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
42,297
2303.01743v1
2,303.01743
1
A Laplace-inspired Distribution on SO(3) for Probabilistic Rotation Estimation
Estimating the 3DoF rotation from a single RGB image is an important yet challenging problem. Probabilistic rotation regression has raised more and more attention with the benefit of expressing uncertainty information along with the prediction. Though modeling noise using Gaussian-resembling Bingham distribution and matrix Fisher distribution is natural, they are shown to be sensitive to outliers for the nature of quadratic punishment to deviations. In this paper, we draw inspiration from multivariate Laplace distribution and propose a novel Rotation Laplace distribution on SO(3). Rotation Laplace distribution is robust to the disturbance of outliers and enforces much gradient to the low-error region, resulting in a better convergence. Our extensive experiments show that our proposed distribution achieves state-of-the-art performance for rotation regression tasks over both probabilistic and non-probabilistic baselines. Our project page is at https://pku-epic.github.io/RotationLaplace.
2023-03-03
{'id': '2303.01743v1', 'title': 'A Laplace-inspired Distribution on SO(3) for Probabilistic Rotation Estimation', 'abstract': 'Estimating the 3DoF rotation from a single RGB image is an important yet\nchallenging problem. Probabilistic rotation regression has raised more and more\nattention with the benefit of expressing uncertainty information along with the\nprediction. Though modeling noise using Gaussian-resembling Bingham\ndistribution and matrix Fisher distribution is natural, they are shown to be\nsensitive to outliers for the nature of quadratic punishment to deviations. In\nthis paper, we draw inspiration from multivariate Laplace distribution and\npropose a novel Rotation Laplace distribution on SO(3). Rotation Laplace\ndistribution is robust to the disturbance of outliers and enforces much\ngradient to the low-error region, resulting in a better convergence. Our\nextensive experiments show that our proposed distribution achieves\nstate-of-the-art performance for rotation regression tasks over both\nprobabilistic and non-probabilistic baselines. Our project page is at\nhttps://pku-epic.github.io/RotationLaplace.', 'authors': ['Yingda Yin', 'Yang Wang', 'He Wang', 'Baoquan Chen'], 'published': '2023-03-03 07:10:02+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2303.01743v1'}
42,298
2302.05807v1
2,302.05807
1
Pushing the Accuracy-Group Robustness Frontier with Introspective Self-play
Standard empirical risk minimization (ERM) training can produce deep neural network (DNN) models that are accurate on average but under-perform in under-represented population subgroups, especially when there are imbalanced group distributions in the long-tailed training data. Therefore, approaches that improve the accuracy-group robustness trade-off frontier of a DNN model (i.e. improving worst-group accuracy without sacrificing average accuracy, or vice versa) is of crucial importance. Uncertainty-based active learning (AL) can potentially improve the frontier by preferentially sampling underrepresented subgroups to create a more balanced training dataset. However, the quality of uncertainty estimates from modern DNNs tend to degrade in the presence of spurious correlations and dataset bias, compromising the effectiveness of AL for sampling tail groups. In this work, we propose Introspective Self-play (ISP), a simple approach to improve the uncertainty estimation of a deep neural network under dataset bias, by adding an auxiliary introspection task requiring a model to predict the bias for each data point in addition to the label. We show that ISP provably improves the bias-awareness of the model representation and the resulting uncertainty estimates. On two real-world tabular and language tasks, ISP serves as a simple "plug-in" for AL model training, consistently improving both the tail-group sampling rate and the final accuracy-fairness trade-off frontier of popular AL methods.
2023-02-11
{'id': '2302.05807v1', 'title': 'Pushing the Accuracy-Group Robustness Frontier with Introspective Self-play', 'abstract': 'Standard empirical risk minimization (ERM) training can produce deep neural\nnetwork (DNN) models that are accurate on average but under-perform in\nunder-represented population subgroups, especially when there are imbalanced\ngroup distributions in the long-tailed training data. Therefore, approaches\nthat improve the accuracy-group robustness trade-off frontier of a DNN model\n(i.e. improving worst-group accuracy without sacrificing average accuracy, or\nvice versa) is of crucial importance. Uncertainty-based active learning (AL)\ncan potentially improve the frontier by preferentially sampling\nunderrepresented subgroups to create a more balanced training dataset. However,\nthe quality of uncertainty estimates from modern DNNs tend to degrade in the\npresence of spurious correlations and dataset bias, compromising the\neffectiveness of AL for sampling tail groups. In this work, we propose\nIntrospective Self-play (ISP), a simple approach to improve the uncertainty\nestimation of a deep neural network under dataset bias, by adding an auxiliary\nintrospection task requiring a model to predict the bias for each data point in\naddition to the label. We show that ISP provably improves the bias-awareness of\nthe model representation and the resulting uncertainty estimates. On two\nreal-world tabular and language tasks, ISP serves as a simple "plug-in" for AL\nmodel training, consistently improving both the tail-group sampling rate and\nthe final accuracy-fairness trade-off frontier of popular AL methods.', 'authors': ['Jeremiah Zhe Liu', 'Krishnamurthy Dj Dvijotham', 'Jihyeon Lee', 'Quan Yuan', 'Martin Strobel', 'Balaji Lakshminarayanan', 'Deepak Ramachandran'], 'published': '2023-02-11 22:59:08+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2302.05807v1'}
42,299
2303.00613v1
2,303.00613
1
Diffusing Graph Attention
The dominant paradigm for machine learning on graphs uses Message Passing Graph Neural Networks (MP-GNNs), in which node representations are updated by aggregating information in their local neighborhood. Recently, there have been increasingly more attempts to adapt the Transformer architecture to graphs in an effort to solve some known limitations of MP-GNN. A challenging aspect of designing Graph Transformers is integrating the arbitrary graph structure into the architecture. We propose Graph Diffuser (GD) to address this challenge. GD learns to extract structural and positional relationships between distant nodes in the graph, which it then uses to direct the Transformer's attention and node representation. We demonstrate that existing GNNs and Graph Transformers struggle to capture long-range interactions and how Graph Diffuser does so while admitting intuitive visualizations. Experiments on eight benchmarks show Graph Diffuser to be a highly competitive model, outperforming the state-of-the-art in a diverse set of domains.
2023-03-01
{'id': '2303.00613v1', 'title': 'Diffusing Graph Attention', 'abstract': "The dominant paradigm for machine learning on graphs uses Message Passing\nGraph Neural Networks (MP-GNNs), in which node representations are updated by\naggregating information in their local neighborhood. Recently, there have been\nincreasingly more attempts to adapt the Transformer architecture to graphs in\nan effort to solve some known limitations of MP-GNN. A challenging aspect of\ndesigning Graph Transformers is integrating the arbitrary graph structure into\nthe architecture. We propose Graph Diffuser (GD) to address this challenge. GD\nlearns to extract structural and positional relationships between distant nodes\nin the graph, which it then uses to direct the Transformer's attention and node\nrepresentation. We demonstrate that existing GNNs and Graph Transformers\nstruggle to capture long-range interactions and how Graph Diffuser does so\nwhile admitting intuitive visualizations. Experiments on eight benchmarks show\nGraph Diffuser to be a highly competitive model, outperforming the\nstate-of-the-art in a diverse set of domains.", 'authors': ['Daniel Glickman', 'Eran Yahav'], 'published': '2023-03-01 16:11:05+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2303.00613v1'}
42,300
2303.02343v1
2,303.02343
1
What Is Missing in IRM Training and Evaluation? Challenges and Solutions
Invariant risk minimization (IRM) has received increasing attention as a way to acquire environment-agnostic data representations and predictions, and as a principled solution for preventing spurious correlations from being learned and for improving models' out-of-distribution generalization. Yet, recent works have found that the optimality of the originally-proposed IRM optimization (IRM) may be compromised in practice or could be impossible to achieve in some scenarios. Therefore, a series of advanced IRM algorithms have been developed that show practical improvement over IRM. In this work, we revisit these recent IRM advancements, and identify and resolve three practical limitations in IRM training and evaluation. First, we find that the effect of batch size during training has been chronically overlooked in previous studies, leaving room for further improvement. We propose small-batch training and highlight the improvements over a set of large-batch optimization techniques. Second, we find that improper selection of evaluation environments could give a false sense of invariance for IRM. To alleviate this effect, we leverage diversified test-time environments to precisely characterize the invariance of IRM when applied in practice. Third, we revisit (Ahuja et al. (2020))'s proposal to convert IRM into an ensemble game and identify a limitation when a single invariant predictor is desired instead of an ensemble of individual predictors. We propose a new IRM variant to address this limitation based on a novel viewpoint of ensemble IRM games as consensus-constrained bi-level optimization. Lastly, we conduct extensive experiments (covering 7 existing IRM variants and 7 datasets) to justify the practical significance of revisiting IRM training and evaluation in a principled manner.
2023-03-04
{'id': '2303.02343v1', 'title': 'What Is Missing in IRM Training and Evaluation? Challenges and Solutions', 'abstract': "Invariant risk minimization (IRM) has received increasing attention as a way\nto acquire environment-agnostic data representations and predictions, and as a\nprincipled solution for preventing spurious correlations from being learned and\nfor improving models' out-of-distribution generalization. Yet, recent works\nhave found that the optimality of the originally-proposed IRM optimization\n(IRM) may be compromised in practice or could be impossible to achieve in some\nscenarios. Therefore, a series of advanced IRM algorithms have been developed\nthat show practical improvement over IRM. In this work, we revisit these recent\nIRM advancements, and identify and resolve three practical limitations in IRM\ntraining and evaluation. First, we find that the effect of batch size during\ntraining has been chronically overlooked in previous studies, leaving room for\nfurther improvement. We propose small-batch training and highlight the\nimprovements over a set of large-batch optimization techniques. Second, we find\nthat improper selection of evaluation environments could give a false sense of\ninvariance for IRM. To alleviate this effect, we leverage diversified test-time\nenvironments to precisely characterize the invariance of IRM when applied in\npractice. Third, we revisit (Ahuja et al. (2020))'s proposal to convert IRM\ninto an ensemble game and identify a limitation when a single invariant\npredictor is desired instead of an ensemble of individual predictors. We\npropose a new IRM variant to address this limitation based on a novel viewpoint\nof ensemble IRM games as consensus-constrained bi-level optimization. Lastly,\nwe conduct extensive experiments (covering 7 existing IRM variants and 7\ndatasets) to justify the practical significance of revisiting IRM training and\nevaluation in a principled manner.", 'authors': ['Yihua Zhang', 'Pranay Sharma', 'Parikshit Ram', 'Mingyi Hong', 'Kush Varshney', 'Sijia Liu'], 'published': '2023-03-04 07:06:24+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2303.02343v1'}
42,302
2111.13073v1
2,111.13073
1
Neuronal Learning Analysis using Cycle-Consistent Adversarial Networks
Understanding how activity in neural circuits reshapes following task learning could reveal fundamental mechanisms of learning. Thanks to the recent advances in neural imaging technologies, high-quality recordings can be obtained from hundreds of neurons over multiple days or even weeks. However, the complexity and dimensionality of population responses pose significant challenges for analysis. Existing methods of studying neuronal adaptation and learning often impose strong assumptions on the data or model, resulting in biased descriptions that do not generalize. In this work, we use a variant of deep generative models called - CycleGAN, to learn the unknown mapping between pre- and post-learning neural activities recorded $in vivo$. We develop an end-to-end pipeline to preprocess, train and evaluate calcium fluorescence signals, and a procedure to interpret the resulting deep learning models. To assess the validity of our method, we first test our framework on a synthetic dataset with known ground-truth transformation. Subsequently, we applied our method to neural activities recorded from the primary visual cortex of behaving mice, where the mice transition from novice to expert-level performance in a visual-based virtual reality experiment. We evaluate model performance on generated calcium signals and their inferred spike trains. To maximize performance, we derive a novel approach to pre-sort neurons such that convolutional-based networks can take advantage of the spatial information that exists in neural activities. In addition, we incorporate visual explanation methods to improve the interpretability of our work and gain insights into the learning process as manifested in the cellular activities. Together, our results demonstrate that analyzing neuronal learning processes with data-driven deep unsupervised methods holds the potential to unravel changes in an unbiased way.
2021-11-25
{'id': '2111.13073v1', 'title': 'Neuronal Learning Analysis using Cycle-Consistent Adversarial Networks', 'abstract': 'Understanding how activity in neural circuits reshapes following task\nlearning could reveal fundamental mechanisms of learning. Thanks to the recent\nadvances in neural imaging technologies, high-quality recordings can be\nobtained from hundreds of neurons over multiple days or even weeks. However,\nthe complexity and dimensionality of population responses pose significant\nchallenges for analysis. Existing methods of studying neuronal adaptation and\nlearning often impose strong assumptions on the data or model, resulting in\nbiased descriptions that do not generalize. In this work, we use a variant of\ndeep generative models called - CycleGAN, to learn the unknown mapping between\npre- and post-learning neural activities recorded $\\textit{in vivo}$. We\ndevelop an end-to-end pipeline to preprocess, train and evaluate calcium\nfluorescence signals, and a procedure to interpret the resulting deep learning\nmodels. To assess the validity of our method, we first test our framework on a\nsynthetic dataset with known ground-truth transformation. Subsequently, we\napplied our method to neural activities recorded from the primary visual cortex\nof behaving mice, where the mice transition from novice to expert-level\nperformance in a visual-based virtual reality experiment. We evaluate model\nperformance on generated calcium signals and their inferred spike trains. To\nmaximize performance, we derive a novel approach to pre-sort neurons such that\nconvolutional-based networks can take advantage of the spatial information that\nexists in neural activities. In addition, we incorporate visual explanation\nmethods to improve the interpretability of our work and gain insights into the\nlearning process as manifested in the cellular activities. Together, our\nresults demonstrate that analyzing neuronal learning processes with data-driven\ndeep unsupervised methods holds the potential to unravel changes in an unbiased\nway.', 'authors': ['Bryan M. Li', 'Theoklitos Amvrosiadis', 'Nathalie Rochefort', 'Arno Onken'], 'published': '2021-11-25 13:24:19+00:00', 'categories': ['q-bio.NC', 'cs.LG'], 'url': 'http://arxiv.org/abs/2111.13073v1'}
42,303
2203.08065v2
2,203.08065
2
Surrogate Gap Minimization Improves Sharpness-Aware Training
The recently proposed Sharpness-Aware Minimization (SAM) improves generalization by minimizing a perturbed loss defined as the maximum loss within a neighborhood in the parameter space. However, we show that both sharp and flat minima can have a low perturbed loss, implying that SAM does not always prefer flat minima. Instead, we define a surrogate gap, a measure equivalent to the dominant eigenvalue of Hessian at a local minimum when the radius of the neighborhood (to derive the perturbed loss) is small. The surrogate gap is easy to compute and feasible for direct minimization during training. Based on the above observations, we propose Surrogate Gap Guided Sharpness-Aware Minimization (GSAM), a novel improvement over SAM with negligible computation overhead. Conceptually, GSAM consists of two steps: 1) a gradient descent like SAM to minimize the perturbed loss, and 2) an ascent step in the orthogonal direction (after gradient decomposition) to minimize the surrogate gap and yet not affect the perturbed loss. GSAM seeks a region with both small loss (by step 1) and low sharpness (by step 2), giving rise to a model with high generalization capabilities. Theoretically, we show the convergence of GSAM and provably better generalization than SAM. Empirically, GSAM consistently improves generalization (e.g., +3.2\% over SAM and +5.4\% over AdamW on ImageNet top-1 accuracy for ViT-B/32). Code is released at https://sites.google.com/view/gsam-iclr22/home.
2022-03-15
{'id': '2203.08065v2', 'title': 'Surrogate Gap Minimization Improves Sharpness-Aware Training', 'abstract': 'The recently proposed Sharpness-Aware Minimization (SAM) improves\ngeneralization by minimizing a \\textit{perturbed loss} defined as the maximum\nloss within a neighborhood in the parameter space. However, we show that both\nsharp and flat minima can have a low perturbed loss, implying that SAM does not\nalways prefer flat minima. Instead, we define a \\textit{surrogate gap}, a\nmeasure equivalent to the dominant eigenvalue of Hessian at a local minimum\nwhen the radius of the neighborhood (to derive the perturbed loss) is small.\nThe surrogate gap is easy to compute and feasible for direct minimization\nduring training. Based on the above observations, we propose Surrogate\n\\textbf{G}ap Guided \\textbf{S}harpness-\\textbf{A}ware \\textbf{M}inimization\n(GSAM), a novel improvement over SAM with negligible computation overhead.\nConceptually, GSAM consists of two steps: 1) a gradient descent like SAM to\nminimize the perturbed loss, and 2) an \\textit{ascent} step in the\n\\textit{orthogonal} direction (after gradient decomposition) to minimize the\nsurrogate gap and yet not affect the perturbed loss. GSAM seeks a region with\nboth small loss (by step 1) and low sharpness (by step 2), giving rise to a\nmodel with high generalization capabilities. Theoretically, we show the\nconvergence of GSAM and provably better generalization than SAM. Empirically,\nGSAM consistently improves generalization (e.g., +3.2\\% over SAM and +5.4\\%\nover AdamW on ImageNet top-1 accuracy for ViT-B/32). Code is released at \\url{\nhttps://sites.google.com/view/gsam-iclr22/home}.', 'authors': ['Juntang Zhuang', 'Boqing Gong', 'Liangzhe Yuan', 'Yin Cui', 'Hartwig Adam', 'Nicha Dvornek', 'Sekhar Tatikonda', 'James Duncan', 'Ting Liu'], 'published': '2022-03-15 16:57:59+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2203.08065v2'}
42,311
2203.08656v1
2,203.08656
1
Learning Representation for Bayesian Optimization with Collision-free Regularization
Bayesian optimization has been challenged by datasets with large-scale, high-dimensional, and non-stationary characteristics, which are common in real-world scenarios. Recent works attempt to handle such input by applying neural networks ahead of the classical Gaussian process to learn a latent representation. We show that even with proper network design, such learned representation often leads to collision in the latent space: two points with significantly different observations collide in the learned latent space, leading to degraded optimization performance. To address this issue, we propose LOCo, an efficient deep Bayesian optimization framework which employs a novel regularizer to reduce the collision in the learned latent space and encourage the mapping from the latent space to the objective value to be Lipschitz continuous. LOCo takes in pairs of data points and penalizes those too close in the latent space compared to their target space distance. We provide a rigorous theoretical justification for LOCo by inspecting the regret of this dynamic-embedding-based Bayesian optimization algorithm, where the neural network is iteratively retrained with the regularizer. Our empirical results demonstrate the effectiveness of LOCo on several synthetic and real-world benchmark Bayesian optimization tasks.
2022-03-16
{'id': '2203.08656v1', 'title': 'Learning Representation for Bayesian Optimization with Collision-free Regularization', 'abstract': 'Bayesian optimization has been challenged by datasets with large-scale,\nhigh-dimensional, and non-stationary characteristics, which are common in\nreal-world scenarios. Recent works attempt to handle such input by applying\nneural networks ahead of the classical Gaussian process to learn a latent\nrepresentation. We show that even with proper network design, such learned\nrepresentation often leads to collision in the latent space: two points with\nsignificantly different observations collide in the learned latent space,\nleading to degraded optimization performance. To address this issue, we propose\nLOCo, an efficient deep Bayesian optimization framework which employs a novel\nregularizer to reduce the collision in the learned latent space and encourage\nthe mapping from the latent space to the objective value to be Lipschitz\ncontinuous. LOCo takes in pairs of data points and penalizes those too close in\nthe latent space compared to their target space distance. We provide a rigorous\ntheoretical justification for LOCo by inspecting the regret of this\ndynamic-embedding-based Bayesian optimization algorithm, where the neural\nnetwork is iteratively retrained with the regularizer. Our empirical results\ndemonstrate the effectiveness of LOCo on several synthetic and real-world\nbenchmark Bayesian optimization tasks.', 'authors': ['Fengxue Zhang', 'Brian Nord', 'Yuxin Chen'], 'published': '2022-03-16 14:44:16+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2203.08656v1'}
42,346
2106.09686v3
2,106.09686
3
How Low Can We Go: Trading Memory for Error in Low-Precision Training
Low-precision arithmetic trains deep learning models using less energy, less memory and less time. However, we pay a price for the savings: lower precision may yield larger round-off error and hence larger prediction error. As applications proliferate, users must choose which precision to use to train a new model, and chip manufacturers must decide which precisions to manufacture. We view these precision choices as a hyperparameter tuning problem, and borrow ideas from meta-learning to learn the tradeoff between memory and error. In this paper, we introduce Pareto Estimation to Pick the Perfect Precision (PEPPP). We use matrix factorization to find non-dominated configurations (the Pareto frontier) with a limited number of network evaluations. For any given memory budget, the precision that minimizes error is a point on this frontier. Practitioners can use the frontier to trade memory for error and choose the best precision for their goals.
2021-06-17
{'id': '2106.09686v3', 'title': 'How Low Can We Go: Trading Memory for Error in Low-Precision Training', 'abstract': 'Low-precision arithmetic trains deep learning models using less energy, less\nmemory and less time. However, we pay a price for the savings: lower precision\nmay yield larger round-off error and hence larger prediction error. As\napplications proliferate, users must choose which precision to use to train a\nnew model, and chip manufacturers must decide which precisions to manufacture.\nWe view these precision choices as a hyperparameter tuning problem, and borrow\nideas from meta-learning to learn the tradeoff between memory and error. In\nthis paper, we introduce Pareto Estimation to Pick the Perfect Precision\n(PEPPP). We use matrix factorization to find non-dominated configurations (the\nPareto frontier) with a limited number of network evaluations. For any given\nmemory budget, the precision that minimizes error is a point on this frontier.\nPractitioners can use the frontier to trade memory for error and choose the\nbest precision for their goals.', 'authors': ['Chengrun Yang', 'Ziyang Wu', 'Jerry Chee', 'Christopher De Sa', 'Madeleine Udell'], 'published': '2021-06-17 17:38:07+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2106.09686v3'}
42,348
2107.10060v5
2,107.1006
5
Conditional GANs with Auxiliary Discriminative Classifier
Conditional generative models aim to learn the underlying joint distribution of data and labels to achieve conditional data generation. Among them, the auxiliary classifier generative adversarial network (AC-GAN) has been widely used, but suffers from the problem of low intra-class diversity of the generated samples. The fundamental reason pointed out in this paper is that the classifier of AC-GAN is generator-agnostic, which therefore cannot provide informative guidance for the generator to approach the joint distribution, resulting in a minimization of the conditional entropy that decreases the intra-class diversity. Motivated by this understanding, we propose a novel conditional GAN with an auxiliary discriminative classifier (ADC-GAN) to resolve the above problem. Specifically, the proposed auxiliary discriminative classifier becomes generator-aware by recognizing the class-labels of the real data and the generated data discriminatively. Our theoretical analysis reveals that the generator can faithfully learn the joint distribution even without the original discriminator, making the proposed ADC-GAN robust to the value of the coefficient hyperparameter and the selection of the GAN loss, and stable during training. Extensive experimental results on synthetic and real-world datasets demonstrate the superiority of ADC-GAN in conditional generative modeling compared to state-of-the-art classifier-based and projection-based conditional GANs.
2021-07-21
{'id': '2107.10060v5', 'title': 'Conditional GANs with Auxiliary Discriminative Classifier', 'abstract': 'Conditional generative models aim to learn the underlying joint distribution\nof data and labels to achieve conditional data generation. Among them, the\nauxiliary classifier generative adversarial network (AC-GAN) has been widely\nused, but suffers from the problem of low intra-class diversity of the\ngenerated samples. The fundamental reason pointed out in this paper is that the\nclassifier of AC-GAN is generator-agnostic, which therefore cannot provide\ninformative guidance for the generator to approach the joint distribution,\nresulting in a minimization of the conditional entropy that decreases the\nintra-class diversity. Motivated by this understanding, we propose a novel\nconditional GAN with an auxiliary discriminative classifier (ADC-GAN) to\nresolve the above problem. Specifically, the proposed auxiliary discriminative\nclassifier becomes generator-aware by recognizing the class-labels of the real\ndata and the generated data discriminatively. Our theoretical analysis reveals\nthat the generator can faithfully learn the joint distribution even without the\noriginal discriminator, making the proposed ADC-GAN robust to the value of the\ncoefficient hyperparameter and the selection of the GAN loss, and stable during\ntraining. Extensive experimental results on synthetic and real-world datasets\ndemonstrate the superiority of ADC-GAN in conditional generative modeling\ncompared to state-of-the-art classifier-based and projection-based conditional\nGANs.', 'authors': ['Liang Hou', 'Qi Cao', 'Huawei Shen', 'Siyuan Pan', 'Xiaoshuang Li', 'Xueqi Cheng'], 'published': '2021-07-21 13:06:32+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2107.10060v5'}
42,352
2105.14080v1
2,105.1408
1
Gotta Go Fast When Generating Data with Score-Based Models
Score-based (denoising diffusion) generative models have recently gained a lot of success in generating realistic and diverse data. These approaches define a forward diffusion process for transforming data to noise and generate data by reversing it (thereby going from noise to data). Unfortunately, current score-based models generate data very slowly due to the sheer number of score network evaluations required by numerical SDE solvers. In this work, we aim to accelerate this process by devising a more efficient SDE solver. Existing approaches rely on the Euler-Maruyama (EM) solver, which uses a fixed step size. We found that naively replacing it with other SDE solvers fares poorly - they either result in low-quality samples or become slower than EM. To get around this issue, we carefully devise an SDE solver with adaptive step sizes tailored to score-based generative models piece by piece. Our solver requires only two score function evaluations, rarely rejects samples, and leads to high-quality samples. Our approach generates data 2 to 10 times faster than EM while achieving better or equal sample quality. For high-resolution images, our method leads to significantly higher quality samples than all other methods tested. Our SDE solver has the benefit of requiring no step size tuning.
2021-05-28
{'id': '2105.14080v1', 'title': 'Gotta Go Fast When Generating Data with Score-Based Models', 'abstract': 'Score-based (denoising diffusion) generative models have recently gained a\nlot of success in generating realistic and diverse data. These approaches\ndefine a forward diffusion process for transforming data to noise and generate\ndata by reversing it (thereby going from noise to data). Unfortunately, current\nscore-based models generate data very slowly due to the sheer number of score\nnetwork evaluations required by numerical SDE solvers.\n In this work, we aim to accelerate this process by devising a more efficient\nSDE solver. Existing approaches rely on the Euler-Maruyama (EM) solver, which\nuses a fixed step size. We found that naively replacing it with other SDE\nsolvers fares poorly - they either result in low-quality samples or become\nslower than EM. To get around this issue, we carefully devise an SDE solver\nwith adaptive step sizes tailored to score-based generative models piece by\npiece. Our solver requires only two score function evaluations, rarely rejects\nsamples, and leads to high-quality samples. Our approach generates data 2 to 10\ntimes faster than EM while achieving better or equal sample quality. For\nhigh-resolution images, our method leads to significantly higher quality\nsamples than all other methods tested. Our SDE solver has the benefit of\nrequiring no step size tuning.', 'authors': ['Alexia Jolicoeur-Martineau', 'Ke Li', 'Rémi Piché-Taillefer', 'Tal Kachman', 'Ioannis Mitliagkas'], 'published': '2021-05-28 19:48:51+00:00', 'categories': ['cs.LG', 'cs.CV', 'math.OC', 'stat.ML'], 'url': 'http://arxiv.org/abs/2105.14080v1'}
42,357
2204.07524v1
2,204.07524
1
Neural Structured Prediction for Inductive Node Classification
This paper studies node classification in the inductive setting, i.e., aiming to learn a model on labeled training graphs and generalize it to infer node labels on unlabeled test graphs. This problem has been extensively studied with graph neural networks (GNNs) by learning effective node representations, as well as traditional structured prediction methods for modeling the structured output of node labels, e.g., conditional random fields (CRFs). In this paper, we present a new approach called the Structured Proxy Network (SPN), which combines the advantages of both worlds. SPN defines flexible potential functions of CRFs with GNNs. However, learning such a model is nontrivial as it involves optimizing a maximin game with high-cost inference. Inspired by the underlying connection between joint and marginal distributions defined by Markov networks, we propose to solve an approximate version of the optimization problem as a proxy, which yields a near-optimal solution, making learning more efficient. Extensive experiments on two settings show that our approach outperforms many competitive baselines.
