id
int64
1
42.8k
arxiv_id
stringlengths
10
13
base_arxiv_id
float64
1.11k
2.51k
version
float64
1
14
title
stringlengths
8
224
abstract
stringlengths
51
4.03k
submit_date
stringdate
2011-09-01 00:00:00
2025-08-13 00:00:00
metadata
stringlengths
283
56.2k
42,703
2310.05691v1
2,310.05691
1
Climate-sensitive Urban Planning through Optimization of Tree Placements
Climate change is increasing the intensity and frequency of many extreme weather events, including heatwaves, which results in increased thermal discomfort and mortality rates. While global mitigation action is undoubtedly necessary, so is climate adaptation, e.g., through climate-sensitive urban planning. Among the most promising strategies is harnessing the benefits of urban trees in shading and cooling pedestrian-level environments. Our work investigates the challenge of optimal placement of such trees. Physical simulations can estimate the radiative and thermal impact of trees on human thermal comfort but induce high computational costs. This rules out optimization of tree placements over large areas and considering effects over longer time scales. Hence, we employ neural networks to simulate the point-wise mean radiant temperatures--a driving factor of outdoor human thermal comfort--across various time scales, spanning from daily variations to extended time scales of heatwave events and even decades. To optimize tree placements, we harness the innate local effect of trees within the iterated local search framework with tailored adaptations. We show the efficacy of our approach across a wide spectrum of study areas and time scales. We believe that our approach is a step towards empowering decision-makers, urban designers and planners to proactively and effectively assess the potential of urban trees to mitigate heat stress.
2023-10-09
{'id': '2310.05691v1', 'title': 'Climate-sensitive Urban Planning through Optimization of Tree Placements', 'abstract': 'Climate change is increasing the intensity and frequency of many extreme\nweather events, including heatwaves, which results in increased thermal\ndiscomfort and mortality rates. While global mitigation action is undoubtedly\nnecessary, so is climate adaptation, e.g., through climate-sensitive urban\nplanning. Among the most promising strategies is harnessing the benefits of\nurban trees in shading and cooling pedestrian-level environments. Our work\ninvestigates the challenge of optimal placement of such trees. Physical\nsimulations can estimate the radiative and thermal impact of trees on human\nthermal comfort but induce high computational costs. This rules out\noptimization of tree placements over large areas and considering effects over\nlonger time scales. Hence, we employ neural networks to simulate the point-wise\nmean radiant temperatures--a driving factor of outdoor human thermal\ncomfort--across various time scales, spanning from daily variations to extended\ntime scales of heatwave events and even decades. To optimize tree placements,\nwe harness the innate local effect of trees within the iterated local search\nframework with tailored adaptations. We show the efficacy of our approach\nacross a wide spectrum of study areas and time scales. We believe that our\napproach is a step towards empowering decision-makers, urban designers and\nplanners to proactively and effectively assess the potential of urban trees to\nmitigate heat stress.', 'authors': ['Simon Schrodi', 'Ferdinand Briegel', 'Max Argus', 'Andreas Christen', 'Thomas Brox'], 'published': '2023-10-09 13:07:23+00:00', 'categories': ['cs.CV', 'physics.ao-ph'], 'url': 'http://arxiv.org/abs/2310.05691v1'}
42,700
2205.11787v3
2,205.11787
3
Quadratic models for understanding catapult dynamics of neural networks
While neural networks can be approximated by linear models as their width increases, certain properties of wide neural networks cannot be captured by linear models. In this work we show that recently proposed Neural Quadratic Models can exhibit the "catapult phase" [Lewkowycz et al. 2020] that arises when training such models with large learning rates. We then empirically show that the behaviour of neural quadratic models parallels that of neural networks in generalization, especially in the catapult phase regime. Our analysis further demonstrates that quadratic models can be an effective tool for analysis of neural networks.
2022-05-24
{'id': '2205.11787v3', 'title': 'Quadratic models for understanding catapult dynamics of neural networks', 'abstract': 'While neural networks can be approximated by linear models as their width\nincreases, certain properties of wide neural networks cannot be captured by\nlinear models. In this work we show that recently proposed Neural Quadratic\nModels can exhibit the "catapult phase" [Lewkowycz et al. 2020] that arises\nwhen training such models with large learning rates. We then empirically show\nthat the behaviour of neural quadratic models parallels that of neural networks\nin generalization, especially in the catapult phase regime. Our analysis\nfurther demonstrates that quadratic models can be an effective tool for\nanalysis of neural networks.', 'authors': ['Libin Zhu', 'Chaoyue Liu', 'Adityanarayanan Radhakrishnan', 'Mikhail Belkin'], 'published': '2022-05-24 05:03:06+00:00', 'categories': ['cs.LG', 'math.OC', 'stat.ML'], 'url': 'http://arxiv.org/abs/2205.11787v3'}
42,705
2302.02257v4
2,302.02257
4
Multi-Source Diffusion Models for Simultaneous Music Generation and Separation
In this work, we define a diffusion-based generative model capable of both music synthesis and source separation by learning the score of the joint probability density of sources sharing a context. Alongside the classic total inference tasks (i.e., generating a mixture, separating the sources), we also introduce and experiment on the partial generation task of source imputation, where we generate a subset of the sources given the others (e.g., play a piano track that goes well with the drums). Additionally, we introduce a novel inference method for the separation task based on Dirac likelihood functions. We train our model on Slakh2100, a standard dataset for musical source separation, provide qualitative results in the generation settings, and showcase competitive quantitative results in the source separation setting. Our method is the first example of a single model that can handle both generation and separation tasks, thus representing a step toward general audio models.
2023-02-04
{'id': '2302.02257v4', 'title': 'Multi-Source Diffusion Models for Simultaneous Music Generation and Separation', 'abstract': 'In this work, we define a diffusion-based generative model capable of both\nmusic synthesis and source separation by learning the score of the joint\nprobability density of sources sharing a context. Alongside the classic total\ninference tasks (i.e., generating a mixture, separating the sources), we also\nintroduce and experiment on the partial generation task of source imputation,\nwhere we generate a subset of the sources given the others (e.g., play a piano\ntrack that goes well with the drums). Additionally, we introduce a novel\ninference method for the separation task based on Dirac likelihood functions.\nWe train our model on Slakh2100, a standard dataset for musical source\nseparation, provide qualitative results in the generation settings, and\nshowcase competitive quantitative results in the source separation setting. Our\nmethod is the first example of a single model that can handle both generation\nand separation tasks, thus representing a step toward general audio models.', 'authors': ['Giorgio Mariani', 'Irene Tallini', 'Emilian Postolache', 'Michele Mancusi', 'Luca Cosmo', 'Emanuele Rodolà'], 'published': '2023-02-04 23:18:36+00:00', 'categories': ['cs.SD', 'cs.LG', 'eess.AS'], 'url': 'http://arxiv.org/abs/2302.02257v4'}
42,706
2309.17275v1
2,309.17275
1
Utility-based Adaptive Teaching Strategies using Bayesian Theory of Mind
Good teachers always tailor their explanations to the learners. Cognitive scientists model this process under the rationality principle: teachers try to maximise the learner's utility while minimising teaching costs. To this end, human teachers seem to build mental models of the learner's internal state, a capacity known as Theory of Mind (ToM). Inspired by cognitive science, we build on Bayesian ToM mechanisms to design teacher agents that, like humans, tailor their teaching strategies to the learners. Our ToM-equipped teachers construct models of learners' internal states from observations and leverage them to select demonstrations that maximise the learners' rewards while minimising teaching costs. Our experiments in simulated environments demonstrate that learners taught this way are more efficient than those taught in a learner-agnostic way. This effect gets stronger when the teacher's model of the learner better aligns with the actual learner's state, either using a more accurate prior or after accumulating observations of the learner's behaviour. This work is a first step towards social machines that teach us and each other, see https://teacher-with-tom.github.io.
2023-09-29
{'id': '2309.17275v1', 'title': 'Utility-based Adaptive Teaching Strategies using Bayesian Theory of Mind', 'abstract': "Good teachers always tailor their explanations to the learners. Cognitive\nscientists model this process under the rationality principle: teachers try to\nmaximise the learner's utility while minimising teaching costs. To this end,\nhuman teachers seem to build mental models of the learner's internal state, a\ncapacity known as Theory of Mind (ToM). Inspired by cognitive science, we build\non Bayesian ToM mechanisms to design teacher agents that, like humans, tailor\ntheir teaching strategies to the learners. Our ToM-equipped teachers construct\nmodels of learners' internal states from observations and leverage them to\nselect demonstrations that maximise the learners' rewards while minimising\nteaching costs. Our experiments in simulated environments demonstrate that\nlearners taught this way are more efficient than those taught in a\nlearner-agnostic way. This effect gets stronger when the teacher's model of the\nlearner better aligns with the actual learner's state, either using a more\naccurate prior or after accumulating observations of the learner's behaviour.\nThis work is a first step towards social machines that teach us and each other,\nsee https://teacher-with-tom.github.io.", 'authors': ['Clémence Grislain', 'Hugo Caselles-Dupré', 'Olivier Sigaud', 'Mohamed Chetouani'], 'published': '2023-09-29 14:27:53+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2309.17275v1'}
42,707
2309.16109v2
2,309.16109
2
Feature Normalization Prevents Collapse of Non-contrastive Learning Dynamics
Contrastive learning is a self-supervised representation learning framework, where two positive views generated through data augmentation are made similar by an attraction force in a data representation space, while a repulsive force makes them far from negative examples. Non-contrastive learning, represented by BYOL and SimSiam, further gets rid of negative examples and improves computational efficiency. While learned representations may collapse into a single point due to the lack of the repulsive force at first sight, Tian et al. (2021) revealed through the learning dynamics analysis that the representations can avoid collapse if data augmentation is sufficiently stronger than regularization. However, their analysis does not take into account commonly-used feature normalization, a normalizer before measuring the similarity of representations, and hence excessively strong regularization may collapse the dynamics, which is an unnatural behavior under the presence of feature normalization. Therefore, we extend the previous theory based on the L2 loss by considering the cosine loss, which involves feature normalization. We show that the cosine loss induces sixth-order dynamics (while the L2 loss induces a third-order one), in which a stable equilibrium dynamically emerges even if there are only collapsed solutions with given initial parameters. Thus, we offer a new understanding that feature normalization plays an important role in robustly preventing the dynamics collapse.
2023-09-28
{'id': '2309.16109v2', 'title': 'Feature Normalization Prevents Collapse of Non-contrastive Learning Dynamics', 'abstract': 'Contrastive learning is a self-supervised representation learning framework,\nwhere two positive views generated through data augmentation are made similar\nby an attraction force in a data representation space, while a repulsive force\nmakes them far from negative examples. Non-contrastive learning, represented by\nBYOL and SimSiam, further gets rid of negative examples and improves\ncomputational efficiency. While learned representations may collapse into a\nsingle point due to the lack of the repulsive force at first sight, Tian et al.\n(2021) revealed through the learning dynamics analysis that the representations\ncan avoid collapse if data augmentation is sufficiently stronger than\nregularization. However, their analysis does not take into account\ncommonly-used feature normalization, a normalizer before measuring the\nsimilarity of representations, and hence excessively strong regularization may\ncollapse the dynamics, which is an unnatural behavior under the presence of\nfeature normalization. Therefore, we extend the previous theory based on the L2\nloss by considering the cosine loss, which involves feature normalization. We\nshow that the cosine loss induces sixth-order dynamics (while the L2 loss\ninduces a third-order one), in which a stable equilibrium dynamically emerges\neven if there are only collapsed solutions with given initial parameters. Thus,\nwe offer a new understanding that feature normalization plays an important role\nin robustly preventing the dynamics collapse.', 'authors': ['Han Bao'], 'published': '2023-09-28 02:23:32+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2309.16109v2'}
42,708
2310.01769v3
2,310.01769
3
How Over-Parameterization Slows Down Gradient Descent in Matrix Sensing: The Curses of Symmetry and Initialization
This paper rigorously shows how over-parameterization changes the convergence behaviors of gradient descent (GD) for the matrix sensing problem, where the goal is to recover an unknown low-rank ground-truth matrix from near-isotropic linear measurements. First, we consider the symmetric setting with the symmetric parameterization where $M^* R^n n$ is a positive semi-definite unknown matrix of rank $r n$, and one uses a symmetric parameterization $XX^$ to learn $M^*$. Here $X R^n k$ with $k > r$ is the factor matrix. We give a novel $(1/T^2)$ lower bound of randomly initialized GD for the over-parameterized case ($k >r$) where $T$ is the number of iterations. This is in stark contrast to the exact-parameterization scenario ($k=r$) where the convergence rate is $(-(T))$. Next, we study asymmetric setting where $M^* R^n_1 n_2$ is the unknown matrix of rank $r \n_1,n_2\$, and one uses an asymmetric parameterization $FG^$ to learn $M^*$ where $F R^n_1 k$ and $G R^n_2 k$. Building on prior work, we give a global exact convergence result of randomly initialized GD for the exact-parameterization case ($k=r$) with an $(-(T))$ rate. Furthermore, we give the first global exact convergence result for the over-parameterization case ($k>r$) with an $(-(^2 T))$ rate where $$ is the initialization scale. This linear convergence result in the over-parameterization case is especially significant because one can apply the asymmetric parameterization to the symmetric setting to speed up from $(1/T^2)$ to linear convergence. On the other hand, we propose a novel method that only modifies one step of GD and obtains a convergence rate independent of $$, recovering the rate in the exact-parameterization case.
2023-10-03
{'id': '2310.01769v3', 'title': 'How Over-Parameterization Slows Down Gradient Descent in Matrix Sensing: The Curses of Symmetry and Initialization', 'abstract': 'This paper rigorously shows how over-parameterization changes the convergence\nbehaviors of gradient descent (GD) for the matrix sensing problem, where the\ngoal is to recover an unknown low-rank ground-truth matrix from near-isotropic\nlinear measurements. First, we consider the symmetric setting with the\nsymmetric parameterization where $M^* \\in \\mathbb{R}^{n \\times n}$ is a\npositive semi-definite unknown matrix of rank $r \\ll n$, and one uses a\nsymmetric parameterization $XX^\\top$ to learn $M^*$. Here $X \\in \\mathbb{R}^{n\n\\times k}$ with $k > r$ is the factor matrix. We give a novel $\\Omega (1/T^2)$\nlower bound of randomly initialized GD for the over-parameterized case ($k >r$)\nwhere $T$ is the number of iterations. This is in stark contrast to the\nexact-parameterization scenario ($k=r$) where the convergence rate is $\\exp\n(-\\Omega (T))$. Next, we study asymmetric setting where $M^* \\in\n\\mathbb{R}^{n_1 \\times n_2}$ is the unknown matrix of rank $r \\ll\n\\min\\{n_1,n_2\\}$, and one uses an asymmetric parameterization $FG^\\top$ to\nlearn $M^*$ where $F \\in \\mathbb{R}^{n_1 \\times k}$ and $G \\in \\mathbb{R}^{n_2\n\\times k}$. Building on prior work, we give a global exact convergence result\nof randomly initialized GD for the exact-parameterization case ($k=r$) with an\n$\\exp (-\\Omega(T))$ rate. Furthermore, we give the first global exact\nconvergence result for the over-parameterization case ($k>r$) with an\n$\\exp(-\\Omega(\\alpha^2 T))$ rate where $\\alpha$ is the initialization scale.\nThis linear convergence result in the over-parameterization case is especially\nsignificant because one can apply the asymmetric parameterization to the\nsymmetric setting to speed up from $\\Omega (1/T^2)$ to linear convergence. On\nthe other hand, we propose a novel method that only modifies one step of GD and\nobtains a convergence rate independent of $\\alpha$, recovering the rate in the\nexact-parameterization case.', 'authors': ['Nuoya Xiong', 'Lijun Ding', 'Simon S. Du'], 'published': '2023-10-03 03:34:22+00:00', 'categories': ['cs.LG', 'math.OC', 'stat.ML'], 'url': 'http://arxiv.org/abs/2310.01769v3'}
42,709
2309.04669v3
2,309.04669
3
Unified Language-Vision Pretraining in LLM with Dynamic Discrete Visual Tokenization
Recently, the remarkable advance of the Large Language Model (LLM) has inspired researchers to transfer its extraordinary reasoning capability to both vision and language data. However, the prevailing approaches primarily regard the visual input as a prompt and focus exclusively on optimizing the text generation process conditioned upon vision content by a frozen LLM. Such an inequitable treatment of vision and language heavily constrains the model's potential. In this paper, we break through this limitation by representing both vision and language in a unified form. Specifically, we introduce a well-designed visual tokenizer to translate the non-linguistic image into a sequence of discrete tokens like a foreign language that LLM can read. The resulting visual tokens encompass high-level semantics worthy of a word and also support dynamic sequence length varying from the image. Coped with this tokenizer, the presented foundation model called LaVIT can handle both image and text indiscriminately under the same generative learning paradigm. This unification empowers LaVIT to serve as an impressive generalist interface to understand and generate multi-modal content simultaneously. Extensive experiments further showcase that it outperforms the existing models by a large margin on massive vision-language tasks. Our code and models are available at https://github.com/jy0205/LaVIT.
2023-09-09
{'id': '2309.04669v3', 'title': 'Unified Language-Vision Pretraining in LLM with Dynamic Discrete Visual Tokenization', 'abstract': "Recently, the remarkable advance of the Large Language Model (LLM) has\ninspired researchers to transfer its extraordinary reasoning capability to both\nvision and language data. However, the prevailing approaches primarily regard\nthe visual input as a prompt and focus exclusively on optimizing the text\ngeneration process conditioned upon vision content by a frozen LLM. Such an\ninequitable treatment of vision and language heavily constrains the model's\npotential. In this paper, we break through this limitation by representing both\nvision and language in a unified form. Specifically, we introduce a\nwell-designed visual tokenizer to translate the non-linguistic image into a\nsequence of discrete tokens like a foreign language that LLM can read. The\nresulting visual tokens encompass high-level semantics worthy of a word and\nalso support dynamic sequence length varying from the image. Coped with this\ntokenizer, the presented foundation model called LaVIT can handle both image\nand text indiscriminately under the same generative learning paradigm. This\nunification empowers LaVIT to serve as an impressive generalist interface to\nunderstand and generate multi-modal content simultaneously. Extensive\nexperiments further showcase that it outperforms the existing models by a large\nmargin on massive vision-language tasks. Our code and models are available at\nhttps://github.com/jy0205/LaVIT.", 'authors': ['Yang Jin', 'Kun Xu', 'Kun Xu', 'Liwei Chen', 'Chao Liao', 'Jianchao Tan', 'Quzhe Huang', 'Bin Chen', 'Chenyi Lei', 'An Liu', 'Chengru Song', 'Xiaoqiang Lei', 'Di Zhang', 'Wenwu Ou', 'Kun Gai', 'Yadong Mu'], 'published': '2023-09-09 03:01:38+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2309.04669v3'}
42,710
2306.08257v1
2,306.08257
1
On the Robustness of Latent Diffusion Models
Latent diffusion models achieve state-of-the-art performance on a variety of generative tasks, such as image synthesis and image editing. However, the robustness of latent diffusion models is not well studied. Previous works only focus on the adversarial attacks against the encoder or the output image under white-box settings, regardless of the denoising process. Therefore, in this paper, we aim to analyze the robustness of latent diffusion models more thoroughly. We first study the influence of the components inside latent diffusion models on their white-box robustness. In addition to white-box scenarios, we evaluate the black-box robustness of latent diffusion models via transfer attacks, where we consider both prompt-transfer and model-transfer settings and possible defense mechanisms. However, all these explorations need a comprehensive benchmark dataset, which is missing in the literature. Therefore, to facilitate the research of the robustness of latent diffusion models, we propose two automatic dataset construction pipelines for two kinds of image editing models and release the whole dataset. Our code and dataset are available at https://github.com/jpzhang1810/LDM-Robustness.
2023-06-14
{'id': '2306.08257v1', 'title': 'On the Robustness of Latent Diffusion Models', 'abstract': 'Latent diffusion models achieve state-of-the-art performance on a variety of\ngenerative tasks, such as image synthesis and image editing. However, the\nrobustness of latent diffusion models is not well studied. Previous works only\nfocus on the adversarial attacks against the encoder or the output image under\nwhite-box settings, regardless of the denoising process. Therefore, in this\npaper, we aim to analyze the robustness of latent diffusion models more\nthoroughly. We first study the influence of the components inside latent\ndiffusion models on their white-box robustness. In addition to white-box\nscenarios, we evaluate the black-box robustness of latent diffusion models via\ntransfer attacks, where we consider both prompt-transfer and model-transfer\nsettings and possible defense mechanisms. However, all these explorations need\na comprehensive benchmark dataset, which is missing in the literature.\nTherefore, to facilitate the research of the robustness of latent diffusion\nmodels, we propose two automatic dataset construction pipelines for two kinds\nof image editing models and release the whole dataset. Our code and dataset are\navailable at \\url{https://github.com/jpzhang1810/LDM-Robustness}.', 'authors': ['Jianping Zhang', 'Zhuoer Xu', 'Shiwen Cui', 'Changhua Meng', 'Weibin Wu', 'Michael R. Lyu'], 'published': '2023-06-14 05:47:15+00:00', 'categories': ['cs.CV', 'cs.CR'], 'url': 'http://arxiv.org/abs/2306.08257v1'}
42,711
2305.19891v4
2,305.19891
4
Dynamic Neighborhood Construction for Structured Large Discrete Action Spaces
Large discrete action spaces (LDAS) remain a central challenge in reinforcement learning. Existing solution approaches can handle unstructured LDAS with up to a few million actions. However, many real-world applications in logistics, production, and transportation systems have combinatorial action spaces, whose size grows well beyond millions of actions, even on small instances. Fortunately, such action spaces exhibit structure, e.g., equally spaced discrete resource units. With this work, we focus on handling structured LDAS (SLDAS) with sizes that cannot be handled by current benchmarks: we propose Dynamic Neighborhood Construction (DNC), a novel exploitation paradigm for SLDAS. We present a scalable neighborhood exploration heuristic that utilizes this paradigm and efficiently explores the discrete neighborhood around the continuous proxy action in structured action spaces with up to $10^73$ actions. We demonstrate the performance of our method by benchmarking it against three state-of-the-art approaches designed for large discrete action spaces across two distinct environments. Our results show that DNC matches or outperforms state-of-the-art approaches while being computationally more efficient. Furthermore, our method scales to action spaces that so far remained computationally intractable for existing methodologies.
2023-05-31
{'id': '2305.19891v4', 'title': 'Dynamic Neighborhood Construction for Structured Large Discrete Action Spaces', 'abstract': 'Large discrete action spaces (LDAS) remain a central challenge in\nreinforcement learning. Existing solution approaches can handle unstructured\nLDAS with up to a few million actions. However, many real-world applications in\nlogistics, production, and transportation systems have combinatorial action\nspaces, whose size grows well beyond millions of actions, even on small\ninstances. Fortunately, such action spaces exhibit structure, e.g., equally\nspaced discrete resource units. With this work, we focus on handling structured\nLDAS (SLDAS) with sizes that cannot be handled by current benchmarks: we\npropose Dynamic Neighborhood Construction (DNC), a novel exploitation paradigm\nfor SLDAS. We present a scalable neighborhood exploration heuristic that\nutilizes this paradigm and efficiently explores the discrete neighborhood\naround the continuous proxy action in structured action spaces with up to\n$10^{73}$ actions. We demonstrate the performance of our method by benchmarking\nit against three state-of-the-art approaches designed for large discrete action\nspaces across two distinct environments. Our results show that DNC matches or\noutperforms state-of-the-art approaches while being computationally more\nefficient. Furthermore, our method scales to action spaces that so far remained\ncomputationally intractable for existing methodologies.', 'authors': ['Fabian Akkerman', 'Julius Luy', 'Wouter van Heeswijk', 'Maximilian Schiffer'], 'published': '2023-05-31 14:26:14+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2305.19891v4'}
42,712
2310.12149v2
2,310.12149
2
Object-aware Inversion and Reassembly for Image Editing
By comparing the original and target prompts, we can obtain numerous editing pairs, each comprising an object and its corresponding editing target. To allow editability while maintaining fidelity to the input image, existing editing methods typically involve a fixed number of inversion steps that project the whole input image to its noisier latent representation, followed by a denoising process guided by the target prompt. However, we find that the optimal number of inversion steps for achieving ideal editing results varies significantly among different editing pairs, owing to varying editing difficulties. Therefore, the current literature, which relies on a fixed number of inversion steps, produces sub-optimal generation quality, especially when handling multiple editing pairs in a natural image. To this end, we propose a new image editing paradigm, dubbed Object-aware Inversion and Reassembly (OIR), to enable object-level fine-grained editing. Specifically, we design a new search metric, which determines the optimal inversion steps for each editing pair, by jointly considering the editability of the target and the fidelity of the non-editing region. We use our search metric to find the optimal inversion step for each editing pair when editing an image. We then edit these editing pairs separately to avoid concept mismatch. Subsequently, we propose an additional reassembly step to seamlessly integrate the respective editing results and the non-editing region to obtain the final edited image. To systematically evaluate the effectiveness of our method, we collect two datasets called OIRBench for benchmarking single- and multi-object editing, respectively. Experiments demonstrate that our method achieves superior performance in editing object shapes, colors, materials, categories, etc., especially in multi-object editing scenarios.
2023-10-18
{'id': '2310.12149v2', 'title': 'Object-aware Inversion and Reassembly for Image Editing', 'abstract': 'By comparing the original and target prompts, we can obtain numerous editing\npairs, each comprising an object and its corresponding editing target. To allow\neditability while maintaining fidelity to the input image, existing editing\nmethods typically involve a fixed number of inversion steps that project the\nwhole input image to its noisier latent representation, followed by a denoising\nprocess guided by the target prompt. However, we find that the optimal number\nof inversion steps for achieving ideal editing results varies significantly\namong different editing pairs, owing to varying editing difficulties.\nTherefore, the current literature, which relies on a fixed number of inversion\nsteps, produces sub-optimal generation quality, especially when handling\nmultiple editing pairs in a natural image. To this end, we propose a new image\nediting paradigm, dubbed Object-aware Inversion and Reassembly (OIR), to enable\nobject-level fine-grained editing. Specifically, we design a new search metric,\nwhich determines the optimal inversion steps for each editing pair, by jointly\nconsidering the editability of the target and the fidelity of the non-editing\nregion. We use our search metric to find the optimal inversion step for each\nediting pair when editing an image. We then edit these editing pairs separately\nto avoid concept mismatch. Subsequently, we propose an additional reassembly\nstep to seamlessly integrate the respective editing results and the non-editing\nregion to obtain the final edited image. To systematically evaluate the\neffectiveness of our method, we collect two datasets called OIRBench for\nbenchmarking single- and multi-object editing, respectively. Experiments\ndemonstrate that our method achieves superior performance in editing object\nshapes, colors, materials, categories, etc., especially in multi-object editing\nscenarios.', 'authors': ['Zhen Yang', 'Ganggui Ding', 'Wen Wang', 'Hao Chen', 'Bohan Zhuang', 'Chunhua Shen'], 'published': '2023-10-18 17:59:02+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2310.12149v2'}
42,713
2308.15812v3
2,308.15812
3
Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models
Aligning large language models (LLMs) with human values and intents critically involves the use of human or AI feedback. While dense feedback annotations are expensive to acquire and integrate, sparse feedback presents a structural design choice between ratings (e.g., score Response A on a scale of 1-7) and rankings (e.g., is Response A better than Response B?). In this work, we analyze the effect of this design choice for the alignment and evaluation of LLMs. We uncover an inconsistency problem wherein the preferences inferred from ratings and rankings significantly disagree 60% for both human and AI annotators. Our subsequent analysis identifies various facets of annotator biases that explain this phenomena, such as human annotators would rate denser responses higher while preferring accuracy during pairwise judgments. To our surprise, we also observe that the choice of feedback protocol also has a significant effect on the evaluation of aligned LLMs. In particular, we find that LLMs that leverage rankings data for alignment (say model X) are preferred over those that leverage ratings data (say model Y), with a rank-based evaluation protocol (is X/Y's response better than reference response?) but not with a rating-based evaluation protocol (score Rank X/Y's response on a scale of 1-7). Our findings thus shed light on critical gaps in methods for evaluating the real-world utility of language models and their strong dependence on the feedback protocol used for alignment. Our code and data are available at https://github.com/Hritikbansal/sparse_feedback.
2023-08-30
{'id': '2308.15812v3', 'title': 'Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models', 'abstract': "Aligning large language models (LLMs) with human values and intents\ncritically involves the use of human or AI feedback. While dense feedback\nannotations are expensive to acquire and integrate, sparse feedback presents a\nstructural design choice between ratings (e.g., score Response A on a scale of\n1-7) and rankings (e.g., is Response A better than Response B?). In this work,\nwe analyze the effect of this design choice for the alignment and evaluation of\nLLMs. We uncover an inconsistency problem wherein the preferences inferred from\nratings and rankings significantly disagree 60% for both human and AI\nannotators. Our subsequent analysis identifies various facets of annotator\nbiases that explain this phenomena, such as human annotators would rate denser\nresponses higher while preferring accuracy during pairwise judgments. To our\nsurprise, we also observe that the choice of feedback protocol also has a\nsignificant effect on the evaluation of aligned LLMs. In particular, we find\nthat LLMs that leverage rankings data for alignment (say model X) are preferred\nover those that leverage ratings data (say model Y), with a rank-based\nevaluation protocol (is X/Y's response better than reference response?) but not\nwith a rating-based evaluation protocol (score Rank X/Y's response on a scale\nof 1-7). Our findings thus shed light on critical gaps in methods for\nevaluating the real-world utility of language models and their strong\ndependence on the feedback protocol used for alignment. Our code and data are\navailable at https://github.com/Hritikbansal/sparse_feedback.", 'authors': ['Hritik Bansal', 'John Dang', 'Aditya Grover'], 'published': '2023-08-30 07:35:32+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CL'], 'url': 'http://arxiv.org/abs/2308.15812v3'}
42,714
2309.10402v2
2,309.10402
2
Minimum width for universal approximation using ReLU networks on compact domain
It has been shown that deep neural networks of a large enough width are universal approximators but they are not if the width is too small. There were several attempts to characterize the minimum width $w_$ enabling the universal approximation property; however, only a few of them found the exact values. In this work, we show that the minimum width for $L^p$ approximation of $L^p$ functions from $[0,1]^d_x$ to $R^d_y$ is exactly $\d_x,d_y,2\$ if an activation function is ReLU-Like (e.g., ReLU, GELU, Softplus). Compared to the known result for ReLU networks, $w_=\d_x+1,d_y\$ when the domain is $R^d_x$, our result first shows that approximation on a compact domain requires smaller width than on $R^d_x$. We next prove a lower bound on $w_$ for uniform approximation using general activation functions including ReLU: $w_d_y+1$ if $d_x<d_y2d_x$. Together with our first result, this shows a dichotomy between $L^p$ and uniform approximations for general activation functions and input/output dimensions.
2023-09-19
{'id': '2309.10402v2', 'title': 'Minimum width for universal approximation using ReLU networks on compact domain', 'abstract': 'It has been shown that deep neural networks of a large enough width are\nuniversal approximators but they are not if the width is too small. There were\nseveral attempts to characterize the minimum width $w_{\\min}$ enabling the\nuniversal approximation property; however, only a few of them found the exact\nvalues. In this work, we show that the minimum width for $L^p$ approximation of\n$L^p$ functions from $[0,1]^{d_x}$ to $\\mathbb R^{d_y}$ is exactly\n$\\max\\{d_x,d_y,2\\}$ if an activation function is ReLU-Like (e.g., ReLU, GELU,\nSoftplus). Compared to the known result for ReLU networks,\n$w_{\\min}=\\max\\{d_x+1,d_y\\}$ when the domain is $\\smash{\\mathbb R^{d_x}}$, our\nresult first shows that approximation on a compact domain requires smaller\nwidth than on $\\smash{\\mathbb R^{d_x}}$. We next prove a lower bound on\n$w_{\\min}$ for uniform approximation using general activation functions\nincluding ReLU: $w_{\\min}\\ge d_y+1$ if $d_x<d_y\\le2d_x$. Together with our\nfirst result, this shows a dichotomy between $L^p$ and uniform approximations\nfor general activation functions and input/output dimensions.', 'authors': ['Namjun Kim', 'Chanho Min', 'Sejun Park'], 'published': '2023-09-19 08:04:48+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2309.10402v2'}
42,715
2309.16298v2
2,309.16298
2
At Which Training Stage Does Code Data Help LLMs Reasoning?
