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
42,380
2209.14267v2
2,209.14267
2
Less is More: Rethinking Few-Shot Learning and Recurrent Neural Nets
The statistical supervised learning framework assumes an input-output set with a joint probability distribution that is reliably represented by the training dataset. The learner is then required to output a prediction rule learned from the training dataset's input-output pairs. In this work, we provide meaningful insights into the asymptotic equipartition property (AEP) Shannon:1948 in the context of machine learning, and illuminate some of its potential ramifications for few-shot learning. We provide theoretical guarantees for reliable learning under the information-theoretic AEP, and for the generalization error with respect to the sample size. We then focus on a highly efficient recurrent neural net (RNN) framework and propose a reduced-entropy algorithm for few-shot learning. We also propose a mathematical intuition for the RNN as an approximation of a sparse coding solver. We verify the applicability, robustness, and computational efficiency of the proposed approach with image deblurring and optical coherence tomography (OCT) speckle suppression. Our experimental results demonstrate significant potential for improving learning models' sample efficiency, generalization, and time complexity, that can therefore be leveraged for practical real-time applications.
2022-09-28
{'id': '2209.14267v2', 'title': 'Less is More: Rethinking Few-Shot Learning and Recurrent Neural Nets', 'abstract': "The statistical supervised learning framework assumes an input-output set\nwith a joint probability distribution that is reliably represented by the\ntraining dataset. The learner is then required to output a prediction rule\nlearned from the training dataset's input-output pairs. In this work, we\nprovide meaningful insights into the asymptotic equipartition property (AEP)\n\\citep{Shannon:1948} in the context of machine learning, and illuminate some of\nits potential ramifications for few-shot learning. We provide theoretical\nguarantees for reliable learning under the information-theoretic AEP, and for\nthe generalization error with respect to the sample size. We then focus on a\nhighly efficient recurrent neural net (RNN) framework and propose a\nreduced-entropy algorithm for few-shot learning. We also propose a mathematical\nintuition for the RNN as an approximation of a sparse coding solver. We verify\nthe applicability, robustness, and computational efficiency of the proposed\napproach with image deblurring and optical coherence tomography (OCT) speckle\nsuppression. Our experimental results demonstrate significant potential for\nimproving learning models' sample efficiency, generalization, and time\ncomplexity, that can therefore be leveraged for practical real-time\napplications.", 'authors': ['Deborah Pereg', 'Martin Villiger', 'Brett Bouma', 'Polina Golland'], 'published': '2022-09-28 17:33:11+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2209.14267v2'}
42,382
2306.03715v1
2,306.03715
1
Unleashing Mask: Explore the Intrinsic Out-of-Distribution Detection Capability
Out-of-distribution (OOD) detection is an indispensable aspect of secure AI when deploying machine learning models in real-world applications. Previous paradigms either explore better scoring functions or utilize the knowledge of outliers to equip the models with the ability of OOD detection. However, few of them pay attention to the intrinsic OOD detection capability of the given model. In this work, we generally discover the existence of an intermediate stage of a model trained on in-distribution (ID) data having higher OOD detection performance than that of its final stage across different settings, and further identify one critical data-level attribution to be learning with the atypical samples. Based on such insights, we propose a novel method, Unleashing Mask, which aims to restore the OOD discriminative capabilities of the well-trained model with ID data. Our method utilizes a mask to figure out the memorized atypical samples, and then finetune the model or prune it with the introduced mask to forget them. Extensive experiments and analysis demonstrate the effectiveness of our method. The code is available at: https://github.com/tmlr-group/Unleashing-Mask.
2023-06-06
{'id': '2306.03715v1', 'title': 'Unleashing Mask: Explore the Intrinsic Out-of-Distribution Detection Capability', 'abstract': 'Out-of-distribution (OOD) detection is an indispensable aspect of secure AI\nwhen deploying machine learning models in real-world applications. Previous\nparadigms either explore better scoring functions or utilize the knowledge of\noutliers to equip the models with the ability of OOD detection. However, few of\nthem pay attention to the intrinsic OOD detection capability of the given\nmodel. In this work, we generally discover the existence of an intermediate\nstage of a model trained on in-distribution (ID) data having higher OOD\ndetection performance than that of its final stage across different settings,\nand further identify one critical data-level attribution to be learning with\nthe atypical samples. Based on such insights, we propose a novel method,\nUnleashing Mask, which aims to restore the OOD discriminative capabilities of\nthe well-trained model with ID data. Our method utilizes a mask to figure out\nthe memorized atypical samples, and then finetune the model or prune it with\nthe introduced mask to forget them. Extensive experiments and analysis\ndemonstrate the effectiveness of our method. The code is available at:\nhttps://github.com/tmlr-group/Unleashing-Mask.', 'authors': ['Jianing Zhu', 'Hengzhuang Li', 'Jiangchao Yao', 'Tongliang Liu', 'Jianliang Xu', 'Bo Han'], 'published': '2023-06-06 14:23:34+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2306.03715v1'}
42,386
2301.11802v1
2,301.11802
1
Decentralized Online Bandit Optimization on Directed Graphs with Regret Bounds
We consider a decentralized multiplayer game, played over $T$ rounds, with a leader-follower hierarchy described by a directed acyclic graph. For each round, the graph structure dictates the order of the players and how players observe the actions of one another. By the end of each round, all players receive a joint bandit-reward based on their joint action that is used to update the player strategies towards the goal of minimizing the joint pseudo-regret. We present a learning algorithm inspired by the single-player multi-armed bandit problem and show that it achieves sub-linear joint pseudo-regret in the number of rounds for both adversarial and stochastic bandit rewards. Furthermore, we quantify the cost incurred due to the decentralized nature of our problem compared to the centralized setting.
2023-01-27
{'id': '2301.11802v1', 'title': 'Decentralized Online Bandit Optimization on Directed Graphs with Regret Bounds', 'abstract': 'We consider a decentralized multiplayer game, played over $T$ rounds, with a\nleader-follower hierarchy described by a directed acyclic graph. For each\nround, the graph structure dictates the order of the players and how players\nobserve the actions of one another. By the end of each round, all players\nreceive a joint bandit-reward based on their joint action that is used to\nupdate the player strategies towards the goal of minimizing the joint\npseudo-regret. We present a learning algorithm inspired by the single-player\nmulti-armed bandit problem and show that it achieves sub-linear joint\npseudo-regret in the number of rounds for both adversarial and stochastic\nbandit rewards. Furthermore, we quantify the cost incurred due to the\ndecentralized nature of our problem compared to the centralized setting.', 'authors': ['Johan Östman', 'Ather Gattami', 'Daniel Gillblad'], 'published': '2023-01-27 15:58:04+00:00', 'categories': ['cs.LG', 'cs.GT'], 'url': 'http://arxiv.org/abs/2301.11802v1'}
42,389
2210.09461v3
2,210.09461
3
Token Merging: Your ViT But Faster
We introduce Token Merging (ToMe), a simple method to increase the throughput of existing ViT models without needing to train. ToMe gradually combines similar tokens in a transformer using a general and light-weight matching algorithm that is as fast as pruning while being more accurate. Off-the-shelf, ToMe can 2x the throughput of state-of-the-art ViT-L @ 512 and ViT-H @ 518 models on images and 2.2x the throughput of ViT-L on video with only a 0.2-0.3% accuracy drop in each case. ToMe can also easily be applied during training, improving in practice training speed up to 2x for MAE fine-tuning on video. Training with ToMe further minimizes accuracy drop, leading to 2x the throughput of ViT-B on audio for only a 0.4% mAP drop. Qualitatively, we find that ToMe merges object parts into one token, even over multiple frames of video. Overall, ToMe's accuracy and speed are competitive with state-of-the-art on images, video, and audio.
2022-10-17
{'id': '2210.09461v3', 'title': 'Token Merging: Your ViT But Faster', 'abstract': "We introduce Token Merging (ToMe), a simple method to increase the throughput\nof existing ViT models without needing to train. ToMe gradually combines\nsimilar tokens in a transformer using a general and light-weight matching\nalgorithm that is as fast as pruning while being more accurate. Off-the-shelf,\nToMe can 2x the throughput of state-of-the-art ViT-L @ 512 and ViT-H @ 518\nmodels on images and 2.2x the throughput of ViT-L on video with only a 0.2-0.3%\naccuracy drop in each case. ToMe can also easily be applied during training,\nimproving in practice training speed up to 2x for MAE fine-tuning on video.\nTraining with ToMe further minimizes accuracy drop, leading to 2x the\nthroughput of ViT-B on audio for only a 0.4% mAP drop. Qualitatively, we find\nthat ToMe merges object parts into one token, even over multiple frames of\nvideo. Overall, ToMe's accuracy and speed are competitive with state-of-the-art\non images, video, and audio.", 'authors': ['Daniel Bolya', 'Cheng-Yang Fu', 'Xiaoliang Dai', 'Peizhao Zhang', 'Christoph Feichtenhofer', 'Judy Hoffman'], 'published': '2022-10-17 22:23:40+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2210.09461v3'}
42,391
2210.00513v2
2,210.00513
2
Gradient Gating for Deep Multi-Rate Learning on Graphs
We present Gradient Gating (G$^2$), a novel framework for improving the performance of Graph Neural Networks (GNNs). Our framework is based on gating the output of GNN layers with a mechanism for multi-rate flow of message passing information across nodes of the underlying graph. Local gradients are harnessed to further modulate message passing updates. Our framework flexibly allows one to use any basic GNN layer as a wrapper around which the multi-rate gradient gating mechanism is built. We rigorously prove that G$^2$ alleviates the oversmoothing problem and allows the design of deep GNNs. Empirical results are presented to demonstrate that the proposed framework achieves state-of-the-art performance on a variety of graph learning tasks, including on large-scale heterophilic graphs.
2022-10-02
{'id': '2210.00513v2', 'title': 'Gradient Gating for Deep Multi-Rate Learning on Graphs', 'abstract': 'We present Gradient Gating (G$^2$), a novel framework for improving the\nperformance of Graph Neural Networks (GNNs). Our framework is based on gating\nthe output of GNN layers with a mechanism for multi-rate flow of message\npassing information across nodes of the underlying graph. Local gradients are\nharnessed to further modulate message passing updates. Our framework flexibly\nallows one to use any basic GNN layer as a wrapper around which the multi-rate\ngradient gating mechanism is built. We rigorously prove that G$^2$ alleviates\nthe oversmoothing problem and allows the design of deep GNNs. Empirical results\nare presented to demonstrate that the proposed framework achieves\nstate-of-the-art performance on a variety of graph learning tasks, including on\nlarge-scale heterophilic graphs.', 'authors': ['T. Konstantin Rusch', 'Benjamin P. Chamberlain', 'Michael W. Mahoney', 'Michael M. Bronstein', 'Siddhartha Mishra'], 'published': '2022-10-02 13:19:48+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2210.00513v2'}
42,401
2302.11893v1
2,302.11893
1
A framework for benchmarking class-out-of-distribution detection and its application to ImageNet
When deployed for risk-sensitive tasks, deep neural networks must be able to detect instances with labels from outside the distribution for which they were trained. In this paper we present a novel framework to benchmark the ability of image classifiers to detect class-out-of-distribution instances (i.e., instances whose true labels do not appear in the training distribution) at various levels of detection difficulty. We apply this technique to ImageNet, and benchmark 525 pretrained, publicly available, ImageNet-1k classifiers. The code for generating a benchmark for any ImageNet-1k classifier, along with the benchmarks prepared for the above-mentioned 525 models is available at https://github.com/mdabbah/COOD_benchmarking. The usefulness of the proposed framework and its advantage over alternative existing benchmarks is demonstrated by analyzing the results obtained for these models, which reveals numerous novel observations including: (1) knowledge distillation consistently improves class-out-of-distribution (C-OOD) detection performance; (2) a subset of ViTs performs better C-OOD detection than any other model; (3) the language--vision CLIP model achieves good zero-shot detection performance, with its best instance outperforming 96% of all other models evaluated; (4) accuracy and in-distribution ranking are positively correlated to C-OOD detection; and (5) we compare various confidence functions for C-OOD detection. Our companion paper, also published in ICLR 2023 (What Can We Learn From The Selective Prediction And Uncertainty Estimation Performance Of 523 Imagenet Classifiers), examines the uncertainty estimation performance (ranking, calibration, and selective prediction performance) of these classifiers in an in-distribution setting.
2023-02-23
{'id': '2302.11893v1', 'title': 'A framework for benchmarking class-out-of-distribution detection and its application to ImageNet', 'abstract': 'When deployed for risk-sensitive tasks, deep neural networks must be able to\ndetect instances with labels from outside the distribution for which they were\ntrained. In this paper we present a novel framework to benchmark the ability of\nimage classifiers to detect class-out-of-distribution instances (i.e.,\ninstances whose true labels do not appear in the training distribution) at\nvarious levels of detection difficulty. We apply this technique to ImageNet,\nand benchmark 525 pretrained, publicly available, ImageNet-1k classifiers. The\ncode for generating a benchmark for any ImageNet-1k classifier, along with the\nbenchmarks prepared for the above-mentioned 525 models is available at\nhttps://github.com/mdabbah/COOD_benchmarking.\n The usefulness of the proposed framework and its advantage over alternative\nexisting benchmarks is demonstrated by analyzing the results obtained for these\nmodels, which reveals numerous novel observations including: (1) knowledge\ndistillation consistently improves class-out-of-distribution (C-OOD) detection\nperformance; (2) a subset of ViTs performs better C-OOD detection than any\nother model; (3) the language--vision CLIP model achieves good zero-shot\ndetection performance, with its best instance outperforming 96% of all other\nmodels evaluated; (4) accuracy and in-distribution ranking are positively\ncorrelated to C-OOD detection; and (5) we compare various confidence functions\nfor C-OOD detection. Our companion paper, also published in ICLR 2023 (What Can\nWe Learn From The Selective Prediction And Uncertainty Estimation Performance\nOf 523 Imagenet Classifiers), examines the uncertainty estimation performance\n(ranking, calibration, and selective prediction performance) of these\nclassifiers in an in-distribution setting.', 'authors': ['Ido Galil', 'Mohammed Dabbah', 'Ran El-Yaniv'], 'published': '2023-02-23 09:57:48+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2302.11893v1'}
42,381
2009.09139v3
2,009.09139
3
Conditionally Adaptive Multi-Task Learning: Improving Transfer Learning in NLP Using Fewer Parameters & Less Data
Multi-Task Learning (MTL) networks have emerged as a promising method for transferring learned knowledge across different tasks. However, MTL must deal with challenges such as: overfitting to low resource tasks, catastrophic forgetting, and negative task transfer, or learning interference. Often, in Natural Language Processing (NLP), a separate model per task is needed to obtain the best performance. However, many fine-tuning approaches are both parameter inefficient, i.e., potentially involving one new model per task, and highly susceptible to losing knowledge acquired during pretraining. We propose a novel Transformer architecture consisting of a new conditional attention mechanism as well as a set of task-conditioned modules that facilitate weight sharing. Through this construction (a hypernetwork adapter), we achieve more efficient parameter sharing and mitigate forgetting by keeping half of the weights of a pretrained model fixed. We also use a new multi-task data sampling strategy to mitigate the negative effects of data imbalance across tasks. Using this approach, we are able to surpass single task fine-tuning methods while being parameter and data efficient (using around 66% of the data for weight updates). Compared to other BERT Large methods on GLUE, our 8-task model surpasses other Adapter methods by 2.8% and our 24-task model outperforms by 0.7-1.0% models that use MTL and single task fine-tuning. We show that a larger variant of our single multi-task model approach performs competitively across 26 NLP tasks and yields state-of-the-art results on a number of test and development sets. Our code is publicly available at https://github.com/CAMTL/CA-MTL.
2020-09-19
{'id': '2009.09139v3', 'title': 'Conditionally Adaptive Multi-Task Learning: Improving Transfer Learning in NLP Using Fewer Parameters & Less Data', 'abstract': 'Multi-Task Learning (MTL) networks have emerged as a promising method for\ntransferring learned knowledge across different tasks. However, MTL must deal\nwith challenges such as: overfitting to low resource tasks, catastrophic\nforgetting, and negative task transfer, or learning interference. Often, in\nNatural Language Processing (NLP), a separate model per task is needed to\nobtain the best performance. However, many fine-tuning approaches are both\nparameter inefficient, i.e., potentially involving one new model per task, and\nhighly susceptible to losing knowledge acquired during pretraining. We propose\na novel Transformer architecture consisting of a new conditional attention\nmechanism as well as a set of task-conditioned modules that facilitate weight\nsharing. Through this construction (a hypernetwork adapter), we achieve more\nefficient parameter sharing and mitigate forgetting by keeping half of the\nweights of a pretrained model fixed. We also use a new multi-task data sampling\nstrategy to mitigate the negative effects of data imbalance across tasks. Using\nthis approach, we are able to surpass single task fine-tuning methods while\nbeing parameter and data efficient (using around 66% of the data for weight\nupdates). Compared to other BERT Large methods on GLUE, our 8-task model\nsurpasses other Adapter methods by 2.8% and our 24-task model outperforms by\n0.7-1.0% models that use MTL and single task fine-tuning. We show that a larger\nvariant of our single multi-task model approach performs competitively across\n26 NLP tasks and yields state-of-the-art results on a number of test and\ndevelopment sets. Our code is publicly available at\nhttps://github.com/CAMTL/CA-MTL.', 'authors': ['Jonathan Pilault', 'Amine Elhattami', 'Christopher Pal'], 'published': '2020-09-19 02:04:34+00:00', 'categories': ['cs.LG', 'stat.ML', 'I.2.7'], 'url': 'http://arxiv.org/abs/2009.09139v3'}
42,385
2103.02200v2
2,103.022
2
Formalizing Generalization and Robustness of Neural Networks to Weight Perturbations
Studying the sensitivity of weight perturbation in neural networks and its impacts on model performance, including generalization and robustness, is an active research topic due to its implications on a wide range of machine learning tasks such as model compression, generalization gap assessment, and adversarial attacks. In this paper, we provide the first integral study and analysis for feed-forward neural networks in terms of the robustness in pairwise class margin and its generalization behavior under weight perturbation. We further design a new theory-driven loss function for training generalizable and robust neural networks against weight perturbations. Empirical experiments are conducted to validate our theoretical analysis. Our results offer fundamental insights for characterizing the generalization and robustness of neural networks against weight perturbations.
2021-03-03
{'id': '2103.02200v2', 'title': 'Formalizing Generalization and Robustness of Neural Networks to Weight Perturbations', 'abstract': 'Studying the sensitivity of weight perturbation in neural networks and its\nimpacts on model performance, including generalization and robustness, is an\nactive research topic due to its implications on a wide range of machine\nlearning tasks such as model compression, generalization gap assessment, and\nadversarial attacks. In this paper, we provide the first integral study and\nanalysis for feed-forward neural networks in terms of the robustness in\npairwise class margin and its generalization behavior under weight\nperturbation. We further design a new theory-driven loss function for training\ngeneralizable and robust neural networks against weight perturbations.\nEmpirical experiments are conducted to validate our theoretical analysis. Our\nresults offer fundamental insights for characterizing the generalization and\nrobustness of neural networks against weight perturbations.', 'authors': ['Yu-Lin Tsai', 'Chia-Yi Hsu', 'Chia-Mu Yu', 'Pin-Yu Chen'], 'published': '2021-03-03 06:17:03+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2103.02200v2'}
42,390
1910.00199v3
1,910.00199
3
Saliency is a Possible Red Herring When Diagnosing Poor Generalization
Poor generalization is one symptom of models that learn to predict target variables using spuriously-correlated image features present only in the training distribution instead of the true image features that denote a class. It is often thought that this can be diagnosed visually using attribution (aka saliency) maps. We study if this assumption is correct. In some prediction tasks, such as for medical images, one may have some images with masks drawn by a human expert, indicating a region of the image containing relevant information to make the prediction. We study multiple methods that take advantage of such auxiliary labels, by training networks to ignore distracting features which may be found outside of the region of interest. This mask information is only used during training and has an impact on generalization accuracy depending on the severity of the shift between the training and test distributions. Surprisingly, while these methods improve generalization performance in the presence of a covariate shift, there is no strong correspondence between the correction of attribution towards the features a human expert has labelled as important and generalization performance. These results suggest that the root cause of poor generalization may not always be spatially defined, and raise questions about the utility of masks as "attribution priors" as well as saliency maps for explainable predictions.
2019-10-01
{'id': '1910.00199v3', 'title': 'Saliency is a Possible Red Herring When Diagnosing Poor Generalization', 'abstract': 'Poor generalization is one symptom of models that learn to predict target\nvariables using spuriously-correlated image features present only in the\ntraining distribution instead of the true image features that denote a class.\nIt is often thought that this can be diagnosed visually using attribution (aka\nsaliency) maps. We study if this assumption is correct. In some prediction\ntasks, such as for medical images, one may have some images with masks drawn by\na human expert, indicating a region of the image containing relevant\ninformation to make the prediction. We study multiple methods that take\nadvantage of such auxiliary labels, by training networks to ignore distracting\nfeatures which may be found outside of the region of interest. This mask\ninformation is only used during training and has an impact on generalization\naccuracy depending on the severity of the shift between the training and test\ndistributions. Surprisingly, while these methods improve generalization\nperformance in the presence of a covariate shift, there is no strong\ncorrespondence between the correction of attribution towards the features a\nhuman expert has labelled as important and generalization performance. These\nresults suggest that the root cause of poor generalization may not always be\nspatially defined, and raise questions about the utility of masks as\n"attribution priors" as well as saliency maps for explainable predictions.', 'authors': ['Joseph D. Viviano', 'Becks Simpson', 'Francis Dutil', 'Yoshua Bengio', 'Joseph Paul Cohen'], 'published': '2019-10-01 04:29:18+00:00', 'categories': ['cs.CV', 'cs.LG', 'eess.IV'], 'url': 'http://arxiv.org/abs/1910.00199v3'}
42,424
2009.14794v4
2,009.14794
4
Rethinking Attention with Performers
We introduce Performers, Transformer architectures which can estimate regular (softmax) full-rank-attention Transformers with provable accuracy, but using only linear (as opposed to quadratic) space and time complexity, without relying on any priors such as sparsity or low-rankness. To approximate softmax attention-kernels, Performers use a novel Fast Attention Via positive Orthogonal Random features approach (FAVOR+), which may be of independent interest for scalable kernel methods. FAVOR+ can be also used to efficiently model kernelizable attention mechanisms beyond softmax. This representational power is crucial to accurately compare softmax with other kernels for the first time on large-scale tasks, beyond the reach of regular Transformers, and investigate optimal attention-kernels. Performers are linear architectures fully compatible with regular Transformers and with strong theoretical guarantees: unbiased or nearly-unbiased estimation of the attention matrix, uniform convergence and low estimation variance. We tested Performers on a rich set of tasks stretching from pixel-prediction through text models to protein sequence modeling. We demonstrate competitive results with other examined efficient sparse and dense attention methods, showcasing effectiveness of the novel attention-learning paradigm leveraged by Performers.
2020-09-30
{'id': '2009.14794v4', 'title': 'Rethinking Attention with Performers', 'abstract': 'We introduce Performers, Transformer architectures which can estimate regular\n(softmax) full-rank-attention Transformers with provable accuracy, but using\nonly linear (as opposed to quadratic) space and time complexity, without\nrelying on any priors such as sparsity or low-rankness. To approximate softmax\nattention-kernels, Performers use a novel Fast Attention Via positive\nOrthogonal Random features approach (FAVOR+), which may be of independent\ninterest for scalable kernel methods. FAVOR+ can be also used to efficiently\nmodel kernelizable attention mechanisms beyond softmax. This representational\npower is crucial to accurately compare softmax with other kernels for the first\ntime on large-scale tasks, beyond the reach of regular Transformers, and\ninvestigate optimal attention-kernels. Performers are linear architectures\nfully compatible with regular Transformers and with strong theoretical\nguarantees: unbiased or nearly-unbiased estimation of the attention matrix,\nuniform convergence and low estimation variance. We tested Performers on a rich\nset of tasks stretching from pixel-prediction through text models to protein\nsequence modeling. We demonstrate competitive results with other examined\nefficient sparse and dense attention methods, showcasing effectiveness of the\nnovel attention-learning paradigm leveraged by Performers.', 'authors': ['Krzysztof Choromanski', 'Valerii Likhosherstov', 'David Dohan', 'Xingyou Song', 'Andreea Gane', 'Tamas Sarlos', 'Peter Hawkins', 'Jared Davis', 'Afroz Mohiuddin', 'Lukasz Kaiser', 'David Belanger', 'Lucy Colwell', 'Adrian Weller'], 'published': '2020-09-30 17:09:09+00:00', 'categories': ['cs.LG', 'cs.CL', 'stat.ML'], 'url': 'http://arxiv.org/abs/2009.14794v4'}
42,434
2106.06981v2
2,106.06981
2
Thinking Like Transformers
What is the computational model behind a Transformer? Where recurrent neural networks have direct parallels in finite state machines, allowing clear discussion and thought around architecture variants or trained models, Transformers have no such familiar parallel. In this paper we aim to change that, proposing a computational model for the transformer-encoder in the form of a programming language. We map the basic components of a transformer-encoder -- attention and feed-forward computation -- into simple primitives, around which we form a programming language: the Restricted Access Sequence Processing Language (RASP). We show how RASP can be used to program solutions to tasks that could conceivably be learned by a Transformer, and how a Transformer can be trained to mimic a RASP solution. In particular, we provide RASP programs for histograms, sorting, and Dyck-languages. We further use our model to relate their difficulty in terms of the number of required layers and attention heads: analyzing a RASP program implies a maximum number of heads and layers necessary to encode a task in a transformer. Finally, we see how insights gained from our abstraction might be used to explain phenomena seen in recent works.
2021-06-13
{'id': '2106.06981v2', 'title': 'Thinking Like Transformers', 'abstract': 'What is the computational model behind a Transformer? Where recurrent neural\nnetworks have direct parallels in finite state machines, allowing clear\ndiscussion and thought around architecture variants or trained models,\nTransformers have no such familiar parallel. In this paper we aim to change\nthat, proposing a computational model for the transformer-encoder in the form\nof a programming language. We map the basic components of a transformer-encoder\n-- attention and feed-forward computation -- into simple primitives, around\nwhich we form a programming language: the Restricted Access Sequence Processing\nLanguage (RASP). We show how RASP can be used to program solutions to tasks\nthat could conceivably be learned by a Transformer, and how a Transformer can\nbe trained to mimic a RASP solution. In particular, we provide RASP programs\nfor histograms, sorting, and Dyck-languages. We further use our model to relate\ntheir difficulty in terms of the number of required layers and attention heads:\nanalyzing a RASP program implies a maximum number of heads and layers necessary\nto encode a task in a transformer. Finally, we see how insights gained from our\nabstraction might be used to explain phenomena seen in recent works.', 'authors': ['Gail Weiss', 'Yoav Goldberg', 'Eran Yahav'], 'published': '2021-06-13 13:04:46+00:00', 'categories': ['cs.LG', 'cs.CL'], 'url': 'http://arxiv.org/abs/2106.06981v2'}
42,462
2302.11068v3
2,302.11068
3
Low Rank Matrix Completion via Robust Alternating Minimization in Nearly Linear Time
Given a matrix $MR^mn$, the low rank matrix completion problem asks us to find a rank-$k$ approximation of $M$ as $UV^$ for $UR^mk$ and $VR^nk$ by only observing a few entries specified by a set of entries $[m][n]$. In particular, we examine an approach that is widely used in practice -- the alternating minimization framework. Jain, Netrapalli, and Sanghavi [JNS13] showed that if $M$ has incoherent rows and columns, then alternating minimization provably recovers the matrix $M$ by observing a nearly linear in $n$ number of entries. While the sample complexity has been subsequently improved [GLZ17], alternating minimization steps are required to be computed exactly. This hinders the development of more efficient algorithms and fails to depict the practical implementation of alternating minimization, where the updates are usually performed approximately in favor of efficiency. In this paper, we take a major step towards a more efficient and error-robust alternating minimization framework. To this end, we develop an analytical framework for alternating minimization that can tolerate a moderate amount of errors caused by approximate updates. Moreover, our algorithm runs in time $O(|| k)$, which is nearly linear in the time to verify the solution while preserving the sample complexity. This improves upon all prior known alternating minimization approaches which require $O(|| k^2)$ time.
2023-02-21
{'id': '2302.11068v3', 'title': 'Low Rank Matrix Completion via Robust Alternating Minimization in Nearly Linear Time', 'abstract': 'Given a matrix $M\\in \\mathbb{R}^{m\\times n}$, the low rank matrix completion\nproblem asks us to find a rank-$k$ approximation of $M$ as $UV^\\top$ for $U\\in\n\\mathbb{R}^{m\\times k}$ and $V\\in \\mathbb{R}^{n\\times k}$ by only observing a\nfew entries specified by a set of entries $\\Omega\\subseteq [m]\\times [n]$. In\nparticular, we examine an approach that is widely used in practice -- the\nalternating minimization framework. Jain, Netrapalli, and Sanghavi [JNS13]\nshowed that if $M$ has incoherent rows and columns, then alternating\nminimization provably recovers the matrix $M$ by observing a nearly linear in\n$n$ number of entries. While the sample complexity has been subsequently\nimproved [GLZ17], alternating minimization steps are required to be computed\nexactly. This hinders the development of more efficient algorithms and fails to\ndepict the practical implementation of alternating minimization, where the\nupdates are usually performed approximately in favor of efficiency.\n In this paper, we take a major step towards a more efficient and error-robust\nalternating minimization framework. To this end, we develop an analytical\nframework for alternating minimization that can tolerate a moderate amount of\nerrors caused by approximate updates. Moreover, our algorithm runs in time\n$\\widetilde O(|\\Omega| k)$, which is nearly linear in the time to verify the\nsolution while preserving the sample complexity. This improves upon all prior\nknown alternating minimization approaches which require $\\widetilde O(|\\Omega|\nk^2)$ time.', 'authors': ['Yuzhou Gu', 'Zhao Song', 'Junze Yin', 'Lichen Zhang'], 'published': '2023-02-21 23:49:36+00:00', 'categories': ['cs.LG', 'cs.DS', 'math.OC', 'stat.ML'], 'url': 'http://arxiv.org/abs/2302.11068v3'}
42,392
2304.04897v1
2,304.04897
1
Neural Image-based Avatars: Generalizable Radiance Fields for Human Avatar Modeling
We present a method that enables synthesizing novel views and novel poses of arbitrary human performers from sparse multi-view images. A key ingredient of our method is a hybrid appearance blending module that combines the advantages of the implicit body NeRF representation and image-based rendering. Existing generalizable human NeRF methods that are conditioned on the body model have shown robustness against the geometric variation of arbitrary human performers. Yet they often exhibit blurry results when generalized onto unseen identities. Meanwhile, image-based rendering shows high-quality results when sufficient observations are available, whereas it suffers artifacts in sparse-view settings. We propose Neural Image-based Avatars (NIA) that exploits the best of those two methods: to maintain robustness under new articulations and self-occlusions while directly leveraging the available (sparse) source view colors to preserve appearance details of new subject identities. Our hybrid design outperforms recent methods on both in-domain identity generalization as well as challenging cross-dataset generalization settings. Also, in terms of the pose generalization, our method outperforms even the per-subject optimized animatable NeRF methods. The video results are available at https://youngjoongunc.github.io/nia
2023-04-10
{'id': '2304.04897v1', 'title': 'Neural Image-based Avatars: Generalizable Radiance Fields for Human Avatar Modeling', 'abstract': 'We present a method that enables synthesizing novel views and novel poses of\narbitrary human performers from sparse multi-view images. A key ingredient of\nour method is a hybrid appearance blending module that combines the advantages\nof the implicit body NeRF representation and image-based rendering. Existing\ngeneralizable human NeRF methods that are conditioned on the body model have\nshown robustness against the geometric variation of arbitrary human performers.\nYet they often exhibit blurry results when generalized onto unseen identities.\nMeanwhile, image-based rendering shows high-quality results when sufficient\nobservations are available, whereas it suffers artifacts in sparse-view\nsettings. We propose Neural Image-based Avatars (NIA) that exploits the best of\nthose two methods: to maintain robustness under new articulations and\nself-occlusions while directly leveraging the available (sparse) source view\ncolors to preserve appearance details of new subject identities. Our hybrid\ndesign outperforms recent methods on both in-domain identity generalization as\nwell as challenging cross-dataset generalization settings. Also, in terms of\nthe pose generalization, our method outperforms even the per-subject optimized\nanimatable NeRF methods. The video results are available at\nhttps://youngjoongunc.github.io/nia', 'authors': ['Youngjoong Kwon', 'Dahun Kim', 'Duygu Ceylan', 'Henry Fuchs'], 'published': '2023-04-10 23:53:28+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2304.04897v1'}
42,403
2202.01344v1
2,202.01344
1
Formal Mathematics Statement Curriculum Learning
We explore the use of expert iteration in the context of language modeling applied to formal mathematics. We show that at same compute budget, expert iteration, by which we mean proof search interleaved with learning, dramatically outperforms proof search only. We also observe that when applied to a collection of formal statements of sufficiently varied difficulty, expert iteration is capable of finding and solving a curriculum of increasingly difficult problems, without the need for associated ground-truth proofs. Finally, by applying this expert iteration to a manually curated set of problem statements, we achieve state-of-the-art on the miniF2F benchmark, automatically solving multiple challenging problems drawn from high school olympiads.
