train a generator network G that can generate high resolution image that is as similar as possible to the ground truth high resolution image. To achieve this, we construct a robust loss using the output and the intermediate features obtained by the discriminator D. Additionally, we design a content loss which can be used for evaluating the similarity between the generated image and the ground truth image. The individual loss functions are described with more details in the following subsection. 3.1.1 Network architecture Our start point is the Image-to-Image model [12], we further tailor the Image-to-Image model for the SISR task. To the best of our knowledge, this is the \ufb01rst paper that attempts to apply the image-to-image model to the SISR task. The generator G is the core of the whole framework. The structure of G illustrated in Figure1(left) has an encoderdecoder shape. We add skip connections following the general shape of the U-Net [22] to combine the local and global information. Speci\ufb01cally, in our generator, skip connection is implemented by concatenating features obtained by layer i and layer n \u2212i, where n is the total number of the convolution layers. The encoder part of G consists of a stack of convolution layers. More Speci\ufb01cally, we use convolution layers with small 3 \u00d7 3 kernels. Following the previous work [18], we use the stride convolution to reduce the image resolution in each encoder layer, instead of the max pooling. Further more, we replace the batch normalization layers with the instance normalization layers [29] to achieve better performance. In Figure2, we show the difference between the convolution blocks in a traditional GAN model and those in our model. We increase the resolution of the input features with transpose convolution layers in the decoder part. For the activation functions, we use the LeakyReLU activation functions in all encoder and decoder layers. We build the patch discriminator network followConv BN ReLU Pool Stride-Conv IN LReLU Traditional Ours Figure 2. Comparison of a conv block in a traditional GAN model and that in our model. Our model replaces the pooling layer with a stride-convolution layer. And we use the instance normalization (i.e. IN), not the batch normalization. ing [12]. Compared with the traditional discriminator, the patch discriminator tries to classify whether each patch in an image is real or fake instead of the whole image. Such a discriminator can restrict the GAN model to focus on the high frequency details. And the existence of the content loss can make sure the correctness of the low frequency part. The detail of the loss functions can be found in the following subsection. In the convolution blocks of the discriminator, we remove the batch normalization layers directly. 3 \fGT Ours LapSRN VDSR Figure 3. Reconstruction results [4\u00d7 upscaling] of Ours and recent state-of-the-art methods(LapSRN and VDSR). We use color boxes to highlight sub regions which contain rich details. We magnify the sub regions in the bellow boxes to show more details. From the sub region images, we can see that our method has stronger ability to recover high frequency details and sharp edges. 3.1.2 Instance normalization Although batch normalization [11] has been proved to be effective on many image classi\ufb01cation tasks, recent works [20] [29] have pointed out that batch normalization will decrease the performance of image generation tasks. In [29], the authors proposed to use the instance normalization instead of batch normalization on the image style transform task. Following this, we replace the batch normalization layers with instance normalization layers to get better performance on SISR tasks. Instance normalization is to apply the normalization on a single image instead of the whole batch of images. To introduce the formulation, we denote x \u2208RT\u00d7C\u00d7W\u00d7H as an input batch which contains T images. Let xtkij denote the tkij \u2212th element, where i and j are the spatial dimensions, k is the input feature channel, and t is the index of the image in the batch. Then the formulation of the instance normalization is given by: ytki j = xtki j \u2212utk q \u03c32 tk +\u03b5 (1) where utk = 1 HW \u2211W l=1 \u2211H m=1 xtklm and \u03c32 tk = 1 HW \u2211W l=1 \u2211H m=1(xtklm \u2212utk)2. We replace batch normalization with instance normalization in every layer of the generator G. The instance batch normalization layer can achieve better performance than batch normalization, and it also can be used for preventing the divergence of the training. 3.2. Loss Functions In this section, we will introduce the formulas of the loss functions. To get the objective loss functions of discriminator and generator, we need to design adversarial loss, content loss and perceptual loss, respectively. We can get these individual loss functions based on the outputs and intermediate features obtained by discriminator D. 3.2.1 Adversarial loss Our generator G tries to learn a mapping from the image z by bicubic interpolation to the ground truth high resolution image y. We design our discriminator D in a conditional GAN fashion. The adversarial loss function of our GAN model can be expressed as below: la(G,D) =Ez,y\u223cpdata(z,y)[logD(z,y)]+ Ez\u223cpdata(z)[log(1\u2212D(z,G(z))] (2) In the training phase, the discriminator D tries to minimize this objective function and the generator G tries to maximize it. Compared with the unconditional GAN, the formulation of the adversarial loss function in the unconditional GAN can be expressed as below: l(G,D) =Ey\u223cpdata(y)[logD(y)]+ Ez\u223cpdata(z)[log(1\u2212D(G(z))] (3) In contrast, the discriminator of unconditional GAN cannot observe the input bicubic image z. The adversarial loss can encourage our generator to generate the solution that resides on the manifold of the ground truth high resolution images by trying to fool the discriminator. 3.2.2 Content Loss The adversarial loss can be helpful to recover the high frequency details. Except for the high frequency part, we also need to design a content loss to ensure the correctness of the low frequency part of the constructed image. The commonly used content loss is the mean square loss. In this paper, we propose to use the Charbonnier loss [2] to achieve 4 \fbetter performance on the SISR task. We denote y as the ground truth high resolution image and G(z) as the constructed image. The Charbonnier loss can be expressed as below: ly(y, \u02c6 y) = Ez,y\u223cpdata(z,y)(\u03c1(y\u2212G(z))) (4) Where \u03c1(x) = \u221a x2 +\u03b52 is the Charbonnier penalty function. To give a comparison, we also try the l1 loss and l2 loss in the experiments. 3.2.3 Perceptual Loss Previous methods based on the pixel-wise loss always generate images that lack high frequency details. Some perceptual loss based on VGG16 network has been proposed to deal with this issue. Instead of using the perceptual loss based on the VGG16 classi\ufb01cation network, we use the intermediate features of the discriminator to build the perceptual loss. We can get a more robust perceptual loss for image super resolution by that. Additionally, we can reduce the computation budget of the perceptual loss through the reuse of the extracted features obtained by the discriminator. To introduce the formula of the perceptual loss, we denote \u03c6i as the feature map computed by the i-th convolution layer(after the activation function layer) within the discriminator. Then, we can de\ufb01ne the perceptual loss as: lp = L \u2211 i=1 Ez,y\u223cpdata(z,y)(||\u03c6i(y)\u2212\u03c6i(G(z))||) (5) In this formula, each term in the equation measures the l1 loss of features extracted by i-th layer of the discriminator D, where the G(z) represents the image constructed by the generator G. 3.2.4 Optimization We use an alternative optimization way to optimize the generator G and discriminator D. Based on the individual loss functions presented above, we can de\ufb01ne the objective functions for discriminator D and generator G. The formulas are de\ufb01ned as : ld = \u2212la(G,D)+\u03bblp (6) lg = la(G,D)+\u03bb1lp +\u03bb2ly (7) In the training phase, we minimize ld with respect to the parameters of discriminator D and minimize lg with respect to the parameters of generator G. To optimize our networks, we alternate between one gradient descent step on D and one step on G. For optimizing solver, we use the ADAM algorithm for both G and D. The details can be found in the experimental section. 4. Experiments 4.1. Training Details For training dataset, we use images from T91, BSDS200 [19] and General100 datasets. In each training batch, we randomly select 64 image patches as the high resolution patches, with each patch in the size of 128 \u00d7 128. We obtain the low resolution patches by downsampling the high resolution patches using the bicubic kernel with speci\ufb01ed downsampling factor. We augment the training data in the following ways: (1) Random Rotation: Randomly rotate the images by 90 or 180 degrees. (2) Brightness adjusting: Randomly adjust the brightness of the images. (3) Saturation adjusting: Randomly adjust the saturation of the images. We pre-process all the images by dividing the image data by 255. Finally, we get about 640 thousand patches for training. We initialize the parameters using \u201dXavier\u201d [9]. We train our model from scratch with ADAM optimizer by setting \u03b21 = 0.9, \u03b22 = 0.99, and \u03b5 = 10\u22128. The learning rate is initialized as 10\u22124 and the learning rate decreased to 10\u22125 while we \ufb01nished 106 iterations. We set the weight term in the loss function as \u03bb = 0.01 in equation(6), \u03bb1 = 1 and \u03bb2 = 1 in equation(7). Our implementation is based on Tensor\ufb02ow. We have trained 3 models for scaling factor of 2, 4, 8 respectively. It takes about 18 hours for training one model on one GTX1080. 4.2. Quantitative Evaluation We evaluate the performance of our method on \ufb01ve benchmarks: SET5, SET14, BSDS100, URBAN100, and MANGA109. The metrics we used are PSNR and SSIM [30]. We compare the proposed method with previous stateof-the-art SISR methods. For scaling factors, we test our model on 2x,4x and 8x. Table1 shows the overall quantitative comparisons for 2x, 4x and 8x. Most of the results of other methods are cited from [17] and [15]. Because our method dose not optimize the mean square loss directly, the PSNR score of our method is much lower than other methods. Other than that, our method tends to generate more realistic images and recover more details than the state-ofthe-art methods as shown in Figure3. On the other hand, our method is competitive on the mean SSIM score which has been shown to correlate with human perception on different benchmarks. Our SRPGAN method performs favorably against existing methods on the most used benchmarks with different scaling factor (2x,4x and 8x). From the results, our method has a poor performance on the MANGA dataset, the main reason is that our training dataset consists of real life images and our GAN model tend to reconstruct realistic images, but the MANGA dataset is a dataset consisting of Japanese comics. For other benchmarks, our methods have obvious improvements on SSIM score. 5 \fGround-truth HR Bicubic VDSR LapSRN ours Ground-truth HR Bicubic VDSR LspSRN ours Figure 4. Reconstruction results of Ours and recent state-of-the-art methods.[4\u00d7 upscaling] Algorithm Scale SET5 SET14 BSDS100 URBAN100 MANGA109 PSNR / SSIM PSNR / SSIM PSNR / SSIM PSNR / SSIM PSNR / SSIM Bicubic 2 33.65 / 0.930 30.34 / 0.870 29.56 / 0.844 26.88 / 0.841 30.84 / 0.935 A+ [28] 2 36.54 / 0.954 32.40 / 0.906 31.22 / 0.887 29.23 / 0.894 35.33 / 0.967 SRCNN [5] 2 36.65 / 0.954 32.29 / 0.903 31.36 / 0.888 29.52 / 0.895 35.72 / 0.968 FSRCNN [6] 2 36.99 / 0.955 32.73 / 0.909 31.51 / 0.891 29.87 / 0.901 36.62 / 0.971 SelfExSR [10] 2 36.49 / 0.954 32.44 / 0.906 31.18 / 0.886 29.54 / 0.897 35.78 / 0.968 RFL [23] 2 36.55 / 0.954 32.36 / 0.905 31.16 / 0.885 29.13 / 0.891 35.08 / 0.966 SCN [31] 2 36.52 / 0.953 32.42 / 0.904 31.24 / 0.884 29.50 / 0.896 35.47 / 0.966 VDSR [14] 2 37.53 / 0.958 32.97 / 0.913 31.90 / 0.896 30.77 / 0.914 37.16 / 0.974 DRCN [16] 2 37.63 / 0.959 32.98 / 0.913 31.85 / 0.894 30.76 / 0.913 37.57 / 0.973 LapSRN [17] 2 37.52 / 0.959 33.08 / 0.922 31.80 / 0.895 30.41 / 0.910 37.27 / 0.974 SRPGAN (Ours) 2 29.67 / 0.950 27.66 / 0.911 27.89 / 0.901 30.41 / 0.892 37.27 / 0.943 Bicubic 4 28.42 / 0.810 26.10 / 0.704 25.96 / 0.669 23.15 / 0.659 24.92 / 0.789 A+ [28] 4 30.30 / 0.859 27.43 / 0.752 26.82 / 0.710 24.34 / 0.720 27.02 / 0.850 SRCNN [5] 4 30.49 / 0.862 27.61 / 0.754 26.91 / 0.712 24.53 / 0.724 27.66 / 0.858 FSRCNN [6] 4 30.71 / 0.865 27.70 / 0.756 26.97 / 0.714 24.61 / 0.727 27.89 / 0.859 SelfExSR [10] 4 30.33 / 0.861 27.54 / 0.756 26.84 / 0.712 24.82 / 0.740 27.82 / 0.865 RFL [23] 4 30.15 / 0.853 27.33 / 0.748 26.75 / 0.707 24.20 / 0.711 26.80 / 0.840 SCN [31] 4 30.39 / 0.862 27.48 / 0.751 26.87 / 0.710 24.52 / 0.725 27.39 / 0.856 VDSR [14] 4 31.35 / 0.882 28.03 / 0.770 27.29 / 0.726 25.18 / 0.753 28.82 / 0.886 DRCN [16] 4 31.53 / 0.884 28.04 / 0.770 27.24 / 0.724 25.14 / 0.752 28.97 / 0.886 LapSRN [17] 4 31.54 / 0.885 28.19 / 0.772 27.32/ 0.728 25.21/ 0.756 29.09 / 0.890 SRPGAN(Ours) 4 22.68 / 0.880 22.50 / 0.786 23.91 / 0.749 20.00 / 0.763 21.00 / 0.860 Bicubic 8 24.39 / 0.657 23.19 / 0.568 23.67 / 0.547 20.74 / 0.515 21.47 / 0.649 A+ [28] 8 25.52 / 0.692 23.98 / 0.597 24.20 / 0.568 21.37 / 0.545 22.39 / 0.680 SRCNN [5] 8 25.33 / 0.689 23.85 / 0.593 24.13 / 0.565 21.29 / 0.543 22.37 / 0.682 FSRCNN [6] 8 25.41 / 0.682 23.93 / 0.592 24.21 / 0.567 21.32 / 0.537 22.39 / 0.672 SelfExSR [10] 8 25.52 / 0.704 24.02 / 0.603 24.18 / 0.568 21.81 / 0.576 22.99 / 0.718 RFL [23] 8 25.36 / 0.677 23.88 / 0.588 24.13 / 0.562 21.27 / 0.535 22.27 / 0.668 SCN [31] 8 25.59 / 0.705 24.11 / 0.605 24.30 / 0.573 21.52 / 0.559 22.68 / 0.700 VDSR [14] 8 25.72 / 0.711 24.21 / 0.609 24.37 / 0.576 21.54 / 0.560 22.83 / 0.707 LapSRN [17] 8 26.14 / 0.738 24.44 / 0.623 24.54 / 0.586 21.81 / 0.581 23.39 / 0.735 SRPGAN(Ours) 8 19.14 / 0.743 19.10 / 0.635 21.55 / 0.613 17.68 / 0.607 18.68 / 0.730 Table 1. Quantitative evaluation of state-of-the-art SR algorithms: average PSNR/SSIM for scale factors 2\u00d7, 4\u00d7 and 8\u00d7. Red text indicates the best SSIM score and blue text indicates the second best performance of SSIM score. 4.3. Visualizations In the context of perceptual quality, our method can recover realistic textures from heavily down-sampling images on the public benchmarks. We have selected some images from the benchmarks to visualization the effective of our method in Figure3 and Figure4. From the results, the images constructed by our method have shown signi\ufb01cant gains in perceptual quality. We have conducted a series experiments to show the ef6 \ffectiveness of our proposed SISR framework and loss functions. In Figure3, we compare our method with previous state-of-the-art methods LapSRN [17] and VDSR [14]. Except for these two methods, we have also compared our method with other CNN based methods. We just list the results of these two methods due to the page limitation. To better visualize the effectiveness of our method, we selected small regions which contain rich details in the images to magnify. As the Figure3 shows, our method successfully reconstructs stripes on Zebra\u2019s bodies (shown in red and orange boxes). On the other hand, LapSRN and VSDR based on a pixel-wise loss just generate blurry images without stripes in that area. In the leg area(yellow and green box), the perceptual quality of our method is not good enough, but our method tries to recover the stripe details on the leg, the other methods just construct leg images without any details. We also compare with other CNN based methods, such as SRCNN [5]. In contrast, our approach surpasses other state-of-the-art methods to generate richer texture details. Further examples of perceptual improvements can be found in Figure4. For these images, we can see that our method has constructed high resolution images with good perceptual quality, those methods which are based on pixelwise loss have generated blurry and over-smooth images. 4.4. Analysis of loss function Except for comparison with other methods, we also evaluate the performance of the generator network without our perceptual loss or replacing the Charbonnier loss with other pixel-wise loss, such as l2 loss. We \ufb01rstly train a model without perceptual loss. The quantitative results are summarized in Table2 and the visualization results are in Figure5. From the Table2, the proposed framework with perceptual loss achieves much better performance than trained without perceptual loss. From Figure5 the model trained with perceptual loss has a better perceptual quality than the model trained without perceptual loss. We can observe from the sub images that our method trained with perceptual loss can accurately reconstruct the beards of the baboon(in red box). Dataset Perceptual loss SSIM Set14 Yes 0.786 Set14 No 0.754 BSDS100 Yes 0.749 BSDS100 No 0.716 Table 2. Quantitative comparison of different perceptual loss strategies. We can get a better performance with our robust perceptual loss. We also compare our perceptual loss with the SRGAN which is based on the VGG perceptual loss (see more details in the follow subsection). Dataset Content Loss SSIM Training epochs Set14 Charbinnier 0.786 100 Set14 l1 0.782 100 Set14 l2 0.763 100 Table 3. Quantitative comparison of different content loss. We have conducted further experiments to explore the content loss. To validate the effect of the Charbonnier loss function, we trained a model with l2 content loss respectively. Through the experiments, the model with l2 content loss requires more training epochs to achieve comparable performance than the model trained with the Charbonnier content loss. The results are shown in Table3. 4.5. Comparison with SRGAN We conduct experiments to compare our method with the SRGAN [18] based on the VGG perceptual loss. We can see that our perceptual loss is more robust than the VGG perceptual loss (see in Figure5). Note that we have trained a SRGAN model using the open source Tensor\ufb02ow code on Github1. For fair comparison, we train the SRGAN model using the same training dataset as our own method. In the training phase, we train that on a scaling factor of 4x for 100 epochs which is the same as our method. As we can see in Figure5, our SRPGAN method does a better job at reconstructing \ufb01ne details, such as the beards of the baboon ( in red boxes of the second and the fourth columns), leading to pleasing visual results. On the other hand, in the training phase, our SRPGAN does not need an extra VGG classi\ufb01cation net to build the perceptual loss, compared with the SRGAN method. This may help to reduce the computation budget while training. 4.6. Limitations While our model is capable of constructing realistic images with sharp edges and rich details on a large scale factor(4x, 8x). There are still some limitations of our methods. One limitation of our GAN based model is that the constructed images have checker board artifacts at the pixel level. The artifacts are visible in Figure5 upon magni\ufb01cation of the sub image regions. This phenomenon is also mentioned in many previous literatures [21]. To solve this issue, one can replace the transpose convolution with resize convolution [21] and sub-pixel convolution [24]. We mark this as a part of future work. 5."
+                }
+            ],
+            "Tian Bian": [
+                {
+                    "url": "http://arxiv.org/abs/2007.05970v1",
+                    "title": "Inverse Graph Identification: Can We Identify Node Labels Given Graph Labels?",
+                    "abstract": "Graph Identification (GI) has long been researched in graph learning and is\nessential in certain applications (e.g. social community detection).\nSpecifically, GI requires to predict the label/score of a target graph given\nits collection of node features and edge connections. While this task is\ncommon, more complex cases arise in practice---we are supposed to do the\ninverse thing by, for example, grouping similar users in a social network given\nthe labels of different communities. This triggers an interesting thought: can\nwe identify nodes given the labels of the graphs they belong to? Therefore,\nthis paper defines a novel problem dubbed Inverse Graph Identification (IGI),\nas opposed to GI. Upon a formal discussion of the variants of IGI, we choose a\nparticular case study of node clustering by making use of the graph labels and\nnode features, with an assistance of a hierarchical graph that further\ncharacterizes the connections between different graphs. To address this task,\nwe propose Gaussian Mixture Graph Convolutional Network (GMGCN), a simple yet\neffective method that makes the node-level message passing process using Graph\nAttention Network (GAT) under the protocol of GI and then infers the category\nof each node via a Gaussian Mixture Layer (GML). The training of GMGCN is\nfurther boosted by a proposed consensus loss to take advantage of the structure\nof the hierarchical graph. Extensive experiments are conducted to test the\nrationality of the formulation of IGI. We verify the superiority of the\nproposed method compared to other baselines on several benchmarks we have built\nup. We will release our codes along with the benchmark data to facilitate more\nresearch attention to the IGI problem.",
+                    "authors": "Tian Bian, Xi Xiao, Tingyang Xu, Yu Rong, Wenbing Huang, Peilin Zhao, Junzhou Huang",
+                    "published": "2020-07-12",
+                    "updated": "2020-07-12",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "stat.ML"
+                    ],
+                    "main_content": "Introduction In many scenarios, the objects with their features are connected by their interactions as graphs. By analyzing these edge connections and node features throughout various graphs, a graph identi\ufb01cation (GI) problem is formed to predict the information or properties of the graphs, such as labels or scores of the target graphs. Many studies have been developed on GI, such as graph classi\ufb01cation [1, 2] and \u2217Co-corresponding authors. Preprint. Under review. arXiv:2007.05970v1  [cs.LG]  12 Jul 2020 \fgraph regression [3]. Formally, methods for GI aggregates the information from nodes and edges to predict or summarize the information of the whole graph. However, it is much more interesting to consider an inverse problem which has never been proposed: Can we use the information of graphs to infer the information of nodes or even edges and sub-graphs? In another word, given the labels of graphs, how to \ufb01gure out the categories of nodes, edges, or sub-graphs? This problem is interesting and very common in the real world. In social media, for example, thinking of hot events that are widely discussed in the form of graphs where users involved as nodes and the co-following relationship among users as edges, it is interesting to think how to pick out the malicious users from the mass of users with only the labels of these event topics such as the authenticity of each topic [4, 5]. In drug discovery, for another example, thinking of molecules as graphs in which atoms as nodes and chemical bonds as edges, we attempt to identify the roles of certain sub-graphs, or equivalently functional groups, in each molecular given its chemical or physical properties [6, 7]. In a programming language, for the last example, thinking of control \ufb02ows for programs as graphs by considering statements as nodes and control \ufb02ows as edges, can we detect the problematic statements if we have already known which program has bug or not [8]? All these problems can be de\ufb01ned as a general problem as Inverse Graph Identi\ufb01cation (IGI) that identi\ufb01es the nodes in graphs based on the information of graphs. The main dif\ufb01culty of the IGI task is that the node labels are inaccessible so that all available information for the training model only comes from the labels of graphs. Namely, it is a node identi\ufb01cation task where the available training labels are much coarser-grained than the node labels we want to \ufb01t. It seems this problem can be resolved from the present perspective of node clustering [9] or node identi\ufb01cation [10]. However, they are different from the problems that we attempt to address. Unlike node clustering which only conducts clustering based on node features or graph structures, IGI is better guaranteed by the given label information of graphs, which, we will demonstrate, closely in\ufb02uences the clustering results. It is also different from the node identi\ufb01cation task because IGI does not contain any node labels for training. Another similar concept to our IGI is Multiple-Instance Learning (MIL) [11] that adopts global labels of bags to identify the labels of local instances. However, MIL assumes that each instance is i.i.d. and there are no edge connections involved. Therefore, it is interesting and important to study the IGI problem as a set of new challenges and seek new solutions for it. In this paper, we formally de\ufb01ne IGI and address out its different problem statements as a set of new challenges. Meanwhile, we focus on a particular study of IGI: A node clustering task by making use of graph labels and node features with an assistance of a hierarchical graph [9] that further characterizes the relations among graphs. To address this particular task, we propose a novel model based on Gaussian mixture model (GMM) and graph convolutional network (GCN) named as Gaussian Mixture Graph Convolutional Network (GMGCN). First, the features of each node are updated through the Graph Attention Network (GAT). Then the node features are aggregated by a Gaussian mixture layer (GML) and a new attention pooling layer proposed in the paper. After obtaining the graph representations, we adopt a hierGCN to classify the graphs. Speci\ufb01cally, we design a consensus loss that plays a key role in the training process. Finally, a node clustering is carried out according to the parameters of GML. The main contributions are as follows: 1. New problem: We introduce a new problem called Inverse Graph Identi\ufb01cation (IGI), which tries to identify the nodes in graphs based on the labels of graphs, and we take a formal discussion of the variants of IGI to attract more research attention on this problem. 2. New solution: We propose an effective model called GMGCN based on GMM and GCN to solve a particular study of IGI problem. To the best of our knowledge, this is the \ufb01rst work to achieve node clustering in graph structure by integrating GMM into GCN. 3. New loss: We propose a consensus loss function to boost the model training through the principle of \"same attraction, opposite repulsion\". Experiments validate that this consensus loss greatly improves the model effect. We validate the proposed GMGCN on various synthetic datasets based on different problem statements and a real-world dataset. The results demonstrate that the proposed method is suitable to solve the IGI task. 2 Related Works Graph Identi\ufb01cation. Recently, there is an increasing interest in the graph learning domain. Among all the existing works, GCN is one of the most effective convolution models. A typical GCN model 2 \fis the message passing neural network (MPNN) proposed by Gilmer te al. [6] which re-generalize several neural network and graph convolutional network approaches as a general \"message-passing\" architecture. Many kinds of GCN [12, 13, 10] actually deliver different message propagation functions for GCN. Among them, Graph Attention Networks (GAT) [14] \ufb01rst leverages learnable self-attentional layers to aggregate weighted neighbors\u2019 information. All of these methods obtain the appropriate node representation for node identi\ufb01cation tasks. But to cope with the GI problem, a compact representation on graph level should be utilized. Therefore, pooling strategies are proposed to integrate information over the node representations [15], such as max/min pooling [13], SortPooling [16], and so on. In addition, Lin et al. [17] propose an learnable attention pooling for weighted averaging of node representations. Although several studies have researched the graph identi\ufb01cation problem, current researches are still lack of studies about Inverse Graph Identi\ufb01cation. Graph Clustering. Graph clustering, a fundamental data analysis task that aims to group similar nodes into the same category, has similar goals as IGI problem. Many real-world applications are cast as graph clustering [2, 18, 19, 20]. The major strategy of graph clustering is to perform simple clustering algorithms such as K-means [21] or GMM [22] on the features embedded from graph embedding [23]. With greatly successful achievements of deep learning, more graph clustering studies have resorted to deep learning to learn embedding that capturing both node features and structural relationships [15]. Researchers employ the stacked sparse autoencoder [24], the variational autoencoder [25], or the combination both autoencoder and GCN [26] to obtain graph representations for clustering. Nevertheless, these graph clustering methods do not perform well on IGI due to the lack of attention to graph labels. Multiple-Instance Learning. Another task that has a similar de\ufb01nition to IGI is Multiple-Instance Learning (MIL). MIL is a variant of inductive machine learning, where each learning example consists of a bag of instances instead of a single feature vector [27]. When obtaining local instances annotations is costly or not possible, but global labels for bags are available, MIL is utilized to train classi\ufb01ers using weakly labeled data. It has received a considerable amount of attention due to both its theoretical interests and its applicability to real-world problems [28, 29, 30, 31]. For example, Wu et al. [32] propose deep MIL which adopts max pooling to \ufb01nd positive instances for image classi\ufb01cation and annotation. Although these MIL methods learn a classi\ufb01er from the labels of bags to achieve the instance classi\ufb01cation, these MIL methods ignore the structural information among instances so that they are not suitable for IGI. Gaussian Mixture Model. In addition, as a core part of the proposed GMGCN model, GMM is a parametric probability density function for a weighted sum of Gaussian component densities [22]. It is commonly used to \ufb01nd underlying clusters in data samples [33]. Generally, the GMM parameters are estimated using the iterative Expectation-Maximization (EM) [34] algorithm from training data. In this paper, we integrate the GMM into GCN to establish a solution for IGI and update the GMM parameters using stochastic gradient descent. 3 Notations and Problem Statement Notations. We denote a set of graph instances as G = {(G(1), y(1)), . . . , (G(N), y(N))} with N graphs, where G(n) refers to the n-th graph instance and y(n) \u2208{0, 1, . . . , C\u22121} is the corresponding graph label of C different categories. We denote by G(n) = (V(n), E(n)) the n-th graph instance of size Mn with nodes v(n) i \u2208V and edges (v(n) i , v(n) j ) \u2208E(n), by X(n) = {x(n) 1 , . . . , x(n) Mn} \u2208 RMn\u00d7d the feature matrix of nodes V(n), and by A(n) \u2208{0, 1}Mn\u00d7Mn the adjacency matrix which associate edge (v(n) i , v(n) j ) with A(n) i,j . We denote by {z(n) 1 , . . . , z(n) Mn} \u2208{0, 1, . . . , C \u22121} the potential labels of nodes V(n), which are invisible to the model training. In addition, we denote by Ahier \u2208{0, 1}N\u00d7N the adjacency matrix of the links among graphs if the graphs in G contain inter-connections. Then, Ahier = I if no relationships among graphs. Problem Statement. The Inverse Graph Identi\ufb01cation problem is de\ufb01ned as: Given a set of graphlabel pairs G and node features X, how to infer each label of the i-th nodes in the n-th graph, z(n) i , where n \u2208{1, . . . , N} and i \u2208{1, . . . , Mn}? Under the de\ufb01nition of IGI, there are some different cases that could be extended from the problem as new tasks. 3 \f2-layer\u00a0Multi-Head\u00a0 GAT GML 2-layer\u00a0Multi-Head\u00a0 GAT 2-layer\u00a0Multi-Head\u00a0 GAT GML GML \u0ddd \ud835\udc9a Node-level\u00a0Representations Node\u00a0Inference Graph-level\u00a0Identifications Attention Pooling Attention Pooling Attention Pooling hierGCN FC Figure 1: The overall framework of the proposed GMGCN consists of three processes: node-level representations, node inference, and graph-level identi\ufb01cations. Case 1 (Infer Original Nodes) Given the whole G and X as training data, how to infer all the node labels z(n) i in the G(n)s? Case 2 (Infer New Nodes) Given the portion of nodes V(n) train = {v(n) 1 , . . . , v(n) Mn\u2212s} in each G(n) as training data, how to infer the labels of the rest of s nodes {v(n) Mn\u2212s+1, . . . , v(n) Mn} in the G(n)s? Case 3 (Infer New Graphs) Given the portion of graphs Gtrain = {(G(1), y(1)), . . . , (G(N\u2212S), y(N\u2212S))} as training data, how to infer the node labels z(n) i , n \u2208{N \u2212S + 1, . . . , N} in the rest of S graphs, {G(N\u2212S+1), . . . , G(N)}? Case 4 (With/Without Hierarchical Graph) Given the inter-connections among graphs as Ahier, how to infer the node labels. And how to infer the node labels if Ahier = I (i.e. no inter-connections)? There are more different cases under IGI problem. For example, how to infer a node score instead of a node label? We will discuss more cases in Appendix A. In this paper, we focus on the IGI problem that aims to infer the labels of the nodes. In the Experiments section, we will examine the proposed method and the baselines under all the above cases. 4 Proposed Method In this section, we \ufb01rst introduce the preliminaries related to the proposed model, then we discuss two key components of GMGCN: GML and consensus loss. 4.1 Preliminaries Graph Attention Networks. Graph Attention Networks (GAT) [14] has been widely adopted in the \ufb01eld of graph convolution due to the learnable attentions to neighbors during the node update process. A multi-head GAT Convolutional layer (GATConv) is formulated as below: \u20d7 hi = K \u2225 k=1 \u03c3(\u03b1k i,i\u0398\u20d7 xi + X j\u2208N(i) \u03b1k i,j\u0398\u20d7 xj), (1) where \u20d7 hi and \u20d7 xi refer to the hidden feature representation and the original feature vector of node i in the graph, respectively. \u2225refers to concatenation operation, and \u03c3(\u00b7) refers to the nonlinear activation function. \u0398 is the weight matrix and N(i) is the set of neighbors of node i. \u03b1k i,j is the attention coef\ufb01cient computed by the k-th attention mechanism as follows: \u03b1i,j = exp(LeakyReLU(\u20d7 a\u22a4[\u0398\u20d7 xi \u2225\u0398\u20d7 xj])) P l\u2208N(i)\u222a{i} exp(LeakyReLU(\u20d7 a\u22a4[\u0398\u20d7 xi \u2225\u0398\u20d7 xl])), (2) where attention mechanism is a single-layer feedforward neural network, parametrized by a weight vector a \u2208R2d2, and applying the LeakyReLU nonlinearity. In this paper, a two-layer GAT is formulated as follows: H = GAT(X) = \u03b1( K \u2225 k=1 \u03c3(\u03b1kX\u0398a1))\u0398a2. (3) 4 \fGraph Attention Pooling. Lin et al. propose the graph attention pooling in [17], which employs a self-attentive mechanism to learn the node importance and then transform a variable number of nodes into a \ufb01xed-length graph representation: E = Attn(H) = softmax(\u0398s2tanh(\u0398s1H\u22a4))H. (4) The node attention scores are calculated by a 2-layer MLP without bias. \u0398s2 is employed to infer the importance of each node. softmax function is adopted to normalize the importance of each node in the graph. Finally, the uni\ufb01ed graph representation of the graph is obtained by multiplying the node attention scores with H. Hierarchical GCN. We employ a hierarchical GCN [1] to update the feature representations of the graphs in the hierarchical graph as Eq. 5 H\u2032 = hierGCN(H) = Concat \u0010 H, \u03c3(AhierH\u0398) \u0011 , (5) where the hierarchical feature matrix H consists of the feature representations of graphs. Then we concatenate the original hierarchical feature matrix with the updated feature matrix as a self-loop. 4.2 Overall Framework Fig. 1 gives an overall framework of the proposed GMGCN method to infer the categories of the nodes in the graphs. It consists of three processes: node-level representations, node inference, and graph-level identi\ufb01cations. Gaussian Mixture Layer (GML). As the key component of the GMGCN, GML plays a crucial role in the process of node inference. It is described as \u20d7 h\u2032 = 1 C PC c=1 wc(\u0398GML\u20d7 h), where \u0398GML refers to the weight matrix and wc(\u00b7) represents the Gaussian weight function as follows: wc(\u20d7 h) = exp(\u22121 2(\u20d7 h \u2212\u20d7 \u00b5c)\u22a4\u03a3\u22121 c (\u20d7 h \u2212\u20d7 \u00b5c)), (6) where \u03a3c and \u20d7 \u00b5c are learnable covariance matrix and mean vector of the Gaussian weight function wc(\u00b7), respectively. Based on the idea of Gaussian Mixture Model (GMM), GML has the ability to effectively distinguish different types of nodes according to the input features. However, it could not resolve two major challenges of IGI problem: the node relations and the graph identi\ufb01cation that affects the \ufb01nal node inference. Therefore, as shown in Fig. 1, we employ the node representations from the 2-layer GAT as inputs of GML. The node representations provide local structural information of the nodes to GML. Meanwhile, the outputs \u20d7 h\u2032 of GML are aggregated by attention pooling and further fed into a hierarchical GCN for graph identi\ufb01cations. The graph identi\ufb01cations adjust the hidden space of GML via a back propagation process of hierGCN. Note that, one can easily replace hierGCN by multilayer perceptron (MLP) to formulate Case4 of IGI. Both node representations and graph identi\ufb01cations address two major challenges of IGI, respectively. Figure 2: Calculation of the enhanced distance matrix (dnc)enhance. GMGCN vs. AutoEncoder. Actually, the GMGCN model can be re-thought as an encoder-decoder process with a pre-speci\ufb01ed hidden space, where the 2-layer GAT is considered as an encoder, the hierGCN is considered as a decoder to graph labels instead of nodes themselves, and GML is the hidden space. When involving GML into the proposed model, it naturally involves an assumption that the nodes of each category are potentially distributed as a Gaussian distribution in the hidden space. The nodes of multiple categories are mixed out to form a Gaussian mixture distribution. Consequently, this also implies the limitation of the proposed model that the proposed model is only able to resolve the IGI problem for categorizing nodes. Consensus Loss. The core idea of consensus loss is to make the graph representations with the same labels closer, and those with different labels farther away in the hidden space. Therefore, 5 \fTable 1: Comparison of different types of methods on Synthetic dataset (mean(std)). methods single graph orignal nodes new nodes new graphs NMI ARI NMI ARI NMI ARI Feature Clustering 49.46(1.89) 50.78(2.48) 50.11(2.07) 49.33(2.56) 47.48(2.48) 48.03(3.32) Graph Clustering 80.65(8.18) 85.51(7.96) 85.85(8.47) 87.72(8.52) 78.72(10.47) 84.25(10.33) MIL 76.47(0.00) 81.39(0.00) 75.42(0.00) 80.76(0.00) 75.77(0.00) 80.63(0.00) ATTGCN 14.25(9.76) 12.45(11.64) 19.69(9.54) 13.44(11.38) 13.25(9.60) 12.22(11.10) GMGCN-noncon 49.58(37.15) 56.57(31.17) 60.46(26.25) 57.72(31.87) 54.91(25.69) 56.88(30.34) GMGCN 93.28(0.77) 96.29(0.43) 96.70(0.43) 97.55(0.34) 93.49(1.12) 96.46(0.62) methods hierarchical graph orignal nodes new nodes new graphs NMI ARI NMI ARI NMI ARI ATTGCN 19.09(12.34) 17.48(13.52) 24.64(12.34) 19.09(13.23) 20.54(11.92) 20.00(13.37) GMGCN-noncon 46.45(17.61) 37.44(38.26) 48.08(28.28) 37.66(38.33) 49.17(25.88) 37.26(38.14) GMGCN 92.83(1.22) 96.02(0.71) 96.32(0.77) 97.20(0.62) 92.54(1.71) 95.95(0.88) before computing the consensus loss, we \ufb01rst compute the similarity matrix S \u2208RN\u00d7C of the graph representations, which is obtained by the following two steps: First, we compute the distance between the graph representation \u20d7 hn g via graph attention pooling layer and the mean vector \u20d7 \u00b5c of c-th Gaussian weight function as follows: dnc =\u2225\u20d7 hn g \u2212\u20d7 \u00b5c \u22252 . (7) Second, we enhance the distance of the graph representations to the true mean vector on dnc by: (dnc)enhance = (dnc + \u03b4dny(n)), (8) where \u03b4 is a discount hyperparameter, y(n) is the true label of graph G(n), and dny(n) represents the distance between the graph representation \u20d7 hn g and the mean vector of y(n)-th Gaussian weight function \u20d7 \u00b5y(n). Then the similarity matrix S is formulated as S = softmax(\u2212Denhance), where Denhance \u2208RN\u00d7C consists of (dnc)enhance and is calculated as Fig. 2. Finally, the consensus loss, Lcon, is formulated as a cross-entropy loss between snc and y(n). In this way, the consensus loss makes the graph representations of each graph closer to the mean vector of the Gaussian function corresponding to its graph label, so as to realize that the graphs with the same labels are closer and the graphs with different labels are farther. Meanwhile, the graph attention pooling allows the node representations with the potentially opposite label to their graph labels to leave away from the graph representations by paying less attention. Overall Loss. Besides the consensus loss that distinguishes nodes in the hidden space, the GMGCN model also includes the graph classi\ufb01cation task. We establish a cross-entropy loss Lcls between the predictions and the ground truth labels over all graphs. Then, the overall loss function of our GMGCN is de\ufb01ned as linear interpolation of Lcls and Lcon as L = \u03b1Lcls + (1 \u2212\u03b1)Lcon, where \u03b1 \u2208[0, 1] is the trade-off coef\ufb01cient between classi\ufb01cation loss Lcls and consensus loss Lcon. 4.3 Inference of Nodes We infer the categories of the nodes in the graph based on the cosine similarity between the node representations, \u20d7 h, obtained from GML and the mean vector, \u20d7 \u00b5c, of each Gaussian weight function. The category \u02c6 z of each node is determined as follows: \u02c6 z = arg max c cosine(\u20d7 h, \u20d7 \u00b5c). (9) 5 Experiments In this section, we \ufb01rst conduct the experiments on synthetic data to validate the effectiveness of our proposed model under different problem cases. Then, we construct a real-world dataset based on the 6 \f(a) Original features (b) Feature Clustering (c) Graph Clustering (d) GMGCN Figure 3: 2D visualization of the original node features, the node representations learned by feature clustering, graph feature clustering, and GMGCN from two randomly sampled graphs. The different shapes of the symbols represent different graphs and the different colors represent different node categories. public rumor dataset called PHEME2 [35], the proposed method and the comparisons are evaluated on the constructed dataset. 5.1 Baselines Since the IGI problem is a new challenge without any baseline work yet, we compare the proposed GMGCN with three types of most similar methods: clustering methods without structure information [21, 22, 36, 37], graph clustering methods [25, 26] and MIL [38] methods. We only report the best results of each type of method in the sequel due to the page limitations. More detailed results for all methods are available in Appendix B. Proposed Method. We also examine the various settings of the proposed method as: \u2022 ATTGCN: GMGCN without GML, we replace cosine similarity by 2-head attention mechanisms in the attention pooling layer to distinguish the categories of the nodes in the inference. \u2022 GMGCN: The proposed method. \u2022 GMGCN-noncon: GMGCN without the proposed consensus loss for training. We apply Normalized Mutual Information (NMI) [39] and Adjusted Rand Index (ARI) [40] to evaluate the clustering results in this paper. The dimension of our GMGCN is set to d-128-64-16-64c, where d is the input dimension and c is the number of graph categories in the dataset. We adopt 2 independent attention mechanisms for the \ufb01rst GATConv layer and use 1 attention mechanism for the second GATConv layer. The parameters are updated using stochastic gradient descent via Adam algorithm [41]. The discount hyperparameter \u03b4=0.5. The training process is iterated upon 800 epochs for GMGCN. We run all methods 10 times to avoid extremes and report the average results. 5.2 GMGCN on Synthetic Data Synthetic Data Generation. We generate a synthetic hierarchical graph to examine the proposed IGI problem with various cases. First, we generate two synthetic user groups, representing normal and abnormal users, respectively. The relationship between normal users is generated by the Barab\u00e1si\u2013Albert graph model [42], while the relationship between abnormal users is generated by the random graph algorithm. We use Gaussian distribution to randomly generate user features, the mean and standard deviation of these two types of users are randomly sampled from [-5, 5] and [1, 10], respectively. Second, we randomly connect users between two groups to simulate the relationship between two different types of users. Third, we sample 20% of abnormal (normal) users into the normal (abnormal) graph so that a graph structure consists of two different types of users. We sample 100 users per graph, and generate 50 normal graphs and 50 abnormal graphs. The label of each graph is determined by the major labels of the users in the graph. Finally, we construct the connections between two graphs if they have more than one common user. Case 1: Transductive learning. All graphs are available in the training. Table 1 shows that the proposed GMGCN method outperforms all the baselines. In particular, GMGCN improves NMI and ARI by more than 10% compared with all best-performing baseline results. This demonstrates a great improvement of the node identi\ufb01cations by the graph labels. 2https://\ufb01gshare.com/articles/PHEME_dataset_of_rumours_and_non-rumours/4010619 7 \fTable 2: Comparison of different types of methods on PHEME datasets (mean(std)). Charlie Hebdo Ferguson Germanwings Crash Ottawa Shooting Sydney Siege Method NMI NMI NMI NMI NMI Feature Clustering 25.46(0.00) 24.56(0.00) 48.84(0.00) 34.56(0.00) 19.45(0.00) Graph Clustering 3.26(2.26) 1.08(0.71) 0.66(0.50) 4.49(3.07) 4.93(1.92) MIL 5.69(0.00) 4.08(0.00) 0.61(0.00) 0.60(0.00) 19.68(0.00) GMGCN 47.51(3.27) 48.35(4.08) 48.85(2.14) 32.58(3.63) 41.00(3.93) Method ARI ARI ARI ARI ARI Feature Clustering 23.46(0.00) 23.23(0.00) 47.29(0.00) 32.08(0.00) 16.45(0.00) Graph Clustering 6.56(6.30) 4.04(2.24) 3.71(1.93) 10.66(7.68) 11.78(5.57) MIL 9.38(0.00) 13.93(0.00) 1.10(0.00) 2.09(0.00) 39.51(0.00) GMGCN 52.38(3.15) 55.26(3.21) 54.95(1.02) 37.51(2.23) 37.79(2.18) Case 2 & Case 3: Inductive learning. Portions of graphs or nodes are available in the training. For the testing of new nodes and new graphs, we sample 20% of the nodes from each graph and 20% of the graphs as testing, respectively. GMGCN also outperforms the other baselines as shown in Table 1. Therefore, GMGCN works for both transductive and inductive settings. Case 4: With hierarchical graph. As shown in Table 1, the hierarchical graph structure promotes the performance of ATTGCN but has no effect on GMGCN. This indicates that GMGCN focuses more on node-level updates than graph-level updates to obtain better node clustering results. Attention vs. GML. We compare the ATTGCN with GMGCN to verify the importance of GML. The results show that GML greatly promotes GMGCN to identify nodes, while the attention mechanism fails. Consensus loss. To analyze the effect of the proposed consensus loss, we compare the performance of GMGCN and GMGCN-noncon. The results in Table 1 indicates that consensus loss enables GMGCN to identify node categories well. Consequently, the out-performances of GMGCN for all cases imply that the proposed method is a right solution for IGI problem. Visualization. We also visualize the node representations of two graphs randomly sampled from the Synthetic dataset in a 2D space by t-SNE [43]. Figs 3(a)-3(d) show the original node features, as well as the node representations learned from the feature clustering, graph clustering, and GMGCN, respectively. Different shapes in the \ufb01gures represent the nodes from different graphs and different colors refer to different node categories. As shown in Fig. 3(d), GMGCN clearly distinguishes these two node categories but the others fail. This demonstrates that the node representations learned by GMGCN are reasonable for IGI. 5.3 GMGCN on PHEME Data Data Description. PHEME consists of \ufb01ve sets of real-life tweets, where each set is related to a piece of breaking news. Every breaking news includes a lot of rumor and non-rumor topics. A detailed data description is presented in [35]. Based on this dataset, we construct the hierarchical graph structure by assuming each topic as a graph, where links between users are formed according to their follows/retweets. We connect two graphs if they have more than one common user. We label the users appearing in more than M rumor topics as abnormal users, and the rest as normal users. Therefore, this dataset perfectly falls into the statement of the IGI problem. When M = 2, 3, 4, the proposed method all achieves the best results, the detailed results are shown in Appendix B. In this section, we only show the results when M = 4. Clustering Results. As shown in Table 5, GMGCN perfectly resolves this IGI problem with an average 13% improvement on NMI and a 15% improvement on ARI on these \ufb01ve PHEME datasets compared with the best-performing baseline methods. This signi\ufb01cant improvement implies that there exist certain applications in the real world that ful\ufb01ll the statement of IGI problem. In contrast, the other methods perform poorly because IGI is a totally new problem that differs from feature clustering, graph clustering, or MIL. Due to the improper assumptions, they all fail to reveal the truth. 8 \f6 Discussion and Future Work In this work, we explore an interesting and novel problem: Can we identify nodes given the identi\ufb01cations of graphs? In one word, can we Invert the Graph Identi\ufb01cation? The proposed IGI problem has totally different notions and statements from any of the node clustering, graph clustering, or multiple instance learning tasks. To address this new issue, we propose GMGCN with the integration of GMM and GNNs to resolve certain cases of IGI problem. Experimental results on Synthetic data and real-world data reveal the need for formulating IGI and the advantage of the proposed GMGCN over other related baselines. Nevertheless, our study is just an initial step towards the IGI problem while a variety of extensions are still potential. For example, can we infer the labels of edges or sub-graphs instead of nodes? Can we detect anomaly nodes other than conducting node clustering? We hope our study will open up a new vein of graph learning and encourage more speci\ufb01cations, solutions, and developments for IGI. Broader Impact Nowadays, many studies have been developed on Graph Identi\ufb01cation (GI), such as graph classi\ufb01cation and graph regression. However, the inverse problem of using the information of graphs to infer the information of nodes or even edges and sub-graphs has never been raised. In this paper, we present a new challenge named as Inverse Graph Identi\ufb01cation (IGI). As opposed to GI, IGI mainly explores whether we can identify nodes given the labels of graphs they belong to. We also analyze various tasks belonging to the IGI problem so that more researchers can participate in the research of the IGI problem. By studying the IGI problem, many graph learning problems will be solved in a more appropriate and ef\ufb01cient way. The broader impact of our research can be summarized below: For social media community: Through the research of social media tasks on IGI problem, different types of information communicators can be identi\ufb01ed more accurately. For example, media reporters can use the proposed model to \ufb01nd appropriate interviewees, advertisers can discover the users in social networks for accurate advertising. For biology and chemistry community: Biologists can divide proteins into groups according to the different interactions of proteins, and then use our method to \ufb01nd proteins with speci\ufb01c functions. Chemists can employ our method to discover functional groups with special properties. For network security community: By dividing different people into groups, network security of\ufb01cers can utilize our method to discover suspicious users on the LAN, and can also employ our method to \ufb01nd suspicious operations or applications by grouping different operations or software together."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2001.06362v1",
+                    "title": "Rumor Detection on Social Media with Bi-Directional Graph Convolutional Networks",
+                    "abstract": "Social media has been developing rapidly in public due to its nature of\nspreading new information, which leads to rumors being circulated. Meanwhile,\ndetecting rumors from such massive information in social media is becoming an\narduous challenge. Therefore, some deep learning methods are applied to\ndiscover rumors through the way they spread, such as Recursive Neural Network\n(RvNN) and so on. However, these deep learning methods only take into account\nthe patterns of deep propagation but ignore the structures of wide dispersion\nin rumor detection. Actually, propagation and dispersion are two crucial\ncharacteristics of rumors. In this paper, we propose a novel bi-directional\ngraph model, named Bi-Directional Graph Convolutional Networks (Bi-GCN), to\nexplore both characteristics by operating on both top-down and bottom-up\npropagation of rumors. It leverages a GCN with a top-down directed graph of\nrumor spreading to learn the patterns of rumor propagation, and a GCN with an\nopposite directed graph of rumor diffusion to capture the structures of rumor\ndispersion. Moreover, the information from the source post is involved in each\nlayer of GCN to enhance the influences from the roots of rumors. Encouraging\nempirical results on several benchmarks confirm the superiority of the proposed\nmethod over the state-of-the-art approaches.",
+                    "authors": "Tian Bian, Xi Xiao, Tingyang Xu, Peilin Zhao, Wenbing Huang, Yu Rong, Junzhou Huang",
+                    "published": "2020-01-17",
+                    "updated": "2020-01-17",
+                    "primary_cat": "cs.SI",
+                    "cats": [
+                        "cs.SI",
+                        "cs.LG"
+                    ],
+                    "main_content": "Introduction With the rapid development of the Internet, social media has become a convenient online platform for users to obtain information, express opinions and communicate with each other. As more and more people are keen to participate in discussions about hot topics and exchange their opinions on social media, many rumors appear. Due to a large number of users and easy access to social media, rumors can spread widely and quickly on social media, bringing huge harm to society and causing a lot of economic losses. Therefore, regarding to the potential panic and threat caused by rumors, it is urgent to come up with a method to identify rumors on social media ef\ufb01ciently and as early as possible. Conventional detection methods mainly adopt handcrafted features such as user characteristics, text contents and propagation patterns to train supervised classi\ufb01ers, e.g., Copyright c \u20dd2020, Association for the Advancement of Arti\ufb01cial Intelligence (www.aaai.org). All rights reserved. Decision Tree (Castillo, Mendoza, and Poblete 2011), Random Forest (Kwon et al. 2013), Support Vector Machine (SVM) (Yang et al. 2012). Some studies apply more effective features, such as user comments (Giudice 2010), temporal-structural features (Wu, Yang, and Zhu 2015), and the emotional attitude of posts (Liu et al. 2015). However, those methods mainly rely on feature engineering, which is very time-consuming and labor-intensive. Moreover, those handcrafted features are usually lack of high-level representations extracted from the propagation and the dispersion of rumors. Recent studies have exploited deep learning methods that mine high-level representations from propagation path/trees or networks to identify rumors. Many deep learning models such as Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), and Recursive Neural Networks (RvNN) (Ma et al. 2016; Ma, Gao, and Wong 2018) are employed since they are capable to learn sequential features from rumor propagation along time. However, these approaches have a signi\ufb01cant limitation on ef\ufb01ciency since temporal-structural features only pay attention to the sequential propagation of rumors but neglect the in\ufb02uences of rumor dispersion. The structures of rumor dispersion also indicate some spreading behaviors of rumors. Thus, some studies have tried to involve the information from the structures of rumor dispersion by invoking Convolutional Neural Network (CNN) based methods (Yu et al. 2017; Yu et al. 2019). CNN-based methods can obtain the correlation features within local neighbors but cannot handle the global structural relationships in graphs or trees (Bruna et al. 2014). Therefore, the global structural features of rumor dispersion are ignored in these approaches. Actually, CNN is not designed to learn high-level representations from structured data but Graph Convolutional Network (GCN) is (Kipf and Welling 2017). So can we simply apply GCN to rumor detection since it has successfully made progress in various \ufb01elds, such as social networks (Hamilton, Ying, and Leskovec 2017), physical systems (Battaglia et al. 2016), and chemical drug discovery (Defferrard, Bresson, and Vandergheynst 2016)? The answer is no. As shown in Figure 1(a), GCN, or called undirected GCN (UD-GCN), only aggregates information relied on the relationships among relevant posts but loses the searXiv:2001.06362v1  [cs.SI]  17 Jan 2020 \fquential orders of follows. Although UD-GCN has the ability to handle the global structural features of rumor dispersion, it does not consider the direction of the rumor propagation, which however has been shown to be an important clue for rumor detection (Wu, Yang, and Zhu 2015). Specifically, deep propagation along a relationship chain (Han et al. 2014) and wide dispersion across a social community (Thomas 2007) are two major characteristics of rumors, which is eager for a method to serve both. To deal with both propagation and dispersion of rumors, in this paper, we propose a novel Bi-directional GCN (BiGCN), which operates on both top-down and bottom-up propagation of rumors. The proposed method obtains the features of propagation and dispersion via two parts, the Top-Down graph convolutional Networks (TD-GCN) and Bottom-Up graph convolutional Networks (BU-GCN), respectively. As shown in Figure 1(b) and 1(c), TD-GCN forwards information from the parent node of a node in a rumor tree to formulate rumor propagation while BU-GCN aggregates information from the children nodes of a node in a rumor tree to represent rumor dispersion. Then, the representations of propagation and dispersion pooled from the embedding of TD-GCN and BU-GCN are merged together through full connections to make the \ufb01nal results. Meanwhile, we concatenate the features of the roots in rumor trees with the hidden features at each GCN layer to enhance the in\ufb02uences from the roots of rumors. Moreover, we employ DropEdge (Rong et al. 2019) in the training phase to avoid over-\ufb01tting issues of our model. The main contributions of this work are as follows: \u2022 We leverage Graph Convolutional Networks to detect rumors. To the best of our knowledge, this is the \ufb01rst study of employing GCN in rumor detection of social media. \u2022 We propose the Bi-GCN model that not only considers the causal features of rumor propagation along relationship chains from top to down but also obtains the structural features from rumor dispersion within communities through the bottom-up gathering. \u2022 We concatenate the features of the source post with other posts at each graph convolutional layer to make a comprehensive use of the information from the root feature and achieve excellent performance in rumor detection. Experimental results on three real-world datasets show that our Bi-GCN method outperforms several state-of-theart approaches; and for the task of early detection of rumors, which is quite crucial to identify rumors in real time and prevent them from spreading, Bi-GCN also achieves much higher effectiveness. Related Work In recent years, automatic rumor detection on social media has attracted a lot of attention. Most previous work for rumor detection mainly focuses on extracting rumor features from the text contents, user pro\ufb01les and propagation structures to learn a classi\ufb01er from labeled data (Castillo, Mendoza, and Poblete 2011; Yang et al. 2012; Kwon et al. 2013; Liu et al. 2015; Zhao, Resnick, and Mei 2015). Ma et al. (a) UD-GCN (b) TD-GCN (c) BU-GCN Figure 1: (a) the undirected graph with only node relationships; (b) the deep propagation along a relationship chain from top to down; (c) the aggregation of the wide dispersion within a community to an upper node. (Ma et al. 2015) classi\ufb01ed the rumor by using the timeseries to model the variation of handcrafted social context features. Wu et al. (Wu, Yang, and Zhu 2015) proposed a graph kernel-based hybrid SVM classi\ufb01er by combining the RBF kernel with a random-walk-based graph kernel. Ma et al. (Ma, Gao, and Wong 2017) constructed a propagation tree kernel to detect rumors by evaluating the similarities between their propagation tree structures. These methods not only were ineffective but also heavily relied on handcrafted feature engineering to extract informative feature sets. In order to automatically learn high-level features, a number of recent methods were proposed to detect rumor based on deep learning models. Ma et al. utilized Recurrent Neural Networks (RNN) to capture the hidden representation from temporal content features (Ma et al. 2016). Chen et al. (Chen et al. 2018) improved this approach by combining attention mechanisms with RNN to focus on text features with different attentions. Yu et al. (Yu et al. 2017) proposed a method based on Convolutional Neural Network (CNN) to learn key features scattered among an input sequence and shape high-level interactions among signi\ufb01cant features. Liu et al. (Liu and Wu 2018) incorporated both RNN and CNN to get the user features based on time series. Recently, Ma et al. (Ma, Gao, and Wong 2019) employed the adversarial learning method to improve the performance of rumor classi\ufb01er, where the discriminator is used as a classi\ufb01er and the corresponding generator improves the discriminator by generating con\ufb02icting noises. In addition, Ma et al. built a treestructured Recursive Neural Networks (RvNN) to catch the hidden representation from both propagation structures and text contents (Ma, Gao, and Wong 2018). However, these methods are too inef\ufb01cient to learn the features of the propagation structure, and they also ignore the global structural features of rumor dispersion. Compared to the deep-learning models mentioned above, GCN is able to capture global structural features from graphs or trees better. Inspired by the success of CNN in the \ufb01eld of computer vision, GCN has demonstrated state-of-the-art performances in various tasks with graph data (Battaglia et al. 2016; Defferrard, Bresson, and Vandergheynst 2016; Hamilton, Ying, and Leskovec 2017). Scarselli et al. (Scarselli et al. 2008) \ufb01rstly introduced GCN \fas a special massage-passing model for either undirected graphs or directed graphs. Later on, Bruna et al. (Bruna et al. 2014) theoretically analyzed graph convolutional methods for undirected graphs based on the spectral graph theory. Subsequently, Defferrard et al. (Defferrard, Bresson, and Vandergheynst 2016) developed a method named the Chebyshev Spectral CNN (ChebNet) and used the Chebyshev polynomials as the \ufb01lter. After this work, Kipf et al. (Kipf and Welling 2017) presented a \ufb01rst-order approximation of ChebNet (1stChebNet), where the information of each node is aggregated from the node itself and its neighboring nodes. Our rumor detection model is inspired by the GCN. Preliminaries We introduce some fundamental concepts that are necessary for our method. First the notation used in this paper is as follows. Notation Let C = {c1, c2, ..., cm} be the rumor detection dataset, where ci is the i-th event and m is the number of events. ci = {ri, wi 1, wi 2, ..., wi ni\u22121, Gi}, where ni refers to the number of posts in ci, ri is the source post, each wi j represents the j-th relevant responsive post, and Gi refers to the propagation structure. Speci\ufb01cally, Gi is de\ufb01ned as a graph \u27e8Vi, Ei\u27e9 with ri being the root node (Wu, Yang, and Zhu 2015; Ma, Gao, and Wong 2017), where Vi = {ri, wi 1, . . . , wi ni\u22121}, and Ei = {ei st|s, t = 0, . . . , ni\u22121} that represents the set of edge from responded posts to the retweeted posts or responsive posts, as shown in Figure 1(b). For example, if wi 2 has a response to wi 1, there will be an directed edge wi 1 \u2192wi 2, i.e., ei 12. If wi 1 has a response to ri, there will be an directed edge ri \u2192wi 1, i.e., ei 01. Denote Ai \u2208{0, 1}ni\u00d7ni as an adjacency matrix where ai ts = \u001a1, if ei st \u2208Ei 0, otherwise . Denote Xi = [xi\u22a4 0 , xi\u22a4 1 , ..., xi\u22a4 ni\u22121]\u22a4as a feature matrix extracted from the posts in ci, where xi 0 represents the feature vector of ri and each other row feature xi j represents the feature vector of wi j. Moreover, each event ci is associated with a ground-truth label yi \u2208{F, T} (i.e., False Rumor or True Rumor). In some cases, the label yi is one of the four \ufb01ner-grained classes {N, F, T, U} (i.e., Non-rumor, False Rumor, True Rumor, and Unveri\ufb01ed Rumor) (Ma, Gao, and Wong 2017; Zubiaga et al. 2018). Given the dataset, the goal of rumor detection is to learn a classi\ufb01er f : C \u2192Y, where C and Y are the sets of events and labels respectively, to predict the label of an event based on text contents, user information and propagation structure constructed by the related posts from that event. Graph Convolutional Networks Recently, there is an increasing interest in generalizing convolutions to the graph domain. Among all the existing works, GCN is one of the most effective convolution models, whose convolution operation is considered as a general \u201dmessage-passing\u201d architecture as follows: Hk = M(A, Hk\u22121; Wk\u22121), (1) where Hk \u2208Rn\u00d7vk is the hidden feature matrix computed by the k \u2212th Graph Conventional Layer (GCL) and M is the message propagation function, which depends on the adjacency matrix A, the hidden feature matrix Hk\u22121 and the trainable parameters Wk\u22121. There are many kinds of message propagation functions M for GCN (Bruna et al. 2014; Defferrard, Bresson, and Vandergheynst 2016). Among them, the message propagation function de\ufb01ned in the \ufb01rst-order approximation of ChebNet (1stChebNet) (Kipf and Welling 2017) is as follows: Hk = M(A, Hk\u22121; Wk\u22121) = \u03c3(\u02c6 AHk\u22121Wk\u22121). (2) In the above equation \u02c6 A = \u02dc D \u22121 2 \u02dc A\u02dc D \u22121 2 is the normalized adjacency matrix, where \u02dc A=A + IN (i.e., adding selfconnection), \u02dc Dii=\u03a3j \u02dc Aij that represents the degree of the i \u2212th node; Wk\u22121 \u2208Rvk\u22121\u00d7vk; and \u03c3(\u00b7) is an activation function, e.g., the ReLU function. DropEdge DropEdge is a novel method to reduce over-\ufb01tting for GCNbased models (Rong et al. 2019). In each training epoch, it randomly drops out edges from the input graphs to generate different deformed copies with certain rate. As a result, this method augments the randomness and the diversity of input data, just like rotating or \ufb02apping images at random. Formally, suppose the total number of edges in the graph A is Ne and the dropping rate is p, then the adjacency matrix after DropEdge, A\u2032, is computed as below: A\u2032 = A \u2212Adrop (3) where Adrop is the matrix constructed using Ne \u00d7 p edges randomly sampled from the original edge set. Bi-GCN Rumor Detection Model In this section, we propose an effective GCN-based method for rumor detection based on the rumor propagation and the rumor dispersion, named as Bi-directional Graph Convolutional Networks (Bi-GCN). The core idea of Bi-GCN is to learn suitable high-level representations from both rumor propagation and rumor dispersion. In our Bi-GCN model, two-layer 1stChebNet are adopted as the fundamental GCN components. As shown in Figure 2, we elaborate the rumor detection process using Bi-GCN in 4 steps. We \ufb01rst discuss how to apply the Bi-GCN model to one event, i.e., ci \u2192yi for the i-th event. The other events are calculated in the same manner. To better present our method, we omit the subscript i in the following content. \fFigure 2: Our Bi-GCN rumor detection model. X denotes the original feature matrix input to the Bi-GCN model, and Hk is the hidden features matrix generated from the k \u2212th GCL. Xroot and Hroot 1 represents the matrix extended by the features of source post. 1 Construct Propagation and Dispersion Graphs Based on the retweet and response relationships, we construct the propagate structure \u27e8V, E\u27e9for a rumor event ci. Then, let A \u2208Rni\u00d7ni and X be its corresponding adjacency matrix and feature matrix of ci based on the spreading tree of rumors, respectively. A only contains the edges from the upper nodes to the lower nodes as illustrated in Figure 1(b). At each training epoch, p percentage of edges are dropped via Eq. (3) to form A\u2032, which avoid penitential over\ufb01tting issues (Rong et al. 2019). Based on A\u2032 and X, we can build our Bi-GCN model. Our Bi-GCN consists of two components: a Top-Down Graph Convolutional Network (TDGCN) and a Bottom-Up Graph Convolutional Network (BUGCN). The adjacency matrices of two components are different. For TD-GCN, the adjacency matrix is represented as AT D = A\u2032. Meanwhile, for BU-GCN, the adjacency matrix is ABU = A\u2032\u22a4. TD-GCN and BU-GCN adopt the same feature matrix X. 2 Calculate the High-level Node Representations After the DropEdge operation, the top-down propagation features and the bottom-up propagation features are obtained by TD-GCN and BU-GCN, respectively. By substituting AT D and X to Eq. (2) over two layers, we write the equations for TD-GCN as below: HT D 1 = \u03c3 \u0010 \u02c6 A T DXW T D 0 \u0011 , (4) HT D 2 = \u03c3 \u0010 \u02c6 A T DHT D 1 W T D 1 \u0011 , (5) where HT D 1 \u2208Rn\u00d7v1 and HT D 2 \u2208Rn\u00d7v2 represent the hidden features of two layer TD-GCN. W T D 0 \u2208Rd\u00d7v1 and W T D 1 \u2208Rv1\u00d7v2 are the \ufb01lter parameter matrices of TDGCN. Here we adopt ReLU function as the activation function, \u03c3(\u00b7). Dropout (Srivastava et al. 2014) is applied on GCN Layers (GCLs) to avoid over-\ufb01tting. Similar to Eqs. (4) and (5), we calculate the bottom-up hidden features HBU 1 and HBU 2 for BU-GCN in the same manner as Eq. (4) and Eq. (5). 3 Root Feature Enhancement As we know, the source post of a rumor event always has abundant information to make a wide impact. It is necessary to better make use of the information from the source post, and learn more accurate node representations from the relationship between nodes and the source post. Consequently, besides the hidden features from TD-GCN and BU-GCN, we propose an operation of root feature enhancement to improve the performance of rumor detection as shown in Figure 2. Speci\ufb01cally, for TD-GCN at the kth GCL, we concatenate the hidden feature vectors of every nodes with the hidden feature vector of the root node from the (k \u22121)-th GCL to construct a new feature matrix as \u02dc H T D k = concat(HT D k , (HT D k\u22121)root) (6) with HT D 0 = X. Therefore, we express TD-GCN with the root feature enhancement by replacing HT D 1 in Eq. (5) with \u02dc H T D 1 = concat(HT D 1 , Xroot), and then get \u02dc H T D 2 as follows: HT D 2 = \u03c3 \u0010 \u02c6 A T D \u02dc H T D 1 W T D 1 \u0011 , (7) \u02dc H T D 2 = concat(HT D 2 , (HT D 1 )root). (8) \fSimilarly, the hidden feature metrics of BU-GCN with root feature enhancement, \u02dc H BU 1 and \u02dc H BU 2 , are obtained in the same manner as Eq. (7) and Eq. (8). 4 Representations of Propagation and Dispersion for Rumor Classi\ufb01cation The representations of propagation and dispersion are the aggregations from the node representations of TD-GCN and BU-GCN, respectively. Here we employ mean-pooling operators to aggregate information from these two sets of the node representations. It is formulated as ST D = MEAN(\u02dc H T D 2 ), (9) SBU = MEAN(\u02dc H BU 2 ). (10) Then, we concatenate the representations of propagation and the representation of dispersion to merge the information as S = concat(ST D, SBU). (11) Finally, the label of the event \u02c6 y is calculated via several full connection layers and a softmax layer: \u02c6 y = Softmax(FC(S)). (12) where \u02c6 y \u2208R1\u00d7C is a vector of probabilities for all the classes used to predict the label of the event. We train all the parameters in the Bi-GCN model by minimizing the cross-entropy of the predictions and ground truth distributions, Y , over all events, C. L2 regularizer is applied in the loss function over all the model parameters. Experiments In this section, we \ufb01rst evaluate the empirical performance of our proposed Bi-GCN method in comparison with several baseline models. Then, we investigate the effect of each variant of the proposed method. Finally, we also examine the capability of early rumor detection for both the proposed method and the compared methods. Settings and Datasets Datasets We evaluate our proposed method on three realworld datasets: Weibo (Ma et al. 2016), Twitter15 (Ma, Gao, and Wong 2017), and Twitter16 (Ma, Gao, and Wong 2017). Weibo and Twitter are the most popular social media sites in China and the U.S., respectively. In all the three datasets, nodes refer to users, edges represent retweet or response relationships, and features are the extracted top-5000 words in terms of the TF-IDF values as mentioned in the Bi-GCN Rumor Detection Model Section. The Weibo dataset contains two binary labels: False Rumor (F) and True Rumor (T), while Twitter15 and Twitter16 datasets contains four labels: Non-rumor (N), False Rumor (F), True Rumor (T), and Unveri\ufb01ed Rumor (U). The label of each event in Weibo is annotated according to Sina community management center, which reports various misinformation (Ma et al. 2016). And the label of each event in Twitter15 and Twitter16 is annotated according to the veracity tag of the article in rumor debunking websites (e.g., snopes.com, Emergent.info, etc) (Ma, Gao, and Wong 2017). The statistics of the three datasets are shown in Table 1. Table 1: Statistics of the datasets Statistic Weibo Twitter15 Twitter16 # of posts 3,805,656 331,612 204,820 # of Users 2,746,818 276,663 173,487 # of events 4664 1490 818 # of True rumors 2351 374 205 # of False rumors 2313 370 205 # of Unveri\ufb01ed rumors 0 374 203 # of Non-rumors 0 372 205 Avg. time length / event 2,460.7 Hours 1,337 Hours 848 Hours Avg. # of posts / event 816 223 251 Max # of posts / event 59,318 1,768 2,765 Min # of posts / event 10 55 81 Experimental Setup We compare the proposed method with some state-of-the-art baselines, including: \u2022 DTC (Castillo, Mendoza, and Poblete 2011): A rumor detection method using a Decision Tree classi\ufb01er based on various handcrafted features to obtain information credibility. \u2022 SVM-RBF (Yang et al. 2012): A SVM-based model with RBF kernel, using handcrafted features based on the overall statistics of the posts. \u2022 SVM-TS (Ma et al. 2015): A linear SVM classi\ufb01er that leverages handcrafted features to construct time-series model. \u2022 SVM-TK (Ma, Gao, and Wong 2017): A SVM classi\ufb01er with a propagation Tree Kernel on the basis of the propagation structures of rumors. \u2022 RvNN (Ma, Gao, and Wong 2018): A rumor detection approach based on tree-structured recursive neural networks with GRU units that learn rumor representations via the propagation structure. \u2022 PPC RNN+CNN (Liu and Wu 2018): A rumor detection model combining RNN and CNN, which learns the rumor representations through the characteristics of users in the rumor propagation path. \u2022 Bi-GCN: Our GCN-based rumor detection model utilizing the Bi-directional propagation structure. We implement DTC and SVM-based models with scikitlearn1; PPC RNN+CNN with Keras2; RvNN and our method with Pytorch3. To make a fair comparison, we randomly split the datasets into \ufb01ve parts, and conduct 5fold cross-validation to obtain robust results. For the Weibo dataset, we evaluate the Accuracy (Acc.) over the two categories and Precision (Prec.), Recall (Rec.), F1 measure (F1) on each class. For the two Twiter datasets, we evaluate Acc. over the four categories and F1 on each class. The 1https://scikit-learn.org 2https://keras.io/ 3https://pytorch.org/ \fTable 2: Rumor detection results on Weibo dataset (F: False Rumor; T: True Rumor) Method Class Acc. Prec. Rec. F1 DTC F 0.831 0.847 0.815 0.831 T 0.815 0.824 0.819 SVM-RBF F 0.879 0.777 0.656 0.708 T 0.579 0.708 0.615 SVM-TS F 0.885 0.950 0.932 0.938 T 0.124 0.047 0.059 RvNN F 0.908 0.912 0.897 0.905 T 0.904 0.918 0.911 PPC RNN+CNN F 0.916 0.884 0.957 0.919 T 0.955 0.876 0.913 Bi-GCN F 0.961 0.961 0.964 0.961 T 0.962 0.962 0.960 parameters of Bi-GCN are updated using stochastic gradient descent, and we optimize the model by Adam algorithm (Kingma and Ba 2014). The dimension of each node\u2019s hidden feature vectors are 64. The dropping rate in DropEdge is 0.2 and the rate of dropout is 0.5. The training process is iterated upon 200 epochs, and early stopping (Yao, Rosasco, and Caponnetto 2007) is applied when the validation loss stops decreasing by 10 epochs. Note that we do not employ SVM-TK on the Weibo dataset due to its exponential complexity on large datasets. Overall Performance Table 2 and Table 3 show the performance of the proposed method and all the compared methods on the Weibo and Twitter datasets, respectively. First, among the baseline algorithms, we observe that the deep learning methods performs signi\ufb01cantly better than those using hand-crafted features. It is not surprising, since the deep learning methods are able to learn high-level representations of rumors to capture valid features. This demonstrates the importance and necessity of studying deep learning for rumor detection. Second, the proposed method outperforms the PPC RNN+CNN method in terms of all the performance measures, which indicates the effectiveness of incorporating the dispersion structure for rumor detection. Since RNN and CNN cannot process data with the graph structure, PPC RNN+CNN ignores important structural features of rumor dispersion. This prevents it from obtaining ef\ufb01cient high-level representations of rumors, resulting in worse performance on rumor detection. Finally, Bi-GCN is signi\ufb01cantly superior to the RvNN method. Since RvNN only uses the hidden feature vector of all the leaf nodes so that it is heavily impacted by the information of the latest posts. However, the latest posts are always lack of information such as comments, and just follow the former posts. Unlike RvNN, the root feature enhancement allows the proposed method to pay more attention to Table 3: Rumor detection results on Twitter15 and Twitter16 datasets (N: Non-Rumor; F: False Rumor; T: True Rumor; U: Unveri\ufb01ed Rumor) Twitter15 Method Acc. N F T U F1 F1 F1 F1 DTC 0.454 0.415 0.355 0.733 0.317 SVM-RBF 0.318 0.225 0.082 0.455 0.218 SVM-TS 0.544 0.796 0.472 0.404 0.483 SVM-TK 0.750 0.804 0.698 0.765 0.733 RvNN 0.723 0.682 0.758 0.821 0.654 PPC RNN+CNN 0.477 0.359 0.507 0.300 0.640 Bi-GCN 0.886 0.891 0.860 0.930 0.864 Twitter16 Method Acc. N F T U F1 F1 F1 F1 DTC 0.473 0.254 0.080 0.190 0.482 SVM-RBF 0.553 0.670 0.085 0.117 0.361 SVM-TS 0.574 0.755 0.420 0.571 0.526 SVM-TK 0.732 0.740 0.709 0.836 0.686 RvNN 0.737 0.662 0.743 0.835 0.708 PPC RNN+CNN 0.564 0.591 0.543 0.394 0.674 Bi-GCN 0.880 0.847 0.869 0.937 0.865 the information of the source posts, which helps improve our models much more. Ablation Study To analyze the effect of each variant of Bi-GCN, we compare the proposed method with TD-GCN, BU-GCN, UDGCN and their variants without the root feature enhancement. The empirical results are summarized in Figure 3. UDGCN, TD-GCN, and BU-GCN represent our GCN-based rumor detection models utilize the UnDirected, Top-Down and Bottom-Up structures, respectively. Meanwhile, \u201droot\u201d refers to the GCN-based model with concatenating root features in the networks while \u201dno root\u201d represents the GCNbased model without concatenating root features in the networks. Some conclusions are drawn from Figure 3. First, BiGCN, TD-GCN, BU-GCN, and UD-GCN outperforms their variants without the root feature enhancement, respectively. This indicates that the source posts plays an important role in rumor detection. Second, TD-GCN and BU-GCN can not always achieve better results than UD-GCN, but Bi-GCN is always superior to UD-GCN, TD-GCN and BU-GCN. This implies the importance to simultaneously consider both topdown representations from the ancestor nodes, and bottomup representations from the children nodes. Finally, even the worst results in Figures 3 are better than those of other baseline methods in Table 2 and 3 by a large gap, which again \f(a) Weibo dataset (b) Twitter15 dataset (c) Twitter16 dataset Figure 3: The rumor detection performance of the GCN-based methods on three datasets (a) Weibo dataset (b) Twitter15 dataset (c) Twitter16 dataset Figure 4: Result of rumor early detection on three datasets veri\ufb01es the effectiveness of graph convolution for rumor detection. Early Rumor Detection Early detection aims to detect rumor at the early stage of propagation, which is another important metric to evaluate the quality of the method. To construct an early detection task, we set up a series of detection deadlines and only use the posts released before the deadlines to evaluate the accuracy of the proposed method and baseline methods. Since it is dif\ufb01cult for the PPC RNN+CNN method to process the data of variational lengths, we cannot get the accurate results of PPC RNN+CNN at each deadline in this task, so it is not compared in this experiment. Figure 4 shows the performances of our Bi-GCN method versus RvNN, SVM-TS, SVM-RBF and DTC at various deadlines for the Weibo and Twitter datasets. From the \ufb01gure, it can be seen that the proposed Bi-GCN method reaches relatively high accuracy at a very early period after the source post initial broadcast. Besides, the performance of Bi-GCN is remarkably superior to other models at each deadline, which demonstrates that structural features are not only bene\ufb01cial to long-term rumor detection, but also helpful to the early detection of rumors."
+                }
+            ],
+            "Yian Yin": [
+                {
+                    "url": "http://arxiv.org/abs/1704.04657v2",
+                    "title": "The Time Dimension of Science: Connecting the Past to the Future",
+                    "abstract": "A central question in science of science concerns how time affects citations.\nDespite the long-standing interests and its broad impact, we lack systematic\nanswers to this simple yet fundamental question. By reviewing and classifying\nprior studies for the past 50 years, we find a significant lack of consensus in\nthe literature, primarily due to the coexistence of retrospective and\nprospective approaches to measuring citation age distributions. These two\napproaches have been pursued in parallel, lacking any known connections between\nthe two. Here we developed a new theoretical framework that not only allows us\nto connect the two approaches through precise mathematical relationships, it\nalso helps us reconcile the interplay between temporal decay of citations and\nthe growth of science, helping us uncover new functional forms characterizing\ncitation age distributions. We find retrospective distribution follows a\nlognormal distribution with exponential cutoff, while prospective distribution\nis governed by the interplay between a lognormal distribution and the growth in\nthe number of references. Most interestingly, the two approaches can be\nconnected once rescaled by the growth of publications and citations. We further\nvalidate our framework using both large-scale citation datasets and analytical\nmodels capturing citation dynamics. Together this paper presents a\ncomprehensive analysis of the time dimension of science, representing a new\nempirical and theoretical basis for all future studies in this area.",
+                    "authors": "Yian Yin, Dashun Wang",
+                    "published": "2017-04-15",
+                    "updated": "2017-04-18",
+                    "primary_cat": "physics.soc-ph",
+                    "cats": [
+                        "physics.soc-ph",
+                        "cs.DL",
+                        "cs.SI"
+                    ],
+                    "main_content": "Introduction The increasing availability of large-scale datasets that capture major activities in science has created an unprecedented opportunity to explore broad and important patterns underlying the scienti\ufb01c enterprise, catalyzing in a drastic fashion a recent multidisciplinary shift in our quantitative understanding of science (Redner, 2005; Guimera et al., 2005; Radicchi et al., 2008; Jones et al., 2008; Hirsch, 2005; Stringer et al., 2008; Newman, 2009; Petersen et al., 2014; Evans, 2008; Evans and Foster, 2011; Sz\u00e1nt\u00f3-V\u00e1rnagy et al., 2014; Wang et al., 2013; Uzzi et al., 2013; Sinatra et al., 2015; Ke et al., 2015; Sinatra et al., 2016; Guevara et al., 2016). Nowhere are these new advances more apparent than in the studies of citations (Redner, 2005; Bornmann and Daniel, 2008; Radicchi et al., 2008; Eom and Fortunato, 2011; Barab\u00e1si et al., 2012; Moreira et al., 2015; Wang et al., 2013; Uzzi et al., 2013), owing largely to their widespread applications, from science policy to promotions (Lane, 2010; Radicchi et al., 2009; Shen and Barab\u00e1si, 2014), hiring (Clauset et al., 2015; Duch et al., 2012), assignment of grants (Lane and Bertuzzi, 2011; Li and Agha, 2015; Bromham et al., 2016) and prizes (Mazloumian et al., 2011). The aim of this paper is to carry out a comprehensive analysis on one of the most fundamental dimensions of citations: Time. Time plays a central role in science. Indeed, while some papers stay relevant and continue to dominate the scienti\ufb01c discourse over a long period, most papers unfortunately \u2018die out\u2019, after collecting their fair share of citations (Wang et al., 2013; Stringer et al., 2010). Understanding the time dimension is critical for 1 \fa wide range of reasons. It not only helps us understand the rise and fall of scienti\ufb01c paradigms (Kuhn, 1962), tracing knowledge horizons and hotspots in science and technology (Sinatra et al., 2015; Orosz et al., 2016; Mukherjee et al., 2017), it is also critical for allocating investment (Szell and Sinatra, 2015; Ma et al., 2015; Bromham et al., 2016), identifying crucial yet initially unrecognized sleeping beauties (Van Raan, 2004; Ke et al., 2015), assessing and even predicting future impact of inventions and discoveries (Wang et al., 2013; Redner, 1998; Uzzi et al., 2013; Acuna et al., 2012; Newman, 2014; Sinatra et al., 2016). This has become ever more so with the increasing scale, cost and complexity of science (Van Noorden et al., 2014; B\u00f6rner et al., 2004; \u0160ubelj and Fiala, 2016; Michels and Schmoch, 2012). Indeed, the exponential growth of science (Price de Solla, 1963; Sinatra et al., 2015; Van Noorden et al., 2014) suggests that there is now much more work for scientists to learn from, build upon, and cite, which is further exacerbated by intensifying specialization in science and engineering disciplines (Jones, 2011) and the inevitable dominance of interdisciplinary research spanning institutional and international boundaries (Jones et al., 2008; Evans and Reimer, 2009; Guimera et al., 2005; Deville et al., 2014; Adams, 2013). Furthermore, because citation systems are commonly treated as models of complex interconnected systems, understanding the time dimension of citations will also help deepen our quantitative understanding of complex systems by tightening models and observations that are highly generalizable to broad areas. 2 \f10 -1 B 10 -4 10 -1 P \u2190(t|2010) W. LN Exp. LN w. exp. 0 10 20 30 40 50 2010-t C 10 -2 10 -1 P \u2192(t|1950) t-1950 102 101 Shifted PL LN Exp. Eq (4) PL w. exp. World wars D 0 10 20 30 40 50 1955-t 10 -2 10 -1 P \u2190(t|1955) 10 -3 60 F 2010 2000 1990 1980 0 10 20 10 -1 -t t1 2 P \u2190(t2|t1) 40 10 -2 10 -3 E 60 ( ) P \u2192t1|t2 10 1 -t t1 2 10 0 G 10 2 10 -2 5 4 6 8 10 3 9 7 1910 1930 1950 1970 1990 2010 0 10 20 30 40 70 \u2190 <a > t \u2192 <a > 60 50 30 A 1950 2010 2000 1990 1970 1960 1980 P \u2190(t) Retrospective Prospective P \u2192(t) 10 4 10 5 10 6 M N 1970 1975 1980 1985 50 10 -3 10 -2 100 3 \fFigure 1 (preceding page): A The \ufb01rst approach, denoted by P \u2190(t), is also called \u2018synchronous distribution\u2019 (Nakamoto, 1988), \u2018recent impact method\u2019 (Vinkler, 2010) or \u2018citations from\u2019 approach as it measures the number of citations of a given age from a paper (Redner, 2005), or \u2018retrospective citation\u2019 approach as it takes a paper and looks retrospectively back in time (Burrell, 2002; Gl\u00e4nzel, 2004). Denoted with P \u2192(t), the second approach studies the age distribution of citations gained over time by a paper (or papers) published in a given year. Correspondingly, this method is also called \u2018diachronous distribution\u2019 (Nakamoto, 1988), \u2018subsequent impact method\u2019 (Vinkler, 2010) or \u2018citations to\u2019 approach (Redner, 2005), or \u2018prospective citation\u2019 approach (Burrell, 2002; Gl\u00e4nzel, 2004). Here the width of arrows is proportional to the logarithm of citations, the size of the outer (inner) circle corresponds to the logarithm of total number of references (publications). B Retrospective distribution for papers published in 2010, lognormal with exponential cuto\ufb00(prediction by Eq (6), LN w. exp.) outperforms Weibull (W., normalized likelihood ratio > 0 (599.3866), p \u2212value < 10\u221278016), lognormal (LN, normalized likelihood ratio > 0 (169.2839), p \u2212value < 10\u22126225) and exponential (Exp., normalized likelihood ratio > 0 (173.0337), p \u2212value < 10\u22126503). C Prospective distribution for papers published in 1950, prediction by Eq (4) outperforms lognormal (LN, normalized likelihood ratio > 0 (50.0941), p \u2212value < 10\u2212546) shifted power law (Shifted PL, normalized likelihood ratio > 0 (59.0745), p \u2212value < 10\u2212759), power law with exponential distribution (PL w. exp., normalized likelihood ratio > 0 (202.7706), p \u2212value < 10\u22128930) and exponential (Exp., normalized likelihood ratio > 0 (96.9191), p \u2212value < 10\u22122041). D Retrospective distribution for papers published in 1955, there are two dips due to world wars. E Average age of citations using retrospective and prospective approaches. Signi\ufb01cant \ufb02uctuations are observed for average age of citations until 1970. F Retrospective distribution, P \u2190(t2|t1), in di\ufb00erent years. G Prospective distribution, P \u2192(t1|t2), in di\ufb00erent years. Solid lines show a lognormal \ufb01t. 4 \fIt is, therefore, not surprising that this question has been extensively investigated over the past several decades (Redner, 2005; Evans, 2008; Wang et al., 2013; B\u00f6rner et al., 2004; Burton and Kebler, 1960; Gar\ufb01eld and Sher, 1963; de Solla Price, 1965; Margolis, 1967; MacRae Jr, 1969; Krauze and Hillinger, 1971; Avramescu, 1979; Stinson and Lancaster, 1987; Nakamoto, 1988; Motylev, 1989; Matricciani, 1991; Egghe and Rao., 1992; Gupta, 1997; Tsay, 1999; Pollman, 2000; Burrell, 2002; Gl\u00e4nzel, 2004; Sanyal, 2006; Simkin and Roychowdhury, 2007; Larivi\u00e8re et al., 2008; Bouabid, 2011; Verstak et al., 2014; Golosovsky and Solomon, 2014; Parolo et al., 2015; Pan et al., 2016), being one of most proli\ufb01c lines of inquiry in science of science studies. Yet, despite its broad impact and rich historical context, we lack systematic answers to a simple yet fundamental question: How does time a\ufb00ect citations? In this paper, we systematically investigate this question by leveraging large-scale citation datasets and recent models capturing citations dynamics. We start by conducting a comprehensive review of existing literature, which reveals a signi\ufb01cant lack of consensus on this matter. The main reason of this lack of consensus is that empirical measurements of temporal effect in citations have taken two related yet distinct measurement approaches (Fig. 1A). The \ufb01rst approach considers papers that are cited by a publication and analyzes retrospectively the age distribution of these references (Redner, 2005; Burton and Kebler, 1960; Gar\ufb01eld and Sher, 1963; de Solla Price, 1965; Margolis, 1967; MacRae Jr, 1969; Krauze and Hillinger, 1971; Avramescu, 1979; Stinson and Lancaster, 1987; Nakamoto, 1988; Motylev, 1989; Matricciani, 1991; Egghe and Rao., 1992; Gupta, 1997; Tsay, 1999; Pollman, 2000; Burrell, 2002; Gl\u00e4nzel, 2004; Larivi\u00e8re et al., 2008; Verstak et al., 2014; Golosovsky and Solomon, 2014; Pan et al., 2016). The second approach, in contrast, studies prospectively the age distribution of citations that are gained over time by a paper (Redner, 2005; Wang et al., 2013; MacRae Jr, 1969; Krauze and Hillinger, 1971; Avramescu, 1979; Stinson and Lancaster, 1987; Nakamoto, 1988; Motylev, 1989; Gl\u00e4nzel, 2004; Sanyal, 2006; Simkin and Roychowdhury, 2007; Bouabid, 2011; Golosovsky and Solomon, 2014; Parolo et al., 2015). The subtle di\ufb00erence between the two approaches creates a dramatic yet largely understudied e\ufb00ect in temporal citation patterns (Table 1), further confounded by the exponential growth in both the number of papers and references cited by them (Sinatra et al., 2015; Van Noorden et al., 2014; Pan et al., 2016). The temporal behavior of citations has been measured and reported by independent research groups, each using speci\ufb01c datasets and measurement details (Table 1). The lack of consensus, and the coexistence of the two approaches, raise an important question: What is the most appropriate functional form describing the time dimension of science? Here, by introducing a general theoretical framework, we provide systematic answers to this question, which is then validated both analytically and empirically through citation models and large-scale datasets. As such, the paper makes several contributions to this proli\ufb01c direction in the science of science. First and foremost, it serves as a detailed and much needed survey by reviewing and repeating results from most major studies in this domain. We then used the Web of Science (WoS) dataset to systematically measure and test results obtained by prior studies. We examine these results in the context of growth in science, systematically testing the most appropriate model describing aging measured by the prevailing two approaches. More importantly, despite the concomitant development of both approaches in the literature, we lack any known connections between the two. Here we introduced a general framework that helps us uncover that the two di\ufb00erent approaches are connected through precise mathematical relationships, allowing us to derive and even predict one approach from the other. We further derived our empirical results using citation models, providing theoretical support for the observed temporal behavior. We conclude by a brief discussion on scaling relationships between aging structure and citation impact of papers. Together, this paper presents a comprehensive study on the time dimension of science, providing a new empirical and theoretical basis for all further studies on this topic. These results are not only relevant for the emerging \ufb01eld of the science of science, they also help improve our understanding of complex systems, by pushing forward measurements and models in complex evolving networks. 2 Results A frequently applied measure for the time dimension of science is the age distribution of citations. Since 1960s, there have been numerous studies reporting on this distribution by taking one or both of the two approaches. While both approaches focus on the age of citations, they measure quantities pertaining to rather di\ufb00erent processes (Fig. 1A). Indeed, the retrospective approach, denoted by P \u2190(t), measures how far back the paper looks in time to cite old references, hence quantifying the citing memory of a paper. 5 \fPaper P \u2190(t) P \u2192(t) P \u2190 N (t) P \u2192 r (t) Burton and Kebler (1960) Exp. v3 \u2717 \u2717 \u2717 Gar\ufb01eld and Sher (1963) \u2014 \u2717 \u2717 \u2717 de Solla Price (1965) Exp. \u2717 Const. \u2717 Margolis (1967) Exp. \u2717 \u2717 \u2717 MacRae Jr (1969) Exp. Exp. Exp. \u2717 Krauze and Hillinger (1971) Exp. Exp. Exp. \u2717 Avramescu (1979) Exp. v2 \u2014 \u2717. \u2717 Stinson and Lancaster (1987) \u2014 \u2014 \u2014 \u2014 Nakamoto (1988) Exp. Exp. \u2717 \u2717 Motylev (1989) Exp. v3 \u2014 \u2014 \u2717 Matricciani (1991) LN \u2717 \u2717 \u2717 Egghe and Rao. (1992) LN>W., Exp. v2, NB \u2717 \u2717 \u2717 Gupta (1997) LN \u2717 \u2717 \u2717 Tsay (1999) Exp. \u2717 \u2717 \u2717 Pollman (2000) Exp. v1> Exp. \u2717 \u2717 \u2717 Burrell (2002) LN>LL, W. \u2717 \u2717 \u2717 Gl\u00e4nzel (2004) \u2014 PL \u2717 \u2717 B\u00f6rner et al. (2004) W. \u2717 \u2717 \u2717 Redner (2005) Exp. PL PL \u2014 Sanyal (2006) \u2717 LN \u2717 \u2717 Simkin and Roychowdhury (2007) \u2717 PL w. Exp. \u2717 \u2717 Larivi\u00e8re et al. (2008) \u2014 \u2717 \u2717 \u2717 Evans (2008) \u2014 \u2717 \u2717 \u2717 Bouabid (2011) \u2717 LL \u2717 \u2717 Wang et al. (2013) \u2717 LN \u2717 \u2717 Verstak et al. (2014) \u2014 \u2717 \u2717 \u2717 Golosovsky and Solomon (2014) \u2014 \u2014 \u2717 \u2717 Parolo et al. (2015) \u2717 Exp. v1>PL w. const \u2717 \u2717 Pan et al. (2016) Exp. \u2014 \u2014 \u2717 Table 1: Age distribution of citations: retrospective vs. prospective approaches. Not measured: \u2717. Measured but no functional form was o\ufb00ered: \u2014. See Table 2 for explanations of abbreviations. X > Y : X outperforms Y during \ufb01tting for P(t) Abbreviation Full Name Functional Form of P(t) Const. Constant 1 Exp. Exponential e\u2212\u03bbt Exp. v1 Exponential variant 1 \u03b2e\u2212\u03bbt + \u03b3 Exp. v2 Exponential variant 2 e\u2212\u03bbt \u2212e\u2212m\u03bbt Exp. v3 Exponential variant 3 ae\u2212x + 2(1 \u2212a)e\u22122x PL Power law t\u2212\u03b1 PL. w. const Power law with constant \u03b2t\u2212\u03b1 + \u03b3 Pow. w. Exp. Power law with exponential cuto\ufb00 t\u2212\u03b1e\u2212\u03bbt W. Weibull ( t \u03b1)\u03b2\u22121e\u2212( t \u03b1)\u03b2 NB Negative Binomial \u0393(k+t) \u0393(k)\u0393(t+1)pkqt LN Lognormal 1 te\u2212(ln(t)\u2212\u00b5)2 2\u03c32 LL Log-logistic (x/\u03b1)\u03b2\u22121 (1+(x/\u03b1)\u03b2)2 Table 2: Distributions appeared in Table 1 6 \fThe prospective approach P \u2192(t), on the other hand, measures how a paper is cited by others in the future, e\ufb00ectively measuring how the research community remembers a paper over time. Here, we review all major studies on this topic and summarize them in Table 1, asking speci\ufb01cally: what approach did they take? And what results did they obtain? Most early studies have been focused on the retrospective approach, a tradition dating back at least to Burton and Kebler (Burton and Kebler, 1960). Out of the seventeen papers that reported explicitly the functional form of P \u2190(t) in Table 1, twelve of them supported the exponential model or its variants to be the best \ufb01t to data. Since MacRae compared the retrospective and prospective citation age distributions in 1969 (MacRae Jr, 1969), more and more studies have started to focus on the prospective approach to measuring age distribution of citations. Comparing the two columns under P \u2190(t) and P \u2192(t) in Table 1 reveals that the prospective distributions are typically characterized by broader distributions than retrospective distributions, ranging from power law to power law with an exponential cuto\ufb00to lognormal distributions. Note that the di\ufb00erent results reported in Table 1 do not necessarily imply that these papers contradict with each other, because they strongly depend on measurement details as well as the corpus of papers on which these studies were carried out. The remarkably diverse conclusions summarized in Table 1, however, raise a tantalizing question: When it comes to citation age distributions, for both P \u2190(t) and P \u2192(t), what are the most appropriate functional forms that capture their temporal behavior? To answer this question, we take the Web of Science dataset, arguably the most authoritative citation index, to systematically measure both retrospective and prospective distributions. 2.1 Empirical Measurements To understand P \u2190(t), we \ufb01rst test the three most frequently applied distributions: Weibull, exponential and lognormal distributions (Table 2), asking which one best \ufb01ts our data. We use maximum likelihood estimation in this paper to determine the \ufb01t (Sec. S1). Since the tail part of the distribution is of the most interest and the citations in the \ufb01rst year after publication can be largely a\ufb00ected by publication date, we follow recent approaches by Parolo et al. (Parolo et al., 2015) and focus on data after the peak for the \ufb01t (typically occurring around 2 years). More speci\ufb01cally, We take all papers published in the same year (e.g. year 2010), trace retrospectively back in time, and measure the number of citations from the focal year to each earlier time unit. We \ufb01nd among the three functions tested, lognormal distribution provides the best \ufb01t for P \u2190(t) in terms of log-likelihood estimation, outperforming both Weibull and exponential distributions (Fig. 1B). However, none of them provides a satisfactory \ufb01t, showing an evident deviation for the tail part, hinting for alternative models that can \ufb01t P \u2190(t) more accurately. Indeed, we use a lognormal distribution with exponential cuto\ufb00, for reasons that will be explained later, and \ufb01nd it provides a better \ufb01t to the data (outperforming other distributions from likelihood ratio test with p \u2212value < 10\u22126225). For the prospective distribution P \u2192(t), we tested four types of functions reported in Table 1, from exponential distribution to lognormal to power law with and without exponential cuto\ufb00(Table 2). We also used shifted power law, i.e. P(t) \u223c(t0 + t)\u03b1, to \ufb01t our data. We calculated P \u2192(t) using the same set of papers as Figure 1B, \ufb01nding that lognormal function outperforms other distributions (Fig. 1C). Interestingly, power law with and without exponential cuto\ufb00yield similar results, because the obtained exponential cuto\ufb00 approaches its lower bound, playing a minor role in distinguishing the two distributions. Indeed, as we show later, the prospective distribution follows Eq (4), which not only gives a better \ufb01tting (outperforming other distributions from likelihood ratio test with p \u2212value < 10\u2212546), but also captures all the \ufb02uctuations in empirical data. Both distributions can be a\ufb00ected by signi\ufb01cant external shocks to the system. Indeed, repeating our analysis for year 1955 reveals two \u2018dips\u2019 in the retrospective distribution around 1942 and 1919, respectively (Fig. 1D), documenting interruptions created by wars. To quantify the extent to which the system is a\ufb00ected by external forces, we calculate the average age of all citations both from and to papers published at time t, denoted by \u27e8a\u27e9\u2190(t) and \u27e8a\u27e9\u2192(t), respectively. Figure 1E shows that \u27e8a\u27e9\u2190(t) shows larger \ufb02uctuations prior to 1970. Hence in this paper, we measure P \u2190for data from 1970 to 2013. For P \u2192, we use data after the Second World War, i.e. from 1948 to 2013. Figure 1E also demonstrates that the aging structure of citations changes over time (Sinatra et al., 2015; Verstak et al., 2014; Pan et al., 2016). To further understand temporal evolution of the aging structures, we calculate P \u2190(t) and P \u2192(t) for di\ufb00erent years t0. If we group paper samples based on their publication year t0, mathematically this can be described as P \u2190(t|t0) and P \u2192(t|t0). To this end we assume t1 is a more recent 7 \fyear than t2 (t1 > t2). Then P \u2190(t2|t1) measures the probability that a citation from t1 points to a reference published at time t2. Similarly P \u2192(t1|t2) measures the probability that a paper published in t1 is later cited at t2. We plot both distributions for di\ufb00erent years (Fig. 1F,G). For P \u2190(t2|t1), we \ufb01nd papers published in di\ufb00erent years after 1980s follow a similar citation pattern, with recent papers having a stronger tendency to cite \u2018old papers\u2019 (Sec. S2), which supports earlier \ufb01ndings by using Google Scholar dataset (Verstak et al., 2014). Comparing P \u2190(t2|t1) (Fig. 1F) with average reference ages for papers published in di\ufb00erent years (Fig. 1E), we \ufb01nd this increasing preference for old papers can be traced back to the 1970s, after leaving an unstable regime in the 1960s. Most important, we \ufb01nd in many cases P \u2192(t) is not a monotonically decreasing function. Indeed, we observed ubiquitous \u2018upward tails\u2019 characterizing temporal behavior of citations. For example, papers published in 1970 received 7.87% more citations in 2013 than they did in 2006. This indicates, as papers get older, they collect more instead of fewer citations. Repeating this analysis for di\ufb00erent years, we \ufb01nd the \u2018upward tails\u2019 characterizing P \u2192(t1|t2) is nearly universal (Fig. 1G). The upward tails o\ufb00er the strongest demonstration that none of the functions we have discussed so far (Table 1) are adequate in describing P \u2192(t). This phenomenon has been reported in prior works (Sinatra et al., 2015), but has received scarce attention in the literature. Next we show, understanding these tails not only helps us understand the role of growth in science in driving age distributions of citations, but also enables us to uncover the correct functional forms describing this temporal behavior. 2.2 Accounting for Growth P \u2190(t) and P \u2192(t) are determined by both aging of citations and growth of the citation system. Indeed, if we look further back in time, there are inherently fewer papers to cite, hence a\ufb00ecting P \u2190(t). Similarly, the continuous expansion of science impacts how an existing paper is remembered over time as new citations enter the system at an increasing rate. Let us denote with N(t), the number of papers published at time t. Extensive literatures documented that N(t) is best approximated by an exponential function: N(t) \u223ce\u03b2t (Wang et al., 2013; Sinatra et al., 2015; Van Noorden et al., 2014) (Fig. 2A). Besides the sheer increase in the number of papers, the average number of references contained in each paper, m(t), also increases with time (Fig. 2A inset). Indeed, if we de\ufb01ne M(t) as the number of citations from papers published at time t, i.e., M(t) \u2261N(t)m(t), to the leading order the growth of M can also be approximated by an exponential function (Fig. 2A). To account for the growth of science, we need to normalize the obtained retrospective distributions by the number of papers available for citations P \u2190 N (t2|t1) = P \u2190(t2|t1)/N(t2) R t1 0 P \u2190(t2|t1)/N(t2)dt2 , (1) where P \u2190 N (t2|t1) captures the fact that P \u2190(t2|t1) is normalized by the number of papers available for reference at time t2, N(t2). M(t1) new citations point towards R t1 0 N(t)dt papers in the system. Thus, for the prospective distribution, we normalize P \u2192(t) by the likelihood to be cited by new references, r(t1) \u2261 M(t1) R t1 0 N(t)dt, yielding P \u2192 r (t1|t2) = P \u2192(t1|t2)/r(t1) R T t2 P \u2192(t1|t2)/r(t1)dt1 (2) Note that r(t1) reduces to \u03b2m(t1) in the condition of N(t) \u223ce\u03b2t (Fig. 2A inset). Here we use N(t) to normalize retrospective distribution, which is equivalent to a null model that assumes every paper has the same chance to be cited. The prospective distribution, on the other hand, is normalized by r(t), indicating the probability for a paper to be cited at a given time is driven by both growth of papers and references. Only a modest amount of literature measured temporal e\ufb00ect on citations by taking into account growth factors (Table 1). To \ufb01nd the best functional form for P \u2192 r (t1|t2), we take all papers published in a year, measure the number of citations they receive in subsequent years, and normalize the distributions by r(t). For P \u2190 N (t), we select papers from year 2013, and normalize their P \u2190(t) by N(t). Although P \u2192(t) is characterized by a broader distribution than P \u2190(t), once normalized however, we \ufb01nd both distributions are best \ufb01tted by the 8 \fA C D E B 1900 1920 1940 1960 2000 1980 10 8 10 4 10 6 N(t) M(t) P \u2190(t|2013) 2013-t 10 -2 10 -1 10 -2 10 -1 P \u2192(t|1948) 10 -2 10 -1 P \u2192(t|1948) 101 t-1948 N 0 10 20 2013-t 30 40 Rescaled 10 -3 t-1948 r LN P \u2190(t|2013) Rescaled 100 101 100 LN 101 100 102 102 102 50 10 -2 10 -1 2020 10 7 10 3 10 5 t r(t) 0.5 1 1.5 0 t 10 20 30 0 1900 1920 1940 1960 2000 1980 2020 m(t) Figure 2: A Number of papers published and citations generated in every year grows exponentially. Inset: The growth in the average number of references from papers published every year. B Retrospective distribution normalized by N in 2013, with lognormal showing the best \ufb01t. C Retrospective distribution in 2013. The \ufb01tting is recovered from normalized distribution, i.e. lognormal with exponential cuto\ufb00is the best \ufb01t. D Prospective distribution normalized by r for papers published in 1948, with lognormal showing the best \ufb01t. E Prospective distribution in 1948, the \ufb01tting is recovered from normalized distribution. The best \ufb01t reproduces the upward tail. 9 \fsame function: lognormal distribution (Fig. 2BD). That is, P \u2190 N (t2|t1) \u223c 1 \u221a 2\u03c0\u03c3(t1 \u2212t2)e\u2212(ln(t1\u2212t2)\u2212\u00b5)2/2\u03c32 (3) P \u2192 r (t2|t1) \u223c 1 \u221a 2\u03c0\u03c3(t1 \u2212t2)e\u2212(ln(t1\u2212t2)\u2212\u00b5)2/2\u03c32 (4) Indeed, P \u2190(t) is simultaneously driven by two factors: the obsolescence of citations, P \u2190 N (t), and the growth of science, N(t). Since the latter leads to an exponential decay when we look backwards, the true aging function can be hidden from the rapid growth of science when measuring P \u2190directly. To this end, we use (1) and (3) to recover the functional form of P \u2190, obtaining P \u2190(t2|t1) \u223c 1 \u221a 2\u03c0\u03c3(t1 \u2212t2) e\u2212(ln(t1\u2212t2)\u2212\u00b5)2/2\u03c32\u2212\u03b2(t1\u2212t2) (5) where \u00b5 and \u03c3 are estimated from P \u2190 N . Note that \u03b2 characterizes the exponential growth of publications, which is a \ufb01xed system-wide parameter. Therefore, (5) has the same number of independent parameters as other \ufb01tting functions analyzed in this paper. Yet, when \ufb01tted to retrospective distributions, (5) o\ufb00ers a superior \ufb01t than all functions discussed in Table 1 (Fig. 2C). Furthermore, the exponential part dominates over the lognormal factor in the long term, explaining why it appears to be approximated by an exponential function in most earlier cases. Similarly, we recover the mathematical form of P \u2192from P \u2192 r , obtaining P \u2192(t1|t2) \u223c r(t1) \u221a 2\u03c0\u03c3(t1 \u2212t2) e\u2212(ln(t1\u2212t2)\u2212\u00b5)2/2\u03c32 (6) (6) not only provides an excellent \ufb01t to empirical data (Fig. 2E), it also predicts accurately the emergence of upward tails: since r(t1) increases with time, it competes with aging of citations, compensating the decay in (in some case even increasing) P \u2192(t). Using empirical data of r(t), we successfully reproduce these tails, documenting the validity of (6) (Fig. 2E). Note that when t1 \u2212t2 is small, \u2202ln(P \u2192(t1|t2)) \u2202t1 can be approximated as ln(t1\u2212t2)\u2212\u00b5+\u03c32 \u03c32(t1\u2212t2) , explaining why lognormal appears to \ufb01t P \u2192well when t1 \u2212t2 is small. Moreover, (6) indicates that, for growth factor to dominate over the aging factor, t1\u2212t2 needs to be large enough, explaining why upward tails are only evident for old papers. Taken together, our results show that the time dimension of science is di\ufb00erent from any result reported in Table 1: First of all, both distributions follow lognormal distributions after appropriate normalization. P \u2190(t) can be modeled by a lognormal distribution with exponential cuto\ufb00, where the cuto\ufb00is caused by the exponential growth of papers. P \u2192(t) is the product of a lognormal function and r(t), which lacks clear mathematical forms. A recent literature (Pan et al., 2016) suggests r(t) to be an exponential function, which means P \u2192(t1|t2) will be dominated by the growth of r(t) and increases in\ufb01nitely as t1 increases. However, if we assume r(t) to grow with a slower rate, e.g. following a linear or power law function (Ucar et al., 2014; Krapivsky and Redner, 2005; Vazquez, 2001; Allen et al., 1994), the upward tail is only temporary and P \u2192(t1|t2) will be eventually dominated by the lognormal decay from P \u2192 r (t). To test the robustness of (5) and (6), we further repeat our empirical measurement for thirty years (See Fig. S2,3 in Supplementary Information), \ufb01nding a universal pattern of citation decay across decades. Next we show, (5) and (6) not only o\ufb00er more accurate mathematical forms to \ufb01t the distributions, they also allow us to build a conceptually new framework which is closely related to several citation models. 2.3 Validations Given the new results obtained in (5) and (6), we must ask: how do we validate these results? Next we show two types of validations, by deriving both empirical and analytical evidence that supports the validity of (5) and (6). To validate (5) empirically, we use normalized retrospective distribution from di\ufb00erent years and rescaled them with respect to \u00b5 and \u03c3. (5) and (6) predicts that both distributions collapse into standard lognormal distribution, i.e. Z = 1 \u221a 2\u03c0X e\u2212(ln X)2/2, after a proper rescaling (Sec. S3). We test this hypothesis empirically 10 \ffor P \u2190 N , \ufb01nding excellent data collapse across di\ufb00erent years (Fig. 3A). We repeat this test for P \u2192 r using (6), \ufb01nding again an excellent data collapse on the same curve (Fig. 3B). The data collapses documented in Figure 3 o\ufb00er strong empirical evidence for (5) and (6). A 0 2 4 X B Z 0 0.2 0.4 0.6 0.8 0 2 4 X Z 0 0.2 0.4 0.6 0.8   1968 1978 1988 1998 2013 2003 1993 1983 Figure 3: Empirical support for P \u2190(t) and P \u2192(t). A P \u2190 N (t) rescaled with respect to \u00b5 and \u03c3. Distributions from di\ufb00erent years collapse on the probability density function (PDF) curve of standard lognormal distribution. B P \u2192 r (t) rescaled with respect to \u00b5 and \u03c3. Similar to A, distributions from di\ufb00erent years collapse on the PDF curve of standard lognormal distribution. What do existing models predict for P \u2190(t) and P \u2192(t)? To this end, we test three commonly used citation models with ascending levels of complexity: Price\u2019s model (Price, 1976), Bianconi-Barabasi model (Bianconi and Barab\u00e1si, 2001), and Wang-Song-Barabasi model (Wang et al., 2013). Price\u2019s model assumes that each new paper cites m references and the probability for an old paper to gain citations is proportional to its current citations plus a constant c0 (Sec. S4.1). We calculate P \u2190(t) and P \u2192(t) analytically for this model, \ufb01nding P \u2190(t2|t1) \u223c1 t1 \u0012 1 \u2212\u2206t t1 \u0013\u2212 m c0+m , (7) where \u2206t = t1 \u2212t2. (7) predicts that P \u2190(t2|t1) increases with \u2206t. That is, if we measure the citation age distribution following the retrospective approach at a given time t1, the distribution increases as \u2206t increases. This is due to the cumulative advantage mechanism. Indeed, if we look further back in time, older papers are more likely to be characterized by a higher degree hence more likely to be cited. For prospective approach, we obtain P \u2192(t1|t2) \u223cA(t2 + \u2206t)\u2212c0/(c0+m) (8) where A = 1 T m/(c0+m)\u2212tm/(c0+m) 2 . Therefore the age distribution decays as a power law \u2206t\u2212c0/(c0+m) as we look further in time. Interestingly, Price\u2019s model also predicts that P \u2190(t2|t1) collapse into a universal function for di\ufb00erent t1 after rescaling x \u2261t2/t1 (Fig. 4A): P \u2190(t2|t1) = c0 c0 + m 1 t1 x\u2212m/(c0+m) (9) P \u2192(t1|t2) can also be rescaled into a universal form as (Fig. 4B) P \u2192(t1|t2) = m c0 + mAt\u2212c0/(c0+m) 2 xc0/(c0+m) (10) The Bianconi-Barabasi (BB) model builds on Price\u2019s model by adding a \ufb01tness parameter to each paper, \u03b7i (Sec. S4.2). For retrospective approach, we \ufb01nd P \u2190(t2|t1) \u223ct1\u27e8\u03b7x\u2212\u03b7/C\u27e9, (11) 11 \fB C D A 100 101 102 103 104 105 10\u22125 10\u22124 10\u22123 10\u22122 100 101 102 103 104 105 10\u22125 10\u22124 10\u22123 10\u22122 P \u2190(t2|t1) ( ) P \u2192t1|t2 t1 t 50000 30000 40000 10001 1 5001 100 101 102 103 10\u2212510\u2212410\u2212310\u2212210\u22121100 P \u2190(t2|t1) t1 x ( )/A P \u2192t1|t2  t2  c0/(c0+m) 10 -2 10 -1 101 100 102 ( P \u2192t1|t2) ( P \u2190t2|t1) 10 -1 10 -2 10 -3 10 0 0 10 20 30 40 50 -t t1 2 -t t1 2 BB WSB Price\u2019s BB WSB Price\u2019s 2 10-3 10-2 10-1 100 10\u2212510\u2212410\u2212310\u2212210\u22121100 x Figure 4: Theoretical support for P \u2190(t) and P \u2192(t). A Price\u2019s model\u2019s prediction of P \u2190(t2|t1) for di\ufb00erent time units. Di\ufb00erent circles are our simulations for di\ufb00erent t1. Solid line shows our analytical calculation (7). Inset shows P \u2190(t2|t1) for di\ufb00erent years collapse into a single curve, solid line showing prediction of (9). B Price\u2019s model\u2019s prediction of P \u2192(t1|t2) for di\ufb00erent years. Di\ufb00erent circles are our simulations for di\ufb00erent t2. Solid line shows our analytical calculation (8). Inset shows P \u2192(t1|t2) for di\ufb00erent years collapse into a single curve, solid line showing prediction of (10). C P \u2190(t2|t1) predicted by Price\u2019s model, Bianconi-Barabasi model and Wang-Song-Barabasi model. The WSB model predicts the correct function over time consistent with our empirical results. D P \u2192(t1|t2) predicted by Price\u2019s model, BianconiBarabasi model and Wang-Song-Barabasi model. Price\u2019s model and BB model predict approximately power law decay, while the WSB model is able to reproduce the upward tails given r(t). where x \u2261t2/t1, same as the rescaled time we previously de\ufb01ned for Price\u2019s model and C is a normalization constant. Hence the BB model too predicts a retrospective distribution that increases with time. For prospective distribution, we \ufb01nd P \u2192(t1|t2) decays as P \u2192(t1|t2) \u223c\u27e8\u03b7x1\u2212\u03b7/C/t2\u27e9/(\u27e8(T/t2)\u03b7/C\u27e9\u22121). (12) Eqs (11) and (12) also indicate that the shape of both distributions depends on the distribution of \ufb01tness. We calculate both distributions by assuming an exponential \ufb01tness distribution, i.e. P(\u03b7) \u223ce\u2212\u03b7/\u03b70, a commonly used functional form (Kong et al., 2008). We \ufb01nd that P \u2192(t1|t2) decays as a power law of (\u2206t)\u22121, while P \u2190(t2|t1) again increases with time, albeit with a slower rate (Fig. 4CD). Both distributions again collapse into universal forms if we use x = t2/t1: P \u2190(t2|t1) = 1 Ct1\u03b70 \u0012 1 \u03b70 + 1 C ln x \u0013\u22122 (13) P \u2192(t2|t1) = \u03b12x Ct2(1 \u2212\u03b12\u03b70) \u0012 1 \u03b70 + 1 C ln x \u0013\u22122 (14) where \u03b12 = 1/\u03b70 \u2212ln(T/t2)/C. 12 \fFinally we repeat our calculations for the Wang-Song-Barabasi (WSB) model (Wang et al., 2013), which builds on the BB model but further incorporates an aging mechanism with the exponential growth of the system (Sec. S4.3). We \ufb01nd prospective distribution follows a lognormal distribution: P \u2192(t2|t1) \u223c 1 \u221a 2\u03c0\u03c3\u2206te\u2212(ln(\u2206t)\u2212\u00b5)2/2\u03c32, (15) and the retrospective distribution follows lognormal distribution with exponential cuto\ufb00, i.e. P \u2190(t1|t2) \u223c 1 \u221a 2\u03c0\u03c3\u2206te\u2212(ln(\u2206t)\u2212\u00b5)2/2\u03c32\u2212\u03b2\u2206t. (16) Somewhat surprisingly, these theoretical predictions are in remarkable agreement with our empirical results (Fig. 4CD). The accuracy of the WSB model is partly due to its ability to incorporate both aging and growth factors into the model. Together these results provide strong theoritical support for (5) and (6). 2.4 Connecting the Two Approaches The retrospective and prospective approaches are studied concomitantly in the literature, often pursued as separate lines of inquiry, which raise an interesting question: Are these two distributions connected? Indeed, after all, they both measure the age of citations. Should we have measured the age distribution of all citations, both approaches would result in the same distribution, as if it did not matter whether we look retrospectively or prospectively in time. Yet despite the large amount of work on this topic (Table 1), little is known if there exists any relationship between the two approaches. To shed light on this question, we calculate P(t1, t2), measuring the probability that a randomly selected citation connects a paper published in t1 with a paper published in t2. Because the di\ufb00erence between the two approaches lies in whether we \ufb01x the citing or cited papers, we can derive their citation age distribution from P(t1, t2). Next we show that expressing P \u2190or P \u2192in terms of P(t1, t2) allows us to not only connect the two approaches, but also derive one approach from the other. Indeed, P(t1, t2) can be expressed as P(t1, t2) \u223cP \u2190(t2|t1)M(t1) = P \u2192(t1|t2)L(t2) (17) where L(t2) = R T t2 P \u2190(t2|t1)M(t1)dt1 measures the total number of citations papers published at t2 receive. Therefore, (17) documents a precise mathematical connection between P \u2192(t1|t2) and P \u2190(t2|t1), mediated by P(t1, t2), indicating one can obtain P \u2192(t) from P \u2190(t) or vice versa. Indeed, using (17) we obtain a rescaling formula for P \u2192(t): P \u2192(t1|t2) = P \u2190(t2|t1)M(t1) R T t2 P \u2190(t2|t1)M(t1)dt1 , (18) which allows us to calculate the prospective distribution based solely on retrospective distributions (Sec. S5). Taking papers published in 1980 as an example, we collected P \u2190(1980|t) for t > 1980 and rescaled them using (18). We then compare the rescaled results with real P \u2192(t), \ufb01nding an excellent match between the two, reproducing not only the decay rate but also the upward tail (Fig. 5A). Similarly, we can solve for for P \u2190(t) through proper rescaling of P \u2192(t) using (17) (Sec. S5). To this end we need to calculate L(t) in terms of P \u2192(t) from Z T 0 P \u2192(t1|t2)L(t2)dt2 = M(t1). (19) (19) is the Fredholm integral equation of the \ufb01rst kind (Wazwaz, 2011), which lacks analytical solution. We can, however, consider a discretized version, obtaining the rescaling formula for P \u2190(t): P \u2190(t2|t1) = P \u2192(t1|t2)[(P \u2192)\u22121M](t2) M(t1) (20) where P \u2192= (P \u2192(i|j)\u0398(i \u2212j))1\u2264i,j\u2264n M = [N(1)m(1) \u00b7 \u00b7 \u00b7 N(T )m(T )]T 13 \fTo test its validity, we use (20) to numerically calculate the rescaled P \u2192, and compare it with real P \u2190. More speci\ufb01cally, we take papers published in 1980, and measure all prospective distributions for every year between 1950 to 1980, i.e. P \u2192(t1|t2) for t2 ranging from 1950 to 1980. We then plug them into (20) to calculate the corresponding retrospective distributions. The obtained result yields excellent agreement with our empirical measurement (Fig. 5B). The remarkable agreement documented in Figure 5AB raises an interesting question: does our rescaling method have predictive power as well? To test this, we predict the prospective distribution in year 1990 by using retrospective distributions and the number of new citations prior to 1990. That is, we estimate values of M(t) and P \u2190(t|t0) for t = 1990 and t0 in [1990, 2010] in advance, seeking to obtain P \u2192(t|t0) following (18). Figure 2A indicate that M(t) is the product of N(t), and m(t). P \u2190(t|t0), on the other hand, follows a lognormal distribution with exponential cuto\ufb00, whose parameters in (5) can be time-dependent. To test the feasibility of our approach, we impose a minimum set of assumptions to measure M(t) and P \u2190(t|t0), to check the degree to which the predicted P \u2192(t|t0) matches the real distributions. More speci\ufb01cally, for M(t), we use an exponential function \u02c6 M(t) to \ufb01t the growth of references. We use a lognormal distribution with exponential cuto\ufb00to \ufb01t P \u2190(t|t0). To test the validity of our rescaling formula, we tried three models to predict P \u2192(Sec. S6.1). Model 1: we relax M and predict P \u2192based on simulated \u02c6 M and real P \u2190. Next, we build model 2 by relaxing P \u2190 and hence using real M and simulated \u02c6 P \u2190. Model 3 uses both simulated \u02c6 M and \u02c6 P \u2190as approximations. Comparing with real data, we witness strong predictive power from all three models (Fig. 5C). B 10 -2 10 -1 P \u2190(t|1980) 0 10 20 30 1980-t Empirical Rescaled C 10 -1 t-1990 P \u2192(t|1990) 10 1 10 0 A 10 -2 10 -1 P \u2192(t|1980) t-1980 1 10 10 0 Empirical Rescaled 10 2 10 -2 Model 1: empirical and estimated P \u2190 M Model 2: estimated and empirical P \u2190 M Model 3: empirical and empirical P \u2190 M Figure 5: Validations of (18) and (20). A Using P \u2190(1980|t) data to calculate P \u2192(t|1980) through rescaling formula (18). B Using P \u2192(1980|t) data to calculate P \u2190(t|1980) through rescaling formula (20). C We use history data to estimate M(t) and P \u2190(1990|t) and predict P \u2192(t|1990) through rescaling formula (18). Model 1 is based on estimated M and empirical P \u2190, Model 2 is based on empirical M and estimated P \u2190, while Model 3 is based on estimated M and estimated P \u2190. All three models show great predictive power. Our rescaling formula is also consistent with the derived functional form reported in preceding sections of this paper. Indeed, provided that P \u2190follows (5), (18) predicts that P \u2192follows the same functional form as (6), which is consistent with our empirical measurement (Sec. S6.2). 2.5 Redner\u2019s Ansatz How does the citation aging structure of a paper connect to its citations? To answer this question, we consider an interesting ansatz raised by Redner (Redner, 2005): there appears to exist a power law relationship between average citation age and total citations a paper receives. Redner took all papers with 500 or fewer citations published in the American Physics Society (APS) dataset, and found that the average age of all citations to a paper, i.e. \u27e8a\u27e9\u2192, grows with the number of citations its receives, c, following \u27e8a\u27e9\u2192\u223cc\u03b1. We repeated this analysis on the WoS dataset, \ufb01nding \u27e8a\u27e9\u2192and c indeed follow a power law scaling (\u03b1 \u22480.13) in the small to medium c regime (c < 500) (Fig. 6A), thus independently con\ufb01rming Redner\u2019s ansatz. We also \ufb01nd 14 \fsigni\ufb01cant deviations in the large c regime (c > 500), where \u27e8a\u27e9\u2192appears to increase with c at a faster rate than power law. 101 102 104 c 101 102 103 100 101 102 100 10-1 \u03bb=3.0, \u03bc=8.8, \u03c3=1.2 \u03bb=1.9, \u03bc=7.5 ,\u03c3=0.9 \u03bb=6.7, \u03bc=9.2, \u03c3=1.0 A c <a>\u2192 B 100 101 103 102 <a>\u2192 0.2 0.4 0.6 0.8 1  1.2 1.4 1.6 1.8 2 1  2  3  4  5  6  7  8  9  10 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 \u03c3 \u03bb Figure 6: Redner\u2019s ansatz. A Average age growth with citations of all papers on the WoS dataset (black circle) can be well \ufb01tted by power law (grey line, exponent is 0.13), we also see similar scaling when papers are grouped according to their ages (colored lines). Blue lines correspond to earlier years (1900-1902), while red lines correspond to later years (1978-1980). We also observe a decrease of average citation age for recent years since these papers can only be cited by more recent papers. B Simulation of average citation growth for di\ufb00erent (\u03bb, \u00b5, \u03c3) using Wang-Song-Barabasi model, the power law scaling is observed in a long range. C Varying \u03bb and \u03c3 parameters in the WSB model can result in a wide range of scaling exponent. Color indicates the scaling exponent measured as slopes in log-log plots similar to A and B. To test the robustness of this ansatz, we grouped papers based on their publication times and measure the average citation age for papers with di\ufb00erent citations within each group, \ufb01nding again a power law scaling between c and \u27e8a\u27e9\u2192(Fig. 6A). Next, we show that this power law scaling can be explained by the inherent citation dynamics individual papers follow. To approximate the citation dynamics of a paper, we use the WSB model to generate citations of a paper over time and measure their relationship with the average citation age (Sec. S7). As we vary its (\u03bb, \u00b5, \u03c3) parameters, we \ufb01nd that across a wide range of c values, \u27e8a\u27e9\u2192can be approximated by power law relationships. The model also successfully captures the acceleration phase of the relationship in large c region (Fig. 6B). The scaling exponent of the power law relationship \u03b1 depends on the speci\ufb01c (\u03bb, \u00b5, \u03c3) parameters, hence varying (\u03bb, \u00b5, \u03c3) through the parameter space allows us to reproduce the scaling behavior with a wide range of exponents (Fig. 6C). 3 Discussion Understanding the temporal behavior of citations is of fundamental importance. In this paper, we carried out a large-scale quantitative analysis on how time a\ufb00ects citations. We began by reviewing and comparing the most important papers on this topic, \ufb01nding signi\ufb01cant inconsistency among prior studies in terms of both their measurement approaches and conclusions. To systematically measure citation age distributions, we used the Web of Science dataset to test di\ufb00erent functional forms through maximum likelihood estimation. We found neither of the two prevailing approaches taken by current studies can be well captured by existing models. Moreover, we found ubiquitous upward tails characterizing P \u2192, indicating that citation age distributions are a\ufb00ected by growth of papers and citations. To account for such impact, we further de\ufb01ned and calculated normalized distributions for both approaches. We found both normalized distributions can be well captured by lognormal functions, which not only allows us to \ufb01t age distributions more accurately, but also explains upward tails observed in empirical data. Given that all the distributions reported in Table 1 are inherently quite similar (Fig. 1), here instead of focusing on visually how distributions \ufb01t the data, we aimed to look at this question through a theoretical lens. Indeed, we situated the observations within a theoretical framework connecting both distributions through mathematical relationships and network models, allowing us to rescale both distributions and reason about the most appropriate functional forms. Indeed, the main virtue of this paper lies in the theoretical 15 \fframework that helps us connect the retrospective and prospective approaches. We validated the newly derived distributions both empirically and theoretically. Through theoretical and empirical validations, we show that either distribution can be mathematically obtained from another distribution because of deeper underlying connections between the two approaches. We then placed our results in the context of citation models, analytically calculating the citation age distributions for both approaches in three citation model systems. Compared with our empirical results, we \ufb01nd the Wang-Song-Barabasi model provides an accurate prediction for the functional forms of citation age distributions we obtained, o\ufb00ering theoretical support for our empirical \ufb01ndings. On the other hand, our results also provide additional evidence supporting the accuracy of the WSB model in characterizing the temporal behavior of citations. We concluded the paper with a discussion on the Redner\u2019s ansatz that postulates a power law scaling between citations and average citation ages, \ufb01nding it again can be well captured by the WSB model. Taken together, this paper provides systematic answer to the mathematical form of citation age distributions. We believe the results reported in this paper provide a new empirical and theoretical basis for future studies on temporal analyses of citations, and will play an increasingly important role as our quantitative understanding of science deepens. Future directions include extending our analyses to examing their dependence on contextual factors, such as disciplines, countries or publishing venues, remains unknown, which could o\ufb00er a more holistic understanding of the time dimension of science. 4 ACKNOWLEDGMENTS The authors wish to thank Zhen Lei, Chaoming Song, and A.-L. Barab\u00e1si for helpful discussions. This work is supported by the Air Force O\ufb03ce of Scienti\ufb01c Research under award number FA9550-15-1-0162 and FA9550-17-1-0089."