2022-04-15
{'id': '2204.07524v1', 'title': 'Neural Structured Prediction for Inductive Node Classification', 'abstract': 'This paper studies node classification in the inductive setting, i.e., aiming\nto learn a model on labeled training graphs and generalize it to infer node\nlabels on unlabeled test graphs. This problem has been extensively studied with\ngraph neural networks (GNNs) by learning effective node representations, as\nwell as traditional structured prediction methods for modeling the structured\noutput of node labels, e.g., conditional random fields (CRFs). In this paper,\nwe present a new approach called the Structured Proxy Network (SPN), which\ncombines the advantages of both worlds. SPN defines flexible potential\nfunctions of CRFs with GNNs. However, learning such a model is nontrivial as it\ninvolves optimizing a maximin game with high-cost inference. Inspired by the\nunderlying connection between joint and marginal distributions defined by\nMarkov networks, we propose to solve an approximate version of the optimization\nproblem as a proxy, which yields a near-optimal solution, making learning more\nefficient. Extensive experiments on two settings show that our approach\noutperforms many competitive baselines.', 'authors': ['Meng Qu', 'Huiyu Cai', 'Jian Tang'], 'published': '2022-04-15 15:50:27+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2204.07524v1'}
42,304
2306.01800v1
2,306.018
1
The ethical ambiguity of AI data enrichment: Measuring gaps in research ethics norms and practices
The technical progression of artificial intelligence (AI) research has been built on breakthroughs in fields such as computer science, statistics, and mathematics. However, in the past decade AI researchers have increasingly looked to the social sciences, turning to human interactions to solve the challenges of model development. Paying crowdsourcing workers to generate or curate data, or data enrichment, has become indispensable for many areas of AI research, from natural language processing to reinforcement learning from human feedback (RLHF). Other fields that routinely interact with crowdsourcing workers, such as Psychology, have developed common governance requirements and norms to ensure research is undertaken ethically. This study explores how, and to what extent, comparable research ethics requirements and norms have developed for AI research and data enrichment. We focus on the approach taken by two leading conferences: ICLR and NeurIPS, and journal publisher Springer. In a longitudinal study of accepted papers, and via a comparison with Psychology and CHI papers, this work finds that leading AI venues have begun to establish protocols for human data collection, but these are are inconsistently followed by authors. Whilst Psychology papers engaging with crowdsourcing workers frequently disclose ethics reviews, payment data, demographic data and other information, similar disclosures are far less common in leading AI venues despite similar guidance. The work concludes with hypotheses to explain these gaps in research ethics practices and considerations for its implications.
2023-06-01
{'id': '2306.01800v1', 'title': 'The ethical ambiguity of AI data enrichment: Measuring gaps in research ethics norms and practices', 'abstract': 'The technical progression of artificial intelligence (AI) research has been\nbuilt on breakthroughs in fields such as computer science, statistics, and\nmathematics. However, in the past decade AI researchers have increasingly\nlooked to the social sciences, turning to human interactions to solve the\nchallenges of model development. Paying crowdsourcing workers to generate or\ncurate data, or data enrichment, has become indispensable for many areas of AI\nresearch, from natural language processing to reinforcement learning from human\nfeedback (RLHF). Other fields that routinely interact with crowdsourcing\nworkers, such as Psychology, have developed common governance requirements and\nnorms to ensure research is undertaken ethically. This study explores how, and\nto what extent, comparable research ethics requirements and norms have\ndeveloped for AI research and data enrichment. We focus on the approach taken\nby two leading conferences: ICLR and NeurIPS, and journal publisher Springer.\nIn a longitudinal study of accepted papers, and via a comparison with\nPsychology and CHI papers, this work finds that leading AI venues have begun to\nestablish protocols for human data collection, but these are are inconsistently\nfollowed by authors. Whilst Psychology papers engaging with crowdsourcing\nworkers frequently disclose ethics reviews, payment data, demographic data and\nother information, similar disclosures are far less common in leading AI venues\ndespite similar guidance. The work concludes with hypotheses to explain these\ngaps in research ethics practices and considerations for its implications.', 'authors': ['Will Hawkins', 'Brent Mittelstadt'], 'published': '2023-06-01 16:12:55+00:00', 'categories': ['cs.CY', 'cs.AI', 'N/A', 'K.4.1'], 'url': 'http://arxiv.org/abs/2306.01800v1'}
42,305
2311.18684v1
2,311.18684
1
Handling Cost and Constraints with Off-Policy Deep Reinforcement Learning
By reusing data throughout training, off-policy deep reinforcement learning algorithms offer improved sample efficiency relative to on-policy approaches. For continuous action spaces, the most popular methods for off-policy learning include policy improvement steps where a learned state-action ($Q$) value function is maximized over selected batches of data. These updates are often paired with regularization to combat associated overestimation of $Q$ values. With an eye toward safety, we revisit this strategy in environments with "mixed-sign" reward functions; that is, with reward functions that include independent positive (incentive) and negative (cost) terms. This setting is common in real-world applications, and may be addressed with or without constraints on the cost terms. We find the combination of function approximation and a term that maximizes $Q$ in the policy update to be problematic in such environments, because systematic errors in value estimation impact the contributions from the competing terms asymmetrically. This results in overemphasis of either incentives or costs and may severely limit learning. We explore two remedies to this issue. First, consistent with prior work, we find that periodic resetting of $Q$ and policy networks can be used to reduce value estimation error and improve learning in this setting. Second, we formulate novel off-policy actor-critic methods for both unconstrained and constrained learning that do not explicitly maximize $Q$ in the policy update. We find that this second approach, when applied to continuous action spaces with mixed-sign rewards, consistently and significantly outperforms state-of-the-art methods augmented by resetting. We further find that our approach produces agents that are both competitive with popular methods overall and more reliably competent on frequently-studied control problems that do not have mixed-sign rewards.
2023-11-30
{'id': '2311.18684v1', 'title': 'Handling Cost and Constraints with Off-Policy Deep Reinforcement Learning', 'abstract': 'By reusing data throughout training, off-policy deep reinforcement learning\nalgorithms offer improved sample efficiency relative to on-policy approaches.\nFor continuous action spaces, the most popular methods for off-policy learning\ninclude policy improvement steps where a learned state-action ($Q$) value\nfunction is maximized over selected batches of data. These updates are often\npaired with regularization to combat associated overestimation of $Q$ values.\nWith an eye toward safety, we revisit this strategy in environments with\n"mixed-sign" reward functions; that is, with reward functions that include\nindependent positive (incentive) and negative (cost) terms. This setting is\ncommon in real-world applications, and may be addressed with or without\nconstraints on the cost terms. We find the combination of function\napproximation and a term that maximizes $Q$ in the policy update to be\nproblematic in such environments, because systematic errors in value estimation\nimpact the contributions from the competing terms asymmetrically. This results\nin overemphasis of either incentives or costs and may severely limit learning.\nWe explore two remedies to this issue. First, consistent with prior work, we\nfind that periodic resetting of $Q$ and policy networks can be used to reduce\nvalue estimation error and improve learning in this setting. Second, we\nformulate novel off-policy actor-critic methods for both unconstrained and\nconstrained learning that do not explicitly maximize $Q$ in the policy update.\nWe find that this second approach, when applied to continuous action spaces\nwith mixed-sign rewards, consistently and significantly outperforms\nstate-of-the-art methods augmented by resetting. We further find that our\napproach produces agents that are both competitive with popular methods overall\nand more reliably competent on frequently-studied control problems that do not\nhave mixed-sign rewards.', 'authors': ['Jared Markowitz', 'Jesse Silverberg', 'Gary Collins'], 'published': '2023-11-30 16:31:04+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2311.18684v1'}
42,306
2311.15100v2
2,311.151
2
Unbalancedness in Neural Monge Maps Improves Unpaired Domain Translation
In optimal transport (OT), a Monge map is known as a mapping that transports a source distribution to a target distribution in the most cost-efficient way. Recently, multiple neural estimators for Monge maps have been developed and applied in diverse unpaired domain translation tasks, e.g. in single-cell biology and computer vision. However, the classic OT framework enforces mass conservation, which makes it prone to outliers and limits its applicability in real-world scenarios. The latter can be particularly harmful in OT domain translation tasks, where the relative position of a sample within a distribution is explicitly taken into account. While unbalanced OT tackles this challenge in the discrete setting, its integration into neural Monge map estimators has received limited attention. We propose a theoretically grounded method to incorporate unbalancedness into any Monge map estimator. We improve existing estimators to model cell trajectories over time and to predict cellular responses to perturbations. Moreover, our approach seamlessly integrates with the OT flow matching (OT-FM) framework. While we show that OT-FM performs competitively in image translation, we further improve performance by incorporating unbalancedness (UOT-FM), which better preserves relevant features. We hence establish UOT-FM as a principled method for unpaired image translation.
2023-11-25
{'id': '2311.15100v2', 'title': 'Unbalancedness in Neural Monge Maps Improves Unpaired Domain Translation', 'abstract': 'In optimal transport (OT), a Monge map is known as a mapping that transports\na source distribution to a target distribution in the most cost-efficient way.\nRecently, multiple neural estimators for Monge maps have been developed and\napplied in diverse unpaired domain translation tasks, e.g. in single-cell\nbiology and computer vision. However, the classic OT framework enforces mass\nconservation, which makes it prone to outliers and limits its applicability in\nreal-world scenarios. The latter can be particularly harmful in OT domain\ntranslation tasks, where the relative position of a sample within a\ndistribution is explicitly taken into account. While unbalanced OT tackles this\nchallenge in the discrete setting, its integration into neural Monge map\nestimators has received limited attention. We propose a theoretically grounded\nmethod to incorporate unbalancedness into any Monge map estimator. We improve\nexisting estimators to model cell trajectories over time and to predict\ncellular responses to perturbations. Moreover, our approach seamlessly\nintegrates with the OT flow matching (OT-FM) framework. While we show that\nOT-FM performs competitively in image translation, we further improve\nperformance by incorporating unbalancedness (UOT-FM), which better preserves\nrelevant features. We hence establish UOT-FM as a principled method for\nunpaired image translation.', 'authors': ['Luca Eyring', 'Dominik Klein', 'Théo Uscidda', 'Giovanni Palla', 'Niki Kilbertus', 'Zeynep Akata', 'Fabian Theis'], 'published': '2023-11-25 18:58:15+00:00', 'categories': ['cs.CV', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2311.15100v2'}
42,307
2310.08337v3
2,310.08337
3
Neural Diffusion Models
Diffusion models have shown remarkable performance on many generative tasks. Despite recent success, most diffusion models are restricted in that they only allow linear transformation of the data distribution. In contrast, broader family of transformations can potentially help train generative distributions more efficiently, simplifying the reverse process and closing the gap between the true negative log-likelihood and the variational approximation. In this paper, we present Neural Diffusion Models (NDMs), a generalization of conventional diffusion models that enables defining and learning time-dependent non-linear transformations of data. We show how to optimise NDMs using a variational bound in a simulation-free setting. Moreover, we derive a time-continuous formulation of NDMs, which allows fast and reliable inference using off-the-shelf numerical ODE and SDE solvers. Finally, we demonstrate the utility of NDMs with learnable transformations through experiments on standard image generation benchmarks, including CIFAR-10, downsampled versions of ImageNet and CelebA-HQ. NDMs outperform conventional diffusion models in terms of likelihood and produce high-quality samples.
2023-10-12
{'id': '2310.08337v3', 'title': 'Neural Diffusion Models', 'abstract': 'Diffusion models have shown remarkable performance on many generative tasks.\nDespite recent success, most diffusion models are restricted in that they only\nallow linear transformation of the data distribution. In contrast, broader\nfamily of transformations can potentially help train generative distributions\nmore efficiently, simplifying the reverse process and closing the gap between\nthe true negative log-likelihood and the variational approximation. In this\npaper, we present Neural Diffusion Models (NDMs), a generalization of\nconventional diffusion models that enables defining and learning time-dependent\nnon-linear transformations of data. We show how to optimise NDMs using a\nvariational bound in a simulation-free setting. Moreover, we derive a\ntime-continuous formulation of NDMs, which allows fast and reliable inference\nusing off-the-shelf numerical ODE and SDE solvers. Finally, we demonstrate the\nutility of NDMs with learnable transformations through experiments on standard\nimage generation benchmarks, including CIFAR-10, downsampled versions of\nImageNet and CelebA-HQ. NDMs outperform conventional diffusion models in terms\nof likelihood and produce high-quality samples.', 'authors': ['Grigory Bartosh', 'Dmitry Vetrov', 'Christian A. Naesseth'], 'published': '2023-10-12 13:54:55+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2310.08337v3'}
42,308
2310.04560v1
2,310.0456
1
Talk like a Graph: Encoding Graphs for Large Language Models
Graphs are a powerful tool for representing and analyzing complex relationships in real-world applications such as social networks, recommender systems, and computational finance. Reasoning on graphs is essential for drawing inferences about the relationships between entities in a complex system, and to identify hidden patterns and trends. Despite the remarkable progress in automated reasoning with natural text, reasoning on graphs with large language models (LLMs) remains an understudied problem. In this work, we perform the first comprehensive study of encoding graph-structured data as text for consumption by LLMs. We show that LLM performance on graph reasoning tasks varies on three fundamental levels: (1) the graph encoding method, (2) the nature of the graph task itself, and (3) interestingly, the very structure of the graph considered. These novel results provide valuable insight on strategies for encoding graphs as text. Using these insights we illustrate how the correct choice of encoders can boost performance on graph reasoning tasks inside LLMs by 4.8% to 61.8%, depending on the task.
2023-10-06
{'id': '2310.04560v1', 'title': 'Talk like a Graph: Encoding Graphs for Large Language Models', 'abstract': 'Graphs are a powerful tool for representing and analyzing complex\nrelationships in real-world applications such as social networks, recommender\nsystems, and computational finance. Reasoning on graphs is essential for\ndrawing inferences about the relationships between entities in a complex\nsystem, and to identify hidden patterns and trends. Despite the remarkable\nprogress in automated reasoning with natural text, reasoning on graphs with\nlarge language models (LLMs) remains an understudied problem. In this work, we\nperform the first comprehensive study of encoding graph-structured data as text\nfor consumption by LLMs. We show that LLM performance on graph reasoning tasks\nvaries on three fundamental levels: (1) the graph encoding method, (2) the\nnature of the graph task itself, and (3) interestingly, the very structure of\nthe graph considered. These novel results provide valuable insight on\nstrategies for encoding graphs as text. Using these insights we illustrate how\nthe correct choice of encoders can boost performance on graph reasoning tasks\ninside LLMs by 4.8% to 61.8%, depending on the task.', 'authors': ['Bahare Fatemi', 'Jonathan Halcrow', 'Bryan Perozzi'], 'published': '2023-10-06 19:55:21+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.04560v1'}
42,309
2310.00545v1
2,310.00545
1
Implicit Neural Representations and the Algebra of Complex Wavelets
Implicit neural representations (INRs) have arisen as useful methods for representing signals on Euclidean domains. By parameterizing an image as a multilayer perceptron (MLP) on Euclidean space, INRs effectively represent signals in a way that couples spatial and spectral features of the signal that is not obvious in the usual discrete representation, paving the way for continuous signal processing and machine learning approaches that were not previously possible. Although INRs using sinusoidal activation functions have been studied in terms of Fourier theory, recent works have shown the advantage of using wavelets instead of sinusoids as activation functions, due to their ability to simultaneously localize in both frequency and space. In this work, we approach such INRs and demonstrate how they resolve high-frequency features of signals from coarse approximations done in the first layer of the MLP. This leads to multiple prescriptions for the design of INR architectures, including the use of complex wavelets, decoupling of low and band-pass approximations, and initialization schemes based on the singularities of the desired signal.
2023-10-01
{'id': '2310.00545v1', 'title': 'Implicit Neural Representations and the Algebra of Complex Wavelets', 'abstract': 'Implicit neural representations (INRs) have arisen as useful methods for\nrepresenting signals on Euclidean domains. By parameterizing an image as a\nmultilayer perceptron (MLP) on Euclidean space, INRs effectively represent\nsignals in a way that couples spatial and spectral features of the signal that\nis not obvious in the usual discrete representation, paving the way for\ncontinuous signal processing and machine learning approaches that were not\npreviously possible. Although INRs using sinusoidal activation functions have\nbeen studied in terms of Fourier theory, recent works have shown the advantage\nof using wavelets instead of sinusoids as activation functions, due to their\nability to simultaneously localize in both frequency and space. In this work,\nwe approach such INRs and demonstrate how they resolve high-frequency features\nof signals from coarse approximations done in the first layer of the MLP. This\nleads to multiple prescriptions for the design of INR architectures, including\nthe use of complex wavelets, decoupling of low and band-pass approximations,\nand initialization schemes based on the singularities of the desired signal.', 'authors': ['T. Mitchell Roddenberry', 'Vishwanath Saragadam', 'Maarten V. de Hoop', 'Richard G. Baraniuk'], 'published': '2023-10-01 02:01:28+00:00', 'categories': ['eess.SP', 'cs.CV', 'eess.IV'], 'url': 'http://arxiv.org/abs/2310.00545v1'}
42,310
2310.08056v4
2,310.08056
4
Learning from Label Proportions: Bootstrapping Supervised Learners via Belief Propagation
Learning from Label Proportions (LLP) is a learning problem where only aggregate level labels are available for groups of instances, called bags, during training, and the aim is to get the best performance at the instance-level on the test data. This setting arises in domains like advertising and medicine due to privacy considerations. We propose a novel algorithmic framework for this problem that iteratively performs two main steps. For the first step (Pseudo Labeling) in every iteration, we define a Gibbs distribution over binary instance labels that incorporates a) covariate information through the constraint that instances with similar covariates should have similar labels and b) the bag level aggregated label. We then use Belief Propagation (BP) to marginalize the Gibbs distribution to obtain pseudo labels. In the second step (Embedding Refinement), we use the pseudo labels to provide supervision for a learner that yields a better embedding. Further, we iterate on the two steps again by using the second step's embeddings as new covariates for the next iteration. In the final iteration, a classifier is trained using the pseudo labels. Our algorithm displays strong gains against several SOTA baselines (up to 15%) for the LLP Binary Classification problem on various dataset types - tabular and Image. We achieve these improvements with minimal computational overhead above standard supervised learning due to Belief Propagation, for large bag sizes, even for a million samples.
2023-10-12
{'id': '2310.08056v4', 'title': 'Learning from Label Proportions: Bootstrapping Supervised Learners via Belief Propagation', 'abstract': "Learning from Label Proportions (LLP) is a learning problem where only\naggregate level labels are available for groups of instances, called bags,\nduring training, and the aim is to get the best performance at the\ninstance-level on the test data. This setting arises in domains like\nadvertising and medicine due to privacy considerations. We propose a novel\nalgorithmic framework for this problem that iteratively performs two main\nsteps. For the first step (Pseudo Labeling) in every iteration, we define a\nGibbs distribution over binary instance labels that incorporates a) covariate\ninformation through the constraint that instances with similar covariates\nshould have similar labels and b) the bag level aggregated label. We then use\nBelief Propagation (BP) to marginalize the Gibbs distribution to obtain pseudo\nlabels. In the second step (Embedding Refinement), we use the pseudo labels to\nprovide supervision for a learner that yields a better embedding. Further, we\niterate on the two steps again by using the second step's embeddings as new\ncovariates for the next iteration. In the final iteration, a classifier is\ntrained using the pseudo labels. Our algorithm displays strong gains against\nseveral SOTA baselines (up to 15%) for the LLP Binary Classification problem on\nvarious dataset types - tabular and Image. We achieve these improvements with\nminimal computational overhead above standard supervised learning due to Belief\nPropagation, for large bag sizes, even for a million samples.", 'authors': ['Shreyas Havaldar', 'Navodita Sharma', 'Shubhi Sareen', 'Karthikeyan Shanmugam', 'Aravindan Raghuveer'], 'published': '2023-10-12 06:09:26+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2310.08056v4'}
42,318
2305.12723v2
2,305.12723
2
Enhancing Small Medical Learners with Privacy-preserving Contextual Prompting
Large language models (LLMs) demonstrate remarkable medical expertise, but data privacy concerns impede their direct use in healthcare environments. Although offering improved data privacy protection, domain-specific small language models (SLMs) often underperform LLMs, emphasizing the need for methods that reduce this performance gap while alleviating privacy concerns. In this paper, we present a simple yet effective method that harnesses LLMs' medical proficiency to boost SLM performance in medical tasks under privacy-restricted scenarios. Specifically, we mitigate patient privacy issues by extracting keywords from medical data and prompting the LLM to generate a medical knowledge-intensive context by simulating clinicians' thought processes. This context serves as additional input for SLMs, augmenting their decision-making capabilities. Our method significantly enhances performance in both few-shot and full training settings across three medical knowledge-intensive tasks, achieving up to a 22.57% increase in absolute accuracy compared to SLM fine-tuning without context, and sets new state-of-the-art results in two medical tasks within privacy-restricted scenarios. Further out-of-domain testing and experiments in two general domain datasets showcase its generalizability and broad applicability. Our code can be found at https://github.com/XZhang97666/PrivacyBoost-SLM.
2023-05-22
{'id': '2305.12723v2', 'title': 'Enhancing Small Medical Learners with Privacy-preserving Contextual Prompting', 'abstract': "Large language models (LLMs) demonstrate remarkable medical expertise, but\ndata privacy concerns impede their direct use in healthcare environments.\nAlthough offering improved data privacy protection, domain-specific small\nlanguage models (SLMs) often underperform LLMs, emphasizing the need for\nmethods that reduce this performance gap while alleviating privacy concerns. In\nthis paper, we present a simple yet effective method that harnesses LLMs'\nmedical proficiency to boost SLM performance in medical tasks under\nprivacy-restricted scenarios. Specifically, we mitigate patient privacy issues\nby extracting keywords from medical data and prompting the LLM to generate a\nmedical knowledge-intensive context by simulating clinicians' thought\nprocesses. This context serves as additional input for SLMs, augmenting their\ndecision-making capabilities. Our method significantly enhances performance in\nboth few-shot and full training settings across three medical\nknowledge-intensive tasks, achieving up to a 22.57% increase in absolute\naccuracy compared to SLM fine-tuning without context, and sets new\nstate-of-the-art results in two medical tasks within privacy-restricted\nscenarios. Further out-of-domain testing and experiments in two general domain\ndatasets showcase its generalizability and broad applicability. Our code can be\nfound at https://github.com/XZhang97666/PrivacyBoost-SLM.", 'authors': ['Xinlu Zhang', 'Shiyang Li', 'Xianjun Yang', 'Chenxin Tian', 'Yao Qin', 'Linda Ruth Petzold'], 'published': '2023-05-22 05:14:38+00:00', 'categories': ['cs.CL', 'cs.AI'], 'url': 'http://arxiv.org/abs/2305.12723v2'}
42,320
2401.16444v1
2,401.16444
1
Enhancing Human Experience in Human-Agent Collaboration: A Human-Centered Modeling Approach Based on Positive Human Gain
Existing game AI research mainly focuses on enhancing agents' abilities to win games, but this does not inherently make humans have a better experience when collaborating with these agents. For example, agents may dominate the collaboration and exhibit unintended or detrimental behaviors, leading to poor experiences for their human partners. In other words, most game AI agents are modeled in a "self-centered" manner. In this paper, we propose a "human-centered" modeling scheme for collaborative agents that aims to enhance the experience of humans. Specifically, we model the experience of humans as the goals they expect to achieve during the task. We expect that agents should learn to enhance the extent to which humans achieve these goals while maintaining agents' original abilities (e.g., winning games). To achieve this, we propose the Reinforcement Learning from Human Gain (RLHG) approach. The RLHG approach introduces a "baseline", which corresponds to the extent to which humans primitively achieve their goals, and encourages agents to learn behaviors that can effectively enhance humans in achieving their goals better. We evaluate the RLHG agent in the popular Multi-player Online Battle Arena (MOBA) game, Honor of Kings, by conducting real-world human-agent tests. Both objective performance and subjective preference results show that the RLHG agent provides participants better gaming experience.
2024-01-28
{'id': '2401.16444v1', 'title': 'Enhancing Human Experience in Human-Agent Collaboration: A Human-Centered Modeling Approach Based on Positive Human Gain', 'abstract': 'Existing game AI research mainly focuses on enhancing agents\' abilities to\nwin games, but this does not inherently make humans have a better experience\nwhen collaborating with these agents. For example, agents may dominate the\ncollaboration and exhibit unintended or detrimental behaviors, leading to poor\nexperiences for their human partners. In other words, most game AI agents are\nmodeled in a "self-centered" manner. In this paper, we propose a\n"human-centered" modeling scheme for collaborative agents that aims to enhance\nthe experience of humans. Specifically, we model the experience of humans as\nthe goals they expect to achieve during the task. We expect that agents should\nlearn to enhance the extent to which humans achieve these goals while\nmaintaining agents\' original abilities (e.g., winning games). To achieve this,\nwe propose the Reinforcement Learning from Human Gain (RLHG) approach. The RLHG\napproach introduces a "baseline", which corresponds to the extent to which\nhumans primitively achieve their goals, and encourages agents to learn\nbehaviors that can effectively enhance humans in achieving their goals better.\nWe evaluate the RLHG agent in the popular Multi-player Online Battle Arena\n(MOBA) game, Honor of Kings, by conducting real-world human-agent tests. Both\nobjective performance and subjective preference results show that the RLHG\nagent provides participants better gaming experience.', 'authors': ['Yiming Gao', 'Feiyu Liu', 'Liang Wang', 'Zhenjie Lian', 'Dehua Zheng', 'Weixuan Wang', 'Wenjin Yang', 'Siqin Li', 'Xianliang Wang', 'Wenhui Chen', 'Jing Dai', 'Qiang Fu', 'Wei Yang', 'Lanxiao Huang', 'Wei Liu'], 'published': '2024-01-28 05:05:57+00:00', 'categories': ['cs.HC', 'cs.AI'], 'url': 'http://arxiv.org/abs/2401.16444v1'}
42,323
2404.03037v3
2,404.03037
3
Model-based Reinforcement Learning for Parameterized Action Spaces
We propose a novel model-based reinforcement learning algorithm -- Dynamics Learning and predictive control with Parameterized Actions (DLPA) -- for Parameterized Action Markov Decision Processes (PAMDPs). The agent learns a parameterized-action-conditioned dynamics model and plans with a modified Model Predictive Path Integral control. We theoretically quantify the difference between the generated trajectory and the optimal trajectory during planning in terms of the value they achieved through the lens of Lipschitz Continuity. Our empirical results on several standard benchmarks show that our algorithm achieves superior sample efficiency and asymptotic performance than state-of-the-art PAMDP methods.
2024-04-03
{'id': '2404.03037v3', 'title': 'Model-based Reinforcement Learning for Parameterized Action Spaces', 'abstract': 'We propose a novel model-based reinforcement learning algorithm -- Dynamics\nLearning and predictive control with Parameterized Actions (DLPA) -- for\nParameterized Action Markov Decision Processes (PAMDPs). The agent learns a\nparameterized-action-conditioned dynamics model and plans with a modified Model\nPredictive Path Integral control. We theoretically quantify the difference\nbetween the generated trajectory and the optimal trajectory during planning in\nterms of the value they achieved through the lens of Lipschitz Continuity. Our\nempirical results on several standard benchmarks show that our algorithm\nachieves superior sample efficiency and asymptotic performance than\nstate-of-the-art PAMDP methods.', 'authors': ['Renhao Zhang', 'Haotian Fu', 'Yilin Miao', 'George Konidaris'], 'published': '2024-04-03 19:48:13+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2404.03037v3'}
42,312
2311.14770v1
2,311.1477
1
Learning to Cooperate and Communicate Over Imperfect Channels
Information exchange in multi-agent systems improves the cooperation among agents, especially in partially observable settings. In the real world, communication is often carried out over imperfect channels. This requires agents to handle uncertainty due to potential information loss. In this paper, we consider a cooperative multi-agent system where the agents act and exchange information in a decentralized manner using a limited and unreliable channel. To cope with such channel constraints, we propose a novel communication approach based on independent Q-learning. Our method allows agents to dynamically adapt how much information to share by sending messages of different sizes, depending on their local observations and the channel's properties. In addition to this message size selection, agents learn to encode and decode messages to improve their jointly trained policies. We show that our approach outperforms approaches without adaptive capabilities in a novel cooperative digit-prediction environment and discuss its limitations in the traffic junction environment.
2023-11-24
{'id': '2311.14770v1', 'title': 'Learning to Cooperate and Communicate Over Imperfect Channels', 'abstract': "Information exchange in multi-agent systems improves the cooperation among\nagents, especially in partially observable settings. In the real world,\ncommunication is often carried out over imperfect channels. This requires\nagents to handle uncertainty due to potential information loss. In this paper,\nwe consider a cooperative multi-agent system where the agents act and exchange\ninformation in a decentralized manner using a limited and unreliable channel.\nTo cope with such channel constraints, we propose a novel communication\napproach based on independent Q-learning. Our method allows agents to\ndynamically adapt how much information to share by sending messages of\ndifferent sizes, depending on their local observations and the channel's\nproperties. In addition to this message size selection, agents learn to encode\nand decode messages to improve their jointly trained policies. We show that our\napproach outperforms approaches without adaptive capabilities in a novel\ncooperative digit-prediction environment and discuss its limitations in the\ntraffic junction environment.", 'authors': ['Jannis Weil', 'Gizem Ekinci', 'Heinz Koeppl', 'Tobias Meuser'], 'published': '2023-11-24 12:15:48+00:00', 'categories': ['cs.MA', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2311.14770v1'}
42,351
2203.10190v3
2,203.1019
3
Fair Federated Learning via Bounded Group Loss
Fair prediction across protected groups is an important constraint for many federated learning applications. However, prior work studying group fair federated learning lacks formal convergence or fairness guarantees. In this work we propose a general framework for provably fair federated learning. In particular, we explore and extend the notion of Bounded Group Loss as a theoretically-grounded approach for group fairness. Using this setup, we propose a scalable federated optimization method that optimizes the empirical risk under a number of group fairness constraints. We provide convergence guarantees for the method as well as fairness guarantees for the resulting solution. Empirically, we evaluate our method across common benchmarks from fair ML and federated learning, showing that it can provide both fairer and more accurate predictions than baseline approaches.