2022-02-03
{'id': '2202.01344v1', 'title': 'Formal Mathematics Statement Curriculum Learning', 'abstract': 'We explore the use of expert iteration in the context of language modeling\napplied to formal mathematics. We show that at same compute budget, expert\niteration, by which we mean proof search interleaved with learning,\ndramatically outperforms proof search only. We also observe that when applied\nto a collection of formal statements of sufficiently varied difficulty, expert\niteration is capable of finding and solving a curriculum of increasingly\ndifficult problems, without the need for associated ground-truth proofs.\nFinally, by applying this expert iteration to a manually curated set of problem\nstatements, we achieve state-of-the-art on the miniF2F benchmark, automatically\nsolving multiple challenging problems drawn from high school olympiads.', 'authors': ['Stanislas Polu', 'Jesse Michael Han', 'Kunhao Zheng', 'Mantas Baksys', 'Igor Babuschkin', 'Ilya Sutskever'], 'published': '2022-02-03 00:17:00+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2202.01344v1'}
42,404
2302.11346v1
2,302.11346
1
Task-Aware Information Routing from Common Representation Space in Lifelong Learning
Intelligent systems deployed in the real world suffer from catastrophic forgetting when exposed to a sequence of tasks. Humans, on the other hand, acquire, consolidate, and transfer knowledge between tasks that rarely interfere with the consolidated knowledge. Accompanied by self-regulated neurogenesis, continual learning in the brain is governed by a rich set of neurophysiological processes that harbor different types of knowledge, which are then integrated by conscious processing. Thus, inspired by the Global Workspace Theory of conscious information access in the brain, we propose TAMiL, a continual learning method that entails task-attention modules to capture task-specific information from the common representation space. We employ simple, undercomplete autoencoders to create a communication bottleneck between the common representation space and the global workspace, allowing only the task-relevant information to the global workspace, thus greatly reducing task interference. Experimental results show that our method outperforms state-of-the-art rehearsal-based and dynamic sparse approaches and bridges the gap between fixed capacity and parameter isolation approaches while being scalable. We also show that our method effectively mitigates catastrophic forgetting while being well-calibrated with reduced task-recency bias.
2023-02-14
{'id': '2302.11346v1', 'title': 'Task-Aware Information Routing from Common Representation Space in Lifelong Learning', 'abstract': 'Intelligent systems deployed in the real world suffer from catastrophic\nforgetting when exposed to a sequence of tasks. Humans, on the other hand,\nacquire, consolidate, and transfer knowledge between tasks that rarely\ninterfere with the consolidated knowledge. Accompanied by self-regulated\nneurogenesis, continual learning in the brain is governed by a rich set of\nneurophysiological processes that harbor different types of knowledge, which\nare then integrated by conscious processing. Thus, inspired by the Global\nWorkspace Theory of conscious information access in the brain, we propose\nTAMiL, a continual learning method that entails task-attention modules to\ncapture task-specific information from the common representation space. We\nemploy simple, undercomplete autoencoders to create a communication bottleneck\nbetween the common representation space and the global workspace, allowing only\nthe task-relevant information to the global workspace, thus greatly reducing\ntask interference. Experimental results show that our method outperforms\nstate-of-the-art rehearsal-based and dynamic sparse approaches and bridges the\ngap between fixed capacity and parameter isolation approaches while being\nscalable. We also show that our method effectively mitigates catastrophic\nforgetting while being well-calibrated with reduced task-recency bias.', 'authors': ['Prashant Bhat', 'Bahram Zonooz', 'Elahe Arani'], 'published': '2023-02-14 16:25:46+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CV'], 'url': 'http://arxiv.org/abs/2302.11346v1'}
42,406
2203.02006v1
2,203.02006
1
Why adversarial training can hurt robust accuracy
Machine learning classifiers with high test accuracy often perform poorly under adversarial attacks. It is commonly believed that adversarial training alleviates this issue. In this paper, we demonstrate that, surprisingly, the opposite may be true -- Even though adversarial training helps when enough data is available, it may hurt robust generalization in the small sample size regime. We first prove this phenomenon for a high-dimensional linear classification setting with noiseless observations. Our proof provides explanatory insights that may also transfer to feature learning models. Further, we observe in experiments on standard image datasets that the same behavior occurs for perceptible attacks that effectively reduce class information such as mask attacks and object corruptions.
2022-03-03
{'id': '2203.02006v1', 'title': 'Why adversarial training can hurt robust accuracy', 'abstract': 'Machine learning classifiers with high test accuracy often perform poorly\nunder adversarial attacks. It is commonly believed that adversarial training\nalleviates this issue. In this paper, we demonstrate that, surprisingly, the\nopposite may be true -- Even though adversarial training helps when enough data\nis available, it may hurt robust generalization in the small sample size\nregime. We first prove this phenomenon for a high-dimensional linear\nclassification setting with noiseless observations. Our proof provides\nexplanatory insights that may also transfer to feature learning models.\nFurther, we observe in experiments on standard image datasets that the same\nbehavior occurs for perceptible attacks that effectively reduce class\ninformation such as mask attacks and object corruptions.', 'authors': ['Jacob Clarysse', 'Julia Hörrmann', 'Fanny Yang'], 'published': '2022-03-03 20:41:38+00:00', 'categories': ['cs.LG', 'cs.CR', 'cs.CV', 'stat.ML'], 'url': 'http://arxiv.org/abs/2203.02006v1'}
42,425
2310.08470v1
2,310.0847
1
Strategies and impact of learning curve estimation for CNN-based image classification
Learning curves are a measure for how the performance of machine learning models improves given a certain volume of training data. Over a wide variety of applications and models it was observed that learning curves follow -- to a large extent -- a power law behavior. This makes the performance of different models for a given task somewhat predictable and opens the opportunity to reduce the training time for practitioners, who are exploring the space of possible models and hyperparameters for the problem at hand. By estimating the learning curve of a model from training on small subsets of data only the best models need to be considered for training on the full dataset. How to choose subset sizes and how often to sample models on these to obtain estimates is however not researched. Given that the goal is to reduce overall training time strategies are needed that sample the performance in a time-efficient way and yet leads to accurate learning curve estimates. In this paper we formulate the framework for these strategies and propose several strategies. Further we evaluate the strategies for simulated learning curves and in experiments with popular datasets and models for image classification tasks.
2023-10-12
{'id': '2310.08470v1', 'title': 'Strategies and impact of learning curve estimation for CNN-based image classification', 'abstract': 'Learning curves are a measure for how the performance of machine learning\nmodels improves given a certain volume of training data. Over a wide variety of\napplications and models it was observed that learning curves follow -- to a\nlarge extent -- a power law behavior. This makes the performance of different\nmodels for a given task somewhat predictable and opens the opportunity to\nreduce the training time for practitioners, who are exploring the space of\npossible models and hyperparameters for the problem at hand. By estimating the\nlearning curve of a model from training on small subsets of data only the best\nmodels need to be considered for training on the full dataset. How to choose\nsubset sizes and how often to sample models on these to obtain estimates is\nhowever not researched. Given that the goal is to reduce overall training time\nstrategies are needed that sample the performance in a time-efficient way and\nyet leads to accurate learning curve estimates. In this paper we formulate the\nframework for these strategies and propose several strategies. Further we\nevaluate the strategies for simulated learning curves and in experiments with\npopular datasets and models for image classification tasks.', 'authors': ['Laura Didyk', 'Brayden Yarish', 'Michael A. Beck', 'Christopher P. Bidinosti', 'Christopher J. Henry'], 'published': '2023-10-12 16:28:25+00:00', 'categories': ['cs.LG', 'cs.NE'], 'url': 'http://arxiv.org/abs/2310.08470v1'}
42,402
2111.15363v2
2,111.15363
2
Voint Cloud: Multi-View Point Cloud Representation for 3D Understanding
Multi-view projection methods have demonstrated promising performance on 3D understanding tasks like 3D classification and segmentation. However, it remains unclear how to combine such multi-view methods with the widely available 3D point clouds. Previous methods use unlearned heuristics to combine features at the point level. To this end, we introduce the concept of the multi-view point cloud (Voint cloud), representing each 3D point as a set of features extracted from several view-points. This novel 3D Voint cloud representation combines the compactness of 3D point cloud representation with the natural view-awareness of multi-view representation. Naturally, we can equip this new representation with convolutional and pooling operations. We deploy a Voint neural network (VointNet) to learn representations in the Voint space. Our novel representation achieves performance on 3D classification, shape retrieval, and robust 3D part segmentation on standard benchmarks ( ScanObjectNN, ShapeNet Core55, and ShapeNet Parts).
2021-11-30
{'id': '2111.15363v2', 'title': 'Voint Cloud: Multi-View Point Cloud Representation for 3D Understanding', 'abstract': 'Multi-view projection methods have demonstrated promising performance on 3D\nunderstanding tasks like 3D classification and segmentation. However, it\nremains unclear how to combine such multi-view methods with the widely\navailable 3D point clouds. Previous methods use unlearned heuristics to combine\nfeatures at the point level. To this end, we introduce the concept of the\nmulti-view point cloud (Voint cloud), representing each 3D point as a set of\nfeatures extracted from several view-points. This novel 3D Voint cloud\nrepresentation combines the compactness of 3D point cloud representation with\nthe natural view-awareness of multi-view representation. Naturally, we can\nequip this new representation with convolutional and pooling operations. We\ndeploy a Voint neural network (VointNet) to learn representations in the Voint\nspace. Our novel representation achieves \\sota performance on 3D\nclassification, shape retrieval, and robust 3D part segmentation on standard\nbenchmarks ( ScanObjectNN, ShapeNet Core55, and ShapeNet Parts).', 'authors': ['Abdullah Hamdi', 'Silvio Giancola', 'Bernard Ghanem'], 'published': '2021-11-30 13:08:19+00:00', 'categories': ['cs.CV', 'cs.LG', '68T45'], 'url': 'http://arxiv.org/abs/2111.15363v2'}
42,426
2302.03933v1
2,302.03933
1
Graph Signal Sampling for Inductive One-Bit Matrix Completion: a Closed-form Solution
Inductive one-bit matrix completion is motivated by modern applications such as recommender systems, where new users would appear at test stage with the ratings consisting of only ones and no zeros. We propose a unified graph signal sampling framework which enjoys the benefits of graph signal analysis and processing. The key idea is to transform each user's ratings on the items to a function (signal) on the vertices of an item-item graph, then learn structural graph properties to recover the function from its values on certain vertices -- the problem of graph signal sampling. We propose a class of regularization functionals that takes into account discrete random label noise in the graph vertex domain, then develop the GS-IMC approach which biases the reconstruction towards functions that vary little between adjacent vertices for noise reduction. Theoretical result shows that accurate reconstructions can be achieved under mild conditions. For the online setting, we develop a Bayesian extension, i.e., BGS-IMC which considers continuous random Gaussian noise in the graph Fourier domain and builds upon a prediction-correction update algorithm to obtain the unbiased and minimum-variance reconstruction. Both GS-IMC and BGS-IMC have closed-form solutions and thus are highly scalable in large data. Experiments show that our methods achieve state-of-the-art performance on public benchmarks.
2023-02-08
{'id': '2302.03933v1', 'title': 'Graph Signal Sampling for Inductive One-Bit Matrix Completion: a Closed-form Solution', 'abstract': "Inductive one-bit matrix completion is motivated by modern applications such\nas recommender systems, where new users would appear at test stage with the\nratings consisting of only ones and no zeros. We propose a unified graph signal\nsampling framework which enjoys the benefits of graph signal analysis and\nprocessing. The key idea is to transform each user's ratings on the items to a\nfunction (signal) on the vertices of an item-item graph, then learn structural\ngraph properties to recover the function from its values on certain vertices --\nthe problem of graph signal sampling. We propose a class of regularization\nfunctionals that takes into account discrete random label noise in the graph\nvertex domain, then develop the GS-IMC approach which biases the reconstruction\ntowards functions that vary little between adjacent vertices for noise\nreduction. Theoretical result shows that accurate reconstructions can be\nachieved under mild conditions. For the online setting, we develop a Bayesian\nextension, i.e., BGS-IMC which considers continuous random Gaussian noise in\nthe graph Fourier domain and builds upon a prediction-correction update\nalgorithm to obtain the unbiased and minimum-variance reconstruction. Both\nGS-IMC and BGS-IMC have closed-form solutions and thus are highly scalable in\nlarge data. Experiments show that our methods achieve state-of-the-art\nperformance on public benchmarks.", 'authors': ['Chao Chen', 'Haoyu Geng', 'Gang Zeng', 'Zhaobing Han', 'Hua Chai', 'Xiaokang Yang', 'Junchi Yan'], 'published': '2023-02-08 08:17:43+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2302.03933v1'}
42,427
2210.03945v2
2,210.03945
2
Understanding HTML with Large Language Models
Large language models (LLMs) have shown exceptional performance on a variety of natural language tasks. Yet, their capabilities for HTML understanding -- i.e., parsing the raw HTML of a webpage, with applications to automation of web-based tasks, crawling, and browser-assisted retrieval -- have not been fully explored. We contribute HTML understanding models (fine-tuned LLMs) and an in-depth analysis of their capabilities under three tasks: (i) Semantic Classification of HTML elements, (ii) Description Generation for HTML inputs, and (iii) Autonomous Web Navigation of HTML pages. While previous work has developed dedicated architectures and training procedures for HTML understanding, we show that LLMs pretrained on standard natural language corpora transfer remarkably well to HTML understanding tasks. For instance, fine-tuned LLMs are 12% more accurate at semantic classification compared to models trained exclusively on the task dataset. Moreover, when fine-tuned on data from the MiniWoB benchmark, LLMs successfully complete 50% more tasks using 192x less data compared to the previous best supervised model. Out of the LLMs we evaluate, we show evidence that T5-based models are ideal due to their bidirectional encoder-decoder architecture. To promote further research on LLMs for HTML understanding, we create and open-source a large-scale HTML dataset distilled and auto-labeled from CommonCrawl.
2022-10-08
{'id': '2210.03945v2', 'title': 'Understanding HTML with Large Language Models', 'abstract': 'Large language models (LLMs) have shown exceptional performance on a variety\nof natural language tasks. Yet, their capabilities for HTML understanding --\ni.e., parsing the raw HTML of a webpage, with applications to automation of\nweb-based tasks, crawling, and browser-assisted retrieval -- have not been\nfully explored. We contribute HTML understanding models (fine-tuned LLMs) and\nan in-depth analysis of their capabilities under three tasks: (i) Semantic\nClassification of HTML elements, (ii) Description Generation for HTML inputs,\nand (iii) Autonomous Web Navigation of HTML pages. While previous work has\ndeveloped dedicated architectures and training procedures for HTML\nunderstanding, we show that LLMs pretrained on standard natural language\ncorpora transfer remarkably well to HTML understanding tasks. For instance,\nfine-tuned LLMs are 12% more accurate at semantic classification compared to\nmodels trained exclusively on the task dataset. Moreover, when fine-tuned on\ndata from the MiniWoB benchmark, LLMs successfully complete 50% more tasks\nusing 192x less data compared to the previous best supervised model. Out of the\nLLMs we evaluate, we show evidence that T5-based models are ideal due to their\nbidirectional encoder-decoder architecture. To promote further research on LLMs\nfor HTML understanding, we create and open-source a large-scale HTML dataset\ndistilled and auto-labeled from CommonCrawl.', 'authors': ['Izzeddin Gur', 'Ofir Nachum', 'Yingjie Miao', 'Mustafa Safdari', 'Austin Huang', 'Aakanksha Chowdhery', 'Sharan Narang', 'Noah Fiedel', 'Aleksandra Faust'], 'published': '2022-10-08 07:27:17+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2210.03945v2'}
42,443
2210.04492v2
2,210.04492
2
Unified Detoxifying and Debiasing in Language Generation via Inference-time Adaptive Optimization
Warning: this paper contains model outputs exhibiting offensiveness and biases. Recently pre-trained language models (PLMs) have prospered in various natural language generation (NLG) tasks due to their ability to generate fairly fluent text. Nevertheless, these models are observed to capture and reproduce harmful contents in training corpora, typically toxic language and social biases, raising severe moral issues. Prior works on ethical NLG tackle detoxifying and debiasing separately, which is problematic since we find debiased models still exhibit toxicity while detoxified ones even exacerbate social biases. To address such a challenge, we propose the first unified framework of detoxifying and debiasing called UDDIA, which jointly formalizes these two problems as rectifying the output space. We theoretically interpret our framework as learning a text distribution mixing weighted attributes. Besides, UDDIA conducts adaptive optimization of only a few parameters during decoding based on a parameter-efficient tuning schema without any training data. This leads to minimal generation quality loss and improved rectification performance with acceptable computational cost. Experimental results demonstrate that compared to several strong baselines, UDDIA achieves debiasing and detoxifying simultaneously and better balances efficiency and effectiveness, taking a further step towards practical ethical NLG.
2022-10-10
{'id': '2210.04492v2', 'title': 'Unified Detoxifying and Debiasing in Language Generation via Inference-time Adaptive Optimization', 'abstract': 'Warning: this paper contains model outputs exhibiting offensiveness and\nbiases. Recently pre-trained language models (PLMs) have prospered in various\nnatural language generation (NLG) tasks due to their ability to generate fairly\nfluent text. Nevertheless, these models are observed to capture and reproduce\nharmful contents in training corpora, typically toxic language and social\nbiases, raising severe moral issues. Prior works on ethical NLG tackle\ndetoxifying and debiasing separately, which is problematic since we find\ndebiased models still exhibit toxicity while detoxified ones even exacerbate\nsocial biases. To address such a challenge, we propose the first unified\nframework of detoxifying and debiasing called UDDIA, which jointly formalizes\nthese two problems as rectifying the output space. We theoretically interpret\nour framework as learning a text distribution mixing weighted attributes.\nBesides, UDDIA conducts adaptive optimization of only a few parameters during\ndecoding based on a parameter-efficient tuning schema without any training\ndata. This leads to minimal generation quality loss and improved rectification\nperformance with acceptable computational cost. Experimental results\ndemonstrate that compared to several strong baselines, UDDIA achieves debiasing\nand detoxifying simultaneously and better balances efficiency and\neffectiveness, taking a further step towards practical ethical NLG.', 'authors': ['Zonghan Yang', 'Xiaoyuan Yi', 'Peng Li', 'Yang Liu', 'Xing Xie'], 'published': '2022-10-10 08:45:25+00:00', 'categories': ['cs.CL', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2210.04492v2'}
42,446
2207.01566v1
2,207.01566
1
General Policy Evaluation and Improvement by Learning to Identify Few But Crucial States
Learning to evaluate and improve policies is a core problem of Reinforcement Learning (RL). Traditional RL algorithms learn a value function defined for a single policy. A recently explored competitive alternative is to learn a single value function for many policies. Here we combine the actor-critic architecture of Parameter-Based Value Functions and the policy embedding of Policy Evaluation Networks to learn a single value function for evaluating (and thus helping to improve) any policy represented by a deep neural network (NN). The method yields competitive experimental results. In continuous control problems with infinitely many states, our value function minimizes its prediction error by simultaneously learning a small set of `probing states' and a mapping from actions produced in probing states to the policy's return. The method extracts crucial abstract knowledge about the environment in form of very few states sufficient to fully specify the behavior of many policies. A policy improves solely by changing actions in probing states, following the gradient of the value function's predictions. Surprisingly, it is possible to clone the behavior of a near-optimal policy in Swimmer-v3 and Hopper-v3 environments only by knowing how to act in 3 and 5 such learned states, respectively. Remarkably, our value function trained to evaluate NN policies is also invariant to changes of the policy architecture: we show that it allows for zero-shot learning of linear policies competitive with the best policy seen during training. Our code is public.
2022-07-04
{'id': '2207.01566v1', 'title': 'General Policy Evaluation and Improvement by Learning to Identify Few But Crucial States', 'abstract': "Learning to evaluate and improve policies is a core problem of Reinforcement\nLearning (RL). Traditional RL algorithms learn a value function defined for a\nsingle policy. A recently explored competitive alternative is to learn a single\nvalue function for many policies. Here we combine the actor-critic architecture\nof Parameter-Based Value Functions and the policy embedding of Policy\nEvaluation Networks to learn a single value function for evaluating (and thus\nhelping to improve) any policy represented by a deep neural network (NN). The\nmethod yields competitive experimental results. In continuous control problems\nwith infinitely many states, our value function minimizes its prediction error\nby simultaneously learning a small set of `probing states' and a mapping from\nactions produced in probing states to the policy's return. The method extracts\ncrucial abstract knowledge about the environment in form of very few states\nsufficient to fully specify the behavior of many policies. A policy improves\nsolely by changing actions in probing states, following the gradient of the\nvalue function's predictions. Surprisingly, it is possible to clone the\nbehavior of a near-optimal policy in Swimmer-v3 and Hopper-v3 environments only\nby knowing how to act in 3 and 5 such learned states, respectively. Remarkably,\nour value function trained to evaluate NN policies is also invariant to changes\nof the policy architecture: we show that it allows for zero-shot learning of\nlinear policies competitive with the best policy seen during training. Our code\nis public.", 'authors': ['Francesco Faccio', 'Aditya Ramesh', 'Vincent Herrmann', 'Jean Harb', 'Jürgen Schmidhuber'], 'published': '2022-07-04 16:34:53+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2207.01566v1'}
42,409
2210.00583v1
2,210.00583
1
The Dynamic of Consensus in Deep Networks and the Identification of Noisy Labels
Deep neural networks have incredible capacity and expressibility, and can seemingly memorize any training set. This introduces a problem when training in the presence of noisy labels, as the noisy examples cannot be distinguished from clean examples by the end of training. Recent research has dealt with this challenge by utilizing the fact that deep networks seem to memorize clean examples much earlier than noisy examples. Here we report a new empirical result: for each example, when looking at the time it has been memorized by each model in an ensemble of networks, the diversity seen in noisy examples is much larger than the clean examples. We use this observation to develop a new method for noisy labels filtration. The method is based on a statistics of the data, which captures the differences in ensemble learning dynamics between clean and noisy data. We test our method on three tasks: (i) noise amount estimation; (ii) noise filtration; (iii) supervised classification. We show that our method improves over existing baselines in all three tasks using a variety of datasets, noise models, and noise levels. Aside from its improved performance, our method has two other advantages. (i) Simplicity, which implies that no additional hyperparameters are introduced. (ii) Our method is modular: it does not work in an end-to-end fashion, and can therefore be used to clean a dataset for any other future usage.
2022-10-02
{'id': '2210.00583v1', 'title': 'The Dynamic of Consensus in Deep Networks and the Identification of Noisy Labels', 'abstract': 'Deep neural networks have incredible capacity and expressibility, and can\nseemingly memorize any training set. This introduces a problem when training in\nthe presence of noisy labels, as the noisy examples cannot be distinguished\nfrom clean examples by the end of training. Recent research has dealt with this\nchallenge by utilizing the fact that deep networks seem to memorize clean\nexamples much earlier than noisy examples. Here we report a new empirical\nresult: for each example, when looking at the time it has been memorized by\neach model in an ensemble of networks, the diversity seen in noisy examples is\nmuch larger than the clean examples. We use this observation to develop a new\nmethod for noisy labels filtration. The method is based on a statistics of the\ndata, which captures the differences in ensemble learning dynamics between\nclean and noisy data. We test our method on three tasks: (i) noise amount\nestimation; (ii) noise filtration; (iii) supervised classification. We show\nthat our method improves over existing baselines in all three tasks using a\nvariety of datasets, noise models, and noise levels. Aside from its improved\nperformance, our method has two other advantages. (i) Simplicity, which implies\nthat no additional hyperparameters are introduced. (ii) Our method is modular:\nit does not work in an end-to-end fashion, and can therefore be used to clean a\ndataset for any other future usage.', 'authors': ['Daniel Shwartz', 'Uri Stern', 'Daphna Weinshall'], 'published': '2022-10-02 17:47:23+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2210.00583v1'}
42,423
2311.18817v2
2,311.18817
2
Dichotomy of Early and Late Phase Implicit Biases Can Provably Induce Grokking
Recent work by Power et al. (2022) highlighted a surprising "grokking" phenomenon in learning arithmetic tasks: a neural net first "memorizes" the training set, resulting in perfect training accuracy but near-random test accuracy, and after training for sufficiently longer, it suddenly transitions to perfect test accuracy. This paper studies the grokking phenomenon in theoretical setups and shows that it can be induced by a dichotomy of early and late phase implicit biases. Specifically, when training homogeneous neural nets with large initialization and small weight decay on both classification and regression tasks, we prove that the training process gets trapped at a solution corresponding to a kernel predictor for a long time, and then a very sharp transition to min-norm/max-margin predictors occurs, leading to a dramatic change in test accuracy.
2023-11-30
{'id': '2311.18817v2', 'title': 'Dichotomy of Early and Late Phase Implicit Biases Can Provably Induce Grokking', 'abstract': 'Recent work by Power et al. (2022) highlighted a surprising "grokking"\nphenomenon in learning arithmetic tasks: a neural net first "memorizes" the\ntraining set, resulting in perfect training accuracy but near-random test\naccuracy, and after training for sufficiently longer, it suddenly transitions\nto perfect test accuracy. This paper studies the grokking phenomenon in\ntheoretical setups and shows that it can be induced by a dichotomy of early and\nlate phase implicit biases. Specifically, when training homogeneous neural nets\nwith large initialization and small weight decay on both classification and\nregression tasks, we prove that the training process gets trapped at a solution\ncorresponding to a kernel predictor for a long time, and then a very sharp\ntransition to min-norm/max-margin predictors occurs, leading to a dramatic\nchange in test accuracy.', 'authors': ['Kaifeng Lyu', 'Jikai Jin', 'Zhiyuan Li', 'Simon S. Du', 'Jason D. Lee', 'Wei Hu'], 'published': '2023-11-30 18:55:38+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2311.18817v2'}
42,440
2306.05292v2
2,306.05292
2
Safe Collaborative Filtering
Excellent tail performance is crucial for modern machine learning tasks, such as algorithmic fairness, class imbalance, and risk-sensitive decision making, as it ensures the effective handling of challenging samples within a dataset. Tail performance is also a vital determinant of success for personalized recommender systems to reduce the risk of losing users with low satisfaction. This study introduces a "safe" collaborative filtering method that prioritizes recommendation quality for less-satisfied users rather than focusing on the average performance. Our approach minimizes the conditional value at risk (CVaR), which represents the average risk over the tails of users' loss. To overcome computational challenges for web-scale recommender systems, we develop a robust yet practical algorithm that extends the most scalable method, implicit alternating least squares (iALS). Empirical evaluation on real-world datasets demonstrates the excellent tail performance of our approach while maintaining competitive computational efficiency.
2023-06-08
{'id': '2306.05292v2', 'title': 'Safe Collaborative Filtering', 'abstract': 'Excellent tail performance is crucial for modern machine learning tasks, such\nas algorithmic fairness, class imbalance, and risk-sensitive decision making,\nas it ensures the effective handling of challenging samples within a dataset.\nTail performance is also a vital determinant of success for personalized\nrecommender systems to reduce the risk of losing users with low satisfaction.\nThis study introduces a "safe" collaborative filtering method that prioritizes\nrecommendation quality for less-satisfied users rather than focusing on the\naverage performance. Our approach minimizes the conditional value at risk\n(CVaR), which represents the average risk over the tails of users\' loss. To\novercome computational challenges for web-scale recommender systems, we develop\na robust yet practical algorithm that extends the most scalable method,\nimplicit alternating least squares (iALS). Empirical evaluation on real-world\ndatasets demonstrates the excellent tail performance of our approach while\nmaintaining competitive computational efficiency.', 'authors': ['Riku Togashi', 'Tatsushi Oka', 'Naoto Ohsaka', 'Tetsuro Morimura'], 'published': '2023-06-08 15:36:02+00:00', 'categories': ['cs.IR', 'cs.LG'], 'url': 'http://arxiv.org/abs/2306.05292v2'}
42,457
2506.02749v1
2,506.02749
1
Knowledge Graph Completion by Intermediate Variables Regularization
Knowledge graph completion (KGC) can be framed as a 3-order binary tensor completion task. Tensor decomposition-based (TDB) models have demonstrated strong performance in KGC. In this paper, we provide a summary of existing TDB models and derive a general form for them, serving as a foundation for further exploration of TDB models. Despite the expressiveness of TDB models, they are prone to overfitting. Existing regularization methods merely minimize the norms of embeddings to regularize the model, leading to suboptimal performance. Therefore, we propose a novel regularization method for TDB models that addresses this limitation. The regularization is applicable to most TDB models and ensures tractable computation. Our method minimizes the norms of intermediate variables involved in the different ways of computing the predicted tensor. To support our regularization method, we provide a theoretical analysis that proves its effect in promoting low trace norm of the predicted tensor to reduce overfitting. Finally, we conduct experiments to verify the effectiveness of our regularization technique as well as the reliability of our theoretical analysis. The code is available at https://github.com/changyi7231/IVR.