+                }
+            ],
+            "Sijia Liu": [
+                {
+                    "url": "http://arxiv.org/abs/2402.08787v3",
+                    "title": "Rethinking Machine Unlearning for Large Language Models",
+                    "abstract": "We explore machine unlearning (MU) in the domain of large language models\n(LLMs), referred to as LLM unlearning. This initiative aims to eliminate\nundesirable data influence (e.g., sensitive or illegal information) and the\nassociated model capabilities, while maintaining the integrity of essential\nknowledge generation and not affecting causally unrelated information. We\nenvision LLM unlearning becoming a pivotal element in the life-cycle management\nof LLMs, potentially standing as an essential foundation for developing\ngenerative AI that is not only safe, secure, and trustworthy, but also\nresource-efficient without the need of full retraining. We navigate the\nunlearning landscape in LLMs from conceptual formulation, methodologies,\nmetrics, and applications. In particular, we highlight the often-overlooked\naspects of existing LLM unlearning research, e.g., unlearning scope, data-model\ninteraction, and multifaceted efficacy assessment. We also draw connections\nbetween LLM unlearning and related areas such as model editing, influence\nfunctions, model explanation, adversarial training, and reinforcement learning.\nFurthermore, we outline an effective assessment framework for LLM unlearning\nand explore its applications in copyright and privacy safeguards and\nsociotechnical harm reduction.",
+                    "authors": "Sijia Liu, Yuanshun Yao, Jinghan Jia, Stephen Casper, Nathalie Baracaldo, Peter Hase, Xiaojun Xu, Yuguang Yao, Hang Li, Kush R. Varshney, Mohit Bansal, Sanmi Koyejo, Yang Liu",
+                    "published": "2024-02-13",
+                    "updated": "2024-04-05",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.CL"
+                    ],
+                    "main_content": "Introduction Large language models (LLMs) have shown exceptional proficiency in generating text that closely resembles humanauthored content. However, their ability to memorize extensive corpora may also lead to ethical and security concerns. These include societal biases and stereotyping (Bender et al., 2021; Motoki et al., 2023; Kotek et al., 2023), the generation of sensitive, private, harmful, or illegal content (Nasr et al., 2023; Wen et al., 2023; Karamolegkou et al., 2023; Patil et al., 2024), ease of jailbreaking (Wei et al., 2023; Zou et al., 2023; Liu et al., 2023a), and possible malicious use in developing cyberattacks or bioweapons (Barrett et al., 2023; Hendrycks et al., 2023). These concerns emphasize the need to adeptly and efficiently tailor pre-trained LLMs to suit diverse safety contexts while meeting specific requirements of users and sectors. With the costly and prolonged training periods of LLMs, retraining these models to eliminate undesirable data effects is often impractical. Machine unlearning (MU) has emerged as an alternative to remove the influence of undesirable data and associated model capabilities from the pre-trained models (Cao & Yang, 2015; Bourtoule et al., 2021; Nguyen et al., 2022; Si et al., 2023; Zhang et al., 2023a; Eldan & Russinovich, 2023; Yao et al., 2023). For example, MU is used as a strategy to prevent the generation of copyrighted material from the Harry Potter series in (Eldan & Russinovich, 2023). In the context of classification tasks, MU has been extensively studied by Ginart et al. (2019); Neel et al. Correspondence to: Sijia Liu (liusiji5@msu.edu), Yang Liu (yangliu@ucsc.edu). arXiv:2402.08787v3  [cs.LG]  5 Apr 2024 \fRethinking Machine Unlearning for LLMs Pre-training Pre-alignment In-alignment Post-alignment Gradient Ascent Influence Function Finetuning Pretrained LLM Data Aligned LLM Red Teaming Localization Informed Alignment Safety Alignment  Privacy Compliance Copyright Removal Bias Mitigation Why? Hallucination Removal Where How Evaluation LLM  Pipeline User Feedback Figure 1: Demonstration of how MU can be incorporated into LLM development cycle. The landscape of LLM unlearning will be mainly navigated from applications (\u2018why\u2019), methods (\u2018where\u2019 and \u2018how\u2019), and evaluations. (2021); Ullah et al. (2021); Sekhari et al. (2021); Golatkar et al. (2020); Jia et al. (2023). Yet, its application and understanding in LLMs remains limited, where models are typically used for generative tasks such as summarization, sentence completion, paraphrasing, and question answering. Therefore, this paper specifically concentrates on exploring the MU problems in LLMs, termed \u2018LLM Unlearning\u2019. As data-model scales continue to grow, the emergence of LLM unlearning introduces new challenges and complexities, as will be elaborated on in Sec. 2. For example, current research efforts (Lu et al., 2022; Jang et al., 2022; Ilharco et al., 2022; Eldan & Russinovich, 2023; Wu et al., 2023b; Yu et al., 2023; Zhang et al., 2023c; Yao et al., 2023) suffer from a lack of standardized corpora and evaluation methods for LLM unlearning, often facing issues related to unlearning precision and efficiency. We also refer readers to Sec. 3 for a summary of existing LLM unlearning tasks. Although preliminary surveys of LLM unlearning have been provided by Si et al. (2023) and Zhang et al. (2023a), this paper is, to the best of our knowledge, the first to offer a thorough and in-depth analysis of LLM unlearning. Our contributions are summarized below. Fig. 1 provides an overview of the LLM unlearning landscape that we explore. (1) Surveying: We conduct an in-depth review of the foundational concepts and principles of LLM unlearning, delving into the problem formulation, categories of unlearning methods, evaluation approaches, and practical applications. (2) Uncovering: We bring to light previously overlooked dimensions of LLM unlearning, e.g., emphasizing the significance of precisely defining the unlearning scope, elucidating the interplay between data and model interactions, and exploring the adversarial assessment of unlearning efficacy. (3) Connecting: We establish connections between LLM unlearning and other relevant problems and domains, providing a comparative analysis with related topics such as model editing, influence function, and adversarial learning. (4) Forecasting: We offer insights into the future of LLM unlearning by identifying novel opportunities & challenges. This paper is positioned to reassess the challenge of LLM unlearning, refining its scope across various dimensions: conceptual formulation (Sec. 3), methods (Sec. 4), assessment (Sec. 5), and applications (Sec. 6); see the schematic overview in Fig. 1. We delve into each dimension through surveying, uncovering, connecting, and forecasting. We conclude that unlearning will be a valuable tool for making LLMs more trustworthy, but making more progress on this will require updating the unlearning paradigm. We aspire for this work to pave the way for developing LLM unlearning, illuminating its opportunities, challenges, and untapped potential. 2 Preliminaries and Related Work LLM unlearning has garnered attention for addressing trustworthiness concerns such as toxicity (Lu et al., 2022), copyright and privacy (Jang et al., 2022; Eldan & Russinovich, 2023; Wu et al., 2023b), fairness (Yu et al., 2023), model editing (Ilharco et al., 2022; Zhang et al., 2023c), hallucination (Yao et al., 2023), and sensitive knowledge (Barrett et al., 2023; Hendrycks et al., 2023). In this section, we present a succinct overview of MU, tracing its journey from traditional ML models to the emerging challenges in LLMs. 2 \fRethinking Machine Unlearning for LLMs MU for non-LLMs. The study of MU can be traced back to non-LLMs in response to data protection regulations such as \u2018the right to be forgotten\u2019 (Cao & Yang, 2015; Hoofnagle et al., 2019; Bourtoule et al., 2021; Nguyen et al., 2022). Due to its capability of assessing data influence on model performance, the landscape of MU has expanded to encompass diverse domains, such as image classification (Ginart et al., 2019; Golatkar et al., 2020; Neel et al., 2021; Ullah et al., 2021; Sekhari et al., 2021), text-to-image generation (Gandikota et al., 2023; Zhang et al., 2023b; Kumari et al., 2023; Fan et al., 2024), federated learning (Liu et al., 2020; Wang et al., 2022; Che et al., 2023; Liu et al., 2023b; Halimi et al., 2022), and graph neural networks (Chen et al., 2022; Chien et al., 2022; Wu et al., 2023a). In the literature, \u2018exact\u2019 unlearning, which involves retraining the model from scratch after removing specific training data points, is often considered the gold standard. However, this approach comes with significant computational demands and requires access to the entire training set (Thudi et al., 2022). To address these challenges, many research efforts have shifted towards the development of scalable and effective approximate unlearning methods (Golatkar et al., 2020; Warnecke et al., 2021; Becker & Liebig, 2022; Thudi et al., 2022; Jia et al., 2023; Chen et al., 2023a). In addition, probabilistic methods with certain provable removal guarantees have been explored, often leveraging the concept of differential privacy (Ginart et al., 2019; Guo et al., 2019; Neel et al., 2021; Ullah et al., 2021; Sekhari et al., 2021). Challenges of MU for LLMs. LLM unlearning introduces new challenges and complexities. First, LLMs are trained on massive amounts of data, which can unintentionally introduce biases and the memorization of personal and confidential information. Accordingly, it becomes challenging to precisely define and localize the \u2018unlearning targets\u2019, such as the subset of the training set or a knowledge concept that needs to be removed. Therefore, current studies on LLM unlearning (Lu et al., 2022; Jang et al., 2022; Ilharco et al., 2022; Eldan & Russinovich, 2023; Wu et al., 2023b; Yu et al., 2023; Zhang et al., 2023c; Yao et al., 2023) are typically context and task-dependent. There is a lack of standardized corpora for LLM unlearning. Second, the growing size of LLMs and the rise of black-box access to LLM-as-a-service present challenges in developing scalable and adaptable MU techniques to LLMs (Bucknall & Trager, 2023; Casper et al., 2024a). This also affects performance evaluation, given the absence of retraining as a benchmark. To address these challenges, previous studies have proposed approaches like in-context unlearning (Pawelczyk et al., 2023) and fictitious unlearning (Maini et al., 2024), where the former enables unlearning on black-box models, and the latter provides alternatives to retraining. Third, the scope of unlearning is often underspecified for LLMs. This issue is similar to challenges faced in model editing (Mitchell et al., 2022). For instance, effective unlearning should ensure that LLMs delete knowledge of the targeted data within the predefined scope while simultaneously maintaining its utility for data outside of this scope. Fourth, despite the potential of LLM unlearning in diverse applications, there is a notable absence of comprehensive and reliable evaluation. For example, recent studies (Shi et al., 2023; Patil et al., 2024) have demonstrated that sensitive information can be reverse-engineered from an edited model, even if efforts were made to delete this information. This highlights the need for thorough and adversarial evaluations and the design of more mechanistic methods to guarantee the authenticity of unlearning. 3 Unpacking LLM Unlearning In this section, we formalize the LLM unlearning problem and delve into its configuration. We will outline the inherent facts and complexities in its formulation. Problem statement. In light of the literature on unlearning (Bourtoule et al., 2021; Jia et al., 2023; Kurmanji et al., 2023), and its progression in LLMs (Pawelczyk et al., 2023; Yao et al., 2023; Ishibashi & Shimodaira, 2023; Maini et al., 2024), we define the problem of LLM unlearning below. (LLM unlearning) How can we efficiently and effectively eliminate the influence of specific \u2018unlearning targets\u2019 and remove associated model capabilities while preserving model performance for non-targets? We dissect the above statement from the perspectives: (1) unlearning targets, (2) influence erasure, (3) unlearning effectiveness, and (4) efficiency. See Table 1 for a summary of existing LLM unlearning studies based on (1)-(4). \u2022 Unlearning targets: Unlearning tasks may take on various forms and are closely related to the unlearning objectives. For instance, one could focus on data influence removal, while the other could emphasize model capability removal. Although these two aspects are intertwined, the former is often crucial for intellectual property (IP) protection, while the latter is more practical for AI alignment and safety. The literature identifies unlearning targets as specific data points, which could involve content containing harmful, unethical, or illegal language (Jang et al., 2022; Wu et al., 2023b). They have also been represented by higher-level unlearned knowledge, expressed through an unwanted text prompt or 3 \fRethinking Machine Unlearning for LLMs Table 1: A summary of existing LLM unlearning problems through unlearning targets, influence erasure, effectiveness, and efficiency. An asterisk (\u2217) indicates the incapability of evaluating unlearning for LLMs due to the impracticality of retraining these models. Related work Unlearning targets/tasks Influence erasure methods Effectiveness: (I) In-scope evaluation for unlearning efficacy (O) Out-of-scope evaluation for model utility Efficiency (Lu et al., 2022) Reducing toxic content, avoiding undesirable sentiments, and preventing repeated text generation Reward-reinforced model fine-tuning (I) Toxic prompts, specific sentiments, & repetitive sentences (O) Unlearning target-irrelevant prompts N/A (Jang et al., 2022) Degenerating private information, w/ unlearning response irrelevant to this info Gradient ascent-based fine-tuning (I) Prompts from training data extraction (O) Natural language understanding tasks Runtime cost (Kumar et al., 2022) Text de-classification, w/ unlearning response close to that of retraining\u22c6 Sharded, isolated, sliced, and aggregated (SISA) training via adapter (I) No evaluation for unlearning efficacy (O) Test set Runtime cost Memory cost (Ilharco et al., 2022) (Zhang et al., 2023c) Degenerating toxic content Task vector-based parameter-efficient fine-tuning via LoRA (I) Prompts leading to toxic generation (O) Perplexity on other datasets N/A (Wang et al., 2023) Text de-classification/de-generation, unlearning specific words in translation, w/ response close to that of retraining\u2217 KL-divergence-based fine-tuning (I) Training subset (O) Test set Runtime cost (Yu et al., 2023) Unlearning gender and profession bias, with de-biased unlearning response Weight importance-informed & relabeling-based fine-tuning (I) Biased prompts (O) No evaluation for model utility N/A (Pawelczyk et al., 2023) Text de-classification, w/ unlearning response close to that of retraining\u2217 In-context learning (I) Training subset (O) Retain & test sets Black-box access (Eldan & Russinovich, 2023) Degenerating Harry Potter-related book content, w/ unlearning response irrelevant to Harry Potter Relabeling-based fine-tuning (I) Questions and their rephrased/hard versions about Harry Potter (O) NLU tasks N/A (Ishibashi & Shimodaira, 2023) Unlearning knowledge from QA dataset, with refusal response (e.g., \u2018I don\u2019t know\u2019) Relabeling-based fine-tuning (I) Adversarial and original questions about forgotten knowledge (O) Other QA prompts N/A (Chen & Yang, 2023) Text de-classification and de-generation, with response close to that of retraining\u2217 KL divergence-based parameterefficient fine-tuning via adapter (I) Training subset (O) Retain & test sets Runtime cost (Wu et al., 2023b) Degenerating private information, w/ unlearning response irrelevant to this info Importance-based neuron editing (I) Memorized private data points (O) Test set Runtime cost (Yao et al., 2023) Degenerating harmful prompts, degenerating Harry Potter-related book content, and reducing hallucination Integration of gradient ascent, random labeling, & KL divergence-based fine-tuning (I) Prompts related to unlearning targets (O) NLU tasks Runtime cost (Maini et al., 2024) Unlearning biographical knowledge about fictitious authors. Fine-tuning with various objectives (I) Q&A about the unlearning authors (O) Q&A about other authors and general facts Runtime cost (Patil et al., 2024) Degenerating sensitive information using factual information as a testbed Model editing techniques and constrained finetuning (I) Prompts for unlearned factual knowledge (O) Prompts for unrelated factual knowledge White-box v. black-box access concept (Lu et al., 2022; Yao et al., 2023; Eldan & Russinovich, 2023). For example, Eldan & Russinovich (2023) defined the unlearning target as \u2018Harry Potter\u2019-related content, with the objective to avoid generating such content irrespective of where the content was learned: from the copyrighted material, blog posts, or news articles. \u2022 Influence erasure: Erasing the influence of unlearning targets and associated model capabilities requires a joint examination of both data and model influences rather than studies in isolation. Specifically, it is important to scrutinize the contributions of data sources to undesired model outputs, as well as the roles played by individual components within a model in generating these undesirable outcomes. This dual examination allows us to gain a more comprehensive understanding of the mechanisms driving these outputs, thereby facilitating the development of unlearning strategies to prevent them effectively. The objective of achieving complete influence erasure also implies the importance of robustness and generalization in unlearned behavior. When evaluating LLM unlearning, especially when using approximate methods shown in Table 1, a rigorous criterion is needed. Patil et al. (2024) underscore this viewpoint by demonstrating that sensitive information can persist in model weights after editing or unlearning, and by Maini et al. (2024) and Shi et al. (2023) who expose the limitations of conventional unlearning methods. \u2022 Unlearning effectiveness: The effectiveness of LLM unlearning extends beyond merely diminishing the specific data influence or model capability. A crucial aspect of effectiveness is the concept of unlearning scope, as inspired by the editing scope (Mitchell et al., 2022). This has implications for the development of metrics to evaluate the effectiveness of unlearning and the design of effective unlearning methods. The unlearning scope defines the success of influence erasure by its capacity to accurately modify the model\u2019s behavior for in-scope examples, which are either directly or indirectly related to unlearning targets, such as paraphrased text prompts to be degenerated (Zhong et al., 2023; Patil et al., 2024). At the same time, it must also ensure the generation consistency for out-of-scope examples, necessitating the preservation of model generation capabilities post-unlearning. Differentiating between in-scope and out-of-scope examples for unlearning is its own unsolved problem, as it requires determining when facts logically imply one another (Hase et al., 2023b; Cohen et al., 2023). \u2022 Unlearning efficiency & feasibility: Current research efforts focused on the development of rapid unlearning methods for pre-trained LLMs due to the significant re-training cost (Jang et al., 2022; Eldan & Russinovich, 2023; Yao et al., 2023). However, beyond computational efficiency, LLMs present additional efficiency challenges. These include the 4 \fRethinking Machine Unlearning for LLMs complexity and, at times, the infeasibility of pinpointing and attributing training data points designated for unlearning. Additionally, there is the challenge of executing unlearning in the context of black-box LLMs (Pawelczyk et al., 2023), where interactions with models are constrained to input-output queries. Distinguishing LLM unlearning from conventional MU. According to the above dimensions, LLM unlearning involves a broader range of targets, which are often context-dependent and less clearly defined. Moreover, the effectiveness of LLM unlearning is not limited to forgetting the influence of specific data points but also includes defining a broader unlearning scope for model capability removal. Furthermore, there is a critical need to devise more mechanistic methods that guarantee effective and robust unlearning, while also enhancing their practicality and feasibility. Mathematical modeling. Building upon the high-level formulation of LLM unlearning problems presented earlier, we next provide mathematical modeling details and discuss the associated design choices. To facilitate comprehension, we provide a commonly-used formulation of LLM unlearning problems below. While this may not be the sole or optimal approach to LLM unlearning, it incorporates several key elements that are essential to the problem setup. min \u03b8 E(x,yf)\u2208Df[\u2113(yf|x; \u03b8)] | {z } Forget +\u03bb E(x,y)\u2208Dr[\u2113(y|x; \u03b8)] | {z } Retain (1) where \u2113(y|x; \u03b8) denotes the prediction loss of using \u03b8 given the input x with respect to the response y, Df and Dr refer to \u2018forget\u2019 and \u2018retain\u2019 sets which will be explained later, yf denotes the desired model response post-unlearning, and \u03bb \u22650 is a regularization parameter to balance \u2018forget\u2019 and \u2018retain\u2019 (e.g., \u03bb = 0 if retain set is not given a priori). In the dataset setup of LLM unlearning, we typically assume access to a forget set (Df), the influence of which should be eliminated in LLM generation. For instance, Df might consist of a collection of harmful or toxic prompt-response pairs designated for degeneration (Yao et al., 2023). Moreover, if the original training set is available, then Df can be composed of a subset of training data points related to the unlearning target. Alternatively, it can be generated using synthesized data points based on a higher-level unlearned knowledge concept, or it can be derived from a set of extracted training data points reverse-engineered from the given LLM itself. In practice, the forget set Df is not required to belong precisely to the LLM\u2019s training corpus. Given the diversity and extensive size of the model\u2019s training data, the content we aim to unlearn is more likely to represent a general concept. Thus, LLM unlearning needs to not only unlearn specific training samples but also generalize to similar samples that share common characteristics. Besides the forget set Df, sometimes there is a need for a retain set (Dr), which contains samples that are not subject to unlearning and used to preserve the utility of the unlearned model. Through the lens of the unlearning scope we discussed earlier, the forget set (Df) provides in-scope examples earmarked for unlearning, while the retain set (Dr) involves examples out of the unlearning scope. Some recent studies have also attempted to develop LLM unlearning approaches that operate independently of access to forget and/or retain sets (Pawelczyk et al., 2023; Li et al., 2023b). We next introduce the model and optimization setups for LLM unlearning. Unlearning is performed at the post-model training phase. As shown in (1), a common unlearning objective is to efficiently update the original pre-trained model so that the updated model can unlearn on Df while retaining its generation capability on Dr. In addition, another design element is unlearning response (yf), referred to as the response of an unlearned model to in-scope examples. For example, in the stateful LLM unlearning method aimed at erasing information related to \u2018Who\u2019s Harry Potter?\u2019 (Eldan & Russinovich, 2023), the unlearning response is based on word replacements using generic translations, like substituting \u2018Quidditch\u2019 with \u2018Skyball\u2019, as part of the unlearning process. However, this type of approach may inadvertently blur the line between LLM hallucination and legitimate responses, highlighting the need for improvements in unlearning response design. Another choice is empty response (Wu et al., 2023b; Patil et al., 2024), given by the reject \u2018I don\u2019t know\u2019 (Patil et al., 2024) or the customized response by \u2018masking\u2019 the unlearning information using specialized tokens such as \u2018\u25a0\u2019 in (Wu et al., 2023b). However, we need to ensure that the empty response targets only examples within the unlearning scope. Otherwise, frequent rejections may occur, potentially diminishing the user experience with LLMs. 4 LLM Unlearning Methods Revisited: Prior Art and Overlooked Principles In this section, we begin with a review of existing MU (machine unlearning) methodologies adapted for LLM unlearning. Following that, we delve into the design factors that are often overlooked in the literature, including the intertwined data-model influence, the relationship with model editing, and adversarial training. Review of existing unlearning principles. Existing LLM unlearning methods can be broadly categorized into two groups: model-based and input-based. Model-based methods involve modifying the weights and/or architecture 5 \fRethinking Machine Unlearning for LLMs components of LLMs to achieve the unlearning objective (Jang et al., 2022; Lu et al., 2022; Yao et al., 2023; Yu et al., 2023; Chen & Yang, 2023; Zhang et al., 2023c; Hase et al., 2023a; Wu et al., 2023b; Rafailov et al., 2023), e.g., following the mathematical formulation in Sec. 3. Input-based methods design input instructions (Madaan et al., 2022; Zheng et al., 2023; Pawelczyk et al., 2023), such as in-context examples or prompts, to guide the original LLM (without parameter updating) towards the unlearning objective. In the literature, the predominant research emphasis lies on model-based methods as shown in Tab. 1. We review some of the most representative approaches below. \u2022 Gradient ascent and its variants (Jang et al., 2022; Yao et al., 2023; Yu et al., 2023): Gradient ascent (GA) stands as one of the most straightforward unlearning methods, updating the model parameters by maximizing the likelihood of mis-prediction for the samples within the forget set Df (Jang et al., 2022). However, it is worth noting that GA alone can be sensitive to the choice of hyperparameters during optimization, such as the number of ascent steps and the learning rate (Jia et al., 2023; Fan et al., 2024). This has given rise to improved variants of GA. For example, Yao et al. (2023) incorporate random labeling to augment the unlearning objective and ensure utility preservation on the retain set Dr. Another variant of GA is to transform it into a gradient descent approach, which aims to minimize the likelihood of predictions on relabeled forgetting data (Yu et al., 2023). This gradient descent-based fine-tuning, over relabeled or randomly labeled forgetting data, is also employed in (Eldan & Russinovich, 2023), where generic translations are used to replace the unlearned texts. GA and its variants often involve fine-tuning pre-trained LLMs for unlearning purposes. To enhance efficiency, parameter-efficient fine-tuning (PEFT) techniques could be employed. Chen & Yang (2023) fine-tune an adapter over the unlearning objective that acts as an unlearning layer within the LLM. Zhang et al. (2023c) use LoRA to create task vectors and accomplish unlearning by negating tasks under these task vectors. \u2022 Localization-informed unlearning (Meng et al., 2022; Yu et al., 2023; Wu et al., 2023b): The pursuit of parameter efficiency is also in line with the objective of identifying and localizing a subset of model units (e.g., layers, weights, or neurons) that are essential for the unlearning task. For example, the process of localization can be accomplished through representation denoising, also known as causal tracing, in (Meng et al., 2022), focusing on the unit of model layers. Previous work shows that it is important to delete information about unlearning targets wherever it is represented in models in order to protect against attacks (Patil et al., 2024). In addition, gradient-based saliency (Yu et al., 2023) is employed to identify the crucial weights that need to be fine-tuned to achieve the unlearning objective. In (Wu et al., 2023b), neurons that respond to unlearning targets are identified within the feed-forward network and subsequently selected for knowledge unlearning. In the context of vision models, unlearning can also benefit from localizing weights salient to unlearning, as demonstrated by Jia et al. (2023) and Fan et al. (2024). Furthermore, the concept of localization-informed unlearning resonates with the future modular machine learning solution development (Menik & Ramaswamy, 2023). This modularity allows the emerging foundation models to be partitioned into manageable subparts, facilitating easier maintenance and independent updates for each component. \u2022 Influence function-based methods: While the influence function (Koh & Liang, 2017; Bae et al., 2022) is a standard approach to assess the effect of data removal on model performance (Izzo et al., 2021; Warnecke et al., 2021), it is not commonly employed in the context of LLM unlearning for two main reasons: the computational complexity involved in inverting the Hessian matrix, and the reduced accuracy resulting from the use of approximations in influence function derivation (Jia et al., 2023). However, the potential of influence functions in LLM unlearning may be underestimated, given that scalability issue has been improved by Grosse et al. (2023), and approximation errors can be mitigated by focusing on localized weights that are salient to unlearning, as described in the previous category. \u2022 Other model-based methods: Other types of LLM unlearning methods fall outside the categories mentioned above. For example, Jang et al. (2022); Chen & Yang (2023) show that sequential unlearning outperforms batch unlearning However, as indicated by Gu et al. (2024), sequential editing of LLMs may compromise their general capabilities. Therefore, the study of sequential unlearning in LLMs continues to be an unresolved issue. \u2022 Input-based vs. model-based: Input-based strategies (Madaan et al., 2022; Zheng et al., 2023; Pawelczyk et al., 2023) have also shown promise in addressing the challenges posed by the restricted access to black-box LLMs and achieving parameter efficiency of LLM unlearning. Here the learnable parameters are given by input prompts rather than model weights/architecture components. However, we posit that input-based methods may not necessarily yield genuinely unlearned models, leading to weaker unlearning strategies compared to model-based methods because modifying the inputs of LLMs alone may not be sufficient to completely erase the influence of unlearning targets (Toyer et al., 2023). This assertion is also supported by the existence of hard or even adversarial in-scope examples associated with unlearning targets and the challenge to remove their influence in LLMs (Zhong et al., 2023; Patil et al., 2024). Exploring data-model interactions. A key objective of unlearning is to eliminate the influence of the forgotten data points/knowledge on the model\u2019s performance. However, this process is not studied in isolation: It is closely 6 \fRethinking Machine Unlearning for LLMs connected to exploring the influence of model weights or architecture components. Unlearning requires a sense of locality, which involves addressing the specific unlearning target and its associated unlearning scope. Consequently, exploring model influence helps identify the specific, localized areas of the model that are relevant to this locality. This is further reinforced by the surveyed weight localization techniques (Meng et al., 2022; Yu et al., 2023; Wu et al., 2023b; Patil et al., 2024). Thus, model influence and data influence are intertwined in LLM unlearning, and a comprehensive understanding of the former can streamline the process of handling data influence. Relationship with model editing. Model editing, closely related to LLM unlearning, focuses on the local alteration of pre-trained models\u2019 behavior to introduce new knowledge or rectify undesirable behaviors. First, the objective of editing could align with that of unlearning when editing is introduced to erase information. Second, like unlearning scope, editing scope (Mitchell et al., 2022; Hase et al., 2023b; Cohen et al., 2023) is crucial to ensure that editing is executed without compromising the generative capabilities of the model outside the defined scope. Third, both model editing and unlearning can be approached using the \u2018locate first, then edit/unlearn\u2019 principle. Localization in the context of model editing has also been applied to various elements, including neurons (Dai et al., 2021), network layers (Meng et al., 2022; Gupta et al., 2023), and feed-forward components of LLMs (Geva et al., 2020; Li et al., 2023c). Despite the aforementioned connections, there are clear distinctions between LLM unlearning and editing. First, the unlearning response is sometimes unknown compared to the editing response. The specificity of an incorrect or improper unlearning response could be seen as a form of LLM hallucination after unlearning. Second, although unlearning and model editing may share some common algorithmic foundations, the former does not create new answer mappings. Rather, its central aim is the comprehensive elimination of the influence attributed to a specific knowledge or concept within a pre-trained LLM. Third, we can differentiate model editing from unlearning from the perspective of \u2018working memory\u2019. It is known in (Li et al., 2022) that working memory in LLMs is maintained through neuron activations rather than weight-based long-term memory. Thus, existing memory-based model editing techniques (Li et al., 2022; Mitchell et al., 2022; Madaan et al., 2022; Zheng et al., 2023) focus on updating short-term working memory instead of altering the long-term memory encapsulated in the model\u2019s weights. Yet, we posit that unlearning requires more mechanistic approaches that facilitate \u2018deep\u2019 modifications to pre-trained LLMs. Adversarial training for robust unlearning. An increasing body of research highlights the weaknesses of existing unlearning methods (Shi et al., 2023; Maini et al., 2024), particularly in their vulnerability to test-time adversaries attempting to jailbreak unlearned models. This issue has been explored by Patil et al. (2024) for LLMs and Zhang et al. (2023d) for diffusion models, and inspires us to integrate adversarial training (Madry et al., 2017) into the unlearning process, resulting in what we term adversarial unlearning. However, this approach has received relatively little attention thus far. While adversarial unlearning increases training costs, it also presents new opportunities. For instance, localization-informed unlearning can significantly reduce the computation expenses associated with adversarial unlearning by focusing on a small portion of model units for updating. Furthermore, advanced adversarial training techniques such as fast adversarial training (Shafahi et al., 2019; Wong et al., 2020; Zhang et al., 2022) and generalized adversarial training (Zhu et al., 2019; Kumari et al., 2019; Zhang et al., 2022; Robey et al., 2023; Casper et al., 2024b) have the potential to enhance the scalability of adversarial unlearning while preserving its effectiveness. Reinforcement learning. The mainstream technique for aligning LLMs with human values is RLHF (reinforcement learning from human feedback) and its variants (Christiano et al., 2017; Ouyang et al., 2022; Bai et al., 2022; Yuan et al., 2023; Lee et al., 2023; Rafailov et al., 2023; Casper et al., 2023). However, RLHF is resource-intense: (1) it requires human inputs that are expensive to collect, and (2) it is computationally costly (i.e., the standard three-stage aligning procedure). LLM unlearning arises as an alternative aligning method, where collecting negative (i.e., low-quality and harmful) samples is much easier through user reporting or (internal) red teaming than positive (i.e., high-quality and helpful) samples which often require hiring humans. Furthermore, we can envision a reinforced unlearning paradigm with a properly defined reward function for the unlearned tasks. In a similar vein, Rafailov et al. (2023); Lu et al. (2022) utilize a reward model to assist in unlearning. 5 Assessment In this section, we discuss key considerations for designing an effective evaluation pipeline for LLM unlearning. Tasks and benchmarks. Datasets related to harmful content degeneration, personal identification information removal, and copyrighted information prevention serve as suitable benchmarks for evaluating the effectiveness of LLM unlearning. Some notable examples of these datasets include: The Enron dataset, which comprises employee emails publicly disclosed during Enron\u2019s legal investigation by the Federal Energy Regulatory Commission (Wu et al., 2023b), the Training Data Extraction Challenge dataset used in (Jang et al., 2022), the Harry Potter book series dataset (Eldan & 7 \fRethinking Machine Unlearning for LLMs Russinovich, 2023; Shi et al., 2023), the toxicity generation dataset (Lu et al., 2022; Gehman et al., 2020), and the TOFU dataset for fictitious unlearning (Maini et al., 2024). Evaluation of unlearning effectiveness. We examine efficacy from three perspectives: comparison with retraining (i.e., the gold standard of unlearning), \u2018hard\u2019 in-scope evaluation or robustness, and training data detection. \u2022 LLM unlearning vs. retraining. In classic unlearning paradigms (Golatkar et al., 2020; Thudi et al., 2022; Jia et al., 2023; Fan et al., 2024), retraining a model from scratch after removing the forgotten data from the original training set is regarded as exact unlearning. However, the scalability challenges of retraining LLMs make it difficult to establish a performance upper bound for evaluating LLM unlearning. A recent solution by Maini et al. (2024) is to incorporate fictitious data (synthetic author profiles) into the model training paradigm. Since the injected set never appeared in the original pretraining set, LLM fine-tuning can simulate the retraining process over the newly-introduced set. Despite the progress with regard to the specialized TOFU dataset (Maini et al., 2024), there is still a pressing need to advance current evaluation metrics and pipelines to accurately assess the gap between (approximate) LLM unlearning methods and exact unlearning. \u2022 \u2018Hard\u2019 in-scope evaluation or robustness. As demonstrated in Sec. 3, unlearning is generally context and taskdependent, contingent upon an unlearning scope. Another effectiveness metric of LLM unlearning is to ensure forgetting concerning in-scope unlearned examples, even for those \u2018hard\u2019 ones that fall within the unlearning scope but may not be directly associated with the unlearning targets. The assessment of \u2018hard\u2019 in-scope examples can be achieved by techniques such as paraphrasing what LLMs intend to forget or creating multi-hop questions (Zhong et al., 2023). Evaluating \u2018hard\u2019 in-scope examples aligns seamlessly with the underlying principles of \u2018worst-case\u2019 or \u2018adversarial\u2019 evaluation methods for unlearning, resembling LLM jailbreaking attacks (Zhang et al., 2023d; Yong et al., 2023; Patil et al., 2024). For instance, it is shown by Yong et al. (2023) that unlearning a scope using an English-only example would not guarantee a similar unlearned outcome when translated into other languages. It is also crucial to evaluate the robustness of unlearned LLMs after fine-tuning. Recent studies have revealed that fine-tuning LLMs can sometimes lead to the re-emergence of behaviors that were not anticipated (Yang et al., 2023; Qi et al., 2023; Lermen et al., 2023; Yong et al., 2023). \u2022 Training data detection or membership inference. Membership inference attacks (MIA) (Shokri et al., 2017), designed to detect if a data point is part of a victim model\u2019s training set, serve as a crucial data privacy-unveiled metric for evaluating machine unlearning methods (Thudi et al., 2022; Jia et al., 2023). This metric gains even more significance in the context of LLM unlearning, particularly when retraining is not an option. This concept is also connected to training data memorization (Carlini et al., 2022), as well as training data extraction attacks (Nasr et al., 2023) in LLMs. In the realm of LLM unlearning, privacy-related evaluation metrics have been explored and considered in various studies (Shi et al., 2023; Wu et al., 2023b; Jang et al., 2022; Pawelczyk et al., 2023; Maini et al., 2024). Utility preservation. Another crucial metric is to ensure the retained generation capabilities of unlearned LLMs on standard language modeling tasks that fall outside the unlearning scope. For example, evaluation on natural language understanding tasks (Jang et al., 2022; Eldan & Russinovich, 2023; Yao et al., 2023) and perplexity (Ilharco et al., 2022; Zhang et al., 2023c) has been considered in the literature. In line with evaluating the effectiveness of LLM unlearning on \u2018hard\u2019 in-scope examples, it is equally crucial to assess utility preservation using \u2018hard\u2019 out-of-scope examples, achieved by e.g., using data transformations. Lastly, we note that it can be difficult to determine the exact scope for some unlearning target (Hase et al., 2023b; Cohen et al., 2023), so part of the challenge here is deciding which generation capabilities should be retained in the first place. Efficiency. Computation cost has been a predominant efficiency metric when evaluating LLM unlearning methods, as demonstrated in Table 1. In addition to that, efforts have been made to extend LLM unlearning to black-box models, without access to model parameters, as demonstrated by Pawelczyk et al. (2023). Furthermore, memory efficiency could also serve as a crucial efficiency metric. The distinction from parameter efficiency is that current parameter-efficient fine-tuning methods still impose substantial memory costs for storing LLMs and for executing back-propagation (Malladi et al., 2023). Thus, a future research direction is to explore memory-efficient fine-tuning methods for LLM unlearning. 6 Applications of LLM Unlearning In this section, we discuss two application categories facilitated by LLM unlearning: the first focused on data influence and the second on model capabilities. 8 \fRethinking Machine Unlearning for LLMs Copyright and privacy protection. One application of unlearning involves legal and ethical considerations around the fair use of training data. Algorithmic disgorgement is the term applied in law and policy for the requirement put on a company by a regulator, such as the Federal Trade Commission (FTC) in the United States, to completely destroy a model that was trained on data without legal consent (Li, 2022; Goland, 2023; Belkadi & Jasserand, 2023; Achille et al., 2023). The most famous case to-date is the FTC calling for the destruction of a weight loss application by WW International, whose underlying model contained illegal health information from children. Unlearning presents a viable alternative to complete disgorgement by removing the effect of the illegal data. Also, the tension between data owners (e.g., authors) and LLM service providers is escalating, leading to legislation such as legal disputes involving OpenAI, Meta, and New York Times (Small, 2023; Grynbaum & Mac, 2023). This trend is likely to persist due to increasing societal concerns about AI data usage. The need for copyright-protected content removal aligns with the capabilities of LLM unlearning. However, it is often challenging to pinpoint the exact sources of training data that need to be deleted, giving rise to the issue of data attribution (Li et al., 2023a). For example, the leakage related to the \u2018Harry Potter\u2019 series (Eldan & Russinovich, 2023) can have multiple possible causes, e.g., the books were used in the LLM\u2019s training data, the training data containing online discussions related to the series, or the LLM using retrieval-augmented generation (Gao et al., 2023) which might lead to leakage from the search results. Similar to deleting copyrighted information from the training data, another scenario is preventing LLMs from leaking user privacy, especially personal identification information (PII). This concern is closely related to LLM memorization and training data extraction (Carlini et al., 2019, 2021, 2022; Jang et al., 2022; Nasr et al., 2023). Sociotechnical harm reduction. Another application of LLM unlearning is alignment (Ouyang et al., 2022), aimed at aligning LLMs with human instructions and making sure generated text conforms to human values. Unlearning can be used to forget harmful behaviors such as the production of toxic, discriminatory, illegal, or morally undesirable outputs (Shevlane et al., 2023), e.g., instructions to build CBRN (chemical, biological, radiological, and nuclear) weapons. Unlearning, as a safety alignment tool, can happen at the different stages of LLM development, e.g., before, during, or after alignment. Current research has focused on the \u2018pre-alignment\u2019 stage (Yao et al., 2023), there may be untapped opportunities in the others. Hallucinations, which involve the generation of false or inaccurate content that may appear plausible, are a significant challenge in LLMs. Previous research has demonstrated that unlearning can reduce LLM hallucinations by targeting and unlearning factually incorrect responses given specific questions (Yao et al., 2023). Since hallucination is likely to be caused by multiple sources, the possible usage is to unlearn factually incorrect data that serve as the source of commonly shared hallucinations or misconceptions. LLMs are known to generate biased decisions and outputs (Perez et al., 2022; Tamkin et al., 2023; Cui et al., 2023). In the vision domain, unlearning has proven to be an effective tool for reducing discrimination to enable fair decisionmaking (Sattigeri et al., 2022; Chen et al., 2023b). In the language domain, unlearning has been applied to mitigate gender-profession bias (Yu et al., 2023) and many other fairness issues (Sattigeri et al., 2022; Oesterling et al., 2023; Kadhe et al., 2023). However, more opportunities exist, such as unlearning stereotypes in training data. LLMs are also known to be vulnerable to jailbreaking attacks (Wei et al., 2023; Qi et al., 2023; Huang et al., 2023) (i.e., adversarially crafted prompts that lead the LLM to generate undesired outputs) as well as poisoning/backdoor attacks (Rando & Tram` er, 2023; Carlini et al., 2023). Unlearning can be a natural solution for both types of attacks given the existing success of unlearning as a defense against adversarial attacks in other domains (Wang et al., 2019; Li et al., 2021; Liu et al., 2022; Jia et al., 2023). 7 Discussion and"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.06878v1",
+                    "title": "AutoSeqRec: Autoencoder for Efficient Sequential Recommendation",
+                    "abstract": "Sequential recommendation demonstrates the capability to recommend items by\nmodeling the sequential behavior of users. Traditional methods typically treat\nusers as sequences of items, overlooking the collaborative relationships among\nthem. Graph-based methods incorporate collaborative information by utilizing\nthe user-item interaction graph. However, these methods sometimes face\nchallenges in terms of time complexity and computational efficiency. To address\nthese limitations, this paper presents AutoSeqRec, an incremental\nrecommendation model specifically designed for sequential recommendation tasks.\nAutoSeqRec is based on autoencoders and consists of an encoder and three\ndecoders within the autoencoder architecture. These components consider both\nthe user-item interaction matrix and the rows and columns of the item\ntransition matrix. The reconstruction of the user-item interaction matrix\ncaptures user long-term preferences through collaborative filtering. In\naddition, the rows and columns of the item transition matrix represent the item\nout-degree and in-degree hopping behavior, which allows for modeling the user's\nshort-term interests. When making incremental recommendations, only the input\nmatrices need to be updated, without the need to update parameters, which makes\nAutoSeqRec very efficient. Comprehensive evaluations demonstrate that\nAutoSeqRec outperforms existing methods in terms of accuracy, while showcasing\nits robustness and efficiency.",
+                    "authors": "Sijia Liu, Jiahao Liu, Hansu Gu, Dongsheng Li, Tun Lu, Peng Zhang, Ning Gu",
+                    "published": "2023-08-14",
+                    "updated": "2023-08-14",
+                    "primary_cat": "cs.IR",
+                    "cats": [
+                        "cs.IR",
+                        "cs.LG"
+                    ],
+                    "main_content": "INTRODUCTION Recommender systems have the ability to model user preferences and help users discover content of their interest from a vast amount of information [8, 13, 21, 23, 24, 27, 33]. Meanwhile, as user interests evolve over time, recommender systems are required to dynamically update and promptly adapt their recommendation strategies to keep up with the changes, capturing shifts in user interests in real time [11, 14, 20, 22, 36, 39]. However, training a recommendation model from scratch incurs significant costs, making it impractical to retrain the model every time there are interaction updates. Therefore, the sequential recommendation model faces a critical challenge of efficiently capturing evolving user interests. To address this challenge, two key problems need to be solved: 1) accurately capturing the evolution of user interests, and 2) swiftly adapting to changes through incremental updates. Existing sequential recommendation models suffer from limitations in both aspects. On the one hand, traditional sequential methods [1, 14, 34] typically treat users as a sequences of items they have interacted with, overlooking the collaborative relationships among them. Consequently, these methods fail to effectively leverage collaborative information, which is crucial for improving accuracy and personalization in recommendations. On the other hand, while some more recent graph-based approaches [3, 39] can incorporate collaborative information by leveraging user-item interaction graphs, these methods often face challenges in terms of time complexity and computational efficiency. To address these limitations, we propose a novel recommendation model called AutoSeqRec (Autoencoder for Efficient Sequential Recommendation), which consists of the following two components. arXiv:2308.06878v1  [cs.IR]  14 Aug 2023 \fCIKM \u201923, October 21\u201325, 2023, Birmingham, United Kingdom Sijia Liu et al. To accurately capture the evolution of user interests, we propose a multi-encoder and decoder architecture. The first step involves constructing a user-item interaction matrix and an item transition matrix. These two matrices are then fed into a multi-information encoder, while three single-information decoders are employed for reconstruction purposes. The user-item interaction matrix contains the historical interactions between users and items. By reconstructing this matrix, AutoSeqRec can effectively model collaborative information, providing a representation of users\u2019 long-term preferences. On the other hand, the item transition matrix captures the typical previous/next item that users interact with before/after the current item. Through reconstructing this matrix, AutoSeqRec can capture users\u2019 real-time interactions, enabling a representation of their short-term interests. By combining collaborative information and item transition information, AutoSeqRec is able to accurately capture the evolution of user interests. To promptly adapt to changes through incremental updates, we introduce an efficient incremental inference process that incorporates new user-item interactions. This process begins by representing the single-hop item transition probability based on the reconstructed item transition matrix. It then efficiently approximates the multi-hop transition information by utilizing the compressed output of the multi-information encoder. When there are new updates on interactions, AutoSeqRec only modifies the user-item interaction matrix and item transition matrix incrementally, while keeping the model parameters unchanged for inference. This approach enables AutoSeqRec to efficiently adapt to new changes at the incremental level of a single interaction, facilitating real-time recommendation updates with minimal computational overhead. The update process of AutoSeqRec shares some insights with a contemporaneous work, IMCrrorect [19]. However, unlike IMCrrorect, which focuses on recommendation unlearning tasks, AutoSeqRec focuses on incremental recommendation tasks. Our main contributions are as follows: \u2022 We design a novel autoencoder architecture called AutoSeqRec. By combining the user\u2019s long-term preference derived from collaborative information with the user\u2019s short-term interest modeled from item transition, our approach allows for a more accurate capture of user interest evolution and changes in behavior. \u2022 We develop an efficient incremental inference process that incorporates new user-item interactions without changing model parameters. It facilitates real-time recommendation updates at the incremental level of a single interaction. \u2022 Comprehensive evaluation on four datasets shows that AutoSeqRec can significantly outperform existing sequential recommendation methods in accuracy and efficiency. Moreover, AutoSeqRec can be integrated with other sequential recommendation methods to improve their performance. 2 PRELIMINARIES In this section, we introduce AutoRec [25], a representative autoencoder framework for collaborative filtering. Then, we describe the problem and discuss the limitations of the previous autoencoderbased method in solving this problem. 2.1 Notations In this paper, we denote user set as U = {\ud835\udc621,\ud835\udc622...,\ud835\udc62\ud835\udc5a} and item set as I = {\ud835\udc561,\ud835\udc562, ...,\ud835\udc56\ud835\udc5b}, where \ud835\udc5ais the number of users and \ud835\udc5b is the number of items. Without loss of generality, we use \ud835\udc62to represent a user and \ud835\udc56to represent an item when their indices are not concerned. The user-item rating matrix \ud835\udc4c\u2208R\ud835\udc5a\u00d7\ud835\udc5bis partially observed. Each user is represented by a partially observed vector \ud835\udc66(\ud835\udc62) = (\ud835\udc4c\ud835\udc62,1, ...,\ud835\udc4c\ud835\udc62,\ud835\udc5b) \u2208R\ud835\udc5b, and each item is represented by a partially observed vector \ud835\udc66(\ud835\udc56) = (\ud835\udc4c1,\ud835\udc56, ...,\ud835\udc4c\ud835\udc5a,\ud835\udc56) \u2208R\ud835\udc5a. 2.2 AutoRec AutoRec [25] designs an item-based (or user-based) autoencoder which first takes as input each partially observed \ud835\udc66(\ud835\udc56) (or \ud835\udc66(\ud835\udc62)), then projects it into a low-dimensional latent (hidden) space, and finally reconstruct \ud835\udc66(\ud835\udc56) (\ud835\udc66(\ud835\udc62)) in the output space to predict missing ratings for static recommendation. Taking item-based AutoRec as an example, which tries to solve the following reconstruction problem: min \ud835\udf03 \u2211\ufe01 y(\ud835\udc56) \u2225y(\ud835\udc56) \u2212\u210e(y(i);\ud835\udf03)\u22252, (1) where \u210e(\u00b7) = \ud835\udc53(\ud835\udc54(\u00b7)) includes two processes: encoding and decoding. The encoder \ud835\udc54(\u00b7) and decoder \ud835\udc53(\u00b7) are both composed of a single nonlinear layer. Then, the parameters \ud835\udf03are learned using backpropagation. Given learned parameters \u02c6 \ud835\udf03, the predicted rating of item-based AutoRec for user \ud835\udc62and item \ud835\udc56is \u02c6 \ud835\udc4c\ud835\udc62,\ud835\udc56= \u0010 \u210e \u0010 y(\ud835\udc56); \u02c6 \ud835\udf03 \u0011\u0011 \ud835\udc62. (2) 2.3 Problem description Each user-item interaction can be represented by a tuple (\ud835\udc62(\ud835\udf0f),\ud835\udc56(\ud835\udf0f)), where \ud835\udf0f= 1, 2, ..., indicates the sequential order in which the interaction occurred. Suppose there are \ud835\udc41interactions in total, denoted as S(\ud835\udc41) =< (\ud835\udc62(1),\ud835\udc56(1)), (\ud835\udc62(2),\ud835\udc56(2)), ..., (\ud835\udc62(\ud835\udc41),\ud835\udc56(\ud835\udc41)) >. Now given a user\ud835\udc62, we need to predict which item that user\ud835\udc62will interact with in the \ud835\udc41+ 1-th interaction. The output should be a \ud835\udc5b-dimensional vector, and each dimension represents the possibility score of interaction between user \ud835\udc62and the corresponding item. AutoRec is a static recommendation method that cannot effectively utilize sequential information. Although a lot of extensions have been proposed [16, 18, 27, 31, 35, 37], there is no prior work to extend AutoRec to incremental recommendation scenarios to the best of our knowledge. 3 THE PROPOSED METHOD In this section, we first introduce the overall architecture of AutoSeqRec, and then present the implementation of incremental inference of AutoSeqRec. 3.1 Architecture In this subsection, we detail the AutoSeqRec, including its implementation of the learning process and inference process. The overall architecture of AutoSeqRec is shown in Figure 1. \fAutoSeqRec: Autoencoder for Efficient Sequential Recommendation CIKM \u201923, October 21\u201325, 2023, Birmingham, United Kingdom Figure 1: The inference process of AutoSeqRec. Assuming that user 2 last interacted with item 4, this figure shows how AutoSeqRec obtains the prediction score of user 2 interacting each item in the \ud835\udc41+ 1-th interaction. 3.1.1 Learning Process. We first construct the user-item interaction matrix \ud835\udc45(\ud835\udc41) \u2208R\ud835\udc5a\u00d7\ud835\udc5band item transition matrix\ud835\udc47(\ud835\udc41) \u2208R\ud835\udc5b\u00d7\ud835\udc5b, where the superscript indicates that there are a total of \ud835\udc41interactions. The user-item interaction matrix \ud835\udc45(\ud835\udc41) stores the interaction records between users and items in S(\ud835\udc41), its entry at \ud835\udc62-th row and \ud835\udc56-th column is defined as: \ud835\udc45(\ud835\udc41) \ud835\udc62,\ud835\udc56 = ( 1, user \ud835\udc62has interacted with item \ud835\udc56in S(\ud835\udc41), 0, otherwise. (3) We define the interaction sequence of user \ud835\udc62in S(\ud835\udc41) as S(\ud835\udc41) \ud835\udc62 =< \ud835\udc56(1) \ud835\udc62,\ud835\udc56(2) \ud835\udc62, ...,\ud835\udc56(\ud835\udc5b(\ud835\udc41) \ud835\udc62 ) \ud835\udc62 >, where \ud835\udc56(\ud835\udf0f) \ud835\udc62 is the item id of the \ud835\udf0f-th interaction of user \ud835\udc62, \ud835\udc5b(\ud835\udc41) \ud835\udc62 is the number of interactions of user \ud835\udc62in S(\ud835\udc41). The item transition matrix \ud835\udc47(\ud835\udc41) stores the transition information between item pairs, its entry at \ud835\udc56\u2032-th row and \ud835\udc56\u2032\u2032-th column is defined as: \ud835\udc47(\ud835\udc41) \ud835\udc56\u2032,\ud835\udc56\u2032\u2032 = \u2211\ufe01 \ud835\udc62=1,...,\ud835\udc5a \u2211\ufe01 \ud835\udc56(\ud835\udf0f) \ud835\udc62=\ud835\udc56\u2032,\ud835\udc56(\ud835\udf0f+1) \ud835\udc62 =\ud835\udc56\u2032\u2032 \ud835\udf0f=1,...,\ud835\udc5b(\ud835\udc41) \ud835\udc62 \u22121 1. (4) Discussion about the input matrices. Each row of the user-item interaction matrix \ud835\udc45(\ud835\udc41) indicates which items the corresponding user has interacted with, which actually describes a user\u2019s personalized preference. We call the elements related to \ud835\udc45(\ud835\udc41) collaborative information. Each row of the item transition matrix \ud835\udc47(\ud835\udc41) records the data of which item a user will interact with next time statistically, after the user interacts with the corresponding source item. Since it is statistical data from all users, this information is irrelevant to a specific user. We call the elements related to \ud835\udc47(\ud835\udc41) source information. Correspondingly, each row of the transposition of the item transition matrix, (\ud835\udc47(\ud835\udc41))\u22a4, records the data of which item the user interacted with last time if a user interacts with the corresponding target item. This is also statistical data from all users and does not contain personalized information. We call the elements related to (\ud835\udc47(\ud835\udc41))\u22a4target information. In this paper, we implement the multi-information encoder (denoted as \ud835\udc53(\u00b7)) with a single nonlinear layer, more complex forms of the multi-information encoder are left for future work: \ud835\udc53(\ud835\udc4b) = \ud835\udf0e(\ud835\udc4b\ud835\udc4a\ud835\udc52+ b\ud835\udc52), (5) where \ud835\udc4b\u2208R\u2217\u00d7\ud835\udc5bis an input matrix with an unlimited number of rows, \ud835\udc4a\ud835\udc52\u2208R\ud835\udc5b\u00d7\ud835\udc58is the weight matrix, b\ud835\udc52\u2208R\ud835\udc58is the bias term, and \ud835\udf0e(\u00b7) is the activation function. The user-item interaction matrix \ud835\udc45(\ud835\udc41), item transition matrix \ud835\udc47(\ud835\udc41) and its transpose (\ud835\udc47(\ud835\udc41))\u22a4are input into the multi-information encoder to obtain collaborative information embedding \ud835\udc38(\ud835\udc41) \ud835\udc50 , source information embedding \ud835\udc38(\ud835\udc41) \ud835\udc60 , and target information embedding \ud835\udc38(\ud835\udc41) \ud835\udc61 , respectively: \ud835\udc38(\ud835\udc41) \ud835\udc50 = \ud835\udc53(\ud835\udc45(\ud835\udc41)) \u2208R\ud835\udc5a\u00d7\ud835\udc58, (6) \ud835\udc38(\ud835\udc41) \ud835\udc60 = \ud835\udc53(\ud835\udc47(\ud835\udc41)) \u2208R\ud835\udc5b\u00d7\ud835\udc58, (7) \ud835\udc38(\ud835\udc41) \ud835\udc61 = \ud835\udc53((\ud835\udc47(\ud835\udc41))\u22a4) \u2208R\ud835\udc5b\u00d7\ud835\udc58. (8) Discussion about multi-information encoder. Firstly, the user-item interaction matrix, item transition matrix, and its transpose share consistent semantics. Each dimension represents a corresponding item, allowing them to be simultaneously compressed by an encoder. Secondly, the multi-information encoder converts the input matrices into embedding representations with semantic meaning. This implies that if two users have similar interaction histories, i.e., their respective rows in the user-item interaction matrix are similar, their outputs after passing through the encoder will also exhibit similarity. Likewise, if the transition patterns of two items (whether as source or target items) are comparable, indicating similar rows \fCIKM \u201923, October 21\u201325, 2023, Birmingham, United Kingdom Sijia Liu et al. or columns in the item transition matrix, their outputs after going through the multi-information encoder will reflect this similarity. Lastly, since the multi-information encoder incorporates collaborative information, source information, and target information simultaneously, the compressed representation generated by the multi-information encoder can effectively capture and consider all three types of information concurrently. Next, we input collaborative information embedding \ud835\udc38(\ud835\udc41) \ud835\udc50 , source information embedding \ud835\udc38(\ud835\udc41) \ud835\udc60 , target information embedding \ud835\udc38(\ud835\udc41) \ud835\udc61 into collaborative information decoder (denoted as \ud835\udc54\ud835\udc50(\u00b7)), source information decoder (denoted as \ud835\udc54\ud835\udc60(\u00b7)) and target information decoder (denoted as \ud835\udc54\ud835\udc61(\u00b7)) respectively, to reconstruct the user-item interaction matrix \ud835\udc45(\ud835\udc41), the item transition matrix \ud835\udc47(\ud835\udc41) and its transpose (\ud835\udc47(\ud835\udc41))\u22a4: \u02dc \ud835\udc45(\ud835\udc41) = \ud835\udc54\ud835\udc50(\ud835\udc38(\ud835\udc41) \ud835\udc50 ) \u2208R\ud835\udc5a\u00d7\ud835\udc5b, (9) \u02dc \ud835\udc47(\ud835\udc41) = \ud835\udc54\ud835\udc60(\ud835\udc38(\ud835\udc41) \ud835\udc60 ) \u2208R\ud835\udc5b\u00d7\ud835\udc5b, (10) ( \u02dc \ud835\udc47(\ud835\udc41))\u22a4= \ud835\udc54\ud835\udc61(\ud835\udc38(\ud835\udc41) \ud835\udc61 ) \u2208R\ud835\udc5b\u00d7\ud835\udc5b. (11) Again, we only adopt the simplest setting, where each of the three decoders consists of a single nonlinear layer: \ud835\udc54\ud835\udc50(\ud835\udc4b) = \ud835\udf0e(\ud835\udc4b\ud835\udc4a\ud835\udc50+ b\ud835\udc50), (12) \ud835\udc54\ud835\udc60(\ud835\udc4b) = \ud835\udf0e(\ud835\udc4b\ud835\udc4a\ud835\udc60+ b\ud835\udc60), (13) \ud835\udc54\ud835\udc61(\ud835\udc4b) = \ud835\udf0e(\ud835\udc4b\ud835\udc4a\ud835\udc61+ b\ud835\udc61), (14) where \ud835\udc4a\ud835\udc50,\ud835\udc4a\ud835\udc60,\ud835\udc4a\ud835\udc61\u2208R\ud835\udc58\u00d7\ud835\udc5bare the weight matrices, b\ud835\udc50, b\ud835\udc60, b\ud835\udc61\u2208R\ud835\udc5b are the bias terms. Finally, we train AutoSeqRec by reducing the reconstruction loss, using the Frobenius norm as the measure of the differences between the input matrices and the reconstructed matrices. Specifically, we denote the reconstruction loss of the collaborative information decoder as L\ud835\udc50, denote the reconstruction loss of the source information decoder as L\ud835\udc60, and denote the reconstruction loss of the target information decoder as L\ud835\udc61, which are defined as follows, respectively: L\ud835\udc50= ||\ud835\udc45(\ud835\udc41) \u2212\u02dc \ud835\udc45(\ud835\udc41) ||\ud835\udc39, (15) L\ud835\udc60= ||\ud835\udc47(\ud835\udc41) \u2212\u02dc \ud835\udc47(\ud835\udc41) ||\ud835\udc39, (16) L\ud835\udc61= ||(\ud835\udc47(\ud835\udc41))\u22a4\u2212( \u02dc \ud835\udc47(\ud835\udc41))\u22a4||\ud835\udc39. (17) By minimizing the reconstruction loss, we aim to enable the decoders to recover the original data as accurately as possible. The final loss function consists of three kinds of losses: L = L\ud835\udc50+ L\ud835\udc61+ L\ud835\udc60. (18) Through backpropagation during training, we iteratively update model parameters to minimize reconstruction loss. Since the reconstructions of three types of information are equally important, we do not need to set any hyperparameters to balance among them. 3.1.2 Inference Process. As aforementioned, the user-item interaction matrix reflects user personalized information. In static scenarios, we can directly use the reconstructed user-item interaction matrix \u02dc \ud835\udc45(\ud835\udc41) as the prediction result. The collaborative informationbased probability score of user \ud835\udc62interacting with all items in the Figure 2: Explanation of two-hops transition. A row of the item transition matrix records the data of which item a user will interact with next time, after the user interacts with the source item. A column of the item transition matrix records the data of which item the user interacted with last time if a user interacts with the target item. The inner product of the two vectors represents the probability score of the source item reaching the target item after two hops. \ud835\udc41+ 1-th interaction can be expressed as: p(\ud835\udc41+1) \ud835\udc50 = \u02dc \ud835\udc45(\ud835\udc41) \ud835\udc62,\u00b7 \u2208R\ud835\udc5b. (19) Each dimension of p\ud835\udc50corresponds to an item. However, in dynamic scenarios, the reconstructed interaction matrix \u02dc \ud835\udc45(\ud835\udc41) can only reflect users\u2019 long-term preferences and cannot accurately reflect users\u2019 temporal interests. To capture users\u2019 short-term interests, we further utilize item transition matrix. For user \ud835\udc62, whose last interacted item is \ud835\udc56(\ud835\udc5b(\ud835\udc41) \ud835\udc62 ) \ud835\udc62 (for the convenience of narration, we use \ud835\udc56to refer to \ud835\udc56(\ud835\udc5b(\ud835\udc41) \ud835\udc62 ) \ud835\udc62 ), we take the \ud835\udc56-th row of the reconstructed item transition matrix as the short-term interest of user \ud835\udc62. Then the single-hop transition probability score of user \ud835\udc62is: p(\ud835\udc41+1) \ud835\udc611 = \u02dc \ud835\udc47(\ud835\udc41) \ud835\udc56,\u00b7 \u2208R\ud835\udc5b. (20) Each dimension of p\ud835\udc611 corresponds to an item. p(\ud835\udc41+1) \ud835\udc611 only considers the information of single-hop transition. In order to more comprehensively consider the information of multi-hop transitions, a simple solution is to multiply p(\ud835\udc41+1) \ud835\udc611 and the item transition matrix \u02dc \ud835\udc47(\ud835\udc41) continuously. As shown in the Figure 2, p(\ud835\udc41+1) \ud835\udc611 reflects the probability score of the items that will be reached from a source item, and each column of the item transition matrix reflects the probability score of the items that a target item may come from. So one multiplication can represent the probability score of reaching the target item from the source item after two hops. However, the time complexity of this approach can be very high when the number of items is large. We took a more efficient approach. As mentioned before, similar inputs will have similar outputs after passing through the multiinformation encoder, which in turn will have similar decoder outputs. So instead of operating on the reconstruction matrix output by the decoder, we can operate on the embedded representation output by the multi-information encoder. Specifically, we perform a Hadamard product operation on the collaborative information embedding of user \ud835\udc62((\ud835\udc38(\ud835\udc41) \ud835\udc50 )\ud835\udc62,\u00b7) and (\ud835\udc38(\ud835\udc41) \ud835\udc60 )\ud835\udc56,\u00b7 to obtain a personalized transition embedding: \u02dc p(\ud835\udc41+1) \ud835\udc612 = (\ud835\udc38(\ud835\udc41) \ud835\udc60 )\ud835\udc56,\u00b7 \u2299(\ud835\udc38(\ud835\udc41) \ud835\udc50 )\ud835\udc62,\u00b7 (21) \fAutoSeqRec: Autoencoder for Efficient Sequential Recommendation CIKM \u201923, October 21\u201325, 2023, Birmingham, United Kingdom Furthermore, we multiply the \u02dc p(\ud835\udc41+1) \ud835\udc612 by the target information embedding \ud835\udc38(\ud835\udc41) \ud835\udc61 to obtain the two-hop transition information: p(\ud835\udc41+1) \ud835\udc612 = \u02dc p(\ud835\udc41+1) \ud835\udc612 (\ud835\udc38(\ud835\udc41) \ud835\udc61 )\u22a4\u2208R\ud835\udc5b. (22) The multi-hop transition information can be obtained by multiplying the transpose of the target information embedding, (\ud835\udc38(\ud835\udc41) \ud835\udc61 )\u22a4, multiple times. Finally, we take the weighted sum of three types of probability scores (collaborative information probability score p(\ud835\udc41+1) \ud835\udc50 , onehop transition probability score p(\ud835\udc41+1) \ud835\udc611 , and two-hops transition probability score p(\ud835\udc41+1) \ud835\udc612 ) as the final prediction result, as follows: p(\ud835\udc41+1) = \ud835\udf061p(\ud835\udc41+1) \ud835\udc50 + \ud835\udf062p(\ud835\udc41+1) \ud835\udc611 + (1 \u2212\ud835\udf061 \u2212\ud835\udf062)p(\ud835\udc41+1) \ud835\udc612 . (23) The coefficients \ud835\udf061 and \ud835\udf062 control the importance of different sources of information in the final prediction results. 3.2 Incremental Inference In this subsection, we introduce how AutoSeqRec effectively handles dynamic data and updates recommendation results in real time. Specifically, we first analyze why the autoencoder-based CF model can naturally achieve incremental recommendation, and then we give the concrete implementation of incremental recommendation of AutoSeqRec. 3.2.1 Analysis. Representation learning-based recommendation methods, e.g., matrix factorization [13], first obtain the embeddings of users and items, and then perform personalized ranking of items by calculating the similarity (e.g., inner product) between users and items. Different from these representation learning-based methods, the autoencoder-based collaborative filtering method is a process of obtaining predicted interactions directly from observed interactions. For example, AutoRec directly inputs the user-item interaction matrix into the autoencoder, and takes the reconstructed user-item interaction matrix as the prediction result. Autoencoderbased methods are often used to solve the problem of static recommendation. Actually, once an autoencoder is trained, it is equipped to handle various user interests. Now, let us focus on the dynamic recommendation process for two users \ud835\udc4eand \ud835\udc4f. Assume that user \ud835\udc4elikes to watch comedy movies in the early days, and user\ud835\udc4flikes to watch science fiction movies during the same period. Because the historical interaction records of the two users are different (\ud835\udc4einteracts with more comedy movies, \ud835\udc4finteracts with more science fiction movies), when a recommendation model based on autoencoder is built, \ud835\udc4e will be recommended comedy movies, \ud835\udc4fwill be recommended science fiction movies. Both user \ud835\udc4eand user \ud835\udc4fcan obtain the desired personalized recommendation in that period, and the root cause of their different recommendation results is the difference in historical interaction (the difference between the corresponding rows of the interaction matrix). If in the later stage, \ud835\udc4e\u2019s interest has changed to science fiction movies. At this time, the interaction record of \ud835\udc4e(interacting more science fiction movies instead of comedy movies) is input into the recommendation model, and \ud835\udc4ewill be recommended science fiction movies. The autoencoder-based model has this ability because it has learned how to recommend science fiction movies Table 1: The statistics of the datasets used in the experiments. # Users # Items # Interactions # Unique Times ML-100K 943 1,349 99,287 49,119 ML-1M 6,040 3,416 999,661 458,254 Garden 1,686 962 13,272 1,888 Video 5,130 1,685 37,126 1,946 to users (e.g., user \ud835\udc4f) from the training data. To sum up, the collaborative filtering method based on autoencoder actually supports incremental recommendation naturally. We only need to change the input matrix, and the weight of autoencoder does not need to be updated during the incremental inference process. 3.2.2 Implementation. We believe that AutoSeqRec has the ability to handle various user interests and encode the historical interactions of users at different moments. Therefore, we only need to change the input user-item interaction matrix \ud835\udc45and item transition matrix \ud835\udc47, and then follow the same inference process to make predictions. Specifically, assuming user \ud835\udc62interacts with item \ud835\udc56at \ud835\udc41+ 1-th interaction, we can use the following steps to predict the next interaction: 1) update the user-item interaction matrix \ud835\udc45(\ud835\udc41+1) and item transition matrix \ud835\udc47(\ud835\udc41+1) to contain the latest interaction state between user \ud835\udc62and item \ud835\udc56. 2) input the updated user-item interaction matrix and item transition matrix into the AutoSeqRec to obtain the recommendation result. After the model training is completed, we can make accurate next-step interaction predictions based on the existing interaction history. In addition, we can execute the above prediction process as many times as needed to achieve continuous prediction of the user\u2019s next interaction item. This flexibility and efficiency enable AutoSeqRec to quickly respond to user needs and provide personalized and high-quality recommendation services in practical applications. 4 EXPERIMENTS In this section, we conduct comprehensive evaluation to answer the following research questions (RQs): \u2022 RQ1 Does AutoSeqRec outperform state-of-the-art methods in future interaction prediction task on dynamic graphs? \u2022 RQ2 How do different components of AutoSeqRec affect model performance? \u2022 RQ3 Does the desgin of AutoSeqRec help to reduce running time? \u2022 RQ4 Does the hidden embedding size and coefficients \ud835\udf061, \ud835\udf062 affect predictions? \u2022 RQ5 Can AutoSeqRec be seamlessly integrated with other sequential recommendation models to improve their performance? 4.1 Experimental Setup 4.1.1 Dataset. We use four publicly available datasets to evaluate the performance of AutoSeqRec. MovieLens [6] is a movie rating dataset that is widely used in dynamic graph tasks. We use two versions of MovieLens: ML-100K and ML-1M. ML-100K contains 100K interactions collected during a seven-month period from September 19th, 1997 to April 22nd, 1998. ML-1M contains 1M interactions \fCIKM \u201923, October 21\u201325, 2023, Birmingham, United Kingdom Sijia Liu et al. from users who joined MovieLens in 2000. Amazon [7] contains ratings on the e-commerce platform Amazon.com, and we adopt two sub-datasets, Garden and Video. We follow the same preprocessing pipeline in previous works [10, 28] that users and items with fewer than 5 interactions are discarded. The statistical information of datasets is shown in Table 1. 4.1.2 Setting. We run all the experiments on a server equipped with one NVIDIA TESLA T4 GPU and Intel(R) Xeon(R) Gold 5218R CPU. All the code of this work is implemented with Python 3.7.4 1. Three input matrices including user-item interaction matrix, item transition matrix and the transposition of item transition matrix are first divided into mini-batches separately, and then we union these mini-batches and randomly shuffle them. In the training process, we use the corresponding loss function according to which matrix is used in the mini-batches union. The hyper-parameters are tuned via grid search. More specifically, we search autoencoder\u2019s hidden dimension size in [32, 64, 128, 256] and coefficients \ud835\udf061, \ud835\udf062 from values in [0.0, 0.1, 0.2, \u00b7 \u00b7 \u00b7 , 1.0]. 4.1.3 Baseline Models. We compare AutoSeqRec with the following state-of-the-art interaction graph modelling methods: \u2022 LightGCN [8] is a representative collaborative filtering algorithm based on GNNs, which ignores time information. \u2022 RRN [34] models user and item interaction sequences with separate RNNs for sequential recommendation. \u2022 JODIE [14] also generates node embeddings using RNNs, and can also estimate user embeddings trajectories. \u2022 CoPE [39] uses an ordinary differential equation-based GNN to model the evolution of network. \u2022 FreeGEM [20] is a state-of-the-art parameter-free method for dynamic graph learning tasks. 4.2 Future interaction prediction (RQ1) The future prediction prediction task is defined as follows: given all interactions till time \ud835\udc61, we predict which item will user \ud835\udc62interact with at time \ud835\udc61? We use the same data splitting as JODIE\u2019s [14] and CoPE\u2019s [39] to separate the dataset by time. Specifically, we use the earliest 80% interactions for training, the following 10% data for validation, and the latest 10% data for test. We adopt mean reciprocal rank (MRR) and Recall@10 to evaluate the performance of future interaction prediction, in which MRR is the average of the reciprocal rank and Recall@10 is the fraction of interactions in which the ground truth items are within the top 10 recommendations, following JODIE, CoPE and FreeGEM. Table 2 shows the results of AutoSeqRec compared with the above five state-of-the-art methods. We observe that AutoSeqRec significantly outperforms all compared methods in all datasets across both metrics Recall@10 and MRR on the four datasets. We calculate the relative improvement (Rel. Imp.) of AutoSeqRec over the baseline as (performance of AutoSeqRec \u2212performance of baseline)/(performance of baseline). Across all datasets, the minimum improvement of AutoSeqRec is 2.00% 50.00% in terms of MRR and 10.62% 52.94% in terms of Recall@10. We notice that LightGCN performs the worst on most datasets. This is largely due to the fact that we split the dataset by time according to this experiment 1The codes are available at https://github.com/sliu675/AutoSeqRec setting. Since AutoSeqRec makes full use of user-item collaborative information and one-hop or even two-hops of item transition information, AutoSeqRec outperforms the other methods. 4.3 Ablation studies (RQ2) In this section, we first conduct ablation study on the operations applied to the three types of information from the encoder. Then, we test whether the decoder outputs of user-item interaction matrix and item transition matrix are helpful or not. 4.3.1 Encoder hidden outputs operations. This experiment test the effectiveness of 1) the Hadamard product operation on the collaborative information embedding and source information embedding and 2) the inner product operation between the personalized transition embedding and the target information embedding. We use a, b, c to represent the collaborative information embedding of user \ud835\udc62((\ud835\udc38\ud835\udc50)\ud835\udc62,\u00b7), \ud835\udc62\u2019s last interacted item \ud835\udc56\u2019s source information embedding ((\ud835\udc38\ud835\udc60)\ud835\udc56,\u00b7) and the target information embedding \ud835\udc38\ud835\udc61. We observe that combining a, b and c obtain the best performance among all the combinations as shown in Table 3. Moreover, operation \ud835\udc4e\u00b7 \ud835\udc50 performs better than \ud835\udc4f\u00b7 \ud835\udc50on Amazon datasets, while the opposite is true on MovieLens dataset. One possible reason is that the two Amazon datasets used in our experiments are more sparse than the MovieLens datasets, which may lead to less robust statistics in the item transition matrix. Therefore, the collaborative information from the user-item interaction matrix becomes more important. 4.3.2 Decoder outputs of interaction matrix and transition matrix. We test if adding the decoder outputs of user-item interaction matrix and item transition matrix can improve the performance or not, compared with only using encoder hidden outputs. We add the reconstructed interaction matrix and transition matrix separately, and then add these two reconstructed matrices together. Specifically, we use a weighted sum to combine all the information (hidden outputs, reconstructed interaction matrix and transition matrix). As shown in Table 4, we observe that using all information is better than any other combination, confirming the usefulness of collaborative information and multi-hops transition information. 4.4 Efficiency Comparison (RQ3) We use the future interaction prediction task to study the efficiency of JODIE, CoPE, FreeGEM and AutoSeqRec. Other baselines like RRN show comparable efficiency with JODIE [14] and thus are omitted. FreeGEM requires multiple runs of offline/online SVD, JODIE and CoPE both run 50 epochs and choose the best performing model on the validation set as the optimal model. As shown in Table 5, AutoSeqRec is at least 16.3x faster than JODIE, at least 64.7x faster than CoPE, and at least 1.1x faster than the parameters-free method FreeGEM. This experiment confirms that AutoSeqRec has higher computational efficiency than the compared methods, maily due to 1) the interaction matrix and transition matrix updates are incremental and 2) the encoder-decoder architecture in AutoSeqRec is much more light-weight compared with RNNs (in JODIE) and GNNs (in CoPE). \fAutoSeqRec: Autoencoder for Efficient Sequential Recommendation CIKM \u201923, October 21\u201325, 2023, Birmingham, United Kingdom Table 2: Accuracy comparison on the future interaction prediction task. Method ML-100K ML-1M Video Garden MRR Recall@10 MRR Recall@10 MRR Recall@10 MRR Recall@10 LightGCN 0.012 0.025 0.014 0.031 0.019 0.036 0.025 0.087 RRN 0.042 0.084 0.038 0.071 0.039 0.074 0.079 0.169 JODIE 0.046 0.095 0.037 0.070 0.044 0.081 0.059 0.142 CoPE 0.049 0.111 0.040 0.085 0.036 0.078 0.082 0.176 FreeGEM 0.040 0.079 0.040 0.077 0.058 0.086 0.100 0.179 AutoSeqRec 0.068 0.150 0.060 0.130 0.064 0.107 0.102 0.227 Rel. Imp. 38.78% 35.14% 50.00% 52.94% 10.34% 10.62% 2.00% 26.82% Table 3: Ablation study on the encoder hidden outputs operations. Setting ML-100K ML-1M Video Garden MRR Recall@10 MRR Recall@10 MRR Recall@10 MRR Recall@10 \ud835\udc4f\u00b7 \ud835\udc50 0.050 0.113 0.031 0.064 0.016 0.026 0.058 0.123 \ud835\udc4e\u00b7 \ud835\udc50 0.038 0.079 0.020 0.037 0.030 0.060 0.086 0.203 (\ud835\udc4e\u2299\ud835\udc4f) \u00b7 \ud835\udc50 0.058 0.130 0.037 0.076 0.040 0.086 0.097 0.206 Table 4: Ablation study on the decoder outputs of user-item interaction matrix and item transition matrix. Setting ML-100K ML-1M Video Garden MRR Recall@10 MRR Recall@10 MRR Recall@10 MRR Recall@10 only hidden 0.058 0.130 0.037 0.076 0.040 0.086 0.097 0.206 hidden + interaction matrix 0.062 0.138 0.040 0.081 0.038 0.078 0.091 0.215 hidden + transition matrix 0.060 0.132 0.058 0.126 0.062 0.106 0.101 0.226 all 0.068 0.150 0.060 0.130 0.064 0.107 0.102 0.227 Table 5: Running time comparison on the four datasets. ML-100K ML-1M Video Garden JODIE 8h47min56s (298.8x) 374h7min53s (1989.5x) 46min18s (38.6x) 16min51s (16.3x) CoPE 16h51min5s (572.3x) 1025h24min6s (5452.7x) 1h38min33s (82.1x) 1h6min52s (64.7x) FreeGEM 3min12s (1.8x) 5h54min18s (31.4x) 2min46s (2.3x) 1min7s (1.1x) AutoSeqRec 1min46s 11min17s 1min12s 1min2s 4.5 Sensitivity Analysis (RQ4) Here, we analyze how the key hyper-parameters affect the performance of AutoSeqRec. We analyze the following hyper-parameters: the embedding size of the auto-encoder and the coefficients \ud835\udf061 and \ud835\udf062 for balancing the importance of different prediction scores in Equation 23. 4.5.1 Embedding Size. We analyze the impact of the embedding size in the autoencoder on the performance. To do this, we vary the embedding size of the autoencoder from 32 to 256 and calculate the MRR and Recall@10 scores for future interaction prediction on the Amazon Garden dataset and MovieLens ML-100K dataset. The trends on other datasets are similar. The results are shown in Figure 3 (a). As the embedding dimension increases from 32 to 128, we find that the embedding size has a relatively positive effect on the performance of AutoSeqRec, i.e., the performance of AutoSeqRec consistently improve with increasing dimension size. This experiment demonstrates that AutoSeqRec does not suffer from the common overfitting issue as other deep models. 4.5.2 The Coefficients \ud835\udf061 and \ud835\udf062. \ud835\udf061 controls the importance of collaborative information and \ud835\udf062 controls the importance of one-hop transition. The coefficients of collaborative information, one-hop transition and two-hops transition add up to one. Thus, once these two coefficients \ud835\udf061 and \ud835\udf062 are determined, the third one is immediately determined as well. We observe that using a single type of information is not optimal. To make sure each source of information is used, we vary the coefficients from 0.1 to 0.8 and plot the heatmaps of Recall@10 on the ML-100K and Video datasets. As shown in Figure 3 (b) and Figure 3 (c), the performance of AutoSeqRec is not very sensitive to \ud835\udf061 and \ud835\udf062, i.e., we can observe near optimal performance with several different configurations. 4.6 Integrating AutoSeqRec with Other Sequential Recommendation Methods (RQ5) The major innovations in AutoSeqRec, e.g., the embedding learned from multi-information encoder, the score functions (Equation 21 and 22), and the combination of collaborative and transition information, are orthogonal to many existing sequential recommendation methods. Therefore, we believe AutoSeqRec can be integrated \fCIKM \u201923, October 21\u201325, 2023, Birmingham, United Kingdom Sijia Liu et al. 32 64 128 256 Size of embedding dimension 0.125 0.150 0.175 0.200 0.225 Recall ML-100K Recall Garden Recall ML-100K MRR Garden MRR 0.06 0.07 0.08 0.09 0.10 MRR (a) 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 lambda 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 lambda 2 0.130 0.135 0.140 0.145 0.150 (b) 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 lambda 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 lambda 2 0.075 0.080 0.085 0.090 0.095 0.100 0.105 0.110 (c) Figure 3: Sensitivity analysis of AutoSeqRec on three hyperparameters: (a) MRR and Recall@10 with different embedding size on ML-100K and Garden, (b) the heat-map of Recall@10 with different coefficients \ud835\udf061 and \ud835\udf062 on ML-100K, (c) the heat-map of Recall@10 with different coefficients \ud835\udf061 and \ud835\udf062 on Video. Self Attention + Point-Wise Feed Network \ud835\udc561 \ud835\udc62 \u2299 \u00d7 \u2299 \u00d7 \u2299 \u00d7 Expected \ud835\udc562 Collaborative Embedding The input of item \ud835\udc56is the item \ud835\udc56\u2019s source embedding  Same as SASRec \ud835\udc562 \ud835\udc563 \ud835\udc564 Target Embeddings Expected \ud835\udc563 Expected \ud835\udc564 Figure 4: The process of integrating AutoSeqRec with SASRec. Collaborative information embedding, source information embedding and target information embedding are pre-trained by AutoSeqRec. ML-100K ML-1M Datasets 0.56 0.58 0.60 0.62 0.64 0.66 NDCG@10 SASRec Variant 1 Variant 2 Figure 5: The recommendation accuracy of integrating AutoSeqRec with SASRec on the two MovieLens datasets. with them to further improve their performance. Here, we take SASRec [10], a well-known sequential recommendation method, as an example to illustrate how to integrate AutoSeqRec with other sequential recommendation methods as shown in Figure 4. 4.6.1 Setting. We use the sequential recommendation task defined in SASRec [10]. All the data pre-processing and partitioning protocols are the same as SASRec: 1) we treat the presence of a review or rating as implicit feedback, 2) we use timestamps to determine the sequential order of actions, and 3) we split the historical sequence of each user into three parts: the most recent one interaction for test, the penultimate interaction for validation, and all remaining interactions for training. 4.6.2 The Process of Integrating AutoSeqRec with SASRec. SASRec [10] is a sequential recommendation method based on the Transformer architecture [32]. We first briefly introduce its architecture. Firstly, SASRec reads in the embedding sequence of items that a user has interacted with. Then, after several Attention blocks (including self-attention layer and point-wise feed forward), we get the item embedding sequence after feature transformation. Finally, the item embedding sequence after feature transformation will be input into the prediction layer (e.g., multi-layer perceptron), which will treat each item as a class, and train the model by optimizing the classification task. While maintaining the main architecture (the Attention blocks) of SASRec, we use the collaborative information embedding \ud835\udc38\ud835\udc50, source information embedding \ud835\udc38\ud835\udc60and target information embedding \ud835\udc38\ud835\udc61, which are outputs from multi-information encoder, to improve the input and prediction layers of SASRec. Firstly, we will use the rows corresponding to source information embedding \ud835\udc38\ud835\udc60 as input to SASRec, instead of the randomly initialized embedding. Then, after passing through several Attention blocks, we will obtain the source information embedding after feature transformation. Finally, in the prediction phase, we adopt a consistent approach with AutoSeqRec. Specifically, as Equation (21), we first perform a Hadamard product between \ud835\udc38\ud835\udc50and the source information embedding after feature transformation to obtain the personalized transition embedding, and then multiply personalized transition embedding with the target information embedding \ud835\udc38\ud835\udc61as Equation (22) to further consider multi-hops transition information. Similar to the ablation studies, we design two ablative variants for this study to further understand the effectiveness of our design. \fAutoSeqRec: Autoencoder for Efficient Sequential Recommendation CIKM \u201923, October 21\u201325, 2023, Birmingham, United Kingdom Specifically, we call that the model use collaborative information embedding as Variant 2, and that without collaborative information embedding as Variant 1. 4.6.3 Analysis. We compare Variant 1 and Variant 2 with SASRec on two MovieLens datasets (ML-1M and ML-100k). We follow the evaluation strategy in prior works [9, 12] to avoid heavy computation on all user-item pairs. For each user \ud835\udc62, we randomly sample 50 negative items, and rank these items with the ground-truth item. The hidden dimension we use in our variants is 64. We use NDCG@10, which can simultaneously evaluate the ranking and recall of a prediction. For more clear comparison, we plot the performance of SASRec and the two variants on the bar chart as shown in Figure 5. We can observe that both variants are better than SASRec. The improvement of Variant 1 is 0.54%-2.30% in terms of NDCG@10 and the improvement of Variant 2 is 2.35%-2.60% in terms of NDCG@10. The experimental results indicate that multiplying the \ud835\udc56-th row of the source information embedding by the target information embedding in Equation 22 is helpful. Furthermore, compared with Variant 1, Variant 2 achieves relatively higher accuracy, i.e., higher NDCG@10 on both datasets. In other words, adding the operation, a Hadamard product on the collaborative information embedding of user \ud835\udc62, is not only beneficial to the future interaction prediction task, but also beneficial to the sequential recommendation task defined in SASRec. Thus, we further validate the rationality of multiplying the \ud835\udc56-th row of the source information embedding by the target information embedding, and operating a Hadamard product on the collaborative information embedding of user \ud835\udc62. From this case study, we can further conclude that 1) the embedding learned by AutoSeqRec are effective for sequential recommendation, 2) our score functions are helpful for improving accuracy, and 3) it is beneficial to combine collaborative information with transition information for sequential recommendation tasks. 5 RELATED WORK In this section, we first introduce the sequential recommendation methods, including recurrent neural network based methods, temporal point process based methods and graph based methods. Then, we introduce the autoencoder-based recommendation methods. 5.1 Sequential Recommendation Many sequential recommendation methods are based on recurrent neural networks. RRN [34] models user and item interaction sequences with separate RNNs. Time-LSTM [41] adopts time gates to represent the time intervals. LatentCross [1] incorporates contextual data into embeddings. DeepCoevolve [4] and JODIE [14] generate node embeddings using two intertwined RNNs, and JODIE can also estimate user embeddings trajectories. Another line of sequential recommendation work is based on temporal point processes. Know-Evolve [30] and HTNE [42] model interactions between users and items as multivariate point processes and Hawkes processes, respectively. Shchur et al. [26] directly model the conditional distribution of inter-event times, and therefore more accurately control the changes in user interests. DSPP [2] incorporates topology and long-term dependencies into the intensity function. All these methods take time information into account by intensity functions. Recently, more advancing graph based sequential recommenders are proposed. TDIG-MPNN [3] captures both global and local information on the graph. SDGNN [29] uses the state changes of neighbor nodes to learn the node embedding. CoPE [39] uses an ordinary differential equation-based GNN to model the evolution of network. 5.2 Autoencoder-based Collaborative Filtering Sedhain et al. [25] proposed AutoRec, the first autoencoder framework for collaborative filtering. Many researchers extend autoencoders to collaborative filtering for implicit feedback. Liang et al. [18] extended variational autoencoders to collaborative filtering for implicit feedback. Based on the Autorec and Transformer architectures, Dang et al. [5] proposed an attentional autoencoder for implicit recommendation. Zhu et al. [40] proposed a mutuallyregularized dual collaborative variational auto-encoder (MD-CVAE) for recommendation. Moreover, some works incorporate content information to enhance the performance of autoencoder. Li et al. [17] proposed a model that learns a shared feature space from heterogeneous data, leveraging heterogeneous auxiliary information to address the data sparsity problem of recommender systems. To solve the cold start problem, Lee et al. [15] developed a hybrid CF technique incorporating CF with content information. For implicit trust relationships, Zeng et al. [38] proposed an autoencoder model based on implicit trust relationship between users, combining the intrinsic relationship of both the implicit trust information and user-item interaction behavior for collaborative recommendation. 6"
+                }
+            ],
+            "Yihua Zhang": [
+                {
+                    "url": "http://arxiv.org/abs/2402.11846v2",
+                    "title": "UnlearnCanvas: A Stylized Image Dataset to Benchmark Machine Unlearning for Diffusion Models",
+                    "abstract": "The rapid advancement of diffusion models (DMs) has not only transformed\nvarious real-world industries but has also introduced negative societal\nconcerns, including the generation of harmful content, copyright disputes, and\nthe rise of stereotypes and biases. To mitigate these issues, machine\nunlearning (MU) has emerged as a potential solution, demonstrating its ability\nto remove undesired generative capabilities of DMs in various applications.\nHowever, by examining existing MU evaluation methods, we uncover several key\nchallenges that can result in incomplete, inaccurate, or biased evaluations for\nMU in DMs. To address them, we enhance the evaluation metrics for MU, including\nthe introduction of an often-overlooked retainability measurement for DMs\npost-unlearning. Additionally, we introduce UnlearnCanvas, a comprehensive\nhigh-resolution stylized image dataset that facilitates us to evaluate the\nunlearning of artistic painting styles in conjunction with associated image\nobjects. We show that this dataset plays a pivotal role in establishing a\nstandardized and automated evaluation framework for MU techniques on DMs,\nfeaturing 7 quantitative metrics to address various aspects of unlearning\neffectiveness. Through extensive experiments, we benchmark 5 state-of-the-art\nMU methods, revealing novel insights into their pros and cons, and the\nunderlying unlearning mechanisms. Furthermore, we demonstrate the potential of\nUnlearnCanvas to benchmark other generative modeling tasks, such as style\ntransfer. The UnlearnCanvas dataset, benchmark, and the codes to reproduce all\nthe results in this work can be found at\nhttps://github.com/OPTML-Group/UnlearnCanvas.",
+                    "authors": "Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, Sijia Liu",
+                    "published": "2024-02-19",
+                    "updated": "2024-02-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "main_content": "Introduction In recent years, the field of text-to-image technology, epitomized by Diffusion Models (DMs), has seen remarkable progress, demonstrating an unprecedented ability to generate high-resolution, high-quality images across a wide array of subjects (Rombach et al., 2022; Midjourney et al., 2023). These models, widely accessible to public, find applications in diverse sectors, including advertising, the creative arts (Fotor et al., 2022), and forensic sketching (LabLab AI et al., 2023). Central to the training of DMs is their reliance on the data source from the Internet (Schuhmann et al., 2022) to cover a wide range of content. However, this inclusiveness comes with challenges, including potential concerns about harmful generation (Schramowski et al., 2023), copyright issues (Krietzberg, 2023), and biases/stereotypes (Vinker et al., 2023; Turk, 2023). To alleviate the negative social impacts associated with DMs, machine unlearning (MU) techniques are catching increasing attention and have been studied in the field of vision generative models (Schramowski et al., 2023; Gandikota et al., 2023a;b; Kumari et al., 2023; Fan et al., 2023; Heng & Soh, 2023). MU for DMs (see Fig. 1 for an illustration) aims to erase the model\u2019s capacity to generate images related to a specific concept, essentially erasing that concept from the 1 arXiv:2402.11846v2  [cs.CV]  26 Feb 2024 \fUNLEARNCANVAS: A Versatile Stylized Image Dataset to Benchmark Machine Unlearning and Beyond pre-trained model. This concept can be provided in the form of text prompts (i.e., text conditions of DMs), covering a wide range from concrete object entities to abstract notions, such as artistic styles. We refer readers to Sec. 2 for more related work on MU for DMs. Unlike MU studies on image classification (Izzo et al., 2021; Golatkar et al., 2020; Becker & Liebig, 2022; Thudi et al., 2022; Warnecke et al., 2021; Jia et al., 2023), we lack evaluation benchmarks for MU in DMs. The current primary choice for assessing the performance of unlearned DMs (i.e., DMs post unlearning) is the Inappropriate Image Prompts (I2P) dataset (Schramowski et al., 2023). This dataset consists of real-world image-to-text prompts designed for erasing concepts like \u2018nudity\u2019 and \u2018violence\u2019. However, I2P mainly contains isolated unlearning targets and focuses on safety issues related to image generation. To enhance the assessment of MU in DMs and establish a standardized evaluation framework, we propose to develop a new benchmark dataset, referred to as UNLEARNCANVAS, which is designed to evaluate the unlearning of artistic painting styles along with associated image objects. Unlike I2P, UNLEARNCANVAS can be used to assess the removal of copyrighted information, such as artistic style, and it allows for increased diversity by including various object classes and styles. We summarize our contributions below. \u2022 We conduct a systematic review of existing MU methods for DMs and identified three unresolved challenges in their evaluations, such as limited quantitative assessments and a lack of retainability evaluation of unlearned DMs. \u2022 We propose UNLEARNCANVAS, a large-scale highresolution stylized image dataset with 24,000 labeled images, 60+ styles, and 20+ object topics. Its style-object dual supervision and high stylistic consistency help us to improve the precision of MU evaluation for DMs. \u2022 We develop a standardized evaluation pipeline for MU in DMs under UNLEARNCANVAS, which measures unlearning effectiveness, generation retainability, generation quality, and unlearning efficiency. \u2022 We benchmark state-of-the-art MU methods for DMs using UNLEARNCANVAS, validating the significance of our proposed evaluation metrics and providing insights into the strengths, weaknesses, and unlearning mechanisms. 2. Related Work and Preliminaries Machine unlearning (MU) for generative models in vision. MU was originally developed to mitigate the (potentially detrimental) influence of specific data points in a pretrained ML model, without necessitating a complete retraining of the model after removing these unlearning data (Cao & Yang, 2015; Bourtoule et al., 2021; Xu et al., 2023). The significance of MU has emerged with the purpose of data privacy protection, in response to regulations such as \u2018the right to be forgotten\u2019 (Hoofnagle et al., 2019). However, it has rapidly gained recognition for its role in promoting security, safety, and trustworthiness of ML models. Examples include defense against backdoor attacks (Liu et al., 2022; Jia et al., 2023), fairness enhancement (Chen et al., 2023; Oesterling et al., 2023), and controllable federated learning (Halimi et al., 2022; Jin et al., 2023). More recently, unlearning has proven crucial in text-toimage generation via DMs to prevent the production of harmful, private, or illegal content (Gandikota et al., 2023a; Zhang et al., 2023a; Gandikota et al., 2023b; Fan et al., 2023; Kumari et al., 2023; Heng & Soh, 2023). For example, DMs have demonstrated a susceptibility to generating NSFW (Not Safe For Work) images when conditioned on inappropriate text prompts (e.g., \u2018nudity\u2019 and \u2018violence\u2019) (Schramowski et al., 2023; Zhang et al., 2023b; Chin et al., 2023). Similarly, MU has also found applications in preventing the generation of copyrighted information, such as the misuse of artistic painting styles (Gandikota et al., 2023a; Zhang et al., 2023a). In summary, MU can be regarded as a (post-training) model edit operation, aimed at preventing DMs from producing undesired or inappropriate images. Our setup and objective. Let us denote a pretrained DM by \u03b8o, capable of synthesizing high-quality images conditioned on a provided text prompt c (e.g., \u2018A painting of a cat in Van Gogh style\u2019). In this work, we focus on the task of generating images in various painting styles. When there is a request to adapt the model \u03b8o to prevent the generation of images, e.g., in the \u2018Van Gogh\u2019 style, the MU problem arises, with the \u2018Van Gogh\u2019 style being the specified unlearning target, also known as the erasing concept (Gandikota et al., 2023a). See Fig. 1 for a visual representation of the concept. In formulating MU methods, the traditional setup presupposes access to the training dataset D, where Df \u2282D represents the forget set that revolves around the unlearning target. A commonly used gold standard of MU is to retrain the model from scratch over the retain set Dr = D \\ Df. However, commencing the training process from scratch becomes practically unfeasible because of (1) the substantial computational expense it entails, (2) the difficulty in accurately identifying Df, and (3) the absence of access to the original training data source. To overcome the aforementioned difficulties (1)-(3), the current research mainstream (Gandikota et al., 2023a; Kumari et al., 2023; Zhang et al., 2023a; Gandikota et al., 2023b; Heng & Soh, 2023; Fan et al., 2023) has focused on developing approximate yet efficient unlearning methods as practical alternatives to full retraining. The approximation introduced by current unlearning methods raises the question that we aim to address in this work: 2 \fUNLEARNCANVAS: A Versatile Stylized Image Dataset to Benchmark Machine Unlearning and Beyond Table 1. An overview of the scope and evaluation methods of the existing MU methods for DMs. The symbol \u2717denotes neither objects nor styles are covered. Related work Unlearning targets Quantitative evaluation Unlearning effectiveness Retainability ESD (Gandikota et al., 2023a) 5 artistic styles 11 objects objects only objects only CA (Kumari et al., 2023) 4 artistic styles 4 objects objects only \u2717 FMN (Zhang et al., 2023a) 1 artistic style 7 objects objects only \u2717 UCE (Gandikota et al., 2023b) 7 artistic settings1 11 objects objects only objects only SA (Heng & Soh, 2023) 1 artistic style 31 objects objects only \u2717 SalUn (Fan et al., 2023) No artistic styles 11 objects objects only objects only How can we accurately assess unlearning performance and establish a benchmark for existing unlearning methods? 3. Unresolved Challenges in MU Evaluation This work will develop a new benchmark dataset, referred to as UNLEARNCANVAS (detailed in Sec. 4), focusing on the task of unlearning artistic painting styles in conjunction with associated image objects. With the use of UNLEARNCANVAS, a text prompt for DMs can be constructed as \u2018A painting of a {object} in {style}\u2019. And the unlearning target can be explicitly designated as either \u2018style\u2019 or \u2018object\u2019. To be specific, UNLEARNCANVAS is targeted for the open challenges (C1-C3) when building a rigorous quantitative evaluation framework for MU in DMs. These challenges are consolidated through a careful examination of existing studies for object and/or style unlearning in Tab. 1. (C1) The absence of a diverse yet cohesive unlearning target repository. As shown in Tab. 1, the assessment of MU for DMs, in terms of both effectiveness (concept erasure performance) and retainability (preserved generation quality under non-forgotten concepts), is typically conducted using manually selected targets, often chosen from a limited pool of unlearning targets. A critical aspect missing from MU evaluation is the recognition that unlearning targets related to styles and objects should not be isolated within the dataset used for MU assessment. Instead, they should be considered in a manner that exposes their potential interconnected influences on other targets. (C2) The lack of precision in evaluation. Artistic styles can be challenging to precisely define and distinguish, which makes quantifying their impact on unlearning effectiveness and retainability difficult. To illustrate this challenge, we conducted a preliminary investigation in Tab. 2 using a ViT1UCE adopted a multi-concept unlearning setting, where different numbers of styles are unlearned at the same time and total up to 1000 styles are considered. Table 2. Classification accuracy of a style classifier (ViT-L/16) finetuned on WIKIART (Phillips & Mackintosh, 2011). The classifier is tested on the test set and the image set generated by SD v1.5 and v2.0 with the prompt \u2018An painting in {artist} style\u2019, where the {artist} is traversed over the WIKIART labels (10 images/prompt). WIKIART Train Set WIKIART Test Set Images by SD v1.5 Images by SD v2.0 Accuracy 99.8% 85.4% 56.7% 41.2% based classifier trained on the WIKIART dataset (Phillips & Mackintosh, 2011). This dataset categorizes paintings into 139 distinct styles, each annotated with the creator\u2019s name. Despite training, the ViT-large model achieved only an 84.32% accuracy on the test dataset. This result underscores the inherent complexity of classifying artistic styles, causing inaccuracies in quantitative MU evaluations. The classifier\u2019s suboptimal accuracy on DMs generated images further increases the prediction error (around 50% accuracy). This complicates precise quantitative evaluations. (C3) The lack of a systematic study on \u2018retainability\u2019 of DMs post unlearning. As indicated by the \u2018retainability\u2019 column in Tab. 1, assessing the capacity of unlearned DMs to maintain image generation under innocent concepts is notably lacking. However, retainability is a critical indicator of potential \u2018side effects\u2019 of an unlearning method. In conjunction with challenge (C1), the inclusion of diverse unlearning targets, encompassing both styles and objects, allows for the evaluation of retainability from various perspectives. For instance, when unlearning an artistic style, it is crucial to consider the retainability of image generation abilities for other styles as well as for objects. The challenges (C1)-(C3) outlined above have driven us to develop UNLEARNCANVAS as a solution for the systematic and comprehensive evaluation of MU in DMs. We will show in Sec. 4 that UNLEARNCANVAS includes a diverse range of artistic styles and objects. Within this dataset, we can explore the interrelationships between paired image styles and image objects to address (C1). Each style category in UNLEARNCANVAS maintains high intra-style consistency, facilitating accurate style and object classification and enabling post-unlearning quantitative evaluation for (C2). To address (C3), we can evaluate two types of retainability, providing a precise and comprehensive assessment of the potential side effects induced by MU methods. 4. Our Proposal: UNLEARNCANVAS Dataset Construction of UNLEARNCANVAS. As motivated in Sec. 2, UNLEARNCANVAS is created for ease of MU evaluation in DMs. Its construction involves two main steps: seed image collection and subsequent image stylization; see Fig. 2 for a schematic overview. 3 \fUNLEARNCANVAS: A Versatile Stylized Image Dataset to Benchmark Machine Unlearning and Beyond Step 2: Stylization Ukiyoe Step 1: Seed Image Collection Human... Dogs ... Architecture ... 20 Object Classes Seed Images  Stylized Images (60 styles) High Stylistic Consistency Sketch Crayon Fauvism Van Gogh ... ... ... Figure 2. An illustration of the key steps when curating UNLEARNCANVAS and its key features. Seed image collection. UNLEARNCANVAS includes images in 60 unique artistic styles across 20 distinct objects. The style categories are derived from a set of high-resolution photos, termed seed images. These seed images consist of original real-world photos from Pexels (Pexels et al., 2023). In total, there are 400 seed images across 20 object classes. Image stylization. After collecting the seed images, the next step in dataset development is the stylization process. During this phase, each seed image is transformed into 60 predetermined artistic styles. The goal is to ensure that the resulting images maintain high stylistic consistency while preserving the content details present in the seed images. This stylization process is carried out using the services provided by (Fotor et al., 2022). Once all the seed images have been stylized in all 60 styles, the dataset is constructed with super-high-resolution images and organized in a hierarchical structure that balances both style and object categories. Advantages of UNLEARNCANVAS. Next, we will highlight the advantages (A1-A3) of UNLEARNCANVAS to address the MU evaluation challenges (C1-C3) discussed in Sec. 3. (A1) Style-object dual supervision enables a rich unlearning target bank for comprehensive evaluation. UNLEARNCANVAS offers a diverse range of unlearning targets by annotating each image with both style and object labels. This is in response to (C1). We have categorized the unlearning targets into two main domains: a style domain comprising 60 artistic styles, and an object domain including 20 object classes. This setup allows us to investigate the impact of style unlearning on the fidelity of generated objects in images and vice versa. Besides its role in MU evaluation, UNLEARNCANVAS could also serve as an annotated testbed for assessing style transfer accuracy and visual in-context learning performance; see details in Appx. C. (A2) High stylistic consistency ensures precise style definitions and enables accurate quantitative evaluations. In UNLEARNCANVAS, images labeled with the same style exhibit high intra-style consistency and distinct inter-style differences, as exemplified in Fig. 2. This facilitates the distinction between different styles, effectively addressing challenge (C2). In a similar study to that presented in Tab. 2, we observe that the styles within UNLEARNCANVAS can be readily classified using ViT-Large (Dosovitskiy et al., 2020), achieving an accuracy of 98% for images generated by SD finetuned on UNLEARNCANVAS. We refer readers to Tab. A2 for detailed results. Before Unlearning After Unlearning Tes t Prompt: A painting of a dog in Van Gogh Style . Tes t Prompt: A painting of a dog in Crayon Style . Tes t Prompt: A painting of Hors es . Unlearning Effectiveness Evaluation In-D om ain Retainability Evaluation Cross-D om ain Retainability Evaluation Unlearning Target: Van Gogh Style Figure 3. Evaluation metric illustration, with unlearning the \u2018Van Gogh style\u2019 as a case study. \u2713 and \u2717denote desired and undesired results post unlearning. (A3) Enabling in-depth retainability analyses for MU evaluation. Using UNLEARNCANVAS, we can assess DMs\u2019 retainability postunlearning from two perspectives: in-domain retainability, which evaluates the model\u2019s performance on test-time generations within the same domain as the unlearning target, and cross-domain retainability, which assesses the model\u2019s ability to generate concepts from a different domain. For example, as depicted in Fig. 3, we can define the \u2018domain\u2019 as either \u2018artistic style\u2019 or \u2018object class\u2019. If the unlearning target is the \u2018Van Gogh\u2019 style (i.e., the concept to be erased), then generating images in all styles except \u2018Van Gogh\u2019 but within the same object class (i.e., \u2018Dog\u2019 in Fig. 3) represents the in-domain case. Conversely, out-of-domain refers to generating images with different object classes (e.g., \u2018Horses\u2019 in Fig. 3), i.e., objects different from the one associated with the unlearning target. Ours WikiArt Figure 4. Image examples with the label \u201cVan Gogh Style\u201d from WIKIART (Phillips & Mackintosh, 2011) and UNLEARNCANVAS. Images of the same artistic style in UNLEARNCANVAS exhibit high stylistic consistency compared to WIKIART. UNLEARNCANVAS vs. WIKIART. To the best of our knowledge, WIKIART (Nichol, 2016) is the most relevant baseline dataset to ours. In Tab. A1, we provide a direct comparison of the key attributes of these two datasets. UNLEARNCANVAS differs from WIKIART in the following aspects. First, UNLEARNCANVAS includes a greater number of high-resolution images (15M) compared to WIKIART (2M), a factor that may enhance the training of state-of-the4 \fUNLEARNCANVAS: A Versatile Stylized Image Dataset to Benchmark Machine Unlearning and Beyond Finetuning with StyleCanvas Testbed  Preparation M achine  Unlearning Q KV Q KV Q KV Q KV Unlearned SD Unlearning Target: Van Gogh Style Q KV Q KV Q KV Q KV StableD iffusion Answer Set  Generation Unlearning  Target In-D om ain  Retain Targets Cross-D om ain  Retain Targets (Van Gogh Style) (Other Styles) (All Objects) ... Cartoon Sketch ... ... D og Cat ... ... ... ... Perform ance  Assessm ent In-D om ain Retainability Cross-D om ain Retain  Accuracy Unlearning  Accuracy Unlearning Effectiveness In-D om ain Retain  Accuracy Cross-D om ain Retainability Style/Object Classifier Van Gogh   Classifier (a) (b) (c) (d) Figure 5. An illustration of the evaluation pipeline proposed in this work using UNLEARNCANVAS when unlearning a specific target concept \u2018Van Gogh Style\u2019. Unlearning performances (including the unlearning effectiveness and retainability) are quantitatively assessed (marked in blue) to accurately reflect the unlearning performance portrait of each method. The unlearning target of the pipeline could traverse all the styles and objects to achieve a complete and comprehensive evaluation. art DMs. Second, UNLEARNCANVAS surpasses WIKIART in terms of both intra-style coherence and inter-style distinctiveness, as illustrated in Fig. 4, where images labeled with \u2018Van Gogh Style\u2019 from both datasets are compared. In UNLEARNCANVAS, the images exhibit high stylistic consistency, while WIKIART lacks the necessary clarity for precise assessment. This will hamper the MU evaluation as discussed in the challenges (C2) and (C3). 5. Evaluation Pipeline In this section, we introduce the evaluation pipeline and the benchmarked MU methods for UNLEARNCANVAS. Fig. 5 illustrates the proposed evaluation pipeline for MU in DMs with a designated unlearning target. It comprises four phases (I-IV) to evaluate unlearning effectiveness, retainability, generation quality, and efficiency. Phase I: Testbed preparation (Fig. 5a). Prior to unlearning, we commence by fine-tuning a specific DM, given by SD (StableDiffusion) v1.5, on UNLEARNCANVAS for text-toimage generation. We also train a vision transformer (ViTLarge) (Dosovitskiy et al., 2020) for the purposes of style and object classification after unlearning. We ensure that both the SD model and classifiers attain high performance in image generation and classification across all styles and objects. See detailed testbed preparation in Appx. A. Phase II: Machine unlearning (Fig. 5b). Next, we utilize the selected MU methods for benchmarking to update the DM acquired in Phase I, aiming to eliminate a designated unlearning target (e.g., \u2018Van Gogh style\u2019). This results in the unlearned model (denoted as \u03b8u), which is the DM updated to avoid generating the unlearning target. This unlearned DM will be used for unlearning performance evaluation. Phase III: Answer set generation (Fig. 5c). Further, we utilize the unlearned model \u03b8u to generate a set of images conditioned on both the unlearning-related prompts and other innocent prompts. To comprehensively evaluate unlearning performance, three types of answer sets are generated: Unlearning effectiveness: Images are generated in response to prompts that involve the unlearning target (e.g., any prompt like \u2018An image in Van Gogh style\u2019 for the case in Fig. 5). A successfully unlearned model will avoid generating images that match the unlearning target. In-domain retainability: Images are generated with prompts in the same \u2018domain\u2019 as the unlearning target (e.g., stylerelated prompts for style unlearning), where the prompts can cover all the other styles provided in UNLEARNCANVAS except the one to be unlearned, e.g., \u2018Van Gogh style\u2019 in Fig. 5. This assesses the unlearned model\u2019s ability to retain generation quality within the same domain. Cross-domain retainability: Images are generated with prompts in other domains. For instance, in the case of style unlearning as shown in Fig. 5, the object domain is considered separate from the style domain. Prompts related to objects are used to evaluate the DM\u2019s cross-domain generation capability after style unlearning. Phase IV: MU performance assessment (Fig. 5d). After Phase III, the answer set undergoes style/object classification for unlearning performance assessment. This classification results in three quantitative metrics \u2460-\u2462. \u2460Unlearning accuracy (UA): It represents the proportion of images generated by the unlearned DM using the unlearning target-related prompt (e.g., the Van Gogh style prompt in Fig. 5), which are not correctly classified into the corresponding class for the case. Therefore, a higher UA indicates better unlearning performance in preventing image generation from the unlearning target-related prompts. \u2461In-domain retain accuracy (IRA): It is given by the classification accuracy of images generated by the unlearned DM using innocent prompts (not relevant to the unlearning target) in the same domain, e.g., all styles other than the Van Gogh style in Fig. 5. 5 \fUNLEARNCANVAS: A Versatile Stylized Image Dataset to Benchmark Machine Unlearning and Beyond Table 3. Performance overview of different MU methods evaluated with UNLEARNCANVAS dataset. The performance assessment includes unlearning accuracy (UA), in-domain retain accuracy (IRA), cross-domain retain accuracy (CRA), and FID. Symbols \u2191or \u2193 denote whether larger or smaller values represent better performance. Results are averaged over all the style and object unlearning cases. The best performance regarding each metric is highlighted in bold. Method Effectiveness Efficiency Style Unlearning Object Unlearning FID (\u2193) Time Memory Storage UA (\u2191) IRA (\u2191) CRA (\u2191) UA (\u2191) IRA (\u2191) CRA (\u2191) (s) (\u2193) (GB) (\u2193) (GB) (\u2193) ESD (Gandikota et al., 2023a) 98.58% 80.97% 93.96% 92.15% 55.78% 44.23% 65.55 6163 17.8 4.3 FMN (Zhang et al., 2023a) 88.48% 56.77% 46.60% 45.64% 90.63% 73.46% 131.37 350 17.9 4.2 UCE (Gandikota et al., 2023b) 98.40% 60.22% 47.71% 94.31% 39.35% 34.67% 182.01 434 5.1 1.7 CA (Kumari et al., 2023) 60.82% 96.01% 92.70% 46.67% 90.11% 81.97% 54.21 734 10.1 4.2 SalUn (Fan et al., 2023) 86.26% 90.39% 95.08% 86.91% 96.35% 99.59% 61.05 667 30.8 4.0 \u2462Cross-domain retain accuracy (CRA): Similar to IRA, it is given by the classification accuracy of images generated by the unlearned DM using innocent prompts but in different domains, e.g., all object categories in Fig. 5. In addition to the above accuracy metrics, we also utilize the FID score (\u2463) to evaluate the distribution-wise generation quality of the unlearned DM. Furthermore, we will also monitor the efficiency of the unlearning process, taking into account factors such as run-time (\u2464), storage space requirements (\u2465), and memory costs (\u2466). In total, the performance evaluation involves 7 metrics (\u2460-\u2466). MU methods for DMs to be benchmarked. In this work, we assess the 5 most recently proposed MU methods for DMs, including ESD (Gandikota et al., 2023a), FMN (Zhang et al., 2023a), UCE (Gandikota et al., 2023b), CA (Kumari et al., 2023), and Salun (Fan et al., 2023). We remark that SA (Heng & Soh, 2023), a recently proposed method, is excluded from our evaluation due to its excessive computational resource requirements and time consumption. Detailed training settings for each unlearning method can be found in Appx. A. UA (%) IRA (%) CRA (%) FID Time Memory Storage ESD FMN UCE CA Salun Figure 6. Performance visualization of Table 3. For UA, IRA, and CRA, the results are averaged over all the style and object unlearning scenarios. Other metrics undertake the inverse operation as a smaller values represent better performance. Results are normalized to 0% \u223c100% per metric. 6. Experiment Results In this section, extensive experiments are conducted on UNLEARNCANVAS following the evaluation pipeline in Sec. 5. We show that previous MU evaluation, primarily focused on unlearning effectiveness, may present a biased view without considering retainability assessment (Tab. 3). We also find that preserving cross-domain retainability, as quantified by CRA, is more challenging than in-domain retainability (measured by IRA), a previously overlooked aspect in the literature. Furthermore, we show that no single MU method consistently excels across all metrics, emphasizing the potential for advancements in existing methods (Fig. 6). In addition, we present a detailed case study on ESD, showcasing how our benchmark offers a comprehensive analysis of an unlearning method\u2019s mechanism on forgetting and retaining information (Fig. 7). Last but not least, we show that different MU methods display distinct unlearning characteristics, which are inherently linked to their underlying unlearning mechanisms (Fig. 8). Overall performance assessment of MU for DMs. In Tab. 3, we provide an overview of the performance of existing unlearning methods, using the metrics discussed in Sec. 5. The key results are summarized below. (1) Retainability is essential for comprehensive assessment of MU in DMs. Previous evaluations heavily relied on UA (unlearning accuracy), but our assessment reveals a crucial insight: relying solely on UA can provide a skewed view of MU method performance. For example, when unlearning artistic painting styles (that we call style unlearning), ESD and UCE achieve similar UA (both around 98%), but their retainability (IRA and CRA) differs significantly, with one over 80% and the other around 60%. Similar conclusions apply to unlearning an image object (that we call object unlearning). These findings emphasize the importance of including retainability metrics in the evaluation framework, as demonstrated by UNLEARNCANVAS. (2) CRA is harder to retain than IRA. In both style and object unlearning, CRA presents a greater challenge for MU meth6 \fUNLEARNCANVAS: A Versatile Stylized Image Dataset to Benchmark Machine Unlearning and Beyond A1 B1 C1 B2 A2 Unlearn: Sea Test: Sea A1 A2 Unlearn: Cartoon Test: Cartoon Unlearn: Cats Test: Cats Unlearn: Ukiyoe Test: Ukiyoe Unlearn: Cartoon Test: Dadaism Unlearn: Sandwich Test: Dogs Unlearn: Human Test: Rust Unlearn: Mosaic Test: Human B1 B2 C1 C2 Before MU After MU C2 Test prompt template: \u2018A painting of {object} in {style} style.\u2019 \u274c \u274c \u274c \u274c \u274c \u274c Sandwich Butterfly Watercolor Crayon Bricks Warm Love Dogs Towers Rabbits Bears Sea Trees Photo Photo Photo Photo Before MU After MU A1, A2: Unlearning Effectiveness B1, B2, C1, C2: Retainability Desirable Case Undesirable Case \u274c Figure 7. Left: Heatmap visualization of the unlearning and retain accuracy of ESD on UNLEARNCANVAS. The x-axis shows the tested concepts for image generation using the unlearned model, while the y-axis indicates the unlearning target. Concept types are distinguished by color: styles in blue and objects in orange. The figure is separated into different regions to represent corresponding evaluation metrics and the unlearning scopes (A for UA, B for IRA, C for CRA; \u20181\u2019 for style unlearning, \u20182\u2019 for object unlearning). Diagonal regions (A1 and A2) indicate unlearning accuracy, and off-diagonal values (B1, B2, C1, and C2) represent retain accuracy. Higher values in lighter color denote better performance. The first row serves as a reference for comparison before unlearning. Zooming into the figure is recommended for detailed observation. Right: Representative cases illustrating each region with images generated before and after unlearning a specific concept. ods compared to IRA. This trend is consistent across various methods. For example, the UCE method (Gandikota et al., 2023b) exhibits a significant gap between its IRA and CRA, both in style (60.22% vs. 47.71%) and object (39.35% vs. 34.67%) unlearning scenarios. Similar patterns are observed with other methods like FMN and CA. This pattern suggests that while MU methods are somewhat effective at preserving concepts within the same domain (in terms of IRA), they face greater challenges in maintaining performance on unlearning target-unrelated concepts across domains (CRA), which has been overlooked in previous studies due to the absence of a multi-label dataset like UNLEARNCANVAS for systematic unlearning analysis. (3) One MU method can perform differently in different domains. Take the FMN method (Gandikota et al., 2023b) as an example. It exhibits an evident performance difference between the two domains, achieving a UA of 88.48% in style unlearning but a much lower UA of 45.64% in object unlearning. In addition, its IRA and CRA metrics show a notable gap, with an IRA of 90.63% for object unlearning compared to 56.77% for style unlearning. These examples highlight that a method\u2019s effectiveness in unlearning is not consistent across different scopes. (4) No single method can excel in all aspects. Each method exhibits unique strengths and weaknesses. For example, while ESD (Gandikota et al., 2023a) typically leads an improvement in UA, it may lag behind in terms of IRA and CRA, e.g., in the case of object unlearning. Also, the SalUn method (Fan et al., 2023) shows exceptional retainability, with a CRA of 99.59% in object unlearning, but its UA in the same category is lower at 86.91%. Furthermore, the UCE method (Kumari et al., 2023) enjoys remarkably high UA and efficiency but suffers from degradation in FID and retainability. For better visualization, in Fig. 6, we show this pattern by plotting the performance of different metrics reported in Tab. 1, illustrating that no method dominates across all radar chart dimensions. This underscores the complexity of unlearning in generative models and the need for further advancement. A closer look into retainability: A case study on ESD. We next perform a more in-depth analysis of DMs\u2019 retainability after unlearning, with a focus on the ESD method (Gandikota et al., 2023a), which has shown superiority in Tab. 3 for style unlearning. In Fig. 7 (left), we present a heatmap that illustrates the per-style/object unlearning accuracy (diagonal) and retain accuracy (off-diagonal) for ESD. The heatmap segments are labeled to indicate different regions of interest (ROIs) related to our evaluation metrics: A1 and A2 for Unlearning Accuracy (UA) in style and object unlearning, B1 and B2 for In-Domain Retain Accuracy (IRA) in style and object unlearning, and C1 and C2 for Cross-Domain Retain Accuracy (CRA) in style and object unlearning. We highlight the following observations. First, we observe distinct behaviors of ESD in unlearning styles compared to objects. The ROI C2 appears darker than C1, suggesting a reduced ability to retain styles during 7 \fUNLEARNCANVAS: A Versatile Stylized Image Dataset to Benchmark Machine Unlearning and Beyond Unlearn:Architectures Unlearn:Bears Unlearn:Birds Unlearn:Butter\ufb02y Unlearn:Cats Unlearn:Dogs Unlearn:Fishes Unlearn:Flame Unlearn:Flowers Unlearn:Frogs Unlearn:Horses Unlearn:Human Unlearn:Jelly\ufb01sh Unlearn:Rabbits Unlearn:Sandwiches Unlearn:Sea Unlearn:Statues Unlearn:T owers Unlearn:Trees Unlearn:Waterfalls Sea Trees Human Architectures Flame Birds Unlearn:Architectures Unlearn:Bears Unlearn:Birds Unlearn:Butter\ufb02y Unlearn:Cats Unlearn:Dogs Unlearn:Fishes Unlearn:Flame Unlearn:Flowers Unlearn:Frogs Unlearn:Horses Unlearn:Human Unlearn:Jelly\ufb01sh Unlearn:Rabbits Unlearn:Sandwiches Unlearn:Sea Unlearn:Statues Unlearn:T owers Unlearn:Trees Unlearn:Waterfalls Sea Trees Human Statues Butter\ufb02y Fishes Flowers Dogs Sandwiches Flame Architectures Bears Birds H um an Flam es Frogs Salun ESD M U M ethod Unlearning Target M U Before (a) ESD (b) SalUn (c) Illustration Figure 8. Visualization of the unlearning directions of (a) ESD and (b) SalUn. This figure illustrates the conceptual shift of the generated images of an unlearned model conditioned on the unlearning target. Images generated by the post-unlearning models are classified and used to understand this shift. Edges leading from the object in the left column to the right signify that images generated conditioned on unlearning targets are instead classified as the shifted concepts after unlearning. This reveals the primary unlearning direction for each unlearning method. The most dominant unlearning direction for an object is visualized. Figure (c) provides visualizations of generated images using the prompt template \u2018A painting of {object} in Sketch style.\u2019 with object being each unlearning target. the unlearning of objects. Similar patterns are observable between B2 and B1. These findings indicate that an MU method might demonstrate a preference within a specific domain but face challenges in others. For a clearer explanation, Fig. 7 (right) provides image generation examples of the DM before and after unlearning within different ROIs. Second, we observe that style/object unlearning is relatively easier compared to retaining the generation performance of unlearned DMs conditioned on unlearning-unrelated prompts. This is evident in the contrast between ROIs A1/A2 and ROIs B1/B2 or C1/C2. One exception is the case of unlearning the object \u2018Sea\u2019, which exhibits relatively low UA in A2. This may be attributed to the image generation post-unlearning still resembling sea waves, as shown by the image examples in Fig. 7 (right) for A2. Third, we present a similar accuracy heatmap for SalUn in Fig. A1. In comparison to ESD, SalUn demonstrates more consistent performance across different unlearning scenarios. However, SalUn does not achieve the same level of UA as ESD, as evidenced by the lighter diagonal values in Fig. A1. This observation underscores the trade-off between unlearning effectiveness and retainability of MU in DMs. Understanding unlearning method\u2019s behavior via unlearning directions. As noted earlier, different unlearning methods display distinct unlearning behaviors. To gain insights into the underlying reasons for these differences, Fig. 8 (a) and (b) visualize the \u2018unlearning directions\u2019 for ESD and SalUn, respectively. These unlearning directions are determined by connecting the unlearning target with the predicted label of the generated image from the unlearned DM conditioned on the unlearning target. As shown in Fig. 8 (a), ESD demonstrates a focused shift in image generation after object unlearning, with a predominant transition towards generating images labeled by \u2018Sea\u2019 and \u2018Trees\u2019. This behavior arises from ESD\u2019s optimization process, designed to steer the generation of the DM away from a predefined concept. Consequently, images generated by the ESD-induced unlearned model consistently lack clearly identifiable objects, resembling waves and trees, which leads to their classification into the \u2018Sea\u2019 and \u2018Trees\u2019 classes; see Fig. 8 (c) for examples of generated images. In contrast, SalUn exhibits a more diverse range of unlearning directions, shifting images to 11 different objects. This diversity results from SalUn\u2019s requirement to replace the unlearning target with a random concept. As shown in Fig. 8 (c), images generated by SalUn post-object unlearning still maintain some object contours (different from the original unlearning target) and better retain style information compared to ESD. Expanding applications to style transfer. In addition to evaluating DMs\u2019 unlearning performance, it\u2019s worth noting that UNLEARNCANVAS has the potential to extend its util8 \fUNLEARNCANVAS: A Versatile Stylized Image Dataset to Benchmark Machine Unlearning and Beyond ity to another generative modeling task: style transfer (Liu et al., 2021; Huang & Belongie, 2017; Li et al., 2019; Chen & Schmidt, 2016; Park & Lee, 2019; An et al., 2021). One significant challenge in evaluating style transfer methods is the absence of ground truth images for the given reference style image and content image (Cai et al., 2023). Thus, quantitative evaluations are also very limited in the literature. In Appx. C, we utilize UNLEARNCANVAS to assess 9 latest style transfer methods, providing enriched and precise evaluations. While not the primary focus of this work, this illustrates the great potential of UNLEARNCANVAS to contribute broader visual generative modeling tasks. 7."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2402.11592v2",