2022-03-18
{'id': '2203.10190v3', 'title': 'Fair Federated Learning via Bounded Group Loss', 'abstract': 'Fair prediction across protected groups is an important constraint for many\nfederated learning applications. However, prior work studying group fair\nfederated learning lacks formal convergence or fairness guarantees. In this\nwork we propose a general framework for provably fair federated learning. In\nparticular, we explore and extend the notion of Bounded Group Loss as a\ntheoretically-grounded approach for group fairness. Using this setup, we\npropose a scalable federated optimization method that optimizes the empirical\nrisk under a number of group fairness constraints. We provide convergence\nguarantees for the method as well as fairness guarantees for the resulting\nsolution. Empirically, we evaluate our method across common benchmarks from\nfair ML and federated learning, showing that it can provide both fairer and\nmore accurate predictions than baseline approaches.', 'authors': ['Shengyuan Hu', 'Zhiwei Steven Wu', 'Virginia Smith'], 'published': '2022-03-18 23:11:54+00:00', 'categories': ['cs.LG', 'cs.CY'], 'url': 'http://arxiv.org/abs/2203.10190v3'}
42,354
2301.01702v2
2,301.01702
2
Automating Nearest Neighbor Search Configuration with Constrained Optimization
The approximate nearest neighbor (ANN) search problem is fundamental to efficiently serving many real-world machine learning applications. A number of techniques have been developed for ANN search that are efficient, accurate, and scalable. However, such techniques typically have a number of parameters that affect the speed-recall tradeoff, and exhibit poor performance when such parameters aren't properly set. Tuning these parameters has traditionally been a manual process, demanding in-depth knowledge of the underlying search algorithm. This is becoming an increasingly unrealistic demand as ANN search grows in popularity. To tackle this obstacle to ANN adoption, this work proposes a constrained optimization-based approach to tuning quantization-based ANN algorithms. Our technique takes just a desired search cost or recall as input, and then generates tunings that, empirically, are very close to the speed-recall Pareto frontier and give leading performance on standard benchmarks.
2023-01-04
{'id': '2301.01702v2', 'title': 'Automating Nearest Neighbor Search Configuration with Constrained Optimization', 'abstract': "The approximate nearest neighbor (ANN) search problem is fundamental to\nefficiently serving many real-world machine learning applications. A number of\ntechniques have been developed for ANN search that are efficient, accurate, and\nscalable. However, such techniques typically have a number of parameters that\naffect the speed-recall tradeoff, and exhibit poor performance when such\nparameters aren't properly set. Tuning these parameters has traditionally been\na manual process, demanding in-depth knowledge of the underlying search\nalgorithm. This is becoming an increasingly unrealistic demand as ANN search\ngrows in popularity. To tackle this obstacle to ANN adoption, this work\nproposes a constrained optimization-based approach to tuning quantization-based\nANN algorithms. Our technique takes just a desired search cost or recall as\ninput, and then generates tunings that, empirically, are very close to the\nspeed-recall Pareto frontier and give leading performance on standard\nbenchmarks.", 'authors': ['Philip Sun', 'Ruiqi Guo', 'Sanjiv Kumar'], 'published': '2023-01-04 16:56:36+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2301.01702v2'}
42,368
2210.12257v2
2,210.12257
2
Efficient Automatic Machine Learning via Design Graphs
Despite the success of automated machine learning (AutoML), which aims to find the best design, including the architecture of deep networks and hyper-parameters, conventional AutoML methods are computationally expensive and hardly provide insights into the relations of different model design choices. To tackle the challenges, we propose FALCON, an efficient sample-based method to search for the optimal model design. Our key insight is to model the design space of possible model designs as a design graph, where the nodes represent design choices, and the edges denote design similarities. FALCON features 1) a task-agnostic module, which performs message passing on the design graph via a Graph Neural Network (GNN), and 2) a task-specific module, which conducts label propagation of the known model performance information on the design graph. Both modules are combined to predict the design performances in the design space, navigating the search direction. We conduct extensive experiments on 27 node and graph classification tasks from various application domains, and an image classification task on the CIFAR-10 dataset. We empirically show that FALCON can efficiently obtain the well-performing designs for each task using only 30 explored nodes. Specifically, FALCON has a comparable time cost with the one-shot approaches while achieving an average improvement of 3.3% compared with the best baselines.
2022-10-21
{'id': '2210.12257v2', 'title': 'Efficient Automatic Machine Learning via Design Graphs', 'abstract': 'Despite the success of automated machine learning (AutoML), which aims to\nfind the best design, including the architecture of deep networks and\nhyper-parameters, conventional AutoML methods are computationally expensive and\nhardly provide insights into the relations of different model design choices.\nTo tackle the challenges, we propose FALCON, an efficient sample-based method\nto search for the optimal model design. Our key insight is to model the design\nspace of possible model designs as a design graph, where the nodes represent\ndesign choices, and the edges denote design similarities. FALCON features 1) a\ntask-agnostic module, which performs message passing on the design graph via a\nGraph Neural Network (GNN), and 2) a task-specific module, which conducts label\npropagation of the known model performance information on the design graph.\nBoth modules are combined to predict the design performances in the design\nspace, navigating the search direction. We conduct extensive experiments on 27\nnode and graph classification tasks from various application domains, and an\nimage classification task on the CIFAR-10 dataset. We empirically show that\nFALCON can efficiently obtain the well-performing designs for each task using\nonly 30 explored nodes. Specifically, FALCON has a comparable time cost with\nthe one-shot approaches while achieving an average improvement of 3.3% compared\nwith the best baselines.', 'authors': ['Shirley Wu', 'Jiaxuan You', 'Jure Leskovec', 'Rex Ying'], 'published': '2022-10-21 21:25:59+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2210.12257v2'}
42,370
2206.10012v1
2,206.10012
1
Limitations of the NTK for Understanding Generalization in Deep Learning
The ``Neural Tangent Kernel'' (NTK) (Jacot et al 2018), and its empirical variants have been proposed as a proxy to capture certain behaviors of real neural networks. In this work, we study NTKs through the lens of scaling laws, and demonstrate that they fall short of explaining important aspects of neural network generalization. In particular, we demonstrate realistic settings where finite-width neural networks have significantly better data scaling exponents as compared to their corresponding empirical and infinite NTKs at initialization. This reveals a more fundamental difference between the real networks and NTKs, beyond just a few percentage points of test accuracy. Further, we show that even if the empirical NTK is allowed to be pre-trained on a constant number of samples, the kernel scaling does not catch up to the neural network scaling. Finally, we show that the empirical NTK continues to evolve throughout most of the training, in contrast with prior work which suggests that it stabilizes after a few epochs of training. Altogether, our work establishes concrete limitations of the NTK approach in understanding generalization of real networks on natural datasets.
2022-06-20
{'id': '2206.10012v1', 'title': 'Limitations of the NTK for Understanding Generalization in Deep Learning', 'abstract': "The ``Neural Tangent Kernel'' (NTK) (Jacot et al 2018), and its empirical\nvariants have been proposed as a proxy to capture certain behaviors of real\nneural networks. In this work, we study NTKs through the lens of scaling laws,\nand demonstrate that they fall short of explaining important aspects of neural\nnetwork generalization. In particular, we demonstrate realistic settings where\nfinite-width neural networks have significantly better data scaling exponents\nas compared to their corresponding empirical and infinite NTKs at\ninitialization. This reveals a more fundamental difference between the real\nnetworks and NTKs, beyond just a few percentage points of test accuracy.\nFurther, we show that even if the empirical NTK is allowed to be pre-trained on\na constant number of samples, the kernel scaling does not catch up to the\nneural network scaling. Finally, we show that the empirical NTK continues to\nevolve throughout most of the training, in contrast with prior work which\nsuggests that it stabilizes after a few epochs of training. Altogether, our\nwork establishes concrete limitations of the NTK approach in understanding\ngeneralization of real networks on natural datasets.", 'authors': ['Nikhil Vyas', 'Yamini Bansal', 'Preetum Nakkiran'], 'published': '2022-06-20 21:23:28+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2206.10012v1'}
42,377
2202.04414v2
2,202.04414
2
Agree to Disagree: Diversity through Disagreement for Better Transferability
Gradient-based learning algorithms have an implicit simplicity bias which in effect can limit the diversity of predictors being sampled by the learning procedure. This behavior can hinder the transferability of trained models by (i) favoring the learning of simpler but spurious features -- present in the training data but absent from the test data -- and (ii) by only leveraging a small subset of predictive features. Such an effect is especially magnified when the test distribution does not exactly match the train distribution -- referred to as the Out of Distribution (OOD) generalization problem. However, given only the training data, it is not always possible to apriori assess if a given feature is spurious or transferable. Instead, we advocate for learning an ensemble of models which capture a diverse set of predictive features. Towards this, we propose a new algorithm D-BAT (Diversity-By-disAgreement Training), which enforces agreement among the models on the training data, but disagreement on the OOD data. We show how D-BAT naturally emerges from the notion of generalized discrepancy, as well as demonstrate in multiple experiments how the proposed method can mitigate shortcut-learning, enhance uncertainty and OOD detection, as well as improve transferability.
2022-02-09
{'id': '2202.04414v2', 'title': 'Agree to Disagree: Diversity through Disagreement for Better Transferability', 'abstract': 'Gradient-based learning algorithms have an implicit simplicity bias which in\neffect can limit the diversity of predictors being sampled by the learning\nprocedure. This behavior can hinder the transferability of trained models by\n(i) favoring the learning of simpler but spurious features -- present in the\ntraining data but absent from the test data -- and (ii) by only leveraging a\nsmall subset of predictive features. Such an effect is especially magnified\nwhen the test distribution does not exactly match the train distribution --\nreferred to as the Out of Distribution (OOD) generalization problem. However,\ngiven only the training data, it is not always possible to apriori assess if a\ngiven feature is spurious or transferable. Instead, we advocate for learning an\nensemble of models which capture a diverse set of predictive features. Towards\nthis, we propose a new algorithm D-BAT (Diversity-By-disAgreement Training),\nwhich enforces agreement among the models on the training data, but\ndisagreement on the OOD data. We show how D-BAT naturally emerges from the\nnotion of generalized discrepancy, as well as demonstrate in multiple\nexperiments how the proposed method can mitigate shortcut-learning, enhance\nuncertainty and OOD detection, as well as improve transferability.', 'authors': ['Matteo Pagliardini', 'Martin Jaggi', 'François Fleuret', 'Sai Praneeth Karimireddy'], 'published': '2022-02-09 12:03:02+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2202.04414v2'}
42,313
2208.08386v2
2,208.08386
2
Neural Embeddings for Text
We propose a new kind of embedding for natural language text that deeply represents semantic meaning. Standard text embeddings use the outputs from hidden layers of a pretrained language model. In our method, we let a language model learn from the text and then literally pick its brain, taking the actual weights of the model's neurons to generate a vector. We call this representation of the text a neural embedding. We confirm the ability of this representation to reflect semantics of the text by an analysis of its behavior on several datasets, and by a comparison of neural embedding with state of the art sentence embeddings.
2022-08-17
{'id': '2208.08386v2', 'title': 'Neural Embeddings for Text', 'abstract': "We propose a new kind of embedding for natural language text that deeply\nrepresents semantic meaning. Standard text embeddings use the outputs from\nhidden layers of a pretrained language model. In our method, we let a language\nmodel learn from the text and then literally pick its brain, taking the actual\nweights of the model's neurons to generate a vector. We call this\nrepresentation of the text a neural embedding. We confirm the ability of this\nrepresentation to reflect semantics of the text by an analysis of its behavior\non several datasets, and by a comparison of neural embedding with state of the\nart sentence embeddings.", 'authors': ['Oleg Vasilyev', 'John Bohannon'], 'published': '2022-08-17 16:26:13+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2208.08386v2'}
42,316
2401.12617v2
2,401.12617
2
The Joint Effect of Task Similarity and Overparameterization on Catastrophic Forgetting -- An Analytical Model
In continual learning, catastrophic forgetting is affected by multiple aspects of the tasks. Previous works have analyzed separately how forgetting is affected by either task similarity or overparameterization. In contrast, our paper examines how task similarity and overparameterization jointly affect forgetting in an analyzable model. Specifically, we focus on two-task continual linear regression, where the second task is a random orthogonal transformation of an arbitrary first task (an abstraction of random permutation tasks). We derive an exact analytical expression for the expected forgetting - and uncover a nuanced pattern. In highly overparameterized models, intermediate task similarity causes the most forgetting. However, near the interpolation threshold, forgetting decreases monotonically with the expected task similarity. We validate our findings with linear regression on synthetic data, and with neural networks on established permutation task benchmarks.
2024-01-23
{'id': '2401.12617v2', 'title': 'The Joint Effect of Task Similarity and Overparameterization on Catastrophic Forgetting -- An Analytical Model', 'abstract': 'In continual learning, catastrophic forgetting is affected by multiple\naspects of the tasks. Previous works have analyzed separately how forgetting is\naffected by either task similarity or overparameterization. In contrast, our\npaper examines how task similarity and overparameterization jointly affect\nforgetting in an analyzable model. Specifically, we focus on two-task continual\nlinear regression, where the second task is a random orthogonal transformation\nof an arbitrary first task (an abstraction of random permutation tasks). We\nderive an exact analytical expression for the expected forgetting - and uncover\na nuanced pattern. In highly overparameterized models, intermediate task\nsimilarity causes the most forgetting. However, near the interpolation\nthreshold, forgetting decreases monotonically with the expected task\nsimilarity. We validate our findings with linear regression on synthetic data,\nand with neural networks on established permutation task benchmarks.', 'authors': ['Daniel Goldfarb', 'Itay Evron', 'Nir Weinberger', 'Daniel Soudry', 'Paul Hand'], 'published': '2024-01-23 10:16:44+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2401.12617v2'}
42,330
2207.07696v1
2,207.07696
1
Algorithmic Determination of the Combinatorial Structure of the Linear Regions of ReLU Neural Networks
We algorithmically determine the regions and facets of all dimensions of the canonical polyhedral complex, the universal object into which a ReLU network decomposes its input space. We show that the locations of the vertices of the canonical polyhedral complex along with their signs with respect to layer maps determine the full facet structure across all dimensions. We present an algorithm which calculates this full combinatorial structure, making use of our theorems that the dual complex to the canonical polyhedral complex is cubical and it possesses a multiplication compatible with its facet structure. The resulting algorithm is numerically stable, polynomial time in the number of intermediate neurons, and obtains accurate information across all dimensions. This permits us to obtain, for example, the true topology of the decision boundaries of networks with low-dimensional inputs. We run empirics on such networks at initialization, finding that width alone does not increase observed topology, but width in the presence of depth does. Source code for our algorithms is accessible online at https://github.com/mmasden/canonicalpoly.
2022-07-15
{'id': '2207.07696v1', 'title': 'Algorithmic Determination of the Combinatorial Structure of the Linear Regions of ReLU Neural Networks', 'abstract': 'We algorithmically determine the regions and facets of all dimensions of the\ncanonical polyhedral complex, the universal object into which a ReLU network\ndecomposes its input space. We show that the locations of the vertices of the\ncanonical polyhedral complex along with their signs with respect to layer maps\ndetermine the full facet structure across all dimensions. We present an\nalgorithm which calculates this full combinatorial structure, making use of our\ntheorems that the dual complex to the canonical polyhedral complex is cubical\nand it possesses a multiplication compatible with its facet structure. The\nresulting algorithm is numerically stable, polynomial time in the number of\nintermediate neurons, and obtains accurate information across all dimensions.\nThis permits us to obtain, for example, the true topology of the decision\nboundaries of networks with low-dimensional inputs. We run empirics on such\nnetworks at initialization, finding that width alone does not increase observed\ntopology, but width in the presence of depth does. Source code for our\nalgorithms is accessible online at https://github.com/mmasden/canonicalpoly.', 'authors': ['Marissa Masden'], 'published': '2022-07-15 18:36:12+00:00', 'categories': ['cs.LG', 'cs.CG', 'math.AT'], 'url': 'http://arxiv.org/abs/2207.07696v1'}
42,332
2205.14309v2
2,205.14309
2
Federated Neural Bandits
Recent works on neural contextual bandits have achieved compelling performances due to their ability to leverage the strong representation power of neural networks (NNs) for reward prediction. Many applications of contextual bandits involve multiple agents who collaborate without sharing raw observations, thus giving rise to the setting of federated contextual bandits. Existing works on federated contextual bandits rely on linear or kernelized bandits, which may fall short when modeling complex real-world reward functions. So, this paper introduces the federated neural-upper confidence bound (FN-UCB) algorithm. To better exploit the federated setting, FN-UCB adopts a weighted combination of two UCBs: $UCB^a$ allows every agent to additionally use the observations from the other agents to accelerate exploration (without sharing raw observations), while $UCB^b$ uses an NN with aggregated parameters for reward prediction in a similar way to federated averaging for supervised learning. Notably, the weight between the two UCBs required by our theoretical analysis is amenable to an interesting interpretation, which emphasizes $UCB^a$ initially for accelerated exploration and relies more on $UCB^b$ later after enough observations have been collected to train the NNs for accurate reward prediction (i.e., reliable exploitation). We prove sub-linear upper bounds on both the cumulative regret and the number of communication rounds of FN-UCB, and empirically demonstrate its competitive performance.
2022-05-28
{'id': '2205.14309v2', 'title': 'Federated Neural Bandits', 'abstract': 'Recent works on neural contextual bandits have achieved compelling\nperformances due to their ability to leverage the strong representation power\nof neural networks (NNs) for reward prediction. Many applications of contextual\nbandits involve multiple agents who collaborate without sharing raw\nobservations, thus giving rise to the setting of federated contextual bandits.\nExisting works on federated contextual bandits rely on linear or kernelized\nbandits, which may fall short when modeling complex real-world reward\nfunctions. So, this paper introduces the federated neural-upper confidence\nbound (FN-UCB) algorithm. To better exploit the federated setting, FN-UCB\nadopts a weighted combination of two UCBs: $\\text{UCB}^{a}$ allows every agent\nto additionally use the observations from the other agents to accelerate\nexploration (without sharing raw observations), while $\\text{UCB}^{b}$ uses an\nNN with aggregated parameters for reward prediction in a similar way to\nfederated averaging for supervised learning. Notably, the weight between the\ntwo UCBs required by our theoretical analysis is amenable to an interesting\ninterpretation, which emphasizes $\\text{UCB}^{a}$ initially for accelerated\nexploration and relies more on $\\text{UCB}^{b}$ later after enough observations\nhave been collected to train the NNs for accurate reward prediction (i.e.,\nreliable exploitation). We prove sub-linear upper bounds on both the cumulative\nregret and the number of communication rounds of FN-UCB, and empirically\ndemonstrate its competitive performance.', 'authors': ['Zhongxiang Dai', 'Yao Shu', 'Arun Verma', 'Flint Xiaofeng Fan', 'Bryan Kian Hsiang Low', 'Patrick Jaillet'], 'published': '2022-05-28 02:58:37+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2205.14309v2'}
42,335
2302.12553v1
2,302.12553
1
Lower Bounds on the Depth of Integral ReLU Neural Networks via Lattice Polytopes
We prove that the set of functions representable by ReLU neural networks with integer weights strictly increases with the network depth while allowing arbitrary width. More precisely, we show that $_2(n)$ hidden layers are indeed necessary to compute the maximum of $n$ numbers, matching known upper bounds. Our results are based on the known duality between neural networks and Newton polytopes via tropical geometry. The integrality assumption implies that these Newton polytopes are lattice polytopes. Then, our depth lower bounds follow from a parity argument on the normalized volume of faces of such polytopes.
2023-02-24
{'id': '2302.12553v1', 'title': 'Lower Bounds on the Depth of Integral ReLU Neural Networks via Lattice Polytopes', 'abstract': 'We prove that the set of functions representable by ReLU neural networks with\ninteger weights strictly increases with the network depth while allowing\narbitrary width. More precisely, we show that $\\lceil\\log_2(n)\\rceil$ hidden\nlayers are indeed necessary to compute the maximum of $n$ numbers, matching\nknown upper bounds. Our results are based on the known duality between neural\nnetworks and Newton polytopes via tropical geometry. The integrality assumption\nimplies that these Newton polytopes are lattice polytopes. Then, our depth\nlower bounds follow from a parity argument on the normalized volume of faces of\nsuch polytopes.', 'authors': ['Christian Haase', 'Christoph Hertrich', 'Georg Loho'], 'published': '2023-02-24 10:14:53+00:00', 'categories': ['cs.LG', 'cs.DM', 'cs.NE', 'math.CO', 'stat.ML'], 'url': 'http://arxiv.org/abs/2302.12553v1'}
42,336
2310.11759v2
2,310.11759
2
Perceptual Scales Predicted by Fisher Information Metrics
Perception is often viewed as a process that transforms physical variables, external to an observer, into internal psychological variables. Such a process can be modeled by a function coined perceptual scale. The perceptual scale can be deduced from psychophysical measurements that consist in comparing the relative differences between stimuli (i.e. difference scaling experiments). However, this approach is often overlooked by the modeling and experimentation communities. Here, we demonstrate the value of measuring the perceptual scale of classical (spatial frequency, orientation) and less classical physical variables (interpolation between textures) by embedding it in recent probabilistic modeling of perception. First, we show that the assumption that an observer has an internal representation of univariate parameters such as spatial frequency or orientation while stimuli are high-dimensional does not lead to contradictory predictions when following the theoretical framework. Second, we show that the measured perceptual scale corresponds to the transduction function hypothesized in this framework. In particular, we demonstrate that it is related to the Fisher information of the generative model that underlies perception and we test the predictions given by the generative model of different stimuli in a set a of difference scaling experiments. Our main conclusion is that the perceptual scale is mostly driven by the stimulus power spectrum. Finally, we propose that this measure of perceptual scale is a way to push further the notion of perceptual distances by estimating the perceptual geometry of images i.e. the path between images instead of simply the distance between those.
2023-10-18
{'id': '2310.11759v2', 'title': 'Perceptual Scales Predicted by Fisher Information Metrics', 'abstract': 'Perception is often viewed as a process that transforms physical variables,\nexternal to an observer, into internal psychological variables. Such a process\ncan be modeled by a function coined perceptual scale. The perceptual scale can\nbe deduced from psychophysical measurements that consist in comparing the\nrelative differences between stimuli (i.e. difference scaling experiments).\nHowever, this approach is often overlooked by the modeling and experimentation\ncommunities. Here, we demonstrate the value of measuring the perceptual scale\nof classical (spatial frequency, orientation) and less classical physical\nvariables (interpolation between textures) by embedding it in recent\nprobabilistic modeling of perception. First, we show that the assumption that\nan observer has an internal representation of univariate parameters such as\nspatial frequency or orientation while stimuli are high-dimensional does not\nlead to contradictory predictions when following the theoretical framework.\nSecond, we show that the measured perceptual scale corresponds to the\ntransduction function hypothesized in this framework. In particular, we\ndemonstrate that it is related to the Fisher information of the generative\nmodel that underlies perception and we test the predictions given by the\ngenerative model of different stimuli in a set a of difference scaling\nexperiments. Our main conclusion is that the perceptual scale is mostly driven\nby the stimulus power spectrum. Finally, we propose that this measure of\nperceptual scale is a way to push further the notion of perceptual distances by\nestimating the perceptual geometry of images i.e. the path between images\ninstead of simply the distance between those.', 'authors': ['Jonathan Vacher', 'Pascal Mamassian'], 'published': '2023-10-18 07:31:47+00:00', 'categories': ['q-bio.NC', 'cs.CV', 'cs.IT', 'math.IT'], 'url': 'http://arxiv.org/abs/2310.11759v2'}
42,317
2011.06258v2
2,011.06258
2
Toward Trainability of Quantum Neural Networks
Quantum Neural Networks (QNNs) have been recently proposed as generalizations of classical neural networks to achieve the quantum speed-up. Despite the potential to outperform classical models, serious bottlenecks exist for training QNNs; namely, QNNs with random structures have poor trainability due to the vanishing gradient with rate exponential to the input qubit number. The vanishing gradient could seriously influence the applications of large-size QNNs. In this work, we provide a viable solution with theoretical guarantees. Specifically, we prove that QNNs with tree tensor and step controlled architectures have gradients that vanish at most polynomially with the qubit number. We numerically demonstrate QNNs with tree tensor and step controlled structures for the application of binary classification. Simulations show faster convergent rates and better accuracy compared to QNNs with random structures.
2020-11-12
{'id': '2011.06258v2', 'title': 'Toward Trainability of Quantum Neural Networks', 'abstract': 'Quantum Neural Networks (QNNs) have been recently proposed as generalizations\nof classical neural networks to achieve the quantum speed-up. Despite the\npotential to outperform classical models, serious bottlenecks exist for\ntraining QNNs; namely, QNNs with random structures have poor trainability due\nto the vanishing gradient with rate exponential to the input qubit number. The\nvanishing gradient could seriously influence the applications of large-size\nQNNs. In this work, we provide a viable solution with theoretical guarantees.\nSpecifically, we prove that QNNs with tree tensor and step controlled\narchitectures have gradients that vanish at most polynomially with the qubit\nnumber. We numerically demonstrate QNNs with tree tensor and step controlled\nstructures for the application of binary classification. Simulations show\nfaster convergent rates and better accuracy compared to QNNs with random\nstructures.', 'authors': ['Kaining Zhang', 'Min-Hsiu Hsieh', 'Liu Liu', 'Dacheng Tao'], 'published': '2020-11-12 08:32:04+00:00', 'categories': ['quant-ph'], 'url': 'http://arxiv.org/abs/2011.06258v2'}
42,319
1912.04871v4
1,912.04871
4
Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients
Discovering the underlying mathematical expressions describing a dataset is a core challenge for artificial intelligence. This is the problem of $symbolic regression$. Despite recent advances in training neural networks to solve complex tasks, deep learning approaches to symbolic regression are underexplored. We propose a framework that leverages deep learning for symbolic regression via a simple idea: use a large model to search the space of small models. Specifically, we use a recurrent neural network to emit a distribution over tractable mathematical expressions and employ a novel risk-seeking policy gradient to train the network to generate better-fitting expressions. Our algorithm outperforms several baseline methods (including Eureqa, the gold standard for symbolic regression) in its ability to exactly recover symbolic expressions on a series of benchmark problems, both with and without added noise. More broadly, our contributions include a framework that can be applied to optimize hierarchical, variable-length objects under a black-box performance metric, with the ability to incorporate constraints in situ, and a risk-seeking policy gradient formulation that optimizes for best-case performance instead of expected performance.
2019-12-10
{'id': '1912.04871v4', 'title': 'Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients', 'abstract': 'Discovering the underlying mathematical expressions describing a dataset is a\ncore challenge for artificial intelligence. This is the problem of\n$\\textit{symbolic regression}$. Despite recent advances in training neural\nnetworks to solve complex tasks, deep learning approaches to symbolic\nregression are underexplored. We propose a framework that leverages deep\nlearning for symbolic regression via a simple idea: use a large model to search\nthe space of small models. Specifically, we use a recurrent neural network to\nemit a distribution over tractable mathematical expressions and employ a novel\nrisk-seeking policy gradient to train the network to generate better-fitting\nexpressions. Our algorithm outperforms several baseline methods (including\nEureqa, the gold standard for symbolic regression) in its ability to exactly\nrecover symbolic expressions on a series of benchmark problems, both with and\nwithout added noise. More broadly, our contributions include a framework that\ncan be applied to optimize hierarchical, variable-length objects under a\nblack-box performance metric, with the ability to incorporate constraints in\nsitu, and a risk-seeking policy gradient formulation that optimizes for\nbest-case performance instead of expected performance.', 'authors': ['Brenden K. Petersen', 'Mikel Landajuela', 'T. Nathan Mundhenk', 'Claudio P. Santiago', 'Soo K. Kim', 'Joanne T. Kim'], 'published': '2019-12-10 18:25:48+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/1912.04871v4'}
42,321
2104.09732v1
2,104.09732
1
Knowledge Distillation as Semiparametric Inference
A popular approach to model compression is to train an inexpensive student model to mimic the class probabilities of a highly accurate but cumbersome teacher model. Surprisingly, this two-step knowledge distillation process often leads to higher accuracy than training the student directly on labeled data. To explain and enhance this phenomenon, we cast knowledge distillation as a semiparametric inference problem with the optimal student model as the target, the unknown Bayes class probabilities as nuisance, and the teacher probabilities as a plug-in nuisance estimate. By adapting modern semiparametric tools, we derive new guarantees for the prediction error of standard distillation and develop two enhancements -- cross-fitting and loss correction -- to mitigate the impact of teacher overfitting and underfitting on student performance. We validate our findings empirically on both tabular and image data and observe consistent improvements from our knowledge distillation enhancements.