2025-06-03
{'id': '2506.02749v1', 'title': 'Knowledge Graph Completion by Intermediate Variables Regularization', 'abstract': 'Knowledge graph completion (KGC) can be framed as a 3-order binary tensor\ncompletion task. Tensor decomposition-based (TDB) models have demonstrated\nstrong performance in KGC. In this paper, we provide a summary of existing TDB\nmodels and derive a general form for them, serving as a foundation for further\nexploration of TDB models. Despite the expressiveness of TDB models, they are\nprone to overfitting. Existing regularization methods merely minimize the norms\nof embeddings to regularize the model, leading to suboptimal performance.\nTherefore, we propose a novel regularization method for TDB models that\naddresses this limitation. The regularization is applicable to most TDB models\nand ensures tractable computation. Our method minimizes the norms of\nintermediate variables involved in the different ways of computing the\npredicted tensor. To support our regularization method, we provide a\ntheoretical analysis that proves its effect in promoting low trace norm of the\npredicted tensor to reduce overfitting. Finally, we conduct experiments to\nverify the effectiveness of our regularization technique as well as the\nreliability of our theoretical analysis. The code is available at\nhttps://github.com/changyi7231/IVR.', 'authors': ['Changyi Xiao', 'Yixin Cao'], 'published': '2025-06-03 11:11:33+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2506.02749v1'}
42,460
2310.06474v3
2,310.06474
3
Multilingual Jailbreak Challenges in Large Language Models
While large language models (LLMs) exhibit remarkable capabilities across a wide range of tasks, they pose potential safety concerns, such as the ``jailbreak'' problem, wherein malicious instructions can manipulate LLMs to exhibit undesirable behavior. Although several preventive measures have been developed to mitigate the potential risks associated with LLMs, they have primarily focused on English. In this study, we reveal the presence of multilingual jailbreak challenges within LLMs and consider two potential risky scenarios: unintentional and intentional. The unintentional scenario involves users querying LLMs using non-English prompts and inadvertently bypassing the safety mechanisms, while the intentional scenario concerns malicious users combining malicious instructions with multilingual prompts to deliberately attack LLMs. The experimental results reveal that in the unintentional scenario, the rate of unsafe content increases as the availability of languages decreases. Specifically, low-resource languages exhibit about three times the likelihood of encountering harmful content compared to high-resource languages, with both ChatGPT and GPT-4. In the intentional scenario, multilingual prompts can exacerbate the negative impact of malicious instructions, with astonishingly high rates of unsafe output: 80.92\% for ChatGPT and 40.71\% for GPT-4. To handle such a challenge in the multilingual context, we propose a novel Self-Defense framework that automatically generates multilingual training data for safety fine-tuning. Experimental results show that ChatGPT fine-tuned with such data can achieve a substantial reduction in unsafe content generation. Data is available at https://github.com/DAMO-NLP-SG/multilingual-safety-for-LLMs.
2023-10-10
{'id': '2310.06474v3', 'title': 'Multilingual Jailbreak Challenges in Large Language Models', 'abstract': "While large language models (LLMs) exhibit remarkable capabilities across a\nwide range of tasks, they pose potential safety concerns, such as the\n``jailbreak'' problem, wherein malicious instructions can manipulate LLMs to\nexhibit undesirable behavior. Although several preventive measures have been\ndeveloped to mitigate the potential risks associated with LLMs, they have\nprimarily focused on English. In this study, we reveal the presence of\nmultilingual jailbreak challenges within LLMs and consider two potential risky\nscenarios: unintentional and intentional. The unintentional scenario involves\nusers querying LLMs using non-English prompts and inadvertently bypassing the\nsafety mechanisms, while the intentional scenario concerns malicious users\ncombining malicious instructions with multilingual prompts to deliberately\nattack LLMs. The experimental results reveal that in the unintentional\nscenario, the rate of unsafe content increases as the availability of languages\ndecreases. Specifically, low-resource languages exhibit about three times the\nlikelihood of encountering harmful content compared to high-resource languages,\nwith both ChatGPT and GPT-4. In the intentional scenario, multilingual prompts\ncan exacerbate the negative impact of malicious instructions, with\nastonishingly high rates of unsafe output: 80.92\\% for ChatGPT and 40.71\\% for\nGPT-4. To handle such a challenge in the multilingual context, we propose a\nnovel \\textsc{Self-Defense} framework that automatically generates multilingual\ntraining data for safety fine-tuning. Experimental results show that ChatGPT\nfine-tuned with such data can achieve a substantial reduction in unsafe content\ngeneration. Data is available at\n\\url{https://github.com/DAMO-NLP-SG/multilingual-safety-for-LLMs}.", 'authors': ['Yue Deng', 'Wenxuan Zhang', 'Sinno Jialin Pan', 'Lidong Bing'], 'published': '2023-10-10 09:44:06+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2310.06474v3'}
42,465
2311.06295v2
2,311.06295
2
Gradual Optimization Learning for Conformational Energy Minimization
Molecular conformation optimization is crucial to computer-aided drug discovery and materials design. Traditional energy minimization techniques rely on iterative optimization methods that use molecular forces calculated by a physical simulator (oracle) as anti-gradients. However, this is a computationally expensive approach that requires many interactions with a physical simulator. One way to accelerate this procedure is to replace the physical simulator with a neural network. Despite recent progress in neural networks for molecular conformation energy prediction, such models are prone to distribution shift, leading to inaccurate energy minimization. We find that the quality of energy minimization with neural networks can be improved by providing optimization trajectories as additional training data. Still, it takes around $5 10^5$ additional conformations to match the physical simulator's optimization quality. In this work, we present the Gradual Optimization Learning Framework (GOLF) for energy minimization with neural networks that significantly reduces the required additional data. The framework consists of an efficient data-collecting scheme and an external optimizer. The external optimizer utilizes gradients from the energy prediction model to generate optimization trajectories, and the data-collecting scheme selects additional training data to be processed by the physical simulator. Our results demonstrate that the neural network trained with GOLF performs on par with the oracle on a benchmark of diverse drug-like molecules using $50$x less additional data.
2023-11-05
{'id': '2311.06295v2', 'title': 'Gradual Optimization Learning for Conformational Energy Minimization', 'abstract': "Molecular conformation optimization is crucial to computer-aided drug\ndiscovery and materials design. Traditional energy minimization techniques rely\non iterative optimization methods that use molecular forces calculated by a\nphysical simulator (oracle) as anti-gradients. However, this is a\ncomputationally expensive approach that requires many interactions with a\nphysical simulator. One way to accelerate this procedure is to replace the\nphysical simulator with a neural network. Despite recent progress in neural\nnetworks for molecular conformation energy prediction, such models are prone to\ndistribution shift, leading to inaccurate energy minimization. We find that the\nquality of energy minimization with neural networks can be improved by\nproviding optimization trajectories as additional training data. Still, it\ntakes around $5 \\times 10^5$ additional conformations to match the physical\nsimulator's optimization quality. In this work, we present the Gradual\nOptimization Learning Framework (GOLF) for energy minimization with neural\nnetworks that significantly reduces the required additional data. The framework\nconsists of an efficient data-collecting scheme and an external optimizer. The\nexternal optimizer utilizes gradients from the energy prediction model to\ngenerate optimization trajectories, and the data-collecting scheme selects\nadditional training data to be processed by the physical simulator. Our results\ndemonstrate that the neural network trained with GOLF performs on par with the\noracle on a benchmark of diverse drug-like molecules using $50$x less\nadditional data.", 'authors': ['Artem Tsypin', 'Leonid Ugadiarov', 'Kuzma Khrabrov', 'Alexander Telepov', 'Egor Rumiantsev', 'Alexey Skrynnik', 'Aleksandr I. Panov', 'Dmitry Vetrov', 'Elena Tutubalina', 'Artur Kadurin'], 'published': '2023-11-05 11:48:08+00:00', 'categories': ['physics.chem-ph', 'cs.LG'], 'url': 'http://arxiv.org/abs/2311.06295v2'}
42,412
2405.02766v1
2,405.02766
1
Beyond Unimodal Learning: The Importance of Integrating Multiple Modalities for Lifelong Learning
While humans excel at continual learning (CL), deep neural networks (DNNs) exhibit catastrophic forgetting. A salient feature of the brain that allows effective CL is that it utilizes multiple modalities for learning and inference, which is underexplored in DNNs. Therefore, we study the role and interactions of multiple modalities in mitigating forgetting and introduce a benchmark for multimodal continual learning. Our findings demonstrate that leveraging multiple views and complementary information from multiple modalities enables the model to learn more accurate and robust representations. This makes the model less vulnerable to modality-specific regularities and considerably mitigates forgetting. Furthermore, we observe that individual modalities exhibit varying degrees of robustness to distribution shift. Finally, we propose a method for integrating and aligning the information from different modalities by utilizing the relational structural similarities between the data points in each modality. Our method sets a strong baseline that enables both single- and multimodal inference. Our study provides a promising case for further exploring the role of multiple modalities in enabling CL and provides a standard benchmark for future research.
2024-05-04
{'id': '2405.02766v1', 'title': 'Beyond Unimodal Learning: The Importance of Integrating Multiple Modalities for Lifelong Learning', 'abstract': 'While humans excel at continual learning (CL), deep neural networks (DNNs)\nexhibit catastrophic forgetting. A salient feature of the brain that allows\neffective CL is that it utilizes multiple modalities for learning and\ninference, which is underexplored in DNNs. Therefore, we study the role and\ninteractions of multiple modalities in mitigating forgetting and introduce a\nbenchmark for multimodal continual learning. Our findings demonstrate that\nleveraging multiple views and complementary information from multiple\nmodalities enables the model to learn more accurate and robust representations.\nThis makes the model less vulnerable to modality-specific regularities and\nconsiderably mitigates forgetting. Furthermore, we observe that individual\nmodalities exhibit varying degrees of robustness to distribution shift.\nFinally, we propose a method for integrating and aligning the information from\ndifferent modalities by utilizing the relational structural similarities\nbetween the data points in each modality. Our method sets a strong baseline\nthat enables both single- and multimodal inference. Our study provides a\npromising case for further exploring the role of multiple modalities in\nenabling CL and provides a standard benchmark for future research.', 'authors': ['Fahad Sarfraz', 'Bahram Zonooz', 'Elahe Arani'], 'published': '2024-05-04 22:02:58+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CV'], 'url': 'http://arxiv.org/abs/2405.02766v1'}
42,430
2310.16597v3
2,310.16597
3
Beyond IID weights: sparse and low-rank deep Neural Networks are also Gaussian Processes
The infinitely wide neural network has been proven a useful and manageable mathematical model that enables the understanding of many phenomena appearing in deep learning. One example is the convergence of random deep networks to Gaussian processes that allows a rigorous analysis of the way the choice of activation function and network weights impacts the training dynamics. In this paper, we extend the seminal proof of Matthews et al. (2018) to a larger class of initial weight distributions (which we call PSEUDO-IID), including the established cases of IID and orthogonal weights, as well as the emerging low-rank and structured sparse settings celebrated for their computational speed-up benefits. We show that fully-connected and convolutional networks initialized with PSEUDO-IID distributions are all effectively equivalent up to their variance. Using our results, one can identify the Edge-of-Chaos for a broader class of neural networks and tune them at criticality in order to enhance their training. Moreover, they enable the posterior distribution of Bayesian Neural Networks to be tractable across these various initialization schemes.
2023-10-25
{'id': '2310.16597v3', 'title': 'Beyond IID weights: sparse and low-rank deep Neural Networks are also Gaussian Processes', 'abstract': 'The infinitely wide neural network has been proven a useful and manageable\nmathematical model that enables the understanding of many phenomena appearing\nin deep learning. One example is the convergence of random deep networks to\nGaussian processes that allows a rigorous analysis of the way the choice of\nactivation function and network weights impacts the training dynamics. In this\npaper, we extend the seminal proof of Matthews et al. (2018) to a larger class\nof initial weight distributions (which we call PSEUDO-IID), including the\nestablished cases of IID and orthogonal weights, as well as the emerging\nlow-rank and structured sparse settings celebrated for their computational\nspeed-up benefits. We show that fully-connected and convolutional networks\ninitialized with PSEUDO-IID distributions are all effectively equivalent up to\ntheir variance. Using our results, one can identify the Edge-of-Chaos for a\nbroader class of neural networks and tune them at criticality in order to\nenhance their training. Moreover, they enable the posterior distribution of\nBayesian Neural Networks to be tractable across these various initialization\nschemes.', 'authors': ['Thiziri Nait-Saada', 'Alireza Naderi', 'Jared Tanner'], 'published': '2023-10-25 12:38:36+00:00', 'categories': ['stat.ML', 'cs.LG'], 'url': 'http://arxiv.org/abs/2310.16597v3'}
42,444
2312.07145v1
2,312.07145
1
Contextual Bandits with Online Neural Regression
Recent works have shown a reduction from contextual bandits to online regression under a realizability assumption [Foster and Rakhlin, 2020, Foster and Krishnamurthy, 2021]. In this work, we investigate the use of neural networks for such online regression and associated Neural Contextual Bandits (NeuCBs). Using existing results for wide networks, one can readily show a $O(T)$ regret for online regression with square loss, which via the reduction implies a $O(K T^3/4)$ regret for NeuCBs. Departing from this standard approach, we first show a $O(T)$ regret for online regression with almost convex losses that satisfy QG (Quadratic Growth) condition, a generalization of the PL (Polyak-ojasiewicz) condition, and that have a unique minima. Although not directly applicable to wide networks since they do not have unique minima, we show that adding a suitable small random perturbation to the network predictions surprisingly makes the loss satisfy QG with unique minima. Based on such a perturbed prediction, we show a $O(T)$ regret for online regression with both squared loss and KL loss, and subsequently convert these respectively to $O(KT)$ and $O(KL^* + K)$ regret for NeuCB, where $L^*$ is the loss of the best policy. Separately, we also show that existing regret bounds for NeuCBs are $(T)$ or assume i.i.d. contexts, unlike this work. Finally, our experimental results on various datasets demonstrate that our algorithms, especially the one based on KL loss, persistently outperform existing algorithms.
2023-12-12
{'id': '2312.07145v1', 'title': 'Contextual Bandits with Online Neural Regression', 'abstract': 'Recent works have shown a reduction from contextual bandits to online\nregression under a realizability assumption [Foster and Rakhlin, 2020, Foster\nand Krishnamurthy, 2021]. In this work, we investigate the use of neural\nnetworks for such online regression and associated Neural Contextual Bandits\n(NeuCBs). Using existing results for wide networks, one can readily show a\n${\\mathcal{O}}(\\sqrt{T})$ regret for online regression with square loss, which\nvia the reduction implies a ${\\mathcal{O}}(\\sqrt{K} T^{3/4})$ regret for\nNeuCBs. Departing from this standard approach, we first show a\n$\\mathcal{O}(\\log T)$ regret for online regression with almost convex losses\nthat satisfy QG (Quadratic Growth) condition, a generalization of the PL\n(Polyak-\\L ojasiewicz) condition, and that have a unique minima. Although not\ndirectly applicable to wide networks since they do not have unique minima, we\nshow that adding a suitable small random perturbation to the network\npredictions surprisingly makes the loss satisfy QG with unique minima. Based on\nsuch a perturbed prediction, we show a ${\\mathcal{O}}(\\log T)$ regret for\nonline regression with both squared loss and KL loss, and subsequently convert\nthese respectively to $\\tilde{\\mathcal{O}}(\\sqrt{KT})$ and\n$\\tilde{\\mathcal{O}}(\\sqrt{KL^*} + K)$ regret for NeuCB, where $L^*$ is the\nloss of the best policy. Separately, we also show that existing regret bounds\nfor NeuCBs are $\\Omega(T)$ or assume i.i.d. contexts, unlike this work.\nFinally, our experimental results on various datasets demonstrate that our\nalgorithms, especially the one based on KL loss, persistently outperform\nexisting algorithms.', 'authors': ['Rohan Deb', 'Yikun Ban', 'Shiliang Zuo', 'Jingrui He', 'Arindam Banerjee'], 'published': '2023-12-12 10:28:51+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2312.07145v1'}
42,445
2309.16671v5
2,309.16671
5
Demystifying CLIP Data
Contrastive Language-Image Pre-training (CLIP) is an approach that has advanced research and applications in computer vision, fueling modern recognition systems and generative models. We believe that the main ingredient to the success of CLIP is its data and not the model architecture or pre-training objective. However, CLIP only provides very limited information about its data and how it has been collected, leading to works that aim to reproduce CLIP's data by filtering with its model parameters. In this work, we intend to reveal CLIP's data curation approach and in our pursuit of making it open to the community introduce Metadata-Curated Language-Image Pre-training (MetaCLIP). MetaCLIP takes a raw data pool and metadata (derived from CLIP's concepts) and yields a balanced subset over the metadata distribution. Our experimental study rigorously isolates the model and training settings, concentrating solely on data. MetaCLIP applied to CommonCrawl with 400M image-text data pairs outperforms CLIP's data on multiple standard benchmarks. In zero-shot ImageNet classification, MetaCLIP achieves 70.8% accuracy, surpassing CLIP's 68.3% on ViT-B models. Scaling to 1B data, while maintaining the same training budget, attains 72.4%. Our observations hold across various model sizes, exemplified by ViT-H achieving 80.5%, without any bells-and-whistles. Curation code and training data distribution on metadata is made available at https://github.com/facebookresearch/MetaCLIP.
2023-09-28
{'id': '2309.16671v5', 'title': 'Demystifying CLIP Data', 'abstract': "Contrastive Language-Image Pre-training (CLIP) is an approach that has\nadvanced research and applications in computer vision, fueling modern\nrecognition systems and generative models. We believe that the main ingredient\nto the success of CLIP is its data and not the model architecture or\npre-training objective. However, CLIP only provides very limited information\nabout its data and how it has been collected, leading to works that aim to\nreproduce CLIP's data by filtering with its model parameters. In this work, we\nintend to reveal CLIP's data curation approach and in our pursuit of making it\nopen to the community introduce Metadata-Curated Language-Image Pre-training\n(MetaCLIP). MetaCLIP takes a raw data pool and metadata (derived from CLIP's\nconcepts) and yields a balanced subset over the metadata distribution. Our\nexperimental study rigorously isolates the model and training settings,\nconcentrating solely on data. MetaCLIP applied to CommonCrawl with 400M\nimage-text data pairs outperforms CLIP's data on multiple standard benchmarks.\nIn zero-shot ImageNet classification, MetaCLIP achieves 70.8% accuracy,\nsurpassing CLIP's 68.3% on ViT-B models. Scaling to 1B data, while maintaining\nthe same training budget, attains 72.4%. Our observations hold across various\nmodel sizes, exemplified by ViT-H achieving 80.5%, without any\nbells-and-whistles. Curation code and training data distribution on metadata is\nmade available at https://github.com/facebookresearch/MetaCLIP.", 'authors': ['Hu Xu', 'Saining Xie', 'Xiaoqing Ellen Tan', 'Po-Yao Huang', 'Russell Howes', 'Vasu Sharma', 'Shang-Wen Li', 'Gargi Ghosh', 'Luke Zettlemoyer', 'Christoph Feichtenhofer'], 'published': '2023-09-28 17:59:56+00:00', 'categories': ['cs.CV', 'cs.CL'], 'url': 'http://arxiv.org/abs/2309.16671v5'}
42,447
2501.03132v1
2,501.03132
1
Communication Bounds for the Distributed Experts Problem
In this work, we study the experts problem in the distributed setting where an expert's cost needs to be aggregated across multiple servers. Our study considers various communication models such as the message-passing model and the broadcast model, along with multiple aggregation functions, such as summing and taking the $_p$ norm of an expert's cost across servers. We propose the first communication-efficient protocols that achieve near-optimal regret in these settings, even against a strong adversary who can choose the inputs adaptively. Additionally, we give a conditional lower bound showing that the communication of our protocols is nearly optimal. Finally, we implement our protocols and demonstrate empirical savings on the HPO-B benchmarks.
2025-01-06
{'id': '2501.03132v1', 'title': 'Communication Bounds for the Distributed Experts Problem', 'abstract': "In this work, we study the experts problem in the distributed setting where\nan expert's cost needs to be aggregated across multiple servers. Our study\nconsiders various communication models such as the message-passing model and\nthe broadcast model, along with multiple aggregation functions, such as summing\nand taking the $\\ell_p$ norm of an expert's cost across servers. We propose the\nfirst communication-efficient protocols that achieve near-optimal regret in\nthese settings, even against a strong adversary who can choose the inputs\nadaptively. Additionally, we give a conditional lower bound showing that the\ncommunication of our protocols is nearly optimal. Finally, we implement our\nprotocols and demonstrate empirical savings on the HPO-B benchmarks.", 'authors': ['Zhihao Jia', 'Qi Pang', 'Trung Tran', 'David Woodruff', 'Zhihao Zhang', 'Wenting Zheng'], 'published': '2025-01-06 16:50:35+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2501.03132v1'}
42,414
2210.06891v4
2,210.06891
4
Experimental Design for Multi-Channel Imaging via Task-Driven Feature Selection
This paper presents a data-driven, task-specific paradigm for experimental design, to shorten acquisition time, reduce costs, and accelerate the deployment of imaging devices. Current approaches in experimental design focus on model-parameter estimation and require specification of a particular model, whereas in imaging, other tasks may drive the design. Furthermore, such approaches often lead to intractable optimization problems in real-world imaging applications. Here we present a new paradigm for experimental design that simultaneously optimizes the design (set of image channels) and trains a machine-learning model to execute a user-specified image-analysis task. The approach obtains data densely-sampled over the measurement space (many image channels) for a small number of acquisitions, then identifies a subset of channels of prespecified size that best supports the task. We propose a method: TADRED for TAsk-DRiven Experimental Design in imaging, to identify the most informative channel-subset whilst simultaneously training a network to execute the task given the subset. Experiments demonstrate the potential of TADRED in diverse imaging applications: several clinically-relevant tasks in magnetic resonance imaging; and remote sensing and physiological applications of hyperspectral imaging. Results show substantial improvement over classical experimental design, two recent application-specific methods within the new paradigm, and state-of-the-art approaches in supervised feature selection. We anticipate further applications of our approach. Code is available: https://github.com/sbb-gh/experimental-design-multichannel
2022-10-13
{'id': '2210.06891v4', 'title': 'Experimental Design for Multi-Channel Imaging via Task-Driven Feature Selection', 'abstract': 'This paper presents a data-driven, task-specific paradigm for experimental\ndesign, to shorten acquisition time, reduce costs, and accelerate the\ndeployment of imaging devices. Current approaches in experimental design focus\non model-parameter estimation and require specification of a particular model,\nwhereas in imaging, other tasks may drive the design. Furthermore, such\napproaches often lead to intractable optimization problems in real-world\nimaging applications. Here we present a new paradigm for experimental design\nthat simultaneously optimizes the design (set of image channels) and trains a\nmachine-learning model to execute a user-specified image-analysis task. The\napproach obtains data densely-sampled over the measurement space (many image\nchannels) for a small number of acquisitions, then identifies a subset of\nchannels of prespecified size that best supports the task. We propose a method:\nTADRED for TAsk-DRiven Experimental Design in imaging, to identify the most\ninformative channel-subset whilst simultaneously training a network to execute\nthe task given the subset. Experiments demonstrate the potential of TADRED in\ndiverse imaging applications: several clinically-relevant tasks in magnetic\nresonance imaging; and remote sensing and physiological applications of\nhyperspectral imaging. Results show substantial improvement over classical\nexperimental design, two recent application-specific methods within the new\nparadigm, and state-of-the-art approaches in supervised feature selection. We\nanticipate further applications of our approach. Code is available:\nhttps://github.com/sbb-gh/experimental-design-multichannel', 'authors': ['Stefano B. Blumberg', 'Paddy J. Slator', 'Daniel C. Alexander'], 'published': '2022-10-13 10:36:24+00:00', 'categories': ['cs.LG', 'cs.AI', 'q-bio.NC'], 'url': 'http://arxiv.org/abs/2210.06891v4'}
42,416
2311.11321v3
2,311.11321
3
Bounds on Representation-Induced Confounding Bias for Treatment Effect Estimation
State-of-the-art methods for conditional average treatment effect (CATE) estimation make widespread use of representation learning. Here, the idea is to reduce the variance of the low-sample CATE estimation by a (potentially constrained) low-dimensional representation. However, low-dimensional representations can lose information about the observed confounders and thus lead to bias, because of which the validity of representation learning for CATE estimation is typically violated. In this paper, we propose a new, representation-agnostic refutation framework for estimating bounds on the representation-induced confounding bias that comes from dimensionality reduction (or other constraints on the representations) in CATE estimation. First, we establish theoretically under which conditions CATE is non-identifiable given low-dimensional (constrained) representations. Second, as our remedy, we propose a neural refutation framework which performs partial identification of CATE or, equivalently, aims at estimating lower and upper bounds of the representation-induced confounding bias. We demonstrate the effectiveness of our bounds in a series of experiments. In sum, our refutation framework is of direct relevance in practice where the validity of CATE estimation is of importance.
2023-11-19
{'id': '2311.11321v3', 'title': 'Bounds on Representation-Induced Confounding Bias for Treatment Effect Estimation', 'abstract': 'State-of-the-art methods for conditional average treatment effect (CATE)\nestimation make widespread use of representation learning. Here, the idea is to\nreduce the variance of the low-sample CATE estimation by a (potentially\nconstrained) low-dimensional representation. However, low-dimensional\nrepresentations can lose information about the observed confounders and thus\nlead to bias, because of which the validity of representation learning for CATE\nestimation is typically violated. In this paper, we propose a new,\nrepresentation-agnostic refutation framework for estimating bounds on the\nrepresentation-induced confounding bias that comes from dimensionality\nreduction (or other constraints on the representations) in CATE estimation.\nFirst, we establish theoretically under which conditions CATE is\nnon-identifiable given low-dimensional (constrained) representations. Second,\nas our remedy, we propose a neural refutation framework which performs partial\nidentification of CATE or, equivalently, aims at estimating lower and upper\nbounds of the representation-induced confounding bias. We demonstrate the\neffectiveness of our bounds in a series of experiments. In sum, our refutation\nframework is of direct relevance in practice where the validity of CATE\nestimation is of importance.', 'authors': ['Valentyn Melnychuk', 'Dennis Frauen', 'Stefan Feuerriegel'], 'published': '2023-11-19 13:31:30+00:00', 'categories': ['stat.ML', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2311.11321v3'}
42,418
2309.09814v3
2,309.09814
3
Convolutional Deep Kernel Machines
Standard infinite-width limits of neural networks sacrifice the ability for intermediate layers to learn representations from data. Recent work (A theory of representation learning gives a deep generalisation of kernel methods, Yang et al. 2023) modified the Neural Network Gaussian Process (NNGP) limit of Bayesian neural networks so that representation learning is retained. Furthermore, they found that applying this modified limit to a deep Gaussian process gives a practical learning algorithm which they dubbed the deep kernel machine (DKM). However, they only considered the simplest possible setting: regression in small, fully connected networks with e.g. 10 input features. Here, we introduce convolutional deep kernel machines. This required us to develop a novel inter-domain inducing point approximation, as well as introducing and experimentally assessing a number of techniques not previously seen in DKMs, including analogues to batch normalisation, different likelihoods, and different types of top-layer. The resulting model trains in roughly 77 GPU hours, achieving around 99% test accuracy on MNIST, 72% on CIFAR-100, and 92.7% on CIFAR-10, which is SOTA for kernel methods.
2023-09-18
{'id': '2309.09814v3', 'title': 'Convolutional Deep Kernel Machines', 'abstract': 'Standard infinite-width limits of neural networks sacrifice the ability for\nintermediate layers to learn representations from data. Recent work (A theory\nof representation learning gives a deep generalisation of kernel methods, Yang\net al. 2023) modified the Neural Network Gaussian Process (NNGP) limit of\nBayesian neural networks so that representation learning is retained.\nFurthermore, they found that applying this modified limit to a deep Gaussian\nprocess gives a practical learning algorithm which they dubbed the deep kernel\nmachine (DKM). However, they only considered the simplest possible setting:\nregression in small, fully connected networks with e.g. 10 input features.\nHere, we introduce convolutional deep kernel machines. This required us to\ndevelop a novel inter-domain inducing point approximation, as well as\nintroducing and experimentally assessing a number of techniques not previously\nseen in DKMs, including analogues to batch normalisation, different\nlikelihoods, and different types of top-layer. The resulting model trains in\nroughly 77 GPU hours, achieving around 99% test accuracy on MNIST, 72% on\nCIFAR-100, and 92.7% on CIFAR-10, which is SOTA for kernel methods.', 'authors': ['Edward Milsom', 'Ben Anson', 'Laurence Aitchison'], 'published': '2023-09-18 14:36:17+00:00', 'categories': ['stat.ML', 'cs.LG'], 'url': 'http://arxiv.org/abs/2309.09814v3'}
42,420
2508.00563v1
2,508.00563
1
Weakly Supervised Virus Capsid Detection with Image-Level Annotations in Electron Microscopy Images
Current state-of-the-art methods for object detection rely on annotated bounding boxes of large data sets for training. However, obtaining such annotations is expensive and can require up to hundreds of hours of manual labor. This poses a challenge, especially since such annotations can only be provided by experts, as they require knowledge about the scientific domain. To tackle this challenge, we propose a domain-specific weakly supervised object detection algorithm that only relies on image-level annotations, which are significantly easier to acquire. Our method distills the knowledge of a pre-trained model, on the task of predicting the presence or absence of a virus in an image, to obtain a set of pseudo-labels that can be used to later train a state-of-the-art object detection model. To do so, we use an optimization approach with a shrinking receptive field to extract virus particles directly without specific network architectures. Through a set of extensive studies, we show how the proposed pseudo-labels are easier to obtain, and, more importantly, are able to outperform other existing weak labeling methods, and even ground truth labels, in cases where the time to obtain the annotation is limited.
2025-08-01
{'id': '2508.00563v1', 'title': 'Weakly Supervised Virus Capsid Detection with Image-Level Annotations in Electron Microscopy Images', 'abstract': 'Current state-of-the-art methods for object detection rely on annotated\nbounding boxes of large data sets for training. However, obtaining such\nannotations is expensive and can require up to hundreds of hours of manual\nlabor. This poses a challenge, especially since such annotations can only be\nprovided by experts, as they require knowledge about the scientific domain. To\ntackle this challenge, we propose a domain-specific weakly supervised object\ndetection algorithm that only relies on image-level annotations, which are\nsignificantly easier to acquire. Our method distills the knowledge of a\npre-trained model, on the task of predicting the presence or absence of a virus\nin an image, to obtain a set of pseudo-labels that can be used to later train a\nstate-of-the-art object detection model. To do so, we use an optimization\napproach with a shrinking receptive field to extract virus particles directly\nwithout specific network architectures. Through a set of extensive studies, we\nshow how the proposed pseudo-labels are easier to obtain, and, more\nimportantly, are able to outperform other existing weak labeling methods, and\neven ground truth labels, in cases where the time to obtain the annotation is\nlimited.', 'authors': ['Hannah Kniesel', 'Leon Sick', 'Tristan Payer', 'Tim Bergner', 'Kavitha Shaga Devan', 'Clarissa Read', 'Paul Walther', 'Timo Ropinski'], 'published': '2025-08-01 12:02:25+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2508.00563v1'}
42,433
2402.02340v2
2,402.0234
2
Learning Semantic Proxies from Visual Prompts for Parameter-Efficient Fine-Tuning in Deep Metric Learning
Deep Metric Learning (DML) has long attracted the attention of the machine learning community as a key objective. Existing solutions concentrate on fine-tuning the pre-trained models on conventional image datasets. As a result of the success of recent pre-trained models trained from larger-scale datasets, it is challenging to adapt the model to the DML tasks in the local data domain while retaining the previously gained knowledge. In this paper, we investigate parameter-efficient methods for fine-tuning the pre-trained model for DML tasks. In particular, we propose a novel and effective framework based on learning Visual Prompts (VPT) in the pre-trained Vision Transformers (ViT). Based on the conventional proxy-based DML paradigm, we augment the proxy by incorporating the semantic information from the input image and the ViT, in which we optimize the visual prompts for each class. We demonstrate that our new approximations with semantic information are superior to representative capabilities, thereby improving metric learning performance. We conduct extensive experiments to demonstrate that our proposed framework is effective and efficient by evaluating popular DML benchmarks. In particular, we demonstrate that our fine-tuning method achieves comparable or even better performance than recent state-of-the-art full fine-tuning works of DML while tuning only a small percentage of total parameters.