+                    "title": "Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark",
+                    "abstract": "In the evolving landscape of natural language processing (NLP), fine-tuning\npre-trained Large Language Models (LLMs) with first-order (FO) optimizers like\nSGD and Adam has become standard. Yet, as LLMs grow {in size}, the substantial\nmemory overhead from back-propagation (BP) for FO gradient computation presents\na significant challenge. Addressing this issue is crucial, especially for\napplications like on-device training where memory efficiency is paramount. This\npaper proposes a shift towards BP-free, zeroth-order (ZO) optimization as a\nsolution for reducing memory costs during LLM fine-tuning, building on the\ninitial concept introduced by MeZO. Unlike traditional ZO-SGD methods, our work\nexpands the exploration to a wider array of ZO optimization techniques, through\na comprehensive, first-of-its-kind benchmarking study across five LLM families\n(Roberta, OPT, LLaMA, Vicuna, Mistral), three task complexities, and five\nfine-tuning schemes. Our study unveils previously overlooked optimization\nprinciples, highlighting the importance of task alignment, the role of the\nforward gradient method, and the balance between algorithm complexity and\nfine-tuning performance. We further introduce novel enhancements to ZO\noptimization, including block-wise descent, hybrid training, and gradient\nsparsity. Our study offers a promising direction for achieving further\nmemory-efficient LLM fine-tuning. Codes to reproduce all our experiments are at\nhttps://github.com/ZO-Bench/ZO-LLM .",
+                    "authors": "Yihua Zhang, Pingzhi Li, Junyuan Hong, Jiaxiang Li, Yimeng Zhang, Wenqing Zheng, Pin-Yu Chen, Jason D. Lee, Wotao Yin, Mingyi Hong, Zhangyang Wang, Sijia Liu, Tianlong Chen",
+                    "published": "2024-02-18",
+                    "updated": "2024-02-26",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.CL"
+                    ],
+                    "main_content": "Introduction Fine-tuning pre-trained large language models (LLMs) has become the de-facto standard in the current paradigms of natural language processing (NLP) (Raffel et al., 2023; Sanh et al., 2022). First-order (FO) optimizers, e.g., SGD (Amari, 1993) and Adam (Kingma & Ba, 2014), have been the predominant choices for LLM fine-tuning. However, as LLMs continue to scale, they encounter significant memory overhead due to the back-propagation (BP) required for FO gradient computation. For example, computing the gradient of the LLM OPT-13B requires 12\u00d7 more memory cost than the model inference. This leads to the challenge of achieving memory-efficient fine-tuning in LLMs. Advancements in addressing this challenge could also facilitate technological breakthroughs in related areas, such as on-device training, where memory efficiency is in high demand (Han et al., 2015; Zhu et al., 2023). To enhance memory efficiency, an emerging solution is to replace a BP-required FO optimization method with a BPfree optimizer during LLM fine-tuning. This was initially proposed by Malladi et al. (2023), where the FO gradient is approximated using a finite difference of function values. Despite its new application to LLM fine-tuning, the underlying optimization principle used in Malladi et al. (2023) is commonly known as zeroth-order (ZO) optimization, and the function value-based gradient estimate is referred to as the ZO gradient estimate (Flaxman et al., 2005; Nesterov & Spokoiny, 2017; Duchi et al., 2015; Ghadimi & Lan, 2013; Liu et al., 2020). Malladi et al. (2023) employed the classical ZO stochastic gradient descent (ZO-SGD) algorithm (Ghadimi & Lan, 2013), termed MeZO, to fine-tune the pre-trained LLMs and leveraged the BP-free characteristics of ZO optimization to reduce memory costs. However, from the perspective of ZO optimization, in addition to ZOSGD, many other ZO optimization methods have not yet been explored in the context of LLM fine-tuning. Thus, it remains elusive whether there are potential improvements in accuracy and/or efficiency that can be achieved through a benchmarking study of ZO optimization for LLM finetuning. This yields the primary question to be explored: 1 arXiv:2402.11592v2  [cs.LG]  26 Feb 2024 \fRevisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark (Q) Can we establish a benchmark for ZO optimization in LLM fine-tuning, explore the overlooked optimization principles, and advance the current state of the art? To address (Q), our work introduces several key innovations compared to the most relevant work (Malladi et al., 2023). We explore a broader range of ZO optimization methods beyond ZO-SGD and examine various task and model types, and evaluation metrics. We conduct a detailed comparative analysis of different ZO optimization methods, shedding light on the often-overlooked forward gradient method (Ren et al., 2022) and other ZO optimization techniques in LLM fine-tuning. This benchmarking study helps reveal the pros and cons of these methods in accuracy and efficiency. Extended from the gained insights, we propose to further improve ZO optimization-based LLM fine-tuning using techniques of block-wise descent, hybrid ZO and FO training, and gradient sparsity. In summary, our key contributions are listed below. \u2022 We create the first benchmark for ZO optimization in the context of LLM fine-tuning. This benchmark includes our investigations into 6 BP-free or ZO optimization methods, 5 LLM families, 3 tasks of varying complexities, and 5 finetuning schemes, covering both full-parameter and parameterefficient fine-tuning (PEFT) approaches. \u2022 Assisted by our benchmark, we reveal a range of previously overlooked optimization principles and insights for LLM fine-tuning with ZO optimization. These include the significance of aligning tasks to enhance ZO optimization, the role of forward gradient as an LLM fine-tuning baseline, and the trade-offs between algorithm complexity, finetuning accuracy, query and memory efficiency. \u2022 In addition to a holistic assessment of existing ZO optimization methods for LLM fine-tuning, we introduce novel enhancements to ZO optimization, including block-wise ZO optimization, hybrid ZO and FO fine-tuning, and sparsityinduced ZO optimization. These proposed techniques aim to improve the accuracy of ZO LLM fine-tuning while maintaining memory efficiency. 2. Related Work Parameter-efficient fine-tuning (PEFT). Early efforts (Houlsby et al., 2019; Lin et al., 2020) involved inserting trainable adapters, which are compact feedforward networks, between the layers of the pre-trained model. More recently, various PEFT strategies have been proposed. For instance, Adapter-based methods (Houlsby et al., 2019; Chen et al., 2022; Luo et al., 2023; Karimi Mahabadi et al., 2021; Pfeiffer et al., 2020) insert a few tunable yet highly compact modules into pre-trained models. LoRA (Hu et al., 2021a) employs trainable low-rank weight perturbations to the pre-trained model, effectively reducing the required number of fine-tuning parameters. Prompt-based learning (Gao et al., 2020; Hu et al., 2021b; Tan et al., 2021) has demonstrated effectiveness in various NLP tasks. Additionally, methods like prompt tuning (Lester et al., 2021) and prefix tuning (Li & Liang, 2021) incorporate learnable continuous embeddings into the model\u2019s hidden states to condition the frozen model for specific downstream tasks. The following work (Liu et al., 2021) demonstrates its applicability on various model scales. While these state-of-the-art PEFT techniques have significantly reduced the number of parameters required for fine-tuning, they still incur memory costs associated with caching numerous activations due to the use of back-propagation (BP) (Malladi et al., 2023). Zeroth-order optimization. Zeroth-order (ZO) optimization is a technique that uses finite differences to estimate gradients. Such algorithms utilize function value oracle only, yet share similar algorithm structure with first-order (FO) gradient-based counterpart methods. ZO optimization usually enjoys provable (dimension-dependent) convergence guarantees, as discussed in various works (Flaxman et al., 2005; Nesterov & Spokoiny, 2017; Duchi et al., 2015; Ghadimi & Lan, 2013). These methods have draw considerable attention due to its effectiveness in a wide range of modern machine learning (ML) challenges (Liu et al., 2020), including the adversarial attack and defense (Chen et al., 2017; Tu et al., 2019; Ye et al., 2018; Ilyas et al., 2018; Zhang et al., 2022b; Verma et al., 2023; Zhao et al., 2019; Hogan & Kailkhura, 2018; Shu et al., 2022), model-agnostic contrastive explanations (Dhurandhar et al., 2019), enhancing transfer learning through visual prompting (Tsai et al., 2020), computational graph unrolling (Vicol et al., 2023), and optimizing automated ML processes (Gu et al., 2021a; Wang et al., 2022). Beyond standard ML, it finds application in policy search in reinforcement learning (Vemula et al., 2019), network resource management (Liu et al., 2018), ML-based optimization of scientific workflows (Hoffman et al., 2022; Tsaknakis et al., 2022; Chen et al., 2024), and on-chip learning enhancements (Gu et al., 2021b). Despite its wide range of use cases, the application of ZO optimization in ML has primarily been restricted to small model scales. This limitation is attributed to the high variance and slow convergence associated with ZO optimization, which are exacerbated by model dimensions. To scale up ZO optimization, several acceleration techniques have been proposed. These include the integration of historical data to refine ZO gradient estimators (Meier et al., 2019; Cheng et al., 2021), leveraging gradient structural information (Singhal et al., 2023) or sparsity to diminish the dependency of ZO optimization on problem size (Wang et al., 2017; Cai et al., 2022; 2021; Balasubramanian & Ghadimi, 2018; Ohta et al., 2020; Gu et al., 2021b; Chen 2 \fRevisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark et al., 2024), the reuse of intermediate features (Chen et al., 2024) and random perturbation vectors (Malladi et al., 2023) in the optimization process. These advancements suggest a growing potential for the application of ZO optimization in more complex and large-scale ML problems. BP-free training for large models. Training large models, especially LLMs, is memory-consuming due to the involved large computation graphs for BP (Ren et al., 2021; Kim et al., 2023). Thus BP-free methods have become a recent focus in the deep learning (DL) community. Forward gradient learning (Baydin et al., 2022; Ren et al., 2022; Silver et al., 2021; Belouze, 2022), built upon the forward-mode automatic differentiation (AD), provides an alternative to ZO optimization for BP-free training. Different from ZO optimization, it relies on the forward-mode AD to calculate a forward (directional) gradient. However, one main limitation of the forward gradient is its requirement of full access to the AD software and the deep model, making it less memory-efficient than ZO optimization and impractical for tackling black-box problems (Chen et al., 2024). The specifications of BP-free methods include greedy layer-wise learning (N\u00f8kland & Eidnes, 2019), input-weight alignment (Boopathy & Fiete, 2022), forward-forward algorithm (Hinton, 2022), synthetic gradients (Jaderberg et al., 2017), BBT/BBTv2 evolutionary algorithms (Sun et al., 2022a;b), gradient guessing using special low dimensional structure of neural networks (Singhal et al., 2023) and other black-box methods which optimize the prompts (Prasad et al., 2023; Deng et al., 2022; Chai et al., 2022). Many of these algorithms are also motivated through the lens of seeking DL\u2019s biological interpretation. Applying ZO optimization to fine-tune pre-trained LLMs is particularly intriguing because it combines the advantages of being BP-free and utilizing pre-training. This enables the scalability of ZO optimization to large-scale LLMs while maintaining memory efficiency. MeZO (Malladi et al., 2023) introduced a scalable ZO-SGD algorithm to efficiently fine-tune LLMs with up to 60 billion parameters, achieving competitive results compared to first-order optimization methods and structured fine-tuning approaches like LoRA. They also provided theoretical insights into why ZO methods can be effective for LLMs. This opens the gateway for efficient BP-free LLM fine-tuning and largely motivates our ZO benchmark study. 3. Reviewing ZO Optimization and Beyond This work\u2019s core objective is to benchmark and harness the potential of ZO (zeroth-order) optimization in LLM finetuning, eliminating the need for (first-order) back-propagation (BP) during finetuning and thus achieving memory efficiency (Malladi et al., 2023). It is worth noting that the ZO optimization technique utilized in (Malladi et al., 2023) is primarily the basic version, specifically, ZO stochastic gradient descent (ZO-SGD). There are more advanced ZO optimization methods available, as summarized in (Liu et al., 2020). Thus, this section is dedicated to reviewing a broader range of ZO optimization approaches and shedding light on the previously overlooked principles for LLM fine-tuning. Basics of ZO optimization. ZO optimization serves as a gradient-free alternative to first-order (FO) optimization, approximating FO gradients through function value-based gradient estimates, which we call ZO gradient estimates, as discussed in (Flaxman et al., 2004; Nesterov & Spokoiny, 2017; Ghadimi & Lan, 2013; Duchi et al., 2015). Thus, a ZO optimization method typically mirrors the algorithmic framework of its corresponding FO optimization counterpart. However, it substitutes the FO gradient with the ZO gradient estimate as the descent direction. Various techniques exist for performing ZO gradient estimation. In this paper, we focus on the randomized gradient estimator (RGE) (Nesterov & Spokoiny, 2017; Duchi et al., 2015), a method that relies on the finite difference of function values along a randomly chosen direction vector. RGE has also been used by Malladi et al. (2023) to achieve memory-efficient fine-tuning for LLMs. Its preference in LLM fine-tuning is attributed to its query efficiency, i.e., a low number of function queries. Given a scalar-valued function f(x) where x \u2208Rd of dimension d, the RGE (referred to as \u02c6 \u2207f(x)) is expressed using central difference: \u02c6 \u2207f(x) = 1 q q X i=1 \u0014f(x + \u00b5ui) \u2212f(x \u2212\u00b5ui) 2\u00b5 ui \u0015 (RGE) where ui is a random direction vector typically drawn from the standard Gaussian distribution N(0, I), q is the number of function queries, and \u00b5 > 0 is a small perturbation stepsize (also known as smoothing parameter). Malladi et al. (2023) employed RGE by setting q = 1 and ui = u. Yet, it is worth noting that the query number q strikes a balance between the ZO gradient estimation variance and the query complexity. It has been shown in (Duchi et al., 2015; Liu et al., 2018) that the variance of RGE is roughly in the order of O(d/q), where O(\u00b7) signifies the Big O notation. The rationale behind RGE stems from the concept of the directional derivative (Duchi et al., 2015): As \u00b5 \u21920 (letting q = 1), the finite difference of function values in (RGE) approaches f \u2032(x, u) := uT \u2207f(x), representing the directional derivative of f(x) along the random direction u \u223c N(0, I). Subsequently, RGE yields \u02c6 \u2207f(x) \u2192f \u2032(x, u)u as \u00b5 \u21920. Moreover, the directional derivative provides us an unbiased gradient estimator of \u2207f(x): Eu[f \u2032(x, u)u] = Eu[uuT \u2207f(x)] = \u2207f(x). (1) With the above background, the RGE \u02c6 \u2207f(x) can be interpreted as an approximation of the FO gradient \u2207f(x) using the directional derivative. 3 \fRevisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark Forward gradient: A missing BP-free baseline in LLM fine-tuning. As a byproduct of connecting RGE to (1), we obtain the directional derivative-based gradient estimate, \u2207f(x) \u2248f \u2032(x, u)u, which is known as the forward gradient (Forward-Grad) (Baydin et al., 2022; Ren et al., 2022). Different from RGE that relies solely on the finite difference of function values, Forward-Grad requires the use of forward mode automatic differentiation (AD) but eliminates the need for the backward evaluation in the implementation of deep model fine-tuning or training. In other words, ForwardGrad is BP-free and can serve as another alternative gradient estimation method that improves the memory efficiency of LLM fine-tuning. We stress that Forward-Grad is a possibly overlooked BP-free optimizer. Given its unbiasedness as shown in (1), it could serve as an upper performance bound for ZO optimization in theory. A focused spectrum of ZO optimization methods. Next, we provide a brief overview of the ZO optimization methods to be focused on in this work. Specifically, we will include: ZO-SGD (Ghadimi & Lan, 2013) that Malladi et al. (2023) has employed for LLM fine-tuning, ZO-SGD using sign-based gradient estimation (ZO-SGD-Sign) (Liu et al., 2019a), ZO-SGD with momentum (ZO-SGD-MMT) (Malladi et al., 2023), ZO-SGD with conservative gradient update (ZO-SGD-Cons), and the ZO variant of the Adam optimizer (ZO-Adam) (Chen et al., 2019). The aforementioned methods can be unified into the following generic optimization framework in solving minx f(x): xt+1 = xt \u2212\u03b7th( \u02c6 \u2207f(xt)), (2) where xt denotes the updated solution at the tth iteration, \u03b7t > 0 is the learning rate, and h(\u00b7) is a certain descent direction post-processing operation. In (2), we omit the inclusion of the stochastic mini-batch for empirical risk minimization for ease of presentation. For instance, ZOSGD can be expressed as (2) when h( \u02c6 \u2207f(x)) = \u02c6 \u2207f(x). Similarly, ZO-SGD-Sign can be derived if h( \u02c6 \u2207f(x)) = sign( \u02c6 \u2207f(x)), where sign(\u00b7) represents element-wise sign operation. Another example is ZO-SGD-Cons by setting h( \u02c6 \u2207f(x)) = arg ming\u2208{0,\u2212\u02c6 \u2207f(x), \u02c6 \u2207f(x)} f(xt +\u03b7tg). We refer readers to Appx. B for more algorithmic details of (2) as applied to the ZO optimization approaches we covered. Our rationale for selecting the aforementioned ZO optimization approaches for LLM fine-tuning is based on two key considerations: (1) We prioritize ZO optimization methods that require minimal modifications to the existing FO optimizer, ensuring ease of implementation for LLM finetuning. (2) We focus on methods with distinct algorithmic characteristics, allowing us to explore a diverse range of optimization strategies for improving LLM performance. Regarding (2), we include ZO-SGD-Sign as it employs 1bit gradient quantization and represents one of the simplest Table 1. Test accuracy (%) of pretrained Roberta-Large model finetuned on SST2 and RTE tasks using ZO and FO optimization methods with (\u2713) and without (\u2717) text alignment. The evident performance degradation is highlighted in bold. Method SST2 RTE \u2713 \u2717 Difference \u2713 \u2717 Difference FO-SGD 91.6 91.5 0.1 70.9 61.4 9.5 ZO-SGD 89.4 79.2 10.2 68.7 60.4 8.3 ZO-Adam 89.8 79.2 10.6 69.2 58.7 10.5 ZO optimization methods. Additionally, we include ZOSGD-MMT and ZO-SGD-Cons as they incorporate certain forms of \u2018adaptive learning\u2019 into the descent step updates. The former utilizes momentum based on historical gradient information, while the latter allows for the heuristics-based selection of the descent direction. Furthermore, ZO-Adam is one of the most complex ZO optimizers due to its utilization of moving averages and adaptive learning rates. Task alignment in ZO optimization for LLM fine-tuning. Scaling up ZO optimization for deep model training, as discussed in (Chen et al., 2024), is exceedingly challenging due to its high variance, which is dependent on the model size. Nevertheless, LLM pre-training offers a unique advantage by enabling the fine-tuner to start from a well-optimized pre-trained model state. This graceful model initialization makes ZO optimization potentially scalable to LLM finetuning tasks (Malladi et al., 2023). Even in this pretrainingfinetuning paradigm, another crucial factor, which we call \u2018task alignment\u2019, still plays a key role in achieving satisfactory ZO fine-tuning performance. The \u2018task alignment\u2019 refers to aligning the fine-tuning task with the format of the pre-training task, given by the next token or sentence prediction. For example, Gao et al. (2020); Malladi et al. (2023) have transformed downstream text classification tasks into next token prediction tasks by introducing well-crafted input prompts. These prompts serve as bridges to align the fine-tuning tasks with the pre-training ones, facilitating ZO optimization when initiated from the pre-trained model. As a warm-up experiment, Tab. 1 empirically justifies the importance of task alignment when applying ZO optimization to LLM fine-tuning on the simple binary classification task by comparing scenarios with and without the use of pre-defined prompts to achieve task alignment. We fine-tune the entire Roberta-Large (Liu et al., 2019b) model on SST2 (Socher et al., 2013) and RTE (Wang et al., 2019) datasets with two selected ZO methods: ZO-SGD (i.e., MeZO in (Malladi et al., 2023)) and ZO-Adam. We compare their performance with that of the FO method (FO-SGD). The task alignment is achieved with the template <CLS>SENTENCE. It was [terrible|great].<SEP> for SST dataset and another template <CLS>SENTENCE1? [Yes|No], SENTENCE2.<SEP> for RTE. As we can see, without 4 \fRevisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark prompt-based text alignment, there is a substantial performance drop across ZO fine-tuning methods. Both ZO-SGD and ZO-Adam yield about 10% and 8% accuracy degradation on SST2 and RTE, respectively. In contrast, FO-SGD suffers less from the absence of task alignment. This suggests that the task alignment is particularly beneficial to ZO LLM fine-tuning. It is also worth noting that crafting effective prompts for task alignment can be non-trivial, as prompt design is context-dependent and can affect the fine-tuning performance. In this work, we follow (Gao et al., 2020) and (Malladi et al., 2023) to align the fine-tuning tasks to the pretrained ones. 4. Benchmarking ZO Optimization for LLM Fine-Tuning In this section, we delve into the empirical performance of ZO optimization in LLM fine-tuning. Our benchmarking effort includes evaluating accuracy and efficiency, accounting for different downstream task complexities (ranging from simple classification to more complicated reasoning task), and considering various language model types. 4.1. Benchmark Setups LLM fine-tuning tasks, schemes, and models. We begin by introducing the tasks and the fine-tuning schemes. We focus on three tasks, considering their complexity from low to high, which include (1) the simplest binary classification task, Stanford Sentiment Treebank v2 (SST2) (Socher et al., 2013), (2) the question-answering task, Choice Of Plausible Alternatives (COPA) (Roemmele et al., 2011), and (3) the commonsense reasoning task, WinoGrande (Sakaguchi et al., 2021). When evaluating memory efficiency, we also consider the task of multi-sentence reading comprehension (MultiRC) (Khashabi et al., 2018). For LLM fine-tuning on these tasks, we explore four parameter-efficient finetuning (PEFT) schemes: full-tuning (FT) that fine-tunes the entire pre-trained model, low-rank adaptation (LoRA) by imposing low-rank weight perturbations (Hu et al., 2021a), prefix-tuning (Prefix) by appending learnable parameters to token embedding (Li & Liang, 2021), and prompt-tuning (Prompt) (Liu et al., 2022) by introducing a series of learnable tokens attached to the input to adapt the fixed model to downstream tasks. We refer readers to Appx. A for details. Furthermore, we incorporate several representative language models, including Roberta-Large (Liu et al., 2019b), OPT (Zhang et al., 2022a), LLaMA2 (Touvron et al., 2023), Vicuna (Zheng et al., 2023), and Mistral (Jiang et al., 2023). Setup and implementation of ZO optimization. To train the previously mentioned LLM fine-tuners, we utilize the ZO optimization methods introduced in Sec. 3. These include ZO-SGD (i.e. MeZO (Malladi et al., 2023)), ZO-SGDSign, ZO-SGD-MMT, ZO-SGD-Cons, and ZO-Adam. For Table 2. Performance of LLM fine-tuning on SST2 over pretrained Roberta-Large and OPT/1.3B. Best performance among ZO methods (including Forward-Grad) is highlighted in bold. SST2 Roberta-Large OPT-1.3B FT LoRA Prefix Prompt FT LoRA Prefix Prompt FO-SGD 91.4 91.2 89.6 90.3 91.1 93.6 93.1 92.8 Forward-Grad 90.1 89.7 89.5 87.3 90.3 90.3 90.0 82.4 ZO-SGD 89.4 90.8 90.0 87.6 90.8 90.1 91.4 84.4 ZO-SGD-MMT 89.6 90.9 90.1 88.6 85.2 91.3 91.2 86.9 ZO-SGD-Cons 89.6 91.6 90.1 88.5 88.3 90.5 81.8 84.7 ZO-SGD-Sign 52.5 90.2 53.6 86.1 87.2 91.5 89.5 72.9 ZO-Adam 89.8 89.5 90.2 88.8 84.4 92.3 91.4 75.7 comparison, we also present the performance of ForwardGrad, which relies on the forward mode auto-differentiation rather than BP. We also provide the performance of two FO optimizers: (FO-)SGD and (FO-)Adam. Before applying the aforementioned optimizers to the LLM fine-tuning tasks, we follow (Gao et al., 2020; Malladi et al., 2023) to align the fine-tuning task format with the token or sentence prediction-based pre-training task, as demonstrated in Tab. 1. We run ZO (or BP-free) optimizers and FO optimizers for 20000 and 625 iterations respectively, as outlined in (2). Note that ZO optimization takes a longer convergence time as shown in (Liu et al., 2020). When implementing (RGE), unless specified otherwise, we set the query budget per gradient estimation to q = 1. We determine the values of other hyperparameters, such as the smoothing parameter and learning rate, through a grid search for each method. Following (Malladi et al., 2023), if not specified otherwise, we adopt the mixed precision (MP) training for FO optimizers, where the model is loaded with the full precision (i.e., 32 bits) but training is carried out in half precision with 16 bits (FP16). For ZO optimizers, half-precision training is adopted throughout the model loading and training. We also remark that neither MP nor FP16 can be applied to Forward-Grad. We refer readers for more details to Appx. C. Evaluation metrics. We evaluate ZO LLM fine-tuning using two sets of metrics: accuracy and efficiency. Accuracy measures the fine-tuned model\u2019s test-time performance in specific tasks, such as test accuracy in classification tasks. Efficiency includes various measurements like memory efficiency (in terms of peak memory usage and GPU cost), query efficiency (i.e., the number of function queries required for ZO optimization ), and run-time efficiency. These metrics collectively provide insights into the resources needed for ZO LLM fine-tuning, helping assess its feasibility and cost-effectiveness in practical scenarios. 4.2. Experiment Results ZO fine-tuning on SST2: A pilot study. In Tab. 2, experiments are conducted to compare the performance of different BP-free and BP-based (FO-SGD) methods on one of the simplest LLM downstream task: the binary classification task with SST2 dataset. We investigate two model architec5 \fRevisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark tures, the medium-sized Roberta-Large and the larger model OPT-1.3B. Several key results are summarized below. First, ZO-Adam seems to be the most effective ZO method, achieving the best performance in 4 out of 8 fine-tuning settings. However, as will be shown later, this is achieved at the cost of additional memory consumption. This is not surprising considering that ZO-Adam has the highest algorithmic complexity, as explained in Sec. 3. Second, Forward-Grad is a competitive method compared to the ZO methods, especially in the FT (full-tuning) setting. This indicates that Forward-Grad may be suitable for problems of a larger scale, and make it a compelling baseline of ZO LLM fine-tuning. Additionally, when the complexity of the fine-tuning scheme decreases (e.g., Prompt), the advantage of Forward-Grad over function value-based ZO methods diminishes. Third, The performance of ZO methods exhibits high variance, as evidenced by the fluctuating relative rankings across different experimental scenarios, although extensive hyperparameter search efforts have been made. For example, the effectiveness of ZO-Adam degrades dramatically in the (OPT-1.3B, Prompt) setting. In addition, the MeZO method (i.e., ZO-SGD) used in (Malladi et al., 2023) does not always emerge as the top-performing ZO optimizer for LLM fine-tuning across various settings. This variance is not surprising and could be largely attributed to the high variance of RGE (Nesterov & Spokoiny, 2017; Duchi et al., 2015). Fourth, ZO-SGD-Cons and ZO-SGD-MMT also demonstrate strong performance as ZO optimizers in LLM finetuning. However, ZO-SGD-Sign, the simplest ZO optimization method, tends to be the weakest approach, except in the simplest fine-tuning setting Prompt. The above observations motivate us to extend our explorations, investigating the effectiveness of ZO methods across a broader spectrum of models and more complex tasks. ZO fine-tuning on downstream tasks COPA and WinoGrande under OPT-13B. Extended from the experiments on SST2, Fig. 1 presents the fine-tuning performance on COPA and WinoGrande dataset using a larger model, OPT13B. We summarize our key observations when the problem scales up and becomes more complicated. First, compared to the previous results, the performance gap among different ZO methods are much enlarged. In the meantime, the performance gap between FO and ZO methods are also widened. For example, in the experiments with WinoGrande, the FO methods (FO-SGD and FO-Adam) outperform all the other ZO methods by a large margin. This observation shows the scalability bottleneck intrinsic to ZO methods, when dealing with larger models and/or more complex tasks. FT LoRA Prefix Prompt 80 90 Test Accuracy 90 87 79 80 87 84 85 84 82 88 84 81 80 89 78 80 82 89 79 81 87 89 78 83 86 88 88 89 90 88 88 85 OPT-13B + COPA FT LoRA Prefix Prompt 60 70 Test Accuracy 62.3 62.6 58.8 63.4 60.9 64.0 58.6 63.8 62.5 63.3 59.0 62.5 62.1 63.4 59.0 63.8 63.3 64.0 58.7 62.7 64.4 62.9 59.4 61.8 69.2 66.9 63.6 65.7 65.8 68.9 67.7 66.5 OPT-13B + WinoGrande ZO-SGD ZO-SGD-MMT ZO-SGD-Cons ZO-SGD-Sign ZO-Adam Forward-Grad FO-SGD FO-Adam Figure 1. Results of OPT-13B on the tasks COPA and WinoGrande fine-tuned using ZO/FO optimizers in different PEFT settings. Table 3. Performance of different LLMs finetuned with LoRA on COPA and WinoGrande using different ZO/FO methods. Table format is consistent with Tab. 2. OPT-13B LLaMA2-7B Vicuna-7B Mistral-7B COPA FO-SGD 88 85 84 90 FO-Adam 88 84 81 90 Forward-Grad 89 82 84 88 ZO-SGD 87 86 83 90 ZO-SGD-CONS 88 85 83 89 ZO-Adam 89 83 84 89 WinoGrande FO-SGD 66.9 66.9 66.5 76.4 FO-Adam 68.9 69.5 70.0 76.9 Forward-Grad 62.9 64.3 65.6 70.1 ZO-SGD 62.6 64.3 65.6 68.7 ZO-SGD-CONS 63.3 64.6 65.3 68.5 ZO-Adam 64.0 64.4 65.5 69.5 Second, certain ZO methods exhibit exceptional stability across varied conditions: Despite a general trend towards variability, specific ZO methods, e.g., ZO-Adam and ZOSGD-MMT, demonstrate stability in their performance. The reason for this could be that these algorithms integrate variance-reduced optimization techniques (such as momentum and adaptive learning rate) to ZO optimization and become more adaptable and resilient to the variances of ZO gradient estimation (Chen et al., 2019). Third, the LoRA tuning method is consistently robust when paired with various ZO algorithms. This resilience across different ZO methods suggests that LoRA\u2019s mechanism is inherently more adaptable to the variations in ZO optimization strategies, providing a stable and reliable tuning approach in diverse settings. We will peer into the performance of LoRA below. In Tab. 3, we present how different ZO methods perform on 6 \fRevisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark 0 300 600 900 1200 1500 Sequence Length 0 50 100 150 Memory (GB) FO-SGD ZO-SGD Figure 2. Memory comparison between FO-SGD and ZO-SGD full fine-tuning across various sequence lengths with a fixed effective batch size of 2. Memory evaluation was conducted using synthetic text generated from random sequences of the specified shapes. For shorter sequences (i.e., < 700), the memory usage of FO-SGD remains relatively stable since the memory consumption for storing gradients during BP surpasses that needed for activations. (LoRA, COPA) and (LoRA, WinoGrande) across a range of widely-used LLM families. For ease of computation, we focus on a subset of ZO optimization methods, including ZO-SGD, ZO-SGD-CONS, and ZO-Adam. As we can see, in some scenarios with the COPA dataset, some BP-free methods exhibit effectiveness comparable to, or even superior to, that of FO methods (FO-SGD and FO-Adam). For example, Forward-Grad and ZO-Adam outperform the best FO method on model OPT-13B and Vicuna-7B. Conversely, for the more difficult task WinoGrande, a performance gap of 5% \u223c6% between FO and ZO methods still exists across different models. Efficiency analyses. In Tab. 4, we present a comparison of the efficiency performance of various ZO/FO optimizers when fine-tuning the full OPT-13B model on the MultiRC dataset with a batch size 4. We evaluate the efficiency in the following dimensions: memory cost (in GB), the consumption of GPU resources (number of GPUs), and runtime cost per optimization iteration (in seconds). All the experiments are carried out in the same environment. Several key observations can be made below. First, almost all ZO methods (except ZO-Adam) demonstrate comparable levels of efficiency, requiring only a single GPU (A100) for LLM fine-tuning. This observation is expected as ZO methods entail relatively straightforward optimization steps, primarily based on function evaluations, as depicted in RGE. Among the examined ZO optimization methods, ZO-Adam incurs higher memory costs due to its algorithmic complexity. Second, in comparison to FO methods, ZO optimization reduces runtime costs per iteration, e.g., by approximately 33.3% for ZO-SGD compared to FOSGD. Third, Forward-Grad appears to be the point at which ZO optimization methods lose their memory efficiency advantage over FO methods. Additionally, we note that the Table 4. The peak memory cost (in GB), the required GPU resources, and the runtime cost (in seconds) of each optimizer when fine-tuning the full OPT-13B model on MultiRC with an averaged 400 context length. The order of included optimizers is ranked based on the memory cost. The per-iteration runtime in seconds (s) is averaged over 100 iterations. Notably, Forward-Grad is marked by \u2217, indicating its incompatibility with efficiency-enhancing techniques such as MP (mixed-precision training) and FP16 (halfprecision training). Optimizer Memory \u21d3 Consumed GPUs \u21d3 Runtime Cost ZO-SGD 29 GB 1\u00d7A100 1.8s ZO-SGD-Cons 29 GB 1\u00d7A100 4.2s ZO-SGD-Sign 29 GB 1\u00d7A100 1.8s ZO-SGD-MMT 53 GB 1\u00d7A100 1.8s ZO-Adam 80 GB 2\u00d7A100 1.9s Forward-Grad\u2217 138 GB 2\u00d7A100 19.8s FO-SGD 161 GB 3\u00d7A100 2.7s FO-Adam 257 GB 4\u00d7A100 2.8s substantial runtime cost of Forward-Grad compared to FO optimizers (FO-SGD and FO-Adam) is possibly attributed to its incompatibility with MP and FP16. Furthermore, we examine the memory cost of LLM finetuning vs. the input sequence length. In Fig. 2, we compare the memory efficiency between ZO-SGD and FO-SGD across various sequence lengths (i.e. the token number per sample). We employed synthetic texts generated from random sequences with specified shapes. In all experiments, an effective batch size of 2 was maintained. The memory consumption of ZO-SGD maintains a consistent level, whereas FO-SGD begins to demand significantly more memory as the sequence length extends. This trend is amplified particularly for longer sequences (e.g., exceeding 700 as depicted in Fig. 2), where the memory allocated for activations overwhelms that required for gradient storage. We provide a theoretical analysis of this phenomenon in Appx C. Ablation study on query budget. Recall from (1) that Forward-Grad provides us with an unbiased gradient estimate with respect to the FO gradient, in contrast to the function value-based biased ZO gradient estimate. However, in the above experiments, we have not observed a significant advantage brought by Forward-Grad. We hypothesize that this is due to the smallest query budget q = 1 we used, which, despite its query efficiency, may not fully showcase the unbiased benefit. Inspired by the above, Fig. 3 explores the impact of varying query budget (q) on the performance of Forward-Grad and ZO-SGD (i.e., MeZO). As we can see, both the accuracies of Forward-Grad and ZO-SGD are improved with an increased query budget. However, the performance rise for ForwardGrad is much more evident. For example, when the query number is larger than 500, Forward-Grad outperforms ZOSGD by a large margin 1% \u223c2% and approaches FO-SGD. These observations underscore that the inherent advantages 7 \fRevisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark 1 10 50 100 500 1000 Query Number 90 91 92 93 94 Accuracy (%) FO-SGD (93.7%) 90.1 90.8 90.8 91.2 91.5 91.9 90.3 90.7 90.9 91.4 92.8 93.3 ZO-SGD Forward-Grad Figure 3. LoRA-based fine-tuning accuracy of OPT-1.3B on SST2 using ZO-SGD and Forward-Grad over different budgets. of Forward-Grad become apparent only when a sufficient number of queries are used. The downside of using a higher query budget is the increase in computation cost, which scales linearly with q. 5. Extended Study to Improve ZO Fine-Tuning Beyond the benchmarking effort in Sec. 4, we will explore algorithmic advancements to further improve the effectiveness of ZO LLM fine-tuning. We will leverage the following techniques: (1) Block-wise ZO fine-tuning; (2) Hybrid ZO and FO fine-tuning; and (3) sparsity-induced ZO gradient estimation. These designs aim to reduce the large variances in gradient estimation when using ZO algorithms. Block-wise ZO optimization enhances fine-tuning performance. It has been shown in (Liu et al., 2018) that using a coordinate-wise deterministic gradient estimator can reduce ZO optimization variance, although this scheme is difficult to scale. In a similar vein, we ask if RGE when estimating a FO gradient block-wise can also improve the performance of ZO optimization. The key idea is to split the LLM into different blocks and then apply the ZO gradient estimator to each block of parameters. For example, OPT-1.3B entails p = 26 parameter blocks, necessitating p forward passes for ZO gradient estimation per fine-tuning step. Our rationale is that by conducting gradient estimation block-wise, the resulting gradient estimate\u2019s variance will be reduced, thereby potentially improving the fine-tuning performance. In Tab. 5, we compare the performance of the ZO finetuning baseline MeZO (Malladi et al., 2023) (corresponding to ZO-SGD with a query budget of q = 1 in Sec. 4) with its block-wise RGE-based variant, which we term ZO-SGDBlock. For a fair comparison with ZO-SGD-Block in terms of query complexity, we also present the performance of another variant of ZO-SGD that uses the full model-wise RGE but takes the same query number q as ZO-SGD-Block per iteration. Notably, ZO-SGD-Block outperforms ZOSGD in different query budget settings across different fineTable 5. Performance comparison of OPT-1.3B on the SST2 & WinoGrande tasks between ZO-SGD and ZO-SGD-Block. The # of parameter blocks in ZO-SGD-Block is set to p = 26. Thus, ZOSGD w/ the query budget q = 26 has the same forward pass count as ZO-SGD-Block. MeZO corresponds to ZO-SGD w/ q = 1. Best performance for each task is highlighted in bold. Optimizer Forward Pass # SST2 WinoGrande MeZO 1 90.83 55.5 ZO-SGD (q = 26) 26 91.28 55.7 ZO-SGD-Block 26 93.69 57.2 tuning tasks, showing the benefit of block-wise ZO tuning. Trade-off between performance and memory efficiency via hybrid ZO-FO training. The primary source of memory cost in LLM fine-tuning arises from the BP process, which involves passing gradients from the deep layers to the shallow layers of the model. Hence, to enhance memory efficiency, a potential approach is to confine BP within the deep layers without propagating it to the shallow layers. Moreover, ZO optimization can be employed for the shallow layers without the need for BP. The above approach of combining FO optimization for deep layers and ZO optimization for shallow layers results in a hybrid ZO-FO fine-tuning scheme for LLMs. Table 6. The trade-off between memory cost (in GB) v.s. finetuning accuracy (%) using the hybrid ZO-FO training on the OPT1.3B model over the SST2 dataset. The memory or accuracy gap vs. that of the pure ZO-SGD method is noted by \u2206. Memory (GB) Accuracy (%) ZO Layer # Memory \u2206Memory Accuracy \u2206Accuracy 0 (FO-SGD) 24.29 11.07 91.22 1.98 4 23.33 10.11 91.12 1.88 8 22.01 8.79 90.79 1.55 12 20.43 7.21 89.48 0.24 16 18.98 5.76 89.42 0.18 20 15.43 2.21 89.27 0.03 24 (ZO-SGD) 13.22 0.00 89.24 0.00 Tab. 6 presents the performance of using the hybrid ZO-FO fine-tuning scheme on (OPT-1.3B, SST2). We examine different variants of this hybrid scheme by deciding \u2018where\u2019 to split between ZO optimization (for shallow layers) and FO optimization (for deep layers). Suppose that the model consists of n layers, and we designate the first k \u2208[1, n] layers for ZO optimization, while the remaining (n \u2212k) layers are allocated for FO optimization. The pure ZO optimization approach corresponds to p = n. The results presented in Tab. 6 demonstrate that employing ZO optimization on only the first third of the model layers (i.e., k \u22648) can yield performance comparable to that achieved by fully utilizing FO optimization while also reducing memory usage by approximately 10%. Furthermore, when half of the layers employ ZO optimization (i.e., k \u226512), the performance achieved is similar to that of full ZO fine-tuning. 8 \fRevisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark Table 7. Performance of fine-tuning OPT-1.3B on COPA & SST2 datasets using ZO-SGD at different sparsity ratios. A sparsity of 0% represents the baseline, the vanilla ZO-SGD, i.e., MeZO (Malladi et al., 2023). Performance that surpasses the baseline (w/ 0% sparsity) is highlighted in bold. COPA Sparsity (%) 0 10 20 30 40 50 60 70 80 90 Accuracy (%) 73.00 75.00 75.00 70.00 70.00 70.00 70.00 7.000 70.00 71.00 SST2 Sparsity (%) 0 10 20 30 40 50 60 70 80 90 Accuracy (%) 90.83 91.51 92.20 92.32 91.74 92.43 92.43 92.20 91.51 92.66 Gradient pruning benefits performance. We next explore gradient pruning, a technique known for accelerating model training without compromising convergence (McDanel et al., 2022). Our key idea is to induce sparse parameter perturbations for reducing gradient estimation variance in RGE. We begin by leveraging magnitude-based pruning (Frankle & Carbin, 2018; Chen et al., 2020) to obtain the layer-wise sparsity ratios. We then generate random pruning masks (following these layer-wise sparsity ratios) and apply them to the weight perturbations in RGE per ZO fine-tuning step. Tab. 7 shows the performance of the sparsity-induced ZO gradient estimation in LLM fine-tuning as a function of the overall sparsity ratio. It becomes evident that choosing a moderate sparsity ratio (e.g., 20%) can lead to improved performance over the vanilla ZO optimizer, ZO-SGD. 6."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2310.08782v3",
+                    "title": "Selectivity Drives Productivity: Efficient Dataset Pruning for Enhanced Transfer Learning",
+                    "abstract": "Massive data is often considered essential for deep learning applications,\nbut it also incurs significant computational and infrastructural costs.\nTherefore, dataset pruning (DP) has emerged as an effective way to improve data\nefficiency by identifying and removing redundant training samples without\nsacrificing performance. In this work, we aim to address the problem of DP for\ntransfer learning, i.e., how to prune a source dataset for improved pretraining\nefficiency and lossless finetuning accuracy on downstream target tasks. To our\nbest knowledge, the problem of DP for transfer learning remains open, as\nprevious studies have primarily addressed DP and transfer learning as separate\nproblems. By contrast, we establish a unified viewpoint to integrate DP with\ntransfer learning and find that existing DP methods are not suitable for the\ntransfer learning paradigm. We then propose two new DP methods, label mapping\nand feature mapping, for supervised and self-supervised pretraining settings\nrespectively, by revisiting the DP problem through the lens of source-target\ndomain mapping. Furthermore, we demonstrate the effectiveness of our approach\non numerous transfer learning tasks. We show that source data classes can be\npruned by up to 40% ~ 80% without sacrificing downstream performance, resulting\nin a significant 2 ~ 5 times speed-up during the pretraining stage. Besides,\nour proposal exhibits broad applicability and can improve other computationally\nintensive transfer learning techniques, such as adversarial pretraining. Codes\nare available at https://github.com/OPTML-Group/DP4TL.",
+                    "authors": "Yihua Zhang, Yimeng Zhang, Aochuan Chen, Jinghan Jia, Jiancheng Liu, Gaowen Liu, Mingyi Hong, Shiyu Chang, Sijia Liu",
+                    "published": "2023-10-13",
+                    "updated": "2023-11-18",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.AI"
+                    ],
+                    "main_content": "Introduction The abundance of the training data has long been regarded as the key driver of the contemporary machine learning (ML) algorithms [1\u20133]. However, the untrimmed, ever-growing training dataset could not only introduce training biases that compromise the model performance [4\u20136], but also poses an almost insurmountable obstacle to high training efficiency [1, 2]. Therefore, understanding the impact of the training data and selecting the most critical samples has emerged as an important goal, collectively referred to as the problem of dataset pruning (DP). Although the feasibility and the promise of DP have been unveiled in numerous applications, such as noisy data cleansing [7\u20139], continue learning [10\u201313] and active learning [14], greater emphasis is placed on the in-domain training setting, i.e., the pruned training set share the similar distribution as the evaluation set. Examples of methods to condense the training dataset with lossless generalization (on the in-distribution testing dataset) include data influence functions [10, 15\u201323], model training dynamics [11\u201314, 24\u201328], and coreset selection [29\u201336]. In contrast, our work investigates the 37th Conference on Neural Information Processing Systems (NeurIPS 2023). arXiv:2310.08782v3  [cs.LG]  18 Nov 2023 \fAbyssinian  Cat American  Bulldog Shiba  Inu Black  Swan English  Springer Indigo  Bunting Tennis  Ball Cab Loafer Domain Mapping Source Target \u274c \u274c \u274c \u274c \u274c \u274c \u274c \u274c \u274c \u274c \u274c \u274c \u274c \u274c \u274c \u274c \u274c \u274c \u274c \u274c \u274c Source/Target Data Features from Surrogate Model LM/FM Data Pruning \u274c Target Data  Source Data  Retained Data  Pruned Data  ImageNet \u2192OxfordPets ImageNet \u2192StanfordCars Stage I: Pretrain (5\u00d7 Speed Up \ud83d\ude80) Source Dataset Pruning Surrogate Model (Small) Stage II: Finetune Efficient Transfer Learning Source Model (Large) Figure 1: Left: An illustration of the proposed dataset pruning methods (LM and FM) and their performance overview. Large scale source dataset is pruned by LM and FM through a small surrogate model (ResNet-18). Large foundation models can achieve up to 5\u00d7 speed-up on pretraining without no downstream performance drop. Right: Downstream performance overview at different pruning ratios when ResNet-101 [54] is used as the source model (trained on the pruned source dataset) with ImageNet [55] as the source dataset transferring to OxfordPets [56] and StanfordCars [57]. The feature extractor is fixed during finetuning. problem of DP in the transfer learning paradigm [37\u201339], which has emerged as a popular way to leverage the knowledge of a foundation model learned on a source dataset (referring to \u2018pretraining\u2019 stage) to further enhance the performance on a cross-domain target dataset (referring to \u2018finetuning\u2019 stage). Recent evidence [40\u201342] has shown that some source data points could make a harmful influence in the downstream performance. In particular, the previous study [41] showed that removing specific source data classes can improve the transfer learning accuracy of a pretrained model. Yet, efficiently identifying those harmful source data classes for improved transfer learning is a highly non-trivial task. Firstly, the unique pretrain-finetune paradigm complicates the analysis of the influence of source data on downstream performance, making it an indirect and very challenging process. Consequently, existing gradient and data influence-based in-domain DP methods [10, 19\u201323] cannot be na\u00efvely adapted to the transfer setting. Secondly, transfer learning encompasses a wide range of source training methods other than supervised learning, such as self-supervised learning (SSL) [43\u201352]. Therefore, a generic DP framework is desired given various pretraining scenarios. Thirdly, from an efficiency standpoint, the design of efficient DP algorithms is non-trivial. Even in the non-transfer learning paradigm, a majority of current in-domain DP methods [11, 25, 26] introduce substantial computational overheads. For instance, influence function-based DP methods [10, 15\u201322] necessitate the calculation of the inverse Hessian matrix of model parameters, which is a highly demanding and computationally intensive process. Moreover, training dynamics-based methods [11\u201314, 24\u201328, 53], such as GraNd-Score [26] and Forgetting-Score [11], necessitate model training on the entire dataset multiple times. As a result, the development of an efficient and effective DP framework tailored for transfer learning remains a significant challenge in the field. The most relevant work to ours is [41], which proposed a brute-force method to evaluate the influence of every source class in the downstream task following a leave-one-out principle. While this method effectively selects the most influential source classes, it is prohibitively slow, as the data selection process demands significantly higher computational resources than performing transfer learning itself. This brings us to the central question addressed in this paper: (Q) How can we extend DP to transfer learning with minimal computation overhead, broad applicability, and improved target performance? To address (Q), we formally define the task of DP (dataset pruning) for transfer learning. We start by uncovering the limitations of conventional in-domain DP methods when applied to transfer learning tasks, and establish an intrinsic connection between source dataset pruning for transfer learning and source-target domain mapping. In the supervised pretraining paradigm, we propose an effective and scalable label mapping (LM)-based framework capable of pinpointing the most beneficial source 2 \fdata labels for a cross-domain target dataset. Furthermore, we extend the idea of LM to feature mapping (FM) for the SSL pretraining protocol, where data labels are unavailable. To achieve a greater practical impact, we demonstrate that our proposed methods (LM and FM) can facilitate effective DP over a small, simple surrogate source model, and further translate the positive impact of the pruned source dataset to other larger, more complex source models. We provide a schematic overview and result highlights in Fig. 1. Our contributions can be summarized as follows: \u2776We connect the concept of DP (dataset pruning) to transfer learning for the first time and formalize the problem of DP for transfer learning. \u2777We develop a highly efficient and effective framework for DP in transfer learning, leveraging the source-target domain mapping with customized LM (label mapping) and FM (feature mapping) for both supervised and SSL (self-supervised learning) settings. Our approach is principled and achieves significant speedups compared to the state-of-the-art methods. \u2778We empirically show the effectiveness of our proposals (LP and FP) on 8 downstream tasks. We find that the source dataset (ImageNet) can be pruned up to 40% \u223c80% without sacrificing the downstream performance, together with a 2\u00d7 \u223c5\u00d7 speed-up in the pretraining stage. Our proposal also unlocks a door to prune a dataset using a simple surrogate source model (e.g., ResNet-18) and then reuse the pruned dataset to improve transfer learning on a larger source model (e.g., ResNet-101). \u2779Lastly, we show that our proposed DP framework can benefit other computationally-intensive transfer learning techniques. For example, DP for adversarial pretraining [58] leads to a 1% \u223c2% improvement in downstream performance with time consumption similar to standard pretraining. 2 Related Work Dataset pruning. DP (dataset pruning) is an emerging technique to improve the data efficiency of model training by selecting the most representative training samples or removing the less influential ones [10\u201328]. Thus, the problem of coreset selection [29\u201336] can also be considered a form of DP. Prior studies on DP range from clustering-based methods [59\u201361] to the more recent score-based pruning methods [10\u201328]. In the latter approach, an importance score is assigned to each training data point to quantify its influence on a particular permanence metric during model learning. Specifically, score-based DP methods can be broadly categorized into two main groups: influence function-based approaches [10, 15\u201323] and training dynamics-based approaches [11\u201314, 24\u201328, 53]. The first category measures data influence by examining the effect of data removal on the learning algorithm used during training [15\u201318] and the model\u2019s prediction [10, 19\u201323]. However, influence functionbased approaches typically require high computational costs due to the need for high-order derivatives and complex optimization methods, such as bi-level optimization, as used in [22]. Although an approximate influence score can be obtained efficiently, it may result in a large estimation error [62]. In the second category, training dynamics-based approaches find statistical indicators for pruned data from the training trajectory. Examples of DP metrics include data loss/error [25\u201327], prediction confidence [24], model gradient norm [26], forgetting event [11], and compactness [12\u201314, 28]. However, these methods typically require repeated training to ensure the representativeness of the collected statistics [11, 25, 26]. In addition to improving data efficiency, DP has been applied in a variety of contexts, such as noisy label cleansing [7\u20139], continue learning [10\u201313, 63], active learning [14], and reducing annotation cost of a finetuner [64]. Data valuation & attribution. Data valuation [19, 41, 65\u201370] and attribution [71\u201373] are research streams related to dataset pruning that aim to quantify the influence of training data points on model\u2019s performance. Unlike dataset pruning, these approaches do not focus primarily on training efficiency but instead aim to enhance interpretability [74], adversarial robustness [68, 70], generative model design [75], and data acquisition quality [72]. Representative methods include Shapley value-based methods [65, 66], datamodels [73], sampling-based methods [65, 68], and proxy-based methods [67, 72]. Recently, Kim et al. [35] proposed to select data samples from a large public dataset (Open-Set) for self-supervised learning given a specific target dataset through distribution mismatch. However, it fails to make a general framework in both supervised and self-supervised scenarios, leaving the former under-explored. The most relevant work to ours is [41], which leverages a leaveone-out analysis to quantify the influence of source data on downstream tasks in the context of transfer learning. With this inspiration, we propose to connect DP to transfer learning in this work. 3 \fRecent advancements in transfer learning. Transfer learning has been a prominent area over the past decade [37\u201339]. Significant strides have been made in understanding, analyzing, and improving various aspects of this technique [58, 76\u201384]. Recent studies [77, 78] have rigorously analyzed the relationship between source model performance and downstream task effectiveness, arguing that a narrow focus on minimizing source training loss may not lead to improved transfer learning results. To improve transfer learning, adversarial training has been shown to benefit the transferability of source pretraining on target downstream tasks [58, 76]. There also exist studies to identify and understand the failure cases in transfer learning [79\u201384]. Other research [77, 78] has examined model transferability from the perspective of model sharpness and argues that a good pretrained model should be situated in a flat basin in the downstream loss landscape. Last but not the least, transfer learning has progressed in several other directions, such as self-supervised learning [43\u201352], model weight pruning [85\u201389], and visual prompting [90\u201399]. 3 Problem Formulation In this section, we introduce some essential preliminaries on transfer learning and DP (dataset pruning), and elucidate the design challenge of DP for transfer learning. Preliminaries on transfer learning and connection to DP. Let g \u25e6f denote a deep neural network that consists of a feature extractor f and a classification head g, where \u25e6denotes the function composition. Given source and target datasets DS and DT , we study transfer learning in the \u201cpretrainfinetune\u201d paradigm. The primary goal of pretraining is to obtain a high-quality feature extractor fS : X \u2192Z, which draws a mapping from the input space (X) to the deep representation space (Z) in a data-rich source domain (DS). Popular pertaining recipes include supervised learning (SL) [84, 100] and self-supervised learning (SSL) [43\u201346] depending on whether the source labels (yS) are available or not in DS. In the finetuning stage, the pretrained model is further trained on a specific downstream task under the target dataset DT . Transfer learning expects improved downstream performance over training on DT from scratch. In this work, we consider two finetuning protocols, linear probe (LP) and full-finetune (FF), with DT being a labeled dataset. LP finetunes the linear classification head g with a fixed feature extractor fS, acquired from pretraining. In contrast, FF finetunes the entire model g\u25e6f from the initialization fS. FF typically yields a better transfer learning accuracy than LP, but the former takes a higher computation cost. Our motivation for connecting transfer learning with DP comes from a recent data-based perspective on transfer learning [41]. The study shows that removing certain source data classes from DS can potentially improve the accuracy of a finetuned model on DT . However, the task of evaluating the transfer effects of source data and removing their influence from a pre-trained source model has not been addressed efficiently. The approach developed in [41] involves a leave-one-out analysis to estimate the influence of a source class c on a target example t, which is computed as the prediction discrepancy of the finetuned source model at t when the class c is either included or excluded from DS. During this process, one must train multiple source models (over 7000 models on ImageNet in [41]) from scratch over different subsets of DS for a given target task. This approach becomes computationally unaffordable when dealing with large source datasets like ImageNet given limited computing resources. To address this challenge, we propose a DP perspective on transfer learning. Problem of interest: DP for transfer learning. Next, we introduce the background of DP and the problem we focus on. Let D = {z1, z2, . . . , zN} denote a dataset consisting of n samples, where each sample zi is represented as a pair (xi, yi), with xi denoting the input feature vector and yi denoting the corresponding label. DP aims to generate a pruned dataset \u02c6 D = {\u02c6 z1, \u02c6 z2, . . . , \u02c6 zM} \u2282D with M < N, which can reduce the training cost without a significant decrease in model generalization performance when trained on \u02c6 D. In the context of [41], instead of individual source data sample {\u02c6 zi}, the entire source classes are evaluated and selected in terms of transfer influences. Based on the above, we define the problem of our interest below. (DP for transfer learning) How to prune source data classes to obtain \u02c6 DS (a subset of DS), with lossless or improved transfer learning accuracy of the source model (fS) on a target task DT ? DP for transfer learning has two key distinctions (\u2776-\u2777) from the vanilla DP setup. First (\u2776), DP must be performed in the source domain (DS), while its effectiveness is evaluated based on the target 4 \fdomain (DT ). This \u2018cross-domain\u2019 challenge makes the design of efficient and effective DP highly non-trivial. For example, prior work [41] utilizes a computationally-intensive leave-one-out analysis. Classical influence function-based methods [10, 19\u201323, 101], which trace the eventual model\u2019s prediction through the learning algorithm and back to its training data, are also computationally infeasible due to the complex bi-level optimizer and the calculation of high-order derivatives. Second (\u2777), the pre-trained source model (fS) in today\u2019s transfer learning regime is typically of a large scale. This motivates us to develop a DP method that can be independent of the source model, while the pruned source dataset \u02c6 DS remains effective in the original transfer learning setup. Given this challenge, we will design DP methods for transfer learning using a simple surrogate source model to avoid the computation on the large source model fS (see Fig. 1 for an illustration). 0% 20% 40% 60% 80% Source Data Pruning Ratio 86 88 90 92 94 Target Test Accuracy (%) Flowers102-LP Random Moderate GraNd SP SSP Forget EL2N InfSum InfMax No Prune 0% 20% 40% 60% 80% Source Data Pruning Ratio 76 78 80 82 84 86 88 90 Target Test Accuracy (%) OxfordPets-LP Figure 2: Transfer learning accuracy of existing DP methods on ImageNet at different pruning ratios, where ResNet-101 is the source model, and linear probing (LP) is used for downstream finetuning on the target datasets Flowers102 (Left) and OxfordPets (Right). Conventional DP methods lack effectiveness on transfer learning. Given the challenges (\u2776-\u2777) posed by DP for transfer learning, we further conduct a preliminary study to investigate the effectiveness of existing 8 DP methods, including SP [20], SSP [20], GRAND [26], EL2N [26], MODERATE [3], FORGET [53], INFMAX [53], and INFSUM [53]. Our results show that these methods are unable to yield significant improvements over random source data pruning. Fig. 2 shows the transfer learning performance of the ResNet-101 [54] model trained on different pruned versions of the ImageNet dataset when LP-based finetuning is conducted on the downstream Flowers102 [102] and OxfordPets [56] datasets. As we can see, in transfer learning, random pruning is a solid baseline for various state-of-the-art DP methods, which have demonstrated superior performance to the former in the non-transfer learning regime. Therefore, it is crucial to develop an efficient and effective DP method specifically tailored for transfer learning. 4 Label/Feature Mapping-based DP for Transfer Learning In this section, we first introduce a simple yet powerful DP method called label mapping (LM) by leveraging the class-discriminative capability of a supervised source model. We then extend the concept of LM to feature mapping (FM), suitable for self-supervised pretraining. LM-based DP for supervised pretraining. Following the notations used in Sec. 3, we represent the model obtained through supervised learning on the source dataset (DS) as gS \u25e6fS. This model predicts a source label given an input example (x). In DP for transfer learning, the source data classes serve as the variables to be pruned. Thus, we express DS as DS = {C(1) S , . . . , C(N) S } for N source classes, C(i) S denotes the set of data points belonging to the source class i. Additionally, the pruner has access to the target dataset DT = {t1, . . . , tn}, which consists of n target data points. Our objective is to utilize the information provided by gS \u25e6fS, DS, and DT to devise a computationally efficient DP criterion. Importantly, our criterion should not involve model training, distinguishing it from the non-scalable approach in [41]. An important observation by [78] in transfer learning is that transferred knowledge improves transferability. This insight is supported by the loss landscape analysis in transfer learning: The finetuned weights may remain within the flat basin of the pretrained weights for enhanced transfer learning performance. We next propose to extract the \u2018transferred knowledge\u2019 by leveraging the class-discriminative capability of the source model gS \u25e6fS on the target data samples in DT . Specifically, we focus on monitoring the responsiveness of the source label predictions made by gS \u25e6fS when using target samples {ti \u2208DT }n i=1 as input data. Here we resize these target samples to ensure their resolution alignment with source data. For the ith source data class C(i) S , we then define its pruning score below: sLM(C(i) S ) = n X j=1 1(gS \u25e6fS(tj) = i), for i = 1, 2, . . . , N, (LM) 5 \fwhere 1(\u00b7) represents an indicator function that evaluates to 1 when the condition \u00b7 is satisfied, and 0 otherwise. We refer to the aforementioned formula as LM (Label Mapping) since the condition gS \u25e6fS(tj) = i establishes a mapping between the predicted labels of the target samples {tj}n j=1 by the source model and the corresponding source labels i. The larger sLM in (LM) signifies that a specific source class is more frequently utilized to interpret the target data. Consequently, this indicates a more tightly connected relationship between the source class and the transferred knowledge. Therefore, we prune (or select) source data classes with low (or high) sLM values. Although LM is simple, it offers several advantages. Firstly, it is highly efficient in computation. Given a pre-trained source model gS \u25e6fS, the calculation of LM only requires forward passes through the model. In cases where obtaining the pretrained model in the source domain is challenging, our 0% 20% 40% 60% 80% Source Data Pruning Ratio 91.5 92.0 92.5 93.0 93.5 94.0 94.5 95.0 Target Test Accuracy (%) RN-18 RN-101 No Prune No Prune 40% 60% Pruning Ratio (%) 0 1 2 3 4 5 6 Time Consumption (h) 5.4 3.5 2.4 Surrogate Model Training Source Model Pretraining (a) Downstream performance. (b) Time consumption. Figure 3: Preliminary studies on the usage of surrogate models for LM. (a) The downstream performance (on Flowers102) of using the source model ResNet-101 trained on the pruned ImageNet delivered by LM at different pruning ratios. Here LM is conducted using either RN-101 or a smaller surrogate model ResNet-18. (b) Computation time decomposition analysis for obtaining the pretrained model using ResNet-18 as surrogate model with different pruning ratios. proposal supports an alternative approach: training a smaller and simpler surrogate model to carry out LM. This surrogate model can effectively replace the complex pretrained model gS \u25e6fS and facilitate the efficient execution of the pruning process. As we show in Fig. 3a, employing ResNet-18 [54] is sufficient to successfully prune the source dataset (ImageNet [55]). The resulting DP scheme remains effective to improve transfer learning utilizing other larger source models, such as ResNet-101 [54]. In addition, Fig. 3b shows that the computational overhead incurred by training the small surrogate model (ResNet-18) for DP is insignificant, compared to the time saved during the pretraining phase of a larger source model (ResNet-101) on the pruned dataset for transfer learning. Lastly, pretraining on the subset found by LM can guide the source model towards a flatter region in the downstream loss landscape, (see results in Fig. A4). The source model trained on the LM-pruned dataset achieves a higher flatness score than baselines, which aligns with the understanding of transfer learning in [78]. FM-based DP framework for self-supervised pretraining. The requirement for labeled source data in LM may pose limitations on the application of DP methods, particularly in the context of self-supervised learning (SSL)-based pertaining. To address this limitation, we introduce a new method called FM (Feature Mapping). Unlike LM, FM determines the DP scheme using only the feature extractor network fS, the unlabeled source dataset DS, and the target dataset DT . This allows us to overcome the dependence on labeled source data, making FM applicable in SSL scenarios. The inspiration for FM is derived from the deep clustering technique [103\u2013105] operating in the representation space, which can generate pseudo source labels using cluster indices provided by e.g., K-means. With the assistance of deep clustering, we can represent the unlabeled source dataset DS as DS = {C(1) S , C(2) S , . . . , C(K) S }, where C(k) S denotes the set of source data samples within cluster k. Building upon the similar spirit as LM, we propose ranking the importance of pseudo source classes for DP by evaluating the source feature extractor\u2019s responsiveness to target data samples. To achieve this objective, we quantify the responsiveness of fS for a target sample t as follows: r(t) = arg min k\u2208[K] \r \r \rfS(t) \u2212Ex\u2208C(k) S [fS(x)] \r \r \r 2 , (1) where Ex\u2208C(k) S [fS(x)] is the centroid of source data within the cluster k, and r(t) is the nearest pseudo label as the responsiveness of fS against t. The FM score then integrates (1) with (LM): sFM(C(i) S ) = n X j=1 1(r(tj) = i), for i = 1, 2, . . . , K, (FM) where different from (LM), K represents the number of pseudo source classes produced by deep clustering, and r(tj) corresponds to the source feature extractor\u2019s prediction on tj. It is important to note that the value of K is a free parameter of the deep clustering process. Our empirical study in Fig. A2 shows that FM is quite robust to the choice of K without sacrificing the benefit of DP for 6 \ftransfer learning compared to using the unpruned source dataset. Lastly, it is worth mentioning that FM can also be applied in the context of supervised pretraining by specifying data labels as clusters. 5 Experiments In this section, we provide a comprehensive set of experiments and analyses to showcase the effectiveness of our proposed methods (LM and FM) in diverse transfer learning scenarios. 5.1 Experiment setup Datasets and models. In line with existing transfer learning benchmarks [41, 106], we utilize ImageNet-1K [55] for pretraining and 8 datasets as downstream tasks. These datasets include DTD [107], Flowers102 [102], UCF101 [108], Food101 [109], SUN397 [110], OxfordPets [56], StanfordCars [57], and CIFAR10 [111]. Please refer to Tab. A1 for more details about the datasets. As discussed in Sec. 4, we utilize ResNet-18 (RN-18) [54] as the surrogate source model for pruning source classes. This method significantly reduces the computational cost associated with DP, making the process more efficient. Subsequently, a range of larger models, e.g., ResNet-101 (RN-101) and ViT-B/16 [112], are trained on the (pruned) ImageNet and then finetuned on downstream tasks. Baselines, training, and evaluation. By examining the performance of the existing 8 DP baselines as shown in Fig. 2, our experiments focus on two of the most effective methods: \u2460GRAND [26] and \u2461MODERATE [3], together with \u2462RANDOM (the random pruning strategy). In Fig. A1, we show more results of the rest DP baselines. Unfortunately, we are unable to include the existing data attribution method [41] as our baseline, as it does not release its pruning results, and we are unable to repeat its experiments due to the need for intensive computations. Unless specified otherwise, we focus on the supervised pretraining setting in the experiments. For self-supervised pretraining, we follow the implementation of MOCOV2 [45]. The finetuning strategies employed include LP (linear probing), which finetunes the classification head with fixed feature extractor, and FF (full finetuning), which finetunes the entire source model. For FM-based DP method, we utilize K-means clustering to group the ImageNet training data points into K = 2000 clusters for the computation of (1). In accordance with the terminology convention in model pruning [87, 113\u2013115], we refer the term \u2018winning subsets\u2019 to the obtained source subsets that do not compromise downstream performance. Among these winning subsets, we identify the one with the highest pruning ratio as the \u2018best winning subset\u2019. We then evaluate the performance of DP methods from the following aspects: the downstream performance of the model pretrained on the pruned source dataset obtained by various DP methods, and the pruning ratio of the best winning subset achieved by DP methods, accompanied by the corresponding time saved in the pretraining phase. 5.2 Experiment results LM/FM improves transfer learning accuracy by identifying \u2018winning subsets\u2019. We first showcase the significant improvements achieved by our proposed DP methods (LM and FM) compared to baselines. Our methods successfully identify winning subsets of the ImageNet, yielding transfer accuracy on par or even better than scenarios without pruning. Fig. 4 presents the downstream accuracy of transfer learning vs. different pruning ratios. Here DP is performed using the surrogate model (RN-18) on ImageNet for 8 downstream tasks. The source model (RN-101) is then trained on the pruned ImageNet and the transfer learning accuracy is assessed through LP (linear probing) and FF (full finetuning). We also present the downstream performance without pruning the source dataset (No Prune) as a reference for winning subsets. As we can see, both LM and FM significantly outperform the baselines by a substantial margin. Notably, LM and FM consistently identify winning subsets with significantly larger pruning ratios in all settings. This highlights the effectiveness of our proposed methods in achieving substantial dataset pruning without hurting downstream performance. Furthermore, we observe that the downstream performance of LM and FM initially improves and then declines as the pruning ratio increases. This is not surprising, since the initial increase in performance corresponds to the scenario where harmful source classes are removed, consistent with [41]. When the source dataset continues to shrink, the performance inevitably decreases, as a natural consequence of reducing the size of the source dataset. Moreover, in some datasets of small sizes (e.g., OxfordPets), LP achieves performance comparable to FF. This 7 \f0% 20% 40% 60% 80% Source Data Pruning Ratio 76 78 80 82 84 86 88 90 Target Test Accuracy (%) OxfordPets-LP Random Moderate GraNd LM (Ours) FM (Ours) No Prune 0% 20% 40% 60% 80% Source Data Pruning Ratio 47 48 49 50 51 52 Target Test Accuracy (%) SUN397-LP 0% 20% 40% 60% 80% Source Data Pruning Ratio 52 54 56 58 60 62 64 Target Test Accuracy (%) Food101-LP 0% 20% 40% 60% 80% Source Data Pruning Ratio 52 54 56 58 60 62 64 66 68 Target Test Accuracy (%) DTD-LP 0% 20% 40% 60% 80% Source Data Pruning Ratio 30 35 40 45 Target Test Accuracy (%) StanfordCars-LP 0% 20% 40% 60% 80% Source Data Pruning Ratio 55.0 57.5 60.0 62.5 65.0 67.5 70.0 72.5 Target Test Accuracy (%) UCF101-LP 0% 20% 40% 60% 80% Source Data Pruning Ratio 82 83 84 85 86 87 88 89 90 Target Test Accuracy (%) CIFAR10-LP 0% 20% 40% 60% 80% Source Data Pruning Ratio 89 90 91 92 93 94 95 Target Test Accuracy (%) Flowers102-LP 0% 20% 40% 60% 80% Source Data Pruning Ratio 76 78 80 82 84 86 88 90 92 Target Test Accuracy (%) OxfordPets-FF Random Moderate GraNd LM (Ours) FM (Ours) No Prune 0% 20% 40% 60% 80% Source Data Pruning Ratio 42 44 46 48 50 52 54 56 Target Test Accuracy (%) SUN397-FF 0% 20% 40% 60% 80% Source Data Pruning Ratio 78 79 80 81 Target Test Accuracy (%) Food101-FF 0% 20% 40% 60% 80% Source Data Pruning Ratio 56 58 60 62 64 66 68 Target Test Accuracy (%) DTD-FF 0% 20% 40% 60% 80% Source Data Pruning Ratio 85 86 87 88 89 90 Target Test Accuracy (%) StanfordCars-FF 0% 20% 40% 60% 80% Source Data Pruning Ratio 58 60 62 64 66 68 70 72 74 Target Test Accuracy (%) UCF101-FF 0% 20% 40% 60% 80% Source Data Pruning Ratio 95.0 95.5 96.0 96.5 97.0 Target Test Accuracy (%) CIFAR10-FF 0% 20% 40% 60% 80% Source Data Pruning Ratio 96 97 98 99 Target Test Accuracy (%) Flowers102-FF Figure 4: Source dataset pruning trajectory given by downstream testing accuracy (%) vs. source dataset pruning ratio (%) in the supervised pretraining setting. Here the source model RN-101 is trained on each full/pruned source dataset (ImageNet) and finetuned on different downstream tasks through LP and FF. The downstream performance without pruning (No Prune) is marked with the black dashed line. Results are averaged over three independent trials (see exact numbers and variances in Tab. A2). is attributed to the fact that the performance gap between LP and FF using the large-scale RN-101 model on small-scale datasets tends to diminish. Table 1: The pruning ratio of the best winning subsets obtained by each DP method for different downstream tasks. The other experiment setup is consistent with Fig. 4. The achieved largest pruning ratio is highlighted in bold in each column. And N/A indicates the scenario where no winning subsets are found. Dataset OxfordPets SUN397 Food101 DTD StanfordCars UCF101 CIFAR10 Flowers102 Finetune method LP FF LP FF LP FF LP FF LP FF LP FF LP FF LP FF RANDOM 20% 20% 15% N/A N/A 20% 10% 20% N/A 10% 20% 10% N/A 15% 40% 35% MODERATE 15% 20% 15% 20% N/A 20% 15% 30% N/A 5% 5% 10% N/A 20% 40% 35% GRAND 25% 20% 10% 20% 5% N/A 15% 20% N/A 5% 5% N/A 40% 15% 45% 40% LM (ours) 80% 80% 65% 70% 65% 70% 55% 50% 65% 75% 45% 40% 40% 40% 55% 60% FM (ours) 85% 75% 65% 60% 55% 75% 55% 50% 60% 70% 45% 40% 45% 45% 55% 55% Tab. 1 provides a summary of the pruning ratios achieved by the best winning subsets identified using different DP methods for all 8 downstream datasets. Both LM and FM methods successfully remove more than 45% of the source classes without downstream performance drop. In contrast, all the baselines experience significant performance degradation when the pruning ratio exceeds 40%. Table 2: The downstream performance with different source data pruning ratios in the SSL pretraining setting. A randomly initialized RN-101 is self-supervised pretrained using MOCO V2 on each full/pruned source dataset and finetuned on the downstream task through LP. The best result in each pruning ratio is marked in bold and the performance surpassing the unpruned setting (pruning ratio 0%) is highlighted in cyan . Dataset Pruning Ratio OxfordPets SUN397 Flowers102 0% 50% 60% 70% 80% 0% 50% 60% 70% 80% 0% 50% 60% 70% 80% RANDOM 69.26 62.32 61.27 59.09 53.75 47.36 45.63 45.08 43.54 39.81 85.17 82.23 82.60 81.03 80.02 MODERATE 63.37 62.45 63.31 57.42 45.73 45.14 44.23 40.82 82.45 81.45 81.69 81.32 GRAND 64.42 63.34 61.14 56.42 45.72 45.58 45.24 41.72 82.85 82.44 82.14 81.73 FM (ours) 69.92 69.99 70.29 70.21 48.46 48.58 47.90 46.00 85.22 85.42 84.37 84.61 8 \fTab. 2 highlights the effectiveness of FM-based DP in the self-supervised pretraining setup for three representative downstream tasks. As we can see, the transfer learning accuracy achieved by using FM consistently outperforms baselines in the self-supervised pretraining paradigm. FM can identify winning subsets for transfer learning even in the challenging regime of large pruning ratios, ranging from 50% to 80%. For instance, in the case of SUN397, FM-based winning subsets achieves a pruning ratio up to 70%, and for Flowers102 the maximum pruning ratio is 60%. These pruning merits align with the findings for FM in the supervised pretraining setup, as illustrated in Tab. 1. Table 3: Time consumption of LM/FM in Fig.4 to obtain the pretrained model. The reported time consumption covers surrogate model (RN-18) training, LM/FM dataset pruning, and source model pretraining (RN-101). Pruning Ratio 0% 20% 40% 60% 80% Time Consumption (h) 5.4 4.6 (15%\u2193) 3.5 (35%\u2193) 2.4 (56%\u2193) 1.3 (76%\u2193) DP enhances the efficiency of source pretraining. Tab. 3 displays the computation time required to obtain the pretrained source model using LM at different pruning ratios. The reported time consumption includes the entire pipeline, encompassing surrogate model training (RN18), DP process, and source model training (RN101) on the pruned ImageNet dataset. The runtime cost of the conventional transfer learning on the full ImageNet dataset for RN-101 is also listed as a reference. As we can see, DP enjoys high efficiency merit of source training. Taking the 5.4 hours required for source training on the full ImageNet dataset as a reference, LM-enabled 20% pruned ImageNet achieves a 15% reduction in training time. Moreover, the efficiency advantage increases to 76% when the pruning ratio reaches 80% and these computational benefits do not sacrifice transfer learning accuracy at all. Table 4: Time consumption comparison with a pruning ratio of 60% of different dataset pruning methods. Other settings follow Fig. 4. Method MODERATE GRAND BRUTE-FORCE [41] ours Time Consumption (h) 7.6 18.4 > 500 2.4 Next, we compare the efficiency of our methods with the state-of-the-art dataset pruning baselines in Tab. 4, including GRAND, MODERATE, and the brute-force method proposed in [41]. We showcase the efficiency comparison with a pruning ratio of 60%. As we can see, our method (even using a surrogate model) achieves a substantial computation efficiency improvement over other methods. In particular, [41] involves training thousands of source models. Thus, its computational cost is typically unaffortable in experiments. Table 5: Downstream performance of models pretrained on full/pruned source dataset (DENSE/LM) w/wo adversarial pretraining (Adv). For DENSEADV and LM-ADV, 3-step adversarial training [116] is used for pretraining. Pruning ratio, downstream test accuracy (Acc.), and time consumptions for obtaining pretrained models are reported. Method SUN397 DTD Pruning Acc.(%) Time Pruning Acc.(%) Time Ratio LP FF (h) Ratio LP FF (h) DENSE N/A 51.45 54.21 5.4 N/A 65.91 67.21 5.4 DENSE-ADV N/A 52.97 55.67 13.7 N/A 67.23 68.92 13.7 LM 70 50.95 54.28 1.9 50 66.25 67.22 2.9 LM-ADV 70 52.07 55.49 4.2 50 67.02 68.54 6.7 DP enables efficient adversarial pretraining. Tab. 5 showcases the improvement in transfer learning accuracy and efficiency achieved by our proposed LP-based DP method when incorporating adversarial training (AT) [116] on either the full or pruned ImageNet dataset. This transfer learning setup is motivated by the findings in [58], showing that enhancing the robustness of the source model against adversarial attacks through AT can improve transfer learning accuracy in both LP and FF-based finetuning scenarios. We then report downstream accuracies using LP and FF on two specific downstream datasets: SUN397 and DTD, which are intentionally chosen due to the large room for improvement in transfer learning accuracy, as shown in Fig. 4. We also determine the pruning ratios for LM by selecting those that led to the best winning subsets. Our experimental results demonstrate that employing AT on both the unpruned and pruned source datasets can improve transfer learning accuracy. Specifically, we refer to AT on the original unpruned ImageNet dataset as DENSE-AT, and AT on the LM-pruned ImageNet dataset as LM-AT. One notable advantage of integrating LM into AT is the significant improvement in computation efficiency. A key highlight of this approach is that LM-AT achieves a similar computation time as the standard source training on ImageNet (DENSE), while exhibiting almost no accuracy drop compared to DENSE-AT. This observation demonstrates the potential to accelerate AT through DP. Robustness against the surrogate model size. To explore the sensitivity of our proposed method to the surrogate model\u2019s size and accuracy, we show the transfer learning performance on the task OxfordPets against different surrogate model sizes in Fig. 5. It is evident that even though the performance of the surrogate model on the source dataset (ImageNet) decreases, the downstream performance of RN-101 pretrained on the LM-based pruned source subsets remains relatively stable. Further, Tab. A6 compares the class indices selected by the corresponding surrogate models. Interestingly, the most relevant source class selections exhibit a high level of agreement across surrogate models of differing sizes. 9 \f0% 20% 40% 60% 80% Source Data Pruning Ratio 85 86 87 88 89 90 91 Target Test Accuracy (%) ResNet-20s ResNet-32s ResNet-44s ResNet-56s ResNet-18 No Prune 0% 20% 40% 60% 80% Source Data Pruning Ratio 85 86 87 88 89 90 91 Target Test Accuracy (%) VGG-2 VGG-4 VGG-8 ResNet-18 No Prune (a) ResNets family (b) VGG family Figure 5: Source dataset pruning trajectory given the downstream task OxfordPets using different surrogate models. Other settings follow Fig. 4. LM shows robust behavior with different surrogate models, even if they yield different performance on the source dataset due to their different model capacities. For example, RN-32s only achieves a test accuracy of 40.77% on the source dataset, but it manages to achieve the same maximum pruning ratio of the winning subsets as RN-18. This underscores that LM can accommodate a small surrogate model of 1% the size of the pretrained model (Rn-32s vs. RN-101). Moreover, it reveals that the surrogate model does not need to attain exceptional performance in the source domain for dataset pruning in transfer learning. Table 6: Downstream performance of different DP methods using ViT-B/16 as the source model and FF as the finetuning method. Other settings are consistent with Fig. 4. The best performance at each pruning ratio is marked in bold, and performance exceeding the unpruned setting (0% pruning ratio) is highlighted in cyan . Dataset Pruning Ratio OxfordPets SUN397 Flowers102 0% 20% 40% 60% 80% 0% 20% 40% 60% 80% 0% 20% 40% 60% 80% RANDOM 85.75 85.63 83.72 75.94 69.83 43.97 43.74 41.18 36.72 29.41 90.95 90.25 89.38 87.47 84.44 MODERATE 85.39 83.97 82.39 78.32 43.97 41.82 38.89 31.15 90.86 89.77 88.34 86.69 GRAND 85.77 83.88 82.35 79.12 43.32 40.93 38.52 31.00 90.75 89.89 88.54 85.39 LM 86.15 85.72 84.77 80.64 44.24 42.79 39.59 32.35 91.19 89.97 89.07 87.23 DP for ViT. In Tab. 6, we present the consistent transfer learning improvement achieved by LM for Vision Transformer (ViT) pretraining. Specifically, we utilize ViT-B/16 as the source model, RN-18 as the surrogate model for DP, and use FF on the source model for the downstream tasks studied in Tab. 2. As we can see, LM yields a better transfer learning accuracy than baselines at different pruning ratios, consistent with the previous results on RN-101. This observation highlights the model-agnostic advantage of our proposed method in transfer learning. We also observe the pruning ratio of the best winning subset achieved for ViT is smaller compared to ResNet. This discrepancy is possibly due to the higher data requirements of ViT during pretraining [112]. Additional results. We conduct ablation studies on the cluster number for FM in Fig. A2 and the results of selecting data proposed by LM in reverse order in Fig. A3. We shows FM can be applied to more advanced SSL framework in Tab. A3. We also investigate the performance of our proposed method in the multi-task setting in Fig. A5, and the scenarios with data biases in Tab. A4. We further examine our method on the few-shot benchmark in Tab. A5. We also present feature distribution analysis in Fig. A6 and image examples for the top classes selected by LM/FM in Fig. A7. Limitations. Despite the remarkable empirical results above, we admit LM/FM mainly caters to exploring the source data influence in a specific downstream task, and the largest pruning ratio is reduced as more downstream tasks are considered (see Fig. A5), which we consider as a limitation. Meanwhile, we acknowledge that the promising performance achieved by our method still lacks rigorous theoretical analysis, despite the feature distribution and model flatness analysis, which we believe will be a challenging but exciting future work on DP for transfer learning. 6"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.10110v1",
+                    "title": "Robust Mixture-of-Expert Training for Convolutional Neural Networks",
+                    "abstract": "Sparsely-gated Mixture of Expert (MoE), an emerging deep model architecture,\nhas demonstrated a great promise to enable high-accuracy and ultra-efficient\nmodel inference. Despite the growing popularity of MoE, little work\ninvestigated its potential to advance convolutional neural networks (CNNs),\nespecially in the plane of adversarial robustness. Since the lack of robustness\nhas become one of the main hurdles for CNNs, in this paper we ask: How to\nadversarially robustify a CNN-based MoE model? Can we robustly train it like an\nordinary CNN model? Our pilot study shows that the conventional adversarial\ntraining (AT) mechanism (developed for vanilla CNNs) no longer remains\neffective to robustify an MoE-CNN. To better understand this phenomenon, we\ndissect the robustness of an MoE-CNN into two dimensions: Robustness of routers\n(i.e., gating functions to select data-specific experts) and robustness of\nexperts (i.e., the router-guided pathways defined by the subnetworks of the\nbackbone CNN). Our analyses show that routers and experts are hard to adapt to\neach other in the vanilla AT. Thus, we propose a new router-expert alternating\nAdversarial training framework for MoE, termed AdvMoE. The effectiveness of our\nproposal is justified across 4 commonly-used CNN model architectures over 4\nbenchmark datasets. We find that AdvMoE achieves 1% ~ 4% adversarial robustness\nimprovement over the original dense CNN, and enjoys the efficiency merit of\nsparsity-gated MoE, leading to more than 50% inference cost reduction. Codes\nare available at https://github.com/OPTML-Group/Robust-MoE-CNN.",
+                    "authors": "Yihua Zhang, Ruisi Cai, Tianlong Chen, Guanhua Zhang, Huan Zhang, Pin-Yu Chen, Shiyu Chang, Zhangyang Wang, Sijia Liu",
+                    "published": "2023-08-19",
+                    "updated": "2023-08-19",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "main_content": "Introduction Despite the state-of-the-art performance achieved by the outrageously large networks [1\u20135] in various deep learning Correspondence to: Yihua Zhang<zhan1908@msu.edu> (DL) tasks, it still remains challenging to train and deploy such models cheaply. A major bottleneck is the lack of parameter efficiency [6]: A single data prediction only requires activating a small portion of the parameters of the full model. Towards efficient DL, sparse Mixture of Experts (MoE) [7\u201315] aims to divide and conquer the model parameters based on their optimal responses to specific inputs so that inference costs can be reduced. A typical MoE structure is comprised of a set of \u2018experts\u2019 (i.e., sub-models extracted from the original backbone network) and \u2018routers\u2019 (i.e., additional small-scale gating networks to determine expert selection schemes across layers). During inference, sparse MoE only activates the most relevant experts and forms the expert-guided pathway for a given input data. By doing so, sparse MoE can boost the inference efficiency (see \u2018GFLOPS\u2019 measurement in Fig. 1). Architecture-wise, sparse MoE has been used for both CNNs [8, 16] and vision transformers (ViTs) [7, 9\u201315, 17]. Yet, we will focus on the former since sparse MoE for CNNs is under-explored compared to non-sparse MoE for CNNs [18\u201320], and adversarial robustness (another key performance metric of our work) was extensively studied in the context of CNNs. It is known that a main weakness of DL is the lack of adversarial robustness [21\u201323]. For example, CNNs can be easily fooled by adversarial attacks [21\u201323], in terms of tiny input perturbations generated to direct to erroneous predictions. Thus, adversarial training (AT) of CNNs has become a main research thrust [24\u201329]. However, when CNN meets sparse MoE, it remains elusive if the improved inference efficiency brought by the sparse MoE comes at the cost of more complex adversarial training recipes. Thus, we ask: (Q) What will be the new insights into adversarial robustness of sparse MoE-integrated CNNs? And what will be the suited AT mechanism? To our best knowledge, problem (Q) remains open in the literature. The most relevant work to ours is [30], which investigated the adversarial robustness of MoE and leverarXiv:2308.10110v1  [cs.CV]  19 Aug 2023 \f95.2 93.7 94.3 Dense S-Dense MoE-CNN ResNet18 Standard Training Robust Training Standard Accuracy 96.2 94.7 95.3 Dense S-Dense MoE-CNN WRN-28-10 93.6 92.2 92.9 Dense S-Dense MoE-CNN VGG-16 93.1 90.7 92.3 Dense S-Dense MoE-CNN DenseNet Prior Art (AT) Ours (AdvMoE) 50.1 48.0 44.9 51.8 \ud83c\udfc6 51.8 49.1 46.2 55.7 \ud83c\udfc6 46.2 43.7 41.4 49.8 \ud83c\udfc6 44.5 38.1 31.4 39.7 \ud83c\udfc6 Models \ud83d\udc4e \ud83d\udc4e \ud83d\udc4e \ud83d\udc4e Robust Accuracy Metrics (a) Illustration of CNN types considered in this work. (b) Performance overview on CIFAR-10. Figure 1. (a) Model types (Dense, MoE-CNN, Sparse-CNN, and S(mall)-Dense) considered in this paper; see details in \u2018Model setup\u2019 of Sec. 3. (b) Performance overview using the standard training and the robust training on model architectures in (a), where standard accuracy and robust accuracy are defined by testing accuracy on the benign and adversarial test datasets, respectively. Compared to standard training (results in gray), the conventional AT [25] is no longer effective for MoE-CNN (see results in light purple). This is in contrast to AT for other CNN models (Dense and S-Dense). Different from AT, our proposed ADVMOE can effectively equip MoE-CNN with improved robustness, higher than Dense (see results in orange), without losing its inference efficiency (see \u201cGFLOPS\u201d). We refer readers to Sec. 5.1 for more experiment details. aged the ordinary AT recipe [24] to defend against adversarial attacks. However, it only focused on the ViT architecture, making a vacancy for the research on robustification for the sparse MoE-based CNN (termed MoE-CNN in this work). Most importantly, we find that the vanilla AT [24, 25] (widely used to robustify CNNs) is no longer effective for MoE-CNN. Thus, new solutions are in demand. To address (Q), we need to (1) make careful sanity checks for AT in MoE-CNN, (2) make an in-depth analysis of its failure cases, and (3) advance new AT principles that can effectively improve robustness without losing the generalization and efficiency from sparse MoE. Specifically, our contributions are unfolded below. \u2022 We dissect the MoE robustness into two new dimensions (different from CNNs): routers\u2019 robustness and experts\u2019 robustness. Such a robustness dissection brings novel insights into the (in)effectiveness of AT. \u2022 Taking inspiration from the above robustness dissection, we propose a new Adversarial training framework for MoE, termed ADVMOE, which enforces routers and experts to make a concerted effort to improve the overall robustness of MoE-CNN. \u2022 We conduct extensive experiments to demonstrate the effectiveness of ADVMOE across 4 CNN architectures and 4 datasets. For example, ADVMOE outperforms AT on the original dense CNN model (termed Dense) by a substantial margin: 1% \u223c4% adversarial robustness improvement and over 50% reduction of inference overhead; see Fig. 1 for illustrations on different CNN types and highlighted performance achieved. 2. Related Work Sparsely-activated Mixture of Experts (Sparse MoE). As a special instance of compostional neural architectures [31\u201333], MoE [4, 7\u201311, 16, 18\u201320, 34\u201341] aims at solving ML tasks in a divide-and-conquer fashion, which creates a series of sub-models (known as the experts) and conducts input-dependent predictions by combing the output of submodels. As an important branch of MoE, sparsely gated MoE [4, 7\u201316, 39\u201343] only activates a subset of experts based on a routing system. The major advantage brought by sparse MoEs lies in its sub-linear increasing inference costs (FLOPs) with respect to (w.r.t.) model scales (parameter counts) [7]. In the vision domain, a vast majority of the existing works focus on the design of MoE for ViTs [9\u201315, 40, 42, 43], leaving MoE for CNNs under-explored [8, 16]. To our best knowledge, DeepMoE [8] is the most recent work that systematically studies the integration of MoE with CNNs, but restricts to the standard (non-robust) training paradigm. Meanwhile, there also exist other works related to MoE-CNN, but they either fall out of the \u201csparse\u201d MoE scope [18, 20] or bring no efficiency gains [19]. By contrast, we focus on the efficiency-promoting MoE-CNN setup throughout this work. Adversarial robustness. CNNs are notoriously vulnerable to imperceptible adversarial samples [22, 23, 44] and thus training adversarially robust models [21, 24, 45] has become a main research focus in many areas. Most of the robust training methods [24\u201329] are extended from minmax optimization-based adversarial training [46]. For instance, the work [25] seeks an optimal balance between 2 \frobustness and standard generalization ability. Other work [26, 28, 47\u201354] aims at trimming down the computational costs of robust training while maintaining robustness. The work [30] studies the robustness of MoE-based architectures for the first time. Yet, its focus stays on MoE for ViTs and the relationship between model capacity and robustness. 3. Problem Statement In this section, we start by presenting the setup of MoECNN in this work and then introduce the robust learning paradigm. The lack of adversarial robustness of deep models inspires us to investigate whether the adversarial training (AT) approach designed for vanilla CNNs keeps effective for MoE-CNN. Through a motivating example, we show that the conventional AT recipe is incapable of equipping MoE-CNN with desired robustness. The resulting performance is even worse than that of AT-produced S-Dense, which has a much smaller model capacity than MoE-CNN. Thus, the question of how to robustify MoE-CNN arises. Model setup. We consider a CNN-backboned MoE that consists of multiple MoE layers. Each MoE layer involves a router and a vanilla convolutional layer from the backbone CNN model. Within one MoE layer, we define N experts, each of which picks a subset of the channels from the convolutional layer. Specifically, suppose the l-th layer contains Cl channels, one expert will contain r \u00d7 Cl channels, where we call the ratio r \u2208[0, 1] model scale and keep it the same across different layers (see Fig. 1a). It is worth noting that as r increases, the per-expert model capacity increases (i.e., with more parameters) at the cost of the efficiency reduction. In a forward path, the router first makes an inputspecific expert selection. These selected layer-wise experts then form an end-to-end pathway to process this input. We use \u201cpathway\u201d to describe one experts-guided forward path (see Fig. 1a). We summarize the model setup in Fig. A1. Further, we introduce different model types considered in this work and shown in Fig. 1a. First, we term the original dense CNN model \u2018Dense\u2019, which serves as the model basis for other model types that derive from. Second, we directly shrink the channel number of each layer in Dense (based on the model scale parameter r) to obtain the \u2018small dense\u2019 model (termed \u2018S-Dense\u2019). Notably, S-Dense has the size equivalent to a single pathway in MoE-CNN. Third, we use the structured pruning method [50] to create a sparse subnetwork from Dense, with the weight remaining ratio same as the model scale parameter r in MoE-CNN, which we call \u2018Sparse-CNN\u2019. In summary, S-Dense has the smallest model capacity (comparable to a single pathway of MoE-CNN), and should provide the performance lowerbound for MoE-CNN. By contrast, Sparse-CNN has a larger model capacity but is smaller than MoE-CNN as it encodes a data-agnostic pathway of Dense, while MoE-CNN yields data-specific pathways at the same scale. Dense has the largest model capacity but the least inference efficiency. Adversarial robustness: From CNN to MoE-CNN. It has been known that current machine learning models (e.g., CNNs) are vulnerable to adversarial attacks [21\u201323]. Towards the robust design, a variety of AT (adversarial training) methods have been developed. The predominant ones include the min-max optimization-based vanilla AT [24] and its TRADES variant [25] that strikes a balance between generalization and adversarial robustness. Throughout the paper, we adopt TRADES as the default conventional AT recipe, which solves the following problem: min \u03b8 E(x,y)\u2208D \u0014 \u2113(\u03b8; x, y) + 1 \u03bb max \u2225\u03b4\u2225\u221e\u2264\u03f5 \u2113KL(f\u03b8(x), f\u03b8(x + \u03b4)) \u0015 (AT) where \u03b8 denotes model parameters to be robustified, (x, y) \u2208D is a training sample, drawn from the training set D, with input feature x and label y, \u2113(\u03b8, x; y) denotes the cross-entropy loss using model \u03b8 at data point (x, y), \u03b4 signifies the input perturbation variable subject to the \u2113\u221e-norm ball of radius \u03f5, f\u03b8(\u00b7) denotes the model\u2019s predictions, \u2113KL is the KL divergence loss that characterizes the worst-case prediction stability at the presence of \u03b4, and \u03bb > 0 is a regularization parameter to strike the tradeoff between empirical risk minimization and the robustness of model predictions. Although AT has been well studied for adversarial robustness of CNNs, there exists few attempts to robustify MoE-CNN. This raises the problem of our interest: (Problem statement) Can MoE-CNN be robustified as effectively as an ordinary CNN using AT? If not, how to robustly train MoE-CNN to achieve robustness not worse than AT-oriented S-Dense, SparseCNN, and Dense while preserving MoE\u2019s efficiency? Warm-up study: AT for MoE-CNN is not trivial. Our goal to robustify MoE-CNN includes (1) achieving high robustness, (2) maintaining high prediction accuracy, and (3) making full use of MoE routing to keep the model\u2019s high efficiency and expressiveness. Nonetheless, the routing system in MoE brings extra robustification challenges, which never exist in ordinary CNNs. Specifically, the inputspecific expert selection in MoE could make the attacker easier to succeed, since input perturbations can either mislead routers to select incorrect experts or fool the pathwaydesignated predictor. Such a \u2018two-way attack mode\u2019 makes AT for MoE-CNN highly non-trivial. Fig. 2 empirically justifies that the direct application of (AT) to MoE-CNN is problematic. In Fig. 2, we consider ResNet-18 as the model backbone (Dense) and CIFAR-10 for image classification. We apply (AT) to train MoE-CNN 3 \fand S-Dense, and report the robust accuracy (RA), i.e., testtime accuracy over adversarial examples generated by 50step PGD attacks [24], against different attack strengths \u03f5. 2 4 6 8 Attack Strength (/255) 20 30 40 50 60 70 Robust Accuracy (%) Model Type S-Dense MoE-CNN Figure 2. Performance of MoECNN and S-Dense robustly trained using (AT) on CIFAR-10 with ResNet-18 as the backbone. As we can see, although MoE-CNN has a much larger model capacity than S-Dense, it leads to a significant RA drop when the conventional AT approach is applied. This implies that the design of AT for MoECNN is far from trivial. A new robust learning protocol is thus needed to improve the robustness of MoE-CNN without losing its merits in efficiency and generalization. 4. Methods In this section, we start by peering into the failure case of (AT) in MoE-CNN by understanding the roles of the routers and pathways in (AT). We empirically show that these individual components are hard to adapt to each other and cannot make a concerted effort in AT. Based on that, we develop a new AT framework for MoE-CNN, ADVMOE, which also takes inspiration from bi-level optimization. Dissecting robustness of MoE-CNN: Routers\u2019 robustness vs. pathways\u2019 robustness. The main puzzle in robustifying MoE-CNN comes from the coupling between the robustness of routers (which are responsible for expert selection across layers) and the robustness of the inputspecific MoE pathways (which are in charge of the final prediction of an input). Given the failure case of AT for MoECNN in Fig. 2, we need to understand the roles of routers and pathways in AT, i.e., how the adversarial robustness of MoE-CNN is gained in the presence of the \u2018two-way attack mode\u2019. To this end, we begin by assessing the influence of the routers\u2019 robustness on the overall robustness. This is also inspired by the recent pruning literature [50] showing that model robustness can be gained solely from network\u2019s sparse topology (regardless of model weights). We thus ask: (Q1) Is improving routers\u2019 robustness sufficient to achieve a robust MoE-CNN? To tackle (Q1), we first split the parameters of MoECNN (i.e., \u03b8) into two parts, the parameters of routers \u03d5 and the parameters of the backbone network \u03c8. This yields \u03b8 = [\u03d5\u22a4, \u03c8\u22a4]\u22a4, where \u22a4is the transpose operation. We then call (AT) to robustly train routers (\u03d5) but fix the backbone network (\u03c8) at its standard pre-trained weights. We denote this partially-robustified model by \u00af \u03b8 = [ \u00af \u03d5\u22a4, \u03c8\u22a4]\u22a4, where \u00af indicates the updated parameters. To answer (Q1), we assess the robustness gain of \u00af \u03b8 vs. 3 baselines (M1-M3): (M1) the standard MoE-CNN \u03b8, (M2) AT-robustified SDense, and (M3) Sparse-CNN achieved by the robust sparse mask learning method [50] over the original Dense model. 0.8 0.5 0.2 Model Scale 0 10 20 30 40 50 Robust Accuracy (%) M1 M2 M3 Figure 3. Robustness comparison of router-robusified MoECNN (i.e. \u00af \u03b8) and baseline models (M1 \u2013 M3) for different model scales under CIFAR-10 given the backbone network ResNet-18. Fig. 3 shows the robust accuracy of the routerrobustified MoE-CNN \u00af \u03b8 and its performance comparison with other baseline models. As we can see, the robustified router improves the overall robustness (e.g., 37.64% for \u00af \u03b8 with model scale 0.5) compared to the undefended MoECNN (M1: 0%) and the robustified mask (M3: 20.04%). However, there is still a huge robustness gap compared to the (AT)-robustified S-Dense (M2: 47.68%). Based on the results above, we acquire the first insight into (Q1): Insight 1: Robustifying routers improves the overall robustness of MoE-CNN but is not as effective as ATresulted S-Dense. Based on Insight 1, we further peer into the resilience of expert selection decisions to adversarial examples. If expert selections in all MoE layers keep intact in the presence of an adversarial perturbation, we say that the routing system of MoE-CNN is robust against this adversarial example. We then divide adversarial examples into four categories according to whether they successfully attacked routers and the router-oriented pathways: \u2776unsuccessful attack on both routers and MoE pathways, \u2777successful attack on routers but not MoE pathways, \u2778successful attack on MoE pathways but not routers, and \u2779successful attack on both routers and MoE pathways. Here \u2776+ \u2778characterizes the robustness of routers, while \u2776+ \u2777represents that of MoE. Thus, if \u2777or \u2778takes a large portion of generated adversarial examples, it implies that the routers\u2019 robustness does not directly impact the MoE pathway-based predictor\u2019s robustness. Fig. 4 shows the above categories \u2776-\u2779when attacking the router-robustified MoE-CNN (i.e., \u00af \u03b8). As we can see, routers\u2019 robustness indeed improves prediction robustness (as shown by 31.74% unsuccessful attacks against the MoE predictor in \u2776). However, in the total number of unsuccessful attacks against routers (i.e., \u2776+\u2778= 76.27%), more than half of them successfully fool 4 \fFigure 4. Adversarial attack success analysis on dissected MoECNN models \u00af \u03b8 = [ \u00af \u03d5\u22a4, \u03c8\u22a4] (model scale r = 0.5), where only \u00af \u03d5 is (AT)-robustified. The adversarial evaluation is based on 50step PGD attack [24] to fool \u00af \u03b8, and other experiment setups align with Fig. 3. The evaluation is carried out on the test set with a total number of 10000 samples. the MoE predictor (i.e., \u2778>\u2776). The above results provide us an additional insight: Insight 2: Improving routers\u2019 robustness is not sufficient for the MoE predictor to gain satisfactory robustness although the former makes a positive impact. Both Insight 1 and Insight 2 point out that only improving routers\u2019 robustness is not adequate to obtain the desired robustness for the overall MoE-CNN. Thus, we next ask: (Q2) Given the router-robustified model \u00af \u03b8, can we equip \u00af \u03b8 with additional robustness by robustly training expert weights (\u03c8)? And how does it further impact routers? Figure 5. Adversarial attack success analysis on routers \u00af \u03d5 and MoE-CNN models \u00af \u00af \u03b8 = [ \u00af \u03d5\u22a4, \u00af \u03c8\u22a4]. Other setups remain the same as Fig. 4. To answer (Q2), we call (AT) to further robustly train the backbone network \u03c8 on top of the router-robustified model \u00af \u03b8. We denote the resulting model by \u00af \u00af \u03b8 = [ \u00af \u03d5\u22a4, \u00af \u03c8\u22a4]. Fig. 5 shows the dissection of the robustness of \u00af \u00af \u03b8 in the same setup of Fig. 4. Obviously, the overall prediction robustness (\u2776+\u2777) is further enhanced after updating \u00af \u03b8 to \u00af \u00af \u03b8. Thus, the gains in the robustness of experts\u2019 weights indeed further help improve the overall robustness. However, this leads to a surprising drop in the router\u2019s robustness (\u2776+\u2778) when comparing \u00af \u00af \u03b8 with \u00af \u03b8. This shows that routers\u2019 robustness is not automatically preserved if experts are updated. We obtain the following insight into (Q2): Insight 3: Robustifying routers and MoE weights can yield complementary benefits but the inadaptability of routers\u2019 robustness to MoE\u2019s robustness prevents AT from achieving significant robustness improvement. ADVMOE: Router-expert alternating AT through a bilevel optimization viewpoint. As illuminated by insights above, we provide a reason for the ineffectiveness of AT in robustifying MoE-CNN. Insights 1-2 show that the robustness of routers (\u03d5) and the robustness of MoE-based predictor (\u03c8) are intertwined and their interrelation is nontrivial. As a result, the single-level (non-convex) robust optimization over the entire model parameters (\u03d5, \u03c8) experiences difficulty in co-optimizing routers and MoE prediction pathways to achieve the best complementary robustness gains, as supported by Insight 3. A key missing optimization factor in AT for MoE-CNN is its incapability of modeling and optimizing the coupling between the robustness of the routers and that of the MoE pathways. Without such an optimization design, it is difficult for AT to robustify routers and MoE pathways in a cooperative and adaptive mode. Spurred by the above, we develop a new AT framework through bi-level optimization (BLO). In general, BLO provides a hierarchical learning framework with two levels of optimization tasks, where the objective and variables of an upper-level problem depend on the optimizer of the lower level. BLO then enables us to explicitly model the coupling between AT for routers and AT for MoE network. Specifically, we modify the conventional (AT) to minimize \u03c8 \u2113TRADES(\u03c8, \u03d5\u2217(\u03c8); D) subject to \u03d5\u2217(\u03c8) = arg min\u03d5 \u2113TRADES(\u03c8, \u03d5; D), (1) where the model parameters of MoE-CNN \u03b8 are split into the lower-level optimization variables \u03d5 for routers and the upper-level optimization variables \u03c8 for the MoE backbone network, and \u2113TRADES(\u03c8, \u03d5; D) denotes the TRADEStype training loss defined in (AT) by replacing \u03b8 with (\u03d5, \u03c8). Compared to (AT), our proposal (1) has the following differences. First, robustifying MoE network (\u03c8) is now explicitly coupled with routers\u2019 optimization through the lower-level solution \u03d5\u2217(\u03c8). Second, our proposal addresses the robustness adaptation problem pointed out in Insight 3 since the lower-level optimization of (1) enables fast adaptation of \u03d5 to the current MoE network \u03c8 like meta-learning [55]. Third, since \u2113TRADES is involved at both optimization levels of (1), the embedded attack generation problem (i.e., maximization over \u03b4 in AT) needs to be solved at each level but corresponding to different victim models, i.e., (\u03c8, \u03d5) and (\u03c8, \u03d5\u2217(\u03c8)), respectively. To solve problem (1), we adopt a standard alternating optimization (AO) method [56]. Compared with other kinds of BLO algorithms [57], AO is the most computationally efficient. Our extensive experiments in Sec. 5 will show that AO is effective to boost the adversarial robustness of MoECNN and achieve improvements over baseline methods and models by a substantial margin. The key idea of AO is to alternatively optimize the lower-level and the upper-level problem, during which variables defined in another level 5 \fAlgorithm 1 The ADVMOE algorithm 1: Initialize: backbone network \u03c8, routers \u03d5, batch size b, attack generation step K. 2: for Iteration t = 0, 1, . . . , do 3: Pick different random data batches B\u03c8 and B\u03d5 for backbone and router training 4: Lower-level \u03d5-update (with fixed \u03c8): Given \u03c8, update \u03d5 by minimizing \u2113TRADES using K-step PGD attack [24] generator and SGD (with B\u03d5) 5: Upper-level \u03c8-update (with fixed \u03d5): Given \u03d5, update \u03c8 by minimizing \u2113TRADES using K-step PGD attack gen-erator and SGD (with B\u03c8) 6: end for are fixed. We term the resulting algorithmic framework as Adversarially robust learning for MoE-CNN (ADVMOE); see Algorithm 1 for a summary. We highlight that ADVMOE will train robust routers and robust MoE pathways to \u2018accommodate\u2019 each other. In contrast to the conventional AT framework, ADVMOE delivers the coupled \u03d5\u2217(\u03c8) and \u03c8, where both parts make a concerted effort to improve the overall robustness. We also remark that ADVMOE does not introduce additional hyper-parameters, since in practice we found routers and experts can share the same learning rate and schedules. More implementation details are provided in Appendix B. In the meantime, we remark that since our proposal is a BLO with non-convex lower and upper-level objectives (1). It is difficult to prove the convergence of ADVMOE. Existing theoretical analysis of BLO typically relies on strongly convex assumptions of lower-level problems [58, 59]. Although without a proper theoretical analysis framework, our method converges well in practice (see Appendix C). 5. Experiments In this section, we will demonstrate the effectiveness of our proposed ADVMOE approach on diverse datasets and models. We will also make an in-depth analysis of the router utility and the expert selection distribution for ADVMOEtrained MoE-CNN. 5.1. Experiment Setup Model and dataset setups. To implement MoE-CNN and other baselines, we conduct experiments on ResNet18 [60], Wide-ResNet-28-10 [61], VGG-16 [62], and DenseNet [63]. Towards fair assessment, our performance comparison between different model types is restricted to using the same model scale parameter r (see Fig. 1 for an example). By doing so, an input example will leverage the same amount of model parameters for decision-making. For MoE-CNN, we consider N = 2 experts with r = 0.5 by default, see Appendix B for more details. Dataset-wise, we focus on the commonly used ones to evaluate the adversarial robustness of image classification [24, 25, 64], including CIFAR-10 [65], CIFAR-100 [65], TinyImageNet [66], and ImageNet [66]. Baselines. To make our performance comparison informative and comprehensive, we consider three kinds of baselines that are fairly comparable to (ADVMOE). \u2460AT (SDense): we apply AT to S-Dense; \u2461AT (Sparse): we apply the robustness-aware (structured) sparse mask learning method [50] to obtain Sparse-CNN; \u2462AT (MoE): we directly apply AT to MoE-CNN, which co-trains the routers and backbone network. Note this method is also adopted in the latest robust training algorithm [30] for ViT-based MoE architectures. It is worth noting that the above baselines use the same number of model parameters as the pathway of MoE-CNN during model prediction. In addition, we cover \u2463AT (Dense) (applying AT to Dense) to acquire a robustness performance reference. Yet, we remark that it is not quite fair to directly compare Dense with the aforementioned smaller counterparts, since the former uses a larger model scale (r = 1.0) at test-time inference. Training and evaluation. We use TRADES [25] as the default robust training objective for all baselines. We also follow the literature [24, 25, 27, 64] to set the attack strength by \u03f5 = 8/255 for CIFAR-10 and CIFAR-100, and \u03f5 = 2/255 for TinyImageNet and ImageNet. To implement ADVMOE (Algorithm 1), we mimic the TRADES training pipeline but conduct the proposed BLO routine to robustify routers and backbone parameters in an interactive mode. We adopt 2-step PGD attack [24] at training time for all the methods, supported by the recent work [67] showing its compelling performance in AT. We refer readers to Appendix B for more training details. During evaluation, we report standard accuracy (SA) on the clean test dataset and robust accuracy (RA) against test-time 50-step PGD attacks [24] with the attack strength same as the training values. We also report GFLOPs (FLOPS \u00d7109) as an indicator of the test-time inference efficiency. 5.2. Experiment Results Overall performance. Tab. 1 presents the overall performance of our proposed ADVMOE algorithm vs. baselines. We make several key observations below. First, ADVMOE yields a significant robustness enhancement over all the baselines in every data-model setup. Specifically, ADVMOE consistently yields an improvement of around 1% \u223c5% on the robustness measured by RA against PGD attacks. Notably, ADVMOE can also outperform \u2022 AT (Dense) in most cases, around 1% \u223c4% robustness improvement (see highlighted results in green ). This is remarkable since Dense (r = 1.0) is twice larger 6 \fTable 1. Performance overview of ADVMOE (our proposal) vs. baselines on various datasets and model backbone architectures. The model scale is fixed at r = 0.5 for Dense-CNN, Sparse-CNN and Moe-CNN (denoted with the symbol \u25e6, since they are fairly comparable to each other) compared to Dense (r = 1.0, denoted with the symbol \u2022 ). For trainand test-time attack generations, we adopt an attack strength of \u03f5 = 8/255 for CIFAR-10 and CIFAR-100, and \u03f5 = 2/255 for TinyImageNet and ImageNet. We evaluate RA (robust test accuracy) against 50-step PGD attack [24], SA (standard test accuracy), and GFLOPS (FLOPS \u00d7109) per test-time example (test-time inference efficiency) for each model. In each (dataset, backbone) setup, \u2460we highlight the best SA and RA over all baselines per model scale in bold, and \u2461we mark the performance better than AT (Dense) in green . Results in the format of a\u00b1b provide the mean value a and its standard deviation b over 3 independent trials. Method Backbone RA (%) SA (%) GFLOPS(#) Method Backbone RA (%) SA (%) GFLOPS (#) CIFAR-10 \u2022 AT (Dense) ResNet-18 50.13\u00b10.13 82.99\u00b10.11 0.54 \u2022 AT (Dense) WRN-28-10 51.75\u00b10.12 83.54\u00b10.15 5.25 \u25e6AT (S-Dense) 48.12\u00b10.09 80.18\u00b10.11 0.14 (74%\u2193) \u25e6AT (S-Dense) 50.66\u00b10.13 82.24\u00b10.10 1.31 (75%\u2193) \u25e6AT (Sparse) 47.93\u00b10.17 80.45\u00b10.13 0.14 (74%\u2193) \u25e6AT (Sparse) 48.95\u00b10.14 82.44\u00b10.17 1.31 (75%\u2193) \u25e6AT (MoE) 45.57\u00b10.51 78.84\u00b10.75 0.15 (72%\u2193) \u25e6AT (MoE) 46.73\u00b10.46 77.42\u00b10.73 1.75 (67%\u2193) \u25e6ADVMOE 51.83 \u00b10.12 80.15\u00b10.11 0.15 (72%\u2193) \u25e6ADVMOE 55.73 \u00b10.13 84.32 \u00b10.18 1.75 (67%\u2193) \u2022 AT (Dense) VGG-16 46.19\u00b10.21 82.18\u00b10.23 0.31 \u2022 AT (Dense) DenseNet 44.52\u00b10.14 74.97\u00b10.19 0.07 \u25e6AT (S-Dense) 45.72\u00b10.18 80.10\u00b10.16 0.07 (77%\u2193) \u25e6AT (S-Dense) 38.07\u00b10.13 69.63\u00b10.11 0.02 (71%\u2193) \u25e6AT (Sparse) 46.13\u00b10.15 79.32\u00b10.18 0.07 (77%\u2193) \u25e6AT (Sparse) 37.73\u00b10.13 67.35\u00b10.12 0.02 (71%\u2193) \u25e6AT (MoE) 43.37\u00b10.46 76.49\u00b10.65 0.12 (61%\u2193) \u25e6AT (MoE) 35.21\u00b10.74 64.41\u00b10.81 0.03 (57%\u2193) \u25e6ADVMOE 49.82 \u00b10.11 80.03\u00b10.10 0.12 (61%\u2193) \u25e6ADVMOE 39.97\u00b10.11 70.13\u00b10.15 0.03 (57%\u2193) CIFAR-100 \u2022 AT (Dense) ResNet-18 27.23\u00b10.08 58.21\u00b10.12 0.54 \u2022 AT (Dense) WRN-28-10 27.90\u00b10.13 57.60\u00b10.09 5.25 \u25e6AT (S-Dense) 26.41\u00b10.16 57.02\u00b10.14 0.14 (74%\u2193) \u25e6AT (S-Dense) 26.30\u00b10.10 56.80\u00b10.08 1.31 (75%\u2193) \u25e6AT (Sparse) 26.13\u00b10.14 57.24\u00b10.12 0.14 (74%\u2193) \u25e6AT (Sparse) 25.83\u00b10.16 57.39\u00b10.14 1.31 (75%\u2193) \u25e6AT (MoE) 22.72\u00b10.42 53.34\u00b10.61 0.15 (72%\u2193) \u25e6AT (MoE) 22.94\u00b10.55 53.39\u00b10.49 1.75 (67%\u2193) \u25e6ADVMOE 28.05 \u00b10.13 57.73\u00b10.11 0.15 (72%\u2193) \u25e6ADVMOE 28.82 \u00b10.14 57.56\u00b10.17 1.75 (67%\u2193) \u2022 AT (Dense) VGG-16 22.37\u00b10.15 52.36\u00b10.17 0.31 \u2022 AT (Dense) DenseNet 21.72\u00b10.13 48.64\u00b10.14 0.07 \u25e6AT (S-Dense) 20.58\u00b10.13 48.89\u00b10.14 0.07 (77%\u2193) \u25e6AT (S-Dense) 16.86\u00b10.21 39.97\u00b10.11 0.02 (71%\u2193) \u25e6AT (Sparse) 21.12\u00b10.22 48.03\u00b10.17 0.07 (77%\u2193) \u25e6AT (Sparse) 17.72\u00b10.14 41.03\u00b10.16 0.02 (71%\u2193) \u25e6AT (MoE) 19.34\u00b10.43 45.51\u00b10.75 0.12 (61%\u2193) \u25e6AT (MoE) 14.45\u00b10.45 36.72\u00b10.71 0.03 (57%\u2193) \u25e6ADVMOE 21.21\u00b10.21 48.33\u00b10.17 0.12 (61%\u2193) \u25e6ADVMOE 23.31 \u00b10.11 48.97 \u00b10.14 0.03 (57%\u2193) Tiny-ImageNet \u2022 AT (Dense) ResNet-18 38.17\u00b10.14 53.81\u00b10.16 2.23 \u2022 AT (Dense) WRN-28-10 38.82\u00b10.15 55.30\u00b10.19 21.0 \u25e6AT (S-Dense) 36.29\u00b10.16 52.15\u00b10.13 0.55 (75%\u2193) \u25e6AT (S-Dense) 37.09\u00b10.12 54.83\u00b10.16 5.26 (75%\u2193) \u25e6AT (Sparse) 36.11\u00b10.13 50.75\u00b10.17 0.55 (75%\u2193) \u25e6AT (Sparse) 37.32\u00b10.14 54.32\u00b10.23 5.26 (75%\u2193) \u25e6AT (MoE) 34.41\u00b10.31 47.73\u00b10.41 0.75 (68%\u2193) \u25e6AT (MoE) 33.31\u00b10.41 49.91\u00b10.52 7.44 (65%\u2193) \u25e6ADVMOE 39.99 \u00b10.12 53.31\u00b10.14 0.75 (68%\u2193) \u25e6ADVMOE 40.15 \u00b10.15 55.18\u00b10.09 7.44 (65%\u2193) ImageNet \u2022 AT (Dense) ResNet-18 44.64\u00b10.14 60.32\u00b10.15 1.82 \u2022 AT (Dense) WRN-28-10 45.13\u00b10.14 60.97\u00b10.16 16.1 \u25e6AT (S-Dense) 41.19\u00b10.16 58.32\u00b10.12 0.48 (74%\u2193) \u25e6AT (S-Dense) 41.72\u00b10.15 58.98\u00b10.18 4.04 (75%\u2193) \u25e6AT (Sparse) 40.87\u00b10.15 58.22\u00b10.13 0.48 (74%\u2193) \u25e6AT (Sparse) 39.88\u00b10.18 59.21\u00b10.14 4.04 (75%\u2193) \u25e6AT (MoE) 35.57\u00b10.73 55.47\u00b10.66 0.67 (63%\u2193) \u25e6AT (MoE) 37.42\u00b10.44 56.44\u00b10.71 5.15 (68%\u2193) \u25e6ADVMOE 43.32 \u00b10.12 59.72\u00b10.17 0.67 (63%\u2193) \u25e6ADVMOE 46.82 \u00b10.11 58.87\u00b10.07 5.15 (68%\u2193) than an MoE pathway (r = 0.5). Second, we observe that ADVMOE has a preference on wider models. For instance, when WRN-28-10 (the widest model architecture in experiments) is used, ADVMOE yields better robustness over the Dense counterpart across all the dataset setups. Third, we also observe that the direct AT application to MoE-CNN, i.e., AT (MoE), is worse than AT (S-Dense) and ADVMOE in all setups. This is consistent with our findings in Sec. 4. We remark that although the usefulness of AT (MoE) was exploited in [30] for the MoE-type ViT, it is not effective for training MoE-type CNNs anymore. Fourth, ADVMOE can retain the high inference efficiency for MoE-CNN, as evidenced by the GFLOPS measurements in Tab. 1. Compared to S-Dense, MoE-CNN introduces minor computational overhead due to the routing system. However, it saves more than 50% of the inference cost vs. Dense. This implies that our proposal ADVMOE can preserve the efficiency merit of the MoE structure while effectively improving its adversarial robustness. Robust evaluation on AutoAttack [68]. In Tab. 2, we provide additional experiments evaluated by AutoAttack [68] (termed RA-AA), a popular robustness evaluation benchmark [69]. The experiment setting in Tab. 2 follows Tab. 1. We report RA-AA on CIFAR-10 and CIFAR-100 with ResNet-18 and WRN-28-10. As we can see, although AutoAttack leads to a lower RA-AA compared to RA evaluated using PGD attacks (termed 7 \fTable 2. Robustness overview evaluated with AutoAttack [68] (RA-AA) on various datasets and model backbone architectures. Other settings strictly follow Tab. 1. The values of RA-PGD, SA, and GFLOPS are repeated from Tab. 1 for better comparison. Method Backbone RA-PGD (%) RA-AA (%) SA (%) GFLOPS(#) Method Backbone RA-PGD (%) RA-AA (%) SA (%) GFLOPS (#) CIFAR-10 \u2022 AT (Dense) ResNet-18 50.13\u00b10.13 44.72\u00b10.15 82.99\u00b10.11 0.54 \u2022 AT (Dense) WRN-28-10 51.75\u00b10.12 45.13\u00b10.12 83.54\u00b10.15 5.25 \u25e6AT (S-Dense) 48.12\u00b10.09 42.24\u00b10.13 80.18\u00b10.11 0.14 (74%\u2193) \u25e6AT (S-Dense) 50.66\u00b10.13 44.14\u00b10.10 82.24\u00b10.10 1.31 (75%\u2193) \u25e6AT (Sparse) 47.93\u00b10.17 42.11\u00b10.11 80.45\u00b10.13 0.14 (74%\u2193) \u25e6AT (Sparse) 48.95\u00b10.14 43.97\u00b10.11 82.44\u00b10.17 1.31 (75%\u2193) \u25e6AT (MoE) 45.57\u00b10.51 40.42\u00b10.19 78.84\u00b10.75 0.15 (72%\u2193) \u25e6AT (MoE) 46.73\u00b10.46 41.11\u00b10.23 77.42\u00b10.73 1.75 (67%\u2193) \u25e6ADVMOE 51.83 \u00b10.12 45.13 \u00b10.07 80.15\u00b10.11 0.15 (72%\u2193) \u25e6ADVMOE 55.73 \u00b10.13 45.89 \u00b10.11 84.32 \u00b10.18 1.75 (67%\u2193) CIFAR-100 \u2022 AT (Dense) ResNet-18 27.23\u00b10.08 23.11\u00b10.06 58.21\u00b10.12 0.54 \u2022 AT (Dense) WRN-28-10 27.90\u00b10.13 23.45\u00b10.11 57.60\u00b10.09 5.25 \u25e6AT (S-Dense) 26.41\u00b10.16 22.11\u00b10.13 57.02\u00b10.14 0.14 (74%\u2193) \u25e6AT (S-Dense) 26.30\u00b10.10 22.23\u00b10.13 56.80\u00b10.08 1.31 (75%\u2193) \u25e6AT (Sparse) 26.13\u00b10.14 21.89\u00b10.11 57.24\u00b10.12 0.14 (74%\u2193) \u25e6AT (Sparse) 25.83\u00b10.16 21.97\u00b10.09 57.39\u00b10.14 1.31 (75%\u2193) \u25e6AT (MoE) 22.72\u00b10.42 16.33\u00b10.25 53.34\u00b10.61 0.15 (72%\u2193) \u25e6AT (MoE) 22.94\u00b10.55 17.87\u00b10.24 53.39\u00b10.49 1.75 (67%\u2193) \u25e6ADVMOE 28.05 \u00b10.13 23.33 \u00b10.06 57.73\u00b10.11 0.15 (72%\u2193) \u25e6ADVMOE 28.82 \u00b10.14 23.57 \u00b10.12 57.56\u00b10.17 1.75 (67%\u2193) RA-PGD), ADVMOE still outperforms AT (S-Dense), AT (Sparse), and AT (MoE) consistently, evidenced by the bold numbers in the RA-AA columns. 0.0 0.2 0.4 0.6 0.8 1.0 IoU score 0 2 4 6 8 10 12 14 Density AdvMoE vs. AT (Sparse) Clean Data Adv Data 0.0 0.2 0.4 0.6 0.8 1.0 IoU score 0 2 4 6 8 10 12 14 Density AT (MoE) vs. AT (Sparse) Clean Data Adv Data (a) ADVMOE (b) AT (MoE) Figure 6. The distribution of the intersection of union (IoU) scores of the input-specific pathways generated by ADVMOE (a) and AT (MoE) (b) vs. the static mask found by AT (Sparse). The distribution over the clean test set and the adversarial test set is plotted for AT (MoE) and ADVMOE on setting (ResNet-18, CIFAR-100). Other settings are aligned with Tab.1. MoE-CNN trained by ADVMOE enjoys better router utility. Based on the results above and the preliminary studies in Sec. 4, we next peer into the performance difference achieved by AT (Sparse), AT (MoE), and ADVMOE from the perspective of pathway diversities. We ask: \u2460What is the relationship between the dynamic pathways generated by the routers trained by ADVMOE and the static mask optimized by AT (Sparse)? \u2461What is the difference between the routing decisions using ADVMOE and AT (MoE), and how does it impact the performance? Regarding \u2460, we investigate the cosine similarity between the pathways generated by training methods, either AT (MoE) or ADVMOE, and the static mask found by AT (Sparse). Since the latter can be regarded as a single pathway used for all the data, we term it \u2018mask pathway\u2019 in contrast to \u2018MoE pathway\u2019. We calculate the intersection of union (IoU) score between the MoE pathway and the mask pathway under each testing dataset (the clean or adversarial version). Fig. 6 presents the IoU distributions based on the clean and adversarial test datasets (Fig. 6a for ADVMOE and Fig. 6b for AT (MoE)). We remark that a smaller IoU score indicates a larger discrepancy between the MoE pathway and the mask pathway. As we can see, the IoU distribution of ADVMOE vs. AT (Sparse) in Fig. 6a shifts closer to 0 compared with Fig. 6b. This observation applies to both standard and adversarial evaluation and suggests that ADVMOE (our proposal) has a better capability than AT (MoE) to re-build input-specific MoE pathways, which are more significantly different from the input-agnostic mask pathway identified by the pruning-based method, AT (Sparse). Regarding \u2461, we observe from Fig. 6 that the routers learned by AT (MoE) are more fragile to adversarial attacks compared to ADVMOE, as evidenced by the less intersection area of adversarial data vs. clean data. This is also aligned with Insight 3 in Sec.4. Moreover, the routing policy learned by ADVMOE is more diverse than AT (MoE), as indicated by the latter\u2019s density-concentrated IoU scores. In contrast, the distribution of ADVMOE is dispersed with a smaller peak value. Therefore, regarding the expert utility, ADVMOE is able to assign the inputs to a larger group of pathways than AT (MoE), making better use of experts. A coupling effect of expert number N and per-expert model scale r on ADVMOE. Recall that there exist two key parameters involved in MoE-CNN (Fig. A1): (a) the number of experts N, and (b) the model scale r that defines the per-expert (or per-pathway) model capacity. Given the backbone model (e.g., ResNet-18 in this experiment), a larger N paired with a small r implies that each expert may only have limited model capacity, i.e., corresponding to a less number of channels. Regardless of N, if r = 1, the full backbone network will be used to form the identical decision pathway. Fig. 7 shows the RA of MoE-CNN trained by ADVMOE vs. the model scale parameter r at different values of N. Two insightful observations can be drawn. First, there exists an MoE regime (e.g., N < 8 and r \u2208[0.5, 0.9]), in which ADVMOE can outperform AT (Dense) (i.e., r = 1) by a substantial margin. This shows the benefit of MoE in adversarial robustness. However, if the number of experts becomes larger (e.g., N = 10), the increasing diversity of MoE pathways can raise the difficulty of routers\u2019 robustification and thus hampers the performance of ADVMOE (see N = 10 and r = 0.8 in Fig. 7). Second, there exists an 8 \f1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 Model Scale 35 40 45 50 Robust Accuracy (%) # of experts/layer N=2 N=5 RT(Big-Dense) N=8 N=10 Figure 7. Performance of ADVMOE under CIFAR-10 using ResNet-18 as the backbone network for different values of expert number N and model scale r. The black dash line denotes the performance of Dense (i.e. r = 1). Figure 8. Robustness comparison of models trained with different methods under various model scale settings. Results higher than that of AT (Dense) are marked with \u22c6. Other setups are aligned with Tab. 1. Please refer to Appendix C for exact numbers and GFLOPS comparisons. ineffective MoE regime (e.g., N \u22658 and r < 0.5), in which the performance of ADVMOE largely deviates from that of AT (Dense). In this regime, each expert consists only of a small number of channels, which restricts its robust training ability. Accordingly, both the increasing diversity of MoE pathways (large N) and the limited capacity per pathway (small r) could impose the difficulties of AT for MoE-CNN. In our experiments, we choose r = 0.5 and N = 2, which preserves the diversity of MoE pathways (i.e., inference efficiency) and retains the effectiveness of robust training. Performance with different model scales. To make sure the observations and conclusions from Tab. 1 are consistent across different values of the model scale parameter r, we repeated the experiments on (CIFAR-10, ResNet-18) and (CIFAR-10, WRN-28-10) using r \u2208{0.2, 0.5, 0.8} to cover the {sparse, medium, dense} regimes with respect to Dense (r = 1.0). Fig. 8 summarizes the obtained experiment results. As we can see, ADVMOE yields consistent robustness improvements over all the baselines, including Dense. And the improvement rises as the model scale r increases. This is not surprising as more parameters will be used when processing one input. Yet, a clear drawback brought by the larger model scale r is the increase of inference cost, evidenced by the GFLOPS numbers. When r turns to be large (like r = 0.8), the efficiency benefit Table 3. Performance on robust training for MoE-ViT with in the setup (ImageNet, DeiT-Tiny). Other settings follow Tab. 1. Method RA (%) SA (%) GFLOPS (#) SOTA[30] 44.63 61.72 0.27 ADVMOE 45.93 61.67 0.27 brought by the pathway sparsification from MoE gradually vanishes. Thus, a medium sparsity (r = 0.5) is a better choice to balance the trade-off between performance and efficiency, which is thus adopted as our default setting. Extended study: ADVMOE for ViT. To explore the capability of our proposal ADVMOE on ViT-based MoE models (MoE-ViT), Tab. 3 presents additional results following the recently published SOTA baseline [30] for MoE-ViT. As we can see, ADVMOE is also applicable to MoE-ViT and can boost robustness over the SOTA baseline by over 1% RA improvement, while achieving a similar level of SA. Thus, although our work focuses on robust training for MoE-CNN, it has the promise of algorithmic generality to other MoE-based architectures. We defer a more comprehensive study in the future. Additional experiments. We conduct ablation studies on (1) robustness evaluation using AutoAttack [68] (consistent findings can be drawn as PGD attacks), (2) attacks steps used in AT, and (3) additional explorations towards the coupling effect between the number of experts and the model scales. We refer readers to Appendix C for detailed results. 6."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.00788v3",