2021-04-20
{'id': '2104.09732v1', 'title': 'Knowledge Distillation as Semiparametric Inference', 'abstract': 'A popular approach to model compression is to train an inexpensive student\nmodel to mimic the class probabilities of a highly accurate but cumbersome\nteacher model. Surprisingly, this two-step knowledge distillation process often\nleads to higher accuracy than training the student directly on labeled data. To\nexplain and enhance this phenomenon, we cast knowledge distillation as a\nsemiparametric inference problem with the optimal student model as the target,\nthe unknown Bayes class probabilities as nuisance, and the teacher\nprobabilities as a plug-in nuisance estimate. By adapting modern semiparametric\ntools, we derive new guarantees for the prediction error of standard\ndistillation and develop two enhancements -- cross-fitting and loss correction\n-- to mitigate the impact of teacher overfitting and underfitting on student\nperformance. We validate our findings empirically on both tabular and image\ndata and observe consistent improvements from our knowledge distillation\nenhancements.', 'authors': ['Tri Dao', 'Govinda M Kamath', 'Vasilis Syrgkanis', 'Lester Mackey'], 'published': '2021-04-20 03:00:45+00:00', 'categories': ['stat.ML', 'cs.LG'], 'url': 'http://arxiv.org/abs/2104.09732v1'}
42,322
2406.12769v1
2,406.12769
1
Latent Intuitive Physics: Learning to Transfer Hidden Physics from A 3D Video
We introduce latent intuitive physics, a transfer learning framework for physics simulation that can infer hidden properties of fluids from a single 3D video and simulate the observed fluid in novel scenes. Our key insight is to use latent features drawn from a learnable prior distribution conditioned on the underlying particle states to capture the invisible and complex physical properties. To achieve this, we train a parametrized prior learner given visual observations to approximate the visual posterior of inverse graphics, and both the particle states and the visual posterior are obtained from a learned neural renderer. The converged prior learner is embedded in our probabilistic physics engine, allowing us to perform novel simulations on unseen geometries, boundaries, and dynamics without knowledge of the true physical parameters. We validate our model in three ways: (i) novel scene simulation with the learned visual-world physics, (ii) future prediction of the observed fluid dynamics, and (iii) supervised particle simulation. Our model demonstrates strong performance in all three tasks.
2024-06-18
{'id': '2406.12769v1', 'title': 'Latent Intuitive Physics: Learning to Transfer Hidden Physics from A 3D Video', 'abstract': 'We introduce latent intuitive physics, a transfer learning framework for\nphysics simulation that can infer hidden properties of fluids from a single 3D\nvideo and simulate the observed fluid in novel scenes. Our key insight is to\nuse latent features drawn from a learnable prior distribution conditioned on\nthe underlying particle states to capture the invisible and complex physical\nproperties. To achieve this, we train a parametrized prior learner given visual\nobservations to approximate the visual posterior of inverse graphics, and both\nthe particle states and the visual posterior are obtained from a learned neural\nrenderer. The converged prior learner is embedded in our probabilistic physics\nengine, allowing us to perform novel simulations on unseen geometries,\nboundaries, and dynamics without knowledge of the true physical parameters. We\nvalidate our model in three ways: (i) novel scene simulation with the learned\nvisual-world physics, (ii) future prediction of the observed fluid dynamics,\nand (iii) supervised particle simulation. Our model demonstrates strong\nperformance in all three tasks.', 'authors': ['Xiangming Zhu', 'Huayu Deng', 'Haochen Yuan', 'Yunbo Wang', 'Xiaokang Yang'], 'published': '2024-06-18 16:37:44+00:00', 'categories': ['cs.AI', 'cs.CV'], 'url': 'http://arxiv.org/abs/2406.12769v1'}
42,324
2309.14525v1
2,309.14525
1
Aligning Large Multimodal Models with Factually Augmented RLHF
Large Multimodal Models (LMM) are built across modalities and the misalignment between two modalities can result in "hallucination", generating textual outputs that are not grounded by the multimodal information in context. To address the multimodal misalignment issue, we adapt the Reinforcement Learning from Human Feedback (RLHF) from the text domain to the task of vision-language alignment, where human annotators are asked to compare two responses and pinpoint the more hallucinated one, and the vision-language model is trained to maximize the simulated human rewards. We propose a new alignment algorithm called Factually Augmented RLHF that augments the reward model with additional factual information such as image captions and ground-truth multi-choice options, which alleviates the reward hacking phenomenon in RLHF and further improves the performance. We also enhance the GPT-4-generated training data (for vision instruction tuning) with previously available human-written image-text pairs to improve the general capabilities of our model. To evaluate the proposed approach in real-world scenarios, we develop a new evaluation benchmark MMHAL-BENCH with a special focus on penalizing hallucinations. As the first LMM trained with RLHF, our approach achieves remarkable improvement on the LLaVA-Bench dataset with the 94% performance level of the text-only GPT-4 (while previous best methods can only achieve the 87% level), and an improvement by 60% on MMHAL-BENCH over other baselines. We opensource our code, model, data at https://llava-rlhf.github.io.
2023-09-25
{'id': '2309.14525v1', 'title': 'Aligning Large Multimodal Models with Factually Augmented RLHF', 'abstract': 'Large Multimodal Models (LMM) are built across modalities and the\nmisalignment between two modalities can result in "hallucination", generating\ntextual outputs that are not grounded by the multimodal information in context.\nTo address the multimodal misalignment issue, we adapt the Reinforcement\nLearning from Human Feedback (RLHF) from the text domain to the task of\nvision-language alignment, where human annotators are asked to compare two\nresponses and pinpoint the more hallucinated one, and the vision-language model\nis trained to maximize the simulated human rewards. We propose a new alignment\nalgorithm called Factually Augmented RLHF that augments the reward model with\nadditional factual information such as image captions and ground-truth\nmulti-choice options, which alleviates the reward hacking phenomenon in RLHF\nand further improves the performance. We also enhance the GPT-4-generated\ntraining data (for vision instruction tuning) with previously available\nhuman-written image-text pairs to improve the general capabilities of our\nmodel. To evaluate the proposed approach in real-world scenarios, we develop a\nnew evaluation benchmark MMHAL-BENCH with a special focus on penalizing\nhallucinations. As the first LMM trained with RLHF, our approach achieves\nremarkable improvement on the LLaVA-Bench dataset with the 94% performance\nlevel of the text-only GPT-4 (while previous best methods can only achieve the\n87% level), and an improvement by 60% on MMHAL-BENCH over other baselines. We\nopensource our code, model, data at https://llava-rlhf.github.io.', 'authors': ['Zhiqing Sun', 'Sheng Shen', 'Shengcao Cao', 'Haotian Liu', 'Chunyuan Li', 'Yikang Shen', 'Chuang Gan', 'Liang-Yan Gui', 'Yu-Xiong Wang', 'Yiming Yang', 'Kurt Keutzer', 'Trevor Darrell'], 'published': '2023-09-25 20:59:33+00:00', 'categories': ['cs.CV', 'cs.CL'], 'url': 'http://arxiv.org/abs/2309.14525v1'}
42,331
2012.06718v1
2,012.06718
1
Learning Consistent Deep Generative Models from Sparse Data via Prediction Constraints
We develop a new framework for learning variational autoencoders and other deep generative models that balances generative and discriminative goals. Our framework optimizes model parameters to maximize a variational lower bound on the likelihood of observed data, subject to a task-specific prediction constraint that prevents model misspecification from leading to inaccurate predictions. We further enforce a consistency constraint, derived naturally from the generative model, that requires predictions on reconstructed data to match those on the original data. We show that these two contributions -- prediction constraints and consistency constraints -- lead to promising image classification performance, especially in the semi-supervised scenario where category labels are sparse but unlabeled data is plentiful. Our approach enables advances in generative modeling to directly boost semi-supervised classification performance, an ability we demonstrate by augmenting deep generative models with latent variables capturing spatial transformations.
2020-12-12
{'id': '2012.06718v1', 'title': 'Learning Consistent Deep Generative Models from Sparse Data via Prediction Constraints', 'abstract': 'We develop a new framework for learning variational autoencoders and other\ndeep generative models that balances generative and discriminative goals. Our\nframework optimizes model parameters to maximize a variational lower bound on\nthe likelihood of observed data, subject to a task-specific prediction\nconstraint that prevents model misspecification from leading to inaccurate\npredictions. We further enforce a consistency constraint, derived naturally\nfrom the generative model, that requires predictions on reconstructed data to\nmatch those on the original data. We show that these two contributions --\nprediction constraints and consistency constraints -- lead to promising image\nclassification performance, especially in the semi-supervised scenario where\ncategory labels are sparse but unlabeled data is plentiful. Our approach\nenables advances in generative modeling to directly boost semi-supervised\nclassification performance, an ability we demonstrate by augmenting deep\ngenerative models with latent variables capturing spatial transformations.', 'authors': ['Gabriel Hope', 'Madina Abdrakhmanova', 'Xiaoyin Chen', 'Michael C. Hughes', 'Michael C. Hughes', 'Erik B. Sudderth'], 'published': '2020-12-12 04:18:50+00:00', 'categories': ['cs.LG', 'cs.CV', 'stat.ML'], 'url': 'http://arxiv.org/abs/2012.06718v1'}
42,325
2410.14069v1
2,410.14069
1
Rethinking Optimal Transport in Offline Reinforcement Learning
We propose a novel algorithm for offline reinforcement learning using optimal transport. Typically, in offline reinforcement learning, the data is provided by various experts and some of them can be sub-optimal. To extract an efficient policy, it is necessary to stitch the best behaviors from the dataset. To address this problem, we rethink offline reinforcement learning as an optimal transportation problem. And based on this, we present an algorithm that aims to find a policy that maps states to a partial distribution of the best expert actions for each given state. We evaluate the performance of our algorithm on continuous control problems from the D4RL suite and demonstrate improvements over existing methods.
2024-10-17
{'id': '2410.14069v1', 'title': 'Rethinking Optimal Transport in Offline Reinforcement Learning', 'abstract': 'We propose a novel algorithm for offline reinforcement learning using optimal\ntransport. Typically, in offline reinforcement learning, the data is provided\nby various experts and some of them can be sub-optimal. To extract an efficient\npolicy, it is necessary to \\emph{stitch} the best behaviors from the dataset.\nTo address this problem, we rethink offline reinforcement learning as an\noptimal transportation problem. And based on this, we present an algorithm that\naims to find a policy that maps states to a \\emph{partial} distribution of the\nbest expert actions for each given state. We evaluate the performance of our\nalgorithm on continuous control problems from the D4RL suite and demonstrate\nimprovements over existing methods.', 'authors': ['Arip Asadulaev', 'Rostislav Korst', 'Alexander Korotin', 'Vage Egiazarian', 'Andrey Filchenkov', 'Evgeny Burnaev'], 'published': '2024-10-17 22:36:43+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2410.14069v1'}
42,326
2312.11460v3
2,312.1146
3
Hybrid Internal Model: Learning Agile Legged Locomotion with Simulated Robot Response
Robust locomotion control depends on accurate state estimations. However, the sensors of most legged robots can only provide partial and noisy observations, making the estimation particularly challenging, especially for external states like terrain frictions and elevation maps. Inspired by the classical Internal Model Control principle, we consider these external states as disturbances and introduce Hybrid Internal Model (HIM) to estimate them according to the response of the robot. The response, which we refer to as the hybrid internal embedding, contains the robot's explicit velocity and implicit stability representation, corresponding to two primary goals for locomotion tasks: explicitly tracking velocity and implicitly maintaining stability. We use contrastive learning to optimize the embedding to be close to the robot's successor state, in which the response is naturally embedded. HIM has several appealing benefits: It only needs the robot's proprioceptions, i.e., those from joint encoders and IMU as observations. It innovatively maintains consistent observations between simulation reference and reality that avoids information loss in mimicking learning. It exploits batch-level information that is more robust to noises and keeps better sample efficiency. It only requires 1 hour of training on an RTX 4090 to enable a quadruped robot to traverse any terrain under any disturbances. A wealth of real-world experiments demonstrates its agility, even in high-difficulty tasks and cases never occurred during the training process, revealing remarkable open-world generalizability.
2023-12-18
{'id': '2312.11460v3', 'title': 'Hybrid Internal Model: Learning Agile Legged Locomotion with Simulated Robot Response', 'abstract': "Robust locomotion control depends on accurate state estimations. However, the\nsensors of most legged robots can only provide partial and noisy observations,\nmaking the estimation particularly challenging, especially for external states\nlike terrain frictions and elevation maps. Inspired by the classical Internal\nModel Control principle, we consider these external states as disturbances and\nintroduce Hybrid Internal Model (HIM) to estimate them according to the\nresponse of the robot. The response, which we refer to as the hybrid internal\nembedding, contains the robot's explicit velocity and implicit stability\nrepresentation, corresponding to two primary goals for locomotion tasks:\nexplicitly tracking velocity and implicitly maintaining stability. We use\ncontrastive learning to optimize the embedding to be close to the robot's\nsuccessor state, in which the response is naturally embedded. HIM has several\nappealing benefits: It only needs the robot's proprioceptions, i.e., those from\njoint encoders and IMU as observations. It innovatively maintains consistent\nobservations between simulation reference and reality that avoids information\nloss in mimicking learning. It exploits batch-level information that is more\nrobust to noises and keeps better sample efficiency. It only requires 1 hour of\ntraining on an RTX 4090 to enable a quadruped robot to traverse any terrain\nunder any disturbances. A wealth of real-world experiments demonstrates its\nagility, even in high-difficulty tasks and cases never occurred during the\ntraining process, revealing remarkable open-world generalizability.", 'authors': ['Junfeng Long', 'Zirui Wang', 'Quanyi Li', 'Jiawei Gao', 'Liu Cao', 'Jiangmiao Pang'], 'published': '2023-12-18 18:59:06+00:00', 'categories': ['cs.RO', 'cs.AI', 'cs.CV', 'cs.LG', 'cs.SY', 'eess.SY'], 'url': 'http://arxiv.org/abs/2312.11460v3'}
42,328
2311.04465v2
2,311.04465
2
Solving High Frequency and Multi-Scale PDEs with Gaussian Processes
Machine learning based solvers have garnered much attention in physical simulation and scientific computing, with a prominent example, physics-informed neural networks (PINNs). However, PINNs often struggle to solve high-frequency and multi-scale PDEs, which can be due to spectral bias during neural network training. To address this problem, we resort to the Gaussian process (GP) framework. To flexibly capture the dominant frequencies, we model the power spectrum of the PDE solution with a student $t$ mixture or Gaussian mixture. We apply the inverse Fourier transform to obtain the covariance function (by Wiener-Khinchin theorem). The covariance derived from the Gaussian mixture spectrum corresponds to the known spectral mixture kernel. Next, we estimate the mixture weights in the log domain, which we show is equivalent to placing a Jeffreys prior. It automatically induces sparsity, prunes excessive frequencies, and adjusts the remaining toward the ground truth. Third, to enable efficient and scalable computation on massive collocation points, which are critical to capture high frequencies, we place the collocation points on a grid, and multiply our covariance function at each input dimension. We use the GP conditional mean to predict the solution and its derivatives so as to fit the boundary condition and the equation itself. As a result, we can derive a Kronecker product structure in the covariance matrix. We use Kronecker product properties and multilinear algebra to promote computational efficiency and scalability, without low-rank approximations. We show the advantage of our method in systematic experiments. The code is released at https://github.com/xuangu-fang/Gaussian-Process-Slover-for-High-Freq-PDE.
2023-11-08
{'id': '2311.04465v2', 'title': 'Solving High Frequency and Multi-Scale PDEs with Gaussian Processes', 'abstract': 'Machine learning based solvers have garnered much attention in physical\nsimulation and scientific computing, with a prominent example, physics-informed\nneural networks (PINNs). However, PINNs often struggle to solve high-frequency\nand multi-scale PDEs, which can be due to spectral bias during neural network\ntraining. To address this problem, we resort to the Gaussian process (GP)\nframework. To flexibly capture the dominant frequencies, we model the power\nspectrum of the PDE solution with a student $t$ mixture or Gaussian mixture. We\napply the inverse Fourier transform to obtain the covariance function (by\nWiener-Khinchin theorem). The covariance derived from the Gaussian mixture\nspectrum corresponds to the known spectral mixture kernel. Next, we estimate\nthe mixture weights in the log domain, which we show is equivalent to placing a\nJeffreys prior. It automatically induces sparsity, prunes excessive\nfrequencies, and adjusts the remaining toward the ground truth. Third, to\nenable efficient and scalable computation on massive collocation points, which\nare critical to capture high frequencies, we place the collocation points on a\ngrid, and multiply our covariance function at each input dimension. We use the\nGP conditional mean to predict the solution and its derivatives so as to fit\nthe boundary condition and the equation itself. As a result, we can derive a\nKronecker product structure in the covariance matrix. We use Kronecker product\nproperties and multilinear algebra to promote computational efficiency and\nscalability, without low-rank approximations. We show the advantage of our\nmethod in systematic experiments. The code is released at\n\\url{https://github.com/xuangu-fang/Gaussian-Process-Slover-for-High-Freq-PDE}.', 'authors': ['Shikai Fang', 'Madison Cooley', 'Da Long', 'Shibo Li', 'Robert Kirby', 'Shandian Zhe'], 'published': '2023-11-08 05:26:58+00:00', 'categories': ['cs.LG', 'cs.CE'], 'url': 'http://arxiv.org/abs/2311.04465v2'}
42,344
2312.05583v2
2,312.05583
2
Better Neural PDE Solvers Through Data-Free Mesh Movers
Recently, neural networks have been extensively employed to solve partial differential equations (PDEs) in physical system modeling. While major studies focus on learning system evolution on predefined static mesh discretizations, some methods utilize reinforcement learning or supervised learning techniques to create adaptive and dynamic meshes, due to the dynamic nature of these systems. However, these approaches face two primary challenges: (1) the need for expensive optimal mesh data, and (2) the change of the solution space's degree of freedom and topology during mesh refinement. To address these challenges, this paper proposes a neural PDE solver with a neural mesh adapter. To begin with, we introduce a novel data-free neural mesh adaptor, called Data-free Mesh Mover (DMM), with two main innovations. Firstly, it is an operator that maps the solution to adaptive meshes and is trained using the Monge-Amp\`ere equation without optimal mesh data. Secondly, it dynamically changes the mesh by moving existing nodes rather than adding or deleting nodes and edges. Theoretical analysis shows that meshes generated by DMM have the lowest interpolation error bound. Based on DMM, to efficiently and accurately model dynamic systems, we develop a moving mesh based neural PDE solver (MM-PDE) that embeds the moving mesh with a two-branch architecture and a learnable interpolation framework to preserve information within the data. Empirical experiments demonstrate that our method generates suitable meshes and considerably enhances accuracy when modeling widely considered PDE systems. The code can be found at: https://github.com/Peiyannn/MM-PDE.git.
2023-12-09
{'id': '2312.05583v2', 'title': 'Better Neural PDE Solvers Through Data-Free Mesh Movers', 'abstract': "Recently, neural networks have been extensively employed to solve partial\ndifferential equations (PDEs) in physical system modeling. While major studies\nfocus on learning system evolution on predefined static mesh discretizations,\nsome methods utilize reinforcement learning or supervised learning techniques\nto create adaptive and dynamic meshes, due to the dynamic nature of these\nsystems. However, these approaches face two primary challenges: (1) the need\nfor expensive optimal mesh data, and (2) the change of the solution space's\ndegree of freedom and topology during mesh refinement. To address these\nchallenges, this paper proposes a neural PDE solver with a neural mesh adapter.\nTo begin with, we introduce a novel data-free neural mesh adaptor, called\nData-free Mesh Mover (DMM), with two main innovations. Firstly, it is an\noperator that maps the solution to adaptive meshes and is trained using the\nMonge-Amp\\`ere equation without optimal mesh data. Secondly, it dynamically\nchanges the mesh by moving existing nodes rather than adding or deleting nodes\nand edges. Theoretical analysis shows that meshes generated by DMM have the\nlowest interpolation error bound. Based on DMM, to efficiently and accurately\nmodel dynamic systems, we develop a moving mesh based neural PDE solver\n(MM-PDE) that embeds the moving mesh with a two-branch architecture and a\nlearnable interpolation framework to preserve information within the data.\nEmpirical experiments demonstrate that our method generates suitable meshes and\nconsiderably enhances accuracy when modeling widely considered PDE systems. The\ncode can be found at: https://github.com/Peiyannn/MM-PDE.git.", 'authors': ['Peiyan Hu', 'Yue Wang', 'Zhi-Ming Ma'], 'published': '2023-12-09 14:05:28+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.NA', 'math.NA'], 'url': 'http://arxiv.org/abs/2312.05583v2'}
42,327
2305.16948v1
2,305.16948
1
Meta-prediction Model for Distillation-Aware NAS on Unseen Datasets
Distillation-aware Neural Architecture Search (DaNAS) aims to search for an optimal student architecture that obtains the best performance and/or efficiency when distilling the knowledge from a given teacher model. Previous DaNAS methods have mostly tackled the search for the neural architecture for fixed datasets and the teacher, which are not generalized well on a new task consisting of an unseen dataset and an unseen teacher, thus need to perform a costly search for any new combination of the datasets and the teachers. For standard NAS tasks without KD, meta-learning-based computationally efficient NAS methods have been proposed, which learn the generalized search process over multiple tasks (datasets) and transfer the knowledge obtained over those tasks to a new task. However, since they assume learning from scratch without KD from a teacher, they might not be ideal for DaNAS scenarios. To eliminate the excessive computational cost of DaNAS methods and the sub-optimality of rapid NAS methods, we propose a distillation-aware meta accuracy prediction model, DaSS (Distillation-aware Student Search), which can predict a given architecture's final performances on a dataset when performing KD with a given teacher, without having actually to train it on the target task. The experimental results demonstrate that our proposed meta-prediction model successfully generalizes to multiple unseen datasets for DaNAS tasks, largely outperforming existing meta-NAS methods and rapid NAS baselines. Code is available at https://github.com/CownowAn/DaSS
2023-05-26
{'id': '2305.16948v1', 'title': 'Meta-prediction Model for Distillation-Aware NAS on Unseen Datasets', 'abstract': "Distillation-aware Neural Architecture Search (DaNAS) aims to search for an\noptimal student architecture that obtains the best performance and/or\nefficiency when distilling the knowledge from a given teacher model. Previous\nDaNAS methods have mostly tackled the search for the neural architecture for\nfixed datasets and the teacher, which are not generalized well on a new task\nconsisting of an unseen dataset and an unseen teacher, thus need to perform a\ncostly search for any new combination of the datasets and the teachers. For\nstandard NAS tasks without KD, meta-learning-based computationally efficient\nNAS methods have been proposed, which learn the generalized search process over\nmultiple tasks (datasets) and transfer the knowledge obtained over those tasks\nto a new task. However, since they assume learning from scratch without KD from\na teacher, they might not be ideal for DaNAS scenarios. To eliminate the\nexcessive computational cost of DaNAS methods and the sub-optimality of rapid\nNAS methods, we propose a distillation-aware meta accuracy prediction model,\nDaSS (Distillation-aware Student Search), which can predict a given\narchitecture's final performances on a dataset when performing KD with a given\nteacher, without having actually to train it on the target task. The\nexperimental results demonstrate that our proposed meta-prediction model\nsuccessfully generalizes to multiple unseen datasets for DaNAS tasks, largely\noutperforming existing meta-NAS methods and rapid NAS baselines. Code is\navailable at https://github.com/CownowAn/DaSS", 'authors': ['Hayeon Lee', 'Sohyun An', 'Minseon Kim', 'Sung Ju Hwang'], 'published': '2023-05-26 14:00:35+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2305.16948v1'}
42,329
2303.01158v1
2,303.01158
1
Iterative Circuit Repair Against Formal Specifications
We present a deep learning approach for repairing sequential circuits against formal specifications given in linear-time temporal logic (LTL). Given a defective circuit and its formal specification, we train Transformer models to output circuits that satisfy the corresponding specification. We propose a separated hierarchical Transformer for multimodal representation learning of the formal specification and the circuit. We introduce a data generation algorithm that enables generalization to more complex specifications and out-of-distribution datasets. In addition, our proposed repair mechanism significantly improves the automated synthesis of circuits from LTL specifications with Transformers. It improves the state-of-the-art by $6.8$ percentage points on held-out instances and $11.8$ percentage points on an out-of-distribution dataset from the annual reactive synthesis competition.
2023-03-02
{'id': '2303.01158v1', 'title': 'Iterative Circuit Repair Against Formal Specifications', 'abstract': 'We present a deep learning approach for repairing sequential circuits against\nformal specifications given in linear-time temporal logic (LTL). Given a\ndefective circuit and its formal specification, we train Transformer models to\noutput circuits that satisfy the corresponding specification. We propose a\nseparated hierarchical Transformer for multimodal representation learning of\nthe formal specification and the circuit. We introduce a data generation\nalgorithm that enables generalization to more complex specifications and\nout-of-distribution datasets. In addition, our proposed repair mechanism\nsignificantly improves the automated synthesis of circuits from LTL\nspecifications with Transformers. It improves the state-of-the-art by $6.8$\npercentage points on held-out instances and $11.8$ percentage points on an\nout-of-distribution dataset from the annual reactive synthesis competition.', 'authors': ['Matthias Cosler', 'Frederik Schmitt', 'Christopher Hahn', 'Bernd Finkbeiner'], 'published': '2023-03-02 11:05:10+00:00', 'categories': ['cs.LG', 'cs.LO'], 'url': 'http://arxiv.org/abs/2303.01158v1'}
42,337
2310.02533v1
2,310.02533
1
Quantifying and mitigating the impact of label errors on model disparity metrics
Errors in labels obtained via human annotation adversely affect a model's performance. Existing approaches propose ways to mitigate the effect of label error on a model's downstream accuracy, yet little is known about its impact on a model's disparity metrics. Here we study the effect of label error on a model's disparity metrics. We empirically characterize how varying levels of label error, in both training and test data, affect these disparity metrics. We find that group calibration and other metrics are sensitive to train-time and test-time label error -- particularly for minority groups. This disparate effect persists even for models trained with noise-aware algorithms. To mitigate the impact of training-time label error, we present an approach to estimate the influence of a training input's label on a model's group disparity metric. We empirically assess the proposed approach on a variety of datasets and find significant improvement, compared to alternative approaches, in identifying training inputs that improve a model's disparity metric. We complement the approach with an automatic relabel-and-finetune scheme that produces updated models with, provably, improved group calibration error.
2023-10-04
{'id': '2310.02533v1', 'title': 'Quantifying and mitigating the impact of label errors on model disparity metrics', 'abstract': "Errors in labels obtained via human annotation adversely affect a model's\nperformance. Existing approaches propose ways to mitigate the effect of label\nerror on a model's downstream accuracy, yet little is known about its impact on\na model's disparity metrics. Here we study the effect of label error on a\nmodel's disparity metrics. We empirically characterize how varying levels of\nlabel error, in both training and test data, affect these disparity metrics. We\nfind that group calibration and other metrics are sensitive to train-time and\ntest-time label error -- particularly for minority groups. This disparate\neffect persists even for models trained with noise-aware algorithms. To\nmitigate the impact of training-time label error, we present an approach to\nestimate the influence of a training input's label on a model's group disparity\nmetric. We empirically assess the proposed approach on a variety of datasets\nand find significant improvement, compared to alternative approaches, in\nidentifying training inputs that improve a model's disparity metric. We\ncomplement the approach with an automatic relabel-and-finetune scheme that\nproduces updated models with, provably, improved group calibration error.", 'authors': ['Julius Adebayo', 'Melissa Hall', 'Bowen Yu', 'Bobbie Chern'], 'published': '2023-10-04 02:18:45+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2310.02533v1'}
42,343
2204.06601v4
2,204.06601
4
Causal Confusion and Reward Misidentification in Preference-Based Reward Learning
Learning policies via preference-based reward learning is an increasingly popular method for customizing agent behavior, but has been shown anecdotally to be prone to spurious correlations and reward hacking behaviors. While much prior work focuses on causal confusion in reinforcement learning and behavioral cloning, we focus on a systematic study of causal confusion and reward misidentification when learning from preferences. In particular, we perform a series of sensitivity and ablation analyses on several benchmark domains where rewards learned from preferences achieve minimal test error but fail to generalize to out-of-distribution states -- resulting in poor policy performance when optimized. We find that the presence of non-causal distractor features, noise in the stated preferences, and partial state observability can all exacerbate reward misidentification. We also identify a set of methods with which to interpret misidentified learned rewards. In general, we observe that optimizing misidentified rewards drives the policy off the reward's training distribution, resulting in high predicted (learned) rewards but low true rewards. These findings illuminate the susceptibility of preference learning to reward misidentification and causal confusion -- failure to consider even one of many factors can result in unexpected, undesirable behavior.
2022-04-13
{'id': '2204.06601v4', 'title': 'Causal Confusion and Reward Misidentification in Preference-Based Reward Learning', 'abstract': "Learning policies via preference-based reward learning is an increasingly\npopular method for customizing agent behavior, but has been shown anecdotally\nto be prone to spurious correlations and reward hacking behaviors. While much\nprior work focuses on causal confusion in reinforcement learning and behavioral\ncloning, we focus on a systematic study of causal confusion and reward\nmisidentification when learning from preferences. In particular, we perform a\nseries of sensitivity and ablation analyses on several benchmark domains where\nrewards learned from preferences achieve minimal test error but fail to\ngeneralize to out-of-distribution states -- resulting in poor policy\nperformance when optimized. We find that the presence of non-causal distractor\nfeatures, noise in the stated preferences, and partial state observability can\nall exacerbate reward misidentification. We also identify a set of methods with\nwhich to interpret misidentified learned rewards. In general, we observe that\noptimizing misidentified rewards drives the policy off the reward's training\ndistribution, resulting in high predicted (learned) rewards but low true\nrewards. These findings illuminate the susceptibility of preference learning to\nreward misidentification and causal confusion -- failure to consider even one\nof many factors can result in unexpected, undesirable behavior.", 'authors': ['Jeremy Tien', 'Jerry Zhi-Yang He', 'Zackory Erickson', 'Anca D. Dragan', 'Daniel S. Brown'], 'published': '2022-04-13 18:41:41+00:00', 'categories': ['cs.LG', 'cs.RO'], 'url': 'http://arxiv.org/abs/2204.06601v4'}
42,361
2206.01202v1
2,206.01202
1
Unveiling The Mask of Position-Information Pattern Through the Mist of Image Features
Recent studies show that paddings in convolutional neural networks encode absolute position information which can negatively affect the model performance for certain tasks. However, existing metrics for quantifying the strength of positional information remain unreliable and frequently lead to erroneous results. To address this issue, we propose novel metrics for measuring (and visualizing) the encoded positional information. We formally define the encoded information as PPP (Position-information Pattern from Padding) and conduct a series of experiments to study its properties as well as its formation. The proposed metrics measure the presence of positional information more reliably than the existing metrics based on PosENet and a test in F-Conv. We also demonstrate that for any extant (and proposed) padding schemes, PPP is primarily a learning artifact and is less dependent on the characteristics of the underlying padding schemes.