Large Language Models (LLMs) have exhibited remarkable reasoning capabilities and become the foundation of language technologies. Inspired by the great success of code data in training LLMs, we naturally wonder at which training stage introducing code data can really help LLMs reasoning. To this end, this paper systematically explores the impact of code data on LLMs at different stages. Concretely, we introduce the code data at the pre-training stage, instruction-tuning stage, and both of them, respectively. Then, the reasoning capability of LLMs is comprehensively and fairly evaluated via six reasoning tasks in five domains. We critically analyze the experimental results and provide conclusions with insights. First, pre-training LLMs with the mixture of code and text can significantly enhance LLMs' general reasoning capability almost without negative transfer on other tasks. Besides, at the instruction-tuning stage, code data endows LLMs the task-specific reasoning capability. Moreover, the dynamic mixing strategy of code and text data assists LLMs to learn reasoning capability step-by-step during training. These insights deepen the understanding of LLMs regarding reasoning ability for their application, such as scientific question answering, legal support, etc. The source code and model parameters are released at the link:~https://github.com/yingweima2022/CodeLLM.
2023-09-28
{'id': '2309.16298v2', 'title': 'At Which Training Stage Does Code Data Help LLMs Reasoning?', 'abstract': "Large Language Models (LLMs) have exhibited remarkable reasoning capabilities\nand become the foundation of language technologies. Inspired by the great\nsuccess of code data in training LLMs, we naturally wonder at which training\nstage introducing code data can really help LLMs reasoning. To this end, this\npaper systematically explores the impact of code data on LLMs at different\nstages. Concretely, we introduce the code data at the pre-training stage,\ninstruction-tuning stage, and both of them, respectively. Then, the reasoning\ncapability of LLMs is comprehensively and fairly evaluated via six reasoning\ntasks in five domains. We critically analyze the experimental results and\nprovide conclusions with insights. First, pre-training LLMs with the mixture of\ncode and text can significantly enhance LLMs' general reasoning capability\nalmost without negative transfer on other tasks. Besides, at the\ninstruction-tuning stage, code data endows LLMs the task-specific reasoning\ncapability. Moreover, the dynamic mixing strategy of code and text data assists\nLLMs to learn reasoning capability step-by-step during training. These insights\ndeepen the understanding of LLMs regarding reasoning ability for their\napplication, such as scientific question answering, legal support, etc. The\nsource code and model parameters are released at the\nlink:~\\url{https://github.com/yingweima2022/CodeLLM}.", 'authors': ['Yingwei Ma', 'Yue Liu', 'Yue Yu', 'Yuanliang Zhang', 'Yu Jiang', 'Changjian Wang', 'Shanshan Li'], 'published': '2023-09-28 09:50:27+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2309.16298v2'}
42,716
2310.00841v3
2,310.00841
3
Drug Discovery with Dynamic Goal-aware Fragments
Fragment-based drug discovery is an effective strategy for discovering drug candidates in the vast chemical space, and has been widely employed in molecular generative models. However, many existing fragment extraction methods in such models do not take the target chemical properties into account or rely on heuristic rules. Additionally, the existing fragment-based generative models cannot update the fragment vocabulary with goal-aware fragments newly discovered during the generation. To this end, we propose a molecular generative framework for drug discovery, named Goal-aware fragment Extraction, Assembly, and Modification (GEAM). GEAM consists of three modules, each responsible for goal-aware fragment extraction, fragment assembly, and fragment modification. The fragment extraction module identifies important fragments contributing to the desired target properties with the information bottleneck principle, thereby constructing an effective goal-aware fragment vocabulary. Moreover, GEAM can explore beyond the initial vocabulary with the fragment modification module, and the exploration is further enhanced through the dynamic goal-aware vocabulary update. We experimentally demonstrate that GEAM effectively discovers drug candidates through the generative cycle of the three modules in various drug discovery tasks. Our code is available at https://github.com/SeulLee05/GEAM.
2023-10-02
{'id': '2310.00841v3', 'title': 'Drug Discovery with Dynamic Goal-aware Fragments', 'abstract': 'Fragment-based drug discovery is an effective strategy for discovering drug\ncandidates in the vast chemical space, and has been widely employed in\nmolecular generative models. However, many existing fragment extraction methods\nin such models do not take the target chemical properties into account or rely\non heuristic rules. Additionally, the existing fragment-based generative models\ncannot update the fragment vocabulary with goal-aware fragments newly\ndiscovered during the generation. To this end, we propose a molecular\ngenerative framework for drug discovery, named Goal-aware fragment Extraction,\nAssembly, and Modification (GEAM). GEAM consists of three modules, each\nresponsible for goal-aware fragment extraction, fragment assembly, and fragment\nmodification. The fragment extraction module identifies important fragments\ncontributing to the desired target properties with the information bottleneck\nprinciple, thereby constructing an effective goal-aware fragment vocabulary.\nMoreover, GEAM can explore beyond the initial vocabulary with the fragment\nmodification module, and the exploration is further enhanced through the\ndynamic goal-aware vocabulary update. We experimentally demonstrate that GEAM\neffectively discovers drug candidates through the generative cycle of the three\nmodules in various drug discovery tasks. Our code is available at\nhttps://github.com/SeulLee05/GEAM.', 'authors': ['Seul Lee', 'Seanie Lee', 'Kenji Kawaguchi', 'Sung Ju Hwang'], 'published': '2023-10-02 01:30:42+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.00841v3'}
42,717
2310.01041v2
2,310.01041
2
Language Model Decoding as Direct Metrics Optimization
Despite the remarkable advances in language modeling, current mainstream decoding methods still struggle to generate texts that align with human texts across different aspects. In particular, sampling-based methods produce less-repetitive texts which are often disjunctive in discourse, while search-based methods maintain topic coherence at the cost of increased repetition. Overall, these methods fall short in achieving holistic alignment across a broad range of aspects. In this work, we frame decoding from a language model as an optimization problem with the goal of strictly matching the expected performance with human texts measured by multiple metrics of desired aspects simultaneously. The resulting decoding distribution enjoys an analytical solution that scales the input language model distribution via a sequence-level energy function defined by these metrics. And most importantly, we prove that this induced distribution is guaranteed to improve the perplexity on human texts, which suggests a better approximation to the underlying distribution of human texts. To facilitate tractable sampling from this globally normalized distribution, we adopt the Sampling-Importance-Resampling technique. Experiments on various domains and model scales demonstrate the superiority of our method in metrics alignment with human texts and human evaluation over strong baselines.
2023-10-02
{'id': '2310.01041v2', 'title': 'Language Model Decoding as Direct Metrics Optimization', 'abstract': 'Despite the remarkable advances in language modeling, current mainstream\ndecoding methods still struggle to generate texts that align with human texts\nacross different aspects. In particular, sampling-based methods produce\nless-repetitive texts which are often disjunctive in discourse, while\nsearch-based methods maintain topic coherence at the cost of increased\nrepetition. Overall, these methods fall short in achieving holistic alignment\nacross a broad range of aspects. In this work, we frame decoding from a\nlanguage model as an optimization problem with the goal of strictly matching\nthe expected performance with human texts measured by multiple metrics of\ndesired aspects simultaneously. The resulting decoding distribution enjoys an\nanalytical solution that scales the input language model distribution via a\nsequence-level energy function defined by these metrics. And most importantly,\nwe prove that this induced distribution is guaranteed to improve the perplexity\non human texts, which suggests a better approximation to the underlying\ndistribution of human texts. To facilitate tractable sampling from this\nglobally normalized distribution, we adopt the Sampling-Importance-Resampling\ntechnique. Experiments on various domains and model scales demonstrate the\nsuperiority of our method in metrics alignment with human texts and human\nevaluation over strong baselines.', 'authors': ['Haozhe Ji', 'Pei Ke', 'Hongning Wang', 'Minlie Huang'], 'published': '2023-10-02 09:35:27+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2310.01041v2'}
42,719
2308.07314v2
2,308.07314
2
Dual Associated Encoder for Face Restoration
Restoring facial details from low-quality (LQ) images has remained a challenging problem due to its ill-posedness induced by various degradations in the wild. The existing codebook prior mitigates the ill-posedness by leveraging an autoencoder and learned codebook of high-quality (HQ) features, achieving remarkable quality. However, existing approaches in this paradigm frequently depend on a single encoder pre-trained on HQ data for restoring HQ images, disregarding the domain gap between LQ and HQ images. As a result, the encoding of LQ inputs may be insufficient, resulting in suboptimal performance. To tackle this problem, we propose a novel dual-branch framework named DAEFR. Our method introduces an auxiliary LQ branch that extracts crucial information from the LQ inputs. Additionally, we incorporate association training to promote effective synergy between the two branches, enhancing code prediction and output quality. We evaluate the effectiveness of DAEFR on both synthetic and real-world datasets, demonstrating its superior performance in restoring facial details. Project page: https://liagm.github.io/DAEFR/
2023-08-14
{'id': '2308.07314v2', 'title': 'Dual Associated Encoder for Face Restoration', 'abstract': 'Restoring facial details from low-quality (LQ) images has remained a\nchallenging problem due to its ill-posedness induced by various degradations in\nthe wild. The existing codebook prior mitigates the ill-posedness by leveraging\nan autoencoder and learned codebook of high-quality (HQ) features, achieving\nremarkable quality. However, existing approaches in this paradigm frequently\ndepend on a single encoder pre-trained on HQ data for restoring HQ images,\ndisregarding the domain gap between LQ and HQ images. As a result, the encoding\nof LQ inputs may be insufficient, resulting in suboptimal performance. To\ntackle this problem, we propose a novel dual-branch framework named DAEFR. Our\nmethod introduces an auxiliary LQ branch that extracts crucial information from\nthe LQ inputs. Additionally, we incorporate association training to promote\neffective synergy between the two branches, enhancing code prediction and\noutput quality. We evaluate the effectiveness of DAEFR on both synthetic and\nreal-world datasets, demonstrating its superior performance in restoring facial\ndetails. Project page: https://liagm.github.io/DAEFR/', 'authors': ['Yu-Ju Tsai', 'Yu-Lun Liu', 'Lu Qi', 'Kelvin C. K. Chan', 'Ming-Hsuan Yang'], 'published': '2023-08-14 17:58:33+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2308.07314v2'}
42,733
2310.06824v3
2,310.06824
3
The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets
Large Language Models (LLMs) have impressive capabilities, but are prone to outputting falsehoods. Recent work has developed techniques for inferring whether a LLM is telling the truth by training probes on the LLM's internal activations. However, this line of work is controversial, with some authors pointing out failures of these probes to generalize in basic ways, among other conceptual issues. In this work, we use high-quality datasets of simple true/false statements to study in detail the structure of LLM representations of truth, drawing on three lines of evidence: 1. Visualizations of LLM true/false statement representations, which reveal clear linear structure. 2. Transfer experiments in which probes trained on one dataset generalize to different datasets. 3. Causal evidence obtained by surgically intervening in a LLM's forward pass, causing it to treat false statements as true and vice versa. Overall, we present evidence that at sufficient scale, LLMs linearly represent the truth or falsehood of factual statements. We also show that simple difference-in-mean probes generalize as well as other probing techniques while identifying directions which are more causally implicated in model outputs.
2023-10-10
{'id': '2310.06824v3', 'title': 'The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets', 'abstract': "Large Language Models (LLMs) have impressive capabilities, but are prone to\noutputting falsehoods. Recent work has developed techniques for inferring\nwhether a LLM is telling the truth by training probes on the LLM's internal\nactivations. However, this line of work is controversial, with some authors\npointing out failures of these probes to generalize in basic ways, among other\nconceptual issues. In this work, we use high-quality datasets of simple\ntrue/false statements to study in detail the structure of LLM representations\nof truth, drawing on three lines of evidence: 1. Visualizations of LLM\ntrue/false statement representations, which reveal clear linear structure. 2.\nTransfer experiments in which probes trained on one dataset generalize to\ndifferent datasets. 3. Causal evidence obtained by surgically intervening in a\nLLM's forward pass, causing it to treat false statements as true and vice\nversa. Overall, we present evidence that at sufficient scale, LLMs linearly\nrepresent the truth or falsehood of factual statements. We also show that\nsimple difference-in-mean probes generalize as well as other probing techniques\nwhile identifying directions which are more causally implicated in model\noutputs.", 'authors': ['Samuel Marks', 'Max Tegmark'], 'published': '2023-10-10 17:54:39+00:00', 'categories': ['cs.AI'], 'url': 'http://arxiv.org/abs/2310.06824v3'}
42,734
2305.11624v2
2,305.11624
2
Efficient ConvBN Blocks for Transfer Learning and Beyond
Convolution-BatchNorm (ConvBN) blocks are integral components in various computer vision tasks and other domains. A ConvBN block can operate in three modes: Train, Eval, and Deploy. While the Train mode is indispensable for training models from scratch, the Eval mode is suitable for transfer learning and beyond, and the Deploy mode is designed for the deployment of models. This paper focuses on the trade-off between stability and efficiency in ConvBN blocks: Deploy mode is efficient but suffers from training instability; Eval mode is widely used in transfer learning but lacks efficiency. To solve the dilemma, we theoretically reveal the reason behind the diminished training stability observed in the Deploy mode. Subsequently, we propose a novel Tune mode to bridge the gap between Eval mode and Deploy mode. The proposed Tune mode is as stable as Eval mode for transfer learning, and its computational efficiency closely matches that of the Deploy mode. Through extensive experiments in object detection, classification, and adversarial example generation across $5$ datasets and $12$ model architectures, we demonstrate that the proposed Tune mode retains the performance while significantly reducing GPU memory footprint and training time, thereby contributing efficient ConvBN blocks for transfer learning and beyond. Our method has been integrated into both PyTorch (general machine learning framework) and MMCV/MMEngine (computer vision framework). Practitioners just need one line of code to enjoy our efficient ConvBN blocks thanks to PyTorch's builtin machine learning compilers.
2023-05-19
{'id': '2305.11624v2', 'title': 'Efficient ConvBN Blocks for Transfer Learning and Beyond', 'abstract': "Convolution-BatchNorm (ConvBN) blocks are integral components in various\ncomputer vision tasks and other domains. A ConvBN block can operate in three\nmodes: Train, Eval, and Deploy. While the Train mode is indispensable for\ntraining models from scratch, the Eval mode is suitable for transfer learning\nand beyond, and the Deploy mode is designed for the deployment of models. This\npaper focuses on the trade-off between stability and efficiency in ConvBN\nblocks: Deploy mode is efficient but suffers from training instability; Eval\nmode is widely used in transfer learning but lacks efficiency. To solve the\ndilemma, we theoretically reveal the reason behind the diminished training\nstability observed in the Deploy mode. Subsequently, we propose a novel Tune\nmode to bridge the gap between Eval mode and Deploy mode. The proposed Tune\nmode is as stable as Eval mode for transfer learning, and its computational\nefficiency closely matches that of the Deploy mode. Through extensive\nexperiments in object detection, classification, and adversarial example\ngeneration across $5$ datasets and $12$ model architectures, we demonstrate\nthat the proposed Tune mode retains the performance while significantly\nreducing GPU memory footprint and training time, thereby contributing efficient\nConvBN blocks for transfer learning and beyond. Our method has been integrated\ninto both PyTorch (general machine learning framework) and MMCV/MMEngine\n(computer vision framework). Practitioners just need one line of code to enjoy\nour efficient ConvBN blocks thanks to PyTorch's builtin machine learning\ncompilers.", 'authors': ['Kaichao You', 'Guo Qin', 'Anchang Bao', 'Meng Cao', 'Ping Huang', 'Jiulong Shan', 'Mingsheng Long'], 'published': '2023-05-19 12:06:34+00:00', 'categories': ['cs.AI'], 'url': 'http://arxiv.org/abs/2305.11624v2'}
42,736
2405.02044v1
2,405.02044
1
Zero-Sum Positional Differential Games as a Framework for Robust Reinforcement Learning: Deep Q-Learning Approach
Robust Reinforcement Learning (RRL) is a promising Reinforcement Learning (RL) paradigm aimed at training robust to uncertainty or disturbances models, making them more efficient for real-world applications. Following this paradigm, uncertainty or disturbances are interpreted as actions of a second adversarial agent, and thus, the problem is reduced to seeking the agents' policies robust to any opponent's actions. This paper is the first to propose considering the RRL problems within the positional differential game theory, which helps us to obtain theoretically justified intuition to develop a centralized Q-learning approach. Namely, we prove that under Isaacs's condition (sufficiently general for real-world dynamical systems), the same Q-function can be utilized as an approximate solution of both minimax and maximin Bellman equations. Based on these results, we present the Isaacs Deep Q-Network algorithms and demonstrate their superiority compared to other baseline RRL and Multi-Agent RL algorithms in various environments.
2024-05-03
{'id': '2405.02044v1', 'title': 'Zero-Sum Positional Differential Games as a Framework for Robust Reinforcement Learning: Deep Q-Learning Approach', 'abstract': "Robust Reinforcement Learning (RRL) is a promising Reinforcement Learning\n(RL) paradigm aimed at training robust to uncertainty or disturbances models,\nmaking them more efficient for real-world applications. Following this\nparadigm, uncertainty or disturbances are interpreted as actions of a second\nadversarial agent, and thus, the problem is reduced to seeking the agents'\npolicies robust to any opponent's actions. This paper is the first to propose\nconsidering the RRL problems within the positional differential game theory,\nwhich helps us to obtain theoretically justified intuition to develop a\ncentralized Q-learning approach. Namely, we prove that under Isaacs's condition\n(sufficiently general for real-world dynamical systems), the same Q-function\ncan be utilized as an approximate solution of both minimax and maximin Bellman\nequations. Based on these results, we present the Isaacs Deep Q-Network\nalgorithms and demonstrate their superiority compared to other baseline RRL and\nMulti-Agent RL algorithms in various environments.", 'authors': ['Anton Plaksin', 'Vitaly Kalev'], 'published': '2024-05-03 12:21:43+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.GT', 'cs.SY', 'eess.SY', 'math.OC', '68T07, 49N70'], 'url': 'http://arxiv.org/abs/2405.02044v1'}
42,746
2501.03264v1
2,501.03264
1
Bridge the Inference Gaps of Neural Processes via Expectation Maximization
The neural process (NP) is a family of computationally efficient models for learning distributions over functions. However, it suffers from under-fitting and shows suboptimal performance in practice. Researchers have primarily focused on incorporating diverse structural inductive biases, e.g. attention or convolution, in modeling. The topic of inference suboptimality and an analysis of the NP from the optimization objective perspective has hardly been studied in earlier work. To fix this issue, we propose a surrogate objective of the target log-likelihood of the meta dataset within the expectation maximization framework. The resulting model, referred to as the Self-normalized Importance weighted Neural Process (SI-NP), can learn a more accurate functional prior and has an improvement guarantee concerning the target log-likelihood. Experimental results show the competitive performance of SI-NP over other NPs objectives and illustrate that structural inductive biases, such as attention modules, can also augment our method to achieve SOTA performance. Our code is available at https://github.com/hhq123gogogo/SI_NPs.
2025-01-04
{'id': '2501.03264v1', 'title': 'Bridge the Inference Gaps of Neural Processes via Expectation Maximization', 'abstract': 'The neural process (NP) is a family of computationally efficient models for\nlearning distributions over functions. However, it suffers from under-fitting\nand shows suboptimal performance in practice. Researchers have primarily\nfocused on incorporating diverse structural inductive biases, \\textit{e.g.}\nattention or convolution, in modeling. The topic of inference suboptimality and\nan analysis of the NP from the optimization objective perspective has hardly\nbeen studied in earlier work. To fix this issue, we propose a surrogate\nobjective of the target log-likelihood of the meta dataset within the\nexpectation maximization framework. The resulting model, referred to as the\nSelf-normalized Importance weighted Neural Process (SI-NP), can learn a more\naccurate functional prior and has an improvement guarantee concerning the\ntarget log-likelihood. Experimental results show the competitive performance of\nSI-NP over other NPs objectives and illustrate that structural inductive\nbiases, such as attention modules, can also augment our method to achieve SOTA\nperformance. Our code is available at\n\\url{https://github.com/hhq123gogogo/SI_NPs}.', 'authors': ['Qi Wang', 'Marco Federici', 'Herke van Hoof'], 'published': '2025-01-04 03:28:21+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.NE'], 'url': 'http://arxiv.org/abs/2501.03264v1'}
42,721
2303.12803v2
2,303.12803
2
Evolving Populations of Diverse RL Agents with MAP-Elites
Quality Diversity (QD) has emerged as a powerful alternative optimization paradigm that aims at generating large and diverse collections of solutions, notably with its flagship algorithm MAP-ELITES (ME) which evolves solutions through mutations and crossovers. While very effective for some unstructured problems, early ME implementations relied exclusively on random search to evolve the population of solutions, rendering them notoriously sample-inefficient for high-dimensional problems, such as when evolving neural networks. Follow-up works considered exploiting gradient information to guide the search in order to address these shortcomings through techniques borrowed from either Black-Box Optimization (BBO) or Reinforcement Learning (RL). While mixing RL techniques with ME unlocked state-of-the-art performance for robotics control problems that require a good amount of exploration, it also plagued these ME variants with limitations common among RL algorithms that ME was free of, such as hyperparameter sensitivity, high stochasticity as well as training instability, including when the population size increases as some components are shared across the population in recent approaches. Furthermore, existing approaches mixing ME with RL tend to be tied to a specific RL algorithm, which effectively prevents their use on problems where the corresponding RL algorithm fails. To address these shortcomings, we introduce a flexible framework that allows the use of any RL algorithm and alleviates the aforementioned limitations by evolving populations of agents (whose definition include hyperparameters and all learnable parameters) instead of just policies. We demonstrate the benefits brought about by our framework through extensive numerical experiments on a number of robotics control problems, some of which with deceptive rewards, taken from the QD-RL literature.
2023-03-09
{'id': '2303.12803v2', 'title': 'Evolving Populations of Diverse RL Agents with MAP-Elites', 'abstract': 'Quality Diversity (QD) has emerged as a powerful alternative optimization\nparadigm that aims at generating large and diverse collections of solutions,\nnotably with its flagship algorithm MAP-ELITES (ME) which evolves solutions\nthrough mutations and crossovers. While very effective for some unstructured\nproblems, early ME implementations relied exclusively on random search to\nevolve the population of solutions, rendering them notoriously\nsample-inefficient for high-dimensional problems, such as when evolving neural\nnetworks. Follow-up works considered exploiting gradient information to guide\nthe search in order to address these shortcomings through techniques borrowed\nfrom either Black-Box Optimization (BBO) or Reinforcement Learning (RL). While\nmixing RL techniques with ME unlocked state-of-the-art performance for robotics\ncontrol problems that require a good amount of exploration, it also plagued\nthese ME variants with limitations common among RL algorithms that ME was free\nof, such as hyperparameter sensitivity, high stochasticity as well as training\ninstability, including when the population size increases as some components\nare shared across the population in recent approaches. Furthermore, existing\napproaches mixing ME with RL tend to be tied to a specific RL algorithm, which\neffectively prevents their use on problems where the corresponding RL algorithm\nfails. To address these shortcomings, we introduce a flexible framework that\nallows the use of any RL algorithm and alleviates the aforementioned\nlimitations by evolving populations of agents (whose definition include\nhyperparameters and all learnable parameters) instead of just policies. We\ndemonstrate the benefits brought about by our framework through extensive\nnumerical experiments on a number of robotics control problems, some of which\nwith deceptive rewards, taken from the QD-RL literature.', 'authors': ['Thomas Pierrot', 'Arthur Flajolet'], 'published': '2023-03-09 19:05:45+00:00', 'categories': ['cs.NE', 'cs.AI'], 'url': 'http://arxiv.org/abs/2303.12803v2'}
42,722
2209.15097v2
2,209.15097
2
Likelihood Adjusted Semidefinite Programs for Clustering Heterogeneous Data
Clustering is a widely deployed unsupervised learning tool. Model-based clustering is a flexible framework to tackle data heterogeneity when the clusters have different shapes. Likelihood-based inference for mixture distributions often involves non-convex and high-dimensional objective functions, imposing difficult computational and statistical challenges. The classic expectation-maximization (EM) algorithm is a computationally thrifty iterative method that maximizes a surrogate function minorizing the log-likelihood of observed data in each iteration, which however suffers from bad local maxima even in the special case of the standard Gaussian mixture model with common isotropic covariance matrices. On the other hand, recent studies reveal that the unique global solution of a semidefinite programming (SDP) relaxed $K$-means achieves the information-theoretically sharp threshold for perfectly recovering the cluster labels under the standard Gaussian mixture model. In this paper, we extend the SDP approach to a general setting by integrating cluster labels as model parameters and propose an iterative likelihood adjusted SDP (iLA-SDP) method that directly maximizes the exact observed likelihood in the presence of data heterogeneity. By lifting the cluster assignment to group-specific membership matrices, iLA-SDP avoids centroids estimation -- a key feature that allows exact recovery under well-separateness of centroids without being trapped by their adversarial configurations. Thus iLA-SDP is less sensitive than EM to initialization and more stable on high-dimensional data. Our numeric experiments demonstrate that iLA-SDP can achieve lower mis-clustering errors over several widely used clustering methods including $K$-means, SDP and EM algorithms.
2022-09-29
{'id': '2209.15097v2', 'title': 'Likelihood Adjusted Semidefinite Programs for Clustering Heterogeneous Data', 'abstract': 'Clustering is a widely deployed unsupervised learning tool. Model-based\nclustering is a flexible framework to tackle data heterogeneity when the\nclusters have different shapes. Likelihood-based inference for mixture\ndistributions often involves non-convex and high-dimensional objective\nfunctions, imposing difficult computational and statistical challenges. The\nclassic expectation-maximization (EM) algorithm is a computationally thrifty\niterative method that maximizes a surrogate function minorizing the\nlog-likelihood of observed data in each iteration, which however suffers from\nbad local maxima even in the special case of the standard Gaussian mixture\nmodel with common isotropic covariance matrices. On the other hand, recent\nstudies reveal that the unique global solution of a semidefinite programming\n(SDP) relaxed $K$-means achieves the information-theoretically sharp threshold\nfor perfectly recovering the cluster labels under the standard Gaussian mixture\nmodel. In this paper, we extend the SDP approach to a general setting by\nintegrating cluster labels as model parameters and propose an iterative\nlikelihood adjusted SDP (iLA-SDP) method that directly maximizes the exact\nobserved likelihood in the presence of data heterogeneity. By lifting the\ncluster assignment to group-specific membership matrices, iLA-SDP avoids\ncentroids estimation -- a key feature that allows exact recovery under\nwell-separateness of centroids without being trapped by their adversarial\nconfigurations. Thus iLA-SDP is less sensitive than EM to initialization and\nmore stable on high-dimensional data. Our numeric experiments demonstrate that\niLA-SDP can achieve lower mis-clustering errors over several widely used\nclustering methods including $K$-means, SDP and EM algorithms.', 'authors': ['Yubo Zhuang', 'Xiaohui Chen', 'Yun Yang'], 'published': '2022-09-29 21:03:13+00:00', 'categories': ['stat.ML', 'cs.LG', 'math.OC'], 'url': 'http://arxiv.org/abs/2209.15097v2'}
42,723
2204.02892v4
2,204.02892
4
Sub-Task Decomposition Enables Learning in Sequence to Sequence Tasks
The field of Natural Language Processing has experienced a dramatic leap in capabilities with the recent introduction of huge Language Models. Despite this success, natural language problems that involve several compounded steps are still practically unlearnable, even by the largest LMs. This complies with experimental failures for end-to-end learning of composite problems that were demonstrated in a variety of domains. An effective mitigation is to introduce intermediate supervision for solving sub-tasks of the compounded problem. Recently, several works have demonstrated high gains by taking a straightforward approach for incorporating intermediate supervision in compounded natural language problems: the sequence-to-sequence LM is fed with an augmented input, in which the decomposed tasks' labels are simply concatenated to the original input. In this paper, we prove a positive learning result that motivates these recent efforts. We show that when concatenating intermediate supervision to the input and training a sequence-to-sequence model on this modified input, unlearnable composite problems can become learnable. We show that this is true for any family of tasks which on the one hand, are unlearnable, and on the other hand, can be decomposed into a polynomial number of simple sub-tasks, each of which depends only on O(1) previous sub-task results. Beyond motivating contemporary empirical efforts for incorporating intermediate supervision in sequence-to-sequence language models, our positive theoretical result is the first of its kind in the landscape of results on the benefits of intermediate supervision for neural-network learning: Until now, all theoretical results on the subject are negative, i.e., show cases where learning is impossible without intermediate supervision, while our result is positive, showing that learning is facilitated in the presence of intermediate supervision.
2022-04-06
{'id': '2204.02892v4', 'title': 'Sub-Task Decomposition Enables Learning in Sequence to Sequence Tasks', 'abstract': "The field of Natural Language Processing has experienced a dramatic leap in\ncapabilities with the recent introduction of huge Language Models. Despite this\nsuccess, natural language problems that involve several compounded steps are\nstill practically unlearnable, even by the largest LMs. This complies with\nexperimental failures for end-to-end learning of composite problems that were\ndemonstrated in a variety of domains. An effective mitigation is to introduce\nintermediate supervision for solving sub-tasks of the compounded problem.\nRecently, several works have demonstrated high gains by taking a\nstraightforward approach for incorporating intermediate supervision in\ncompounded natural language problems: the sequence-to-sequence LM is fed with\nan augmented input, in which the decomposed tasks' labels are simply\nconcatenated to the original input. In this paper, we prove a positive learning\nresult that motivates these recent efforts. We show that when concatenating\nintermediate supervision to the input and training a sequence-to-sequence model\non this modified input, unlearnable composite problems can become learnable. We\nshow that this is true for any family of tasks which on the one hand, are\nunlearnable, and on the other hand, can be decomposed into a polynomial number\nof simple sub-tasks, each of which depends only on O(1) previous sub-task\nresults. Beyond motivating contemporary empirical efforts for incorporating\nintermediate supervision in sequence-to-sequence language models, our positive\ntheoretical result is the first of its kind in the landscape of results on the\nbenefits of intermediate supervision for neural-network learning: Until now,\nall theoretical results on the subject are negative, i.e., show cases where\nlearning is impossible without intermediate supervision, while our result is\npositive, showing that learning is facilitated in the presence of intermediate\nsupervision.", 'authors': ['Noam Wies', 'Yoav Levine', 'Amnon Shashua'], 'published': '2022-04-06 15:16:27+00:00', 'categories': ['cs.CL', 'cs.LG'], 'url': 'http://arxiv.org/abs/2204.02892v4'}
42,724
2302.01497v3
2,302.01497
3
Gradient Estimation for Unseen Domain Risk Minimization with Pre-Trained Models
Domain generalization aims to build generalized models that perform well on unseen domains when only source domains are available for model optimization. Recent studies have shown that large-scale pre-trained models can enhance domain generalization by leveraging their generalization power. However, these pre-trained models lack target task-specific knowledge yet due to discrepancies between the pre-training objectives and the target task. Although the task-specific knowledge could be learned from source domains by fine-tuning, this hurts the generalization power of pre-trained models due to gradient bias toward the source domains. To alleviate this problem, we propose a new domain generalization method that estimates unobservable gradients that reduce potential risks in unseen domains using a large-scale pre-trained model. These estimated unobservable gradients allow the pre-trained model to learn task-specific knowledge further while preserving its generalization ability by relieving the gradient bias. Our experimental results show that our method outperforms baseline methods on DomainBed, a standard benchmark in domain generalization. We also provide extensive analyses to demonstrate that the pre-trained model can learn task-specific knowledge without sacrificing its generalization power.