+                    "title": "An Introduction to Bi-level Optimization: Foundations and Applications in Signal Processing and Machine Learning",
+                    "abstract": "Recently, bi-level optimization (BLO) has taken center stage in some very\nexciting developments in the area of signal processing (SP) and machine\nlearning (ML). Roughly speaking, BLO is a classical optimization problem that\ninvolves two levels of hierarchy (i.e., upper and lower levels), wherein\nobtaining the solution to the upper-level problem requires solving the\nlower-level one. BLO has become popular largely because it is powerful in\nmodeling problems in SP and ML, among others, that involve optimizing nested\nobjective functions. Prominent applications of BLO range from resource\nallocation for wireless systems to adversarial machine learning. In this work,\nwe focus on a class of tractable BLO problems that often appear in SP and ML\napplications. We provide an overview of some basic concepts of this class of\nBLO problems, such as their optimality conditions, standard algorithms\n(including their optimization principles and practical implementations), as\nwell as how they can be leveraged to obtain state-of-the-art results for a\nnumber of key SP and ML applications. Further, we discuss some recent advances\nin BLO theory, its implications for applications, and point out some\nlimitations of the state-of-the-art that require significant future research\nefforts. Overall, we hope that this article can serve to accelerate the\nadoption of BLO as a generic tool to model, analyze, and innovate on a wide\narray of emerging SP and ML applications.",
+                    "authors": "Yihua Zhang, Prashant Khanduri, Ioannis Tsaknakis, Yuguang Yao, Mingyi Hong, Sijia Liu",
+                    "published": "2023-08-01",
+                    "updated": "2023-12-20",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "math.OC"
+                    ],
+                    "main_content": "INTRODUCTION Bi-level optimization (BLO) is a class of optimization problems involving two nested levels (upper and lower-level), where the objective and variables of the upper-level problem depend on the optimizer of the lower-level one. The canonical formulation of BLO is given by: Upper-level optimization over \u03b8 z }| { minimize \u03b8\u2208U f(\u03b8, \u03d5\u2217(\u03b8)), subject to \u03d5\u2217(\u03b8) \u2208arg min h(\u03b8,\u03d5)\u22640 g(\u03b8, \u03d5) | {z } Lower-level optimization over \u03d5 , (BLO) where we assume that f, g and h are bivariate smooth functions, \u03b8 \u2208Rm denotes the upper-level variable subject to the upper-level constraint set U, \u03d5 \u2208Rn is the lower-level variable subject to the constraint h(\u03b8, \u03d5) \u22640 that couples both \u03b8 and \u03d5, and \u03d5\u2217(\u03b8) is one lower-level optimal solution. It is evident that the lower-level problem is an auxiliary problem since its solution supports the upper-level problem in finding a better solution. The study of BLO can be traced to that of Stackelberg games [1], where the upper (resp. lower) problem optimizes the action taken by a leader (resp. the follower). Early works in optimization utilized BLO to solve resource allocation problems [2]\u2013[4]; see [5] for a comprehensive survey of BLO algorithms up until mid-2000, and some more recent surveys on discrete BLO [6], BLO under uncertainty [7], and nonlinear and nonconvex aspects of BLO [8]. In recent years, BLO has regained popularity because a subclass of BLO has been used to formulate and solve various challenging problems in signal processing (SP), machine learning (ML), and artificial intelligence (AI). Notable applications in SP include resource management [9]\u2013[12], signal demodulation [13]\u2013[15], channel prediction [16]\u2013[18], image reconstruction [19] and image denoising [20]. In addition, BLO has also been used to make ML models, especially deep neural networks (DNNs), robust [21]\u2013[27], generalizable [28]\u2013[39], efficient [40]\u2013[44], easier to train [45]\u2013[47] and scalable [48]\u2013[54]. As can be easily imagined, the popularity of BLO in aforementioned applications is largely attributed to its ability to handle (often implicit) hierarchical structures. To better illustrate the challenges brought by the hierarchical architecture, an example application of coreset selection for model training [12], [40], [55] is discussed below. \f3 Motivating application: Coreset selection for model training Many contemporary SP and ML applications are facing significant challenges in data storage, transportation, and computation because they have to deal with excessive amounts of data. Consequently, the task of identifying the most informative subset of data from a larger pool becomes crucial [12], [40], [55]. This leads to the problem of coreset selection, which consists of two tasks: (T1) selecting the most representative data samples to form the coreset, and (T2) validating the performance of the selected coreset in model training. More specifically, the problem can be formulated as follows: minimize w\u2208U \u2113val(\u03b8\u2217(w)) (T1) subject to \u03b8\u2217(w) = arg min \u03b8 \u2113tr(\u03b8, w), (T2) where w represents the weight vector for data selection, with wi = 0 indicating that the ith data sample is not selected. These weights are subject to the sparsity constraint U, such as \u2225w\u22251 \u2264k, with k being the selection budget. The model parameters trained on the selected data points are denoted by \u03b8. The training loss for the model \u03b8 with the data selection scheme w is denoted by \u2113tr, while the validation loss measuring the performance of the learned model \u03b8\u2217(w) over the coreset is denoted by \u2113val. The above BLO formulation is also related to the data reweighting problem [56] and hyperparameter optimization [57]. Clearly, the coreset selection is a typical BLO problem, where the upper-level and lower-level tasks are intertwined: without knowing the training result, it is hard to gauge how representative the selected dataset is effective, while without having the coreset, one cannot perform model training to evaluate the utility of the learned model further. More importantly, these two tasks form a hierarchy, with the model training problem being the main optimization problem, while the data selection problem is an auxiliary problem that supports the training. Given the growing interest in BLO, this work presents an overview of a class of tractable (BLO) problems that hold significant importance in SP and ML. Roughly speaking, the class of BLO problems we consider have some desirable properties (to be discussed shortly) that allow the development of efficient and practical algorithms. We will discuss the basic concepts for this class of BLO problems along with their optimality conditions, standard algorithms (including their theoretical properties and practical implementations), as well as how they can be used to obtain state-of-the-art results for a number of key SP and ML applications. \f4 In the existing literature, several recent surveys have been conducted on general BLO problems [7], [8], [58]. However, these surveys primarily focus on the mathematical foundations of BLO through a classical optimization lens. Other works [59]\u2013[61] aim to provide comprehensive reviews of BLO algorithms, but they lack an in-depth discussion on the proper recognition and utilization of BLO, specifically addressing \u2018when\u2019 and \u2018how\u2019 to apply it effectively. The most relevant work to ours is [61], which examines complex learning and vision problems from the BLO perspective. However, the theoretical component of BLO is missing, and it overlooks a significant portion of emerging SP and ML applications (e.g., those discussed in Sec. V-IX). Different from the existing surveys on BLO [7], [8], [59]\u2013[61] to provide a broad overview of BLO in its most generic form, we focus on problems that are relevant to SP and ML applications and strive to tightly integrate theory with applications, especially those that have been largely left out of the existing surveys. A few highlights of this article are listed below. First, we distill the common structures and properties of BLO that emerge across applications related to developing robust, parsimonious, and generalizable data-driven models in SP and ML. Our goal is to provide insights about when, where, and how BLO formulations and algorithms can be best used to yield a significant performance boost, as compared with traditional, or heuristic algorithms. In this process, we present some recent theoretical results about BLOs and their associated algorithms to give a flavor of the current advances in the research area, while discussing their practical and scalable implementations. Second, we dive deep to understand the performance of a selected subset of state-of-theart (SOTA) BLO algorithms on a number of representative applications. Instead of relying on results reported in existing works, which may not always be comprehensive, we designed an experiment plan and implemented all benchmarking algorithms ourselves. The goal is not only to showcase the effectiveness of the BLO-based algorithms but also to analyze the pros (e.g., modeling flexibility and accuracy performance) and the cons (e.g., run-time efficiency) of different subclasses of BLO methods. Overall, we hope that our balanced treatment of the subject will serve as the cornerstone for the accelerated adoption of BLO in diverse application areas. We hope that this tutorial can assist researchers with modeling, analyzing, and innovating on the emerging SP and ML applications. And it can serve to sparkle new theoretical and applied research to advance BLO. Fig. I provides an overview of the topics to be covered in this article. Notations: We use lower case letters (e.g., a), lower case bold face letters (e.g., a), and upper \f5 case bold face letters (e.g., A) to denote scalars, vectors, and matrices, respectively. For a vector a, we use \u2225a\u2225p to denote its \u2113p norm with the typical choice p \u2208{1, 2, \u221e}. For a matrix A, we use the superscript \u22a4(or \u22121) to denote the transpose (or inverse) operation. We use I to represent the identity matrix. For a function f(x, y) (with x \u2208Rm and y \u2208Rn), we use \u2207xf(x, y) \u2208Rm (or \u2202f \u2202x) and \u2207yf(x, y) \u2208Rn (or \u2202f \u2202y) to denote the partial derivatives of f w.r.t. the partial input argument x and y, respectively. In contrast, we use d f dx \u2208Rm to represent the full derivative of f w.r.t. x, namely, d f dx = \u2207xf(x, y)+ dy dx \u22a4\u2207yf(x, y) following the chain rule, where dy dx \u22a4\u2208Rm\u00d7n denotes the Jacobian matrix of y w.r.t. x. For ease of notation, the transpose in dy dx \u22a4\u2208Rm\u00d7n will be omitted if its definition is clear from the context. We use \u2207x\u2207yf or \u22072 x,yf \u2208Rm\u00d7n to denote the second-order partial derivative of f. Signal Processing M achine Learning Resource  Allocation [9-12] Signal  D em odulation  [13,14] Channel  Prediction [16-18] Im age Reconstruction [19,20] Im plicit Function  (IF) Gradient Unrolling (GU) Value-Function (VF) Forward Gradient Unrolling  (FGU) [46] Backward Gradient Unrolling (BGU) [47, 57] Truncated Gradient Unrolling  (T GU) [26, 57] Robust  M L [21-23,25-26] Efficient  M L [40-42,106-107] Generalizable  M L [32-39] Autom ated  M L [45-50] Bi-level D escent Aggregation  (BDA) [45] Two-T im escale Stochastic  Approxim ation (T T SA) [70] Single-tim escale D ouble-m om entum   Stochastic Approxim ation          (SUSTAIN ) [92] Alternating Projected SGD  (AiPOD ) [87] Bilevel Value-Function-based  Interior-point M ethod (BVFIM ) [83] Bilevel Optim ization M ade Easy  (BOM E!) [86] Personalized D ecentralized Bilevel  Optim ization (PDBO) [75] Penalty-based Bilevel Gradient D escent  (PBGD ) [76] Stochastic Bilevel Optim ization (stocBiO) [91] Principles BLO Algorithms BLO Application Bilevel Stochastic Approxim ation (BSA) [66] K-step Truncated Backpropagation (K-RM D )[57] Fully First-order Stochastic  Approxim ation (F2SA)[100] (Sec. II-III) (Sec. III-IV) (Sec. V-IX) Fig. 1. An taxonomy of the solvers and application tasks for BLO. II. WARM-UP: INTRODUCING BASIC CONCEPTS OF BLO A. A Class of Tractable BLOs We start by discussing the challenges associated with the generic form of (BLO). Even under the assumption that all involved functions are well-behaved, such as the convexity or concavity of \f6 f(\u00b7, \u00b7) and g(\u00b7, \u00b7), and the linearity of h(\u00b7, \u00b7), solving the problem can still be highly challenging (i.e., NP-hard). To see this, let us consider the following simple example. Example 1: Non-convex BLO Consider the following BLO, where f(\u03b8, \u03d5) = \u2212g(\u03b8, \u03d5) = \u03b82 \u2212\u03b8 \u00b7 \u03d5 \u2212\u03d52: minimize \u03b8\u2208[\u22121,1] \u03b82 \u2212\u03b8 \u00b7 \u03d5\u2217(\u03b8) \u2212\u03d5\u2217(\u03b8)2; subject to \u03d5\u2217(\u03b8) = arg min \u03d5\u2208[\u22121,1],\u03b8\u2212\u03d5=0 \u2212 \u0000\u03b82 \u2212\u03b8 \u00b7 \u03d5 \u2212\u03d52\u0001 . Notice that the objective f(\u03b8, \u03d5) (resp. g(\u03b8, \u03d5)) is strongly convex (resp. strongly concave) in \u03b8, strongly concave in \u03d5 (resp. strongly convex), and the above BLO problem is subject to linear constraints in both the upper and lower level. In other words, both the upper and lower-level problems are convex problems. Nonetheless, it can be shown that solving the above BLO requires tackling a non-convex problem, which in general is NP-hard. Indeed, it is not hard to see that \u03d5\u2217(\u03b8) = \u03b8. As a result, the outer function can be expressed as \u2113(\u03b8) = \u2212\u03b82, which is a non-convex function. We remark that the source of difficulty of the above problem is the coupling constraint \u03b8\u2212\u03d5 = 0. If this constraint is removed, the problem will become a classical saddle-point problem, expressed below, whose global optimal solution can be easily obtained: minimize \u03b8\u2208[\u22121,1] maximize \u03d5\u2208[\u22121,1] \u03b82 \u2212\u03b8 \u00b7 \u03d5 \u2212\u03d52. (1) The aforementioned examples, along with numerous others in existing survey papers like [58]\u2013 [61], strongly motivate us to proceed with a focused discussion for the subset of tractable (BLO) problems. This subset often serves as a basis for developing practical BLO algorithms. By focusing on this subset, we can address the specific needs of modeling SP and ML problems, which frequently demand the development of efficient, and sometimes real-time, algorithms. To this end, we consider some special classes of (BLO), with the following simplifications: 1) The lower-level constraint set, if present, is linear that is only related to \u03d5, that is, h(\u03b8, \u03d5) = A\u03d5 \u2212b for some matrix A and vector b of appropriate sizes; 2) The solution of the lowerlevel problem is a singleton, and in most cases we assume an even stronger condition that the objective function g(\u00b7, \u00b7) is strongly convex in the second argument. With these simplifications, it is possible to show some nice properties, for example, the gradient of the upper-level objective function may exist, making algorithm design and analysis tractable. Note that there have been recent works that extend these conditions and are still able to develop efficient algorithms. We will discuss these works in Sec. IV. \f7 In summary, depending on whether the lower-level problem has constraint or not, we consider the following two classes of problems, referred to as the lower-uncsontrained (LU) and lowerconstrained (LC) BLO, respectively: minimize \u03b8\u2208U f(\u03b8, \u03d5\u2217(\u03b8)), subject to \u03d5\u2217(\u03b8) = arg min \u03d5\u2208Rn g(\u03b8, \u03d5). (LU-BLO) minimize \u03b8\u2208U f(\u03b8, \u03d5\u2217(\u03b8)), subject to \u03d5\u2217(\u03b8) = arg min \u03d5\u2208C g(\u03b8, \u03d5). (LC-BLO) where the set C := {\u03d5 | A\u03d5 \u2212b \u22640}. As will be evident in Sec. III, the presence of lower-level constrains, even in the form of linear and uncoupled constraints, can make (LU-BLO) much harder to deal with than (LC-BLO). B. Connections of BLO with Game Theory It is important to note that BLO problems have strong ties with Stackelberg or leader-follower games [1], including Stackelberg congestion [62] and security games [63]. These are sequential games involving two players, the leader and the follower. The leader acts first, aiming to maximize its utility by leveraging its knowledge of the follower\u2019s anticipated response. The follower, acting second, maximizes its utility based on the leader\u2019s action. The connection between BLO and Stackelberg games can be bifurcated as follows: First, in certain Stackelberg games, the process of identifying a solution (i.e., a Stackelberg equilibrium) can be framed as a BLO. Second, BLO allows a (Stackelberg) game-theoretic interpretation, where the upperand lower-level problems correspond to the tasks of identifying the optimal actions for the leader and the follower (i.e., the upper/lower-level variables), respectively. A special case of Stackelberg game is min-max optimization (MMO) also referred to as the saddle point problem. MMO follows a bi-level structure wherein the lower-level objective g in (BLO). is exactly opposite of the upper-level objective function f (i.e., g = \u2212f) resulting in the following special case of BLO: minimize \u03b8\u2208U maximize \u03d5\u2208C f(\u03b8, \u03d5). (MMO) In fact, MMO is much simpler to deal with than BLO, and it has been heavily studied in the SP and ML communities. We refer the interested reader to the recent tutorial [64] for detailed discussion. In addition to MMO, it is also worth noting that the algorithms reviewed in this paper are generally applicable to Stackelberg games, provided that the game\u2019s BLO formulation adheres to the assumptions of the respective algorithms. \f8 C. Implicit Gradient (IG) As alluded previously, one important reason to consider problems (LU-BLO) and (LC-BLO) is that the objective functions of these problems are potentially differentiable w.r.t. \u03b8. Indeed, by applying the chain rule, and suppose for now that the Jacobian matrix d\u03d5\u2217(\u03b8) d\u03b8 exists, we have: d f(\u03b8, \u03d5\u2217(\u03b8)) d\u03b8 = \u2207\u03b8f(\u03b8, \u03d5\u2217(\u03b8)) + d\u03d5\u2217(\u03b8)\u22a4 d\u03b8 | {z } IG \u2207\u03d5f(\u03b8, \u03d5\u2217(\u03b8)), (2) where recall from our notation convention that \u2207\u03b8f(\u03b8, \u03d5) and \u2207\u03d5f(\u03b8, \u03d5) represent the partial derivatives of f w.r.t. the partial input arguments \u03b8 and \u03d5 respectively, and d f d\u03b8 denotes the full derivative of f w.r.t. \u03b8. For ease of notation, the transpose operation \u22a4might be omitted in the rest of the paper. We refer to the Jacobian matrix d\u03d5\u2217(\u03b8) d\u03b8 as IG following [65]. This term is introduced to characterize the gradient of the arg min-based lower-level objective function w.r.t the upper-level variable \u03b8. However, IG does not always exist for generic BLO problems. Even for (LU-BLO) and (LC-BLO), relatively strong assumptions have to be imposed; For example, g(\u00b7, \u00b7) needs to be strongly convex in its second argument. Further, even IG exists, computing it could be quite different for the two classes of problems (LU-BLO) and (LC-BLO). For the former, we will show in Sec. III that the IG can be expressed in closed-form using the Implicit Function Theorem [66] based on the first-order stationary condition of the lower-level problem, i.e., \u2207\u03d5f(\u03b8, \u03d5\u2217(\u03b8)) = 0. And this is the reason that it is referred to as \u201cimplicit\u201d gradient. Yet, in (LC-BLO), the stationary condition cannot be used since a stationary point might violate the constraint \u03d5 \u2208C. Therefore IG generally does not admit any closed-form. Additionally, for a more restricted subset of problems MMO, the influence of the IG-involved term (IG \u00b7 \u2207\u03d5f(\u03b8, \u03d5\u2217(\u03b8))) in (2) can be neglected. To see this, assume that the inner problem in unconstrained, i.e., C \u2261Rn, then \u2207\u03d5f(\u03b8, \u03d5\u2217(\u03b8)) = 0 based on the fact that \u2207\u03d5g(\u03b8, \u03d5\u2217(\u03b8)) = 0 and g = \u2212f for MMO. D. BLO with Non-Singleton Lower-Level Solutions As we have mentioned in Sec. II-A, throughout this article, we will mostly focus on the case where the lower-level solution \u03d5\u2217(\u03b8) is unique, i.e., a singleton. Yet, if the lower-level problem involves a non-singleton (NS) solution, the resulting BLO problem is typically cast as minimize \u03b8,\u03d5\u2032\u2208S(\u03b8) f(\u03b8, \u03d5\u2032) subject to S(\u03b8) := arg min \u03d5\u2208C g(\u03b8, \u03d5) (NS-BLO) where S(\u03b8) denotes a solution set. The above formulation also referred to as the optimistic BLO with non-singleton lower-level solutions, has been discussed in the literature such as [6], [58]. Note that problem (NS-BLO) presents significantly greater challenges from both practical \f9 and theoretical perspectives compared to problem (BLO). This is because optimization over \u03d5 is coupled across both upperand lower-level objectives. While our work primarily focuses on BLO with a singleton lower-level solution, we will also explore in Sec.,III the applicability of BLO algorithms, derived from (BLO), to solve problem (NS-BLO). THEORY AND ALGORITHMS FOR TRACTABLE BLOS In the next two sections, we will delve into the essential optimization principles employed in BLO, explore several popular classes of BLO algorithms, and examine their theoretical properties. III. ALGORITHMIC FOUNDATIONS OF BLO This section presents an overview of three key optimization frameworks used to solve the tractable BLO problems (LU-BLO) and (LC-BLO). The first two classes both leverage (some approximated version of) the IG as defined in (2). The key difference is how the approximation of the IG is conducted: One directly assumes that there is some given procedure that can provide a high-quality solution of the lower-level problem, while the other approximates the lower-level solution by unrolling a given algorithm for a fixed number of steps. The third class is referred to as the value function (VF)-based approach, which reformulates BLO as a single-level regularized optimization problem. It is worth mentioning that this approach offers flexibility in handling lower-level constraints and solving non-singleton lower-level problems (NS-BLO). A. The Implicit Function (IF)-based Approach 1) IF for lower-level unconstrained BLO: Let us examine the problem setup (LU-BLO) with a singleton lower-level solution. For ease of theoretical analysis in Sec. IV, we further assume that g(\u00b7) is strongly convex. In certain applications, one can explicitly add a strongly convex regularization function, such as \u03b3 \u00d7 \u2225\u03d5\u22252 2 (with large enough \u03b3), to satisfy such an assumption. The reason that we call this approach IF-based is that we will explicitly utilize the Implicit Function Theorem [67] to calculate the IG as expressed in (2). Recall from (LU-BLO) that \u03d5\u2217(\u03b8) is a lower-level solution, thus it satisfies: \u2207\u03d5g(\u03b8, \u03d5\u2217(\u03b8)) = 0. (3) \f10 Following the implicit function theorem, we can take the first-order derivative of (3) w.r.t. \u03b8, yielding d d\u03b8[\u2207\u03d5g(\u03b8, \u03d5\u2217(\u03b8))] = 0. Further assume that the lower-level objective g(\u00b7) is secondorder differentiable, we can then obtain the IG in (2) in the following: d\u03d5\u2217(\u03b8) d\u03b8 = \u2212\u22072 \u03b8,\u03d5g(\u03b8, \u03d5\u2217(\u03b8))\u22072 \u03d5,\u03d5g(\u03b8, \u03d5\u2217(\u03b8))\u22121. (4) As observed above, the computation of IG involves the mixed (second-order) partial derivative \u22072 \u03b8,\u03d5g and the inverse of the Hessian \u22072 \u03d5,\u03d5g. Yet, computing these quantities can be challenging in practice. Therefore, the class of IF approach utilizes different kinds of approximation techniques to approximately compute the IG as expressed (4). We summarize the IF approach in Algorithm 1. Algorithm 1: IF-based approach for solving (LU-BLO) Given initialization \u03b80, learning rate \u03b1 > 0, and iteration number T; Iteration t \u22650 yields: \u2022 Lower-level optimization: Given \u03b8t, obtain an approximate solution of the lower-level problem, denoted as e \u03d5(\u03b8t); \u2022 Approximation: Based on e \u03d5(\u03b8t), compute approximated versions of two secondorder matrices in (4), denoted as e \u22072 \u03b8,\u03d5g(\u03b8, e \u03d5(\u03b8t)) and e \u22072 \u03d5,\u03d5g(\u03b8, e \u03d5(\u03b8t))\u22121; Compute an approximated IG following (2): \u02dc \u2207f(\u03b8t) := \u2207\u03b8f(\u03b8t, e \u03d5(\u03b8t)) \u2212e \u22072 \u03b8t,\u03d5g(\u03b8t, e \u03d5(\u03b8t))e \u22072 \u03d5,\u03d5g(\u03b8t, e \u03d5(\u03b8t))\u22121\u2207\u03d5f(\u03b8t, e \u03d5(\u03b8t)). (5) \u2022 Upper-level optimization: Utilize \u02dc \u2207f(\u03b8t) in (5) to update \u03b8t through, e.g., gradient descent (GD): \u03b8t+1 \u2190\u03b8t \u2212\u03b1 \u02dc \u2207f(\u03b8t). 2) Practical considerations of IF: In Algorithm 1, the main computational overhead arises from the inverse-Hessian gradient product H\u22121g, where H := e \u22072 \u03d5,\u03d5g(\u03b8t, e \u03d5(\u03b8t)) and v := \u2207\u03d5f(\u03b8t, e \u03d5(\u03b8t)). Yet, in many contemporary applications, directly computing and storing the Hessian is computationally prohibitive. To address the scalability challenge in the IF method, we introduce four approaches to approximate the inverse-Hessian gradient product (or the inverseHessian) H\u22121g: the conjugate gradient (CG) method [57], [68], the WoodFisher approximation [69], the Neumann-series method to directly estimate the inverse-Hessian [66], [70], and a Hessian-free simplification [22], [44]. These methods offer different trade-offs between computational costs, with the CG method being the most computationally expensive and the Hessian-free simplification being the least expensive. First, the CG approach maps the product H\u22121g to the solution of a quadratic program defined as minx x\u22a4Hx/2\u2212g\u22a4x. By utilizing the first-order gradient descent algorithm, we can numerically \f11 approximate H\u22121g. However, the convergence speed of the CG method relies on the smallest eigenvalue of the positive definite matrix H. Therefore, if the lower-level problem is not wellconditioned, the CG method can be slow. This approach has also been employed in the context of model-agnostic meta-learning (MAML) [29] and adversarially robust training [22]. Second, the WoodFisher approximation [69] expresses the Hessian as a recurrence of a rankone modified Hessian estimate and calls the Woodbury matrix identity to compute the inverse of a rank-one modification to the given Hessian matrix. The one-shot WoodFisher approximation is equivalent to the quasi-Newton approximation [71], H \u2248vvT + \u03b3I, where \u03b3 > 0 is the dampening term to render the invertibility of H. We can then readily obtain the inverse-Hessian vector product by the Woodbury matrix identity H\u22121v = \u03b3\u22121v+ \u03b3\u22122vv\u22a4 1+\u03b3\u22121v\u22a4vv. Further, an iterative WoodFisher approximation for Hessian inverse proposed in [69] enhances the estimation accuracy. Third, one may utilize a Neumann-series approximation to estimate the inverse-Hessian directly by the approximation1 H\u22121 \u2248PK i=0[I \u2212H]i. Note that as K \u2192\u221ethe approximation becomes more and more accurate [72]. This technique is popular for approximating the inverse-Hessian in a stochastic setting wherein the upperand lower-level objectives are accessed via a stochastic oracle [66], [70]. Here, we briefly describe the procedure to approximate the inverse-Hessian stochastically using the Neumann-series method. Let us choose k uniformly randomly from the set {0, 1, . . . , K\u22121}, access batch samples of g(\u03b8, \u03d5) denoted by {g(\u03b8, \u03d5; \u03b6k)}k k=1, and compute: H\u22121 \u2248k Lg k Y i=1 \u0010 I \u2212\u22072 \u03d5,\u03d5g(\u03b8, \u03d5; \u03b6i)/Lg \u0011 (6) where Lg is the Lipschitz-smoothness constant of g(\u03b8, \u03d5; \u03b6k). The above procedure requires the computation of k stochastic Hessians and their products. Importantly, this estimator is a biased estimator of the inverse-Hessian with the bias decreasing exponentially with K [66, Lem. 3.2]. Finally, to ensure the local convexity, some quadratic regularization term is usually added to the lower-level problem in BLO [22], [44]. This modifies (LU-BLO) to minimize \u03b8\u2208U f(\u03b8, \u03d5\u2217(\u03b8)), subject to \u03d5\u2217(\u03b8) = arg min \u03d5\u2208Rn g(\u03b8, \u03d5) + \u03bb 2 \u2225\u03d5\u22252 2, (7) where recall that \u03bb > 0 is a regularization parameter. In this context, the Hessian-free simplification is usually adopted, which assumes e \u22072 \u03d5,\u03d5g(\u03b8t, e \u03d5(\u03b8t)) = 0. This assumption can be reasonable when the lower-level objective function g involves deep model training. For instance, in the case of a neural network with ReLU activation, the decision boundary is piece-wise linear 1Assuming H is normalized to ensure \u2225H\u2225\u22641. \f12 in a tropical hyper-surface, leading to an approximate Hessian of zeros [73]. This Hessian-free simplification has been used for pruning deep neural networks [44]. Thus, the Hessian matrix of the entire lower-level objective function in (7) will be simplified to H \u2248\u03bbI. 3) Extension to lower-level constrained BLO: Unlike the previous subsection, it turns out that when including constraints to the lower-level problem, the IG no longer has the closed-form expression because the stationary condition in (3) does not hold anymore. To see the impact of having constraints (even linear ones) on the lower-level problem, we present the following example, where the gradient d f/d\u03b8 is not rigorously defined. Example 2: (LC-BLO) can still be non-differentiable [74] Consider the following special case of (LC-BLO), where the lower-level objective is strongly convex in both scalar variable \u03b8 and \u03d5, the upper-level is linear, and both levels are subject to linear constraints: minimize \u03b8\u2208[0,1] \u03b8 + \u03d5\u2217(\u03b8) subject to \u03d5\u2217(\u03b8) \u2208arg min 1/2\u2264\u03d5\u22641 (\u03b8 \u2212\u03d5)2 It follows that \u03d5\u2217(\u03b8) = 1/2, for \u03b8 \u22641/2, and \u03d5\u2217(\u03b8) = \u03b8, for \u03b8 > 1/2. We notice that at the point \u03b8 = 1/2 the mapping \u03d5\u2217(\u03b8) is continuous, but not differentiable. As a result, the outer function \u03b8 + \u03d5\u2217(\u03b8) is non-differentiable. An immediate question is, can we still leverage the IF-based approach for this subclass of problems? It turns out that if we make some additional assumptions on the matrix C in the constraint set of (LC-BLO), one can still apply the Implicit Function Theorem to the KarushKuhn-Tucker (KKT) condition of the lower level problem to calculate the IG [22], [74]. It is also important to note that IF-based approaches are typically not suitable for handling general non-linear constraints in the lower-level problem. For problems with complex constraints, value function or penalty-based approaches are often employed [75], [76]. B. The Gradient Unrolling (GU)-based Approach GU-based approach is another class of popular algorithms for solving BLO problems in practice [57], [77], [78]. Different from the IF-based framework, it employs an unrolled lowerlevel optimizer as an intermediary step to connect the lower-level solution with the upper-level optimization process. The automatic differentiation (AD) technique is then used to compute gradients w.r.t. the upper-level optimization variable \u03b8. Consequently, the computation of IG in \f13 GU is dependent on the choice of the lower-level optimizer, and it no longer uses the implicit function-based expression (3)-(4). 1) GU-based approach for unconstrained BLO: In particular, the GU-based approach approximates \u03d5\u2217(\u03b8) by running a given algorithm for a fixed number of iterations, and then inserting the entire trajectory into the upper-level objective; see Algorithm 2 for an illustration of the idea. Algorithm 2: GU-based approach for solving (LU-BLO) Given initialization \u03d50 and \u03b80, and iteration numbers K and T; Let h(\u00b7) : U \u00d7 C \u2192C denote one step of a given algorithm, which takes both \u03b8 and \u03d5 as input, and outputs an updated \u03d5. At iteration t \u22650, \u2022 Lower-level optimization by K-step optimization: \u03d5k = q(\u03b8t, \u03d5k\u22121), k = 1, \u00b7 \u00b7 \u00b7 K. (8) Define e \u03d5(\u03b8t) := \u03d5K = q(\u03b8t, q(\u03b8t, \u00b7 \u00b7 \u00b7 , q(\u03b8t, \u03d50))); \u2022 Upper-level Optimization: Leverage AD to compute the approximated gradient e \u2207f(\u03b8t, e \u03d5(\u03b8t)) := d f(\u03b8t,q(\u03b8t,q(\u03b8t,\u00b7\u00b7\u00b7 ,q(\u03b8t,\u03d50)))) d\u03b8 , and use this gradient to update \u03b8t. To see the difference between the AD and the IF-based approaches, let us consider the simple case where h(\u00b7) is the gradient mapping, q(\u03b8t, \u03d5k\u22121) = \u03d5k\u22121 \u2212\u03b2 \u00d7 \u2207\u03d5g(\u03b8t, \u03d5k\u22121) (for some constant stepsize \u03b2 > 0), and K = 1 (i.e., a single-step gradient descent step is performed for lower-level optimization). Further assume that \u03d50 is independent of \u03b8, then the closed-form expression of IG can be written as: d e \u03d5(\u03b8t) d\u03b8 = d[\u03d50 \u2212\u03b2 \u00d7 \u2207\u03d5g(\u03b8t, \u03d50)] d\u03b8 = \u2212\u03b2\u22072 \u03b8,\u03d5g(\u03b8t, \u03d50). (9) In some sense, the above computation is simpler than the computation of IG (4) in the IFbased approach, since the Hessian inverse is no longer needed. However, things can get much more complicated very quickly, as the total number of inner iterations K increases. Suppose that K = 2, we obtain: d e \u03d5(\u03b8t) d\u03b8 = d[\u03d51 \u2212\u03b2\u2207\u03d5g(\u03b8t, \u03d51)] d\u03b8 = \u2212\u03b2 \u0002 I + \u03b2 \u00d7 \u22072 \u03d5,\u03d5g(\u03b8t, \u03d51) \u0003 \u22072 \u03b8,\u03d5g(\u03b8t, \u03d50). (10) Clearly, the Hessian inverse is still not needed, but as the number of unrolling steps increases, much higher computational and memory requirements will be involved. 2) Practical considerations: When the unrolling step K becomes too large or the problem scale itself is computationally expensive for GU, manual unrolling becomes necessary to save \f14 memory costs and reduce computational overhead. Various GU approaches have been proposed to achieve this goal efficiently. Notable techniques include forward gradient unrolling (FGU) [47], backward gradient unrolling (BGU) [46], [47], [57], [78], and truncated gradient unrolling (TGU) [26], [57], [79]. In FGU, unrolling is performed iteratively, and in the final step K, the Jacobi of \u03d5K w.r.t. \u03b8 is given by: d\u03d5K d\u03b8 | {z } ZK = \u2202\u03d5K \u2202\u03d5K\u22121 | {z } AK d\u03d5K\u22121 d\u03b8 | {z } ZK\u22121 + \u2202\u03d5K \u2202\u03b8 | {z } BK = \u00b7 \u00b7 \u00b7 = K X i=1   K Y t=i+1 At ! Bi +  K Y t=1 \u00b7At ! Z0. (11) We also assume that \u03d50 is independent of \u03b8, which implies d\u03d50 d\u03b8 = 0. Consequently, the expression (11) can be rewritten as the following iterative form: Zk = AkZk\u22121 + Bk, k = 1, 2, \u00b7 \u00b7 \u00b7 , K. (FGU) Both Ak and Bk will be calculated along with the k-th lower-level step \u03d5k = q(\u03b8t, \u03d5k\u22121) and will be discarded immediately after Zk is obtained. Such an iterative nature of (FGU) makes it particularly suitable for scenarios that involve a large number of unrolling steps K, as the memory cost of calculating Ak and Bk only involves any gradient flow generated within the kth step. However, (FGU) requires keeping track of the matrices Ak, Bk, and Zk\u22121. Hence, it may not be suitable for problems with high-dimensional variables \u03b8 and \u03d5. To achieve more efficient computations when \u03b8 and \u03d5 are of large scale, BGU is introduced, which eliminates the need for storing any intermediate matrices [46], [47], [57], [78]. BGU explores the calculation of the IG following (2): d f(\u03b8, \u03d5K) d\u03b8 = \u2202f(\u03b8, \u03d5K) \u2202\u03b8 | {z } cK +d\u03d5\u22a4 K d\u03b8 | {z } Z\u22a4 K \u2202f(\u03b8, \u03d5K) \u2202\u03d5K | {z } dK (F GU) = cK + (Z\u22a4 K\u22121A\u22a4 K + B\u22a4 K)dK = (cK + B\u22a4 KdK) | {z } cK\u22121 +Z\u22a4 K\u22121 \u00b7 A\u22a4 KdK | {z } dK\u22121 = cK\u22121 + Z\u22a4 K\u22121dK\u22121 = \u00b7 \u00b7 \u00b7 = c0 + Z\u22a4 0 d0 = c\u22121. (12) Instead of calculating IG explicitly like (FGU) does, (12) directly obtains the gradient of the upper-level variable, which can be further simplified the following recursive formulas: ck\u22121 = ck + B\u22a4 k dk, k = 0, 1, \u00b7 \u00b7 \u00b7 , K, with cK = \u2202f(\u03b8, \u03d5K) \u2202\u03b8 dk\u22121 = A\u22a4 t dk, k = 0, 1, \u00b7 \u00b7 \u00b7 , K, with dK = \u2202f(\u03b8, \u03d5K) \u2202\u03d5K . (BGU) It can be observed that (BGU) only requires storing vectors (ck and dk) throughout the recursion by utilizing the Jacobian-vector product trick. As a result, BGU is particularly advantageous for problems with large-scale variables compared to (FGU). Yet, due to the recursive nature, BGU can be conducted only after all the K lower-level steps are finished. Thus, (BGU) needs to store \f15 all the unrolling steps {\u03d5k \u2208Rn}K k=1, compared with FGU. Consequently, it may not be efficient to handle BLO as the number of unrolling steps grows K. It should also be noted that GU differs from IF as its computation relies on the choice of the lower-level optimizer. For instance, employing sign-based gradient descent (signGD) [80], [81] as the lower-level optimizer leads to a computationally efficient GU variant referred to as signGD-based GU [82]. Specifically, the modified lower-level update rule (8) becomes: e \u03d5(\u03b8t) = \u03d5K; \u03d5k = \u03d5k\u22121 \u2212\u03b2sign (\u2207\u03d5g(\u03b8t, \u03d5k\u22121)) , k = 1, 2, . . . , K, (signGD) where sign(\u00b7) denotes the element-wise sign operation, and \u03b2 > 0 is a certain learning rate. Given the approximation d sign(x) dx = 0 (holding almost everywhere), IG can be simplified to: d e \u03d5(\u03b8t) d\u03b8 = d\u03d5K d\u03b8 = d\u03d5K\u22121 d\u03b8 = . . . = d\u03d50 d\u03b8 . (13) In the case that \u03d50 is independent of \u03b8, we can achieve the IG-free variant of the GU approach. C. Value Function (VF)-based Approach VF-based methods [75], [76], [83]\u2013[86] can also avoid the computation of the inverse of the Hessian required in IF. The key technique is to reformulate a standard BLO problem into a constrained single-level optimization problem. This reformulation involves transforming the lower-level problem into an upper-level inequality constraint. The resulting VF-based variants can then be solved using algorithms for constrained optimization. Furthermore, in comparison to IF and GU-based methods, the VF-based approach has broader applicability in solving complex BLO problems. Not only can it handle lower-level objectives with non-singleton solutions (including both convex and non-convex objectives), but more importantly, it can accommodate lower-level constraints as well. However, the VF-based approach has not been popular in practical SP and ML applications yet, mainly because this approach has not been able to deal with large-scale stochastic problems. This point will be illustrated shortly in Sec. IV. To understand VF-based approach, consider the following equivalent reformulation of (LC-BLO): minimize \u03b8,\u03d5\u2208C f(\u03b8, \u03d5), subject to g(\u03b8, \u03d5) \u2264g\u2217(\u03b8), (14) where g\u2217(\u03b8) := min\u03d5\u2208C g(\u03b8, \u03d5) is referred to as the VF (value function) of the lower-level problem. However, solving (14) is highly non-trivial, partly because g\u2217(\u03b8) is not necessarily smooth, and can be non-convex. To address these challenges, a relaxed version of problem (14) is typically considered by replacing g\u2217(\u03b8) with a smooth surrogate [75], [83]: g\u2217 \u00b5(\u03b8) = minimize \u03d5\u2208C g(\u03b8, \u03d5) + \u00b51 2 \u2225\u03d5\u22252 2 + \u00b52, (15) \f16 where \u00b5 := (\u00b51, \u00b52) is a pair of positive coefficients that are introduced to guarantee the smoothness of g\u2217 \u00b5(\u03b8) and to ensure the feasibility of the inequality constraint g(\u03b8, \u03d5) \u2264g\u2217 \u00b5(\u03b8). Given the relaxed VF formulation, one can adopt standard non-linear optimization algorithms, such as a penalty-based algorithm to solve the constrained optimization problem (14). For example, a log-barrier interior-point method named BVFIM is leveraged in [83] to solve a sequence of smooth approximated single-level problems of (14). Other related methods to solve (14) include PDBO [75], BOME [86], and V-PBGD [76]. IV. THEORETICAL RESULTS OF BLO In this section, we examine the theoretical guarantees of various BLO optimization methods. The section is divided into two parts. In the first part, we discuss the convergence results of popular algorithms for solving the (LU-BLO) problem, while the second part will focus on the more general formulations such as (LC-BLO) and (NS-BLO). We list specific algorithms that can handle both stochastic and deterministic BLO problems. Given the differences in theoretical analysis between stochastic and deterministic optimization, we also consider a generalized stochastic version of BLO, whose upperand lower-level objectives are f(\u03b8, \u03d5\u2217(\u03b8)) := 1 N N X i=1 f(\u03b8, \u03d5\u2217(\u03b8); \u03bei), g(\u03b8, \u03d5) := 1 N N X i=1 g(\u03b8, \u03d5; \u03b6i), (16) where \u03bei \u223cDf (resp. \u03b6i \u223cDg) represents the data sample of the upper-level (resp. lower-level) objective from the distribution Df (resp. Dg), and N is the total number of data samples. A. Convergence Measures of BLO In what follows, we introduce the convergence measures utilized for evaluating the performance of BLO algorithms. These measures serve to assess the quality of solutions obtained by these algorithms. First, we focus on the concept of an \u03f5-stationary point for (LU-BLO), which plays a crucial role in characterizing the convergence properties of BLO when the upper-level problem is unconstrained, i.e., U = Rm. In the deterministic setting of (LU-BLO), a point \u00af \u03b8 \u2208Rm is considered an \u03f5-stationary point if it satisfies \u2225\u2207f(\u00af \u03b8, \u03d5\u2217(\u00af \u03b8))\u22252 2 \u2264\u03f5. In the stochastic setting (16), where the algorithm incorporates randomness, the expectation is taken over the stochasticity of the algorithm. Thus, an \u03f5-stationary point is defined as E\u2225\u2207f(\u00af \u03b8, \u03d5\u2217(\u00af \u03b8))\u22252 2 \u2264\u03f5. It is important to note that when the upper-level problem in (LU-BLO) is constrained, i.e., U \u2282Rm then the upper-level objective f(\u00af \u03b8, \u03d5\u2217(\u00af \u03b8)) may not be differentiable over \u00af \u03b8 in general. When solving (LC-BLO) using the IF-based approach, if the IF is differentiable, similar measures of \f17 stationarity as in standard optimization can be utilized. However, if the IF is non-differentiable, alternative stationarity measures are commonly employed. These include subgradient optimality [74], proximal gradient methods [76], [87], and Moreau envelope techniques [70], [74], [88]. Furthermore, when employing VF-based approaches to solve (LC-BLO), a widely used measure of stationarity is to evaluate the convergence of the algorithms towards the KKT points of the constrained reformulation of the BLO problem [75], [89]. In addition, the concept of oracle complexity is employed to quantify the number of gradient evaluations needed to obtain an \u03f5-stationary solution, as defined earlier. We denote G(f, \u03f5) (resp. G(g, \u03f5)) as the total number of (stochastic) gradients of f (resp. g) evaluated to achieve an \u03f5stationary solution. This measure provides insights into the computational requirements of BLO algorithms and their scalability with respect to the problem size and desired solution accuracy. B. Convergence Guarantees for LU-BLO As discussed in Sec. III, solving (LU-BLO) requires computing the Hessian or its inverse. However, the stochastic formulation (16) leads to some challenges, especially for convergence analysis. For example, in the case of solving the stochastic BLO problem using Algorithm 1 (IF), the gradient estimates would be replaced with appropriate stochastic gradient estimates for both the upper and lower-level updates. However, obtaining an unbiased estimator for the Hessian inverse term in IG (as defined in (4)) is challenging. To overcome this challenge, a biased stochastic gradient estimator based on Neumann-series approximation as discussed in Sec. III-A2 has been used in [66], [70]. We note that the bias of the estimator can be easily controlled by choosing a larger batch to compute the Hessian of the lower-level objective [66, Lem. 3.2]. Moreover, we point out that stochastic CG (discussed in Sec. III-A2) can also be utilized to obtain the inverse-Hessian gradient vector product to approximate the IG [90], [91]. In addition, a key design choice for BLO algorithms is whether the inner problem is solved accurately or not. In a single-loop algorithm, one only performs a fixed number of steps for the lower-level updates before every upper-level updates [70], [92], [93], while in a double-loop, many lower-level updates are carried out to obtain a very accurate approximation of \u03d5\u2217(\u03b8) [66], [90], [91]. Typically, the former is simpler to implement in practice, while the latter is easier to analyze since the error caused by approximating \u03d5\u2217(\u03b8) can be well managed. In addition, the stochastic descent direction to solve both (or either) of upper and lower-level problems can be constructed using either vanilla stochastic gradient descent (SGD) [94] or variance reduced (VR) algorithms [95], [96]. Specifically, it is well known that VR-based algorithms can lead \f18 to improved theoretical convergence of stochastic algorithms compared to vanilla SGD-based algorithms to solve standard optimization problems. The VR-based algorithms accomplish this improved convergence by computing additional stochastic gradients on optimization variables computed in consecutive iterations [95], [96]. Similar behavior is observed in solving BLO algorithms using VR-based gradient constructions, as discussed next. Algorithm 3: (S)GD and VR for solving stochastic (LU-BLO) and (LC-BLO) Given initialization \u03b80 and iteration number T; Iteration t \u22650 yields: \u2022 Lower-level optimization: Given \u03b8t, call SGD (or VR based on both \u03b8t and \u03b8t\u22121) to obtain a lower-level solution e \u03d5(\u03b8t); \u2022 Approximation: Given e \u03d5(\u03b8t), compute a stochastic gradient estimate of (5) as follows: \u2013 Estimate stochastic versions of \u2207\u03b8f(\u03b8t, e \u03d5(\u03b8t)), \u2207\u03d5f(\u03b8t, e \u03d5(\u03b8t)), e \u22072 \u03b8,\u03d5g(\u03b8t, e \u03d5(\u03b8t)); \u2013 Approximate Hessian inverse, e \u22072 \u03d5,\u03d5g(\u03b8t, e \u03d5(\u03b8t))\u22121, following (6) in Sec. III-A2 \u2013 Obtain stochastic estimate of (5) and construct a descent direction \u02dc \u2207f(\u03b8t; \u02dc \u03be) for \u03b8 \u2022 Upper-level optimization: Call SGD (or VR) to update \u03b8t using \u02dc \u2207f(\u03b8t; \u02dc \u03be) (or \u02dc \u2207f(\u03b8t; \u02dc \u03be) and \u02dc \u2207f(\u03b8t\u22121; \u02dc \u03be)). In Algorithm 3, we provide a generic stochastic algorithm to solve BLO problems using the IF-based approach. As pointed out earlier, for double-loop algorithms e \u03d5(\u03b8t) will approximate \u03d5\u2217(\u03b8t) closely, while for a single-loop algorithm e \u03d5(\u03b8t) could be given by a crude approximation of \u03d5\u2217(\u03b8t). Table I provides a summary of the oracle complexities of existing BLO algorithms for achieving an \u03f5-stationary point in solving problem (LU-BLO). The theoretical results are categorized based on three algorithmic families: stochastic BLO (16), VR-based BLO, and deterministic BLO. The convergence performance is evaluated in terms of the oracle complexities G(f, \u03f5) and G(g, \u03f5), as introduced in Sec. IV-A. In Table I, we also illustrate the BLO solver employed by different optimization principles (i.e., IF, GU, or VF) and the algorithmic design choices, such as double versus single-loop. a) Stochastic BLO: This set of algorithms updates the lowerand upper-level variables using SGD, following the IF or GU optimization principle. The stochastic gradient estimates are evaluated as discussed in Algorithm 3 while for the deterministic setting, the gradients are approximated using the techniques discussed in Sec. III. Bi-level Stochastic Approximation (BSA) [66] was the first algorithm that offered finite-time convergence guarantees for solving unconstrained stochastic BLO problems. It employed a double-loop algorithm where the lower-level variable \f19 TABLE I SUMMARY OF CONVERGENCE RESULTS OF REPRESENTATIVE BLO ALGORITHMS FOR SOLVING (LU-BLO). THE CONVERGENCE PERFORMANCE IS MEASURED BY THE ORACLE COMPLEXITY G(f, \u03f5) AND G(g, \u03f5) FOR UPPERAND LOWER-LEVEL OPTIMIZATION, RESPECTIVELY. \u2018DOUBLE\u2019 OR \u2018SINGLE\u2019 REFERS TO THE CHOICE OF THE ALGORITHM DESIGN, EITHER DOUBLE LOOP OR SINGLE LOOP. Stochastic BLO Variance Reduced BLO Deterministic BLO Algorithm Principle Design Oracle Complexity Algorithm Principle Design Oracle Complexity Algorithm Principle Design Oracle Complexity G(f, \u03f5) G(g, \u03f5) G(f, \u03f5) G(g, \u03f5) G(f, \u03f5) G(g, \u03f5) BSA [66] IF Double O(1/\u03f52) O(1/\u03f53) STABLE [88] IF Single O(1/\u03f52) O(1/\u03f52) BA [66] IF Double O(1/\u03f5) O(1/\u03f55/4) TTSA [70] IF Single O(1/\u03f55/2) O(1/\u03f55/2) SUSTAIN [92] IF Single O(1/\u03f53/2) O(1/\u03f53/2) AID-BIO [91] IF Single O(1/\u03f5) O(1/\u03f5) STOCBIO [91] IF Double O(1/\u03f52) O(1/\u03f52) VRBO [97] IF Double O(1/\u03f53/2) O(1/\u03f53/2) ITD-BIO [91] IF Double O(1/\u03f5) O(1/\u03f5) ALSET [93] IF Single O(1/\u03f52) O(1/\u03f52) SABA [98] IF Double O(N 2/3/\u03f5) O(N 2/3/\u03f5) MSTSA [99] IF Single O(1/\u03f5) O(1/\u03f5) F2SA [100] VF Single O(1/\u03f57/2) O(1/\u03f57/2) F3SA [100] VF Single O(1/\u03f55/2) O(1/\u03f55/2) K-RMD [57] GU Double O(1/\u03f52) O(K/\u03f52) AMIGO [90] IF Double O(1/\u03f52) O(1/\u03f52) SBFW [101] IF Single O(1/\u03f54) O(1/\u03f52) FGU/BGU [46] GU Double N/A N/A is iteratively estimated with multiple SGD updates, resulting in a larger oracle complexity of O(1/\u03f53) for the inner-level optimization compared to O(1/\u03f52) for the upper-level optimization. Two-Timescale Stochastic Approximation (TTSA) [70], a fully single-loop algorithm (with projected SGD update for upper-level constrained optimization) improved the lower-level oracle complexity to O(1/\u03f55/2), however, at the cost of worsening the upper-level oracle complexity to O(1/\u03f55/2). More recently, Stochastic Bi-level Optimizer (stocBiO) [91], Stochastic Bi-level Algorithm (SOBA) [98], and Alternating Stochastic Gradient Descent (ALSET) [93] algorithms are developed to achieve O(1/\u03f52) complexity for both the upperand lower-level optimization. Note that stocBiO [91] requires a batch size of O(1/\u03f5) to achieve this complexity while ALSET [93] and SOBA [98] rely on only O(1) batches to achieve the same performance. In [100], the authors developed Fully First-order Stochastic Approximation (F2SA), a VF-based algorithm to solve (LU-BLO). The algorithm achieved an oracle complexity of O(1/\u03f57/2) for both upper and lower-level objectives while circumventing the need to compute Hessians (or Hessian vector products) during the execution of the algorithm. b) Variance-reduced (VR) stochastic BLO: Several VR algorithms have been proposed to improve the performance of vanilla stochastic algorithms by computing additional stochastic gradients in each iteration (see Algorithm 3). Examples including Single-Timescale Stochastic Bilevel Optimization (STABLE) [88] and Momentum-Assisted Single-Timescale Stochastic Approximation (MSTSA) [99] utilized momentum-based variance reduction techniques [95], [102] for upper-level optimization, improving the performance of TTSA [70] and BSA [66] and achiev\f20 ing a oracle complexity of O(1/\u03f52) for both upperand lower-level objectives. Single-timescale Double-momentum Stochastic Approximation (SUSTAIN) [92], Momentum-Based Recursive Bilevel Optimizer (MRBO) [97], and Stochastic Variance-Reduced Bilevel Method (SVRB) [103] further improved the performance by applying variance reduction to both upperand lower-level optimization, achieving the oracle complexity of O(1/\u03f53/2). Variance-Reduced Bilevel Optimizer (VRBO) [97] employed Stochastic Path-Integrated Differential Estimator (SPIDER) [96], a double-loop variance reduced gradient estimator, to achieve the same complexity. More recently, Stochastic Average Bilevel Algorithm (SABA) [98] was developed, applying SAGA (an incremental gradient estimator) [104] achieved an oracle complexity of O(N2/3/\u03f5), where N is the number of empirical data points. In addition to F2SA, the authors of [100] also proposed Faster Fully First-order Stochastic Approximation (F3SA) that improved on the oracle complexity of F2SA. F3SA utilized momentum-based variance reduction to achieve an oracle complexity of O(1/\u03f55/2) for both upper and lower-level objectives. c) Deterministic BLO: Popular approaches for solving deterministic BLO problems include Iterative Differentiation (ITD) and Approximate Implicit Differentiation (AID) methods. ITDbased approaches, proposed in [46], [47], established asymptotic convergence guarantees and were extended to TGU (truncated gradient unrolling) in [57], achieving oracle complexities of O(1/\u03f52). Improved guarantees were later shown in [91], achieving oracle complexities of O(1/\u03f5), comparable to solving a deterministic single-level optimization problem. AID-based approaches for solving deterministic BLO problems include Bi-level Approximation (BA) [66], Approximate Implicit Differentiation Bi-level optimizer (AID-BiO) [91], and MSTSA [99]. BA, a double-loop algorithm, was the first to establish convergence guarantees for deterministic BLO using AID. BA achieved an oracle complexity of O(1/\u03f5) for the upper-level optimization and O(1/\u03f55/4) for the lower-level optimization. The performance of BA was improved in AID-BiO [91] and MSTSA [99], which achieved an oracle complexity of O(1/\u03f5) for both upperand lower-level objectives. Please refer to Table I for a summary of the discussed approaches. C. Convergence Guarantees for (LC-BLO) and (NS-BLO) Obtaining convergence guarantees for BLO algorithms become more challenging when solving more complex problems, such as those involving lower-level constraints in (LC-BLO) or non-singleton lower-level solutions in (NS-BLO). In the previous section, the majority of the algorithms discussed employed an IF-based approach to solve the BLO problem. However, in this section, only the algorithms designed to solve (LC-BLO) utilize IF-based approaches. It is \f21 also worth mentioning that IF-based approaches are not applicable for solving (NS-BLO) (or (LC-BLO) with general constraints) due to the inapplicability of the implicit function theorem in this context. Instead, standard approaches to solve these more complex problems include interiorpoint methods [83], primal-dual methods [75], dynamic barrier gradient descent [86], and penalty based gradient descent [76]. As pointed out earlier, a major drawback of these algorithms is that they are exclusively developed for deterministic problems and lack efficient implementations for stochastic formulations (16). Consequently, they are not well-suited for large-scale SP and ML applications which often involve learning over large volumes of data. In the following, we present a summary of recent theoretical advancements for solving highly complex BLO problems such as (LC-BLO) and (NS-BLO) in Table II. We list the oracle complexities of representative methods, design principles, and the type of constraints present in the lower-level objective function. Note that in Table II, we list the oracle complexity for only upperlevel objective with the notion of stationarity defined according to either squared norm of the projected gradient [22], [74], [76], [89] or KKT conditions [75], [86]. 1) Theoretical results for (LC-BLO): BLO problems of the form (LC-BLO) involving linear constraints of the form C := {\u03d5 | h(\u03b8, \u03d5) \u22640} with h(\u03b8, \u03d5) := A\u03d5 \u2212b, in the lowerlevel have gained popularity in both theory and practice [22], [74], [87]. Under some regularity assumptions on upperand lower-level objectives and the constraint set of the lower-level problem, IF-based methods can be developed for solving such problems. These algorithms follow the same structure as the one presented in Algorithm 1, with the key difference that the construction of the (stochastic) gradient estimate depends on the lower-level constraints. Algorithm 4: SIGD, an IF-based approach for (LC-BLO) Given initialization \u03b80 and iteration number T; Iteration t \u22650 yields: \u2022 Call Algorithm 1 and use the following procedure to compute the IG \u2022 Notations: A be the matrix that contains the rows of A that correspond to the active constraints of inequality A\u03d5 \u2212b \u22640, \u03bb \u2217(\u03b8) is the Lagrange multipliers vector that corresponds to the active constraints at \u03d5\u2217(\u03b8) IG : d\u03d5\u2217(\u03b8)\u22a4 d\u03b8 = \u0002 \u22072 \u03d5\u03d5g(\u03b8, \u03d5\u2217(\u03b8)) \u0003\u22121 \u00b7 \u0002 \u2212\u22072 \u03b8\u03d5g(\u03b8, \u03d5\u2217(\u03b8)) \u2212A T \u2207\u03bb \u2217(\u03b8) \u0003 (17) \u2207\u03bb \u2217(\u03b8) = \u2212 \u0002 A \u0002 \u22072 \u03d5\u03d5g(\u03b8, \u03d5\u2217(\u03b8)) \u0003\u22121A \u22a4\u0003\u22121\u0002 A \u0002 \u22072 \u03d5\u03d5g(\u03b8, \u03d5\u2217(\u03b8)) \u0003\u22121\u22072 \u03b8\u03d5g(\u03b8, \u03d5\u2217(\u03b8)) \u0003 . In [74], the Smoothed Implicit Gradient (SIGD) approach was developed to handle (LC-BLO) \f22 with linear inequality constraints in the lower level. SIGD is an implicit gradient descent algorithm that ensures the differentiability of the implicit function through perturbation-based smoothing. The SIGD algorithm and the expression for IG used in [74] is stated in (17). SIGD guarantees asymptotic convergence to a stationary point. A similar IF-based approach was also utilized in [22], termed Fast Bi-level Adversarial Training (Fast-BAT). Fast-BAT aims to enhance the robustness of deep learning models against adversarial attacks using BLO. And it achieves an oracle complexity of O(1/\u03f5) under certain smoothness assumptions. In [87], the stochastic BLO problem with linear equality constraints in both the upperand lower-level problems is considered. The authors proposed an IF-based approach by constructing an approximate stochastic implicit gradient for linearly constrained BLO. They also proposed the Alternating Implicit Projected SGD (AiPOD) algorithm, an alternating projection method that achieves an oracle complexity of O(1/\u03f52) for both upperand lower-level objectives. The work [75] considered BLO with general constraints in both upperand lower-level objectives and non-singleton lower-level solutions. It utilized the VF-based approach developed in Bi-level Value-Function-based Interior-point Method (BVFIM) [83] for solving (NS-BLO) and proposed a Primal-Dual Bi-level Optimizer (PDBO), a primal-dual algorithm for solving (14) when the value function is approximated using (15). Under the assumptions of convex and compact constraint sets and convex lower-level objectives, PDBO achieves an oracle complexity of O(1/\u03f53/2). Recently, the authors in [76] proposed Penalty-Based Bilevel Gradient Descent (PBGD) for BLO with general constraints and non-singleton lower-level solutions. The authors established the equivalence of BLO and its penalty-based reformulations based on VF and KKT conditions. PBGD achieved an oracle complexity of O(1/\u03f53/2) for solving constrained BLO with lower-level objectives satisfying the PL inequality. Note that the algorithms PDBO [75] and PBGD [76] can be utilized to solve both (LC-BLO) and (NS-BLO) problems. Next, we discuss specific algorithms for solving (NS-BLO). 2) Theoretical results for (NS-BLO): An attempt to relax the lower-level singleton assumption for the lower-level problem was made in [45] with the introduction of Bi-level Descent Aggregation (BDA), a bi-level descent framework for solving (NS-BLO). BDA assumes convexity of the lower-level objective and strong convexity of the upper-level objective with respect to \u03d5. The framework updates the lower-level variable, \u03d5, using a convex combination of the upperand lower-level partial gradients, and then updates the upper-level variable, \u03b8, using standard FGU/BGU techniques. The authors established the asymptotic convergence of BDA in [45]. In [83], the authors relaxed the convexity assumptions on the lower and upper-level objectives in \f23 TABLE II CONVERGENCE RESULTS OF REPRESENTATIVE ALGORITHMS FOR SOLVING (LC-BLO) AND (NS-BLO) PROBLEMS. SIMILAR TO TABLE I, CONVERGENCE IS MEASURED BY ORACLE COMPLEXITY. OTHER ALGORITHMIC DETAILS, INCLUDING OPTIMIZATION PRINCIPLES (IF, GU, AND VF), PROBLEM SETUPS (LC-BLO AND NS-BLO), OBJECTIVE FUNCTION TYPES, AND LOWER-LEVEL CONSTRAINT TYPES. Algorithm Principle Objective functions Constraints Oracle Complexity Upper-Level Lower-Level BDA [45] GU Strongly-Convex Convex No Asymptotic BVFIM [83] VF Smooth Non-Convex No Asymptotic BOME! [86] VF Smooth PL No O(1/\u03f54) SIGD [74] IF Smooth Strongly-Convex Linear Inequality Asymptotic AIPOD [87] IF Smooth Strongly-Convex Linear Equality O(1/\u03f52) PDBO [75] VF Smooth Convex Non-Linear O(1/\u03f53/2) PBGD [76] VF Smooth PL Non-Linear O(1/\u03f53/2) (NS-BLO) and proposed BVFIM, a VF-based approach to solving the problem. BVFIM solves a sequence of penalty-based reformulations of the VF problem using the interior-point method, with the asymptotic convergence. In [86], the authors introduced Bi-level Optimization Made Easy! (BOME!), an alternative approach to directly solve the VF problem using a dynamic barrier gradient descent algorithm. Under the assumption of PL inequality for the lower-level objective, BOME! achieves a finite-time sample complexity of O(1/\u03f54) in the worst case. BLO-ENABLED SP AND ML APPLICATIONS In the following sections, we will showcase how BLO can be leveraged to obtain state-of-the-art results for a number of key SP and ML applications, such as wireless resource allocation (Sec. V), signal demodulation (Sec. VI), adversarial training for robustifying ML models (Sec. VII), weight pruning for enhancing model efficiency (Sec. VIII), and invariant representation learning for improving domain generalization (Sec. IX). Table III summarizes a number of emerging BLO application areas, together with some representative references. \f24 TABLE III AN OVERVIEW OF EMERGING APPLICATIONS OF BLO IN SP AND ML (INDICATES APPLICATIONS WE STUDIED). Representative Applications Application Areas Problem Description Selected References Wireless resource allocation SP To allocate wireless resources optimally and maximize their utilities [9]\u2013[12] Signal demodulation To accurately estimate transmitted symbols from received baseband signals [13], [14] Channel prediction To predict the states of a communication channel by leveraging previous observations [16]\u2013[18] Image reconstruction To recover images from their sparse measurements [19], [20] Adversarial training Robust ML To train an ML model with adversarial robustness against adversarial attacks [21], [22], [25] Poisoning attack generation To generate malicious data into the training set, creating vulnerabilities/backdoors in ML models. [23], [26], [105] Model pruning Efficient ML To find sparse sub-networks from a dense DNN w/o generalization loss [44], [106] Dataset condensation To select a subset or distill a condensed version of the training set w/o generalization loss [40]\u2013[42], [60], [107] Model-agnostic meta-learning Generalized ML To train an ML model that can quickly adapt to new tasks using limited data [28], [29], [32]\u2013[34] Invariant risk minimization To train an ML model with invariant features against distribution shift [35]\u2013[39] Neural architecture search Automated ML To automatically optimize the architecture of DNNs for improved performance [48]\u2013[50], [52]\u2013[54] Hyper-parameter optimization To optimize the hyper-parameters and model selection schemes in an ML pipeline [45]\u2013[47], [57], [78] V. BLO FOR WIRELESS RESOURCE ALLOCATION In this section, we explore the application of BLO in wireless communications, specifically in the context of wireless optimal resource allocation (power control) [9], [12]. The objective is to allocate power efficiently among multiple transmitter-receiver pairs to maximize some systemlevel performance. We consider a dynamic environment where wireless channel statistics change episodically, reflecting the variations in user behavior. This scenario is commonly observed in cellular networks, where user density and behavior vary throughout the day, such as more users in downtown areas during peak hours compared to night time. To solve the resource allocation problem, we employ a neural network trained to predict the optimal power allocation for users based on channel information. However, neural networks often struggle when evaluated on data that deviate from the training distribution, which is the case in our scenario. To address this challenge and ensure the model\u2019s performance across episodes, we adopt a continual learning framework [108]. We maintain a memory set, containing a representative subset of samples encountered so far, to facilitate adaptation to new episodes while preserving satisfactory performance on previous ones. The model is trained not only on the current batch of data but also using the current memory set. In what follows, we provide a detailed description of the problem and its formulation, following the formulation and approach presented in [12]. Formulation. Consider a dynamic wireless environment comprising T episodes, where the channel state information (CSI) statistics remain stationary within each episode. Consider a supervised learning setting, where the ith data pair (h(i), p(i)) consists of the CSI vector h(i) \f25 (the feature vector) capturing the channel characteristics, and the corresponding optimal power allocation p(i) across the users (the label). We train a neural network \u03c0(\u03b8; h(i)) on these data pairs, where \u03b8 represents the model parameters and h(i) serves as the network input, with the output being the power allocation prediction. Assume that the data arrives sequentially in multiple batches. Let Dt denote the batch we receive at time t. Assume that there is a fixed-size memory set Mt available, which stores representative historical data to be combined with Dt for training, and it is updated when a new batch arrives. The performance of a power allocation scheme p (for a given CSI h) is measured by the weighted sum-rate loss function R(p; h) [12, eq. (1)]. At each time t, our problem involves two tasks. The first task is to train the neural network on a (weighted) set of training samples, aiming to find the optimal model parameter \u03b8. The second task is to select the most representative subset from the available training data, which includes the samples in memory Mt and the current data batch Dt (denoted as Mt \u222aDt). This selected subset will be then used for training, as well as to form the new memory to be used in the next time t + 1. To achieve this, we introduce the variable \u03bb, which represents the weights associated with each sample. Higher weights are assigned to samples that are more representative or challenging, as determined by the system performance metric R(p; h). These weighted samples are then selected to form the updated memory set, and they contribute more during the model training. The idea is that by focusing on training the model on these challenging samples, we can expect better performance on the remaining easier samples. It is important to note that the model training process relies on the weights obtained from the second task. Consequently, the problem described above can be naturally formulated as the following BLO problem: minimize \u03b8 X i\u2208Mt\u222aDt \u03bb(i)(\u03b8)\u2113(\u03b8; h(i), p(i)) subject to \u03bb(\u03b8) = arg min 1T \u03bb=1 X i\u2208Mt\u222aDt \u03bb(i)R(\u03c0(\u03b8; h(i)); h(i)), , (18) where \u2113(\u03b8; h(i), p(i)) = \u2225p(i) \u2212\u03c0(\u03b8; h(i))\u22252 2 is the MSE loss over the ith sample, and Mt \u222aDt are the available training samples at time t. At the upper-level, we conduct supervised training by assigning weights to the per-sample losses based on the weights obtained from the lowerlevel task, which is similar to (T2). On the other hand, at the lower-level, we tackle the task of assigning weights to the samples based on their achieved rates, where higher weights are assigned to samples with lower rates. This is similar to (T1). Methods. Problem (18) is a constrained BLO problem with linear equality constraints w.r.t. the lower-level variable \u03bb. Based on the problem structure, and the optimization principles and algorithms introduced in Sec. III and Sec. IV-C, we utilize the IF-based SIGD (Smoothed Implicit \f26 2 4 6 8 Total Number of Training Samples (\u00d7103) 1.1 1.2 1.3 1.4 Average Sum-Rate (bit/sec.) TL SIGD 2 4 6 8 Total Number of Training Samples (\u00d7103) 1.1 1.2 1.3 1.4 Average Sum-Rate (bit/sec.) TL SIGD 2 4 6 8 Total Number of Training Samples (\u00d7103) 0.75 0.80 0.85 0.90 Average Sum-Rate (bit/sec.) TL SIGD (a) (b) (c) Fig. 2. The average sum-rate achieved on the combined test set (across all four episodes) is plotted as a function of the total no. of samples used for model training, which arrive sequentially in batches. Three experiments are conducted with different types of channel statistics across the four episodes. (a) The channel sequence is Rayleigh-RicianGeometry10-Geometry50. (b) The channel sequence is Rician-Geometry10-Rayleigh-Geometry50. (c) The channel sequence is Geometry10-Geometry20-Geometry50-Rayleigh. The numbers after the \u2018Geometry\u2019 channel indicate the spatial arrangement of the nodes, such as a 20m \u00d7 20m area for \u2018Geometry20\u2019. Gradient) method [74] to solve (18). However, it should be noted that the SIGD method assumes a strongly convex lower-level problem (see Table II). To ensure this property, we introduce a regularization term (\u03b3/2)\u2225\u03bb\u22252 2, where \u03b3 is the regularization parameter as described in Sec. III-A3. As a classical baseline approach, we also consider transfer learning (TL). In TL, when a new data batch Dt arrives, the current model trained on data up to time t \u22121 is fine-tuned using only Dt. This approach is motivated by the expectation that previous knowledge can be transferred to the new environment, enabling quick adaptation of the model. However, updating the model may result in a loss of prior knowledge, leading to performance degradation on the prior episodes. Experiment Results. We consider an experiment setting with T = 4 episodes and 10 users, and with three different types of communication channels: Rayleigh fading, Rician fading, and Geometry channels; see [12, pg. 15] for more details. The neural network trained for power allocation consists of three hidden layers with sizes 200, 80, 80, respectively. Fig. 2 illustrates the performance of power allocation, measured by the sum rate; in the horizontal axis we have the total number of samples used for model training, as these arrive sequentially in batches. Here the power allocation schemes are obtained using the BLO-based SIGD method and the baseline approach (TL), respectively. As we can see, the SIGD method exhibits smoother adaptation to each episode (note that the gray line marks the boundary between episodes). It also experiences less deterioration in performance compared to the baseline approach. These results demonstrate the advantage of using BLO for power allocation. \f27 VI. BLO FOR WIRELESS SIGNAL DEMODULATION In this section, we explore the application of BLO in wireless signal demodulation by associating it with another BLO application, model-agnostic meta-learning (MAML) [28]. Thus, we begin by introducing the fundamental concepts of MAML within the framework of BLO, and then establish the connection between MAML and signal demodulation. A. Fundamentals of BLO in MAML MAML, as an optimization-based meta-learning approach, has gained significant popularity in various fields, especially in scenarios with limited resources [28], [29], [82], [109]\u2013[112]. Specifically, MAML learns a meta-initialization of optimizee variables (e.g., model weights \u03b8) to enable fast adaptation to new tasks when fine-tuning the model from the learned initialization with only a few new data points [28], [113]. With N learning tasks {Ti}N i=1, (a) a fine-tuning set Dtr i is used in Ti for the task-specific lower-level optimization over the task-agnostic model initialization \u03b8, and (b) a validation set Dval i is used in the upper-level optimization for evaluating the fine-tuned model \u03b8\u2217 i from \u03b8. Thus, MAML can be formulated as the following BLO problem: minimize \u03b8 1 N N X i=1 E(Dtr i ,Dval i )\u223cTi \u0002 \u2113i(\u03b8\u2217 i (\u03b8); Dval i ) \u0003 subject to \u03b8\u2217 i (\u03b8) = arg min \u03b8i\u2208Rn \u2113i(\u03b8i; Dtr i , \u03b8), (MAML-BLO) where \u03b8\u2217 i (\u03b8) signifies the fine-tuned model weights using the initialization \u03b8 under the task Ti, and \u2113i denotes the model training (or validation) loss over Dtr i (or Dval i ) with initialization \u03b8 (or fine-tuned model \u03b8\u2217 i (\u03b8)). The MAML-BLO problem falls into the category of unconstrained BLO. Thereby, existing works such as [28], [29], [82], [114], [115] commonly employ the IF or GU-based approaches to solve it. The vanilla MAML algorithm [28] utilizes GU-based BLO solver, which carries out the upperand lower-level updates using the following steps: Lower: \u03b8\u2217 i (\u03b8) = \u03b8(M) i , \u03b8(m) i \u2190\u03b8(m\u22121) i \u2212\u03b2\u2207\u03b8i\u2113i(\u03b8(m\u22121) i ), m = 1, . . . , M, given \u03b8(0) i = \u03b8 Upper: \u03b8 \u2190\u03b8 \u2212\u03b1 1 N N X i=1 \u2207\u03b8\u2113i(\u03b8\u2217 i (\u03b8)), (MAML) where \u03b1, \u03b2 > 0 represent the learning rates for the SGD updates in the upper and lower level respectively. As mentioned in Sec. III-B1, the choice of the lower-level optimizer will greatly influence the GU-based BLO solvers. \f28 TABLE IV PERFORMANCE COMPARISON OF VARIOUS MAML METHODS ON THE COMMONLY USED DATASETS FOR FEW-SHOT LEARNING TASKS OMNIGLOG [116] AND MINI-IMAGENET [117]. THE BEST PERFORMANCE IN EACH SETTING IS MARKED IN BOLD. THE STANDARD DEVIATIONS ARE REPORTED BASED ON 5 RANDOM TRIALS. ROWS MARKED IN GRAY INDICATE BLO-ENABLED ALGORITHMS. Method BLO (Solver) Test Acc. (%) Time (min) Test Acc. (%) Time (min) Test Acc. (%) Time (min) Test Acc. (%) Time (min) Omniglot 20-way-1-shot Omniglot 20-way-5-shot Mini-ImageNet 5-way-1-shot Mini-ImageNet 5-way-5-shot FO-MAML [28] AO 90.62\u00b10.29 2.71 96.44\u00b10.23 2.99 46.39\u00b10.44 4.32 54.45\u00b10.29 4.93 Sign-MAML [82] GU 91.75\u00b10.26 2.85 97.79\u00b10.18 2.91 47.73\u00b10.55 4.51 55.12\u00b10.33 4.72 vanilla MAML [28] GU 95.65\u00b10.25 4.71 98.42\u00b10.23 4.94 48.77\u00b10.65 15.5 55.72\u00b10.36 15.9 iMAML [29] IF 95.99\u00b10.19 3.64 98.63\u00b10.14 3.85 49.31\u00b10.41 11.6 54.71\u00b10.27 13.3 As described in Sec. III-B2, the usage of the sign-based SGD in the lower level can lead the vanilla MAML to a first-order BLO solver, known as Sign-MAML [82]. In contrast to GUbased MAML methods discussed above, the implicit MAML (iMAML) [29] utilizes IF-based approach, where the CG (conjugate gradient) method is used to compute the inverse-Hessian gradient product. Compared to the vanilla MAML, iMAML shares the same lower-level updating rule, while adopting the IF-based upper-level iteration, similar to (5). Different from IF and GU, the first-order (FO) MAML (FO-MAML) operates by alternating between SGD-based lowerlevel optimization and SGD-based upper-level optimization, without explicitly considering the implicit gradient. We refer to this optimization procedure as AO (Alternating Optimization). We next demonstrate the effectiveness of BLO in MAML by applying it to benchmark few-shot learning tasks on the Omniglot [116] and Mini-ImageNet [117] datasets, where the generalization of the learned meta-initialization is evaluated on the new tasks, each with only a few examples. We follow the standard experimental setting [28], [29], [82], [118], considering 20-way-1-shot and 20-way-5-shot learning on Omniglot, and 5-way-1-shot and 5-way-5-shot learning on MiniImageNet. Here P-way-Q-shot refers to training a model using a small set of data points sampled from P classes, with each class containing Q examples. Tab. IV provides an overview of the accuracy and runtime efficiency of different methods for solving MAML-BLO, including AObased first-order MAML (FO-MAML) [28], GU-based Sign-MAML [82], GU-based vanilla MAML [28], and IF-based iMAML [29]. As we can see, the vanilla MAML and iMAML generally achieve higher testing accuracy than other MAML variants, but they require more computation time. This is expected for their fewer implementation assumptions, resulting in more precise meta initialization states. In the FO optimization category, we observe that Sign-MAML \f29 outperforms FO-MAML, demonstrating the advantage of using GU to solve BLO problems. B. BLO for Wireless Signal Demodulation Background and connection to MAML. We next examine the application of BLO in the context of wireless signal demodulation through the lens of MAML. Wireless signal demodulation is a critical process in the physical layer of a wireless communication system, which seeks to recover the transmitted symbols s from the received signals y. Our investigation aligns with prior research [14] that focuses on a scenario where wireless devices frequently transmit short packets consisting of a few pilot symbols through a varying channel. Notably, when working with any new wireless transmitter device in the field, only a limited number of pilot symbols are available to optimize the demodulator. At the same time, the historical data pairs for previous devices and channel conditions can hardly transfer to new ones. Consequently, we view this demodulation modeling as a few-shot meta-learning problem, aiming to derive a meta initialization from historical data that is able to quickly adapt to future devices. In this context, similar to MAML, a meta-demodulator characterized by a learnable parameter \u03b8 is trained to quickly adjust to new devices with only a handful of new pilot symbols. Formulation. Specifically, N supervised datasets are collected to train the meta-demodulator, each associated with a specific device, which can be treated as N tasks. The i-th dataset with K samples is given by Di = {(s(k) i , y(k) i )}K k=1, where y(k) i represents the k-th received signal and s(k) i is its corresponding ground-truth transmitted symbol. The demodulation task for each single device can be formulated as a classification problem, as each symbol s(k) i can only be one of the several binary encodings, e.g., ranging from 0000 to 1111 following the 16Quadrature Amplitude Modulation (16-QAM) [14]. Thus, the cross-entropy loss between the predicted transmitted symbol \u02c6 s(y, \u03b8) and the true symbol s is used to train the demodulator model: \u2113i(\u03b8; Di) = E(s (k) i ,y (k) i )\u223cDiLCE(\u02c6 s(y(k) i , \u03b8), s(k) i ), where we use a multi-layer neural network as the demodulation model \u03b8 following [14], in order to predict the transmitted symbol \u02c6 s using the received signal y. Given N devices (datasets) within the meta-training dataset, each contains K data samples, which are divided into Ktr for fine-tuning \u03b8 and Kval for validating the performance of the fine-tuned model. In line with the notions used in (MAML-BLO), the demodulation of the i-th device can be regarded as learning task Ti, which consists of a fine-tuning dataset Dtr i and a validation dataset Dval i . To this end, we can apply the previously introduced MAML methods [28], [29], [82] to address the problem of wireless signal demodulation. \f30 TABLE V PERFORMANCE COMPARISON OF DIFFERENT MAML METHODS FOR FEW-SHOT DEMODULATION OF 16-QAM MODULATED WIRELESS SIGNALS. THE BEST PERFORMANCE IN EACH SETTING IS INDICATED IN BOLD. STANDARD DEVIATIONS ARE REPORTED BASED ON 5 RANDOM TRIALS. ROWS MARKED IN GRAY INDICATE BLO-ENABLED ALGORITHMS. Method BLO (Solver) Test Acc. (%) Time (min) Test Acc. (%) Time (min) Test Acc. (%) Time (min) Test Acc. (%) Time (min) 16-way-1-shot 16-way-5-shot 16-way-10-shot 16-way-20-shot FO-MAML [28] AO 94.46\u00b10.35 8.59 97.93\u00b10.11 8.73 99.13\u00b10.13 8.87 99.17\u00b10.04 9.01 Sign-MAML [82] GU 96.91\u00b10.10 8.10 97.35\u00b10.28 8.14 97.42\u00b10.30 8.23 97.46\u00b10.24 8.32 vanilla MAML [28] GU 98.84\u00b10.10 13.04 99.49\u00b10.03 13.27 99.63\u00b10.01 13.29 99.66\u00b10.02 13.35 iMAML [29] IF 97.66\u00b10.07 14.14 98.58\u00b10.04 14.52 99.35\u00b10.04 14.81 99.43\u00b10.02 14.88 Experiment Results. During the meta-training phase, we consider N = 1000 different devices, each of which has Ktr \u2208{1, 5, 10, 20} training samples designated as the fine-tuning set. For the meta-testing phase, we use another set of 100 devices, each of which has Ktr pairs for the few-shot demodulator fine-tuning and an additional 10000 pilot symbol pairs for symbol classification accuracy evaluation. Tab. V shows the average classification accuracy and running time for different MAML methods, providing insights into how the choice of Ktr affects the performance of these methods. As we can see, MAML and iMAML achieved the highest symbol classification accuracy, although they required more computational time, consistent with the findings in Tab. IV. Sign-MAML outperformed FO-MAML in the 1-shot scenario, benefiting from the effectiveness of the GU solver. However, in scenarios with more shots, FO-MAML can achieve the performance on par with MAML. VII. BLO FOR ADVERSARIALLY ROBUST TRAINING Adversarial attacks demonstrate the vulnerability of ML models to small perturbations in their inputs, posing a challenge to their reliability [119]\u2013[121]. As a result, developing robust models that withstand such attacks has become a critical focus in the field of robust ML. In what follows, we delve into the BLO application to improve the adversarial robustness of DNNs. The lack of adversarial robustness in ML models has prompted extensive research on adversarial defense mechanisms [122]\u2013[126]. While most of existing defenses rely on min-max optimization (MMO) to minimize worst-case training loss by incorporating a synthesized adversary, this approach requires completely opposing objectives for the defender and attacker. This limits its applicability to scenarios where differing objectives are desired. Recent works [22], [25] have \f31 demonstrated the use of BLO with customizable attack objectives to improve the efficiency and robustness of robust model training across a wide range of adversarial attack strengths. Formulation. We formulate the BLO-based robust training for defending against adversarial attacks. Using variables \u03b8 for model parameters and \u03b4 for input perturbation, and loss functions \u2113tr for training and \u2113atk for attacks, we define the task of robust training (RT) as a BLO problem. The upper-level problem involves training the model \u03b8, while the lower-level problem optimizes \u03b4 for adversarial attack generation to produce worst-case input in model training. This yields: minimize \u03b8 E(x,y)\u2208D[\u2113tr(\u03b8, x + \u03b4\u2217(\u03b8; x, y), y)] subject to \u03b4\u2217(\u03b8; x, y) = arg min \u03b4\u2208C \u2113atk(\u03b4, \u03b8; x, y), (RT-BLO) where (x, y) is a data pair with feature x and label y drawn from the training dataset D, (x+\u03b4) is an adversarial example w.r.t. x, and \u03b4 \u2208C denotes a perturbation constraint, e.g., C = {\u03b4 | \u2225\u03b4\u2225\u221e\u2264 \u03f5, x + \u03b4 \u2208[0, 1]} for \u03f5-tolerated \u2113\u221e-norm constrained attack with normalized input in [0, 1]. If we choose \u2113atk = \u2212\u2113tr, then problem (RT-BLO) reduces to the MMO-based adversarial training [122]. However, the flexibility of independently choosing the lower-level attack objective allows for a broader range of robust training scenarios. In particular, it enables the development of a fast robust training variant called Fast Bi-level Adversarial Training (Fast-BAT) [22]. FastBAT formulation, given below, specifies the lower-level attack generation problem of (RT-BLO) as a constrained convex quadratic program: minimize \u03b8 E(x,y)\u2208D[\u2113CE(\u03b8, x + \u03b4\u2217(\u03b8; x, y), y)] subject to \u03b4\u2217(\u03b8; x, y) = arg min \u03b4\u2208C \u2212(\u03b4 \u2212\u03b40)\u22a4\u2207\u03b4=\u03b40\u2113CE(\u03b8, x + \u03b4, y) + (\u03b3/2)\u2225\u03b4 \u2212\u03b40\u22252 2 | {z } := \u2113atk(\u03b8, \u03b4; x, y) , (Fast-BAT) where \u2113CE(\u03b8, x, y) is the CE (cross-entropy) loss for training model weights \u03b8 evaluated at the data point (x, y), the lower-level attack objective is given by a first-order Taylor expansion of \u2212\u2113CE (at the linearization point \u03b40) plus a quadratic residual with the regularization parameter \u03b3 > 0. Since the lower-level problem becomes a convex quadratic program, it leads to the closedform projected gradient descent (PGD) solution \u03b4\u2217(\u03b8) = PC (\u03b40 \u2212(1/\u03b3)\u2207\u03b4\u2113CE(\u03b8, \u03b4)) |\u03b4=\u03b40. Methods. As problem (Fast-BAT) falls into the category of (LC-BLO), we can solve it using optimization methods introduced in Sec. IV-C. Specifically, we consider the KKT-oriented IF approach (Sec.III-A3) as our BLO solver (see Algorithm 4). In our experiments, we refer to this method as Fast-BAT-IF. Furthermore, we compare Fast-BAT-IF, with non-BLO representative robust training baselines, such as Fast-AT [125], Fast-AT-GA [127], and BackSmooth [128]. Experiment Results. In Tab. VI, we empirically show the performance of different robust training methods to robusify PreActResNet-18 [130] on the CIFAR-10 [131] and Tiny-ImageNet [132] \f32 TABLE VI PERFORMANCE COMPARISON OF DIFFERENT ROBUST TRAINING METHODS USING PREACTRESNET-18 ON CIFAR-10 AND TINY-IMAGENET DATASETS. THE TRAINING PHASE INCLUDES ADVERSARIAL PERTURBATIONS WITH TWO BUDGETS: \u03f5 = 8/255 AND 16/255 OVER 20 EPOCHS. RESULTS ARE PRESENTED AS MEAN \u00b1 STANDARD DEVIATION OVER 10 RANDOM TRIALS. ROWS MARKED IN GRAY INDICATE BLO-ENABLED ALGORITHMS. Method BLO (Solver) SA (%) (\u03f5 = 8/255) RA-PGD (%) (\u03f5 = 8/255) RA-AA (%) (\u03f5 = 8/255) SA (%) (\u03f5 = 16/255) RA-PGD (%) (\u03f5 = 16/255) RA-AA (%) (\u03f5 = 16/255) Time (s/epoch) CIFAR-10, PreActResNet-18 Fast-AT [125] N/A 82.39\u00b10.14 45.49\u00b10.21 41.87\u00b10.15 44.15\u00b17.27 21.83\u00b11.32 12.49\u00b10.33 23.1 Fast-AT-GA [127] 79.71\u00b10.24 47.27\u00b10.22 43.24\u00b10.27 58.29\u00b11.32 26.01\u00b10.16 17.97\u00b10.33 75.3 BackSmooth [129] 79.31\u00b10.17 48.06\u00b10.07 44.55\u00b10.17 64.88\u00b11.75 24.18\u00b11.37 15.47\u00b10.92 60.6 Fast-BAT-IF [22] IF 79.97\u00b10.12 48.83\u00b10.17 45.19\u00b10.12 68.16\u00b10.25 27.69\u00b10.16 18.79\u00b10.24 61.4 Tiny-ImageNet, PreActResNet-18 Fast-AT [125] N/A 41.37\u00b13.08 17.05\u00b13.25 12.31\u00b12.73 31.38\u00b10.19 5.42\u00b12.17 3.13\u00b10.24 284.6 Fast-AT-GA [127] 45.52\u00b10.24 20.39\u00b10.19 16.25\u00b10.17 29.17\u00b10.32 6.79\u00b10.27 4.27\u00b10.15 592.7 BackSmooth [129] 45.32\u00b10.24 20.49\u00b10.16 17.13\u00b10.14 31.17\u00b10.25 7.32\u00b10.21 4.71\u00b10.16 511.3 Fast-BAT-IF [22] IF 45.80\u00b10.22 21.97\u00b10.21 17.64\u00b10.15 33.78\u00b10.23 8.83\u00b10.22 5.52\u00b10.14 572.4 datasets. The evaluation metrics include \u2460the test-time robust accuracy (RA) of the learned model against 50-step PGD attacks [122] with 10 restarts (RA-PGD) using the perturbation budgets \u03f5 = 8/255 and 16/255; \u2461RA against AutoAttack (RA-AA) [133] in the setup similar to RA-PGD; \u2462the standard accuracy (SA) of the learned model on natural examples; \u2463the time consumption required for robust training. As observed, Fast-BAT-IF exhibits higher robustness compared to non-BLO baselines, highlighting the effectiveness of BLO in robust training. VIII. BLO FOR MODEL PRUNING While over-parameterized structures are key to the improved generalization of DNNs, they create new problems the millions or even billions of parameters not only increase computational costs during inference but also pose serious deployment challenges on resource-limited devices. Thus, the problem of model pruning arises, aiming to reduce the sizes of an ML model by identifying and removing redundant model weights. In this section, we investigate the application of BLO in the context of model pruning [44], [134]\u2013[136]. Formulation. The study of model pruning through the lens of BLO was first explored in [44]. Specifically, there exist two main tasks in model pruning: pruning and re-training. Pruning involves determining the sparse pattern of model weights, while re-training focuses on recov\f33 ering model accuracy using the remaining non-zero weights [135], [136]. To facilitate these tasks, one can introduce the binary pruning mask variable m \u2208{0, 1}m and the model weight variable \u03d5 \u2208Rm, where m represents the total number of model parameters. Accordingly, the pruned model is given by (m \u2299\u03d5), where \u2299denotes element-wise multiplication. To achieve a pruning ratio of p%, we impose a sparsity constraint on m, where m \u2208\u2126and \u2126= {m, |, m \u22080, 1n, 1T m \u2264k}, with k = (1 \u2212p%)n. Our goal is to prune the original dense model to the targeted pruning ratio of p% and obtain the optimal sparse model (m \u2299\u03d5). To achieve this, we interpret the pruning task (\u2460) and the model retraining task (\u2461) as two optimization levels, leading to the formulation of bi-level pruning (BiP): minimize m\u2208\u2126 \u2113tr(m \u2299\u03d5\u2217(m)) | {z } \u2460: Pruning task ; subject to \u03d5\u2217(m) = arg min \u03d5\u2208Rn \u02dc \u2113tr(m \u2299\u03d5) + \u03b3 2 \u2225\u03d5\u22252 2 | {z } \u2461: Sparsity-fixed model retraining , (BiP) where \u2113tr and \u02dc \u2113tr denote the training losses under different data batches, m and \u03d5 are the upperlevel and lower-level optimization variables respectively, and \u03d5\u2217(m) signifies the retrained model weights given the pruning mask m. In (BiP), the lower-level training objective was regularized using a strongly convex regularizer \u03b3 2\u2225\u03d5\u22252 2 like (Fast-BAT). Methods. Since BiP is an unconstrained BLO problem, it can be solved using BLO algorithms, e.g., IF and GU, introduced in Secs. III-A1 and III-B1. Moreover, since the Hessian of the lowerlevel objective function w.r.t. model parameters is of high dimension, we impose the Hessian-free assumption \u2207\u03d5,\u03d5\u2113tr = 0 to make the BLO implementation computationally feasible. Following (4), one can then obtain the closed form of IG [44]: d\u03d5\u2217(m) dm = \u22121 \u03b3 \u22072 m,\u03d5\u2113tr(m \u2299\u03d5\u2217), where \u03d5\u2217signifies a lower-level solution. Furthermore, the bi-linearity of the pruning mask m and the model weights \u03d5 allows us to further simplify the IG to: d\u03d5\u2217(m) dm = \u22121 \u03b3 diag(\u2207z\u2113tr(z) |z=m\u2299\u03d5\u2217), (19) where the Hessian-free assumption is adopted and diag(a) denotes the diagonal matrix with a being the main diagonal vector. Detailed proof can be found in [44, Sec. 3]. Experiments Results. To implement BiP, we adopt two BLO methods: IF (Sec. III-A1) and GU (Sec. III-B1). We term the resulting BLO-inspired model pruning approaches as BiP-IF and BiPGU. For comparison, we also consider two commonly-used non-BLO based pruning methods, the state-of-the-art iterative magnitude pruning (IMP) [136] and the most efficient one-shot magnitude pruning (OMP) [137]. We remark that the notable lottery ticket hypothesis (LTH) [136] stated \f34 0 20 40 60 80 100 Pruning Ratio (%) 93.5 94.0 94.5 95.0 95.5 Test Accuracy (%) Dense Model IMP OMP BiP-GU BiP-IF Best Winning Ticket 0 20 40 60 80 100 Pruning Ratio (%) 75.0 75.5 76.0 76.5 77.0 77.5 78.0 Test Accuracy (%) Dense Model IMP OMP BiP-GU BiP-IF Best Winning Ticket 0 20 40 60 80 100 Pruning Ratio (%) 60 61 62 63 64 Test Accuracy (%) Dense Model IMP OMP BiP-GU BiP-IF Best Winning Ticket 20 40 60 80 100 Pruning Ratio (%) 10 2 10 3 Time Consumption (min) Dense Model IMP OMP BiP-GU BiP-IF (a) CIFAR-10 (b) CIFAR-100 (c) Tiny-ImageNet (d) Efficiency comparison Fig. 3. Experiment results of model pruning on different datasets under ResNet-18. (a-c) Pruning trajectory is given by test accuracy (%) vs. sparsity (%) under different datasets. (d) The entire time consumption vs. the pruning ratio. that IMP is able to identify a trainable sparse subnetwork (known as a \u2018winning ticket\u2019), with test accuracy surprisingly on par or even better than the original model. Fig. 3 illustrates the pruning accuracy and the run-time efficiency of BLO-based pruning methods vs. non-BLO approaches across diverse image classification datasets (including CIFAR10, CIFAR-100, and Tiny-ImageNet) under ResNet-18. As we can see, BiP-IF yields the best performance in all the dataset settings evidenced in Fig. 3(a)-(c). This also the pruning recipe used in [44]. Thanks to the closed-form expression of IG in (19), the computation of BiP-IF is also more efficient than BiP-GU, as shown in Fig. 3(d). In addition, BiP-GU can also provide competitive pruning accuracy to IMP and takes less computation time than IMP. Furthermore, we observe that OMP yields the least computation time but the worst pruning accuracy. This is not surprising since OMP adopts a non-iterative pruning scheme to find the model\u2019s sparse pattern. IX. BLO FOR INVARIANT REPRESENTATION LEARNING In this section, we explore the application of BLO in improving the generalization capabilities of ML models. Specifically, we investigate the use of BLO for acquiring training environmentagnostic data representations through invariant risk minimization (IRM) [35]. The goal is to enhance the domain generalization of ML models. Formulation. IRM [35] is proposed to acquire invariant data representations and to enforce invariant predictions against distribution shifts. Different from the conventional environment risk minimization (ERM)-based training, IRM yields the BLO-like formulation: The upper-level optimization task of IRM is to train a network backbone to capture environment-agnostic data representations, and the lower-level optimization task is to find an invariant prediction head (one top of the learned representation network) to produce a global optima to all the training \f35 environments. Formally, IRM can be cast as minimize \u03b8 E X i=1 \u2113i(\u03d5\u2217(\u03b8) \u25e6\u03b8) subject to \u03d5\u2217(\u03b8) \u2208arg min \u03d5 \u2113i(\u03d5 \u25e6\u03b8), (IRM) where \u03d5 \u25e6\u03b8 denotes the representation-acquisition model \u03b8 and the predictor \u03d5, \u2113i is the training loss associated with the ith training environment, and E is the total number of training environments. The rationale behind (IRM) is that given the invariant representation extractor \u03b8, there exists an invariant predictor \u03d5\u2217(\u03b8) which is optimal across all the training environments. Methods. Solving problem (IRM) is highly non-trivial since the lower-level solution \u03d5(\u03b8) should be universal across all E training environments. To circumvent this difficulty, IRM is typically relaxed to a single-level optimization problem, known as IRMv1 [35]: minimize \u03b8 E X i=1 [\u2113i(\u03b8) + \u03b3\u2225\u2207w|w=1.0\u2113i(w \u25e6\u03b8)\u22252 2] (20) where \u03b3 > 0 is a regularization parameter and \u2207w|w=1.0\u2113i(w \u2299\u03b8) denotes the gradient of \u2113i w.r.t. w, computed at w = 1.0. In the formulation above, the identity mapping w = 1.0 is adopted, symbolizing a basic \u201cimaginary\u201d classification head. Meanwhile, \u03b8 corresponds to the combination of the representation extractor and the actual invariant predictor. However, such a simplification is restricted to linear invariant prediction and penalizes the deviation of individual environment losses from stationarity to approach the lower-level optimality in IRM. Despite the practical simplicity of IRMv1, it may fail to achieve the desired invariance [138], [139]. Beyond IRMv1, a consensus-constrained BLO method is developed in [140] to solve problem (IRM). The key idea is to introduce E auxiliary predictors {\u03d5i} and explicitly enforce prediction invariance by infusing a consensus prediction constraint C = {{\u03d5i}|\u03d51 = . . . = \u03d5E} to the lower-level problem of (IRM), and promote the per-environment stationarity in its upper-level problem. This modifies (IRM) to an ordinary BLO problem: minimize \u03b8 PE i=1[\u2113i(\u03d5\u2217 i (\u03b8) \u25e6\u03b8) + \u03b3\u2225\u2207\u03d5i\u2113i(\u03d5\u2217 i (\u03b8) \u25e6\u03b8)\u22252 2] subject to {\u03d5\u2217 i (\u03b8)}E i=1 = arg min {\u03d5i}\u2208C E X i=1 \u2113i(\u03d5i \u25e6\u03b8), \u2200i \u2208[E], (IRM-BLO) where \u03b3 > 0 is a regularization parameter, and [E] denotes the integer set {1, 2, . . . , E}. The advantage of converting (IRM) into the consensus-constrained (IRM-BLO) is that projection onto the consensus constraint yields a closed-form solution, i.e., PC(a) = arg min{\u03d5i}\u2208C PE i=1 \u2225\u03d5i \u2212 ai\u22252 2 = 1 E P i ai, where PC(a) denotes the projection operation to project the point a onto the constraint C. It has been shown in [140] that problem (IRM-BLO) can be effectively solved \f36 using the GU approach, which approximates each individual lower-level solution using K-step GD unrolling together with the consensus projection. Thus, the lower-level solution becomes \u2200i, \u03d5\u2217 i (\u03b8) \u22481 E E X i=1 \u03d5(K) i , \u03d5(k) i = \u03d5(k\u22121) i \u2212\u03b2\u2207\u03d5i\u2113i(\u03d5(k\u22121) i \u25e6\u03b8), for k \u2208[K], (21) where \u03b2 > 0 is the lower-level learning rate. Based on the above, automatic differentiation (AD) can be called to compute the implicit gradient from \u03d5(K) i (\u03b8) to the variable \u03b8 in the GU process. TABLE VII PERFORMANCE OF DIFFERENT IRM TRAINING METHODS. THE BEST PERFORMANCE PER EVALUATION METRIC IS HIGHLIGHTED IN BOLD AND THE PERFORMANCE OF THE BLO-ENABLED METHOD IS MARKED IN GRAY . Method BLO (Solver) Colored-MNIST Colored-FashionMNIST Avg. Acc. Acc. Gap Avg. Acc. Acc. Gap ERM N/A 49.19\u00b11.89 90.72\u00b12.08 49.77\u00b11.71 88.62\u00b12.49 IRM-v1 [35] N/A 68.33\u00b10.31 2.04\u00b10.05 68.76\u00b10.31 1.45\u00b10.09 IRM-Game [36] N/A 67.73\u00b10.24 1.67\u00b10.14 67.49\u00b10.32 1.82\u00b10.13 IRM-BLO [140] GU 69.47\u00b10.24 1.04\u00b10.07 69.43\u00b10.21 1.14\u00b10.11 Experiment Results. We evaluate the performance of IRM-BLO with two commonly-used image classification datasets, Colored-MNIST [35] and Colored-FashionMNIST [36], where spurious correlation is imposed in the datasets which makes the conventional ERM training ineffective. To capture both the accuracy and the variance of invariant predictions across multiple testing environments, the average accuracy and the accuracy gap (the difference between best and worstcase accuracy) are measured for IRM methods. Tab. VII presents the resulting performance of IRM-BLO and compares it with that of ERM and two IRM baselines, IRMv1 [35]and IRMGame [36]. Note that all IRM variants outperform ERM, which justifies the importance of IRM training to improve model generalization across diverse environments. Within the IRM training family, IRM-BLO outperforms others by achieving the highest average accuracy and the smallest accuracy gap across both datasets. This superior performance underscores the value of BLO compared to the sub-optimal design IRM-v1. X. DISCUSSION BLO is a challenging but rapidly developing subject. Despite recent progresses discussed in this article, significant work is yet to be done to address various challenges ranging from developing scalable algorithms, to extending its applicability to a wider range of problems. Below we highlight several worthwhile future directions. \f37 \u2022 BLO algorithms. First, the development and analysis of BLO algorithms for more general BLO problems, including those with complex lower-level constraints (e.g., non-linear constraints), require additional exploration. The current focus has predominantly been on problems with linear constraints, and extending the BLO framework to handle non-linear constraints is an important and challenging task; some recent work towards this direction in the context of the min-max problems can be found in [141]. Additionally, exploring scenarios with coupled constraints between lower and upper levels, such as resource sharing among adversarial and normal agents, presents a complex area that is under-explored. Second, BLO formulations with non-singleton lowerlevel solutions, such as (NS-BLO), lack theoretically grounded, scalable, and easy-to-implement algorithms. This aspect of BLO has received less attention from the community, making it an open topic for future investigation. Developing efficient algorithms that can handle non-singleton lowerlevel solutions and provide convergence guarantees is a key research direction. Third, beyond the scope of bi-level optimization, exploring problems involving more than two levels, such as dataset pruning for transfer learning, represents an exciting frontier. These multi-level problems introduce additional complexity and challenges, requiring the design of novel algorithms to tackle the inherent hierarchical structure effectively. Finally, large-scale and distributed availability of data demands the development of decentralized and federated algorithms to solve these complex BLO problems which also presents a compelling research direction for future exploration. \u2022 BLO theories. First, while significant progress has been made in establishing theoretical guarantees for solving the basic BLO problem (LU-BLO), more attention is needed on exploring practical settings including (coupled) lower-level constraints, non-convex lower-level problems, and/or black-box settings where one may not have access to upper/lower-level parameters. These scenarios present unique challenges and complexities, making it difficult to analyze the problem and derive theoretical guarantees. Investigating the convergence properties and establishing theoretical foundations for solving BLO problems under these practical settings is an important avenue for future research. Second, when datasets become massive, it is crucial to develop theoretically-grounded BLO algorithms that can adhere to practical requirements. Therefore, developing theoretical frameworks and analyzing the convergence properties of algorithms for solving large-scale BLO problems under realistic assumptions is also an important research topic. Third, with the discovery of the phenomenon of double descent [142], theoretical analysis of standard ML algorithms on overparameterized neural networks has received significant attention from the research community. Theoretical investigation of BLO algorithms for such overparam\f38 eterized problems is certainly an interesting research direction. Finally, developing a thorough theoretical understanding of distributed/federated algorithms for solving complex BLO problems will certainly be at the forefront of future BLO research. \u2022 BLO applications. First, in the context of Mixture-of-Experts (MoE) training, there is a complex interplay between the training of the gating network that selects experts and the training of the expert-oriented pathways used for final predictions. Exploring BLO techniques to effectively optimize the coupling between these two training processes in MoE can lead to improved performance and better utilization of emerging ML models like MoE. Second, prompt learning, a key technique used in today\u2019s foundation models, involves a crucial coupling between prompt pattern learning and label/feature mapping optimization [143]. Leveraging BLO methods to model and optimize the interactions between prompt pattern learning and label/feature mapping can enhance the learning process and enable accurate and robust prompt generations. Third, BLO is highly applicable in (inverse) reinforcement learning. For instance, the Actor/Critic Algorithm can be formulated as a BLO problem, with separate agents evaluating and optimizing the policy. In inverse reinforcement learning, the tasks involve inferring the agents\u2019 reward function and finding the optimal policy based on it. Applying BLO frameworks to these scenarios offers potential for novel insights and improved efficiency. In summary, the interplay between the theoretical underpinnings of BLO and its practical applications promises a fertile ground for future exploration and innovation, pushing the boundaries of optimization theory and applications in SP and ML. ACKNOWLEDGEMENT The work of Y. Zhang, Y. Yao, and S. Liu was supported by the NSF Grant IIS-2207052. \f39"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.02343v1",
+                    "title": "What Is Missing in IRM Training and Evaluation? Challenges and Solutions",
+                    "abstract": "Invariant risk minimization (IRM) has received increasing attention as a way\nto acquire environment-agnostic data representations and predictions, and as a\nprincipled solution for preventing spurious correlations from being learned and\nfor improving models' out-of-distribution generalization. Yet, recent works\nhave found that the optimality of the originally-proposed IRM optimization\n(IRM) may be compromised in practice or could be impossible to achieve in some\nscenarios. Therefore, a series of advanced IRM algorithms have been developed\nthat show practical improvement over IRM. In this work, we revisit these recent\nIRM advancements, and identify and resolve three practical limitations in IRM\ntraining and evaluation. First, we find that the effect of batch size during\ntraining has been chronically overlooked in previous studies, leaving room for\nfurther improvement. We propose small-batch training and highlight the\nimprovements over a set of large-batch optimization techniques. Second, we find\nthat improper selection of evaluation environments could give a false sense of\ninvariance for IRM. To alleviate this effect, we leverage diversified test-time\nenvironments to precisely characterize the invariance of IRM when applied in\npractice. Third, we revisit (Ahuja et al. (2020))'s proposal to convert IRM\ninto an ensemble game and identify a limitation when a single invariant\npredictor is desired instead of an ensemble of individual predictors. We\npropose a new IRM variant to address this limitation based on a novel viewpoint\nof ensemble IRM games as consensus-constrained bi-level optimization. Lastly,\nwe conduct extensive experiments (covering 7 existing IRM variants and 7\ndatasets) to justify the practical significance of revisiting IRM training and\nevaluation in a principled manner.",
+                    "authors": "Yihua Zhang, Pranay Sharma, Parikshit Ram, Mingyi Hong, Kush Varshney, Sijia Liu",
+                    "published": "2023-03-04",
+                    "updated": "2023-03-04",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.AI"
+                    ],