2024-02-04
{'id': '2402.02340v2', 'title': 'Learning Semantic Proxies from Visual Prompts for Parameter-Efficient Fine-Tuning in Deep Metric Learning', 'abstract': 'Deep Metric Learning (DML) has long attracted the attention of the machine\nlearning community as a key objective. Existing solutions concentrate on\nfine-tuning the pre-trained models on conventional image datasets. As a result\nof the success of recent pre-trained models trained from larger-scale datasets,\nit is challenging to adapt the model to the DML tasks in the local data domain\nwhile retaining the previously gained knowledge. In this paper, we investigate\nparameter-efficient methods for fine-tuning the pre-trained model for DML\ntasks. In particular, we propose a novel and effective framework based on\nlearning Visual Prompts (VPT) in the pre-trained Vision Transformers (ViT).\nBased on the conventional proxy-based DML paradigm, we augment the proxy by\nincorporating the semantic information from the input image and the ViT, in\nwhich we optimize the visual prompts for each class. We demonstrate that our\nnew approximations with semantic information are superior to representative\ncapabilities, thereby improving metric learning performance. We conduct\nextensive experiments to demonstrate that our proposed framework is effective\nand efficient by evaluating popular DML benchmarks. In particular, we\ndemonstrate that our fine-tuning method achieves comparable or even better\nperformance than recent state-of-the-art full fine-tuning works of DML while\ntuning only a small percentage of total parameters.', 'authors': ['Li Ren', 'Chen Chen', 'Liqiang Wang', 'Kien Hua'], 'published': '2024-02-04 04:42:05+00:00', 'categories': ['cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2402.02340v2'}
42,435
2302.05737v3
2,302.05737
3
A Reparameterized Discrete Diffusion Model for Text Generation
This work studies discrete diffusion probabilistic models with applications to natural language generation. We derive an alternative yet equivalent formulation of the sampling from discrete diffusion processes and leverage this insight to develop a family of reparameterized discrete diffusion models. The derived generic framework is highly flexible, offers a fresh perspective of the generation process in discrete diffusion models, and features more effective training and decoding techniques. We conduct extensive experiments to evaluate the text generation capability of our model, demonstrating significant improvements over existing diffusion models.
2023-02-11
{'id': '2302.05737v3', 'title': 'A Reparameterized Discrete Diffusion Model for Text Generation', 'abstract': 'This work studies discrete diffusion probabilistic models with applications\nto natural language generation. We derive an alternative yet equivalent\nformulation of the sampling from discrete diffusion processes and leverage this\ninsight to develop a family of reparameterized discrete diffusion models. The\nderived generic framework is highly flexible, offers a fresh perspective of the\ngeneration process in discrete diffusion models, and features more effective\ntraining and decoding techniques. We conduct extensive experiments to evaluate\nthe text generation capability of our model, demonstrating significant\nimprovements over existing diffusion models.', 'authors': ['Lin Zheng', 'Jianbo Yuan', 'Lei Yu', 'Lingpeng Kong'], 'published': '2023-02-11 16:26:57+00:00', 'categories': ['cs.CL', 'cs.LG'], 'url': 'http://arxiv.org/abs/2302.05737v3'}
42,419
2212.02582v1
2,212.02582
1
Rethinking Backdoor Data Poisoning Attacks in the Context of Semi-Supervised Learning
Semi-supervised learning methods can train high-accuracy machine learning models with a fraction of the labeled training samples required for traditional supervised learning. Such methods do not typically involve close review of the unlabeled training samples, making them tempting targets for data poisoning attacks. In this paper we investigate the vulnerabilities of semi-supervised learning methods to backdoor data poisoning attacks on the unlabeled samples. We show that simple poisoning attacks that influence the distribution of the poisoned samples' predicted labels are highly effective - achieving an average attack success rate as high as 96.9%. We introduce a generalized attack framework targeting semi-supervised learning methods to better understand and exploit their limitations and to motivate future defense strategies.
2022-12-05
{'id': '2212.02582v1', 'title': 'Rethinking Backdoor Data Poisoning Attacks in the Context of Semi-Supervised Learning', 'abstract': "Semi-supervised learning methods can train high-accuracy machine learning\nmodels with a fraction of the labeled training samples required for traditional\nsupervised learning. Such methods do not typically involve close review of the\nunlabeled training samples, making them tempting targets for data poisoning\nattacks. In this paper we investigate the vulnerabilities of semi-supervised\nlearning methods to backdoor data poisoning attacks on the unlabeled samples.\nWe show that simple poisoning attacks that influence the distribution of the\npoisoned samples' predicted labels are highly effective - achieving an average\nattack success rate as high as 96.9%. We introduce a generalized attack\nframework targeting semi-supervised learning methods to better understand and\nexploit their limitations and to motivate future defense strategies.", 'authors': ['Marissa Connor', 'Vincent Emanuele'], 'published': '2022-12-05 20:21:31+00:00', 'categories': ['cs.LG', 'cs.CR'], 'url': 'http://arxiv.org/abs/2212.02582v1'}
42,421
2212.04461v1
2,212.04461
1
Leveraging Unlabeled Data to Track Memorization
Deep neural networks may easily memorize noisy labels present in real-world data, which degrades their ability to generalize. It is therefore important to track and evaluate the robustness of models against noisy label memorization. We propose a metric, called susceptibility, to gauge such memorization for neural networks. Susceptibility is simple and easy to compute during training. Moreover, it does not require access to ground-truth labels and it only uses unlabeled data. We empirically show the effectiveness of our metric in tracking memorization on various architectures and datasets and provide theoretical insights into the design of the susceptibility metric. Finally, we show through extensive experiments on datasets with synthetic and real-world label noise that one can utilize susceptibility and the overall training accuracy to distinguish models that maintain a low memorization on the training set and generalize well to unseen clean data.
2022-12-08
{'id': '2212.04461v1', 'title': 'Leveraging Unlabeled Data to Track Memorization', 'abstract': 'Deep neural networks may easily memorize noisy labels present in real-world\ndata, which degrades their ability to generalize. It is therefore important to\ntrack and evaluate the robustness of models against noisy label memorization.\nWe propose a metric, called susceptibility, to gauge such memorization for\nneural networks. Susceptibility is simple and easy to compute during training.\nMoreover, it does not require access to ground-truth labels and it only uses\nunlabeled data. We empirically show the effectiveness of our metric in tracking\nmemorization on various architectures and datasets and provide theoretical\ninsights into the design of the susceptibility metric. Finally, we show through\nextensive experiments on datasets with synthetic and real-world label noise\nthat one can utilize susceptibility and the overall training accuracy to\ndistinguish models that maintain a low memorization on the training set and\ngeneralize well to unseen clean data.', 'authors': ['Mahsa Forouzesh', 'Hanie Sedghi', 'Patrick Thiran'], 'published': '2022-12-08 18:36:41+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2212.04461v1'}
42,442
2211.08422v3
2,211.08422
3
Mechanistic Mode Connectivity
We study neural network loss landscapes through the lens of mode connectivity, the observation that minimizers of neural networks retrieved via training on a dataset are connected via simple paths of low loss. Specifically, we ask the following question: are minimizers that rely on different mechanisms for making their predictions connected via simple paths of low loss? We provide a definition of mechanistic similarity as shared invariances to input transformations and demonstrate that lack of linear connectivity between two models implies they use dissimilar mechanisms for making their predictions. Relevant to practice, this result helps us demonstrate that naive fine-tuning on a downstream dataset can fail to alter a model's mechanisms, e.g., fine-tuning can fail to eliminate a model's reliance on spurious attributes. Our analysis also motivates a method for targeted alteration of a model's mechanisms, named connectivity-based fine-tuning (CBFT), which we analyze using several synthetic datasets for the task of reducing a model's reliance on spurious attributes.
2022-11-15
{'id': '2211.08422v3', 'title': 'Mechanistic Mode Connectivity', 'abstract': "We study neural network loss landscapes through the lens of mode\nconnectivity, the observation that minimizers of neural networks retrieved via\ntraining on a dataset are connected via simple paths of low loss. Specifically,\nwe ask the following question: are minimizers that rely on different mechanisms\nfor making their predictions connected via simple paths of low loss? We provide\na definition of mechanistic similarity as shared invariances to input\ntransformations and demonstrate that lack of linear connectivity between two\nmodels implies they use dissimilar mechanisms for making their predictions.\nRelevant to practice, this result helps us demonstrate that naive fine-tuning\non a downstream dataset can fail to alter a model's mechanisms, e.g.,\nfine-tuning can fail to eliminate a model's reliance on spurious attributes.\nOur analysis also motivates a method for targeted alteration of a model's\nmechanisms, named connectivity-based fine-tuning (CBFT), which we analyze using\nseveral synthetic datasets for the task of reducing a model's reliance on\nspurious attributes.", 'authors': ['Ekdeep Singh Lubana', 'Eric J. Bigelow', 'Robert P. Dick', 'David Krueger', 'Hidenori Tanaka'], 'published': '2022-11-15 18:58:28+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2211.08422v3'}
42,448
2303.00326v1
2,303.00326
1
Empowering Networks With Scale and Rotation Equivariance Using A Similarity Convolution
The translational equivariant nature of Convolutional Neural Networks (CNNs) is a reason for its great success in computer vision. However, networks do not enjoy more general equivariance properties such as rotation or scaling, ultimately limiting their generalization performance. To address this limitation, we devise a method that endows CNNs with simultaneous equivariance with respect to translation, rotation, and scaling. Our approach defines a convolution-like operation and ensures equivariance based on our proposed scalable Fourier-Argand representation. The method maintains similar efficiency as a traditional network and hardly introduces any additional learnable parameters, since it does not face the computational issue that often occurs in group-convolution operators. We validate the efficacy of our approach in the image classification task, demonstrating its robustness and the generalization ability to both scaled and rotated inputs.
2023-03-01
{'id': '2303.00326v1', 'title': 'Empowering Networks With Scale and Rotation Equivariance Using A Similarity Convolution', 'abstract': 'The translational equivariant nature of Convolutional Neural Networks (CNNs)\nis a reason for its great success in computer vision. However, networks do not\nenjoy more general equivariance properties such as rotation or scaling,\nultimately limiting their generalization performance. To address this\nlimitation, we devise a method that endows CNNs with simultaneous equivariance\nwith respect to translation, rotation, and scaling. Our approach defines a\nconvolution-like operation and ensures equivariance based on our proposed\nscalable Fourier-Argand representation. The method maintains similar efficiency\nas a traditional network and hardly introduces any additional learnable\nparameters, since it does not face the computational issue that often occurs in\ngroup-convolution operators. We validate the efficacy of our approach in the\nimage classification task, demonstrating its robustness and the generalization\nability to both scaled and rotated inputs.', 'authors': ['Zikai Sun', 'Thierry Blu'], 'published': '2023-03-01 08:43:05+00:00', 'categories': ['cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2303.00326v1'}
42,451
2302.09347v1
2,302.09347
1
Closed-Loop Transcription via Convolutional Sparse Coding
Autoencoding has achieved great empirical success as a framework for learning generative models for natural images. Autoencoders often use generic deep networks as the encoder or decoder, which are difficult to interpret, and the learned representations lack clear structure. In this work, we make the explicit assumption that the image distribution is generated from a multi-stage sparse deconvolution. The corresponding inverse map, which we use as an encoder, is a multi-stage convolution sparse coding (CSC), with each stage obtained from unrolling an optimization algorithm for solving the corresponding (convexified) sparse coding program. To avoid computational difficulties in minimizing distributional distance between the real and generated images, we utilize the recent closed-loop transcription (CTRL) framework that optimizes the rate reduction of the learned sparse representations. Conceptually, our method has high-level connections to score-matching methods such as diffusion models. Empirically, our framework demonstrates competitive performance on large-scale datasets, such as ImageNet-1K, compared to existing autoencoding and generative methods under fair conditions. Even with simpler networks and fewer computational resources, our method demonstrates high visual quality in regenerated images. More surprisingly, the learned autoencoder performs well on unseen datasets. Our method enjoys several side benefits, including more structured and interpretable representations, more stable convergence, and scalability to large datasets. Our method is arguably the first to demonstrate that a concatenation of multiple convolution sparse coding/decoding layers leads to an interpretable and effective autoencoder for modeling the distribution of large-scale natural image datasets.
2023-02-18
{'id': '2302.09347v1', 'title': 'Closed-Loop Transcription via Convolutional Sparse Coding', 'abstract': 'Autoencoding has achieved great empirical success as a framework for learning\ngenerative models for natural images. Autoencoders often use generic deep\nnetworks as the encoder or decoder, which are difficult to interpret, and the\nlearned representations lack clear structure. In this work, we make the\nexplicit assumption that the image distribution is generated from a multi-stage\nsparse deconvolution. The corresponding inverse map, which we use as an\nencoder, is a multi-stage convolution sparse coding (CSC), with each stage\nobtained from unrolling an optimization algorithm for solving the corresponding\n(convexified) sparse coding program. To avoid computational difficulties in\nminimizing distributional distance between the real and generated images, we\nutilize the recent closed-loop transcription (CTRL) framework that optimizes\nthe rate reduction of the learned sparse representations. Conceptually, our\nmethod has high-level connections to score-matching methods such as diffusion\nmodels. Empirically, our framework demonstrates competitive performance on\nlarge-scale datasets, such as ImageNet-1K, compared to existing autoencoding\nand generative methods under fair conditions. Even with simpler networks and\nfewer computational resources, our method demonstrates high visual quality in\nregenerated images. More surprisingly, the learned autoencoder performs well on\nunseen datasets. Our method enjoys several side benefits, including more\nstructured and interpretable representations, more stable convergence, and\nscalability to large datasets. Our method is arguably the first to demonstrate\nthat a concatenation of multiple convolution sparse coding/decoding layers\nleads to an interpretable and effective autoencoder for modeling the\ndistribution of large-scale natural image datasets.', 'authors': ['Xili Dai', 'Ke Chen', 'Shengbang Tong', 'Jingyuan Zhang', 'Xingjian Gao', 'Mingyang Li', 'Druv Pai', 'Yuexiang Zhai', 'XIaojun Yuan', 'Heung-Yeung Shum', 'Lionel M. Ni', 'Yi Ma'], 'published': '2023-02-18 14:40:07+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2302.09347v1'}
42,428
2307.07107v2
2,307.07107
2
Graph Positional and Structural Encoder
Positional and structural encodings (PSE) enable better identifiability of nodes within a graph, rendering them essential tools for empowering modern GNNs, and in particular graph Transformers. However, designing PSEs that work optimally for all graph prediction tasks is a challenging and unsolved problem. Here, we present the Graph Positional and Structural Encoder (GPSE), the first-ever graph encoder designed to capture rich PSE representations for augmenting any GNN. GPSE learns an efficient common latent representation for multiple PSEs, and is highly transferable: The encoder trained on a particular graph dataset can be used effectively on datasets drawn from markedly different distributions and modalities. We show that across a wide range of benchmarks, GPSE-enhanced models can significantly outperform those that employ explicitly computed PSEs, and at least match their performance in others. Our results pave the way for the development of foundational pre-trained graph encoders for extracting positional and structural information, and highlight their potential as a more powerful and efficient alternative to explicitly computed PSEs and existing self-supervised pre-training approaches. Our framework and pre-trained models are publicly available at https://github.com/G-Taxonomy-Workgroup/GPSE. For convenience, GPSE has also been integrated into the PyG library to facilitate downstream applications.
2023-07-14
{'id': '2307.07107v2', 'title': 'Graph Positional and Structural Encoder', 'abstract': 'Positional and structural encodings (PSE) enable better identifiability of\nnodes within a graph, rendering them essential tools for empowering modern\nGNNs, and in particular graph Transformers. However, designing PSEs that work\noptimally for all graph prediction tasks is a challenging and unsolved problem.\nHere, we present the Graph Positional and Structural Encoder (GPSE), the\nfirst-ever graph encoder designed to capture rich PSE representations for\naugmenting any GNN. GPSE learns an efficient common latent representation for\nmultiple PSEs, and is highly transferable: The encoder trained on a particular\ngraph dataset can be used effectively on datasets drawn from markedly different\ndistributions and modalities. We show that across a wide range of benchmarks,\nGPSE-enhanced models can significantly outperform those that employ explicitly\ncomputed PSEs, and at least match their performance in others. Our results pave\nthe way for the development of foundational pre-trained graph encoders for\nextracting positional and structural information, and highlight their potential\nas a more powerful and efficient alternative to explicitly computed PSEs and\nexisting self-supervised pre-training approaches. Our framework and pre-trained\nmodels are publicly available at https://github.com/G-Taxonomy-Workgroup/GPSE.\nFor convenience, GPSE has also been integrated into the PyG library to\nfacilitate downstream applications.', 'authors': ['Semih Cantürk', 'Renming Liu', 'Olivier Lapointe-Gagné', 'Vincent Létourneau', 'Guy Wolf', 'Dominique Beaini', 'Ladislav Rampášek'], 'published': '2023-07-14 01:04:18+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2307.07107v2'}
42,429
2406.18562v1
2,406.18562
1
Views Can Be Deceiving: Improved SSL Through Feature Space Augmentation
Supervised learning methods have been found to exhibit inductive biases favoring simpler features. When such features are spuriously correlated with the label, this can result in suboptimal performance on minority subgroups. Despite the growing popularity of methods which learn from unlabeled data, the extent to which these representations rely on spurious features for prediction is unclear. In this work, we explore the impact of spurious features on Self-Supervised Learning (SSL) for visual representation learning. We first empirically show that commonly used augmentations in SSL can cause undesired invariances in the image space, and illustrate this with a simple example. We further show that classical approaches in combating spurious correlations, such as dataset re-sampling during SSL, do not consistently lead to invariant representations. Motivated by these findings, we propose LateTVG to remove spurious information from these representations during pre-training, by regularizing later layers of the encoder via pruning. We find that our method produces representations which outperform the baselines on several benchmarks, without the need for group or label information during SSL.
2024-05-28
{'id': '2406.18562v1', 'title': 'Views Can Be Deceiving: Improved SSL Through Feature Space Augmentation', 'abstract': 'Supervised learning methods have been found to exhibit inductive biases\nfavoring simpler features. When such features are spuriously correlated with\nthe label, this can result in suboptimal performance on minority subgroups.\nDespite the growing popularity of methods which learn from unlabeled data, the\nextent to which these representations rely on spurious features for prediction\nis unclear. In this work, we explore the impact of spurious features on\nSelf-Supervised Learning (SSL) for visual representation learning. We first\nempirically show that commonly used augmentations in SSL can cause undesired\ninvariances in the image space, and illustrate this with a simple example. We\nfurther show that classical approaches in combating spurious correlations, such\nas dataset re-sampling during SSL, do not consistently lead to invariant\nrepresentations. Motivated by these findings, we propose LateTVG to remove\nspurious information from these representations during pre-training, by\nregularizing later layers of the encoder via pruning. We find that our method\nproduces representations which outperform the baselines on several benchmarks,\nwithout the need for group or label information during SSL.', 'authors': ['Kimia Hamidieh', 'Haoran Zhang', 'Swami Sankaranarayanan', 'Marzyeh Ghassemi'], 'published': '2024-05-28 18:42:13+00:00', 'categories': ['cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2406.18562v1'}
42,431
2407.03442v1
2,407.03442
1
Fisher-aware Quantization for DETR Detectors with Critical-category Objectives
The impact of quantization on the overall performance of deep learning models is a well-studied problem. However, understanding and mitigating its effects on a more fine-grained level is still lacking, especially for harder tasks such as object detection with both classification and regression objectives. This work defines the performance for a subset of task-critical categories, i.e. the critical-category performance, as a crucial yet largely overlooked fine-grained objective for detection tasks. We analyze the impact of quantization at the category-level granularity, and propose methods to improve performance for the critical categories. Specifically, we find that certain critical categories have a higher sensitivity to quantization, and are prone to overfitting after quantization-aware training (QAT). To explain this, we provide theoretical and empirical links between their performance gaps and the corresponding loss landscapes with the Fisher information framework. Using this evidence, we apply a Fisher-aware mixed-precision quantization scheme, and a Fisher-trace regularization for the QAT on the critical-category loss landscape. The proposed methods improve critical-category metrics of the quantized transformer-based DETR detectors. They are even more significant in case of larger models and higher number of classes where the overfitting becomes more severe. For example, our methods lead to 10.4% and 14.5% mAP gains for, correspondingly, 4-bit DETR-R50 and Deformable DETR on the most impacted critical classes in the COCO Panoptic dataset.
2024-07-03
{'id': '2407.03442v1', 'title': 'Fisher-aware Quantization for DETR Detectors with Critical-category Objectives', 'abstract': 'The impact of quantization on the overall performance of deep learning models\nis a well-studied problem. However, understanding and mitigating its effects on\na more fine-grained level is still lacking, especially for harder tasks such as\nobject detection with both classification and regression objectives. This work\ndefines the performance for a subset of task-critical categories, i.e. the\ncritical-category performance, as a crucial yet largely overlooked fine-grained\nobjective for detection tasks. We analyze the impact of quantization at the\ncategory-level granularity, and propose methods to improve performance for the\ncritical categories. Specifically, we find that certain critical categories\nhave a higher sensitivity to quantization, and are prone to overfitting after\nquantization-aware training (QAT). To explain this, we provide theoretical and\nempirical links between their performance gaps and the corresponding loss\nlandscapes with the Fisher information framework. Using this evidence, we apply\na Fisher-aware mixed-precision quantization scheme, and a Fisher-trace\nregularization for the QAT on the critical-category loss landscape. The\nproposed methods improve critical-category metrics of the quantized\ntransformer-based DETR detectors. They are even more significant in case of\nlarger models and higher number of classes where the overfitting becomes more\nsevere. For example, our methods lead to 10.4% and 14.5% mAP gains for,\ncorrespondingly, 4-bit DETR-R50 and Deformable DETR on the most impacted\ncritical classes in the COCO Panoptic dataset.', 'authors': ['Huanrui Yang', 'Yafeng Huang', 'Zhen Dong', 'Denis A Gudovskiy', 'Tomoyuki Okuno', 'Yohei Nakata', 'Yuan Du', 'Kurt Keutzer', 'Shanghang Zhang'], 'published': '2024-07-03 18:35:53+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2407.03442v1'}
42,441
2310.00873v2
2,310.00873
2
Deep Neural Networks Tend To Extrapolate Predictably
Conventional wisdom suggests that neural network predictions tend to be unpredictable and overconfident when faced with out-of-distribution (OOD) inputs. Our work reassesses this assumption for neural networks with high-dimensional inputs. Rather than extrapolating in arbitrary ways, we observe that neural network predictions often tend towards a constant value as input data becomes increasingly OOD. Moreover, we find that this value often closely approximates the optimal constant solution (OCS), i.e., the prediction that minimizes the average loss over the training data without observing the input. We present results showing this phenomenon across 8 datasets with different distributional shifts (including CIFAR10-C and ImageNet-R, S), different loss functions (cross entropy, MSE, and Gaussian NLL), and different architectures (CNNs and transformers). Furthermore, we present an explanation for this behavior, which we first validate empirically and then study theoretically in a simplified setting involving deep homogeneous networks with ReLU activations. Finally, we show how one can leverage our insights in practice to enable risk-sensitive decision-making in the presence of OOD inputs.
2023-10-02
{'id': '2310.00873v2', 'title': 'Deep Neural Networks Tend To Extrapolate Predictably', 'abstract': 'Conventional wisdom suggests that neural network predictions tend to be\nunpredictable and overconfident when faced with out-of-distribution (OOD)\ninputs. Our work reassesses this assumption for neural networks with\nhigh-dimensional inputs. Rather than extrapolating in arbitrary ways, we\nobserve that neural network predictions often tend towards a constant value as\ninput data becomes increasingly OOD. Moreover, we find that this value often\nclosely approximates the optimal constant solution (OCS), i.e., the prediction\nthat minimizes the average loss over the training data without observing the\ninput. We present results showing this phenomenon across 8 datasets with\ndifferent distributional shifts (including CIFAR10-C and ImageNet-R, S),\ndifferent loss functions (cross entropy, MSE, and Gaussian NLL), and different\narchitectures (CNNs and transformers). Furthermore, we present an explanation\nfor this behavior, which we first validate empirically and then study\ntheoretically in a simplified setting involving deep homogeneous networks with\nReLU activations. Finally, we show how one can leverage our insights in\npractice to enable risk-sensitive decision-making in the presence of OOD\ninputs.', 'authors': ['Katie Kang', 'Amrith Setlur', 'Claire Tomlin', 'Sergey Levine'], 'published': '2023-10-02 03:25:32+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.00873v2'}
42,459
2309.06599v1
2,309.06599
1
Reasoning with Latent Diffusion in Offline Reinforcement Learning
Offline reinforcement learning (RL) holds promise as a means to learn high-reward policies from a static dataset, without the need for further environment interactions. However, a key challenge in offline RL lies in effectively stitching portions of suboptimal trajectories from the static dataset while avoiding extrapolation errors arising due to a lack of support in the dataset. Existing approaches use conservative methods that are tricky to tune and struggle with multi-modal data (as we show) or rely on noisy Monte Carlo return-to-go samples for reward conditioning. In this work, we propose a novel approach that leverages the expressiveness of latent diffusion to model in-support trajectory sequences as compressed latent skills. This facilitates learning a Q-function while avoiding extrapolation error via batch-constraining. The latent space is also expressive and gracefully copes with multi-modal data. We show that the learned temporally-abstract latent space encodes richer task-specific information for offline RL tasks as compared to raw state-actions. This improves credit assignment and facilitates faster reward propagation during Q-learning. Our method demonstrates state-of-the-art performance on the D4RL benchmarks, particularly excelling in long-horizon, sparse-reward tasks.
2023-09-12
{'id': '2309.06599v1', 'title': 'Reasoning with Latent Diffusion in Offline Reinforcement Learning', 'abstract': 'Offline reinforcement learning (RL) holds promise as a means to learn\nhigh-reward policies from a static dataset, without the need for further\nenvironment interactions. However, a key challenge in offline RL lies in\neffectively stitching portions of suboptimal trajectories from the static\ndataset while avoiding extrapolation errors arising due to a lack of support in\nthe dataset. Existing approaches use conservative methods that are tricky to\ntune and struggle with multi-modal data (as we show) or rely on noisy Monte\nCarlo return-to-go samples for reward conditioning. In this work, we propose a\nnovel approach that leverages the expressiveness of latent diffusion to model\nin-support trajectory sequences as compressed latent skills. This facilitates\nlearning a Q-function while avoiding extrapolation error via\nbatch-constraining. The latent space is also expressive and gracefully copes\nwith multi-modal data. We show that the learned temporally-abstract latent\nspace encodes richer task-specific information for offline RL tasks as compared\nto raw state-actions. This improves credit assignment and facilitates faster\nreward propagation during Q-learning. Our method demonstrates state-of-the-art\nperformance on the D4RL benchmarks, particularly excelling in long-horizon,\nsparse-reward tasks.', 'authors': ['Siddarth Venkatraman', 'Shivesh Khaitan', 'Ravi Tej Akella', 'John Dolan', 'Jeff Schneider', 'Glen Berseth'], 'published': '2023-09-12 20:58:21+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2309.06599v1'}
42,461
2311.11822v1
2,311.11822
1
Zero redundancy distributed learning with differential privacy
Deep learning using large models have achieved great success in a wide range of domains. However, training these models on billions of parameters is very challenging in terms of the training speed, memory cost, and communication efficiency, especially under the privacy-preserving regime with differential privacy (DP). On the one hand, DP optimization has comparable efficiency to the standard non-private optimization on a single GPU, but on multiple GPUs, existing DP distributed learning (such as pipeline parallel) has suffered from significantly worse efficiency. On the other hand, the Zero Redundancy Optimizer (ZeRO) is a state-of-the-art solution to the standard distributed learning, exhibiting excellent training efficiency on large models, but to work compatibly with DP is technically complicated. In this work, we develop a new systematic solution, DP-ZeRO, (I) to scale up the trainable DP model size, e.g. to GPT-100B, (II) to obtain the same computation and communication efficiency as the standard ZeRO, and (III) to enable mixed-precision DP training. Our DP-ZeRO, like the standard ZeRO, has the potential to train models with arbitrary size and is evaluated on the world's largest DP models in terms of the number of trainable parameters.
2023-11-20
{'id': '2311.11822v1', 'title': 'Zero redundancy distributed learning with differential privacy', 'abstract': "Deep learning using large models have achieved great success in a wide range\nof domains. However, training these models on billions of parameters is very\nchallenging in terms of the training speed, memory cost, and communication\nefficiency, especially under the privacy-preserving regime with differential\nprivacy (DP). On the one hand, DP optimization has comparable efficiency to the\nstandard non-private optimization on a single GPU, but on multiple GPUs,\nexisting DP distributed learning (such as pipeline parallel) has suffered from\nsignificantly worse efficiency. On the other hand, the Zero Redundancy\nOptimizer (ZeRO) is a state-of-the-art solution to the standard distributed\nlearning, exhibiting excellent training efficiency on large models, but to work\ncompatibly with DP is technically complicated. In this work, we develop a new\nsystematic solution, DP-ZeRO, (I) to scale up the trainable DP model size, e.g.\nto GPT-100B, (II) to obtain the same computation and communication efficiency\nas the standard ZeRO, and (III) to enable mixed-precision DP training. Our\nDP-ZeRO, like the standard ZeRO, has the potential to train models with\narbitrary size and is evaluated on the world's largest DP models in terms of\nthe number of trainable parameters.", 'authors': ['Zhiqi Bu', 'Justin Chiu', 'Ruixuan Liu', 'Sheng Zha', 'George Karypis'], 'published': '2023-11-20 14:58:56+00:00', 'categories': ['cs.LG', 'cs.CC', 'cs.CR', 'cs.DC'], 'url': 'http://arxiv.org/abs/2311.11822v1'}
42,432
2203.13457v2
2,203.13457
2
Chaos is a Ladder: A New Theoretical Understanding of Contrastive Learning via Augmentation Overlap
Recently, contrastive learning has risen to be a promising approach for large-scale self-supervised learning. However, theoretical understanding of how it works is still unclear. In this paper, we propose a new guarantee on the downstream performance without resorting to the conditional independence assumption that is widely adopted in previous work but hardly holds in practice. Our new theory hinges on the insight that the support of different intra-class samples will become more overlapped under aggressive data augmentations, thus simply aligning the positive samples (augmented views of the same sample) could make contrastive learning cluster intra-class samples together. Based on this augmentation overlap perspective, theoretically, we obtain asymptotically closed bounds for downstream performance under weaker assumptions, and empirically, we propose an unsupervised model selection metric ARC that aligns well with downstream accuracy. Our theory suggests an alternative understanding of contrastive learning: the role of aligning positive samples is more like a surrogate task than an ultimate goal, and the overlapped augmented views (i.e., the chaos) create a ladder for contrastive learning to gradually learn class-separated representations. The code for computing ARC is available at https://github.com/zhangq327/ARC.