2023-02-03
{'id': '2302.01497v3', 'title': 'Gradient Estimation for Unseen Domain Risk Minimization with Pre-Trained Models', 'abstract': 'Domain generalization aims to build generalized models that perform well on\nunseen domains when only source domains are available for model optimization.\nRecent studies have shown that large-scale pre-trained models can enhance\ndomain generalization by leveraging their generalization power. However, these\npre-trained models lack target task-specific knowledge yet due to discrepancies\nbetween the pre-training objectives and the target task. Although the\ntask-specific knowledge could be learned from source domains by fine-tuning,\nthis hurts the generalization power of pre-trained models due to gradient bias\ntoward the source domains. To alleviate this problem, we propose a new domain\ngeneralization method that estimates unobservable gradients that reduce\npotential risks in unseen domains using a large-scale pre-trained model. These\nestimated unobservable gradients allow the pre-trained model to learn\ntask-specific knowledge further while preserving its generalization ability by\nrelieving the gradient bias. Our experimental results show that our method\noutperforms baseline methods on DomainBed, a standard benchmark in domain\ngeneralization. We also provide extensive analyses to demonstrate that the\npre-trained model can learn task-specific knowledge without sacrificing its\ngeneralization power.', 'authors': ['Byounggyu Lew', 'Donghyun Son', 'Buru Chang'], 'published': '2023-02-03 02:12:09+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2302.01497v3'}
42,725
2210.07499v2
2,210.07499
2
Bayes risk CTC: Controllable CTC alignment in Sequence-to-Sequence tasks
Sequence-to-Sequence (seq2seq) tasks transcribe the input sequence to a target sequence. The Connectionist Temporal Classification (CTC) criterion is widely used in multiple seq2seq tasks. Besides predicting the target sequence, a side product of CTC is to predict the alignment, which is the most probable input-long sequence that specifies a hard aligning relationship between the input and target units. As there are multiple potential aligning sequences (called paths) that are equally considered in CTC formulation, the choice of which path will be most probable and become the predicted alignment is always uncertain. In addition, it is usually observed that the alignment predicted by vanilla CTC will drift compared with its reference and rarely provides practical functionalities. Thus, the motivation of this work is to make the CTC alignment prediction controllable and thus equip CTC with extra functionalities. The Bayes risk CTC (BRCTC) criterion is then proposed in this work, in which a customizable Bayes risk function is adopted to enforce the desired characteristics of the predicted alignment. With the risk function, the BRCTC is a general framework to adopt some customizable preference over the paths in order to concentrate the posterior into a particular subset of the paths. In applications, we explore one particular preference which yields models with the down-sampling ability and reduced inference costs. By using BRCTC with another preference for early emissions, we obtain an improved performance-latency trade-off for online models. Experimentally, the proposed BRCTC reduces the inference cost of offline models by up to 47% without performance degradation and cuts down the overall latency of online systems to an unseen level.
2022-10-14
{'id': '2210.07499v2', 'title': 'Bayes risk CTC: Controllable CTC alignment in Sequence-to-Sequence tasks', 'abstract': 'Sequence-to-Sequence (seq2seq) tasks transcribe the input sequence to a\ntarget sequence. The Connectionist Temporal Classification (CTC) criterion is\nwidely used in multiple seq2seq tasks. Besides predicting the target sequence,\na side product of CTC is to predict the alignment, which is the most probable\ninput-long sequence that specifies a hard aligning relationship between the\ninput and target units. As there are multiple potential aligning sequences\n(called paths) that are equally considered in CTC formulation, the choice of\nwhich path will be most probable and become the predicted alignment is always\nuncertain. In addition, it is usually observed that the alignment predicted by\nvanilla CTC will drift compared with its reference and rarely provides\npractical functionalities. Thus, the motivation of this work is to make the CTC\nalignment prediction controllable and thus equip CTC with extra\nfunctionalities. The Bayes risk CTC (BRCTC) criterion is then proposed in this\nwork, in which a customizable Bayes risk function is adopted to enforce the\ndesired characteristics of the predicted alignment. With the risk function, the\nBRCTC is a general framework to adopt some customizable preference over the\npaths in order to concentrate the posterior into a particular subset of the\npaths. In applications, we explore one particular preference which yields\nmodels with the down-sampling ability and reduced inference costs. By using\nBRCTC with another preference for early emissions, we obtain an improved\nperformance-latency trade-off for online models. Experimentally, the proposed\nBRCTC reduces the inference cost of offline models by up to 47% without\nperformance degradation and cuts down the overall latency of online systems to\nan unseen level.', 'authors': ['Jinchuan Tian', 'Brian Yan', 'Jianwei Yu', 'Chao Weng', 'Dong Yu', 'Shinji Watanabe'], 'published': '2022-10-14 03:55:36+00:00', 'categories': ['cs.CL', 'cs.SD', 'eess.AS'], 'url': 'http://arxiv.org/abs/2210.07499v2'}
42,726
2301.08842v1
2,301.08842
1
Limitations of Piecewise Linearity for Efficient Robustness Certification
Certified defenses against small-norm adversarial examples have received growing attention in recent years; though certified accuracies of state-of-the-art methods remain far below their non-robust counterparts, despite the fact that benchmark datasets have been shown to be well-separated at far larger radii than the literature generally attempts to certify. In this work, we offer insights that identify potential factors in this performance gap. Specifically, our analysis reveals that piecewise linearity imposes fundamental limitations on the tightness of leading certification techniques. These limitations are felt in practical terms as a greater need for capacity in models hoped to be certified efficiently. Moreover, this is in addition to the capacity necessary to learn a robust boundary, studied in prior work. However, we argue that addressing the limitations of piecewise linearity through scaling up model capacity may give rise to potential difficulties -- particularly regarding robust generalization -- therefore, we conclude by suggesting that developing smooth activation functions may be the way forward for advancing the performance of certified neural networks.
2023-01-21
{'id': '2301.08842v1', 'title': 'Limitations of Piecewise Linearity for Efficient Robustness Certification', 'abstract': 'Certified defenses against small-norm adversarial examples have received\ngrowing attention in recent years; though certified accuracies of\nstate-of-the-art methods remain far below their non-robust counterparts,\ndespite the fact that benchmark datasets have been shown to be well-separated\nat far larger radii than the literature generally attempts to certify. In this\nwork, we offer insights that identify potential factors in this performance\ngap. Specifically, our analysis reveals that piecewise linearity imposes\nfundamental limitations on the tightness of leading certification techniques.\nThese limitations are felt in practical terms as a greater need for capacity in\nmodels hoped to be certified efficiently. Moreover, this is in addition to the\ncapacity necessary to learn a robust boundary, studied in prior work. However,\nwe argue that addressing the limitations of piecewise linearity through scaling\nup model capacity may give rise to potential difficulties -- particularly\nregarding robust generalization -- therefore, we conclude by suggesting that\ndeveloping smooth activation functions may be the way forward for advancing the\nperformance of certified neural networks.', 'authors': ['Klas Leino'], 'published': '2023-01-21 01:20:43+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2301.08842v1'}
42,727
2207.01190v1
2,207.0119
1
Pareto Optimization for Active Learning under Out-of-Distribution Data Scenarios
Pool-based Active Learning (AL) has achieved great success in minimizing labeling cost by sequentially selecting informative unlabeled samples from a large unlabeled data pool and querying their labels from oracle/annotators. However, existing AL sampling strategies might not work well in out-of-distribution (OOD) data scenarios, where the unlabeled data pool contains some data samples that do not belong to the classes of the target task. Achieving good AL performance under OOD data scenarios is a challenging task due to the natural conflict between AL sampling strategies and OOD sample detection. AL selects data that are hard to be classified by the current basic classifier (e.g., samples whose predicted class probabilities have high entropy), while OOD samples tend to have more uniform predicted class probabilities (i.e., high entropy) than in-distribution (ID) data. In this paper, we propose a sampling scheme, Monte-Carlo Pareto Optimization for Active Learning (POAL), which selects optimal subsets of unlabeled samples with fixed batch size from the unlabeled data pool. We cast the AL sampling task as a multi-objective optimization problem, and thus we utilize Pareto optimization based on two conflicting objectives: (1) the normal AL data sampling scheme (e.g., maximum entropy), and (2) the confidence of not being an OOD sample. Experimental results show its effectiveness on both classical Machine Learning (ML) and Deep Learning (DL) tasks.
2022-07-04
{'id': '2207.01190v1', 'title': 'Pareto Optimization for Active Learning under Out-of-Distribution Data Scenarios', 'abstract': 'Pool-based Active Learning (AL) has achieved great success in minimizing\nlabeling cost by sequentially selecting informative unlabeled samples from a\nlarge unlabeled data pool and querying their labels from oracle/annotators.\nHowever, existing AL sampling strategies might not work well in\nout-of-distribution (OOD) data scenarios, where the unlabeled data pool\ncontains some data samples that do not belong to the classes of the target\ntask. Achieving good AL performance under OOD data scenarios is a challenging\ntask due to the natural conflict between AL sampling strategies and OOD sample\ndetection. AL selects data that are hard to be classified by the current basic\nclassifier (e.g., samples whose predicted class probabilities have high\nentropy), while OOD samples tend to have more uniform predicted class\nprobabilities (i.e., high entropy) than in-distribution (ID) data. In this\npaper, we propose a sampling scheme, Monte-Carlo Pareto Optimization for Active\nLearning (POAL), which selects optimal subsets of unlabeled samples with fixed\nbatch size from the unlabeled data pool. We cast the AL sampling task as a\nmulti-objective optimization problem, and thus we utilize Pareto optimization\nbased on two conflicting objectives: (1) the normal AL data sampling scheme\n(e.g., maximum entropy), and (2) the confidence of not being an OOD sample.\nExperimental results show its effectiveness on both classical Machine Learning\n(ML) and Deep Learning (DL) tasks.', 'authors': ['Xueying Zhan', 'Zeyu Dai', 'Qingzhong Wang', 'Qing Li', 'Haoyi Xiong', 'Dejing Dou', 'Antoni B. Chan'], 'published': '2022-07-04 04:11:44+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2207.01190v1'}
42,728
2301.09254v1
2,301.09254
1
Learning to Linearize Deep Neural Networks for Secure and Efficient Private Inference
The large number of ReLU non-linearity operations in existing deep neural networks makes them ill-suited for latency-efficient private inference (PI). Existing techniques to reduce ReLU operations often involve manual effort and sacrifice significant accuracy. In this paper, we first present a novel measure of non-linearity layers' ReLU sensitivity, enabling mitigation of the time-consuming manual efforts in identifying the same. Based on this sensitivity, we then present SENet, a three-stage training method that for a given ReLU budget, automatically assigns per-layer ReLU counts, decides the ReLU locations for each layer's activation map, and trains a model with significantly fewer ReLUs to potentially yield latency and communication efficient PI. Experimental evaluations with multiple models on various datasets show SENet's superior performance both in terms of reduced ReLUs and improved classification accuracy compared to existing alternatives. In particular, SENet can yield models that require up to ~2x fewer ReLUs while yielding similar accuracy. For a similar ReLU budget SENet can yield models with ~2.32% improved classification accuracy, evaluated on CIFAR-100.
2023-01-23
{'id': '2301.09254v1', 'title': 'Learning to Linearize Deep Neural Networks for Secure and Efficient Private Inference', 'abstract': "The large number of ReLU non-linearity operations in existing deep neural\nnetworks makes them ill-suited for latency-efficient private inference (PI).\nExisting techniques to reduce ReLU operations often involve manual effort and\nsacrifice significant accuracy. In this paper, we first present a novel measure\nof non-linearity layers' ReLU sensitivity, enabling mitigation of the\ntime-consuming manual efforts in identifying the same. Based on this\nsensitivity, we then present SENet, a three-stage training method that for a\ngiven ReLU budget, automatically assigns per-layer ReLU counts, decides the\nReLU locations for each layer's activation map, and trains a model with\nsignificantly fewer ReLUs to potentially yield latency and communication\nefficient PI. Experimental evaluations with multiple models on various datasets\nshow SENet's superior performance both in terms of reduced ReLUs and improved\nclassification accuracy compared to existing alternatives. In particular, SENet\ncan yield models that require up to ~2x fewer ReLUs while yielding similar\naccuracy. For a similar ReLU budget SENet can yield models with ~2.32% improved\nclassification accuracy, evaluated on CIFAR-100.", 'authors': ['Souvik Kundu', 'Shunlin Lu', 'Yuke Zhang', 'Jacqueline Liu', 'Peter A. Beerel'], 'published': '2023-01-23 03:33:38+00:00', 'categories': ['cs.CV', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2301.09254v1'}
42,729
2206.01816v2
2,206.01816
2
From $t$-SNE to UMAP with contrastive learning
Neighbor embedding methods $t$-SNE and UMAP are the de facto standard for visualizing high-dimensional datasets. Motivated from entirely different viewpoints, their loss functions appear to be unrelated. In practice, they yield strongly differing embeddings and can suggest conflicting interpretations of the same data. The fundamental reasons for this and, more generally, the exact relationship between $t$-SNE and UMAP have remained unclear. In this work, we uncover their conceptual connection via a new insight into contrastive learning methods. Noise-contrastive estimation can be used to optimize $t$-SNE, while UMAP relies on negative sampling, another contrastive method. We find the precise relationship between these two contrastive methods and provide a mathematical characterization of the distortion introduced by negative sampling. Visually, this distortion results in UMAP generating more compact embeddings with tighter clusters compared to $t$-SNE. We exploit this new conceptual connection to propose and implement a generalization of negative sampling, allowing us to interpolate between (and even extrapolate beyond) $t$-SNE and UMAP and their respective embeddings. Moving along this spectrum of embeddings leads to a trade-off between discrete / local and continuous / global structures, mitigating the risk of over-interpreting ostensible features of any single embedding. We provide a PyTorch implementation.
2022-06-03
{'id': '2206.01816v2', 'title': 'From $t$-SNE to UMAP with contrastive learning', 'abstract': 'Neighbor embedding methods $t$-SNE and UMAP are the de facto standard for\nvisualizing high-dimensional datasets. Motivated from entirely different\nviewpoints, their loss functions appear to be unrelated. In practice, they\nyield strongly differing embeddings and can suggest conflicting interpretations\nof the same data. The fundamental reasons for this and, more generally, the\nexact relationship between $t$-SNE and UMAP have remained unclear. In this\nwork, we uncover their conceptual connection via a new insight into contrastive\nlearning methods. Noise-contrastive estimation can be used to optimize $t$-SNE,\nwhile UMAP relies on negative sampling, another contrastive method. We find the\nprecise relationship between these two contrastive methods and provide a\nmathematical characterization of the distortion introduced by negative\nsampling. Visually, this distortion results in UMAP generating more compact\nembeddings with tighter clusters compared to $t$-SNE. We exploit this new\nconceptual connection to propose and implement a generalization of negative\nsampling, allowing us to interpolate between (and even extrapolate beyond)\n$t$-SNE and UMAP and their respective embeddings. Moving along this spectrum of\nembeddings leads to a trade-off between discrete / local and continuous /\nglobal structures, mitigating the risk of over-interpreting ostensible features\nof any single embedding. We provide a PyTorch implementation.', 'authors': ['Sebastian Damrich', 'Jan Niklas Böhm', 'Fred A. Hamprecht', 'Dmitry Kobak'], 'published': '2022-06-03 20:50:54+00:00', 'categories': ['cs.LG', 'cs.HC'], 'url': 'http://arxiv.org/abs/2206.01816v2'}
42,730
2209.14855v2
2,209.14855
2
Continuous PDE Dynamics Forecasting with Implicit Neural Representations
Effective data-driven PDE forecasting methods often rely on fixed spatial and / or temporal discretizations. This raises limitations in real-world applications like weather prediction where flexible extrapolation at arbitrary spatiotemporal locations is required. We address this problem by introducing a new data-driven approach, DINo, that models a PDE's flow with continuous-time dynamics of spatially continuous functions. This is achieved by embedding spatial observations independently of their discretization via Implicit Neural Representations in a small latent space temporally driven by a learned ODE. This separate and flexible treatment of time and space makes DINo the first data-driven model to combine the following advantages. It extrapolates at arbitrary spatial and temporal locations; it can learn from sparse irregular grids or manifolds; at test time, it generalizes to new grids or resolutions. DINo outperforms alternative neural PDE forecasters in a variety of challenging generalization scenarios on representative PDE systems.
2022-09-29
{'id': '2209.14855v2', 'title': 'Continuous PDE Dynamics Forecasting with Implicit Neural Representations', 'abstract': "Effective data-driven PDE forecasting methods often rely on fixed spatial and\n/ or temporal discretizations. This raises limitations in real-world\napplications like weather prediction where flexible extrapolation at arbitrary\nspatiotemporal locations is required. We address this problem by introducing a\nnew data-driven approach, DINo, that models a PDE's flow with continuous-time\ndynamics of spatially continuous functions. This is achieved by embedding\nspatial observations independently of their discretization via Implicit Neural\nRepresentations in a small latent space temporally driven by a learned ODE.\nThis separate and flexible treatment of time and space makes DINo the first\ndata-driven model to combine the following advantages. It extrapolates at\narbitrary spatial and temporal locations; it can learn from sparse irregular\ngrids or manifolds; at test time, it generalizes to new grids or resolutions.\nDINo outperforms alternative neural PDE forecasters in a variety of challenging\ngeneralization scenarios on representative PDE systems.", 'authors': ['Yuan Yin', 'Matthieu Kirchmeyer', 'Jean-Yves Franceschi', 'Alain Rakotomamonjy', 'Patrick Gallinari'], 'published': '2022-09-29 15:17:50+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.NE', 'stat.ML'], 'url': 'http://arxiv.org/abs/2209.14855v2'}
42,731
2210.03022v3
2,210.03022
3
Stateful active facilitator: Coordination and Environmental Heterogeneity in Cooperative Multi-Agent Reinforcement Learning
In cooperative multi-agent reinforcement learning, a team of agents works together to achieve a common goal. Different environments or tasks may require varying degrees of coordination among agents in order to achieve the goal in an optimal way. The nature of coordination will depend on the properties of the environment -- its spatial layout, distribution of obstacles, dynamics, etc. We term this variation of properties within an environment as heterogeneity. Existing literature has not sufficiently addressed the fact that different environments may have different levels of heterogeneity. We formalize the notions of coordination level and heterogeneity level of an environment and present HECOGrid, a suite of multi-agent RL environments that facilitates empirical evaluation of different MARL approaches across different levels of coordination and environmental heterogeneity by providing a quantitative control over coordination and heterogeneity levels of the environment. Further, we propose a Centralized Training Decentralized Execution learning approach called Stateful Active Facilitator (SAF) that enables agents to work efficiently in high-coordination and high-heterogeneity environments through a differentiable and shared knowledge source used during training and dynamic selection from a shared pool of policies. We evaluate SAF and compare its performance against baselines IPPO and MAPPO on HECOGrid. Our results show that SAF consistently outperforms the baselines across different tasks and different heterogeneity and coordination levels. We release the code for HECOGrid as well as all our experiments.
2022-10-04
{'id': '2210.03022v3', 'title': 'Stateful active facilitator: Coordination and Environmental Heterogeneity in Cooperative Multi-Agent Reinforcement Learning', 'abstract': 'In cooperative multi-agent reinforcement learning, a team of agents works\ntogether to achieve a common goal. Different environments or tasks may require\nvarying degrees of coordination among agents in order to achieve the goal in an\noptimal way. The nature of coordination will depend on the properties of the\nenvironment -- its spatial layout, distribution of obstacles, dynamics, etc. We\nterm this variation of properties within an environment as heterogeneity.\nExisting literature has not sufficiently addressed the fact that different\nenvironments may have different levels of heterogeneity. We formalize the\nnotions of coordination level and heterogeneity level of an environment and\npresent HECOGrid, a suite of multi-agent RL environments that facilitates\nempirical evaluation of different MARL approaches across different levels of\ncoordination and environmental heterogeneity by providing a quantitative\ncontrol over coordination and heterogeneity levels of the environment. Further,\nwe propose a Centralized Training Decentralized Execution learning approach\ncalled Stateful Active Facilitator (SAF) that enables agents to work\nefficiently in high-coordination and high-heterogeneity environments through a\ndifferentiable and shared knowledge source used during training and dynamic\nselection from a shared pool of policies. We evaluate SAF and compare its\nperformance against baselines IPPO and MAPPO on HECOGrid. Our results show that\nSAF consistently outperforms the baselines across different tasks and different\nheterogeneity and coordination levels. We release the code for HECOGrid as well\nas all our experiments.', 'authors': ['Dianbo Liu', 'Vedant Shah', 'Oussama Boussif', 'Cristian Meo', 'Anirudh Goyal', 'Tianmin Shu', 'Michael Mozer', 'Nicolas Heess', 'Yoshua Bengio'], 'published': '2022-10-04 18:17:01+00:00', 'categories': ['cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2210.03022v3'}
42,732
2302.00854v1
2,302.00854
1
Learning PDE Solution Operator for Continuous Modeling of Time-Series
Learning underlying dynamics from data is important and challenging in many real-world scenarios. Incorporating differential equations (DEs) to design continuous networks has drawn much attention recently, however, most prior works make specific assumptions on the type of DEs, making the model specialized for particular problems. This work presents a partial differential equation (PDE) based framework which improves the dynamics modeling capability. Building upon the recent Fourier neural operator, we propose a neural operator that can handle time continuously without requiring iterative operations or specific grids of temporal discretization. A theoretical result demonstrating its universality is provided. We also uncover an intrinsic property of neural operators that improves data efficiency and model generalization by ensuring stability. Our model achieves superior accuracy in dealing with time-dependent PDEs compared to existing models. Furthermore, several numerical pieces of evidence validate that our method better represents a wide range of dynamics and outperforms state-of-the-art DE-based models in real-time-series applications. Our framework opens up a new way for a continuous representation of neural networks that can be readily adopted for real-world applications.
2023-02-02
{'id': '2302.00854v1', 'title': 'Learning PDE Solution Operator for Continuous Modeling of Time-Series', 'abstract': 'Learning underlying dynamics from data is important and challenging in many\nreal-world scenarios. Incorporating differential equations (DEs) to design\ncontinuous networks has drawn much attention recently, however, most prior\nworks make specific assumptions on the type of DEs, making the model\nspecialized for particular problems. This work presents a partial differential\nequation (PDE) based framework which improves the dynamics modeling capability.\nBuilding upon the recent Fourier neural operator, we propose a neural operator\nthat can handle time continuously without requiring iterative operations or\nspecific grids of temporal discretization. A theoretical result demonstrating\nits universality is provided. We also uncover an intrinsic property of neural\noperators that improves data efficiency and model generalization by ensuring\nstability. Our model achieves superior accuracy in dealing with time-dependent\nPDEs compared to existing models. Furthermore, several numerical pieces of\nevidence validate that our method better represents a wide range of dynamics\nand outperforms state-of-the-art DE-based models in real-time-series\napplications. Our framework opens up a new way for a continuous representation\nof neural networks that can be readily adopted for real-world applications.', 'authors': ['Yesom Park', 'Jaemoo Choi', 'Changyeon Yoon', 'Chang hoon Song', 'Myungjoo Kang'], 'published': '2023-02-02 03:47:52+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2302.00854v1'}
42,735
2211.06530v2
2,211.0653
2
Multi-Epoch Matrix Factorization Mechanisms for Private Machine Learning
We introduce new differentially private (DP) mechanisms for gradient-based machine learning (ML) with multiple passes (epochs) over a dataset, substantially improving the achievable privacy-utility-computation tradeoffs. We formalize the problem of DP mechanisms for adaptive streams with multiple participations and introduce a non-trivial extension of online matrix factorization DP mechanisms to our setting. This includes establishing the necessary theory for sensitivity calculations and efficient computation of optimal matrices. For some applications like $>\!\! 10,000$ SGD steps, applying these optimal techniques becomes computationally expensive. We thus design an efficient Fourier-transform-based mechanism with only a minor utility loss. Extensive empirical evaluation on both example-level DP for image classification and user-level DP for language modeling demonstrate substantial improvements over all previous methods, including the widely-used DP-SGD . Though our primary application is to ML, our main DP results are applicable to arbitrary linear queries and hence may have much broader applicability.
2022-11-12
{'id': '2211.06530v2', 'title': 'Multi-Epoch Matrix Factorization Mechanisms for Private Machine Learning', 'abstract': 'We introduce new differentially private (DP) mechanisms for gradient-based\nmachine learning (ML) with multiple passes (epochs) over a dataset,\nsubstantially improving the achievable privacy-utility-computation tradeoffs.\nWe formalize the problem of DP mechanisms for adaptive streams with multiple\nparticipations and introduce a non-trivial extension of online matrix\nfactorization DP mechanisms to our setting. This includes establishing the\nnecessary theory for sensitivity calculations and efficient computation of\noptimal matrices. For some applications like $>\\!\\! 10,000$ SGD steps, applying\nthese optimal techniques becomes computationally expensive. We thus design an\nefficient Fourier-transform-based mechanism with only a minor utility loss.\nExtensive empirical evaluation on both example-level DP for image\nclassification and user-level DP for language modeling demonstrate substantial\nimprovements over all previous methods, including the widely-used DP-SGD .\nThough our primary application is to ML, our main DP results are applicable to\narbitrary linear queries and hence may have much broader applicability.', 'authors': ['Christopher A. Choquette-Choo', 'H. Brendan McMahan', 'Keith Rush', 'Abhradeep Thakurta'], 'published': '2022-11-12 00:41:11+00:00', 'categories': ['cs.LG', 'cs.CR', 'cs.DS', 'stat.ML'], 'url': 'http://arxiv.org/abs/2211.06530v2'}
42,737
2304.08487v1
2,304.08487
1
Hyper-Decision Transformer for Efficient Online Policy Adaptation
Decision Transformers (DT) have demonstrated strong performances in offline reinforcement learning settings, but quickly adapting to unseen novel tasks remains challenging. To address this challenge, we propose a new framework, called Hyper-Decision Transformer (HDT), that can generalize to novel tasks from a handful of demonstrations in a data- and parameter-efficient manner. To achieve such a goal, we propose to augment the base DT with an adaptation module, whose parameters are initialized by a hyper-network. When encountering unseen tasks, the hyper-network takes a handful of demonstrations as inputs and initializes the adaptation module accordingly. This initialization enables HDT to efficiently adapt to novel tasks by only fine-tuning the adaptation module. We validate HDT's generalization capability on object manipulation tasks. We find that with a single expert demonstration and fine-tuning only 0.5% of DT parameters, HDT adapts faster to unseen tasks than fine-tuning the whole DT model. Finally, we explore a more challenging setting where expert actions are not available, and we show that HDT outperforms state-of-the-art baselines in terms of task success rates by a large margin.
2023-04-17
{'id': '2304.08487v1', 'title': 'Hyper-Decision Transformer for Efficient Online Policy Adaptation', 'abstract': "Decision Transformers (DT) have demonstrated strong performances in offline\nreinforcement learning settings, but quickly adapting to unseen novel tasks\nremains challenging. To address this challenge, we propose a new framework,\ncalled Hyper-Decision Transformer (HDT), that can generalize to novel tasks\nfrom a handful of demonstrations in a data- and parameter-efficient manner. To\nachieve such a goal, we propose to augment the base DT with an adaptation\nmodule, whose parameters are initialized by a hyper-network. When encountering\nunseen tasks, the hyper-network takes a handful of demonstrations as inputs and\ninitializes the adaptation module accordingly. This initialization enables HDT\nto efficiently adapt to novel tasks by only fine-tuning the adaptation module.\nWe validate HDT's generalization capability on object manipulation tasks. We\nfind that with a single expert demonstration and fine-tuning only 0.5% of DT\nparameters, HDT adapts faster to unseen tasks than fine-tuning the whole DT\nmodel. Finally, we explore a more challenging setting where expert actions are\nnot available, and we show that HDT outperforms state-of-the-art baselines in\nterms of task success rates by a large margin.", 'authors': ['Mengdi Xu', 'Yuchen Lu', 'Yikang Shen', 'Shun Zhang', 'Ding Zhao', 'Chuang Gan'], 'published': '2023-04-17 17:59:32+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CV', 'cs.RO'], 'url': 'http://arxiv.org/abs/2304.08487v1'}
42,740
2202.12166v1
2,202.12166
1
Attention Enables Zero Approximation Error
Deep learning models have been widely applied in various aspects of daily life. Many variant models based on deep learning structures have achieved even better performances. Attention-based architectures have become almost ubiquitous in deep learning structures. Especially, the transformer model has now defeated the convolutional neural network in image classification tasks to become the most widely used tool. However, the theoretical properties of attention-based models are seldom considered. In this work, we show that with suitable adaptations, the single-head self-attention transformer with a fixed number of transformer encoder blocks and free parameters is able to generate any desired polynomial of the input with no error. The number of transformer encoder blocks is the same as the degree of the target polynomial. Even more exciting, we find that these transformer encoder blocks in this model do not need to be trained. As a direct consequence, we show that the single-head self-attention transformer with increasing numbers of free parameters is universal. These surprising theoretical results clearly explain the outstanding performances of the transformer model and may shed light on future modifications in real applications. We also provide some experiments to verify our theoretical result.
2022-02-24
{'id': '2202.12166v1', 'title': 'Attention Enables Zero Approximation Error', 'abstract': 'Deep learning models have been widely applied in various aspects of daily\nlife. Many variant models based on deep learning structures have achieved even\nbetter performances. Attention-based architectures have become almost\nubiquitous in deep learning structures. Especially, the transformer model has\nnow defeated the convolutional neural network in image classification tasks to\nbecome the most widely used tool. However, the theoretical properties of\nattention-based models are seldom considered. In this work, we show that with\nsuitable adaptations, the single-head self-attention transformer with a fixed\nnumber of transformer encoder blocks and free parameters is able to generate\nany desired polynomial of the input with no error. The number of transformer\nencoder blocks is the same as the degree of the target polynomial. Even more\nexciting, we find that these transformer encoder blocks in this model do not\nneed to be trained. As a direct consequence, we show that the single-head\nself-attention transformer with increasing numbers of free parameters is\nuniversal. These surprising theoretical results clearly explain the outstanding\nperformances of the transformer model and may shed light on future\nmodifications in real applications. We also provide some experiments to verify\nour theoretical result.', 'authors': ['Zhiying Fang', 'Yidong Ouyang', 'Ding-Xuan Zhou', 'Guang Cheng'], 'published': '2022-02-24 16:06:01+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2202.12166v1'}
42,742
2208.06193v3
2,208.06193
3
Diffusion Policies as an Expressive Policy Class for Offline Reinforcement Learning
Offline reinforcement learning (RL), which aims to learn an optimal policy using a previously collected static dataset, is an important paradigm of RL. Standard RL methods often perform poorly in this regime due to the function approximation errors on out-of-distribution actions. While a variety of regularization methods have been proposed to mitigate this issue, they are often constrained by policy classes with limited expressiveness that can lead to highly suboptimal solutions. In this paper, we propose representing the policy as a diffusion model, a recent class of highly-expressive deep generative models. We introduce Diffusion Q-learning (Diffusion-QL) that utilizes a conditional diffusion model to represent the policy. In our approach, we learn an action-value function and we add a term maximizing action-values into the training loss of the conditional diffusion model, which results in a loss that seeks optimal actions that are near the behavior policy. We show the expressiveness of the diffusion model-based policy, and the coupling of the behavior cloning and policy improvement under the diffusion model both contribute to the outstanding performance of Diffusion-QL. We illustrate the superiority of our method compared to prior works in a simple 2D bandit example with a multimodal behavior policy. We then show that our method can achieve state-of-the-art performance on the majority of the D4RL benchmark tasks.