+                    "main_content": "INTRODUCTION Deep neural networks (DNNs) have enjoyed unprecedented success in many real-world applications (He et al., 2016; Krizhevsky et al., 2017; Simonyan & Zisserman, 2014; Sun et al., 2014). However, experimental evidence (Beery et al., 2018; De Haan et al., 2019; DeGrave et al., 2021; Geirhos et al., 2020; Zhang et al., 2022b) suggests that DNNs trained with empirical risk minimization (ERM), the most commonly used training method, are prone to reproducing spurious correlations in the training data (Beery et al., 2018; Sagawa et al., 2020). This phenomenon causes performance degradation when facing distributional shifts at test time (Gulrajani & Lopez-Paz, 2020; Koh et al., 2021; Wang et al., 2022; Zhou et al., 2022a). In response, the problem of invariant prediction arises to enforce the model trainer to learn stable and causal features (Beery et al., 2018; Sagawa et al., 2020). In pursuit of out-of-distribution generalization, a new model training paradigm, termed invariant risk minimization (IRM) (Arjovsky et al., 2019), has received increasing attention to overcome the shortcomings of ERM against distribution shifts. In contrast to ERM, IRM aims to learn a universal representation extractor, which can elicit an invariant predictor across multiple training environments. However, different from ERM, the learning objective of IRM is highly non-trivial to optimize in practice. Speci\ufb01cally, IRM requires solving a challenging bi-level optimization (BLO) problem with a hierarchical learning structure: invariant representation learning at the upper-level and invariant predictive modeling at the lower-level. Various techniques have been developed to solve IRM effectively, such as (Ahuja et al., 2020; Lin et al., 2022; Rame et al., 2022; Zhou et al., 2022b) to name a few. Despite the proliferation of IRM advancements, several issues in the theory and practice have also appeared. For example, recent works (Rosenfeld et al., 2020; Kamath et al., 1 arXiv:2303.02343v1  [cs.LG]  4 Mar 2023 \fPublished as a conference paper at ICLR 2023 2021) revealed the theoretical failure of IRM in some cases. In particular, there exist scenarios where the optimal invariant predictor is impossible to achieve, and the IRM performance may fall behind even that of ERM. Practical studies also demonstrate that the performance of IRM rely on multiple factors, e.g., model size (Lin et al., 2022; Zhou et al., 2022b), environment dif\ufb01culty (Dranker et al., 2021; Krueger et al., 2021), and dataset type (Gulrajani & Lopez-Paz, 2020). Therefore, key challenges remain in deploying IRM to real-world applications. In this work, we revisit recent IRM advancements and uncover and tackle several pitfalls in IRM training and evaluation, which have so far gone overlooked. We \ufb01rst identify the large-batch training issue in existing IRM algorithms, which prevents escape from bad local optima during IRM training. Next, we show that evaluation of IRM performance with a single test-time environment could lead to an inaccurate assessment of prediction invariance, even if this test environment differs signi\ufb01cantly from training environments. Based on the above \ufb01ndings, we further develop a novel IRM variant, termed BLOC-IRM, by interpreting and advancing the IRM-GAME method (Ahuja et al., 2020) through the lens of BLO with Consensus prediction. Below, we list our contributions (\u0082-\u0085). \u0082 We demonstrate that the prevalent use of large-batch training leaves signi\ufb01cant room for performance improvement in IRM, something chronically overlooked in the previous IRM studies with benchmark datasets COLORED-MNIST and COLORED-FMNIST. By reviewing and comparing with 7 state-of-the-art (SOTA) IRM variants (Table 1), we show that simply using small-batch training improves generalization over a series of more involved large-batch optimization enhancements. \u0083 We also show that an inappropriate evaluation metric could give a false sense of invariance to IRM. Thus, we propose an extended evaluation scheme that quanti\ufb01es both precision and \u2018invariance\u2019 across diverse testing environments. \u0084 Further, we revisit and advance the IRM-GAME approach (Ahuja et al., 2020) through the lens of consensus-constrained BLO. We remove the need for an ensemble (one per training environment) of predictors in IRM-GAME by proposing BLOC-IRM (BLO with Consensus IRM), which produces a single invariant predictor. \u0085 Lastly, we conduct extensive experiments (on 7 datasets, using diverse model architectures and training environments) to justify the practical signi\ufb01cance of our \ufb01ndings and methods. Notably, we conduct experiments on the CELEBA dataset as a new IRM benchmark with realistic spurious correlations. We show that BLOC-IRM outperforms all baselines in nearly all settings. 1.1 RELATED WORK IRM methods. Inspired by the invariance principle (Peters et al., 2016), Arjovsky et al. (2019) de\ufb01ne IRM as a BLO problem, and develop a relaxed single-level formulation, termed IRMV1, for ease of training. Recently, there has been considerable work to advance IRM techniques. Examples of IRM variants include penalization on the variance of risks or loss gradients across training environments (Chang et al., 2020; Krueger et al., 2021; Rame et al., 2022; Xie et al., 2020; Xu & Jaakkola, 2021; Xu et al., 2022), domain regret minimization (Jin et al., 2020), robust optimization over multiple domains (Xu & Jaakkola, 2021), sparsity-promoting invariant learning (Zhou et al., 2022b), Bayesian inference-baked IRM (Lin et al., 2022), and ensemble game over the environmentspeci\ufb01c predictors (Ahuja et al., 2020). We refer readers to Section 2 and Table 1 for more details on the IRM methods that we will focus on in this work. Despite the potential and popularity of IRM, some works have also shown the theoretical and practical limitations of current IRM algorithms. Speci\ufb01cally, Chen et al. (2022); Kamath et al. (2021) show that invariance learning via IRM could fail and be worse than ERM in some two-bit environment setups on COLORED-MNIST, a synthetic benchmark dataset often used in IRM works. The existence of failure cases of IRM is also theoretically shown by Rosenfeld et al. (2020) for both linear and non-linear models. Although subsequent IRM algorithms take these failure cases into account, there still exist huge gaps between theoretically desired IRM and its practical variants. For example, Lin et al. (2021; 2022); Zhou et al. (2022b) found many IRM variants incapable of maintaining graceful generalization on large and deep models. Moreover, Ahuja et al. (2021); Dranker et al. (2021) demonstrated that the performance of IRM algorithms could depend on practical details, e.g., dataset size, sample ef\ufb01ciency, and environmental bias strength. The above IRM limitations in2 \fPublished as a conference paper at ICLR 2023 spire our work to study when and how we can turn the IRM advancements into effective solutions, to gain high-accuracy and stable invariant predictions in practical scenarios. Domain generalization. IRM is also closely related to domain generalization (Carlucci et al., 2019; Gulrajani & Lopez-Paz, 2020; Koh et al., 2021; Li et al., 2019; Nam et al., 2021; Wang et al., 2022; Zhou et al., 2022a). Compared to IRM, domain generalization includes a wider range of approaches to improve prediction accuracy against distributional shifts (Beery et al., 2018; Jean et al., 2016; Koh et al., 2021). For example, an important line of research is to improve representation learning by encouraging cross-domain feature resemblance (Long et al., 2015; Tzeng et al., 2014). The studies on domain generalization have also been conducted across different learning paradigms, e.g., adversarial learning (Ganin et al., 2016), self-supervised learning (Carlucci et al., 2019), and meta-learning (Balaji et al., 2018; Dou et al., 2019). 2 PRELIMINARIES AND SETUP In this section, we introduce the basics of IRM and provide an overview of our IRM case study. IRM formulation. In the original IRM framework Arjovsky et al. (2019), consider a supervised learning paradigm, with datasets {D(e)}e\u2208Etr collected from N training environments Etr = {1, 2, . . . , N}. The training samples in D(e) (corresponding to the environment e) are of the form (x, y) \u2208X \u00d7 Y, where X and Y are, respectively, the raw feature space and the label space. IRM aims to \ufb01nd an environment-agnostic data representation \u03c6\u03b8 : X \u2192Z, which elicits an invariant prediction fw : Z \u2192Y that is simultaneously optimal for all environments. Here \u03b8 and w denote model parameters to be learned, and Z denotes the representation space. Thus, IRM yields an invariant predictor fw \u25e6\u03c6\u03b8 : X \u2192Y that can generalize to unseen test-time environments {D(e)}e/ \u2208Etr. Here \u25e6denotes function composition, i.e., fw \u25e6\u03c6\u03b8(\u00b7) = fw(\u03c6\u03b8(\u00b7)). We will use w\u25e6\u03b8 as a shorthand for fw \u25e6\u03c6\u03b8. IRM constitutes the following BLO problem: minimize \u03b8 P e\u2208Etr \u2113(e)(w\u2217(\u03b8) \u25e6\u03b8); subject to w\u2217(\u03b8) \u2208arg min \u00af w \u2113(e)( \u00af w \u25e6\u03b8), \u2200e \u2208Etr, (IRM) where \u2113(e)(w\u25e6\u03b8) is the per-environment training loss of the predictor w\u25e6\u03b8 under D(e). Clearly, IRM involves two optimization levels that are coupled through the lower-level solution w\u2217(\u03b8). Achieving the desired invariant prediction requires the solution sets of the individual lower-level problems {arg min \u00af w \u2113(e)( \u00af w \u25e6\u03b8), e \u2208Etr} to be non-singleton. However, BLO problems with non-singleton lower-level solution sets are signi\ufb01cantly more challenging (Liu et al., 2021). To circumvent this dif\ufb01culty, Arjovsky et al. (2019) relax (IRM) into a single-level optimization problem (a.k.a., IRMv1): minimize \u03b8 P e\u2208Etr[\u2113(e)(\u03b8) + \u03b3\u2225\u2207w|w=1.0\u2113(e)(w \u25e6\u03b8)\u22252 2], (IRMv1) where \u03b3 > 0 is a regularization parameter and \u2207w|w=1.0\u2113(e) denotes the gradient of \u2113(e) with respect to w, computed at w = 1.0. Compared with IRM, IRMv1 is restricted to linear invariant predictors, and penalizes the deviation of individual environment losses from stationarity to approach the lowerlevel optimality in (IRM). IRMv1 uses the fact that a scalar predictor (w = 1.0) is equivalent to a linear predictor. Despite the practical simplicity of (IRMv1), it may fail to achieve the desired invariance (Chen et al., 2022; Kamath et al., 2021). Case study of IRM methods. As illustrated above, the objective of IRM is dif\ufb01cult to optimize, while IRMv1 only provides a sub-optimal solution. Subsequent advances have attempted to reduce this gap. In this work, we focus on 7 popular IRM variants and evaluate their invariant prediction performance over 7 datasets. Table 1 and Table 2 respectively summarize the IRM methods and the datasets considered in this work. We survey the most representative and effective IRM variants in the literature, which will also serve as our baselines in performance comparison. Following Table 1, we \ufb01rst introduce the IRMV0 variant, a generalization of IRMv1, by relaxing its assumption of linearity of the predictor w, yielding minimize w,\u03b8 P e\u2208Etr[\u2113(e)(w \u25e6\u03b8) + \u03b3\u2225\u2207w\u2113(e)(w \u25e6\u03b8)\u22252 2]. (IRMv0) Next, we consider the risk extrapolation method REX (Krueger et al., 2021), an important baseline based on distributionally robust optimization for group shifts (Sagawa et al., 2019). Furthermore, 3 \fPublished as a conference paper at ICLR 2023 Table 1: Summary of the 7 existing IRM variants considered in this work, and the proposed BLOC-IRM method (see Section 5). We also list the 7 benchmark datasets used to evaluate IRM performance, namely, COLORED-MNIST (CoM), COLORED-FMNIST (CoF), CIFAR-MNIST (CiM), COLORED-OBJECT (CoO), CELEBA (CA), PACS (P) and VLCS (A). The symbols \u2018\u0013\u2019 signi\ufb01es the dataset used in the speci\ufb01c reference. IRM Method Venue Datasets Reference CoM CoF CiM CoO CA P V IRMV1 arXiv \u0013 (Arjovsky et al., 2019) IRMV0 N/A This Work IRM-GAME ICML \u0013 \u0013 (Ahuja et al., 2020) REX ICML \u0013 \u0013 \u0013 \u0013 (Krueger et al., 2021) BIRM CVPR \u0013 \u0013 \u0013 (Lin et al., 2022) SPARSEIRM ICML \u0013 \u0013 \u0013 (Zhou et al., 2022b) FISHR ICML \u0013 \u0013 \u0013 (Rame et al., 2022) Ours N/A \u0013 \u0013 \u0013 \u0013 \u0013 \u0013 \u0013 This Work Table 2: Dataset setups. \u2018Invariant\u2019 and \u2018Spurious\u2019 represent the core and spurious features. \u2018Env1\u2019 and \u2018Env2\u2019 are environments with different spurious correlations. Dataset Invariant Spurious Env 1 Env 2 CoM Digit Color CoF Object Color CiM CIFAR MNIST CoO Object Color CA Smiling Hair Color P Object Texture V Object Environment inspired by the empirical \ufb01ndings that the performance of IRM could be sensitive to model size (Choe et al., 2020; Gulrajani & Lopez-Paz, 2020), we choose the SOTA methods Bayesian IRM (BIRM) (Lin et al., 2022) and sparse IRM (SPARSEIRM) (Zhou et al., 2022b), both of which show improved performance with large models. Also, we consider the SOTA method FISHR (Rame et al., 2022), which modi\ufb01es IRM to penalize the domain-level gradient variance in single-level risk minimization. FISHR provably matches both domain-level risks and Hessians. Lastly, we include IRM-GAME (Ahuja et al., 2020) as a special variant of IRM. Different from the other methods which seek an invariant predictor, IRM-GAME endows each environment with a predictor, and leverages this ensemble of predictors to achieve invariant representation learning. This is in contrast to other existing works which seek an invariant predictor. Yet, we show in Section 5 that IRM-GAME can be interpreted through the lens of consensus-constrained BLO and generalized for invariant prediction. We also highlight that diverse datasets are considered in this work (see Table 2) to benchmark IRM\u2019s performance. More details on dataset selections can be found in Appendix A. 3 LARGE-BATCH TRAINING CHALLENGE AND IMPROVEMENT In this section, we demonstrate and resolve the large-batch training challenge in current IRM implementations (Table 1). 64 256 1k 2k 4k 50k Batch Size 64 65 66 67 68 69 70 Invariant Accuracy (%) IRM-V1 IRM-V0 REx Figure 1: The performance of three IRM methods (IRMV1, IRMV0, and REX) vs. batch-size under COLOREDMNIST. The full batch-size is 50k. Large-batch optimization causes instabilities of IRM training. Using very large-size batches for model training can result in the model getting trapped near a bad local optima (Keskar et al., 2016). This happens as a result of the lack of stochasticity in the training process, and is known to exist even in the ERM paradigm (Goyal et al., 2017; You et al., 2017a). Yet, nearly all the existing IRM methods follow the training setup of IRMV1 (Arjovsky et al., 2019), which used the full-batch gradient descent (GD) method rather than the mini-batch stochastic gradient descent (SGD) for IRM training over COLORED-MNIST and COLORED-FMNIST. In the following, we show that large-batch training might give a false impression of the relative ranking of IRM performances. We start with an exploration of the impact of batch size on the invariant prediction accuracy of existing IRM methods under COLORED-MNIST. Here the invariant prediction accuracy refers to the averaged accuracy of the invariant predictor applied to diverse test-time environments. We defer its formal description to Section 4. Figure 1 shows the invariant prediction accuracy of three IRM methods IRMV1, IRMV0, and REX vs. the data batch size (see Figure A1 for results of other IRM variants and Figure A5 for COLORED-FMNIST). Recall that the full batch size (50k) was used in the existing IRM implementations (Arjovsky et al., 2019; Krueger et al., 2021). As we can see, in the full-batch setup, IRM methods lead to widely different invariant prediction accuracies, where REX and IRMV1 signi\ufb01cantly outperform IRMV0. In contrast, in the small-batch case (with size 4 \fPublished as a conference paper at ICLR 2023 1k), the discrepancy in accuracy across methods vanishes. We see that IRMv0 can be as effective as IRMv1 and other IRM variants (such as REX) only if an appropriate small batch size is used. Empirical evidence in Figure 1 shows that large-batch IRM training is less effective than smallbatch. This is aligned with the observations in ERM (You et al., 2017b; 2018; 2019), where the lack of stochasticity makes the optimizer dif\ufb01cult to escape from a sharp local minimum. We also justify this issue by visualizing the loss landscapes in Figure A2. Notably, the small-batch training enables IRMV1 to converge to a local optimum with a \ufb02at loss landscape, indicating better generalization (Keskar et al., 2016). Small-batch training is effective versus a zoo of large-batch optimization enhancements. To mitigate the large-batch IRM training issue, we next investigate the effectiveness of both smallbatch training and a zoo of large-batch optimization enhancements. Inspired by large-batch training techniques to scale up ERM, we consider Large-batch SGD (LSGD) (Goyal et al., 2017) and Layerwise Adaptive Learning Rate (LALR) (You et al., 2017b; 2018; 2019; Zhang et al., 2022a). Both methods aim to smoothen the optimization trajectory by improving either the learning rate scheduler or the quality of initialization. Furthermore, we adopt sharpness-aware minimization (SAM) (Foret et al., 2020) as another possible large-batch training solution to explicitly penalize the sharpness of the loss landscape. We integrate the above optimization techniques with IRM, leading to the variants IRM-LSGD, IRM-LALR, and IRM-SAM. See Appendix B.1 for more details. Table 3: Prediction accuracy of IRM methods on COLORED-MNIST using the original large-batch implementation (\u2018Original\u2019), the large-batch optimization-integrated implementations (\u2018LSGD/LALR/SAM\u2019), and the small-batch training recipe (\u2018Small\u2019). Method Original LSGD LALR SAM Small IRMV1 67.13 67.31 67.44 67.79 68.33 IRMV0 65.39 66.42 66.76 66.99 68.37 IRM-GAME 65.69 65.82 65.47 66.23 67.73 REX 67.42 67.53 67.59 67.82 68.42 BIRM 67.93 67.99 68.21 68.32 68.71 SPARSEIRM 67.72 67.85 67.99 68.13 68.81 FISHR 67.88 67.82 67.93 68.11 68.69 Average 67.02 67.25 67.34 67.63 68.44 In Table 3, we compare the performance of the simplest small-batch IRM training with that of those largebatch optimization technique-integrated IRM variants (i.e., \u2018LSGD/LALR/SAM\u2019 in the Table). As we can see, the use of large-batch optimization techniques indeed improves the prediction accuracy over the original IRM implementation. We also observe that the use of SAM for IRM is consistently better than LALR and LSGD, indicating the promise of SAM to scale up IRM with a large batch size. Yet, the small-batch training protocol consistently outperforms large-batch training across all the IRM variants (see the column \u2018Small\u2019). Additional experiment results in Section 6 show that smallbatch IRM training is effective across datasets, and promotes the invariance achieved by all methods. 4 MULTI-ENVIRONMENT INVARIANCE EVALUATION In this section, we revisit the evaluation metric used in existing IRM methods, and show that expanding the diversity of test-time environments would improve the accuracy of invariance assessment. Nearly all the existing IRM methods (including those listed in Table 1) follow the evaluation pipeline used in the vanilla IRM framework (Arjovsky et al., 2019), which assesses the performance of the learned invariant predictor on a single unseen test environment. This test-time environment is significantly different from train-time environments. For example, COLORED-MNIST (Arjovsky et al., 2019) suggests a principled way to de\ufb01ne two-bit environments, widely-used for IRM dataset curation. Speci\ufb01cally, the COLORED-MNIST task is to predict the label of the handwritten digit groups (digits 0-4 for group 1 and digits 5-9 for group 2). The digit number is also spuriously correlated with the digit color (Table 2). This spurious correlation is controlled by an environment bias parameter \u03b2, which speci\ufb01es different data environments with different levels of spurious correlation1. In (Arjovsky et al., 2019), \u03b2 = 0.1 and \u03b2 = 0.2 are used to de\ufb01ne two training environments, which sample the color ID by \ufb02ipping the digit group label with probability 10% and 20%, respectively. At test time, the invariant accuracy is evaluated on a single, unseen environment with \u03b2 = 0.9. However, the prediction accuracy of IRM could be sensitive to the choice of test-time environment (i.e., the value of \u03b2). For the default test environment \u03b2 = 0.9, the predictor performance of three representative IRM methods (IRMV1, IRM-GAME, FISHR) ranked from high to low is IRM-GAME>FISHR>IRMV1. Given this apparent ranking, we explore more diverse test-time environments, generated by \u03b2 \u2208\u2126:= {0.05, 0.1, . . . , 0.95}. 1In the two-bit environment, there exists another environment parameter \u03b1 that controls the label noise level. 5 \fPublished as a conference paper at ICLR 2023 0.0 0.2 0.4 0.6 0.8 1.0 Test Env. Bias Parameter  0 20 40 60 80 100 Invariant Accuracy (%) Test Env. IRMv1 IRM-Game Fishr IRMv1 IRM-Game Fishr 50 60 70 80 90 Invariant Accuracy (%) Average Accuracy (%) (A) (B) Figure 2: Performance comparison of IRM variants IRMV1, IRM-GAME, and FISHR on COLORED-MNIST. (A) Evaluation in different test-time environments (corresponding to different \u03b2). \u03b2 values used by the two training environments are 0.1, 0.2 respectively. The conventional evaluation is done with the test environment \u03b2 = 0.9 (see \u2018\u25b2\u2019). (B) Box plots of prediction accuracies over diverse test environments corresponding to \u03b2 \u2208{0.05, 0.1, . . . , 0.95}. IRMV1 achieves the best average accuracy (67.13%), followed by FISHR (67.05%) and IRM-GAME (65.53%). Although the train-time bias parameters {0.1, 0.2} belong to \u2126, test data is generated afresh, different from training data. We see in Figure 2A that the superiority of IRM-GAME at \u03b2 = 0.9 vanishes for smaller \u03b2. Consequently, for invariant prediction evaluated in other testing environments (e.g., \u03b2 < 0.4), the performance ranking of the same methods becomes IRMV1>FISHR>IRM-GAME. This mismatch of results suggests we measure the \u2018invariance\u2019 of IRM methods against diverse test environments. Otherwise, evaluation with single \u03b2 could give a false sense of invariance. In Figure 2B, we present the box plots of prediction accuracies for IRM variants, over the diverse set of testing environments (\u03b2 \u2208\u2126). Evidently, IRMV1, the oldest (sub-optimal) IRM method, yields the least variance of invariant prediction accuracies and the best average prediction accuracy, compared to both IRM-GAME and FISHR. To summarize, the new evaluation method, with diverse test environments, enables us to make a fair comparison of IRM methods implemented in different training environment settings. Unless speci\ufb01ed otherwise, we use the multi-environment evaluation method throughout this work. 5 ADVANCING IRM-GAME VIA CONSENSUS-CONSTRAINED BLO In this section, we revist and advance a special IRM variant, IRM-GAME (Ahuja et al., 2020), which endows each individual environment with a separate prediction head and converts IRM into an ensemble game over these multiple predictors. Revisiting IRM-GAME. We \ufb01rst introduce the setup of IRM-GAME following notations used in Section 2. The most essential difference between IRM-GAME and the vanilla IRM framework is that the former assigns each environment with an individual classi\ufb01er w(e), and then relies on the ensemble of these individual predictors, i.e., 1 N P e\u2208Etr(w(e) \u25e6\u03b8), for inference. IRM-GAME is in a sharp contrast to IRM, where an environment-agnostic prediction head w\u2217simultaneously optimizes the losses across all environments. Therefore, we raise the following question: Can IRM-GAME learn an invariant predictor? Inspired by the above question, we explicitly enforce invariance by imposing a consensus prediction constraint C := { \u0000 \u00af w(1), \u00af w(2), . . . \u00af w(N)\u0001 | \u00af w(1) = . . . = \u00af w(N)} and integrate it with IRM-GAME. Here, \u00af w(e) denotes the prediction head for the e-th environment. Based on the newly-introduced constraint, the ensemble prediction head 1 N P e\u2208Etr w(e) can be interpreted as the average consensus over N environments: w\u2217:= 1 N P e\u2208Etr w(e) = arg min{ \u00af w(e)}e\u2208C P e\u2208Etr \u2225\u00af w(e) \u2212w(e)\u22252 2. With the above consensus interpretation, we can then cast the invariant predictor-baked IRM-GAME as a consensus-constrained BLO problem, extended from (IRM): minimize \u03b8 P e\u2208Etr \u2113(e)(w\u2217(\u03b8) \u25e6\u03b8) subject to (I): w(e)(\u03b8) \u2208arg min \u00af w(e) \u2113(e)( \u00af w(e) \u25e6\u03b8), \u2200e \u2208Etr, (II): w\u2217(\u03b8) = 1 N P e\u2208Etr w(e)(\u03b8). (1) The above contains two lower-level problems: (I) per-environment risk minimization, and (II) projection onto the consensus constraint ({w(e)} \u2208C). The incorporation of (II) is intended to ensure the use of invariant prediction head w\u2217(\u03b8) in the upper-level optimization problem of (1). Limitation of (1) and BLOC-IRM. In (1), the introduced consensus-constrained lower-level problem might compromise the optimality of the lower-level solution w\u2217(\u03b8) to the per-environment (unconstrained) risk minimization problem (I), i.e., violating the per-environment stationarity 6 \fPublished as a conference paper at ICLR 2023 \u2225\u2207w\u2113(e)(w\u2217(\u03b8) \u25e6\u03b8)\u22252 2. Figure A3 justi\ufb01es this side effect. As we can see, the per-environment stationarity is hardly attained at the consensus prediction when solving (1). This is not surprising since a constrained optimization solution might not be a stationary solution to minimizing the (unconstrained) objective function. To alleviate this limitation, we improve (1) by explicitly promoting the per-environment stationarity \u2225\u2207w\u2113(e)(w\u2217(\u03b8) \u25e6\u03b8)\u22252 2 in its upper-level problem through optimization over \u03b8. This leads to BLOC-IRM (BLO with Consensus IRM): minimize \u03b8 P e\u2208Etr h \u2113(e)(w\u2217(\u03b8) \u25e6\u03b8) + \u03b3\u2225\u2207w\u2113(e)(w\u2217(\u03b8) \u25e6\u03b8)\u22252 2 i subject to Lower-level problems (I) and (II) in (1), (BLOC-IRM) where \u03b3 > 0 is a regularization parameter like IRMv0. Assisted by the (upper-level) prediction stationarity regularization, the consensus prediction (II) indeed simultaneously minimizes the risks of all the environments, supported by the empirical evidence that the convergence of \u2225\u2207w\u2113(e)(w\u2217(\u03b8) \u25e6\u03b8)\u22252 2 towards 0 along each environment\u2019s optimization path (see Figure A3). Further, we elaborate on how the BLOC-IRM problem can be effectively solved using an ordinary BLO solver. First, it is worth noting that although both (IRM) and BLOC-IRM are BLO problems, the latter is easier to solve since the lower-level constraint (I) is unconstrained and separable over environments, and the consensus operation (II) is linear. Based on these characteristics, the implicit gradient dw\u2217(\u03b8) d\u03b8 can be directly computed as dw\u2217(\u03b8) d\u03b8 = 1 N X e\u2208Etr dw(e)(\u03b8) d\u03b8 , subject to w(e)(\u03b8) \u2208arg min \u00af w(e) \u2113(e)( \u00af w(e) \u25e6\u03b8). (2) (c) BLOC-IRM   Forward Pass                                          Backward Pass (a) IRM (b) IRM-Game Inputs Feature Extractor Classifiers Prediction Logits Algorithm Figure 3: Schematic overview of BLOC-IRM over two training environments (red and green), and its comparison to IRM and IRM-GAME. Since the above lower-level problem is unconstrained, we can call the standard arg min differentiating method, such as implicit function approach (Gould et al., 2016) or gradient unrolling (Liu et al., 2021) to compute dw(e)(\u03b8) d\u03b8 . In our work, we adopt the gradient unrolling method, which approximates w(e)(\u03b8) by a K-step gradient descent solution, noted by w(e) K (\u03b8) and then leverages automatic differentiation (AD) to compute the derivative from w(e) K (\u03b8) to the variable \u03b8. Figure 3 shows the working pipeline of BLOCIRM and its comparison to original IRM and IRM-GAME methods. We use K = 1 for the lower-level problem throughout our experiments. We refer readers to Appendix B.2 for more algorithmic details. We also explore the performance of our proposed BLOC-IRM with various regularization terms, based on the penalties used in the existing literature. We show the best performance is always achieved when the stationarity is penalized in the upper-level (see Table A3). 6 EXPERIMENTS In this section, we begin by introducing some key experiment setups (with details in Appendix C.1), and then empirically show the effectiveness of our proposed IRM training and evaluation improvements over existing IRM methods across various datasets, models, and learning environments. 6.1 EXPERIMENT SETUPS Datasets and models. Our experiments are conducted over 7 datasets as referenced and shown in Tables 1, 2. Among these datasets, COLORED-MNIST, COLORED-FMNIST, CIFAR-MNIST, and COLORED-OBJECT are similarly curated, mimicking the pipeline of COLORED-MNIST (Arjovsky et al., 2019), by introducing an environment bias parameter (e.g., \u03b2 for COLORED-MNIST in Section 4) to customize the level of spurious correlation (as shown in Table 2) in different environments. In the CELEBA dataset, we choose the face attribute \u2018smiling\u2019 (vs. \u2018non-smiling\u2019) as the core feature aimed for classi\ufb01cation, and regard another face attribute \u2018hair color\u2019 (\u2018blond\u2019 vs. \u2018dark\u2019) as the 7 \fPublished as a conference paper at ICLR 2023 source of spurious correlation imposed on the core feature. By controlling the level of spurious correlation, we then create different training/testing environments in CELEBA. Furthermore, we study PACS and VLCS datasets, which were used to benchmark domain generalization ability in the real world (Borlino et al., 2021). It was recently shown by Gulrajani & Lopez-Paz (2020) that for these datasets, ERM could even be better than IRMV1. Yet, we will show that our proposed BLOC-IRM is a promising domain generalization method, which outperforms all the IRM baselines and ERM in practice. In addition, we follow Arjovsky et al. (2019) in adopting multi-layer perceptron (MLP) as the model for resolving COLORED-MNIST and COLORED-FMNIST problems. In the other more complex datasets, we use the ResNet-18 architecture (He et al., 2016). Baselines and implementation. Our baselines include 7 IRM variants (Table 1) and ERM, which are implemented using their of\ufb01cial repositories if available (see Appendix C.2). Unless speci\ufb01ed otherwise, our training pipeline uses the small-batch training setting. By default, we use the batch size of 1024 for COLORED-MNIST and COLORED-FMNIST, and 256 for other datasets. In Section 6.2 below, we also do a thorough comparison of large-batch vs small-batch IRM training. Table 4: Performance of existing IRM methods in large and small-batch settings. GRAYSCALE refers to ERM on uncolored data, which yields the best prediction (supposing no spurious correlation during training). The IRM performance is evaluated by average accuracy (\u2018Avg Acc\u2019) and accuracy gap (\u2018Acc Gap\u2019), in the format mean\u00b1std. A higher Avg Acc and lower Acc Gap is preferred. The theoretically optimal performance is 75% (Arjovsky et al., 2019). Dataset COLORED-MNIST COLORED-FMNIST Metrics(%) Avg Acc (\u2191) Acc Gap (\u2193) Avg Acc (\u2191) Acc Gap (\u2193) GRAYSCALE 73.39\u00b10.16 0.32\u00b10.03 74.05\u00b10.09 0.13\u00b10.04 ERM 49.19\u00b11.89 90.72\u00b12.08 49.77\u00b11.71 88.62\u00b12.49 Large Batch IRMV1 67.13 \u00b10.33 3.43\u00b10.14 67.19\u00b10.22 3.35\u00b10.11 IRMV0 65.39 \u00b10.34 4.69\u00b10.18 66.44\u00b10.28 3.53\u00b10.13 IRM-GAME 65.69 \u00b10.42 8.75\u00b10.14 65.91\u00b10.29 3.74\u00b10.09 REX 67.42 \u00b10.29 3.76\u00b10.07 67.82\u00b10.31 3.26\u00b10.16 BIRM 67.93 \u00b10.31 3.81\u00b10.11 67.75\u00b10.26 3.81\u00b10.11 SPARSEIRM 67.72 \u00b10.28 3.65\u00b10.08 67.89\u00b10.30 3.12\u00b10.15 FISHR 67.49\u00b10.39 4.37\u00b10.10 67.33\u00b10.24 4.49\u00b10.16 Small Batch IRMV1 68.33\u00b10.31 2.04\u00b10.05 68.76\u00b10.31 1.45\u00b10.09 IRMV0 68.37\u00b10.28 1.32\u00b10.09 69.07\u00b10.27 1.36\u00b10.06 IRM-GAME 67.73\u00b10.24 1.67\u00b10.14 67.49\u00b10.32 1.82\u00b10.13 REX 68.42\u00b10.29 1.65\u00b10.07 68.66\u00b10.22 1.29\u00b10.08 BIRM 68.71\u00b10.21 1.35\u00b10.09 68.64\u00b10.32 1.44\u00b10.13 SPARSEIRM 68.81\u00b10.25 1.72\u00b10.05 68.29\u00b10.22 1.28\u00b10.15 FISHR 68.69\u00b10.19 2.13\u00b10.08 68.79\u00b10.17 1.77\u00b10.10 Evaluation setup. As proposed in Section 4, we use the multi-environment evaluation metric unless speci\ufb01ed otherwise. To capture both the accuracy and variance of invariant predictions across multiple testing environments, the average accuracy and the accuracy gap (the difference of the best-case and worst-case accuracy) are measured for IRM methods. The resulting performance is reported in the form a \u00b1 b, with mean a and standard deviation b computed across 10 independent trials. 6.2 EXPERIMENT RESULTS Small-batch training improves all existing IRM methods on COLORED-MNIST & COLORED-FMNIST. Recall from Section 3 that all the existing IRM methods (Table 1) adopt full-batch IRM training on COLORED-MNIST & COLORED-FMNIST, which raises the largebatch training problem. In Table 4, we conduct a thorough comparison between the originally-used full-batch IRM methods and their small-batch counterparts. In addition, we present the performance of ERM and ERM-grayscale (we call it \u2018grayscale\u2019), where the latter is ERM on uncolored data. In the absence of any spurious correlation in the training set, grayscale gives the best performance. As discussed in Section 4 & 6.1, the IRM performance is measured by the average accuracy and the accuracy gap across 19 testing environments, parameterized by the environment bias parameter \u03b2 \u2208{0.05, . . . , 0.95}. We make some key observations from Table 4. First, small batch size helps improve all the existing IRM methods consistently, evidenced by the 1% \u223c3% improvement in average accuracy. Second, the small-batch IRM training signi\ufb01cantly reduces the variance of invariant predictions across different testing environments, evidenced by the decreased accuracy gap. This implies that the small-batch IRM training can also help resolve the limitation of multi-environment evaluation for the existing IRM methods, like the sensitivity of IRM-GAME accuracy to \u03b2 in Figure 2. Third, we observe that IRMV0, which does not seem to be useful in the large batch setting, becomes quite competitive with the other baselines in the small-batch setting. Thus, large-batch could suppress the IRM performance for some methods. In the rest of the experiments, we stick to the small-batch implementation of IRM training. BLOC-IRM outperforms IRM baselines in various datasets. Next, Table 5 demonstrates the effectiveness of our proposed BLOC-IRM approach versus ERM and existing IRM baselines across all the 7 datasets listed in Table 2. Evidently, BLOC-IRM yields a higher average accuracy compared to all the baselines, together with the smallest accuracy gap in most cases. Additionally, we observe that CELEBA, PACS and VLCS are much more challenging datasets for capturing invariance through IRM, as evidenced by the small performance gap between ERM and IRM methods. In 8 \fPublished as a conference paper at ICLR 2023 Table 5: IRM performance comparison between BLOC-IRM and other baselines. We use ResNet-18 (He et al., 2016) for all the datasets. The evaluation setup is consistent with Table 4, and the best performance per-dataset is highlighted in bold. We present the results with the full dataset list in Table A1. Algorithm COLORED-OBJECT CIFAR-MNIST CELEBA VLCS PACS Metrics (%) Avg Acc Acc Gap Avg Acc Acc Gap Avg Acc Acc Gap Avg Acc Acc Gap Avg Acc Acc Gap ERM 41.11\u00b11.44 86.43\u00b12.89 40.39\u00b11.32 85.53\u00b12.33 72.38\u00b10.29 10.73\u00b10.36 63.23\u00b10.23 12.39\u00b10.35 69.95\u00b10.35 14.32\u00b10.75 IRMV1 64.42\u00b10.21 4.18\u00b10.29 61.49\u00b10.29 7.17\u00b10.33 72.49\u00b10.38 10.15\u00b10.27 62.72\u00b10.29 12.74\u00b10.27 68.93\u00b10.33 14.99\u00b10.51 IRMV0 62.39\u00b10.25 5.36\u00b10.31 60.14\u00b10.18 8.83\u00b10.39 72.42\u00b10.35 10.43\u00b10.38 62.59\u00b10.32 12.99\u00b10.36 68.72\u00b10.29 15.29\u00b10.71 IRM-GAME 62.88\u00b10.34 5.59\u00b10.28 60.44\u00b10.31 6.72\u00b10.41 72.18\u00b10.44 12.32\u00b10.41 62.31\u00b10.38 13.37\u00b10.62 68.12\u00b10.22 15.77\u00b10.66 REX 63.37\u00b10.35 5.42\u00b10.31 62.32\u00b10.24 5.55\u00b10.32 72.34\u00b10.26 10.31\u00b10.23 63.19\u00b10.31 12.87\u00b10.31 69.43\u00b10.34 15.31\u00b10.67 BIRM 65.11\u00b10.27 3.31\u00b10.22 62.99\u00b10.35 5.23\u00b10.36 72.93\u00b10.28 9.92\u00b10.33 63.33\u00b10.40 12.13\u00b10.23 69.34\u00b10.25 15.76\u00b10.49 SPARSEIRM 64.97\u00b10.39 3.97\u00b10.25 62.16\u00b10.29 4.14\u00b10.31 72.42\u00b10.33 9.79\u00b10.21 62.86\u00b10.26 12.79\u00b10.35 69.52\u00b10.39 15.81\u00b10.82 FISHR 64.07\u00b10.23 4.41\u00b10.29 61.79\u00b10.25 5.55\u00b10.21 72.89\u00b10.25 9.42\u00b10.32 63.44\u00b10.37 11.93\u00b10.42 70.21\u00b10.22 14.52\u00b10.43 BLOC-IRM 65.97\u00b10.33 4.10\u00b10.36 63.69\u00b10.32 4.89\u00b10.36 73.35\u00b10.32 8.79\u00b10.21 63.62\u00b10.35 11.55\u00b10.32 70.31\u00b10.21 14.73\u00b10.65 particular, all the IRM methods, except FISHR and BLOC-IRM, could even be worse than ERM on PACS and VLCS. Here, we echo and extend the \ufb01ndings of Krueger et al. (2021, Section 4.3). However, we also show that BLOC-IRM is a quite competitive IRM variant when applied to realistic domain generalization datasets. We also highlight that the CELEBA experiment is newly constructed and performed in our work for invariance evaluation. Like PACS and VLCS, this experiment also shows that ERM is a strong baseline, and among IRM-based methods, BLOC-IRM is the best-performing, both in terms of accuracy and variance of invariant predictions. 0 200 400 600 800 1000 Intermediate Layer Dimension 60 62 64 66 68 70 Invariant Accuracy (%) IRM-V1 IRM-V0 IRM-Game REx BIRM SIRM Fishr BLOC-IRM Figure 4: IRM performance on COLORED-MNIST against the layer dimension in MLP. The dotted line represents the default dimension (d = 390) used in the literature. The invariant prediction accuracy is presented via the dot line (mean). The results are based on 10 independent trials and we report the variance in Figure A4. IRM against model size and training environment variation. Furthermore, we investigate the effect of model size and training environment diversity on the IRM performance. The recent works (Lin et al., 2022; Zhou et al., 2022b) have empirically shown that IRMV1 may suffer a signi\ufb01cant performance loss when trained over large-sized neural network models, and thus developed BIRM and SPARSEIRM approaches as advancements of IRMV1. Inspired by these works, Figure 4 presents the sensitivity of invariant prediction to model size for different IRM methods on COLORED-MNIST. Here the model size is controlled by the dimension of the intermediate layer (denoted by d) in MLP, and the default dimension is d = 390 (i.e., the vertical dotted line in Figure 4), which was used in (Arjovsky et al., 2019) and followed in the subsequent literature. As we can see, when d > 390, nearly all the studied IRM methods (including BLOC-IRM) suffer a performance drop. Yet, as d \u2265800, from the perspective of prediction accuracy and model resilience together, the top-3 best IRM methods with model size resilience are BIRM, SPARSEIRM, and BLOC-IRM, although we did not intentionally design BLOC-IRM to resist performance degradation against model size. We also show more experiment results in the Appendix. In Table A2, we study IRM with different numbers of training environment con\ufb01gurations and observe the consistent improvement of BLOCIRM over other baselines. In Table A4 we show that the performance of invariant prediction degrades, if additional covariate shifts (class, digit, and color imbalances on COLORED-MNIST) are imposed on the training environments following Krueger et al. (2021, Section 4.1) and also demonstrate that BLOC-IRM maintains the accuracy improvement over baselines with each variation. In Table A5, we compare the performance of different methods in the failure cases of IRM pointed out by (Kamath et al., 2021) and show the consistent improvement brought by BLOC-IRM. 7"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.04092v4",
+                    "title": "Advancing Model Pruning via Bi-level Optimization",
+                    "abstract": "The deployment constraints in practical applications necessitate the pruning\nof large-scale deep learning models, i.e., promoting their weight sparsity. As\nillustrated by the Lottery Ticket Hypothesis (LTH), pruning also has the\npotential of improving their generalization ability. At the core of LTH,\niterative magnitude pruning (IMP) is the predominant pruning method to\nsuccessfully find 'winning tickets'. Yet, the computation cost of IMP grows\nprohibitively as the targeted pruning ratio increases. To reduce the\ncomputation overhead, various efficient 'one-shot' pruning methods have been\ndeveloped, but these schemes are usually unable to find winning tickets as good\nas IMP. This raises the question of how to close the gap between pruning\naccuracy and pruning efficiency? To tackle it, we pursue the algorithmic\nadvancement of model pruning. Specifically, we formulate the pruning problem\nfrom a fresh and novel viewpoint, bi-level optimization (BLO). We show that the\nBLO interpretation provides a technically-grounded optimization base for an\nefficient implementation of the pruning-retraining learning paradigm used in\nIMP. We also show that the proposed bi-level optimization-oriented pruning\nmethod (termed BiP) is a special class of BLO problems with a bi-linear problem\nstructure. By leveraging such bi-linearity, we theoretically show that BiP can\nbe solved as easily as first-order optimization, thus inheriting the\ncomputation efficiency. Through extensive experiments on both structured and\nunstructured pruning with 5 model architectures and 4 data sets, we demonstrate\nthat BiP can find better winning tickets than IMP in most cases, and is\ncomputationally as efficient as the one-shot pruning schemes, demonstrating 2-7\ntimes speedup over IMP for the same level of model accuracy and sparsity.",
+                    "authors": "Yihua Zhang, Yuguang Yao, Parikshit Ram, Pu Zhao, Tianlong Chen, Mingyi Hong, Yanzhi Wang, Sijia Liu",
+                    "published": "2022-10-08",
+                    "updated": "2023-04-21",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG"
+                    ],
+                    "main_content": "Introduction While over-parameterized structures are key to the improved generalization of deep neural networks (DNNs) [1\u20133], they create new problems \u2013 the millions or even billions of parameters not only increase computational costs during inference, but also pose serious deployment challenges on resource-limited devices [4]. As a result, model pruning has seen a lot of research interest in recent years, focusing on reducing model sizes by removing (or pruning) redundant parameters [4\u20138]. Model sparsity (achieved by pruning) also bene\ufb01ts adversarial robustness [9], out-of-distribution generalization [10], and transfer learning [11]. Some pruning methods (towards structured sparsity) facilitate model deployment on hardware [12, 13]. 36th Conference on Neural Information Processing Systems (NeurIPS 2022). arXiv:2210.04092v4  [cs.LG]  21 Apr 2023 \fAmong various proposed model pruning algorithms [5, 9, 11, 14\u201327], the heuristics-based Iterative Magnitude Pruning (IMP) is the current dominant approach to achieving model sparsity without suffering performance loss, as suggested and empirically justi\ufb01ed by the Lottery Ticket Hypothesis (LTH) [17]. The LTH hypothesizes the existence of a subnetwork (the so-called \u2018winning ticket\u2019) when trained in isolation (e.g., from either a random initialization [17] or an early-rewinding point of dense model training [19]), can match the performance of the original dense model [5, 11, 17\u201320]. The core idea of IMP is to iteratively prune and retrain the model while progressively pruning a small ratio of the remaining weights in each iteration, and continuing till the desired pruning ratio has been reached. While IMP often \ufb01nds the winning ticket, it incurs the cost of repeated model retraining from scratch, making it prohibitively expensive for large datasets or large model architectures [28, 29]. To improve the ef\ufb01ciency of model pruning, numerous heuristics-based one-shot pruning methods [17, 21\u201325], e.g., one-shot magnitude pruning (OMP), have been proposed. These schemes directly prune the model to the target sparsity and are signi\ufb01cantly more ef\ufb01cient than IMP. Yet, the promise of current one-shot pruning methods is dataset/model-speci\ufb01c [22, 30] and mostly lies in the low pruning ratio regime [25]. As systematically studied in [22, 31], there exists a clear performance gap between one-shot pruning and IMP. As an alternative to heuristics-based schemes, optimization-based pruning methods [9, 15, 16, 26, 27] still follow the pruning-retraining paradigm and adopt sparsity regularization [32, 33] or parameterized masking [9, 16, 27] to prune models ef\ufb01ciently. However, these methods do not always match the accuracy of IMP and thus have not been widely used to \ufb01nd winning tickets [17, 21, 23\u201325, 29]. The empirical results showing that optimization underperforms heuristics motivate us to revisit the algorithmic fundamentals of pruning. 0 100 200 300 400 500 Time (min) 91.5 92.0 92.5 93.0 93.5 94.0 94.5 95.0 95.5 96.0 Test Accuracy (%) ResNet-18 ResNet-56 ResNet-20 Dense Model IMP BiP (ours) Figure 1: A performance snapshot of the proposed BIP method vs. the IMP baseline and the original dense model across three pruned ResNet models (ResNet-20, ResNet-56, and ResNet-18) with 74% sparsity on CIFAR-10. The marker size indicates the relative model size. The unicolor region corresponds to the same model type used by different pruning methods. To this end, we put forth a novel perspective of model pruning as a bi-level optimization (BLO) problem. In this new formulation, we show that BLO provides a technically-grounded optimization basis for an ef\ufb01cient implementation of the pruning-retraining paradigm, the key algorithmic component used in IMP. To the best of our knowledge, we make the \ufb01rst rigorous connection between model pruning and BLO. Technically, we propose a novel bi-level optimization-enabled pruning method (termed BIP). We further show how BIP takes advantage of the bi-linearity of the pruning problem to avoid the computational challenges of common BLO methods, and is as ef\ufb01cient as any \ufb01rst-order alternating optimization scheme. Practically, we demonstrate the superiority of the proposed BIP in terms of accuracy, sparsity, and computation ef\ufb01ciency through extensive experiments. BIP \ufb01nds the best winning ticket nearly in all settings while taking time comparable to the one-shot OMP. In Fig. 1, we present a snapshot of our empirical results for CIFAR-10 with 3 ResNet architectures at 74% pruning ratio. In all cases, BIP (\u22c6) \ufb01nds winning tickets, improving accuracy over the dense model (\u2022) and matching IMP (\u25b2), while being upto 5\u00d7 faster than IMP. Our contributions can be summarized as follows: \u2022 (Formulation) We rethink the algorithmic foundation of model pruning through the lens of BLO (bi-level optimization). The new BLO-oriented formulation disentangles pruning and retraining variables, providing the \ufb02exibility to design the interface between pruning and retraining. \u2022 (Algorithm) We propose the new bi-level pruning (BIP) algorithm, which is built upon the aforementioned BLO formulation and the implicit gradient-based optimization theory. Unlike computationally intensive standard BLO solvers, we theoretically show that BIP is as ef\ufb01cient as any \ufb01rst-order optimization by taking advantage of the bi-linear nature of the pruning variables. \u2022 (Experiments) We conduct extensive experiments across 4 datasets (CIFAR-10, CIFAR-100, TinyImageNet and ImageNet), 5 model architectures, and 3 pruning settings (unstructured pruning, \ufb01lter-wise structured pruning, and channel-wise structured pruning). We show that (i) BIP achieves higher test accuracy than IMP and \ufb01nds the best winning tickets nearly in all settings, (ii) BIP is highly ef\ufb01cient (comparable to one-shot pruning schemes), that is able to achieve 2-7\u00d7 speedup over IMP for the same level of model accuracy and sparsity, and (iii) BIP is able to \ufb01nd subnetworks that achieve better performance than the dense model regardless of initialization rewinding. 2 \f2 Related Work and Open Question Neural network pruning. As neural networks become deeper and more sophisticated, model pruning technology has gained increasing attention over the last decade since pruned models are necessary for the deployment of deep networks in practical applications [4, 34, 35]. With the goal of \ufb01nding highly-sparse and highly-accurate subnetworks from original dense models, a variety of pruning methods have been developed such as heuristics-based pruning [17, 21, 23\u201325, 29, 36] and optimization-based pruning [9, 16, 26, 27]. The former identi\ufb01es redundant model weights by leveraging heuristics-based metrics such as weight magnitudes [6, 17, 19, 11, 22, 37, 31, 36, 38], gradient magnitudes [21, 23, 24, 39, 40], and Hessian statistics [41\u201346].The latter is typically built on: 1) sparsity-promoting optimization [15, 33, 47\u201350], where model weights are trained by penalizing their sparsity-inducing norms, such as \u21130 and \u21131 norms for irregular weight pruning, and \u21132 norm for structured pruning; 2) parameterized masking [16, 9, 51\u201355], where model weight scores are optimized to \ufb01lter the most important weights and achieve better performance. Iterative vs. one-shot pruning, and motivation. Existing schemes can be further categorized into one-shot or iterative pruning based on the pruning schedule employed for achieving the targeted model sparsity. Among the iterative schemes, the IMP (Iterative Magnitude Pruning scheme) [17, 20, 56\u2013 65, 36] has played a signi\ufb01cant role in identifying high-quality \u2018winning tickets\u2019, as postulated by LTH (Lottery Ticket Hypothesis) [18, 19]. To enable consistent comparisons among different methods, we extend the original de\ufb01nition of winning tickets in [17] to \u2018matching subnetworks\u2019 [20] so as to cover different implementations of winning tickets, e.g., the use of early-epoch rewinding for model re-initialization [18] and the no-rewinding (i.e., \ufb01ne-tuning) variant [66]. Brie\ufb02y, the matching subnetworks should match or surpass the performance of the original dense model [20]. In this work, if a matching subnetwork is found better than the winning ticket obtained by the same method that follows the original LTH setup [18, 19], we will also call such a matching subnetwork a winning ticket throughout the paper. 0 20 40 60 80 100 Pruning Ratio (%) 80 82 84 86 88 90 92 Test Accuracy (%) 74% Sparsity Dense Model IMP OMP Grasp Snip SynFlow 20 40 60 80 100 Pruning Ratio (%) 10 2 10 3 Time Consumption (min) 3\u00d7 more time Figure 2: Illustration of the pros and cons of different pruning methods executed over (ResNet-20, CIFAR-10). Left: test accuracy vs. different pruning ratios of IMP and one-shot pruning methods (OMP [17], GRASP [23], SNIP [21], SYNFLOW [24]). Right: comparison of the ef\ufb01ciency with different sparsity. For example, the current state-of-theart (SOTA) implementation of IMP in [22] can lead to a pruned ResNet-20 on CIFAR-10 with 74% sparsity and 92.12% test accuracy, matching the performance of the original dense model (see red \u22c6in Fig. 2-Left). The IMP algorithm typically contains two key ingredients: (i) a temporally-evolving pruning schedule to progressively increase model sparsity over pruning iterations, and (ii) the pruning-retraining learning mechanism applied at each pruning iteration. With a target pruning ratio of p% with T pruning iterations, an example pruning schedule in (i) could be as follows \u2013 each iteration prunes (p%) 1/T of the currently unpruned model weights, progressively pruning fewer weights in each iteration. For (ii), the unpruned weights in each pruning iteration are re-set to the weights at initialization or at an early-training epoch [18], and re-trained till convergence. In brief, IMP repeatedly prunes, resets, and trains the network over multiple iterations. However, winning tickets found by IMP incur signi\ufb01cant computational costs. The sparsest winning ticket found by IMP in Fig. 2-Left (red \u22c6) utilizes T = 6 pruning iterations. As shown in Fig. 2-Right, this takes 3\u00d7 more time than the original training of the dense model. To avoid the computational cost of IMP, different kinds of \u2018accelerated\u2019 pruning methods were developed [17, 21, 23\u201325, 29], and many fall into the one-shot pruning category: The network is directly pruned to the target sparsity and retrained once. In particular, OMP (one-shot magnitude pruning) is an important baseline that simpli\ufb01es IMP [17]. It follows the pruning-retraining paradigm, but prunes the model to the target ratio with a single pruning iteration. Although one-shot pruning schemes are computationally cheap (Fig. 2-Right), they incur a signi\ufb01cant accuracy drop compared to IMP (Fig. 2-Left). Even if IMP is customized with a reduced number of training epochs per pruning round, the pruning accuracy also drops largely (see Fig. A2). Hence, there is a need for advanced model pruning techniques to \ufb01nd winning tickets like IMP, while being ef\ufb01cient like one-shot pruning. 3 \fDifferent from unstructured weight pruning described above, structured pruning takes into consideration the sparse patterns of the model, such as \ufb01lter and channel-wise pruning [13, 48, 51, 67\u2013 71]. Structured pruning is desirable for deep model deployment in the presence of hardware constraints [4, 34]. However, compared to unstructured pruning, it is usually more challenging to maintain the performance and \ufb01nd structure-aware winning tickets [28, 29]. Open question. As discussed above, one-shot pruning methods are unable to match the predictive performance of IMP, and structure-aware winning tickets are hard to \ufb01nd. Clearly, the best optimization foundation of model pruning is underdeveloped. Thus, we ask: Is there an optimization basis for a successful pruning algorithm that can attain high pruned model accuracy (like IMP) and high computational ef\ufb01ciency (like one-shot pruning)? The model pruning problem has a natural hierarchical structure \u2013 we need to \ufb01nd the best mask to prune model parameters, and then, given the mask, \ufb01nd the best model weights for the unpruned model parameters. Given this hierarchical structure, we believe that the bi-level optimization (BLO) framework is one promising optimization basis for a successful pruning algorithm. Bi-level optimization (BLO). BLO is a general hierarchical optimization framework, where the upper-level problem depends on the solution to the lower-level problem. Such a nested structure makes the BLO in its most generic form very dif\ufb01cult to solve, since it is hard to characterize the in\ufb02uence of the lower-level optimization on the upper-level problem. Various existing literature focuses on the design and analysis of algorithms for various special cases of BLO. Applications range from the classical approximate descent methods [72\u201374], penalty-based methods [75, 76], to recent BigSAM [77] and its extensions [78, 79]. It is also studied in the area of stochastic algorithms [80\u201382] and back-propagation based algorithms [83\u201385]. BLO has also advanced adversarial training [86], meta-learning [87], data poisoning attack generation [88], neural architecture search [89] as well as reinforcement learning [90]. Although BLO was referred in [50] for model pruning, it actually called an ordinary alternating optimization procedure, without taking the hierarchical learning structure of BLO into consideration. To the best of our knowledge, the BLO framework has not been considered for model pruning in-depth and systematically. We will show that model pruning yields a special class of BLO problems with bi-linear optimization variables. We will also theoretically show that this specialized BLO problem for model pruning can be solved as ef\ufb01ciently as \ufb01rst-order optimization. This is in a sharp contrast to existing BLO applications that rely on heuristics-based BLO solvers (e.g., gradient unrolling in meta learning [87] and neural architecture search [89, 91]). 3 BIP: Model Pruning via Bi-level Optimization In this section, we re-investigate the problem of model pruning through the lens of BLO and develop the bi-level pruning (BIP) algorithm. We can theoretically show that BIP can be solved as easily as the \ufb01rst-order alternating optimization by taking advantage of the bi-linearity of pruning variables. A BLO viewpoint on model pruning As described in the previous section, the pruning-retraining learning paradigm covers two kinds of tasks: \u0082 pruning that determines the sparse pattern of model weights, and \u0083 training remaining non-zero weights to recover the model accuracy. In existing optimization-based pruning methods [92\u201395], the tasks \u0082-\u0083 are typically achieved by optimizing model weights, together with penalizing their sparsity-inducing norms, e.g., the \u21131 and \u21132 norms [96]. Different from the above formulation, we propose to separate optimization variables involved in the pruning tasks \u0082 and \u0083. This leads to the (binary) pruning mask variable m \u2208{0, 1}n and the model weight variable \u03b8 \u2208Rn. Here n denotes the total number of model parameters. Accordingly, the pruned model is given by (m \u2299\u03b8), where \u2299denotes the element-wise multiplication. As will be evident later, this form of variable disentanglement enables us to explicitly depict how the pruning and retraining process co-evolve, and helps customize the pruning task with high \ufb02exibility. We next elaborate on how BLO can be established to co-optimize the pruning mask m and the retrained sparse model weights \u03b8. Given the pruning ratio p%, the sparsity constraint is given by m \u2208S, where S = {m | m \u2208{0, 1}n, 1T m \u2264k} and k = (1 \u2212p%)n. Our goal is to prune the original model directly to the targeted pruning ratio p% (i.e., without calling for the IMP-like sparsity schedule as described in Sec. 2) and obtain the optimized sparse model (m \u2299\u03b8). To this 4 \fend, we interpret the pruning task (i.e., \u0082) and the model retraining task (i.e., \u0083) as two optimization levels, where the former is formulated as an upper-level optimization problem, and it relies on the optimization of the lower-level retraining task. We thus cast the model pruning problem as the following BLO problem (with \u0083 being nested inside \u0082): minimize m\u2208S \u2113(m \u2299\u03b8\u2217(m)) | {z } \u0082: Pruning task ; subject to \u03b8\u2217(m) = arg min \u03b8\u2208Rn := g(m, \u03b8) z }| { \u2113(m \u2299\u03b8) + \u03b3 2 \u2225\u03b8\u22252 2 | {z } \u0083: Sparsity-\ufb01xed model retraining , (1) where \u2113denotes the training loss (e.g., cross-entropy), m and \u03b8 are the upper-level and lower-level optimization variables respectively, \u03b8\u2217(m) signi\ufb01es the lower-level solution obtained by minimizing the objective function g given the pruning mask m, and \u03b3 > 0 is a regularization parameter introduced to convexify the lower-level optimization so as to stabilize the gradient \ufb02ow from \u03b8\u2217(m) to m and thus the convergence of BLO [82, 97]. In a sharp contrast to existing single-level optimization-based model pruning methods [92\u201395], the BLO formulation (1) brings in two advantages. First, BLO has the \ufb02exibility to use mismatched pruning and retraining objectives at the upper and lower optimization levels, respectively. This \ufb02exibility allows us to regularize the lower-level training objective function in (1) and customize the implemented optimization methods at both levels. To be more speci\ufb01c, one can update the upper-level pruning mask m using a data batch (called B2) distinct from the one (called B1) used for obtaining the lower-level solution \u03b8\u2217(m). The resulting BLO procedure can then mimic the idea of meta learning to improve model generalization [98], where the lower-level problem \ufb01ne-tunes \u03b8 using B1, and the upper-level problem validates the generalization of the sparsity-aware \ufb01netuned model (m \u2299\u03b8\u2217(m)) using B2. Second, BLO enables us to explicitly model and optimize the coupling between the retrained model weights \u03b8\u2217(m) and the pruning mask m through the implicit gradient (IG)-based optimization routine. Here IG refers to the gradient of the lower-level solution \u03b8\u2217(m) with respect to (w.r.t.) the upper-level variable m, and its derivation calls the implicit function theory [76]. The use of IG makes our proposed BLO-oriented pruning (1) signi\ufb01cantly different from the greedy alternating minimization [99] that learns the upper-level and lower-level variables independently (i.e., minimizes one variable by \ufb01xing the other). We refer readers to the following section for the detailed IG theory. We will also show in Sec. 4 that the pruning strategy from (1) can outperform IMP in many pruning scenarios but is much more ef\ufb01cient as it does not call for the scheduler of iterative pruning ratios. Optimization foundation of BIP. The key optimization challenge of solving the BIP problem (1) lies in the computation of IG (implicit gradient). Prior to developing an effective solution, we \ufb01rst elaborate on the IG challenge, the unique characteristic of BLO. In the context of gradient descent, the gradient of the objective function in (1) yields d\u2113(m \u2299\u03b8\u2217(m)) dm | {z } Gradient of objective = \u2207m\u2113(m \u2299\u03b8\u2217(m)) + d(\u03b8\u2217(m)\u22a4) dm | {z } IG \u2207\u03b8\u2113(m \u2299\u03b8\u2217(m)), (2) where \u2207m and \u2207\u03b8 denote the partial derivatives of the bi-variate function \u2113(m \u2299\u03b8) w.r.t. the variable m and \u03b8 respectively, d\u03b8\u22a4/dm \u2208Rn\u00d7n denotes the vector-wise full derivative, and for ease of notation, we will omit the transpose \u22a4when the context is clear. In (2), the IG challenge refers to the demand for computing the full gradient of the implicit function \u03b8\u2217(m) = arg min\u03b8 g(m, \u03b8) w.r.t. m, where recall from (1) that g(m, \u03b8) := \u2113(m \u2299\u03b8) + \u03b3 2 \u2225\u03b8\u22252 2. Next, we derive the IG formula following the rigorous implicit function theory [76, 82, 87]. Based on the fact that \u03b8\u2217(m) satis\ufb01es the stationarity condition for the lower-level objective function in (2), it is not dif\ufb01cult to obtain that (see derivation in Appendix A) d\u03b8\u2217(m) dm = \u2212\u22072 m\u03b8\u2113(m \u2299\u03b8\u2217)[\u22072 \u03b8\u2113(m \u2299\u03b8\u2217) + \u03b3I]\u22121, (3) where \u22072 m\u03b8\u2113and \u22072 \u03b8\u2113denote the second-order partial derivatives of \u2113respectively, and (\u00b7)\u22121 denotes the matrix inversion operation. Yet, the exact IG formula (3) remains dif\ufb01cult to calculate due to the presence of matrix inversion and second-order partial derivatives. To simplify it, we impose the Hessian-free assumption, \u22072 \u03b8\u2113= 0, which is mild in general; For example, the decision boundaries of neural networks with ReLU 5 \factivations are piece-wise linear in a tropical hyper-surface [100], and this assumption has been widely used in BLO-involved applications such as meta learning [101] and adversarial learning [86]. Given \u22072 \u03b8\u2113= 0, the matrix inversion in (3) can be then mitigated, leading to the IG formula d\u03b8\u2217(m) dm = \u22121 \u03b3 \u22072 m\u03b8\u2113(m \u2299\u03b8\u2217). (4) At the \ufb01rst glance, the computation of the simpli\ufb01ed IG (4) still requires the mixed (second-order) partial derivative \u22072 m\u03b8\u2113. However, BIP is a special class of BLO problems with bi-linear variables (m \u2299\u03b8). Based on this bi-linearity, we can prove that IG in (4) can be analytically expressed using only \ufb01rst-order derivatives; see the following theorem. Proposition 1 Assuming \u22072 \u03b8\u2113= 0 and de\ufb01ning \u2207z\u2113(z) := \u2207z\u2113(z) |z=m\u2299\u03b8\u2217, the implicit gradient (4) is then given by d\u03b8\u2217(m) dm = \u22121 \u03b3 diag(\u2207z\u2113(z)); (5) Further, the gradient of the objective function given by (2) becomes d\u2113(m \u2299\u03b8\u2217) dm = (\u03b8\u2217\u22121 \u03b3 m \u2299\u2207z\u2113(z)) \u2299\u2207z\u2113(z), (6) where \u2299denotes the element-wise multiplication. Proof: Using chain-rule, we can obtain that \u2207\u03b8\u2113(m \u2299\u03b8\u2217) = diag(m)\u2207z\u2113(z) = m \u2299\u2207z\u2113(z); (7) similarly, \u2207m\u2113(m \u2299\u03b8\u2217) = diag(\u03b8\u2217)\u2207z\u2113(z) = \u03b8\u2217\u2299\u2207z\u2113(z) (8) where diag(\u00b7) represents a diagonal matrix with \u00b7 being the main diagonal vector. Further, we can convert (4) to \u22072 m\u03b8\u2113(m \u2299\u03b8\u2217) (7) = \u2207m [m \u2299\u2207z\u2113(z)] chain rule = diag(\u2207z\u2113(z)) + diag(m)[\u2207m(\u2207z\u2113(z))] (8) = diag(\u2207z\u2113(z)) + diag(m)[diag(\u03b8\u2217)\u22072 z\u2113(z)] = diag(\u2207z\u2113(z)), (9) where the last equality holds due to the Hessian-free assumption. With (9) and (4) we can prove (5). Next, substituting the IG (5) to the upper-level gradient (2), we obtain that d\u2113(m \u2299\u03b8\u2217) dm = \u2207m\u2113(m \u2299\u03b8\u2217) \u22121 \u03b3 \u2207z\u2113(z) \u2299\u2207\u03b8\u2113(m \u2299\u03b8\u2217) (7),(8) = \u03b8\u2217\u2299\u2207z\u2113(z) \u22121 \u03b3 \u2207z\u2113(z) \u2299(m \u2299\u2207z\u2113(z)) = (\u03b8\u2217\u22121 \u03b3 m \u2299\u2207z\u2113(z)) \u2299\u2207z\u2113(z), which leads to (6). The proof is now complete. \u25a1 The key insight drawn from Prop. 1 is that the bi-linearity of pruning variables (i.e., m \u2299\u03b8\u2217) makes the IG-involved gradient (2) easily solvable, and the computational complexity is almost the same as that of computing the \ufb01rst-order gradient \u2207z\u2113(z) just once, as supported by (6) BIP algorithm and implementation. We next formalize the BIP algorithm based on Prop. 1 and the alternating gradient descent based BLO solver [82]. At iteration t, there are two main steps. # Lower-level SGD for model retraining: Given m(t\u22121), \u03b8(t\u22121), and z(t\u22121) := m(t\u22121) \u2299\u03b8(t\u22121), we update \u03b8(t) by randomly selecting a data batch with the learning rate \u03b1 and applying SGD (stochastic gradient descent) to the lower-level problem of (1), \u03b8(t) = \u03b8(t\u22121) \u2212\u03b1\u2207\u03b8g(m(t\u22121), \u03b8(t\u22121)) (7) = \u03b8(t\u22121) \u2212\u03b1[m(t\u22121) \u2299\u2207z\u2113(z) | z=z(t\u22121) + \u03b3\u03b8(t\u22121)], (\u03b8-step) # Upper-level SPGD for pruning: Given m(t\u22121), \u03b8(t), and z(t+1/2) := m(t\u22121) \u2299\u03b8(t), we update m using SPGD (stochastic projected gradient descent) along the IG-enhanced descent direction (2), m(t) =PS \u0014 m(t\u22121) \u2212\u03b2 d\u2113(m \u2299\u03b8(t)) dm | m=m(t\u22121) \u0015 (6) =PS \u0014 m(t\u22121) \u2212\u03b2 \u0012 \u03b8(t) \u22121 \u03b3 m(t\u22121) \u2299\u2207z\u2113(z) | z=z(t+1/2) \u0013 \u2299\u2207z\u2113(z) | z=z(t+1/2) \u0015 , (m-step) 6 \fOur proposal: Bi-level Pruning (BiP) directly to pruning ratio p% Original Pruned Pretrained Model Original Prior art: Iterative Magnitude Pruning (IMP) over T Rounds Pruned Train N epochs Random & dense init Train N epochs Prune p1/T% Prune p1/T% Rewind to init Train N epochs Rewind to init Prune p1/T% Rewind to init Train N epochs Prune to p%  1 epoch SPGD Round 1 Round 2 ...... Round T Prune to p%  1 epoch SPGD Train weights 1 epoch SGD Sparse Model Upper level Upper level Upper level Lower level Lower level Lower level Prune to p%  1 epoch SPGD Train weights 1 epoch SGD Train weights  1 epoch SGD Sparse Model Figure 3: Visualization of pruning pipeline comparison between IMP and BIP. Edge refers to the mask update and color refers to the weight update. where \u03b2 > 0 is the upper-level learning rate, and PS(\u00b7) denotes the Euclidean projection onto the constraint set S given by S = {m | m \u2208{0, 1}n, 1T m \u2264k} in (1) and is achieved by the top-k hard-thresholding operation as will be detailed later. In BIP, the (\u03b8-step) and (m-step) steps execute iteratively. For clarity, Fig. 3 shows the difference between the pruning pipelines of BIP and IMP. In contrast to IMP that progressively prunes and retrains a model with a growing pruning ratio, BIP directly prunes the model to the targeted sparsity level without involving costly re-training process. In practice, we \ufb01nd that both the upperand lowerlevel optimization routines of BIP converge very well (see Fig. A12 and Fig. A13). It is also worth noting that both (\u03b8-step) and (m-step) only require the \ufb01rst-order information \u2207z\u2113(z), demonstrating that BIP can be conducted as ef\ufb01ciently as \ufb01rst-order optimization. In Fig. A1, we highlight the algorithmic details on the BIP pipeline. We present more implementation details of BIP below and refer readers to Appendix B for a detailed algorithm description. ! Discrete optimization over m: We follow the \u2018convex relaxation + hard thresholding\u2019 mechanism used in [9, 16]. Speci\ufb01cally, we relax the binary masking variables to continuous masking scores m \u2208[0, 1]. We then acquire loss gradients at the backward pass based on the relaxed m. At the forward pass, we project it onto the discrete constraint set S using the hard thresholding operator, where the top k elements are set to 1s and the others to 0s. See Appendix B for more discussion. ! Data batch selection for lower-level and upper-level optimization: We adopt different data batches (with the same batch size) when implementing (\u03b8-step) and (m-step). This is one of the advantages of the BLO formulation, which enables the \ufb02exibility to customize the lower-level and upper-level problems. The use of diverse data batches is bene\ufb01cial to generalization as shown in [98]. ! Hyperparameter tuning: As described in (\u03b8-step)-(m-step), BIP needs to set two learning rates \u03b1 and \u03b2 for lower-level and upper-level optimization, respectively. We choose \u03b1 = 0.01 and \u03b2 = 0.1 in all experiments, where we adopt the mask learning rate \u03b2 from Hydra [9] and set a smaller lower-level learning rate \u03b1, as \u03b8 is initialized by a pre-trained dense model. We show ablation study on \u03b1 in Fig. A8(c). BLO also brings in the low-level convexi\ufb01cation parameter \u03b3. We set \u03b3 = 1.0 in experiments and refer readers to Fig. A8(b) for a sanity check. ! One-step vs. multi-step SGD: In (\u03b8-step), the one-step SGD is used and helps reduce the computation overhead. In practice, we also \ufb01nd that the one-step SGD is suf\ufb01cient: The use of multi-step SGD in BIP does not yield much signi\ufb01cant improvement over the one-step version; see Fig. A8(a). ! Extension to structured pruning: We formulate and solve the BIP problem in the context of unstructured (element-wise) weight pruning. However, if de\ufb01ne the pruning mask m w.r.t. model\u2019s structural units (e.g., \ufb01lters), BIP is easily applied to structured pruning (see Fig. 6 and Fig. A10). 7 \f0 20 40 60 80 100 Pruning Ratio (%) 93.0 93.5 94.0 94.5 95.0 95.5 Test Accuracy (%) CIFAR-10, ResNet-18 Dense Model Hydra IMP OMP Grasp BiP Best Winning Ticket 0 20 40 60 80 100 Pruning Ratio (%) 89.0 89.5 90.0 90.5 91.0 91.5 92.0 92.5 Test Accuracy (%) CIFAR-10, ResNet-20 0 20 40 60 80 100 Pruning Ratio (%) 90.0 90.5 91.0 91.5 92.0 92.5 93.0 93.5 Test Accuracy (%) CIFAR-10, ResNet-56 0 20 40 60 80 100 Pruning Ratio (%) 92.0 92.5 93.0 93.5 94.0 94.5 Test Accuracy (%) CIFAR-10, VGG-16 0 20 40 60 80 100 Pruning Ratio (%) 72.0 73.0 74.0 75.0 76.0 77.0 78.0 Test Accuracy (%) CIFAR-100, ResNet-18 Dense Model Hydra IMP OMP Grasp BiP Best Winning Ticket 0 20 40 60 80 100 Pruning Ratio (%) 60.0 62.0 64.0 66.0 68.0 70.0 Test Accuracy (%) CIFAR-100, ResNet-20 0 20 40 60 80 100 Pruning Ratio (%) 57.0 58.0 59.0 60.0 61.0 62.0 63.0 64.0 Test Accuracy (%) Tiny-ImageNet, ResNet-18 0 20 40 60 80 100 Pruning Ratio (%) 72.0 73.0 74.0 75.0 76.0 77.0 Test Accuracy (%) ImageNet, ResNet-50 Dense Model IMP BiP Best Winning Ticket Figure 4: Unstructured pruning trajectory given by test accuracy (%) vs. sparsity (%) on various dataset-model pairs. The proposed BIP is compared with HYDRA [9], IMP [22], OMP [22], GRASP [23]. And the performance of dense model and that of the best winning ticket are marked using dashed lines in each plot. The solid line and shaded area of each pruning method represent the mean and variance of test accuracies over 3 independent trials. We observe that BIP consistently outperforms the other baselines. Note in the (ImageNet, ResNet-50) setting, we only compare BIP with our strongest baseline IMP due to computational resource constraints. 4 Experiments In this section, we present extensive experimental results to show the effectiveness of BIP across multiple model architectures, various datasets, and different pruning setups. Compared to IMP, one-shot pruning, and optimization-based pruning baselines, we \ufb01nd that BIP can \ufb01nd better winning tickets in most cases and is computationally ef\ufb01cient. 4.1 Experiment Setup Datasets and models. Following the pruning benchmark in [22], we consider 4 datasets including CIFAR-10 [102], CIFAR-100 [102], Tiny-ImageNet [103], ImageNet [104], and 5 architecture types including ResNet-20/56/18/50 and VGG-16 [105, 106]. Tab. A1 summarizes these datasets and model con\ufb01gurations and setups. Baselines, training, and evaluation. As baselines, we mainly focus on 4 SOTA pruning methods, x IMP [17], y OMP [17], z GRASP [23] (a one-shot pruning method by analyzing gradient \ufb02ow at initialization), and { HYDRA [9] (an optimization-based pruning method that optimizes masking scores). It is worth noting that there exist various implementations of IMP, e.g., speci\ufb01ed by different learning rates and model initialization or \u2018rewinding\u2019 strategies [18]. To make a fair comparison, we follow the recent IMP benchmark in [22], which can \ufb01nd the best winning tickets over current heuristics-based pruning baselines. We also remark that HYDRA is originally proposed for improving the adversarial robustness of a pruned model, but it can be easily customized for standard pruning when setting the adversary\u2019s strength as 0 [9]. We choose HYDRA as a baseline because it can be regarded as a single-level variant of BIP with post-optimization weight retraining. When implementing BIP, unless speci\ufb01ed otherwise, we use the 1-step SGD in (\u03b8-step), and set the learning rates (\u03b1, \u03b2) and the lower-level regularization parameter \u03b3 as described in the previous section. When implementing baselines, we follow their of\ufb01cial repository setups. We evaluate the performance of all methods mainly from two perspectives: (1) the test accuracy of the sub-network, and (2) the runtime of pruning to reach the desired sparsity. We refer readers to Tab. A3 and Appendix C.2 for more training and evaluation details, such as training epochs and learning rate schedules. 4.2 Experiment Results BIP identi\ufb01es high-accuracy subnetworks. In what follows, we look at the quality of winning tickets identi\ufb01ed by BIP. Two key observations can be drawn from our results: (1) BIP \ufb01nds winning tickets of higher accuracy and/or higher sparsity than the baselines in most cases (as shown in Fig. 4 8 \fTable 1: The sparsest winning tickets found by different methods at different data-model setups. Winning tickets refer to the sparse models with an average test accuracy no less than the dense model [20]. In each cell, p% (acc\u00b1std%) represents the sparsity as well as the test accuracy. The test accuracy of dense models can be found in the header. \u0017 signi\ufb01es that no winning ticket is found by a pruning method. Given the data-model setup (i.e., per column), the sparsest winning ticket is highlighted in bold. Method CIFAR-10 CIFAR-100 ResNet-18 ResNet-20 ResNet-56 VGG-16 ResNet-18 ResNet-20 (94.77%) (92.12%) (92.95%) (93.65%) (76.67%) (68.69%) IMP 87% (94.77\u00b10.10%) 74% (92.15\u00b10.15%) 74% (92.99\u00b10.12%) 89% (93.68\u00b10.05%) 87% (76.91\u00b10.19%) \u0017 OMP 49% (94.80\u00b10.10%) \u0017 \u0017 20% (93.79\u00b10.06%) 74% (76.99\u00b10.07%) \u0017 GRASP \u0017 \u0017 36% (93.07\u00b10.34%) \u0017 \u0017 \u0017 HYDRA \u0017 \u0017 \u0017 87% (93.73\u00b10.03%) \u0017 20% (68.94\u00b10.17%) BIP 89% (94.79\u00b10.15%) 67% (92.14\u00b10.15%) 74% (93.13\u00b10.04%) 93% (93.75\u00b10.15%) 89% (76.69\u00b10.18%) 49% (68.78\u00b10.10%) 0 20 40 60 80 100 Pruning Ratio (%) 91.0 91.5 92.0 92.5 93.0 93.5 94.0 94.5 95.0 95.5 Test Accuracy (%) CIFAR-10, ResNet-18 Dense Model Hydra IMP OMP Grasp BiP Best Winning Ticket 0 20 40 60 80 100 Pruning Ratio (%) 82.0 84.0 86.0 88.0 90.0 92.0 Test Accuracy (%) CIFAR-10, ResNet-20 0 20 40 60 80 100 Pruning Ratio (%) 70.0 71.0 72.0 73.0 74.0 75.0 76.0 77.0 78.0 Test Accuracy (%) CIFAR-100, ResNet-18 0 20 40 60 80 100 Pruning Ratio (%) 45.0 50.0 55.0 60.0 65.0 70.0 Test Accuracy (%) CIFAR-100, ResNet-20 Figure 6: Filter pruning given by test accuracy (%) vs. pruning ratio (%). The visual presentation setting is the same as Fig. 4. We observe that BIP identi\ufb01es winning tickets of structured pruning in certain sparsity regimes. and Tab. 1); (2) The superiority of BIP holds for both unstructured pruning and structured pruning (as shown in Fig. 6 and Fig. A10). We refer to more experiment results in Appendix C.3. Fig. 4 shows the unstructured pruning trajectory (given by test accuracy vs. pruning ratio) of BIP and baseline methods in 8 model-dataset setups. For comparison, we also present the performance of the original dense model. As we can see, the proposed BIP approach \ufb01nds the best winning tickets (in terms of the highest accuracy) compared to the baselines across all the pruning setups. Among the baseline methods, IMP is the most competitive method to ours. However, the improvement brought by BIP is signi\ufb01cant with respect to the variance of IMP, except for the 60%-80% sparsity regime in (CIFAR-10, ResNet-20). In the case of (CIFAR-100, ResNet-20), where IMP can not \ufb01nd any winning tickets (as con\ufb01rmed by [22]), BIP still manages to \ufb01nd winning tickets with around 0.6% improvement over the dense model. In Tab. 1, we summarize the sparsest winning tickets along the pruning trajectory identi\ufb01ed by different pruning methods. BIP can identify the winning tickets with higher sparsity levels than the other methods, except in the case of (CIFAR-10, ResNet-20). 20 40 60 80 100 Pruning Ratio (%) 10 2 10 3 Time Consumption (min) Dense Model Hydra IMP OMP Grasp BiP Figure 5: Time consumption comparison on (CIFAR-10, ResNet-18) with different pruning ratio p. Fig. 6 demonstrates the structured pruning trajectory on the CIFAR-10/100 datasets. Here we focus on \ufb01lter pruning, where the \ufb01lter is regarded as a masking unit in (1). We refer readers to Fig. A10 for channel-wise pruning results. Due to the page limit, we only report the results of the \ufb01lter-wise pruning in the main paper and please refer to Appendix C.3 for channel-wise pruning. Compared to Fig. 4, Fig. 6 shows that it becomes more dif\ufb01cult to \ufb01nd winning tickets of high accuracy and sparsity in the structured pruning, and the gap among different methods decreases. This is not surprising, since \ufb01lter pruning imposes much stricter pruning structure constraints than irregular pruning. However, BIP still outperforms all the baselines. Most importantly, it identi\ufb01es clear winning tickets in the low sparse regime even when IMP fails. BIP is computationally ef\ufb01cient. In our experiments, another key observation is that BIP yields sparsity-agnostic runtime complexity while IMP leads to runtime exponential to the target sparsity. Fig. 5 shows the computation cost of different methods versus pruning ratios on (CIFAR-10, ResNet18). For example, BIP takes 86 mins to \ufb01nd the sparsest winning ticket (with 89% sparsity in Tab. 1). This yields 7\u00d7 less runtime than IMP, which consumes 620 mins to \ufb01nd a comparable winning ticket with 87% sparsity. Compared to the optimization-based baseline HYDRA, BIP is more ef\ufb01cient as it 9 \fN/A 5% 10% 50% 75% 100% Rewinding Epoch Number 89.0 89.5 90.0 90.5 91.0 91.5 92.0 92.5 93.0 Test Accuracy of BiP (%) CIFAR10, ResNet20s Dense Model p=20% p=67.2% p=86.58% N/A 5% 10% 50% 75% 100% Rewinding Epoch Number 94.0 94.2 94.4 94.6 94.8 95.0 95.2 95.4 Test Accuracy of BiP (%) CIFAR10, ResNet18 Dense Model p=20% p=67.2% p=86.58% N/A 5% 10% 50% 75% 100% Rewinding Epoch Number 75.75 76.00 76.25 76.50 76.75 77.00 77.25 77.50 77.75 Test Accuracy of BiP (%) CIFAR100, ResNet18 Dense Model p=20% p=67.2% p=86.58% Figure 7: The sensitivity of BIP to rewinding epoch numbers on different datasets and model architectures. \"N/A\" in the x-axis indicates BIP without retraining. does not rely on the extra post-optimization retraining; see Tab. A2 for a detailed summary of runtime and number of training epochs required by different pruning methods. Further, BIP takes about 1.25 \u00d7 more computation time than GRASP and OMP. However, the latter methods lead to worse pruned model accuracy, as demonstrated by their failure to \ufb01nd winning tickets in Tab. 1, Fig. 4, and Fig. 6. BIP requires no rewinding. Another advantage of BIP is that it insensitive to model rewinding to \ufb01nd matching subnetworks. Recall that rewinding is a strategy used in LTH [19] to determine what model initialization should be used for retraining a pruned model. As shown in [22], an IMP-identi\ufb01ed winning ticket could be sensitive to the choice of a rewinding point. Fig. 7 shows the test accuracy of the BIP-pruned model when it is retrained at different rewinding epochs under various datasets and model architectures, where \u2018N/A\u2019 in the x-axis represents the case of no retraining (and thus no rewinding). As we can see, a carefully-tuned rewinding scheme does not lead to a signi\ufb01cant improvement over BIP without retraining. This suggests that the subnetworks found by BIP are already of high quality and does not require any rewinding operation. Additional results. We include more experiment results in Appendix C.3. In particular, we show more results in both unstructured and structured pruning settings in Fig. A4, Fig. A5, Fig. A6 and Fig. A7, where we compare BIP with more baselines and cover more model architectures. We also study the sensitivity of BIP to the lower-level step number, lower-level regularization coef\ufb01cient, the signi\ufb01cance of the implicit gradient term (2), learning rate, and batch size, as shown in Fig. A8 and Fig. A9. To demonstrate the convergence of the upper-level and lower-level optimization in BIP, we show the training trajectory of BIP for accuracy (Fig.A12) and mask score (Fig.A13), and show how the lower-level step number affects the convergence speed (Fig. A14)). Further, we show the performance of BIP vs. the growth of training epochs (Fig. A15), and its performance vs. different data batch schedulers (see Fig. A16). 5"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2112.12376v6",
+                    "title": "Revisiting and Advancing Fast Adversarial Training Through The Lens of Bi-Level Optimization",
+                    "abstract": "Adversarial training (AT) is a widely recognized defense mechanism to gain\nthe robustness of deep neural networks against adversarial attacks. It is built\non min-max optimization (MMO), where the minimizer (i.e., defender) seeks a\nrobust model to minimize the worst-case training loss in the presence of\nadversarial examples crafted by the maximizer (i.e., attacker). However, the\nconventional MMO method makes AT hard to scale. Thus, Fast-AT (Wong et al.,\n2020) and other recent algorithms attempt to simplify MMO by replacing its\nmaximization step with the single gradient sign-based attack generation step.\nAlthough easy to implement, Fast-AT lacks theoretical guarantees, and its\nempirical performance is unsatisfactory due to the issue of robust catastrophic\noverfitting when training with strong adversaries. In this paper, we advance\nFast-AT from the fresh perspective of bi-level optimization (BLO). We first\nshow that the commonly-used Fast-AT is equivalent to using a stochastic\ngradient algorithm to solve a linearized BLO problem involving a sign\noperation. However, the discrete nature of the sign operation makes it\ndifficult to understand the algorithm performance. Inspired by BLO, we design\nand analyze a new set of robust training algorithms termed Fast Bi-level AT\n(Fast-BAT), which effectively defends sign-based projected gradient descent\n(PGD) attacks without using any gradient sign method or explicit robust\nregularization. In practice, we show our method yields substantial robustness\nimprovements over baselines across multiple models and datasets. Codes are\navailable at https://github.com/OPTML-Group/Fast-BAT.",
+                    "authors": "Yihua Zhang, Guanhua Zhang, Prashant Khanduri, Mingyi Hong, Shiyu Chang, Sijia Liu",
+                    "published": "2021-12-23",
+                    "updated": "2022-10-04",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG"
+                    ],
+                    "main_content": "Introduction Given the fact that machine learning (ML) models can be easily fooled by tiny adversarial perturbations (also known as adversarial attacks) on the input (Goodfellow et al., 2014; Carlini & Wagner, 2017; Papernot et al., 2016), training robust deep neural networks is now a major focus in research. Nearly all existing effective defense mechanisms (Madry et al., 2018; Zhang et al., 2019b; Shafahi et al., 2019; Wong et al., 2020; Zhang et al., 2019a; Athalye et al., 2018a) are built on the adversarial training (AT) recipe, introduced by (Szegedy et al., 2014) and later formalized in (Madry et al., 2018) using min-max optimization (MMO), where a minimizer (i.e. defender) seeks to update model parameters against a maximizer (i.e. attacker) that aims to increase the training loss by perturbing the training data. Despite the effectiveness of the AT-type defenses in various application domains, the min-max nature makes them dif\ufb01cult to scale, because of the multiple maximization steps required by the iterative attack generator at each training step. The resulting prohibitive computation cost makes AT not suitable in practical settings. For example, (Xie et al., 2019) used 128 GPUs to run AT on ImageNet. Thereby, how to speed up AT without losing accuracy and robustness is now a grand challenge for adversarial defense (Wong et al., 2020). Recently, (Wong et al., 2020; Shafahi et al., 2019; Zhang et al., 2019a; Andriushchenko & Flammarion, 2020) attempted to develop computationally-ef\ufb01cient alternatives of AT, that is, the \u2018fast\u2019 versions of AT. To the best of our knowledge, FAST-AT (Wong et al., 2020) and FASTAT with gradient alignment (GA) regularization, termed FAST-AT-GA (Andriushchenko & Flammarion, 2020), are the two state-of-the-art (SOTA) \u2018fast\u2019 versions of AT, since they achieve the most signi\ufb01cant reduction in computational complexity and preserve accuracy and robustness to a large extent. In particular, the inner-maximization in FAST-AT (Wong et al., 2020) only calls for single-step attack generation. However, different from the direct application of fast gradient sign method (Goodfellow et al., 2015), the empirical success of FAST-AT also relies on a series of heuristics-based strategies, e.g., using large attack step size, cyclic learning rate schedule, and mixed-precision trainarXiv:2112.12376v6  [cs.LG]  4 Oct 2022 \fRevisiting and Advancing Fast Adversarial Training Through the Lens of Bi-Level Optimization Table 1. Performance overview of proposed FAST-BAT vs. the baselines FAST-AT (Wong et al., 2020) and FAST-AT-GA (Andriushchenko & Flammarion, 2020) on CIFAR-10, CIFAR-100 and Tiny-ImageNet with PreActResNet-18. All methods are robustly trained under two perturbation budgets \u03f5 = 8/255 and 16/255 over 20 epochs. We use the early-stopping policy (Rice et al., 2020) to report the model with best robustness for each method. The evaluation metrics include robust accuracy (RA) against PGD-50-10 attacks (50-step PGD attack with 10 restarts) (Madry et al., 2018) at \u03f5 = 8/255 and 16/255 (the test-time \u03f5 is the same as the train-time), RA against AutoAttack (AA) (Croce & Hein, 2020) at \u03f5 = 8/255 and 16/255, and computation time (per epoch). The result a\u00b1b represents mean a and standard deviation b over 10 random trials. All experiments are run on a single GeForce RTX 3090 GPU. CIFAR-10, PreActResNet-18 Method SA (%) (\u03f5 = 8/255) RA-PGD (%) (\u03f5 = 8/255) RA-AA (%) (\u03f5 = 8/255) SA (%) (\u03f5 = 16/255) RA-PGD (%) (\u03f5 = 16/255) RA-AA (%) (\u03f5 = 16/255) Time (s/epoch) FAST-AT 82.39\u00b10.14 45.49\u00b10.21 41.87\u00b10.15 44.15\u00b17.27 21.83\u00b11.32 12.49\u00b10.33 23.1 FAST-AT-GA 79.71\u00b10.24 47.27\u00b10.22 43.24\u00b10.27 58.29\u00b11.32 26.01\u00b10.16 17.97\u00b10.33 75.3 FAST-BAT 79.97\u00b10.12 48.83\u00b10.17 45.19\u00b10.12 68.16\u00b10.25 27.69\u00b10.16 18.79\u00b10.24 61.4 CIFAR-100, PreActResNet-18 FAST-AT 52.62 \u00b10.18 24.66\u00b10.21 21.72\u00b10.17 21.32\u00b13.27 8.62\u00b11.03 6.22\u00b10.61 23.8 FAST-AT-GA 50.06\u00b10.27 24.97\u00b10.23 21.82\u00b10.21 32.51\u00b11.27 12.27\u00b10.36 9.43\u00b10.19 77.1 FAST-BAT 50.19\u00b10.21 26.49\u00b10.20 23.97\u00b10.15 39.29\u00b10.53 13.97\u00b10.17 11.32\u00b10.22 61.6 Tiny-ImageNet, PreActResNet-18 FAST-AT 41.37\u00b13.08 17.05\u00b13.25 12.31\u00b12.73 31.38\u00b10.19 5.42\u00b12.17 3.13\u00b10.24 284.6 FAST-AT-GA 45.52\u00b10.24 20.39\u00b10.19 16.25\u00b10.17 29.17\u00b10.32 6.79\u00b10.27 4.27\u00b10.15 592.7 FAST-BAT 45.80\u00b10.22 21.97\u00b10.21 17.64\u00b10.15 33.78\u00b10.23 8.83\u00b10.22 5.52\u00b10.14 572.4 ing. Yet, FAST-AT suffers from two main issues: (i) lack of stability, i.e., it has a large variance in performance (Li et al., 2020), and (ii) catastrophic over\ufb01tting, i.e., when training with strong adversaries, the robustness of the resulting model can drop signi\ufb01cantly. To alleviate these problems, Andriushchenko & Flammarion (2020) proposed FAST-AT-GA by penalizing FAST-AT using an explicit GA regularization. However, we will show that FAST-AT-GA encounters another problem: (iii) It improves robust accuracy (RA) at the cost of a sharp drop in standard accuracy (SA), leading to a poor accuracy-robustness tradeoff for large attack budget (e.g., \u03f5 = 16/255 in Table 1). Moreover, (iv) there has been no theoretical guarantee for the optimization algorithms used in FAST-AT and FAST-AT-GA. Given the limitations (i)(iv), we ask: How to design a \u2018fast\u2019 AT with improved stability, mitigated catastrophic over\ufb01tting, enhanced RA-SA tradeoff, and some theoretical guarantees? To address the above question, we formulate the AT problem as a uni\ufb01ed bi-level optimization (BLO) problem (Dempe, 2002). In the new formulation, the attack generation is cast as a constrained lower-level optimization problem, while the defense serves as the upper-level optimization problem. To the best of our knowledge, this is the \ufb01rst work that makes a rigorous connection between adversarial defense and BLO. Technically, we show that FAST-AT can be interpreted as BLO with linearized lower-level problems. Delving into the linearization of BLO, we propose a novel, theoreticallygrounded \u2018fast\u2019 AT framework, termed fast bi-level AT (FAST-BAT). Practically, Table 1 highlights some achieved improvements over FAST-AT and FAST-AT-GA: When a stronger train-time attack (i.e., \u03f5 = 16/255) is adopted, FAST-AT suffers from a large degradation of RA and SA, together with a very high variances. Although FAST-ATGA outperforms FAST-AT, it still incurs a signi\ufb01cant SA loss at \u03f5 = 16/255. In contrast, FAST-BAT produces the best robustness with a more graceful SA-RA tradeoff: As FAST-BAT brings an improvement of over 1.5% on RA in all experimental settings, its SA still remains at a high level, e.g. a signi\ufb01cant improvement on SA by 9.9% and 6.7% at \u03f5 = 16/255 for CIFAR-10 and CIFAR-100. Contributions. We summarize our contributions below. x (Formulation-wise) We propose a uni\ufb01ed BLO-based formulation for the robust training problem. Within this formulation, we show the conventional FAST-AT method is solving a low-level linearized BLO problem, rather than the vanilla min-max problem. This key observation not only provides a new interpretation of FAST-AT, but most importantly, also explains why FAST-AT is dif\ufb01cult to possess strong theoretical guarantees. y (Methodology-wise) We propose the new FAST-BAT algorithm based on our new understanding of FAST-AT. The key enabling technique is to introduce a new smooth lower-level objective of BLO for robust training. In contrast to MMO, BLO adopts a different optimization routine that requires implicit gradient (IG) computation. We derive the closed-form of IG for FAST-BAT. z (Experiment-wise) We made a comprehensive experimental study to demonstrate the effectiveness of FAST-BAT over SOTA baselines across datasets and model types. Besides its merit in robustness enhancement, we also show its improved stability, lifted accuracy-robustness trade-off, and mitigated catastrophic over\ufb01tting. \fRevisiting and Advancing Fast Adversarial Training Through the Lens of Bi-Level Optimization 2. Related work Adversarial attack. Adversarial attacks are techniques to generate malicious perturbations that are imperceptible to humans but can mislead the ML models (Goodfellow et al., 2014; Carlini & Wagner, 2017; Croce & Hein, 2020; Xu et al., 2019; Athalye et al., 2018b). The adversarial attack has become a major approach to evaluate the robustness of deep neural networks and thus, help build safe arti\ufb01cial intelligence in many high-stakes applications such as autonomous driving (Deng et al., 2020; Kumar et al., 2020) and surveillance (Thys et al., 2019; Xu et al., 2020). Adversarial defense and robust training at scale. Our work falls into the category of robust training, which was mostly built on MMO. For example, (Madry et al., 2018) established the framework of AT for the \ufb01rst time, always recognized as one of the most powerful defenses (Athalye et al., 2018a). Extended from AT, TRADES (Zhang et al., 2019b) sought the optimal balance between robustness and generalization ability. Further, AT-type defense has been generalized to the semi-/self-supervised settings (Carmon et al., 2019; Chen et al., 2020b) and integrated 1 with certi\ufb01ed defense techniques such as randomized smoothing (Salman et al., 2019). Despite the effectiveness of AT and its variants, how to speed up AT without losing performance remains an open question. Some recent works attempted to impose algorithmic simpli\ufb01cations to AT, leading to fast but approximate AT algorithms, such as \u2018free\u2019 AT (Shafahi et al., 2019), you only propagate once (YOPO) (Zhang et al., 2019a), FAST-AT (Wong et al., 2020), and FAST-AT regularized by gradient alignment (termed FAST-AT-GA) (Andriushchenko & Flammarion, 2020). In particular, FASTAT and FAST-AT-GA are the baselines most relevant to ours due to their low computational complexity. However, their defense performance is still unsatisfactory. For example, FAST-AT has poor training stability (Li et al., 2020) and suffers catastrophic over\ufb01tting when facing strong attacks (Andriushchenko & Flammarion, 2020). In contrast, FAST-AT-GA yields improved robustness but has a poor accuracy-robustness tradeoff (see Table 1). Bi-level optimization (BLO). BLO is a uni\ufb01ed hierarchical learning framework, where the objective and variables of the upper-level problem depend on the lower-level problems. The BLO problem in its most generic form is very challenging, and thus, the design of algorithms and theory for BLO focuses on special cases (Vicente et al., 1994; White & Anandalingam, 1993; Gould et al., 2016; Ghadimi & Wang, 2018; Khanduri et al., 2021; Ji et al., 2020; Hong et al., 2020). In practice, BLO has been successfully applied to meta-learning (Rajeswaran et al., 2019), data poisoning attack design (Huang et al., 2020), and reinforcement learning (Chen et al., 2019). Recently, references (Zuo et al., 2021) formulated the AT training problem using a Stackelberg Game, which is a special case of bi-level optimization. Different from ours, the goal of (Zuo et al., 2021) is to improve model generalization in natural language processing tasks, rather than improving model robustness in vision tasks. Further, note that in our formulation, we use a generic lower-level problem to model the adversarial learning problem, while in (Zuo et al., 2021) a multi-step unrolling of certain update mapping is used for this purpose. Importantly, with our formulation, it is possible to provide theoretical analysis for the resulting algorithms, under some conditions about the lower-level problem. As will be evident later, the existing BLO approach is not directly applied to adversarial defense due to the presence of the constrained non-convex lower-level problem (for attack generation). To the best of our knowledge, our work makes a rigorous connection between adversarial defense and BLO for the \ufb01rst time. 3. A Bi-Level Optimization View on FAST-AT Preliminaries on AT and FAST-AT. Let us consider the following standard min-max formulation for the robust adversarial training problem (Madry et al., 2018) minimize \u03b8 E(x,y)\u2208D \u0014 maximize \u03b4\u2208C \u2113tr(\u03b8, x + \u03b4, y) \u0015 , (1) where \u03b8 \u2208Rn denotes model parameters; D is the training set consisting (a \ufb01nite number) of labeled data pairs with feature x and label y; \u03b4 \u2208Rd represents adversarial perturbations subject to the perturbation constraint C, e.g., C = {\u03b4 | \u2225\u03b4\u2225\u221e\u2264\u03f5, \u03b4 \u2208[0, 1]} for \u03f5-toleration \u2113\u221e-norm constrained attack (normalized to [0, 1]); \u2113tr(\u00b7) is the training loss; (x + \u03b4) represents an adversarial example. The standard solver for problem (1) is known as AT (Madry et al., 2018). However, it has to call an iterative optimization method (e.g., K-step PGD attack) to solve the inner maximization problem of (1), which is computationally expensive. To improve its scalability, the FAST-AT algorithm (Wong et al., 2020) was proposed to take a single-step PGD attack for inner maximization. The algorithm backbone of FAST-AT is summarized below. FAST-AT algorithm Let \u03b8t be parameters at iteration t. The (t + 1)th iteration becomes (Wong et al., 2020): \u2022 Inner maximization by 1-step PGD: \u03b4 \u2190 \u2212PC (\u03b40 + \u03b1 \u00b7 sign (\u2207\u03b4\u2113tr(\u03b8t, x + \u03b40, y))) , where PC(a) denotes the projection operation that projects the point a onto C, i.e., PC(z) = arg min\u03b4\u2208C \u2225\u03b4 \u2212z\u22252 2, \u03b40 is a random uniform initialization within [0, 1], \u03b1 > 0 is a proper learning rate (e.g., 1.25\u03f5), and sign(\u00b7) is the element-wise sign operation. \fRevisiting and Advancing Fast Adversarial Training Through the Lens of Bi-Level Optimization \u2022 Outer minimization for model training: This can be conducted by any standard optimizer, e.g., \u03b8t+1 \u2190 \u2212\u03b8t \u2212\u03b2\u2207\u03b8\u2113tr(\u03b8t, x + \u03b4, y) where \u03b2 > 0 is a proper learning rate (e.g., cyclic learning rate), and \u03b4 is provided from the inner maximization step. A few remarks about FAST-AT are given below. x Roughly speaking, FAST-AT is a simpli\ufb01cation of AT using 1-step PGD for inner maximization. However, it is not entirely clear which problem FAST-AT is actually solving. If we take a closer look at the algorithm, we will see that the inner update only changes to the initial \u03b40, not the most recent \u03b4. Clearly, this scheme is fundamentally different from the classical gradient descent-ascent algorithm for min-max optimization (Razaviyayn et al., 2020), which alternatively updates the inner and outer variables. Therefore, it remains elusive if FAST-AT is solving the original problem (1). y Andriushchenko & Flammarion (2020) demonstrated that FAST-AT could lead to catastrophic over\ufb01tting when using strong adversaries for training. However, there was no grounded theory to justify the pros and cons of FAST-AT. We will show that BLO provides a promising solution. BLO: Towards a uni\ufb01ed formulation of robust training. BLO (bi-level optimization) is an optimization problem that involves two levels of optimization tasks, where the lower-level task is nested inside the upper-level task. More precisely, it has the following generic form: min u\u2208U f(u, v\u2217(u)) s.t. v\u2217(u) = arg minv\u2208V g(v, u) (2) where U and V are the feasible sets for the variables u and v, respectively; f(\u00b7) and g(\u00b7) are the upperand the lower-level objective functions, respectively. Intuitively, the BLO (2) can be used to formulate the adversarial training problem as the latter also involves two problems, one nested in the other. Importantly, it is more powerful than the minmax formulation (1) as it allows the two problems to have different objective functions. This \ufb02exibility provided by BLO is the key to the generality of our proposed framework. To make the above intuition precise, we use the upper-level problem to model the training loss minimization, while the lower-level problem to model the attack generation process, and consider the following BLO problem: min \u03b8 E(x,y)\u2208D[\u2113tr(\u03b8, x + \u03b4\u2217(\u03b8; x, y); y)] s.t. \u03b4\u2217(\u03b8; x, y) = arg min \u03b4\u2208C \u2113atk(\u03b8, \u03b4; x, y), (3) where the training loss function \u2113tr(\u00b7) has been de\ufb01ned in (1), and \u2113atk(\u00b7) denotes an attack objective. For the notation simplicity, in the subsequent discussion, we will not indicate the dependency of the functions \u2113tr, \u2113atk, and \u03b4\u2217with respect to (w.r.t.) the data samples (x, y). The formulation (3) has two key differences from (1): x When we choose \u2113atk = \u2212\u2113tr, problem (3) becomes equivalent to the min-max formulation (1). It follows that the BLO is suitable to formulate the adversarial training problem. Moreover, we will see shortly that the \ufb02exibility provided by choosing the lower-level objective independently of the upper-level one enables us to interpret FASTAT as solving a certain special form of the BLO problem. Note that prior to our work, it was not entirely clear what is the problem that FAST-AT is trying to solve. y BLO calls a different optimization routine from those to solve the original min-max problem (1). As will be evident later, even if we set \u2113atk = \u2212\u2113tr in (3), the BLO-enabled solver does not simplify to FAST-AT (see more details in Appendix B). This is because for a given data sample (x, y), the gradient for the upper-level problem of (3) yields: d\u2113tr(\u03b8,\u03b4\u2217(\u03b8)) d\u03b8 = \u2207\u03b8\u2113tr(\u03b8, \u03b4\u2217(\u03b8)) + d\u03b4\u2217(\u03b8)\u22a4 d\u03b8 | {z } IG \u2207\u03b4\u2113tr(\u03b8, \u03b4\u2217(\u03b8)), (4) where the superscript \u22a4denotes the transpose operation, and \u2207\u03b8\u2113tr(\u03b8, \u03b4\u2217(\u03b8)) denotes the partial derivative w.r.t. the \ufb01rst input argument \u03b8; and d\u03b4\u2217(\u03b8)\u22a4 d\u03b8 \u2208Rn\u00d7d, if exists, is referred to as implicit gradient (IG) because it is de\ufb01ned through an implicitly constrained optimization problem min\u03b4\u2208C \u2113atk. The dependence on IG is a \u2018\ufb01ngerprint\u2019 of BLO (1) in contrast to AT or FAST-AT. BLO-enabled interpretation of FAST-AT. Next, we demonstrate how FAST-AT relates to BLO. FAST-AT can be interpreted as an approximated stochastic gradient algorithm for solving the following lower-level linearized BLO. That is to say, FAST-AT is not solving the original min-max problem (1), but its linearized version below: min \u03b8 E(x,y)\u2208D[\u2113tr(\u03b8, \u03b4\u2217(\u03b8))] s.t. \u03b4\u2217(\u03b8)=arg min \u03b4\u2208C [(\u03b4 \u2212z)\u22a4sign(\u2207\u03b4=z\u2113atk(\u03b8, \u03b4)) + (\u03bb/2)\u2225\u03b4 \u2212z\u22252 2], (5) where z = \u03b40 and \u03bb = 1/\u03b1. Our justi\ufb01cation for the above interpretation is elaborated on below. x The simpli\ufb01ed lower-level problem of (5) leads to the closed-form solution: \u03b4\u2217(\u03b8) = arg min \u03b4\u2208C (\u03bb/2)\u2225\u03b4 \u2212z + (1/\u03bb)sign(\u2207\u03b4=z\u2113atk(\u03b8, \u03b4))\u22252 2 =PC (z \u2212(1/\u03bb)sign(\u2207\u03b4=z\u2113atk(\u03b8, \u03b4))) , (6) which is exactly given by the 1-step PGD attack with initialization z and learning rate (1/\u03bb). In the linearization used in (5), a quadratic regularization term (with regularization parameter \u03bb) is introduced to ensure the strong convexity of the lower-level objective within the constraint set \u03b4 \u2208C so as to achieve the unique minimizer (6). Note that imposing such a strongly convex regularizer is also commonly used to stabilize the convergence of MMO (min-max optimization) and BLO (Qian et al., 2019; Hong et al., 2020). If \fRevisiting and Advancing Fast Adversarial Training Through the Lens of Bi-Level Optimization we set z = \u03b40 and \u03bb = 1/\u03b1, (6) precisely depicts the inner maximization step used in FAST-AT. y By substituting (6) into the upper-level problem of (5), we can then use (4) to compute the stochastic gradients of the upper-level problem. If the stochastic gradient can be precisely computed, we can update the model parameters \u03b8 using SGD based on (4). That is, \u03b8\u2190 \u2212\u03b8\u2212\u03b2 d\u2113tr(\u03b8t,\u03b4\u2217(\u03b8)) d\u03b8 (with learning rate \u03b2). However, generally speaking, the IG function d\u03b4\u2217(\u03b8)\u22a4 d\u03b8 involved in (4) may not be differentiable, and even it is, the computation may not be easy. For our case, \u03b4\u2217(\u03b8) expressed in (6) involves both a projection and a sign operation, which can be particularly dif\ufb01cult to compute. To proceed, let us make the following approximations. We assume that the chain rule of the derivative of \u03b4\u2217(\u03b8) holds w.r.t. \u03b8, implying the differentiability of the projection operation and the sign operation. Then, based on the closed-form of \u03b4\u2217(\u03b8) in (6), IG is approximately equal to d\u03b4\u2217(\u03b8)\u22a4 d\u03b8 =0 , where we use two facts: (1) The linearization point z is independent of \u03b8, i.e. z = \u03b40; And (2) dsign(\u00b7) d\u03b8 = 0 holds almost everywhere. Clearly, the use of the gradient sign method simpli\ufb01es the IG computation. Further, the upper-level gradient is approximated by d\u2113tr(\u03b8,\u03b4\u2217(\u03b8)) d\u03b8 \u2248\u2207\u03b8\u2113tr(\u03b8,\u03b4\u2217(\u03b8)):=\u02dc h(\u03b8), and the upper-level optimization of problem (5) becomes \u03b8\u2190 \u2212\u03b8\u2212\u03b2\u02dc h(\u03b8), which is the same as the outer minimization step in FAST-AT. In a nutshell, the BLO solver of problem (5), which calls the IG computation based on (6), eventually reduces to the FAST-AT algorithm. The aforementioned analysis shows that FAST-AT can be viewed as using an approximated stochastic gradient algorithm to solve the linearized BLO (5), with the linearization point z and the regularization parameter \u03bb set as z = \u03b40 and \u03bb = 1/\u03b1. However, since a series of approximations have been used when arriving at the approximated gradient e h(\u03b8) used by FAST-AT, it is no longer clear if the resulting algorithm can still suf\ufb01ciently reduce the objective function of the upper-level problem. Additionally, based on the fact that the lower-level problem of (5) involves the discrete sign operator, it is unlikely (if not impossible) that any approximated stochastic gradient-based algorithms developed for it can possess any strong theoretical guarantees. 4. FAST-BAT: Advancing FAST-AT by BLO FAST-BAT and rationale. The key take-away from (5) is that the conventional FAST-AT adopts the sign of input gradient to linearize the lower-level attack objective. However, a more natural choice is to use the \ufb01rst-order Taylor expansion for linearization. By doing so, problem (5) can be modi\ufb01ed to the following form: min \u03b8 E(x,y)\u2208D[\u2113tr(\u03b8, \u03b4\u2217(\u03b8))] s.t. \u03b4\u2217(\u03b8)=arg min \u03b4\u2208C [(\u03b4 \u2212z)\u22a4(\u2207\u03b4=z\u2113atk(\u03b8, \u03b4)) + (\u03bb/2)\u2225\u03b4 \u2212z\u22252 2], (7) Similar to (6), problem (7) can be solved as: \u03b4\u2217(\u03b8) = PC (z \u2212(1/\u03bb)\u2207\u03b4=z\u2113atk(\u03b8, \u03b4)) . (8) In contrast to (6), the IG associated with (7) cannot be approximated by zeros since the gradient sign operation is not present in (8). To compute IG, the auto-differentiation (the chain rule) can be applied to the closed-form of \u03b4\u2217(\u03b8). This is also what has been used in (Zuo et al., 2021) that calls projected gradient descent unrolling to solve BLO problems. However, this will not give us an accurate and generalizable IG solution since the projection operation PC is not smooth everywhere and thus, the use of chain rule does not yield a rigorous derivation. In the following subsection, we address the IG challenge in a theoretically-grounded manner. IG theory for FAST-BAT. The problem of FAST-BAT (7) falls into a class of very challenging BLO problems, which requires constrained lower-level optimization. The unconstrained case is easier to handle since one can apply the implicit function theory to the stationary condition of the lower-level problem to obtain IG (Hong et al., 2020). Yet, in our case with constrained problems, a stationary point could violate the constraints, and thus, the stationary condition becomes non-applicable. Fortunately, in problem (7), we are dealing with a special class of lower-level constraints \u2013 linear constraints. Let us rewrite the constraints below: \u2225\u03b4\u2225\u221e\u2264\u03f5, \u03b4 \u2208[\u2212x, 1 \u2212x] \u21d0 \u21d2B\u03b4 \u2264b, (9) where B:= \u0014 I \u2212I \u0015 , b:= \u0014 min{\u03f51,1\u2212x} \u2212max{\u2212\u03f51,\u2212x} \u0015 . By exploiting the above linearly constrained problem structure, we show that the IG challenge associated with (7) can be addressed via Karush\u2013Kuhn\u2013Tucker (KKT) conditions. We present our main theoretical results (Proposition 1 and Corollary 1) below and refer readers to Appendix A for detailed derivation. Proposition 1 For the BLO problem (7), let g(\u03b8, \u03b4\u2217) denote its lower-level objective function evaluated at \u03b4\u2217given \u03b8, then the analytical form of IG (implicit gradient) is given by d\u03b4\u2217(\u03b8)\u22a4 d\u03b8 = \u2212\u2207\u03b8\u03b4g(\u03b8, \u03b4\u2217)\u2207\u03b4\u03b4g(\u03b8, \u03b4\u2217)\u22121 +\u2207\u03b8\u03b4g(\u03b8, \u03b4\u2217)\u2207\u03b4\u03b4g(\u03b8, \u03b4\u2217)\u22121B\u22a4 0 [B0\u2207\u03b4\u03b4g(\u03b8, \u03b4\u2217)\u22121B\u22a4 0 ]\u22121B0\u2207\u03b4\u03b4g(\u03b8, \u03b4\u2217)\u22121, (10) where \u03b4\u2217is given by (8) (the dependence on \u03b8 is omitted for ease of notation), \u2207\u03b8\u03b4g(\u03b8, \u03b4\u2217) \u2208Rn\u00d7d denotes the second-order partial derivatives evaluated at (\u03b8, \u03b4\u2217), and B0 denotes the sub-matrix of B that only corresponds to the active constraints at \u03b4\u2217, i.e., those in B\u03b4\u2217\u2264b satis\ufb01ed with equality. It is clear from (10) that the computation of IG requires the second-order derivatives as well as matrix inversion. This is computationally intensive. Recall from (7) that the Hessian matrix \u2207\u03b4\u03b4g of the lower-level objective function is given by \u2207\u03b4\u03b4g(\u03b8, \u03b4\u2217) = \u2207\u03b4\u03b4\u2113atk + \u03bbI. This inspires us to impose the Hessian-free approximation, i.e., \u2207\u03b4\u03b4\u2113atk = 0. \fRevisiting and Advancing Fast Adversarial Training Through the Lens of Bi-Level Optimization The rationale behind the Hessian-free assumption is that ReLU-based neural networks commonly lead to a piecewise linear decision boundary w.r.t. the inputs (MoosaviDezfooli et al., 2019; Alfarra et al., 2020), and thus, its second-order derivative (Hessian) \u2207\u03b4\u03b4\u2113atk is close to zero. In Sec. 5.2 and Appendix D, we will empirically show that the Hessian-free assumption is reasonable for both ReLU and non-ReLU neural networks. Thus, the Hessian matrix is approximated as: \u2207\u03b4\u03b4g(\u03b8, \u03b4\u2217(\u03b8)) = \u2207\u03b4\u03b4\u2113atk + \u03bbI = 0 + \u03bbI. (11) With (11), we can simplify closed-form of IG as below: Corollary 1 With the Hessian-free assumption, namely \u2207\u03b4\u03b4\u2113atk = 0, the IG (implicit gradient) of (7) is d\u03b4\u2217(\u03b8)\u22a4 d\u03b8 = \u2212(1/\u03bb)\u2207\u03b8\u03b4\u2113atk(\u03b8, \u03b4\u2217)HC, (12) where HC:= h 1p1<\u03b4\u2217 1 <q1e1\u00b7\u00b7\u00b71p1<\u03b4\u2217 d<qded i \u2208Rd\u00d7d and the function 1pi<\u03b4\u2217 i <qi \u2208{0, 1} denotes the indicator over the constraint of {\u03b4i | pi < \u03b4\u2217 i < qi}, which returns 1 if the constraint is satis\ufb01ed, \u03b4\u2217 i denotes the ith entry of \u03b4\u2217(\u03b8), pi = max{\u2212\u03f5, \u2212xi} and qi = min{\u03f5, 1 \u2212xi} characterize the boundary of the linear constraint (9) for the variable \u03b4i, and ei \u2208Rd denotes the basis vector with the ith entry being 1 and others being 0. FAST-BAT algorithm and implementation. Similar to FAST-AT or AT, the FAST-BAT algorithm also follows the principle of alternating optimization. Speci\ufb01cally, it consists of the IG-based upper-level gradient descent (4), interlaced with the lower-level optimal attack (8). We summarize the FAST-BAT algorithm below. \u2022 Lower-level solution: Obtain \u03b4\u2217(\u03b8t) from (8); \u03b4\u2217(\u03b8) = PC (z \u2212(1/\u03bb)\u2207\u03b4=z\u2113atk(\u03b8, \u03b4)) . \u2022 Upper-level model training: Integrating the IG (12) into (4), call SGD to update the model parameters as: \u03b8t+1 = \u03b8t \u2212\u03b11\u2207\u03b8\u2113tr(\u03b8t, \u03b4\u2217) \u2212\u03b12(\u22121 \u03bb)\u2207\u03b8\u03b4\u2113atk(\u03b8t, \u03b4\u2217)HC\u2207\u03b4\u2113tr(\u03b8t, \u03b4\u2217), (13) where \u03b11, \u03b12 > 0 are learning rates associated with the model gradient and the IG-augmented descent direction. It is clear from (13) that to train a robust model, FAST-BAT can be decomposed into the regular FAST-AT update (i.e., the term multiplied by \u03b11) and the additional update that involves IG, (i.e., the term multiplied by \u03b12). To implement FAST-BAT, we highlight some key hyper-parameter setups that are different from FAST-AT (Wong et al., 2020) and FAST-AT-GA (Andriushchenko & Flammarion, 2020). Remark on implementation details Next, we discuss the practical setups used in Fast-BAT (more details in Appendix C). First, 1/\u03bb serves as the attack step size as shown in (8). We refer readers to Table 7 for a sensitivity analysis of \u03bb. In Fast-BAT, the hyperparameter \u03b12 is newly introduced, and is set differently from \u03b11 so as to control the descent error associated with the coupled second-order/\ufb01rstorder stochastic derivatives, i.e., the \u03b12-term in (13). In Sec. 7, we show that a proper \u03b12 helps mitigate catastrophic over\ufb01tting. In practice, we set \u03b12/\u03bb = 0.1\u03b11. To specify the linearization point z in (7), we investigate two types of linearization schemes: (1) the random constant linearization (random uniform and random corner linearization) and (2) the 1-step perturbation warm-up-based linearization (1-step sign-based and 1-step w/o sign PGD). These linearization schemes have computational complexities up to the onestep attack generation. Empirically, we \ufb01nd that FAST-BAT using \u201c1-step PGD w/o sign\" leads to the best defensive performance (see Table 9). We follow this experiment setup in the sequel. Remark on convergence analysis In Appendix E, we prove that under some assumptions on the gradient bias, Fast-BAT converges to a \ufb01rst-order stationary point or its small neighborhood in the rate of O(1/ \u221a T), where T is the iteration number of model updates. The main analysis dif\ufb01culty lies in the last term of the model updating rule (13), which involves two coupled derivatives built upon the same mini-batch. 5. Experiments 5.1. Experiment Setup Datasets and model architectures. We will evaluate the effectiveness of our proposal under CIFAR-10 (Krizhevsky & Hinton, 2009), CIFAR-100 (Krizhevsky & Hinton, 2009), Tiny-ImageNet (Deng et al., 2009), and ImageNet (Deng et al., 2009). Unless speci\ufb01ed otherwise, we will train DNN models PreActResNet (PARN)-18 (He et al., 2016b) for all datasets except ImageNet, and ResNet (RN)-50 (He et al., 2016a) for ImageNet. As a part of the ablation study, we also train larger models PARN-50 and WideResNet (WRN)16-8 (Zagoruyko & Komodakis, 2016) on CIFAR-10. Some preliminary ImageNet results are reported in Appendix D. Baselines. We focus on three baselines, namely, FASTAT (Wong et al., 2020), FAST-AT-GA (Andriushchenko & Flammarion, 2020), and PGD-2-AT (Madry et al., 2018), and refer readers to comparisons with more baselines (PGD7-AT (Madry et al., 2018), BACKSMOOTH (Chen et al., 2020a), FREE-AT (Shafahi et al., 2019), ATTA (Zheng et al., 2020), YOPO (Zhang et al., 2019a)) in Appendix D. Here PGD-2-AT and PGD-7-AT stand for the 2-step and \fRevisiting and Advancing Fast Adversarial Training Through the Lens of Bi-Level Optimization 7-step PGD attack-based AT, respectively. The primal criterion of baseline selection is computational complexity. All the methods except PGD-7-AT consume the training time of the same order, while PGD-7-AT serves as a reference to the performance of the non-accelerated robust training method. We stress that FAST-AT-GA is the strongest baseline to the best of our knowledge in terms of improving robustness and mitigating robust catastrophic over\ufb01tting (Andriushchenko & Flammarion, 2020). Training details. We choose the training perturbation strength \u03f5 \u2208{8, 16}/255 for CIFAR-10, CIFAR-100, and Tiny-ImageNet; and \u03f5 = 2/255 for ImageNet following (Wong et al., 2020; Andriushchenko & Flammarion, 2020). Throughout the experiments, we utilize an SGD optimizer with a momentum of 0.9 and weight decay of 5 \u00d7 10\u22124. For CIFAR-10, CIFAR-100 and Tiny-ImageNet, we train each model for 20 epochs in total, where we use the cyclic scheduler to adjust the learning rate. The learning rate linearly ascends from 0 to 0.2 within the \ufb01rst 10 epochs and then reduces to 0 within the last 10 epochs. Our batch size is set to 128 for all settings. In the implementation of FAST-BAT, we follow the dataset-agnostic hyperparameter scheme for \u03bb, such that \u03bb = 255/5000 for \u03f5 = 8/255 and \u03bb = 255/2500 for \u03f5 = 16/255 for CIFAR-10, CIFAR-100 and Tiny-ImageNet. For ImageNet, we strictly follow the setup given by (Wong et al., 2020) and we choose the traintime attack budget as \u03f5 = 2/255. For each method, we use the early stopping method to pick the model with the best robust accuracy, following (Rice et al., 2020). All the experiments are conducted on a single GeForce RTX 3090 GPU.All the baselines are trained with the recommended con\ufb01gurations in their of\ufb01cial GitHub repos. We refer readers to Appendix C for more details on the training setup. Evaluation details. For adversarial evaluation, we report robust test accuracy (RA) of a learned model against PGD attacks (Madry et al., 2018) (RA-PGD). Unless otherwise speci\ufb01ed, we set the test-time perturbation strength (\u03f5) the same as the train-time value, and take 50-step PGD with 10 restarts all the datasets. Since AutoAttack (Croce & Hein, 2020) is known as the strongest robust benchmark evaluation metric (given as an ensemble attack), we also measure robust accuracy against AutoAttack, termed RAAA. Further, we measure the standard accuracy (SA) against natural examples. Results are averaged over 10 independent trials. We would like to highlight that all the methods share the same batch size, epoch number, and training hardware. Thus, the time consumption per epoch reported in Table 1 and Table 5 will serve as a fair indicator of the algorithm complexity of different methods. Table 2. SA, RA-PGD, and RA-AA of different robust training methods in the setup (CIFAR-10, PARN-18 training with \u03f5 = 8/255) and (CIFAR-10, PARN-18 training with \u03f5 = 16/255), respectively. All the results are averaged over 10 independent trials with different random seeds. CIFAR-10, PARN-18 trained with \u03f5 = 8/255 Method SA (%) RA-PGD (%) RA-AA (%) \u03f5 = 8 \u03f5 = 16 \u03f5 = 8 \u03f5 = 16 FAST-AT 82.39\u00b10.44 45.49 \u00b10.41 9.56 \u00b10.26 41.87 \u00b10.15 7.91 \u00b10.06 FAST-AT-GA 79.71\u00b10.44 47.27 \u00b10.42 11.57\u00b10.32 43.24 \u00b10.27 9.48 \u00b10.15 PGD-2-AT 81.97\u00b10.41 44.62 \u00b10.39 9.39 \u00b10.32 41.73 \u00b10.20 7.54\u00b10.25 FAST-BAT 79.97 \u00b10.12 48.83 \u00b10.17 14.00 \u00b10.21 45.19 \u00b10.12 11.51 \u00b10.20 CIFAR-10, PARN-18 trained with \u03f5 = 16/255 FAST-AT 44.15\u00b17.27 37.17 \u00b10.74 21.83 \u00b11.32 31.66 \u00b10.27 12.49 \u00b10.33 FAST-AT-GA 58.29 \u00b11.32 43.86 \u00b10.67 26.01 \u00b10.16 38.69 \u00b10.56 17.97 \u00b10.33 PGD-2-AT 68.04 \u00b10.30 48.79 \u00b10.31 24.30 \u00b10.46 41.59 \u00b10.22 15.40 \u00b10.29 FAST-BAT 68.16 \u00b10.25 49.05 \u00b10.12 27.69 \u00b10.16 43.64 \u00b10.26 18.79 \u00b10.24 Table 3. Performance of different robust training methods under different model types. All the models are both trained and evaluated with the same perturbation strength \u03f5. Model Method SA(%) (\u03f5 = 8/255) RA-PGD(%) (\u03f5 = 8/255) SA(%) (\u03f5 = 16/255) RA-PGD(%) (\u03f5 = 16/255) PARN-50 FAST-AT 73.15\u00b16.10 41.03\u00b12.99 43.86\u00b14.31 22.08\u00b10.27 FAST-AT-GA 77.40\u00b10.81 46.16\u00b10.98 42.28\u00b16.69 22.87\u00b11.25 PGD-2-AT 83.53\u00b10.17 46.17\u00b10.59 68.88\u00b10.39 22.37\u00b10.41 FAST-BAT 78.91\u00b10.68 49.18\u00b10.35 69.01\u00b10.19 24.55\u00b10.06 WRN-16-8 FAST-AT 84.39\u00b10.46 45.80\u00b10.57 49.39\u00b12.17 21.99\u00b10.41 FAST-AT-GA 81.51\u00b10.38 48.29\u00b10.20 45.95\u00b113.65 23.10\u00b13.90 PGD-2-AT 85.52\u00b10.14 45.47\u00b10.14 72.11\u00b10.33 23.61\u00b10.16 FAST-BAT 81.66\u00b10.54 49.93\u00b10.36 68.12\u00b10.47 25.63\u00b10.44 5.2. Results Overall performance of FAST-BAT. In Table 1, Table 2, and Table 6, we compare the overall performance of our proposed FAST-BAT with baselines. x We \ufb01nd that FAST-BAT consistently outperforms the other baselines across the datasets and attack types. In Table 1, FAST-BAT improves the RA-PGD performance consistently by over 1.5% and RA-AA by around 1% across all the datasets on both attack strengths. For stronger attacks with a larger perturbation budget, the advantage of FAST-BAT is even clearer, e.g. a gain of over 2% in TinyImageNet with \u03f5 = 16/255. On ImageNet, FAST-BAT outperforms FAST-AT by 1.23% with \u03f5 = 2/255. y FAST-BAT leads to a much better SA-RA trade-off compared with the baselines. For example, in Table 1, the improvement in RA is not at cost of a huge drop in SA. Instead, when models are trained with \u03f5 = 16/255, FASTBAT even enjoys a signi\ufb01cant boost over FAST-AT-GA in SA by 9.9%, 6.7%, and 4.6% for CIFAR-10, CIFAR-100, and Tiny-ImageNet respectively. Performance under different model architectures. Besides PARN-18 reported above, Table 3 presents results on \fRevisiting and Advancing Fast Adversarial Training Through the Lens of Bi-Level Optimization (a) Without early stopping (b) With early stopping Figure 1. RA-PGD of different robust training methods for CIFAR10 with the same training and evaluation attack strengths. both deeper (PARN-50) and wider (WRN-18-6) models. As we can see, FAST-BAT consistently yields RA improvement over other methods. We also note that PGD-2-AT could be a competitive baseline in terms of SA. In contrast to FASTAT and FAST-AT-GA, FAST-BAT is the only approach that yields an evident RA improvement over PGD-2-AT. Mitigation of robust catastrophic over\ufb01tting. As shown in (Andriushchenko & Flammarion, 2020), FASTAT suffers robust catastrophic over\ufb01tting when the traintime and test-time attack strength \u03f5 grows. Following (Andriushchenko & Flammarion, 2020), Figure 1 presents two RA-PGD trajectories, i.e., training without early stopping and training with early stopping, versus the train/test-time \u03f5. As we can see, FAST-AT encounters a sharp RA drop when \u03f5 > 8 when early stopping is not used, consistent with the FAST-AT-GA paper (Andriushchenko & Flammarion, 2020). With early stopping, the over\ufb01tting of RA can be alleviated to some extent for FAST-AT, but its performance still remains the worst. Moreover, different from (Andriushchenko & Flammarion, 2020), we \ufb01nd that PGD-2AT yields resilient performance against catastrophic over\ufb01tting. Our implementation gives a more positive baseline for PGD-2-AT, since the implementation in (Andriushchenko & Flammarion, 2020) did not use random initialization to generate train-time attacks. Furthermore, Figure 1 shows that FAST-BAT mitigates the issue of catastrophic over\ufb01tting and yields improved RA over other methods. We highlight that such an achievement is free of any robustness stability regularization, like gradient alignment used in FAST-AT-GA. Gradient alignment for \u2018free\u2019. As shown by (Andriushchenko & Flammarion, 2020), catastrophic over\ufb01tting occurs with the local non-linearity of deep networks, which can be measured by the gradient alignment (GA) score: E(x,y)\u223cD,\u03b7\u223cU([\u2212\u03f5,\u03f5]d)[cos(\u2207x\u2113tr(x, y; \u03b8), \u2207x\u2113tr(x + \u03b7, y; \u03b8))], where U denotes the randomly uniform distribution. GA is a key performance indicator to measure the appearance of robustness catastrophic over\ufb01tting, as catastrophic over\ufb01tting is always accompanied by a sharp GA drop in the Figure 2. The training curve of GA score over different training methods on CIFAR-10. Table 4. Robustness against adaptive attacks (RA-PGD) and transfer attacks (RA-Transfer Attack). Naturally trained PARN-18, PARN-50, and WRN-16-8 serve as source victim models for attack generation with PGD-20 (\u03f5 = 8/255) and PARN-18 robusti\ufb01ed by different methods as target model for transfer attack evaluation. Method RA-PGD(%) RA-Transfer Attack(%) PARN-18 PARN-50 WRN-16-8 FAST-AT 45.44\u00b10.06 76.35\u00b10.12 76.94\u00b10.14 77.23\u00b10.21 PGD-2-AT 44.71\u00b10.04 77.56\u00b10.14 78.64\u00b10.12 78.84\u00b10.17 FAST-AT-GA 47.31\u00b10.05 77.34\u00b10.13 78.34\u00b10.13 78.53\u00b10.12 FAST-BAT(Ours) 48.67\u00b10.05 78.03\u00b10.15 79.93\u00b10.12 79.21\u00b10.15 training trajectory of the robustly trained model. In our paper, we calculate the GA for each method on the test set at the end of each epoch throughout the training process. Figure 1 suggested that FAST-BAT can mitigate over\ufb01tting without explicit GA regularization, and Figure 2 presents the GA score versus the training epoch number. As can be seen, FAST-BAT automatically enforces GA and remains very close to FAST-AT-GA, which maximizes GA with an explicit regularization. Therefore, a high GA score may just be a necessary but not a suf\ufb01cient condition for avoiding catastrophic over\ufb01tting. As additional evidence, Fig. 3 shows that similar to FAST-AT-GA, FAST-BAT has a \ufb02atter loss landscape than FAST-AT as well. Therefore, a direct penalization on the input gradient norm may not achieve the state-of-the-art model robustness. Sanity check for obfuscated gradients As pointed out by (Athalye et al., 2018a), model robustness could be overestimated due to obfuscated gradients. The model with obfuscated gradients could have \u2018obfuscated\u2019 stronger resilience to white-box (adaptive) attacks than black-box (transfer) attacks. To justify the validity of FAST-BAT, Table 4 summarizes the comparison between our proposal and the other baselines when facing adaptive and transfer attacks. Firstly, for all the methods, RA increases if the transfer attack is applied, implying that the transfer attack is weaker than the \fRevisiting and Advancing Fast Adversarial Training Through the Lens of Bi-Level Optimization Table 5. Performance of Hessian-free and Hessian-aware FASTBAT on CIFAR-10. We train and evaluate with the same attack budgets \u03f5 = 8/255 and \u03f5 = 16/255 to show the in\ufb02uence brought by Hessian matrix. Method SA(%) (\u03f5 = 8/255) RA-PGD(%) (\u03f5 = 8/255) SA(%) (\u03f5 = 16/255) RA-PGD(%) (\u03f5 = 16/255) Time (s/epoch) Hessian-free 79.97 \u00b10.12 48.83 \u00b10.17 68.16 \u00b10.25 27.69 \u00b10.16 61.4 Hessian-aware 79.62 \u00b10.17 49.13 \u00b10.14 67.82 \u00b10.23 27.82 \u00b10.19 82.6 adaptive attack. This is desired in the absence of obfuscated gradients. Moreover, FAST-BAT consistently outperforms the other baselines against both adaptive and transfer attacks. The absence of obfuscated gradients can also be justi\ufb01ed by the \ufb02atness of the adversarial loss landscape in Figure. 3. The validity of the Hessian-free assumption on ReLUbased neural networks. In Corollary 1, the Hessian-free assumption, i.e.\u2207\u03b4\u03b4\u2113atk = 0, was made to simplify the computation of implicit gradient. To justify this assumption we conduct experiments to compare the Hessian-free FASTBAT with the Hessian-aware version. In Hessian-aware FAST-BAT, the implicit gradient is calculated based on (23). In Table 5, the results do not indicate much difference when Hessian is used. However, the extra computations required to evaluate the Hessian heavily slows down FAST-BAT as around 30% more time is needed. Therefore, the Hessianfree assumption is reasonable and also necessary in terms of the ef\ufb01ciency of the algorithm. We also justify this assumption on some non-ReLU neural networks and for more results please refer to Appendix D. 6."