2022-03-25
{'id': '2203.13457v2', 'title': 'Chaos is a Ladder: A New Theoretical Understanding of Contrastive Learning via Augmentation Overlap', 'abstract': 'Recently, contrastive learning has risen to be a promising approach for\nlarge-scale self-supervised learning. However, theoretical understanding of how\nit works is still unclear. In this paper, we propose a new guarantee on the\ndownstream performance without resorting to the conditional independence\nassumption that is widely adopted in previous work but hardly holds in\npractice. Our new theory hinges on the insight that the support of different\nintra-class samples will become more overlapped under aggressive data\naugmentations, thus simply aligning the positive samples (augmented views of\nthe same sample) could make contrastive learning cluster intra-class samples\ntogether. Based on this augmentation overlap perspective, theoretically, we\nobtain asymptotically closed bounds for downstream performance under weaker\nassumptions, and empirically, we propose an unsupervised model selection metric\nARC that aligns well with downstream accuracy. Our theory suggests an\nalternative understanding of contrastive learning: the role of aligning\npositive samples is more like a surrogate task than an ultimate goal, and the\noverlapped augmented views (i.e., the chaos) create a ladder for contrastive\nlearning to gradually learn class-separated representations. The code for\ncomputing ARC is available at https://github.com/zhangq327/ARC.', 'authors': ['Yifei Wang', 'Qi Zhang', 'Yisen Wang', 'Jiansheng Yang', 'Zhouchen Lin'], 'published': '2022-03-25 05:36:26+00:00', 'categories': ['cs.LG', 'cs.CV', 'stat.ML'], 'url': 'http://arxiv.org/abs/2203.13457v2'}
42,438
2205.13152v1
2,205.13152
1
Transferable Adversarial Attack based on Integrated Gradients
The vulnerability of deep neural networks to adversarial examples has drawn tremendous attention from the community. Three approaches, optimizing standard objective functions, exploiting attention maps, and smoothing decision surfaces, are commonly used to craft adversarial examples. By tightly integrating the three approaches, we propose a new and simple algorithm named Transferable Attack based on Integrated Gradients (TAIG) in this paper, which can find highly transferable adversarial examples for black-box attacks. Unlike previous methods using multiple computational terms or combining with other methods, TAIG integrates the three approaches into one single term. Two versions of TAIG that compute their integrated gradients on a straight-line path and a random piecewise linear path are studied. Both versions offer strong transferability and can seamlessly work together with the previous methods. Experimental results demonstrate that TAIG outperforms the state-of-the-art methods. The code will available at https://github.com/yihuang2016/TAIG
2022-05-26
{'id': '2205.13152v1', 'title': 'Transferable Adversarial Attack based on Integrated Gradients', 'abstract': 'The vulnerability of deep neural networks to adversarial examples has drawn\ntremendous attention from the community. Three approaches, optimizing standard\nobjective functions, exploiting attention maps, and smoothing decision\nsurfaces, are commonly used to craft adversarial examples. By tightly\nintegrating the three approaches, we propose a new and simple algorithm named\nTransferable Attack based on Integrated Gradients (TAIG) in this paper, which\ncan find highly transferable adversarial examples for black-box attacks. Unlike\nprevious methods using multiple computational terms or combining with other\nmethods, TAIG integrates the three approaches into one single term. Two\nversions of TAIG that compute their integrated gradients on a straight-line\npath and a random piecewise linear path are studied. Both versions offer strong\ntransferability and can seamlessly work together with the previous methods.\nExperimental results demonstrate that TAIG outperforms the state-of-the-art\nmethods. The code will available at https://github.com/yihuang2016/TAIG', 'authors': ['Yi Huang', 'Adams Wai-Kin Kong'], 'published': '2022-05-26 04:59:28+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2205.13152v1'}
42,474
2110.05841v1
2,110.05841
1
Relative Molecule Self-Attention Transformer
Self-supervised learning holds promise to revolutionize molecule property prediction - a central task to drug discovery and many more industries - by enabling data efficient learning from scarce experimental data. Despite significant progress, non-pretrained methods can be still competitive in certain settings. We reason that architecture might be a key bottleneck. In particular, enriching the backbone architecture with domain-specific inductive biases has been key for the success of self-supervised learning in other domains. In this spirit, we methodologically explore the design space of the self-attention mechanism tailored to molecular data. We identify a novel variant of self-attention adapted to processing molecules, inspired by the relative self-attention layer, which involves fusing embedded graph and distance relationships between atoms. Our main contribution is Relative Molecule Attention Transformer (R-MAT): a novel Transformer-based model based on the developed self-attention layer that achieves state-of-the-art or very competitive results across a~wide range of molecule property prediction tasks.
2021-10-12
{'id': '2110.05841v1', 'title': 'Relative Molecule Self-Attention Transformer', 'abstract': 'Self-supervised learning holds promise to revolutionize molecule property\nprediction - a central task to drug discovery and many more industries - by\nenabling data efficient learning from scarce experimental data. Despite\nsignificant progress, non-pretrained methods can be still competitive in\ncertain settings. We reason that architecture might be a key bottleneck. In\nparticular, enriching the backbone architecture with domain-specific inductive\nbiases has been key for the success of self-supervised learning in other\ndomains. In this spirit, we methodologically explore the design space of the\nself-attention mechanism tailored to molecular data. We identify a novel\nvariant of self-attention adapted to processing molecules, inspired by the\nrelative self-attention layer, which involves fusing embedded graph and\ndistance relationships between atoms. Our main contribution is Relative\nMolecule Attention Transformer (R-MAT): a novel Transformer-based model based\non the developed self-attention layer that achieves state-of-the-art or very\ncompetitive results across a~wide range of molecule property prediction tasks.', 'authors': ['Łukasz Maziarka', 'Dawid Majchrowski', 'Tomasz Danel', 'Piotr Gaiński', 'Jacek Tabor', 'Igor Podolak', 'Paweł Morkisz', 'Stanisław Jastrzębski'], 'published': '2021-10-12 09:05:26+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2110.05841v1'}
42,484
2110.06207v2
2,110.06207
2
Open-Set Recognition: a Good Closed-Set Classifier is All You Need?
The ability to identify whether or not a test sample belongs to one of the semantic classes in a classifier's training set is critical to practical deployment of the model. This task is termed open-set recognition (OSR) and has received significant attention in recent years. In this paper, we first demonstrate that the ability of a classifier to make the 'none-of-above' decision is highly correlated with its accuracy on the closed-set classes. We find that this relationship holds across loss objectives and architectures, and further demonstrate the trend both on the standard OSR benchmarks as well as on a large-scale ImageNet evaluation. Second, we use this correlation to boost the performance of a maximum logit score OSR 'baseline' by improving its closed-set accuracy, and with this strong baseline achieve state-of-the-art on a number of OSR benchmarks. Similarly, we boost the performance of the existing state-of-the-art method by improving its closed-set accuracy, but the resulting discrepancy with the strong baseline is marginal. Our third contribution is to present the 'Semantic Shift Benchmark' (SSB), which better respects the task of detecting semantic novelty, in contrast to other forms of distribution shift also considered in related sub-fields, such as out-of-distribution detection. On this new evaluation, we again demonstrate that there is negligible difference between the strong baseline and the existing state-of-the-art. Project Page: https://www.robots.ox.ac.uk/~vgg/research/osr/
2021-10-12
{'id': '2110.06207v2', 'title': 'Open-Set Recognition: a Good Closed-Set Classifier is All You Need?', 'abstract': "The ability to identify whether or not a test sample belongs to one of the\nsemantic classes in a classifier's training set is critical to practical\ndeployment of the model. This task is termed open-set recognition (OSR) and has\nreceived significant attention in recent years. In this paper, we first\ndemonstrate that the ability of a classifier to make the 'none-of-above'\ndecision is highly correlated with its accuracy on the closed-set classes. We\nfind that this relationship holds across loss objectives and architectures, and\nfurther demonstrate the trend both on the standard OSR benchmarks as well as on\na large-scale ImageNet evaluation. Second, we use this correlation to boost the\nperformance of a maximum logit score OSR 'baseline' by improving its closed-set\naccuracy, and with this strong baseline achieve state-of-the-art on a number of\nOSR benchmarks. Similarly, we boost the performance of the existing\nstate-of-the-art method by improving its closed-set accuracy, but the resulting\ndiscrepancy with the strong baseline is marginal. Our third contribution is to\npresent the 'Semantic Shift Benchmark' (SSB), which better respects the task of\ndetecting semantic novelty, in contrast to other forms of distribution shift\nalso considered in related sub-fields, such as out-of-distribution detection.\nOn this new evaluation, we again demonstrate that there is negligible\ndifference between the strong baseline and the existing state-of-the-art.\nProject Page: https://www.robots.ox.ac.uk/~vgg/research/osr/", 'authors': ['Sagar Vaze', 'Kai Han', 'Andrea Vedaldi', 'Andrew Zisserman'], 'published': '2021-10-12 17:58:59+00:00', 'categories': ['cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2110.06207v2'}
42,490
2109.03839v3
2,109.03839
3
Sqrt(d) Dimension Dependence of Langevin Monte Carlo
This article considers the popular MCMC method of unadjusted Langevin Monte Carlo (LMC) and provides a non-asymptotic analysis of its sampling error in 2-Wasserstein distance. The proof is based on a refinement of mean-square analysis in Li et al. (2019), and this refined framework automates the analysis of a large class of sampling algorithms based on discretizations of contractive SDEs. Using this framework, we establish an $O(d/)$ mixing time bound for LMC, without warm start, under the common log-smooth and log-strongly-convex conditions, plus a growth condition on the 3rd-order derivative of the potential of target measures. This bound improves the best previously known $O(d/)$ result and is optimal (in terms of order) in both dimension $d$ and accuracy tolerance $$ for target measures satisfying the aforementioned assumptions. Our theoretical analysis is further validated by numerical experiments.
2021-09-08
{'id': '2109.03839v3', 'title': 'Sqrt(d) Dimension Dependence of Langevin Monte Carlo', 'abstract': 'This article considers the popular MCMC method of unadjusted Langevin Monte\nCarlo (LMC) and provides a non-asymptotic analysis of its sampling error in\n2-Wasserstein distance. The proof is based on a refinement of mean-square\nanalysis in Li et al. (2019), and this refined framework automates the analysis\nof a large class of sampling algorithms based on discretizations of contractive\nSDEs. Using this framework, we establish an $\\tilde{O}(\\sqrt{d}/\\epsilon)$\nmixing time bound for LMC, without warm start, under the common log-smooth and\nlog-strongly-convex conditions, plus a growth condition on the 3rd-order\nderivative of the potential of target measures. This bound improves the best\npreviously known $\\tilde{O}(d/\\epsilon)$ result and is optimal (in terms of\norder) in both dimension $d$ and accuracy tolerance $\\epsilon$ for target\nmeasures satisfying the aforementioned assumptions. Our theoretical analysis is\nfurther validated by numerical experiments.', 'authors': ['Ruilin Li', 'Hongyuan Zha', 'Molei Tao'], 'published': '2021-09-08 18:00:05+00:00', 'categories': ['cs.LG', 'cs.NA', 'math.NA', 'math.PR', 'math.ST', 'stat.ML', 'stat.TH'], 'url': 'http://arxiv.org/abs/2109.03839v3'}
42,492
2201.08549v1
2,201.08549
1
Fair Node Representation Learning via Adaptive Data Augmentation
Node representation learning has demonstrated its efficacy for various applications on graphs, which leads to increasing attention towards the area. However, fairness is a largely under-explored territory within the field, which may lead to biased results towards underrepresented groups in ensuing tasks. To this end, this work theoretically explains the sources of bias in node representations obtained via Graph Neural Networks (GNNs). Our analysis reveals that both nodal features and graph structure lead to bias in the obtained representations. Building upon the analysis, fairness-aware data augmentation frameworks on nodal features and graph structure are developed to reduce the intrinsic bias. Our analysis and proposed schemes can be readily employed to enhance the fairness of various GNN-based learning mechanisms. Extensive experiments on node classification and link prediction are carried out over real networks in the context of graph contrastive learning. Comparison with multiple benchmarks demonstrates that the proposed augmentation strategies can improve fairness in terms of statistical parity and equal opportunity, while providing comparable utility to state-of-the-art contrastive methods.
2022-01-21
{'id': '2201.08549v1', 'title': 'Fair Node Representation Learning via Adaptive Data Augmentation', 'abstract': 'Node representation learning has demonstrated its efficacy for various\napplications on graphs, which leads to increasing attention towards the area.\nHowever, fairness is a largely under-explored territory within the field, which\nmay lead to biased results towards underrepresented groups in ensuing tasks. To\nthis end, this work theoretically explains the sources of bias in node\nrepresentations obtained via Graph Neural Networks (GNNs). Our analysis reveals\nthat both nodal features and graph structure lead to bias in the obtained\nrepresentations. Building upon the analysis, fairness-aware data augmentation\nframeworks on nodal features and graph structure are developed to reduce the\nintrinsic bias. Our analysis and proposed schemes can be readily employed to\nenhance the fairness of various GNN-based learning mechanisms. Extensive\nexperiments on node classification and link prediction are carried out over\nreal networks in the context of graph contrastive learning. Comparison with\nmultiple benchmarks demonstrates that the proposed augmentation strategies can\nimprove fairness in terms of statistical parity and equal opportunity, while\nproviding comparable utility to state-of-the-art contrastive methods.', 'authors': ['O. Deniz Kose', 'Yanning Shen'], 'published': '2022-01-21 05:49:15+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2201.08549v1'}
42,436
2307.12851v2
2,307.12851
2
Early Neuron Alignment in Two-layer ReLU Networks with Small Initialization
This paper studies the problem of training a two-layer ReLU network for binary classification using gradient flow with small initialization. We consider a training dataset with well-separated input vectors: Any pair of input data with the same label are positively correlated, and any pair with different labels are negatively correlated. Our analysis shows that, during the early phase of training, neurons in the first layer try to align with either the positive data or the negative data, depending on its corresponding weight on the second layer. A careful analysis of the neurons' directional dynamics allows us to provide an $O(n)$ upper bound on the time it takes for all neurons to achieve good alignment with the input data, where $n$ is the number of data points and $$ measures how well the data are separated. After the early alignment phase, the loss converges to zero at a $O(1t)$ rate, and the weight matrix on the first layer is approximately low-rank. Numerical experiments on the MNIST dataset illustrate our theoretical findings.
2023-07-24
{'id': '2307.12851v2', 'title': 'Early Neuron Alignment in Two-layer ReLU Networks with Small Initialization', 'abstract': "This paper studies the problem of training a two-layer ReLU network for\nbinary classification using gradient flow with small initialization. We\nconsider a training dataset with well-separated input vectors: Any pair of\ninput data with the same label are positively correlated, and any pair with\ndifferent labels are negatively correlated. Our analysis shows that, during the\nearly phase of training, neurons in the first layer try to align with either\nthe positive data or the negative data, depending on its corresponding weight\non the second layer. A careful analysis of the neurons' directional dynamics\nallows us to provide an $\\mathcal{O}(\\frac{\\log n}{\\sqrt{\\mu}})$ upper bound on\nthe time it takes for all neurons to achieve good alignment with the input\ndata, where $n$ is the number of data points and $\\mu$ measures how well the\ndata are separated. After the early alignment phase, the loss converges to zero\nat a $\\mathcal{O}(\\frac{1}{t})$ rate, and the weight matrix on the first layer\nis approximately low-rank. Numerical experiments on the MNIST dataset\nillustrate our theoretical findings.", 'authors': ['Hancheng Min', 'Enrique Mallada', 'René Vidal'], 'published': '2023-07-24 14:51:54+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2307.12851v2'}
42,473
2311.01205v2
2,311.01205
2
Attacking Graph Neural Networks with Bit Flips: Weisfeiler and Lehman Go Indifferent
Prior attacks on graph neural networks have mostly focused on graph poisoning and evasion, neglecting the network's weights and biases. Traditional weight-based fault injection attacks, such as bit flip attacks used for convolutional neural networks, do not consider the unique properties of graph neural networks. We propose the Injectivity Bit Flip Attack, the first bit flip attack designed specifically for graph neural networks. Our attack targets the learnable neighborhood aggregation functions in quantized message passing neural networks, degrading their ability to distinguish graph structures and losing the expressivity of the Weisfeiler-Lehman test. Our findings suggest that exploiting mathematical properties specific to certain graph neural network architectures can significantly increase their vulnerability to bit flip attacks. Injectivity Bit Flip Attacks can degrade the maximal expressive Graph Isomorphism Networks trained on various graph property prediction datasets to random output by flipping only a small fraction of the network's bits, demonstrating its higher destructive power compared to a bit flip attack transferred from convolutional neural networks. Our attack is transparent and motivated by theoretical insights which are confirmed by extensive empirical results.
2023-11-02
{'id': '2311.01205v2', 'title': 'Attacking Graph Neural Networks with Bit Flips: Weisfeiler and Lehman Go Indifferent', 'abstract': "Prior attacks on graph neural networks have mostly focused on graph poisoning\nand evasion, neglecting the network's weights and biases. Traditional\nweight-based fault injection attacks, such as bit flip attacks used for\nconvolutional neural networks, do not consider the unique properties of graph\nneural networks. We propose the Injectivity Bit Flip Attack, the first bit flip\nattack designed specifically for graph neural networks. Our attack targets the\nlearnable neighborhood aggregation functions in quantized message passing\nneural networks, degrading their ability to distinguish graph structures and\nlosing the expressivity of the Weisfeiler-Lehman test. Our findings suggest\nthat exploiting mathematical properties specific to certain graph neural\nnetwork architectures can significantly increase their vulnerability to bit\nflip attacks. Injectivity Bit Flip Attacks can degrade the maximal expressive\nGraph Isomorphism Networks trained on various graph property prediction\ndatasets to random output by flipping only a small fraction of the network's\nbits, demonstrating its higher destructive power compared to a bit flip attack\ntransferred from convolutional neural networks. Our attack is transparent and\nmotivated by theoretical insights which are confirmed by extensive empirical\nresults.", 'authors': ['Lorenz Kummer', 'Samir Moustafa', 'Nils N. Kriege', 'Wilfried N. Gansterer'], 'published': '2023-11-02 12:59:32+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CR', 'cs.NE'], 'url': 'http://arxiv.org/abs/2311.01205v2'}
42,509
2403.01189v1
2,403.01189
1
Training Unbiased Diffusion Models From Biased Dataset
With significant advancements in diffusion models, addressing the potential risks of dataset bias becomes increasingly important. Since generated outputs directly suffer from dataset bias, mitigating latent bias becomes a key factor in improving sample quality and proportion. This paper proposes time-dependent importance reweighting to mitigate the bias for the diffusion models. We demonstrate that the time-dependent density ratio becomes more precise than previous approaches, thereby minimizing error propagation in generative learning. While directly applying it to score-matching is intractable, we discover that using the time-dependent density ratio both for reweighting and score correction can lead to a tractable form of the objective function to regenerate the unbiased data density. Furthermore, we theoretically establish a connection with traditional score-matching, and we demonstrate its convergence to an unbiased distribution. The experimental evidence supports the usefulness of the proposed method, which outperforms baselines including time-independent importance reweighting on CIFAR-10, CIFAR-100, FFHQ, and CelebA with various bias settings. Our code is available at https://github.com/alsdudrla10/TIW-DSM.
2024-03-02
{'id': '2403.01189v1', 'title': 'Training Unbiased Diffusion Models From Biased Dataset', 'abstract': 'With significant advancements in diffusion models, addressing the potential\nrisks of dataset bias becomes increasingly important. Since generated outputs\ndirectly suffer from dataset bias, mitigating latent bias becomes a key factor\nin improving sample quality and proportion. This paper proposes time-dependent\nimportance reweighting to mitigate the bias for the diffusion models. We\ndemonstrate that the time-dependent density ratio becomes more precise than\nprevious approaches, thereby minimizing error propagation in generative\nlearning. While directly applying it to score-matching is intractable, we\ndiscover that using the time-dependent density ratio both for reweighting and\nscore correction can lead to a tractable form of the objective function to\nregenerate the unbiased data density. Furthermore, we theoretically establish a\nconnection with traditional score-matching, and we demonstrate its convergence\nto an unbiased distribution. The experimental evidence supports the usefulness\nof the proposed method, which outperforms baselines including time-independent\nimportance reweighting on CIFAR-10, CIFAR-100, FFHQ, and CelebA with various\nbias settings. Our code is available at https://github.com/alsdudrla10/TIW-DSM.', 'authors': ['Yeongmin Kim', 'Byeonghu Na', 'Minsang Park', 'JoonHo Jang', 'Dongjun Kim', 'Wanmo Kang', 'Il-Chul Moon'], 'published': '2024-03-02 12:06:42+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2403.01189v1'}
42,511
2403.06392v1
2,403.06392
1
Towards Robust Out-of-Distribution Generalization Bounds via Sharpness
Generalizing to out-of-distribution (OOD) data or unseen domain, termed OOD generalization, still lacks appropriate theoretical guarantees. Canonical OOD bounds focus on different distance measurements between source and target domains but fail to consider the optimization property of the learned model. As empirically shown in recent work, the sharpness of learned minima influences OOD generalization. To bridge this gap between optimization and OOD generalization, we study the effect of sharpness on how a model tolerates data change in domain shift which is usually captured by "robustness" in generalization. In this paper, we give a rigorous connection between sharpness and robustness, which gives better OOD guarantees for robust algorithms. It also provides a theoretical backing for "flat minima leads to better OOD generalization". Overall, we propose a sharpness-based OOD generalization bound by taking robustness into consideration, resulting in a tighter bound than non-robust guarantees. Our findings are supported by the experiments on a ridge regression model, as well as the experiments on deep learning classification tasks.
2024-03-11
{'id': '2403.06392v1', 'title': 'Towards Robust Out-of-Distribution Generalization Bounds via Sharpness', 'abstract': 'Generalizing to out-of-distribution (OOD) data or unseen domain, termed OOD\ngeneralization, still lacks appropriate theoretical guarantees. Canonical OOD\nbounds focus on different distance measurements between source and target\ndomains but fail to consider the optimization property of the learned model. As\nempirically shown in recent work, the sharpness of learned minima influences\nOOD generalization. To bridge this gap between optimization and OOD\ngeneralization, we study the effect of sharpness on how a model tolerates data\nchange in domain shift which is usually captured by "robustness" in\ngeneralization. In this paper, we give a rigorous connection between sharpness\nand robustness, which gives better OOD guarantees for robust algorithms. It\nalso provides a theoretical backing for "flat minima leads to better OOD\ngeneralization". Overall, we propose a sharpness-based OOD generalization bound\nby taking robustness into consideration, resulting in a tighter bound than\nnon-robust guarantees. Our findings are supported by the experiments on a ridge\nregression model, as well as the experiments on deep learning classification\ntasks.', 'authors': ['Yingtian Zou', 'Kenji Kawaguchi', 'Yingnan Liu', 'Jiashuo Liu', 'Mong-Li Lee', 'Wynne Hsu'], 'published': '2024-03-11 02:57:27+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2403.06392v1'}
42,514
2310.09676v2
2,310.09676
2
Mastering Robot Manipulation with Multimodal Prompts through Pretraining and Multi-task Fine-tuning
Prompt-based learning has been demonstrated as a compelling paradigm contributing to large language models' tremendous success (LLMs). Inspired by their success in language tasks, existing research has leveraged LLMs in embodied instruction following and task planning. In this work, we tackle the problem of training a robot to understand multimodal prompts, interleaving vision signals with text descriptions. This type of task poses a major challenge to robots' capability to understand the interconnection and complementarity between vision and language signals. In this work, we introduce an effective framework that learns a policy to perform robot manipulation with multimodal prompts from multi-task expert trajectories. Our methods consist of a two-stage training pipeline that performs inverse dynamics pretraining and multi-task finetuning. To facilitate multimodal understanding, we design our multimodal prompt encoder by augmenting a pretrained LM with a residual connection to the visual input and model the dependencies among action dimensions. Empirically, we evaluate the efficacy of our method on the VIMA-BENCH and establish a new state-of-the-art (10% improvement in success rate). Moreover, we demonstrate that our model exhibits remarkable in-context learning ability. Project page: https://midas-icml.github.io/.
2023-10-14
{'id': '2310.09676v2', 'title': 'Mastering Robot Manipulation with Multimodal Prompts through Pretraining and Multi-task Fine-tuning', 'abstract': "Prompt-based learning has been demonstrated as a compelling paradigm\ncontributing to large language models' tremendous success (LLMs). Inspired by\ntheir success in language tasks, existing research has leveraged LLMs in\nembodied instruction following and task planning. In this work, we tackle the\nproblem of training a robot to understand multimodal prompts, interleaving\nvision signals with text descriptions. This type of task poses a major\nchallenge to robots' capability to understand the interconnection and\ncomplementarity between vision and language signals. In this work, we introduce\nan effective framework that learns a policy to perform robot manipulation with\nmultimodal prompts from multi-task expert trajectories. Our methods consist of\na two-stage training pipeline that performs inverse dynamics pretraining and\nmulti-task finetuning. To facilitate multimodal understanding, we design our\nmultimodal prompt encoder by augmenting a pretrained LM with a residual\nconnection to the visual input and model the dependencies among action\ndimensions. Empirically, we evaluate the efficacy of our method on the\nVIMA-BENCH and establish a new state-of-the-art (10% improvement in success\nrate). Moreover, we demonstrate that our model exhibits remarkable in-context\nlearning ability. Project page: \\url{https://midas-icml.github.io/}.", 'authors': ['Jiachen Li', 'Qiaozi Gao', 'Michael Johnston', 'Xiaofeng Gao', 'Xuehai He', 'Suhaila Shakiah', 'Hangjie Shi', 'Reza Ghanadan', 'William Yang Wang'], 'published': '2023-10-14 22:24:58+00:00', 'categories': ['cs.RO', 'cs.AI'], 'url': 'http://arxiv.org/abs/2310.09676v2'}
42,516
2401.14110v1
2,401.1411
1
Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators
The majority of the research on the quantization of Deep Neural Networks (DNNs) is focused on reducing the precision of tensors visible by high-level frameworks (e.g., weights, activations, and gradients). However, current hardware still relies on high-accuracy core operations. Most significant is the operation of accumulating products. This high-precision accumulation operation is gradually becoming the main computational bottleneck. This is because, so far, the usage of low-precision accumulators led to a significant degradation in performance. In this work, we present a simple method to train and fine-tune high-end DNNs, to allow, for the first time, utilization of cheaper, $12$-bits accumulators, with no significant degradation in accuracy. Lastly, we show that as we decrease the accumulation precision further, using fine-grained gradient approximations can improve the DNN accuracy.
2024-01-25
{'id': '2401.14110v1', 'title': 'Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators', 'abstract': 'The majority of the research on the quantization of Deep Neural Networks\n(DNNs) is focused on reducing the precision of tensors visible by high-level\nframeworks (e.g., weights, activations, and gradients). However, current\nhardware still relies on high-accuracy core operations. Most significant is the\noperation of accumulating products. This high-precision accumulation operation\nis gradually becoming the main computational bottleneck. This is because, so\nfar, the usage of low-precision accumulators led to a significant degradation\nin performance. In this work, we present a simple method to train and fine-tune\nhigh-end DNNs, to allow, for the first time, utilization of cheaper, $12$-bits\naccumulators, with no significant degradation in accuracy. Lastly, we show that\nas we decrease the accumulation precision further, using fine-grained gradient\napproximations can improve the DNN accuracy.', 'authors': ['Yaniv Blumenfeld', 'Itay Hubara', 'Daniel Soudry'], 'published': '2024-01-25 11:46:01+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.AR'], 'url': 'http://arxiv.org/abs/2401.14110v1'}
42,437
2309.16014v3
2,309.16014
3
Graph-level Representation Learning with Joint-Embedding Predictive Architectures
Joint-Embedding Predictive Architectures (JEPAs) have recently emerged as a novel and powerful technique for self-supervised representation learning. They aim to learn an energy-based model by predicting the latent representation of a target signal y from the latent representation of a context signal x. JEPAs bypass the need for negative and positive samples, traditionally required by contrastive learning while avoiding the overfitting issues associated with generative pretraining. In this paper, we show that graph-level representations can be effectively modeled using this paradigm by proposing a Graph Joint-Embedding Predictive Architecture (Graph-JEPA). In particular, we employ masked modeling and focus on predicting the latent representations of masked subgraphs starting from the latent representation of a context subgraph. To endow the representations with the implicit hierarchy that is often present in graph-level concepts, we devise an alternative prediction objective that consists of predicting the coordinates of the encoded subgraphs on the unit hyperbola in the 2D plane. Through multiple experimental evaluations, we show that Graph-JEPA can learn highly semantic and expressive representations, as shown by the downstream performance in graph classification, regression, and distinguishing non-isomorphic graphs. The code is available at https://github.com/geriskenderi/graph-jepa.
2023-09-27
{'id': '2309.16014v3', 'title': 'Graph-level Representation Learning with Joint-Embedding Predictive Architectures', 'abstract': 'Joint-Embedding Predictive Architectures (JEPAs) have recently emerged as a\nnovel and powerful technique for self-supervised representation learning. They\naim to learn an energy-based model by predicting the latent representation of a\ntarget signal y from the latent representation of a context signal x. JEPAs\nbypass the need for negative and positive samples, traditionally required by\ncontrastive learning while avoiding the overfitting issues associated with\ngenerative pretraining. In this paper, we show that graph-level representations\ncan be effectively modeled using this paradigm by proposing a Graph\nJoint-Embedding Predictive Architecture (Graph-JEPA). In particular, we employ\nmasked modeling and focus on predicting the latent representations of masked\nsubgraphs starting from the latent representation of a context subgraph. To\nendow the representations with the implicit hierarchy that is often present in\ngraph-level concepts, we devise an alternative prediction objective that\nconsists of predicting the coordinates of the encoded subgraphs on the unit\nhyperbola in the 2D plane. Through multiple experimental evaluations, we show\nthat Graph-JEPA can learn highly semantic and expressive representations, as\nshown by the downstream performance in graph classification, regression, and\ndistinguishing non-isomorphic graphs. The code is available at\nhttps://github.com/geriskenderi/graph-jepa.', 'authors': ['Geri Skenderi', 'Hang Li', 'Jiliang Tang', 'Marco Cristani'], 'published': '2023-09-27 20:42:02+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2309.16014v3'}
42,469
2402.14817v3
2,402.14817
3
Cameras as Rays: Pose Estimation via Ray Diffusion
Estimating camera poses is a fundamental task for 3D reconstruction and remains challenging given sparsely sampled views (<10). In contrast to existing approaches that pursue top-down prediction of global parametrizations of camera extrinsics, we propose a distributed representation of camera pose that treats a camera as a bundle of rays. This representation allows for a tight coupling with spatial image features improving pose precision. We observe that this representation is naturally suited for set-level transformers and develop a regression-based approach that maps image patches to corresponding rays. To capture the inherent uncertainties in sparse-view pose inference, we adapt this approach to learn a denoising diffusion model which allows us to sample plausible modes while improving performance. Our proposed methods, both regression- and diffusion-based, demonstrate state-of-the-art performance on camera pose estimation on CO3D while generalizing to unseen object categories and in-the-wild captures.