2022-08-12
{'id': '2208.06193v3', 'title': 'Diffusion Policies as an Expressive Policy Class for Offline Reinforcement Learning', 'abstract': 'Offline reinforcement learning (RL), which aims to learn an optimal policy\nusing a previously collected static dataset, is an important paradigm of RL.\nStandard RL methods often perform poorly in this regime due to the function\napproximation errors on out-of-distribution actions. While a variety of\nregularization methods have been proposed to mitigate this issue, they are\noften constrained by policy classes with limited expressiveness that can lead\nto highly suboptimal solutions. In this paper, we propose representing the\npolicy as a diffusion model, a recent class of highly-expressive deep\ngenerative models. We introduce Diffusion Q-learning (Diffusion-QL) that\nutilizes a conditional diffusion model to represent the policy. In our\napproach, we learn an action-value function and we add a term maximizing\naction-values into the training loss of the conditional diffusion model, which\nresults in a loss that seeks optimal actions that are near the behavior policy.\nWe show the expressiveness of the diffusion model-based policy, and the\ncoupling of the behavior cloning and policy improvement under the diffusion\nmodel both contribute to the outstanding performance of Diffusion-QL. We\nillustrate the superiority of our method compared to prior works in a simple 2D\nbandit example with a multimodal behavior policy. We then show that our method\ncan achieve state-of-the-art performance on the majority of the D4RL benchmark\ntasks.', 'authors': ['Zhendong Wang', 'Jonathan J Hunt', 'Mingyuan Zhou'], 'published': '2022-08-12 09:54:11+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2208.06193v3'}
42,744
2210.06820v2
2,210.0682
2
Personalized Federated Hypernetworks for Privacy Preservation in Multi-Task Reinforcement Learning
Multi-Agent Reinforcement Learning currently focuses on implementations where all data and training can be centralized to one machine. But what if local agents are split across multiple tasks, and need to keep data private between each? We develop the first application of Personalized Federated Hypernetworks (PFH) to Reinforcement Learning (RL). We then present a novel application of PFH to few-shot transfer, and demonstrate significant initial increases in learning. PFH has never been demonstrated beyond supervised learning benchmarks, so we apply PFH to an important domain: RL price-setting for energy demand response. We consider a general case across where agents are split across multiple microgrids, wherein energy consumption data must be kept private within each microgrid. Together, our work explores how the fields of personalized federated learning and RL can come together to make learning efficient across multiple tasks while keeping data secure.
2022-10-13
{'id': '2210.06820v2', 'title': 'Personalized Federated Hypernetworks for Privacy Preservation in Multi-Task Reinforcement Learning', 'abstract': 'Multi-Agent Reinforcement Learning currently focuses on implementations where\nall data and training can be centralized to one machine. But what if local\nagents are split across multiple tasks, and need to keep data private between\neach? We develop the first application of Personalized Federated Hypernetworks\n(PFH) to Reinforcement Learning (RL). We then present a novel application of\nPFH to few-shot transfer, and demonstrate significant initial increases in\nlearning. PFH has never been demonstrated beyond supervised learning\nbenchmarks, so we apply PFH to an important domain: RL price-setting for energy\ndemand response. We consider a general case across where agents are split\nacross multiple microgrids, wherein energy consumption data must be kept\nprivate within each microgrid. Together, our work explores how the fields of\npersonalized federated learning and RL can come together to make learning\nefficient across multiple tasks while keeping data secure.', 'authors': ['Doseok Jang', 'Larry Yan', 'Lucas Spangher', 'Costas J. Spanos'], 'published': '2022-10-13 08:11:12+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.SY', 'eess.SY'], 'url': 'http://arxiv.org/abs/2210.06820v2'}
42,738
2310.08587v3
2,310.08587
3
Pseudo-Generalized Dynamic View Synthesis from a Video
Rendering scenes observed in a monocular video from novel viewpoints is a challenging problem. For static scenes the community has studied both scene-specific optimization techniques, which optimize on every test scene, and generalized techniques, which only run a deep net forward pass on a test scene. In contrast, for dynamic scenes, scene-specific optimization techniques exist, but, to our best knowledge, there is currently no generalized method for dynamic novel view synthesis from a given monocular video. To answer whether generalized dynamic novel view synthesis from monocular videos is possible today, we establish an analysis framework based on existing techniques and work toward the generalized approach. We find a pseudo-generalized process without scene-specific appearance optimization is possible, but geometrically and temporally consistent depth estimates are needed. Despite no scene-specific appearance optimization, the pseudo-generalized approach improves upon some scene-specific methods.
2023-10-12
{'id': '2310.08587v3', 'title': 'Pseudo-Generalized Dynamic View Synthesis from a Video', 'abstract': 'Rendering scenes observed in a monocular video from novel viewpoints is a\nchallenging problem. For static scenes the community has studied both\nscene-specific optimization techniques, which optimize on every test scene, and\ngeneralized techniques, which only run a deep net forward pass on a test scene.\nIn contrast, for dynamic scenes, scene-specific optimization techniques exist,\nbut, to our best knowledge, there is currently no generalized method for\ndynamic novel view synthesis from a given monocular video. To answer whether\ngeneralized dynamic novel view synthesis from monocular videos is possible\ntoday, we establish an analysis framework based on existing techniques and work\ntoward the generalized approach. We find a pseudo-generalized process without\nscene-specific appearance optimization is possible, but geometrically and\ntemporally consistent depth estimates are needed. Despite no scene-specific\nappearance optimization, the pseudo-generalized approach improves upon some\nscene-specific methods.', 'authors': ['Xiaoming Zhao', 'Alex Colburn', 'Fangchang Ma', 'Miguel Angel Bautista', 'Joshua M. Susskind', 'Alexander G. Schwing'], 'published': '2023-10-12 17:59:58+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2310.08587v3'}
42,739
2311.02124v1
2,311.02124
1
Sliced Denoising: A Physics-Informed Molecular Pre-Training Method
While molecular pre-training has shown great potential in enhancing drug discovery, the lack of a solid physical interpretation in current methods raises concerns about whether the learned representation truly captures the underlying explanatory factors in observed data, ultimately resulting in limited generalization and robustness. Although denoising methods offer a physical interpretation, their accuracy is often compromised by ad-hoc noise design, leading to inaccurate learned force fields. To address this limitation, this paper proposes a new method for molecular pre-training, called sliced denoising (SliDe), which is based on the classical mechanical intramolecular potential theory. SliDe utilizes a novel noise strategy that perturbs bond lengths, angles, and torsion angles to achieve better sampling over conformations. Additionally, it introduces a random slicing approach that circumvents the computationally expensive calculation of the Jacobian matrix, which is otherwise essential for estimating the force field. By aligning with physical principles, SliDe shows a 42\% improvement in the accuracy of estimated force fields compared to current state-of-the-art denoising methods, and thus outperforms traditional baselines on various molecular property prediction tasks.
2023-11-03
{'id': '2311.02124v1', 'title': 'Sliced Denoising: A Physics-Informed Molecular Pre-Training Method', 'abstract': 'While molecular pre-training has shown great potential in enhancing drug\ndiscovery, the lack of a solid physical interpretation in current methods\nraises concerns about whether the learned representation truly captures the\nunderlying explanatory factors in observed data, ultimately resulting in\nlimited generalization and robustness. Although denoising methods offer a\nphysical interpretation, their accuracy is often compromised by ad-hoc noise\ndesign, leading to inaccurate learned force fields. To address this limitation,\nthis paper proposes a new method for molecular pre-training, called sliced\ndenoising (SliDe), which is based on the classical mechanical intramolecular\npotential theory. SliDe utilizes a novel noise strategy that perturbs bond\nlengths, angles, and torsion angles to achieve better sampling over\nconformations. Additionally, it introduces a random slicing approach that\ncircumvents the computationally expensive calculation of the Jacobian matrix,\nwhich is otherwise essential for estimating the force field. By aligning with\nphysical principles, SliDe shows a 42\\% improvement in the accuracy of\nestimated force fields compared to current state-of-the-art denoising methods,\nand thus outperforms traditional baselines on various molecular property\nprediction tasks.', 'authors': ['Yuyan Ni', 'Shikun Feng', 'Wei-Ying Ma', 'Zhi-Ming Ma', 'Yanyan Lan'], 'published': '2023-11-03 07:58:05+00:00', 'categories': ['q-bio.BM', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2311.02124v1'}
42,741
2310.06161v1
2,310.06161
1
Mitigating Simplicity Bias in Deep Learning for Improved OOD Generalization and Robustness
Neural networks (NNs) are known to exhibit simplicity bias where they tend to prefer learning 'simple' features over more 'complex' ones, even when the latter may be more informative. Simplicity bias can lead to the model making biased predictions which have poor out-of-distribution (OOD) generalization. To address this, we propose a framework that encourages the model to use a more diverse set of features to make predictions. We first train a simple model, and then regularize the conditional mutual information with respect to it to obtain the final model. We demonstrate the effectiveness of this framework in various problem settings and real-world applications, showing that it effectively addresses simplicity bias and leads to more features being used, enhances OOD generalization, and improves subgroup robustness and fairness. We complement these results with theoretical analyses of the effect of the regularization and its OOD generalization properties.
2023-10-09
{'id': '2310.06161v1', 'title': 'Mitigating Simplicity Bias in Deep Learning for Improved OOD Generalization and Robustness', 'abstract': "Neural networks (NNs) are known to exhibit simplicity bias where they tend to\nprefer learning 'simple' features over more 'complex' ones, even when the\nlatter may be more informative. Simplicity bias can lead to the model making\nbiased predictions which have poor out-of-distribution (OOD) generalization. To\naddress this, we propose a framework that encourages the model to use a more\ndiverse set of features to make predictions. We first train a simple model, and\nthen regularize the conditional mutual information with respect to it to obtain\nthe final model. We demonstrate the effectiveness of this framework in various\nproblem settings and real-world applications, showing that it effectively\naddresses simplicity bias and leads to more features being used, enhances OOD\ngeneralization, and improves subgroup robustness and fairness. We complement\nthese results with theoretical analyses of the effect of the regularization and\nits OOD generalization properties.", 'authors': ['Bhavya Vasudeva', 'Kameron Shahabi', 'Vatsal Sharan'], 'published': '2023-10-09 21:19:39+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2310.06161v1'}
42,743
2311.04829v2
2,311.04829
2
Functional Bayesian Tucker Decomposition for Continuous-indexed Tensor Data
Tucker decomposition is a powerful tensor model to handle multi-aspect data. It demonstrates the low-rank property by decomposing the grid-structured data as interactions between a core tensor and a set of object representations (factors). A fundamental assumption of such decomposition is that there are finite objects in each aspect or mode, corresponding to discrete indexes of data entries. However, real-world data is often not naturally posed in this setting. For example, geographic data is represented as continuous indexes of latitude and longitude coordinates, and cannot fit tensor models directly. To generalize Tucker decomposition to such scenarios, we propose Functional Bayesian Tucker Decomposition (FunBaT). We treat the continuous-indexed data as the interaction between the Tucker core and a group of latent functions. We use Gaussian processes (GP) as functional priors to model the latent functions. Then, we convert each GP into a state-space prior by constructing an equivalent stochastic differential equation (SDE) to reduce computational cost. An efficient inference algorithm is developed for scalable posterior approximation based on advanced message-passing techniques. The advantage of our method is shown in both synthetic data and several real-world applications. We release the code of FunBaT at https://github.com/xuangu-fang/Functional-Bayesian-Tucker-Decomposition.
2023-11-08
{'id': '2311.04829v2', 'title': 'Functional Bayesian Tucker Decomposition for Continuous-indexed Tensor Data', 'abstract': 'Tucker decomposition is a powerful tensor model to handle multi-aspect data.\nIt demonstrates the low-rank property by decomposing the grid-structured data\nas interactions between a core tensor and a set of object representations\n(factors). A fundamental assumption of such decomposition is that there are\nfinite objects in each aspect or mode, corresponding to discrete indexes of\ndata entries. However, real-world data is often not naturally posed in this\nsetting. For example, geographic data is represented as continuous indexes of\nlatitude and longitude coordinates, and cannot fit tensor models directly. To\ngeneralize Tucker decomposition to such scenarios, we propose Functional\nBayesian Tucker Decomposition (FunBaT). We treat the continuous-indexed data as\nthe interaction between the Tucker core and a group of latent functions. We use\nGaussian processes (GP) as functional priors to model the latent functions.\nThen, we convert each GP into a state-space prior by constructing an equivalent\nstochastic differential equation (SDE) to reduce computational cost. An\nefficient inference algorithm is developed for scalable posterior approximation\nbased on advanced message-passing techniques. The advantage of our method is\nshown in both synthetic data and several real-world applications. We release\nthe code of FunBaT at\n\\url{https://github.com/xuangu-fang/Functional-Bayesian-Tucker-Decomposition}.', 'authors': ['Shikai Fang', 'Xin Yu', 'Zheng Wang', 'Shibo Li', 'Mike Kirby', 'Shandian Zhe'], 'published': '2023-11-08 16:54:23+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2311.04829v2'}
42,745
2311.16160v1
2,311.1616
1
Protein-ligand binding representation learning from fine-grained interactions
The binding between proteins and ligands plays a crucial role in the realm of drug discovery. Previous deep learning approaches have shown promising results over traditional computationally intensive methods, but resulting in poor generalization due to limited supervised data. In this paper, we propose to learn protein-ligand binding representation in a self-supervised learning manner. Different from existing pre-training approaches which treat proteins and ligands individually, we emphasize to discern the intricate binding patterns from fine-grained interactions. Specifically, this self-supervised learning problem is formulated as a prediction of the conclusive binding complex structure given a pocket and ligand with a Transformer based interaction module, which naturally emulates the binding process. To ensure the representation of rich binding information, we introduce two pre-training tasks, i.e.~atomic pairwise distance map prediction and mask ligand reconstruction, which comprehensively model the fine-grained interactions from both structure and feature space. Extensive experiments have demonstrated the superiority of our method across various binding tasks, including protein-ligand affinity prediction, virtual screening and protein-ligand docking.
2023-11-09
{'id': '2311.16160v1', 'title': 'Protein-ligand binding representation learning from fine-grained interactions', 'abstract': 'The binding between proteins and ligands plays a crucial role in the realm of\ndrug discovery. Previous deep learning approaches have shown promising results\nover traditional computationally intensive methods, but resulting in poor\ngeneralization due to limited supervised data. In this paper, we propose to\nlearn protein-ligand binding representation in a self-supervised learning\nmanner. Different from existing pre-training approaches which treat proteins\nand ligands individually, we emphasize to discern the intricate binding\npatterns from fine-grained interactions. Specifically, this self-supervised\nlearning problem is formulated as a prediction of the conclusive binding\ncomplex structure given a pocket and ligand with a Transformer based\ninteraction module, which naturally emulates the binding process. To ensure the\nrepresentation of rich binding information, we introduce two pre-training\ntasks, i.e.~atomic pairwise distance map prediction and mask ligand\nreconstruction, which comprehensively model the fine-grained interactions from\nboth structure and feature space. Extensive experiments have demonstrated the\nsuperiority of our method across various binding tasks, including\nprotein-ligand affinity prediction, virtual screening and protein-ligand\ndocking.', 'authors': ['Shikun Feng', 'Minghao Li', 'Yinjun Jia', 'Weiying Ma', 'Yanyan Lan'], 'published': '2023-11-09 01:33:09+00:00', 'categories': ['q-bio.BM', 'cs.LG'], 'url': 'http://arxiv.org/abs/2311.16160v1'}
42,747
2310.13822v2
2,310.13822
2
Adversarial Attacks on Fairness of Graph Neural Networks
Fairness-aware graph neural networks (GNNs) have gained a surge of attention as they can reduce the bias of predictions on any demographic group (e.g., female) in graph-based applications. Although these methods greatly improve the algorithmic fairness of GNNs, the fairness can be easily corrupted by carefully designed adversarial attacks. In this paper, we investigate the problem of adversarial attacks on fairness of GNNs and propose G-FairAttack, a general framework for attacking various types of fairness-aware GNNs in terms of fairness with an unnoticeable effect on prediction utility. In addition, we propose a fast computation technique to reduce the time complexity of G-FairAttack. The experimental study demonstrates that G-FairAttack successfully corrupts the fairness of different types of GNNs while keeping the attack unnoticeable. Our study on fairness attacks sheds light on potential vulnerabilities in fairness-aware GNNs and guides further research on the robustness of GNNs in terms of fairness.
2023-10-20
{'id': '2310.13822v2', 'title': 'Adversarial Attacks on Fairness of Graph Neural Networks', 'abstract': 'Fairness-aware graph neural networks (GNNs) have gained a surge of attention\nas they can reduce the bias of predictions on any demographic group (e.g.,\nfemale) in graph-based applications. Although these methods greatly improve the\nalgorithmic fairness of GNNs, the fairness can be easily corrupted by carefully\ndesigned adversarial attacks. In this paper, we investigate the problem of\nadversarial attacks on fairness of GNNs and propose G-FairAttack, a general\nframework for attacking various types of fairness-aware GNNs in terms of\nfairness with an unnoticeable effect on prediction utility. In addition, we\npropose a fast computation technique to reduce the time complexity of\nG-FairAttack. The experimental study demonstrates that G-FairAttack\nsuccessfully corrupts the fairness of different types of GNNs while keeping the\nattack unnoticeable. Our study on fairness attacks sheds light on potential\nvulnerabilities in fairness-aware GNNs and guides further research on the\nrobustness of GNNs in terms of fairness.', 'authors': ['Binchi Zhang', 'Yushun Dong', 'Chen Chen', 'Yada Zhu', 'Minnan Luo', 'Jundong Li'], 'published': '2023-10-20 21:19:54+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.13822v2'}
42,748
2303.00257v1
2,303.00257
1
Hidden Markov Transformer for Simultaneous Machine Translation
Simultaneous machine translation (SiMT) outputs the target sequence while receiving the source sequence, and hence learning when to start translating each target token is the core challenge for SiMT task. However, it is non-trivial to learn the optimal moment among many possible moments of starting translating, as the moments of starting translating always hide inside the model and can only be supervised with the observed target sequence. In this paper, we propose a Hidden Markov Transformer (HMT), which treats the moments of starting translating as hidden events and the target sequence as the corresponding observed events, thereby organizing them as a hidden Markov model. HMT explicitly models multiple moments of starting translating as the candidate hidden events, and then selects one to generate the target token. During training, by maximizing the marginal likelihood of the target sequence over multiple moments of starting translating, HMT learns to start translating at the moments that target tokens can be generated more accurately. Experiments on multiple SiMT benchmarks show that HMT outperforms strong baselines and achieves state-of-the-art performance.
2023-03-01
{'id': '2303.00257v1', 'title': 'Hidden Markov Transformer for Simultaneous Machine Translation', 'abstract': 'Simultaneous machine translation (SiMT) outputs the target sequence while\nreceiving the source sequence, and hence learning when to start translating\neach target token is the core challenge for SiMT task. However, it is\nnon-trivial to learn the optimal moment among many possible moments of starting\ntranslating, as the moments of starting translating always hide inside the\nmodel and can only be supervised with the observed target sequence. In this\npaper, we propose a Hidden Markov Transformer (HMT), which treats the moments\nof starting translating as hidden events and the target sequence as the\ncorresponding observed events, thereby organizing them as a hidden Markov\nmodel. HMT explicitly models multiple moments of starting translating as the\ncandidate hidden events, and then selects one to generate the target token.\nDuring training, by maximizing the marginal likelihood of the target sequence\nover multiple moments of starting translating, HMT learns to start translating\nat the moments that target tokens can be generated more accurately. Experiments\non multiple SiMT benchmarks show that HMT outperforms strong baselines and\nachieves state-of-the-art performance.', 'authors': ['Shaolei Zhang', 'Yang Feng'], 'published': '2023-03-01 06:28:33+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2303.00257v1'}
42,751
2305.19982v1
2,305.19982
1
Adam Accumulation to Reduce Memory Footprints of both Activations and Gradients for Large-scale DNN Training
Running out of GPU memory has become a main bottleneck for large-scale DNN training. How to reduce the memory footprint during training has received intensive research attention. We find that previous gradient accumulation reduces activation memory but fails to be compatible with gradient memory reduction due to a contradiction between preserving gradients and releasing gradients. To address this issue, we propose a novel optimizer accumulation method for Adam, named Adam Accumulation (AdamA), which enables reducing both activation and gradient memory. Specifically, AdamA directly integrates gradients into optimizer states and accumulates optimizer states over micro-batches, so that gradients can be released immediately after use. We mathematically and experimentally demonstrate AdamA yields the same convergence properties as Adam. Evaluated on transformer-based models, AdamA achieves up to 23% memory reduction compared to gradient accumulation with less than 2% degradation in training throughput. Notably, AdamA can work together with memory reduction methods for optimizer states to fit 1.26x~3.14x larger models over PyTorch and DeepSpeed baseline on GPUs with different memory capacities.
2023-05-31
{'id': '2305.19982v1', 'title': 'Adam Accumulation to Reduce Memory Footprints of both Activations and Gradients for Large-scale DNN Training', 'abstract': 'Running out of GPU memory has become a main bottleneck for large-scale DNN\ntraining. How to reduce the memory footprint during training has received\nintensive research attention. We find that previous gradient accumulation\nreduces activation memory but fails to be compatible with gradient memory\nreduction due to a contradiction between preserving gradients and releasing\ngradients. To address this issue, we propose a novel optimizer accumulation\nmethod for Adam, named Adam Accumulation (AdamA), which enables reducing both\nactivation and gradient memory. Specifically, AdamA directly integrates\ngradients into optimizer states and accumulates optimizer states over\nmicro-batches, so that gradients can be released immediately after use. We\nmathematically and experimentally demonstrate AdamA yields the same convergence\nproperties as Adam. Evaluated on transformer-based models, AdamA achieves up to\n23% memory reduction compared to gradient accumulation with less than 2%\ndegradation in training throughput. Notably, AdamA can work together with\nmemory reduction methods for optimizer states to fit 1.26x~3.14x larger models\nover PyTorch and DeepSpeed baseline on GPUs with different memory capacities.', 'authors': ['Yijia Zhang', 'Yibo Han', 'Shijie Cao', 'Guohao Dai', 'Youshan Miao', 'Ting Cao', 'Fan Yang', 'Ningyi Xu'], 'published': '2023-05-31 16:06:50+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2305.19982v1'}
42,753
2304.06738v1
2,304.06738
1
A Study of Biologically Plausible Neural Network: The Role and Interactions of Brain-Inspired Mechanisms in Continual Learning
Humans excel at continually acquiring, consolidating, and retaining information from an ever-changing environment, whereas artificial neural networks (ANNs) exhibit catastrophic forgetting. There are considerable differences in the complexity of synapses, the processing of information, and the learning mechanisms in biological neural networks and their artificial counterparts, which may explain the mismatch in performance. We consider a biologically plausible framework that constitutes separate populations of exclusively excitatory and inhibitory neurons that adhere to Dale's principle, and the excitatory pyramidal neurons are augmented with dendritic-like structures for context-dependent processing of stimuli. We then conduct a comprehensive study on the role and interactions of different mechanisms inspired by the brain, including sparse non-overlapping representations, Hebbian learning, synaptic consolidation, and replay of past activations that accompanied the learning event. Our study suggests that the employing of multiple complementary mechanisms in a biologically plausible architecture, similar to the brain, may be effective in enabling continual learning in ANNs.
2023-04-13
{'id': '2304.06738v1', 'title': 'A Study of Biologically Plausible Neural Network: The Role and Interactions of Brain-Inspired Mechanisms in Continual Learning', 'abstract': "Humans excel at continually acquiring, consolidating, and retaining\ninformation from an ever-changing environment, whereas artificial neural\nnetworks (ANNs) exhibit catastrophic forgetting. There are considerable\ndifferences in the complexity of synapses, the processing of information, and\nthe learning mechanisms in biological neural networks and their artificial\ncounterparts, which may explain the mismatch in performance. We consider a\nbiologically plausible framework that constitutes separate populations of\nexclusively excitatory and inhibitory neurons that adhere to Dale's principle,\nand the excitatory pyramidal neurons are augmented with dendritic-like\nstructures for context-dependent processing of stimuli. We then conduct a\ncomprehensive study on the role and interactions of different mechanisms\ninspired by the brain, including sparse non-overlapping representations,\nHebbian learning, synaptic consolidation, and replay of past activations that\naccompanied the learning event. Our study suggests that the employing of\nmultiple complementary mechanisms in a biologically plausible architecture,\nsimilar to the brain, may be effective in enabling continual learning in ANNs.", 'authors': ['Fahad Sarfraz', 'Elahe Arani', 'Bahram Zonooz'], 'published': '2023-04-13 16:34:12+00:00', 'categories': ['cs.NE', 'cs.AI', 'cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2304.06738v1'}
42,756
2301.05217v3
2,301.05217
3
Progress measures for grokking via mechanistic interpretability
Neural networks often exhibit emergent behavior, where qualitatively new capabilities arise from scaling up the amount of parameters, training data, or training steps. One approach to understanding emergence is to find continuous progress measures that underlie the seemingly discontinuous qualitative changes. We argue that progress measures can be found via mechanistic interpretability: reverse-engineering learned behaviors into their individual components. As a case study, we investigate the recently-discovered phenomenon of ``grokking'' exhibited by small transformers trained on modular addition tasks. We fully reverse engineer the algorithm learned by these networks, which uses discrete Fourier transforms and trigonometric identities to convert addition to rotation about a circle. We confirm the algorithm by analyzing the activations and weights and by performing ablations in Fourier space. Based on this understanding, we define progress measures that allow us to study the dynamics of training and split training into three continuous phases: memorization, circuit formation, and cleanup. Our results show that grokking, rather than being a sudden shift, arises from the gradual amplification of structured mechanisms encoded in the weights, followed by the later removal of memorizing components.
2023-01-12
{'id': '2301.05217v3', 'title': 'Progress measures for grokking via mechanistic interpretability', 'abstract': "Neural networks often exhibit emergent behavior, where qualitatively new\ncapabilities arise from scaling up the amount of parameters, training data, or\ntraining steps. One approach to understanding emergence is to find continuous\n\\textit{progress measures} that underlie the seemingly discontinuous\nqualitative changes. We argue that progress measures can be found via\nmechanistic interpretability: reverse-engineering learned behaviors into their\nindividual components. As a case study, we investigate the recently-discovered\nphenomenon of ``grokking'' exhibited by small transformers trained on modular\naddition tasks. We fully reverse engineer the algorithm learned by these\nnetworks, which uses discrete Fourier transforms and trigonometric identities\nto convert addition to rotation about a circle. We confirm the algorithm by\nanalyzing the activations and weights and by performing ablations in Fourier\nspace. Based on this understanding, we define progress measures that allow us\nto study the dynamics of training and split training into three continuous\nphases: memorization, circuit formation, and cleanup. Our results show that\ngrokking, rather than being a sudden shift, arises from the gradual\namplification of structured mechanisms encoded in the weights, followed by the\nlater removal of memorizing components.", 'authors': ['Neel Nanda', 'Lawrence Chan', 'Tom Lieberum', 'Jess Smith', 'Jacob Steinhardt'], 'published': '2023-01-12 18:56:49+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2301.05217v3'}
42,758
2207.12600v2
2,207.126
2
Learning Hierarchical Protein Representations via Complete 3D Graph Networks
We consider representation learning for proteins with 3D structures. We build 3D graphs based on protein structures and develop graph networks to learn their representations. Depending on the levels of details that we wish to capture, protein representations can be computed at different levels, e.g., the amino acid, backbone, or all-atom levels. Importantly, there exist hierarchical relations among different levels. In this work, we propose to develop a novel hierarchical graph network, known as ProNet, to capture the relations. Our ProNet is very flexible and can be used to compute protein representations at different levels of granularity. By treating each amino acid as a node in graph modeling as well as harnessing the inherent hierarchies, our ProNet is more effective and efficient than existing methods. We also show that, given a base 3D graph network that is complete, our ProNet representations are also complete at all levels. Experimental results show that ProNet outperforms recent methods on most datasets. In addition, results indicate that different downstream tasks may require representations at different levels. Our code is publicly available as part of the DIG library (https://github.com/divelab/DIG).
2022-07-26
{'id': '2207.12600v2', 'title': 'Learning Hierarchical Protein Representations via Complete 3D Graph Networks', 'abstract': 'We consider representation learning for proteins with 3D structures. We build\n3D graphs based on protein structures and develop graph networks to learn their\nrepresentations. Depending on the levels of details that we wish to capture,\nprotein representations can be computed at different levels, \\emph{e.g.}, the\namino acid, backbone, or all-atom levels. Importantly, there exist hierarchical\nrelations among different levels. In this work, we propose to develop a novel\nhierarchical graph network, known as ProNet, to capture the relations. Our\nProNet is very flexible and can be used to compute protein representations at\ndifferent levels of granularity. By treating each amino acid as a node in graph\nmodeling as well as harnessing the inherent hierarchies, our ProNet is more\neffective and efficient than existing methods. We also show that, given a base\n3D graph network that is complete, our ProNet representations are also complete\nat all levels. Experimental results show that ProNet outperforms recent methods\non most datasets. In addition, results indicate that different downstream tasks\nmay require representations at different levels. Our code is publicly available\nas part of the DIG library (\\url{https://github.com/divelab/DIG}).', 'authors': ['Limei Wang', 'Haoran Liu', 'Yi Liu', 'Jerry Kurtin', 'Shuiwang Ji'], 'published': '2022-07-26 01:55:25+00:00', 'categories': ['cs.LG', 'q-bio.QM'], 'url': 'http://arxiv.org/abs/2207.12600v2'}
42,749
2308.12519v3
2,308.12519
3
Rational Decision-Making Agent with Internalized Utility Judgment
Large language models (LLMs) have demonstrated remarkable advancements and have attracted significant efforts to develop LLMs into agents capable of executing intricate multi-step decision-making tasks beyond traditional NLP applications. Existing approaches to LLM-based decision-making predominantly build upon the manually-designed external performance metrics to guide the decision-making process. However, reliance on the external performance metrics as prior is problematic in real-world scenarios, where such prior may be unavailable, flawed, or even erroneous. For genuine autonomous decision making, it is imperative for the agent to develop its rationality from its posterior experiences to judge decisions independently. Central to the development of rationality is the construction of an internalized utility judgment, capable of assigning numerical utilities to each decision. This paper proposes RadAgent (Rational Decision-Making Agent), which fosters the development of its rationality through an iterative framework involving Experience Exploration and Utility Learning. Within this framework, Elo-based Utility Construction is devised to assign Elo scores to individual decision steps to judge their utilities via pairwise comparisons. Consequently, these Elo scores guide the decision-making process to derive optimal outcomes. Experimental results on the ToolBench dataset demonstrate RadAgent's superiority over baselines, achieving over 10% improvement in Pass Rate on diverse tasks. It offers higher-quality solutions and reduces costs (ChatGPT API calls), highlighting its effectiveness and efficiency.
2023-08-24
{'id': '2308.12519v3', 'title': 'Rational Decision-Making Agent with Internalized Utility Judgment', 'abstract': "Large language models (LLMs) have demonstrated remarkable advancements and\nhave attracted significant efforts to develop LLMs into agents capable of\nexecuting intricate multi-step decision-making tasks beyond traditional NLP\napplications. Existing approaches to LLM-based decision-making predominantly\nbuild upon the manually-designed external performance metrics to guide the\ndecision-making process. However, reliance on the external performance metrics\nas prior is problematic in real-world scenarios, where such prior may be\nunavailable, flawed, or even erroneous. For genuine autonomous decision making,\nit is imperative for the agent to develop its rationality from its posterior\nexperiences to judge decisions independently. Central to the development of\nrationality is the construction of an internalized utility judgment, capable of\nassigning numerical utilities to each decision. This paper proposes RadAgent\n(Rational Decision-Making Agent), which fosters the development of its\nrationality through an iterative framework involving Experience Exploration and\nUtility Learning. Within this framework, Elo-based Utility Construction is\ndevised to assign Elo scores to individual decision steps to judge their\nutilities via pairwise comparisons. Consequently, these Elo scores guide the\ndecision-making process to derive optimal outcomes. Experimental results on the\nToolBench dataset demonstrate RadAgent's superiority over baselines, achieving\nover 10% improvement in Pass Rate on diverse tasks. It offers higher-quality\nsolutions and reduces costs (ChatGPT API calls), highlighting its effectiveness\nand efficiency.", 'authors': ['Yining Ye', 'Xin Cong', 'Shizuo Tian', 'Yujia Qin', 'Chong Liu', 'Yankai Lin', 'Zhiyuan Liu', 'Maosong Sun'], 'published': '2023-08-24 03:11:45+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2308.12519v3'}
42,750
2401.11611v1
2,401.11611
1
Continuous Field Reconstruction from Sparse Observations with Implicit Neural Networks
Reliably reconstructing physical fields from sparse sensor data is a challenge that frequently arises in many scientific domains. In practice, the process generating the data often is not understood to sufficient accuracy. Therefore, there is a growing interest in using the deep neural network route to address the problem. This work presents a novel approach that learns a continuous representation of the physical field using implicit neural representations (INRs). Specifically, after factorizing spatiotemporal variability into spatial and temporal components using the separation of variables technique, the method learns relevant basis functions from sparsely sampled irregular data points to develop a continuous representation of the data. In experimental evaluations, the proposed model outperforms recent INR methods, offering superior reconstruction quality on simulation data from a state-of-the-art climate model and a second dataset that comprises ultra-high resolution satellite-based sea surface temperature fields.