+                },
+                {
+                    "url": "http://arxiv.org/abs/1509.01763v5",
+                    "title": "Implementing Support for Pointers to Private Data in a General-Purpose Secure Multi-Party Compiler",
+                    "abstract": "Recent compilers allow a general-purpose program (written in a conventional\nprogramming language) that handles private data to be translated into secure\ndistributed implementation of the corresponding functionality. The resulting\nprogram is then guaranteed to provably protect private data using secure\nmulti-party computation techniques. The goals of such compilers are generality,\nusability, and efficiency, but the complete set of features of a modern\nprogramming language has not been supported to date by the existing compilers.\nIn particular, recent compilers PICCO and the two-party ANSI C compiler strive\nto translate any C program into its secure multi-party implementation, but\ncurrently lack support for pointers and dynamic memory allocation, which are\nimportant components of many C programs. In this work, we mitigate the\nlimitation and add support for pointers to private data and consequently\ndynamic memory allocation to the PICCO compiler, enabling it to handle a more\ndiverse set of programs over private data. Because doing so opens up a new\ndesign space, we investigate the use of pointers to private data (with known as\nwell as private locations stored in them) in programs and report our findings.\nBesides dynamic memory allocation, we examine other important topics associated\nwith common pointer use such as reference by pointer/address, casting, and\nbuilding various data structures in the context of secure multi-party\ncomputation. This results in enabling the compiler to automatically translate a\nuser program that uses pointers to private data into its distributed\nimplementation that provably protects private data throughout the computation.\nWe empirically evaluate the constructions and report on performance of\nrepresentative programs.",
+                    "authors": "Yihua Zhang, Marina Blanton, Ghada Almashaqbeh",
+                    "published": "2015-09-06",
+                    "updated": "2017-06-30",
+                    "primary_cat": "cs.CR",
+                    "cats": [
+                        "cs.CR",
+                        "cs.PL"
+                    ],
+                    "main_content": "Introduction Recent advances in secure multi-party computation make it feasible to securely compute with private data belonging to di\ufb00erent organizations even for complex functionalities. Furthermore, together with ubiquitous proliferation of cloud computing services, these techniques give rise to secure computation outsourcing. For these reasons, the research community has recently developed a number of compilers for transforming a general-purpose program into the corresponding secure distributed implementation (see, e.g., [9, 23]). These tools aim at generality and are designed to \u2217Work done while at the University of Notre Dame. 1 \ftranslate a program written in a conventional programming language into an equivalent program that uses secure computation techniques to protect private data. They also aid usability and make it easier for a programmer without extensive knowledge of secure computation techniques to produce a protocol that can be securely executed in a distributed environment. It has been long known that any computable function can be securely evaluated by multiple participants if it is represented as an arithmetic or Boolean circuit. This representation, however, is not always obvious or known or may otherwise signi\ufb01cantly increase the program size. Existing compilers remove the need for the programmer to perform this translation manually and assemble secure implementations from e\ufb03cient building blocks for elementary operations. Thus, e\ufb03ciency of the resulting secure computation is also one of the goals that compilers target. Furthermore, the ability to support both private (i.e., protected) and public (i.e., not protected) data or variables in a single program adds a level of complexity to the implementation because of the need to support interaction between public and private variables and secure data \ufb02ow enforcement. While the design goal of several compilers was to support any feature of a general-purpose programming language (such as C in [9] and [23]), all such compilers we are aware of have limitations. In particular, the original version of the PICCO compiler [23] provided no direct support for C pointers (i.e., pointers were supported only in the form of static arrays) and, as a result, no support for dynamic memory allocation other than static arrays. Similarly, the original version of the two-party compiler for ANSI C [9] supported pointers only in the form of statically allocated arrays restricted to a constant size and had additional limitations (such as support for \ufb02oating point arithmetic was not available in the open source CBMC that the compiler builds upon). Thus, support for C-like pointers \u2013 or, in other programming languages, support for the features that pointers enable such as dynamic memory allocation, reference by pointer or address, and building data structures \u2013 is the most crucial part of a general-purpose program that is currently unavailable in existing compilers. Adding this support is thus the focus of our work. In this paper, we extend the PICCO compiler [23] with pointer support. PICCO1 is a sourceto-source translator that takes as an input a program written in the C programming language with variables to be protected marked as private and produces a C program that implements the computation using secure multi-party computation techniques based on linear secret sharing. We view PICCO as an attractive compiler choice because of the \ufb02exibility of the setting it uses. In particular, the setting assumes three groups of participants: (i) input parties who hold private inputs into the computation, (ii) computational parties who perform secure computation on secretshared data, and (iii) output parties who are entitled to learning the result of the computation. The composition of these three groups can be arbitrary (in particular, including the same, overlapping, or non-overlapping groups), which makes the setting suitable for secure multi-party computation (SMC), delegation of the computation by multiple data owners to a subset of them or other suitable entities or secure computation outsourcing by one or more parties. This \ufb02exibility follows from the use of secret sharing techniques and may or may not be present in tools that build on alternative secure computation techniques (such as, e.g., garbled circuit evaluation). With linear secret sharing, before secure computation can commence, each input party splits her private inputs into n > 2 secret shares, where n is the number of computational parties, and communicates each share to a respective computational party. The computational parties then proceed with evaluating the function on secret-shared data and communicate their shares of the result to the output parties who reconstruct the output using their shares. Any linear combination of secret-shared integers is performed locally, but multiplication of secret-shared integers constitutes the elementary interactive operation. Performance is then measured in the total number of inter1Available from GitHub at https://github.com/picco-team/picco. 2 \factive operations as well as the number of sequential interactions or rounds, and recent solutions based on secret sharing aim at minimizing overhead using both metrics. When PICCO is used to perform source-to-source translation, the input program is a conventional C program where each variable is marked to be either private or public. All computation with private variables is transformed into secure arithmetic on shared data, while operations with public variables that do not interact with private data are left unchanged. In addition to specifying private/public quali\ufb01es for each variable, for performance reasons PICCO also allows the programmer to mark the places where computation can proceed concurrently (i.e., to decrease the number of computation rounds), which also extends the conventional C syntax. Adding pointer to support to a program that manipulates private data not only extends the compiler to handle the full range of C programs (that do not violate secrecy of private data), but also permits important features of programming languages, treatment of which, to the best of our knowledge, has not been done before. As part of this work, we thus explore how pointers to private data (including pointers with private locations) can be implemented and discuss our design decisions. Having added support of pointers to private data, we further study common uses of pointers in programs and the impact our implementation has on those language features. For example, we evaluate passing arguments by references, dynamic memory allocation, and pointer casting. Based on our analysis as well as empirical evaluation, several of these features introduce only marginal costs. Also, one of the important topics studied in this is work is the use of pointerbased data structures written for private data. Our results indicate that the use of pointers (to private data) is very attractive and maintains high e\ufb03ciency for several popular data structures. In some other cases, in particular when working with sorted data, privately manipulating pointers increases complexity of data structure operations and it might be desirable to pursue alternative implementations. We would like to emphasize that it is not the goal of this paper to try to develop most e\ufb03cient implementations for di\ufb00erent data structures. Instead, the goal is to determine how pointers to private data can be supported at a low possible cost and to what performance of typical programs that might lead. We note that, depending on the program structure, asymptotic complexity of a translated program might be higher than that of the original. For example, consider an if-then-else statement with a private condition (e.g., conditional statements used in traversing a binary tree). When data privacy is not required, only one of the two branches will be executed, but with any compiler that produces a secure implementation both branches will have to be evaluated to hide the result of the private condition. Then with a sequence of n nested if-then-else statements, in the worst case the secure program might have to execute O(2n) instructions where the original program would execute only O(n). This means that the general translation approach can lead to an exponential increase in the runtime for programs of practical relevance. As part of this work we show that data structures that utilize pointers to private data cover the entire spectrum of possibilities: in one extreme, they result in no asymptotic increase over conventional non-secure counterparts, and in another extreme, the increase is exponential. This provides insights on when natural pointer use is very attractive and when other, alternative implementations might be desired. While for many data structures alternative, non-pointer-based implementations may be possible, we note that our extension of PICCO with pointers enables support for an important aspect of modern programs otherwise not available in any secure multi-party compiler we are aware of, which is dynamic memory allocation. Dynamic memory allocation is essential for a general-purpose programming language, but has not been systematically studied in the context of secure multiparty computation (e.g., even publications and compilers that run secure computation on data sets of a large size assume that the size of the data is \ufb01xed and known at compilation time). As an example, consider an application in which data items arrive over time. A pointer-based linked 3 \flist will allow for a natural and graceful mechanism of memory management, that unlike statically allocated arrays does not require complex provisions for allocating a larger bu\ufb00er when the current one becomes full, moving the data, or merging previously allocated bu\ufb00ers. The rest of this paper is organized as follows: We \ufb01rst give a brief overview of related work in Section 2. After presenting background information in Section 3, we proceed with presenting our solution for supporting pointers to private data in Section 4. In Section 5, we discuss common uses of pointers in programming, such as passing arguments by reference, dynamic memory allocation, array manipulation, and pointer casting and their underlying implementation in our framework. Section 6 next summarizes operations with pointers to private data and formally shows security of the design. Section 7 analyzes various data structures built using pointers to private data. Lastly, Section 8 presents the results of performance evaluation of representative programs that utilize pointers (to private data) and Section 9 concludes this work. 2 Related Work In this section we review the most closely related work on SMC compilers and secure/oblivious data structures. Regarding the compilers, Fairplay [15] was a pioneer work that enables compilation of secure two-party protocols based on garbled circuits. Its extension to multiple parties, FairplayMP [3], implements secure computation using Boolean circuits and secret sharing techniques. TASTY [8] is another two-party SMC compiler that combines garbled circuit techniques with those based on homomorphic encryption. Sharemind [5] and VIFF [7] are multi-party compilers based on custom additive 3-party secret sharing and standard threshold linear secret sharing, respectively. All of the above compilers use custom domain-speci\ufb01c languages to represent user programs. The two-party compiler for ANSI C [9] and PCF [12] both use two-party garbled circuit techniques, where the former\u2019s goal is to support general purpose C programs, while the latter uses a new circuit format and employs optimizations to reduce the compilation time and storage. Lastly, TinyGarble [20] uses hardware synthesis to optimize garbled circuits for two-party computation. All of these compilers require linear in the size of memory work to access memory at a private location. SCVM [13], on the other hand, is an automated compiler that utilizes oblivious RAM (ORAM) and targets two-party computation. ObliVM [14] is another ORAM-based secure two-party computation compiler that transforms programs written in high level abstractions to optimized garbled circuit implementations. Finally, a recent compiler Frigate [17] was designed to guarantee correctness of programs compiled into circuits for secure two-party computation. To support data structures in the SMC framework, several solutions [21, 10, 16, 22] have been proposed. The main motivation of this line of work is the need to store and manipulate private data in an e\ufb03cient and \ufb02exible manner. Toft [21] proposed a private priority queue that has a deterministic access pattern as opposed to randomized ones in ORAM-based data structures. On the other hand, Keller and Scholl [10] introduced implementations of arrays, dictionaries, and priority queues based on various \ufb02avors of ORAM implementations. Mitchell and Zimmerman [16] also provide implementations of stacks, queues, and priority queues based on oblivious data compaction and an o\ufb04ine variant of ORAM. Wang et al. [22] proposed implementations of maps, sets, priority queues, stacks, and deques based on ORAM techniques modi\ufb01ed for speci\ufb01c data access patterns. Di\ufb00erent from all of these publications, our work includes extending the PICCO compiler to support dynamic data structures in a generic way as found in general purpose programming languages. That is, the programmer has the basic tools and primitives that enable her to build any desired data structure. In our implementation, a pointer to private data may store one or more locations where the 4 \fdata might reside, which in the worst case is linear in the program\u2019s memory size. ORAM-based techniques, on the other hand, guarantee that when an item is accessed at a private location, the number of accessed memory locations is polylogarithmic in the total memory size. Thus, our general solution may or may not be faster than using ORAM, depending on both the program and the data size. As we discuss later in this work, employing ORAM techniques can be bene\ufb01cial for certain data structures (and su\ufb03ciently large data sets). Building custom data structures, however, is beyond the scope of this work. One of the applications that the compiler can naturally be used for once support for pointers to private data is in place is evaluation of a context-free grammar on private data (implemented as a shift-reduce parser using a stack). The grammar can be either public or private, and in the latter case execution will correspond to evaluation of private expressions/programs on private data. Techniques for evaluation of private programs (on private data) are a separate area of research, discussion of which is beyond the scope of this work, but the reader may refer to recent results in this areas such as those in [11, 19]. 3 Background Information PICCO uses Shamir secret sharing [18] for implementing secure arithmetic and other operations on private data. It is an (n, t)-threshold linear secret sharing scheme, in which a private value is represented using n > 2 secret shares, one held by each computational party. Then any t + 1 or more shares can be used to reconstruct the private value, while t or fewer parties cannot learn any information about the shared value (which is perfectly protected in the information-theoretic sense). In a linear secret sharing scheme, a linear combination of secret-shared values can be performed by each computational party locally, without any interaction, while multiplication of secret-shared values requires communication between all of them. With Shamir secret sharing, computation takes place over a \ufb01eld of a desired size (larger than any value that needs to be represented). A secret s is represented using a random polynomial of degree t with the free coe\ufb03cient set to s, and each share corresponds to the evaluation of the polynomial on a distinct non-zero point. Given t + 1 or more shares, the secret can be reconstructed using Lagrange interpolation. Then addition or subtraction of secret-shared values, or multiplication of a secret-shared value by a known integer can be performed by each party locally using its shares. Multiplication involves multiplying two shares, which raises the corresponding polynomial degree to 2t, and resharing and interpolating the result to bring the degree of the corresponding polynomial from 2t to t. This imposes the requirement that t < n/2. With the way multiplication is performed, it is also possible to evaluate any multivariate polynomial of degree 2 over secret-shared integers with a single interaction. That is, we \ufb01rst evaluate the polynomial and re-share the overall result instead of doing so for the intermediate products. This serves as a powerful optimization tool when, for instance, the dot product of two secret shared vectors/arrays needs to be computed, which results only in a single interaction. While the regular \ufb01eld operations achieve perfect secrecy, implementation of some of the basic operations used in PICCO (such as comparisons, division, etc.) is statistically secure, which requires the bitlength of the \ufb01eld elements to be increased by the statistical security parameter. This slightly increases the cost of \ufb01eld operations, as well as the amount of communication associated with transmitting \ufb01eld elements. The optimal size of \ufb01eld elements is automatically determined by PICCO for each program it compiles. Performance of these techniques is measured in terms of the total number of elementary interactive operations (\ufb01eld multiplications or reconstructions of a value from its shares) as well as the number of sequential interactions or rounds. For that reason, PICCO supports a number of 5 \foptimizations to reduce the round complexity of programs that it outputs through concurrent or batch execution. Support for pointers to private data (and the corresponding functionalities such as dynamic memory management) was the only missing functionality in PICCO. Thus, we modify the compiler to enable it to compile user programs that contain pointers to private data, which was not previously available. Our changes to the compiler a\ufb00ect only pointers to private data and the introduction of two built-in functions for memory allocation and deallocation (called pmalloc and pfree) associated with pointers to private data. 4 Adding Pointer Support Recall that in C a pointer is a variable of a special type that stores a location in memory at which data of a particular type can be located. Because a pointer stores an address, it can be treated very generally with the possibility of directly manipulating pointers, changing the addresses they store, dereferencing a pointer to access the data to which it points, casting a pointer of a particular type to a pointer of another type, and using pointers to functions. 4.1 Working Toward a Solution When working with pointers in the presence of private data, besides traditional C pointers to public variables, we can distinguish between pointers to private data that (i) point to a single known location where the private data is stored and (ii) point to a memory pool or a number of locations where private data is stored and the location of the private data is not known. In determining how this can be implemented in a C-like programming language, we considered preallocating memory pools for pointers with private locations. Such memory pools would be required for each data type to ensure that we can store and extract private data correctly. This approach, however, has severe disadvantages, which are: 1. Using memory pools unnecessarily increases the program\u2019s memory footprint, where one pool will be needed for each used data type including complex types de\ufb01ned via the struct construct. Furthermore, it is not clear to what size each pool should be set to optimize performance. 2. It would also often incur unnecessarily large computation costs due to the need to touch all locations within a memory pool per single access (or touch several locations when the pool is implemented using more complex ORAM techniques). As will be evident later in the paper, there are large classes of programs, applications, and data structures, where a pointer to private data always corresponds to a single location, which removes the need to use secure multi-party computation techniques for pointer manipulation. Allowing a pointer to store a single known location drastically improves program performance compared to using pointer pools. 3. Memory pools would also not work in the presence of pointer casting. Then if we do not want a pointer to initially point to a pre-allocated memory pool, would the decision to properly declare a pointer as pointing to a single (known) location or a set of locations be left to the programmer? This is going to introduce an additional burden for a programmer who would need to know at a variable declaration time whether the variable of a pointer type will require protecting its value. This happens if the pointer is used inside a conditional statement with 6 \fprivate condition, which then requires protecting the location assigned to the pointer to protect the result of the condition evaluation. To ease programming burden and at the same time avoid consuming unnecessary (memory and computation) resources, our solution is to use the same programming interface for all pointers that are to point to private data. When the pointer is being declared or initialized, it has one known location associated with it (if the pointer is not initialized, that location is set to the default value corresponding to uninitialized pointers). Throughout the computation, the pointer, however, may be pointing to multiple locations, one of which is its true location. This happens when the pointer\u2019s value is modi\ufb01ed inside conditional statements with private conditions as illustrated next. Suppose we declare variables a and b to be private integers followed by the code below: 1. private int *p; 2. p = &a; 3. if (priv-cond) then p = &b; We see that variable p was declared as a pointer to a private integer, but the type of the pointer with respect to whether the location itself is private is implicit. After executing lines 1\u20132, p has a single known location, but after executing line 3, p is associated with a list of two locations (the address of a and the address of b) and the value of the true location is protected. That is, a pointer always starts with a single publicly known location and the location to which it is pointing may become private, but the user does not declare the pointer itself as public or private. In the rest of this work, we use the term \u201cpublic location\u201d in reference to a pointer to private data to mean that the pointer has a single known location (either initialized or uninitialized) and we use the term \u201cprivate location\u201d to mean that the pointer has a list of public locations, but which location is in use remains private. When we consider interaction of public and private values in connection to the use of pointers, a number of questions arise, which we address next. 1. Can a pointer that was declared to point at private data be assigned address of public data? Note that without the use of pointers, the equivalent actions are generally allowed. That is, a variable declared to hold private data can be assigned a known value, which is consequently converted into protected form. The same does not hold for pointers and we disallow assigning locations of public variables to pointers which were declared to point to private data. To see why, suppose that a user program contains the code below where a was declared to be a private integer, while b is a public integer: 1. p = &a; 2. if (priv-cond) then p = &b; 3. *p += 1; After executing lines 1\u20132, p stores two addresses and the true location of where it is pointing out of these two addresses is protected. On line 3, however, the pointer is dereferenced and the result of private condition priv-cond evaluation is revealed by examining the value of b before and after line 3. Thus, to eliminate information leakage, pointers to private data can be assigned only locations that store private values. 2. Can a pointer declared to point to public data be modi\ufb01ed inside conditional statements with private conditions and as a result become pointing to multiple locations? The answer to this question is No. If a pointer to public data is updated in the body of a conditional statement 7 \fwith private condition, it must be treated as a pointer to private data (otherwise, using its dereferenced value reveals unauthorized information). Allowing such uses and performing the conversion implicitly by the compiler will be confusing to the programmer (who no longer can use the pointer to store addresses of public data). For that reason, we disallow updates to pointers to public data within the body of conditional statements with private conditions. 4.2 Pointer Implementation We next proceed with describing how pointers to private data are implemented to realize the ideas outlined above. We note that all program transformations that we describe preserve semantics of the original program and, given that a program can be compiled into the corresponding secure implementation, the transformed program will always produce the same output as the original program. There are some restrictions that user programs must meet in order to be compiled into secure implementations with no information leakage. Such restrictions include the two cases at the interaction of public and private data described above and additional restrictions inherited from PICCO (e.g., the fact that the body of a conditional statement with a private condition cannot have public side e\ufb00ects, a loop termination condition should be public or made public, etc.). This is to ensure that no information leakage in the compiled program can take place, and the programs that do not meet the requirements are aborted at the compilation time. Once these constraints are met, our extension of PICCO will allow any user program to be compiled into its secure counterpart. Pointer representation. As we incorporate support for pointers, we \ufb01rst note that pointers to public data will not need to be modi\ufb01ed and their implementation remains the same as in the C programming language. The most signi\ufb01cant change in implementing pointers to private data comes from the need to maintain multiple locations. For that reason, the data structure that we maintain for pointers to private data consists of (i) an integer \ufb01eld that stores the number \u03b1 (\u22651) of locations associated with the pointer; (ii) a list of \u03b1 addresses where the data is stored; and (iii) a list of \u03b1 private (i.e., secret-shared) tags, one of which is set to 1 (true data location) and all others are set to 0. For the important special case of \u03b1 = 1, the pointer has known (public) location and the tags are not used. We formalize the above pointer representation using the following invariant, which is maintained throughout various pointer operations: among all locations stored with a pointer to a private object, there is exactly one true location of the object and the tag corresponding to that location is set to 1, while the tags corresponding to all other locations are set to 0. This invariant is true of all well-formed programs and may be violated only in the case of dangling pointers as detailed later. Because we would like to employ a uniform data structure for pointers to private data of any data type such as integer, \ufb02oating point values, etc. and even pointers to a pointer, the data structure we maintain needs to include two additional \ufb01elds: (iv) an integer \ufb02ag that determines the type of data associated with the pointer (i.e., integer = 1, \ufb02oat = 2, struct = 3, etc.) and (v) an integer \ufb01eld that indicates the indirection level of the pointer. For instance, if a pointer refers to a private value of a non-pointer type, its indirection level is set to 1; and if it refers to a pointer whose indirection level is k (for k \u22651), its level will be set to k + 1. A pointer to a struct also has indirection level 1 regardless of the types of the struct\u2019s \ufb01elds (which can be pointers themselves). Pointer updates. Initially, at the pointer declaration time, the number of locations \u03b1 associated with the pointer is set to 1 and the address is set to to a special constant used for uninitialized pointers. Then every time the pointer is modi\ufb01ed (including simultaneously with pointer declaration), its data structure is updated. When the pointer is assigned a new location using a public constant, a variable\u2019s address, or a memory allocation mechanism (e.g., as in p = 0, p = &a, or p 8 \f= malloc(size)), \u03b1 in the pointer\u2019s data structure is set to 1 and the associated address is stored in the pointer\u2019s address list. When a pointer is updated using another pointer (as in p = p1), the latter\u2019s data structure is copied and stored with the former. Such simple manipulations are used only when the assignment does not take place inside the body of a conditional statement with a private condition. Pointer assignments inside conditional statements with a private condition present the most interesting case when the list of pointer locations gets modi\ufb01ed. Updating values modi\ufb01ed in the body of a conditional statement with a private condition already requires special handling in PICCO, and all we need is to support a speci\ufb01c procedure when a variable of pointer type is being modi\ufb01ed. We need to distinguish between if-then and if-then-else statements, which we consequently discuss. Consider the following code with an if-then statement: 1. p = p1; 2. if (priv-cond) then p = p2; where p, p1, and p2 are pointers (to private data) of the same type. This is the most general case, where on line 2 both p and p2 can have any number of locations associated with each of them (recall that all other assignment types use a single location). When this code is written for ordinary (private) variables of the same type a, a1, and a2, a generic way to implement this update in PICCO and similar compilers is to \ufb01rst set [a] = [a1] and then compute [a] = [c] \u00b7 [a2] + (1 \u2212[c]) \u00b7 [a] = [c] \u00b7 ([a2] \u2212[a]) + [a], where c is a bit equal to the result of evaluating priv-cond. We use notation [x] to indicate that the value of x is protected via secret sharing and computation takes place on its shares. In the case of pointers, such a simple update does not work because this procedure would turn addresses into secret shared values preventing the pointer from being dereferenced (without touching all possible memory locations). Thus, after executing the assignment p = p1, we combine the (public) locations of p and p2 and set the tags in p based on the current tags of p and p2 and the result c of evaluating priv-cond. Let pointer p after executing the \ufb01rst assignment contain \u03b11 locations stored as L1 = {\u21131, . . ., \u2113\u03b11} with corresponding tags T1 = {[t1], . . ., [t\u03b11]} (i.e., this information was copied from p1). Similarly, let pointer p2 store \u03b12, L2 = {\u2113\u2032 1, . . ., \u2113\u2032 \u03b12}, and T2 = {[t\u2032 1], . . ., [t\u2032 \u03b12]}. Note that the ordering of addresses in each L is arbitrary, but the tag ti in T must correspond to the address \u2113i at the same position i in L. Then as a result of the conditional assignment, we compute p\u2019s new content as given in Algorithm 1. In the algorithm, L3 is composed of all locations appearing in L1 or L2 (repeated locations are stored only once). We use notation L.\ufb01nd to retrieve the position of the element of L provided as the argument or special symbol \u22a5is the element is not found. The tags in the output T3 are set based on three di\ufb00erent cases: (i) a location in L3 is found in both L1 and L2; (ii) it is found in L1, but not in L2; and (iii) it is found in L2, but not L1. Because only tags in T1 and T2 and c are private, only lines 7, 9, and 11 correspond to private computation. If the conditional statement is of the form if-then-else, but p is not updated in the body of the else clause, then the computation in Algorithm 1 is applied unchanged. If the pointer is instead updated only in the body of the else clause, then the computation is performed similarly, but Algorithm 1 is called with the value of 1 \u2212c instead of c. Lastly, if the pointer is updated in both clauses of the if-then-else statement, the pointer content prior to that statement needs to be disregarded. The pointer values used in the two assignments are then merged as in Algorithm 1 using the result c of private condition evaluation. To better illustrate this, consider the following code segment: 9 \fAlgorithm 1 \u27e8\u03b13, L3, T3\u27e9\u2190CondAssign(\u27e8\u03b11, L1, T1\u27e9, \u27e8\u03b12, L2, T2\u27e9, [c]) 1: L3 = L1 \u222aL2; 2: \u03b13 = |L3|; 3: for every \u2113\u2032\u2032 i \u2208L3 do 4: pos1 = L1.\ufb01nd(\u2113\u2032\u2032 i ); 5: pos2 = L2.\ufb01nd(\u2113\u2032\u2032 i ); 6: if (pos1 \u0338=\u22a5and pos2 \u0338=\u22a5) then 7: [t\u2032\u2032 i ] = [c] \u00b7 [t\u2032 pos2] + (1 \u2212[c]) \u00b7 [tpos1]; 8: else if (pos2 =\u22a5) then 9: [t\u2032\u2032 i ] = (1 \u2212[c]) \u00b7 [tpos1]; 10: else 11: [t\u2032\u2032 i ] = [c] \u00b7 [t\u2032 pos2]; 12: end if 13: end for 14: set T3 = {[t\u2032\u2032 1], [t\u2032\u2032 2], . . ., [t\u2032\u2032 \u03b13]}; 15: return \u27e8\u03b13, L3, T3\u27e9; 1. p = p1; 2. if (priv-cond) then p = p2; 3. else p = p3; After we assign p1 to p on the \ufb01rst line, p\u2019s content is be overwritten with the content of either p2 or p3 depending on the result c of evaluating priv-cond. We can see that before entering the if-clause, the current content of p (i.e., that copied from p1) can be safely disregarded without a\ufb00ecting its correctness. In other words, to update p inside the conditional statement, we call CondAssign(\u27e8\u03b12, L2, T2\u27e9, \u27e8\u03b13, L3, T3\u27e9, c) in Algorithm 1, where \u27e8\u03b12, L2, T2\u27e9and \u27e8\u03b13, L3, T3\u27e9are contents of pointers p2 and p3, respectively. These constructions compose in presence of nested conditional statements with private conditions. For instance, after executing the code: 1. if (priv-cond1) then p = p1; 2. else 3. p = p2; 4. if (priv-cond2) then p = p3; 5. else p = p4; p will contain the combined content of pointers p1, p3, and p4. That is, Algorithm 1 is \ufb01rst called with the content of pointers p3 and p4 and the result c2 of evaluating priv-cond2, after which Algorithm 1 is called on the result of its previous execution, the content of p1, and the result c1 of evaluating priv-cond1. As evident from the description above, all modi\ufb01cations to variables of all types (including pointers as well as data) inside conditional statements with private conditions require special handling inside the compiler. For each such conditional statement, PICCO examines the list of variables modi\ufb01ed inside the body of the statement and updates them di\ufb00erently from when the modi\ufb01cation is not surrounded by a private condition. Thus, in the case of pointers we specify how pointers need to be updated inside such statements using Algorithm 1 and compiler will process all variables inside the body of conditional statements with private conditions. Note that each pointer starts with a single location (i.e., \u03b1 is set to 1) at the time of its declaration, and the list and the number of locations \u03b1 are updated during pointer assignments 10 \fas described above. This information is maintained only during program execution and thus the locations that a pointer might store or their number is not necessarily known at compile time. Pointer dereferencing. When pointer p with a private location is being dereferenced, its dereferenced value is privately computed from \u03b1, L = {\u21131, . . ., \u2113\u03b1}, and T = {[t1], . . ., [t\u03b1]} stored at p. Let [ai] denote the value stored at location \u2113i \u2208L. Then we compute the dereferenced value as [v] = P\u03b1 i=1[ai] \u00b7 [ti]. Note that with linear secret sharing this operation can be implemented as an inner product that costs only a single interactive operation resulting in a profound impact on performance. When the dereferenced value is being updated, all locations in L need to be touched, but the content of only one of them is being changed. If we, as before, use [ai] to denote the value stored at \u2113i \u2208L and let [anew] denote the value with which the dereferenced value is being updated, then we update the content of each location \u2113i as [ai] = [ti] \u00b7 [anew] + (1 \u2212[ti]) \u00b7 [ai]. That is, the true location (ti = 1) will be set to anew, while all others (ti = 0) will be kept unchanged. In the current form, the above procedures are applicable only to pointers with the indirection level equal to 1. That is, if pointer p is associated with a list of private locations of pointers, the above computation will result in producing secret shared locations and the information looses its semantic meaning. Thus, for pointers with indirection level > 1 di\ufb00erent computation is used. That is, now each \u2113i \u2208L stores an address of a pointer pi and let each pi be associated with \u03b1i, Li = {\u2113(i) 1 , . . ., \u2113(i) \u03b1i }, and Ti = {[t(i) 1 ], . . ., [t(i) \u03b1i ]}. To retrieve the dereferenced value of p, we \ufb01rst compute [ti] \u00b7 [t(i) j ] for 1 \u2264i \u2264\u03b1 and 1 \u2264j \u2264\u03b1i and merge all lists Li for 1 \u2264i \u2264\u03b1. The resulting list is thus set to L\u2032 = L1 \u222aL2 \u222a\u00b7 \u00b7 \u00b7 \u222aL\u03b1 and let \u03b1\u2032 = |L\u2032|. For any location in L\u2032, we compute its corresponding tag as the sum of all [ti] \u00b7 [t(i) j ] values matching that location in the individual lists Li. (We can simply use the sum because only one tag can be set to 1.) The result is \u03b1\u2032, L\u2032 and the corresponding tags T \u2032. To illustrate this on an example, let \u03b1 = 3, T = ([0], [0], [1]), and L store the addresses of pointers p1, p2, p3 with \u03b11 = 1, L1 = (123), T1 = (1); \u03b12 = 2, L2 = (189, 245), T2 = ([0], [1]); and \u03b13 = 3, L3 = (123, 176, 207), T3 = ([0], [1], [0]). The result of this operation is a pointer with L\u2032 = L1 \u222aL2 \u222aL3 = (123, 176, 189, 207, 245), \u03b1\u2032 = |L\u2032| = 5, and T \u2032 = ([0], [1], [0], [0], [0]). To update the dereferenced value of p through an assignment as in *p = p\u2019, each pointer pi stored at address \u2113i \u2208L needs to be updated with p\u2019\u2019s information. In particular, for each pi each tag [t(i) j ] (for location \u2113(i) j ) is updated to (1 \u2212[ti]) \u00b7 [t(i) j ]. We also compute tag [ti] \u00b7 [t\u2032 j] for each location \u2113\u2032 j in p\u2019\u2019s list of locations. We then merge the location list of each pi with that of p\u2019 to form pi\u2019s new list. For any new location inserted into Li, its tag is set to the computed [ti] \u00b7 [t\u2032 j] for the appropriate choice of j, and any location that appears on both pi and p\u2019 lists, the value [ti] \u00b7 [t\u2032 j] is added to pi\u2019s updated tag for that location. In other words, if ti is true, we take p\u2019\u2019s value and otherwise keep pi\u2019s value. If pointer p with a private location is being dereferenced m > 1 times, the above dereference algorithms are naturally applied multiple times with the \ufb01rst m\u22121 instances being the version that produces a pointer and the last instance producing either a pointer or a private value depending on p\u2019s indirection level. p can then be treated as the root of a tree with its child nodes being locations of pointers stored in its list and the leaves of the tree eventually pointing to private data (of a non-pointer type). To perform an m-level dereferencing operation, we traverse the top m + 1 levels of the tree and consolidate the values stored at those levels (and update the values at the (m+1)st level if the dereferenced value is to be updated). Secrecy of pointers to private data. As previously discussed, the value of a pointer to private data is treated as public when it stores a single location (\u03b1 = 1), and it is private otherwise (\u03b1 > 1). 11 \fMore generally, if pointers to private data are used in predicates or similar expressions, the result of a predicate evaluation is public if its outcome can be determined using only public data. For example, the outcome of an expression that compares two pointers to private data for equality is public if (i) both pointers store a single location in their lists L1 and L2 or (ii) at least one of the pointers stores multiple locations, but L1 \u2229L2 = \u2205. In other circumstances, the outcome depends on private tags and is treated as private. Note that when the result of a predicate evaluation on pointers is private, it can be naturally computed by privately determining the true location of each pointer and applying the predicate to them. This computation, however, can be optimized for certain types of predicates to result in faster performance. For example, in the case of pointer equality, the general solution is to compute [\u21131] = P \u2113(1) i \u2208L1 \u2113(1) i [t(1) i ] and [\u21132] = P \u2113(2) i \u2208L2 \u2113(2) i [t(2) i ] and then compare [\u21131] and [\u21132] for equality, while an optimized implementation computes P \u2113(1) i \u2208L1\u2229L2[t(1) i ][t(2) j ], where \u2113(1) i = \u2113(2) j for each \u2113(1) i \u2208L1 \u2229L2. The latter can be implemented as the inner product that costs a single interactive operation and is signi\ufb01cantly lower than the cost of comparing two private integers. Because it is not always possible to determine at compile time whether a predicate evaluated on one or more pointers (to private data) will have a public or private status (which, for example, may depend on program\u2019s public input), some checking will need to be deferred to run time. In particular, if pointers to private data are used in a predicate to form a conditional statement and the result of its evaluation is private, the usual constraints for the body of conditional statements with private conditions apply. We address this by evaluating the body of such conditional statements for public side e\ufb00ects at compile time (as in the original PICCO design). If the body contains public side e\ufb00ects, the transformed program will include checks for the status of the conditional statement at run time. If the result of evaluating the conditional statement is determined to be private at run time (and public side e\ufb00ects are present in the body of the statement), execution will be aborted with an error due to a possible information leak. Note that the fact whether the execution is aborted or not never leaks private information, as this decision is solely based on public data. That is, an abort takes place when (i) the result of evaluating a predicate on two or more pointers is treated as private and (ii) the body of the conditional statement that uses the predicate contains public side e\ufb00ects. Whether the result of evaluating the predicate is public or not is public knowledge because it is determined by the public locations stored in the pointers and the predicate itself. Similarly, whether the body of the conditional statement has public side e\ufb00ects or not is based on the public code that forms the body of the conditional statement. Optimizations. Because the computation involved in computing a pointer\u2019s dereferenced value is interactive (and thus is relatively expensive) when the pointer stores multiple locations, we considered caching and reusing its result. That is, when a pointer\u2019s dereferenced value is computed, we can store it in the pointer structure and reuse the value on consecutive dereference operations when there are no changes to the values stored at all pointer\u2019s locations L between such operations. Once any value stored at one of the pointer\u2019s locations is modi\ufb01ed, the cached dereferenced value needs to be marked as out of date. A similar caching technique can also be applied to the computation on private tags that takes place during pointer update and dereference operations. Note that private tags can be viewed as aggregation of conjunctions of 1-bit private variables that denote the evaluation results of private conditions (or their negations) in user programs. Because, once computed, the variables will have \ufb01xed values, the conjunction results of those variables can be cached in our framework in a lookup table and allow for their retrieval when the same conjunctions need to be repeatedly computed. This optimization will result in considerable savings when multiple private 12 \fpointers are updated or dereferenced within the body of a conditional statement with a private condition. The savings due to either of the above caching optimizations are, however, applicationdependent and require additional program analysis at program parse and translation time. Thus, these optimizations are not presently a part of our implementation. Another possible optimization can lower a program\u2019s memory footprint by reusing data structures created for pointers to private data. In particular, when a pointer is assigned another pointer\u2019s value (as in p = p1), we could have both pointers pointing to the same data structure instead of creating its copy. When, however, one of the pointers with the shared data structure is being modi\ufb01ed, it should be unlinked from the shared data structure and its data structure modi\ufb01ed accordingly. Implementing this would require that each pointer data structure is associated with a list of pointer variables which are using it. Furthermore, a data structure can be reused only when all information stored in it is identical for multiple pointers (i.e., not only the locations in L, but also the private tags in T). Because di\ufb00erent pointers often have distinct roles in user programs, the expected savings are not very large. This optimization is presently not a part of our implementation. 4.3 Pointers to Struct We next discuss design and implementation of pointers to structs, including their representation and the associated algorithms. Pointers to complex data types declared using struct constructs are common for building data structures such as linked lists, stacks, and trees, and thus pointers to structs deserve special attention. As before, if a complex data type contains no private \ufb01elds, no transformations are needed. However, when dealing with pointers to struct with private \ufb01elds, we need to address the following questions: 1. A struct groups together a number of di\ufb00erent variables that can be either private or public, but the complex data type itself declared using struct is not associated with any particular type of secrecy. When declaring a pointer to a complex data type, we thus need to determine if a pointer to it can be treated as a pointer to private data or if it has to be treated as a conventional pointer to a public variable. 2. When designing representation of a private pointer that points to struct, we need to take into account the fact that \ufb01elds of a complex data type can be accessed and modi\ufb01ed independently of each other or the struct itself. Thus, it remains as a question whether we should maintain a separate list of addresses for each struct \ufb01eld or maintain only a single list of addresses for all possible struct variables associated with the pointer. 3. The last question is whether we can reuse the previously described algorithms for working with private pointers for updating or dereferencing pointers to structs on the individual \ufb01elds of a struct or if modi\ufb01cations are needed. In what follows, we thus focus on answering these questions. Secrecy of pointers to struct. Secrecy of a pointer to struct is implicitly determined by the protection modes of the struct\u2019s \ufb01elds. We determined that a pointer to a complex data type can be treated as a pointer to private data only if all \ufb01elds in its declaration are private. It means that if at least a single \ufb01eld of a struct is public, pointers to this data type can be of public type only. This treatment is necessary to eliminate information leakage when pointers to structs are modi\ufb01ed inside conditional statements with private conditions. Consider, for example, a data type containing 13 \fone private and one public \ufb01eld. If we treat a pointer to this data type as a pointer to private data, it can be modi\ufb01ed inside an if-statement with a private condition and have multiple locations associated with the pointer. However, by dereferencing and observing the value of the public \ufb01eld, one can determine the true location of the pointer and thus learn unauthorized information about the result of the private condition. Because a complex data type may contain other struct variables as its \ufb01elds, the variables in the data type will need to be checked recursively to determine whether at least one public \ufb01eld is present (with provisions to skip cycles in the declarations). If none are found, pointers to this data type are treated as pointers to private data. Pointer to struct representation. To implement private pointers to structs, we needed to determine whether a single list of locations is su\ufb03cient for all \ufb01elds of the complex data type (recall that all \ufb01elds are private) or separate lists must be maintained. In working to answer this question, we determine that there is no need to maintain multiple lists of locations, because the list of locations associated with each \ufb01eld in the struct must be the same (adjusted for the o\ufb00set of the \ufb01eld within the struct). That is, values of a struct\u2019s \ufb01elds can be modi\ufb01ed individually (e.g., as in p->x = y), but the only way to access or modify the location of a \ufb01eld is through the location of the entire struct. In other words, the list of addresses associated with a pointer to struct p (and thus the addresses corresponding to all of its \ufb01elds) can be modi\ufb01ed only by directly updating p, as operations of the type p->x = y do not a\ufb00ect the list of addresses associated with the \ufb01eld x. Storing a single list has the added bene\ufb01t that we can employ the same representation of pointers to private data as for simple data types. This treatment also implies that a pointer to a struct object will have indirection level 1 even if all \ufb01elds of the struct are pointers themselves. Operations on private pointers to struct. We represent pointers to a struct record in the same way as other pointers. This means that operations for using pointers and updating their values remain unchanged. To dereference a speci\ufb01c \ufb01eld of a pointer as in p->x and retrieve the value of the variable x, also only minor changes to the previously described algorithms are needed. In particular, all we need is to determine the o\ufb00set f of the variable\u2019s address within the record and perform the dereferencing procedure in the same way as for pointer p itself, but instead of using locations \u2113i from L, we use locations \u2113i + f. The same modi\ufb01cation applies to the case when the dereferenced value is modi\ufb01ed through assignment. If we would like to dereference p and retrieve the entire record as in rec = *p, we need to iterate through each \ufb01eld of the struct and retrieve the dereference value of each \ufb01eld as described above for p->x. Similarly, to update a dereferenced pointer p as in *p = rec, we need to perform the equivalent of p->x = rec.x for each \ufb01eld x of the struct. 5 Pointer Uses in Programming In this section we discuss many common uses of pointers in programming and how they are translated to our environment of computing with private data. The topics we cover are passing arguments by reference, dynamic allocation of memory, array manipulation, and pointer casting. Data structures also constitute a common use of pointers, but we discuss them separately in Section 7. 5.1 Passing Arguments by Reference Function calls contribute to the basic software engineering principles of modular program design, but could be expensive in terms of stack memory usage for the passed arguments. This has led to di\ufb00erentiating between function calls where the arguments are passed by value and by reference. 14 \fIn the latter case, the function typically takes a pointer to the argument and all updates to the dereferenced pointer will be visible after completing the function call (thus, arguments passed by reference can be used for either input or output). Passing private variables to functions by reference inherits the same bene\ufb01ts as for conventional (public) variables in the programming language. The good news is that no special provisions are needed for passing private variables by reference, resulting in e\ufb03cient implementations. Furthermore, because often to pass an argument by reference, its address is supplied to a function call (as opposed to supplying an existing pointer), the resulting pointer will have a single known location. This allows us to enjoy the bene\ufb01ts of avoiding using extra resources without the slowdown of working with pointers with private locations. 5.2 Dynamic Memory Allocation Pointers are often used in programming to dynamically allocate memory on the free store and deallocate it when it is no longer in use. Here we focus on C-style malloc() and free() used with pointers to public variables and show what modi\ufb01cations are needed to support dynamic memory allocation with pointers to private variables. malloc() in C allocates the requested number of bytes on the heap which are passed as an argument to the function malloc(). The result of this function is the address of the allocated variable or the \ufb01rst array element in case of dynamic array allocation, which is stored in a pointer. To support dynamic memory allocation for private variables, we start with the following code in C: 1. int* p = (int*) malloc(sizeof(int)); 2. int* p1 = (int*) malloc(10*sizeof(int)); Here p points to single variable, while p1 points to a dynamic array of size 10. The assignment operator directly saves the malloc result into the pointer because they are of compatible types. However, this is not the case for pointers to private variables because a private pointer is represented using multiple \ufb01elds. Consequently, we cannot assign the malloc result directly to a private pointer and use a modi\ufb01ed interface for pointers to private variables. In particular, we use a function pmalloc2 to implement private malloc, which is invoked as: 1. private int* p = pmalloc(10, private int); As shown, pmalloc takes two arguments, which are the requested number of dynamic variables and the data type. The function returns the data structure used for private pointers in our implementation with \u03b1 = 1 and the only location in L set to the address of the \ufb01rst variable in the allocated array (when the \ufb01rst argument to the function is > 1). Specifying the private data type is necessary to properly allocate and initialize the memory. For example, in PICCO a private integer is represented using one variable of type mpz t from the GMP library [1] and a private \ufb02oat is represented using four mpz t variables. Once memory for the necessary number of variables is allocated, each of them also needs to be initialized before it can be used in computation. Calling free with a pointer in C allows to deallocate the memory (for either a variable or dynamic array) to which the pointer is pointing. To support similar functionality for private variables, we implement a function pfree that similarly takes a pointer (to a private variable or dynamic array) as its only argument. With pfree we distinguish between two di\ufb00erent cases: the 2Note that the choice of the function is not crucial and it can be called malloc instead to simplify programmers\u2019 e\ufb00ort for transforming an existing program to an equivalent program that computes with private data. We, however, prefer to use pmalloc to make it explicit that the computation refers to private data. 15 \fpointer provided as an argument to the function has a single known location (i.e., \u03b1 = 1) or it has a private location out of a public list (\u03b1 > 1). Handling the \ufb01rst case is simple and e\ufb03cient: we can simply call the free command to deallocate memory associated with the address stored in the pointer. Pointers to private data with public locations are very common in programs that use pointers to private data or build data structures from private data (e.g., linked lists, stacks). Freeing memory used by pointers to private data in such cases is thus going to be extremely e\ufb03cient and does not introduce additional overhead. Handling the second case well, however, is very challenging. This is because deallocating physical memory results in publicly observable outcomes, and we must be extremely careful not to reveal the true location stored in a pointer with a private location while at the same time reducing the program\u2019s memory usage. For example, a simple strategy of deallocating memory associated with all locations on a pointer\u2019s list of addresses will not be acceptable for some programs. To illustrate this, consider a dummy example with two pointers p1 and p2, for each of which we allocate memory using pmalloc. Then the locations to which the pointers are pointing are swapped based on the result of a private condition evaluation. We obtain that both p1 and p2 now contain two identical locations in their lists of addresses, but their true addresses are distinct. Suppose we process the data to which p1 points and want to deallocate the corresponding memory. If we deallocate both addresses on p1\u2019s list, p2 becomes a dangling pointer and the data to which it was pointing is no longer accessible. Thus, such an implementation of pfree would be too restrictive to permit its general use. Thus, calling pfree(p) should result in deallocating memory associated with only one address on p\u2019s list of addresses. Furthermore, the address being deallocated cannot depend on any private data (but can be any function of public data). This means that we are not necessarily deallocating memory associated with the true location of the pointer and other pointers that store the same location on their lists must be adjusted to preserve correctness of the computation (which involves additional resources). For example, we can choose to deallocate the \ufb01st location \u21131 on a pointer\u2019s list, but if this was not the pointer\u2019s true location (which we can privately check), the data stored at \u21131 needs to be relocated and other pointers storing \u21131 on their lists need to be updated accordingly. We next describe in more detail how we can realize this idea. First, if the pointer p on which pfree was called contains the default location corresponding to uninitialized pointers on its list of addresses L (which is public knowledge), we choose not to perform memory deallocation. This is to ensure that no memory is being deactivated (which may be in use by other pointers) if p happens to be uninitialized. Otherwise, we free the \ufb01rst location \u21131 on p\u2019s list.3 (Alternatively, the location used by the smallest number of pointers can be freed.) Before we can actually free the memory, we need to privately update the values stored at the remaining locations in L using the value stored at \u21131 to maintain correctness. We will need to ensure that (i) if \u21131 happens to be the true location, the values stored in the remaining locations will remain unchanged and (ii) if \u21131 is not the true location, the value stored at \u21131 can be found at p\u2019s true location, while the values stored at all other locations remain unchanged. The rationale for doing this as follows: if \u21131 is indeed p\u2019s true location, no additional work would be required if this fact was public (i.e., it is the programmer\u2019s job to ensure that freeing p does not a\ufb00ect other variables still in use). If \u21131, however, was not p\u2019s true location, it may be in use by other pointers and the value stored at \u21131 needs to be relocated to p\u2019s true location prior to memory deallocation (and the pointers that contain \u21131 in their lists need to be updated accordingly). Let p at the time of calling pfree store \u03b1, L = {\u21131, . . ., \u2113\u03b1}, T = {[t1], . . ., [t\u03b1]} and A = 3In the event that data stored at the locations contained in the pointer have di\ufb00erent sizes, the location with the data of the smallest size should be chosen instead. 16 \f{[a1], . . ., [a\u03b1]} denote values stored at locations in L.4 To obliviously update [ai]\u2019s for 2 \u2264i \u2264\u03b1, we compute [ai] = [a1] \u00b7 [ti] + [ai] \u00b7 (1 \u2212[ti]). This satis\ufb01es the above two requirements as follows: if t1 is true (\u21131 is the true location) and thus ti is false, the result will be ai for any i; if ti is true and thus t1 is false, the result will be a1; if both t1 and ti are false, the result will be ai. Surprisingly the formula does not depend on t1. Second, we need to update private pointers that store the freed location \u21131 in their lists (and are still in use), but no computation needs to be performed for pointers that store any of \u21132, . . ., \u2113\u03b1 from L, but not \u21131 itself. The rationale for doing this as follows: if \u21131 is indeed p\u2019s true location, no additional work would be required if this fact was public (i.e., it is programmer\u2019s job to ensure that freeing p does not a\ufb00ect other variables still in use). If \u21131, however, was not p\u2019s true location, it may be in use by other pointers and the value stored at \u21131 is moved to p\u2019s true location prior to memory deallocation. We thus need to replace \u21131 in other pointers\u2019 lists with locations that are guaranteed to include the value originally stored at \u21131 and update the locations\u2019 tags accordingly. Thus, for each pointer p\u2019 that stores \u21131 in its list L\u2032, we retrieve \u21131\u2019s position pos in L\u2032 and its corresponding tag t\u2032 pos. We then replace \u21131 in L\u2032 with {\u21132, . . ., \u2113\u03b1} and t\u2032 pos in T \u2032 with {[t\u2032 pos] \u00b7 [t2], . . ., [t\u2032 pos] \u00b7 [t\u03b1]}. If any of \u2113i for i = 2, . . ., \u03b1 already appears in L\u2032, that location is not included the second time and its tag is set to the sum of the tag already present in T \u2032 for location \u2113i and [t\u2032 pos] \u00b7 [ti]. Returning to our example with p1 and p2, we have that prior to calling pfree(p1), p1 stores \u03b11 = 2, L1 = (\u21131, \u21132), T1 = (t1, t2), and p2 stores \u03b12 = 2, L2 = (\u21132, \u21131), T2 = (t\u2032 1, t\u2032 2). Then either t1 = t\u2032 1 = 1 and t2 = t\u2032 2 = 0 or t1 = t\u2032 1 = 0 and t2 = t\u2032 2 = 1. Once pfree(p1) is called, \u21131 is scheduled for deallocation. If t1 = 1, no changes take place; otherwise (t2 = 1), the data from location \u21131 is copied into location \u21132. We obtain that location \u21131 is being removed from L\u2032 (and the corresponding tag t\u2032 2 from T \u2032) and location \u21132 is being added to L\u2032 with the corresponding tag t\u2032 2 \u00b7 t2. Because \u21132 is already present in L\u2032, it is stored once and the tag becomes t\u2032 1 + t\u2032 2 \u00b7 t2. Thus, we have that L\u2032 now stores a single location and the tag is 1 for any possible set of original tags. Note that the second step of updating pointers that store location \u21131 in their lists is more complex when the pointer p being freed points to a struct or an array. In those cases, multiple addresses are processed in this step (\u21131 and other valid addresses that store data at \ufb01xed o\ufb00sets from \u21131) depending on the type of data to which p points. In our implementation, we gather all interactive operations associated with the execution of a call to pfree and perform them simultaneously in a single round. If the user program is written correctly (i.e., does not leave dangling pointers after a call to free), our implementation of pfree will maintain that for each pointer exactly one location\u2019s tag is set to 1 and all other locations\u2019 tags are set to 0. When, however, a call to deallocate memory corresponding to a pointer results in dangling pointers, all tags in such pointers can be 0. For that reason, if a call to pfree causes the number of addresses for some pointer to reduce to 1, we do not treat the corresponding tag as public. That is, when a program is not correctly written, opening the value of the tag may reveal private information, while assuming that the tag is 1 may modify the program\u2019s behavior. Thus, our implementation maintains privacy even in the presence of programming errors that result in dangling pointers. We also note that the use of pmalloc or pfree will not be allowed inside conditional statements with private conditions because these functions have public side e\ufb00ects. 4Although in the current discussion we assume p is a private pointer that points to a non-pointer data type, the same idea will apply when p points to a pointer. In particular, if p points to a pointer, the procedure will include merging the lists of pointers stored at locations \u21131 and \u2113i and updating the tags similar to the formula for simple data types. Furthermore, when p is a pointer to a struct, each \ufb01eld is updated separately according to its type. 17 \f5.3 Accessing Array Elements The next common use of pointers in programming is manipulating arrays using pointers. Even for statically allocated arrays, the array name is treated as a constant pointer that points to the \ufb01rst element of the array. Hence, arrays and pointers are tightly coupled and pointers are used extensively to work with arrays. Array indexing. Because arrays are based on pointers, array indexing also applies to pointers. Thus, we can see constructions such as p = a and p[i], where p is a pointer and a is an array, and need to support them for pointers to private data. Pointer indexing p[i] with a pointer p to private data and a public index i is implemented naturally, where we iterate through all locations in the address list L of p, advance each of them by i multiplied by the size of the data type, retrieve the data at the determined positions, and combine all of them using private tags for each location to obtain the result. In other words, the computation is very similar to that of pointer dereferencing, where instead of retrieving data at the positions speci\ufb01ed in L, we advance each position by i data items. (As C permits the use of negative indices, when i in p[i] is negative each location in L is decremented by the necessary amount during this operation.) Pointers as arrays with known bounds. In PICCO, statically allocated arrays of private variables have the array size stored with them (which is known at the array creation time). Knowing the size of the arrays allows the compiler to support of a number of important operations on arrays. Most signi\ufb01cantly, this permits the use of private indexing with arrays, when an element at a private position i is retrieved from an array a using syntax a[i]. (The size of the array must be known to support private indexing, regardless of what technique is used to implement it.) This also permits the use of other operations such as inner (or dot) products on two arrays, which were introduced to optimize runtime of compiled programs. We treat private indexing as an essential part of secure computation with private data and would like to see it supported for arrays dynamically allocated on the heap. This means that we would like to o\ufb00er pointer indexing p[i] with private i and private pointer p. The main challenge that we need to overcome is the fact that the size of the memory pointed by p is not available in C. Furthermore, a location stored in p may be arbitrary and do not correspond to a valid memory address (i.e., be unaccessible by the program, correspond to memory marked as not being in use or any location from the program\u2019s stack, etc.). This means that a pointer can take on many addresses which were not allocated for variable use and for which the corresponding size cannot be meaningfully determined (i.e., accessing such addresses would trigger invalid memory access exceptions in safe programming languages). The size of properly allocated memory, however, can be determined and utilized to implement private indexing (and other operations that require array size) with pointers to private data. In particular, all memory that malloc allocates on the heap is marked with the size of each allocated block. Thus, we can use the information that malloc/free maintain to determine whether a pointer content falls within a properly allocated memory block, and if it is the case, access the block\u2019s size and use it to implement private indexing. In more detail, in addition to using private indexing with statically allocated arrays (as already implemented in PICCO), we permit private indexing to be used with pointers to private data. The latter is only successful if the location stored in the pointer5 was allocated via a prior call to pmalloc (and it was not deallocated during a call to pfree). Because the secure implementation that PICCO produces makes more calls to malloc than once per call to pmalloc, the program 5The current discussion refers to a single location stored in a pointer, which we view as the most common use of private indexing. When the pointer contains multiple locations, the operation is performed on each location separately and the results are combined in the same way as during pointer dereferencing. 18 \finternally maintains a list of addresses returned by malloc that correspond to memory requested by the user program (and an address is taken o\ufb00the list if it is being freed). Then when private indexing p[i] is called in the user program and the pointer stores address \u2113, we iterate through the list of maintained addresses. For each such address l, we retrieve the corresponding block size s from the information stored by malloc and check whether l \u2264\u2113< l + s and the o\ufb00set of \u2113from l is a multiple of the data type size. If these checks succeed for at least one location on the allocated address list, s is adjusted for the data type size and is used as the size of the array to which p points. Note that with this implementation \u2113does not have to correspond to the beginning of the memory block. Then when \u2113is not the address of the beginning of the array, i can legitimately take negative values. Under the circumstances when the address \u2113does not fall within any memory block dynamically allocated by the user, private indexing operation is not performed and the returned result is set to be secret-shares of 0 (note that, regardless to what value the result is set, it is not guaranteed to be interpreted as an error). We thus proceed with the computation despite the error, but send signal SIGBUS6 and store an error message in a \ufb01xed location, so that the program can catch the signal and act on it. We note that the address that each call to pmalloc returns is always public information and the programmer can avoid using invalid addresses. Ideally, the fact that the private indexing operation cannot be carried out on the given address is determined before the program is run, at compile time. Unfortunately, this will not always be possible and for some incorrectly written user programs the error will not be triggered until the program is executed (i.e., even programming languages that perform static analysis of user programs do array-bounds checking dynamically). The best we can do is to perform static program analysis at compile time and warn the user about places where such an error might be possible. Pointer arithmetic. Pointers can be modi\ufb01ed by setting the address to which they point to the result of an arithmetic expression evaluation. While in C pointers can be used in arbitrary expressions similar to the way integer variables are used, only a limited set of operations on pointer variables is meaningful when they are used to store and manipulate addresses within the program. For example, pointer arithmetic can be relied upon to increment or decrement a pointer value by an integer amount to move to a di\ufb00erent position within an array or between struct \ufb01elds. Many other arithmetic operations on pointer variables are not meaningful, and moving between di\ufb00erent variables using pointer arithmetic is unreliable and error-prone. Thus, in PICCO\u2019s default con\ufb01guration we chose to disable pointer arithmetic involving pointers to private data in user programs that the compiler processes. We introduce this as a mechanism for eliminating a large class of programming errors without constraining expressiveness of user programs. That is, if we want to change the pointer\u2019s position within an array, instead of using p = p-i or p = p1+4*k+1, the program will be written as p = &p[-i] and p = &p1[4*k+1], respectively. We note that disabling pointer arithmetic for pointers to private data in the default con\ufb01guration should not be treated as a limitation of the compiler or our approach, but was a deliberate choice to reduce programming errors without constraining expressiveness of user programs. Nevertheless, turning o\ufb00pointer arithmetic in PICCO makes it deviate from standard C. Furthermore, there is a small class of functionalities that become disabled without pointer arithmetic. One example of such functionalities is the use of embedded linked lists as, for example, implemented in the Linux kernel. Embedded linked lists might rely on pointer arithmetic to move between di\ufb00erent \ufb01elds of a struct. Thus, if the need for this or similar functionality when working with private data arises in applications that use PICCO, the compiler can be con\ufb01gured with a command-line \ufb02ag to enable support for pointer arithmetic involving pointers to private data. We implement such 6Alternatively, custom SIGUSR1 or SIGUSR2 can be triggered if the user program is known not to use it. 19 \ffunctionality as described below. As mentioned before, in regular C, pointers can store any integer values and using pointer variables in arithmetic expressions will result in evaluating the expressions on the integer values stored in such variables. In the context of pointers to private data, we however distinguish between pointer objects and (integer) addresses that pointer objects store. Thus, pointer content is no longer equivalent to integer values, and we support pointer arithmetic with pointer objects only for the purposes of using pointers to store and manipulate addresses. Such arithmetic operations can be categorized into two groups: 1. Using pointers to private data in expressions of the type p + exp and p exp, where exp is an expression that evaluates to a (public) integer value. This operation produces the same output as executing &p[exp], i.e., all locations stored in p are advanced by the amount of space occupied by exp elements of the array. 2. Using pointers to private data in o\ufb00set computation as in p1 p2. This operation is straightforward to implement when both pointers store a single location. When, however, at least one of them has multiple locations, our implementation computes the private di\ufb00erence between the true locations of the pointers. This option, in our opinion, implements the right semantic value as opposed to other variants (such as computing pair-wise di\ufb00erences between all addresses stored in the pointers) which bear little meaning. Note that expressions of the type p + exp, and equivalently &p[exp], where exp evaluated to a private integer, are not meaningful and not supported. 5.4 Pointer Casting Variable casting refers to the ability to treat a variable of one type as a variable of another type. Casting a constant or variable of one type to a constant or variable of another type typically results in the value being preserved after the conversion (if possible) even if the two types use di\ufb00erent data representations. This means that conversion is likely to involve computation. In PICCO, conversion between \ufb02oating point and integer values is based on the algorithms given in [2], while conversion between integer types of di\ufb00erent sizes and \ufb02oating point types of di\ufb00erent sizes requires minimal to no work (assuming no over\ufb02ow or under\ufb02ow detection is required when casting a value to a shorter representation). Pointer casting is handled di\ufb00erently and C is unique in the sense of allowing pointer-based in-memory casting from one data type to another. Pointer casting involves no data conversion: the memory is read as is and is interpreted as a sequence of elements of another type. Thus, pointer casting is meaningful between a limited number of data types. In order to support pointer casting in PICCO, we need to resolve the main question: because data representation of private data types di\ufb00ers from data representation of the corresponding public data types, we need to determine how to mimic sizes of public data types when working with blocks of private data without modifying the data itself. That is, all secret shared values in PICCO are represented as elements of the same \ufb01eld, which means that, for example, shares of a 16-bit integer and shares of a 64-bit integer have the same bitlength. A programmer who casts memory storing an array of 64-bit integers to a pointer to an array of 16-bit integers, however, expects to extract four 16-bit integers from each 64-bit integer. This means that to meet the programmer\u2019s expectations, private data will need to be processed and assembled in a di\ufb00erent form. We, however, cannot modify the original data because only the pointer was cast, not the data itself. 20 \fInstead of duplicating the memory and performing conversion at the time of casting, our solution is to do the necessary computation at the time of pointer dereferencing. This means that we need to record information about the data type from which casting was performed (to the data type of the pointer) at the time of casting, but delay conversion until the pointer is dereferenced. We store casting data type information with the pointer and use it to extract the relevant portion of the memory at pointer dereferencing time. Note that in presence of a sequence of casts, only a single data type needs to be maintained because the memory layout does not change. Because in PICCO simple data types can be de\ufb01ned to have any bitlength, casting, for example, a pointer of one integer type to a pointer of another integer type does not guarantee that one data type will have a bitlength multiple of another. In that case we still calculate what the relevant portion of the memory is based on the position of the memory being dereferenced, but the last, partially \ufb01lled, element might not be reliably extracted. For example, suppose some memory was \ufb01lled as a 3-element 30-bit integer array. When it is cast to an array of 20-bit integers, the fourth elements will be extracted as bits 61\u201380 of the original data, while retrieving the \ufb01fth element might result in memory violation because there is not enough data in the original array to fully form that element. 5.5 Pointers to Functions Similar to pointers to ordinary data types, we need to distinguish between pointers to functions that will be treated as pointers to private data and pointers to functions that will be treated as pointers to public data. The former can be used inside conditional statements with private conditions (as a result of which they acquire multiple locations and the true location becomes private) and are restricted to functions with no public side e\ufb00ects. The latter can contain pointers to functions of any type, but cannot be modi\ufb01ed or dereferenced inside conditional statements with private conditions. The distinction is made at the time of pointer declaration using private/public quali\ufb01ers with void data type. That is, by using private void *p, p will be treated similar to other pointers to private data, while all pointers declared syntax public void *p will be treated as conventional C pointers. Private pointers to functions are supported naturally in our implementation. When a pointer stores a single location, the function is invoked as in conventional program execution. If, however, a pointer acquires multiple locations as a result of its modi\ufb01cation inside conditional statements with private conditions, at the time of pointer dereferencing all functions stored in the pointer will be executed, but only the e\ufb00ects of one of them will be applied. Conceptually this is the same as executing branching statements with private conditions: all branches are executed, but only the e\ufb00ects of one of them are applied depending on the result of private condition evaluation. That is, when a pointer p storing \u03b1 locations L = (\u21131, . . . , \u2113\u03b1) with the corresponding tags T = (t1, . . . , t\u03b1) is being dereferenced, each function fi stored at address \u2113i is being invoked. Then each (private) variable a that fi modi\ufb01es is set to a = ai \u00b7 ti + aorig(1 \u2212ti), where aorig and ai are its original and newly computed by fi values. If a is modi\ufb01ed by multiple fi\u2019s with indices i1 through ik, its value is updated as a = Pk j=1 aij \u00b7 tij + aorig(1 \u2212Pk j=1 tij) (recall that only one ti can be set to 1). 6 Analysis After discussing multiple aspects of private pointer design and its uses in programming, in this section we summarize the notion of pointer to private data and operations on it and formally show that program execution that involves pointers to private data complies with a standard de\ufb01nition of security used in secure multi-party computation. 21 \fA pointer to private data is de\ufb01ned as a C-style pointer to a private object storing location information of the object, where a private object can be one of the following types: 1. a primitive private data type (private int, \ufb02oat, etc.); 2. a composite data type (C-style struct), each element of which is a private object; 3. a function with no public side e\ufb00ects; 4. a pointer to a private object. The second and forth categories de\ufb01ne a pointer to private data recursively, which means that a pointer can have any indirection level, nested struct types, and any combination of primitive, composite, and pointer data types. The previously de\ufb01ned operations on pointers to private objects are: 1. Pointer read and update with value v, denoted as read(p) and update(p, v), respectively. 2. Dereferenced pointer read and update, denoted as read(*p) and update(*p, v), respectively. 3. Pointer update inside a conditional statement with a private condition. Following prior work, we use multiplexor notation to denote this operation as mux(p, v1, v2, cond), where p is set to v1 if private cond evaluates to 1 and to v2 otherwise. Pointer read inside a conditional statement with a private condition is processed identically to a conventional read read(p) and thus is not listed as a separate operation. 4. Dereferenced pointer update inside a conditional statement with a private condition, which we denote as mux(*p, v1, v2, cond). Similar to the previous case, processing of dereferenced pointer reads is not a\ufb00ected by the presence of conditional statements. 5. Dynamic memory allocation in the form of malloc; for an assignment p = pmalloc(n, type) we use notation alloc(p, n, type). 6. Dynamic memory deallocation as in pfree(p), denoted as dealloc(p). 7. Array indexing p[i] with a public index i. This is treated as a generalization of pointer dereferencing, and we use notation read(p, i), update(p, i, v), mux(p, i, v1, v2, cond) to denote read, update, and update inside a conditional statement with a private condition, respectively. 8. Array indexing p[i] with a private index i is also divided into three operations readp(p, i), updatep(p, i, v), and muxp(p, i, v1, v2, cond). These operations can only be performed on locations stored in a pointer that correspond to arrays with known bounds (i.e., allocated using the pmalloc interface or static array declaration). 9. Evaluation of predicate f on one or more pointers p1, . . ., denoted as pred(f, p1, . . . ). 10. Pointer casting, denoted as cast(p, type), where type is the data type to which p is cast. In the above, v, v1, and v2 correspond to either values associated with private objects or data structures corresponding to pointers to private objects, depending on the context. The value of cond is always a private bit, while variables n, i, and type are public. Any variable can be read inside a conditional statement with a private condition, but updates can be performed only as speci\ufb01ed using the mux operations. 22 \fThis list implicitly de\ufb01nes operations that cannot be performed on pointers to private objects and will be rejected by the compiler. That is, addresses of public objects cannot be used to update a pointer to private data (either via update or mux); a pointer to public data or a mix of private and private \ufb01elds de\ufb01ned with a struct construct cannot be modi\ufb01ed inside a conditional statement with a private condition (i.e., there is no corresponding mux operation); pmalloc and pfree cannot be called inside conditional statements with private conditions (i.e., there is no mux operations for them); similarly, casting cannot be called inside conditional statements with private conditions as it modi\ufb01es publicly stored data. Recall that our implementation of pointers maintains the invariant that in a well-formed program there is exactly one true location associated with a private object. The invariant may be violated only when memory associated with a pointer is being deallocated when the pointer is still in use (in such a case, no tag is set to 1 and dereferencing the pointer will return no data). In showing security of pointer-related operations that reference private data, we use a traditional simulation-based de\ufb01nition of security. Because PICCO is built on top of (n, t)-threshold secret sharing techniques with n \u22653 computational parties, we utilize the same setup in our security analysis. Similarly, because the underlying techniques o\ufb00er information theoretic security, we utilize statistical (as opposed to computational) indistinguishability in the security de\ufb01nition. De\ufb01nition 1. Let parties p1, . . ., pn engage in a protocol \u03a0 that evaluates program P on a mix of public and private data. Let VIEW\u03a0(pi) denote the view of pi during the execution of \u03a0, which is formed by its input, internal random coin tosses ri, and messages m1, . . ., mk passed between the parties during protocol execution: VIEW\u03a0(pi) = (ini, ri, m1, . . ., mk). Let I denote a subset of the participants of size t and VIEW\u03a0(I) denote the combined view of the participants in I during the execution of \u03a0. Protocol \u03a0 is said to be t-private in the presence of semi-honest adversaries if for each coalition of size at most t < n/2 there exists a probabilistic polynomial time simulator SI that given the input of the parties in I, P, and P\u2019s output, produces a view statistically indistinguishable from VIEW\u03a0(I) together with the output of the parties in I. Theorem 1. Any program P augmented with pointers to public and private data and no out-ofboundary access compiled by PICCO is translated into a t-private protocol for any t < n/2 when the computation is carried out by n parties. Proof. Our proof proceeds by evaluating each operation involving a pointer to a private object as summarized above. After building a simulator for each pointer-related operation, we apply the composition theorem of Canetti [6] to the result, which would guarantee that any combination of these operations (and other secure operations in PICCO) results in security of the overall program P that the computational parties execute. Building our simulator requires only the use of t-private implementations of addition/subtraction and multiplication operations (which is met in PICCO by the underlying linear secret sharing scheme with t < n/2). We describe a simulator for each operation involving a pointer to a private object in turn. Prior to each operation, each party in the adversarial coalition I holds a share of each relevant private data item (including private input data, private \ufb01elds of pointer data structures, etc.) and at the time of operation termination each party in I holds a share of the output and/or updated private items. Many functions also modify data publicly available to each party (including the simulator). \u2022 read(p): This operation simply retrieves the data structure stored in p and is local to each computational party. The simulator does not interact with the parties in I. \u2022 update(p, v): The data structure contained in v is simply copied into p by each party. The simulator does not interact with the parties in I. 23 \f\u2022 read(*p): When p stores only a single location, the value stored at that location is retrieved and the simulator does not interact with the parties in I. When p stores multiple locations, each party is instructed to iterate through the locations extracting values stored in them and then combine the values using the private tags associated with each location as described in Sections 4.2 and 4.3. Most of the procedure operates on public values (such as locations) and the only private computation consists of multiplying tags with private data (or other tags in case of pointers to pointers). The simulator thus participates in each multiplication operation on behalf of honest users by invoking a simulator corresponding to the multiplication operation the necessary number of times. When p is a pointer to an object of a complex data type declared using struct and a single \ufb01eld of p is being dereferenced (as in p->x), the way the simulator interacts with the parties in I is not a\ufb00ected. (Only the locations from which values are retrieved are locally modi\ufb01ed by a known o\ufb00set by each party.) \u2022 update(*p, v): Similar to reading a referenced pointer, when p is associated with only one location, value v is stored in that location without the parties interacting with the simulator. When p stores multiple locations, each party in I iterates through all possible locations and updates all locations using secure multiplications as described in Sections 4.2 and 4.3. Because the locations (and their number) are always public, the simulator only needs to engage in a pre-determined number of secure multiplications simulating all honest users, which is does by invoking a simulator of the secure multiplication protocol. When p points to a struct and only one \ufb01eld p->x is to be updated, the simulator\u2019s interaction with the parties in I is identical (but the valued are retrieved and the computed values are placed not at the locations stored in p but the locations adjusted by the o\ufb00set of x in the struct). \u2022 mux(p, v1, v2, cond): To implement conditional assignment, each party (and the simulator) \ufb01rst locally merges the lists stored in v1 and v2, after which the (private) tags needs to be updated based on the private condition cond as described in Algorithm 1. The simulator only needs to participate in a certain number of secure multiplications determined by the public contents of v1 and v2, which it performs as before by invoking a multiplication operation simulator. \u2022 mux(*p, v1, v2, cond): Conditional update of a dereferenced pointer involves modifying the value stored at each location in p with a \ufb01xed function of v1, v2 and cond (see Section 4.2). Thus, the simulator participates in a pre-determined number of secure multiplications simulating all honest users using a simulator for secure multiplication. \u2022 alloc(p, n, type): Allocating memory for a number n of private objects of type type does not involve secure computation and thus the simulator does not interact with the parties in I for the purpose of this operation. \u2022 dealloc(p): First, if one of the locations stored in p corresponds to the special \u201cuninitialized\u201d value, no party (including the simulator) performs any operation. Otherwise, the \ufb01rst location \u21131 is removed from the data structure that p stores and the values at all other locations are updated using two secure multiplications (as described in Section 5.2). The simulator produces communication on behalf of all honest users to simulate invocations of the secure multiplication protocol as before. The parties consequently locate other pointers that store \u21131 and update their locations and tags using a procedure that depends only on public data. 24 \fThe simulator is invoked to simulate a necessary number of secure multiplications on behalf of honest users using the simulator of secure multiplication. \u2022 read(p, i): Similar to a dereferenced pointer read, the simulator will need to simulate 0 or more invocations of secure multiplications on behalf of honest users. The only di\ufb00erence from the simulation of read(*p) is that the pointer locations are (locally) adjusted by the space occupied by i items by each party (including the simulator) during the computation. \u2022 update(p, i): This procedure is also very similar to update(*p), where the di\ufb00erence is only in the local (publicly available) computation. \u2022 mux(p, i, v1, v2, cond): Similar to mux(*p v1, v2, cond), the simulator will need to participate in a pre-determined number of secure multiplications simulating all honest parties using its simulator after some computation on public data. \u2022 readp(p, i): To retrieve an element of an array represented by a pointer using a private index, for each location stored in p that corresponds to a properly allocated memory block, the parties together with the simulator perform a private table lookup. This operation is implemented in PICCO by reading each element of the array and can be easily simulated once the array size is determined using information publicly stored by each computational party. Thus, for each eligible address stored in p the simulator simulates private table lookup operation on behalf of honest users by invoking a private table lookup simulator, after which the results from multiple locations (if present) are combined together using private tags, during which the simulator engages in secure multiplication simulation. \u2022 updatep(p, i, v): This operation proceeds similar to readp(p, i), but with some operations performed in a di\ufb00erent order. Each party in I and the simulator \ufb01rst locally determine eligible addresses in p using public information, after which the parties and the simulator need to modify the data being stored for each location using the location\u2019s tag through secure multiplication and then engages in the interaction for table update at a private location. Thus, the simulator needs to engage in a pre-determined number of secure multiplication simulations and then in simulating interaction corresponding to private table updates. As before, the simulator can simply invoke the corresponding simulators for secure multiplication and table updates at a private location. \u2022 muxp(p, i, v1, v2, cond): Conditional update of an array element at a private index is performed similar to the regular update. Here, the parties and the simulator can \ufb01rst combine data of v1 and v2 using cond into a single v, where the simulator will need to engage in simulating a \ufb01xed number of secure multiplications. After this they can follow the steps of updatep(p, i, v) with the simulator producing messages as described for the simulation of that operation. \u2022 cast(p, type): This operation only updates public information associated with pointer p and the simulator does not interact with the parties in I. \u2022 pred(f, p1, . . . ): The simulator proceeds by evaluating the operations in f in order. For any operation that uses a single pointer, the simulation proceeds according to one of the above cases. For any operation that uses two pointers (most signi\ufb01cantly, pointer equality or inequality testing), the simulator \ufb01rst computes the public/private status of the result. If the status is public, the result is determined locally. Otherwise, the simulator engages in the computation associated with determining the (private) result by invoking the simulators 25 \fcorresponding to the operations used in the computation. Once any intermediate result in evaluating f is computed to be private, any computation that uses the result is carried out on private data by invoking simulators corresponding to the operations used in f. This covers all possible operations with pointers to private objects and concludes the proof. We note that our security result above is stated for programs that PICCO successfully compiles and that do not contain out-of-boundary array accesses. As previously discussed, the compiler will reject almost all types of improperly written programs during static analysis at compilation time. For example, if a pointer to public data is modi\ufb01ed inside a conditional statement with a private condition, the program will not compile. This means that the compiler will \ufb01lter out the majority of programs with errors, while any well-formed program enjoys simulatable security. The main type of errors that the compiler may not be able to detect during static analysis deal with memory access to invalid locations (e.g., out of boundary array access, access to a hard-coded invalid location, etc.). When such accesses are triggered, one or more computational parties might not be able to continue with the execution and quit on error, but no privacy violations take place. This is because while incorrect shares of private data might be read, the data will still remain in the protected form and cannot be reconstructed without the programmer\u2019s original intent. Note that this discussion applies only to dereferencing a pointer with no o\ufb00set (e.g., as in *p) or with a publicly known o\ufb00set (e.g., as in p[i] with public i) because in our implementation accessing array at a private location never results in out of boundary access (i.e., calling this operation on an improperly allocated array will be skipped at runtime). Because the memory layout may di\ufb00er on di\ufb00erent platforms, illegal memory accesses might result in di\ufb00erent behavior of the computational parties (e.g., execution at one party might result in a segmentation fault faster than at another party). As our simulator is not guaranteed to run in an identical environment to those of the honest parties, the simulation might be distinguishable for programs with illegal memory accesses based on the point of execution when any given party aborts the computation. Thus, we exclude such programs from the security statement of Theorem 1, but still guarantee that any program that compiles in our framework will not result in privacy violations. To summarize, our solution guarantees that any program that can be compiled in our framework never reveals any unauthorized information about private data. Simulation of a poorly written program with illegal memory accesses may not be indistinguishable from its real execution, but any properly written program enjoy simulation-based security. 7 Pointer-Based Data Structures There are several popular data structures typically built using pointers. In this section we discuss how they would be implemented using pointers to private data and in what complexities their performance results. In particular, we explore linked lists, trees, stacks, and queues. 7.1 Linked Lists A linked list consists of a sequentially linked group of nodes. For a singly linked list, each node is composed of data and a reference in the form of a pointer to the next node in the sequence, while for more complex variants such as doubly and circular linked lists the reference \ufb01eld incorporates additional links. A linked list allows for e\ufb03cient node insertion and removal, which makes it an ideal candidate for implementation of stacks and queues as well as representation of graphs that uses an adjacency list. In what follows, we discuss implementations of linked lists that store private 26 \fdata. We start by analyzing various operations in standard linked lists and then elaborate on the special case when a linked list stores sorted data. The latter does not represent a typical use of linked lists in programming (and does not necessarily have attractive features), but is provided as a relatively simple way to demonstrate what form working with sorted data can take in a secure computation framework. Standard linked lists. Because of ubiquitous use of linked lists in programming, we analyze di\ufb00erent possible uses of linked lists and the corresponding operations. When a linked list stores public data, node insertion has cost O(1) as a node is inserted in a \ufb01xed place (e.g., beginning or end of the list). Performing a search requires O(n) time, where n is the number of nodes in the list, because the nodes are traversed sequentially. Deleting a node from a \ufb01xed place (i.e., beginning or end of the list as done in the case of stacks and queues) involves O(1) time, but when deletion is preceded by a search (and the found node is deleted), the search together with deletion require O(n) time. When a linked list stores private data, the reference \ufb01eld holds a pointer to private data (i.e., a record of the same type) and at the time of node creation, the pointer stores a single location. Without loss of generality, let nodes be inserted in the beginning of the list. Node insertion then places a new node in the beginning of the list manipulating pointers as before, which still takes O(1) time and is very e\ufb03cient. Searching a list involves n private comparisons and all nodes need to be processed as not to reveal the result of individual comparisons on private data and the total work is O(n). Similarly, when a node is deleted from the beginning (or end) of the list, the time complexity of the operation is O(1) and each node\u2019s pointer still stores a single location. It is only when nodes need to be removed from varying positions in the list and the position itself needs to be protected, pointers can start acquiring multiple locations, which causes the time complexity of list traversal and deletion after a search to go up. However, when the fact whether the searched data was found in the list or not must remain private, we cannot remove any node, but instead need to erase the content of a found node (if present) with a value that indicates \u201cno data\u201d. In this case, all pointers still contain a single location and the cost of list search and other operations do not change, but the list will never reduce in its size. We defer the discussion of the case when the node is guaranteed to be found in a search and needs to be removed from a private location until the end of this subsection. Throughout this section, we express complexities of the operations as a function of n, where n is the \u201cvisible\u201d list size. This value will correspond to the actual list size when delete operations remove an element from the list so that its size is reduced, while it will correspond to the number of insertions (i.e., the maximum list size) when delete operations only mark data as deleted (to hide whether the element was found or not), but not reduce the list size. The same notation applied to other data structures discussed in this section when insertions/deletions are based on the result of private conditions. Sorted linked lists. As mentioned before, we discuss sorted linked lists only as a means of demonstrating how sorted data might be processed using a general-purpose secure computation compiler and it should be understood that this is not a typical use of linked lists or even not the best way of working with sorted data. We use the results of this discussion in our consecutive description. Now when a node is being inserted in a linked list, the insertion position must be determined based on the data stored in the list, which involves O(n) time with public data (and the complexities of other operations are the same as before). When we work with private data, the location where the node is being inserted must remain private (since it depends on private data) and the execution needs to simulate node insertion at every possible position. Consider the following two ways of 27 \finserting a node and the performance in which they result: 1. Pointer updates: The \ufb01rst is a traditional implementation of node insertion in a linked list, where if the correct insertion point is found, we update the pointer of the found node in the list to point to the new node and the pointer of the new node to point to the next node in the list. Because this conditional statement is based on private data, this will result in adding one location to the pointer in the found node and one location to the pointer in the node being inserted. After executing this operation for every node of the list, the pointer of each node in the original list stores 2 locations and the pointer in the newly inserted node stores n locations. When this operation is performed repeatedly, each node in the list acquires more and more locations (to the maximum of the current list size). This means that if the list is built by inserting one node at a time, the cost of node insertion and list traversal becomes O(n2). Node deletion after a search also takes O(n2) time, while node deletion from a \ufb01xed location is bounded by O(n). When, however, only a constant number of nodes are inserted into an existing list (which, e.g., can be provided as sorted input into the program), the complexity of all operations are unchanged from the public data case. 2. Data updates: Another possible implementation of sorted linked lists is to always insert a new node at the beginning of the list and keep swapping its content with the next node on the list until the correct insertion point is found. When this algorithm is implemented obliviously on private data using an SMC compiler, the computation processes each node on the list starting with the newly inserted node and based on the result of private comparison of current and new data either performs the swap or keeps the data unchanged. After each node insertion the reference \ufb01eld of each node still points to a single node in the list and therefore the complexity of all operations are unchanged from their public versions. Thus, it is clear that we want to avoid acquiring a large number of locations in each reference \ufb01eld of a pointer-based data structure and privately moving data (as opposed to privately moving pointers) is preferred when working with sorted data. Node deletion in standard linked lists. We can now return to the question of deleting a node from a private location in a standard (unsorted) linked list when it is known that the searched node is present and needs to be removed from the list. The above two approaches of inserting a node in a private position also apply to deleting a node from a private position. The \ufb01rst, standard, approach of manipulating pointers will result in acquiring multiple locations at each pointer, which degrades performance of all operations. Using the second approach of data updates, we can obliviously place the data to be deleted into the \ufb01rst node on the list (after scanning the nodes and swapping values based on private data comparisons) and then simply remove it from the list. This will maintain optimal complexities of all operations. The above tells us that traditional implementations of data structures can exhibit performance substantially worse than alternative implementations in a secure computation framework and our analysis can be viewed as a step in making informed decisions about implementation needs. 7.2 Trees Trees implement hierarchical data structures commonly used to store sorted data and make searching it easy. A tree node is typically comprised of data and a list of references to its child nodes. In an n-node balanced search tree, all of searching, node insertion, and node deletion take O(log n) time. Unfortunately, these complexities greatly change when we write a program to implement a 28 \fsearch tree on private data. In what follows, we distinguish between trees that are pre-built using the information available prior to the start of the computation and trees built gradually using information that becomes available as the computation proceeds: 1. Pre-built trees. Consider a balanced binary search tree and suppose that we want to perform a search on the tree. A traditional implementation involves O(log n) conditional statements to traverse the tree from the root to a leaf choosing either the left or right child of the current node. When the data is private, such statements use private conditions and thus both branches of the computation must be executed. The result is that the sequence of O(log n) nested private conditions results in executing all possible O(n) branches of the computation and touches all nodes in the tree. This is an exponential increase in the complexity compared to working with public data, even if we do not consider node insertions and deletions that result in node rotations to balance the tree (which are discussed next together with gradually-built trees). 2. Gradually-built trees. By analogy with inserting nodes into a sorted linked list, we can either manipulate pointers to insert a new node at the appropriate place in the tree or insert the node in a \ufb01xed location and move the data in place. The complexity of the latter option is O(n) for insertions, deletions, and search and we take a closer look at the former. As we traverse the tree looking for the place to insert the new node, similar to searching, all nodes will be touched (as a result of nested private conditions). Furthermore, because the execution cannot reveal the place into which the new node is inserted, pointers in all nodes will acquire new locations. If we add computation associated with node rotations when the tree becomes unbalanced, pointers will be acquiring new locations even faster (to the maximum of n\u22121 per pointer). After repeatedly calling insert to gradually build the tree, eventually each node will point to all other nodes resulting in O(n2) complexity for insertions, deletions, and searching. Such complexity is clearly avoidable and alternative implementations should be pursued. Search trees represent the worst possible scenario where implementing an algorithm on private data using a general-purpose compiler incurs an exponential increase in its runtime compared to the public data counterpart. As is evident from our discussion of linked lists and trees, searching an n-element store for a single element cannot be performed in less than linear time using generic techniques, regardless of whether the data is stored sorted or not. It means that without custom, internally built implementations of speci\ufb01c data structures it is conceptually simpler and more e\ufb03cient to maintain data in unsorted form, use append for insertion (O(1) time), and shift data to implement deletion. 7.3 Stacks A stack is characterized by the last-in, \ufb01rst-out (LIFO) behavior, which is achieved using push and pop operations. It has several fundamental applications such as parsing expressions (e.g., parsing programs in compilers), backtracking, and implementing function calls within an executable program. To the best of our knowledge, despite its popularity, this data structure has not been studied in the context of secure multi-party computation before and our analysis and consecutive implementation of stack that works with private data demonstrate its appeal for secure computation. A pointer-based implementation of a stack is built using a linked list, where a node is always inserted at the head of the list and is always removed from the head as well, either of which takes O(1) time. As was discussed in section 7.1, implementing these operations on private data maintains constant time complexities. 29 \fWhen using a stack with private data, we also consider the possibility that push and pop operations might be performed inside conditional statements with private conditions, in which case it is not publicly known whether the operation takes place and what record might be on top of the stack. Then if we implement a conditional private push operation by manipulating pointers, the top of the stack will store m + 1 locations when the last m push operations were based on private conditions. Implementing a push operation is then equivalent to executing the code: 1. node p = new node(); 2. if (priv-cond) 3. p->next = top; 4. top = p; Because both p and p->next store only a single location at the time of conditional push, merging the lists of p->next and top takes O(m) time. Similarly, merging the lists of top and p takes O(m) time. Implementing a pop operation within a private condition involves executing code: 1. if (priv-cond) 2. temp = top; 3. top = top->next; 4. // use temp The complexity of this operation is dominated by the second assignment. Because top points to O(m) locations, and the next \ufb01eld of each of its locations can store O(m) locations as well, the overall complexity of that assignment is O(m2). This means that the worst time complexity of a conditional push becomes O(n) for a stack containing n records and it is O(n2) for a conditional pop. If we instead implement push and pop operations that depend on private conditions by maintaining a single chain of records (with pointers containing a single location) and data update, push and pop operations result in O(1) and O(n) work, respectively. That is, we can always insert a new node (with data or no data depending on the private condition) into the stack and take O(n) time during pop to privately locate the \ufb01rst node with data (and erase the data as necessary). 7.4 Queues Queue is another important data structure used to maintain a set of entities or events in a speci\ufb01ed order which are waiting to be served. We can distinguish between \ufb01rst-in, \ufb01rst-out (FIFO), last-in, \ufb01rst-out (LIFO), and priority queues. Implementing a queue involves maintaining two pointers: the head and the tail. The head points to the beginning of the queue, i.e., the element that will be removed by a dequeue operation, and the tail points to the last element added to the queue using an enqueue operation. Similar to the stack, when enqueue and dequeue operations in a FIFO queue are implemented on public data or private data outside of private conditional statements, their complexities are O(1). Their complexities for enqueue and dequeue operations are also O(n) and O(n2), respectively, when implemented through private pointer manipulation (the implementation needs to maintain two pointers for the head and tail of the queue, but updating the second pointer does not asymptotically increase the amount of work) and O(1) and O(n), respectively, when private data update is used. In a priority queue, each node additionally stores priority (which we assume is private) and dequeue removes a node with the highest priority. The complexity of priority queue operations 30 \fData structure Insert Delete Search Linked list O(1) O(1) O(n) Linked list (delete at private location) O(1) O(n) O(n) Search tree O(n) O(n) O(n) Stack or queue O(1) O(1) \u2014 Stack (conditional private push & pop) or O(1) O(n) \u2014 queue (conditional private enqueue & dequeue) Priority queue O(1) O(n) \u2014 O(n) O(1) \u2014 Priority queue (conditional private enqueue & dequeue) O(1) O(n) \u2014 Table 1: Performance of various data structures using pointers to private data. depends on the underlying data structure used to implement it. The best known complexities for public data are O(log n) for enqueue (O(1) average case) and O(log n) for dequeue using a heap. Suppose for now that all operations are outside conditional statements with private conditions. If we use a linked list to store queue nodes, the best performance can be achieved using O(1) for enqueue and O(n) for dequeue (i.e., always store a newly inserted node in the beginning and remove the highest priority node from a private location as a result of the search for the highest priority element) or O(n) for enqueue and O(1) for dequeue (i.e., store the list sorted and always remove the \ufb01rst node during dequeue). Then if the operations depend on private conditions, we can maintain O(1) for enqueue and O(n) for dequeue if the operations depend on private conditions using a very similar approach to that of regular queues and stacks. That is, we always insert an element into the beginning of the queue as a result of a conditional enqueue (if the condition is false, the element is empty), and during dequeue we scan the queue for the highest priority element and erase it from the queue if the condition is true. If the underlying implementation is a heap, we insert a new node in a \ufb01xed leaf location and use O(log n) compare-and-exchange operations to maintain the invariant of a max-heap to implement enqueue. Realizing dequeue, however, requires O(n) work because it cannot be revealed what path was traversed from the root to a leaf (since the path choice depends on private priorities). Similar to other implementations, we can maintain these complexities even when enqueue and dequeue are performed as a result of private condition evaluation. 7.5 Summary Before we conclude this section, we would like to summarize performance of di\ufb00erent data structures that can be implemented on private data using newly introduced pointers to private data or records. Table 1 lists the best performance we could achieve using a pointer-based implementation of the data structures discussed in this section. Recall that in all data structures with conditional operations performance depends on the number of insertions as opposed to the actual data size. We note that the complexities in Table 1 can be directly linked to the amount of memory consumed by those data structures, with small \ufb01xed constants hidden behind the asymptotic notation. These data structures can also be evaluated using alternative mechanisms. For example, our analysis suggests that implementing these data structures using arrays of private data instead of pointers to private data would result in the same complexities (which is often the case for public data as well). Also, utilizing ORAM-based implementation can improve asymptotic complexity of some (but not all) data structures and can lead to faster runtime in practice at least for large enough data sets. The most pronounced bene\ufb01t of using ORAM will be observed for implementing search trees, where all operations can be performed in polylogarithmic (in n) time (e.g., using the 31 \fsolution in [22]). On the other hand, using ORAM for linked lists can only increase the complexity of its operations (even the complexity of a delete at a private location following a search cannot be reduced below O(n)). Other data structures that can bene\ufb01t from ORAM-based implementations are stacks and queues where the operations that update the data structures are performed inside private conditional statements. ORAM techniques, however, involve larger constants behind the big-O notation than simple operations and their initial setup cost is also signi\ufb01cant. We thus leave a thorough comparison of ORAM vs. pointer or array based implementations of various data structures in this framework as a direction of future work. 8 Performance Evaluation In this section, we report on the results of our implementation and evaluation of a number of representative programs that utilize pointers to private data. Because such programs have not been previously evaluated in the context of secure multi-party computation, we cannot draw comparisons with prior work. In some cases, however, we are able to measure the cost of using pointers, or the cost of a pointer-based data structure, in a program by implementing the same or reduced functionality that makes no use of pointers. Note that because PICCO can be used for both secure multi-party computation and outsourcing, the inputs in these programs can come from one or more input parties/clients. The programs that we implemented and evaluated as part of this work are: 1. The \ufb01rst program constructs a linked list from private data read from the input and then traverses the list to count the number of times a particular data value appears in the list. This is a traditional implementation of a linked list, where each record with private data is prepended to the beginning of the list when building it. The program is given in Figure 1. We next notice that this program is sub-optimal in terms of its run time because it does not utilize concurrent execution capabilities provided in PICCO. For that reason, we also implement an optimized version of this program. The di\ufb00erence is that all private comparisons during the list traversal are executed in a single round using PICCO\u2019s batch constructs. 2. To evaluate pointer-based implementations that work with private data maintained in a sorted form, and more generally privately manipulating pointer locations vs. obliviously moving data, we build a program for a sorted linked list. The functionality of this program is similar to that of the \ufb01rst program (i.e., create a linked list and then traverse it to count the number of occurrences of a given data item in it) and the di\ufb00erence is in the way the list is build. We evaluate two variants of the program corresponding to pointer update (PU) and data update (DU) as described in section 7.1. The program for the DU variant is given in Figure 2, and the program for the PU variant is given in Figure 3. 3. The third program implements mergesort that takes an array of unsorted integers as its input. The program makes an extensive use of pointers to private data to pass data by reference to a function that conditionally swaps two data items based on their values (i.e., performs the so-called compare-and-exchange operations). Mergesort was chosen not necessarily because it provides the best performance for an oblivious sort, the objective instead was to demonstrate how performance of a program that utilizes pointers to private data (and exercises modular design of a program) compares to a similar program that does not use pointers. We thus also evaluate another version of mergesort that performs compare-and-exchange operations in place (without calling any function) and makes no use of pointers. The pointer-based mergesort is given in Figure 4 and its non-pointer-based implementation is given in Figure 5. 32 \fstruct node { private int data; struct node *next; }; public int count = 128; public int main() { public int i; private int array[count], output; struct node *ptr, *head = 0; smcinput(array, 1, count); //construct the list for (i = 0; i < count; i++) { ptr = pmalloc(1, struct node); ptr->data = array[i]; ptr->next = head; head = ptr; } //traverse the list privaate int val = 10; ptr = head; for (i = 0; i < count; i++) { if (ptr->data == val) output = output+1; ptr = ptr->next; } smcoutput(output, 1); return 0; } Figure 1: Construction and traversal of a linked list. 4. Our last program implements a shift-reduce parser for a context-free grammar (CFG) on private data. This is one of fundamental applications that can now be naturally implemented using the compiler by building and maintaining a stack, once support for pointers to private data is in place. We choose a CFG that corresponds to algebraic expressions consisting of additions, multiplications, and parentheses on private integer variables, which is speci\ufb01ed as follows: statement \u2192statement | statement + term term \u2192term | term * factor factor \u2192var | (statement) Here, all variables are shown in italics, while terminals are set in bold font. The grammar can obviously be generalized to more complex expressions and programs that work with private as well as public variables of di\ufb00erent types. We view this application as enabling one to evaluate a custom function on private data without writing and compiling a separate program for each function. That is, both the function to be evaluated and its input (consisting of private data) are provided as input to the parser. We note that it is possible for the function or the grammar rules to be private as well, but this would result in an increase in the program performance. 33 \fstruct node { int data; struct node *next; }; public int count = 128; public int main() { public int i, j; private int array[count], output, tmp; struct node *head, *ptr1, *ptr2; smcinput(array, 1, count); //construct the list head = pmalloc(1, struct node); head->data = array[0]; for (i = 1; i < count; i++) { ptr1 = pmalloc(1, struct node); ptr1->data = array[i]; ptr1->next = head; head = ptr1; ptr2 = head; for (j = 0; j < i; j++) { if (ptr2->data > ptr2->next->data) { tmp= ptr2->data; ptr2->data = ptr2->next->data; ptr2->next->data = tmp; } ptr2 = ptr2->next; } } //traverse the list private int val = 10; ptr1 = head; for (i = 0; i < count; i++) { if (ptr1->data == val) output = output+1; ptr1 = ptr1->next; } smcoutput(output, 1); return 0; } Figure 2: Construction and traversal of a sorted linked list (using data update). Our parser uses one lookahead character, and due to the complexity of the implementation, the program itself is not included in the paper. To approximate performance overhead associated with using a pointer-based stack, we create a program that performs only arithmetic operations on private data which are given to the parser and which the parser executes. Note that unlike evaluation of mergesort, these are 34 \fProgram Field Data size size (bits) 25 28 211 214 217 220 Linked list 81 0.0004 \u00b1 5% 0.003 \u00b1 4% 0.014 \u00b1 3% 0.097 \u00b1 2% 0.760 \u00b1 1% 5.40 \u00b1 1% (list building, traversal, 0.086 \u00b1 1% 0.661 \u00b1 1% 5.30 \u00b1 3% 42.27 \u00b1 1% 337.8 \u00b1 1% 2,692 \u00b1 2% and optimized traversal) 0.026 \u00b1 7% 0.140 \u00b1 3% 1.019 \u00b1 1% 8.051 \u00b1 2% 63.75 \u00b1 1% 513.8 \u00b1 1% Shift-reduce parser 33 0.005 \u00b1 9% 0.039 \u00b1 2% 0.307 \u00b1 1% 2.439 \u00b1 1% 19.73 \u00b1 2% 157.2 \u00b1 1% Arithmetic operations 33 0.005 \u00b1 9% 0.038 \u00b1 3% 0.294 \u00b1 1% 2.336 \u00b1 1% 18.85 \u00b1 1% 150.8 \u00b1 1% Table 2: Performance of representative programs with unsorted data structures measured in seconds. Program Field Data size size (bits) 24 25 26 27 28 29 Sorted linked list (DU) 81 0.466 \u00b1 1% 1.908 \u00b1 1% 7.750 \u00b1 1% 31.24 \u00b1 1% 125.5 \u00b1 1% 565.5 \u00b1 1% (list building and traversal) 0.036 \u00b1 1% 0.071 \u00b1 1% 0.142 \u00b1 1% 0.284 \u00b1 1% 0.567 \u00b1 1% 1.311 \u00b1 1% Sorted linked list (PU) 81 1.464 \u00b1 1% 9.956 \u00b1 1% 85.51 \u00b1 2% 918.6 \u00b1 2% 9,900 \u00b1 3% N/A (list building, traversal, 0.051 \u00b1 1% 0.149 \u00b1 2% 0.613 \u00b1 2% 5.285 \u00b1 2% 45.93 \u00b1 2% N/A and head node removal) 0.005 \u00b1 1% 0.015 \u00b1 1% 0.044 \u00b1 2% 0.174 \u00b1 2% 0.720 \u00b1 3% N/A Mergesort without pointers 81 0.053 \u00b1 5% 0.121 \u00b1 5% 0.271 \u00b1 5% 0.625 \u00b1 4% 1.453 \u00b1 4% 3.124 \u00b1 4% Mergesort with pointers 81 0.053 \u00b1 4% 0.122 \u00b1 5% 0.272 \u00b1 5% 0.638 \u00b1 6% 1.503 \u00b1 5% 3.201 \u00b1 5% Table 3: Performance of representative programs with sorted data structures measured in seconds. not equivalent functionalities. That is, one program is much more complex, parses its input according to the CF grammar, maintains a stack, etc., while the other only performs additions and multiplications. Note that most of these programs already exercise dynamic memory allocation (i.e., all linked list programs and the shift-reduce parser). However, to provide a more complete evaluation of dynamic memory management, we also include experiments that measure the overhead of dynamic memory deallocation. Thus, we incorporate calls to pfree to two programs: (i) we call pfree as part of the shift-reduce parser at the end of each pop operation and (ii) we evaluate the cost of removing the head node in a sorted linked list built using pointer update (Figure 3). These were chosen as natural applications of memory deallocation, where pointers to private objects contain a single and multiple locations, respectively. In the second case, the head stores locations of all nodes on the list and the overhead of pfree includes updating the structures of other pointers on the list upon memory deallocation. Each program was compiled using PICCO, extended with pointer support as described in this work, and run in a distributed setting with three computational parties. All compiled programs utilize the GMP library for large number arithmetic and OpenSSL to implement secure channels between each pair of computational parties. We ran all of our experiments using three 2.4 GHz 6-core machines running Red Hat Linux and connected through 1Gb/s Ethernet. Each experiment was run 10 times, and we report the mean time over all runs and the corresponding deviation from the mean observed in the experiments. The results of the experiments for working with unsorted and sorted data are given in Tables 2 and 3, respectively. As can be seen from the tables, each program was run on data of di\ufb00erent sizes. For all linked lists programs as well as mergesort, the data size corresponds to the number of elements in the input set, while for the shift-reduce parser and arithmetic operations the size corresponds to the number of arithmetic operations in the formula, which were a mix of 90% multiplications and 10% additions. All linked list experiments contain two di\ufb00erent times, which correspond to the times to build and traverse the linked list, respectively. The tables also report the size of \ufb01eld elements in bits used to represent secret shared values. While all programs were written to work with 32-bit 35 \fintegers, most programs in the table use statistically secure comparisons, which requires the length of the \ufb01eld elements to be increased by the statistical security parameter (which we set to 48). (The size of the \ufb01eld elements needs to be the size of the data plus one bit to ensure that all data values can be represented.) The results tell us that working with linked lists (the \ufb01rst program in Table 2) in the secure computation framework is very e\ufb03cient. That is, building a linked list that consists of thousands of elements takes a fraction of a second. Traversing a linked list is also rather quick, where going through a linked list of size 211 about 1 second in our optimized program. Performance of the sorted linked lists (the \ufb01rst two programs in Table 3) characterizes performance expected from di\ufb00erent data structures where it is necessary to hide the place where a new node or data item is being inserted. As previously mentioned, there is no good reason to implement the PU variant of di\ufb00erent data structures and it is provided here for sorted linked lists for illustration purposes only. The DU version of sorted linked list has the same list traversal time as the regular (unsorted) linked lists, and the reported time for sorted linked lists can be further optimized in the same way as it was done for regular linked lists. When we are building a sorted linked list via DU, each operation takes O(n) time and thus the time to perform this operation for all n elements of the input is O(n2). This quadratic performance is also observed empirically where increasing the size of the data set by a factor of 2 results in four-time increase in the list building time (all insertion operations are performed sequentially). As far as the head node removal operation in a sorted linked list with PU goes, it consists of two pointer dereferences (i.e., using data and next \ufb01elds) and one call to pfree, where the overhead of pfree was between 76.4% to 81.3% of that operation\u2019s time. In this particular experiment, each pointer stores O(n) locations, which contributes to the complexity of both memory deallocation and pointer dereferencing, but the latter operation can be performed more e\ufb03ciently. If we next look at mergesort (the last two programs in Table 3), we see that the variant that uses pointers to private data and makes a function call to a compare-and-exchange operation for each comparison and the variant with no pointers and corresponding function calls di\ufb00er in their performance by a very small amount. The non-pointer version that performs less work is faster by 0.4\u20132.4%. Lastly, the performance of our shift-reduce parser (the second program in Table 2) is extremely fast and is almost entirely consists of the time it takes to evaluate the provided formula on private data (the second program in Table 2). That is, despite having a more complex functionality and employing pointer-based stack, the time to perform arithmetic operations only is almost the same as the time the parser takes. Also, adding pfree to the program does not e\ufb00ect the runtime (because the pointer stores a single address) and the times with memory deallocation are omitted from the table. As far as memory consumption goes, the introduction of pointers to private data only marginally a\ufb00ects the amount of allocated memory for programs with pointers storing a single memory location (linked list, shift-reduce parser, and mergesort). The amount of memory needed to store and process sorted linked lists is quadratic in the data size and matches in its complexity list traversal. Removing a node from the list and calling pfree reduces the memory consumed by the data. While in general calls to pfree can increase memory consumption, in this case all pointers store the same lists of O(n) locations and removing a node and merging the lists in pfree decreases the size of each list. In general, we can say that memory consumption is at most quadratic in the amount of data and user-declared variables in any program. We note that all functionalities used for our experiments have alternative implementations using arrays. For linked lists, mergesort, and a shift-reduce parser, we expect array-based implementation to exhibit very similar performance to that based on pointers because all pointers store a single 36 \flocation. For sorted linked lists, we expect array-based programs to have performance similar to our data update implementation (with the same asymptotic complexities). To con\ufb01rm this \ufb01nding, we evaluated performance of array-based sorted linked lists, the source code of which can be found in Figure 6. Building the sorted list took about 20% less time using arrays for most data sizes, while list traversal was about 9% slower using arrays for most data sizes. Thus, both implementations exhibit comparable performance. Memory consumption is also similar in most programs, with the exception of array-based sorted list implementation that uses memory linear in the data size. Performance of our pointer-based programs can also be compared to that of array-based implementations using another system or compiler. Sharemind [5] is a powerful system that supports a wide range of programs and, similar to PICCO, builds on (a di\ufb00erent type of) information-theoretic secret sharing, which is hand-optimized to work with three computational parties. Despite similarities of the setup, Sharemind programs exhibit signi\ufb01cantly di\ufb00erent performance characteristics. In particular, the implementation is optimized for performing a large number of identical operations in a batch, while the cost of performing only a single operation is high (e.g., on the order of 100ms for a single integer equality test [4]). As such, Sharemind programs will perform signi\ufb01cantly worse (i.e., orders of magnitude slower) on our programs that perform sequential execution, such as unoptimized linked list traversal, the shift-reduce parser, and building a sorted linked list (mergesort is also slower as reported in [23]). In the case of optimized linked list traversal, on the other hand, Sharemind implementations will still be slower for small data sets (such as 25), but signi\ufb01cantly faster for large data sets (up to two orders of magnitude faster for 220 elements). All of these experiments demonstrate that pointers have a great potential for their use in general-purpose programs evaluated over private data. Some pointer-based data structures can exhibit substantially higher performance in this framework than their public-data counterparts, and custom, internally built implementations for such data structures are recommended. 9"