2024-02-22
{'id': '2402.14817v3', 'title': 'Cameras as Rays: Pose Estimation via Ray Diffusion', 'abstract': 'Estimating camera poses is a fundamental task for 3D reconstruction and\nremains challenging given sparsely sampled views (<10). In contrast to existing\napproaches that pursue top-down prediction of global parametrizations of camera\nextrinsics, we propose a distributed representation of camera pose that treats\na camera as a bundle of rays. This representation allows for a tight coupling\nwith spatial image features improving pose precision. We observe that this\nrepresentation is naturally suited for set-level transformers and develop a\nregression-based approach that maps image patches to corresponding rays. To\ncapture the inherent uncertainties in sparse-view pose inference, we adapt this\napproach to learn a denoising diffusion model which allows us to sample\nplausible modes while improving performance. Our proposed methods, both\nregression- and diffusion-based, demonstrate state-of-the-art performance on\ncamera pose estimation on CO3D while generalizing to unseen object categories\nand in-the-wild captures.', 'authors': ['Jason Y. Zhang', 'Amy Lin', 'Moneish Kumar', 'Tzu-Hsuan Yang', 'Deva Ramanan', 'Shubham Tulsiani'], 'published': '2024-02-22 18:59:56+00:00', 'categories': ['cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2402.14817v3'}
42,477
2306.08103v4
2,306.08103
4
Generating Images with 3D Annotations Using Diffusion Models
Diffusion models have emerged as a powerful generative method, capable of producing stunning photo-realistic images from natural language descriptions. However, these models lack explicit control over the 3D structure in the generated images. Consequently, this hinders our ability to obtain detailed 3D annotations for the generated images or to craft instances with specific poses and distances. In this paper, we propose 3D Diffusion Style Transfer (3D-DST), which incorporates 3D geometry control into diffusion models. Our method exploits ControlNet, which extends diffusion models by using visual prompts in addition to text prompts. We generate images of the 3D objects taken from 3D shape repositories (e.g., ShapeNet and Objaverse), render them from a variety of poses and viewing directions, compute the edge maps of the rendered images, and use these edge maps as visual prompts to generate realistic images. With explicit 3D geometry control, we can easily change the 3D structures of the objects in the generated images and obtain ground-truth 3D annotations automatically. This allows us to improve a wide range of vision tasks, e.g., classification and 3D pose estimation, in both in-distribution (ID) and out-of-distribution (OOD) settings. We demonstrate the effectiveness of our method through extensive experiments on ImageNet-100/200, ImageNet-R, PASCAL3D+, ObjectNet3D, and OOD-CV. The results show that our method significantly outperforms existing methods, e.g., 3.8 percentage points on ImageNet-100 using DeiT-B.
2023-06-13
{'id': '2306.08103v4', 'title': 'Generating Images with 3D Annotations Using Diffusion Models', 'abstract': 'Diffusion models have emerged as a powerful generative method, capable of\nproducing stunning photo-realistic images from natural language descriptions.\nHowever, these models lack explicit control over the 3D structure in the\ngenerated images. Consequently, this hinders our ability to obtain detailed 3D\nannotations for the generated images or to craft instances with specific poses\nand distances. In this paper, we propose 3D Diffusion Style Transfer (3D-DST),\nwhich incorporates 3D geometry control into diffusion models. Our method\nexploits ControlNet, which extends diffusion models by using visual prompts in\naddition to text prompts. We generate images of the 3D objects taken from 3D\nshape repositories (e.g., ShapeNet and Objaverse), render them from a variety\nof poses and viewing directions, compute the edge maps of the rendered images,\nand use these edge maps as visual prompts to generate realistic images. With\nexplicit 3D geometry control, we can easily change the 3D structures of the\nobjects in the generated images and obtain ground-truth 3D annotations\nautomatically. This allows us to improve a wide range of vision tasks, e.g.,\nclassification and 3D pose estimation, in both in-distribution (ID) and\nout-of-distribution (OOD) settings. We demonstrate the effectiveness of our\nmethod through extensive experiments on ImageNet-100/200, ImageNet-R,\nPASCAL3D+, ObjectNet3D, and OOD-CV. The results show that our method\nsignificantly outperforms existing methods, e.g., 3.8 percentage points on\nImageNet-100 using DeiT-B.', 'authors': ['Wufei Ma', 'Qihao Liu', 'Jiahao Wang', 'Angtian Wang', 'Xiaoding Yuan', 'Yi Zhang', 'Zihao Xiao', 'Guofeng Zhang', 'Beijia Lu', 'Ruxiao Duan', 'Yongrui Qi', 'Adam Kortylewski', 'Yaoyao Liu', 'Alan Yuille'], 'published': '2023-06-13 19:48:56+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2306.08103v4'}
42,483
2305.13300v4
2,305.133
4
Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts
By providing external information to large language models (LLMs), tool augmentation (including retrieval augmentation) has emerged as a promising solution for addressing the limitations of LLMs' static parametric memory. However, how receptive are LLMs to such external evidence, especially when the evidence conflicts with their parametric memory? We present the first comprehensive and controlled investigation into the behavior of LLMs when encountering knowledge conflicts. We propose a systematic framework to elicit high-quality parametric memory from LLMs and construct the corresponding counter-memory, which enables us to conduct a series of controlled experiments. Our investigation reveals seemingly contradicting behaviors of LLMs. On the one hand, different from prior wisdom, we find that LLMs can be highly receptive to external evidence even when that conflicts with their parametric memory, given that the external evidence is coherent and convincing. On the other hand, LLMs also demonstrate a strong confirmation bias when the external evidence contains some information that is consistent with their parametric memory, despite being presented with conflicting evidence at the same time. These results pose important implications that are worth careful consideration for the further development and deployment of tool- and retrieval-augmented LLMs. Resources are available at https://github.com/OSU-NLP-Group/LLM-Knowledge-Conflict.
2023-05-22
{'id': '2305.13300v4', 'title': 'Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts', 'abstract': "By providing external information to large language models (LLMs), tool\naugmentation (including retrieval augmentation) has emerged as a promising\nsolution for addressing the limitations of LLMs' static parametric memory.\nHowever, how receptive are LLMs to such external evidence, especially when the\nevidence conflicts with their parametric memory? We present the first\ncomprehensive and controlled investigation into the behavior of LLMs when\nencountering knowledge conflicts. We propose a systematic framework to elicit\nhigh-quality parametric memory from LLMs and construct the corresponding\ncounter-memory, which enables us to conduct a series of controlled experiments.\nOur investigation reveals seemingly contradicting behaviors of LLMs. On the one\nhand, different from prior wisdom, we find that LLMs can be highly receptive to\nexternal evidence even when that conflicts with their parametric memory, given\nthat the external evidence is coherent and convincing. On the other hand, LLMs\nalso demonstrate a strong confirmation bias when the external evidence contains\nsome information that is consistent with their parametric memory, despite being\npresented with conflicting evidence at the same time. These results pose\nimportant implications that are worth careful consideration for the further\ndevelopment and deployment of tool- and retrieval-augmented LLMs. Resources are\navailable at https://github.com/OSU-NLP-Group/LLM-Knowledge-Conflict.", 'authors': ['Jian Xie', 'Kai Zhang', 'Jiangjie Chen', 'Renze Lou', 'Yu Su'], 'published': '2023-05-22 17:57:41+00:00', 'categories': ['cs.CL', 'cs.AI'], 'url': 'http://arxiv.org/abs/2305.13300v4'}
42,488
2309.14859v2
2,309.14859
2
Navigating Text-To-Image Customization: From LyCORIS Fine-Tuning to Model Evaluation
Text-to-image generative models have garnered immense attention for their ability to produce high-fidelity images from text prompts. Among these, Stable Diffusion distinguishes itself as a leading open-source model in this fast-growing field. However, the intricacies of fine-tuning these models pose multiple challenges from new methodology integration to systematic evaluation. Addressing these issues, this paper introduces LyCORIS (Lora beYond Conventional methods, Other Rank adaptation Implementations for Stable diffusion) [https://github.com/KohakuBlueleaf/LyCORIS], an open-source library that offers a wide selection of fine-tuning methodologies for Stable Diffusion. Furthermore, we present a thorough framework for the systematic assessment of varied fine-tuning techniques. This framework employs a diverse suite of metrics and delves into multiple facets of fine-tuning, including hyperparameter adjustments and the evaluation with different prompt types across various concept categories. Through this comprehensive approach, our work provides essential insights into the nuanced effects of fine-tuning parameters, bridging the gap between state-of-the-art research and practical application.
2023-09-26
{'id': '2309.14859v2', 'title': 'Navigating Text-To-Image Customization: From LyCORIS Fine-Tuning to Model Evaluation', 'abstract': 'Text-to-image generative models have garnered immense attention for their\nability to produce high-fidelity images from text prompts. Among these, Stable\nDiffusion distinguishes itself as a leading open-source model in this\nfast-growing field. However, the intricacies of fine-tuning these models pose\nmultiple challenges from new methodology integration to systematic evaluation.\nAddressing these issues, this paper introduces LyCORIS (Lora beYond\nConventional methods, Other Rank adaptation Implementations for Stable\ndiffusion) [https://github.com/KohakuBlueleaf/LyCORIS], an open-source library\nthat offers a wide selection of fine-tuning methodologies for Stable Diffusion.\nFurthermore, we present a thorough framework for the systematic assessment of\nvaried fine-tuning techniques. This framework employs a diverse suite of\nmetrics and delves into multiple facets of fine-tuning, including\nhyperparameter adjustments and the evaluation with different prompt types\nacross various concept categories. Through this comprehensive approach, our\nwork provides essential insights into the nuanced effects of fine-tuning\nparameters, bridging the gap between state-of-the-art research and practical\napplication.', 'authors': ['Shih-Ying Yeh', 'Yu-Guan Hsieh', 'Zhidong Gao', 'Bernard B W Yang', 'Giyeong Oh', 'Yanmin Gong'], 'published': '2023-09-26 11:36:26+00:00', 'categories': ['cs.CV', 'cs.AI', 'cs.GR', 'cs.LG'], 'url': 'http://arxiv.org/abs/2309.14859v2'}
42,497
2401.12588v2
2,401.12588
2
Interpreting Equivariant Representations
Latent representations are used extensively for downstream tasks, such as visualization, interpolation or feature extraction of deep learning models. Invariant and equivariant neural networks are powerful and well-established models for enforcing inductive biases. In this paper, we demonstrate that the inductive bias imposed on the by an equivariant model must also be taken into account when using latent representations. We show how not accounting for the inductive biases leads to decreased performance on downstream tasks, and vice versa, how accounting for inductive biases can be done effectively by using an invariant projection of the latent representations. We propose principles for how to choose such a projection, and show the impact of using these principles in two common examples: First, we study a permutation equivariant variational auto-encoder trained for molecule graph generation; here we show that invariant projections can be designed that incur no loss of information in the resulting invariant representation. Next, we study a rotation-equivariant representation used for image classification. Here, we illustrate how random invariant projections can be used to obtain an invariant representation with a high degree of retained information. In both cases, the analysis of invariant latent representations proves superior to their equivariant counterparts. Finally, we illustrate that the phenomena documented here for equivariant neural networks have counterparts in standard neural networks where invariance is encouraged via augmentation. Thus, while these ambiguities may be known by experienced developers of equivariant models, we make both the knowledge as well as effective tools to handle the ambiguities available to the broader community.
2024-01-23
{'id': '2401.12588v2', 'title': 'Interpreting Equivariant Representations', 'abstract': 'Latent representations are used extensively for downstream tasks, such as\nvisualization, interpolation or feature extraction of deep learning models.\nInvariant and equivariant neural networks are powerful and well-established\nmodels for enforcing inductive biases. In this paper, we demonstrate that the\ninductive bias imposed on the by an equivariant model must also be taken into\naccount when using latent representations. We show how not accounting for the\ninductive biases leads to decreased performance on downstream tasks, and vice\nversa, how accounting for inductive biases can be done effectively by using an\ninvariant projection of the latent representations. We propose principles for\nhow to choose such a projection, and show the impact of using these principles\nin two common examples: First, we study a permutation equivariant variational\nauto-encoder trained for molecule graph generation; here we show that invariant\nprojections can be designed that incur no loss of information in the resulting\ninvariant representation. Next, we study a rotation-equivariant representation\nused for image classification. Here, we illustrate how random invariant\nprojections can be used to obtain an invariant representation with a high\ndegree of retained information. In both cases, the analysis of invariant latent\nrepresentations proves superior to their equivariant counterparts. Finally, we\nillustrate that the phenomena documented here for equivariant neural networks\nhave counterparts in standard neural networks where invariance is encouraged\nvia augmentation. Thus, while these ambiguities may be known by experienced\ndevelopers of equivariant models, we make both the knowledge as well as\neffective tools to handle the ambiguities available to the broader community.', 'authors': ['Andreas Abildtrup Hansen', 'Anna Calissano', 'Aasa Feragen'], 'published': '2024-01-23 09:43:30+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2401.12588v2'}
42,439
2105.10536v1
2,105.10536
1
Semi-Supervised Audio Representation Learning for Modeling Beehive Strengths
Honey bees are critical to our ecosystem and food security as a pollinator, contributing 35% of our global agriculture yield. In spite of their importance, beekeeping is exclusively dependent on human labor and experience-derived heuristics, while requiring frequent human checkups to ensure the colony is healthy, which can disrupt the colony. Increasingly, pollinator populations are declining due to threats from climate change, pests, environmental toxicity, making their management even more critical than ever before in order to ensure sustained global food security. To start addressing this pressing challenge, we developed an integrated hardware sensing system for beehive monitoring through audio and environment measurements, and a hierarchical semi-supervised deep learning model, composed of an audio modeling module and a predictor, to model the strength of beehives. The model is trained jointly on audio reconstruction and prediction losses based on human inspections, in order to model both low-level audio features and circadian temporal dynamics. We show that this model performs well despite limited labels, and can learn an audio embedding that is useful for characterizing different sound profiles of beehives. This is the first instance to our knowledge of applying audio-based deep learning to model beehives and population size in an observational setting across a large number of hives.
2021-05-21
{'id': '2105.10536v1', 'title': 'Semi-Supervised Audio Representation Learning for Modeling Beehive Strengths', 'abstract': 'Honey bees are critical to our ecosystem and food security as a pollinator,\ncontributing 35% of our global agriculture yield. In spite of their importance,\nbeekeeping is exclusively dependent on human labor and experience-derived\nheuristics, while requiring frequent human checkups to ensure the colony is\nhealthy, which can disrupt the colony. Increasingly, pollinator populations are\ndeclining due to threats from climate change, pests, environmental toxicity,\nmaking their management even more critical than ever before in order to ensure\nsustained global food security. To start addressing this pressing challenge, we\ndeveloped an integrated hardware sensing system for beehive monitoring through\naudio and environment measurements, and a hierarchical semi-supervised deep\nlearning model, composed of an audio modeling module and a predictor, to model\nthe strength of beehives. The model is trained jointly on audio reconstruction\nand prediction losses based on human inspections, in order to model both\nlow-level audio features and circadian temporal dynamics. We show that this\nmodel performs well despite limited labels, and can learn an audio embedding\nthat is useful for characterizing different sound profiles of beehives. This is\nthe first instance to our knowledge of applying audio-based deep learning to\nmodel beehives and population size in an observational setting across a large\nnumber of hives.', 'authors': ['Tony Zhang', 'Szymon Zmyslony', 'Sergei Nozdrenkov', 'Matthew Smith', 'Brandon Hopkins'], 'published': '2021-05-21 18:59:29+00:00', 'categories': ['cs.SD', 'cs.LG', 'eess.AS'], 'url': 'http://arxiv.org/abs/2105.10536v1'}
42,453
2008.06043v3
2,008.06043
3
Offline Meta-Reinforcement Learning with Advantage Weighting
This paper introduces the offline meta-reinforcement learning (offline meta-RL) problem setting and proposes an algorithm that performs well in this setting. Offline meta-RL is analogous to the widely successful supervised learning strategy of pre-training a model on a large batch of fixed, pre-collected data (possibly from various tasks) and fine-tuning the model to a new task with relatively little data. That is, in offline meta-RL, we meta-train on fixed, pre-collected data from several tasks in order to adapt to a new task with a very small amount (less than 5 trajectories) of data from the new task. By nature of being offline, algorithms for offline meta-RL can utilize the largest possible pool of training data available and eliminate potentially unsafe or costly data collection during meta-training. This setting inherits the challenges of offline RL, but it differs significantly because offline RL does not generally consider a) transfer to new tasks or b) limited data from the test task, both of which we face in offline meta-RL. Targeting the offline meta-RL setting, we propose Meta-Actor Critic with Advantage Weighting (MACAW), an optimization-based meta-learning algorithm that uses simple, supervised regression objectives for both the inner and outer loop of meta-training. On offline variants of common meta-RL benchmarks, we empirically find that this approach enables fully offline meta-reinforcement learning and achieves notable gains over prior methods.
2020-08-13
{'id': '2008.06043v3', 'title': 'Offline Meta-Reinforcement Learning with Advantage Weighting', 'abstract': 'This paper introduces the offline meta-reinforcement learning (offline\nmeta-RL) problem setting and proposes an algorithm that performs well in this\nsetting. Offline meta-RL is analogous to the widely successful supervised\nlearning strategy of pre-training a model on a large batch of fixed,\npre-collected data (possibly from various tasks) and fine-tuning the model to a\nnew task with relatively little data. That is, in offline meta-RL, we\nmeta-train on fixed, pre-collected data from several tasks in order to adapt to\na new task with a very small amount (less than 5 trajectories) of data from the\nnew task. By nature of being offline, algorithms for offline meta-RL can\nutilize the largest possible pool of training data available and eliminate\npotentially unsafe or costly data collection during meta-training. This setting\ninherits the challenges of offline RL, but it differs significantly because\noffline RL does not generally consider a) transfer to new tasks or b) limited\ndata from the test task, both of which we face in offline meta-RL. Targeting\nthe offline meta-RL setting, we propose Meta-Actor Critic with Advantage\nWeighting (MACAW), an optimization-based meta-learning algorithm that uses\nsimple, supervised regression objectives for both the inner and outer loop of\nmeta-training. On offline variants of common meta-RL benchmarks, we empirically\nfind that this approach enables fully offline meta-reinforcement learning and\nachieves notable gains over prior methods.', 'authors': ['Eric Mitchell', 'Rafael Rafailov', 'Xue Bin Peng', 'Sergey Levine', 'Chelsea Finn'], 'published': '2020-08-13 17:57:14+00:00', 'categories': ['cs.LG', 'cs.AI', 'stat.ML'], 'url': 'http://arxiv.org/abs/2008.06043v3'}
42,463
2006.16469v2
2,006.16469
2
Model-Targeted Poisoning Attacks with Provable Convergence
In a poisoning attack, an adversary with control over a small fraction of the training data attempts to select that data in a way that induces a corrupted model that misbehaves in favor of the adversary. We consider poisoning attacks against convex machine learning models and propose an efficient poisoning attack designed to induce a specified model. Unlike previous model-targeted poisoning attacks, our attack comes with provable convergence to any attainable target classifier. The distance from the induced classifier to the target classifier is inversely proportional to the square root of the number of poisoning points. We also provide a lower bound on the minimum number of poisoning points needed to achieve a given target classifier. Our method uses online convex optimization, so finds poisoning points incrementally. This provides more flexibility than previous attacks which require a priori assumption about the number of poisoning points. Our attack is the first model-targeted poisoning attack that provides provable convergence for convex models, and in our experiments, it either exceeds or matches state-of-the-art attacks in terms of attack success rate and distance to the target model.
2020-06-30
{'id': '2006.16469v2', 'title': 'Model-Targeted Poisoning Attacks with Provable Convergence', 'abstract': 'In a poisoning attack, an adversary with control over a small fraction of the\ntraining data attempts to select that data in a way that induces a corrupted\nmodel that misbehaves in favor of the adversary. We consider poisoning attacks\nagainst convex machine learning models and propose an efficient poisoning\nattack designed to induce a specified model. Unlike previous model-targeted\npoisoning attacks, our attack comes with provable convergence to {\\it any}\nattainable target classifier. The distance from the induced classifier to the\ntarget classifier is inversely proportional to the square root of the number of\npoisoning points. We also provide a lower bound on the minimum number of\npoisoning points needed to achieve a given target classifier. Our method uses\nonline convex optimization, so finds poisoning points incrementally. This\nprovides more flexibility than previous attacks which require a priori\nassumption about the number of poisoning points. Our attack is the first\nmodel-targeted poisoning attack that provides provable convergence for convex\nmodels, and in our experiments, it either exceeds or matches state-of-the-art\nattacks in terms of attack success rate and distance to the target model.', 'authors': ['Fnu Suya', 'Saeed Mahloujifar', 'Anshuman Suri', 'David Evans', 'Yuan Tian'], 'published': '2020-06-30 01:56:35+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CR', 'stat.ML'], 'url': 'http://arxiv.org/abs/2006.16469v2'}
42,466
2010.14498v2
2,010.14498
2
Implicit Under-Parameterization Inhibits Data-Efficient Deep Reinforcement Learning
We identify an implicit under-parameterization phenomenon in value-based deep RL methods that use bootstrapping: when value functions, approximated using deep neural networks, are trained with gradient descent using iterated regression onto target values generated by previous instances of the value network, more gradient updates decrease the expressivity of the current value network. We characterize this loss of expressivity via a drop in the rank of the learned value network features, and show that this typically corresponds to a performance drop. We demonstrate this phenomenon on Atari and Gym benchmarks, in both offline and online RL settings. We formally analyze this phenomenon and show that it results from a pathological interaction between bootstrapping and gradient-based optimization. We further show that mitigating implicit under-parameterization by controlling rank collapse can improve performance.
2020-10-27
{'id': '2010.14498v2', 'title': 'Implicit Under-Parameterization Inhibits Data-Efficient Deep Reinforcement Learning', 'abstract': 'We identify an implicit under-parameterization phenomenon in value-based deep\nRL methods that use bootstrapping: when value functions, approximated using\ndeep neural networks, are trained with gradient descent using iterated\nregression onto target values generated by previous instances of the value\nnetwork, more gradient updates decrease the expressivity of the current value\nnetwork. We characterize this loss of expressivity via a drop in the rank of\nthe learned value network features, and show that this typically corresponds to\na performance drop. We demonstrate this phenomenon on Atari and Gym benchmarks,\nin both offline and online RL settings. We formally analyze this phenomenon and\nshow that it results from a pathological interaction between bootstrapping and\ngradient-based optimization. We further show that mitigating implicit\nunder-parameterization by controlling rank collapse can improve performance.', 'authors': ['Aviral Kumar', 'Rishabh Agarwal', 'Dibya Ghosh', 'Sergey Levine'], 'published': '2020-10-27 17:55:16+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2010.14498v2'}
42,478
2010.14134v3
2,010.14134
3
Selective Classification Can Magnify Disparities Across Groups
Selective classification, in which models can abstain on uncertain predictions, is a natural approach to improving accuracy in settings where errors are costly but abstentions are manageable. In this paper, we find that while selective classification can improve average accuracies, it can simultaneously magnify existing accuracy disparities between various groups within a population, especially in the presence of spurious correlations. We observe this behavior consistently across five vision and NLP datasets. Surprisingly, increasing abstentions can even decrease accuracies on some groups. To better understand this phenomenon, we study the margin distribution, which captures the model's confidences over all predictions. For symmetric margin distributions, we prove that whether selective classification monotonically improves or worsens accuracy is fully determined by the accuracy at full coverage (i.e., without any abstentions) and whether the distribution satisfies a property we call left-log-concavity. Our analysis also shows that selective classification tends to magnify full-coverage accuracy disparities. Motivated by our analysis, we train distributionally-robust models that achieve similar full-coverage accuracies across groups and show that selective classification uniformly improves each group on these models. Altogether, our results suggest that selective classification should be used with care and underscore the importance of training models to perform equally well across groups at full coverage.
2020-10-27
{'id': '2010.14134v3', 'title': 'Selective Classification Can Magnify Disparities Across Groups', 'abstract': "Selective classification, in which models can abstain on uncertain\npredictions, is a natural approach to improving accuracy in settings where\nerrors are costly but abstentions are manageable. In this paper, we find that\nwhile selective classification can improve average accuracies, it can\nsimultaneously magnify existing accuracy disparities between various groups\nwithin a population, especially in the presence of spurious correlations. We\nobserve this behavior consistently across five vision and NLP datasets.\nSurprisingly, increasing abstentions can even decrease accuracies on some\ngroups. To better understand this phenomenon, we study the margin distribution,\nwhich captures the model's confidences over all predictions. For symmetric\nmargin distributions, we prove that whether selective classification\nmonotonically improves or worsens accuracy is fully determined by the accuracy\nat full coverage (i.e., without any abstentions) and whether the distribution\nsatisfies a property we call left-log-concavity. Our analysis also shows that\nselective classification tends to magnify full-coverage accuracy disparities.\nMotivated by our analysis, we train distributionally-robust models that achieve\nsimilar full-coverage accuracies across groups and show that selective\nclassification uniformly improves each group on these models. Altogether, our\nresults suggest that selective classification should be used with care and\nunderscore the importance of training models to perform equally well across\ngroups at full coverage.", 'authors': ['Erik Jones', 'Shiori Sagawa', 'Pang Wei Koh', 'Ananya Kumar', 'Percy Liang'], 'published': '2020-10-27 08:51:30+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2010.14134v3'}
42,486
2007.07400v1
2,007.074
1
Anatomy of Catastrophic Forgetting: Hidden Representations and Task Semantics
A central challenge in developing versatile machine learning systems is catastrophic forgetting: a model trained on tasks in sequence will suffer significant performance drops on earlier tasks. Despite the ubiquity of catastrophic forgetting, there is limited understanding of the underlying process and its causes. In this paper, we address this important knowledge gap, investigating how forgetting affects representations in neural network models. Through representational analysis techniques, we find that deeper layers are disproportionately the source of forgetting. Supporting this, a study of methods to mitigate forgetting illustrates that they act to stabilize deeper layers. These insights enable the development of an analytic argument and empirical picture relating the degree of forgetting to representational similarity between tasks. Consistent with this picture, we observe maximal forgetting occurs for task sequences with intermediate similarity. We perform empirical studies on the standard split CIFAR-10 setup and also introduce a novel CIFAR-100 based task approximating realistic input distribution shift.
2020-07-14
{'id': '2007.07400v1', 'title': 'Anatomy of Catastrophic Forgetting: Hidden Representations and Task Semantics', 'abstract': 'A central challenge in developing versatile machine learning systems is\ncatastrophic forgetting: a model trained on tasks in sequence will suffer\nsignificant performance drops on earlier tasks. Despite the ubiquity of\ncatastrophic forgetting, there is limited understanding of the underlying\nprocess and its causes. In this paper, we address this important knowledge gap,\ninvestigating how forgetting affects representations in neural network models.\nThrough representational analysis techniques, we find that deeper layers are\ndisproportionately the source of forgetting. Supporting this, a study of\nmethods to mitigate forgetting illustrates that they act to stabilize deeper\nlayers. These insights enable the development of an analytic argument and\nempirical picture relating the degree of forgetting to representational\nsimilarity between tasks. Consistent with this picture, we observe maximal\nforgetting occurs for task sequences with intermediate similarity. We perform\nempirical studies on the standard split CIFAR-10 setup and also introduce a\nnovel CIFAR-100 based task approximating realistic input distribution shift.', 'authors': ['Vinay V. Ramasesh', 'Ethan Dyer', 'Maithra Raghu'], 'published': '2020-07-14 23:31:14+00:00', 'categories': ['cs.LG', 'cs.CV', 'stat.ML'], 'url': 'http://arxiv.org/abs/2007.07400v1'}
42,449
2308.15911v1
2,308.15911
1
Cyclophobic Reinforcement Learning
In environments with sparse rewards, finding a good inductive bias for exploration is crucial to the agent's success. However, there are two competing goals: novelty search and systematic exploration. While existing approaches such as curiosity-driven exploration find novelty, they sometimes do not systematically explore the whole state space, akin to depth-first-search vs breadth-first-search. In this paper, we propose a new intrinsic reward that is cyclophobic, i.e., it does not reward novelty, but punishes redundancy by avoiding cycles. Augmenting the cyclophobic intrinsic reward with a sequence of hierarchical representations based on the agent's cropped observations we are able to achieve excellent results in the MiniGrid and MiniHack environments. Both are particularly hard, as they require complex interactions with different objects in order to be solved. Detailed comparisons with previous approaches and thorough ablation studies show that our newly proposed cyclophobic reinforcement learning is more sample efficient than other state of the art methods in a variety of tasks.
2023-08-30
{'id': '2308.15911v1', 'title': 'Cyclophobic Reinforcement Learning', 'abstract': "In environments with sparse rewards, finding a good inductive bias for\nexploration is crucial to the agent's success. However, there are two competing\ngoals: novelty search and systematic exploration. While existing approaches\nsuch as curiosity-driven exploration find novelty, they sometimes do not\nsystematically explore the whole state space, akin to depth-first-search vs\nbreadth-first-search. In this paper, we propose a new intrinsic reward that is\ncyclophobic, i.e., it does not reward novelty, but punishes redundancy by\navoiding cycles. Augmenting the cyclophobic intrinsic reward with a sequence of\nhierarchical representations based on the agent's cropped observations we are\nable to achieve excellent results in the MiniGrid and MiniHack environments.\nBoth are particularly hard, as they require complex interactions with different\nobjects in order to be solved. Detailed comparisons with previous approaches\nand thorough ablation studies show that our newly proposed cyclophobic\nreinforcement learning is more sample efficient than other state of the art\nmethods in a variety of tasks.", 'authors': ['Stefan Sylvius Wagner', 'Peter Arndt', 'Jan Robine', 'Stefan Harmeling'], 'published': '2023-08-30 09:38:44+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.RO'], 'url': 'http://arxiv.org/abs/2308.15911v1'}
42,452
2305.05239v1
2,305.05239
1
Learnable Behavior Control: Breaking Atari Human World Records via Sample-Efficient Behavior Selection
The exploration problem is one of the main challenges in deep reinforcement learning (RL). Recent promising works tried to handle the problem with population-based methods, which collect samples with diverse behaviors derived from a population of different exploratory policies. Adaptive policy selection has been adopted for behavior control. However, the behavior selection space is largely limited by the predefined policy population, which further limits behavior diversity. In this paper, we propose a general framework called Learnable Behavioral Control (LBC) to address the limitation, which a) enables a significantly enlarged behavior selection space via formulating a hybrid behavior mapping from all policies; b) constructs a unified learnable process for behavior selection. We introduce LBC into distributed off-policy actor-critic methods and achieve behavior control via optimizing the selection of the behavior mappings with bandit-based meta-controllers. Our agents have achieved 10077.52% mean human normalized score and surpassed 24 human world records within 1B training frames in the Arcade Learning Environment, which demonstrates our significant state-of-the-art (SOTA) performance without degrading the sample efficiency.