2022-06-02
{'id': '2206.01202v1', 'title': 'Unveiling The Mask of Position-Information Pattern Through the Mist of Image Features', 'abstract': 'Recent studies show that paddings in convolutional neural networks encode\nabsolute position information which can negatively affect the model performance\nfor certain tasks. However, existing metrics for quantifying the strength of\npositional information remain unreliable and frequently lead to erroneous\nresults. To address this issue, we propose novel metrics for measuring (and\nvisualizing) the encoded positional information. We formally define the encoded\ninformation as PPP (Position-information Pattern from Padding) and conduct a\nseries of experiments to study its properties as well as its formation. The\nproposed metrics measure the presence of positional information more reliably\nthan the existing metrics based on PosENet and a test in F-Conv. We also\ndemonstrate that for any extant (and proposed) padding schemes, PPP is\nprimarily a learning artifact and is less dependent on the characteristics of\nthe underlying padding schemes.', 'authors': ['Chieh Hubert Lin', 'Hsin-Ying Lee', 'Hung-Yu Tseng', 'Maneesh Singh', 'Ming-Hsuan Yang'], 'published': '2022-06-02 17:59:57+00:00', 'categories': ['cs.CV', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2206.01202v1'}
42,363
2302.11978v1
2,302.11978
1
Does Deep Learning Learn to Abstract? A Systematic Probing Framework
Abstraction is a desirable capability for deep learning models, which means to induce abstract concepts from concrete instances and flexibly apply them beyond the learning context. At the same time, there is a lack of clear understanding about both the presence and further characteristics of this capability in deep learning models. In this paper, we introduce a systematic probing framework to explore the abstraction capability of deep learning models from a transferability perspective. A set of controlled experiments are conducted based on this framework, providing strong evidence that two probed pre-trained language models (PLMs), T5 and GPT2, have the abstraction capability. We also conduct in-depth analysis, thus shedding further light: (1) the whole training phase exhibits a "memorize-then-abstract" two-stage process; (2) the learned abstract concepts are gathered in a few middle-layer attention heads, rather than being evenly distributed throughout the model; (3) the probed abstraction capabilities exhibit robustness against concept mutations, and are more robust to low-level/source-side mutations than high-level/target-side ones; (4) generic pre-training is critical to the emergence of abstraction capability, and PLMs exhibit better abstraction with larger model sizes and data scales.
2023-02-23
{'id': '2302.11978v1', 'title': 'Does Deep Learning Learn to Abstract? A Systematic Probing Framework', 'abstract': 'Abstraction is a desirable capability for deep learning models, which means\nto induce abstract concepts from concrete instances and flexibly apply them\nbeyond the learning context. At the same time, there is a lack of clear\nunderstanding about both the presence and further characteristics of this\ncapability in deep learning models. In this paper, we introduce a systematic\nprobing framework to explore the abstraction capability of deep learning models\nfrom a transferability perspective. A set of controlled experiments are\nconducted based on this framework, providing strong evidence that two probed\npre-trained language models (PLMs), T5 and GPT2, have the abstraction\ncapability. We also conduct in-depth analysis, thus shedding further light: (1)\nthe whole training phase exhibits a "memorize-then-abstract" two-stage process;\n(2) the learned abstract concepts are gathered in a few middle-layer attention\nheads, rather than being evenly distributed throughout the model; (3) the\nprobed abstraction capabilities exhibit robustness against concept mutations,\nand are more robust to low-level/source-side mutations than\nhigh-level/target-side ones; (4) generic pre-training is critical to the\nemergence of abstraction capability, and PLMs exhibit better abstraction with\nlarger model sizes and data scales.', 'authors': ['Shengnan An', 'Zeqi Lin', 'Bei Chen', 'Qiang Fu', 'Nanning Zheng', 'Jian-Guang Lou'], 'published': '2023-02-23 12:50:02+00:00', 'categories': ['cs.LG', 'cs.CL'], 'url': 'http://arxiv.org/abs/2302.11978v1'}
42,333
2006.02615v3
2,006.02615
3
Double Generative Adversarial Networks for Conditional Independence Testing
In this article, we study the problem of high-dimensional conditional independence testing, a key building block in statistics and machine learning. We propose an inferential procedure based on double generative adversarial networks (GANs). Specifically, we first introduce a double GANs framework to learn two generators of the conditional distributions. We then integrate the two generators to construct a test statistic, which takes the form of the maximum of generalized covariance measures of multiple transformation functions. We also employ data-splitting and cross-fitting to minimize the conditions on the generators to achieve the desired asymptotic properties, and employ multiplier bootstrap to obtain the corresponding $p$-value. We show that the constructed test statistic is doubly robust, and the resulting test both controls type-I error and has the power approaching one asymptotically. Also notably, we establish those theoretical guarantees under much weaker and practically more feasible conditions compared to the existing tests, and our proposal gives a concrete example of how to utilize some state-of-the-art deep learning tools, such as GANs, to help address a classical but challenging statistical problem. We demonstrate the efficacy of our test through both simulations and an application to an anti-cancer drug dataset. A Python implementation of the proposed procedure is available at https://github.com/tianlinxu312/dgcit.
2020-06-03
{'id': '2006.02615v3', 'title': 'Double Generative Adversarial Networks for Conditional Independence Testing', 'abstract': 'In this article, we study the problem of high-dimensional conditional\nindependence testing, a key building block in statistics and machine learning.\nWe propose an inferential procedure based on double generative adversarial\nnetworks (GANs). Specifically, we first introduce a double GANs framework to\nlearn two generators of the conditional distributions. We then integrate the\ntwo generators to construct a test statistic, which takes the form of the\nmaximum of generalized covariance measures of multiple transformation\nfunctions. We also employ data-splitting and cross-fitting to minimize the\nconditions on the generators to achieve the desired asymptotic properties, and\nemploy multiplier bootstrap to obtain the corresponding $p$-value. We show that\nthe constructed test statistic is doubly robust, and the resulting test both\ncontrols type-I error and has the power approaching one asymptotically. Also\nnotably, we establish those theoretical guarantees under much weaker and\npractically more feasible conditions compared to the existing tests, and our\nproposal gives a concrete example of how to utilize some state-of-the-art deep\nlearning tools, such as GANs, to help address a classical but challenging\nstatistical problem. We demonstrate the efficacy of our test through both\nsimulations and an application to an anti-cancer drug dataset. A Python\nimplementation of the proposed procedure is available at\nhttps://github.com/tianlinxu312/dgcit.', 'authors': ['Chengchun Shi', 'Tianlin Xu', 'Wicher Bergsma', 'Lexin Li'], 'published': '2020-06-03 16:14:15+00:00', 'categories': ['stat.ML', 'cs.LG'], 'url': 'http://arxiv.org/abs/2006.02615v3'}
42,334
2002.07650v5
2,002.0765
5
Uncertainty Estimation in Autoregressive Structured Prediction
Uncertainty estimation is important for ensuring safety and robustness of AI systems. While most research in the area has focused on un-structured prediction tasks, limited work has investigated general uncertainty estimation approaches for structured prediction. Thus, this work aims to investigate uncertainty estimation for autoregressive structured prediction tasks within a single unified and interpretable probabilistic ensemble-based framework. We consider: uncertainty estimation for sequence data at the token-level and complete sequence-level; interpretations for, and applications of, various measures of uncertainty; and discuss both the theoretical and practical challenges associated with obtaining them. This work also provides baselines for token-level and sequence-level error detection, and sequence-level out-of-domain input detection on the WMT'14 English-French and WMT'17 English-German translation and LibriSpeech speech recognition datasets.
2020-02-18
{'id': '2002.07650v5', 'title': 'Uncertainty Estimation in Autoregressive Structured Prediction', 'abstract': "Uncertainty estimation is important for ensuring safety and robustness of AI\nsystems. While most research in the area has focused on un-structured\nprediction tasks, limited work has investigated general uncertainty estimation\napproaches for structured prediction. Thus, this work aims to investigate\nuncertainty estimation for autoregressive structured prediction tasks within a\nsingle unified and interpretable probabilistic ensemble-based framework. We\nconsider: uncertainty estimation for sequence data at the token-level and\ncomplete sequence-level; interpretations for, and applications of, various\nmeasures of uncertainty; and discuss both the theoretical and practical\nchallenges associated with obtaining them. This work also provides baselines\nfor token-level and sequence-level error detection, and sequence-level\nout-of-domain input detection on the WMT'14 English-French and WMT'17\nEnglish-German translation and LibriSpeech speech recognition datasets.", 'authors': ['Andrey Malinin', 'Mark Gales'], 'published': '2020-02-18 15:40:13+00:00', 'categories': ['stat.ML', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2002.07650v5'}
42,338
2106.12764v1
2,106.12764
1
Density Constrained Reinforcement Learning
We study constrained reinforcement learning (CRL) from a novel perspective by setting constraints directly on state density functions, rather than the value functions considered by previous works. State density has a clear physical and mathematical interpretation, and is able to express a wide variety of constraints such as resource limits and safety requirements. Density constraints can also avoid the time-consuming process of designing and tuning cost functions required by value function-based constraints to encode system specifications. We leverage the duality between density functions and Q functions to develop an effective algorithm to solve the density constrained RL problem optimally and the constrains are guaranteed to be satisfied. We prove that the proposed algorithm converges to a near-optimal solution with a bounded error even when the policy update is imperfect. We use a set of comprehensive experiments to demonstrate the advantages of our approach over state-of-the-art CRL methods, with a wide range of density constrained tasks as well as standard CRL benchmarks such as Safety-Gym.
2021-06-24
{'id': '2106.12764v1', 'title': 'Density Constrained Reinforcement Learning', 'abstract': 'We study constrained reinforcement learning (CRL) from a novel perspective by\nsetting constraints directly on state density functions, rather than the value\nfunctions considered by previous works. State density has a clear physical and\nmathematical interpretation, and is able to express a wide variety of\nconstraints such as resource limits and safety requirements. Density\nconstraints can also avoid the time-consuming process of designing and tuning\ncost functions required by value function-based constraints to encode system\nspecifications. We leverage the duality between density functions and Q\nfunctions to develop an effective algorithm to solve the density constrained RL\nproblem optimally and the constrains are guaranteed to be satisfied. We prove\nthat the proposed algorithm converges to a near-optimal solution with a bounded\nerror even when the policy update is imperfect. We use a set of comprehensive\nexperiments to demonstrate the advantages of our approach over state-of-the-art\nCRL methods, with a wide range of density constrained tasks as well as standard\nCRL benchmarks such as Safety-Gym.', 'authors': ['Zengyi Qin', 'Yuxiao Chen', 'Chuchu Fan'], 'published': '2021-06-24 04:22:03+00:00', 'categories': ['cs.LG', 'cs.SY', 'eess.SY'], 'url': 'http://arxiv.org/abs/2106.12764v1'}
42,340
2010.01155v2
2,010.01155
2
Representational aspects of depth and conditioning in normalizing flows
Normalizing flows are among the most popular paradigms in generative modeling, especially for images, primarily because we can efficiently evaluate the likelihood of a data point. This is desirable both for evaluating the fit of a model, and for ease of training, as maximizing the likelihood can be done by gradient descent. However, training normalizing flows comes with difficulties as well: models which produce good samples typically need to be extremely deep -- which comes with accompanying vanishing/exploding gradient problems. A very related problem is that they are often poorly conditioned: since they are parametrized as invertible maps from $R^d R^d$, and typical training data like images intuitively is lower-dimensional, the learned maps often have Jacobians that are close to being singular. In our paper, we tackle representational aspects around depth and conditioning of normalizing flows: both for general invertible architectures, and for a particular common architecture, affine couplings. We prove that $(1)$ affine coupling layers suffice to exactly represent a permutation or $1 1$ convolution, as used in GLOW, showing that representationally the choice of partition is not a bottleneck for depth. We also show that shallow affine coupling networks are universal approximators in Wasserstein distance if ill-conditioning is allowed, and experimentally investigate related phenomena involving padding. Finally, we show a depth lower bound for general flow architectures with few neurons per layer and bounded Lipschitz constant.
2020-10-02
{'id': '2010.01155v2', 'title': 'Representational aspects of depth and conditioning in normalizing flows', 'abstract': 'Normalizing flows are among the most popular paradigms in generative\nmodeling, especially for images, primarily because we can efficiently evaluate\nthe likelihood of a data point. This is desirable both for evaluating the fit\nof a model, and for ease of training, as maximizing the likelihood can be done\nby gradient descent. However, training normalizing flows comes with\ndifficulties as well: models which produce good samples typically need to be\nextremely deep -- which comes with accompanying vanishing/exploding gradient\nproblems. A very related problem is that they are often poorly conditioned:\nsince they are parametrized as invertible maps from $\\mathbb{R}^d \\to\n\\mathbb{R}^d$, and typical training data like images intuitively is\nlower-dimensional, the learned maps often have Jacobians that are close to\nbeing singular.\n In our paper, we tackle representational aspects around depth and\nconditioning of normalizing flows: both for general invertible architectures,\nand for a particular common architecture, affine couplings. We prove that\n$\\Theta(1)$ affine coupling layers suffice to exactly represent a permutation\nor $1 \\times 1$ convolution, as used in GLOW, showing that representationally\nthe choice of partition is not a bottleneck for depth. We also show that\nshallow affine coupling networks are universal approximators in Wasserstein\ndistance if ill-conditioning is allowed, and experimentally investigate related\nphenomena involving padding. Finally, we show a depth lower bound for general\nflow architectures with few neurons per layer and bounded Lipschitz constant.', 'authors': ['Frederic Koehler', 'Viraj Mehta', 'Andrej Risteski'], 'published': '2020-10-02 18:15:45+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2010.01155v2'}
42,350
2011.10039v2
2,011.10039
2
Creative Sketch Generation
Sketching or doodling is a popular creative activity that people engage in. However, most existing work in automatic sketch understanding or generation has focused on sketches that are quite mundane. In this work, we introduce two datasets of creative sketches -- Creative Birds and Creative Creatures -- containing 10k sketches each along with part annotations. We propose DoodlerGAN -- a part-based Generative Adversarial Network (GAN) -- to generate unseen compositions of novel part appearances. Quantitative evaluations as well as human studies demonstrate that sketches generated by our approach are more creative and of higher quality than existing approaches. In fact, in Creative Birds, subjects prefer sketches generated by DoodlerGAN over those drawn by humans! Our code can be found at https://github.com/facebookresearch/DoodlerGAN and a demo can be found at http://doodlergan.cloudcv.org.
2020-11-19
{'id': '2011.10039v2', 'title': 'Creative Sketch Generation', 'abstract': 'Sketching or doodling is a popular creative activity that people engage in.\nHowever, most existing work in automatic sketch understanding or generation has\nfocused on sketches that are quite mundane. In this work, we introduce two\ndatasets of creative sketches -- Creative Birds and Creative Creatures --\ncontaining 10k sketches each along with part annotations. We propose DoodlerGAN\n-- a part-based Generative Adversarial Network (GAN) -- to generate unseen\ncompositions of novel part appearances. Quantitative evaluations as well as\nhuman studies demonstrate that sketches generated by our approach are more\ncreative and of higher quality than existing approaches. In fact, in Creative\nBirds, subjects prefer sketches generated by DoodlerGAN over those drawn by\nhumans! Our code can be found at https://github.com/facebookresearch/DoodlerGAN\nand a demo can be found at http://doodlergan.cloudcv.org.', 'authors': ['Songwei Ge', 'Vedanuj Goswami', 'C. Lawrence Zitnick', 'Devi Parikh'], 'published': '2020-11-19 18:57:00+00:00', 'categories': ['cs.CV', 'cs.AI'], 'url': 'http://arxiv.org/abs/2011.10039v2'}
42,353
2009.13447v3
2,009.13447
3
Why resampling outperforms reweighting for correcting sampling bias with stochastic gradients
A data set sampled from a certain population is biased if the subgroups of the population are sampled at proportions that are significantly different from their underlying proportions. Training machine learning models on biased data sets requires correction techniques to compensate for the bias. We consider two commonly-used techniques, resampling and reweighting, that rebalance the proportions of the subgroups to maintain the desired objective function. Though statistically equivalent, it has been observed that resampling outperforms reweighting when combined with stochastic gradient algorithms. By analyzing illustrative examples, we explain the reason behind this phenomenon using tools from dynamical stability and stochastic asymptotics. We also present experiments from regression, classification, and off-policy prediction to demonstrate that this is a general phenomenon. We argue that it is imperative to consider the objective function design and the optimization algorithm together while addressing the sampling bias.
2020-09-28
{'id': '2009.13447v3', 'title': 'Why resampling outperforms reweighting for correcting sampling bias with stochastic gradients', 'abstract': 'A data set sampled from a certain population is biased if the subgroups of\nthe population are sampled at proportions that are significantly different from\ntheir underlying proportions. Training machine learning models on biased data\nsets requires correction techniques to compensate for the bias. We consider two\ncommonly-used techniques, resampling and reweighting, that rebalance the\nproportions of the subgroups to maintain the desired objective function. Though\nstatistically equivalent, it has been observed that resampling outperforms\nreweighting when combined with stochastic gradient algorithms. By analyzing\nillustrative examples, we explain the reason behind this phenomenon using tools\nfrom dynamical stability and stochastic asymptotics. We also present\nexperiments from regression, classification, and off-policy prediction to\ndemonstrate that this is a general phenomenon. We argue that it is imperative\nto consider the objective function design and the optimization algorithm\ntogether while addressing the sampling bias.', 'authors': ['Jing An', 'Lexing Ying', 'Yuhua Zhu'], 'published': '2020-09-28 16:12:38+00:00', 'categories': ['cs.LG', 'cs.NA', 'math.NA', 'stat.ML'], 'url': 'http://arxiv.org/abs/2009.13447v3'}
42,339
2302.07350v2
2,302.0735
2
Graph schemas as abstractions for transfer learning, inference, and planning
Transferring latent structure from one environment or problem to another is a mechanism by which humans and animals generalize with very little data. Inspired by cognitive and neurobiological insights, we propose graph schemas as a mechanism of abstraction for transfer learning. Graph schemas start with latent graph learning where perceptually aliased observations are disambiguated in the latent space using contextual information. Latent graph learning is also emerging as a new computational model of the hippocampus to explain map learning and transitive inference. Our insight is that a latent graph can be treated as a flexible template -- a schema -- that models concepts and behaviors, with slots that bind groups of latent nodes to the specific observations or groundings. By treating learned latent graphs (schemas) as prior knowledge, new environments can be quickly learned as compositions of schemas and their newly learned bindings. We evaluate graph schemas on two previously published challenging tasks: the memory & planning game and one-shot StreetLearn, which are designed to test rapid task solving in novel environments. Graph schemas can be learned in far fewer episodes than previous baselines, and can model and plan in a few steps in novel variations of these tasks. We also demonstrate learning, matching, and reusing graph schemas in more challenging 2D and 3D environments with extensive perceptual aliasing and size variations, and show how different schemas can be composed to model larger and more complex environments. To summarize, our main contribution is a unified system, inspired and grounded in cognitive science, that facilitates rapid transfer learning of new environments using schemas via map-induction and composition that handles perceptual aliasing.
2023-02-14
{'id': '2302.07350v2', 'title': 'Graph schemas as abstractions for transfer learning, inference, and planning', 'abstract': 'Transferring latent structure from one environment or problem to another is a\nmechanism by which humans and animals generalize with very little data.\nInspired by cognitive and neurobiological insights, we propose graph schemas as\na mechanism of abstraction for transfer learning. Graph schemas start with\nlatent graph learning where perceptually aliased observations are disambiguated\nin the latent space using contextual information. Latent graph learning is also\nemerging as a new computational model of the hippocampus to explain map\nlearning and transitive inference. Our insight is that a latent graph can be\ntreated as a flexible template -- a schema -- that models concepts and\nbehaviors, with slots that bind groups of latent nodes to the specific\nobservations or groundings. By treating learned latent graphs (schemas) as\nprior knowledge, new environments can be quickly learned as compositions of\nschemas and their newly learned bindings. We evaluate graph schemas on two\npreviously published challenging tasks: the memory & planning game and one-shot\nStreetLearn, which are designed to test rapid task solving in novel\nenvironments. Graph schemas can be learned in far fewer episodes than previous\nbaselines, and can model and plan in a few steps in novel variations of these\ntasks. We also demonstrate learning, matching, and reusing graph schemas in\nmore challenging 2D and 3D environments with extensive perceptual aliasing and\nsize variations, and show how different schemas can be composed to model larger\nand more complex environments. To summarize, our main contribution is a unified\nsystem, inspired and grounded in cognitive science, that facilitates rapid\ntransfer learning of new environments using schemas via map-induction and\ncomposition that handles perceptual aliasing.', 'authors': ['J. Swaroop Guntupalli', 'Rajkumar Vasudeva Raju', 'Shrinu Kushagra', 'Carter Wendelken', 'Danny Sawyer', 'Ishan Deshpande', 'Guangyao Zhou', 'Miguel Lázaro-Gredilla', 'Dileep George'], 'published': '2023-02-14 21:23:22+00:00', 'categories': ['cs.AI', 'cs.LG', 'q-bio.NC'], 'url': 'http://arxiv.org/abs/2302.07350v2'}
42,342
2302.02931v2
2,302.02931
2
Bitrate-Constrained DRO: Beyond Worst Case Robustness To Unknown Group Shifts
Training machine learning models robust to distribution shifts is critical for real-world applications. Some robust training algorithms (e.g., Group DRO) specialize to group shifts and require group information on all training points. Other methods (e.g., CVaR DRO) that do not need group annotations can be overly conservative, since they naively upweight high loss points which may form a contrived set that does not correspond to any meaningful group in the real world (e.g., when the high loss points are randomly mislabeled training points). In this work, we address limitations in prior approaches by assuming a more nuanced form of group shift: conditioned on the label, we assume that the true group function (indicator over group) is simple. For example, we may expect that group shifts occur along low bitrate features (e.g., image background, lighting). Thus, we aim to learn a model that maintains high accuracy on simple group functions realized by these low bitrate features, that need not spend valuable model capacity achieving high accuracy on contrived groups of examples. Based on this, we consider the two-player game formulation of DRO where the adversary's capacity is bitrate-constrained. Our resulting practical algorithm, Bitrate-Constrained DRO (BR-DRO), does not require group information on training samples yet matches the performance of Group DRO on datasets that have training group annotations and that of CVaR DRO on long-tailed distributions. Our theoretical analysis reveals that in some settings BR-DRO objective can provably yield statistically efficient and less conservative solutions than unconstrained CVaR DRO.
2023-02-06
{'id': '2302.02931v2', 'title': 'Bitrate-Constrained DRO: Beyond Worst Case Robustness To Unknown Group Shifts', 'abstract': "Training machine learning models robust to distribution shifts is critical\nfor real-world applications. Some robust training algorithms (e.g., Group DRO)\nspecialize to group shifts and require group information on all training\npoints. Other methods (e.g., CVaR DRO) that do not need group annotations can\nbe overly conservative, since they naively upweight high loss points which may\nform a contrived set that does not correspond to any meaningful group in the\nreal world (e.g., when the high loss points are randomly mislabeled training\npoints). In this work, we address limitations in prior approaches by assuming a\nmore nuanced form of group shift: conditioned on the label, we assume that the\ntrue group function (indicator over group) is simple. For example, we may\nexpect that group shifts occur along low bitrate features (e.g., image\nbackground, lighting). Thus, we aim to learn a model that maintains high\naccuracy on simple group functions realized by these low bitrate features, that\nneed not spend valuable model capacity achieving high accuracy on contrived\ngroups of examples. Based on this, we consider the two-player game formulation\nof DRO where the adversary's capacity is bitrate-constrained. Our resulting\npractical algorithm, Bitrate-Constrained DRO (BR-DRO), does not require group\ninformation on training samples yet matches the performance of Group DRO on\ndatasets that have training group annotations and that of CVaR DRO on\nlong-tailed distributions. Our theoretical analysis reveals that in some\nsettings BR-DRO objective can provably yield statistically efficient and less\nconservative solutions than unconstrained CVaR DRO.", 'authors': ['Amrith Setlur', 'Don Dennis', 'Benjamin Eysenbach', 'Aditi Raghunathan', 'Chelsea Finn', 'Virginia Smith', 'Sergey Levine'], 'published': '2023-02-06 17:07:16+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2302.02931v2'}
42,360
2205.05869v2
2,205.05869
2
View Synthesis with Sculpted Neural Points
We address the task of view synthesis, generating novel views of a scene given a set of images as input. In many recent works such as NeRF (Mildenhall et al., 2020), the scene geometry is parameterized using neural implicit representations (i.e., MLPs). Implicit neural representations have achieved impressive visual quality but have drawbacks in computational efficiency. In this work, we propose a new approach that performs view synthesis using point clouds. It is the first point-based method that achieves better visual quality than NeRF while being 100x faster in rendering speed. Our approach builds on existing works on differentiable point-based rendering but introduces a novel technique we call "Sculpted Neural Points (SNP)", which significantly improves the robustness to errors and holes in the reconstructed point cloud. We further propose to use view-dependent point features based on spherical harmonics to capture non-Lambertian surfaces, and new designs in the point-based rendering pipeline that further boost the performance. Finally, we show that our system supports fine-grained scene editing. Code is available at https://github.com/princeton-vl/SNP.
2022-05-12
{'id': '2205.05869v2', 'title': 'View Synthesis with Sculpted Neural Points', 'abstract': 'We address the task of view synthesis, generating novel views of a scene\ngiven a set of images as input. In many recent works such as NeRF (Mildenhall\net al., 2020), the scene geometry is parameterized using neural implicit\nrepresentations (i.e., MLPs). Implicit neural representations have achieved\nimpressive visual quality but have drawbacks in computational efficiency. In\nthis work, we propose a new approach that performs view synthesis using point\nclouds. It is the first point-based method that achieves better visual quality\nthan NeRF while being 100x faster in rendering speed. Our approach builds on\nexisting works on differentiable point-based rendering but introduces a novel\ntechnique we call "Sculpted Neural Points (SNP)", which significantly improves\nthe robustness to errors and holes in the reconstructed point cloud. We further\npropose to use view-dependent point features based on spherical harmonics to\ncapture non-Lambertian surfaces, and new designs in the point-based rendering\npipeline that further boost the performance. Finally, we show that our system\nsupports fine-grained scene editing. Code is available at\nhttps://github.com/princeton-vl/SNP.', 'authors': ['Yiming Zuo', 'Jia Deng'], 'published': '2022-05-12 03:54:35+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2205.05869v2'}
42,362
2302.03018v1
2,302.03018
1
DDM$^2$: Self-Supervised Diffusion MRI Denoising with Generative Diffusion Models
Magnetic resonance imaging (MRI) is a common and life-saving medical imaging technique. However, acquiring high signal-to-noise ratio MRI scans requires long scan times, resulting in increased costs and patient discomfort, and decreased throughput. Thus, there is great interest in denoising MRI scans, especially for the subtype of diffusion MRI scans that are severely SNR-limited. While most prior MRI denoising methods are supervised in nature, acquiring supervised training datasets for the multitude of anatomies, MRI scanners, and scan parameters proves impractical. Here, we propose Denoising Diffusion Models for Denoising Diffusion MRI (DDM$^2$), a self-supervised denoising method for MRI denoising using diffusion denoising generative models. Our three-stage framework integrates statistic-based denoising theory into diffusion models and performs denoising through conditional generation. During inference, we represent input noisy measurements as a sample from an intermediate posterior distribution within the diffusion Markov chain. We conduct experiments on 4 real-world in-vivo diffusion MRI datasets and show that our DDM$^2$ demonstrates superior denoising performances ascertained with clinically-relevant visual qualitative and quantitative metrics.