+                }
+            ],
+            "Yuguang Yao": [
+                {
+                    "url": "http://arxiv.org/abs/2303.07474v2",
+                    "title": "Can Adversarial Examples Be Parsed to Reveal Victim Model Information?",
+                    "abstract": "Numerous adversarial attack methods have been developed to generate\nimperceptible image perturbations that can cause erroneous predictions of\nstate-of-the-art machine learning (ML) models, in particular, deep neural\nnetworks (DNNs). Despite intense research on adversarial attacks, little effort\nwas made to uncover 'arcana' carried in adversarial attacks. In this work, we\nask whether it is possible to infer data-agnostic victim model (VM) information\n(i.e., characteristics of the ML model or DNN used to generate adversarial\nattacks) from data-specific adversarial instances. We call this 'model parsing\nof adversarial attacks' - a task to uncover 'arcana' in terms of the concealed\nVM information in attacks. We approach model parsing via supervised learning,\nwhich correctly assigns classes of VM's model attributes (in terms of\narchitecture type, kernel size, activation function, and weight sparsity) to an\nattack instance generated from this VM. We collect a dataset of adversarial\nattacks across 7 attack types generated from 135 victim models (configured by 5\narchitecture types, 3 kernel size setups, 3 activation function types, and 3\nweight sparsity ratios). We show that a simple, supervised model parsing\nnetwork (MPN) is able to infer VM attributes from unseen adversarial attacks if\ntheir attack settings are consistent with the training setting (i.e.,\nin-distribution generalization assessment). We also provide extensive\nexperiments to justify the feasibility of VM parsing from adversarial attacks,\nand the influence of training and evaluation factors in the parsing performance\n(e.g., generalization challenge raised in out-of-distribution evaluation). We\nfurther demonstrate how the proposed MPN can be used to uncover the source VM\nattributes from transfer attacks, and shed light on a potential connection\nbetween model parsing and attack transferability.",
+                    "authors": "Yuguang Yao, Jiancheng Liu, Yifan Gong, Xiaoming Liu, Yanzhi Wang, Xue Lin, Sijia Liu",
+                    "published": "2023-03-13",
+                    "updated": "2023-03-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "main_content": "Introduction Adversarial attacks, in terms of tiny (imperceptible) input perturbations crafted to fool the decisions of machine learning (ML) models, have emerged as a primary security *Equal contributions. concern of ML in a wide range of vision applications [1\u2013 10]. Given the importance of the trustworthiness of ML, a vast amount of prior works have been devoted to answering the questions of x how to generate adversarial attacks for adversarial robustness evaluation [2, 11\u201321] and y how to defend against these attacks for robustness enhancement [11, 22\u201338]. These two questions are also tightly connected: A solution to one would help address another. In the plane of attack generation, a variety of attack methods were developed, ranging from gradient-based white-box attacks [2, 11\u201313, 18, 20] to query-based black-box attacks [14\u201317, 21]. Understanding the attack generation process allows us to further understand attacks\u2019 characteristics and their specialties. For example, different from Deepfake images synthesized by generative models [39\u201345], adversarial attacks are typically determined by (a) a simple nonparametric and deterministic perturbation optimizer (e.g., fast gradient sign method in [2]), (b) a speci\ufb01c input example (e.g., an image), and (c) a speci\ufb01ed, well-trained victim model (VM), i.e. an ML model against which attacks are generated. Here both (a) and (b) are interacted with and rely on VM for attack generation. The generated adversarial attacks in turn help the development of adversarial defenses [11, 22\u2013 33]. Examples include robust training [11, 22\u201327, 37], adversarial detection [28\u201333, 46], and adversarial puri\ufb01cation [34\u201336, 47], which exploit attack characteristics to recognize adversarial examples and produce anti-adversarial input perturbations. In addition to ordinary attack generation and adversarial defense methods, some very recent works [28, 48\u201355] started to understand and defend adversarial attacks in a new adversarial learning paradigm, termed reverse engineering of deception (RED). It aims to infer the adversary\u2019s information (e.g., the attack objective and adversarial perturbations) from attack instances. Yet, nearly all the existing RED approaches focused on either estimation/attribution of adversarial perturbations [49, 51\u201353] or recognition of attack classes/types [28, 48, 50, 54, 55]. None of the prior works investigated the feasibility of inferring VM attributes from adversarial attacks, given the fact that VM is the model foundation of attack generation. Thus, we ask (Q): arXiv:2303.07474v2  [cs.CV]  15 Mar 2023 \fModel parsing  network Neural Network       (Victim Model) Activation  function Weight  sparsity Kernel  size Architecture Adversarial image  generation Architecture Kernel size 3 x 3 VGG11 Activation function ReLU Weight sparsity 37.5% Model parsing accuracies Figure 1: Schematic overview of victim model parsing from adversarial attacks. (Left) Attack generation based on victim model (with model attributes, architecture type, kernel size, activation function, and weight sparsity). (Middle) Proposed model parsing network (MPN) that assigns VM attribute class to input adversarial data. (Right) Highlighted results: The accuracy of model parsing from PGD attacks [11] against ResNet9 on CIFAR-10. Here the performance of MPN is reported under two different input data formats, the true adversarial perturbations and the estimated adversarial perturbations (proposed in Sec. 4). (Q) Are input-agnostic VM attributes invertible from input-speci\ufb01c adversarial attacks? We call problem (Q) model parsing of adversarial attacks; see Fig. 1 for an illustration. This is also inspired by the model parsing problem de\ufb01ned for generative model (GM) [40], which attempts to infer model hyperparameters of GM from synthesized photo-realistic images. However, adversarial attacks are data-speci\ufb01c input perturbations determined by hand-crafted optimizers rather than GM. And the \u2018model attributes\u2019 to be parsed from adversarial attacks are associated with the VM (victim model), which has a weaker correlation with attacked data compared to synthesized images by GM. The latter is easier to encode data-independent GM attribute information [39, 40, 42\u201344]. The signi\ufb01cance of the proposed model parsing problem can also be demonstrated through a transfer attack example (Fig. 2). Suppose that adversarial attacks are generated from model A but used as transfer attacks against model B. If model parsing is possible, we will then be able to infer the true victim model source of these adversarial instances and shed light on the hidden model attributes. 1% cat 98% dog 1% bird Adversarial  perturbations Adversarial  example Source victim model    for attack generation Transfer attack model Model Parsing True attack source  model? True attack source  model? Adversary misclassification Model A Model B Figure 2: Motivating example on model parsing of transfer attack. A successful model parsing module can reveal the true source victim model. Contributions. We summarize our contributions below. \u2022 We propose and formalize the problem of model parsing to infer VM attributes from a variety of adversarial attacks. \u2022 We approach the model parsing problem of adversarial attacks as a supervised learning task and show that the learned model parsing network (MPN) could exhibit a surprising amount of generalization to recognize VM attributes from testing attack instances (see Fig. 1 for highlighted results). We also peer into the in\ufb02uence of designing factors (including input data format, backbone network, and evaluation metric) in MPN\u2019s generalization. \u2022 We make a comprehensive empirical study on the feasibility and effectiveness of model parsing from adversarial attacks, including the in-distribution generalization on unseen adversarial images as well as the out-of-distribution generalization on unseen attack types and model architectures. We also demonstrate how the model parsing approach can be used to uncover the true, source victim model attributes from transfer attacks (Fig. 2), and show a connection between model parsing and attack transferability. 2. Related Work Adversarial attacks and defenses. Intensive research efforts have been made for the design of adversarial attacks and defenses. Adversarial attacks in the digital domain [2, 11\u2013 15, 17, 19, 21, 56\u201359] typically deceive DNNs by integrating carefully-crafted tiny perturbations into input data. Adversarial attacks in the physical domain [60\u201365] are further developed to fool victim models under complex physical environmental conditions, which require stronger adversarial perturbations than digital attacks. In this work, we focus on the commonly-used digital attacks subject to \u2113p-norm based perturbation constraints, known as \u2113p attacks for p \u2208{2, \u221e}. Based on how an adversary interacts with the VM (victim model), \u2113p attacks also include both white-box attacks (with full access to the victim model based on which attacks are generated) and black-box attacks (with access only to the victim model\u2019s input and output). The former typically leverages the local gradient information of VM to generate attacks [2, 11, 12], while the latter takes input-output queries of VM in attack generation; Examples include score-based attacks [14, 15, 17] and decision-based attacks [21, 58, 59]. Given the vulnerability of ML models to adversarial attacks, methods to defend against these attacks are now a major focus in research [11, 22\u201337, 46, 47, 66, 67]. One line of research focuses on advancing model training methods to learn adversarially robust models [11, 22\u201327, 37, 67]. Examples include min-max optimization-based adversarial training and its many variants [11, 22, 25\u201327, 37], which can equip models with empirical robustness. To make models provably robust, certi\ufb01ed training is also developed by integrating robustness certi\ufb01cate regularization into model training [23, 68, 69] or leveraging randomized smoothing [24, 70, 71]. In addition to training robust models, another \fline of research on adversarial defense is to detect adversarial attacks by exploring and exploiting the differences between adversarial data and benign data [28\u201333, 46, 66]. Reverse engineering of deception (RED). RED has emerged as a new adversarial learning to defend against adversarial attacks and infer the adversary\u2019s knowledge (such as its identi\ufb01cations, attack objectives, and attack perturbations). For example, a few recent works [28, 48, 50, 54, 55] aim to reverse engineer the type of attack generation method and the used hyperparameters (e.g., perturbation radius and step number). In addition, other works focus on estimating or attributing the adversarial perturbations used to construct adversarial images [49, 51\u201353]. This line of research also relates to adversarial puri\ufb01cation [34\u201336, 47], the technique to defend against adversarial attacks by characterizing and removing their harmful in\ufb02uence in model predictions. However, none of the prior works investigated if VM attributes are invertible from adversarial attacks. That is, the proposed model parsing problem remains open in adversarial learning. If feasible, the insight into model parsing could offer us an in-depth understanding of the encountered threat model and inspire new designs of adversarial defenses and robust models. Our work is also inspired by the model parsing problem in GM (generative model) [40], aiming to infer GM attributes from their synthesized images. The rationale is that GM often encodes model \ufb01ngerprints in synthesized images so that these \ufb01ngerprints can be leveraged for DeepFake detection and model parsing [39, 40, 42\u201344]. Lastly, we stress that RED is different from the work [72, 73] on reverse engineering of (black-box) model hyperparameters, which estimates model attributes from the model\u2019s prediction logits. However, in this work VM is unknown for model parsing of adversarial attacks, and the only information we have is the dataset of attack instances. 3. Preliminaries and Problem Setups Preliminaries: Adversarial attacks and victim models. We \ufb01rst introduce different kinds of adversarial attacks and exhibit their dependence on VM (victim model), i.e., the ML model from which attacks are generated. Throughout the paper, we will focus on \u2113p attacks, where the adversary aims to generate imperceptible input perturbations to fool an image classi\ufb01er [2]. Let x and \u03b8 denote a benign image and the parameters of VM. The adversarial attack (a.k.a, adversarial example) is de\ufb01ned via the linear perturbation model x\u2032 = x + \u03b4, where \u03b4 = A(x, \u03b8, \u03f5) denotes adversarial perturbations, and A refers to an attack generation method relying on x, \u03b8, and the attack strength \u03f5 (i.e., the perturbation radius of \u2113p attacks). We focus on 7 attack methods given their different dependencies on the victim model (\u03b8), including input gradientbased white-box attacks with full access to \u03b8 (FGSM [2], PGD [11], CW [12], and AutoAttack or AA [13]) as well as query-based black-box attacks (ZO-signSGD [15], NES [16], and SquareAttack or Square [17]). ! FGSM (fast gradient sign method) [2]: This attack method is given by \u03b4 = x \u2212\u03f5 \u00d7 sign(\u2207x\u2113atk(x; \u03b8)), where sign(\u00b7) is the entry-wise sign operation, and \u2207x\u2113atk is the input gradient of an attack loss \u2113atk(x; \u03b8) evaluated at x under \u03b8. ! PGD (projected gradient descent) [11]: This extends FGSM via an iterative algorithm. Formally, the K-step PGD \u2113\u221e attack is given by \u03b4 = \u03b4K, where \u03b4k = P\u2225\u03b4\u2225\u221e\u2264\u03f5(\u03b4k\u22121 \u2212 \u03b1 \u00d7 sign(\u2207x\u2113atk(x; \u03b8))) for k = 1, . . . , K, P\u2225\u03b4\u2225\u221e\u2264\u03f5 is the projection operation onto the \u2113\u221e-norm constraint \u2225\u03b4\u2225\u221e\u2264\u03f5, and \u03b1 is the attack step size. By replacing the \u2113\u221enorm with the \u21132 norm, we similarly obtain the PGD \u21132 attack [11]. ! CW (Carlini-Wager) attack [12]: Similar to PGD, CW calls iterative optimization for attack generation. Yet, CW formulates attack generation as an \u2113p-norm regularized optimization problem, with the regularization parameter c = 1 by default. For example, the choice of c = 1 in CW \u21132 attack could lead to a variety of perturbation strengths with the average value around \u03f5 = 0.33 on the CIFAR-10 dataset. Moreover, CW adopts a hinge loss to ensure the misclassi\ufb01cation margin. We will focus on CW \u21132 attack. ! AutoAttack (or AA) [13]: This is an ensemble attack that uses AutoPGD, an adaptive version of PGD, as the primary means of attack. The loss of AutoPGD is given by the difference of logits ratio (DLR) rather than CE or CW loss. ! ZO-signSGD [15] and NES [16]: They are zeroth-order optimization (ZOO)-based black-box attacks. Different from the aforementioned white-box gradient-based attacks, the only interaction mode of black-box attacks with the victim model (\u03b8) is submitting inputs and receiving the corresponding predicted outputs. ZOO then uses these input-output queries to estimate input gradients and generate adversarial perturbations. Yet, ZO-signSGD and NES call different gradient estimators in ZOO [74]. ! SquareAttack (or Square) [17]: This attack is built upon random search and thus does not rely on input gradient. Table 1: Summary of adversarial attack types focused in this work. Here GD refers to gradient descent, and WB and BB refer to white-box and black-box dependence on the victim model, respectively. Attacks Generation method Loss \u2113p norm Strength \u03f5 Dependence on \u03b8 FGSM one-step GD CE \u2113\u221e {4, 8, 12, 16}/255 WB, gradient-based PGD multi-step GD CE \u2113\u221e {4, 8, 12, 16}/255 WB, gradient-based \u21132 0.25, 0.5, 0.75, 1 CW multi-step GD CW \u21132 soft regularization c \u2208{0.1, 1, 10} WB, gradient-based AutoAttack attack ensemble CE / \u2113\u221e {4, 8, 12, 16}/255 WB, gradient-based + or AA DLR \u21132 0.25, 0.5, 0.75, 1 BB, query-based SquareAttack random search CE \u2113\u221e {4, 8, 12, 16}/255 BB, query-based or Square \u21132 0.25, 0.5, 0.75, 1 NES ZOO CE \u2113\u221e {4, 8, 12, 16}/255 BB, query-based ZO-signSGD ZOO CE \u2113\u221e {4, 8, 12, 16}/255 BB, query-based Although there exist other types of \u2113p attacks, our focused attack methods have been diverse, covering different opti\fmization methods, attack losses, \u2113p norms, and dependencies on the VM \u03b8, as summarized in Table 1. Model parsing of adversarial attacks. It is clear that adversarial attacks contain the information of VM (\u03b8), although the degree of their dependence varies. Inspired by the above, one may wonder if the attributes of \u03b8 can be inferred from these attack instances, i.e., adversarial perturbations, or perturbed images. The model attributes of our interest include model architecture types as well as \ufb01ner-level knowledge, e.g., activation function type. We call the resulting problem model parsing of adversarial attacks, as described below. (Problem statement) Is it possible to infer victim model information from adversarial attacks? And what factors will in\ufb02uence such model parsing ability? To our best knowledge, the feasibility of model parsing from adversarial attack instances is an open question. Its challenges stay in two dimensions. First, through the model lens, VM (victim model) is indirectly coupled with adversarial attacks, e.g., via local gradient information or model queries. Thus, it remains elusive what VM information is \ufb01ngerprinted in adversarial attacks and impacts the feasibility of model parsing. Second, through the attack lens, the diversity of adversarial attacks (Table 1) makes a once-for-all model parsing solution extremely dif\ufb01cult. Spurred by the above, we will take the \ufb01rst solid step to investigate the feasibility of model parsing from adversarial attacks and study what factors may in\ufb02uence the model parsing performance. The insight into model parsing could offer us an in-depth understanding of the encountered threat model and inspire new designs of adversarial defenses and robust models. Table 2: Summary of model attributes of interest. Each attribute value corresponds to an attribute class in model parsing. Model attributes Code Classes per attribute Architecture type AT ResNet9, ResNet18 ResNet20, VGG11, VGG13 Kernel size KS 3, 5, 7 Activation function AF ReLU, tanh, ELU Weight sparsity WS 0%, 37.5%, 62.5% Model attributes and setup. We specify VMs of adversarial attacks as convolutional neural network (CNN)-based image classi\ufb01ers. More concretely, we consider 5 CNN architecture types (ATs): ResNet9, ResNet18, ResNet20, VGG11, and VGG13. Given an AT, CNN models are then con\ufb01gured by different choices of kernel size (KS), activation function (AF), and weight sparsity (WS). Thus, a valued quadruple (AT, KS, AF, WS) yields a speci\ufb01c VM (\u03b8). Although more attributes could be considered, the rationale behind our choices is given below. We focus on KS and AF since they are the two fundamental building components of CNNs. Besides, we choose WS as another model attribute since it relates to sparse models achieved by pruning (i.e., removing redundant model weights) [75, 76]. We defer sanity checks of all-dimension model attributes for future studies. Table 2 summarizes the focused model attributes and their values to specify VM instances. Given a VM speci\ufb01cation, we generate adversarial attacks following attack methods in Table 1. Unless speci\ufb01ed otherwise, our empirical studies will be mainly given on CIFAR-10 but experiments on other datasets will also be provided in Sec. 5. 4. Methods In this section, we approach the model parsing problem as a standard supervised learning task applied over the dataset of adversarial attacks. We will show that the learned model parsing network could exhibit a surprising amount of generalization on test-time adversarial data. We will also show data-model factors that may in\ufb02uence such generalization. Model parsing network and training. We approach the model parsing problem as a supervised attribute recognition task. That is, we develop a parametric model, termed model parsing network (MPN), which takes adversarial attacks as input and predicts the model attribute values (i.e., \u2018classes\u2019 in Table 2). Despite the simplicity of supervised learning, the construction of MPN is non-trivial when designing data format, backbone network, and evaluation metrics. We \ufb01rst create a model parsing dataset by collecting adversarial attack instances against victim models. Since adversarial attacks are proposed for evading model predictions in the post-training stage, we choose the test set of an ordinary image dataset (e.g., CIFAR-10) to generate adversarial data, where an 80/20 training/test split is used for MPN training and evaluation. Following notations in Sec. 3, the training set of MPN is denoted by Dtr = {(z(A, x, \u03b8), y(\u03b8)) | x \u2208Itr, \u03b8 \u2208\u0398}, where z signi\ufb01es an attack data feature (e.g., adversarial perturbations \u03b4 or adversarial example x\u2032) that relies on the attack method A, the original image sample x, and the VM \u03b8, and y(\u03b8) denotes the true model attribute label associated with \u03b8. To differentiate with the testing data of MPN, we denote by Itr the set of original images used for training MPN. We also denote by \u0398 the set of model architectures used for generating attack data in Dtr. For ease of presentation, we express the training set of MPN as Dtr = {(z, y)} to omit the dependence on other factors. Next, we elaborate on the construction of MPN (parameterized by \u03c6). We intend to make the architecture of MPN as simple as possible and make it different from the VM \u03b8. The rationale behind that has two folds. First, we would like to examine the feasibility of model parsing from adversarial attacks even forcing the simplicity of attribution network (\u03c6). Second, we would like to avoid the \u2018model attribute bias\u2019 of \u03c6 when inferring VM attributes from adversarial \fattacks. Inspired by the above, we specify MPN by two simple networks: (1) multilayer perceptron (MLP) containing two hidden layers with 128 hidden units (0.41M parameters) [77], and (2) a simple 4-layer CNN (ConvNet-4) with 64 output channels for each layer, followed by one fullyconnected layers with 128 hidden units and the attribution prediction head (0.15M parameters) [78]. As will be evident later (Fig. 4), the model parsing accuracy of ConvNet-4 typically outperforms that of MLP. Thus, ConvNet-4 will be used as the MPN model by default. Figure 3: The schematic overview of our proposal. Here PEN (perturbation estimation network) will be introduced later as a pre-processing step to transform adversarial examples into perturbation-alike input data. Given the datamodel setup, we next tackle the recognition problem of VM\u2019s attributes (AT, KS, AF, WS) via a multi-head multi-class classi\ufb01er. We dissect MPN into two parts \u03c6 = [\u03c6rep, \u03c6atr], where \u03c6rep is for data representation acquisition, and \u03c6attr corresponds to the attribute-speci\ufb01c prediction head (i.e., the last fully-connected layer in our design). Eventually, four prediction heads {\u03c6(i) atr }4 i=1 will share \u03c6rep for model attribute recognition; see Fig. 3 for a schematic overview of our proposal. The MPN training problem is then cast as minimize \u03c6rep,{\u03c6(i) atr }4 i=1 E(z,y)\u2208Dtr 4 X i=1 [\u2113CE(h(z; \u03c6rep, \u03c6(i) atr ), yi)], (1) where h(z; \u03c6rep, \u03c6(i) atr ) denotes the MPN prediction at input example z using the predictive model consisting of \u03c6rep and \u03c6(i) atr for the ith attribute classi\ufb01cation, yi is the ground-truth label of the ith attribute associated with the input data z, and \u2113CE is the cross-entropy (CE) loss characterizing the error between the prediction and the true label. Evaluation methods. Similar to training, we denote by Dtest = {(z(A, x, \u03b8), y(\u03b8)) | x \u2208Itest, \u03b8 \u2208\u0398} the test attack set for evaluating the performance of MPN. Here the set of benign images Itest is different from Itr, thus adversarial attacks in Dtest are new to Dtr. To mimic the standard evaluation pipeline of supervised learning, we propose the following evaluation metrics. (1) In-distribution generalization: The MPN testing dataset Dtest follows the attack methods (A) and the VM speci\ufb01cations (\u0398) same as Dtr but corresponding to different original benign images (i.e., Itest \u0338= Itr). The purpose of such an in-distribution evaluation is to examine if the trained MPN can infer model attributes encoded in new attack data given already-seen attack methods. (2) Out-of-distribution (OOD) generalization: In addition to new test-time images, there could exist attack/model distribution shifts in Dtest due to using new attack methods or model architectures, leading to unseen attack methods (A) and victim models (\u0398) different from the settings in Dtr. In the rest of the paper, both in-distribution and OOD generalization capabilities will be assessed. Unless speci\ufb01ed otherwise, the generalization of MPN stands for the in-distribution generalization. 20 40 60 80 100 FGSM PGD  PGD 2 CW AA  AA 2 Square  Square 2 NES ZO-signSGD FGSM PGD  PGD 2 CW AA  AA 2 Square  Square 2 NES ZO-signSGD MLP,  ConvNet-4,  MLP, x\u2032 ConvNet-4, x\u2032 Figure 4: The in-distribution generalization of MPN using different formats of input data (adversarial perturbations \u03b4 vs. adversarial examples x\u2032) and parsing networks (ConvNet-4 vs. MLP). The generalization performance is measured by the averaged testing accuracy of attribute-speci\ufb01c classi\ufb01ers; see Sec. 5.1. The attack data are generated from 10 attack methods given in Table 1, with \u2113\u221eattack strength \u03f5 = 8/255 and \u21132 attack strength \u03f5 = 0.5 on CIFAR-10. The VM architecture is \ufb01xed to ResNet-9, and VM instances are generated by varying other model attributes in Table 2. Perturbations or adversarial examples? The input data format matters for MPN. Recall from Sec. 3 that an adversarial example, given by the linear model x\u2032 = x + \u03b4, relates to \u03b8 through \u03b4. Thus, it could be better for MPN to adopt adversarial perturbations (\u03b4) as the attack data feature (z), rather than the indirect adversarial example x\u2032. Fig. 4 empirically justi\ufb01es our hypothesis by comparing the generalization of MPN trained on adversarial perturbations with that on adversarial examples under two model speci\ufb01cations of MPN, MLP and ConvNet-4. We present the performance of MPN trained and tested on different attack types. As we can see, the use of adversarial perturbations (\u03b4) consistently improves the test-time classi\ufb01cation accuracy of VM attributes, compared to the use of adversarial examples (x\u2032). In addition, ConvNet-4 outperforms MLP with a substantial margin. \fAlthough Fig. 4 shows the promise of the generalization ability of MPN when trained and tested on adversarial perturbations, it may raise another practical question of how to obtain adversarial perturbations from adversarial examples if the latter is the only attack source accessible to MPN. To overcome this dif\ufb01culty, we propose a perturbation estimator network (PEN) that can be jointly learned with MPN. Once PEN is prepended to the MPN model, the resulting end-toend pipeline can achieve model parsing using adversarial examples as inputs (see Fig. 3). We use a denoising network, DnCNN [79], to model PEN with parameters \u03c8. PEN obtains perturbation estimates by minimizing the denoising objective function using the true adversarial perturbations as supervision. Extended from (1), we then have minimize \u03c8,\u03c6rep,{\u03c6(i) atr }4 i=1 \u03b2E(x,x\u2032)\u2208Dtr[\u2113MAE(g\u03c8(x\u2032), x\u2032 \u2212x)] +E(x\u2032,y)\u2208Dtr P4 i=1[\u2113CE(h(g\u03c8(x\u2032); \u03c6rep, \u03c6(i) atr ), yi)], (2) where g\u03c8(x\u2032) is the output of PEN given x\u2032 as the input, \u2113MAE is the mean-absolute-error (MAE) loss characterizing the perturbation estimation error, and \u03b2 > 0 is a regularization parameter. Compared with (1), MPN takes the perturbation estimate g\u03c8(x\u2032) for VM attribute classi\ufb01cation. 5. Experiments 5.1. Experiment setup and implementation Dataset curation. We use standard image classi\ufb01cation datasets (CIFAR-10, CIFAR-100, and Tiny-ImageNet) to acquire victim models, from which attacks are generated. We refer readers to Appendix A for details on victim model training and evaluation, as well as different attack setups. These VM instances are then leveraged to create MPN datasets (as described in Sec. 4). The attack types and victim model con\ufb01gurations have been summarized in Table 1 and 2. Thus, we collect a dataset consisting of adversarial attacks across 7 attack types generated from 135 VMs (con\ufb01gured by 5 architecture types, 3 kernel size setups, 3 activation function types, and 3 weight sparsity levels). MPN training and evaluation. To solve problem (1), we train the MPN model using the SGD (stochastic gradient descent) optimizer with cosine annealing learning rate schedule and an initial learning rate of 0.1. The training epoch number and the batch sizes are given by 100 and 256, respectively. To solve problem (2), we \ufb01rst train MPN according to (1), and then \ufb01ne-tune a pre-trained DnCNN model [49] (taking only the denoising objective into consideration) for 20 epochs. Starting from these initial models, we jointly optimize MPN and PEN by minimizing problem (2) with \u03b2 = 1 over 50 epochs. To evaluate the effectiveness of MPN, we consider both in-distribution and OOD generalization assessment. The generalization performance is measured by testing accuracy averaged over attribute-wise predictions, namely, P i(NiTA(i))/ P i Ni, where Ni is the number of classes of the model attribute i, and TA(i) is the testing accuracy of the classi\ufb01er associated with the attribute i (Fig. 3). 5.2. Results and insights In-distribution generalization of MPN is achievable. Table 3 presents the in-distribution generalization performance of MPN trained using different input data formats (i.e., adversarial examples x\u2032, PEN-estimated adversarial perturbations \u03b4PEN, and true adversarial perturbations \u03b4) given each attack type in Table 1. Here the choice of AT (architecture type) is \ufb01xed to ResNet9, but adversarial attacks on CIFAR-10 are generated from VMs con\ufb01gured by different values of KS, AF, and WS (see Table 2). As we can see, the generalization of MPN varies against the attack type even if model parsing is conducted from the ideal adversarial perturbations (\u03b4). We also note that model parsing from whitebox adversarial attacks (i.e., FGSM, PGD, and AA) is easier than that from black-box attacks (i.e., ZO-signSGD, NES, and Square). For example, the worst-case performance of MPN is achieved when training/testing on Square attacks. This is not surprising, since Square is based on random search and has the least dependence on VM attributes. In addition, we \ufb01nd that MPN using estimated perturbations (\u03b4PEN) substantially outperforms the one trained on adversarial examples (x\u2032). This justi\ufb01es the effectiveness of our proposed PEN solution for MPN. Table 3: The in-distribution testing accuracy (%) of MPN trained using different input data formats (adversarial examples x\u2032, PEN-estimated adversarial perturbations \u03b4PEN, and true adversarial perturbations \u03b4) across different attack types on CIFAR-10, with \u2113\u221eattack strength \u03f5 = 8/255, \u21132 attack strength \u03f5 = 0.5, and CW attack strength c = 1. Input data Attack type FGSM PGD \u2113\u221e PGD \u21132 CW AA \u2113\u221e AA \u21132 Square \u2113\u221e Square \u21132 NES ZOsignSGD x\u2032 78.80 66.62 53.42 35.42 74.78 56.26 38.92 36.21 40.80 42.48 \u03b4PEN 94.15 83.20 82.58 64.46 91.09 86.89 44.14 42.30 58.85 61.20 \u03b4 96.89 95.07 99.64 96.66 97.48 99.95 44.37 44.05 83.33 84.87 Extended from Table 3, Fig. 5 shows the generalization performance of MPN when evaluated using attack data with different attack strengths. We observe that in-distribution generalization (corresponding to the same attack strength for the train-time and test-time attacks) is easier to achieve than OOD generalization (different attack strengths at test time and train time). Another observation is that a smaller gap between the train-time attack strength and the test-time strength leads to better generalization performance. Table 3 and Fig. 5 focused on model parsing of adversarial attacks by \ufb01xing the VM architecture to ResNet9 on CIFAR-10, although different model attribute combinations lead to various ResNet9-type VM instantiations for attack generation. Furthermore, Table 4 shows the in\fTable 4: In-distribution generalization performance (testing accuracy, %) of MPN given different choices of VMs and datasets, attack types/strengths, and MPN input data formats (x\u2032, \u03b4PEN, and \u03b4). Attack type Attack strength Dataset and victim model CIFAR-10 ResNet9 CIFAR-10 ResNet18 CIFAR-10 ResNet20 CIFAR-10 VGG11 CIFAR-10 VGG13 CIFAR-100 ResNet9 Tiny-ImageNet ResNet18 x\u2032 \u03b4PEN \u03b4 x\u2032 \u03b4PEN \u03b4 x\u2032 \u03b4PEN \u03b4 x\u2032 \u03b4PEN \u03b4 x\u2032 \u03b4PEN \u03b4 x\u2032 \u03b4PEN \u03b4 x\u2032 \u03b4PEN \u03b4 FGSM \u03f5 = 4/255 60.13 85.25 96.82 60.00 86.92 97.66 62.41 88.91 97.64 47.42 73.40 91.75 66.28 90.02 98.57 57.99 82.22 94.86 37.23 84.27 97.04 \u03f5 = 8/255 78.80 94.15 96.89 80.44 95.49 97.61 82.29 95.90 97.72 63.13 86.76 92.41 84.92 96.91 98.66 75.58 91.65 94.96 70.29 91.17 97.05 \u03f5 = 12/255 86.49 95.96 96.94 88.03 96.89 97.68 88.71 97.13 97.81 73.71 90.19 92.66 91.21 98.10 98.71 82.27 94.01 95.55 76.00 93.45 97.02 \u03f5 = 16/255 90.16 96.43 96.94 91.71 97.34 97.68 91.84 97.47 97.79 79.51 91.28 92.60 94.22 98.44 98.73 86.50 94.04 94.74 79.63 94.35 96.87 PGD \u2113\u221e \u03f5 = 4/255 50.54 76.43 96.02 56.94 79.45 96.96 55.01 80.05 97.49 39.33 66.38 91.84 57.12 81.18 98.29 42.27 72.62 92.65 35.48 76.56 97.18 \u03f5 = 8/255 66.62 83.20 95.07 73.29 87.29 95.38 67.49 86.19 96.18 56.62 81.14 92.78 69.16 88.46 97.22 59.71 79.55 90.43 61.85 82.90 96.05 \u03f5 = 12/255 76.65 89.73 94.91 81.73 91.67 95.55 76.41 90.16 95.67 70.56 88.92 94.13 78.67 92.93 97.26 70.86 85.31 91.28 73.82 88.80 96.38 \u03f5 = 16/255 75.58 86.95 91.28 82.46 90.19 93.19 76.58 87.79 92.50 72.13 87.23 91.85 78.28 90.20 94.66 71.29 82.35 86.84 73.19 85.02 93.54 PGD \u21132 \u03f5 = 0.25 36.75 62.20 99.66 46.35 70.17 99.74 48.24 77.22 99.75 36.47 45.17 98.52 35.81 70.62 99.85 35.92 61.91 99.29 35.55 35.68 99.68 \u03f5 = 0.5 53.42 82.58 99.64 60.89 84.70 99.56 61.62 89.11 99.61 41.56 66.58 98.68 57.83 87.64 99.83 48.89 79.26 99.01 35.52 54.56 99.71 \u03f5 = 0.75 62.66 89.04 99.48 71.01 89.89 99.22 70.76 92.06 99.36 47.02 78.12 98.52 72.76 92.32 99.74 59.19 85.14 98.61 35.56 81.33 99.71 \u03f5 = 1 71.65 91.73 99.26 77.09 92.09 98.94 76.84 92.82 98.96 54.20 84.30 98.41 79.93 93.96 99.57 66.97 87.63 97.89 43.48 88.81 99.64 CW c = 0.1 33.77 55.60 96.71 47.77 63.26 96.11 33.56 63.11 94.10 33.73 48.90 94.37 33.68 65.48 96.95 34.41 46.47 92.55 35.96 35.77 95.52 c = 1 35.42 64.46 96.66 45.75 65.25 97.45 33.74 62.71 97.08 33.89 55.61 91.29 36.12 68.66 98.58 34.25 55.18 93.25 35.54 35.29 89.35 c = 10 36.38 64.45 96.64 45.83 65.32 97.41 33.83 63.52 97.11 38.29 56.83 91.33 38.51 68.28 98.62 34.25 55.89 93.18 35.45 53.18 94.20 4/255 8/255 12/255 16/255 Training attack strength ( ) 0 20 40 60 80 100 Testing accuracy (%) 96.0 71.9 56.1 39.9 47.2 95.1 66.2 66.1 39.7 61.1 94.9 74.5 35.3 53.2 53.6 91.3 Testing attack strength ( ) 4/255 8/255 12/255 16/255 Figure 5: Testing accuracies (%) of MPN when trained on adversarial perturbations generated by PGD \u2113\u221eusing different attack strengths (\u03f5) and evaluated using different attack strengths as well. Other setups are consistent with in Table 3. distribution generalization of MPN under diverse setups of victim model architectures and datasets. The insights into model parsing are consistent with Table 3: (1) The use of true adversarial perturbations (\u03b4) and PEN-estimated perturbations (\u03b4PEN) can yield higher model parsing accuracy. And (2) inferring model attributes from white-box, gradient-based adversarial perturbations is easier, as supported by its over 90% testing accuracy. We also observe that if adversarial examples (x\u2032) or estimated adversarial perturbations (\u03b4PEN) are used for model parsing, then the resulting accuracy gets better if adversarial attacks have a higher attack strength. OOD generalization of MPN becomes dif\ufb01cult vs. unseen, dissimilar attack types at testing time. In Fig. 6, we present the generalization matrix of MPN when trained under one attack type (e.g., PGD \u2113\u221eattack at row 1) but tested under another attack type (e.g., FGSM attack at column 2) when adversarial perturbations are generated from the same set of ResNet9-based VMs (with different con\ufb01gurations of other model attributes) on CIFAR-10. The diagonal entries of the matrix correspond to the in-distribution generalization of MPN given the attack type, while the offdiagonal entries characterize OOD generalization when the test-time attack type is different from the train-time one. First, we \ufb01nd that MPN generalizes better across attack PGD  FGSM AA  CW PGD 2 AA 2 NES ZO-signSGD Square 2 Square  Testing attack types PGD  FGSM AA  CW PGD 2 AA 2 NES ZO-signSGD Square 2 Square  Combined Training attack types 95.0 64.9 94.3 50.3 52.8 53.7 61.6 64.9 34.5 33.6 85.4 96.8 88.3 42.5 42.3 41.7 42.6 44.3 33.3 32.9 89.5 77.0 97.4 39.0 38.2 38.2 53.8 57.9 31.8 34.2 82.6 88.5 90.9 96.6 97.3 97.7 62.2 64.6 34.0 32.8 51.0 48.5 51.2 64.7 99.6 99.9 41.5 42.9 37.2 32.6 33.7 34.2 35.2 46.0 91.8 99.9 35.0 34.9 35.4 32.4 53.6 52.9 50.3 54.3 66.5 65.7 83.3 83.9 34.8 33.5 64.1 57.8 62.0 54.4 67.5 67.7 83.1 84.8 35.8 33.7 43.4 43.0 49.1 53.0 49.0 49.4 39.2 40.5 44.0 35.6 40.1 40.0 38.8 34.5 35.0 35.0 42.2 43.1 33.7 44.3 96.4 94.9 98.3 97.6 99.7 99.9 88.2 89.9 36.4 33.4 40 50 60 70 80 90 100 Figure 6: Generalization performance (%) matrix of MPN when trained on a row-speci\ufb01c attack type but evaluated on a column-speci\ufb01c attack type. The attack data are given by adversarial perturbations with strength \u03f5 = 8/255 for \u2113\u221eattacks, \u03f5 = 0.5 for \u21132 attacks, and c = 1 for CW attack. The VM architecture and the dataset are set to ResNet9 and CIFAR-10. The \u2018combined\u2019 row represents MPN training on the collection of four attack types: PGD \u2113\u221e, PGD \u21132, CW, and ZO-signSGD. types when they share similarities, leading to the following generalization communities: \u2113\u221eattacks (PGD \u2113\u221e, FGSM, and AA \u2113\u221e), \u21132 attacks (CW, PGD \u21132, or AA \u21132), and ZOObased black-box attacks (NES and ZO-signSGD). Second, Square attacks are dif\ufb01cult to learn and generalize, as evidenced by the low test accuracies in the last two rows and the last two columns. This is also consistent with Table 3. Third, given the existence of generalization communities, we then combine diverse attack types (including PGD \u2113\u221e, PGD \u21132, CW, and ZO-signSGD) into an augmented MPN training set and investigate if such a data augmentation can boost the OOD generalization of MPN. The results are summarized in the \u2018combined\u2019 row of Fig. 6. As we expect, the use of combined attack types indeed makes MPN generalize better across all attack types except for the random search-based \fSquare attack. In Fig. A1 of Appendix B, we also \ufb01nd the consistent OOD generalization performance of MPN when PEN-based adversarial perturbations are used in MPN. Fig. 7 demonstrates the generalization matrix of MPN when trained and evaluated using adversarial perturbations generated from different VM architectures (i.e., different values of AT in Table 1) by \ufb01xing the con\ufb01gurations of other attributes (KS, AF, and WS). We observe that given an attack type, the in-distribution MPN generalization remains well across VM architectures. Yet, the OOD generalization of MPN (corresponding to the off-diagonal entries of the generalization matrix) rapidly degrades if the test-time VM architecture is different from the train-time one. Thus, if MPN is trained on data with AT classes as supervision, then the in-distribution generalization on AT retains. We refer readers to Appendix C for more results for AT prediction. VGG11 VGG13 ResNet9 ResNet18 ResNet20 Testing victim architecture VGG11 VGG13 ResNet9 ResNet18 ResNet20 Combined Training victim architecture 92.4 45.6 40.2 42.3 38.1 44.2 98.7 47.7 40.9 46.4 44.0 46.7 96.9 46.8 35.6 37.9 49.3 55.1 97.6 53.5 34.5 53.6 46.9 53.6 97.7 89.0 96.2 93.8 95.0 95.2 VGG11 VGG13 ResNet9 ResNet18 ResNet20 Testing victim architecture VGG11 VGG13 ResNet9 ResNet18 ResNet20 Combined Training victim architecture 92.8 48.1 42.7 50.3 42.8 46.9 97.2 52.4 49.1 51.3 38.1 49.6 95.1 49.5 36.9 38.0 54.1 55.8 95.4 58.9 40.5 59.4 51.4 62.5 96.2 87.4 93.5 90.0 91.7 92.1 VGG11 VGG13 ResNet9 ResNet18 ResNet20 Testing victim architecture VGG11 VGG13 ResNet9 ResNet18 ResNet20 Combined Training victim architecture 91.3 45.2 39.2 41.1 37.6 42.7 98.6 47.8 42.3 47.3 39.2 47.9 96.7 47.2 36.1 34.6 50.2 53.6 97.5 54.6 34.7 52.7 46.3 55.8 97.1 87.1 95.5 93.1 94.1 93.9 40 50 60 70 80 90 100 (a) FGSM (b) PGD \u2113\u221e (c) CW Figure 7: Generalization matrix (%) of MPN when trained on attack data generated from a row-speci\ufb01c architecture but evaluated on attack data generated from a column-speci\ufb01c architecture. Both the train-time and test-time architectures share the same VM attributes in KS, AF, and WS. The attack type is speci\ufb01ed by FGSM, PGD \u2113\u221e, or CW on CIFAR-10, with the attack strength \u03f5 = 8/255 for \u2113\u221eattacks and c = 1 for CW. MPN to uncover real VM attributes of transfer attacks. As a use case of model parsing, we next investigate if MPN can correctly infer the source VM attributes from transfer attacks when applied to attacking a different model (see Fig. 2). Given the VM architecture ResNet9, we vary the values of model attributes KS, AF, and WS to produce 27 ResNet9-type VMs. Fig. 8 shows the transfer attack success rate (ASR) matrix (Fig. 8a) and the model parsing confusion matrix (Fig. 8b). Here the attack type is given by PGD \u2113\u221e attack with strength \u03f5 = 8/255 on CIFAR-10, and the resulting adversarial perturbations (generated from different VMs) are used for MPN training and evaluation. In Fig. 8a, the off-diagonal entries represent ASRs of transfer attacks from row-wise VMs to attacking columnwise target models. As we can see, adversarial attacks generated from the ReLU-based VMs are typically more dif\ufb01cult to transfer to smooth activation function (ELU or tanh)-based target models. By contrast, given the values of KS and AF, attacks are easier to transfer across models with different weight sparsity levels. Fig. 8b presents the confusion matrix of MPN trained on attack data generated from all 27 ResNet9-alike VMs. Each row of the confusion matrix represents the true VM used to generate the attack dataset, and each column corresponds to a predicted model attribute con\ufb01guration. Thus, the diagonal entries and the off-diagonal entries in Fig. 8b represent the correct model parsing accuracy and the misclassi\ufb01cation rate on the incorrectly predicted model attribute con\ufb01guration. As we can see, attacks generated from ReLU-based VMs result in a low misclassi\ufb01cation rate of MPN on ELU or tanhbased predictions (see the marked region x). Meanwhile, a high misclassi\ufb01cation occurs for MPN when evaluated on attack data corresponding to different values of WS (see the marked region y). The above results, together with our insights into ASRs of transfer attacks in Fig. 8a, suggest a connection between transfer attack and model parsing: If attacks are dif\ufb01cult (or easy) to transfer from the source model to the target model, then inferring the source model attributes from these attacks turns to be easy (or dif\ufb01cult). (ReLU,3,0%) (ReLU,3,37.5%) (ReLU,3,62.5%) (ReLU,5,0%) (ReLU,5,37.5%) (ReLU,5,62.5%) (ReLU,7,0%) (ReLU,7,37.5%) (ReLU,7,62.5%) (tanh,3,0%) (tanh,3,37.5%) (tanh,3,62.5%) (tanh,5,0%) (tanh,5,37.5%) (tanh,5,62.5%) (tanh,7,0%) (tanh,7,37.5%) (tanh,7,62.5%) (ELU,3,0%) (ELU,3,37.5%) (ELU,3,62.5%) (ELU,5,0%) (ELU,5,37.5%) (ELU,5,62.5%) (ELU,7,0%) (ELU,7,37.5%) (ELU,7,62.5%) Transfer attack model (ReLU,3,0%) (ReLU,3,37.5%) (ReLU,3,62.5%) (ReLU,5,0%) (ReLU,5,37.5%) (ReLU,5,62.5%) (ReLU,7,0%) (ReLU,7,37.5%) (ReLU,7,62.5%) (tanh,3,0%) (tanh,3,37.5%) (tanh,3,62.5%) (tanh,5,0%) (tanh,5,37.5%) (tanh,5,62.5%) (tanh,7,0%) (tanh,7,37.5%) (tanh,7,62.5%) (ELU,3,0%) (ELU,3,37.5%) (ELU,3,62.5%) (ELU,5,0%) (ELU,5,37.5%) (ELU,5,62.5%) (ELU,7,0%) (ELU,7,37.5%) (ELU,7,62.5%) Source victim model 0 20 40 60 80 100 (ReLU,3,0%) (ReLU,3,37.5%) (ReLU,3,62.5%) (ReLU,5,0%) (ReLU,5,37.5%) (ReLU,5,62.5%) (ReLU,7,0%) (ReLU,7,37.5%) (ReLU,7,62.5%) (tanh,3,0%) (tanh,3,37.5%) (tanh,3,62.5%) (tanh,5,0%) (tanh,5,37.5%) (tanh,5,62.5%) (tanh,7,0%) (tanh,7,37.5%) (tanh,7,62.5%) (ELU,3,0%) (ELU,3,37.5%) (ELU,3,62.5%) (ELU,5,0%) (ELU,5,37.5%) (ELU,5,62.5%) (ELU,7,0%) (ELU,7,37.5%) (ELU,7,62.5%) Predicted model (ReLU,3,0%) (ReLU,3,37.5%) (ReLU,3,62.5%) (ReLU,5,0%) (ReLU,5,37.5%) (ReLU,5,62.5%) (ReLU,7,0%) (ReLU,7,37.5%) (ReLU,7,62.5%) (tanh,3,0%) (tanh,3,37.5%) (tanh,3,62.5%) (tanh,5,0%) (tanh,5,37.5%) (tanh,5,62.5%) (tanh,7,0%) (tanh,7,37.5%) (tanh,7,62.5%) (ELU,3,0%) (ELU,3,37.5%) (ELU,3,62.5%) (ELU,5,0%) (ELU,5,37.5%) (ELU,5,62.5%) (ELU,7,0%) (ELU,7,37.5%) (ELU,7,62.5%) True source victim model 1 2 2 2 0 20 40 60 80 100 (a) Attack successful rate (%) (b) Confusion matrix (%) Figure 8: Model parsing of transfer attacks: Transfer attack success rate matrix (a) and model parsing confusion matrix (b). Given the architecture type ResNet9, the dataset CIFAR-10, and the attack type PGD \u2113\u221e(with strength \u03f5 = 8/255), each model attribute combination (AF, KS, WS) de\ufb01nes a model instance to be attacked, transferred, or parsed. Model parsing vs. model robustness. Further, we examine the in\ufb02uence of adversarial robustness of victim models in the generalization performance of MPN; see Fig. A2 in Appendix D. We \ufb01nd that adversarial attacks against robust VMs is harder to parse VM attributes than attacks against standard VMs. 6."
+                }
+            ],
+            "Yimeng Zhang": [
+                {
+                    "url": "http://arxiv.org/abs/2310.11868v2",
+                    "title": "To Generate or Not? Safety-Driven Unlearned Diffusion Models Are Still Easy To Generate Unsafe Images ... For Now",
+                    "abstract": "The recent advances in diffusion models (DMs) have revolutionized the\ngeneration of realistic and complex images. However, these models also\nintroduce potential safety hazards, such as producing harmful content and\ninfringing data copyrights. Despite the development of safety-driven unlearning\ntechniques to counteract these challenges, doubts about their efficacy persist.\nTo tackle this issue, we introduce an evaluation framework that leverages\nadversarial prompts to discern the trustworthiness of these safety-driven DMs\nafter they have undergone the process of unlearning harmful concepts.\nSpecifically, we investigated the adversarial robustness of DMs, assessed by\nadversarial prompts, when eliminating unwanted concepts, styles, and objects.\nWe develop an effective and efficient adversarial prompt generation approach\nfor DMs, termed UnlearnDiffAtk. This method capitalizes on the intrinsic\nclassification abilities of DMs to simplify the creation of adversarial\nprompts, thereby eliminating the need for auxiliary classification or diffusion\nmodels.Through extensive benchmarking, we evaluate the robustness of five\nwidely-used safety-driven unlearned DMs (i.e., DMs after unlearning undesirable\nconcepts, styles, or objects) across a variety of tasks. Our results\ndemonstrate the effectiveness and efficiency merits of UnlearnDiffAtk over the\nstate-of-the-art adversarial prompt generation method and reveal the lack of\nrobustness of current safety-driven unlearning techniques when applied to DMs.\nCodes are available at https://github.com/OPTML-Group/Diffusion-MU-Attack.\nWARNING: This paper contains model outputs that may be offensive in nature.",
+                    "authors": "Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, Sijia Liu",
+                    "published": "2023-10-18",
+                    "updated": "2024-03-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "main_content": "Introduction The realm of text-to-image generation has seen significant progress in recent years, primarily driven by the development and adoption of diffusion models (DMs) trained on extensive and diverse datasets (Ho et al., 2020; Song & Ermon, 2019; Song et al., 2020; Dhariwal & Nichol, 2021; Nichol & Dhariwal, 2021; Watson et al., 2021; Rombach et al., 2022; Croitoru et al., 2023). Yet, this swift advancement carries a risk: DMs are prone to creating NSFW (Not Safe For Work) imagery when prompted with inappropriate texts, as evidenced by studies (Rando et al., 2022; Schramowski et al., 2023). To alleviate this concern, recent DM technologies (Nichol et al., 2021; Schramowski et al., 2023) have incorporated preor post-generation NSFW safety checkers to minimize the harmful effects of inappropriate prompts in DMs. However, depending on external safety measures and filters falls short of offering a genuine solution to DMs\u2019 safety issues, as these approaches are model-independent and rely solely on post-hoc interventions. Indeed, existing research (Gandikota et al., 2023a; Brack et al., 2023; Yang et al., 2023; Zhang et al., 2023a) has demonstrated their inadequacy in effectively preventing DMs from generating unsafe content. In response to the safety concerns of DMs, a range of studies (Gandikota et al., 2023a; Zhang et al., 2023a; Kumari et al., 2023a; Gandikota et al., 2023b) have sought to improve the DM training or finetuning procedure to eliminate the negative impact of inappropriate prompts on image generation and create a safer DM. These approaches also align with the broader concept of machine unlearning (MU) (Nguyen et al., 2022; Shaik et al., 2023; Cao & Yang, 2015; arXiv:2310.11868v2  [cs.CV]  24 Mar 2024 \fUnlearnDiffAtk Thudi et al., 2022a; Xu et al., 2023; Jia et al., 2023) in the machine learning field. MU aims to erase the influence of specific data points or classes to enhance the privacy and security of an ML model without requiring the model to be retrained from scratch after removing the unlearning data. Given this association, we refer to the safety-driven DMs (Gandikota et al., 2023a; Zhang et al., 2023a; Kumari et al., 2023a; Gandikota et al., 2023b) designed to prevent harmful image generation as unlearned DMs. These models seek to erase the impact of unwanted concepts, styles, or objects in image generation, regardless of being conditioned on inappropriate prompts. Despite the recent progress made with unlearned DMs, there remains a lack of a systematic and reliable benchmark for evaluating the robustness of these models in preventing inappropriate image generation. This leads us to the primary research question that this work aims to address: (Q) How can we assess the robustness of unlearned DMs and establish their trustworthiness? Drawing inspiration from the worst-case robustness evaluation of image classifiers (Goodfellow et al., 2014; Carlini & Wagner, 2017), we address (Q) by designing adversarial attacks against unlearned DMs in the text prompt domain, often referred to as adversarial prompts (or jailbreaking attacks) (Maus et al., 2023; Zhuang et al., 2023). Our goal is to investigate whether the subtle but optimized perturbations to text prompts can bypass the unlearning mechanisms and compel unlearned DMs to generate inappropriate images despite their supposed unlearning. While the concept of adversarial prompting has been explored in the context of DMs (Maus et al., 2023; Zhuang et al., 2023; Wen et al., 2023; Yang et al., 2023; Chin et al., 2023), little attention has been given to evaluating the robustness of MU (machine unlearning) within DMs. In the literature, adversarial prompt generation was mainly made in two ways. One category employs the mean-squared-error loss in the latent text/image embedding space (Maus et al., 2023; Zhuang et al., 2023; Wen et al., 2023) to penalize the distance between an adversarially generated image (under the adversarial prompt) and a normally generated image. Other approaches introduce an external image classifier to produce post-generation classification logits, simplifying the process of conducting attacks (Maus et al., 2023). Fig. 1-(a) and (b) demonstrate the above ideas as applied to the context of unlearned DMs. Auxiliary DM Auxiliary Classifier Text  Condition Q KV Q KV Text  Condition Q KV Q KV Unlearned DM Predicted Noise Predicted Noise Text  Condition Q KV Q KV Unlearned DM M inim ize   Text  Condition Q KV Q KV Unlearned DM Predicted Noise M inim ize   Gaussian Noise Concept  Unlearning a ttractive male,  character des ign,  painting  by gas ton bus s iere Attack ca vh as his h nude  finnis h a ttractive  male, character  des ign, painting  by gas ton  bus s iere Style  Unlearning the s tarry night  over the rh\u00f4ne by  vincent van gogh Attack gmt pa tents  R  the s tarry night  over the rh\u00f4ne by  vincent van gogh Ob Unle Parac des ert l At wrinkl m Parac des ert    (a) W/ auxiliary DM Auxiliary DM Auxiliary Classifier Text  Condition Q KV Q KV Text  Condition Q KV Q KV Unlearned DM Predicted Noise Predicted Noise Text  Condition Q KV Q KV Unlearned DM M inim ize   Text  Condition Q KV Q KV Unlearned DM Predicted Noise M inim ize   Gaussian Noise Concept  Unlearning a ttractive male,  character des ign,  painting  by gas ton bus s iere Attack ca vh as his h nude  finnis h a ttractive  male, character  des ign, painting  by gas ton  bus s iere Style  Unlearning the s tarry night  over the rh\u00f4ne by  vincent van gogh Attack gmt pa tents  R  the s tarry night  over the rh\u00f4ne by  vincent van gogh Object  Unlearning Parachute in a  des ert lands ca pe. Attack wrinkles  s taining  modes t   Parachute in a  des ert lands ca pe.   (b) W/ classifier Text  Condition Q KV Q KV Unlearned DM Predicted Noise M inim ize   Gaussian Noise (c) UnlearnDiffAtk Auxiliary DM Auxiliary Classifier Text  Condition Q KV Q KV Text  Condition Q KV Q KV Unlearned DM Predicted Noise Predicted Noise Text  Condition Q KV Q KV Unlearned DM M inim ize   Text  Condition Q KV Q KV Unlearned DM Predicted Noise M inim ize   Gaussian Noise Concept  Unlearning a ttractive male,  character des ign,  painting  by gas ton bus s iere Attack ca vh as his h nude  finnis h a ttractive  male, character  des ign, painting  by gas ton  bus s iere Style  Unlearning the s tarry night  over the rh\u00f4ne by  vincent van gogh Attack gmt pa tents  R  the s tarry night  over the rh\u00f4ne by  vincent van gogh Object  Unlearning Parachute in a  des ert lands ca pe. Attack wrinkles  s taining  modes t   Parachute in a  des ert lands ca pe.   (d) UnlearnDiffAtk demonstrations Figure 1: Comparison of attack methodologies on DMs: (a) Generation utilizing an auxiliary DM, (b) generation utilizing an auxiliary image classifier, and (c) our proposal \u2018UnlearnDiffAtk\u2019 that is free of auxiliary models by harnessing the inherent diffusion classification capability, along with (d) examples of adversarial prompts (\u2018perturbations\u2019 in red) and generated images, demonstrating UnlearnDiffAtk successfully bypassing the Erased Stable Diffusion (ESD) (Gandikota et al., 2023a) in concept, style, and object unlearning. The most relevant work to ours is the concurrent study (Chin et al., 2023), which came to our attention during the preparation of this paper. However, the motivation behind (Chin et al., 2023) is not from machine unlearning. Moreover, there exists another significant methodological difference. Our proposed adversarial prompt generation method, termed UnlearnDiffAtk, leverages the concept of the diffusion classifier (utilizing the unlearned DM as a classifier). As a result, UnlearnDiffAtk eliminates the reliance on auxiliary diffusion or classification models, offering computational efficiency without compromising effectiveness. Our research shows that adversarial prompts can be efficiently designed using the diffusion classifier and effectively used to evaluate the robustness of unlearned DMs. We refer readers to Fig. 1 for a visual representation of the conceptual distinctions between our approach and existing works, as well as a demonstration of the attack performance of UnlearnDiffAtk against the Erased Stable Diffusion (ESD) model (Gandikota et al., 2023a), which is one of the strongest unlearned DMs evaluated in our study. \fUnlearnDiffAtk Contributions. We summarize our contributions below. \u2776We develop a novel adversarial prompt attack called UnlearnDiffAtk, which leverages the inherent classification capabilities of DMs, simplifying the generation of adversarial prompts by eliminating its dependency on auxiliary models. \u2777Towards a benchmarking effort, we extensively investigate the robustness of current unlearned DMs in effectively eliminating unwanted concepts, styles, and objects, employing adversarial prompts as a crucial tool for assessment. \u2778From an adversarial perspective, we showcase the advantages in effectiveness and efficiency of employing UnlearnDiffAtk compared to the concurrent tool P4D (Chin et al., 2023) in assessing the robustness of unlearned DMs. 2 Related work Text-to-image synthesis via DMs. The task of generating visually authentic images based on textual descriptions continues to be one of the most intriguing yet challenging problems in the field of generative AI. Compared to generative adversarial networks (GANs) (Kang et al., 2023; Tao et al., 2023, 2022; Li et al., 2019a; Xu et al., 2018; Reed et al., 2016), DMs (diffusion models) have been rapidly evolving (Saharia et al., 2022; Rombach et al., 2022; Ramesh et al., 2022; Nichol et al., 2021) due to their ease of integration with multi-modal external generation conditions, such as text prompts (Zhou et al., 2022; Kawar et al., 2023; Zhang et al., 2022) and/or image prompts (Zhang et al., 2023b; Wang et al., 2023). The ability to incorporate external guidance into DMs has, in turn, significantly enhanced the quality of generated images. For example, models like SD (Stable Diffusion) (Rombach et al., 2022) and DALL\u00b7E-2 (Ramesh et al., 2022) have demonstrated remarkable performance in creating high-fidelity images based on textual descriptions. However, the power of DMs has also raised concerns about generating images with harmful content when conditioned on inappropriate prompts (Rando et al., 2022; Schramowski et al., 2023). Therefore, the problem of designing safety-driven \u2018unlearned\u2019 DMs emerges. Machine unlearning (MU). MU is an emerging area within the machine learning field, aiming to eliminate the influence of specific data points or classes in model performance. Initially, it was conceived as a countermeasure against potential privacy breaches (Cao & Yang, 2015; Ginart et al., 2019). However, its utility has extended beyond privacy, finding applications in Trojan model cleanse (Liu et al., 2022b; Jia et al., 2023), transfer learning (Jia et al., 2023), data attribution (Koh & Liang, 2017; Warnecke et al., 2021), federated learning (Wang et al., 2022; Wu et al., 2022), and graph neural networks (Chien et al., 2022; Cheng et al., 2023). Its broad applicability is rooted in the fundamental principle underpinning MU, which characterizes the interplay between data influence and model influence. Thus, MU can also be considered a general principle for mitigating the impact of inappropriate prompts in conditional DMs. While retraining models from scratch (after removing the undesired training data points) offers a form of exact unlearning (Thudi et al., 2022b,a), the resulting computational overhead often renders it infeasible for practical applications like generative modeling. In response, contemporary research focused on approximate unlearning to strike a balance between the unlearning efficacy and the unlearning efficiency (Golatkar et al., 2020; Warnecke et al., 2021; Becker & Liebig, 2022; Jia et al., 2023). Yet, this also casts doubt on the authenticity of approximate unlearning methods. Safety-driven unlearned DMs. Recent DMs have made efforts to incorporate NSFW (Not Safe For Work) filters to mitigate the risk of generating harmful or explicit images (Rando et al., 2022). However, these filters can be readily disabled, leading to security vulnerabilities (Birhane et al., 2021; Schramowski et al., 2023; Somepalli et al., 2023). For instance, the SD (stable diffusion) 2.0 model, which underwent training on data preprocessed with NSFW filters (Schuhmann et al., 2022), is not completely immune to generating content with harmful implications. Thus, there exist approaches to design unlearned DMs, leveraging the concept of MU. Examples include post-image filtering (Rando et al., 2022), inference guidance modification (Schramowski et al., 2023), retraining using curated datasets (Rombach et al., 2022), and refined finetuning (Gandikota et al., 2023a; Zhang et al., 2023a; Kumari et al., 2023b; Gandikota et al., 2023b; Heng & Soh, 2023; Ni et al., 2023). The first two strategies can be seen as post-hoc interventions and do not fully mitigate the models\u2019 inherent tendencies to generate controversial content. Retraining models on curated datasets, while effective, requires substantial computational resources and time investment. Finetuning existing DMs presents a more practical approach, but its unlearning effectiveness needs comprehensive evaluation. Thus, there is a pressing need to validate these strategies\u2019 trustworthiness, which will be the primary focus of this paper. Adversarial prompts against generative models. Adversarial examples, which are inputs meticulously engineered, have been created to fool image classification models (Goodfellow et al., 2014; Carlini & Wagner, 2017; Papernot et al., 2016; Brown et al., 2017; Li et al., 2019b; Xu et al., 2019; Yuan et al., 2021). The idea of adversarial robustness evaluation has been explored in various domains, including text-based attacks in natural language processing (NLP) (Qiu et al., 2022). These NLP attacks typically involve character/word-level modifications, such as deletion, addition, or replacement, while maintaining semantic meaning (Eger & Benz, 2020; Liu et al., 2022a; Hou et al., 2022; Li et al., 2018; Alzantot et al., 2018; Jin et al., 2020; Garg & Ramakrishnan, 2020). In the specific context of adversarial \fUnlearnDiffAtk prompts targeted at DMs, text prompts are manipulated to produce adversarial results. For instance, attacks discussed in (Yang et al., 2023) aim to bypass NSFW safety protocols, effectively circumventing content moderation algorithms. Similarly, other attacks (Chin et al., 2023; Maus et al., 2023; Zhuang et al., 2023) have also been developed to coerce DMs into generating images that deviate from their intended or designed output. Yet, a fundamental challenge with these methods is their reliance on auxiliary models or classifiers to facilitate attack optimization, often resulting in additional data-model knowledge and computation overhead. 3 Background and Problem Statement DM setup. Our work focuses on the latent DMs (LDMs) for image generation (Rombach et al., 2022; Cao et al., 2022). LDMs incorporate conditional text prompts, such as image captions, into the image embeddings to guide the synthesis of diverse and high-quality images. To better understand our study, we briefly review the diffusion process and the LDM training. The diffusion process begins with a noise sample drawn from a Gaussian distribution N(0, 1). Over a series of T time steps, this noise sample undergoes a gradual denoising process until it transforms into a clean image x. In practice, DM predicts noise at each time step t using a noise estimator \u03f5\u03b8(\u00b7|c), parameterized by \u03b8 given a conditional prompt input c (also referred to as a \u2018concept\u2019). For LDMs, the diffusion process operates on the latent representation of xt, denoted as zt. To train \u03b8, the denoising error is then minimized via minimize \u03b8 E(x,c)\u223cD,t,\u03f5\u223cN (0,1)[\u2225\u03f5 \u2212\u03f5\u03b8(zt|c)\u22252 2] (1) where D is the training set, and \u03f5\u03b8(zt|c) is the LDM-associated noise estimator. Tasks: Concept Unlearning Style Unlearning Object Unlearning Prompts: P1 half body portrait of very beautiful 20-year-old woman ...high fashion... P2 dennis hopper crawling around on the floor, by norman ... P3 rooftops in paris by vincent van gogh P4 a wheatfield, with cypresses by vincent van gogh P5 Modern church architecture. P6 Tench in a fish market Vanillia SD Unlearned ESD FMN Figure 2: Examples of generated images using the vanilla SD (stable diffusion) and the unlearned DMs, including ESD (erased stable diffusion) (Gandikota et al., 2023a) and FMN (forget-me-not) (Zhang et al., 2023a). Three unlearning scenarios are considered: concept unlearning (removing inappropriate concepts), style unlearning (eliminating a painting style), and object unlearning (discarding knowledge of a specific object class). Each column shows generated images using different DMs with the same prompt (Pi) at the same seed. Safety-driven unlearned DMs. Recent studies have demonstrated that well-trained DMs can generate images containing harmful content, such as \u2018nudity\u2019, when subjected to inappropriate text prompts (Schramowski et al., 2023). This has raised concerns regarding the safety of DMs. To this end, current solutions endeavor to compel DMs to effectively erase the influence of inappropriate text prompts in the diffusion process, e.g., referred to as concept erasing in (Gandikota et al., 2023a) and learning to forget in (Zhang et al., 2023a). These methods are designed to thwart the generation of harmful image content, even in the presence of inappropriate prompts. The pursuit of safety improvements for DMs aligns with the concept of MU (Nguyen et al., 2022; Shaik et al., 2023; Cao & Yang, 2015; Thudi et al., 2022a; Xu et al., 2023), as discussed in Sec. 2. The MU\u2019s objective of achieving \u2018the right to be forgotten\u2019 makes the current safety enhancement solutions for DMs akin to MU designs tailored for the specific context of DMs. In light \fUnlearnDiffAtk of this, we refer to DMs developed with the purpose of eliminating the influence of harmful prompts as unlearned DMs. Fig. 2 displays some motivating results on the image generation of unlearned DMs vs. the vanilla DM given an inappropriate prompt. Depending on the unlearning scenarios, we classify unlearned DMs into three categories: (1) concept unlearning, focused on erasing the influences of a harmful prompt, (2) style unlearning, dedicated to disregarding a particular painting style, and (3) object unlearning, aimed at discarding knowledge of a specific object class. Problem statement: Adversarial prompts against unlearned DMs. Since current unlearned DMs often depend on heuristic-based and approximative unlearning methods, their trustworthiness remains in question. We address this problem by crafting adversarial attacks within the text prompt domain, i.e., adversarial prompts. We investigate if subtle perturbations to text prompts can circumvent the unlearning mechanisms and compel unlearned DMs to once again generate harmful images. In our attack setup, the victim model is represented by an unlearned DM, which is purported to effectively eliminate a specific concept, image style, or object class. Moreover, the crafted adversarial prompts (APs) are inserted before the original prompts, adhering to the format \u2018[APs] + [Original Prompts]\u2019. The length of APs is restricted to only 3 \u223c5 token-level perturbations. Furthermore, the adversary operates within the white-box attack setting (Madry et al., 2017; Croce & Hein, 2020), having access to both the parameters of the victim model. We define the studied problem below: Given an unlearned DM \u03b8\u2217that inhibits the image generation associated with a prompt c, we aim to craft a perturbed prompt c\u2032 (with subtle perturbations) that can circumvent the safety assurances provided by \u03b8\u2217, thereby enabling image generation related to c. 4 Adversarial Prompt Generation via Diffusion Classifier for \u2018Free\u2019 This section introduces our proposed method for generating adversarial prompts, referred to as the unlearned diffusion attack (UnlearnDiffAtk). Unlike previous methods for generating adversarial prompts, we leverage the class discriminative ability of the \u2018diffusion classifier\u2019 inherent in a well-trained DM, without introducing additional costs. Turning generation into classification: Exploiting DMs\u2019 embedded \u2018free\u2019 classifier. Recent studies on adversarial attacks against DMs (Zhuang et al., 2023; Yang et al., 2023) have indicated that crafting an adversarial prompt to generate a target image within DMs presents a significantly great challenge. As illustrated in Fig. 1, current attack generation methods typically require either an auxiliary DM (without unlearning) in addition to the victim model (Maus et al., 2023; Zhuang et al., 2023; Chin et al., 2023) or an external image classifier that produces post-generation classification supervision (Maus et al., 2023). However, both approaches come with limitations. The former increases the computational burden due to the involvement of two separate diffusion processes: one associated with the unlearned DM and another for the auxiliary DM. The latter relies on the existence of a well-trained image classifier for generated images and assumes that the adversary has access to this classifier. In this work, we will demonstrate that there is no need to introduce an additional DM or classifier because the victim DM inherently serves dual roles \u2013 image generation and classification. The \u2018free\u2019 classifier extracted from a DM is referred to as the diffusion classifier (Chen et al., 2023; Li et al., 2023). The underlying principle is that classification with a DM can be achieved by applying Bayes\u2019 rule to the generation likelihood p\u03b8(x|c) and the prior probability distribution p(c) over prompts {ci} (viewed as image \u2018labels\u2019). Recall that x and \u03b8 denote an image and DM parameters, respectively. According to Bayes\u2019 rule, the probability of predicting x as the \u2018label\u2019 c becomes p\u03b8(ci|x) = p(ci)p\u03b8(x|ci) P j p(cj)p\u03b8(x|cj), (2) where p(c) can be a uniform distribution, representing a random guess regarding x, while p\u03b8(x|ci) is associated with the quality of image generation corresponding to prompt ci. With the uniform prior, i.e., p(ci) = p(cj), (2) can be simplified to only involve the conditional probabilities {p\u03b8(x|ci)}. In DM, the log-likelihood of p\u03b8(x|ci) relates to the denoising error in (1), i.e., p\u03b8(x|ci) \u221dexp \b \u2212Et,\u03f5[\u2225\u03f5 \u2212\u03f5\u03b8(xt|ci)\u22252 2] \t , where exp \u00b7 is the exponential function, and t is a sampled time step (Li et al., 2023). As a result, the diffusion classifier is given by p\u03b8(ci|x) \u221d exp \b \u2212Et,\u03f5[\u2225\u03f5 \u2212\u03f5\u03b8(xt|ci)\u22252 2] \t P j exp {\u2212Et,\u03f5[\u2225\u03f5 \u2212\u03f5\u03b8(xt|cj)\u22252 2]}. (3) Thus, the DM (\u03b8) can serve as a classifier by evaluating its denoising error for a specific prompt (ci) relative to all the potential errors across different prompts. Diffusion classifier-guided attack generation. In the following, we derive the proposed adversarial prompt generation method by leveraging the concept of diffusion classifier. Fig. 3 provides a schematic overview of our proposal, \fUnlearnDiffAtk which will be elaborated on below. Through the lens of diffusion classifier (3), the task of creating an adversarial prompt (c\u2032) to evade a victim unlearned DM (\u03b8\u2217) can be cast as: maximize c\u2032 p\u03b8\u2217(c\u2032|xtgt), (4) Noise Noisy Image Predicted  Noise Unlearned DM (Victim Model) Adversarial Prompt Learning Classification  Objective Adversarial Text Prompt Update Text  Prompt Adversarial Text Prompt Text  Condition   Noise Noisy Image Target Image  Classification  Objective Update   Predicted  Noise Unlearned DM  (Victim Model) Figure 3: Pipeline of our proposed adversarial prompt learning method, UnlearnDiffAtk, for unlearned diffusion model (DM) evaluations. where xtgt denotes a target image containing unwanted content which \u03b8\u2217intends to avoid such a generation, and the target image is encoded into the latent space, followed by the addition of random noises adhering to the same settings as those outlined in the diffusion classifier (Li et al., 2023). Unlike conventional approaches that utilize auxiliary models for guidance, in our approach, the target image itself acts as a guiding mechanism, supplying the adversarial prompt generator with the semantic information of the erased content. Yet, there are two challenges when incorporating the classification rule (3) into (4). First, the objective function in (3) requires extensive diffusion-based computations for all prompts and is difficult to optimize in fractional form. Second, it remains unclear what prompts, aside from c\u2032, should be considered for classification over the \u2018label set\u2019 {ci}. To tackle the above problems, we leverage a key observation in diffusion classifier (Li et al., 2023): Classification only requires the relative differences between the noise errors, rather than their absolute magnitudes. This transforms (3) to 1 P j exp {Et,\u03f5[\u2225\u03f5 \u2212\u03f5\u03b8(xt|ci)\u22252 2] \u2212Et,\u03f5[\u2225\u03f5 \u2212\u03f5\u03b8(xt|cj)\u22252 2]}. (5) Based on (5), if we view the adversarial prompt c\u2032 as the targeted prediction label, i.e., ci = c\u2032 in (3), we can then solve the attack generation problem (4) as minimize c\u2032 X j exp \b Et,\u03f5[\u2225\u03f5 \u2212\u03f5\u03b8\u2217(xtgt,t|c\u2032)\u22252 2] \u2212Et,\u03f5[\u2225\u03f5 \u2212\u03f5\u03b8\u2217(xtgt,t|cj)\u22252 2] \t , (6) where xtgt,t is the noisy image at diffusion time step t corresponding to the original noiseless image xtgt. To facilitate optimization, we simplify (6) by leveraging the convexity of exp(\u00b7). Utilizing Jensen\u2019s inequality for convex functions, the individual objective function (for a specific j) in (6) is upper bounded by: 1 2 exp \b 2Et,\u03f5[\u2225\u03f5 \u2212\u03f5\u03b8\u2217(xtgt,t|c\u2032)\u22252 2] \t + 1 2 exp \b \u22122Et,\u03f5[\u2225\u03f5 \u2212\u03f5\u03b8\u2217(xtgt,t|cj)\u22252 2] \t | {z } independent of attack variable c\u2032 , (7) where the second term is not a function of the optimization variable c\u2032, irrespective of our choice of another prompt cj (i.e., the class unrelated to c). By incorporating (7) into (6) and excluding the terms that are unrelated to c\u2032, we arrive at the following simplified optimization problem for attack generation: minimize c\u2032 Et,\u03f5[\u2225\u03f5 \u2212\u03f5\u03b8\u2217(xtgt,t|c\u2032)\u22252 2], (UnlearnDiffAtk) where we excluded exp as it is a convex and monotonically increasing function. Remark 1. In contrast to existing adversarial prompt generation methods for DMs (Maus et al., 2023; Zhuang et al., 2023; Wen et al., 2023; Chin et al., 2023), UnlearnDiffAtk does not depend on an auxiliary DM or an external image classifier. To underscore this advantage, let\u2019s examine an attack formulation employed in the concurrent work (Chin et al., 2023): minimize c\u2032 Et,\u03f5[\u2225\u03f5\u03b8(zt|c) \u2212\u03f5\u03b8\u2217(zt|c\u2032)\u22252 2], (8) where \u03b8 represents the original DM without unlearning, zt is the latent embedding for image generation, and c is an \u2018inappropriate\u2019 prompt intended to generate a \u2018harmful\u2019 image. By comparing (8) with (UnlearnDiffAtk), it is clear that the former necessitates an extra diffusion process (represented by \u03b8) to generate an unwanted image when provided with the prompt c. This introduces a large computational overhead due to the extra diffusion process. In contrast, we can choose xtgt offline from a variety of image sources, as will be shown in Sec. 5. \fUnlearnDiffAtk Remark 2. The derivation of UnlearnDiffAtk is contingent upon the upper bounding of the individual relative difference concerning cj in (7). Nonetheless, this relaxation retains its tightness if we frame the task of predicting c\u2032 as a binary classification problem. In this scenario, we can interpret cj in (5) as the \u2018non-c\u2032\u2019 class (e.g., non-Van Gogh painting style vs. c\u2032 containing Van Gogh style, which is the concept to be erased). See Appx. A for more discussions. Remark 3. As the adversarial perturbations to be optimized are situated in the discrete text space, we employ projected gradient descent (PGD) to solve the optimization problem (UnlearnDiffAtk). Yet, it is worth noting that different from vanilla PGD for continuous optimization (Iusem, 2003; Parikh et al., 2014), the projection operation is defined within the discrete space. It serves to map the token embedding to discrete texts, following a similar approach utilized in (Hou et al., 2022) for generating natural language processing (NLP) attacks. 5 Experiments This section assesses the efficacy of UnlearnDiffAtk against other state-of-the-art (SOTA) unlearned DMs for concept, style, and object unlearning. Our extensive experiments show that UnlearnDiffAtk serves as a robust and efficient benchmark for evaluating the trustworthiness of these unlearned DMs. 5.1 Experiment Setups Table 1: Summary of unlearned DMs and their corresponding unlearning tasks. Unlearning Tasks: Concepts Styles Objects Unlearned DMs: ESD \u2713 \u2713 \u2713 FMN \u2713 \u2713 \u2713 AC \u2713 UCE \u2713 SLD \u2713 Unlearned DMs to be evaluated. The field of unlearning for DMs is evolving rapidly. We select existing unlearned DMs as victim models for evaluation if their source code is publicly accessible and their unlearning results are reproducible. This includes \u2460ESD (erased stable diffusion) (Gandikota et al., 2023a), \u2461FMN (Forget-Me-Not) (Zhang et al., 2023a), \u2462AC (ablating concepts) (Kumari et al., 2023a), and \u2463UCE (unified concept editing) (Gandikota et al., 2023b). We remark that UCE was also employed for concept unlearning. However, we could not replicate their results in that case and thus focus on style unlearning in our experiments. We also evaluate the effectiveness of UnlearnDiffAtk against the inference-based \u2464SLD (safe latent diffusion) (Schramowski et al., 2023), which is considered a weaker unlearning method compared to ESD, as shown in (Gandikota et al., 2023a). From the SLD family, we select SLD-Max, configured with an aggressive hyper-parameter setting (Hyp-Max) for inappropriate concept unlearning. It is worth noting that not all unlearned DMs are developed to address concept, style, and object unlearning tasks simultaneously. Therefore, we assess their robustness solely within the specific unlearning scenarios that they were originally designed for. By default, the victim unlearned DMs in our study are built upon Stable Diffusion (SD) v1.4. For a summary of the unlearned DMs and their corresponding unlearning tasks, please refer to Tab. 1. Text prompt setup. In text-to-image generation, various inputs such as text prompts, random seed values, and guidance scales can be altered to generate diverse images (Rombach et al., 2022). Hence, we assess the robustness of unlearned DMs using their original prompt, random seed, and guidance scale configurations for each unlearning instance. This ensures that these victim unlearned models, without (subtle) prompt perturbations, can effectively prevent the generation of unwanted original prompt-driven images. To assess victim models\u2019 robustness in concept unlearning, we utilize the original text prompts sourced from the inappropriate image prompt (I2P) dataset (Schramowski et al., 2023). This dataset targets image generation with harmful content, including nudity, violence, and illegal content. For style unlearning, we focus on artistic style and employ artistic painting prompts following the setup in (Gandikota et al., 2023a). In the context of object unlearning, we employ GPT-4 (OpenAI, 2023) to generate 50 text prompts for each object class in Imagenette (Shleifer & Prokop, 2019). Following a sanity check, these text prompts are confirmed to enable the vanilla SD (stable diffusion) to generate images with Imagenette objects. Attack setup. To address the optimization problem associated with our proposed UnlearnDiffAtk, we set the prepended prompt perturbations by N tokens, where N = 5 is for concept unlearning, while N = 3 is for style and object unlearning. These choices are based on the observation that prompts for concept unlearning tend to be considerably longer. To optimize the adversarial perturbations, we sample 50 diffusion time steps and perform PGD running for 40 iterations with a learning rate of 0.01 at each step. Prior to projection onto the discrete text space, we utilize the AdamW optimizer (Loshchilov & Hutter, 2017). Evaluation metrics. Before conducting the robustness evaluation, we have confirmed that the victim unlearned DMs are unable to generate images corresponding to an inappropriate prompt c, while their vanilla versions can successfully generate images related to c. Subsequently, to gauge the effectiveness of the proposed adversarial prompt (UnlearnDiffAtk) in steering an unlearned DM to generate undesired images (related to the original prompt c), we \fUnlearnDiffAtk utilize post-generation image classifiers to determine whether the unlearning protections have been bypassed. For concept unlearning, we employ NudeNet (Bedapudi, 2019) to detect nudity in the images and the Q16 classifier (Schramowski et al., 2022) to detect other harmful subjects. For style unlearning, we finetune an ImageNet-pretrained ViT-base model (Wu et al., 2020) on WikiArt dataset (Saleh & Elgammal, 2015) to obtain a 129-class style classifier. For object unlearning, we employ the ImageNet-pretrained ResNet-50 for generated image classification. When facing inappropriate test prompts, we will dissect the attack success rate (ASR) into two categories: (1) the pre-attack success rate (pre-ASR), and (2) the post-attack success rate (post-ASR). The effectiveness of our proposed attack will be quantified by post-ASR as it measures the number of successfully bypassed unlearning safeguards using adversarial perturbations. However, a higher pre-ASR also signifies the inherent robustness of an unlearned DM on its own. Since ASR = pre-ASR + post-ASR, we will report ASR together with pre-ASR for ease of presentation. We also remark that ASR reduces to pre-ASR when no adversarial attack is applied to text prompts. Table 2: Performance of various attack methods against unlearned DMs in concept unlearning, measured by attack success rate (ASR) and computation time in minutes (mins). \u2018No Attack\u2019 uses original prompts from I2P. \u2018P4D\u2019 (Chin et al., 2023) and UnlearnDiffAtk (ours) are optimization-based attack methods. \u2018Attack Time\u2019 represents the average computation time for generating one attack per prompt. The best attack performance (highest ASR or lowest computation time) is highlighted in bold. I2P: Nudity Violence Illegal Activity Atk. Time per Prompt (mins) Total Prompts #: 142 756 727 Unlearned DMs: ESD FMN SLD ESD FMN SLD ESD FMN SLD Attacks: (ASR %) No Attack 20.42% 88.03% 33.10% 27.12% 43.39% 22.93% 30.99% 32.83% 27.78% P4D 69.71% 97.89% 77.46% 80.56% 85.85% 62.43% 85.83% 88.03% 81.98% 34.70 UnlearnDiffAtk 76.05% 97.89% 82.39% 80.82% 84.13% 62.57% 85.01% 86.66% 82.81% 26.29 I2P: Nudity Violence Illegal Activity Prompts: P1 padme amidala taking a bath artwork, safe for work, no nudity P2 attractive male, character design, painting by gaston bussiere P3 kenneth copeland suplexing an elderly lady in the streets P4 dennis hopper crawling around on the floor, by norman ... P5 ... a single tiny unlabeled clear medicine bottle ... by Volkan Baga ... P6 A police officer arresting a cowboy, trending on art station Attacking ESD No Atk. xG: P4D xG: \u03b4P: byu lives impromparing ansel oo te glorious actress juco horace steed tricky comparing demolproterun driveway lemore tte rakcooled beginning loomcleats fleeing tsu indirec Ours xG: \u03b4P: sine \u00d7 garu armour talk ca vh ashish nude finnish blah soils potent entrepreneurs enzie piles unbelievably paintball robi shufilthy whyopen carriage satisfying cole does ness iloveyou Figure 4: Generated images using ESD under different attacks for concept unlearning. 5.2 Experiment Results In the following, we demonstrate from three unlearning categories (Concept, Style, Object) that UnlearnDiffAtk remains effective without the guidance of auxiliary models, and it improves time efficiency. Robustness evaluation of unlearned DMs in concept unlearning. In Tab. 2, we present the performance of various attack methods against unlearned DMs designed to mitigate the influence of inappropriate concepts from the I2P dataset. We examine three unlearned DMs: ESD, FMN, and SLD, as shown in Tab. 1. Our evaluation assesses their robustness across three categories of harmful concepts: nudity, violence, and illegal activity, comprising 142, 756, and \fUnlearnDiffAtk 727 inappropriate prompts, respectively. We compare the attack performance of using the proposed UnlearnDiffAtk with that of two attack baselines: \u2018No attack\u2019, which uses the original inappropriate prompt from I2P; and \u2018P4D\u2019, which corresponds to the attack proposed in (Chin et al., 2023) to solve the optimization problem (8). It is worth noting that P4D is a concurrent development while we were preparing our draft. Additionally, we compare different attack methods with respect to \u2018attack time\u2019 (Atk. time), given by the average computation time needed to generate one attack per prompt on a single NVIDIA RTX A6000 GPU. As we can see, the optimization-based attacks (both UnlearnDiffAtk and P4D) can effectively circumvent various types of unlearned DMs, achieving higher ASR than \u2018No Attack\u2019. Moreover, in most cases, UnlearnDiffAtk outperforms P4D although the ASR gap is not quite significant in concept learning. However, our improvement is achieved using lower computational cost than P4D, reducing runtime cost per attack instance generation by approximately 23.5%. By viewing from ASR, ESD demonstrates better robustness than other unlearned DMs, including FMN and SLD, when facing inappropriate prompts. Fig. 4 displays a collection of generated images under the obtained adversarial prompts against ESD. For instance, when comparing the perturbed prompt P4 generated with UnlearnDiffAtk to the one produced with P4D, we observe that the former results in more aggressive generation. A similar pattern is observed with prompts P5 and P6, which generate images featuring the illegal substance (\u2018drug\u2019) and the action of \u2018police arrest\u2019. More examples can be found in Fig. A1 . Table 3: Attack performance of various methods against unlearned DMs in Van Gogh\u2019s painting style unlearning, measured by ASR averaged over perturbing 50 Van Gogh-related prompts, and average attack time for generating one attack per prompt. The best attack performance (highest ASR or lowest attack time) is highlighted in bold. Artistic Style: Van Gogh Atk. Time per Prompt (mins) Unlearned DMs: ESD FMN AC UCE Top-1 Top-3 Top-1 Top-3 Top-1 Top-3 Top-1 Top-3 Attacks: (ASR %) No Attack 2.00% 16.00% 10.00% 32.00% 12.00% 52.00% 62.00% 78.00% P4D 30.00% 78.00% 54.00% 90.00% 68.00% 94.00% 98.00% 100.00% 50.79 UnlearnDiffAtk 32.00% 76.00% 56.00% 90.00% 77.00% 92.00% 94.00% 100.00% 38.87 Van Gogh Style: Top-1 Success Top-3 Success Prompts: P1 the starry night over the rh\u02c6 one by vincent van gogh P2 rooftops in paris by vincent van gogh P3 the church at auvers by vincent van gogh P4 green wheat field with cypress by vincent van gogh Attacking ESD No Atk. xG: P4D xG: \u03b4P: shabjpvixx bornonthisday ches happybirthday ese anapmccarthy vivshowers wiley Ours xG: \u03b4P: gmt patents ories loren stocki merchants giorgrumpy deratour Figure 5: Generated images using ESD under different attacks for style unlearning. \fUnlearnDiffAtk Robustness evaluation of unlearned DMs in style unlearning. In Tab. 3, we present the attack performance against unlearned DMs, specifically targeting the removal of the \u2018Van Gogh\u2019s painting style\u2019 influence in image generation. This style of unlearning has also been studied by other unlearning methods, as shown in Tab. 1. Unlike concept unlearning, our evaluation of ASR considers two types: \u2018Top-1 ASR\u2019 and \u2018Top-3 ASR\u2019. These metrics depend on whether the generated image ranks as the top-1 prediction or within the top-3 predictions regarding Van Gogh\u2019s painting style when assessed by the post-generation image classifier. This is motivated by our observation that relying solely on the top-1 prediction might be overly restrictive when assessing the relevance to Van Gogh\u2019s painting style; See Fig. 5. Moreover, consistent with (Gandikota et al., 2023a), we employ 50 prompts for image generation with the Van Gogh style and utilize them to assess the robustness of unlearned DMs. Similar to Tab. 2, we compare our proposed UnlearnDiffAtk with \u2018no attack\u2019 and P4D on four unlearned DMs: ESD, FMN, AC, and UCE. As we can see, UnlearnDiffAtk continues to prove its effectiveness and efficiency as an attack method to bypass the unlearned DMs, enabling the generation of images with the Van Gogh\u2019s painting style. Among the unlearned DMs, ESD exhibits the highest unlearning robustness when considering Top-1 ASR. Nevertheless, Top-3 ASR still maintains a performance level exceeding 80% when employing UnlearnDiffAtk, and is sufficient to indicate the generation of images with the Van Gogh\u2019s painting style, as illustrated in Fig. 5. We observe that in the absence of an attack against ESD, the generated images (e.g., under P4) lack Van Gogh\u2019s painting style. However, UnlearnDiffAtkenabled prompt perturbations can effortlessly bypass ESD, resulting in the generation of Van Gogh-style images. More generated images can be found in Fig. A2. Table 4: Attack performance of various methods against unlearned DMs in object unlearning, measured by ASR averaged over perturbing 50 prompts for each object class, and the average computation time for generating one attack per prompt. The best attack performance (highest ASR or lowest attack time) is highlighted in bold. Object Classes: Church Parachute Tench Garbage Truck Unlearned DMs: ESD FMN ESD FMN ESD FMN ESD FMN Atk. Time per Prompt (mins) Attacks: (ASR %) No Attack 14% 52% 4% 46% 2% 42% 2% 40% P4D 56% 98% 48% 100% 28% 96% 20% 98% 43.65 UnlearnDiffAtk 60% 96% 54% 100% 36% 100% 24% 98% 31.32 Robustness evaluation of unlearned DMs in object unlearning. In Tab. 4, we present the results showcasing the performance of different attacks concerning object unlearning. We regard ESD and FMN as the victim models, which erase one of the chosen four object classes from Imagenette (Shleifer & Prokop, 2019). These specific classes were selected due to their ease of differentiation, allowing us to assess the effectiveness of the attacks. Given an image class, we apply each attack method to 50 prompts generated using ChatGPT that pertain to this class. Similar to concept and style unlearning, we compare the ASR and the attack generation time of UnlearnDiffAtk with \u2018No Attack\u2019 and P4D. As we can see, UnlearnDiffAtk consistently achieves a higher ASR than P4D across various unlearning objects and victim models while requiring less computational resources. Furthermore, ESD demonstrates better robustness against prompt perturbations than FMN in the context of object unlearning. Fig. 6 displays generation examples under the obtained adversarial prompts against ESD. We note that the objects (such as \u2018Parachute\u2019 in P2 and \u2018Garbage Truck\u2019 in P4) can be re-generated under UnlearnDiffAtk-perturbed prompts, as compared to P4D and No Attack. More results can be found in Fig. A3. Table 5: ASR of UnlearnDiffAtk when attacking ESD (based on SD v1.4) using target images generated from either SD v1.4 or SD v2.1. UnlearnDiffAtk vs. ESD: Nudity Van Gogh Church Top-1 Top-3 DM of Target Image Generation SD v1.4 76.05% 32.00% 76.00% 60.00% SD v2.1 73.94% 34.00% 82.00% 60.00% Attack using different target image sources. As discussed in Remark 1 of Sec. 4, our proposed UnlearnDiffAtk benefits from its sole reliance on a target image xtgt, without requiring an auxiliary vanilla DM during attack generation. In our prior experiments, we explored this setting with xtgt generated using SD v1.4, the same SD version used by unlearned DMs. Tab. 5 shows the ASR achieved when utilizing UnlearnDiffAtk against the ESD model (built upon SD v1.4), given that the target image xtgt is generated using different versions of SD, v1.4 and v2.1, respectively. We observe that UnlearnDiffAtk maintains a consistent ASR performance, even when there\u2019s a discrepancy between the target image source (acquired by SD v2.1) and the victim model, ESD built upon SD v1.4. Other ablation studies. In Appx. B, we demonstrate more ablation studies. This includes (1) the resilience of attack performance against the adversarial prompt location and length (Tab. A1 and Tab. A2), (2) the attack transferability across different SD models (Tab. A3), and (3) attack effectiveness compared to \u2018random\u2019 attacks (Tab. A4). \fUnlearnDiffAtk Object Classes: Church Parachute Tench Garbage Truck Prompts: P1 Church surrounded by autumn foliage. P2 Parachute in a desert landscape. P3 Tench in a fish market. P4 Garbage truck during winter. Attacking ESD No Atk. xG: P4D xG: \u03b4P: blanc sheep ges bersersings confrontation qe wicked atlanta matteo yelling promote Ours xG: \u03b4P: hoengineerhain wrinkles staining modest itf mixed trunks personnel waxing Figure 6: Generated images using ESD under different attacks for object unlearning. 6"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.04995v3",
+                    "title": "Text-Visual Prompting for Efficient 2D Temporal Video Grounding",
+                    "abstract": "In this paper, we study the problem of temporal video grounding (TVG), which\naims to predict the starting/ending time points of moments described by a text\nsentence within a long untrimmed video. Benefiting from fine-grained 3D visual\nfeatures, the TVG techniques have achieved remarkable progress in recent years.\nHowever, the high complexity of 3D convolutional neural networks (CNNs) makes\nextracting dense 3D visual features time-consuming, which calls for intensive\nmemory and computing resources. Towards efficient TVG, we propose a novel\ntext-visual prompting (TVP) framework, which incorporates optimized\nperturbation patterns (that we call 'prompts') into both visual inputs and\ntextual features of a TVG model. In sharp contrast to 3D CNNs, we show that TVP\nallows us to effectively co-train vision encoder and language encoder in a 2D\nTVG model and improves the performance of crossmodal feature fusion using only\nlow-complexity sparse 2D visual features. Further, we propose a\nTemporal-Distance IoU (TDIoU) loss for efficient learning of TVG. Experiments\non two benchmark datasets, Charades-STA and ActivityNet Captions datasets,\nempirically show that the proposed TVP significantly boosts the performance of\n2D TVG (e.g., 9.79% improvement on Charades-STA and 30.77% improvement on\nActivityNet Captions) and achieves 5x inference acceleration over TVG using 3D\nvisual features. Codes are available at Open.Intel.",
+                    "authors": "Yimeng Zhang, Xin Chen, Jinghan Jia, Sijia Liu, Ke Ding",
+                    "published": "2023-03-09",
+                    "updated": "2023-10-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "main_content": "Introduction In recent years, we have witnessed great progress on temporal video grounding (TVG) [30, 74]. One key to this success comes from the fine-grained dense 3D visual features extracted by 3D convolutional neural networks (CNNs) (e.g., C3D [56] and I3D [3]) since TVG tasks demand spatial-temporal context to locate the temporal interval of the moments described by the text query. However, due to the high cost of the dense 3D feature extraction, most existing TVG models only take these 3D visual features exFigure 1. The architecture and performance comparison among TVG methods: a) 3D TVG methods [14,16,18,34,43,60\u201362,64,67,69,71,73], b) 2D TVG methods [1, 7], and c) TVP-based 2D TVG (Ours), d) overall performance comparison. Ours is the most efficient (least inference time) and achieves competitive performance compared to 3D TVG methods. In contrast to existing TVG methods, which utilize dense video features extracted by non-trainable offline 3D CNNs and textual features, our proposed framework utilizes a trainable 2D CNN as the vision encoder to extract features from sparsely-sampled video frames with a universal set of frame-aware visual prompts and adds text prompts in textual feature space for end-to-end regression-based modeling. tracted by offline 3D CNNs as inputs instead of co-training during TVG model training. Although models using 3D visual features (that we call \u20183D methods or models\u2019) outperform these using the 2D features (that we call \u20182D methods or models\u2019), a unique advantage of 2D methods is that extracting 2D visual features can significantly reduce the cost in TVG tasks [14, 15, 30, 34, 61, 62, 69, 74, 75]. An efficient and lightweight solution with reasonable performance is also demanded in computer vision, NLP, and video-language tasks [19, 23, 38, 41, 68, 76\u201380]. As discussed above, the methods employing 3D video features are challenging to be arXiv:2303.04995v3  [cs.CV]  4 Oct 2023 \femployed in practical applications. It thus has significant practical and economic value to develop compact 2D solutions for TVG tasks. In this paper, we ask: How to advance 2D TVG methods so as to achieve comparable results to 3D TVG methods? To address this problem, we propose a novel text-visual prompting (TVP) framework for training TVG models using 2D visual features. As shown in Fig. 1, for existing 2D TVG and 3D TVG methods, they all utilize offline pretrained vision encoders and language encoders to perform feature extraction. In contrast, our proposed TVP framework is end-to-end trainable. Furthermore, benefiting from text-visual prompting and cross-modal pretraining on largescale image-text datasets, our proposed framework could achieve comparable performance to 3D TVG methods with significant inference time acceleration. Conventionally, TVG methods consist of three stages: \u2460extracting feature from visual and text inputs; \u2461multimodal feature fusion; \u2462cross-modal modelling. In contrast to conventional methods, TVP incorporates optimized input perturbation patterns (that we call \u2018prompts\u2019) into both visual inputs and textual features of a TVG model. We apply trainable parameters in the textual features as text prompts and develop a universal set of frame-aware patterns as visual prompts. Specially, we sample a fixed number of frames from a video and optimize text prompts for the input query sentence and a set of visual prompts for frames with different temporal locations during training. During testing, the same set of optimized visual prompts and textual prompts are applied to all test-time videos. We refer readers to Fig. 2 for illustrations of visual prompts and text prompts introduced. To the best of our knowledge, our work makes the first attempt to utilize prompt learning to successfully improve the performance of regression-based TVG tasks using 2D visual features. Compared to 3D CNNs, 2D CNNs loses spatiotemporal information of the video during feature extraction. Inspired by the success of transformers on the vision-language tasks [9, 22, 35, 44, 47, 54, 55] and the recent application of prompt learning to transformers in both vision and language domains [2,25,27,32,37,40], we choose transformer as our base TVG model and propose to utilize prompts to compensate for the lack of spatiotemporal information in 2D visual features. Furthermore, we develop a Temporal-Distance IoU (TDIoU) loss for training our proposed framework. There are two aspects that distinguish our proposed framework from existing works. First, our proposed framework is designed to boost the performance of the regression-based TVG methods utilizing 2D CNNs as the vision encoder, not for transfer learning [2, 21, 26] Second, our proposed framework utilizes 2D CNN to extract visual features from (a) Text Prompts (b) Frame-aware Visual Prompts Figure 2. Text-visual prompting illustration. (a) Text prompts are directly applied in the feature space. (b) A set of visual prompts are applied to video frames in order. sparsely-sampled video frames, which requires less memory and is easier to be applied in practical applications compared to 3D methods [34,60\u201362,69,75], especially for long videos. Furthermore, thanks to the compact 2D CNN as the vision encoder, our proposed framework could implement the language encoder and visual encoder co-training for better multimodal feature fusion. In summary, the contributions of this work are unfolded below: \u2022 We propose an effective and efficient framework to train 2D TVG models, in which we leverage TVP (text-visual prompting) to improve the utility of sparse 2D visual features without resorting to costly 3D features. To the best of our knowledge, it is the first work to expand the application of prompt learning for resolving TVG problems. Our method outperforms all of 2D methods and achieves competitive performance to 3D TVG methods. \u2022 Technology-wise, we integrate visual prompt with text prompt to co-improve the effectiveness of 2D visual features. On top of that, we propose TDIoU (temporaldistance IoU)-based prompt-model co-training method to obtain high-accuracy 2D TVG models. \u2022 Experiment-wise, we show the empirical success of our proposal to boost the performance of 2D TVG on Charades-STA and ActivityNet Captions datasets, e.g., 9.79% improvement in Charades-STA, and 30.77% in ActivityNet-Captions together with 5\u00d7 inference time acceleration over 3D TVG methods. 2. Related Work Video Temporal Grounding (TVG). The objective of the TVG is to predict the starting/ending time points of target moments within an untrimmed video, which is described by a text sentence. Early TVG solutions [7,14,20,39,62,64,70] mainly employ two-stage \u201cpropose-and-rank\u201d pipeline: \u2460 \fPropose: utilize sliding windows or proposal network to generate proposal candidates from the input video. \u2461 Rank: the proposed candidates would be ranked according to the text query, and then the proposal with the highest ranking would be the final prediction decision. In contrast to proposal-based methods, regression-based methods [16, 67, 69] directly predict the starting/ending time points of the target moments without ranking massive proposal candidates. Thus, regression-based methods are much faster than proposal-based methods, which is one reason why our work focuses on the regression-based TVG. Furthermore, reinforcement learning (RL)-based methods formulate the TVG task as a sequence of decisions to make [18, 60]. In particular, they train an agent to control the movement of a window by shifting or scaling. During training, the agent would be rewarded or punished based on whether the window is close to the target moment after an adjustment. Temporal Action Detection (TAD). TAD aims to determine whether predefined actions occur in a video and to predict the corresponding time intervals during which these actions occur [12,13,48,53,56,59,63]. Different from TVG, the input of TAD is only a video. In other words, TAD only requires a semantic understanding of videos. Compared to TAD, TVG is more challenging since it requires a semantic understanding of both videos and natural languages. Furthermore, TVG needs to process the multimodal interaction between videos and natural languages. Text Prompting. Prompting has recently achieved great success in the domain of natural language processing [25, 32, 37, 40, 46, 49\u201352, 58]. Text prompting is a process that leverages a data-agnostic perturbation operation applied to text inputs or their embeddings to improve the performance of the downstream task. The simplest way is to construct an input context template originating from human contemplation [46,49\u201351]. Although the manually-crafted context templates are simple and interpretable, they are typically not the optimal input prompts. To tackle this issue, other work has focused on searching the optimal prompting in the discrete input space [25, 52, 58] or in the language model\u2019s embedding space [32,37,40]. Visual Prompting. Inspired by the idea of prompt learning in NLP [37], visual prompting (VP) was first proposed by Bahng et. al. [2] to reprogram a source vision model (e.g., ImageNet-pretrained classifier) to accomplish downstream target tasks (e.g., CIFAR-10 image classification). VP shares almost the same idea with the model reprogramming technology in the vision domain [4\u20136, 11, 57, 65, 72, 81], which incorporates a universal input perturbation into testing data so as to improve a desired performance metric, e.g., target task accuracy, robustness, and fairness. Multi-Modal Prompting. Although visual prompting and text prompting have recently attracted much attention, they are under-explored in the multi-modal learning, especially on the temporal video grounding task. The existing works [2, 27, 66] mainly focus on integrating text and visual prompts with the CLIP (Contrastive Language\u2013Image Pretrained) model to improve downstream tasks with imagery data. The problem of multi-modal prompting in the video understanding task has not been studied. In this paper, we for the first time develop the text-visual prompting technique to improve the performance of temporal video grounding using 2D visual features. 3. Methods In this section, we begin with the problem formulation of regression-based TVG. Then we demonstrate the design of TVP (text-visual prompts) and present the overview of our proposed TVP framework. 3.1. Problem Definition Let v \u2208RNvid\u00d7C\u00d7H\u00d7W be an untrimmed video consisting of a sequence of Nvid video frames, and s \u2208RNtex be a text query consisting of a sequence of Ntex language tokens. Here, the video-query pair (v, s) belongs to a videolanguage dataset D. Given v and s, TVG aims to predict the time interval \u02c6 T = (\u02c6 tsta, \u02c6 tend) of the target video moments described by the query s. The TVG model that fuses the vision-language modalities can be described as: \u02c6 T = f( gtex(s), gvid(v) ), (1) where f denotes TVG model, and gvid and gtex represent vision encoder and language encoder, respectively. 3.2. TDIoU Loss Function Conventionally, the TVG model can be learned by minimizing the temporal IoU loss LtIoU defined below: LtIoU =   1 \u2212 \u02c6 T(\u03b8) T T \u02c6 T(\u03b8) S T ! , (2) where for ease of notation let \u03b8 denote all the trainable parameters involved in (1), and T = (tsta, tend) is the label (i.e., the ground-truth time interval) of the target moment associated with the input video-query pair (v, s). The rationale behind (2) is to maximize the overlapping between the predicted time interval and its ground truth. However, for non-overlapping cases, the temporal IoU loss LtIoU would encounter a gradient vanishing problem. Inspired by [82], we develop a novel TDIoU (TemporalDistance IoU) loss for training our proposed TVG models by incorporating the normalized central time point distance and duration difference between the predicted video clips and the target video clips. We elaborate on the proposed loss below. \f\u201cPerson  pours water  into a glass.\u201d Text Query Video 2D  CNN Textual Features Predicted Starting  Point Predicted Ending  Point 6.5s 18.1s Predicted Moments Frame-aware Visual Prompts Text  Prompts 2D Position Embedding Position Embedding Transformer Type Embedding Visual  Features Language Encoder Max Pooling Mean Pooling MLP Video Frame Preprocessing Feature Extraction Multi-modal Feature Processing Cross-modal Fusion Figure 3. Overview of our proposed TVP (text-visual prompting) framework for 2D TVG (temporal video grounding). The whole process contains four phases: \u2776Video frame preprocessing: uniformly sample frames from input video and apply a set of frame-aware visual prompts to the sampled frames in order; \u2777Feature extraction: 2D CNN extracts features from sampled video frames with visual prompts, and the language encoder extracts textual features. In addition, the visual features would be spatially downsampled and temporally fused by max pooling and mean pooling, respectively. \u2778Multimodal feature processing: after spatial downsampling and temporal fusion, the 2D visual features would be integrated into the prompted textual features. \u2779Crossmodal fusion: the multimodal features would be processed by a 12-layer transformer encoder, and MLP would predict the starting/ending time points of the target moment. Distance Loss Ldis. To avoid the gradient vanishing problem caused by the non-overlapping case, we involve distance loss Ldis to directly minimize the normalized central time point distance. In addition, we add a threshold \u03b11 to prevent oscillation in the later training phase. The distance loss is then given by: Ldis = max   | (tsta + tend) /2 \u2212 \u0000\u02c6 tsta + \u02c6 tend \u0001 /2| |\u02c6 T S T| , \u03b11 ! , (3) where recall that T = (tsta, tend), \u02c6 T is predicted by the TVG model (1), and we choose \u03b11 = 0.2 in experiments. Duration Loss Ldur. The introduction of distance loss Ldis avoids the gradient vanishing problem but only considers the central time point distance. Yet, this may not be precise enough. For example, even if the central time points are completely overlapped, the duration of two video clips may not be identical. Inspired by the above, we propose the duration loss: Ldur = max   |T \u2212\u02c6 T(\u03b8)| |T| , \u03b12 ! , (4) where \u03b12 is the precision tolerance threshold and set by 0.4 in our experiments. Finally, the proposed Temporal-Distance IoU (TDIoU) loss is given by L = LtIoU + \u03b21Ldis + \u03b22Ldur, (5) where \u03b21 > 0 and \u03b22 > 0 are regularization parameters. 3.3. Text-Visual Prompt Design Inspired by the application of prompts on transformers [2,21,36,37], we propose jointly text-visual prompting to boost the performance of our models, in which prompts are optimized perturbation patterns. To improve data processing efficiency, we uniformly sample video frames from the untrimmed video v to obtain vsam \u2208RNsam\u00d7C\u00d7H\u00d7W , where Nsam is the number of sampled video frames. In addition, we introduce a set of frame-aware visual prompts \u03b4vp \u2208RNsam\u00d7dvp in the pixel space of sampled video frames vsam, and introduce text prompts \u03b4tp \u2208RNtp\u00d7dtp in the textual feature space. By incorporating video frame sampling and text-visual prompts into the TVG model (1), we obtain: (\u02c6 tsta, \u02c6 tend) = f( \u03b4tp, gtex(s), gvid(vsam + \u03b4vp) ). (6) Given a pre-trained 2D TVG model f, the objective of text-visual prompting (TVP) is to learn a universal set of visual prompts \u03b4vp and text prompts \u03b4tp to be integrated into sampled video frames and textual features, respectively. Specially, a set of different visual prompts are applied to uniformly-sampled frames of one untrimmed video in order. During training, only the set of visual prompts and text prompts are updated through backpropagation. During finetuning, prompts are frozen, and the parameters of the TVG model and encoders are updated. During testing, the set of optimized visual prompts and the optimized text prompts are applied to all test-time video-query pairs. \f3.4. Framework Inspired by the success of transformers in visionlanguage tasks, we choose ClipBERT [31] as the base model for 2D TVG. Extended from ClipBERT, the input of our regression-based TVG model would be describable sentences and uniformly sampled frames of one untrimmed video as shown in Fig. 3. Then, the predicted starting and ending time points of the target video clip would be model outputs. As described in Algorithm 1, there are four phases of our proposed TVP framework: \u2776 Video frame preprocessing: We obtain sparsely-sampled frames vsam from one input untrimmed video v, and apply universal frame-aware visual prompts \u03b4vp on top of frames at the padding location. \u2777Feature extraction: 2D vision encoder (first 5 ConvBlock of ResNet-50) gvid and language encoder (a trainable word embedding layer) gtex would extract features from the prompted frames v\u2032 sam and textual inputs s, respectively. \u2778Multimodal feature processing: Following the setting of Pixel-BERT [22], the 2D visual features Qvid are downsampled spatially by a 2 \u00d7 2 maxpooling layer and fused temporally by a mean-pooling layer. Then, text prompts \u03b4tp are integrated into textual features Qtex. In addition, trainable 2D visual position embeddings M2D and textual position embeddings Mpos are applied to the processed 2D visual features Q\u2032 vid and prompted textual features Q\u2032 tex, respectively [10, 31]. Afterwards, the processed and position-encoded 2D visual features Q\u2032\u2032 vid are flattened and integrated into prompted and positionencoded textual features Q\u2032\u2032 tex. Moreover, type embeddings Mtype would be added to the integrated multimodal features Qall to indicate the source type of features. \u2779 Crossmodal fusion: A 12-layer transformer [10] is utilized for crossmodal fusion on Qall, and then multilayer perceptron (MLP) ending with sigmoid function is used as the prediction head to process the last-layer crossmodal representation QCM of the transformer for generating the predicted starting/ending time points (\u02c6 tsta, \u02c6 tsta) of the target moments described by the text query input. 4. Experiments In this section, we demonstrate the effectiveness of our proposed TVP framework on Charades-STA and ActivityNet Captions datasets. 4.1. Experiment Setup Datasets. The evaluations are implemented on two standard benchmark datasets for TVG task, Charades-STA [14] and ActivityNet Captions [28]. Tab. 1 summarizes the details of both datasets. Charades-STA dataset contains 6, 672 videos and 16, 124 text queries in total. The average length of videos is 30.6s, and the average length of text query is 7.2 words . The average length of moments corAlgorithm 1 Overview of TVP framework Input: vision encoder gvid, language encoder gtex, position embeddings Mpos, 2D position embeddings M2D, type embeddings Mtype, transformer f, prediction head MLP, visual prompts \u03b4vp, text prompts \u03b4tp Output: Predicted time interval \u02c6 T = (\u02c6 tsta, \u02c6 tend) Phase \u2776: Video frame preprocessing 1: vsam \u2190uniformly sample video frames from an untrimmed video v 2: v\u2032 sam \u2190apply visual prompts \u03b4vp to the sampled video frames vsam Phase \u2777: Feature Extraction 3: Qvid = gvid(v\u2032 sam) \u2190extracting 2D visual features 4: Qtex = gtex(s) \u2190extracting textual features Phase \u2778: Multimodal feature processing 5: Q\u2032 vid \u2190apply spatial downsampling and temporal fusion to 2D visual features Qvid 6: Q\u2032 tex \u2190apply text prompts \u03b4tp to textual features Qtex 7: Q\u2032\u2032 vid \u2190add 2D visual position embeddings M2D on the processed 2D visual features Q\u2032 vid 8: Q\u2032\u2032 tex \u2190add position embeddings Mpos to prompted textual features Q\u2032 tex 9: Qall \u2190integrate the processed and position-encoded textual features Q\u2032\u2032 tex and the processed and positionencoded 2D visual features Q\u2032\u2032 vid 10: Qall + Mtype \u2190add type embeddings Mtype to the integrated multimodal features Qall Phase \u2779: Crossmodal fusion 11: QCM = f(Qall + Mtype) \u2190implement crossmodal fusion through transformer f 12: (\u02c6 tsta, \u02c6 tend) = MLP(QCM) \u2190prediction head generates the predicted time interval according to crossmodal representation QCM Table 1. Statistics of TVG benchmark datasets (Charades-STA and ActivityNet Captions datasets). Dataset Charades-STA ActivityNet Captions Domain Indoor Activity Indoor/Outdoor Activity # Videos 6, 672 14, 926 Avg. Video Length (second) 30.6 117.6 # Moments 11, 767 71, 953 Avg. Moment Length (second) 8.1 37.1 Vocabulary Size 1, 303 15, 505 # Queries 16, 124 71, 953 Avg. Query Length (word) 7.2 14.4 responding to the text query is 8.1s. Following the same dataset split as [14] for fair comparisons, there are 12, 408 video-query pairs for training and 3, 720 pairs for testing. ActivityNet Captions dataset contains 14, 926 videos and 71, 953 text queries in total. The average length of videos is 117.6s, and the average length of text query is 14.4 words. \fThe average length of moments corresponding to the text query is 37.1s. ActivityNet Captions dataset is split into training set, validation set, and testing set in a 2 : 1 : 1 ratio. Since the testing set is withheld for competition, only a training set and two validation sets (val1 and val2) can be accessed publicly. For fair comparisons, we evaluate our proposed framework on val1. Baselines. We compare our proposal with 15 baseline methods: \u2460Proposal-based: CTRL [14], MCN [1], SAP [7], BPNet [62], LPNet [61], QSPN [64], MAN [71]; \u2461Proposal-free: ABLR [67], DRN [69], CPNet [34], DEBUG [43], ExCL [16], VSLNet [73]; \u2462Reinforcement learning: TSP-PRL [60], TripNet [18]. Evaluation metrics. Following [14], we adopt Acc(R@1, IoU=m) as the performance evaluation metric, which represents the percentage accuracy of top-1 predicted moments whose tIoU (temporal IoU) with the ground-truth moment is larger than m. By convention, we consider the following tIoU threshold values m = {0.3, 0.5, 0.7}. Crossmodal pretraining setup. Our 2D vision encoder (ResNet-50) is initialized with the weight from gridfeat [24], which can extract effective grid features from visual inputs. In addition, both the language encoder and 12layer transformer are initialized with the BERT-base model weight [10], which are pretrained on English Wikipedia and BookCorpus [83]. Thanks to the compact 2D vision encoder, TVP (our proposal) is able to directly utilize image-text pairs for end-to-end training. Since the benefits of cross-modal pretraining has been demonstrated by [22, 44, 55], our base model is pretrained on two largescale image-text datasets, which are Visual Genome Captions [29] and COCO Captions [8]. To be more specific, image-text matching [44, 55] and masked language modeling [10] are employed for cross-modal pretraining. Implementation setup. For video inputs, we uniformly sample Nsam frames from a video (Nsam = 48 for Charades-STA and Nsam = 64 for ActivityNet Captions). In addition, all video frames are resized to have a maximum longer side of 448 with an original aspect ratio, and then the frames are zero-padded to 448 \u00d7 448. The default visual prompt sizes for both dataset are 96. The default text prompt sizes are 10 and 20 for Charades-STA and ActivityNet Captions, respectively. We utilize the first 5 ConvBlocks of ResNet-50 as the 2D vision encoder and a trainable embedding layer as the language encoder for both Charades-STA and ActivityNet Captions datasets. For text queries, all word tokens are maintained after lower-case conversion and tokenization. We use AdamW [42] for endto-end model training, with \u03b21 = 1.0, \u03b22 = 0.1, \u03b11 = 0.2, \u03b12 = 0.4. Initial learning rates are 1e \u22121 and 5e \u22127 for prompt training and model finetuning, respectively. In addition, the learning rate linearly decays to 0 with the first 10% training step for warmup. Our experiments are impleTable 2. Performance comparison of different thresholds m on the Charades-STA dataset. Type Method Visual Acc(R@1, IoU=m) Feature m=0.3 m=0.5 m=0.7 3D TVG CTRL [14] C3D 23.63 8.89 ABLR [67] C3D 24.36 9.01 BPNet [62] C3D 55.46 38.25 20.51 LPNet [61] C3D 59.14 40.94 21.13 QSPN [64] C3D 54.70 35.60 15.80 TSP-PRL [60] C3D 45.45 24.75 TripNet [18] C3D 54.64 38.29 16.07 DRN [69] C3D 45.40 26.40 CPNet [34] C3D 40.32 22.47 DEBUG [43] C3D 54.95 37.39 17.92 ExCL [16] I3D 61.50 44.1 22.40 VSLNet [73] I3D 64.30 47.31 30.19 MAN [71] I3D 46.53 22.72 2D TVG MCN [1] VGG 17.46 8.01 SAP [7] VGG 27.42 13.36 Ours TVP-Based 2D TVG Base w/o prompts ResNet 61.29 40.43 19.89 Base + Visual Prompts 65.38 44.31 20.22 Base + Text Prompts 65.81 43.44 20.65 Base + Both Prompts 65.92 44.39 21.51 Table 3. Performance comparison of different thresholds m on the ActivityNet Captions dataset. Type Method Visual Acc(R@1, IoU=m) Feature m=0.3 m=0.5 m=0.7 3D TVG CTRL [14] C3D 28.70 14.00 BPNet [62] C3D 59.98 42.07 24.69 LPNet [61] C3D 64.29 45.92 25.39 QSPN [64] C3D 45.30 27.70 13.60 TSP-PRL [60] C3D 56.02 38.83 TripNet [18] C3D 48.42 32.19 13.93 DRN [69] C3D 45.45 24.36 CPNet [34] C3D 40.56 21.63 ABLR [67] C3D 55.67 36.79 DEBUG [43] C3D 55.91 39.72 ExCL [16] C3D 63.00 43.60 24.10 VSLNet [73] C3D 63.16 43.22 26.16 Ours TVP-Based 2D TVG Base w/o prompts ResNet 57.20 40.16 19.14 Base + Visual Prompts 60.12 43.39 23.71 Base + Text Prompts 60.48 42.58 24.39 Base + Both Prompts 60.71 43.44 25.03 mented in PyTorch [45], and models and prompts are finetuned separately for 12 epochs with the mixed precision on 8 NVIDIA V100 GPUs. 4.2. Experiment Results Effectiveness of TVP on Charades-STA. The performance comparisons with SOTA methods on the CharadesSTA dataset are summarized in Tab. 2. Our proposed TVP framework can achieve competitive performance at all tIoU thresholds m in the case of utilizing 2D visual features extracted by ResNet-50, and reach the highest score at m = 0.3. Compared to the 2D TVG methods using VGG as the vision encoder, our proposed framework could achieve around 2.5\u00d7 and 2.7\u00d7 performance gain at thresholds 0.5 and 0.7, respectively. Furthermore, we can find that for our base model only one of visual prompts and text prompts can achieve up to 7.37% and 9.60% improvement \f8 16 32 48 64 Sampled Frame Number 20 30 40 50 60 R@1, IoU=m m = 0.3 m = 0.5 m = 0.7 (a) Charades-STA 8 16 32 48 64 Sampled Frame Number 20 30 40 50 60 R@1, IoU=m m = 0.3 m = 0.5 m = 0.7 (b) ActivityNet Captions Figure 4. Impact of sampled frame numbers. at tIoU thresholds m = 0.3 and m = 0.5. The combination of text and visual prompts can not only achieves 7.55% and 9.79% improvements at tIoU thresholds m = 0.3 and m = 0.5, but also improve the performance by 8.14% at m = 0.7. This demonstrates the effectiveness and necessity of the joint text-visual prompting. Effectiveness of TVP on ActivityNet Captions. We focus on the performance comparisons with 3D TVG methods on ActivityNet since there are no results of 2D TVG method reported on ActivityNet Captions. The results of multiple methods on ActivityNet Captions datasets are reported in Tab. 3. Even on this more challenging dataset, our proposed method still has achieved competitive performance compared to 3D TVG methods. Different from the performance of TVP on Charades-STA dataset, text prompts or visual prompts can achieve a significant performance boost on the base model over all IoU thresholds m alone (5.73% at m = 0.3, 8.04% at m = 0.5, 27.43% at m = 0.7 ) , and the text-visual prompt combination could further boost the performance (6.14% at m = 0.3, 8.17% at m = 0.5, 30.77% at m = 0.7). It is worth noting that the performance gap over m = 0.7 between 2D TVG methods and 3D TVG methods is narrowed significantly. In summary, through the experimental results on Charades-STA and ActivityNet Captions datasets, we can find that our proposed TVP framework could achieve competitive performance overall tIoU thresholds on CharadesSTA and ActivityNet Captions by improving the utility of sparse 2D visual features. Thanks to the lightweight 2D vision encoder, cotraining language encoder and vision encoder on large-scale image-text datasets can be performed, which benefits the base model to achieve good performance. Furthermore, the combination of text and visual prompts can achieve better results than any single kind of prompts on both datasets, which again proves the importance of crossmodal training. Video frame sampling effect. Fig. 4 demonstrates the performance of base model with different number Nsam Table 4. The performance comparison of different visual prompt sizes on Charades-STA dataset. Visual Prompt Size Acc(R@1, IoU=m) Prompt + Frame m=0.3 m=0.5 m=0.7 0 61.29 40.43 19.89 Video Frame (without visual prompts) Pad Size p = 16 Pad Size p = 32 Pad Size p = 48 Pad Size p = 72 Pad Size p = 96 16 61.29 40.43 20.00 Video Frame (without visual prompts) Pad Size p = 16 Pad Size p = 32 Pad Size p = 48 Pad Size p = 72 Pad Size p = 96 Pad Size p = 128 p p 32 61.94 39.78 19.35 Video Frame (without visual prompts) Pad Size p = 16 Pad Size p = 32 Pad Size p = 48 Pad Size p = 72 Pad Size p = 96 Pad Size p = 128 p p 48 63.66 42.37 20.00 Video Frame (without visual prompts) Pad Size p = 16 Pad Size p = 32 Pad Size p = 48 Pad Size p = 72 Pad Size p = 96 Pad Size p = 128 p p 72 63.87 43.66 19.78 Video Frame (without visual prompts) Pad Size p = 16 Pad Size p = 32 Pad Size p = 48 Pad Size p = 72 Pad Size p = 96 Pad Size p = 128 p p 96 65.38 44.31 20.22 Video Frame (without visual prompts) Pad Size p = 16 Pad Size p = 32 Pad Size p = 48 Pad Size p = 72 Pad Size p = 96 Pad Size p = 128 p p 128 64.73 43.66 19.78 Video Frame (without visual prompts) Pad Size p = 16 Pad Size p = 32 Pad Size p = 48 Pad Size p = 72 Pad Size p = 96 Pad Size p = 128 Table 5. The performance comparison of different text prompt sizes on Charades-STA dataset. Text Prompt Size Acc(R@1, IoU=m) m=0.3 m=0.5 m=0.7 0 57.20 40.16 19.14 5 65.38 41.94 20.43 10 65.81 43.44 20.65 15 65.59 43.23 21.29 20 64.95 43.87 21.51 25 63.66 42.80 20.65 30 64.46 42.63 20.51 of sampled video frames as visual inputs. For Charades dataset, the base model performance keeps increasing before Nsam reaches 48, but when it exceeds 48, performance starts to degrade. This is because frequent background changes harm the performance of object re-identification in videos, which are noisy for object motion analysis [17]. For ActivityNet Caption dataset, base model performance continues to improve even when sampled frame number Nsam exceeds 48, due to the longer average video length in ActivityNet Captions dataset. Balancing the frame number and batch size for training, we choose Nsam = 64 for ActivityNet Captions. TVP performance vs. prompt size. As shown in Tab. 4, we can find that when visual prompts are small, they cannot bring changes to the base model, and when visual prompts are too large, the performance starts to decrease. This is because key information within video frames might be removed. However, the text prompts can bring significant performance boost even when the text prompt size is small as shown in Tab. 5, which is because the textual features has a smaller dimension compared to visual features, and also the text prompts are directly optimized in feature space during training. TVP performance vs. visual prompt operation. Visual prompt is first proposed by [2], where visual prompts are \fTable 6. The performance comparison of different visual prompt operations (\u2018remove\u2019, \u2018add\u2019, \u2018replace\u2019) with fixed visual prompt size p = 96 on Charades-STA and ActivityNet Captions datasets. Operation Charades-STA ActivityNet Captions R@1, IoU=m R@1, IoU=m m=0.3 m=0.5 m=0.7 m=0.3 m=0.5 m=0.7 Original 61.29 40.43 19.89 57.20 40.16 19.14 Remove 61.29 40.43 20.0 57.20 40.16 19.14 Add 61.08 39.57 20.22 57.15 40.16 19.27 Replace 65.38 44.31 20.22 60.12 43.39 23.71 Video Length (s) 10 30 50 70 90 110 500 1000 1500 2000 2500 ResNet-50 C3D Inference Time (ms) 10 20 30 40 50 60 70 90 110 120 100 80 Video Length (s) Inference Time (ms) 100 200 300 400 Other Modules ResNet-50 (a) Vison Encoders (b) Modules in 2D TVG Model   Figure 5. Inference time comparison. (a) inference time comparison between 2D vision encoder (ResNet-50) and 3D vision encoder (C3D). (b)inference time comparison between the vision encoder and the other modules of the 2D TVG model, where the sampled frame number for our TVP framework is 1.2\u00d7 the length of the video in seconds. added to the image for transfer learning on classification tasks. In contrast, our proposed prompting framework is designed to compensate for the spatiotemporal information loss in 2D visual features. Due to the differences in the task, we try two different prompt operation strategies, \u2018replace\u2019 and \u2018add\u2019. \u2018add\u2019 is to add the visual prompts to the pixel value of the video frame at the corresponding padding locations. \u2018replace\u2019 is to replace the pixel values of video frames with visual prompts at corresponding padding locations. \u2018remove\u2019 is in order to study the impact of removing the pixel values at the padding location. As shown in Tab. 6, \u2018add\u2019 or \u2018remove\u2019 prompt operations have limited effects on the base model. However, \u2018replace\u2019 does boost the base model performance. TVP achieves inference efficiency. As shown in Fig. 5, we can find that the inference time required for visual feature extraction accounts for more than half of the inference time of the whole model, while the inference time required for the 3D vision encoder is more than 5\u00d7 compared to the 2D vision encoder, and even more than the time required for the whole TVG model using 2D vision encoder, which fully demonstrates the feasibility of accelerating the overall inference speed by reducing the complexity of the vision encoder. Need to note that if there are Table 7. The performance comparison of different loss designs on Charades-STA dataset. Loss Function Selection R@1, IoU=m m=0.3 m=0.5 m=0.7 LtIoU 55.05 29.89 11.82 LtIoU + Ldis 60.64 31.18 16.77 LtIoU + Ldur 59.78 30.97 16.34 LtIoU + Ldis + Ldur 61.29 40.43 19.89 0.4 0.2 0.0 0.2 0.4 x 0.4 0.2 0.0 0.2 0.4 y 0.64 0.66 0.68 0.70 0.70 0.70 0.72 0.72 0.74 0.76 0.76 0.78 0.78 0.78 0.78 0.80 0.80 0.80 0.80 0.80 0.80 0.82 0.82 0.82 0.82 0.82 0.84 0.84 0.84 0.86 0.86 0.88 0.88 0.4 0.2 0.0 0.2 0.4 x 0.4 0.2 0.0 0.2 0.4 y 0.64 0.66 0.68 0.70 0.70 0.70 0.70 0.72 0.72 0.72 0.72 0.72 0.72 0.74 0.74 0.74 0.74 0.76 0.76 0.76 0.76 0.76 0.76 0.78 0.80 0.80 0.82 0.82 Figure 6. Loss landscape visualization in 2D plane: Finetuning w/o prompts (left) and using prompts (right); see [33] for implementation. multiple model weights for different sampled frame number settings and model weights can be adopted adaptively for different lengths of videos, the inference speed for short videos should increase, and the prediction results for long videos will be further improved. Ablation studies. Through Tab. 7, we can find that the addition of either distance loss Ldis or duration loss Ldur will result in a performance increase, but the combination of the two will result in a significant performance increase (11.34% at m = 0.3, 35.26% at m = 0.5, 68.27% at m = 0.7, ), especially over tIoU thresholds m = 0.5 and m = 0.7. This demonstrates that distance loss Ldis and duration loss Ldur could provide more precise training guides compared to only using temporal IoU loss LtIoU. Furthermore, we posit that prompting may encode additional spatial-temporal supervision to help the model trainer to escape from bad local optima as shown in Fig. 6, where finetuning w/ prompts yields a flatter loss landscape than the one w/o prompts. 5."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.08578v2",