2024-01-21
{'id': '2401.11611v1', 'title': 'Continuous Field Reconstruction from Sparse Observations with Implicit Neural Networks', 'abstract': 'Reliably reconstructing physical fields from sparse sensor data is a\nchallenge that frequently arises in many scientific domains. In practice, the\nprocess generating the data often is not understood to sufficient accuracy.\nTherefore, there is a growing interest in using the deep neural network route\nto address the problem. This work presents a novel approach that learns a\ncontinuous representation of the physical field using implicit neural\nrepresentations (INRs). Specifically, after factorizing spatiotemporal\nvariability into spatial and temporal components using the separation of\nvariables technique, the method learns relevant basis functions from sparsely\nsampled irregular data points to develop a continuous representation of the\ndata. In experimental evaluations, the proposed model outperforms recent INR\nmethods, offering superior reconstruction quality on simulation data from a\nstate-of-the-art climate model and a second dataset that comprises ultra-high\nresolution satellite-based sea surface temperature fields.', 'authors': ['Xihaier Luo', 'Wei Xu', 'Yihui Ren', 'Shinjae Yoo', 'Balu Nadiga'], 'published': '2024-01-21 22:18:29+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2401.11611v1'}
42,752
2305.02139v1
2,305.02139
1
A Curriculum View of Robust Loss Functions
Robust loss functions are designed to combat the adverse impacts of label noise, whose robustness is typically supported by theoretical bounds agnostic to the training dynamics. However, these bounds may fail to characterize the empirical performance as it remains unclear why robust loss functions can underfit. We show that most loss functions can be rewritten into a form with the same class-score margin and different sample-weighting functions. The resulting curriculum view provides a straightforward analysis of the training dynamics, which helps attribute underfitting to diminished average sample weights and noise robustness to larger weights for clean samples. We show that simple fixes to the curriculums can make underfitting robust loss functions competitive with the state-of-the-art, and training schedules can substantially affect the noise robustness even with robust loss functions. Code is available at github.
2023-05-03
{'id': '2305.02139v1', 'title': 'A Curriculum View of Robust Loss Functions', 'abstract': 'Robust loss functions are designed to combat the adverse impacts of label\nnoise, whose robustness is typically supported by theoretical bounds agnostic\nto the training dynamics. However, these bounds may fail to characterize the\nempirical performance as it remains unclear why robust loss functions can\nunderfit. We show that most loss functions can be rewritten into a form with\nthe same class-score margin and different sample-weighting functions. The\nresulting curriculum view provides a straightforward analysis of the training\ndynamics, which helps attribute underfitting to diminished average sample\nweights and noise robustness to larger weights for clean samples. We show that\nsimple fixes to the curriculums can make underfitting robust loss functions\ncompetitive with the state-of-the-art, and training schedules can substantially\naffect the noise robustness even with robust loss functions. Code is available\nat \\url{github}.', 'authors': ['Zebin Ou', 'Yue Zhang'], 'published': '2023-05-03 14:13:03+00:00', 'categories': ['cs.LG', 'cs.CL'], 'url': 'http://arxiv.org/abs/2305.02139v1'}
42,754
2310.04475v2
2,310.04475
2
Demystifying Embedding Spaces using Large Language Models
Embeddings have become a pivotal means to represent complex, multi-faceted information about entities, concepts, and relationships in a condensed and useful format. Nevertheless, they often preclude direct interpretation. While downstream tasks make use of these compressed representations, meaningful interpretation usually requires visualization using dimensionality reduction or specialized machine learning interpretability methods. This paper addresses the challenge of making such embeddings more interpretable and broadly useful, by employing Large Language Models (LLMs) to directly interact with embeddings -- transforming abstract vectors into understandable narratives. By injecting embeddings into LLMs, we enable querying and exploration of complex embedding data. We demonstrate our approach on a variety of diverse tasks, including: enhancing concept activation vectors (CAVs), communicating novel embedded entities, and decoding user preferences in recommender systems. Our work couples the immense information potential of embeddings with the interpretative power of LLMs.
2023-10-06
{'id': '2310.04475v2', 'title': 'Demystifying Embedding Spaces using Large Language Models', 'abstract': 'Embeddings have become a pivotal means to represent complex, multi-faceted\ninformation about entities, concepts, and relationships in a condensed and\nuseful format. Nevertheless, they often preclude direct interpretation. While\ndownstream tasks make use of these compressed representations, meaningful\ninterpretation usually requires visualization using dimensionality reduction or\nspecialized machine learning interpretability methods. This paper addresses the\nchallenge of making such embeddings more interpretable and broadly useful, by\nemploying Large Language Models (LLMs) to directly interact with embeddings --\ntransforming abstract vectors into understandable narratives. By injecting\nembeddings into LLMs, we enable querying and exploration of complex embedding\ndata. We demonstrate our approach on a variety of diverse tasks, including:\nenhancing concept activation vectors (CAVs), communicating novel embedded\nentities, and decoding user preferences in recommender systems. Our work\ncouples the immense information potential of embeddings with the interpretative\npower of LLMs.', 'authors': ['Guy Tennenholtz', 'Yinlam Chow', 'Chih-Wei Hsu', 'Jihwan Jeong', 'Lior Shani', 'Azamat Tulepbergenov', 'Deepak Ramachandran', 'Martin Mladenov', 'Craig Boutilier'], 'published': '2023-10-06 05:27:28+00:00', 'categories': ['cs.CL', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2310.04475v2'}
42,755
2308.07847v2
2,308.07847
2
Robustness Over Time: Understanding Adversarial Examples' Effectiveness on Longitudinal Versions of Large Language Models
Large Language Models (LLMs) undergo continuous updates to improve user experience. However, prior research on the security and safety implications of LLMs has primarily focused on their specific versions, overlooking the impact of successive LLM updates. This prompts the need for a holistic understanding of the risks in these different versions of LLMs. To fill this gap, in this paper, we conduct a longitudinal study to examine the adversarial robustness -- specifically misclassification, jailbreak, and hallucination -- of three prominent LLMs: GPT-3.5, GPT-4, and LLaMA. Our study reveals that LLM updates do not consistently improve adversarial robustness as expected. For instance, a later version of GPT-3.5 degrades regarding misclassification and hallucination despite its improved resilience against jailbreaks, and GPT-4 demonstrates (incrementally) higher robustness overall. Moreover, larger model sizes do not necessarily yield improved robustness. Specifically, larger LLaMA models do not uniformly exhibit improved robustness across all three aspects studied. Importantly, minor updates lacking substantial robustness improvements can exacerbate existing issues rather than resolve them. By providing a more nuanced understanding of LLM robustness over time, we hope our study can offer valuable insights for developers and users navigating model updates and informed decisions in model development and usage for LLM vendors.
2023-08-15
{'id': '2308.07847v2', 'title': "Robustness Over Time: Understanding Adversarial Examples' Effectiveness on Longitudinal Versions of Large Language Models", 'abstract': 'Large Language Models (LLMs) undergo continuous updates to improve user\nexperience. However, prior research on the security and safety implications of\nLLMs has primarily focused on their specific versions, overlooking the impact\nof successive LLM updates. This prompts the need for a holistic understanding\nof the risks in these different versions of LLMs. To fill this gap, in this\npaper, we conduct a longitudinal study to examine the adversarial robustness --\nspecifically misclassification, jailbreak, and hallucination -- of three\nprominent LLMs: GPT-3.5, GPT-4, and LLaMA. Our study reveals that LLM updates\ndo not consistently improve adversarial robustness as expected. For instance, a\nlater version of GPT-3.5 degrades regarding misclassification and hallucination\ndespite its improved resilience against jailbreaks, and GPT-4 demonstrates\n(incrementally) higher robustness overall. Moreover, larger model sizes do not\nnecessarily yield improved robustness. Specifically, larger LLaMA models do not\nuniformly exhibit improved robustness across all three aspects studied.\nImportantly, minor updates lacking substantial robustness improvements can\nexacerbate existing issues rather than resolve them. By providing a more\nnuanced understanding of LLM robustness over time, we hope our study can offer\nvaluable insights for developers and users navigating model updates and\ninformed decisions in model development and usage for LLM vendors.', 'authors': ['Yugeng Liu', 'Tianshuo Cong', 'Zhengyu Zhao', 'Michael Backes', 'Yun Shen', 'Yang Zhang'], 'published': '2023-08-15 15:51:52+00:00', 'categories': ['cs.CR'], 'url': 'http://arxiv.org/abs/2308.07847v2'}
42,757
2310.05986v1
2,310.05986
1
The Unreasonable Effectiveness of Linear Prediction as a Perceptual Metric
We show how perceptual embeddings of the visual system can be constructed at inference-time with no training data or deep neural network features. Our perceptual embeddings are solutions to a weighted least squares (WLS) problem, defined at the pixel-level, and solved at inference-time, that can capture global and local image characteristics. The distance in embedding space is used to define a perceptual similarity metric which we call LASI: Linear Autoregressive Similarity Index. Experiments on full-reference image quality assessment datasets show LASI performs competitively with learned deep feature based methods like LPIPS (Zhang et al., 2018) and PIM (Bhardwaj et al., 2020), at a similar computational cost to hand-crafted methods such as MS-SSIM (Wang et al., 2003). We found that increasing the dimensionality of the embedding space consistently reduces the WLS loss while increasing performance on perceptual tasks, at the cost of increasing the computational complexity. LASI is fully differentiable, scales cubically with the number of embedding dimensions, and can be parallelized at the pixel-level. A Maximum Differentiation (MAD) competition (Wang & Simoncelli, 2008) between LASI and LPIPS shows that both methods are capable of finding failure points for the other, suggesting these metrics can be combined.
2023-10-06
{'id': '2310.05986v1', 'title': 'The Unreasonable Effectiveness of Linear Prediction as a Perceptual Metric', 'abstract': 'We show how perceptual embeddings of the visual system can be constructed at\ninference-time with no training data or deep neural network features. Our\nperceptual embeddings are solutions to a weighted least squares (WLS) problem,\ndefined at the pixel-level, and solved at inference-time, that can capture\nglobal and local image characteristics. The distance in embedding space is used\nto define a perceptual similarity metric which we call LASI: Linear\nAutoregressive Similarity Index. Experiments on full-reference image quality\nassessment datasets show LASI performs competitively with learned deep feature\nbased methods like LPIPS (Zhang et al., 2018) and PIM (Bhardwaj et al., 2020),\nat a similar computational cost to hand-crafted methods such as MS-SSIM (Wang\net al., 2003). We found that increasing the dimensionality of the embedding\nspace consistently reduces the WLS loss while increasing performance on\nperceptual tasks, at the cost of increasing the computational complexity. LASI\nis fully differentiable, scales cubically with the number of embedding\ndimensions, and can be parallelized at the pixel-level. A Maximum\nDifferentiation (MAD) competition (Wang & Simoncelli, 2008) between LASI and\nLPIPS shows that both methods are capable of finding failure points for the\nother, suggesting these metrics can be combined.', 'authors': ['Daniel Severo', 'Lucas Theis', 'Johannes Ballé'], 'published': '2023-10-06 19:02:00+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2310.05986v1'}
42,759
2312.13866v2
2,312.13866
2
Understanding Inter-Session Intentions via Complex Logical Reasoning
Understanding user intentions is essential for improving product recommendations, navigation suggestions, and query reformulations. However, user intentions can be intricate, involving multiple sessions and attribute requirements connected by logical operators such as And, Or, and Not. For instance, a user may search for Nike or Adidas running shoes across various sessions, with a preference for purple. In another example, a user may have purchased a mattress in a previous session and is now looking for a matching bed frame without intending to buy another mattress. Existing research on session understanding has not adequately addressed making product or attribute recommendations for such complex intentions. In this paper, we present the task of logical session complex query answering (LS-CQA), where sessions are treated as hyperedges of items, and we frame the problem of complex intention understanding as an LS-CQA task on an aggregated hypergraph of sessions, items, and attributes. This is a unique complex query answering task with sessions as ordered hyperedges. We also introduce a new model, the Logical Session Graph Transformer (LSGT), which captures interactions among items across different sessions and their logical connections using a transformer structure. We analyze the expressiveness of LSGT and prove the permutation invariance of the inputs for the logical operators. By evaluating LSGT on three datasets, we demonstrate that it achieves state-of-the-art results.
2023-12-21
{'id': '2312.13866v2', 'title': 'Understanding Inter-Session Intentions via Complex Logical Reasoning', 'abstract': 'Understanding user intentions is essential for improving product\nrecommendations, navigation suggestions, and query reformulations. However,\nuser intentions can be intricate, involving multiple sessions and attribute\nrequirements connected by logical operators such as And, Or, and Not. For\ninstance, a user may search for Nike or Adidas running shoes across various\nsessions, with a preference for purple. In another example, a user may have\npurchased a mattress in a previous session and is now looking for a matching\nbed frame without intending to buy another mattress. Existing research on\nsession understanding has not adequately addressed making product or attribute\nrecommendations for such complex intentions. In this paper, we present the task\nof logical session complex query answering (LS-CQA), where sessions are treated\nas hyperedges of items, and we frame the problem of complex intention\nunderstanding as an LS-CQA task on an aggregated hypergraph of sessions, items,\nand attributes. This is a unique complex query answering task with sessions as\nordered hyperedges. We also introduce a new model, the Logical Session Graph\nTransformer (LSGT), which captures interactions among items across different\nsessions and their logical connections using a transformer structure. We\nanalyze the expressiveness of LSGT and prove the permutation invariance of the\ninputs for the logical operators. By evaluating LSGT on three datasets, we\ndemonstrate that it achieves state-of-the-art results.', 'authors': ['Jiaxin Bai', 'Chen Luo', 'Zheng Li', 'Qingyu Yin', 'Yangqiu Song'], 'published': '2023-12-21 14:03:30+00:00', 'categories': ['cs.AI', 'cs.CL'], 'url': 'http://arxiv.org/abs/2312.13866v2'}
42,761
2309.11713v2
2,309.11713
2
Quasi-Monte Carlo for 3D Sliced Wasserstein
Monte Carlo (MC) integration has been employed as the standard approximation method for the Sliced Wasserstein (SW) distance, whose analytical expression involves an intractable expectation. However, MC integration is not optimal in terms of absolute approximation error. To provide a better class of empirical SW, we propose quasi-sliced Wasserstein (QSW) approximations that rely on Quasi-Monte Carlo (QMC) methods. For a comprehensive investigation of QMC for SW, we focus on the 3D setting, specifically computing the SW between probability measures in three dimensions. In greater detail, we empirically evaluate various methods to construct QMC point sets on the 3D unit-hypersphere, including the Gaussian-based and equal area mappings, generalized spiral points, and optimizing discrepancy energies. Furthermore, to obtain an unbiased estimator for stochastic optimization, we extend QSW to Randomized Quasi-Sliced Wasserstein (RQSW) by introducing randomness in the discussed point sets. Theoretically, we prove the asymptotic convergence of QSW and the unbiasedness of RQSW. Finally, we conduct experiments on various 3D tasks, such as point-cloud comparison, point-cloud interpolation, image style transfer, and training deep point-cloud autoencoders, to demonstrate the favorable performance of the proposed QSW and RQSW variants.
2023-09-21
{'id': '2309.11713v2', 'title': 'Quasi-Monte Carlo for 3D Sliced Wasserstein', 'abstract': 'Monte Carlo (MC) integration has been employed as the standard approximation\nmethod for the Sliced Wasserstein (SW) distance, whose analytical expression\ninvolves an intractable expectation. However, MC integration is not optimal in\nterms of absolute approximation error. To provide a better class of empirical\nSW, we propose quasi-sliced Wasserstein (QSW) approximations that rely on\nQuasi-Monte Carlo (QMC) methods. For a comprehensive investigation of QMC for\nSW, we focus on the 3D setting, specifically computing the SW between\nprobability measures in three dimensions. In greater detail, we empirically\nevaluate various methods to construct QMC point sets on the 3D\nunit-hypersphere, including the Gaussian-based and equal area mappings,\ngeneralized spiral points, and optimizing discrepancy energies. Furthermore, to\nobtain an unbiased estimator for stochastic optimization, we extend QSW to\nRandomized Quasi-Sliced Wasserstein (RQSW) by introducing randomness in the\ndiscussed point sets. Theoretically, we prove the asymptotic convergence of QSW\nand the unbiasedness of RQSW. Finally, we conduct experiments on various 3D\ntasks, such as point-cloud comparison, point-cloud interpolation, image style\ntransfer, and training deep point-cloud autoencoders, to demonstrate the\nfavorable performance of the proposed QSW and RQSW variants.', 'authors': ['Khai Nguyen', 'Nicola Bariletto', 'Nhat Ho'], 'published': '2023-09-21 01:32:42+00:00', 'categories': ['stat.ML', 'cs.GR', 'cs.LG'], 'url': 'http://arxiv.org/abs/2309.11713v2'}
42,762
2305.19306v2
2,305.19306
2
A Graph is Worth 1-bit Spikes: When Graph Contrastive Learning Meets Spiking Neural Networks
While contrastive self-supervised learning has become the de-facto learning paradigm for graph neural networks, the pursuit of higher task accuracy requires a larger hidden dimensionality to learn informative and discriminative full-precision representations, raising concerns about computation, memory footprint, and energy consumption burden (largely overlooked) for real-world applications. This work explores a promising direction for graph contrastive learning (GCL) with spiking neural networks (SNNs), which leverage sparse and binary characteristics to learn more biologically plausible and compact representations. We propose SpikeGCL, a novel GCL framework to learn binarized 1-bit representations for graphs, making balanced trade-offs between efficiency and performance. We provide theoretical guarantees to demonstrate that SpikeGCL has comparable expressiveness with its full-precision counterparts. Experimental results demonstrate that, with nearly 32x representation storage compression, SpikeGCL is either comparable to or outperforms many fancy state-of-the-art supervised and self-supervised methods across several graph benchmarks.
2023-05-30
{'id': '2305.19306v2', 'title': 'A Graph is Worth 1-bit Spikes: When Graph Contrastive Learning Meets Spiking Neural Networks', 'abstract': 'While contrastive self-supervised learning has become the de-facto learning\nparadigm for graph neural networks, the pursuit of higher task accuracy\nrequires a larger hidden dimensionality to learn informative and discriminative\nfull-precision representations, raising concerns about computation, memory\nfootprint, and energy consumption burden (largely overlooked) for real-world\napplications. This work explores a promising direction for graph contrastive\nlearning (GCL) with spiking neural networks (SNNs), which leverage sparse and\nbinary characteristics to learn more biologically plausible and compact\nrepresentations. We propose SpikeGCL, a novel GCL framework to learn binarized\n1-bit representations for graphs, making balanced trade-offs between efficiency\nand performance. We provide theoretical guarantees to demonstrate that SpikeGCL\nhas comparable expressiveness with its full-precision counterparts.\nExperimental results demonstrate that, with nearly 32x representation storage\ncompression, SpikeGCL is either comparable to or outperforms many fancy\nstate-of-the-art supervised and self-supervised methods across several graph\nbenchmarks.', 'authors': ['Jintang Li', 'Huizhe Zhang', 'Ruofan Wu', 'Zulun Zhu', 'Baokun Wang', 'Changhua Meng', 'Zibin Zheng', 'Liang Chen'], 'published': '2023-05-30 16:03:11+00:00', 'categories': ['cs.NE', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2305.19306v2'}
42,763
2401.01869v1
2,401.01869
1
On the hardness of learning under symmetries
We study the problem of learning equivariant neural networks via gradient descent. The incorporation of known symmetries ("equivariance") into neural nets has empirically improved the performance of learning pipelines, in domains ranging from biology to computer vision. However, a rich yet separate line of learning theoretic research has demonstrated that actually learning shallow, fully-connected (i.e. non-symmetric) networks has exponential complexity in the correlational statistical query (CSQ) model, a framework encompassing gradient descent. In this work, we ask: are known problem symmetries sufficient to alleviate the fundamental hardness of learning neural nets with gradient descent? We answer this question in the negative. In particular, we give lower bounds for shallow graph neural networks, convolutional networks, invariant polynomials, and frame-averaged networks for permutation subgroups, which all scale either superpolynomially or exponentially in the relevant input dimension. Therefore, in spite of the significant inductive bias imparted via symmetry, actually learning the complete classes of functions represented by equivariant neural networks via gradient descent remains hard.
2024-01-03
{'id': '2401.01869v1', 'title': 'On the hardness of learning under symmetries', 'abstract': 'We study the problem of learning equivariant neural networks via gradient\ndescent. The incorporation of known symmetries ("equivariance") into neural\nnets has empirically improved the performance of learning pipelines, in domains\nranging from biology to computer vision. However, a rich yet separate line of\nlearning theoretic research has demonstrated that actually learning shallow,\nfully-connected (i.e. non-symmetric) networks has exponential complexity in the\ncorrelational statistical query (CSQ) model, a framework encompassing gradient\ndescent. In this work, we ask: are known problem symmetries sufficient to\nalleviate the fundamental hardness of learning neural nets with gradient\ndescent? We answer this question in the negative. In particular, we give lower\nbounds for shallow graph neural networks, convolutional networks, invariant\npolynomials, and frame-averaged networks for permutation subgroups, which all\nscale either superpolynomially or exponentially in the relevant input\ndimension. Therefore, in spite of the significant inductive bias imparted via\nsymmetry, actually learning the complete classes of functions represented by\nequivariant neural networks via gradient descent remains hard.', 'authors': ['Bobak T. Kiani', 'Thien Le', 'Hannah Lawrence', 'Stefanie Jegelka', 'Melanie Weber'], 'published': '2024-01-03 18:24:18+00:00', 'categories': ['cs.LG', 'cs.DS', 'math.ST', 'stat.ML', 'stat.TH'], 'url': 'http://arxiv.org/abs/2401.01869v1'}
42,764
2402.01900v3
2,402.019
3
Distributional Off-policy Evaluation with Bellman Residual Minimization
We study distributional off-policy evaluation (OPE), of which the goal is to learn the distribution of the return for a target policy using offline data generated by a different policy. The theoretical foundation of many existing work relies on the supremum-extended statistical distances such as supremum-Wasserstein distance, which are hard to estimate. In contrast, we study the more manageable expectation-extended statistical distances and provide a novel theoretical justification on their validity for learning the return distribution. Based on this attractive property, we propose a new method called Energy Bellman Residual Minimizer (EBRM) for distributional OPE. We provide corresponding in-depth theoretical analyses. We establish a finite-sample error bound for the EBRM estimator under the realizability assumption. Furthermore, we introduce a variant of our method based on a multi-step extension which improves the error bound for non-realizable settings. Notably, unlike prior distributional OPE methods, the theoretical guarantees of our method do not require the completeness assumption.
2024-02-02
{'id': '2402.01900v3', 'title': 'Distributional Off-policy Evaluation with Bellman Residual Minimization', 'abstract': 'We study distributional off-policy evaluation (OPE), of which the goal is to\nlearn the distribution of the return for a target policy using offline data\ngenerated by a different policy. The theoretical foundation of many existing\nwork relies on the supremum-extended statistical distances such as\nsupremum-Wasserstein distance, which are hard to estimate. In contrast, we\nstudy the more manageable expectation-extended statistical distances and\nprovide a novel theoretical justification on their validity for learning the\nreturn distribution. Based on this attractive property, we propose a new method\ncalled Energy Bellman Residual Minimizer (EBRM) for distributional OPE. We\nprovide corresponding in-depth theoretical analyses. We establish a\nfinite-sample error bound for the EBRM estimator under the realizability\nassumption. Furthermore, we introduce a variant of our method based on a\nmulti-step extension which improves the error bound for non-realizable\nsettings. Notably, unlike prior distributional OPE methods, the theoretical\nguarantees of our method do not require the completeness assumption.', 'authors': ['Sungee Hong', 'Zhengling Qi', 'Raymond K. W. Wong'], 'published': '2024-02-02 20:59:29+00:00', 'categories': ['stat.ML', 'cs.LG'], 'url': 'http://arxiv.org/abs/2402.01900v3'}
42,765
2402.00728v1
2,402.00728
1
Dropout-Based Rashomon Set Exploration for Efficient Predictive Multiplicity Estimation
Predictive multiplicity refers to the phenomenon in which classification tasks may admit multiple competing models that achieve almost-equally-optimal performance, yet generate conflicting outputs for individual samples. This presents significant concerns, as it can potentially result in systemic exclusion, inexplicable discrimination, and unfairness in practical applications. Measuring and mitigating predictive multiplicity, however, is computationally challenging due to the need to explore all such almost-equally-optimal models, known as the Rashomon set, in potentially huge hypothesis spaces. To address this challenge, we propose a novel framework that utilizes dropout techniques for exploring models in the Rashomon set. We provide rigorous theoretical derivations to connect the dropout parameters to properties of the Rashomon set, and empirically evaluate our framework through extensive experimentation. Numerical results show that our technique consistently outperforms baselines in terms of the effectiveness of predictive multiplicity metric estimation, with runtime speedup up to $205000$. With efficient Rashomon set exploration and metric estimation, mitigation of predictive multiplicity is then achieved through dropout ensemble and model selection.
2024-02-01
{'id': '2402.00728v1', 'title': 'Dropout-Based Rashomon Set Exploration for Efficient Predictive Multiplicity Estimation', 'abstract': 'Predictive multiplicity refers to the phenomenon in which classification\ntasks may admit multiple competing models that achieve almost-equally-optimal\nperformance, yet generate conflicting outputs for individual samples. This\npresents significant concerns, as it can potentially result in systemic\nexclusion, inexplicable discrimination, and unfairness in practical\napplications. Measuring and mitigating predictive multiplicity, however, is\ncomputationally challenging due to the need to explore all such\nalmost-equally-optimal models, known as the Rashomon set, in potentially huge\nhypothesis spaces. To address this challenge, we propose a novel framework that\nutilizes dropout techniques for exploring models in the Rashomon set. We\nprovide rigorous theoretical derivations to connect the dropout parameters to\nproperties of the Rashomon set, and empirically evaluate our framework through\nextensive experimentation. Numerical results show that our technique\nconsistently outperforms baselines in terms of the effectiveness of predictive\nmultiplicity metric estimation, with runtime speedup up to $20\\times \\sim\n5000\\times$. With efficient Rashomon set exploration and metric estimation,\nmitigation of predictive multiplicity is then achieved through dropout ensemble\nand model selection.', 'authors': ['Hsiang Hsu', 'Guihong Li', 'Shaohan Hu', 'Chun-Fu', 'Chen'], 'published': '2024-02-01 16:25:00+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2402.00728v1'}
42,760
2211.15897v3
2,211.15897
3
Learning Antidote Data to Individual Unfairness
Fairness is essential for machine learning systems deployed in high-stake applications. Among all fairness notions, individual fairness, deriving from a consensus that `similar individuals should be treated similarly,' is a vital notion to describe fair treatment for individual cases. Previous studies typically characterize individual fairness as a prediction-invariant problem when perturbing sensitive attributes on samples, and solve it by Distributionally Robust Optimization (DRO) paradigm. However, such adversarial perturbations along a direction covering sensitive information used in DRO do not consider the inherent feature correlations or innate data constraints, therefore could mislead the model to optimize at off-manifold and unrealistic samples. In light of this drawback, in this paper, we propose to learn and generate antidote data that approximately follows the data distribution to remedy individual unfairness. These generated on-manifold antidote data can be used through a generic optimization procedure along with original training data, resulting in a pure pre-processing approach to individual unfairness, or can also fit well with the in-processing DRO paradigm. Through extensive experiments on multiple tabular datasets, we demonstrate our method resists individual unfairness at a minimal or zero cost to predictive utility compared to baselines.
2022-11-29
{'id': '2211.15897v3', 'title': 'Learning Antidote Data to Individual Unfairness', 'abstract': "Fairness is essential for machine learning systems deployed in high-stake\napplications. Among all fairness notions, individual fairness, deriving from a\nconsensus that `similar individuals should be treated similarly,' is a vital\nnotion to describe fair treatment for individual cases. Previous studies\ntypically characterize individual fairness as a prediction-invariant problem\nwhen perturbing sensitive attributes on samples, and solve it by\nDistributionally Robust Optimization (DRO) paradigm. However, such adversarial\nperturbations along a direction covering sensitive information used in DRO do\nnot consider the inherent feature correlations or innate data constraints,\ntherefore could mislead the model to optimize at off-manifold and unrealistic\nsamples. In light of this drawback, in this paper, we propose to learn and\ngenerate antidote data that approximately follows the data distribution to\nremedy individual unfairness. These generated on-manifold antidote data can be\nused through a generic optimization procedure along with original training\ndata, resulting in a pure pre-processing approach to individual unfairness, or\ncan also fit well with the in-processing DRO paradigm. Through extensive\nexperiments on multiple tabular datasets, we demonstrate our method resists\nindividual unfairness at a minimal or zero cost to predictive utility compared\nto baselines.", 'authors': ['Peizhao Li', 'Ethan Xia', 'Hongfu Liu'], 'published': '2022-11-29 03:32:39+00:00', 'categories': ['cs.LG', 'cs.CY'], 'url': 'http://arxiv.org/abs/2211.15897v3'}
42,769
2301.12052v1
2,301.12052
1
Leveraging Importance Weights in Subset Selection
We present a subset selection algorithm designed to work with arbitrary model families in a practical batch setting. In such a setting, an algorithm can sample examples one at a time but, in order to limit overhead costs, is only able to update its state (i.e. further train model weights) once a large enough batch of examples is selected. Our algorithm, IWeS, selects examples by importance sampling where the sampling probability assigned to each example is based on the entropy of models trained on previously selected batches. IWeS admits significant performance improvement compared to other subset selection algorithms for seven publicly available datasets. Additionally, it is competitive in an active learning setting, where the label information is not available at selection time. We also provide an initial theoretical analysis to support our importance weighting approach, proving generalization and sampling rate bounds.