2023-02-06
{'id': '2302.03018v1', 'title': 'DDM$^2$: Self-Supervised Diffusion MRI Denoising with Generative Diffusion Models', 'abstract': 'Magnetic resonance imaging (MRI) is a common and life-saving medical imaging\ntechnique. However, acquiring high signal-to-noise ratio MRI scans requires\nlong scan times, resulting in increased costs and patient discomfort, and\ndecreased throughput. Thus, there is great interest in denoising MRI scans,\nespecially for the subtype of diffusion MRI scans that are severely\nSNR-limited. While most prior MRI denoising methods are supervised in nature,\nacquiring supervised training datasets for the multitude of anatomies, MRI\nscanners, and scan parameters proves impractical. Here, we propose Denoising\nDiffusion Models for Denoising Diffusion MRI (DDM$^2$), a self-supervised\ndenoising method for MRI denoising using diffusion denoising generative models.\nOur three-stage framework integrates statistic-based denoising theory into\ndiffusion models and performs denoising through conditional generation. During\ninference, we represent input noisy measurements as a sample from an\nintermediate posterior distribution within the diffusion Markov chain. We\nconduct experiments on 4 real-world in-vivo diffusion MRI datasets and show\nthat our DDM$^2$ demonstrates superior denoising performances ascertained with\nclinically-relevant visual qualitative and quantitative metrics.', 'authors': ['Tiange Xiang', 'Mahmut Yurt', 'Ali B Syed', 'Kawin Setsompop', 'Akshay Chaudhari'], 'published': '2023-02-06 18:56:39+00:00', 'categories': ['eess.IV', 'cs.CV'], 'url': 'http://arxiv.org/abs/2302.03018v1'}
42,365
2201.08115v2
2,201.08115
2
Priors, Hierarchy, and Information Asymmetry for Skill Transfer in Reinforcement Learning
The ability to discover behaviours from past experience and transfer them to new tasks is a hallmark of intelligent agents acting sample-efficiently in the real world. Equipping embodied reinforcement learners with the same ability may be crucial for their successful deployment in robotics. While hierarchical and KL-regularized reinforcement learning individually hold promise here, arguably a hybrid approach could combine their respective benefits. Key to these fields is the use of information asymmetry across architectural modules to bias which skills are learnt. While asymmetry choice has a large influence on transferability, existing methods base their choice primarily on intuition in a domain-independent, potentially sub-optimal, manner. In this paper, we theoretically and empirically show the crucial expressivity-transferability trade-off of skills across sequential tasks, controlled by information asymmetry. Given this insight, we introduce Attentive Priors for Expressive and Transferable Skills (APES), a hierarchical KL-regularized method, heavily benefiting from both priors and hierarchy. Unlike existing approaches, APES automates the choice of asymmetry by learning it in a data-driven, domain-dependent, way based on our expressivity-transferability theorems. Experiments over complex transfer domains of varying levels of extrapolation and sparsity, such as robot block stacking, demonstrate the criticality of the correct asymmetric choice, with APES drastically outperforming previous methods.
2022-01-20
{'id': '2201.08115v2', 'title': 'Priors, Hierarchy, and Information Asymmetry for Skill Transfer in Reinforcement Learning', 'abstract': 'The ability to discover behaviours from past experience and transfer them to\nnew tasks is a hallmark of intelligent agents acting sample-efficiently in the\nreal world. Equipping embodied reinforcement learners with the same ability may\nbe crucial for their successful deployment in robotics. While hierarchical and\nKL-regularized reinforcement learning individually hold promise here, arguably\na hybrid approach could combine their respective benefits. Key to these fields\nis the use of information asymmetry across architectural modules to bias which\nskills are learnt. While asymmetry choice has a large influence on\ntransferability, existing methods base their choice primarily on intuition in a\ndomain-independent, potentially sub-optimal, manner. In this paper, we\ntheoretically and empirically show the crucial expressivity-transferability\ntrade-off of skills across sequential tasks, controlled by information\nasymmetry. Given this insight, we introduce Attentive Priors for Expressive and\nTransferable Skills (APES), a hierarchical KL-regularized method, heavily\nbenefiting from both priors and hierarchy. Unlike existing approaches, APES\nautomates the choice of asymmetry by learning it in a data-driven,\ndomain-dependent, way based on our expressivity-transferability theorems.\nExperiments over complex transfer domains of varying levels of extrapolation\nand sparsity, such as robot block stacking, demonstrate the criticality of the\ncorrect asymmetric choice, with APES drastically outperforming previous\nmethods.', 'authors': ['Sasha Salter', 'Kristian Hartikainen', 'Walter Goodwin', 'Ingmar Posner'], 'published': '2022-01-20 11:12:56+00:00', 'categories': ['cs.AI', 'cs.LG', 'cs.RO', 'stat.ML'], 'url': 'http://arxiv.org/abs/2201.08115v2'}
42,341
2310.05422v1
2,310.05422
1
Reward-Consistent Dynamics Models are Strongly Generalizable for Offline Reinforcement Learning
Learning a precise dynamics model can be crucial for offline reinforcement learning, which, unfortunately, has been found to be quite challenging. Dynamics models that are learned by fitting historical transitions often struggle to generalize to unseen transitions. In this study, we identify a hidden but pivotal factor termed dynamics reward that remains consistent across transitions, offering a pathway to better generalization. Therefore, we propose the idea of reward-consistent dynamics models: any trajectory generated by the dynamics model should maximize the dynamics reward derived from the data. We implement this idea as the MOREC (Model-based Offline reinforcement learning with Reward Consistency) method, which can be seamlessly integrated into previous offline model-based reinforcement learning (MBRL) methods. MOREC learns a generalizable dynamics reward function from offline data, which is subsequently employed as a transition filter in any offline MBRL method: when generating transitions, the dynamics model generates a batch of transitions and selects the one with the highest dynamics reward value. On a synthetic task, we visualize that MOREC has a strong generalization ability and can surprisingly recover some distant unseen transitions. On 21 offline tasks in D4RL and NeoRL benchmarks, MOREC improves the previous state-of-the-art performance by a significant margin, i.e., 4.6% on D4RL tasks and 25.9% on NeoRL tasks. Notably, MOREC is the first method that can achieve above 95% online RL performance in 6 out of 12 D4RL tasks and 3 out of 9 NeoRL tasks.
2023-10-09
{'id': '2310.05422v1', 'title': 'Reward-Consistent Dynamics Models are Strongly Generalizable for Offline Reinforcement Learning', 'abstract': 'Learning a precise dynamics model can be crucial for offline reinforcement\nlearning, which, unfortunately, has been found to be quite challenging.\nDynamics models that are learned by fitting historical transitions often\nstruggle to generalize to unseen transitions. In this study, we identify a\nhidden but pivotal factor termed dynamics reward that remains consistent across\ntransitions, offering a pathway to better generalization. Therefore, we propose\nthe idea of reward-consistent dynamics models: any trajectory generated by the\ndynamics model should maximize the dynamics reward derived from the data. We\nimplement this idea as the MOREC (Model-based Offline reinforcement learning\nwith Reward Consistency) method, which can be seamlessly integrated into\nprevious offline model-based reinforcement learning (MBRL) methods. MOREC\nlearns a generalizable dynamics reward function from offline data, which is\nsubsequently employed as a transition filter in any offline MBRL method: when\ngenerating transitions, the dynamics model generates a batch of transitions and\nselects the one with the highest dynamics reward value. On a synthetic task, we\nvisualize that MOREC has a strong generalization ability and can surprisingly\nrecover some distant unseen transitions. On 21 offline tasks in D4RL and NeoRL\nbenchmarks, MOREC improves the previous state-of-the-art performance by a\nsignificant margin, i.e., 4.6% on D4RL tasks and 25.9% on NeoRL tasks. Notably,\nMOREC is the first method that can achieve above 95% online RL performance in 6\nout of 12 D4RL tasks and 3 out of 9 NeoRL tasks.', 'authors': ['Fan-Ming Luo', 'Tian Xu', 'Xingchen Cao', 'Yang Yu'], 'published': '2023-10-09 05:37:58+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.05422v1'}
42,349
2305.15264v1
2,305.15264
1
Error Feedback Shines when Features are Rare
We provide the first proof that gradient descent $(greenGD)$ with greedy sparsification $(greenTopK)$ and error feedback $(greenEF)$ can obtain better communication complexity than vanilla $greenGD$ when solving the distributed optimization problem $_xR^d f(x)=1n_i=1^n f_i(x)$, where $n$ = # of clients, $d$ = # of features, and $f_1,,f_n$ are smooth nonconvex functions. Despite intensive research since 2014 when $greenEF$ was first proposed by Seide et al., this problem remained open until now. We show that $greenEF$ shines in the regime when features are rare, i.e., when each feature is present in the data owned by a small number of clients only. To illustrate our main result, we show that in order to find a random vector $x$ such that $f(x) ^2 $ in expectation, $greenGD$ with the $greenTop1$ sparsifier and $greenEF$ requires $O (( L+bluer redcn ( redcn _i L_i^2, 1n_i=1^n L_i^2 ) ) 1 )$ bits to be communicated by each worker to the server only, where $L$ is the smoothness constant of $f$, $L_i$ is the smoothness constant of $f_i$, $redc$ is the maximal number of clients owning any feature ($1redc n$), and $bluer$ is the maximal number of features owned by any client ($1bluer d$). Clearly, the communication complexity improves as $redc$ decreases (i.e., as features become more rare), and can be much better than the $O(bluer L 1)$ communication complexity of $greenGD$ in the same regime.
2023-05-24
{'id': '2305.15264v1', 'title': 'Error Feedback Shines when Features are Rare', 'abstract': 'We provide the first proof that gradient descent $\\left({\\color{green}\\sf\nGD}\\right)$ with greedy sparsification $\\left({\\color{green}\\sf TopK}\\right)$\nand error feedback $\\left({\\color{green}\\sf EF}\\right)$ can obtain better\ncommunication complexity than vanilla ${\\color{green}\\sf GD}$ when solving the\ndistributed optimization problem $\\min_{x\\in \\mathbb{R}^d}\n{f(x)=\\frac{1}{n}\\sum_{i=1}^n f_i(x)}$, where $n$ = # of clients, $d$ = # of\nfeatures, and $f_1,\\dots,f_n$ are smooth nonconvex functions. Despite intensive\nresearch since 2014 when ${\\color{green}\\sf EF}$ was first proposed by Seide et\nal., this problem remained open until now. We show that ${\\color{green}\\sf EF}$\nshines in the regime when features are rare, i.e., when each feature is present\nin the data owned by a small number of clients only. To illustrate our main\nresult, we show that in order to find a random vector $\\hat{x}$ such that\n$\\lVert {\\nabla f(\\hat{x})} \\rVert^2 \\leq \\varepsilon$ in expectation,\n${\\color{green}\\sf GD}$ with the ${\\color{green}\\sf Top1}$ sparsifier and\n${\\color{green}\\sf EF}$ requires ${\\cal O} \\left(\\left( L+{\\color{blue}r}\n\\sqrt{ \\frac{{\\color{red}c}}{n} \\min \\left( \\frac{{\\color{red}c}}{n} \\max_i\nL_i^2, \\frac{1}{n}\\sum_{i=1}^n L_i^2 \\right) }\\right) \\frac{1}{\\varepsilon}\n\\right)$ bits to be communicated by each worker to the server only, where $L$\nis the smoothness constant of $f$, $L_i$ is the smoothness constant of $f_i$,\n${\\color{red}c}$ is the maximal number of clients owning any feature ($1\\leq\n{\\color{red}c} \\leq n$), and ${\\color{blue}r}$ is the maximal number of\nfeatures owned by any client ($1\\leq {\\color{blue}r} \\leq d$). Clearly, the\ncommunication complexity improves as ${\\color{red}c}$ decreases (i.e., as\nfeatures become more rare), and can be much better than the ${\\cal\nO}({\\color{blue}r} L \\frac{1}{\\varepsilon})$ communication complexity of\n${\\color{green}\\sf GD}$ in the same regime.', 'authors': ['Peter Richtárik', 'Elnur Gasanov', 'Konstantin Burlachenko'], 'published': '2023-05-24 15:52:07+00:00', 'categories': ['math.OC', 'cs.DC', 'cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2305.15264v1'}
42,356
2312.15918v2
2,312.15918
2
Supervised Knowledge Makes Large Language Models Better In-context Learners
Large Language Models (LLMs) exhibit emerging in-context learning abilities through prompt engineering. The recent progress in large-scale generative models has further expanded their use in real-world language applications. However, the critical challenge of improving the generalizability and factuality of LLMs in natural language understanding and question answering remains under-explored. While previous in-context learning research has focused on enhancing models to adhere to users' specific instructions and quality expectations, and to avoid undesired outputs, little to no work has explored the use of task-Specific fine-tuned Language Models (SLMs) to improve LLMs' in-context learning during the inference stage. Our primary contribution is the establishment of a simple yet effective framework that enhances the reliability of LLMs as it: 1) generalizes out-of-distribution data, 2) elucidates how LLMs benefit from discriminative models, and 3) minimizes hallucinations in generative tasks. Using our proposed plug-in method, enhanced versions of Llama 2 and ChatGPT surpass their original versions regarding generalizability and factuality. We offer a comprehensive suite of resources, including 16 curated datasets, prompts, model checkpoints, and LLM outputs across 9 distinct tasks. The code and data are released at: https://github.com/YangLinyi/Supervised-Knowledge-Makes-Large-Language-Models-Better-In-context-Learners. Our empirical analysis sheds light on the advantages of incorporating discriminative models into LLMs and highlights the potential of our methodology in fostering more reliable LLMs.
2023-12-26
{'id': '2312.15918v2', 'title': 'Supervised Knowledge Makes Large Language Models Better In-context Learners', 'abstract': "Large Language Models (LLMs) exhibit emerging in-context learning abilities\nthrough prompt engineering. The recent progress in large-scale generative\nmodels has further expanded their use in real-world language applications.\nHowever, the critical challenge of improving the generalizability and\nfactuality of LLMs in natural language understanding and question answering\nremains under-explored. While previous in-context learning research has focused\non enhancing models to adhere to users' specific instructions and quality\nexpectations, and to avoid undesired outputs, little to no work has explored\nthe use of task-Specific fine-tuned Language Models (SLMs) to improve LLMs'\nin-context learning during the inference stage. Our primary contribution is the\nestablishment of a simple yet effective framework that enhances the reliability\nof LLMs as it: 1) generalizes out-of-distribution data, 2) elucidates how LLMs\nbenefit from discriminative models, and 3) minimizes hallucinations in\ngenerative tasks. Using our proposed plug-in method, enhanced versions of Llama\n2 and ChatGPT surpass their original versions regarding generalizability and\nfactuality. We offer a comprehensive suite of resources, including 16 curated\ndatasets, prompts, model checkpoints, and LLM outputs across 9 distinct tasks.\nThe code and data are released at:\nhttps://github.com/YangLinyi/Supervised-Knowledge-Makes-Large-Language-Models-Better-In-context-Learners.\nOur empirical analysis sheds light on the advantages of incorporating\ndiscriminative models into LLMs and highlights the potential of our methodology\nin fostering more reliable LLMs.", 'authors': ['Linyi Yang', 'Shuibai Zhang', 'Zhuohao Yu', 'Guangsheng Bao', 'Yidong Wang', 'Jindong Wang', 'Ruochen Xu', 'Wei Ye', 'Xing Xie', 'Weizhu Chen', 'Yue Zhang'], 'published': '2023-12-26 07:24:46+00:00', 'categories': ['cs.CL', 'cs.AI'], 'url': 'http://arxiv.org/abs/2312.15918v2'}
42,395
2310.15149v1
2,310.15149
1
Unlocking the Transferability of Tokens in Deep Models for Tabular Data
Fine-tuning a pre-trained deep neural network has become a successful paradigm in various machine learning tasks. However, such a paradigm becomes particularly challenging with tabular data when there are discrepancies between the feature sets of pre-trained models and the target tasks. In this paper, we propose TabToken, a method aims at enhancing the quality of feature tokens (i.e., embeddings of tabular features). TabToken allows for the utilization of pre-trained models when the upstream and downstream tasks share overlapping features, facilitating model fine-tuning even with limited training examples. Specifically, we introduce a contrastive objective that regularizes the tokens, capturing the semantics within and across features. During the pre-training stage, the tokens are learned jointly with top-layer deep models such as transformer. In the downstream task, tokens of the shared features are kept fixed while TabToken efficiently fine-tunes the remaining parts of the model. TabToken not only enables knowledge transfer from a pre-trained model to tasks with heterogeneous features, but also enhances the discriminative ability of deep tabular models in standard classification and regression tasks.
2023-10-23
{'id': '2310.15149v1', 'title': 'Unlocking the Transferability of Tokens in Deep Models for Tabular Data', 'abstract': 'Fine-tuning a pre-trained deep neural network has become a successful\nparadigm in various machine learning tasks. However, such a paradigm becomes\nparticularly challenging with tabular data when there are discrepancies between\nthe feature sets of pre-trained models and the target tasks. In this paper, we\npropose TabToken, a method aims at enhancing the quality of feature tokens\n(i.e., embeddings of tabular features). TabToken allows for the utilization of\npre-trained models when the upstream and downstream tasks share overlapping\nfeatures, facilitating model fine-tuning even with limited training examples.\nSpecifically, we introduce a contrastive objective that regularizes the tokens,\ncapturing the semantics within and across features. During the pre-training\nstage, the tokens are learned jointly with top-layer deep models such as\ntransformer. In the downstream task, tokens of the shared features are kept\nfixed while TabToken efficiently fine-tunes the remaining parts of the model.\nTabToken not only enables knowledge transfer from a pre-trained model to tasks\nwith heterogeneous features, but also enhances the discriminative ability of\ndeep tabular models in standard classification and regression tasks.', 'authors': ['Qi-Le Zhou', 'Han-Jia Ye', 'Le-Ye Wang', 'De-Chuan Zhan'], 'published': '2023-10-23 17:53:09+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.15149v1'}
42,399
2402.03545v3
2,402.03545
3
Online Feature Updates Improve Online (Generalized) Label Shift Adaptation
This paper addresses the prevalent issue of label shift in an online setting with missing labels, where data distributions change over time and obtaining timely labels is challenging. While existing methods primarily focus on adjusting or updating the final layer of a pre-trained classifier, we explore the untapped potential of enhancing feature representations using unlabeled data at test-time. Our novel method, Online Label Shift adaptation with Online Feature Updates (OLS-OFU), leverages self-supervised learning to refine the feature extraction process, thereby improving the prediction model. By carefully designing the algorithm, theoretically OLS-OFU maintains the similar online regret convergence to the results in the literature while taking the improved features into account. Empirically, it achieves substantial improvements over existing methods, which is as significant as the gains existing methods have over the baseline (i.e., without distribution shift adaptations).
2024-02-05
{'id': '2402.03545v3', 'title': 'Online Feature Updates Improve Online (Generalized) Label Shift Adaptation', 'abstract': 'This paper addresses the prevalent issue of label shift in an online setting\nwith missing labels, where data distributions change over time and obtaining\ntimely labels is challenging. While existing methods primarily focus on\nadjusting or updating the final layer of a pre-trained classifier, we explore\nthe untapped potential of enhancing feature representations using unlabeled\ndata at test-time. Our novel method, Online Label Shift adaptation with Online\nFeature Updates (OLS-OFU), leverages self-supervised learning to refine the\nfeature extraction process, thereby improving the prediction model. By\ncarefully designing the algorithm, theoretically OLS-OFU maintains the similar\nonline regret convergence to the results in the literature while taking the\nimproved features into account. Empirically, it achieves substantial\nimprovements over existing methods, which is as significant as the gains\nexisting methods have over the baseline (i.e., without distribution shift\nadaptations).', 'authors': ['Ruihan Wu', 'Siddhartha Datta', 'Yi Su', 'Dheeraj Baby', 'Yu-Xiang Wang', 'Kilian Q. Weinberger'], 'published': '2024-02-05 22:03:25+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2402.03545v3'}
42,410
2404.06694v2
2,404.06694
2
How to Craft Backdoors with Unlabeled Data Alone?
Relying only on unlabeled data, Self-supervised learning (SSL) can learn rich features in an economical and scalable way. As the drive-horse for building foundation models, SSL has received a lot of attention recently with wide applications, which also raises security concerns where backdoor attack is a major type of threat: if the released dataset is maliciously poisoned, backdoored SSL models can behave badly when triggers are injected to test samples. The goal of this work is to investigate this potential risk. We notice that existing backdoors all require a considerable amount of labeled data that may not be available for SSL. To circumvent this limitation, we explore a more restrictive setting called no-label backdoors, where we only have access to the unlabeled data alone, where the key challenge is how to select the proper poison set without using label information. We propose two strategies for poison selection: clustering-based selection using pseudolabels, and contrastive selection derived from the mutual information principle. Experiments on CIFAR-10 and ImageNet-100 show that both no-label backdoors are effective on many SSL methods and outperform random poisoning by a large margin. Code will be available at https://github.com/PKU-ML/nlb.
2024-04-10
{'id': '2404.06694v2', 'title': 'How to Craft Backdoors with Unlabeled Data Alone?', 'abstract': 'Relying only on unlabeled data, Self-supervised learning (SSL) can learn rich\nfeatures in an economical and scalable way. As the drive-horse for building\nfoundation models, SSL has received a lot of attention recently with wide\napplications, which also raises security concerns where backdoor attack is a\nmajor type of threat: if the released dataset is maliciously poisoned,\nbackdoored SSL models can behave badly when triggers are injected to test\nsamples. The goal of this work is to investigate this potential risk. We notice\nthat existing backdoors all require a considerable amount of \\emph{labeled}\ndata that may not be available for SSL. To circumvent this limitation, we\nexplore a more restrictive setting called no-label backdoors, where we only\nhave access to the unlabeled data alone, where the key challenge is how to\nselect the proper poison set without using label information. We propose two\nstrategies for poison selection: clustering-based selection using pseudolabels,\nand contrastive selection derived from the mutual information principle.\nExperiments on CIFAR-10 and ImageNet-100 show that both no-label backdoors are\neffective on many SSL methods and outperform random poisoning by a large\nmargin. Code will be available at https://github.com/PKU-ML/nlb.', 'authors': ['Yifei Wang', 'Wenhan Ma', 'Stefanie Jegelka', 'Yisen Wang'], 'published': '2024-04-10 02:54:18+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CR'], 'url': 'http://arxiv.org/abs/2404.06694v2'}
42,345
2305.16174v2
2,305.16174
2
From Latent Graph to Latent Topology Inference: Differentiable Cell Complex Module
Latent Graph Inference (LGI) relaxed the reliance of Graph Neural Networks (GNNs) on a given graph topology by dynamically learning it. However, most of LGI methods assume to have a (noisy, incomplete, improvable, ...) input graph to rewire and can solely learn regular graph topologies. In the wake of the success of Topological Deep Learning (TDL), we study Latent Topology Inference (LTI) for learning higher-order cell complexes (with sparse and not regular topology) describing multi-way interactions between data points. To this aim, we introduce the Differentiable Cell Complex Module (DCM), a novel learnable function that computes cell probabilities in the complex to improve the downstream task. We show how to integrate DCM with cell complex message passing networks layers and train it in a end-to-end fashion, thanks to a two-step inference procedure that avoids an exhaustive search across all possible cells in the input, thus maintaining scalability. Our model is tested on several homophilic and heterophilic graph datasets and it is shown to outperform other state-of-the-art techniques, offering significant improvements especially in cases where an input graph is not provided.
2023-05-25
{'id': '2305.16174v2', 'title': 'From Latent Graph to Latent Topology Inference: Differentiable Cell Complex Module', 'abstract': 'Latent Graph Inference (LGI) relaxed the reliance of Graph Neural Networks\n(GNNs) on a given graph topology by dynamically learning it. However, most of\nLGI methods assume to have a (noisy, incomplete, improvable, ...) input graph\nto rewire and can solely learn regular graph topologies. In the wake of the\nsuccess of Topological Deep Learning (TDL), we study Latent Topology Inference\n(LTI) for learning higher-order cell complexes (with sparse and not regular\ntopology) describing multi-way interactions between data points. To this aim,\nwe introduce the Differentiable Cell Complex Module (DCM), a novel learnable\nfunction that computes cell probabilities in the complex to improve the\ndownstream task. We show how to integrate DCM with cell complex message passing\nnetworks layers and train it in a end-to-end fashion, thanks to a two-step\ninference procedure that avoids an exhaustive search across all possible cells\nin the input, thus maintaining scalability. Our model is tested on several\nhomophilic and heterophilic graph datasets and it is shown to outperform other\nstate-of-the-art techniques, offering significant improvements especially in\ncases where an input graph is not provided.', 'authors': ['Claudio Battiloro', 'Indro Spinelli', 'Lev Telyatnikov', 'Michael Bronstein', 'Simone Scardapane', 'Paolo Di Lorenzo'], 'published': '2023-05-25 15:33:19+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.NE'], 'url': 'http://arxiv.org/abs/2305.16174v2'}
42,347
2310.00724v3
2,310.00724
3
Subtractive Mixture Models via Squaring: Representation and Learning
Mixture models are traditionally represented and learned by adding several distributions as components. Allowing mixtures to subtract probability mass or density can drastically reduce the number of components needed to model complex distributions. However, learning such subtractive mixtures while ensuring they still encode a non-negative function is challenging. We investigate how to learn and perform inference on deep subtractive mixtures by squaring them. We do this in the framework of probabilistic circuits, which enable us to represent tensorized mixtures and generalize several other subtractive models. We theoretically prove that the class of squared circuits allowing subtractions can be exponentially more expressive than traditional additive mixtures; and, we empirically show this increased expressiveness on a series of real-world distribution estimation tasks.
2023-10-01
{'id': '2310.00724v3', 'title': 'Subtractive Mixture Models via Squaring: Representation and Learning', 'abstract': 'Mixture models are traditionally represented and learned by adding several\ndistributions as components. Allowing mixtures to subtract probability mass or\ndensity can drastically reduce the number of components needed to model complex\ndistributions. However, learning such subtractive mixtures while ensuring they\nstill encode a non-negative function is challenging. We investigate how to\nlearn and perform inference on deep subtractive mixtures by squaring them. We\ndo this in the framework of probabilistic circuits, which enable us to\nrepresent tensorized mixtures and generalize several other subtractive models.\nWe theoretically prove that the class of squared circuits allowing subtractions\ncan be exponentially more expressive than traditional additive mixtures; and,\nwe empirically show this increased expressiveness on a series of real-world\ndistribution estimation tasks.', 'authors': ['Lorenzo Loconte', 'Aleksanteri M. Sladek', 'Stefan Mengel', 'Martin Trapp', 'Arno Solin', 'Nicolas Gillis', 'Antonio Vergari'], 'published': '2023-10-01 16:51:58+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2310.00724v3'}
42,359
2310.01820v2
2,310.0182
2
Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks
Graph Neural Networks (GNNs) are neural models that leverage the dependency structure in graphical data via message passing among the graph nodes. GNNs have emerged as pivotal architectures in analyzing graph-structured data, and their expansive application in sensitive domains requires a comprehensive understanding of their decision-making processes -- necessitating a framework for GNN explainability. An explanation function for GNNs takes a pre-trained GNN along with a graph as input, to produce a `sufficient statistic' subgraph with respect to the graph label. A main challenge in studying GNN explainability is to provide fidelity measures that evaluate the performance of these explanation functions. This paper studies this foundational challenge, spotlighting the inherent limitations of prevailing fidelity metrics, including $Fid_+$, $Fid_-$, and $Fid_$. Specifically, a formal, information-theoretic definition of explainability is introduced and it is shown that existing metrics often fail to align with this definition across various statistical scenarios. The reason is due to potential distribution shifts when subgraphs are removed in computing these fidelity measures. Subsequently, a robust class of fidelity measures are introduced, and it is shown analytically that they are resilient to distribution shift issues and are applicable in a wide range of scenarios. Extensive empirical analysis on both synthetic and real datasets are provided to illustrate that the proposed metrics are more coherent with gold standard metrics. The source code is available at https://trustai4s-lab.github.io/fidelity.
2023-10-03
{'id': '2310.01820v2', 'title': 'Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks', 'abstract': "Graph Neural Networks (GNNs) are neural models that leverage the dependency\nstructure in graphical data via message passing among the graph nodes. GNNs\nhave emerged as pivotal architectures in analyzing graph-structured data, and\ntheir expansive application in sensitive domains requires a comprehensive\nunderstanding of their decision-making processes -- necessitating a framework\nfor GNN explainability. An explanation function for GNNs takes a pre-trained\nGNN along with a graph as input, to produce a `sufficient statistic' subgraph\nwith respect to the graph label. A main challenge in studying GNN\nexplainability is to provide fidelity measures that evaluate the performance of\nthese explanation functions. This paper studies this foundational challenge,\nspotlighting the inherent limitations of prevailing fidelity metrics, including\n$Fid_+$, $Fid_-$, and $Fid_\\Delta$. Specifically, a formal,\ninformation-theoretic definition of explainability is introduced and it is\nshown that existing metrics often fail to align with this definition across\nvarious statistical scenarios. The reason is due to potential distribution\nshifts when subgraphs are removed in computing these fidelity measures.\nSubsequently, a robust class of fidelity measures are introduced, and it is\nshown analytically that they are resilient to distribution shift issues and are\napplicable in a wide range of scenarios. Extensive empirical analysis on both\nsynthetic and real datasets are provided to illustrate that the proposed\nmetrics are more coherent with gold standard metrics. The source code is\navailable at https://trustai4s-lab.github.io/fidelity.", 'authors': ['Xu Zheng', 'Farhad Shirani', 'Tianchun Wang', 'Wei Cheng', 'Zhuomin Chen', 'Haifeng Chen', 'Hua Wei', 'Dongsheng Luo'], 'published': '2023-10-03 06:25:14+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.01820v2'}
42,366
2404.11599v1
2,404.11599
1
Variational Bayesian Last Layers
We introduce a deterministic variational formulation for training Bayesian last layer neural networks. This yields a sampling-free, single-pass model and loss that effectively improves uncertainty estimation. Our variational Bayesian last layer (VBLL) can be trained and evaluated with only quadratic complexity in last layer width, and is thus (nearly) computationally free to add to standard architectures. We experimentally investigate VBLLs, and show that they improve predictive accuracy, calibration, and out of distribution detection over baselines across both regression and classification. Finally, we investigate combining VBLL layers with variational Bayesian feature learning, yielding a lower variance collapsed variational inference method for Bayesian neural networks.