2023-05-09
{'id': '2305.05239v1', 'title': 'Learnable Behavior Control: Breaking Atari Human World Records via Sample-Efficient Behavior Selection', 'abstract': 'The exploration problem is one of the main challenges in deep reinforcement\nlearning (RL). Recent promising works tried to handle the problem with\npopulation-based methods, which collect samples with diverse behaviors derived\nfrom a population of different exploratory policies. Adaptive policy selection\nhas been adopted for behavior control. However, the behavior selection space is\nlargely limited by the predefined policy population, which further limits\nbehavior diversity. In this paper, we propose a general framework called\nLearnable Behavioral Control (LBC) to address the limitation, which a) enables\na significantly enlarged behavior selection space via formulating a hybrid\nbehavior mapping from all policies; b) constructs a unified learnable process\nfor behavior selection. We introduce LBC into distributed off-policy\nactor-critic methods and achieve behavior control via optimizing the selection\nof the behavior mappings with bandit-based meta-controllers. Our agents have\nachieved 10077.52% mean human normalized score and surpassed 24 human world\nrecords within 1B training frames in the Arcade Learning Environment, which\ndemonstrates our significant state-of-the-art (SOTA) performance without\ndegrading the sample efficiency.', 'authors': ['Jiajun Fan', 'Yuzheng Zhuang', 'Yuecheng Liu', 'Jianye Hao', 'Bin Wang', 'Jiangcheng Zhu', 'Hao Wang', 'Shu-Tao Xia'], 'published': '2023-05-09 08:00:23+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2305.05239v1'}
42,454
2207.02189v2
2,207.02189
2
Accelerating Hamiltonian Monte Carlo via Chebyshev Integration Time
Hamiltonian Monte Carlo (HMC) is a popular method in sampling. While there are quite a few works of studying this method on various aspects, an interesting question is how to choose its integration time to achieve acceleration. In this work, we consider accelerating the process of sampling from a distribution $(x) (-f(x))$ via HMC via time-varying integration time. When the potential $f$ is $L$-smooth and $m$-strongly convex, i.e.\ for sampling from a log-smooth and strongly log-concave target distribution $$, it is known that under a constant integration time, the number of iterations that ideal HMC takes to get an $$ Wasserstein-2 distance to the target $$ is $O( 1 )$, where $:= Lm$ is the condition number. We propose a scheme of time-varying integration time based on the roots of Chebyshev polynomials. We show that in the case of quadratic potential $f$, i.e., when the target $$ is a Gaussian distribution, ideal HMC with this choice of integration time only takes $O( 1 )$ number of iterations to reach Wasserstein-2 distance less than $$; this improvement on the dependence on condition number is akin to acceleration in optimization. The design and analysis of HMC with the proposed integration time is built on the tools of Chebyshev polynomials. Experiments find the advantage of adopting our scheme of time-varying integration time even for sampling from distributions with smooth strongly convex potentials that are not quadratic.
2022-07-05
{'id': '2207.02189v2', 'title': 'Accelerating Hamiltonian Monte Carlo via Chebyshev Integration Time', 'abstract': 'Hamiltonian Monte Carlo (HMC) is a popular method in sampling. While there\nare quite a few works of studying this method on various aspects, an\ninteresting question is how to choose its integration time to achieve\nacceleration. In this work, we consider accelerating the process of sampling\nfrom a distribution $\\pi(x) \\propto \\exp(-f(x))$ via HMC via time-varying\nintegration time. When the potential $f$ is $L$-smooth and $m$-strongly convex,\ni.e.\\ for sampling from a log-smooth and strongly log-concave target\ndistribution $\\pi$, it is known that under a constant integration time, the\nnumber of iterations that ideal HMC takes to get an $\\epsilon$ Wasserstein-2\ndistance to the target $\\pi$ is $O( \\kappa \\log \\frac{1}{\\epsilon} )$, where\n$\\kappa := \\frac{L}{m}$ is the condition number. We propose a scheme of\ntime-varying integration time based on the roots of Chebyshev polynomials. We\nshow that in the case of quadratic potential $f$, i.e., when the target $\\pi$\nis a Gaussian distribution, ideal HMC with this choice of integration time only\ntakes $O( \\sqrt{\\kappa} \\log \\frac{1}{\\epsilon} )$ number of iterations to\nreach Wasserstein-2 distance less than $\\epsilon$; this improvement on the\ndependence on condition number is akin to acceleration in optimization. The\ndesign and analysis of HMC with the proposed integration time is built on the\ntools of Chebyshev polynomials. Experiments find the advantage of adopting our\nscheme of time-varying integration time even for sampling from distributions\nwith smooth strongly convex potentials that are not quadratic.', 'authors': ['Jun-Kun Wang', 'Andre Wibisono'], 'published': '2022-07-05 17:42:22+00:00', 'categories': ['cs.LG', 'stat.CO'], 'url': 'http://arxiv.org/abs/2207.02189v2'}
42,455
2302.14665v1
2,302.14665
1
Parametrizing Product Shape Manifolds by Composite Networks
Parametrizations of data manifolds in shape spaces can be computed using the rich toolbox of Riemannian geometry. This, however, often comes with high computational costs, which raises the question if one can learn an efficient neural network approximation. We show that this is indeed possible for shape spaces with a special product structure, namely those smoothly approximable by a direct sum of low-dimensional manifolds. Our proposed architecture leverages this structure by separately learning approximations for the low-dimensional factors and a subsequent combination. After developing the approach as a general framework, we apply it to a shape space of triangular surfaces. Here, typical examples of data manifolds are given through datasets of articulated models and can be factorized, for example, by a Sparse Principal Geodesic Analysis (SPGA). We demonstrate the effectiveness of our proposed approach with experiments on synthetic data as well as manifolds extracted from data via SPGA.
2023-02-28
{'id': '2302.14665v1', 'title': 'Parametrizing Product Shape Manifolds by Composite Networks', 'abstract': 'Parametrizations of data manifolds in shape spaces can be computed using the\nrich toolbox of Riemannian geometry. This, however, often comes with high\ncomputational costs, which raises the question if one can learn an efficient\nneural network approximation. We show that this is indeed possible for shape\nspaces with a special product structure, namely those smoothly approximable by\na direct sum of low-dimensional manifolds. Our proposed architecture leverages\nthis structure by separately learning approximations for the low-dimensional\nfactors and a subsequent combination. After developing the approach as a\ngeneral framework, we apply it to a shape space of triangular surfaces. Here,\ntypical examples of data manifolds are given through datasets of articulated\nmodels and can be factorized, for example, by a Sparse Principal Geodesic\nAnalysis (SPGA). We demonstrate the effectiveness of our proposed approach with\nexperiments on synthetic data as well as manifolds extracted from data via\nSPGA.', 'authors': ['Josua Sassen', 'Klaus Hildebrandt', 'Martin Rumpf', 'Benedikt Wirth'], 'published': '2023-02-28 15:31:23+00:00', 'categories': ['cs.LG', 'cs.CV', 'cs.GR', 'math.DG', '68T07, 65D18, 65D15'], 'url': 'http://arxiv.org/abs/2302.14665v1'}
42,456
2202.00395v2
2,202.00395
2
Is the Performance of My Deep Network Too Good to Be True? A Direct Approach to Estimating the Bayes Error in Binary Classification
There is a fundamental limitation in the prediction performance that a machine learning model can achieve due to the inevitable uncertainty of the prediction target. In classification problems, this can be characterized by the Bayes error, which is the best achievable error with any classifier. The Bayes error can be used as a criterion to evaluate classifiers with state-of-the-art performance and can be used to detect test set overfitting. We propose a simple and direct Bayes error estimator, where we just take the mean of the labels that show uncertainty of the class assignments. Our flexible approach enables us to perform Bayes error estimation even for weakly supervised data. In contrast to others, our method is model-free and even instance-free. Moreover, it has no hyperparameters and gives a more accurate estimate of the Bayes error than several baselines empirically. Experiments using our method suggest that recently proposed deep networks such as the Vision Transformer may have reached, or is about to reach, the Bayes error for benchmark datasets. Finally, we discuss how we can study the inherent difficulty of the acceptance/rejection decision for scientific articles, by estimating the Bayes error of the ICLR papers from 2017 to 2023.
2022-02-01
{'id': '2202.00395v2', 'title': 'Is the Performance of My Deep Network Too Good to Be True? A Direct Approach to Estimating the Bayes Error in Binary Classification', 'abstract': 'There is a fundamental limitation in the prediction performance that a\nmachine learning model can achieve due to the inevitable uncertainty of the\nprediction target. In classification problems, this can be characterized by the\nBayes error, which is the best achievable error with any classifier. The Bayes\nerror can be used as a criterion to evaluate classifiers with state-of-the-art\nperformance and can be used to detect test set overfitting. We propose a simple\nand direct Bayes error estimator, where we just take the mean of the labels\nthat show \\emph{uncertainty} of the class assignments. Our flexible approach\nenables us to perform Bayes error estimation even for weakly supervised data.\nIn contrast to others, our method is model-free and even instance-free.\nMoreover, it has no hyperparameters and gives a more accurate estimate of the\nBayes error than several baselines empirically. Experiments using our method\nsuggest that recently proposed deep networks such as the Vision Transformer may\nhave reached, or is about to reach, the Bayes error for benchmark datasets.\nFinally, we discuss how we can study the inherent difficulty of the\nacceptance/rejection decision for scientific articles, by estimating the Bayes\nerror of the ICLR papers from 2017 to 2023.', 'authors': ['Takashi Ishida', 'Ikko Yamane', 'Nontawat Charoenphakdee', 'Gang Niu', 'Masashi Sugiyama'], 'published': '2022-02-01 13:22:26+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2202.00395v2'}
42,458
2207.08821v1
2,207.08821
1
The Multiple Subnetwork Hypothesis: Enabling Multidomain Learning by Isolating Task-Specific Subnetworks in Feedforward Neural Networks
Neural networks have seen an explosion of usage and research in the past decade, particularly within the domains of computer vision and natural language processing. However, only recently have advancements in neural networks yielded performance improvements beyond narrow applications and translated to expanded multitask models capable of generalizing across multiple data types and modalities. Simultaneously, it has been shown that neural networks are overparameterized to a high degree, and pruning techniques have proved capable of significantly reducing the number of active weights within the network while largely preserving performance. In this work, we identify a methodology and network representational structure which allows a pruned network to employ previously unused weights to learn subsequent tasks. We employ these methodologies on well-known benchmarking datasets for testing purposes and show that networks trained using our approaches are able to learn multiple tasks, which may be related or unrelated, in parallel or in sequence without sacrificing performance on any task or exhibiting catastrophic forgetting.
2022-07-18
{'id': '2207.08821v1', 'title': 'The Multiple Subnetwork Hypothesis: Enabling Multidomain Learning by Isolating Task-Specific Subnetworks in Feedforward Neural Networks', 'abstract': 'Neural networks have seen an explosion of usage and research in the past\ndecade, particularly within the domains of computer vision and natural language\nprocessing. However, only recently have advancements in neural networks yielded\nperformance improvements beyond narrow applications and translated to expanded\nmultitask models capable of generalizing across multiple data types and\nmodalities. Simultaneously, it has been shown that neural networks are\noverparameterized to a high degree, and pruning techniques have proved capable\nof significantly reducing the number of active weights within the network while\nlargely preserving performance. In this work, we identify a methodology and\nnetwork representational structure which allows a pruned network to employ\npreviously unused weights to learn subsequent tasks. We employ these\nmethodologies on well-known benchmarking datasets for testing purposes and show\nthat networks trained using our approaches are able to learn multiple tasks,\nwhich may be related or unrelated, in parallel or in sequence without\nsacrificing performance on any task or exhibiting catastrophic forgetting.', 'authors': ['Jacob Renn', 'Ian Sotnek', 'Benjamin Harvey', 'Brian Caffo'], 'published': '2022-07-18 15:07:13+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2207.08821v1'}
42,450
2312.11954v2
2,312.11954
2
Adversarial AutoMixup
Data mixing augmentation has been widely applied to improve the generalization ability of deep neural networks. Recently, offline data mixing augmentation, e.g. handcrafted and saliency information-based mixup, has been gradually replaced by automatic mixing approaches. Through minimizing two sub-tasks, namely, mixed sample generation and mixup classification in an end-to-end way, AutoMix significantly improves accuracy on image classification tasks. However, as the optimization objective is consistent for the two sub-tasks, this approach is prone to generating consistent instead of diverse mixed samples, which results in overfitting for target task training. In this paper, we propose AdAutomixup, an adversarial automatic mixup augmentation approach that generates challenging samples to train a robust classifier for image classification, by alternatively optimizing the classifier and the mixup sample generator. AdAutomixup comprises two modules, a mixed example generator, and a target classifier. The mixed sample generator aims to produce hard mixed examples to challenge the target classifier, while the target classifier's aim is to learn robust features from hard mixed examples to improve generalization. To prevent the collapse of the inherent meanings of images, we further introduce an exponential moving average (EMA) teacher and cosine similarity to train AdAutomixup in an end-to-end way. Extensive experiments on seven image benchmarks consistently prove that our approach outperforms the state of the art in various classification scenarios. The source code is available at https://github.com/JinXins/Adversarial-AutoMixup.
2023-12-19
{'id': '2312.11954v2', 'title': 'Adversarial AutoMixup', 'abstract': "Data mixing augmentation has been widely applied to improve the\ngeneralization ability of deep neural networks. Recently, offline data mixing\naugmentation, e.g. handcrafted and saliency information-based mixup, has been\ngradually replaced by automatic mixing approaches. Through minimizing two\nsub-tasks, namely, mixed sample generation and mixup classification in an\nend-to-end way, AutoMix significantly improves accuracy on image classification\ntasks. However, as the optimization objective is consistent for the two\nsub-tasks, this approach is prone to generating consistent instead of diverse\nmixed samples, which results in overfitting for target task training. In this\npaper, we propose AdAutomixup, an adversarial automatic mixup augmentation\napproach that generates challenging samples to train a robust classifier for\nimage classification, by alternatively optimizing the classifier and the mixup\nsample generator. AdAutomixup comprises two modules, a mixed example generator,\nand a target classifier. The mixed sample generator aims to produce hard mixed\nexamples to challenge the target classifier, while the target classifier's aim\nis to learn robust features from hard mixed examples to improve generalization.\nTo prevent the collapse of the inherent meanings of images, we further\nintroduce an exponential moving average (EMA) teacher and cosine similarity to\ntrain AdAutomixup in an end-to-end way. Extensive experiments on seven image\nbenchmarks consistently prove that our approach outperforms the state of the\nart in various classification scenarios. The source code is available at\nhttps://github.com/JinXins/Adversarial-AutoMixup.", 'authors': ['Huafeng Qin', 'Xin Jin', 'Yun Jiang', 'Mounim A. El-Yacoubi', 'Xinbo Gao'], 'published': '2023-12-19 08:55:00+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2312.11954v2'}
42,476
2306.04974v2
2,306.04974
2
Conservative Prediction via Data-Driven Confidence Minimization
In safety-critical applications of machine learning, it is often desirable for a model to be conservative, abstaining from making predictions on unknown inputs which are not well-represented in the training data. However, detecting unknown examples is challenging, as it is impossible to anticipate all potential inputs at test time. To address this, prior work (Hendrycks et al., 2018) minimizes model confidence on an auxiliary outlier dataset carefully curated to be disjoint from the training distribution. We theoretically analyze the choice of auxiliary dataset for confidence minimization, revealing two actionable insights: (1) if the auxiliary set contains unknown examples similar to those seen at test time, confidence minimization leads to provable detection of unknown test examples, and (2) if the first condition is satisfied, it is unnecessary to filter out known examples for out-of-distribution (OOD) detection. Motivated by these guidelines, we propose the Data-Driven Confidence Minimization (DCM) framework, which minimizes confidence on an uncertainty dataset. We apply DCM to two problem settings in which conservative prediction is paramount -- selective classification and OOD detection -- and provide a realistic way to gather uncertainty data for each setting. In our experiments, DCM consistently outperforms existing selective classification approaches on 4 datasets when tested on unseen distributions and outperforms state-of-the-art OOD detection methods on 12 ID-OOD dataset pairs, reducing FPR (at TPR $95\%$) by $6.3\%$ and $58.1\%$ on CIFAR-10 and CIFAR-100 compared to Outlier Exposure.
2023-06-08
{'id': '2306.04974v2', 'title': 'Conservative Prediction via Data-Driven Confidence Minimization', 'abstract': 'In safety-critical applications of machine learning, it is often desirable\nfor a model to be conservative, abstaining from making predictions on unknown\ninputs which are not well-represented in the training data. However, detecting\nunknown examples is challenging, as it is impossible to anticipate all\npotential inputs at test time. To address this, prior work (Hendrycks et al.,\n2018) minimizes model confidence on an auxiliary outlier dataset carefully\ncurated to be disjoint from the training distribution. We theoretically analyze\nthe choice of auxiliary dataset for confidence minimization, revealing two\nactionable insights: (1) if the auxiliary set contains unknown examples similar\nto those seen at test time, confidence minimization leads to provable detection\nof unknown test examples, and (2) if the first condition is satisfied, it is\nunnecessary to filter out known examples for out-of-distribution (OOD)\ndetection. Motivated by these guidelines, we propose the Data-Driven Confidence\nMinimization (DCM) framework, which minimizes confidence on an uncertainty\ndataset. We apply DCM to two problem settings in which conservative prediction\nis paramount -- selective classification and OOD detection -- and provide a\nrealistic way to gather uncertainty data for each setting. In our experiments,\nDCM consistently outperforms existing selective classification approaches on 4\ndatasets when tested on unseen distributions and outperforms state-of-the-art\nOOD detection methods on 12 ID-OOD dataset pairs, reducing FPR (at TPR $95\\%$)\nby $6.3\\%$ and $58.1\\%$ on CIFAR-10 and CIFAR-100 compared to Outlier Exposure.', 'authors': ['Caroline Choi', 'Fahim Tajwar', 'Yoonho Lee', 'Huaxiu Yao', 'Ananya Kumar', 'Chelsea Finn'], 'published': '2023-06-08 07:05:36+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2306.04974v2'}
42,479
2309.14348v3
2,309.14348
3
Defending Against Alignment-Breaking Attacks via Robustly Aligned LLM
Recently, Large Language Models (LLMs) have made significant advancements and are now widely used across various domains. Unfortunately, there has been a rising concern that LLMs can be misused to generate harmful or malicious content. Though a line of research has focused on aligning LLMs with human values and preventing them from producing inappropriate content, such alignments are usually vulnerable and can be bypassed by alignment-breaking attacks via adversarially optimized or handcrafted jailbreaking prompts. In this work, we introduce a Robustly Aligned LLM (RA-LLM) to defend against potential alignment-breaking attacks. RA-LLM can be directly constructed upon an existing aligned LLM with a robust alignment checking function, without requiring any expensive retraining or fine-tuning process of the original LLM. Furthermore, we also provide a theoretical analysis for RA-LLM to verify its effectiveness in defending against alignment-breaking attacks. Through real-world experiments on open-source large language models, we demonstrate that RA-LLM can successfully defend against both state-of-the-art adversarial prompts and popular handcrafted jailbreaking prompts by reducing their attack success rates from nearly 100% to around 10% or less.
2023-09-18
{'id': '2309.14348v3', 'title': 'Defending Against Alignment-Breaking Attacks via Robustly Aligned LLM', 'abstract': 'Recently, Large Language Models (LLMs) have made significant advancements and\nare now widely used across various domains. Unfortunately, there has been a\nrising concern that LLMs can be misused to generate harmful or malicious\ncontent. Though a line of research has focused on aligning LLMs with human\nvalues and preventing them from producing inappropriate content, such\nalignments are usually vulnerable and can be bypassed by alignment-breaking\nattacks via adversarially optimized or handcrafted jailbreaking prompts. In\nthis work, we introduce a Robustly Aligned LLM (RA-LLM) to defend against\npotential alignment-breaking attacks. RA-LLM can be directly constructed upon\nan existing aligned LLM with a robust alignment checking function, without\nrequiring any expensive retraining or fine-tuning process of the original LLM.\nFurthermore, we also provide a theoretical analysis for RA-LLM to verify its\neffectiveness in defending against alignment-breaking attacks. Through\nreal-world experiments on open-source large language models, we demonstrate\nthat RA-LLM can successfully defend against both state-of-the-art adversarial\nprompts and popular handcrafted jailbreaking prompts by reducing their attack\nsuccess rates from nearly 100% to around 10% or less.', 'authors': ['Bochuan Cao', 'Yuanpu Cao', 'Lu Lin', 'Jinghui Chen'], 'published': '2023-09-18 02:07:22+00:00', 'categories': ['cs.CL', 'cs.AI', 'cs.CR', 'cs.LG'], 'url': 'http://arxiv.org/abs/2309.14348v3'}
42,480
2310.02664v2
2,310.02664
2
On Memorization in Diffusion Models
Due to their capacity to generate novel and high-quality samples, diffusion models have attracted significant research interest in recent years. Notably, the typical training objective of diffusion models, i.e., denoising score matching, has a closed-form optimal solution that can only generate training data replicating samples. This indicates that a memorization behavior is theoretically expected, which contradicts the common generalization ability of state-of-the-art diffusion models, and thus calls for a deeper understanding. Looking into this, we first observe that memorization behaviors tend to occur on smaller-sized datasets, which motivates our definition of effective model memorization (EMM), a metric measuring the maximum size of training data at which a learned diffusion model approximates its theoretical optimum. Then, we quantify the impact of the influential factors on these memorization behaviors in terms of EMM, focusing primarily on data distribution, model configuration, and training procedure. Besides comprehensive empirical results identifying the influential factors, we surprisingly find that conditioning training data on uninformative random labels can significantly trigger the memorization in diffusion models. Our study holds practical significance for diffusion model users and offers clues to theoretical research in deep generative models. Code is available at https://github.com/sail-sg/DiffMemorize.
2023-10-04
{'id': '2310.02664v2', 'title': 'On Memorization in Diffusion Models', 'abstract': 'Due to their capacity to generate novel and high-quality samples, diffusion\nmodels have attracted significant research interest in recent years. Notably,\nthe typical training objective of diffusion models, i.e., denoising score\nmatching, has a closed-form optimal solution that can only generate training\ndata replicating samples. This indicates that a memorization behavior is\ntheoretically expected, which contradicts the common generalization ability of\nstate-of-the-art diffusion models, and thus calls for a deeper understanding.\nLooking into this, we first observe that memorization behaviors tend to occur\non smaller-sized datasets, which motivates our definition of effective model\nmemorization (EMM), a metric measuring the maximum size of training data at\nwhich a learned diffusion model approximates its theoretical optimum. Then, we\nquantify the impact of the influential factors on these memorization behaviors\nin terms of EMM, focusing primarily on data distribution, model configuration,\nand training procedure. Besides comprehensive empirical results identifying the\ninfluential factors, we surprisingly find that conditioning training data on\nuninformative random labels can significantly trigger the memorization in\ndiffusion models. Our study holds practical significance for diffusion model\nusers and offers clues to theoretical research in deep generative models. Code\nis available at https://github.com/sail-sg/DiffMemorize.', 'authors': ['Xiangming Gu', 'Chao Du', 'Tianyu Pang', 'Chongxuan Li', 'Min Lin', 'Ye Wang'], 'published': '2023-10-04 09:04:20+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CV'], 'url': 'http://arxiv.org/abs/2310.02664v2'}
42,482
2310.02619v2
2,310.02619
2
Generative Modeling of Regular and Irregular Time Series Data via Koopman VAEs
Generating realistic time series data is important for many engineering and scientific applications. Existing work tackles this problem using generative adversarial networks (GANs). However, GANs are unstable during training, and they can suffer from mode collapse. While variational autoencoders (VAEs) are known to be more robust to the these issues, they are (surprisingly) less considered for time series generation. In this work, we introduce Koopman VAE (KoVAE), a new generative framework that is based on a novel design for the model prior, and that can be optimized for either regular and irregular training data. Inspired by Koopman theory, we represent the latent conditional prior dynamics using a linear map. Our approach enhances generative modeling with two desired features: (i) incorporating domain knowledge can be achieved by leveraging spectral tools that prescribe constraints on the eigenvalues of the linear map; and (ii) studying the qualitative behavior and stability of the system can be performed using tools from dynamical systems theory. Our results show that KoVAE outperforms state-of-the-art GAN and VAE methods across several challenging synthetic and real-world time series generation benchmarks. Whether trained on regular or irregular data, KoVAE generates time series that improve both discriminative and predictive metrics. We also present visual evidence suggesting that KoVAE learns probability density functions that better approximate the empirical ground truth distribution.
2023-10-04
{'id': '2310.02619v2', 'title': 'Generative Modeling of Regular and Irregular Time Series Data via Koopman VAEs', 'abstract': 'Generating realistic time series data is important for many engineering and\nscientific applications. Existing work tackles this problem using generative\nadversarial networks (GANs). However, GANs are unstable during training, and\nthey can suffer from mode collapse. While variational autoencoders (VAEs) are\nknown to be more robust to the these issues, they are (surprisingly) less\nconsidered for time series generation. In this work, we introduce Koopman VAE\n(KoVAE), a new generative framework that is based on a novel design for the\nmodel prior, and that can be optimized for either regular and irregular\ntraining data. Inspired by Koopman theory, we represent the latent conditional\nprior dynamics using a linear map. Our approach enhances generative modeling\nwith two desired features: (i) incorporating domain knowledge can be achieved\nby leveraging spectral tools that prescribe constraints on the eigenvalues of\nthe linear map; and (ii) studying the qualitative behavior and stability of the\nsystem can be performed using tools from dynamical systems theory. Our results\nshow that KoVAE outperforms state-of-the-art GAN and VAE methods across several\nchallenging synthetic and real-world time series generation benchmarks. Whether\ntrained on regular or irregular data, KoVAE generates time series that improve\nboth discriminative and predictive metrics. We also present visual evidence\nsuggesting that KoVAE learns probability density functions that better\napproximate the empirical ground truth distribution.', 'authors': ['Ilan Naiman', 'N. Benjamin Erichson', 'Pu Ren', 'Michael W. Mahoney', 'Omri Azencot'], 'published': '2023-10-04 07:14:43+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.02619v2'}
42,464
2402.13241v2
2,402.13241
2
Federated Causal Discovery from Heterogeneous Data
Conventional causal discovery methods rely on centralized data, which is inconsistent with the decentralized nature of data in many real-world situations. This discrepancy has motivated the development of federated causal discovery (FCD) approaches. However, existing FCD methods may be limited by their potentially restrictive assumptions of identifiable functional causal models or homogeneous data distributions, narrowing their applicability in diverse scenarios. In this paper, we propose a novel FCD method attempting to accommodate arbitrary causal models and heterogeneous data. We first utilize a surrogate variable corresponding to the client index to account for the data heterogeneity across different clients. We then develop a federated conditional independence test (FCIT) for causal skeleton discovery and establish a federated independent change principle (FICP) to determine causal directions. These approaches involve constructing summary statistics as a proxy of the raw data to protect data privacy. Owing to the nonparametric properties, FCIT and FICP make no assumption about particular functional forms, thereby facilitating the handling of arbitrary causal models. We conduct extensive experiments on synthetic and real datasets to show the efficacy of our method. The code is available at https://github.com/lokali/FedCDH.git.
2024-02-20
{'id': '2402.13241v2', 'title': 'Federated Causal Discovery from Heterogeneous Data', 'abstract': 'Conventional causal discovery methods rely on centralized data, which is\ninconsistent with the decentralized nature of data in many real-world\nsituations. This discrepancy has motivated the development of federated causal\ndiscovery (FCD) approaches. However, existing FCD methods may be limited by\ntheir potentially restrictive assumptions of identifiable functional causal\nmodels or homogeneous data distributions, narrowing their applicability in\ndiverse scenarios. In this paper, we propose a novel FCD method attempting to\naccommodate arbitrary causal models and heterogeneous data. We first utilize a\nsurrogate variable corresponding to the client index to account for the data\nheterogeneity across different clients. We then develop a federated conditional\nindependence test (FCIT) for causal skeleton discovery and establish a\nfederated independent change principle (FICP) to determine causal directions.\nThese approaches involve constructing summary statistics as a proxy of the raw\ndata to protect data privacy. Owing to the nonparametric properties, FCIT and\nFICP make no assumption about particular functional forms, thereby facilitating\nthe handling of arbitrary causal models. We conduct extensive experiments on\nsynthetic and real datasets to show the efficacy of our method. The code is\navailable at https://github.com/lokali/FedCDH.git.', 'authors': ['Loka Li', 'Ignavier Ng', 'Gongxu Luo', 'Biwei Huang', 'Guangyi Chen', 'Tongliang Liu', 'Bin Gu', 'Kun Zhang'], 'published': '2024-02-20 18:53:53+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2402.13241v2'}
42,472
2402.12874v1
2,402.12874
1
Skill or Luck? Return Decomposition via Advantage Functions
Learning from off-policy data is essential for sample-efficient reinforcement learning. In the present work, we build on the insight that the advantage function can be understood as the causal effect of an action on the return, and show that this allows us to decompose the return of a trajectory into parts caused by the agent's actions (skill) and parts outside of the agent's control (luck). Furthermore, this decomposition enables us to naturally extend Direct Advantage Estimation (DAE) to off-policy settings (Off-policy DAE). The resulting method can learn from off-policy trajectories without relying on importance sampling techniques or truncating off-policy actions. We draw connections between Off-policy DAE and previous methods to demonstrate how it can speed up learning and when the proposed off-policy corrections are important. Finally, we use the MinAtar environments to illustrate how ignoring off-policy corrections can lead to suboptimal policy optimization performance.
2024-02-20
{'id': '2402.12874v1', 'title': 'Skill or Luck? Return Decomposition via Advantage Functions', 'abstract': "Learning from off-policy data is essential for sample-efficient reinforcement\nlearning. In the present work, we build on the insight that the advantage\nfunction can be understood as the causal effect of an action on the return, and\nshow that this allows us to decompose the return of a trajectory into parts\ncaused by the agent's actions (skill) and parts outside of the agent's control\n(luck). Furthermore, this decomposition enables us to naturally extend Direct\nAdvantage Estimation (DAE) to off-policy settings (Off-policy DAE). The\nresulting method can learn from off-policy trajectories without relying on\nimportance sampling techniques or truncating off-policy actions. We draw\nconnections between Off-policy DAE and previous methods to demonstrate how it\ncan speed up learning and when the proposed off-policy corrections are\nimportant. Finally, we use the MinAtar environments to illustrate how ignoring\noff-policy corrections can lead to suboptimal policy optimization performance.", 'authors': ['Hsiao-Ru Pan', 'Bernhard Schölkopf'], 'published': '2024-02-20 10:09:00+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2402.12874v1'}
42,528
2305.16310v1
2,305.1631
1
Securing Deep Generative Models with Universal Adversarial Signature
Recent advances in deep generative models have led to the development of methods capable of synthesizing high-quality, realistic images. These models pose threats to society due to their potential misuse. Prior research attempted to mitigate these threats by detecting generated images, but the varying traces left by different generative models make it challenging to create a universal detector capable of generalizing to new, unseen generative models. In this paper, we propose to inject a universal adversarial signature into an arbitrary pre-trained generative model, in order to make its generated contents more detectable and traceable. First, the imperceptible optimal signature for each image can be found by a signature injector through adversarial training. Subsequently, the signature can be incorporated into an arbitrary generator by fine-tuning it with the images processed by the signature injector. In this way, the detector corresponding to the signature can be reused for any fine-tuned generator for tracking the generator identity. The proposed method is validated on the FFHQ and ImageNet datasets with various state-of-the-art generative models, consistently showing a promising detection rate. Code will be made publicly available at https://github.com/zengxianyu/genwm.