2023-01-28
{'id': '2301.12052v1', 'title': 'Leveraging Importance Weights in Subset Selection', 'abstract': 'We present a subset selection algorithm designed to work with arbitrary model\nfamilies in a practical batch setting. In such a setting, an algorithm can\nsample examples one at a time but, in order to limit overhead costs, is only\nable to update its state (i.e. further train model weights) once a large enough\nbatch of examples is selected. Our algorithm, IWeS, selects examples by\nimportance sampling where the sampling probability assigned to each example is\nbased on the entropy of models trained on previously selected batches. IWeS\nadmits significant performance improvement compared to other subset selection\nalgorithms for seven publicly available datasets. Additionally, it is\ncompetitive in an active learning setting, where the label information is not\navailable at selection time. We also provide an initial theoretical analysis to\nsupport our importance weighting approach, proving generalization and sampling\nrate bounds.', 'authors': ['Gui Citovsky', 'Giulia DeSalvo', 'Sanjiv Kumar', 'Srikumar Ramalingam', 'Afshin Rostamizadeh', 'Yunjuan Wang'], 'published': '2023-01-28 02:07:31+00:00', 'categories': ['cs.LG', 'cs.AI', 'stat.ML'], 'url': 'http://arxiv.org/abs/2301.12052v1'}
42,771
2206.14754v2
2,206.14754
2
Distilling Model Failures as Directions in Latent Space
Existing methods for isolating hard subpopulations and spurious correlations in datasets often require human intervention. This can make these methods labor-intensive and dataset-specific. To address these shortcomings, we present a scalable method for automatically distilling a model's failure modes. Specifically, we harness linear classifiers to identify consistent error patterns, and, in turn, induce a natural representation of these failure modes as directions within the feature space. We demonstrate that this framework allows us to discover and automatically caption challenging subpopulations within the training dataset. Moreover, by combining our framework with off-the-shelf diffusion models, we can generate images that are especially challenging for the analyzed model, and thus can be used to perform synthetic data augmentation that helps remedy the model's failure modes. Code available at https://github.com/MadryLab/failure-directions
2022-06-29
{'id': '2206.14754v2', 'title': 'Distilling Model Failures as Directions in Latent Space', 'abstract': "Existing methods for isolating hard subpopulations and spurious correlations\nin datasets often require human intervention. This can make these methods\nlabor-intensive and dataset-specific. To address these shortcomings, we present\na scalable method for automatically distilling a model's failure modes.\nSpecifically, we harness linear classifiers to identify consistent error\npatterns, and, in turn, induce a natural representation of these failure modes\nas directions within the feature space. We demonstrate that this framework\nallows us to discover and automatically caption challenging subpopulations\nwithin the training dataset. Moreover, by combining our framework with\noff-the-shelf diffusion models, we can generate images that are especially\nchallenging for the analyzed model, and thus can be used to perform synthetic\ndata augmentation that helps remedy the model's failure modes. Code available\nat https://github.com/MadryLab/failure-directions", 'authors': ['Saachi Jain', 'Hannah Lawrence', 'Ankur Moitra', 'Aleksander Madry'], 'published': '2022-06-29 16:35:24+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2206.14754v2'}
42,773
2210.00092v2
2,210.00092
2
Federated Training of Dual Encoding Models on Small Non-IID Client Datasets
Dual encoding models that encode a pair of inputs are widely used for representation learning. Many approaches train dual encoding models by maximizing agreement between pairs of encodings on centralized training data. However, in many scenarios, datasets are inherently decentralized across many clients (user devices or organizations) due to privacy concerns, motivating federated learning. In this work, we focus on federated training of dual encoding models on decentralized data composed of many small, non-IID (independent and identically distributed) client datasets. We show that existing approaches that work well in centralized settings perform poorly when naively adapted to this setting using federated averaging. We observe that, we can simulate large-batch loss computation on individual clients for loss functions that are based on encoding statistics. Based on this insight, we propose a novel federated training approach, Distributed Cross Correlation Optimization (DCCO), which trains dual encoding models using encoding statistics aggregated across clients, without sharing individual data samples. Our experimental results on two datasets demonstrate that the proposed DCCO approach outperforms federated variants of existing approaches by a large margin.
2022-09-30
{'id': '2210.00092v2', 'title': 'Federated Training of Dual Encoding Models on Small Non-IID Client Datasets', 'abstract': 'Dual encoding models that encode a pair of inputs are widely used for\nrepresentation learning. Many approaches train dual encoding models by\nmaximizing agreement between pairs of encodings on centralized training data.\nHowever, in many scenarios, datasets are inherently decentralized across many\nclients (user devices or organizations) due to privacy concerns, motivating\nfederated learning. In this work, we focus on federated training of dual\nencoding models on decentralized data composed of many small, non-IID\n(independent and identically distributed) client datasets. We show that\nexisting approaches that work well in centralized settings perform poorly when\nnaively adapted to this setting using federated averaging. We observe that, we\ncan simulate large-batch loss computation on individual clients for loss\nfunctions that are based on encoding statistics. Based on this insight, we\npropose a novel federated training approach, Distributed Cross Correlation\nOptimization (DCCO), which trains dual encoding models using encoding\nstatistics aggregated across clients, without sharing individual data samples.\nOur experimental results on two datasets demonstrate that the proposed DCCO\napproach outperforms federated variants of existing approaches by a large\nmargin.', 'authors': ['Raviteja Vemulapalli', 'Warren Richard Morningstar', 'Philip Andrew Mansfield', 'Hubert Eichner', 'Karan Singhal', 'Arash Afkanpour', 'Bradley Green'], 'published': '2022-09-30 21:03:43+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2210.00092v2'}
42,775
2103.07601v3
2,103.07601
3
Approximating How Single Head Attention Learns
Why do models often attend to salient words, and how does this evolve throughout training? We approximate model training as a two stage process: early on in training when the attention weights are uniform, the model learns to translate individual input word `i` to `o` if they co-occur frequently. Later, the model learns to attend to `i` while the correct output is $o$ because it knows `i` translates to `o`. To formalize, we define a model property, Knowledge to Translate Individual Words (KTIW) (e.g. knowing that `i` translates to `o`), and claim that it drives the learning of the attention. This claim is supported by the fact that before the attention mechanism is learned, KTIW can be learned from word co-occurrence statistics, but not the other way around. Particularly, we can construct a training distribution that makes KTIW hard to learn, the learning of the attention fails, and the model cannot even learn the simple task of copying the input words to the output. Our approximation explains why models sometimes attend to salient words, and inspires a toy example where a multi-head attention model can overcome the above hard training distribution by improving learning dynamics rather than expressiveness. We end by discussing the limitation of our approximation framework and suggest future directions.
2021-03-13
{'id': '2103.07601v3', 'title': 'Approximating How Single Head Attention Learns', 'abstract': 'Why do models often attend to salient words, and how does this evolve\nthroughout training? We approximate model training as a two stage process:\nearly on in training when the attention weights are uniform, the model learns\nto translate individual input word `i` to `o` if they co-occur frequently.\nLater, the model learns to attend to `i` while the correct output is $o$\nbecause it knows `i` translates to `o`. To formalize, we define a model\nproperty, Knowledge to Translate Individual Words (KTIW) (e.g. knowing that `i`\ntranslates to `o`), and claim that it drives the learning of the attention.\nThis claim is supported by the fact that before the attention mechanism is\nlearned, KTIW can be learned from word co-occurrence statistics, but not the\nother way around. Particularly, we can construct a training distribution that\nmakes KTIW hard to learn, the learning of the attention fails, and the model\ncannot even learn the simple task of copying the input words to the output. Our\napproximation explains why models sometimes attend to salient words, and\ninspires a toy example where a multi-head attention model can overcome the\nabove hard training distribution by improving learning dynamics rather than\nexpressiveness. We end by discussing the limitation of our approximation\nframework and suggest future directions.', 'authors': ['Charlie Snell', 'Ruiqi Zhong', 'Dan Klein', 'Jacob Steinhardt'], 'published': '2021-03-13 02:32:19+00:00', 'categories': ['cs.CL', 'cs.AI'], 'url': 'http://arxiv.org/abs/2103.07601v3'}
42,776
2208.01545v1
2,208.01545
1
The Curse of Low Task Diversity: On the Failure of Transfer Learning to Outperform MAML and Their Empirical Equivalence
Recently, it has been observed that a transfer learning solution might be all we need to solve many few-shot learning benchmarks -- thus raising important questions about when and how meta-learning algorithms should be deployed. In this paper, we seek to clarify these questions by 1. proposing a novel metric -- the diversity coefficient -- to measure the diversity of tasks in a few-shot learning benchmark and 2. by comparing Model-Agnostic Meta-Learning (MAML) and transfer learning under fair conditions (same architecture, same optimizer, and all models trained to convergence). Using the diversity coefficient, we show that the popular MiniImageNet and CIFAR-FS few-shot learning benchmarks have low diversity. This novel insight contextualizes claims that transfer learning solutions are better than meta-learned solutions in the regime of low diversity under a fair comparison. Specifically, we empirically find that a low diversity coefficient correlates with a high similarity between transfer learning and MAML learned solutions in terms of accuracy at meta-test time and classification layer similarity (using feature based distance metrics like SVCCA, PWCCA, CKA, and OPD). To further support our claim, we find this meta-test accuracy holds even as the model size changes. Therefore, we conclude that in the low diversity regime, MAML and transfer learning have equivalent meta-test performance when both are compared fairly. We also hope our work inspires more thoughtful constructions and quantitative evaluations of meta-learning benchmarks in the future.
2022-08-02
{'id': '2208.01545v1', 'title': 'The Curse of Low Task Diversity: On the Failure of Transfer Learning to Outperform MAML and Their Empirical Equivalence', 'abstract': 'Recently, it has been observed that a transfer learning solution might be all\nwe need to solve many few-shot learning benchmarks -- thus raising important\nquestions about when and how meta-learning algorithms should be deployed. In\nthis paper, we seek to clarify these questions by 1. proposing a novel metric\n-- the diversity coefficient -- to measure the diversity of tasks in a few-shot\nlearning benchmark and 2. by comparing Model-Agnostic Meta-Learning (MAML) and\ntransfer learning under fair conditions (same architecture, same optimizer, and\nall models trained to convergence). Using the diversity coefficient, we show\nthat the popular MiniImageNet and CIFAR-FS few-shot learning benchmarks have\nlow diversity. This novel insight contextualizes claims that transfer learning\nsolutions are better than meta-learned solutions in the regime of low diversity\nunder a fair comparison. Specifically, we empirically find that a low diversity\ncoefficient correlates with a high similarity between transfer learning and\nMAML learned solutions in terms of accuracy at meta-test time and\nclassification layer similarity (using feature based distance metrics like\nSVCCA, PWCCA, CKA, and OPD). To further support our claim, we find this\nmeta-test accuracy holds even as the model size changes. Therefore, we conclude\nthat in the low diversity regime, MAML and transfer learning have equivalent\nmeta-test performance when both are compared fairly. We also hope our work\ninspires more thoughtful constructions and quantitative evaluations of\nmeta-learning benchmarks in the future.', 'authors': ['Brando Miranda', 'Patrick Yu', 'Yu-Xiong Wang', 'Sanmi Koyejo'], 'published': '2022-08-02 15:49:11+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2208.01545v1'}
42,766
2311.14864v2
2,311.14864
2
Effective Structural Encodings via Local Curvature Profiles
Structural and Positional Encodings can significantly improve the performance of Graph Neural Networks in downstream tasks. Recent literature has begun to systematically investigate differences in the structural properties that these approaches encode, as well as performance trade-offs between them. However, the question of which structural properties yield the most effective encoding remains open. In this paper, we investigate this question from a geometric perspective. We propose a novel structural encoding based on discrete Ricci curvature (Local Curvature Profiles, short LCP) and show that it significantly outperforms existing encoding approaches. We further show that combining local structural encodings, such as LCP, with global positional encodings improves downstream performance, suggesting that they capture complementary geometric information. Finally, we compare different encoding types with (curvature-based) rewiring techniques. Rewiring has recently received a surge of interest due to its ability to improve the performance of Graph Neural Networks by mitigating over-smoothing and over-squashing effects. Our results suggest that utilizing curvature information for structural encodings delivers significantly larger performance increases than rewiring.
2023-11-24
{'id': '2311.14864v2', 'title': 'Effective Structural Encodings via Local Curvature Profiles', 'abstract': 'Structural and Positional Encodings can significantly improve the performance\nof Graph Neural Networks in downstream tasks. Recent literature has begun to\nsystematically investigate differences in the structural properties that these\napproaches encode, as well as performance trade-offs between them. However, the\nquestion of which structural properties yield the most effective encoding\nremains open. In this paper, we investigate this question from a geometric\nperspective. We propose a novel structural encoding based on discrete Ricci\ncurvature (Local Curvature Profiles, short LCP) and show that it significantly\noutperforms existing encoding approaches. We further show that combining local\nstructural encodings, such as LCP, with global positional encodings improves\ndownstream performance, suggesting that they capture complementary geometric\ninformation. Finally, we compare different encoding types with\n(curvature-based) rewiring techniques. Rewiring has recently received a surge\nof interest due to its ability to improve the performance of Graph Neural\nNetworks by mitigating over-smoothing and over-squashing effects. Our results\nsuggest that utilizing curvature information for structural encodings delivers\nsignificantly larger performance increases than rewiring.', 'authors': ['Lukas Fesser', 'Melanie Weber'], 'published': '2023-11-24 22:42:37+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2311.14864v2'}
42,767
2310.00144v1
2,310.00144
1
Probabilistic Sampling-Enhanced Temporal-Spatial GCN: A Scalable Framework for Transaction Anomaly Detection in Ethereum Networks
The rapid evolution of the Ethereum network necessitates sophisticated techniques to ensure its robustness against potential threats and to maintain transparency. While Graph Neural Networks (GNNs) have pioneered anomaly detection in such platforms, capturing the intricacies of both spatial and temporal transactional patterns has remained a challenge. This study presents a fusion of Graph Convolutional Networks (GCNs) with Temporal Random Walks (TRW) enhanced by probabilistic sampling to bridge this gap. Our approach, unlike traditional GCNs, leverages the strengths of TRW to discern complex temporal sequences in Ethereum transactions, thereby providing a more nuanced transaction anomaly detection mechanism. Preliminary evaluations demonstrate that our TRW-GCN framework substantially advances the performance metrics over conventional GCNs in detecting anomalies and transaction bursts. This research not only underscores the potential of temporal cues in Ethereum transactional data but also offers a scalable and effective methodology for ensuring the security and transparency of decentralized platforms. By harnessing both spatial relationships and time-based transactional sequences as node features, our model introduces an additional layer of granularity, making the detection process more robust and less prone to false positives. This work lays the foundation for future research aimed at optimizing and enhancing the transparency of blockchain technologies, and serves as a testament to the significance of considering both time and space dimensions in the ever-evolving landscape of the decentralized platforms.
2023-09-29
{'id': '2310.00144v1', 'title': 'Probabilistic Sampling-Enhanced Temporal-Spatial GCN: A Scalable Framework for Transaction Anomaly Detection in Ethereum Networks', 'abstract': 'The rapid evolution of the Ethereum network necessitates sophisticated\ntechniques to ensure its robustness against potential threats and to maintain\ntransparency. While Graph Neural Networks (GNNs) have pioneered anomaly\ndetection in such platforms, capturing the intricacies of both spatial and\ntemporal transactional patterns has remained a challenge. This study presents a\nfusion of Graph Convolutional Networks (GCNs) with Temporal Random Walks (TRW)\nenhanced by probabilistic sampling to bridge this gap. Our approach, unlike\ntraditional GCNs, leverages the strengths of TRW to discern complex temporal\nsequences in Ethereum transactions, thereby providing a more nuanced\ntransaction anomaly detection mechanism. Preliminary evaluations demonstrate\nthat our TRW-GCN framework substantially advances the performance metrics over\nconventional GCNs in detecting anomalies and transaction bursts. This research\nnot only underscores the potential of temporal cues in Ethereum transactional\ndata but also offers a scalable and effective methodology for ensuring the\nsecurity and transparency of decentralized platforms. By harnessing both\nspatial relationships and time-based transactional sequences as node features,\nour model introduces an additional layer of granularity, making the detection\nprocess more robust and less prone to false positives. This work lays the\nfoundation for future research aimed at optimizing and enhancing the\ntransparency of blockchain technologies, and serves as a testament to the\nsignificance of considering both time and space dimensions in the ever-evolving\nlandscape of the decentralized platforms.', 'authors': ['Stefan Kambiz Behfar', 'Jon Crowcroft'], 'published': '2023-09-29 21:08:21+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CR'], 'url': 'http://arxiv.org/abs/2310.00144v1'}
42,768
2311.03755v2
2,311.03755
2
Multilingual Mathematical Autoformalization
Autoformalization is the task of translating natural language materials into machine-verifiable formalisations. Progress in autoformalization research is hindered by the lack of a sizeable dataset consisting of informal-formal pairs expressing the same essence. Existing methods tend to circumvent this challenge by manually curating small corpora or using few-shot learning with large language models. But these methods suffer from data scarcity and formal language acquisition difficulty. In this work, we create $MMA$, a large, flexible, multilingual, and multi-domain dataset of informal-formal pairs, by using a language model to translate in the reverse direction, that is, from formal mathematical statements into corresponding informal ones. Experiments show that language models fine-tuned on $MMA$ produce $16-18\%$ of statements acceptable with minimal corrections on the $miniF2F$ and $ProofNet$ benchmarks, up from $0\%$ with the base model. We demonstrate that fine-tuning on multilingual formal data results in more capable autoformalization models even when deployed on monolingual tasks.
2023-11-07
{'id': '2311.03755v2', 'title': 'Multilingual Mathematical Autoformalization', 'abstract': 'Autoformalization is the task of translating natural language materials into\nmachine-verifiable formalisations. Progress in autoformalization research is\nhindered by the lack of a sizeable dataset consisting of informal-formal pairs\nexpressing the same essence. Existing methods tend to circumvent this challenge\nby manually curating small corpora or using few-shot learning with large\nlanguage models. But these methods suffer from data scarcity and formal\nlanguage acquisition difficulty. In this work, we create $\\texttt{MMA}$, a\nlarge, flexible, multilingual, and multi-domain dataset of informal-formal\npairs, by using a language model to translate in the reverse direction, that\nis, from formal mathematical statements into corresponding informal ones.\nExperiments show that language models fine-tuned on $\\texttt{MMA}$ produce\n$16-18\\%$ of statements acceptable with minimal corrections on the\n$\\texttt{miniF2F}$ and $\\texttt{ProofNet}$ benchmarks, up from $0\\%$ with the\nbase model. We demonstrate that fine-tuning on multilingual formal data results\nin more capable autoformalization models even when deployed on monolingual\ntasks.', 'authors': ['Albert Q. Jiang', 'Wenda Li', 'Mateja Jamnik'], 'published': '2023-11-07 06:42:15+00:00', 'categories': ['cs.CL', 'cs.LG'], 'url': 'http://arxiv.org/abs/2311.03755v2'}
42,770
2310.01798v2
2,310.01798
2
Large Language Models Cannot Self-Correct Reasoning Yet
Large Language Models (LLMs) have emerged as a groundbreaking technology with their unparalleled text generation capabilities across various applications. Nevertheless, concerns persist regarding the accuracy and appropriateness of their generated content. A contemporary methodology, self-correction, has been proposed as a remedy to these issues. Building upon this premise, this paper critically examines the role and efficacy of self-correction within LLMs, shedding light on its true potential and limitations. Central to our investigation is the notion of intrinsic self-correction, whereby an LLM attempts to correct its initial responses based solely on its inherent capabilities, without the crutch of external feedback. In the context of reasoning, our research indicates that LLMs struggle to self-correct their responses without external feedback, and at times, their performance even degrades after self-correction. Drawing from these insights, we offer suggestions for future research and practical applications in this field.
2023-10-03
{'id': '2310.01798v2', 'title': 'Large Language Models Cannot Self-Correct Reasoning Yet', 'abstract': 'Large Language Models (LLMs) have emerged as a groundbreaking technology with\ntheir unparalleled text generation capabilities across various applications.\nNevertheless, concerns persist regarding the accuracy and appropriateness of\ntheir generated content. A contemporary methodology, self-correction, has been\nproposed as a remedy to these issues. Building upon this premise, this paper\ncritically examines the role and efficacy of self-correction within LLMs,\nshedding light on its true potential and limitations. Central to our\ninvestigation is the notion of intrinsic self-correction, whereby an LLM\nattempts to correct its initial responses based solely on its inherent\ncapabilities, without the crutch of external feedback. In the context of\nreasoning, our research indicates that LLMs struggle to self-correct their\nresponses without external feedback, and at times, their performance even\ndegrades after self-correction. Drawing from these insights, we offer\nsuggestions for future research and practical applications in this field.', 'authors': ['Jie Huang', 'Xinyun Chen', 'Swaroop Mishra', 'Huaixiu Steven Zheng', 'Adams Wei Yu', 'Xinying Song', 'Denny Zhou'], 'published': '2023-10-03 04:56:12+00:00', 'categories': ['cs.CL', 'cs.AI'], 'url': 'http://arxiv.org/abs/2310.01798v2'}
42,772
2312.00157v2
2,312.00157
2
Universal Backdoor Attacks
Web-scraped datasets are vulnerable to data poisoning, which can be used for backdooring deep image classifiers during training. Since training on large datasets is expensive, a model is trained once and re-used many times. Unlike adversarial examples, backdoor attacks often target specific classes rather than any class learned by the model. One might expect that targeting many classes through a naive composition of attacks vastly increases the number of poison samples. We show this is not necessarily true and more efficient, universal data poisoning attacks exist that allow controlling misclassifications from any source class into any target class with a small increase in poison samples. Our idea is to generate triggers with salient characteristics that the model can learn. The triggers we craft exploit a phenomenon we call inter-class poison transferability, where learning a trigger from one class makes the model more vulnerable to learning triggers for other classes. We demonstrate the effectiveness and robustness of our universal backdoor attacks by controlling models with up to 6,000 classes while poisoning only 0.15% of the training dataset. Our source code is available at https://github.com/Ben-Schneider-code/Universal-Backdoor-Attacks.
2023-11-30
{'id': '2312.00157v2', 'title': 'Universal Backdoor Attacks', 'abstract': 'Web-scraped datasets are vulnerable to data poisoning, which can be used for\nbackdooring deep image classifiers during training. Since training on large\ndatasets is expensive, a model is trained once and re-used many times. Unlike\nadversarial examples, backdoor attacks often target specific classes rather\nthan any class learned by the model. One might expect that targeting many\nclasses through a naive composition of attacks vastly increases the number of\npoison samples. We show this is not necessarily true and more efficient,\nuniversal data poisoning attacks exist that allow controlling\nmisclassifications from any source class into any target class with a small\nincrease in poison samples. Our idea is to generate triggers with salient\ncharacteristics that the model can learn. The triggers we craft exploit a\nphenomenon we call inter-class poison transferability, where learning a trigger\nfrom one class makes the model more vulnerable to learning triggers for other\nclasses. We demonstrate the effectiveness and robustness of our universal\nbackdoor attacks by controlling models with up to 6,000 classes while poisoning\nonly 0.15% of the training dataset. Our source code is available at\nhttps://github.com/Ben-Schneider-code/Universal-Backdoor-Attacks.', 'authors': ['Benjamin Schneider', 'Nils Lukas', 'Florian Kerschbaum'], 'published': '2023-11-30 19:37:47+00:00', 'categories': ['cs.LG', 'cs.CR', 'cs.CV'], 'url': 'http://arxiv.org/abs/2312.00157v2'}
42,774
2310.17230v1
2,310.1723
1
Codebook Features: Sparse and Discrete Interpretability for Neural Networks
Understanding neural networks is challenging in part because of the dense, continuous nature of their hidden states. We explore whether we can train neural networks to have hidden states that are sparse, discrete, and more interpretable by quantizing their continuous features into what we call codebook features. Codebook features are produced by finetuning neural networks with vector quantization bottlenecks at each layer, producing a network whose hidden features are the sum of a small number of discrete vector codes chosen from a larger codebook. Surprisingly, we find that neural networks can operate under this extreme bottleneck with only modest degradation in performance. This sparse, discrete bottleneck also provides an intuitive way of controlling neural network behavior: first, find codes that activate when the desired behavior is present, then activate those same codes during generation to elicit that behavior. We validate our approach by training codebook Transformers on several different datasets. First, we explore a finite state machine dataset with far more hidden states than neurons. In this setting, our approach overcomes the superposition problem by assigning states to distinct codes, and we find that we can make the neural network behave as if it is in a different state by activating the code for that state. Second, we train Transformer language models with up to 410M parameters on two natural language datasets. We identify codes in these models representing diverse, disentangled concepts (ranging from negative emotions to months of the year) and find that we can guide the model to generate different topics by activating the appropriate codes during inference. Overall, codebook features appear to be a promising unit of analysis and control for neural networks and interpretability. Our codebase and models are open-sourced at https://github.com/taufeeque9/codebook-features.
2023-10-26
{'id': '2310.17230v1', 'title': 'Codebook Features: Sparse and Discrete Interpretability for Neural Networks', 'abstract': 'Understanding neural networks is challenging in part because of the dense,\ncontinuous nature of their hidden states. We explore whether we can train\nneural networks to have hidden states that are sparse, discrete, and more\ninterpretable by quantizing their continuous features into what we call\ncodebook features. Codebook features are produced by finetuning neural networks\nwith vector quantization bottlenecks at each layer, producing a network whose\nhidden features are the sum of a small number of discrete vector codes chosen\nfrom a larger codebook. Surprisingly, we find that neural networks can operate\nunder this extreme bottleneck with only modest degradation in performance. This\nsparse, discrete bottleneck also provides an intuitive way of controlling\nneural network behavior: first, find codes that activate when the desired\nbehavior is present, then activate those same codes during generation to elicit\nthat behavior. We validate our approach by training codebook Transformers on\nseveral different datasets. First, we explore a finite state machine dataset\nwith far more hidden states than neurons. In this setting, our approach\novercomes the superposition problem by assigning states to distinct codes, and\nwe find that we can make the neural network behave as if it is in a different\nstate by activating the code for that state. Second, we train Transformer\nlanguage models with up to 410M parameters on two natural language datasets. We\nidentify codes in these models representing diverse, disentangled concepts\n(ranging from negative emotions to months of the year) and find that we can\nguide the model to generate different topics by activating the appropriate\ncodes during inference. Overall, codebook features appear to be a promising\nunit of analysis and control for neural networks and interpretability. Our\ncodebase and models are open-sourced at\nhttps://github.com/taufeeque9/codebook-features.', 'authors': ['Alex Tamkin', 'Mohammad Taufeeque', 'Noah D. Goodman'], 'published': '2023-10-26 08:28:48+00:00', 'categories': ['cs.LG', 'cs.CL'], 'url': 'http://arxiv.org/abs/2310.17230v1'}
42,778
2301.12245v1
2,301.12245
1
Supervision Complexity and its Role in Knowledge Distillation
Despite the popularity and efficacy of knowledge distillation, there is limited understanding of why it helps. In order to study the generalization behavior of a distilled student, we propose a new theoretical framework that leverages supervision complexity: a measure of alignment between teacher-provided supervision and the student's neural tangent kernel. The framework highlights a delicate interplay among the teacher's accuracy, the student's margin with respect to the teacher predictions, and the complexity of the teacher predictions. Specifically, it provides a rigorous justification for the utility of various techniques that are prevalent in the context of distillation, such as early stopping and temperature scaling. Our analysis further suggests the use of online distillation, where a student receives increasingly more complex supervision from teachers in different stages of their training. We demonstrate efficacy of online distillation and validate the theoretical findings on a range of image classification benchmarks and model architectures.
2023-01-28
{'id': '2301.12245v1', 'title': 'Supervision Complexity and its Role in Knowledge Distillation', 'abstract': "Despite the popularity and efficacy of knowledge distillation, there is\nlimited understanding of why it helps. In order to study the generalization\nbehavior of a distilled student, we propose a new theoretical framework that\nleverages supervision complexity: a measure of alignment between\nteacher-provided supervision and the student's neural tangent kernel. The\nframework highlights a delicate interplay among the teacher's accuracy, the\nstudent's margin with respect to the teacher predictions, and the complexity of\nthe teacher predictions. Specifically, it provides a rigorous justification for\nthe utility of various techniques that are prevalent in the context of\ndistillation, such as early stopping and temperature scaling. Our analysis\nfurther suggests the use of online distillation, where a student receives\nincreasingly more complex supervision from teachers in different stages of\ntheir training. We demonstrate efficacy of online distillation and validate the\ntheoretical findings on a range of image classification benchmarks and model\narchitectures.", 'authors': ['Hrayr Harutyunyan', 'Ankit Singh Rawat', 'Aditya Krishna Menon', 'Seungyeon Kim', 'Sanjiv Kumar'], 'published': '2023-01-28 16:34:47+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2301.12245v1'}
42,779
2111.13420v3
2,111.1342
3
Confounder Identification-free Causal Visual Feature Learning
Confounders in deep learning are in general detrimental to model's generalization where they infiltrate feature representations. Therefore, learning causal features that are free of interference from confounders is important. Most previous causal learning based approaches employ back-door criterion to mitigate the adverse effect of certain specific confounder, which require the explicit identification of confounder. However, in real scenarios, confounders are typically diverse and difficult to be identified. In this paper, we propose a novel Confounder Identification-free Causal Visual Feature Learning (CICF) method, which obviates the need for identifying confounders. CICF models the interventions among different samples based on front-door criterion, and then approximates the global-scope intervening effect upon the instance-level interventions from the perspective of optimization. In this way, we aim to find a reliable optimization direction, which avoids the intervening effects of confounders, to learn causal features. Furthermore, we uncover the relation between CICF and the popular meta-learning strategy MAML, and provide an interpretation of why MAML works from the theoretical perspective of causal learning for the first time. Thanks to the effective learning of causal features, our CICF enables models to have superior generalization capability. Extensive experiments on domain generalization benchmark datasets demonstrate the effectiveness of our CICF, which achieves the state-of-the-art performance.
2021-11-26
{'id': '2111.13420v3', 'title': 'Confounder Identification-free Causal Visual Feature Learning', 'abstract': "Confounders in deep learning are in general detrimental to model's\ngeneralization where they infiltrate feature representations. Therefore,\nlearning causal features that are free of interference from confounders is\nimportant. Most previous causal learning based approaches employ back-door\ncriterion to mitigate the adverse effect of certain specific confounder, which\nrequire the explicit identification of confounder. However, in real scenarios,\nconfounders are typically diverse and difficult to be identified. In this\npaper, we propose a novel Confounder Identification-free Causal Visual Feature\nLearning (CICF) method, which obviates the need for identifying confounders.\nCICF models the interventions among different samples based on front-door\ncriterion, and then approximates the global-scope intervening effect upon the\ninstance-level interventions from the perspective of optimization. In this way,\nwe aim to find a reliable optimization direction, which avoids the intervening\neffects of confounders, to learn causal features. Furthermore, we uncover the\nrelation between CICF and the popular meta-learning strategy MAML, and provide\nan interpretation of why MAML works from the theoretical perspective of causal\nlearning for the first time. Thanks to the effective learning of causal\nfeatures, our CICF enables models to have superior generalization capability.\nExtensive experiments on domain generalization benchmark datasets demonstrate\nthe effectiveness of our CICF, which achieves the state-of-the-art performance.", 'authors': ['Xin Li', 'Zhizheng Zhang', 'Guoqiang Wei', 'Cuiling Lan', 'Wenjun Zeng', 'Xin Jin', 'Zhibo Chen'], 'published': '2021-11-26 10:57:47+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2111.13420v3'}
42,780
2209.15055v4
2,209.15055
4
Implicit Bias of Large Depth Networks: a Notion of Rank for Nonlinear Functions
We show that the representation cost of fully connected neural networks with homogeneous nonlinearities - which describes the implicit bias in function space of networks with $L_2$-regularization or with losses such as the cross-entropy - converges as the depth of the network goes to infinity to a notion of rank over nonlinear functions. We then inquire under which conditions the global minima of the loss recover the `true' rank of the data: we show that for too large depths the global minimum will be approximately rank 1 (underestimating the rank); we then argue that there is a range of depths which grows with the number of datapoints where the true rank is recovered. Finally, we discuss the effect of the rank of a classifier on the topology of the resulting class boundaries and show that autoencoders with optimal nonlinear rank are naturally denoising.
2022-09-29
{'id': '2209.15055v4', 'title': 'Implicit Bias of Large Depth Networks: a Notion of Rank for Nonlinear Functions', 'abstract': "We show that the representation cost of fully connected neural networks with\nhomogeneous nonlinearities - which describes the implicit bias in function\nspace of networks with $L_2$-regularization or with losses such as the\ncross-entropy - converges as the depth of the network goes to infinity to a\nnotion of rank over nonlinear functions. We then inquire under which conditions\nthe global minima of the loss recover the `true' rank of the data: we show that\nfor too large depths the global minimum will be approximately rank 1\n(underestimating the rank); we then argue that there is a range of depths which\ngrows with the number of datapoints where the true rank is recovered. Finally,\nwe discuss the effect of the rank of a classifier on the topology of the\nresulting class boundaries and show that autoencoders with optimal nonlinear\nrank are naturally denoising.", 'authors': ['Arthur Jacot'], 'published': '2022-09-29 18:57:51+00:00', 'categories': ['stat.ML', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2209.15055v4'}
42,781
2206.06720v2
2,206.0672
2
Deep Variational Implicit Processes
Implicit processes (IPs) are a generalization of Gaussian processes (GPs). IPs may lack a closed-form expression but are easy to sample from. Examples include, among others, Bayesian neural networks or neural samplers. IPs can be used as priors over functions, resulting in flexible models with well-calibrated prediction uncertainty estimates. Methods based on IPs usually carry out function-space approximate inference, which overcomes some of the difficulties of parameter-space approximate inference. Nevertheless, the approximations employed often limit the expressiveness of the final model, resulting, e.g., in a Gaussian predictive distribution, which can be restrictive. We propose here a multi-layer generalization of IPs called the Deep Variational Implicit process (DVIP). This generalization is similar to that of deep GPs over GPs, but it is more flexible due to the use of IPs as the prior distribution over the latent functions. We describe a scalable variational inference algorithm for training DVIP and show that it outperforms previous IP-based methods and also deep GPs. We support these claims via extensive regression and classification experiments. We also evaluate DVIP on large datasets with up to several million data instances to illustrate its good scalability and performance.