2024-04-17
{'id': '2404.11599v1', 'title': 'Variational Bayesian Last Layers', 'abstract': 'We introduce a deterministic variational formulation for training Bayesian\nlast layer neural networks. This yields a sampling-free, single-pass model and\nloss that effectively improves uncertainty estimation. Our variational Bayesian\nlast layer (VBLL) can be trained and evaluated with only quadratic complexity\nin last layer width, and is thus (nearly) computationally free to add to\nstandard architectures. We experimentally investigate VBLLs, and show that they\nimprove predictive accuracy, calibration, and out of distribution detection\nover baselines across both regression and classification. Finally, we\ninvestigate combining VBLL layers with variational Bayesian feature learning,\nyielding a lower variance collapsed variational inference method for Bayesian\nneural networks.', 'authors': ['James Harrison', 'John Willes', 'Jasper Snoek'], 'published': '2024-04-17 17:50:24+00:00', 'categories': ['cs.LG', 'cs.CV', 'stat.ML'], 'url': 'http://arxiv.org/abs/2404.11599v1'}
42,387
2309.02130v2
2,309.0213
2
Asymmetric Momentum: A Rethinking of Gradient Descent
Through theoretical and experimental validation, unlike all existing adaptive methods like Adam which penalize frequently-changing parameters and are only applicable to sparse gradients, we propose the simplest SGD enhanced method, Loss-Controlled Asymmetric Momentum(LCAM). By averaging the loss, we divide training process into different loss phases and using different momentum. It not only can accelerates slow-changing parameters for sparse gradients, similar to adaptive optimizers, but also can choose to accelerates frequently-changing parameters for non-sparse gradients, thus being adaptable to all types of datasets. We reinterpret the machine learning training process through the concepts of weight coupling and weight traction, and experimentally validate that weights have directional specificity, which are correlated with the specificity of the dataset. Thus interestingly, we observe that in non-sparse gradients, frequently-changing parameters should actually be accelerated, which is completely opposite to traditional adaptive perspectives. Compared to traditional SGD with momentum, this algorithm separates the weights without additional computational costs. It is noteworthy that this method relies on the network's ability to extract complex features. We primarily use Wide Residual Networks for our research, employing the classic datasets Cifar10 and Cifar100 to test the ability for feature separation and conclude phenomena that are much more important than just accuracy rates. Finally, compared to classic SGD tuning methods, while using WRN on these two datasets and with nearly half the training epochs, we achieve equal or better test accuracy.
2023-09-05
{'id': '2309.02130v2', 'title': 'Asymmetric Momentum: A Rethinking of Gradient Descent', 'abstract': "Through theoretical and experimental validation, unlike all existing adaptive\nmethods like Adam which penalize frequently-changing parameters and are only\napplicable to sparse gradients, we propose the simplest SGD enhanced method,\nLoss-Controlled Asymmetric Momentum(LCAM). By averaging the loss, we divide\ntraining process into different loss phases and using different momentum. It\nnot only can accelerates slow-changing parameters for sparse gradients, similar\nto adaptive optimizers, but also can choose to accelerates frequently-changing\nparameters for non-sparse gradients, thus being adaptable to all types of\ndatasets. We reinterpret the machine learning training process through the\nconcepts of weight coupling and weight traction, and experimentally validate\nthat weights have directional specificity, which are correlated with the\nspecificity of the dataset. Thus interestingly, we observe that in non-sparse\ngradients, frequently-changing parameters should actually be accelerated, which\nis completely opposite to traditional adaptive perspectives. Compared to\ntraditional SGD with momentum, this algorithm separates the weights without\nadditional computational costs. It is noteworthy that this method relies on the\nnetwork's ability to extract complex features. We primarily use Wide Residual\nNetworks for our research, employing the classic datasets Cifar10 and Cifar100\nto test the ability for feature separation and conclude phenomena that are much\nmore important than just accuracy rates. Finally, compared to classic SGD\ntuning methods, while using WRN on these two datasets and with nearly half the\ntraining epochs, we achieve equal or better test accuracy.", 'authors': ['Gongyue Zhang', 'Dinghuang Zhang', 'Shuwen Zhao', 'Donghan Liu', 'Carrie M. Toptan', 'Honghai Liu'], 'published': '2023-09-05 11:16:47+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2309.02130v2'}
42,355
2310.05236v1
2,310.05236
1
Enhancing Kernel Flexibility via Learning Asymmetric Locally-Adaptive Kernels
The lack of sufficient flexibility is the key bottleneck of kernel-based learning that relies on manually designed, pre-given, and non-trainable kernels. To enhance kernel flexibility, this paper introduces the concept of Locally-Adaptive-Bandwidths (LAB) as trainable parameters to enhance the Radial Basis Function (RBF) kernel, giving rise to the LAB RBF kernel. The parameters in LAB RBF kernels are data-dependent, and its number can increase with the dataset, allowing for better adaptation to diverse data patterns and enhancing the flexibility of the learned function. This newfound flexibility also brings challenges, particularly with regards to asymmetry and the need for an efficient learning algorithm. To address these challenges, this paper for the first time establishes an asymmetric kernel ridge regression framework and introduces an iterative kernel learning algorithm. This novel approach not only reduces the demand for extensive support data but also significantly improves generalization by training bandwidths on the available training data. Experimental results on real datasets underscore the remarkable performance of the proposed algorithm, showcasing its superior capability in handling large-scale datasets compared to Nystr\"om approximation-based algorithms. Moreover, it demonstrates a significant improvement in regression accuracy over existing kernel-based learning methods and even surpasses residual neural networks.
2023-10-08
{'id': '2310.05236v1', 'title': 'Enhancing Kernel Flexibility via Learning Asymmetric Locally-Adaptive Kernels', 'abstract': 'The lack of sufficient flexibility is the key bottleneck of kernel-based\nlearning that relies on manually designed, pre-given, and non-trainable\nkernels. To enhance kernel flexibility, this paper introduces the concept of\nLocally-Adaptive-Bandwidths (LAB) as trainable parameters to enhance the Radial\nBasis Function (RBF) kernel, giving rise to the LAB RBF kernel. The parameters\nin LAB RBF kernels are data-dependent, and its number can increase with the\ndataset, allowing for better adaptation to diverse data patterns and enhancing\nthe flexibility of the learned function. This newfound flexibility also brings\nchallenges, particularly with regards to asymmetry and the need for an\nefficient learning algorithm. To address these challenges, this paper for the\nfirst time establishes an asymmetric kernel ridge regression framework and\nintroduces an iterative kernel learning algorithm. This novel approach not only\nreduces the demand for extensive support data but also significantly improves\ngeneralization by training bandwidths on the available training data.\nExperimental results on real datasets underscore the remarkable performance of\nthe proposed algorithm, showcasing its superior capability in handling\nlarge-scale datasets compared to Nystr\\"om approximation-based algorithms.\nMoreover, it demonstrates a significant improvement in regression accuracy over\nexisting kernel-based learning methods and even surpasses residual neural\nnetworks.', 'authors': ['Fan He', 'Mingzhen He', 'Lei Shi', 'Xiaolin Huang', 'Johan A. K. Suykens'], 'published': '2023-10-08 17:08:15+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.05236v1'}
42,376
2310.14344v2
2,310.14344
2
What's in a Prior? Learned Proximal Networks for Inverse Problems
Proximal operators are ubiquitous in inverse problems, commonly appearing as part of algorithmic strategies to regularize problems that are otherwise ill-posed. Modern deep learning models have been brought to bear for these tasks too, as in the framework of plug-and-play or deep unrolling, where they loosely resemble proximal operators. Yet, something essential is lost in employing these purely data-driven approaches: there is no guarantee that a general deep network represents the proximal operator of any function, nor is there any characterization of the function for which the network might provide some approximate proximal. This not only makes guaranteeing convergence of iterative schemes challenging but, more fundamentally, complicates the analysis of what has been learned by these networks about their training data. Herein we provide a framework to develop learned proximal networks (LPN), prove that they provide exact proximal operators for a data-driven nonconvex regularizer, and show how a new training strategy, dubbed proximal matching, provably promotes the recovery of the log-prior of the true data distribution. Such LPN provide general, unsupervised, expressive proximal operators that can be used for general inverse problems with convergence guarantees. We illustrate our results in a series of cases of increasing complexity, demonstrating that these models not only result in state-of-the-art performance, but provide a window into the resulting priors learned from data.
2023-10-22
{'id': '2310.14344v2', 'title': "What's in a Prior? Learned Proximal Networks for Inverse Problems", 'abstract': 'Proximal operators are ubiquitous in inverse problems, commonly appearing as\npart of algorithmic strategies to regularize problems that are otherwise\nill-posed. Modern deep learning models have been brought to bear for these\ntasks too, as in the framework of plug-and-play or deep unrolling, where they\nloosely resemble proximal operators. Yet, something essential is lost in\nemploying these purely data-driven approaches: there is no guarantee that a\ngeneral deep network represents the proximal operator of any function, nor is\nthere any characterization of the function for which the network might provide\nsome approximate proximal. This not only makes guaranteeing convergence of\niterative schemes challenging but, more fundamentally, complicates the analysis\nof what has been learned by these networks about their training data. Herein we\nprovide a framework to develop learned proximal networks (LPN), prove that they\nprovide exact proximal operators for a data-driven nonconvex regularizer, and\nshow how a new training strategy, dubbed proximal matching, provably promotes\nthe recovery of the log-prior of the true data distribution. Such LPN provide\ngeneral, unsupervised, expressive proximal operators that can be used for\ngeneral inverse problems with convergence guarantees. We illustrate our results\nin a series of cases of increasing complexity, demonstrating that these models\nnot only result in state-of-the-art performance, but provide a window into the\nresulting priors learned from data.', 'authors': ['Zhenghan Fang', 'Sam Buchanan', 'Jeremias Sulam'], 'published': '2023-10-22 16:31:01+00:00', 'categories': ['cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2310.14344v2'}
42,378
2307.01715v3
2,307.01715
3
Align With Purpose: Optimize Desired Properties in CTC Models with a General Plug-and-Play Framework
Connectionist Temporal Classification (CTC) is a widely used criterion for training supervised sequence-to-sequence (seq2seq) models. It enables learning the relations between input and output sequences, termed alignments, by marginalizing over perfect alignments (that yield the ground truth), at the expense of imperfect alignments. This binary differentiation of perfect and imperfect alignments falls short of capturing other essential alignment properties that hold significance in other real-world applications. Here we propose $Align With Purpose$, a $general Plug-and-Play framework$ for enhancing a desired property in models trained with the CTC criterion. We do that by complementing the CTC with an additional loss term that prioritizes alignments according to a desired property. Our method does not require any intervention in the CTC loss function, enables easy optimization of a variety of properties, and allows differentiation between both perfect and imperfect alignments. We apply our framework in the domain of Automatic Speech Recognition (ASR) and show its generality in terms of property selection, architectural choice, and scale of training dataset (up to 280,000 hours). To demonstrate the effectiveness of our framework, we apply it to two unrelated properties: emission time and word error rate (WER). For the former, we report an improvement of up to 570ms in latency optimization with a minor reduction in WER, and for the latter, we report a relative improvement of 4.5% WER over the baseline models. To the best of our knowledge, these applications have never been demonstrated to work on a scale of data as large as ours. Notably, our method can be implemented using only a few lines of code, and can be extended to other alignment-free loss functions and to domains other than ASR.
2023-07-04
{'id': '2307.01715v3', 'title': 'Align With Purpose: Optimize Desired Properties in CTC Models with a General Plug-and-Play Framework', 'abstract': 'Connectionist Temporal Classification (CTC) is a widely used criterion for\ntraining supervised sequence-to-sequence (seq2seq) models. It enables learning\nthe relations between input and output sequences, termed alignments, by\nmarginalizing over perfect alignments (that yield the ground truth), at the\nexpense of imperfect alignments. This binary differentiation of perfect and\nimperfect alignments falls short of capturing other essential alignment\nproperties that hold significance in other real-world applications. Here we\npropose $\\textit{Align With Purpose}$, a $\\textbf{general Plug-and-Play\nframework}$ for enhancing a desired property in models trained with the CTC\ncriterion. We do that by complementing the CTC with an additional loss term\nthat prioritizes alignments according to a desired property. Our method does\nnot require any intervention in the CTC loss function, enables easy\noptimization of a variety of properties, and allows differentiation between\nboth perfect and imperfect alignments. We apply our framework in the domain of\nAutomatic Speech Recognition (ASR) and show its generality in terms of property\nselection, architectural choice, and scale of training dataset (up to 280,000\nhours). To demonstrate the effectiveness of our framework, we apply it to two\nunrelated properties: emission time and word error rate (WER). For the former,\nwe report an improvement of up to 570ms in latency optimization with a minor\nreduction in WER, and for the latter, we report a relative improvement of 4.5%\nWER over the baseline models. To the best of our knowledge, these applications\nhave never been demonstrated to work on a scale of data as large as ours.\nNotably, our method can be implemented using only a few lines of code, and can\nbe extended to other alignment-free loss functions and to domains other than\nASR.', 'authors': ['Eliya Segev', 'Maya Alroy', 'Ronen Katsir', 'Noam Wies', 'Ayana Shenhav', 'Yael Ben-Oren', 'David Zar', 'Oren Tadmor', 'Jacob Bitterman', 'Amnon Shashua', 'Tal Rosenwein'], 'published': '2023-07-04 13:34:47+00:00', 'categories': ['cs.CL', 'cs.LG', 'cs.SD', 'eess.AS'], 'url': 'http://arxiv.org/abs/2307.01715v3'}
42,384
2308.13234v3
2,308.13234
3
Decoding Natural Images from EEG for Object Recognition
Electroencephalography (EEG) signals, known for convenient non-invasive acquisition but low signal-to-noise ratio, have recently gained substantial attention due to the potential to decode natural images. This paper presents a self-supervised framework to demonstrate the feasibility of learning image representations from EEG signals, particularly for object recognition. The framework utilizes image and EEG encoders to extract features from paired image stimuli and EEG responses. Contrastive learning aligns these two modalities by constraining their similarity. With the framework, we attain significantly above-chance results on a comprehensive EEG-image dataset, achieving a top-1 accuracy of 15.6% and a top-5 accuracy of 42.8% in challenging 200-way zero-shot tasks. Moreover, we perform extensive experiments to explore the biological plausibility by resolving the temporal, spatial, spectral, and semantic aspects of EEG signals. Besides, we introduce attention modules to capture spatial correlations, providing implicit evidence of the brain activity perceived from EEG data. These findings yield valuable insights for neural decoding and brain-computer interfaces in real-world scenarios. The code will be released on https://github.com/eeyhsong/NICE-EEG.
2023-08-25
{'id': '2308.13234v3', 'title': 'Decoding Natural Images from EEG for Object Recognition', 'abstract': 'Electroencephalography (EEG) signals, known for convenient non-invasive\nacquisition but low signal-to-noise ratio, have recently gained substantial\nattention due to the potential to decode natural images. This paper presents a\nself-supervised framework to demonstrate the feasibility of learning image\nrepresentations from EEG signals, particularly for object recognition. The\nframework utilizes image and EEG encoders to extract features from paired image\nstimuli and EEG responses. Contrastive learning aligns these two modalities by\nconstraining their similarity. With the framework, we attain significantly\nabove-chance results on a comprehensive EEG-image dataset, achieving a top-1\naccuracy of 15.6% and a top-5 accuracy of 42.8% in challenging 200-way\nzero-shot tasks. Moreover, we perform extensive experiments to explore the\nbiological plausibility by resolving the temporal, spatial, spectral, and\nsemantic aspects of EEG signals. Besides, we introduce attention modules to\ncapture spatial correlations, providing implicit evidence of the brain activity\nperceived from EEG data. These findings yield valuable insights for neural\ndecoding and brain-computer interfaces in real-world scenarios. The code will\nbe released on https://github.com/eeyhsong/NICE-EEG.', 'authors': ['Yonghao Song', 'Bingchuan Liu', 'Xiang Li', 'Nanlin Shi', 'Yijun Wang', 'Xiaorong Gao'], 'published': '2023-08-25 08:05:37+00:00', 'categories': ['cs.HC', 'cs.AI', 'eess.SP', 'q-bio.NC'], 'url': 'http://arxiv.org/abs/2308.13234v3'}
42,394
2310.02129v5
2,310.02129
5
Unveiling the Pitfalls of Knowledge Editing for Large Language Models
As the cost associated with fine-tuning Large Language Models (LLMs) continues to rise, recent research efforts have pivoted towards developing methodologies to edit implicit knowledge embedded within LLMs. Yet, there's still a dark cloud lingering overhead -- will knowledge editing trigger butterfly effect? since it is still unclear whether knowledge editing might introduce side effects that pose potential risks or not. This paper pioneers the investigation into the potential pitfalls associated with knowledge editing for LLMs. To achieve this, we introduce new benchmark datasets and propose innovative evaluation metrics. Our results underline two pivotal concerns: (1) Knowledge Conflict: Editing groups of facts that logically clash can magnify the inherent inconsistencies in LLMs-a facet neglected by previous methods. (2) Knowledge Distortion: Altering parameters with the aim of editing factual knowledge can irrevocably warp the innate knowledge structure of LLMs. Experimental results vividly demonstrate that knowledge editing might inadvertently cast a shadow of unintended consequences on LLMs, which warrant attention and efforts for future works. Code and data are available at https://github.com/zjunlp/PitfallsKnowledgeEditing.
2023-10-03
{'id': '2310.02129v5', 'title': 'Unveiling the Pitfalls of Knowledge Editing for Large Language Models', 'abstract': "As the cost associated with fine-tuning Large Language Models (LLMs)\ncontinues to rise, recent research efforts have pivoted towards developing\nmethodologies to edit implicit knowledge embedded within LLMs. Yet, there's\nstill a dark cloud lingering overhead -- will knowledge editing trigger\nbutterfly effect? since it is still unclear whether knowledge editing might\nintroduce side effects that pose potential risks or not. This paper pioneers\nthe investigation into the potential pitfalls associated with knowledge editing\nfor LLMs. To achieve this, we introduce new benchmark datasets and propose\ninnovative evaluation metrics. Our results underline two pivotal concerns: (1)\nKnowledge Conflict: Editing groups of facts that logically clash can magnify\nthe inherent inconsistencies in LLMs-a facet neglected by previous methods. (2)\nKnowledge Distortion: Altering parameters with the aim of editing factual\nknowledge can irrevocably warp the innate knowledge structure of LLMs.\nExperimental results vividly demonstrate that knowledge editing might\ninadvertently cast a shadow of unintended consequences on LLMs, which warrant\nattention and efforts for future works. Code and data are available at\nhttps://github.com/zjunlp/PitfallsKnowledgeEditing.", 'authors': ['Zhoubo Li', 'Ningyu Zhang', 'Yunzhi Yao', 'Mengru Wang', 'Xi Chen', 'Huajun Chen'], 'published': '2023-10-03 15:10:46+00:00', 'categories': ['cs.CL', 'cs.AI', 'cs.CV', 'cs.DB', 'cs.LG'], 'url': 'http://arxiv.org/abs/2310.02129v5'}
42,358
2006.09994v1
2,006.09994
1
Noise or Signal: The Role of Image Backgrounds in Object Recognition
We assess the tendency of state-of-the-art object recognition models to depend on signals from image backgrounds. We create a toolkit for disentangling foreground and background signal on ImageNet images, and find that (a) models can achieve non-trivial accuracy by relying on the background alone, (b) models often misclassify images even in the presence of correctly classified foregrounds--up to 87.5% of the time with adversarially chosen backgrounds, and (c) more accurate models tend to depend on backgrounds less. Our analysis of backgrounds brings us closer to understanding which correlations machine learning models use, and how they determine models' out of distribution performance.
2020-06-17
{'id': '2006.09994v1', 'title': 'Noise or Signal: The Role of Image Backgrounds in Object Recognition', 'abstract': "We assess the tendency of state-of-the-art object recognition models to\ndepend on signals from image backgrounds. We create a toolkit for disentangling\nforeground and background signal on ImageNet images, and find that (a) models\ncan achieve non-trivial accuracy by relying on the background alone, (b) models\noften misclassify images even in the presence of correctly classified\nforegrounds--up to 87.5% of the time with adversarially chosen backgrounds, and\n(c) more accurate models tend to depend on backgrounds less. Our analysis of\nbackgrounds brings us closer to understanding which correlations machine\nlearning models use, and how they determine models' out of distribution\nperformance.", 'authors': ['Kai Xiao', 'Logan Engstrom', 'Andrew Ilyas', 'Aleksander Madry'], 'published': '2020-06-17 16:54:43+00:00', 'categories': ['cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2006.09994v1'}
42,372
2101.08543v2
2,101.08543
2
Boost then Convolve: Gradient Boosting Meets Graph Neural Networks
Graph neural networks (GNNs) are powerful models that have been successful in various graph representation learning tasks. Whereas gradient boosted decision trees (GBDT) often outperform other machine learning methods when faced with heterogeneous tabular data. But what approach should be used for graphs with tabular node features? Previous GNN models have mostly focused on networks with homogeneous sparse features and, as we show, are suboptimal in the heterogeneous setting. In this work, we propose a novel architecture that trains GBDT and GNN jointly to get the best of both worlds: the GBDT model deals with heterogeneous features, while GNN accounts for the graph structure. Our model benefits from end-to-end optimization by allowing new trees to fit the gradient updates of GNN. With an extensive experimental comparison to the leading GBDT and GNN models, we demonstrate a significant increase in performance on a variety of graphs with tabular features. The code is available: https://github.com/nd7141/bgnn.
2021-01-21
{'id': '2101.08543v2', 'title': 'Boost then Convolve: Gradient Boosting Meets Graph Neural Networks', 'abstract': 'Graph neural networks (GNNs) are powerful models that have been successful in\nvarious graph representation learning tasks. Whereas gradient boosted decision\ntrees (GBDT) often outperform other machine learning methods when faced with\nheterogeneous tabular data. But what approach should be used for graphs with\ntabular node features? Previous GNN models have mostly focused on networks with\nhomogeneous sparse features and, as we show, are suboptimal in the\nheterogeneous setting. In this work, we propose a novel architecture that\ntrains GBDT and GNN jointly to get the best of both worlds: the GBDT model\ndeals with heterogeneous features, while GNN accounts for the graph structure.\nOur model benefits from end-to-end optimization by allowing new trees to fit\nthe gradient updates of GNN. With an extensive experimental comparison to the\nleading GBDT and GNN models, we demonstrate a significant increase in\nperformance on a variety of graphs with tabular features. The code is\navailable: https://github.com/nd7141/bgnn.', 'authors': ['Sergei Ivanov', 'Liudmila Prokhorenkova'], 'published': '2021-01-21 10:46:41+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.SI'], 'url': 'http://arxiv.org/abs/2101.08543v2'}
42,373
2006.07409v1
2,006.07409
1
How to Avoid Being Eaten by a Grue: Structured Exploration Strategies for Textual Worlds
Text-based games are long puzzles or quests, characterized by a sequence of sparse and potentially deceptive rewards. They provide an ideal platform to develop agents that perceive and act upon the world using a combinatorially sized natural language state-action space. Standard Reinforcement Learning agents are poorly equipped to effectively explore such spaces and often struggle to overcome bottlenecks---states that agents are unable to pass through simply because they do not see the right action sequence enough times to be sufficiently reinforced. We introduce Q*BERT, an agent that learns to build a knowledge graph of the world by answering questions, which leads to greater sample efficiency. To overcome bottlenecks, we further introduce MC!Q*BERT an agent that uses an knowledge-graph-based intrinsic motivation to detect bottlenecks and a novel exploration strategy to efficiently learn a chain of policy modules to overcome them. We present an ablation study and results demonstrating how our method outperforms the current state-of-the-art on nine text games, including the popular game, Zork, where, for the first time, a learning agent gets past the bottleneck where the player is eaten by a Grue.
2020-06-12
{'id': '2006.07409v1', 'title': 'How to Avoid Being Eaten by a Grue: Structured Exploration Strategies for Textual Worlds', 'abstract': 'Text-based games are long puzzles or quests, characterized by a sequence of\nsparse and potentially deceptive rewards. They provide an ideal platform to\ndevelop agents that perceive and act upon the world using a combinatorially\nsized natural language state-action space. Standard Reinforcement Learning\nagents are poorly equipped to effectively explore such spaces and often\nstruggle to overcome bottlenecks---states that agents are unable to pass\nthrough simply because they do not see the right action sequence enough times\nto be sufficiently reinforced. We introduce Q*BERT, an agent that learns to\nbuild a knowledge graph of the world by answering questions, which leads to\ngreater sample efficiency. To overcome bottlenecks, we further introduce\nMC!Q*BERT an agent that uses an knowledge-graph-based intrinsic motivation to\ndetect bottlenecks and a novel exploration strategy to efficiently learn a\nchain of policy modules to overcome them. We present an ablation study and\nresults demonstrating how our method outperforms the current state-of-the-art\non nine text games, including the popular game, Zork, where, for the first\ntime, a learning agent gets past the bottleneck where the player is eaten by a\nGrue.', 'authors': ['Prithviraj Ammanabrolu', 'Ethan Tien', 'Matthew Hausknecht', 'Mark O. Riedl'], 'published': '2020-06-12 18:24:06+00:00', 'categories': ['cs.AI', 'cs.CL', 'cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2006.07409v1'}
42,374
2006.12800v2
2,006.128
2
Calibration of Neural Networks using Splines
Calibrating neural networks is of utmost importance when employing them in safety-critical applications where the downstream decision making depends on the predicted probabilities. Measuring calibration error amounts to comparing two empirical distributions. In this work, we introduce a binning-free calibration measure inspired by the classical Kolmogorov-Smirnov (KS) statistical test in which the main idea is to compare the respective cumulative probability distributions. From this, by approximating the empirical cumulative distribution using a differentiable function via splines, we obtain a recalibration function, which maps the network outputs to actual (calibrated) class assignment probabilities. The spine-fitting is performed using a held-out calibration set and the obtained recalibration function is evaluated on an unseen test set. We tested our method against existing calibration approaches on various image classification datasets and our spline-based recalibration approach consistently outperforms existing methods on KS error as well as other commonly used calibration measures. Our Code is available at https://github.com/kartikgupta-at-anu/spline-calibration.
2020-06-23
{'id': '2006.12800v2', 'title': 'Calibration of Neural Networks using Splines', 'abstract': 'Calibrating neural networks is of utmost importance when employing them in\nsafety-critical applications where the downstream decision making depends on\nthe predicted probabilities. Measuring calibration error amounts to comparing\ntwo empirical distributions. In this work, we introduce a binning-free\ncalibration measure inspired by the classical Kolmogorov-Smirnov (KS)\nstatistical test in which the main idea is to compare the respective cumulative\nprobability distributions. From this, by approximating the empirical cumulative\ndistribution using a differentiable function via splines, we obtain a\nrecalibration function, which maps the network outputs to actual (calibrated)\nclass assignment probabilities. The spine-fitting is performed using a held-out\ncalibration set and the obtained recalibration function is evaluated on an\nunseen test set. We tested our method against existing calibration approaches\non various image classification datasets and our spline-based recalibration\napproach consistently outperforms existing methods on KS error as well as other\ncommonly used calibration measures. Our Code is available at\nhttps://github.com/kartikgupta-at-anu/spline-calibration.', 'authors': ['Kartik Gupta', 'Amir Rahimi', 'Thalaiyasingam Ajanthan', 'Thomas Mensink', 'Cristian Sminchisescu', 'Richard Hartley'], 'published': '2020-06-23 07:18:05+00:00', 'categories': ['cs.LG', 'cs.CV', 'stat.ML'], 'url': 'http://arxiv.org/abs/2006.12800v2'}
42,379
2010.02075v1
2,010.02075
1
Learned Hardware/Software Co-Design of Neural Accelerators
The use of deep learning has grown at an exponential rate, giving rise to numerous specialized hardware and software systems for deep learning. Because the design space of deep learning software stacks and hardware accelerators is diverse and vast, prior work considers software optimizations separately from hardware architectures, effectively reducing the search space. Unfortunately, this bifurcated approach means that many profitable design points are never explored. This paper instead casts the problem as hardware/software co-design, with the goal of automatically identifying desirable points in the joint design space. The key to our solution is a new constrained Bayesian optimization framework that avoids invalid solutions by exploiting the highly constrained features of this design space, which are semi-continuous/semi-discrete. We evaluate our optimization framework by applying it to a variety of neural models, improving the energy-delay product by 18% (ResNet) and 40% (DQN) over hand-tuned state-of-the-art systems, as well as demonstrating strong results on other neural network architectures, such as MLPs and Transformers.