2023-05-25
{'id': '2305.16310v1', 'title': 'Securing Deep Generative Models with Universal Adversarial Signature', 'abstract': 'Recent advances in deep generative models have led to the development of\nmethods capable of synthesizing high-quality, realistic images. These models\npose threats to society due to their potential misuse. Prior research attempted\nto mitigate these threats by detecting generated images, but the varying traces\nleft by different generative models make it challenging to create a universal\ndetector capable of generalizing to new, unseen generative models. In this\npaper, we propose to inject a universal adversarial signature into an arbitrary\npre-trained generative model, in order to make its generated contents more\ndetectable and traceable. First, the imperceptible optimal signature for each\nimage can be found by a signature injector through adversarial training.\nSubsequently, the signature can be incorporated into an arbitrary generator by\nfine-tuning it with the images processed by the signature injector. In this\nway, the detector corresponding to the signature can be reused for any\nfine-tuned generator for tracking the generator identity. The proposed method\nis validated on the FFHQ and ImageNet datasets with various state-of-the-art\ngenerative models, consistently showing a promising detection rate. Code will\nbe made publicly available at \\url{https://github.com/zengxianyu/genwm}.', 'authors': ['Yu Zeng', 'Mo Zhou', 'Yuan Xue', 'Vishal M. Patel'], 'published': '2023-05-25 17:59:01+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2305.16310v1'}
42,530
2310.02233v2
2,310.02233
2
Generalized Schrödinger Bridge Matching
Modern distribution matching algorithms for training diffusion or flow models directly prescribe the time evolution of the marginal distributions between two boundary distributions. In this work, we consider a generalized distribution matching setup, where these marginals are only implicitly described as a solution to some task-specific objective function. The problem setup, known as the Generalized Schr\"odinger Bridge (GSB), appears prevalently in many scientific areas both within and without machine learning. We propose Generalized Schr\"odinger Bridge Matching (GSBM), a new matching algorithm inspired by recent advances, generalizing them beyond kinetic energy minimization and to account for task-specific state costs. We show that such a generalization can be cast as solving conditional stochastic optimal control, for which efficient variational approximations can be used, and further debiased with the aid of path integral theory. Compared to prior methods for solving GSB problems, our GSBM algorithm better preserves a feasible transport map between the boundary distributions throughout training, thereby enabling stable convergence and significantly improved scalability. We empirically validate our claims on an extensive suite of experimental setups, including crowd navigation, opinion depolarization, LiDAR manifolds, and image domain transfer. Our work brings new algorithmic opportunities for training diffusion models enhanced with task-specific optimality structures. Code available at https://github.com/facebookresearch/generalized-schrodinger-bridge-matching
2023-10-03
{'id': '2310.02233v2', 'title': 'Generalized Schrödinger Bridge Matching', 'abstract': 'Modern distribution matching algorithms for training diffusion or flow models\ndirectly prescribe the time evolution of the marginal distributions between two\nboundary distributions. In this work, we consider a generalized distribution\nmatching setup, where these marginals are only implicitly described as a\nsolution to some task-specific objective function. The problem setup, known as\nthe Generalized Schr\\"odinger Bridge (GSB), appears prevalently in many\nscientific areas both within and without machine learning. We propose\nGeneralized Schr\\"odinger Bridge Matching (GSBM), a new matching algorithm\ninspired by recent advances, generalizing them beyond kinetic energy\nminimization and to account for task-specific state costs. We show that such a\ngeneralization can be cast as solving conditional stochastic optimal control,\nfor which efficient variational approximations can be used, and further\ndebiased with the aid of path integral theory. Compared to prior methods for\nsolving GSB problems, our GSBM algorithm better preserves a feasible transport\nmap between the boundary distributions throughout training, thereby enabling\nstable convergence and significantly improved scalability. We empirically\nvalidate our claims on an extensive suite of experimental setups, including\ncrowd navigation, opinion depolarization, LiDAR manifolds, and image domain\ntransfer. Our work brings new algorithmic opportunities for training diffusion\nmodels enhanced with task-specific optimality structures. Code available at\nhttps://github.com/facebookresearch/generalized-schrodinger-bridge-matching', 'authors': ['Guan-Horng Liu', 'Yaron Lipman', 'Maximilian Nickel', 'Brian Karrer', 'Evangelos A. Theodorou', 'Ricky T. Q. Chen'], 'published': '2023-10-03 17:42:11+00:00', 'categories': ['stat.ML', 'cs.LG', 'math.OC'], 'url': 'http://arxiv.org/abs/2310.02233v2'}
42,534
2405.17404v1
2,405.17404
1
Spectral Greedy Coresets for Graph Neural Networks
The ubiquity of large-scale graphs in node-classification tasks significantly hinders the real-world applications of Graph Neural Networks (GNNs). Node sampling, graph coarsening, and dataset condensation are effective strategies for enhancing data efficiency. However, owing to the interdependence of graph nodes, coreset selection, which selects subsets of the data examples, has not been successfully applied to speed up GNN training on large graphs, warranting special treatment. This paper studies graph coresets for GNNs and avoids the interdependence issue by selecting ego-graphs (i.e., neighborhood subgraphs around a node) based on their spectral embeddings. We decompose the coreset selection problem for GNNs into two phases: a coarse selection of widely spread ego graphs and a refined selection to diversify their topologies. We design a greedy algorithm that approximately optimizes both objectives. Our spectral greedy graph coreset (SGGC) scales to graphs with millions of nodes, obviates the need for model pre-training, and applies to low-homophily graphs. Extensive experiments on ten datasets demonstrate that SGGC outperforms other coreset methods by a wide margin, generalizes well across GNN architectures, and is much faster than graph condensation.
2024-05-27
{'id': '2405.17404v1', 'title': 'Spectral Greedy Coresets for Graph Neural Networks', 'abstract': 'The ubiquity of large-scale graphs in node-classification tasks significantly\nhinders the real-world applications of Graph Neural Networks (GNNs). Node\nsampling, graph coarsening, and dataset condensation are effective strategies\nfor enhancing data efficiency. However, owing to the interdependence of graph\nnodes, coreset selection, which selects subsets of the data examples, has not\nbeen successfully applied to speed up GNN training on large graphs, warranting\nspecial treatment. This paper studies graph coresets for GNNs and avoids the\ninterdependence issue by selecting ego-graphs (i.e., neighborhood subgraphs\naround a node) based on their spectral embeddings. We decompose the coreset\nselection problem for GNNs into two phases: a coarse selection of widely spread\nego graphs and a refined selection to diversify their topologies. We design a\ngreedy algorithm that approximately optimizes both objectives. Our spectral\ngreedy graph coreset (SGGC) scales to graphs with millions of nodes, obviates\nthe need for model pre-training, and applies to low-homophily graphs. Extensive\nexperiments on ten datasets demonstrate that SGGC outperforms other coreset\nmethods by a wide margin, generalizes well across GNN architectures, and is\nmuch faster than graph condensation.', 'authors': ['Mucong Ding', 'Yinhan He', 'Jundong Li', 'Furong Huang'], 'published': '2024-05-27 17:52:12+00:00', 'categories': ['cs.LG', 'cs.AI', 'stat.ML'], 'url': 'http://arxiv.org/abs/2405.17404v1'}
42,565
2101.05982v2
2,101.05982
2
Randomized Ensembled Double Q-Learning: Learning Fast Without a Model
Using a high Update-To-Data (UTD) ratio, model-based methods have recently achieved much higher sample efficiency than previous model-free methods for continuous-action DRL benchmarks. In this paper, we introduce a simple model-free algorithm, Randomized Ensembled Double Q-Learning (REDQ), and show that its performance is just as good as, if not better than, a state-of-the-art model-based algorithm for the MuJoCo benchmark. Moreover, REDQ can achieve this performance using fewer parameters than the model-based method, and with less wall-clock run time. REDQ has three carefully integrated ingredients which allow it to achieve its high performance: (i) a UTD ratio >> 1; (ii) an ensemble of Q functions; (iii) in-target minimization across a random subset of Q functions from the ensemble. Through carefully designed experiments, we provide a detailed analysis of REDQ and related model-free algorithms. To our knowledge, REDQ is the first successful model-free DRL algorithm for continuous-action spaces using a UTD ratio >> 1.
2021-01-15
{'id': '2101.05982v2', 'title': 'Randomized Ensembled Double Q-Learning: Learning Fast Without a Model', 'abstract': 'Using a high Update-To-Data (UTD) ratio, model-based methods have recently\nachieved much higher sample efficiency than previous model-free methods for\ncontinuous-action DRL benchmarks. In this paper, we introduce a simple\nmodel-free algorithm, Randomized Ensembled Double Q-Learning (REDQ), and show\nthat its performance is just as good as, if not better than, a state-of-the-art\nmodel-based algorithm for the MuJoCo benchmark. Moreover, REDQ can achieve this\nperformance using fewer parameters than the model-based method, and with less\nwall-clock run time. REDQ has three carefully integrated ingredients which\nallow it to achieve its high performance: (i) a UTD ratio >> 1; (ii) an\nensemble of Q functions; (iii) in-target minimization across a random subset of\nQ functions from the ensemble. Through carefully designed experiments, we\nprovide a detailed analysis of REDQ and related model-free algorithms. To our\nknowledge, REDQ is the first successful model-free DRL algorithm for\ncontinuous-action spaces using a UTD ratio >> 1.', 'authors': ['Xinyue Chen', 'Che Wang', 'Zijian Zhou', 'Keith Ross'], 'published': '2021-01-15 06:25:58+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2101.05982v2'}
42,467
2403.07955v2
2,403.07955
2
Towards Faithful Explanations: Boosting Rationalization with Shortcuts Discovery
The remarkable success in neural networks provokes the selective rationalization. It explains the prediction results by identifying a small subset of the inputs sufficient to support them. Since existing methods still suffer from adopting the shortcuts in data to compose rationales and limited large-scale annotated rationales by human, in this paper, we propose a Shortcuts-fused Selective Rationalization (SSR) method, which boosts the rationalization by discovering and exploiting potential shortcuts. Specifically, SSR first designs a shortcuts discovery approach to detect several potential shortcuts. Then, by introducing the identified shortcuts, we propose two strategies to mitigate the problem of utilizing shortcuts to compose rationales. Finally, we develop two data augmentations methods to close the gap in the number of annotated rationales. Extensive experimental results on real-world datasets clearly validate the effectiveness of our proposed method.
2024-03-12
{'id': '2403.07955v2', 'title': 'Towards Faithful Explanations: Boosting Rationalization with Shortcuts Discovery', 'abstract': 'The remarkable success in neural networks provokes the selective\nrationalization. It explains the prediction results by identifying a small\nsubset of the inputs sufficient to support them. Since existing methods still\nsuffer from adopting the shortcuts in data to compose rationales and limited\nlarge-scale annotated rationales by human, in this paper, we propose a\nShortcuts-fused Selective Rationalization (SSR) method, which boosts the\nrationalization by discovering and exploiting potential shortcuts.\nSpecifically, SSR first designs a shortcuts discovery approach to detect\nseveral potential shortcuts. Then, by introducing the identified shortcuts, we\npropose two strategies to mitigate the problem of utilizing shortcuts to\ncompose rationales. Finally, we develop two data augmentations methods to close\nthe gap in the number of annotated rationales. Extensive experimental results\non real-world datasets clearly validate the effectiveness of our proposed\nmethod.', 'authors': ['Linan Yue', 'Qi Liu', 'Yichao Du', 'Li Wang', 'Weibo Gao', 'Yanqing An'], 'published': '2024-03-12 07:24:17+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2403.07955v2'}
42,470
2309.03241v2
2,309.03241
2
GPT Can Solve Mathematical Problems Without a Calculator
Previous studies have typically assumed that large language models are unable to accurately perform arithmetic operations, particularly multiplication of >8 digits, and operations involving decimals and fractions, without the use of calculator tools. This paper aims to challenge this misconception. With sufficient training data, a 2 billion-parameter language model can accurately perform multi-digit arithmetic operations with almost 100% accuracy without data leakage, significantly surpassing GPT-4 (whose multi-digit multiplication accuracy is only 4.3%). We also demonstrate that our MathGLM, fine-tuned from GLM-10B on a dataset with additional multi-step arithmetic operations and math problems described in text, achieves similar performance to GPT-4 on a 5,000-samples Chinese math problem test set. Our code and data are public at https://github.com/THUDM/MathGLM.
2023-09-06
{'id': '2309.03241v2', 'title': 'GPT Can Solve Mathematical Problems Without a Calculator', 'abstract': 'Previous studies have typically assumed that large language models are unable\nto accurately perform arithmetic operations, particularly multiplication of >8\ndigits, and operations involving decimals and fractions, without the use of\ncalculator tools. This paper aims to challenge this misconception. With\nsufficient training data, a 2 billion-parameter language model can accurately\nperform multi-digit arithmetic operations with almost 100% accuracy without\ndata leakage, significantly surpassing GPT-4 (whose multi-digit multiplication\naccuracy is only 4.3%). We also demonstrate that our MathGLM, fine-tuned from\nGLM-10B on a dataset with additional multi-step arithmetic operations and math\nproblems described in text, achieves similar performance to GPT-4 on a\n5,000-samples Chinese math problem test set. Our code and data are public at\nhttps://github.com/THUDM/MathGLM.', 'authors': ['Zhen Yang', 'Ming Ding', 'Qingsong Lv', 'Zhihuan Jiang', 'Zehai He', 'Yuyi Guo', 'Jinfeng Bai', 'Jie Tang'], 'published': '2023-09-06 06:18:16+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CL'], 'url': 'http://arxiv.org/abs/2309.03241v2'}
42,471
2402.14393v1
2,402.14393
1
Graph Parsing Networks
Graph pooling compresses graph information into a compact representation. State-of-the-art graph pooling methods follow a hierarchical approach, which reduces the graph size step-by-step. These methods must balance memory efficiency with preserving node information, depending on whether they use node dropping or node clustering. Additionally, fixed pooling ratios or numbers of pooling layers are predefined for all graphs, which prevents personalized pooling structures from being captured for each individual graph. In this work, inspired by bottom-up grammar induction, we propose an efficient graph parsing algorithm to infer the pooling structure, which then drives graph pooling. The resulting Graph Parsing Network (GPN) adaptively learns personalized pooling structure for each individual graph. GPN benefits from the discrete assignments generated by the graph parsing algorithm, allowing good memory efficiency while preserving node information intact. Experimental results on standard benchmarks demonstrate that GPN outperforms state-of-the-art graph pooling methods in graph classification tasks while being able to achieve competitive performance in node classification tasks. We also conduct a graph reconstruction task to show GPN's ability to preserve node information and measure both memory and time efficiency through relevant tests.
2024-02-22
{'id': '2402.14393v1', 'title': 'Graph Parsing Networks', 'abstract': "Graph pooling compresses graph information into a compact representation.\nState-of-the-art graph pooling methods follow a hierarchical approach, which\nreduces the graph size step-by-step. These methods must balance memory\nefficiency with preserving node information, depending on whether they use node\ndropping or node clustering. Additionally, fixed pooling ratios or numbers of\npooling layers are predefined for all graphs, which prevents personalized\npooling structures from being captured for each individual graph. In this work,\ninspired by bottom-up grammar induction, we propose an efficient graph parsing\nalgorithm to infer the pooling structure, which then drives graph pooling. The\nresulting Graph Parsing Network (GPN) adaptively learns personalized pooling\nstructure for each individual graph. GPN benefits from the discrete assignments\ngenerated by the graph parsing algorithm, allowing good memory efficiency while\npreserving node information intact. Experimental results on standard benchmarks\ndemonstrate that GPN outperforms state-of-the-art graph pooling methods in\ngraph classification tasks while being able to achieve competitive performance\nin node classification tasks. We also conduct a graph reconstruction task to\nshow GPN's ability to preserve node information and measure both memory and\ntime efficiency through relevant tests.", 'authors': ['Yunchong Song', 'Siyuan Huang', 'Xinbing Wang', 'Chenghu Zhou', 'Zhouhan Lin'], 'published': '2024-02-22 09:08:36+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2402.14393v1'}
42,481
2412.00020v1
2,412.0002
1
Partitioning Message Passing for Graph Fraud Detection
Label imbalance and homophily-heterophily mixture are the fundamental problems encountered when applying Graph Neural Networks (GNNs) to Graph Fraud Detection (GFD) tasks. Existing GNN-based GFD models are designed to augment graph structure to accommodate the inductive bias of GNNs towards homophily, by excluding heterophilic neighbors during message passing. In our work, we argue that the key to applying GNNs for GFD is not to exclude but to distinguish neighbors with different labels. Grounded in this perspective, we introduce Partitioning Message Passing (PMP), an intuitive yet effective message passing paradigm expressly crafted for GFD. Specifically, in the neighbor aggregation stage of PMP, neighbors with different classes are aggregated with distinct node-specific aggregation functions. By this means, the center node can adaptively adjust the information aggregated from its heterophilic and homophilic neighbors, thus avoiding the model gradient being dominated by benign nodes which occupy the majority of the population. We theoretically establish a connection between the spatial formulation of PMP and spectral analysis to characterize that PMP operates an adaptive node-specific spectral graph filter, which demonstrates the capability of PMP to handle heterophily-homophily mixed graphs. Extensive experimental results show that PMP can significantly boost the performance on GFD tasks.
2024-11-16
{'id': '2412.00020v1', 'title': 'Partitioning Message Passing for Graph Fraud Detection', 'abstract': 'Label imbalance and homophily-heterophily mixture are the fundamental\nproblems encountered when applying Graph Neural Networks (GNNs) to Graph Fraud\nDetection (GFD) tasks. Existing GNN-based GFD models are designed to augment\ngraph structure to accommodate the inductive bias of GNNs towards homophily, by\nexcluding heterophilic neighbors during message passing. In our work, we argue\nthat the key to applying GNNs for GFD is not to exclude but to {\\em\ndistinguish} neighbors with different labels. Grounded in this perspective, we\nintroduce Partitioning Message Passing (PMP), an intuitive yet effective\nmessage passing paradigm expressly crafted for GFD. Specifically, in the\nneighbor aggregation stage of PMP, neighbors with different classes are\naggregated with distinct node-specific aggregation functions. By this means,\nthe center node can adaptively adjust the information aggregated from its\nheterophilic and homophilic neighbors, thus avoiding the model gradient being\ndominated by benign nodes which occupy the majority of the population. We\ntheoretically establish a connection between the spatial formulation of PMP and\nspectral analysis to characterize that PMP operates an adaptive node-specific\nspectral graph filter, which demonstrates the capability of PMP to handle\nheterophily-homophily mixed graphs. Extensive experimental results show that\nPMP can significantly boost the performance on GFD tasks.', 'authors': ['Wei Zhuo', 'Zemin Liu', 'Bryan Hooi', 'Bingsheng He', 'Guang Tan', 'Rizal Fathony', 'Jia Chen'], 'published': '2024-11-16 11:30:53+00:00', 'categories': ['cs.AI', 'cs.LG', 'cs.SI'], 'url': 'http://arxiv.org/abs/2412.00020v1'}
42,485
2305.19871v2
2,305.19871
2
There is more to graphs than meets the eye: Learning universal features with self-supervision
We study the problem of learning features through self-supervision that are generalisable to multiple graphs. State-of-the-art graph self-supervision restricts training to only one graph, resulting in graph-specific models that are incompatible with different but related graphs. We hypothesize that training with more than one graph that belong to the same family can improve the quality of the learnt representations. However, learning universal features from disparate node/edge features in different graphs is non-trivial. To address this challenge, we first homogenise the disparate features with graph-specific encoders that transform the features into a common space. A universal representation learning module then learns generalisable features on this common space. We show that compared to traditional self-supervision with one graph, our approach results in (1) better performance on downstream node classification, (2) learning features that can be re-used for unseen graphs of the same family, (3) more efficient training and (4) compact yet generalisable models. We also show ability of the proposed framework to deliver these benefits for relatively larger graphs. In this paper, we present a principled way to design foundation graph models that learn from more than one graph in an end-to-end manner, while bridging the gap between self-supervised and supervised performance.
2023-05-31
{'id': '2305.19871v2', 'title': 'There is more to graphs than meets the eye: Learning universal features with self-supervision', 'abstract': 'We study the problem of learning features through self-supervision that are\ngeneralisable to multiple graphs. State-of-the-art graph self-supervision\nrestricts training to only one graph, resulting in graph-specific models that\nare incompatible with different but related graphs. We hypothesize that\ntraining with more than one graph that belong to the same family can improve\nthe quality of the learnt representations. However, learning universal features\nfrom disparate node/edge features in different graphs is non-trivial. To\naddress this challenge, we first homogenise the disparate features with\ngraph-specific encoders that transform the features into a common space. A\nuniversal representation learning module then learns generalisable features on\nthis common space. We show that compared to traditional self-supervision with\none graph, our approach results in (1) better performance on downstream node\nclassification, (2) learning features that can be re-used for unseen graphs of\nthe same family, (3) more efficient training and (4) compact yet generalisable\nmodels. We also show ability of the proposed framework to deliver these\nbenefits for relatively larger graphs. In this paper, we present a principled\nway to design foundation graph models that learn from more than one graph in an\nend-to-end manner, while bridging the gap between self-supervised and\nsupervised performance.', 'authors': ['Laya Das', 'Sai Munikoti', 'Nrushad Joshi', 'Mahantesh Halappanavar'], 'published': '2023-05-31 14:08:48+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2305.19871v2'}
42,495
2310.01452v2
2,310.01452
2
Fooling the Textual Fooler via Randomizing Latent Representations
Despite outstanding performance in a variety of NLP tasks, recent studies have revealed that NLP models are vulnerable to adversarial attacks that slightly perturb the input to cause the models to misbehave. Among these attacks, adversarial word-level perturbations are well-studied and effective attack strategies. Since these attacks work in black-box settings, they do not require access to the model architecture or model parameters and thus can be detrimental to existing NLP applications. To perform an attack, the adversary queries the victim model many times to determine the most important words in an input text and to replace these words with their corresponding synonyms. In this work, we propose a lightweight and attack-agnostic defense whose main goal is to perplex the process of generating an adversarial example in these query-based black-box attacks; that is to fool the textual fooler. This defense, named AdvFooler, works by randomizing the latent representation of the input at inference time. Different from existing defenses, AdvFooler does not necessitate additional computational overhead during training nor relies on assumptions about the potential adversarial perturbation set while having a negligible impact on the model's accuracy. Our theoretical and empirical analyses highlight the significance of robustness resulting from confusing the adversary via randomizing the latent space, as well as the impact of randomization on clean accuracy. Finally, we empirically demonstrate near state-of-the-art robustness of AdvFooler against representative adversarial word-level attacks on two benchmark datasets.
2023-10-02
{'id': '2310.01452v2', 'title': 'Fooling the Textual Fooler via Randomizing Latent Representations', 'abstract': "Despite outstanding performance in a variety of NLP tasks, recent studies\nhave revealed that NLP models are vulnerable to adversarial attacks that\nslightly perturb the input to cause the models to misbehave. Among these\nattacks, adversarial word-level perturbations are well-studied and effective\nattack strategies. Since these attacks work in black-box settings, they do not\nrequire access to the model architecture or model parameters and thus can be\ndetrimental to existing NLP applications. To perform an attack, the adversary\nqueries the victim model many times to determine the most important words in an\ninput text and to replace these words with their corresponding synonyms. In\nthis work, we propose a lightweight and attack-agnostic defense whose main goal\nis to perplex the process of generating an adversarial example in these\nquery-based black-box attacks; that is to fool the textual fooler. This\ndefense, named AdvFooler, works by randomizing the latent representation of the\ninput at inference time. Different from existing defenses, AdvFooler does not\nnecessitate additional computational overhead during training nor relies on\nassumptions about the potential adversarial perturbation set while having a\nnegligible impact on the model's accuracy. Our theoretical and empirical\nanalyses highlight the significance of robustness resulting from confusing the\nadversary via randomizing the latent space, as well as the impact of\nrandomization on clean accuracy. Finally, we empirically demonstrate near\nstate-of-the-art robustness of AdvFooler against representative adversarial\nword-level attacks on two benchmark datasets.", 'authors': ['Duy C. Hoang', 'Quang H. Nguyen', 'Saurav Manchanda', 'MinLong Peng', 'Kok-Seng Wong', 'Khoa D. Doan'], 'published': '2023-10-02 06:57:25+00:00', 'categories': ['cs.CL', 'cs.AI'], 'url': 'http://arxiv.org/abs/2310.01452v2'}
42,475
2207.10792v2
2,207.10792
2
Test-Time Adaptation via Self-Training with Nearest Neighbor Information
Test-time adaptation (TTA) aims to adapt a trained classifier using online unlabeled test data only, without any information related to the training procedure. Most existing TTA methods adapt the trained classifier using the classifier's prediction on the test data as pseudo-label. However, under test-time domain shift, accuracy of the pseudo labels cannot be guaranteed, and thus the TTA methods often encounter performance degradation at the adapted classifier. To overcome this limitation, we propose a novel test-time adaptation method, called Test-time Adaptation via Self-Training with nearest neighbor information (TAST), which is composed of the following procedures: (1) adds trainable adaptation modules on top of the trained feature extractor; (2) newly defines a pseudo-label distribution for the test data by using the nearest neighbor information; (3) trains these modules only a few times during test time to match the nearest neighbor-based pseudo label distribution and a prototype-based class distribution for the test data; and (4) predicts the label of test data using the average predicted class distribution from these modules. The pseudo-label generation is based on the basic intuition that a test data and its nearest neighbor in the embedding space are likely to share the same label under the domain shift. By utilizing multiple randomly initialized adaptation modules, TAST extracts useful information for the classification of the test data under the domain shift, using the nearest neighbor information. TAST showed better performance than the state-of-the-art TTA methods on two standard benchmark tasks, domain generalization, namely VLCS, PACS, OfficeHome, and TerraIncognita, and image corruption, particularly CIFAR-10/100C.
2022-07-08
{'id': '2207.10792v2', 'title': 'Test-Time Adaptation via Self-Training with Nearest Neighbor Information', 'abstract': "Test-time adaptation (TTA) aims to adapt a trained classifier using online\nunlabeled test data only, without any information related to the training\nprocedure. Most existing TTA methods adapt the trained classifier using the\nclassifier's prediction on the test data as pseudo-label. However, under\ntest-time domain shift, accuracy of the pseudo labels cannot be guaranteed, and\nthus the TTA methods often encounter performance degradation at the adapted\nclassifier. To overcome this limitation, we propose a novel test-time\nadaptation method, called Test-time Adaptation via Self-Training with nearest\nneighbor information (TAST), which is composed of the following procedures: (1)\nadds trainable adaptation modules on top of the trained feature extractor; (2)\nnewly defines a pseudo-label distribution for the test data by using the\nnearest neighbor information; (3) trains these modules only a few times during\ntest time to match the nearest neighbor-based pseudo label distribution and a\nprototype-based class distribution for the test data; and (4) predicts the\nlabel of test data using the average predicted class distribution from these\nmodules. The pseudo-label generation is based on the basic intuition that a\ntest data and its nearest neighbor in the embedding space are likely to share\nthe same label under the domain shift. By utilizing multiple randomly\ninitialized adaptation modules, TAST extracts useful information for the\nclassification of the test data under the domain shift, using the nearest\nneighbor information. TAST showed better performance than the state-of-the-art\nTTA methods on two standard benchmark tasks, domain generalization, namely\nVLCS, PACS, OfficeHome, and TerraIncognita, and image corruption, particularly\nCIFAR-10/100C.", 'authors': ['Minguk Jang', 'Sae-Young Chung', 'Hye Won Chung'], 'published': '2022-07-08 05:02:15+00:00', 'categories': ['cs.CV', 'cs.AI'], 'url': 'http://arxiv.org/abs/2207.10792v2'}
42,506
2206.13602v2
2,206.13602
2
Molecular Geometry Pretraining with SE(3)-Invariant Denoising Distance Matching
Molecular representation pretraining is critical in various applications for drug and material discovery due to the limited number of labeled molecules, and most existing work focuses on pretraining on 2D molecular graphs. However, the power of pretraining on 3D geometric structures has been less explored. This is owing to the difficulty of finding a sufficient proxy task that can empower the pretraining to effectively extract essential features from the geometric structures. Motivated by the dynamic nature of 3D molecules, where the continuous motion of a molecule in the 3D Euclidean space forms a smooth potential energy surface, we propose GeoSSL, a 3D coordinate denoising pretraining framework to model such an energy landscape. Further by leveraging an SE(3)-invariant score matching method, we propose GeoSSL-DDM in which the coordinate denoising proxy task is effectively boiled down to denoising the pairwise atomic distances in a molecule. Our comprehensive experiments confirm the effectiveness and robustness of our proposed method.
2022-06-27
{'id': '2206.13602v2', 'title': 'Molecular Geometry Pretraining with SE(3)-Invariant Denoising Distance Matching', 'abstract': 'Molecular representation pretraining is critical in various applications for\ndrug and material discovery due to the limited number of labeled molecules, and\nmost existing work focuses on pretraining on 2D molecular graphs. However, the\npower of pretraining on 3D geometric structures has been less explored. This is\nowing to the difficulty of finding a sufficient proxy task that can empower the\npretraining to effectively extract essential features from the geometric\nstructures. Motivated by the dynamic nature of 3D molecules, where the\ncontinuous motion of a molecule in the 3D Euclidean space forms a smooth\npotential energy surface, we propose GeoSSL, a 3D coordinate denoising\npretraining framework to model such an energy landscape. Further by leveraging\nan SE(3)-invariant score matching method, we propose GeoSSL-DDM in which the\ncoordinate denoising proxy task is effectively boiled down to denoising the\npairwise atomic distances in a molecule. Our comprehensive experiments confirm\nthe effectiveness and robustness of our proposed method.', 'authors': ['Shengchao Liu', 'Hongyu Guo', 'Jian Tang'], 'published': '2022-06-27 19:30:53+00:00', 'categories': ['cs.LG', 'q-bio.QM'], 'url': 'http://arxiv.org/abs/2206.13602v2'}
42,512
2212.04084v1
2,212.04084
1
Federated Learning for Inference at Anytime and Anywhere
Federated learning has been predominantly concerned with collaborative training of deep networks from scratch, and especially the many challenges that arise, such as communication cost, robustness to heterogeneous data, and support for diverse device capabilities. However, there is no unified framework that addresses all these problems together. This paper studies the challenges and opportunities of exploiting pre-trained Transformer models in FL. In particular, we propose to efficiently adapt such pre-trained models by injecting a novel attention-based adapter module at each transformer block that both modulates the forward pass and makes an early prediction. Training only the lightweight adapter by FL leads to fast and communication-efficient learning even in the presence of heterogeneous data and devices. Extensive experiments on standard FL benchmarks, including CIFAR-100, FEMNIST and SpeechCommandsv2 demonstrate that this simple framework provides fast and accurate FL while supporting heterogenous device capabilities, efficient personalization, and scalable-cost anytime inference.
2022-12-08
{'id': '2212.04084v1', 'title': 'Federated Learning for Inference at Anytime and Anywhere', 'abstract': 'Federated learning has been predominantly concerned with collaborative\ntraining of deep networks from scratch, and especially the many challenges that\narise, such as communication cost, robustness to heterogeneous data, and\nsupport for diverse device capabilities. However, there is no unified framework\nthat addresses all these problems together. This paper studies the challenges\nand opportunities of exploiting pre-trained Transformer models in FL. In\nparticular, we propose to efficiently adapt such pre-trained models by\ninjecting a novel attention-based adapter module at each transformer block that\nboth modulates the forward pass and makes an early prediction. Training only\nthe lightweight adapter by FL leads to fast and communication-efficient\nlearning even in the presence of heterogeneous data and devices. Extensive\nexperiments on standard FL benchmarks, including CIFAR-100, FEMNIST and\nSpeechCommandsv2 demonstrate that this simple framework provides fast and\naccurate FL while supporting heterogenous device capabilities, efficient\npersonalization, and scalable-cost anytime inference.', 'authors': ['Zicheng Liu', 'Da Li', 'Javier Fernandez-Marques', 'Stefanos Laskaridis', 'Yan Gao', 'Łukasz Dudziak', 'Stan Z. Li', 'Shell Xu Hu', 'Timothy Hospedales'], 'published': '2022-12-08 05:32:33+00:00', 'categories': ['cs.LG', 'cs.DC'], 'url': 'http://arxiv.org/abs/2212.04084v1'}