2022-06-14
{'id': '2206.06720v2', 'title': 'Deep Variational Implicit Processes', 'abstract': 'Implicit processes (IPs) are a generalization of Gaussian processes (GPs).\nIPs may lack a closed-form expression but are easy to sample from. Examples\ninclude, among others, Bayesian neural networks or neural samplers. IPs can be\nused as priors over functions, resulting in flexible models with\nwell-calibrated prediction uncertainty estimates. Methods based on IPs usually\ncarry out function-space approximate inference, which overcomes some of the\ndifficulties of parameter-space approximate inference. Nevertheless, the\napproximations employed often limit the expressiveness of the final model,\nresulting, e.g., in a Gaussian predictive distribution, which can be\nrestrictive. We propose here a multi-layer generalization of IPs called the\nDeep Variational Implicit process (DVIP). This generalization is similar to\nthat of deep GPs over GPs, but it is more flexible due to the use of IPs as the\nprior distribution over the latent functions. We describe a scalable\nvariational inference algorithm for training DVIP and show that it outperforms\nprevious IP-based methods and also deep GPs. We support these claims via\nextensive regression and classification experiments. We also evaluate DVIP on\nlarge datasets with up to several million data instances to illustrate its good\nscalability and performance.', 'authors': ['Luis A. Ortega', 'Simón Rodríguez Santana', 'Daniel Hernández-Lobato'], 'published': '2022-06-14 10:04:41+00:00', 'categories': ['stat.ML', 'cs.LG'], 'url': 'http://arxiv.org/abs/2206.06720v2'}
42,782
2212.08320v2
2,212.0832
2
Autoencoders as Cross-Modal Teachers: Can Pretrained 2D Image Transformers Help 3D Representation Learning?
The success of deep learning heavily relies on large-scale data with comprehensive labels, which is more expensive and time-consuming to fetch in 3D compared to 2D images or natural languages. This promotes the potential of utilizing models pretrained with data more than 3D as teachers for cross-modal knowledge transferring. In this paper, we revisit masked modeling in a unified fashion of knowledge distillation, and we show that foundational Transformers pretrained with 2D images or natural languages can help self-supervised 3D representation learning through training Autoencoders as Cross-Modal Teachers (ACT). The pretrained Transformers are transferred as cross-modal 3D teachers using discrete variational autoencoding self-supervision, during which the Transformers are frozen with prompt tuning for better knowledge inheritance. The latent features encoded by the 3D teachers are used as the target of masked point modeling, wherein the dark knowledge is distilled to the 3D Transformer students as foundational geometry understanding. Our ACT pretrained 3D learner achieves state-of-the-art generalization capacity across various downstream benchmarks, e.g., 88.21% overall accuracy on ScanObjectNN. Codes have been released at https://github.com/RunpeiDong/ACT.
2022-12-16
{'id': '2212.08320v2', 'title': 'Autoencoders as Cross-Modal Teachers: Can Pretrained 2D Image Transformers Help 3D Representation Learning?', 'abstract': 'The success of deep learning heavily relies on large-scale data with\ncomprehensive labels, which is more expensive and time-consuming to fetch in 3D\ncompared to 2D images or natural languages. This promotes the potential of\nutilizing models pretrained with data more than 3D as teachers for cross-modal\nknowledge transferring. In this paper, we revisit masked modeling in a unified\nfashion of knowledge distillation, and we show that foundational Transformers\npretrained with 2D images or natural languages can help self-supervised 3D\nrepresentation learning through training Autoencoders as Cross-Modal Teachers\n(ACT). The pretrained Transformers are transferred as cross-modal 3D teachers\nusing discrete variational autoencoding self-supervision, during which the\nTransformers are frozen with prompt tuning for better knowledge inheritance.\nThe latent features encoded by the 3D teachers are used as the target of masked\npoint modeling, wherein the dark knowledge is distilled to the 3D Transformer\nstudents as foundational geometry understanding. Our ACT pretrained 3D learner\nachieves state-of-the-art generalization capacity across various downstream\nbenchmarks, e.g., 88.21% overall accuracy on ScanObjectNN. Codes have been\nreleased at https://github.com/RunpeiDong/ACT.', 'authors': ['Runpei Dong', 'Zekun Qi', 'Linfeng Zhang', 'Junbo Zhang', 'Jianjian Sun', 'Zheng Ge', 'Li Yi', 'Kaisheng Ma'], 'published': '2022-12-16 07:46:53+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2212.08320v2'}
42,783
2203.07720v3
2,203.0772
3
Revitalize Region Feature for Democratizing Video-Language Pre-training of Retrieval
Recent dominant methods for video-language pre-training (VLP) learn transferable representations from the raw pixels in an end-to-end manner to achieve advanced performance on downstream video-language retrieval. Despite the impressive results, VLP research becomes extremely expensive with the need for massive data and a long training time, preventing further explorations. In this work, we revitalize region features of sparsely sampled video clips to significantly reduce both spatial and temporal visual redundancy towards democratizing VLP research at the same time achieving state-of-the-art results. Specifically, to fully explore the potential of region features, we introduce a novel bidirectional region-word alignment regularization that properly optimizes the fine-grained relations between regions and certain words in sentences, eliminating the domain/modality disconnections between pre-extracted region features and text. Extensive results of downstream video-language retrieval tasks on four datasets demonstrate the superiority of our method on both effectiveness and efficiency, e.g., our method achieves competing results with 80\% fewer data and 85\% less pre-training time compared to the most efficient VLP method so far lei2021less. The code will be available at https://github.com/showlab/DemoVLP.
2022-03-15
{'id': '2203.07720v3', 'title': 'Revitalize Region Feature for Democratizing Video-Language Pre-training of Retrieval', 'abstract': 'Recent dominant methods for video-language pre-training (VLP) learn\ntransferable representations from the raw pixels in an end-to-end manner to\nachieve advanced performance on downstream video-language retrieval. Despite\nthe impressive results, VLP research becomes extremely expensive with the need\nfor massive data and a long training time, preventing further explorations. In\nthis work, we revitalize region features of sparsely sampled video clips to\nsignificantly reduce both spatial and temporal visual redundancy towards\ndemocratizing VLP research at the same time achieving state-of-the-art results.\nSpecifically, to fully explore the potential of region features, we introduce a\nnovel bidirectional region-word alignment regularization that properly\noptimizes the fine-grained relations between regions and certain words in\nsentences, eliminating the domain/modality disconnections between pre-extracted\nregion features and text. Extensive results of downstream video-language\nretrieval tasks on four datasets demonstrate the superiority of our method on\nboth effectiveness and efficiency, \\textit{e.g.}, our method achieves competing\nresults with 80\\% fewer data and 85\\% less pre-training time compared to the\nmost efficient VLP method so far \\cite{lei2021less}. The code will be available\nat \\url{https://github.com/showlab/DemoVLP}.', 'authors': ['Guanyu Cai', 'Yixiao Ge', 'Binjie Zhang', 'Alex Jinpeng Wang', 'Rui Yan', 'Xudong Lin', 'Ying Shan', 'Lianghua He', 'Xiaohu Qie', 'Jianping Wu', 'Mike Zheng Shou'], 'published': '2022-03-15 08:18:27+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2203.07720v3'}
42,784
2303.14969v1
2,303.14969
1
Universal Few-shot Learning of Dense Prediction Tasks with Visual Token Matching
Dense prediction tasks are a fundamental class of problems in computer vision. As supervised methods suffer from high pixel-wise labeling cost, a few-shot learning solution that can learn any dense task from a few labeled images is desired. Yet, current few-shot learning methods target a restricted set of tasks such as semantic segmentation, presumably due to challenges in designing a general and unified model that is able to flexibly and efficiently adapt to arbitrary tasks of unseen semantics. We propose Visual Token Matching (VTM), a universal few-shot learner for arbitrary dense prediction tasks. It employs non-parametric matching on patch-level embedded tokens of images and labels that encapsulates all tasks. Also, VTM flexibly adapts to any task with a tiny amount of task-specific parameters that modulate the matching algorithm. We implement VTM as a powerful hierarchical encoder-decoder architecture involving ViT backbones where token matching is performed at multiple feature hierarchies. We experiment VTM on a challenging variant of Taskonomy dataset and observe that it robustly few-shot learns various unseen dense prediction tasks. Surprisingly, it is competitive with fully supervised baselines using only 10 labeled examples of novel tasks (0.004% of full supervision) and sometimes outperforms using 0.1% of full supervision. Codes are available at https://github.com/GitGyun/visual_token_matching.
2023-03-27
{'id': '2303.14969v1', 'title': 'Universal Few-shot Learning of Dense Prediction Tasks with Visual Token Matching', 'abstract': 'Dense prediction tasks are a fundamental class of problems in computer\nvision. As supervised methods suffer from high pixel-wise labeling cost, a\nfew-shot learning solution that can learn any dense task from a few labeled\nimages is desired. Yet, current few-shot learning methods target a restricted\nset of tasks such as semantic segmentation, presumably due to challenges in\ndesigning a general and unified model that is able to flexibly and efficiently\nadapt to arbitrary tasks of unseen semantics. We propose Visual Token Matching\n(VTM), a universal few-shot learner for arbitrary dense prediction tasks. It\nemploys non-parametric matching on patch-level embedded tokens of images and\nlabels that encapsulates all tasks. Also, VTM flexibly adapts to any task with\na tiny amount of task-specific parameters that modulate the matching algorithm.\nWe implement VTM as a powerful hierarchical encoder-decoder architecture\ninvolving ViT backbones where token matching is performed at multiple feature\nhierarchies. We experiment VTM on a challenging variant of Taskonomy dataset\nand observe that it robustly few-shot learns various unseen dense prediction\ntasks. Surprisingly, it is competitive with fully supervised baselines using\nonly 10 labeled examples of novel tasks (0.004% of full supervision) and\nsometimes outperforms using 0.1% of full supervision. Codes are available at\nhttps://github.com/GitGyun/visual_token_matching.', 'authors': ['Donggyun Kim', 'Jinwoo Kim', 'Seongwoong Cho', 'Chong Luo', 'Seunghoon Hong'], 'published': '2023-03-27 07:58:42+00:00', 'categories': ['cs.CV', 'cs.AI'], 'url': 'http://arxiv.org/abs/2303.14969v1'}
42,785
2210.05572v1
2,210.05572
1
Knowledge-Driven New Drug Recommendation
Drug recommendation assists doctors in prescribing personalized medications to patients based on their health conditions. Existing drug recommendation solutions adopt the supervised multi-label classification setup and only work with existing drugs with sufficient prescription data from many patients. However, newly approved drugs do not have much historical prescription data and cannot leverage existing drug recommendation methods. To address this, we formulate the new drug recommendation as a few-shot learning problem. Yet, directly applying existing few-shot learning algorithms faces two challenges: (1) complex relations among diseases and drugs and (2) numerous false-negative patients who were eligible but did not yet use the new drugs. To tackle these challenges, we propose EDGE, which can quickly adapt to the recommendation for a new drug with limited prescription data from a few support patients. EDGE maintains a drug-dependent multi-phenotype few-shot learner to bridge the gap between existing and new drugs. Specifically, EDGE leverages the drug ontology to link new drugs to existing drugs with similar treatment effects and learns ontology-based drug representations. Such drug representations are used to customize the metric space of the phenotype-driven patient representations, which are composed of a set of phenotypes capturing complex patient health status. Lastly, EDGE eliminates the false-negative supervision signal using an external drug-disease knowledge base. We evaluate EDGE on two real-world datasets: the public EHR data (MIMIC-IV) and private industrial claims data. Results show that EDGE achieves 7.3% improvement on the ROC-AUC score over the best baseline.
2022-10-11
{'id': '2210.05572v1', 'title': 'Knowledge-Driven New Drug Recommendation', 'abstract': 'Drug recommendation assists doctors in prescribing personalized medications\nto patients based on their health conditions. Existing drug recommendation\nsolutions adopt the supervised multi-label classification setup and only work\nwith existing drugs with sufficient prescription data from many patients.\nHowever, newly approved drugs do not have much historical prescription data and\ncannot leverage existing drug recommendation methods. To address this, we\nformulate the new drug recommendation as a few-shot learning problem. Yet,\ndirectly applying existing few-shot learning algorithms faces two challenges:\n(1) complex relations among diseases and drugs and (2) numerous false-negative\npatients who were eligible but did not yet use the new drugs. To tackle these\nchallenges, we propose EDGE, which can quickly adapt to the recommendation for\na new drug with limited prescription data from a few support patients. EDGE\nmaintains a drug-dependent multi-phenotype few-shot learner to bridge the gap\nbetween existing and new drugs. Specifically, EDGE leverages the drug ontology\nto link new drugs to existing drugs with similar treatment effects and learns\nontology-based drug representations. Such drug representations are used to\ncustomize the metric space of the phenotype-driven patient representations,\nwhich are composed of a set of phenotypes capturing complex patient health\nstatus. Lastly, EDGE eliminates the false-negative supervision signal using an\nexternal drug-disease knowledge base. We evaluate EDGE on two real-world\ndatasets: the public EHR data (MIMIC-IV) and private industrial claims data.\nResults show that EDGE achieves 7.3% improvement on the ROC-AUC score over the\nbest baseline.', 'authors': ['Zhenbang Wu', 'Huaxiu Yao', 'Zhe Su', 'David M Liebovitz', 'Lucas M Glass', 'James Zou', 'Chelsea Finn', 'Jimeng Sun'], 'published': '2022-10-11 16:07:52+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.IR'], 'url': 'http://arxiv.org/abs/2210.05572v1'}
42,786
2210.02077v1
2,210.02077
1
Exploring The Role of Mean Teachers in Self-supervised Masked Auto-Encoders
Masked image modeling (MIM) has become a popular strategy for self-supervised learning~(SSL) of visual representations with Vision Transformers. A representative MIM model, the masked auto-encoder (MAE), randomly masks a subset of image patches and reconstructs the masked patches given the unmasked patches. Concurrently, many recent works in self-supervised learning utilize the student/teacher paradigm which provides the student with an additional target based on the output of a teacher composed of an exponential moving average (EMA) of previous students. Although common, relatively little is known about the dynamics of the interaction between the student and teacher. Through analysis on a simple linear model, we find that the teacher conditionally removes previous gradient directions based on feature similarities which effectively acts as a conditional momentum regularizer. From this analysis, we present a simple SSL method, the Reconstruction-Consistent Masked Auto-Encoder (RC-MAE) by adding an EMA teacher to MAE. We find that RC-MAE converges faster and requires less memory usage than state-of-the-art self-distillation methods during pre-training, which may provide a way to enhance the practicality of prohibitively expensive self-supervised learning of Vision Transformer models. Additionally, we show that RC-MAE achieves more robustness and better performance compared to MAE on downstream tasks such as ImageNet-1K classification, object detection, and instance segmentation.
2022-10-05
{'id': '2210.02077v1', 'title': 'Exploring The Role of Mean Teachers in Self-supervised Masked Auto-Encoders', 'abstract': 'Masked image modeling (MIM) has become a popular strategy for self-supervised\nlearning~(SSL) of visual representations with Vision Transformers. A\nrepresentative MIM model, the masked auto-encoder (MAE), randomly masks a\nsubset of image patches and reconstructs the masked patches given the unmasked\npatches. Concurrently, many recent works in self-supervised learning utilize\nthe student/teacher paradigm which provides the student with an additional\ntarget based on the output of a teacher composed of an exponential moving\naverage (EMA) of previous students. Although common, relatively little is known\nabout the dynamics of the interaction between the student and teacher. Through\nanalysis on a simple linear model, we find that the teacher conditionally\nremoves previous gradient directions based on feature similarities which\neffectively acts as a conditional momentum regularizer. From this analysis, we\npresent a simple SSL method, the Reconstruction-Consistent Masked Auto-Encoder\n(RC-MAE) by adding an EMA teacher to MAE. We find that RC-MAE converges faster\nand requires less memory usage than state-of-the-art self-distillation methods\nduring pre-training, which may provide a way to enhance the practicality of\nprohibitively expensive self-supervised learning of Vision Transformer models.\nAdditionally, we show that RC-MAE achieves more robustness and better\nperformance compared to MAE on downstream tasks such as ImageNet-1K\nclassification, object detection, and instance segmentation.', 'authors': ['Youngwan Lee', 'Jeffrey Willette', 'Jonghee Kim', 'Juho Lee', 'Sung Ju Hwang'], 'published': '2022-10-05 08:08:55+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2210.02077v1'}
42,787
2209.14996v1
2,209.14996
1
Multiple Modes for Continual Learning
Adapting model parameters to incoming streams of data is a crucial factor to deep learning scalability. Interestingly, prior continual learning strategies in online settings inadvertently anchor their updated parameters to a local parameter subspace to remember old tasks, else drift away from the subspace and forget. From this observation, we formulate a trade-off between constructing multiple parameter modes and allocating tasks per mode. Mode-Optimized Task Allocation (MOTA), our contributed adaptation strategy, trains multiple modes in parallel, then optimizes task allocation per mode. We empirically demonstrate improvements over baseline continual learning strategies and across varying distribution shifts, namely sub-population, domain, and task shift.
2022-09-29
{'id': '2209.14996v1', 'title': 'Multiple Modes for Continual Learning', 'abstract': 'Adapting model parameters to incoming streams of data is a crucial factor to\ndeep learning scalability. Interestingly, prior continual learning strategies\nin online settings inadvertently anchor their updated parameters to a local\nparameter subspace to remember old tasks, else drift away from the subspace and\nforget. From this observation, we formulate a trade-off between constructing\nmultiple parameter modes and allocating tasks per mode. Mode-Optimized Task\nAllocation (MOTA), our contributed adaptation strategy, trains multiple modes\nin parallel, then optimizes task allocation per mode. We empirically\ndemonstrate improvements over baseline continual learning strategies and across\nvarying distribution shifts, namely sub-population, domain, and task shift.', 'authors': ['Siddhartha Datta', 'Nigel Shadbolt'], 'published': '2022-09-29 17:55:32+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2209.14996v1'}
42,788
2209.15092v1
2,209.15092
1
Improving Generative Flow Networks with Path Regularization
Generative Flow Networks (GFlowNets) are recently proposed models for learning stochastic policies that generate compositional objects by sequences of actions with the probability proportional to a given reward function. The central problem of GFlowNets is to improve their exploration and generalization. In this work, we propose a novel path regularization method based on optimal transport theory that places prior constraints on the underlying structure of the GFlowNets. The prior is designed to help the GFlowNets better discover the latent structure of the target distribution or enhance its ability to explore the environment in the context of active learning. The path regularization controls the flow in GFlowNets to generate more diverse and novel candidates via maximizing the optimal transport distances between two forward policies or to improve the generalization via minimizing the optimal transport distances. In addition, we derive an efficient implementation of the regularization by finding its closed form solutions in specific cases and a meaningful upper bound that can be used as an approximation to minimize the regularization term. We empirically demonstrate the advantage of our path regularization on a wide range of tasks, including synthetic hypergrid environment modeling, discrete probabilistic modeling, and biological sequence design.
2022-09-29
{'id': '2209.15092v1', 'title': 'Improving Generative Flow Networks with Path Regularization', 'abstract': 'Generative Flow Networks (GFlowNets) are recently proposed models for\nlearning stochastic policies that generate compositional objects by sequences\nof actions with the probability proportional to a given reward function. The\ncentral problem of GFlowNets is to improve their exploration and\ngeneralization. In this work, we propose a novel path regularization method\nbased on optimal transport theory that places prior constraints on the\nunderlying structure of the GFlowNets. The prior is designed to help the\nGFlowNets better discover the latent structure of the target distribution or\nenhance its ability to explore the environment in the context of active\nlearning. The path regularization controls the flow in GFlowNets to generate\nmore diverse and novel candidates via maximizing the optimal transport\ndistances between two forward policies or to improve the generalization via\nminimizing the optimal transport distances. In addition, we derive an efficient\nimplementation of the regularization by finding its closed form solutions in\nspecific cases and a meaningful upper bound that can be used as an\napproximation to minimize the regularization term. We empirically demonstrate\nthe advantage of our path regularization on a wide range of tasks, including\nsynthetic hypergrid environment modeling, discrete probabilistic modeling, and\nbiological sequence design.', 'authors': ['Anh Do', 'Duy Dinh', 'Tan Nguyen', 'Khuong Nguyen', 'Stanley Osher', 'Nhat Ho'], 'published': '2022-09-29 20:54:41+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2209.15092v1'}
42,789
2209.15266v2
2,209.15266
2
Data Poisoning Attacks Against Multimodal Encoders
Recently, the newly emerged multimodal models, which leverage both visual and linguistic modalities to train powerful encoders, have gained increasing attention. However, learning from a large-scale unlabeled dataset also exposes the model to the risk of potential poisoning attacks, whereby the adversary aims to perturb the model's training data to trigger malicious behaviors in it. In contrast to previous work, only poisoning visual modality, in this work, we take the first step to studying poisoning attacks against multimodal models in both visual and linguistic modalities. Specially, we focus on answering two questions: (1) Is the linguistic modality also vulnerable to poisoning attacks? and (2) Which modality is most vulnerable? To answer the two questions, we propose three types of poisoning attacks against multimodal models. Extensive evaluations on different datasets and model architectures show that all three attacks can achieve significant attack performance while maintaining model utility in both visual and linguistic modalities. Furthermore, we observe that the poisoning effect differs between different modalities. To mitigate the attacks, we propose both pre-training and post-training defenses. We empirically show that both defenses can significantly reduce the attack performance while preserving the model's utility.
2022-09-30
{'id': '2209.15266v2', 'title': 'Data Poisoning Attacks Against Multimodal Encoders', 'abstract': "Recently, the newly emerged multimodal models, which leverage both visual and\nlinguistic modalities to train powerful encoders, have gained increasing\nattention. However, learning from a large-scale unlabeled dataset also exposes\nthe model to the risk of potential poisoning attacks, whereby the adversary\naims to perturb the model's training data to trigger malicious behaviors in it.\nIn contrast to previous work, only poisoning visual modality, in this work, we\ntake the first step to studying poisoning attacks against multimodal models in\nboth visual and linguistic modalities. Specially, we focus on answering two\nquestions: (1) Is the linguistic modality also vulnerable to poisoning attacks?\nand (2) Which modality is most vulnerable? To answer the two questions, we\npropose three types of poisoning attacks against multimodal models. Extensive\nevaluations on different datasets and model architectures show that all three\nattacks can achieve significant attack performance while maintaining model\nutility in both visual and linguistic modalities. Furthermore, we observe that\nthe poisoning effect differs between different modalities. To mitigate the\nattacks, we propose both pre-training and post-training defenses. We\nempirically show that both defenses can significantly reduce the attack\nperformance while preserving the model's utility.", 'authors': ['Ziqing Yang', 'Xinlei He', 'Zheng Li', 'Michael Backes', 'Mathias Humbert', 'Pascal Berrang', 'Yang Zhang'], 'published': '2022-09-30 06:50:08+00:00', 'categories': ['cs.CR', 'cs.LG'], 'url': 'http://arxiv.org/abs/2209.15266v2'}
42,790
2112.13339v2
2,112.13339
2
Quasi-Taylor Samplers for Diffusion Generative Models based on Ideal Derivatives
Diffusion generative models have emerged as a new challenger to popular deep neural generative models such as GANs, but have the drawback that they often require a huge number of neural function evaluations (NFEs) during synthesis unless some sophisticated sampling strategies are employed. This paper proposes new efficient samplers based on the numerical schemes derived by the familiar Taylor expansion, which directly solves the ODE/SDE of interest. In general, it is not easy to compute the derivatives that are required in higher-order Taylor schemes, but in the case of diffusion models, this difficulty is alleviated by the trick that the authors call ``ideal derivative substitution,'' in which the higher-order derivatives are replaced by tractable ones. To derive ideal derivatives, the authors argue the ``single point approximation,'' in which the true score function is approximated by a conditional one, holds in many cases, and considered the derivatives of this approximation. Applying thus obtained new quasi-Taylor samplers to image generation tasks, the authors experimentally confirmed that the proposed samplers could synthesize plausible images in small number of NFEs, and that the performance was better or at the same level as DDIM and Runge-Kutta methods. The paper also argues the relevance of the proposed samplers to the existing ones mentioned above.
2021-12-26
{'id': '2112.13339v2', 'title': 'Quasi-Taylor Samplers for Diffusion Generative Models based on Ideal Derivatives', 'abstract': "Diffusion generative models have emerged as a new challenger to popular deep\nneural generative models such as GANs, but have the drawback that they often\nrequire a huge number of neural function evaluations (NFEs) during synthesis\nunless some sophisticated sampling strategies are employed. This paper proposes\nnew efficient samplers based on the numerical schemes derived by the familiar\nTaylor expansion, which directly solves the ODE/SDE of interest. In general, it\nis not easy to compute the derivatives that are required in higher-order Taylor\nschemes, but in the case of diffusion models, this difficulty is alleviated by\nthe trick that the authors call ``ideal derivative substitution,'' in which the\nhigher-order derivatives are replaced by tractable ones. To derive ideal\nderivatives, the authors argue the ``single point approximation,'' in which the\ntrue score function is approximated by a conditional one, holds in many cases,\nand considered the derivatives of this approximation. Applying thus obtained\nnew quasi-Taylor samplers to image generation tasks, the authors experimentally\nconfirmed that the proposed samplers could synthesize plausible images in small\nnumber of NFEs, and that the performance was better or at the same level as\nDDIM and Runge-Kutta methods. The paper also argues the relevance of the\nproposed samplers to the existing ones mentioned above.", 'authors': ['Hideyuki Tachibana', 'Mocho Go', 'Muneyoshi Inahara', 'Yotaro Katayama', 'Yotaro Watanabe'], 'published': '2021-12-26 09:38:11+00:00', 'categories': ['stat.ML', 'cs.CV', 'cs.LG', 'cs.SD', 'eess.AS'], 'url': 'http://arxiv.org/abs/2112.13339v2'}
42,791
2301.07605v2
2,301.07605
2
Strong inductive biases provably prevent harmless interpolation
Classical wisdom suggests that estimators should avoid fitting noise to achieve good generalization. In contrast, modern overparameterized models can yield small test error despite interpolating noise -- a phenomenon often called "benign overfitting" or "harmless interpolation". This paper argues that the degree to which interpolation is harmless hinges upon the strength of an estimator's inductive bias, i.e., how heavily the estimator favors solutions with a certain structure: while strong inductive biases prevent harmless interpolation, weak inductive biases can even require fitting noise to generalize well. Our main theoretical result establishes tight non-asymptotic bounds for high-dimensional kernel regression that reflect this phenomenon for convolutional kernels, where the filter size regulates the strength of the inductive bias. We further provide empirical evidence of the same behavior for deep neural networks with varying filter sizes and rotational invariance.
2023-01-18
{'id': '2301.07605v2', 'title': 'Strong inductive biases provably prevent harmless interpolation', 'abstract': 'Classical wisdom suggests that estimators should avoid fitting noise to\nachieve good generalization. In contrast, modern overparameterized models can\nyield small test error despite interpolating noise -- a phenomenon often called\n"benign overfitting" or "harmless interpolation". This paper argues that the\ndegree to which interpolation is harmless hinges upon the strength of an\nestimator\'s inductive bias, i.e., how heavily the estimator favors solutions\nwith a certain structure: while strong inductive biases prevent harmless\ninterpolation, weak inductive biases can even require fitting noise to\ngeneralize well. Our main theoretical result establishes tight non-asymptotic\nbounds for high-dimensional kernel regression that reflect this phenomenon for\nconvolutional kernels, where the filter size regulates the strength of the\ninductive bias. We further provide empirical evidence of the same behavior for\ndeep neural networks with varying filter sizes and rotational invariance.', 'authors': ['Michael Aerni', 'Marco Milanta', 'Konstantin Donhauser', 'Fanny Yang'], 'published': '2023-01-18 15:37:11+00:00', 'categories': ['stat.ML', 'cs.LG'], 'url': 'http://arxiv.org/abs/2301.07605v2'}
42,792
2106.02619v2
2,106.02619
2
Forward Super-Resolution: How Can GANs Learn Hierarchical Generative Models for Real-World Distributions
Generative adversarial networks (GANs) are among the most successful models for learning high-complexity, real-world distributions. However, in theory, due to the highly non-convex, non-concave landscape of the minmax training objective, GAN remains one of the least understood deep learning models. In this work, we formally study how GANs can efficiently learn certain hierarchically generated distributions that are close to the distribution of real-life images. We prove that when a distribution has a structure that we refer to as Forward Super-Resolution, then simply training generative adversarial networks using stochastic gradient descent ascent (SGDA) can learn this distribution efficiently, both in sample and time complexities. We also provide empirical evidence that our assumption "forward super-resolution" is very natural in practice, and the underlying learning mechanisms that we study in this paper (to allow us efficiently train GAN via SGDA in theory) simulates the actual learning process of GANs on real-world problems.
2021-06-04
{'id': '2106.02619v2', 'title': 'Forward Super-Resolution: How Can GANs Learn Hierarchical Generative Models for Real-World Distributions', 'abstract': 'Generative adversarial networks (GANs) are among the most successful models\nfor learning high-complexity, real-world distributions. However, in theory, due\nto the highly non-convex, non-concave landscape of the minmax training\nobjective, GAN remains one of the least understood deep learning models. In\nthis work, we formally study how GANs can efficiently learn certain\nhierarchically generated distributions that are close to the distribution of\nreal-life images. We prove that when a distribution has a structure that we\nrefer to as Forward Super-Resolution, then simply training generative\nadversarial networks using stochastic gradient descent ascent (SGDA) can learn\nthis distribution efficiently, both in sample and time complexities. We also\nprovide empirical evidence that our assumption "forward super-resolution" is\nvery natural in practice, and the underlying learning mechanisms that we study\nin this paper (to allow us efficiently train GAN via SGDA in theory) simulates\nthe actual learning process of GANs on real-world problems.', 'authors': ['Zeyuan Allen-Zhu', 'Yuanzhi Li'], 'published': '2021-06-04 17:33:29+00:00', 'categories': ['cs.LG', 'cs.DS', 'cs.NE', 'math.OC', 'stat.ML'], 'url': 'http://arxiv.org/abs/2106.02619v2'}
42,793
2207.10305v2
2,207.10305
2
Detecting Small Query Graphs in A Large Graph via Neural Subgraph Search
Recent advances have shown the success of using reinforcement learning and search to solve NP-hard graph-related tasks, such as Traveling Salesman Optimization, Graph Edit Distance computation, etc. However, it remains unclear how one can efficiently and accurately detect the occurrences of a small query graph in a large target graph, which is a core operation in graph database search, biomedical analysis, social group finding, etc. This task is called Subgraph Matching which essentially performs subgraph isomorphism check between a query graph and a large target graph. One promising approach to this classical problem is the "learning-to-search" paradigm, where a reinforcement learning (RL) agent is designed with a learned policy to guide a search algorithm to quickly find the solution without any solved instances for supervision. However, for the specific task of Subgraph Matching, though the query graph is usually small given by the user as input, the target graph is often orders-of-magnitude larger. It poses challenges to the neural network design and can lead to solution and reward sparsity. In this paper, we propose NSUBS with two innovations to tackle the challenges: (1) A novel encoder-decoder neural network architecture to dynamically compute the matching information between the query and the target graphs at each search state; (2) A novel look-ahead loss function for training the policy network. Experiments on six large real-world target graphs show that NSUBS can significantly improve the subgraph matching performance.