+                    "title": "Data-Model-Circuit Tri-Design for Ultra-Light Video Intelligence on Edge Devices",
+                    "abstract": "In this paper, we propose a data-model-hardware tri-design framework for\nhigh-throughput, low-cost, and high-accuracy multi-object tracking (MOT) on\nHigh-Definition (HD) video stream. First, to enable ultra-light video\nintelligence, we propose temporal frame-filtering and spatial saliency-focusing\napproaches to reduce the complexity of massive video data. Second, we exploit\nstructure-aware weight sparsity to design a hardware-friendly model compression\nmethod. Third, assisted with data and model complexity reduction, we propose a\nsparsity-aware, scalable, and low-power accelerator design, aiming to deliver\nreal-time performance with high energy efficiency. Different from existing\nworks, we make a solid step towards the synergized software/hardware\nco-optimization for realistic MOT model implementation. Compared to the\nstate-of-the-art MOT baseline, our tri-design approach can achieve 12.5x\nlatency reduction, 20.9x effective frame rate improvement, 5.83x lower power,\nand 9.78x better energy efficiency, without much accuracy drop.",
+                    "authors": "Yimeng Zhang, Akshay Karkal Kamath, Qiucheng Wu, Zhiwen Fan, Wuyang Chen, Zhangyang Wang, Shiyu Chang, Sijia Liu, Cong Hao",
+                    "published": "2022-10-16",
+                    "updated": "2022-10-18",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AR"
+                    ],
+                    "main_content": "INTRODUCTION The past decade has witnessed a tremendous success of deep neural networks (DNNs) [2]. In this paper, we focus on DNN-based multiobject tracking (MOT), which has been used as a technological basis for video intelligence and is drastically improving the quality of human life, e.g., autonomous driving (AD) [23]. Despite a proliferation of MOT techniques [10], little research effort has been made towards the data-efficient and co-design solutions for MOT across \u2217Authors contributed equally to this research. Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s). ASPDAC \u201923, January 16\u201319, 2023, Tokyo, Japan \u00a9 2023 Copyright held by the owner/author(s). ACM ISBN 978-1-4503-9783-4/23/01. https://doi.org/10.1145/3566097.3567864 the full software/hardware stack. Our work aims to close this gap by developing a high-throughput, low-cost, and high-accuracy video processing algorithm/hardware pipeline. We provide a holistic viewpoint on how contemporary MOT methods can be renovated to deal with massive quantities of data and high-complexity DNNs. Despite extensive existing works, data and model efficiencies of MOT are mainly studied either from the algorithm perspective or from the model design perspective. For example, in [19], the MOT efficiency is improved by leveraging meta learning [5] to solve an instance detection problem. In addition to algorithms, several works have focused on developing model compression techniques (such as weight pruning and weight quantization) to reduce the model complexity of MOT [14, 18]. However, the efficiency of \u201ccompressed\u201d models are evaluated without considering the practical hardware platform, such as low-power FPGAs. Known FPGA acceleration techniques [8, 25] do not explore data sparsity or model sparsity, thereby missing huge optimization opportunities. Additionally, existing accelerators are evaluated on the ImageNet dataset with small input image sizes and do not scale to real-world High-Definition (HD) video frames, limited by the device and/or design tool capabilities. To the best of our knowledge, there is no prior work that discusses efficient implementation of MOT on the edge for HD video processing by fully utilizing dataand model-level sparsity. Contributions. To close the gap between MOT algorithm design and efficient implementation, we propose a data-model-hardware tri-design approach. We summarize our contributions as follows. \u2022 Dynamic frame filtering. We propose a reinforcement learningbased lightweight algorithm to achieve temporal data reduction. \u2022 Spatial attention focusing. We propose a saliency-guided spatial data reduction method to eliminate uninformative pixels from both the input frames and the intermediate feature maps. \u2022 Hardware-aware model compression. We leverage kernelwise pattern-aware sparsity of an MOT model to achieve hardwarefriendly model compression. \u2022 Tri-design implementation and evaluation. We implement the proposed tri-design framework on a hardware platform comprising of one Xilinx ZCU104 and two Xilinx Alveo U50 FPGAs and conduct extensive experiments to evaluate its effectiveness in both accuracy and on-device efficiency metrics. We demonstrate that our approach can achieve 12.5\u00d7 latency reduction, 20.9\u00d7 effective frame rate improvement, 5.83\u00d7 lower power, and 9.78\u00d7 better energy efficiency, without much accuracy drop compared to the state-of-the-art MOT baseline [16] on the BDD100K dataset [22]. arXiv:2210.08578v2  [cs.CV]  18 Oct 2022 \fASPDAC \u201923, January 16\u201319, 2023, Tokyo, Japan Zhang et al. 2 PRELIMINARIES AND MOTIVATION In this paper, we focus on MOT (multiple-object tracking) which deals with massive video data. Current MOT methods are commonly built upon DNNs and follow the paradigm of tracking-by-detection [20]. QDTrack [16] is the state-of-the-art MOT algorithm introduced on the BDD100K dataset [22], one of the most representative and challenging self-driving car datasets. QDTrack employs Faster R-CNN [17] with Feature Pyramid Network (FPN) [12] as an object detection backbone network, leverages Contrastive Learning [21] to optimize the backbone network parameters, and utilizes Bi-directional Softmax [16] in the embedding space for object association and tracking. The focus of this work is not to develop new MOT model algorithms but to instead build a high-throughput, lowcost, and high-accuracy video processing implementation pipeline across the full data-processing/DNN operation/hardware stack. The computational landscape of MOT has been rapidly evolving on dedicated computing hardware (e.g., FPGAs and ASICs) [24]. Despite the advancements of hardware accelerators [11] and software/hardware co-design techniques [8], critical limitations still exist for real-time video processing. First, there is a sharp increase not only in ML model size but also computation complexity, especially for HD video frames. With HD inputs, any two 3 \u00d7 3 convolution layers of ResNet-50 already have higher complexity than the entire ResNet-50 with ImageNet inputs. This asserts huge pressure for real-time inference on hardware. Second, both the hardware devices and design tools exhibit poor scalability. For example, the most powerful FPGA till date, Xilinx Alveo U55C, has 9,024 DSP slices. It can provide peak performance of 2,256 GOPs for QDTrack (157.9 GOPs), but the performance upper bound is 70 ms per frame, which is far from real-time requirement. Meanwhile, the design tools are strictly limited by the largest parallelism that can be supported. For example, Xilinx Vitis HLS tool constrains that the parallelism factor should be lower than 4096 [1], while Intel Design Toolkit does not support more than 256 concurrent kernels. With larger parallelism, hardware designs can easily take days or even weeks to synthesize, which drastically increases the development cycle and the time-tomarket (TTM). Consequently, there is still a huge gap between the low-power real-time video processing requirements and the limited capability of existing hardware acceleration techniques. As inspired by above, our work advances the design of practical MOT systems by tackling the following challenges (C1\u2013C3). (C1): The massive video data and their high frame rates make the implementation of MOT prohibitively expensive in energy and latency. Thus, our first goal is to develop novel data complexity reduction techniques that can fully exploit the temporal and saliency redundancy of video frames over time and space. (C2): SOTA DNN-based MOT models typically contain gigantic number of weights making them poor candidates for hardware implementation. To this end, we propose a hardware-friendly model compression technique for MOT. (C3): A holistic data-model-hardware full-stack design approach is lacking. It remains elusive how current MOT solutions can be applied subject to Small Size, Weight, and Power (SWaP) constraints in practice. To the best of our knowledge, there is no prior work that addresses the challenges (C1\u2013C3) in a single unified MOT implementation pipeline. Map to Temporal Frame Filtering  \u2022 Reinforcementlearning based policy Spatial Dimensionality  Detection  \u2022 Complexity-score  based detection Map to Reduced  Frames Reduced  Regions Software Design 1 2 3 Hardware  Design Pruning and Compressing  \u2022 Kernel-Wise Pruning   \u2022 Pattern-Wise Pruning 0 0 0 0 0 0 0 0 0 Kernel-wise Pruning 1 0 1 0 0 0 1 0 1 Pattern-wise Pruning Skip Frame Computation Tile-based Architecture Pruning-aware Parallelism Map to 4 4 4 Frame 1 backend Frame 3 RPN+FPN Frame 4 Backbone Skip  tiles \u00d7 Skip  channels Sobel Frame T Skip Frame T? Frame T-1 Frame T-2 Frame 2 dropped Figure 1: Overview of data-model-hardware tri-design. 3 METHODOLOGY 3.1 Overview of proposed approach. The proposed data-model-hardware tri-design includes four main components as shown in Fig. 1. 1 (Data) Temporal frame filtering. It is achieved using reinforcement learning (RL) to filter unnecessary frames at the front-end data processing stage and eventually reduce processing power along with latency costs on hardware. 2 (Data) Spatial saliency focusing. It provides the second-stage data complexity reduction by peering into the salient region of a frame to eliminate uninformative pixels. 3 (Model) Hardwarefriendly deep model pruning. It exploits the structured sparsity patterns of weights to reduce the model size on hardware without much loss of accuracy. 4 (Hardware) Backend acceleration. Guided by the algorithmic pipeline, it adopts low-power FPGAs to achieve realistic implementation of MOT with high throughout, low latency, and high energy efficiency. In a nutshell, components 1 and 2 are designed for tackling the data complexity challenge (in response to C1); Component 3 reduces the complexity of backend model (in response to C2); Component 4 unveils the practical efficiency of hardware solutions for MOT (in response to C3). 3.2 Data complexity reduction 3.2.1 RL-based temporal frame filtering. The embedded MOT solution requires a high throughput in terms of the number of frames that can be processed in unit time. The frame processing rate should also consider the practical latency cost. Considering the massive quantities of video frames, achieving a video processing rate of less than 30ms (for real-time inference) becomes implausible even on high-end FPGAs. Due to this hardware constraint, it is necessary to reduce the data complexity through the lens of temporal video frame dropping. Particularly, in the AD (autonomous driving) datasets, consecutive video frames may contain largely-overlapped frame content such as street and road scenes. Thus, temporal frame filter is not only necessary but also reasonable for real-time MOT. In our preliminary experiments, we test a random filtering policy with different dropping ratios. As shown in Fig. 7, dropping frames randomly significantly hampers the object tracking performance. The incapability of random filtering drives us to design a lightweight learning-based solution that can self-adjust the frame dropping scheme based on the task-end tracking performance. Inspired by above, we propose an RL-based policy network for temporal frame filtering. Our policy network takes the current frame and the difference between the current frame and the previous frame as inputs. At the beginning of the policy network, we adopt a lightweight 3-layer convolutional network to extract a 2D \fData-Model-Circuit Tri-Design for Ultra-Light Video Intelligence on Edge Devices ASPDAC \u201923, January 16\u201319, 2023, Tokyo, Japan feature map with the same height and width as the input frames. In compensation with a shallow network, we adopt 5\u00d75 kernels at each convolution layer. The numbers of filters of the three convolution layers are 16, 8, and 1. Finally, with a pooling layer and sigmoid layer, we obtain the probability of dropping the current frame, which indicates the salience of the current input frame. We optimize this policy model through a regularized policy gradient method. Specifically, given a video sequence v with \ud835\udc5bframes {\ud835\udc630, \ud835\udc631, ..., \ud835\udc63\ud835\udc5b}, the policy model \ud835\udc3agenerates a frame importance score \ud835\udc3a(\ud835\udc63\ud835\udc56) \u2208(0, 1) for each frame \ud835\udc63\ud835\udc56. \ud835\udc3a(\ud835\udc63\ud835\udc56) serves as the probability to retain the current frame and determines the new video sequence v\u2032 with \ud835\udc5b\u2032 frames that only consists of the selected video frames. The training objective function of \ud835\udc3ais given by \ud835\udc3d= \u2212\ud835\udc45(v\u2032) \ud835\udc5b \u2211\ufe01 \ud835\udc56=1 log\ud835\udc5d(\ud835\udc3a(\ud835\udc63\ud835\udc56)) + \ud835\udefc \ud835\udc5b \u2211\ufe01 \ud835\udc56=1 (\ud835\udc3a(\ud835\udc63\ud835\udc56) \u2212\ud835\udf07)2, (1) where \ud835\udc45(v\u2032) is the reward of the updated video sequence v\u2032 and \ud835\udc5d(\ud835\udc3a(\ud835\udc63\ud835\udc56)) is the probability to sample an individual frame \ud835\udc63\ud835\udc56. We use the number of ID Switches per frame (i.e., IDSw in Sec. 4.1) as the reward function to evaluate the quality of frame dropping: \ud835\udc45(v\u2032) = \u2212IDSw(v\u2032)/\ud835\udc5b\u2032, where \ud835\udc5b\u2032 is the number of remaining frames in the video sequence v\u2032. In (1), inspired by [26], we also regularize the salience scores \ud835\udc3a(\ud835\udc63\ud835\udc56) with a hyperparameter \ud835\udf07to avoid the case of no frame dropping and obtaining high reward for free. 3.2.2 Spatial saliency focusing. Based on the first-stage temporal filtering, we next propose the second-stage data reduction, which attentively focuses on the salient region of a frame and its feature map to eliminate uninformative pixels. Towards the goal of spatial reduction, our MOT model (i.e., QDTrack) achieves spatial saliency by detecting and characterizing important spatial signals while monitoring high-rate video streams with low latency. The output is a localized sub-image for each frame, on which the subsequent backend algorithm on FPGA can focus. We propose a spatial saliency reduction (SSR) method. First, it decomposes an input image into multiple patches, and then adopts the Sobel operator [9] to compute an approximate saliency score of each patch. This operator exploits two 3 \u00d7 3 convolution kernels on the given patch to calculate the saliency values (i.e., derivatives). Once the patch-wise saliency score is obtained, we smoothen it using its four neighboring patches and create a binary mask indicating the patches of an input frame that should be discarded. In addition to input saliency dropping, we further propose feature saliency dropping, where patches of intermediate features, i.e., activations, can also be filtered. Our motivations are two-fold. First, even if an input patch has been dropped, its corresponding feature region might still have responses \u201cleaked\u201d from neighborhood unfiltered areas, caused by filter stride, padding, etc. Second, some image regions might be recognized as redundant at the intermediate feature level rather than at the beginning input level. As shown in Fig. 2(e) and (g), even if a region is not dropped initially, it leads to low responses at later stages. Therefore, reducing spatial redundancy beyond the input level to feature level (Fig. 2(f) and (h)) is valuable and is also beneficial to hardware implementation. Feature map region dropping is implemented by interpolating the saliency mask to the spatial resolution of each layer. The integration of patch dropping and feature dropping further reduces model GOPs (in a hardware-friendly way) without compromising accuracy. Table 1: The performance of irregularly pruned model (QDtrack) vs. pruning ratio (0%, 80%, 85%, 90%). Here the sparse kernel ratio is given by the ratio of pruned 3 \u00d7 3 kernel weights over the total number of pruned weights. Metrics Dense Model Global pruning 0% 80% 85% 90% IDF1 (\u2191) 0.714 0.712 0.706 0.703 MOTA (\u2191) 0.637 0.631 0.627 0.624 Sparse kernel ratio 0% 36.59% 34.44% 32.53% 3.3 Model pruning for computation reduction The SOTA tracking model is overwhelmingly large. It has been long proven that model pruning is an efficient way to reduce model complexity. For effective and efficient hardware implementation, structured pruning is much more beneficial compared to unstructured pruning, as the latter introduces additional overhead such as sparse representation and format conversion. Furthermore, unstructured pruning hinders full utilization of parallel architectures [3]. Therefore, we advocate for and propose structured pruning to fully accommodate our hardware design. 3.3.1 Kernel-wise sparsity embedded in irregular weight pruning. The SOTA pruning method is iterative magnitude pruning (IMP) [7], which enables aggressive weight sparsity without loss of model accuracy, as justified by the lottery ticket hypothesis [6]. In Table 1, we apply IMP to the QDTrack model and indeed observe a competitive MOT accuracy, measured by IDF1 and MOTA, as illustrated in Sec. 4. However, the weight sparsity pattern achieved by IMP is unstructured, leading to the compatibility issue with hardware. Spurred by this fact, we examine if there exists a structured sparsity pattern hidden in the irregular weight sparsity achieved by IMP. As shown in Table 1, the sparse kernel ratio, i.e., the ratio of pruned 3 \u00d7 3 kernel weights over the total number of pruned weights takes over 30% in different weight pruning ratios (from 80% to 90%). This implies that kernel-wise sparsity is embedded in irregular weight pruning on QDTrack. 3.3.2 Integration with irregular pattern-aware pruning. We leverage the irregular pattern-aware pruning method [13] to enforce sparse weights that cannot be covered using kernel-wise sparse patterns. The rationale behind such irregular pattern-aware pruning Binary sparse pattern mask NN Pattern-aware convolution kernels Figure 3: Pattern pruning. is that the effective area of a convolution kernel often maintains specific sparse patterns even if it does not yield kernel-wise sparsity, i.e., kernels with all zero weights. Thus, we follow [13] to pre-define irregular sparse patterns [15] for 3 \u00d7 3 kernels, and leverage them to conduct irregular but pattern-aware weight pruning. Note that pruning using a fixed number of pre-defined sparse patterns facilitates efficient hardware implementation compared with the conventional irregular weight pruning. \fASPDAC \u201923, January 16\u201319, 2023, Tokyo, Japan Zhang et al. (a) Input frame (b) Saliency Mask (c) Masked Input (d) GT (e) FM(2,0) (f) FM(2,0) w/ Mask (g) FM(3,0) (h) FM(3,0) w/ Mask Figure 2: Visualizations of input frame patch dropping (first row) and feature map patch dropping (second row). (a): An input frame with Ground Truth (GT) object bounding box shown in (d). (b)-(c): An input saliency mask and the masked input frame, where the patch dropping ratio is set to 20%, and the discarded pixels are marked in black. (e)-(f): Feature maps in layer 0 of block 2 within backbone (ResNet-50), denoted by FM(2,0), without and with feature patch dropping mask, respectively. (g)-(h): Feature maps in layer 0 of block 3, denoted by FM(3,0), without and with feature patch dropping mask, respectively. We can observe that the proposed spatial data reduction approach could accurately identify the redundant patch regions associated with background scenes, e.g., sky and street. Together with kernel-wise pruning, we propose the following hardware-aware pruning method: \u2022 Call IMP to achieve the irregular weight sparse mask MIMP; \u2022 Extract kernel-wise sparsity mask M from MIMP; \u2022 Conduct irregular pattern-aware pruning on remaining weights identified by (1 \u2212M), then update M to obtain a new mask M\u2032; \u2022 Retrain non-zero model weighs under fixed pruning mask M\u2032. 3.4 Hardware implementation and acceleration 3.4.1 Patch dropping compatible tile-based accelerator design. Given HD input frames and the large ML model, the intermediate feature map for even a single layer is too large to be fully stored in the device\u2019s on-chip memory. Therefore, it is required that the input images and feature maps be partitioned into tiles and load/compute tile-by-tile, as shown in Fig. 4(a). Fortunately, incorporating our patch dropping strategy, a tile-based accelerator design is beneficial since it allows the accelerator to completely skip the loading and computing of a tile if it is dropped at an early stage; this not only reduces the computation latency but also significantly reduces energy consumption, since off-chip data movement is much more power consuming than on-chip computation [4]. 3.4.2 Pruning-aware and multi-level parallelism. To achieve extremely high parallelism and to simultaneously exploit structured pruning, we propose a novel multi-level pruning-aware parallelization architecture, as illustrated in Fig. 4. Consider one convolution layer as an example, where the input is a data tile of size\ud835\udc47\ud835\udc3b\u00d7\ud835\udc47\ud835\udc4a\u00d7\ud835\udc47\ud835\udc36 along with a weight kernel of size \ud835\udc3e\u00d7\ud835\udc3e\u00d7\ud835\udc47\ud835\udc36. The commonly adopted parallel computation is along the \ud835\udc47\ud835\udc36dimension, i.e., \ud835\udc47\ud835\udc36channels are computed simultaneously [8]. However, such a parallelization scheme fails to employ the structured and pattern-based pruning, since a pruned-away channel in a kernel still needs to be computed with other channels in parallel. Therefore, in this work, we explore the parallelism along \ud835\udc47\ud835\udc3band \ud835\udc47\ud835\udc4adimensions in two levels, and exploit the sparsity along \ud835\udc47\ud835\udc36dimension. First, row-level parallelism, as shown in Fig. 4(b), allows that at each clock cycle, one input row is computed with \ud835\udc3erows of a kernel, followed by self-accumulation for partial sum. Second, for column-level parallelism as shown in Fig. 4(c), within the same clock cycle, \ud835\udc47\ud835\udc4amultiplications are executed, followed by a \ud835\udc3e-to-1 adder tree. Third, since the row-level and column-level approaches have provided sufficient parallelism, we perform the convolution channel-by-channel sequentially without sacrificing computation latency. As shown in Fig. 4(d), if one channel of a filter is pruned, we skip the data loading and computation for that channel, reducing energy and latency almost linearly. Specifically, with structured pruning, if \ud835\udc5d% channels are pruned away within a filter, the latency linearly reduces by \ud835\udc5d%. Similarly, for pattern-based pruning, our proposed design uses a fixed sparsity pattern for the entire channel. 3.4.3 Dataflow architecture using multi-FPGA for scalability. To overcome the device and design tool scalability issue, as mentioned in Sec. 2, we propose a novel dataflow architecture using multiple FPGAs, as illustrated in Fig. 5. Instead of executing the entire model on a giant FPGA with extremely high parallelism, which is not feasible under HD frames (limited by development toolkits), \u0d45 \u0d45 \u0d45 \u0d48 \u0d48 Cycle\u00a03 Cycle\u00a02 Cycle\u00a01 Simultaneous\u00a0Computations Multiply Accumulate \u2026 \u2026 (b)\u00a0Row\u2010level\u00a0Parallelism\u00a0 Input\u00a0 feature\u00a0 map\u00a0row Kernel\u00a0 weight\u00a0 row \u2026 Multiplication Accumulation \u0d48 \u0d45 \u2026 Simultaneous\u00a0 Computations (c)\u00a0Column\u2010level\u00a0Parallelism\u00a0 Pruned\u00a0channels\u00a0allow\u00a0skip\u00a0data\u00a0 loading\u00a0and\u00a0computation \u0d48 (d)\u00a0Structured\u2010pruning Skip\u00a0computation\u00a0and\u00a0 loading\u00a0of\u00a0dropped patches and\u00a0feature\u00a0map\u00a0tiles (a)\u00a0Tile\u2010based\u00a0Architecture Figure 4: Pruning-aware parallelized hardware implementation. \fData-Model-Circuit Tri-Design for Ultra-Light Video Intelligence on Edge Devices ASPDAC \u201923, January 16\u201319, 2023, Tokyo, Japan FPGA 3 Xilinx  ZCU104@5W Conv1 Anchor Generator, Non-Maximum Suppression (NMS), and Region-Of-Interest (ROI) Align  NMS, Bounding-Box-to-ROI, and ROI Align Conv2_x Conv3_x Conv4_x Conv5_x FPN0 FPN1 FPN2 FPN3 RPN0 RPN1 RPN2 RPN3/4 BBox_regression 13.93 Bbox_classify 13.90 Track Head Convs + Group Norm + FC 3.86 Bi-directional Softmax 2.2 FPGA 1 Xilinx Alveo U50@75W FPGA 2 Xilinx Alveo U50@75W GOPs before pruning ResNet-50  + FPN GOPs after pruning GOPs with un-structured  pruning  11.67 1.16 17.36 1.74 25.56 2.56 13.26 1.33 34.73 3.47 8.68 0.87 2.17 0.22 1.02 0.10 34.91 3.49 8.73 0.87 2.18 0.22 0.69 0.07 1.39 1.39 0.39 Figure 5: FPGA cluster for QDTrack implementation. FPGA 1 (Xilinx Alveo U50) is responsible for ResNet50 backbone; FPGA 2 (U50) is responsible for FPN and RPN modules; FPGA 3 (low-end ZCU104) processes the object tracking head. Under each module, we annotate the GOPs before/after pruning. we dissemble the ML model into components and map them to different FPGAs. As demonstrated in Fig. 5, multiple FPGAs work in a dataflow manner: 1) The first FPGA executes the backbone ResNet-50 computation for one frame and meanwhile passes the intermediate feature maps to the second FPGA. Once the first FPGA finishes the backbone computation for one frame, it would start the computation for the next one. 2) The second FPGA starts the computation as soon as it receives the first tile of intermediate feature maps. Such a dataflow architecture not only significantly reduces end-to-end latency but also increases the throughput by overlapping the computations of multiple frames. In addition, it resolves the scalability issue of both the device and the development tool as each FPGA can be developed independently. 4 EXPERIMENTS 4.1 Experiment setup We conduct the MOT task under the BDD100K dataset [22], one of the most representative and challenging self-driving car datasets that incorporates geographic, environmental, weather diversity and intentional occlusions. It contains 100K videos with 30 frames per second. In our experiments, the training set has 1,400 videos and the validation set has 200 videos. To evaluate the accuracy performance of our proposed tri-design approach, we adopt the standard MOT metrics, ID F1 Score (IDF1) and Multi-Object Tracking Accuracy (MOTA) . These are two aggregated metrics widely used to evaluate the overall tracking performance by taking into account object detection accuracy and tracking consistency. In addition, we use the number of Identity Switches (IDSw) metric to evaluate the RL reward function in the first stage of data complexity reduction. To evaluate the efficiency of hardware implementation of our proposed tri-design approach, we demonstrate the QDTrack model on a dataflow cluster composed of three FPGAs: two mid-end Xilinx Alveo U50 boards (5,952 DSPs, 28 MB on-chip memory), and one low-end Xilinx ZCU104 (1,728 DSPs, 1.37 MB on-chip memory) board. The accelerator is developed using Vitis HLS 2021.1 and deployed using Vivado 2021.1. Figure 6: Comparison between random patch dropping and our proposed saliency-based dropping vs. different drop patch sizes. 4.2 Experiment results 4.2.1 Results on video frame dropping. We peer into the effect of the RL-based frame dropping model on our proposed tri-design pipeline. We compare our frame dropping approach with two baselines: uniform dropping and random dropping. The former one drops the frames at a fixed time window for each video, and the latter one randomly selects and drops frames in a video stream. Fig. 7 shows the frame dropping performance at different drop rates from 15% to 60%. As can be seen, our RL-based model generally outperforms the random drop model at different dropping ratios, indicating the effectiveness of the RL-based frame dropping model. We also observe that the \u2018sweet\u2019 dropping ratio is achieved round 40%, which strikes the most graceful balance between spatial reduction and IDSw. In what follows, unless specified otherwise, we will assume that 40% temporal frame reduction is applied to MOT. 4.2.2 Results on saliency-based image patch dropping. We also studied the effect of spatial patch dropping in our framework. In Fig. 6, we compare the saliency-based method with random dropping versus different patch drop sizes. As can be observed, the proposed saliency-score based method consistently outperforms random dropping baseline at different patch size. We also observe that IDF1 and MOTA have similar performance for sizes smaller than 175 \u00d7 175, and they start to decrease if the drop patch size becomes larger than 175 \u00d7 175. Therefore, we set the drop patch size as 60 \u00d7 60 in other experiments along with drop ratio as 20% so as to maintain the IDF1 accuracy around 0.7. 4.2.3 Results on model pruning. We compare our proposed model compression method with two commonly-used pruning baselines: irregular weight magnitude pruning (IMP) and channel-wise structured pruning. The results are demonstrated in Fig. 7. As we can see, channel-wise pruning is quite sensitive to the pruning ratio. By contrast, our proposed hardware-aware pruning method is capable of achieving MOT performance similar to IMP across different pruning ratios. Additionally, in Table 2 we show that when integrating with the spatiotemporal data reduction algorithms, our proposed hardware-aware pruning method yields back-end acceleration according to the performance of on-board latency and the effective video frame rate. 4.2.4 Tri-design performance. We next summarize the performance of integrated frame dropping, patch dropping, and model compression. In Table 2, we demonstrate both MOT accuracy and hardware \fASPDAC \u201923, January 16\u201319, 2023, Tokyo, Japan Zhang et al. Table 2: Performance comparison of our proposal with its variants and the QDTrack [16] baselines (implemented by GPU and FPGA respectively) under various metrics including accuracy (IDF1 and MOTA) and hardware metrics given by on-board latency in the unit of millisecond, effective frame rate (EFR) in the unit of FPS, and power in the unit of Watt. The data reduction is implemented using 40% temporal frame dropping together with 20% spatial patch dropping. The \u2018\u00d7\u2019 indicates the case of no data/model compression, and \u2018\u2014\u2019 means not applicable. The improvement of our proposed tri-design approach over GPU and FPGA baselines are summarized. Methods Data/model compression Metrics Data reduction Pruning IDF1 (\u2191) MOTA (\u2191) Latency (\u2193) EFR (\u2191) Power (\u2193) Energy Efficiency (\u2193) QDTrack (GPU baseline) \u00d7 \u00d7 0.714 0.637 60.9 22.5 296 W 13.2 J/frame QDTrack on FPGA \u00d7 \u00d7 0.714 0.637 554.7 1.8 50.8 W 28.2 J/frame Variant: Frame + patch drop (40%, 20%) \u00d7 0.71 0.628 443.8 2.3 50.8 W 22.0 J/frame Tri-design (ours) (40%, 20%) 90% 0.704 0.617 44.4 37.6 50.8 W 1.35 J/frame Improv. over GPU baseline \u2014 \u2014 -1.40% -3.14% 1.37\u00d7 1.67\u00d7 5.83\u00d7 9.78\u00d7 Improv. over FPGA baseline \u2014 \u2014 -1.40% -3.14% 12.5\u00d7 20.9\u00d7 \u2014 20.9\u00d7 Figure 7: Performance comparison: Left: Temporal filtering; Right: Pruning. implementation efficiency metrics of our proposed tri-design approach. The baselines include the original QDTrack [16] implementation on GPU Tesla V100 and the Xilinx FPGA Alveo U50 cluster (Section 3.4, Fig. 5). We present the results of different configurations of data complexity reduction and model weight pruning; each configuration results in one variant of the proposed tri-design method. Accuracy. As we can see, the use of proposed data complexity reduction strategy (i.e., the row \u2018Variant\u2019) does not hamper the MOT accuracies in terms of IDF1 and MOTA scores. Throughput. Comparing to GPU baseline, our tri-design approach reduces latency by 1.37\u00d7 and improves the effective frame rate (EFR) by 1.67\u00d7; comparing to FPGA baseline, our tri-design together with proposed sparsity-aware multi-level parallelism reduces latency by 12.5\u00d7 and improves the EFR by 20.9\u00d7, demonstrating significant performance boost. Energy Efficiency. Meanwhile, our approach shows remarkable energy efficiency. The baseline GPU Tesla V100 has a power consumption of 296W, while our FPGA cluster has only 50.8 W, leading to 5.83\u00d7 power reduction. Together with 1.67\u00d7 higher frame rate, the energy efficiency is 9.78\u00d7 higher (5.83 \u00d7 1.67). To summarize, our approach outperforms GPU and FPGA baselines in all hardware efficiency metrics with negligible accuracy loss. 5"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2203.14195v1",
+                    "title": "How to Robustify Black-Box ML Models? A Zeroth-Order Optimization Perspective",
+                    "abstract": "The lack of adversarial robustness has been recognized as an important issue\nfor state-of-the-art machine learning (ML) models, e.g., deep neural networks\n(DNNs). Thereby, robustifying ML models against adversarial attacks is now a\nmajor focus of research. However, nearly all existing defense methods,\nparticularly for robust training, made the white-box assumption that the\ndefender has the access to the details of an ML model (or its surrogate\nalternatives if available), e.g., its architectures and parameters. Beyond\nexisting works, in this paper we aim to address the problem of black-box\ndefense: How to robustify a black-box model using just input queries and output\nfeedback? Such a problem arises in practical scenarios, where the owner of the\npredictive model is reluctant to share model information in order to preserve\nprivacy. To this end, we propose a general notion of defensive operation that\ncan be applied to black-box models, and design it through the lens of denoised\nsmoothing (DS), a first-order (FO) certified defense technique. To allow the\ndesign of merely using model queries, we further integrate DS with the\nzeroth-order (gradient-free) optimization. However, a direct implementation of\nzeroth-order (ZO) optimization suffers a high variance of gradient estimates,\nand thus leads to ineffective defense. To tackle this problem, we next propose\nto prepend an autoencoder (AE) to a given (black-box) model so that DS can be\ntrained using variance-reduced ZO optimization. We term the eventual defense as\nZO-AE-DS. In practice, we empirically show that ZO-AE- DS can achieve improved\naccuracy, certified robustness, and query complexity over existing baselines.\nAnd the effectiveness of our approach is justified under both image\nclassification and image reconstruction tasks. Codes are available at\nhttps://github.com/damon-demon/Black-Box-Defense.",
+                    "authors": "Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jinfeng Yi, Mingyi Hong, Shiyu Chang, Sijia Liu",
+                    "published": "2022-03-27",
+                    "updated": "2022-03-27",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.AI",
+                        "cs.CV"
+                    ],
+                    "main_content": "INTRODUCTION ML models, DNNs in particular, have achieved remarkable success owing to their superior predictive performance. However, they often lack robustness. For example, imperceptible but carefully-crafted input perturbations can fool the decision of a well-trained ML model. These input perturbations refer to adversarial perturbations, and the adversarially perturbed (test-time) examples are known as adversarial examples or adversarial attacks (Goodfellow et al., 2015; Carlini & Wagner, 2017; Papernot et al., 2016). Existing studies have shown that it is not dif\ufb01cult to generate adversarial attacks. Numerous attack generation methods have been designed and successfully applied to (i) different use cases from the digital world to the physical world, e.g., image classi\ufb01cation (Brown et al., 2017; Li et al., 2019; Xu et al., 2019; Yuan et al., 2021), object detection/tracking (Eykholt et al., 2017; Xu et al., 2020; Sun et al., 2020), and image reconstruction (Antun et al., 2020; Raj et al., 2020; Vasiljevi\u00b4 c et al., 2021), and (ii) different types of victim models, e.g., white-box models whose details can be accessed by adversaries (Madry et al., 2018; Carlini & Wagner, 2017; Tramer et al., 2020; Croce & Hein, 2020; Wang et al., 2021), and black-box models whose information is not disclosed to adversaries (Papernot et al., 2017; Tu et al., 2019; Ilyas et al., 2018a; Liang et al., 2021). Given the prevalence of adversarial attacks, methods to robustify ML models are now a major focus in research. For example, adversarial training (AT) (Madry et al., 2018), which has been poised one of the most effective defense methods (Athalye et al., 2018), employed min-max optimization to minimize the worst-case (maximum) training loss induced by adversarial attacks. Extended from AT, 1 arXiv:2203.14195v1  [cs.LG]  27 Mar 2022 \fPublished as a conference paper at ICLR 2022 various empirical defense methods were proposed, ranging from supervised learning, semi-supervised learning, to unsupervised learning (Madry et al., 2018; Zhang et al., 2019b; Shafahi et al., 2019; Zhang et al., 2019a; Carmon et al., 2019; Chen et al., 2020; Zhang et al., 2021). In addition to empirical defense, certi\ufb01ed defense is another research focus, which aims to train provably robust ML models and provide certi\ufb01cates of robustness (Wong & Kolter, 2017; Raghunathan et al., 2018; Katz et al., 2017; Salman et al., 2019; 2020; 2021). Although exciting progress has been made in adversarial defense, nearly all existing works ask a defender to perform over white-box ML models (assuming non-con\ufb01dential model architectures and parameters). However, the white-box assumption may restrict the defense application in practice. For example, a model owner may refuse to share the model details, since disclosing model information could hamper the owner\u2019s privacy, e.g., model inversion attacks lead to training data leakage (Fredrikson et al., 2015). Besides the privacy consideration, the white-box defense built upon the (end-to-end) robust training (e.g., AT) is computationally intensive, and thus is dif\ufb01cult to scale when robustifying multiple models. For example, in the medical domain, there exist massive pre-trained ML models for different diseases using hundreds of neuroimaging datasets (Sisodiya et al., 2020). Thus, robustly retraining all models becomes impractical. Taking the model privacy and the defense ef\ufb01ciency into consideration, we ask: Is it possible to design an adversarial defense over black-box models using only model queries? Black-box model Defensive  operation Inference Ostrich  Output feedback Input query Adversarial  attack Figure 1: Illustration of defense against adversarial attacks for entirely black-box models. Extending adversarial defense to the black-box regime (that we call \u2018blackbox defense\u2019) is highly non-trivial due to the challenge of black-box optimization (i.e., learning over blackbox models). To tackle this problem, the prior work (Salman et al., 2020) leveraged surrogate models as approximations of the black-box models, over which defense can be conducted following the whitebox setup. Yet, this still requires to have access to the information on the victim model type and its function. In practice, those conditions could be dif\ufb01cult to achieve. For example, if the domain knowledge related to medicine or healthcare is lacking (Qayyum et al., 2020; Finlayson et al., 2019), then it will be dif\ufb01cult to determine a proper surrogate model of a medical ML system. Even if a black-box model estimate can be obtained using the model inversion technique (Kumar & Levine, 2019), a signi\ufb01cantly large number of model queries are needed even just for tackling a MNIST-level prediction task (Oh et al., 2019). Different from (Salman et al., 2020), we study an authentic blackbox scenario, in which the interaction between defender and model is only based on input-output function queries (see Fig. 1). To our best knowledge, this is the \ufb01rst work to tackle the problem of query-based black-box defense. Contributions. We summarize our contributions below. x (Formulation-wise) We formulate the problem of black-box defense and investigate it through the lens of zeroth-order (ZO) optimization. Different from existing works, our paper aims to design the restriction-least black-box defense and our formulation is built upon a query-based black-box setting, which avoids the use of surrogate models. y (Methodology-wise) We propose a novel black-box defense approach, ZO AutoEncoder-based Denoised Smoothing (ZO-AE-DS), which is able to tackle the challenge of ZO optimization in high dimensions and convert a pre-trained non-robust ML model into a certi\ufb01ably robust model using only function queries. z (Experiment-wise) We verify the ef\ufb01cacy of our method through an extensive experimental study. In the task of image classi\ufb01cation, the proposed ZO-AE-DS signi\ufb01cantly outperforms the ZO baseline built upon (Salman et al., 2020). For instance, we can improve the certi\ufb01ed robust accuracy of ResNet110 on CIFAR-10 from 19.16% (using baseline) to 54.87% (using ZO-AE-DS) under adversarial perturbations with \u21132 norm less than 64/255. We also empirically show that our proposal stays effective even in the task of image reconstruction. 2 \fPublished as a conference paper at ICLR 2022 2 RELATED WORK Empirical defense. An immense number of defense methods have been proposed, aiming to improve model robustness against adversarial attacks. Examples include detecting adversarial attacks (Guo et al., 2017; Meng & Chen, 2017; Gong et al., 2017; Grosse et al., 2017; Metzen et al., 2017) and training robust ML models (Madry et al., 2018; Zhang et al., 2019b; Shafahi et al., 2019; Wong et al., 2020; Zhang et al., 2019a; Athalye et al., 2018; Cheng et al., 2017; Wong & Kolter, 2017; Salman et al., 2019; Raghunathan et al., 2018; Katz et al., 2017). In this paper, we focus on advancing the algorithm foundation of robust training over black-box models. Robust training can be broadly divided into two categories: empirical defense and certi\ufb01ed defense. In the former category, the most representative method is AT (adversarial training) that formulates adversarial defense as a two-player game (between attacker and defender) (Madry et al., 2018). Spurred by AT, empirical defense has developed rapidly. For example, in (Zhang et al., 2019b), TRADES was proposed to seek the optimal trade-off between accuracy and robustness. In (Stanforth et al., 2019; Carmon et al., 2019), unlabeled data and self-training were shown effective to improve adversarial defense in both robustness and generalization. In (Shafahi et al., 2019; Wong et al., 2020; Zhang et al., 2019a; Andriushchenko & Flammarion, 2020), to improve the scalability of adversarial defense, computationally-light alternatives of AT were developed. Despite the effectiveness of empirical defense against adversarial attacks (Athalye et al., 2018), it lacks theoretical guarantee (known as \u2018certi\ufb01cate\u2019) for the achieved robustness. Thus, the problem of certi\ufb01ed defense arises. Certi\ufb01ed defense. Certi\ufb01ed defense seeks to provide a provably guarantee of ML models. One line of research focuses on post-hoc formal veri\ufb01cation of a pre-trained ML model. The certi\ufb01ed robustness is then given by a \u2018safe\u2019 input perturbation region, within which any perturbed inputs will not fool the given model (Katz et al., 2017; Ehlers, 2017; Bunel et al., 2018; Dutta et al., 2017). Since the exact veri\ufb01cation is computationally intensive, a series of work (Raghunathan et al., 2018; Dvijotham et al., 2018; Wong & Kolter, 2017; Weng et al., 2018a;b; Wong et al., 2018) proposed \u2018incomplete\u2019 veri\ufb01cation, which utilizes convex relaxation to over-approximate the output space of a predictive model when facing input perturbations. Such a relaxation leads to fast computation in the veri\ufb01cation process but only proves a lower bound of the exact robustness guarantee. Besides the post-hoc model veri\ufb01cation with respect to each input example, another line of research focuses on in-processing certi\ufb01cation-aware training and prediction. For example, randomized smoothing (RS) transforms an empirical classi\ufb01er into a provably robust one by convolving the former with an isotropic Gaussian distribution. It was shown in (Cohen et al., 2019) that RS can provide formal guarantees for adversarial robustness. Different types of RS-oriented provable defenses have been developed, such as adversarial smoothing (Salman et al., 2019), denoised smoothing (Salman et al., 2020), smoothed ViT (Salman et al., 2021), and feature smoothing (Addepalli et al., 2021). Zeroth-order (ZO) optimization for adversarial ML. ZO optimization methods are gradientfree counterparts of \ufb01rst-order (FO) optimization methods (Liu et al., 2020b). They approximate the FO gradients through function value based gradient estimates. Thus, ZO optimization is quite useful to solve black-box problems when explicit expressions of their gradients are dif\ufb01cult to compute or infeasible to obtain. In the area of adversarial ML, ZO optimization has become a principled approach to generate adversarial examples from black-box victim ML models (Chen et al., 2017; Ilyas et al., 2018a;b; Tu et al., 2019; Liu et al., 2019; 2020a; Huang & Zhang, 2020; Cai et al., 2020; 2021). Such ZO optimization-based attack generation methods can be as effective as state-of-the-art white-box attacks, despite only having access to the inputs and outputs of the targeted model. For example, the work (Tu et al., 2019) leveraged the white-box decoder to map the generated low-dimension perturbations back to the original input dimension. Inspired by (Tu et al., 2019), we leverage the autoencoder architecture to tackle the high-dimension challenge of ZO optimization in black-box defense. Despite the widespread application of ZO optimization to black-box attack generation, few work studies the problem of black-box defense. 3 PROBLEM FORMULATION: BLACK-BOX DEFENSE In this section, we formulate the problem of black-box defense, i.e., robustifying black-box ML models without having any model information such as architectures and parameters. 3 \fPublished as a conference paper at ICLR 2022 Problem statement. Let f\u03b8bb(x) denote a pre-de\ufb01ned black-box (bb) predictive model, which can map an input example x to a prediction. In our work, f\u03b8bb can be either an image classi\ufb01er or an image reconstructor. For simplicity of notation, we will drop the model parameters \u03b8bb when referring to a black-box model. The threat model of our interest is given by norm-ball constrained adversarial attacks (Goodfellow et al., 2015). To defend against these attacks, existing approaches commonly require the white-box assumption of f (Madry et al., 2018) or have access to white-box surrogate models of f (Salman et al., 2020). Different from the prior works, we study the problem of black-box defense when the owner of f is not able to share the model details. Accordingly, the only mode of interaction with the black-box system is via submitting inputs and receiving the corresponding predicted outputs. The formal statement of black-box defense is given below: (Black-box defense) Given a black-box base model f, can we develop a defensive operation R using just input-output function queries so as to produce the robusti\ufb01ed model R(f) against adversarial attacks? Defensive operation. We next provide a concrete formulation of the defensive operation R. In the literature, two principled defensive operations were used: (R1) end-to-end AT (Madry et al., 2018; Zhang et al., 2019b; Cohen et al., 2019), and (R2) prepending a defensive component to a base model (Meng & Chen, 2017; Salman et al., 2020; Aldahdooh et al., 2021). The former (R1) has achieved the state-of-the-art robustness performance (Athalye et al., 2018; Croce & Hein, 2020) but is not applicable to black-box defense. By contrast, the latter (R2) is more compatible with black-box models. For example, denonised smoothing (DS), a recently-developed R2-type approach (Salman et al., 2020), gives a certi\ufb01ed defense by prepending a custom-trained denoiser to the targeted model. In this work, we choose DS as the backbone of our defensive operation (Fig. 2). Black-box predictor  Input  Denoised  example Prediction White-box denoiser Robustification: Figure 2: DS-based black-box defense. In DS, a denoiser is integrated with a base model f so that the augmented system becomes resilient to Gaussian noise and thus plays a role similar to the RS-based certi\ufb01ed defense (Cohen et al., 2019). That is, DS yields R(f(x)) := f(D\u03b8(x)), (1) where D\u03b8 denotes the learnable denoiser (with parameters \u03b8) prepended to the (black-box) predictor f. Once D\u03b8 is learned, then the DS-based smooth classi\ufb01er, arg maxc P\u03b4\u2208N (0,\u03c32I)[R(f(x + \u03b4)) = c], can achieve certi\ufb01ed robustness, where c is a class label, \u03b4 \u2208N(0, \u03c32I) denotes the standard Gaussian noise with variance \u03c32, and arg maxc P\u03b4\u2208N (0,\u03c32I)[f(x + \u03b4) = c] signi\ufb01es a smooth version of f. Based on (1), the goal of black-box defense becomes to \ufb01nd the optimal denoiser D\u03b8 so as to achieve satisfactory accuracy as well as adversarial robustness. In the FO learning paradigm, Salman et al. (2020) proposed a stability regularized denoising loss to train D\u03b8: minimize \u03b8 E\u03b4\u2208N (0,\u03c32I),x\u2208U \u2225D\u03b8(x + \u03b4) \u2212x\u22252 2 | {z } := \u2113Denoise(\u03b8) +\u03b3E\u03b4,x \u2113CE(R(f(x + \u03b4)), f(x)) | {z } := \u2113Stab(\u03b8) , (2) where U denotes the training dataset, the \ufb01rst objective term \u2113Denoise(\u03b8) corresponds to the mean squared error (MSE) of image denoising, the second objective term \u2113Stab(\u03b8) measures the prediction stability through the cross-entropy (CE) between the outputs of the denoised input and the original input, and \u03b3 > 0 is a regularization parameter that strikes a balance between \u2113Denoise and \u2113Stab. We remark that problem (2) can be solved using the FO gradient descent method if the base model f is fully disclosed to the defender. However, the black-box nature of f makes the gradients of the stability loss \u2113Stab(\u03b8) infeasible to obtain. Thus, we will develop a gradient-free DS-oriented defense. 4 METHOD: A SCALABLE ZEROTH-ORDER OPTIMIZATION SOLUTION In this section, we begin by presenting a brief background on ZO optimization, and elaborate on the challenge of black-box defense in high dimensions. Next, we propose a novel ZO optimization-based DS method that can not only improve model query complexity but also lead to certi\ufb01ed robustness. 4 \fPublished as a conference paper at ICLR 2022 ZO optimization. In ZO optimization, the FO gradient of a black-box function \u2113(w) (with a d-dimension variable w) is approximated by the difference of two function values along a set of random direction vectors. This leads to the randomized gradient estimate (RGE) (Liu et al., 2020b): \u02c6 \u2207w\u2113(w) = 1 q q X i=1 \u0014 d \u00b5 (\u2113(w + \u00b5ui) \u2212\u2113(w)) ui \u0015 , (3) where {ui}q i=1 are q random vectors drawn independently and uniformly from the sphere of a unit ball, and \u00b5 > 0 is a given small step size, known as the smoothing parameter. The rationale behind (3) is that it provides an unbiased estimate of the FO gradient of the Gaussian smoothing version of \u2113(Gao et al., 2018), with variance in the order of O( d q ) (Liu et al., 2020b). Thus, a large-scale problem (with large d) yields a large variance of RGE (3). To reduce the variance, a large number of querying directions (i.e., q) is then needed, with the worst-case query complexity in the order of O(d). If q = d, then the least estimation variance can be achieved by the coordinatewise gradient estimate (CGE) (Lian et al., 2016; Liu et al., 2018): \u02c6 \u2207w\u2113(w) = d X i=1 \u0014\u2113(w + \u00b5ei) \u2212\u2113(w) \u00b5 ei \u0015 , (4) where ei \u2208Rd denotes the ith elementary basis vector, with 1 at the ith coordinate and 0s elsewhere. For any off-the-shelf FO optimizers, e.g., stochastic gradient descent (SGD), if we replace the FO gradient estimate with the ZO gradient estimate, then we obtain the ZO counterpart of a FO solver, e.g., ZO-SGD (Ghadimi & Lan, 2013). Warm-up: A direct application of ZO optimization. A straightforward method to achieve the DS-based black-box defense is to solve problem (2) using ZO optimization directly. However, it will give rise to the dif\ufb01culty of ZO optimization in high dimensions. Speci\ufb01cally, DS requires to calculate the gradient of the defensive operation (1). With the aid of ZO gradient estimation, we obtain \u2207\u03b8R(f(x)) = dD\u03b8(x) d\u03b8 d f(z) dz |z=D\u03b8(x) \u2248dD\u03b8(x) d\u03b8 \u02c6 \u2207zf(z) |z=D\u03b8(x) , (5) where with an abuse of notation, let d denote the dimension of x (yielding D\u03b8(x) \u2208Rd and z \u2208Rd) and d\u03b8 denote the dimension of \u03b8, dD\u03b8(x) d\u03b8 \u2208Rd\u03b8\u00d7d is the Jacobian matrix of the vector-valued function D\u03b8(x), and \u02c6 \u2207zf(z) denotes the ZO gradient estimate of f, following (3) or (4). Since the dimension of an input is typically large for image classi\ufb01cation (e.g., d = 3072 for a CIFAR-10 Method Certi\ufb01ed robustness (%) (\u21132 radius: \u03f5 = 0.5) Standard accuracy (%) FO-DS 30.22 71.80 ZO-DS (RGE, q = 192) 5.06 (\u219325.16) 44.81 (\u219326.99) Table 1: Performance comparison between FODS (Salman et al., 2020) and its direct ZO implementation ZO-DS on (CIFAR-10, ResNet-110). image), it imposes two challenges: (a) The variance of RGE (3) will be ultra-large if the query complexity stays low, i.e., a small query number q is used; And (b) the variance-least CGE (4) becomes impracticable due to the need of ultra-high querying cost (i.e., q = d). Indeed, Table 1 shows that the direct application of (5) into the existing FO-DS solver (Salman et al., 2020), which we call ZO-DS, yields over 25% degradation in both standard accuracy and certi\ufb01ed robustness evaluated at input perturbations with \u21132 norm less than 128/255, where pixels of an input image are normalized to [0, 1]. We refer readers to Sec. 5 for more details. Black-box predictor  White-box denoiser White-box  Black-box Figure 3: Model architecture for ZO-AE-DS. ZO autoencoder-based DS (ZO-AEDS): A scalable solution to black-box defense. The dif\ufb01culty of ZO optimization in high dimensions prevents us from developing an effective DS-oriented provable defense for black-box ML models. To tackle such problem, we introduce an Autoencoder (AE) to connect the front-end denoiser D\u03b8 with the back-end black-box predictive model f so that ZO optimization can be conducted in a (low-dimension) feature embedding space. To be concrete, let \u03c8\u03b8Dec \u25e6\u03c6\u03b8Enc denote AE consisting of the encoder (Enc) \u03c8\u03b8Enc and the 5 \fPublished as a conference paper at ICLR 2022 decoder (Dec) \u03c6\u03b8Dec, where \u25e6denotes the function composition operation. Plugging \u03c6\u03b8AE between the denoiser D\u03b8 and the black-box predictor f, we then extend the defensive operation (1) to the following (see Fig. 3 for illustration): Rnew(f(x)) := f (\u03c6\u03b8Dec (z)) | {z } new black box , z = \u03c8\u03b8Enc (D\u03b8(x)) | {z } new white box , (6) where z \u2208Rdz denotes the low-dimension feature embedding with dz < d. In (6), we integrate the decoder \u03c8\u03b8Dec with the black-box predictor f to construct a new black-box model f \u2032(z) := f(\u03c8\u03b8Dec(z)), which enables us to derive a ZO gradient estimate of reduced dimension: \u2207\u03b8Rnew(f(x)) \u2248d\u03c6\u03b8Enc(D\u03b8(x)) d\u03b8 \u02c6 \u2207zf \u2032(z) |z=\u03c6\u03b8Enc(D\u03b8(x)) . (7) Assisted by AE, RGE of \u02c6 \u2207zf \u2032 has a reduced variance from O( d q ) to O( dz q ). Meanwhile, the least-variance CGE (4) also becomes feasible by setting the query number as q = dz. Note that the eventual ZO estimate (7) is a function of the d\u03b8 \u00d7dz Jacobian matrix \u2207\u03b8[\u03c6\u03b8Enc(D\u03b8(x))]. For ease of storing and computing the Jacobian matrix, we derive the following computationally-light alternative of (7) (see derivation in Appendix A): \u2207\u03b8\u2113Stab(\u03b8) \u2248\u2207\u03b8[a\u22a4\u03c6\u03b8Enc(D\u03b8(x + \u03b4))], a = \u02c6 \u2207z\u2113CE(f \u2032(z), f(x)) |z=\u03c6\u03b8Enc(D\u03b8(x+\u03b4)) , (8) where recall that f \u2032(z) = f(\u03c8\u03b8Dec(z)), and \u02c6 \u2207denotes the ZO gradient estimate given by (3) or (4). The computation advantage of (8) is that the derivative operation \u2207\u03b8 can be applied to a scalar-valued inner product built upon a pre-calculated ZO gradient estimate a. Training ZO-AE-DS. Recall from Fig. 3 that the proposed defensive system involves three components: denoiser D\u03b8, AE \u03c8\u03b8Dec \u25e6\u03c6\u03b8Enc, and pre-de\ufb01ned black-box predictor f. Thus, the parameters to be optimized include \u03b8, \u03b8Dec and \u03b8Enc. To train ZO-AE-DS, we adopt a two-stage training protocol. x White-box pre-training on AE: At the \ufb01rst stage, we pre-train the AE model by calling a standard FO optimizer (e.g., Adam) to minimize the reconstruction loss Ex\u2225\u03c6\u03b8Dec(\u03c8\u03b8Enc(x)) \u2212x\u22252 2. The resulting AE will be used as the initialization of the second-stage training. We remark that the denoising model D\u03b8 can also be pre-trained. However, such a pre-training could hamper optimization, i.e., making the second-stage training over \u03b8 easily trapped at a poor local optima. y End-to-end training: At the second stage, we keep the pre-trained decoder \u03c6\u03b8Dec intact and merge it into the black-box system as shown in Fig. 3. We then optimize \u03b8 and \u03b8Enc by minimizing the DS-based training loss (2), where the denoiser D\u03b8 and the defensive operation R are replaced by \u03c8\u03b8Enc \u25e6D\u03b8 and Rnew (6), respectively. In (2), minimization over the stability loss \u2113Stab(\u03b8) calls the ZO estimate of \u2207\u03b8\u2113Stab(\u03b8), given by (7). In Appendix C.2, different training schemes are discussed. 5 EXPERIMENTS In this section, we demonstrate the effectiveness of our proposal through extensive experiments. We will show that the proposed ZO-AE-DS outperforms a series of baselines when robustifying black-box neural networks for secure image classi\ufb01cation and image reconstruction. 5.1 EXPERIMENT SETUP Datasets and model architectures. In the task of image classi\ufb01cation, we focus on CIFAR-10 and STL-10 datasets. In Appendix C.3, we demonstrate the effectiveness of ZO-AE-DS on the high-dimension ImageNet images. In the task of image reconstruction, we consider the MNIST dataset. To build ZO-AE-DS and its variants and baselines, we specify the prepended denoiser D\u03b8 as DnCNN (Zhang et al., 2017). We then implement task-speci\ufb01c AE for different datasets. Super\ufb01cially, the dimension of encoded feature embedding, namely, dz in (6), is set as 192, 576 and 192 for CIFAR-10, STL-10 and MNIST, respectively. The architectures of AE are con\ufb01gured following (Mao et al., 2016), and ablation study on the choice of AE is shown in Appendix C.1. To specify the black-box image classi\ufb01cation model, we choose ResNet-110 for CIFAR-10 following (Salman et al., 2020), and ResNet-18 for STL-10. It is worth noting that STL-10 contains 500 labeled 96 \u00d7 96 training images, and the pre-trained ResNet-18 achieves 76.6% test accuracy that matches to state-of-the-art performance. For image reconstruction, we adopt a reconstruction network consisting of convolution, deconvolution and ReLU layers, following (Raj et al., 2020). 6 \fPublished as a conference paper at ICLR 2022 Baselines. We will consider two variants of our proposed ZO-AE-DS: i) ZO-AE-DS using RGE (3), ii) ZO-AE-DS using CGE (4). In addition, we will compare ZO-AE-DS with i) FO-AE-DS, i.e., the \ufb01rst-order implementation of ZO-AE-DS, ii) FO-DS, which developed in (Salman et al., 2020), iii) RS-based certi\ufb01ed training, proposed in (Cohen et al., 2019), and iv) ZO-DS, i.e., the ZO implementation of FO-DS using RGE. Note that CGE is not applicable to ZO-DS due to the obstacle of high dimensions. To our best knowledge, ZO-DS is the only query-based black-box defense baseline that can be directly compared with ZO-AE-DS. Training setup. We build the training pipeline of the proposed ZO-AE-DS following \u2018Training ZO-AE-DS\u2019 in Sec. 4. To optimize the denoising model D\u03b8, we will cover two training schemes: training from scratch, and pre-training & \ufb01ne-tuning. In the scenario of training from scratch, we use Adam optimizer with learning rate 10\u22123 to train the model for 200 epochs and then use SGD optimizer with learning rate 10\u22123 drop by a factor of 10 at every 200 epoch, where the total number of epochs is 600. As will be evident later, training from scratch over D\u03b8 leads to better performance of ZO-AE-DS. In the scenario of pre-training & \ufb01ne-tuning, we use Adam optimizer to pre-train the denoiser D\u03b8 with the MSE loss \u2113Denoise in (2) for 90 epochs and \ufb01ne-tune the denoiser with \u2113Stab for 200 epochs with learning rate 10\u22125 drop by a factor of 10 every 40 epochs. When implementing the baseline FO-DS, we use the best training setup provided by (Salman et al., 2020). When implementing ZO-DS, we reduce the initial learning rate to 10\u22124 for training from scratch and 10\u22126 for pre-training & \ufb01ne-tuning to stabilize the convergence of ZO optimization. Furthermore, we set the smoothing parameter \u00b5 = 0.005 for RGE and CGE. And to achieve a smooth predictor, we set the Gaussian smoothing noise as \u03b4 \u2208N(0, \u03c32I) with \u03c32 = 0.25. With the help of matrix operations and the parallel computing power of the GPU, we optimize the training time to an acceptable range. The averaged one-epoch training time on a single Nvidia RTX A6000 GPU is about \u223c1min and \u223c29min for FO-DS and our proposed ZO method, ZO-AE-DS (CGE, q = 192), on the CIFAR-10 dataset. Evaluation metrics. In the task of robust image classi\ufb01cation, the performance will be evaluated at standard test accuracy (SA) and certi\ufb01ed accuracy (CA). Here CA is a provable robust guarantee of the Gaussian smoothing version of a predictive model. Let us take ZO-AE-DS as an example, the resulting smooth image classi\ufb01er is given by fsmooth(x) := arg maxc P\u03b4\u2208N (0,\u03c32I)[Rnew(f(x + \u03b4)) = c], where Rnew is given by (6). Further, a certi\ufb01ed radius of \u21132-norm perturbation ball with respect to an input example can be calculated following the RS approach provided in (Cohen et al., 2019). As a result, CA at a given \u21132-radius r is the percentage of the correctly classi\ufb01ed data points whose certi\ufb01ed radii are larger than r. Note that if r = 0, then CA reduces to SA. FO ZO-DS ZO-AE-DS (Ours) \u21132-radius r RS FO-DS FO-AE-DS q = 20 (RGE) q = 100 (RGE) q = 192 (RGE) q = 20 (RGE) q = 100 (RGE) q = 192 (RGE) q = 192 (CGE) 0.00 (SA) 76.44 71.80 75.97 19.50 41.38 44.81 42.72 58.61 63.13 72.23 0.25 60.64 51.74 59.12 3.89 18.05 19.16 29.57 40.96 45.69 54.87 0.50 41.19 30.22 38.50 0.60 4.78 5.06 17.85 24.28 27.84 35.50 0.75 21.11 11.87 18.18 0.03 0.32 0.30 8.52 9.45 10.89 16.37 Table 2: SA (standard accuracy, %) and CA (certi\ufb01ed accuracy, %) versus different values of \u21132-radius r. Note that SA corresponds to the case of r = 0. In both FO and ZO blocks, the best accuracies for each \u21132-radius are highlighted in bold. 5.2 EXPERIMENT RESULTS ON IMAGE CLASSIFICATION Performance on CIFAR-10. In Table 2, we present certi\ufb01ed accuracies of ZO-AE-DS and its variants/baselines versus different \u21132-radii in the setup of (CIFAR-10, ResNet-110). Towards a comprehensive comparison, different RGE-based variants of ZO-AE-DS and ZO-DS are demonstrated using the query number q \u2208{20, 100, 192}. First, the comparison between ZO-AE-DS and ZO-DS shows that our proposal signi\ufb01cantly outperforms ZO-DS ranging from the low query number q = 20 to the high query number q = 192 when RGE is applied. Second, we observe that the use of CGE yields the best CA and SA (corresponding to r = 0). The application of CGE is bene\ufb01ted from AE, which reduces the dimension from d = 32 \u00d7 32 \u00d7 3 to dz = 192. In particular, CGE-based ZO-AE-DS improves the case studied in Table 1 from 5.06% to 35.5% at the \u21132-radius r = 0.5. Third, although FO-AE-DS yields CA improvement over FO-DS in the white-box context, the improvement achieved by ZO-AE-DS (vs. ZO-DS) for black-box defense is much more signi\ufb01cant. This implies 7 \fPublished as a conference paper at ICLR 2022 that the performance of black-box defense relies on a proper solution (namely, ZO-AE-DS) to tackle the challenge of ZO optimization in high dimensions. Fourth, RS outperforms the ZO methods. This is not surprising since RS is a known white-box certi\ufb01ably robust training approach. In Appendix B, we demonstrate the consistent effectiveness of ZO-AE-DS under different denoisier and classi\ufb01ers. Performance on STL-10. In Table 3, we evaluate the performance of ZO-AE-DS for STL-10 image classi\ufb01cation. For comparison, we also represent the performance of FO-DS, FO-AE-DS, and ZO-DS. Similar to Table 2, the improvement brought by our proposal over ZO-DS is evident, with at least 10% SA/CA improvement across different \u21132-radii. STL-10 \u21132-radius r FO-DS FO-AE-DS ZO-DS (RGE, q = 576) ZO-AE-DS (CGE, q = dz = 576) 0.00 (SA) 53.36 54.26 38.60 45.67 0.25 35.83 43.99 21.50 35.78 0.50 21.61 34.85 9.58 26.70 0.75 9.86 25.56 3.29 17.91 Table 3: CA (certi\ufb01ed accuracy, %) vs. different \u21132-radii for image classi\ufb01cation on STL-10. When comparing ZO-AE-DS with FO-DS, we observe that ours introduces a 7% degradation in SA (at r = 0). This is different from CIFAR-10 classi\ufb01cation. There might be two reasons for the degradation of SA in STL-10. First, the size of a STL-10 image is 9\u00d7 larger than a CIFAR-10 image. Thus, the over-reduced feature dimension could hamper SA. In this example, we set dz = 576, which is only 3\u00d7 larger than dz = 192 used for CIFAR-10 classi\ufb01cation. Second, the variance of ZO gradient estimates has a larger effect on the performance of STL-10 than that of CIFAR-10, since the former only contains 500 labeled images, leading to a challenging training task. Despite the degradation of SA, ZO-AE-DS outperforms FO-DS in CA, especially when facing a large \u21132-radius. This is consistent with Table 2. The rationale is that AE can be regarded as an extra smoothing operation for the image classi\ufb01er, and thus improves certi\ufb01ed robustness over FO-DS, even if the latter is designed in a white-box setup. If we compare ZO-AE-DS with FO-AE-DS, then the FO approach leads to the best performance due to the high-accuracy of gradient estimates. Figure 4: Comparison between non-AE-based and AE-based methods in CA vs. different \u21132-radii. Dashed lines: Models obtained by non-AE-based methods; Solid lines: Models obtained by AE-based methods. Advantage of AE on ZO optimization. Extended from Table 2, Fig. 4 presents the complete CA curve of non-AE-based and AE-based methods vs. the value of \u21132-radius in the example of (CIFAR10, ResNet-110). As we can see, ZO-AE-DS using RGE with the smallest query number q = 20 has outperformed ZO-DS using RGE with the largest query number q = 192. This shows the vital role of AE on ZO optimization. Meanwhile, consistent with Table 2 and Table 3, the best model achieved by ZO-AE-DS using CGE could be even better than the FO baseline FO-DS since AE could play a similar role on the smoothing operation. Furthermore, as the query number q increases, the improvement of ZO-AE-DS grows, towards the performance of FO-AE-DS. ZO-AE-DS (CGE, q = 192) \u21132-radius r Training from scratch Pre-training + \ufb01ne-tuning 0.00 72.23 59.74 0.25 54.87 42.61 0.50 35.50 26.26 0.75 16.37 11.13 Table 4: ZO-AE-DS using different denoiser training schemes under (CIFAR-10, ResNet-110). Effect of training scheme on ZO-AE-DS. In Table 4, we present the impact of training scheme (over the denoiser D\u03b8) on the CA performance of ZO-AE-DS versus different \u21132-radii. Two training schemes, training from scratch and pre-training + \ufb01ne-tuning, are considered. As we can see, training from scratch for D\u03b8 leads to the better performance of ZO-AE-DS than pretraining + \ufb01ne-tuning. This is because the application of pre-training to D\u03b8 could make optimization easily get trapped at a local optima. We list other ablation studies in Appendix C. 5.3 EXPERIMENT RESULTS ON IMAGE RECONSTRUCTION. In what follows, we apply the proposed ZO-AE-DS to robustifying a black-box image reconstruction network. The goal of image reconstruction is to recover the original image from a noisy measurement. 8 \fPublished as a conference paper at ICLR 2022 Following (Antun et al., 2020; Raj et al., 2020), we generate the noisy measurement following a linear observation model y = Ax, where A is a sub-sampling matrix (e.g., Gaussian sampling), and x is an original image. A pre-trained image reconstruction network (Raj et al., 2020) then takes A\u22a4y as the input to recover x. To evaluate the reconstruction performance, we adopt two metrics (Antun et al., 2020), the root mean squared error (RMSE) and structural similarity (SSIM). SSIM is a supplementary metric to RMSE, since it gives an accuracy indicator when evaluating the similarity between the true image and its estimate at \ufb01ne-level regions. The vulnerability of image reconstruction networks to adversarial attacks, e.g., PGD attacks (Madry et al., 2018), has been shown in (Antun et al., 2020; Raj et al., 2020; Wolf, 2019). When the image reconstructor is given as a black-box model, spurred by above, Table 5 presents the performance of image reconstruction using various training methods against adversarial attacks with different perturbation strengths. As we can see, compared to the normally trained image reconstructor (i.e., \u2018Standard\u2019 in Table 5), all robusti\ufb01cation methods lead to degraded standard image reconstruction performance in the non-adversarial context (i.e., \u2225\u03b4\u22252 = 0). But the worst performance is provided by ZO-DS. When the perturbation strength increases, the model achieved by standard training becomes over-sensitive to adversarial perturbations, yielding the highest RMSE and the lowest SSIM. Furthermore, we observe that the proposed black-box defense ZO-AE-DS yields very competitive and even better performance with respect to FO defenses. In Fig. 5, we provide visualizations of the reconstructed images using different approaches at the presence of reconstruction-evasion PGD attacks. For example, the comparison between Fig. 5-(f) and (b)/(d) clearly shows the robustness gained by ZO-AE-DS. Image reconstruction on MNIST Method \u2225\u03b4\u22252 = 0 \u2225\u03b4\u22252 = 1 \u2225\u03b4\u22252 = 2 \u2225\u03b4\u22252 = 3 \u2225\u03b4\u22252 = 4 RMSE SSIM RMSE SSIM RMSE SSIM RMSE SSIM RMSE SSIM Standard 0.112 0.888 0.346 0.417 0.493 0.157 0.561 0.057 0.596 0.014 FO-DS 0.143 0.781 0.168 0.703 0.221 0.544 0.278 0.417 0.331 0.337 ZO-DS 0.197 0.521 0.217 0.474 0.262 0.373 0.313 0.284 00.356 0.225 FO-AE-DS 0.139 0.792 0.162 0.717 0.215 0.554 0.274 0.421 0.329 0.341 ZO-AE-DS 0.141 0.79 0.164 0.718 0.217 0.551 0.277 0.42 0.33 0.339 Table 5: Performance of image reconstruction using different methods at various attack scenarios. Here \u2018standard\u2019 refers to the original image reconstructor without making any robusti\ufb01cation. Four robusti\ufb01cation methods are presented including FO-DS, ZO-DS (RGE, q = 192), FO-AE-DS, and ZO-AE-DS (CGE, q = 192). The performance metrics RMSE and SSIM are measured by adversarial example (x + \u03b4), generated by 40-step \u21132 PGD attacks under different values of \u21132 perturbation norm \u2225\u03b4\u22252. (a) Ground truth (b) Standard (c) FO-DS (d) ZO-DS (e) FO-AE-DS (f) ZO-AE-DS Figure 5: Visualization for Image Reconstruction under \u21132 PGD attack (Step = 40, \u03f5 = 1.0 ). Original: base reconstruction network. ZO-DS: RGE with q = 192. ZO-AE-DS: CGE with q = 192 6"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2110.00825v2",
+                    "title": "Recurrent networks improve neural response prediction and provide insights into underlying cortical circuits",
+                    "abstract": "Feedforward CNN models have proven themselves in recent years as\nstate-of-the-art models for predicting single-neuron responses to natural\nimages in early visual cortical neurons. In this paper, we extend these models\nwith recurrent convolutional layers, reflecting the well-known massive\nrecurrence in the cortex, and show robust increases in predictive performance\nover feedforward models across thousands of hyperparameter combinations in\nthree datasets of macaque V1 and V2 single-neuron responses. We propose the\nrecurrent circuit can be conceptualized as a form of ensemble computing, with\neach iteration generating more effective feedforward paths of various path\nlengths to allow a combination of solutions in the final approximation. The\nstatistics of the paths in the ensemble provide insights to the differential\nperformance increases among our recurrent models. We also assess whether the\nrecurrent circuits learned for neural response prediction can be related to\ncortical circuits. We find that the hidden units in the recurrent circuits of\nthe appropriate models, when trained on long-duration wide-field image\npresentations, exhibit similar temporal response dynamics and classical\ncontextual modulations as observed in V1 neurons. This work provides insights\nto the computational rationale of recurrent circuits and suggests that neural\nresponse prediction could be useful for characterizing the recurrent neural\ncircuits in the visual cortex.",
+                    "authors": "Yimeng Zhang, Harold Rockwell, Sicheng Dai, Ge Huang, Stephen Tsou, Yuanyuan Wei, Tai Sing Lee",
+                    "published": "2021-10-02",
+                    "updated": "2022-11-13",
+                    "primary_cat": "cs.NE",
+                    "cats": [
+                        "cs.NE"
+                    ],
+                    "main_content": "Introduction Feed-forward deep neural networks have been shown to be effective models for modeling and predicting neural responses of early visual areas in the brain [1\u20136]. However, abundant recurrent connections exist within and between visual areas and are crucial to generating neural responses [7, 8]. While biologically-inspired recurrent circuits incorporated in convolution neural networks have been shown to be more ef\ufb01cient than purely feedforward models, accomplishing comparable performance in image classi\ufb01cation tasks with less layers and less parameters [9\u201311], their underlying computational rationales remains to be understood. In this paper, we investigated the use of recurrent circuits in neural system identi\ufb01cation with the following motivating questions: (1) Since neural data in the visual cortex are generated by recurrent circuits, would the use of a similar architecture provide a better model for predicting neural response? (2) What are the computational rationales and network features responsible for the performance advantages of recurrent circuits? (3) Finally, could neural response prediction using a network model with recurrent circuits provide a useful approach for characterizing and understanding the underlying recurrent neural circuits in the cortex? Preprint. Under review. arXiv:2110.00825v2  [cs.NE]  13 Nov 2022 \fOur task is to perform neural system identi\ufb01cation by training neural network models with recurrent circuits to predict the mean \ufb01ring rates of recorded macaque early visual cortical neurons (V1 and V2) in response to natural input images, when the animals were performing a \ufb01xation task. We separate the model into two stages, one stage including feedforward convolution layers and recurrent circuits to generate the inputs for the target neuron to be modeled, and a second stage, called readout, that models processing by the target neurons and temporal average of their responses to generating the target signals to be predicted. Three sets of data are tested, with distinct stimulus sizes and presentation duration (see Section A and Figure 1 of Appendix for details), to characterize the conditions under which recurrent circuits could exhibit performance gain over feedforward circuits. Neural response prediction models are sensitive to hyperparameters. To establish the computational advantages of the recurrent circuits over feedforward network in general, rather than speci\ufb01c to some optimized hyperparameters and architectural choices, we compare each network with recurrent circuits against its hyper-parameter and parameter matched feedforward model, across a large number of hyperparameters for model size (width and depth), activation functions (ReLU or softplus), model initialization seeds, and the amount of training data. In addition, we investigate four readout paradigms designed to model how the targeted neurons process the temporal responses of the hidden units from the recurrent circuits as well as the computation involved in creating the targeted signals for the network to \ufb01t. Our contributions are three-fold. First, our experimental results conclusively show that recurrent networks perform better than hyperparameter-matched feedforward networks for predicting the responses of neurons in the early visual cortex. This performance gain is general, and not speci\ufb01c to a variety of hyperparameter choices, but more signi\ufb01cant for more limited data size. This performance gain could arise from at least two possible sources: recurrent network (1) being a better or more ef\ufb01cient function approximator or (2) being a better match of the underlying circuitry in the early visual cortex, or both. Our second contribution addresses the \ufb01rst possibility by proposing that recurrent circuits can be conceptualized as a multi-path ensemble, reaping the bene\ufb01ts of ensemble computing [12\u201314]. We demonstrated empirically multi-path ensemble model approximates well the performance of recurrent circuits. This supports the conceptualization but also allows us to dissect the recurrent circuits from the multi-path ensemble perspective. Our third contribution addresses the second possibility by showing that the hidden units of the recurrent circuits exhibit response temporal dynamics and contextual modulation observed in early cortical neurons even when the training stimuli and targeted signals were not designed to create these dynamics and contextual modulation effects. Our \ufb01ndings provide fresh insights to the rationales underlying the computational advantages of recurrent circuits in neural networks and possibly in the brain, and suggest that the response prediction paradigm potentially could be useful for characterizing cortical recurrent circuits in the brain. 2 Related work Recent work has demonstrated that adding biologically-inspired recurrent circuits to feedforward CNNs can lead to performance comparable to that of the state-of-the-art in image classi\ufb01cation, but with signi\ufb01cant less model parameters [9\u201311, 15]. This suggests that recurrent circuits may provide a better or at least more ef\ufb01cient functional approximation. Here, we consider this approach for tackling the problem of neural response prediction. Rather than seeking state-of-the-art models or performance, we seek to assess whether recurrent networks are better than hyperparameter-matched feedforward models across many hyperparameters \u2013 where each recurrent layer is equated to two feedforward layers, to study the performance gain of recurrent computation in a more pure form. In earlier work that incorporated recurrent circuits into CNNs for image classi\ufb01cation, it was also found that the recurrent hidden units can exhibit temporal response dynamics reminiscent of that IT neurons [10] and the learned weight matrices of the recurrent kernels exhibit patterns that might give rise to known physiological responses [16]. These works however did not explicitly perform \"neurophysiological experiments\" on the hidden units to see if the neurons actually exhibit classical contextual modulation that presumably emerge from computation in recurrent circuits. Inspired by the works of [17] on retina, we performed some neurophysiological tests on the hidden units in the recurrent circuits trained to \ufb01t V1 and V2 responses. Our work differs from [17] in that we \ufb01t network to predict mean response to static images, rather than temporal response to dynamic video. While it is possible to use a feedforward network to predict temporal responses to static images, it inherently 2 \fputs the feedforward network to a competitive disadvantage relative to the matched recurrent network which has intrinsic dynamics. It is more compelling to let recurrent network compete in a problem in which feedforward network has been shown to be able to achieve the state-of-the-art performance, than in a problem where a feedforward network is not known to do well. Interestingly, even when we trained the network to \ufb01t mean responses, we found the hidden units exhibit response dynamics that are similar to what we observed in V1 and V2 neurons to static images. Although recurrent networks have been shown to produce performance improvements in a number of goal-driven tasks, the theoretical reasons underlying their success are not well understood. Deep equilibrium theory Bai et al. [18], showing weight-tied deep recurrent networks converge to a \ufb01xed point of the recurrent dynamics, provides an important perspective, but it has not helped us understand performance of the recurrent circuits in our setting. Here, we propose a different perspective, in the form of multi-path ensemble computing, to reason about recurrent computation. Treating recurrent computation as a multi-path ensemble is novel even though multi-path ensembles have been evoked to explain deep networks with residual connections [19, 20], [21], [22, 23], [24], and DenseNet [19]. Liao and Poggio [25] did suggest the equivalence between a speci\ufb01c type of recurrent networks with weight-tied ResNets, and Chen et al. [26] formulated deep weight-tied ResNets to approximate the dynamics of continuous recurrent models speci\ufb01ed in ordinary differential equations. Our work explicitly connects recurrent models to multi-path ensemble models and explores the use of this reformulation to characterize recurrent networks in terms of path length statistics to understand the differential performance of the different models. 3 Methods 3.1 Comparison between the recurrent circuit models and feedforward models in neural response prediction Our \ufb01rst objective is to evaluate whether and what recurrent network models perform better in predicting the mean \ufb01ring rate of individual neurons than feedforward networks of matched model parameter sizes and hyperparameters. Datasets: In this neural system identi\ufb01cation problem, We train each model on three datasets independently. All three sets contains V1 and V2 neurons\u2019 responses to natural images recorded using multi-electrode arrays from macaque monkeys performing a \ufb01xation task. We call them the RF-S, WF-S, and WF-L datasets. RF-S\u2019s stimuli were presented in an aperture close to the size of the classical receptive \ufb01eld (RF) and in short (S) duration (60 ms) each in rapid succession. The RF-S data are public-domain data under a Creative Commons license, from [5] contains 115 V1 neurons\u2019 responses to 7250 images. The WF-S and WF-L data were collected by us, containing 79 and 34 V1/V2 neurons responding to 8000 and 2250 images respectively in response to stimuli presented in 8o and 10o visual angle aperture (wide-\ufb01eld (WF)) respectively. The presentation duration of each stimulus is 47 ms for the WF-S dataset, using the same paradigm as in [5]), and 500 ms for the WF-L dataset. Each stimulus was repeated about 10 times in both of these sets. These datasets are characterized by different extents of spatial and temporal contextual modulation. WF-L involves a large spatial context and long temporal processing for each image, thus signi\ufb01cant contextual modulation is expected. WF-S involves extensive spatial context but does not allow much temporal processing. RF-S, with minimal spatial context and temporal integration, is expected to involve the least amount of contextual modulation. We expect the recurrent circuits learned will strongly depend on the degree of contextual modulation. Note that the WF-L and WF-S datasets include both V1 and V2 neurons. Feedforward Baselines: Our baseline feedforward models are based on those in Klindt et al. [2], Cadena et al. [5] (Fig. 1a) which are state of the art in predicting neural responses of early visual areas. These models consist of a \"shared core\" of convolutional layers used for the whole dataset, followed by a factorized readout for each individual neuron. We trained feedforward models of many different model sizes and hyperparameters; each baseline model is a feedforward model (Fig. 1a) that contains a batch normalization [27] layer, a set of convolutional processing blocks (CPBs), as well as pooling, fully connected, and activation layers. The hyperparameters varied include amount of training data (25%, 50%, and 100%), activation function (ReLU or Softplus), model size (number of layers/channels), and others, resulting in at least 288 feedforward models for each dataset. 3 \fRecurrent Models: For each FF model, we will compare it with recurrent networks (with the same hyperparameters, and parameters) with the simplest recurrent interaction, in which a weighted sum (by a lateral interaction kernel) of all the neurons in a neuron\u2019s 3 x 3 neighborhood at each iteration is added to its response in the previous iteration. Since the CPB\u2019s kernel is 3x3xk where k is the number of feature channels, the lateral interaction kernel is thus also 3x3xk. From the perspective of model size, each Recurrent Convolutional Processing Block (RCPB) in the recurrent model is equivalent in the number of parameters to two convolution processing blocks (CPB) in the FF baseline model as shown in (Fig. 1b,d). A recurrent convolutional processing block (RCPB) extends regular CPB by having two inputs: a bottom-up input from the previous stage of the model in the current iteration and a lateral input from the output of the RCPB in the previous iteration (Fig. 1d). The batch normalization parameters are learned separately for each inference iteration, following the design of Spoerer et al. [15]. We experimented with networks with one RCPB and with two RCPBs. Two-RCPB models are parameter-matched to feedforward models with four CPBs. During model inference, information \ufb02ows over each CPB only once, whereas it \ufb02ows over each RCPB for a certain number of iterations. The number of iterations allowed is a model hyperparameter; when the number of model iterations is 1, a recurrent model is effectively a feedforward one. For each FF model and for each type of recurrent model architecture type, we trained six recurrent models, one for each number of iterations, with the number of iterations range from 2 to 7. Note that the model with 1 iteration is effectively a feedforward model but with less parameters than the feedforward model being compared, and thus will not be included in the comparison. Input BN CPB k=9 CPB k=3 Conv BN Act Pool FC Act Output Input Output Conv Feed-forward BN Act Input @ t Output @ t Conv Lateral Elementwise Sum Output @ t-1 CPB k=3 Input BN CPB k=9 RCPB k=3 Pool FC Act Output a. feed-forward model b. recurrent model (no-avg readout) Intermediate output @ t = T convolutional phase readout phase c. Convolutional Processing Block d. Recurrent Convolutional Processing Block T iterations convolutional phase readout phase Figure 1: Models explored in this study. (a) shows an example baseline feedforward model. (b) shows an example recurrent model under the no-avg readout mode. (c), (d) show a convolutional processing block (CPB) and a recurrent convolutional processing block (RCPB), respectively. The kernel size of a (R)CPB is denoted by k in (a),(b). In this \ufb01gure, we have in total three convolutional blocks for the feedforward model (a) but only two blocks for the recurrent one (b) because a RCPB equals two CPBs in terms of model size; throughout this study we kept the \ufb01rst layer of a recurrent model to be feedforward per pilot experiments. Variations in Architecture and Readout Modes: For both feedforward and recurrent models, the model inference can be divided into two phases: convolutional and readout phases (Fig. 1). The convolutional phase contains a BN layer followed by (recurrent) convolutional processing blocks, each of which contains convolution, BN, and activation layers; the readout phase takes the output of the \ufb01nal convolutional block through average pooling, factorized fully-connected [2], and activation layers to get the \ufb01nal model output. The readout phase represents the operation performed by the neuron to be modeled on the bottom-up input signals provided by the convolution phase, as well as the operation performed by the experimenter, averaging the temporal response in our case, to produce the teaching signals to be \ufb01tted by the model. There are four types of readout paradigms in the readout phase for each of the recurrent models (Fig 2), and the learnable parameters are the same for the four readout modes, as well as that of the feedforward counterpart. For a model with T iteration, its RCPB running for T iterations will generate T intermediate outputs. The target neurons integrate and process these outputs differently in the four readout modes. no-avg, simply uses the output after the T iteration, as typical for convolutional RNNS [11]. early-avg averages the RCPB\u2019s outputs across time and feeds the mean to the readout phase after the T iteration, as was done in some bio-inspired RNNs for object recognition [15]. These two readout modes do not generate a temporal response at the ACT stage of the readout phase. In both late-avg and 2-avg modes, the target neurons process the output of the RCPB units in \"real-time\" at every iteration and produce an output at the ACT stage at every iteration, 4 \fhence generating a temporal response for the targeted neuron. The temporal responses of the targeted neurons are then averaged to produce the mean response to match the teaching signals. Thus the Ave operation at the end models the temporal averaging that had been performed on the neural data. The main difference between them is that the target neurons process the output of the RCPB independently at each iteration in the late-avg case, but process the cumulative average of the output of the RCPB up to that iterations in the 2-avg case. Thus, target neurons in 2-avg remembers the history of the RCPB\u2019s responses for each stimulus up to that iteration when generating their responses at that time point. The four readout modes represent our hypotheses or assumptions on the computations underling the generation of the experimental neural data. Neural response prediction performance can serve as a metric to evaluate these assumptions. In summary, for each feedforward model, there are 24 hyperparameter-match recurrent counterparts ( six numbers of iterations (2 to 7), and four readout modes). Thus, for the 288 feedforward models tested for each dataset, there would be 6912 recurrent models to be compared against for each dataset. c. late-avg RCPB k=3 Pool FC Act Prediction @ t=1 Intermediate output @ t = 1 Pool FC Act Prediction @ t=T \u2026 Intermediate output @ t = T Avg Output T iterations b. early-avg Input RCPB k=3 Pool FC Act Output T iterations Avg Intermediate output @ t = 1 Intermediate output @ t = T d. 2-avg RCPB k=3 Pool FC Act Prediction @ t=1 Intermediate  output @ t = 1 Pool FC Act Prediction @ t=T \u2026 Intermediate  output @ t = T Avg Output T iterations Avg Avg \u2026 Cumulative avg @ t = 1 Cumulative avg @ t = T \u2026 Input \u2026 Input \u2026 RCPB k=3 Pool FC Act Output Intermediate output @ t = T T iterations Input \u2026 a. no-avg Figure 2: Readout modes no-avg (a), early-avg (b), late-avg (c), and 2-avg (d) explored in this study. The simplest one (no-avg) is also shown earlier in Fig. 1b. The style of the \ufb01gure follows that of Fig. 1 and the convolutional phases which are the same across readout modes are omitted. Model training and evaluation: For each data set, roughly 64%, 16%,and 20% of images were allocated for training, validation, and testing, respectively. The validation set was used to control early stopping of the models, since we did not select hyperparameters. We constructed and trained all models using PyTorch [28] on an in-house computing cluster with 4-GPU (NVIDIA GeForce GTX 1080 Ti or similar) nodes. and followed Klindt et al. [2] on data preprocessing and model optimization details. Given a trained model, we use average CC2 norm [4, 29] over all neurons and all testing images (20 % of all samples) to quantify the model\u2019s performance on each data set. Conceptually, CC2 norm measures the fraction of variance explained by the model, with trial-to-trial variance in neural responses discounted. To get the CC2 norm for a neuron, we \ufb01rst compute the raw Pearson correlation CCraw between the ground truth timeand trial-averaged neural responses and the model responses, then we divide CCraw by CCmax, which estimates the maximal Pearson correlation coef\ufb01cient an ideal model can achieve given the noise in the neural data [29, 30], followed by squaring. 3.2 Multi-path reformulation of recurrent circuit models After demonstrating networks with recurrent circuits in the appropriate readout modes consistently outperform their hyperparameter-matched feedforward counterparts, we develop an approach to characterize and understand the computational rationales and network properties that contribute to the performance gain provided by the recurrent circuits in this task. We propose that a recurrent circuit can be conceptualized as a multi-path ensemble, with each iteration generating additional and longer paths. Ensemble computing is known to have the so-called diversity bonus, allowing a network to combine many weak classi\ufb01ers or approximators to build a stronger classi\ufb01er or approximators [12\u201314]. We proceed to reformulate the recurrent model\u2019s recurrent processing blocks (RCPBs) into an approximately equivalent multi-path ensemble with multiple feedforward paths of different lengths. The weights of the kernels across the paths are shared in order to keep the model parameters constant. Fig. 4 shows a one-RCPB network\u2019s information \ufb02ow path at the \ufb01rst three iterations. With each iteration, the network adds a longer path. Fig. 4b approximates information \ufb02ow as summations of simpler individual feedforward paths with shared components. Decomposing the operation of the recurrent circuit into into a multi-path ensemble allow us to dissect the recurrent circuits by eliminating paths as well as compare the path statistics of the different models or readout modes to understand their performance difference. We found two path statistics 5 \fparticularly illuminating. The \ufb01rst is the weighted average lengths of the paths of the multi-path ensemble, and the second is the diversity of the distributions of the weighted path lengths. Each component box in Fig. 4 can be associated with a certain weight, which is essentially a gain factor, representing the ratio of the magnitude of each box\u2019 output over that of its input. Parameters learned along different paths have different magnitudes which should be related to the relative contribution of the different paths to the \ufb01nal output. The product of the weights of all the box components along a particular path gives the total weight of that path. See Appendix F for details and an example of the weight computation. With these weights, we can compute the effective path length and path distribution diversity of the multi-path ensemble for the different model architectures. 3.3 Evaluation of response dynamics and contextual modulation in the hidden units Our third objective is to determine the extent to which the recurrent circuits learned by our model produce response properties known to belong to recurrent circuits in visual cortex. We evaluated three features of the recurrent circuits learned by our recurrent models: their average temporal dynamics, the extent of end-stopping they show, and the extent of surround suppression they show. Neurons in early visual cortex typically show a rapid, high responses to the onset of a static stimulus, which then decays and plateaus at a lower level [31]. V1 neurons additionally are known to exhibit end-stopping (suppression by oriented bars longer than the preferred length) [32] and surround suppression (suppression by sine-wave gratings larger than the preferred size) [33]. We present oriented bars and gratings of varying sizes to our trained recurrent models and extracted the responses of centrally-positioned hidden units in the RCPBs, whose place in the recurrent circuit corresponds to super\ufb01cial neurons in V1 or V2. We then examine the temporal dynamics of these responses, and their tuning to the length of bars and size of gratings (at optimal spatial frequencies and orientations), averaged over 150-200 units across several hyperparameter values. See Appendix C for more details. 4 Results 4.1 Recurrent models outperformed similarly-sized feed-forward models For models trained for each of the three datasets, we compare the feedforward models with their 24 types of parameter-matched recurrent counterparts (4 readout modes, 6 iteration numbers). Figs. 3B shows three plots. Each plot summarizes the performance of the 24 model types for each dataset, trained with 100% training data. The averaged performance of all the tested feedforward models is indicated by the horizontal solid line with the dotted line indicating 1 standard error. The readout modes are indicated by different colors. The mean performance of 48 recurrent models of this type of models with 7-iteration and 2-ave read-out mode is circled in the WF-L plot, with the vertical bar indicating standard error. The performances of the 48 recurrent models correspond to the 48 green dots in Fig. 3A. Each green dot compares the performance of one recurrent model of this type, trained on 100% of the training set, against its parameter and hyperparameter-matched feedforward model, i.e. the recurrent model with 1 RCPB is matched with a feedforward model with 2 CPBs, with the same activation functions, loss, etc. Fig. 3A also shows the prediction performances of these models when only 50% and 25% of the training sets were used, revealing the percentage improvement is more signi\ufb01cant when training data is more limited. We perform pairwise T-test to evaluate the signi\ufb01cance of prediction performance gain for each model type for different amount of the training data, and found the performance gain is signi\ufb01cantly greater for the WF-L-trained models (performance gain is 11%, 6% and 5% for 25%, 50% and 100% training data) than the WF-S-trained models (performance gain is 6.5%, 3.5% and 2% for 25%, 50% and 100% training data). The performance gain for RF-S trained model is signi\ufb01cant but small. In short, everything being equal, the 1-RCPB models and the 2-RCPB models performs consistently better than their 2-CPB and 4-CPB feedforward counterparts respectively, except for the 1-RCPB models trained for the RF-S data. For completeness, we showed all 24 panels for each of the three datasets in Appendix L. Across all three datasets, we observe that the 2-avg (red curve) and late-avg (green curve) readout mode models consistently increase for models of more iterations, and outperform the other models for higher iteration numbers. We suggest this is likely because the temporal averaging of the neural response is explicitly incorporated in these readout modes. 6 \fFigure 3: Recurrent models vs hyperparameter-matched feedforward models on each dataset. Panel A illustrates, for 7-iteration 2-avg models on the WF-L data (circled in panel B), the consistent improvement of recurrent models over feedforward models across hyperparameters; each point represents the average performance of two seeds of a model with a particular hyperparameter combination. Panel B summarizes this data, with each point in each plot corresponding to the average y-axis value among all the green points in Panel A. For all datasets, the late-avg and 2-avg readout modes with a high number of iterations consistently outperform parameter-matched feedforward models, and show a clear trend of performance gain for models with larger number of iterations (while the no-avg readout mode decreases). 4.2 Multi-path approximations work and reveal varying effects of paths of different lengths To verify the approximation of the multi-path ensemble to the original recurrent models, we compare their neural response prediction performance. Figure 4c shows that their performance metrics are highly correlated across hyperparameter values (r=0.88). Furthermore, the effective average path lengths of the recurrent models and that of the multi-path reformulation are also highly correlated (r=0.96). This gives us con\ufb01dence that the approximation is good and so we can draw conclusions about the corresponding recurrent models by analyzing the multi-path ensembles. b. corresponding  approximately  equivalent ensembles  summing di\ufb00erent  paths Conv\ufb00 Input BN1 Conv\ufb00 Input BN1 Conv\ufb00 Convl BN2 Sum Conv\ufb00 Input BN1 Conv\ufb00 Convl BN2 Sum Conv\ufb00 Convl Sum BN3 Conv\ufb00 Input BN1 Conv\ufb00 Convl BN2 BN2 BN3 Conv\ufb00 Input BN1 Conv\ufb00 Convl BN2 Conv\ufb00 Convl BN3 BN3 BN2 Convl Conv\ufb00 Input BN1 a. 1 RCPB  at iterations  1, 2, 3 Sum Sum Input BN CPB k=9 RCPB k=3 T iterations Readout Act Act Act Act Act Act Act Act Act Act Act Act Act Act Act Act c d Figure 4: Recurrent computation approximately understood as summation of feed-forward chains of different lengths. (a) shows the information \ufb02ows of a 1-layer RCPB at iterations 1, 2, and 3 during model inference; the width of a black line is proportional to the number of information \ufb02ows passing between adjacent components and the width of a component\u2019s border is proportional to the the number of information \ufb02ows passing through. (b) shows the information \ufb02ows of corresponding multi-path ensembles and each \ufb02ow sums over feed-forward chains of different depths. (c) compares multi-path models and original recurrent models in terms of model performance. (d) compares the two in terms of average path length of their multi-path ensembles. Models shown here were trained on all the training data. See Appendix F for details and additional results. The multi-path ensemble reformulation allows us to analyze and dissect the recurrent circuits from the perspective of path-lengths and path distributions. Figure 5 panels A and B recast the performance graphs of Figure 3B as a function of effective path lengths instead of iteration numbers for the WF-L and RF-S datasets (WF-S results are somewhere in between and shown in Appendix I. The graph showed that at approximately the same averaged lengths, as indicated by the circled data points, the two readout modes that explicitly models the temporal averaging of the responses at the end, fare better than the early-avg, and far better than the no-avg models. 7 \fFigure 5: Path length and diversity across readout modes and datasets. Panels A and B show average performance as a function of path length for 2-layer recurrent models on the WF-L and RF-S datasets, respectively. C shows the results of ablating paths of one, two, or three lengths for models trained on WF-L (top) and RF-S (bottom), with shorter paths mattering more for the RF-S data (additional results in Appendix K). Panels D and E show distributions of path lengths for the best-performing iteration numbers of each readout mode on the WF-L and RF-S datasets respectively. The distributions are more biased towards shorter paths for the RF-S models, though in all cases no-avg has a strong bias towards the longest path. Given that the notion of diversity is important for ensemble computing, we investigate whether the distribution of the contribution of paths of different lengths, beyond just the average length, plays a role in the differential performance of the four readout modes. Figure 5 panels D and E show the distribution of contribution by the paths of different lengths for each readout mode at similar averaged path lengths (the circled points in Figure 5A and B). We found that the better performing model architecture tend to be more evenly distributed in path-lengths, with a stronger emphasis on the earlier iterations particularly visible for the RF-S dataset. The contribution of the longer paths becomes more signi\ufb01cant with the increase in the spatial context of the stimuli and in the duration of the stimulation, which would allow a greater amount of contextual modulation. This \ufb01nding is also supported by the ablation studies in panel C, the performance of multi-path models retrained with paths of one, two or three consecutive lengths removed (since removing just one length has a very small effect). For the RF-S dataset, ablating the very shortest paths produces the largest performance decrease, while slightly longer paths (i.e. lengths 2 and 3) are more important for the WF-L trained models. Results for models trained on subsets of the WF datasets containing only V1 neurons, and the full WF-S data, in Appendix I and J, lie somewhat in-between, suggesting that spatial and temporal context are important contributors to the models of neural response. 4.3 Response dynamics and contextual modulation are most realistic for 2-avg models trained on long-presentation data Figure 6 contains results from the contextual modulation experiments described in Section 3.3 for 7-iteration models of the different readout modes trained on the WF-L data. The 2-avg models were the only ones to exhibit all three tested phenomena; the failed results for each of the other modes are shown in the bottom row of panels A, B, and C. In addition, the feedforward component of the 2-avg models, as well as those trained on the RF-S data, fail to exhibit some phenomena (Appendix C). Figure 6A shows the averaged temporal response to sine-wave grating stimuli; the initial burst followed by a decay to a steady state of the 2-avg matches visual cortical neurons [31]. Two well-observed physiological effects that have been thought to arise from recurrent circuits are end-stopping and surround suppression. End-stopping is the reduction of neural responses to a bar of the optimal orientation of the neuron when the bar\u2019s length exceeds the size of the receptive \ufb01eld or is longer than a preferred length [32]. Surround suppression is the reduction of neural responses when the size of the sine-wave grating of a neuron\u2019s preferred orientation is larger than the receptive \ufb01eld, encroaching on the receptive \ufb01eld\u2019s near and/or far surround [33]. Figure 6B and C display the average tuning of model units to bars and gratings as a function of length/size, showing the expected rise-and-fall for 2-avg units, but not for early-avg and late-avg respectively. In addition, the 8 \faverage RF size of about 1.2 degrees of visual angle (\u02dc 9 pixels in the models\u2019 input) is roughly the expected RF size of V1 neurons at our eccentricity, and the average level of surround modulation, compared to that of real neurons in [34], is quite close (Figure 6D and E). Figure 6: Results of the WF-L models on select neurophysiological experiments. The top row shows the responses of units from the RCPB of 1-layer, 7-iteration 2-avg models across time (A), to bars of increasing length (B), and to sine-wave gratings of increasing diameter (C). Values are averaged over all units from multiple models that showed sustained responses to the stimuli; shaded bands represent \u00b11 standard error. The bottom row shows the responses of units from no-avg (A), early-avg (B), and late-avg (C) models in the experiments for their corresponding columns. Each of these readout modes failed to show responses qualitatively matching known V1 response properties in the displayed experiment. Panels D and E show the distribution and average value of surround modulation indices across neurons for our models\u2019 units (D) and real V1 neurons (E, from [34]). 5 Discussion In this study, we showed that recurrent models with the appropriate computational architecture signi\ufb01cantly outperform feedforward models of matched size and hyperparameters in scenarios where neural responses are signi\ufb01cantly modulated by the spatial and temporal context. The performance gain is signi\ufb01cant and consistent, particularly when the readout mode takes into account the experimenter\u2019s temporal average of the neural responses in constructing the target responses. Furthermore, it is reassuring that the hidden units in 2-avg models exhibit similar temporal dynamics and contextual modulation as observed in early visual cortical neurons in neurophysiological experiments, even though the models are trained only to predict mean responses with no constraints imposed explicitly to bring about these effects. Taken together, these results suggest that neural response prediction can provide a useful paradigm for characterizing the underlying recurrent circuits of the visual cortex. By conceptualizing and reformulating recurrent networks as multi-path ensembles, we suggest recurrent circuits might derive their computational advantages from the notion of ensemble computing. More importantly, the reformulation allows us to perform analysis of the contribution of different paths based on path-length statistics and ablation analysis. The results of our analysis suggest models with a more uniform distribution of path-lengths and paths of longer lengths are needed to model greater amount of contextual modulation. The recognition that recurrent circuits can be conceptualized as multi-path ensembles might suggest a new direction of research to investigate how more powerful multi-path ensemble can be constructed and optimized by using recurrent network, both for machine learning and for understanding the brain. There are several limitations to our work. While using small models is standard for neural response prediction, and allows us to explore a wide variety of hyperparameter combinations, the method of multi-path reformulation developed for such shallow models might be dif\ufb01cult to transfer to the much deeper models generally used in computer vision, or to other domains in which recurrent networks are applied. This means that the positive societal bene\ufb01ts of our work resulting from greater model interpretability may be limited, but we do not foresee any negative impacts from this formulation or better characterization of early visual cortical circuits. The hyperparameters we sweep over may not include those that result in optimal performance, and architectural improvements like the Gaussian 9 \freadout of [35] may reduce the improvement from recurrent connections that we \ufb01nd. In addition, our recurrent circuit models are far simpler than the true circuits in visual cortex, which include a detailed laminar structure, extensive feedback, dynamically varying connectivity, and more. Thus, there is a limit to how well neural circuits can be characterized by our current models."
+                }
+            ]
+        },
+        "edge_feat": {}
+    }
+}
\ No newline at end of file