2022-07-21
{'id': '2207.10305v2', 'title': 'Detecting Small Query Graphs in A Large Graph via Neural Subgraph Search', 'abstract': 'Recent advances have shown the success of using reinforcement learning and\nsearch to solve NP-hard graph-related tasks, such as Traveling Salesman\nOptimization, Graph Edit Distance computation, etc. However, it remains unclear\nhow one can efficiently and accurately detect the occurrences of a small query\ngraph in a large target graph, which is a core operation in graph database\nsearch, biomedical analysis, social group finding, etc. This task is called\nSubgraph Matching which essentially performs subgraph isomorphism check between\na query graph and a large target graph. One promising approach to this\nclassical problem is the "learning-to-search" paradigm, where a reinforcement\nlearning (RL) agent is designed with a learned policy to guide a search\nalgorithm to quickly find the solution without any solved instances for\nsupervision. However, for the specific task of Subgraph Matching, though the\nquery graph is usually small given by the user as input, the target graph is\noften orders-of-magnitude larger. It poses challenges to the neural network\ndesign and can lead to solution and reward sparsity. In this paper, we propose\nNSUBS with two innovations to tackle the challenges: (1) A novel\nencoder-decoder neural network architecture to dynamically compute the matching\ninformation between the query and the target graphs at each search state; (2) A\nnovel look-ahead loss function for training the policy network. Experiments on\nsix large real-world target graphs show that NSUBS can significantly improve\nthe subgraph matching performance.', 'authors': ['Yunsheng Bai', 'Derek Xu', 'Yizhou Sun', 'Wei Wang'], 'published': '2022-07-21 04:47:21+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2207.10305v2'}
42,794
2210.02083v2
2,210.02083
2
Multi-View Independent Component Analysis with Shared and Individual Sources
Independent component analysis (ICA) is a blind source separation method for linear disentanglement of independent latent sources from observed data. We investigate the special setting of noisy linear ICA where the observations are split among different views, each receiving a mixture of shared and individual sources. We prove that the corresponding linear structure is identifiable, and the source distribution can be recovered. To computationally estimate the sources, we optimize a constrained form of the joint log-likelihood of the observed data among all views. We also show empirically that our objective recovers the sources also in the case when the measurements are corrupted by noise. Furthermore, we propose a model selection procedure for recovering the number of shared sources which we verify empirically. Finally, we apply the proposed model in a challenging real-life application, where the estimated shared sources from two large transcriptome datasets (observed data) provided by two different labs (two different views) lead to recovering (shared) sources utilized for finding a plausible representation of the underlying graph structure.
2022-10-05
{'id': '2210.02083v2', 'title': 'Multi-View Independent Component Analysis with Shared and Individual Sources', 'abstract': 'Independent component analysis (ICA) is a blind source separation method for\nlinear disentanglement of independent latent sources from observed data. We\ninvestigate the special setting of noisy linear ICA where the observations are\nsplit among different views, each receiving a mixture of shared and individual\nsources. We prove that the corresponding linear structure is identifiable, and\nthe source distribution can be recovered. To computationally estimate the\nsources, we optimize a constrained form of the joint log-likelihood of the\nobserved data among all views. We also show empirically that our objective\nrecovers the sources also in the case when the measurements are corrupted by\nnoise. Furthermore, we propose a model selection procedure for recovering the\nnumber of shared sources which we verify empirically. Finally, we apply the\nproposed model in a challenging real-life application, where the estimated\nshared sources from two large transcriptome datasets (observed data) provided\nby two different labs (two different views) lead to recovering (shared) sources\nutilized for finding a plausible representation of the underlying graph\nstructure.', 'authors': ['Teodora Pandeva', 'Patrick Forré'], 'published': '2022-10-05 08:23:05+00:00', 'categories': ['cs.LG', 'eess.SP'], 'url': 'http://arxiv.org/abs/2210.02083v2'}
42,796
2302.12047v1
2,302.12047
1
Domain Generalisation via Domain Adaptation: An Adversarial Fourier Amplitude Approach
We tackle the domain generalisation (DG) problem by posing it as a domain adaptation (DA) task where we adversarially synthesise the worst-case target domain and adapt a model to that worst-case domain, thereby improving the model's robustness. To synthesise data that is challenging yet semantics-preserving, we generate Fourier amplitude images and combine them with source domain phase images, exploiting the widely believed conjecture from signal processing that amplitude spectra mainly determines image style, while phase data mainly captures image semantics. To synthesise a worst-case domain for adaptation, we train the classifier and the amplitude generator adversarially. Specifically, we exploit the maximum classifier discrepancy (MCD) principle from DA that relates the target domain performance to the discrepancy of classifiers in the model hypothesis space. By Bayesian hypothesis modeling, we express the model hypothesis space effectively as a posterior distribution over classifiers given the source domains, making adversarial MCD minimisation feasible. On the DomainBed benchmark including the large-scale DomainNet dataset, the proposed approach yields significantly improved domain generalisation performance over the state-of-the-art.
2023-02-23
{'id': '2302.12047v1', 'title': 'Domain Generalisation via Domain Adaptation: An Adversarial Fourier Amplitude Approach', 'abstract': "We tackle the domain generalisation (DG) problem by posing it as a domain\nadaptation (DA) task where we adversarially synthesise the worst-case target\ndomain and adapt a model to that worst-case domain, thereby improving the\nmodel's robustness. To synthesise data that is challenging yet\nsemantics-preserving, we generate Fourier amplitude images and combine them\nwith source domain phase images, exploiting the widely believed conjecture from\nsignal processing that amplitude spectra mainly determines image style, while\nphase data mainly captures image semantics. To synthesise a worst-case domain\nfor adaptation, we train the classifier and the amplitude generator\nadversarially. Specifically, we exploit the maximum classifier discrepancy\n(MCD) principle from DA that relates the target domain performance to the\ndiscrepancy of classifiers in the model hypothesis space. By Bayesian\nhypothesis modeling, we express the model hypothesis space effectively as a\nposterior distribution over classifiers given the source domains, making\nadversarial MCD minimisation feasible. On the DomainBed benchmark including the\nlarge-scale DomainNet dataset, the proposed approach yields significantly\nimproved domain generalisation performance over the state-of-the-art.", 'authors': ['Minyoung Kim', 'Da Li', 'Timothy Hospedales'], 'published': '2023-02-23 14:19:07+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2302.12047v1'}
42,797
2209.14932v1
2,209.14932
1
Contrastive Unsupervised Learning of World Model with Invariant Causal Features
In this paper we present a world model, which learns causal features using the invariance principle. In particular, we use contrastive unsupervised learning to learn the invariant causal features, which enforces invariance across augmentations of irrelevant parts or styles of the observation. The world-model-based reinforcement learning methods independently optimize representation learning and the policy. Thus naive contrastive loss implementation collapses due to a lack of supervisory signals to the representation learning module. We propose an intervention invariant auxiliary task to mitigate this issue. Specifically, we utilize depth prediction to explicitly enforce the invariance and use data augmentation as style intervention on the RGB observation space. Our design leverages unsupervised representation learning to learn the world model with invariant causal features. Our proposed method significantly outperforms current state-of-the-art model-based and model-free reinforcement learning methods on out-of-distribution point navigation tasks on the iGibson dataset. Moreover, our proposed model excels at the sim-to-real transfer of our perception learning module. Finally, we evaluate our approach on the DeepMind control suite and enforce invariance only implicitly since depth is not available. Nevertheless, our proposed model performs on par with the state-of-the-art counterpart.
2022-09-29
{'id': '2209.14932v1', 'title': 'Contrastive Unsupervised Learning of World Model with Invariant Causal Features', 'abstract': 'In this paper we present a world model, which learns causal features using\nthe invariance principle. In particular, we use contrastive unsupervised\nlearning to learn the invariant causal features, which enforces invariance\nacross augmentations of irrelevant parts or styles of the observation. The\nworld-model-based reinforcement learning methods independently optimize\nrepresentation learning and the policy. Thus naive contrastive loss\nimplementation collapses due to a lack of supervisory signals to the\nrepresentation learning module. We propose an intervention invariant auxiliary\ntask to mitigate this issue. Specifically, we utilize depth prediction to\nexplicitly enforce the invariance and use data augmentation as style\nintervention on the RGB observation space. Our design leverages unsupervised\nrepresentation learning to learn the world model with invariant causal\nfeatures. Our proposed method significantly outperforms current\nstate-of-the-art model-based and model-free reinforcement learning methods on\nout-of-distribution point navigation tasks on the iGibson dataset. Moreover,\nour proposed model excels at the sim-to-real transfer of our perception\nlearning module. Finally, we evaluate our approach on the DeepMind control\nsuite and enforce invariance only implicitly since depth is not available.\nNevertheless, our proposed model performs on par with the state-of-the-art\ncounterpart.', 'authors': ['Rudra P. K. Poudel', 'Harit Pandya', 'Roberto Cipolla'], 'published': '2022-09-29 16:49:24+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CV', 'cs.RO', 'stat.ML'], 'url': 'http://arxiv.org/abs/2209.14932v1'}
42,798
2203.03539v2
2,203.03539
2
Understanding The Robustness of Self-supervised Learning Through Topic Modeling
Self-supervised learning has significantly improved the performance of many NLP tasks. However, how can self-supervised learning discover useful representations, and why is it better than traditional approaches such as probabilistic models are still largely unknown. In this paper, we focus on the context of topic modeling and highlight a key advantage of self-supervised learning - when applied to data generated by topic models, self-supervised learning can be oblivious to the specific model, and hence is less susceptible to model misspecification. In particular, we prove that commonly used self-supervised objectives based on reconstruction or contrastive samples can both recover useful posterior information for general topic models. Empirically, we show that the same objectives can perform on par with posterior inference using the correct model, while outperforming posterior inference using misspecified models.
2022-02-02
{'id': '2203.03539v2', 'title': 'Understanding The Robustness of Self-supervised Learning Through Topic Modeling', 'abstract': 'Self-supervised learning has significantly improved the performance of many\nNLP tasks. However, how can self-supervised learning discover useful\nrepresentations, and why is it better than traditional approaches such as\nprobabilistic models are still largely unknown. In this paper, we focus on the\ncontext of topic modeling and highlight a key advantage of self-supervised\nlearning - when applied to data generated by topic models, self-supervised\nlearning can be oblivious to the specific model, and hence is less susceptible\nto model misspecification. In particular, we prove that commonly used\nself-supervised objectives based on reconstruction or contrastive samples can\nboth recover useful posterior information for general topic models.\nEmpirically, we show that the same objectives can perform on par with posterior\ninference using the correct model, while outperforming posterior inference\nusing misspecified models.', 'authors': ['Zeping Luo', 'Shiyou Wu', 'Cindy Weng', 'Mo Zhou', 'Rong Ge'], 'published': '2022-02-02 06:20:59+00:00', 'categories': ['cs.CL', 'cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2203.03539v2'}
42,799
2209.15283v1
2,209.15283
1
Sparse tree-based initialization for neural networks
Dedicated neural network (NN) architectures have been designed to handle specific data types (such as CNN for images or RNN for text), which ranks them among state-of-the-art methods for dealing with these data. Unfortunately, no architecture has been found for dealing with tabular data yet, for which tree ensemble methods (tree boosting, random forests) usually show the best predictive performances. In this work, we propose a new sparse initialization technique for (potentially deep) multilayer perceptrons (MLP): we first train a tree-based procedure to detect feature interactions and use the resulting information to initialize the network, which is subsequently trained via standard stochastic gradient strategies. Numerical experiments on several tabular data sets show that this new, simple and easy-to-use method is a solid concurrent, both in terms of generalization capacity and computation time, to default MLP initialization and even to existing complex deep learning solutions. In fact, this wise MLP initialization raises the resulting NN methods to the level of a valid competitor to gradient boosting when dealing with tabular data. Besides, such initializations are able to preserve the sparsity of weights introduced in the first layers of the network through training. This fact suggests that this new initializer operates an implicit regularization during the NN training, and emphasizes that the first layers act as a sparse feature extractor (as for convolutional layers in CNN).
2022-09-30
{'id': '2209.15283v1', 'title': 'Sparse tree-based initialization for neural networks', 'abstract': 'Dedicated neural network (NN) architectures have been designed to handle\nspecific data types (such as CNN for images or RNN for text), which ranks them\namong state-of-the-art methods for dealing with these data. Unfortunately, no\narchitecture has been found for dealing with tabular data yet, for which tree\nensemble methods (tree boosting, random forests) usually show the best\npredictive performances. In this work, we propose a new sparse initialization\ntechnique for (potentially deep) multilayer perceptrons (MLP): we first train a\ntree-based procedure to detect feature interactions and use the resulting\ninformation to initialize the network, which is subsequently trained via\nstandard stochastic gradient strategies. Numerical experiments on several\ntabular data sets show that this new, simple and easy-to-use method is a solid\nconcurrent, both in terms of generalization capacity and computation time, to\ndefault MLP initialization and even to existing complex deep learning\nsolutions. In fact, this wise MLP initialization raises the resulting NN\nmethods to the level of a valid competitor to gradient boosting when dealing\nwith tabular data. Besides, such initializations are able to preserve the\nsparsity of weights introduced in the first layers of the network through\ntraining. This fact suggests that this new initializer operates an implicit\nregularization during the NN training, and emphasizes that the first layers act\nas a sparse feature extractor (as for convolutional layers in CNN).', 'authors': ['Patrick Lutz', 'Ludovic Arnould', 'Claire Boyer', 'Erwan Scornet'], 'published': '2022-09-30 07:44:03+00:00', 'categories': ['stat.ML', 'cs.LG'], 'url': 'http://arxiv.org/abs/2209.15283v1'}
42,800
2203.01629v5
2,203.01629
5
Learning Group Importance using the Differentiable Hypergeometric Distribution
Partitioning a set of elements into subsets of a priori unknown sizes is essential in many applications. These subset sizes are rarely explicitly learned - be it the cluster sizes in clustering applications or the number of shared versus independent generative latent factors in weakly-supervised learning. Probability distributions over correct combinations of subset sizes are non-differentiable due to hard constraints, which prohibit gradient-based optimization. In this work, we propose the differentiable hypergeometric distribution. The hypergeometric distribution models the probability of different group sizes based on their relative importance. We introduce reparameterizable gradients to learn the importance between groups and highlight the advantage of explicitly learning the size of subsets in two typical applications: weakly-supervised learning and clustering. In both applications, we outperform previous approaches, which rely on suboptimal heuristics to model the unknown size of groups.
2022-03-03
{'id': '2203.01629v5', 'title': 'Learning Group Importance using the Differentiable Hypergeometric Distribution', 'abstract': 'Partitioning a set of elements into subsets of a priori unknown sizes is\nessential in many applications. These subset sizes are rarely explicitly\nlearned - be it the cluster sizes in clustering applications or the number of\nshared versus independent generative latent factors in weakly-supervised\nlearning. Probability distributions over correct combinations of subset sizes\nare non-differentiable due to hard constraints, which prohibit gradient-based\noptimization. In this work, we propose the differentiable hypergeometric\ndistribution. The hypergeometric distribution models the probability of\ndifferent group sizes based on their relative importance. We introduce\nreparameterizable gradients to learn the importance between groups and\nhighlight the advantage of explicitly learning the size of subsets in two\ntypical applications: weakly-supervised learning and clustering. In both\napplications, we outperform previous approaches, which rely on suboptimal\nheuristics to model the unknown size of groups.', 'authors': ['Thomas M. Sutter', 'Laura Manduchi', 'Alain Ryser', 'Julia E. Vogt'], 'published': '2022-03-03 10:44:50+00:00', 'categories': ['cs.LG', 'stat.CO', 'stat.ML'], 'url': 'http://arxiv.org/abs/2203.01629v5'}
42,801
2206.07136v3
2,206.07136
3
Automatic Clipping: Differentially Private Deep Learning Made Easier and Stronger
Per-example gradient clipping is a key algorithmic step that enables practical differential private (DP) training for deep learning models. The choice of clipping threshold R, however, is vital for achieving high accuracy under DP. We propose an easy-to-use replacement, called automatic clipping, that eliminates the need to tune R for any DP optimizers, including DP-SGD, DP-Adam, DP-LAMB and many others. The automatic variants are as private and computationally efficient as existing DP optimizers, but require no DP-specific hyperparameters and thus make DP training as amenable as the standard non-private training. We give a rigorous convergence analysis of automatic DP-SGD in the non-convex setting, showing that it can enjoy an asymptotic convergence rate that matches the standard SGD, under a symmetric gradient noise assumption of the per-sample gradients (commonly used in the non-DP literature). We demonstrate on various language and vision tasks that automatic clipping outperforms or matches the state-of-the-art, and can be easily employed with minimal changes to existing codebases.
2022-06-14
{'id': '2206.07136v3', 'title': 'Automatic Clipping: Differentially Private Deep Learning Made Easier and Stronger', 'abstract': 'Per-example gradient clipping is a key algorithmic step that enables\npractical differential private (DP) training for deep learning models. The\nchoice of clipping threshold R, however, is vital for achieving high accuracy\nunder DP. We propose an easy-to-use replacement, called automatic clipping,\nthat eliminates the need to tune R for any DP optimizers, including DP-SGD,\nDP-Adam, DP-LAMB and many others. The automatic variants are as private and\ncomputationally efficient as existing DP optimizers, but require no DP-specific\nhyperparameters and thus make DP training as amenable as the standard\nnon-private training. We give a rigorous convergence analysis of automatic\nDP-SGD in the non-convex setting, showing that it can enjoy an asymptotic\nconvergence rate that matches the standard SGD, under a symmetric gradient\nnoise assumption of the per-sample gradients (commonly used in the non-DP\nliterature). We demonstrate on various language and vision tasks that automatic\nclipping outperforms or matches the state-of-the-art, and can be easily\nemployed with minimal changes to existing codebases.', 'authors': ['Zhiqi Bu', 'Yu-Xiang Wang', 'Sheng Zha', 'George Karypis'], 'published': '2022-06-14 19:49:44+00:00', 'categories': ['cs.LG', 'cs.CL', 'cs.CR', 'cs.CV'], 'url': 'http://arxiv.org/abs/2206.07136v3'}
42,802
2312.11309v2
2,312.11309
2
The Ultimate Combo: Boosting Adversarial Example Transferability by Composing Data Augmentations
To help adversarial examples generalize from surrogate machine-learning (ML) models to targets, certain transferability-based black-box evasion attacks incorporate data augmentations (e.g., random resizing). Yet, prior work has explored limited augmentations and their composition. To fill the gap, we systematically studied how data augmentation affects transferability. Specifically, we explored 46 augmentation techniques originally proposed to help ML models generalize to unseen benign samples, and assessed how they impact transferability, when applied individually or composed. Performing exhaustive search on a small subset of augmentation techniques and genetic search on all techniques, we identified augmentation combinations that help promote transferability. Extensive experiments with the ImageNet and CIFAR-10 datasets and 18 models showed that simple color-space augmentations (e.g., color to greyscale) attain high transferability when combined with standard augmentations. Furthermore, we discovered that composing augmentations impacts transferability mostly monotonically (i.e., more augmentations $$ $$transferability). We also found that the best composition significantly outperformed the state of the art (e.g., 91.8% vs. $$82.5% average transferability to adversarially trained targets on ImageNet). Lastly, our theoretical analysis, backed by empirical evidence, intuitively explains why certain augmentations promote transferability.
2023-12-18
{'id': '2312.11309v2', 'title': 'The Ultimate Combo: Boosting Adversarial Example Transferability by Composing Data Augmentations', 'abstract': 'To help adversarial examples generalize from surrogate machine-learning (ML)\nmodels to targets, certain transferability-based black-box evasion attacks\nincorporate data augmentations (e.g., random resizing). Yet, prior work has\nexplored limited augmentations and their composition. To fill the gap, we\nsystematically studied how data augmentation affects transferability.\nSpecifically, we explored 46 augmentation techniques originally proposed to\nhelp ML models generalize to unseen benign samples, and assessed how they\nimpact transferability, when applied individually or composed. Performing\nexhaustive search on a small subset of augmentation techniques and genetic\nsearch on all techniques, we identified augmentation combinations that help\npromote transferability. Extensive experiments with the ImageNet and CIFAR-10\ndatasets and 18 models showed that simple color-space augmentations (e.g.,\ncolor to greyscale) attain high transferability when combined with standard\naugmentations. Furthermore, we discovered that composing augmentations impacts\ntransferability mostly monotonically (i.e., more augmentations $\\rightarrow$\n$\\ge$transferability). We also found that the best composition significantly\noutperformed the state of the art (e.g., 91.8% vs. $\\le$82.5% average\ntransferability to adversarially trained targets on ImageNet). Lastly, our\ntheoretical analysis, backed by empirical evidence, intuitively explains why\ncertain augmentations promote transferability.', 'authors': ['Zebin Yun', 'Achi-Or Weingarten', 'Eyal Ronen', 'Mahmood Sharif'], 'published': '2023-12-18 16:02:43+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2312.11309v2'}
42,803
2210.05995v2
2,210.05995
2
SGDA with shuffling: faster convergence for nonconvex-PŁ minimax optimization
Stochastic gradient descent-ascent (SGDA) is one of the main workhorses for solving finite-sum minimax optimization problems. Most practical implementations of SGDA randomly reshuffle components and sequentially use them (i.e., without-replacement sampling); however, there are few theoretical results on this approach for minimax algorithms, especially outside the easier-to-analyze (strongly-)monotone setups. To narrow this gap, we study the convergence bounds of SGDA with random reshuffling (SGDA-RR) for smooth nonconvex-nonconcave objectives with Polyak-ojasiewicz (P) geometry. We analyze both simultaneous and alternating SGDA-RR for nonconvex-P and primal-P-P objectives, and obtain convergence rates faster than with-replacement SGDA. Our rates extend to mini-batch SGDA-RR, recovering known rates for full-batch gradient descent-ascent (GDA). Lastly, we present a comprehensive lower bound for GDA with an arbitrary step-size ratio, which matches the full-batch upper bound for the primal-P-P case.
2022-10-12
{'id': '2210.05995v2', 'title': 'SGDA with shuffling: faster convergence for nonconvex-PŁ minimax optimization', 'abstract': 'Stochastic gradient descent-ascent (SGDA) is one of the main workhorses for\nsolving finite-sum minimax optimization problems. Most practical\nimplementations of SGDA randomly reshuffle components and sequentially use them\n(i.e., without-replacement sampling); however, there are few theoretical\nresults on this approach for minimax algorithms, especially outside the\neasier-to-analyze (strongly-)monotone setups. To narrow this gap, we study the\nconvergence bounds of SGDA with random reshuffling (SGDA-RR) for smooth\nnonconvex-nonconcave objectives with Polyak-{\\L}ojasiewicz (P{\\L}) geometry. We\nanalyze both simultaneous and alternating SGDA-RR for nonconvex-P{\\L} and\nprimal-P{\\L}-P{\\L} objectives, and obtain convergence rates faster than\nwith-replacement SGDA. Our rates extend to mini-batch SGDA-RR, recovering known\nrates for full-batch gradient descent-ascent (GDA). Lastly, we present a\ncomprehensive lower bound for GDA with an arbitrary step-size ratio, which\nmatches the full-batch upper bound for the primal-P{\\L}-P{\\L} case.', 'authors': ['Hanseul Cho', 'Chulhee Yun'], 'published': '2022-10-12 08:05:41+00:00', 'categories': ['math.OC', 'stat.ML'], 'url': 'http://arxiv.org/abs/2210.05995v2'}
42,804
2210.16315v3
2,210.16315
3
Beyond calibration: estimating the grouping loss of modern neural networks
The ability to ensure that a classifier gives reliable confidence scores is essential to ensure informed decision-making. To this end, recent work has focused on miscalibration, i.e., the over or under confidence of model scores. Yet calibration is not enough: even a perfectly calibrated classifier with the best possible accuracy can have confidence scores that are far from the true posterior probabilities. This is due to the grouping loss, created by samples with the same confidence scores but different true posterior probabilities. Proper scoring rule theory shows that given the calibration loss, the missing piece to characterize individual errors is the grouping loss. While there are many estimators of the calibration loss, none exists for the grouping loss in standard settings. Here, we propose an estimator to approximate the grouping loss. We show that modern neural network architectures in vision and NLP exhibit grouping loss, notably in distribution shifts settings, which highlights the importance of pre-production validation.
2022-10-28
{'id': '2210.16315v3', 'title': 'Beyond calibration: estimating the grouping loss of modern neural networks', 'abstract': 'The ability to ensure that a classifier gives reliable confidence scores is\nessential to ensure informed decision-making. To this end, recent work has\nfocused on miscalibration, i.e., the over or under confidence of model scores.\nYet calibration is not enough: even a perfectly calibrated classifier with the\nbest possible accuracy can have confidence scores that are far from the true\nposterior probabilities. This is due to the grouping loss, created by samples\nwith the same confidence scores but different true posterior probabilities.\nProper scoring rule theory shows that given the calibration loss, the missing\npiece to characterize individual errors is the grouping loss. While there are\nmany estimators of the calibration loss, none exists for the grouping loss in\nstandard settings. Here, we propose an estimator to approximate the grouping\nloss. We show that modern neural network architectures in vision and NLP\nexhibit grouping loss, notably in distribution shifts settings, which\nhighlights the importance of pre-production validation.', 'authors': ['Alexandre Perez-Lebel', 'Marine Le Morvan', 'Gaël Varoquaux'], 'published': '2022-10-28 07:04:20+00:00', 'categories': ['cs.LG', 'cs.AI', 'stat.ML'], 'url': 'http://arxiv.org/abs/2210.16315v3'}
42,805
2209.14863v2
2,209.14863
2
Neural Networks Efficiently Learn Low-Dimensional Representations with SGD
We study the problem of training a two-layer neural network (NN) of arbitrary width using stochastic gradient descent (SGD) where the input $xR^d$ is Gaussian and the target $y R$ follows a multiple-index model, i.e., $y=g(u_1,x,...,u_k,x)$ with a noisy link function $g$. We prove that the first-layer weights of the NN converge to the $k$-dimensional principal subspace spanned by the vectors $u_1,...,u_k$ of the true model, when online SGD with weight decay is used for training. This phenomenon has several important consequences when $k d$. First, by employing uniform convergence on this smaller subspace, we establish a generalization error bound of $O(kd/T)$ after $T$ iterations of SGD, which is independent of the width of the NN. We further demonstrate that, SGD-trained ReLU NNs can learn a single-index target of the form $y=f(u,x) + $ by recovering the principal direction, with a sample complexity linear in $d$ (up to log factors), where $f$ is a monotonic function with at most polynomial growth, and $$ is the noise. This is in contrast to the known $d^(p)$ sample requirement to learn any degree $p$ polynomial in the kernel regime, and it shows that NNs trained with SGD can outperform the neural tangent kernel at initialization. Finally, we also provide compressibility guarantees for NNs using the approximate low-rank structure produced by SGD.
2022-09-29
{'id': '2209.14863v2', 'title': 'Neural Networks Efficiently Learn Low-Dimensional Representations with SGD', 'abstract': 'We study the problem of training a two-layer neural network (NN) of arbitrary\nwidth using stochastic gradient descent (SGD) where the input\n$\\boldsymbol{x}\\in \\mathbb{R}^d$ is Gaussian and the target $y \\in \\mathbb{R}$\nfollows a multiple-index model, i.e.,\n$y=g(\\langle\\boldsymbol{u_1},\\boldsymbol{x}\\rangle,...,\\langle\\boldsymbol{u_k},\\boldsymbol{x}\\rangle)$\nwith a noisy link function $g$. We prove that the first-layer weights of the NN\nconverge to the $k$-dimensional principal subspace spanned by the vectors\n$\\boldsymbol{u_1},...,\\boldsymbol{u_k}$ of the true model, when online SGD with\nweight decay is used for training. This phenomenon has several important\nconsequences when $k \\ll d$. First, by employing uniform convergence on this\nsmaller subspace, we establish a generalization error bound of\n$O(\\sqrt{{kd}/{T}})$ after $T$ iterations of SGD, which is independent of the\nwidth of the NN. We further demonstrate that, SGD-trained ReLU NNs can learn a\nsingle-index target of the form\n$y=f(\\langle\\boldsymbol{u},\\boldsymbol{x}\\rangle) + \\epsilon$ by recovering the\nprincipal direction, with a sample complexity linear in $d$ (up to log\nfactors), where $f$ is a monotonic function with at most polynomial growth, and\n$\\epsilon$ is the noise. This is in contrast to the known $d^{\\Omega(p)}$\nsample requirement to learn any degree $p$ polynomial in the kernel regime, and\nit shows that NNs trained with SGD can outperform the neural tangent kernel at\ninitialization. Finally, we also provide compressibility guarantees for NNs\nusing the approximate low-rank structure produced by SGD.', 'authors': ['Alireza Mousavi-Hosseini', 'Sejun Park', 'Manuela Girotti', 'Ioannis Mitliagkas', 'Murat A. Erdogdu'], 'published': '2022-09-29 15:29:10+00:00', 'categories': ['stat.ML', 'cs.LG'], 'url': 'http://arxiv.org/abs/2209.14863v2'}
42,806
2212.04089v3
2,212.04089
3
Editing Models with Task Arithmetic
Changing how pre-trained models behave -- e.g., improving their performance on a downstream task or mitigating biases learned during pre-training -- is a common practice when developing machine learning systems. In this work, we propose a new paradigm for steering the behavior of neural networks, centered around task vectors. A task vector specifies a direction in the weight space of a pre-trained model, such that movement in that direction improves performance on the task. We build task vectors by subtracting the weights of a pre-trained model from the weights of the same model after fine-tuning on a task. We show that these task vectors can be modified and combined together through arithmetic operations such as negation and addition, and the behavior of the resulting model is steered accordingly. Negating a task vector decreases performance on the target task, with little change in model behavior on control tasks. Moreover, adding task vectors together can improve performance on multiple tasks at once. Finally, when tasks are linked by an analogy relationship of the form ``A is to B as C is to D", combining task vectors from three of the tasks can improve performance on the fourth, even when no data from the fourth task is used for training. Overall, our experiments with several models, modalities and tasks show that task arithmetic is a simple, efficient and effective way of editing models.
2022-12-08
{'id': '2212.04089v3', 'title': 'Editing Models with Task Arithmetic', 'abstract': 'Changing how pre-trained models behave -- e.g., improving their performance\non a downstream task or mitigating biases learned during pre-training -- is a\ncommon practice when developing machine learning systems. In this work, we\npropose a new paradigm for steering the behavior of neural networks, centered\naround \\textit{task vectors}. A task vector specifies a direction in the weight\nspace of a pre-trained model, such that movement in that direction improves\nperformance on the task. We build task vectors by subtracting the weights of a\npre-trained model from the weights of the same model after fine-tuning on a\ntask. We show that these task vectors can be modified and combined together\nthrough arithmetic operations such as negation and addition, and the behavior\nof the resulting model is steered accordingly. Negating a task vector decreases\nperformance on the target task, with little change in model behavior on control\ntasks. Moreover, adding task vectors together can improve performance on\nmultiple tasks at once. Finally, when tasks are linked by an analogy\nrelationship of the form ``A is to B as C is to D", combining task vectors from\nthree of the tasks can improve performance on the fourth, even when no data\nfrom the fourth task is used for training. Overall, our experiments with\nseveral models, modalities and tasks show that task arithmetic is a simple,\nefficient and effective way of editing models.', 'authors': ['Gabriel Ilharco', 'Marco Tulio Ribeiro', 'Mitchell Wortsman', 'Suchin Gururangan', 'Ludwig Schmidt', 'Hannaneh Hajishirzi', 'Ali Farhadi'], 'published': '2022-12-08 05:50:53+00:00', 'categories': ['cs.LG', 'cs.CL', 'cs.CV'], 'url': 'http://arxiv.org/abs/2212.04089v3'}