2020-10-05
{'id': '2010.02075v1', 'title': 'Learned Hardware/Software Co-Design of Neural Accelerators', 'abstract': 'The use of deep learning has grown at an exponential rate, giving rise to\nnumerous specialized hardware and software systems for deep learning. Because\nthe design space of deep learning software stacks and hardware accelerators is\ndiverse and vast, prior work considers software optimizations separately from\nhardware architectures, effectively reducing the search space. Unfortunately,\nthis bifurcated approach means that many profitable design points are never\nexplored. This paper instead casts the problem as hardware/software co-design,\nwith the goal of automatically identifying desirable points in the joint design\nspace. The key to our solution is a new constrained Bayesian optimization\nframework that avoids invalid solutions by exploiting the highly constrained\nfeatures of this design space, which are semi-continuous/semi-discrete. We\nevaluate our optimization framework by applying it to a variety of neural\nmodels, improving the energy-delay product by 18% (ResNet) and 40% (DQN) over\nhand-tuned state-of-the-art systems, as well as demonstrating strong results on\nother neural network architectures, such as MLPs and Transformers.', 'authors': ['Zhan Shi', 'Chirag Sakhuja', 'Milad Hashemi', 'Kevin Swersky', 'Calvin Lin'], 'published': '2020-10-05 15:12:52+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.AR', 'stat.ML'], 'url': 'http://arxiv.org/abs/2010.02075v1'}
42,398
2401.12046v2
2,401.12046
2
Fourier Transporter: Bi-Equivariant Robotic Manipulation in 3D
Many complex robotic manipulation tasks can be decomposed as a sequence of pick and place actions. Training a robotic agent to learn this sequence over many different starting conditions typically requires many iterations or demonstrations, especially in 3D environments. In this work, we propose Fourier Transporter (FourTran) which leverages the two-fold SE(d)xSE(d) symmetry in the pick-place problem to achieve much higher sample efficiency. FourTran is an open-loop behavior cloning method trained using expert demonstrations to predict pick-place actions on new environments. FourTran is constrained to incorporate symmetries of the pick and place actions independently. Our method utilizes a fiber space Fourier transformation that allows for memory-efficient construction. We test our proposed network on the RLbench benchmark and achieve state-of-the-art results across various tasks.
2024-01-22
{'id': '2401.12046v2', 'title': 'Fourier Transporter: Bi-Equivariant Robotic Manipulation in 3D', 'abstract': 'Many complex robotic manipulation tasks can be decomposed as a sequence of\npick and place actions. Training a robotic agent to learn this sequence over\nmany different starting conditions typically requires many iterations or\ndemonstrations, especially in 3D environments. In this work, we propose Fourier\nTransporter (FourTran) which leverages the two-fold SE(d)xSE(d) symmetry in the\npick-place problem to achieve much higher sample efficiency. FourTran is an\nopen-loop behavior cloning method trained using expert demonstrations to\npredict pick-place actions on new environments. FourTran is constrained to\nincorporate symmetries of the pick and place actions independently. Our method\nutilizes a fiber space Fourier transformation that allows for memory-efficient\nconstruction. We test our proposed network on the RLbench benchmark and achieve\nstate-of-the-art results across various tasks.', 'authors': ['Haojie Huang', 'Owen Howell', 'Dian Wang', 'Xupeng Zhu', 'Robin Walters', 'Robert Platt'], 'published': '2024-01-22 15:38:29+00:00', 'categories': ['cs.RO', 'cs.LG'], 'url': 'http://arxiv.org/abs/2401.12046v2'}
42,364
2305.13404v3
2,305.13404
3
Improving Convergence and Generalization Using Parameter Symmetries
In many neural networks, different values of the parameters may result in the same loss value. Parameter space symmetries are loss-invariant transformations that change the model parameters. Teleportation applies such transformations to accelerate optimization. However, the exact mechanism behind this algorithm's success is not well understood. In this paper, we show that teleportation not only speeds up optimization in the short-term, but gives overall faster time to convergence. Additionally, teleporting to minima with different curvatures improves generalization, which suggests a connection between the curvature of the minimum and generalization ability. Finally, we show that integrating teleportation into a wide range of optimization algorithms and optimization-based meta-learning improves convergence. Our results showcase the versatility of teleportation and demonstrate the potential of incorporating symmetry in optimization.
2023-05-22
{'id': '2305.13404v3', 'title': 'Improving Convergence and Generalization Using Parameter Symmetries', 'abstract': "In many neural networks, different values of the parameters may result in the\nsame loss value. Parameter space symmetries are loss-invariant transformations\nthat change the model parameters. Teleportation applies such transformations to\naccelerate optimization. However, the exact mechanism behind this algorithm's\nsuccess is not well understood. In this paper, we show that teleportation not\nonly speeds up optimization in the short-term, but gives overall faster time to\nconvergence. Additionally, teleporting to minima with different curvatures\nimproves generalization, which suggests a connection between the curvature of\nthe minimum and generalization ability. Finally, we show that integrating\nteleportation into a wide range of optimization algorithms and\noptimization-based meta-learning improves convergence. Our results showcase the\nversatility of teleportation and demonstrate the potential of incorporating\nsymmetry in optimization.", 'authors': ['Bo Zhao', 'Robert M. Gower', 'Robin Walters', 'Rose Yu'], 'published': '2023-05-22 18:35:42+00:00', 'categories': ['cs.LG', 'math.OC'], 'url': 'http://arxiv.org/abs/2305.13404v3'}
42,367
2305.03923v2
2,305.03923
2
Active Continual Learning: On Balancing Knowledge Retention and Learnability
Acquiring new knowledge without forgetting what has been learned in a sequence of tasks is the central focus of continual learning (CL). While tasks arrive sequentially, the training data are often prepared and annotated independently, leading to the CL of incoming supervised learning tasks. This paper considers the under-explored problem of active continual learning (ACL) for a sequence of active learning (AL) tasks, where each incoming task includes a pool of unlabelled data and an annotation budget. We investigate the effectiveness and interplay between several AL and CL algorithms in the domain, class and task-incremental scenarios. Our experiments reveal the trade-off between two contrasting goals of not forgetting the old knowledge and the ability to quickly learn new knowledge in CL and AL, respectively. While conditioning the AL query strategy on the annotations collected for the previous tasks leads to improved task performance on the domain and task incremental learning, our proposed forgetting-learning profile suggests a gap in balancing the effect of AL and CL for the class-incremental scenario.
2023-05-06
{'id': '2305.03923v2', 'title': 'Active Continual Learning: On Balancing Knowledge Retention and Learnability', 'abstract': 'Acquiring new knowledge without forgetting what has been learned in a\nsequence of tasks is the central focus of continual learning (CL). While tasks\narrive sequentially, the training data are often prepared and annotated\nindependently, leading to the CL of incoming supervised learning tasks. This\npaper considers the under-explored problem of active continual learning (ACL)\nfor a sequence of active learning (AL) tasks, where each incoming task includes\na pool of unlabelled data and an annotation budget. We investigate the\neffectiveness and interplay between several AL and CL algorithms in the domain,\nclass and task-incremental scenarios. Our experiments reveal the trade-off\nbetween two contrasting goals of not forgetting the old knowledge and the\nability to quickly learn new knowledge in CL and AL, respectively. While\nconditioning the AL query strategy on the annotations collected for the\nprevious tasks leads to improved task performance on the domain and task\nincremental learning, our proposed forgetting-learning profile suggests a gap\nin balancing the effect of AL and CL for the class-incremental scenario.', 'authors': ['Thuy-Trang Vu', 'Shahram Khadivi', 'Mahsa Ghorbanali', 'Dinh Phung', 'Gholamreza Haffari'], 'published': '2023-05-06 04:11:03+00:00', 'categories': ['cs.LG', 'cs.CL'], 'url': 'http://arxiv.org/abs/2305.03923v2'}
42,388
2310.18144v4
2,310.18144
4
Improving Intrinsic Exploration by Creating Stationary Objectives
Exploration bonuses in reinforcement learning guide long-horizon exploration by defining custom intrinsic objectives. Several exploration objectives like count-based bonuses, pseudo-counts, and state-entropy maximization are non-stationary and hence are difficult to optimize for the agent. While this issue is generally known, it is usually omitted and solutions remain under-explored. The key contribution of our work lies in transforming the original non-stationary rewards into stationary rewards through an augmented state representation. For this purpose, we introduce the Stationary Objectives For Exploration (SOFE) framework. SOFE requires identifying sufficient statistics for different exploration bonuses and finding an efficient encoding of these statistics to use as input to a deep network. SOFE is based on proposing state augmentations that expand the state space but hold the promise of simplifying the optimization of the agent's objective. We show that SOFE improves the performance of several exploration objectives, including count-based bonuses, pseudo-counts, and state-entropy maximization. Moreover, SOFE outperforms prior methods that attempt to stabilize the optimization of intrinsic objectives. We demonstrate the efficacy of SOFE in hard-exploration problems, including sparse-reward tasks, pixel-based observations, 3D navigation, and procedurally generated environments.
2023-10-27
{'id': '2310.18144v4', 'title': 'Improving Intrinsic Exploration by Creating Stationary Objectives', 'abstract': "Exploration bonuses in reinforcement learning guide long-horizon exploration\nby defining custom intrinsic objectives. Several exploration objectives like\ncount-based bonuses, pseudo-counts, and state-entropy maximization are\nnon-stationary and hence are difficult to optimize for the agent. While this\nissue is generally known, it is usually omitted and solutions remain\nunder-explored. The key contribution of our work lies in transforming the\noriginal non-stationary rewards into stationary rewards through an augmented\nstate representation. For this purpose, we introduce the Stationary Objectives\nFor Exploration (SOFE) framework. SOFE requires identifying sufficient\nstatistics for different exploration bonuses and finding an efficient encoding\nof these statistics to use as input to a deep network. SOFE is based on\nproposing state augmentations that expand the state space but hold the promise\nof simplifying the optimization of the agent's objective. We show that SOFE\nimproves the performance of several exploration objectives, including\ncount-based bonuses, pseudo-counts, and state-entropy maximization. Moreover,\nSOFE outperforms prior methods that attempt to stabilize the optimization of\nintrinsic objectives. We demonstrate the efficacy of SOFE in hard-exploration\nproblems, including sparse-reward tasks, pixel-based observations, 3D\nnavigation, and procedurally generated environments.", 'authors': ['Roger Creus Castanyer', 'Joshua Romoff', 'Glen Berseth'], 'published': '2023-10-27 13:51:18+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2310.18144v4'}
42,415
2310.06205v1
2,310.06205
1
Fair Classifiers that Abstain without Harm
In critical applications, it is vital for classifiers to defer decision-making to humans. We propose a post-hoc method that makes existing classifiers selectively abstain from predicting certain samples. Our abstaining classifier is incentivized to maintain the original accuracy for each sub-population (i.e. no harm) while achieving a set of group fairness definitions to a user specified degree. To this end, we design an Integer Programming (IP) procedure that assigns abstention decisions for each training sample to satisfy a set of constraints. To generalize the abstaining decisions to test samples, we then train a surrogate model to learn the abstaining decisions based on the IP solutions in an end-to-end manner. We analyze the feasibility of the IP procedure to determine the possible abstention rate for different levels of unfairness tolerance and accuracy constraint for achieving no harm. To the best of our knowledge, this work is the first to identify the theoretical relationships between the constraint parameters and the required abstention rate. Our theoretical results are important since a high abstention rate is often infeasible in practice due to a lack of human resources. Our framework outperforms existing methods in terms of fairness disparity without sacrificing accuracy at similar abstention rates.
2023-10-09
{'id': '2310.06205v1', 'title': 'Fair Classifiers that Abstain without Harm', 'abstract': 'In critical applications, it is vital for classifiers to defer\ndecision-making to humans. We propose a post-hoc method that makes existing\nclassifiers selectively abstain from predicting certain samples. Our abstaining\nclassifier is incentivized to maintain the original accuracy for each\nsub-population (i.e. no harm) while achieving a set of group fairness\ndefinitions to a user specified degree. To this end, we design an Integer\nProgramming (IP) procedure that assigns abstention decisions for each training\nsample to satisfy a set of constraints. To generalize the abstaining decisions\nto test samples, we then train a surrogate model to learn the abstaining\ndecisions based on the IP solutions in an end-to-end manner. We analyze the\nfeasibility of the IP procedure to determine the possible abstention rate for\ndifferent levels of unfairness tolerance and accuracy constraint for achieving\nno harm. To the best of our knowledge, this work is the first to identify the\ntheoretical relationships between the constraint parameters and the required\nabstention rate. Our theoretical results are important since a high abstention\nrate is often infeasible in practice due to a lack of human resources. Our\nframework outperforms existing methods in terms of fairness disparity without\nsacrificing accuracy at similar abstention rates.', 'authors': ['Tongxin Yin', 'Jean-François Ton', 'Ruocheng Guo', 'Yuanshun Yao', 'Mingyan Liu', 'Yang Liu'], 'published': '2023-10-09 23:07:28+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.06205v1'}
42,417
2401.07993v2
2,401.07993
2
Carrying over algorithm in transformers
Addition is perhaps one of the simplest arithmetic tasks one can think of and is usually performed using the carrying over algorithm. This algorithm consists of two tasks: adding digits in the same position and carrying over a one whenever necessary. We study how transformer models implement this algorithm and how the two aforementioned tasks are allocated to different parts of the network. We first focus on two-layer encoder-only models and show that the carrying over algorithm is implemented in a modular fashion. The first layer is mostly responsible for adding digits in the same position. The second layer first decides, in the attention, which positions need a carried one or not, and then performs the carrying of the one in the final MLP. We provide a simple way of precisely identifying which neurons are responsible for that task. This implementation of the carrying over algorithm occurs across a range of hyperparameters for two as well as three-layer models. For small decoder-only models, we observe the same implementation and provide suggestive evidence for its existence in three 7B large language models.
2024-01-15
{'id': '2401.07993v2', 'title': 'Carrying over algorithm in transformers', 'abstract': 'Addition is perhaps one of the simplest arithmetic tasks one can think of and\nis usually performed using the carrying over algorithm. This algorithm consists\nof two tasks: adding digits in the same position and carrying over a one\nwhenever necessary. We study how transformer models implement this algorithm\nand how the two aforementioned tasks are allocated to different parts of the\nnetwork. We first focus on two-layer encoder-only models and show that the\ncarrying over algorithm is implemented in a modular fashion. The first layer is\nmostly responsible for adding digits in the same position. The second layer\nfirst decides, in the attention, which positions need a carried one or not, and\nthen performs the carrying of the one in the final MLP. We provide a simple way\nof precisely identifying which neurons are responsible for that task. This\nimplementation of the carrying over algorithm occurs across a range of\nhyperparameters for two as well as three-layer models. For small decoder-only\nmodels, we observe the same implementation and provide suggestive evidence for\nits existence in three 7B large language models.', 'authors': ['Jorrit Kruthoff'], 'published': '2024-01-15 22:36:11+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2401.07993v2'}
42,422
2305.13293v2
2,305.13293
2
Time Fairness in Online Knapsack Problems
The online knapsack problem is a classic problem in the field of online algorithms. Its canonical version asks how to pack items of different values and weights arriving online into a capacity-limited knapsack so as to maximize the total value of the admitted items. Although optimal competitive algorithms are known for this problem, they may be fundamentally unfair, i.e., individual items may be treated inequitably in different ways. We formalize a practically-relevant notion of time fairness which effectively models a trade off between static and dynamic pricing in a motivating application such as cloud resource allocation, and show that existing algorithms perform poorly under this metric. We propose a parameterized deterministic algorithm where the parameter precisely captures the Pareto-optimal trade-off between fairness (static pricing) and competitiveness (dynamic pricing). We show that randomization is theoretically powerful enough to be simultaneously competitive and fair; however, it does not work well in experiments. To further improve the trade-off between fairness and competitiveness, we develop a nearly-optimal learning-augmented algorithm which is fair, consistent, and robust (competitive), showing substantial performance improvements in numerical experiments.
2023-05-22
{'id': '2305.13293v2', 'title': 'Time Fairness in Online Knapsack Problems', 'abstract': 'The online knapsack problem is a classic problem in the field of online\nalgorithms. Its canonical version asks how to pack items of different values\nand weights arriving online into a capacity-limited knapsack so as to maximize\nthe total value of the admitted items. Although optimal competitive algorithms\nare known for this problem, they may be fundamentally unfair, i.e., individual\nitems may be treated inequitably in different ways. We formalize a\npractically-relevant notion of time fairness which effectively models a trade\noff between static and dynamic pricing in a motivating application such as\ncloud resource allocation, and show that existing algorithms perform poorly\nunder this metric. We propose a parameterized deterministic algorithm where the\nparameter precisely captures the Pareto-optimal trade-off between fairness\n(static pricing) and competitiveness (dynamic pricing). We show that\nrandomization is theoretically powerful enough to be simultaneously competitive\nand fair; however, it does not work well in experiments. To further improve the\ntrade-off between fairness and competitiveness, we develop a nearly-optimal\nlearning-augmented algorithm which is fair, consistent, and robust\n(competitive), showing substantial performance improvements in numerical\nexperiments.', 'authors': ['Adam Lechowicz', 'Rik Sengupta', 'Bo Sun', 'Shahin Kamali', 'Mohammad Hajiesmaili'], 'published': '2023-05-22 17:51:35+00:00', 'categories': ['cs.LG', 'cs.CY', 'cs.DS'], 'url': 'http://arxiv.org/abs/2305.13293v2'}
42,369
2303.00442v1
2,303.00442
1
Re-weighting Based Group Fairness Regularization via Classwise Robust Optimization
Many existing group fairness-aware training methods aim to achieve the group fairness by either re-weighting underrepresented groups based on certain rules or using weakly approximated surrogates for the fairness metrics in the objective as regularization terms. Although each of the learning schemes has its own strength in terms of applicability or performance, respectively, it is difficult for any method in the either category to be considered as a gold standard since their successful performances are typically limited to specific cases. To that end, we propose a principled method, dubbed as , which unifies the two learning schemes by incorporating a well-justified group fairness metric into the training objective using a class wise distributionally robust optimization (DRO) framework. We then develop an iterative optimization algorithm that minimizes the resulting objective by automatically producing the correct re-weights for each group. Our experiments show that FairDRO is scalable and easily adaptable to diverse applications, and consistently achieves the state-of-the-art performance on several benchmark datasets in terms of the accuracy-fairness trade-off, compared to recent strong baselines.
2023-03-01
{'id': '2303.00442v1', 'title': 'Re-weighting Based Group Fairness Regularization via Classwise Robust Optimization', 'abstract': 'Many existing group fairness-aware training methods aim to achieve the group\nfairness by either re-weighting underrepresented groups based on certain rules\nor using weakly approximated surrogates for the fairness metrics in the\nobjective as regularization terms. Although each of the learning schemes has\nits own strength in terms of applicability or performance, respectively, it is\ndifficult for any method in the either category to be considered as a gold\nstandard since their successful performances are typically limited to specific\ncases. To that end, we propose a principled method, dubbed as \\ours, which\nunifies the two learning schemes by incorporating a well-justified group\nfairness metric into the training objective using a class wise distributionally\nrobust optimization (DRO) framework. We then develop an iterative optimization\nalgorithm that minimizes the resulting objective by automatically producing the\ncorrect re-weights for each group. Our experiments show that FairDRO is\nscalable and easily adaptable to diverse applications, and consistently\nachieves the state-of-the-art performance on several benchmark datasets in\nterms of the accuracy-fairness trade-off, compared to recent strong baselines.', 'authors': ['Sangwon Jung', 'Taeeon Park', 'Sanghyuk Chun', 'Taesup Moon'], 'published': '2023-03-01 12:00:37+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CY'], 'url': 'http://arxiv.org/abs/2303.00442v1'}
42,383
2303.04381v1
2,303.04381
1
Automatically Auditing Large Language Models via Discrete Optimization
Auditing large language models for unexpected behaviors is critical to preempt catastrophic deployments, yet remains challenging. In this work, we cast auditing as an optimization problem, where we automatically search for input-output pairs that match a desired target behavior. For example, we might aim to find a non-toxic input that starts with "Barack Obama" that a model maps to a toxic output. This optimization problem is difficult to solve as the set of feasible points is sparse, the space is discrete, and the language models we audit are non-linear and high-dimensional. To combat these challenges, we introduce a discrete optimization algorithm, ARCA, that jointly and efficiently optimizes over inputs and outputs. Our approach automatically uncovers derogatory completions about celebrities (e.g. "Barack Obama is a legalized unborn" -> "child murderer"), produces French inputs that complete to English outputs, and finds inputs that generate a specific name. Our work offers a promising new tool to uncover models' failure-modes before deployment.
2023-03-08
{'id': '2303.04381v1', 'title': 'Automatically Auditing Large Language Models via Discrete Optimization', 'abstract': 'Auditing large language models for unexpected behaviors is critical to\npreempt catastrophic deployments, yet remains challenging. In this work, we\ncast auditing as an optimization problem, where we automatically search for\ninput-output pairs that match a desired target behavior. For example, we might\naim to find a non-toxic input that starts with "Barack Obama" that a model maps\nto a toxic output. This optimization problem is difficult to solve as the set\nof feasible points is sparse, the space is discrete, and the language models we\naudit are non-linear and high-dimensional. To combat these challenges, we\nintroduce a discrete optimization algorithm, ARCA, that jointly and efficiently\noptimizes over inputs and outputs. Our approach automatically uncovers\nderogatory completions about celebrities (e.g. "Barack Obama is a legalized\nunborn" -> "child murderer"), produces French inputs that complete to English\noutputs, and finds inputs that generate a specific name. Our work offers a\npromising new tool to uncover models\' failure-modes before deployment.', 'authors': ['Erik Jones', 'Anca Dragan', 'Aditi Raghunathan', 'Jacob Steinhardt'], 'published': '2023-03-08 05:09:59+00:00', 'categories': ['cs.LG', 'cs.CL'], 'url': 'http://arxiv.org/abs/2303.04381v1'}
42,405
2209.05557v3
2,209.05557
3
Blurring Diffusion Models
Recently, Rissanen et al., (2022) have presented a new type of diffusion process for generative modeling based on heat dissipation, or blurring, as an alternative to isotropic Gaussian diffusion. Here, we show that blurring can equivalently be defined through a Gaussian diffusion process with non-isotropic noise. In making this connection, we bridge the gap between inverse heat dissipation and denoising diffusion, and we shed light on the inductive bias that results from this modeling choice. Finally, we propose a generalized class of diffusion models that offers the best of both standard Gaussian denoising diffusion and inverse heat dissipation, which we call Blurring Diffusion Models.
2022-09-12
{'id': '2209.05557v3', 'title': 'Blurring Diffusion Models', 'abstract': 'Recently, Rissanen et al., (2022) have presented a new type of diffusion\nprocess for generative modeling based on heat dissipation, or blurring, as an\nalternative to isotropic Gaussian diffusion. Here, we show that blurring can\nequivalently be defined through a Gaussian diffusion process with non-isotropic\nnoise. In making this connection, we bridge the gap between inverse heat\ndissipation and denoising diffusion, and we shed light on the inductive bias\nthat results from this modeling choice. Finally, we propose a generalized class\nof diffusion models that offers the best of both standard Gaussian denoising\ndiffusion and inverse heat dissipation, which we call Blurring Diffusion\nModels.', 'authors': ['Emiel Hoogeboom', 'Tim Salimans'], 'published': '2022-09-12 19:16:48+00:00', 'categories': ['cs.LG', 'cs.CV', 'stat.ML'], 'url': 'http://arxiv.org/abs/2209.05557v3'}
42,408
2306.13085v1
2,306.13085
1
Harnessing Mixed Offline Reinforcement Learning Datasets via Trajectory Weighting
Most offline reinforcement learning (RL) algorithms return a target policy maximizing a trade-off between (1) the expected performance gain over the behavior policy that collected the dataset, and (2) the risk stemming from the out-of-distribution-ness of the induced state-action occupancy. It follows that the performance of the target policy is strongly related to the performance of the behavior policy and, thus, the trajectory return distribution of the dataset. We show that in mixed datasets consisting of mostly low-return trajectories and minor high-return trajectories, state-of-the-art offline RL algorithms are overly restrained by low-return trajectories and fail to exploit high-performing trajectories to the fullest. To overcome this issue, we show that, in deterministic MDPs with stochastic initial states, the dataset sampling can be re-weighted to induce an artificial dataset whose behavior policy has a higher return. This re-weighted sampling strategy may be combined with any offline RL algorithm. We further analyze that the opportunity for performance improvement over the behavior policy correlates with the positive-sided variance of the returns of the trajectories in the dataset. We empirically show that while CQL, IQL, and TD3+BC achieve only a part of this potential policy improvement, these same algorithms combined with our reweighted sampling strategy fully exploit the dataset. Furthermore, we empirically demonstrate that, despite its theoretical limitation, the approach may still be efficient in stochastic environments. The code is available at https://github.com/Improbable-AI/harness-offline-rl.
2023-06-22
{'id': '2306.13085v1', 'title': 'Harnessing Mixed Offline Reinforcement Learning Datasets via Trajectory Weighting', 'abstract': 'Most offline reinforcement learning (RL) algorithms return a target policy\nmaximizing a trade-off between (1) the expected performance gain over the\nbehavior policy that collected the dataset, and (2) the risk stemming from the\nout-of-distribution-ness of the induced state-action occupancy. It follows that\nthe performance of the target policy is strongly related to the performance of\nthe behavior policy and, thus, the trajectory return distribution of the\ndataset. We show that in mixed datasets consisting of mostly low-return\ntrajectories and minor high-return trajectories, state-of-the-art offline RL\nalgorithms are overly restrained by low-return trajectories and fail to exploit\nhigh-performing trajectories to the fullest. To overcome this issue, we show\nthat, in deterministic MDPs with stochastic initial states, the dataset\nsampling can be re-weighted to induce an artificial dataset whose behavior\npolicy has a higher return. This re-weighted sampling strategy may be combined\nwith any offline RL algorithm. We further analyze that the opportunity for\nperformance improvement over the behavior policy correlates with the\npositive-sided variance of the returns of the trajectories in the dataset. We\nempirically show that while CQL, IQL, and TD3+BC achieve only a part of this\npotential policy improvement, these same algorithms combined with our\nreweighted sampling strategy fully exploit the dataset. Furthermore, we\nempirically demonstrate that, despite its theoretical limitation, the approach\nmay still be efficient in stochastic environments. The code is available at\nhttps://github.com/Improbable-AI/harness-offline-rl.', 'authors': ['Zhang-Wei Hong', 'Pulkit Agrawal', 'Rémi Tachet des Combes', 'Romain Laroche'], 'published': '2023-06-22 17:58:02+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2306.13085v1'}
42,411
2206.07370v1
2,206.0737
1
Lattice Convolutional Networks for Learning Ground States of Quantum Many-Body Systems
Deep learning methods have been shown to be effective in representing ground-state wave functions of quantum many-body systems. Existing methods use convolutional neural networks (CNNs) for square lattices due to their image-like structures. For non-square lattices, existing method uses graph neural network (GNN) in which structure information is not precisely captured, thereby requiring additional hand-crafted sublattice encoding. In this work, we propose lattice convolutions in which a set of proposed operations are used to convert non-square lattices into grid-like augmented lattices on which regular convolution can be applied. Based on the proposed lattice convolutions, we design lattice convolutional networks (LCN) that use self-gating and attention mechanisms. Experimental results show that our method achieves performance on par or better than existing methods on spin 1/2 $J_1$-$J_2$ Heisenberg model over the square, honeycomb, triangular, and kagome lattices while without using hand-crafted encoding.
2022-06-15
{'id': '2206.07370v1', 'title': 'Lattice Convolutional Networks for Learning Ground States of Quantum Many-Body Systems', 'abstract': 'Deep learning methods have been shown to be effective in representing\nground-state wave functions of quantum many-body systems. Existing methods use\nconvolutional neural networks (CNNs) for square lattices due to their\nimage-like structures. For non-square lattices, existing method uses graph\nneural network (GNN) in which structure information is not precisely captured,\nthereby requiring additional hand-crafted sublattice encoding. In this work, we\npropose lattice convolutions in which a set of proposed operations are used to\nconvert non-square lattices into grid-like augmented lattices on which regular\nconvolution can be applied. Based on the proposed lattice convolutions, we\ndesign lattice convolutional networks (LCN) that use self-gating and attention\nmechanisms. Experimental results show that our method achieves performance on\npar or better than existing methods on spin 1/2 $J_1$-$J_2$ Heisenberg model\nover the square, honeycomb, triangular, and kagome lattices while without using\nhand-crafted encoding.', 'authors': ['Cong Fu', 'Xuan Zhang', 'Huixin Zhang', 'Hongyi Ling', 'Shenglong Xu', 'Shuiwang Ji'], 'published': '2022-06-15 08:24:37+00:00', 'categories': ['quant-ph', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2206.07370v1'}
42,413
2205.15460v2
2,205.1546
2
Critic Sequential Monte Carlo
We introduce CriticSMC, a new algorithm for planning as inference built from a composition of sequential Monte Carlo with learned Soft-Q function heuristic factors. These heuristic factors, obtained from parametric approximations of the marginal likelihood ahead, more effectively guide SMC towards the desired target distribution, which is particularly helpful for planning in environments with hard constraints placed sparsely in time. Compared with previous work, we modify the placement of such heuristic factors, which allows us to cheaply propose and evaluate large numbers of putative action particles, greatly increasing inference and planning efficiency. CriticSMC is compatible with informative priors, whose density function need not be known, and can be used as a model-free control algorithm. Our experiments on collision avoidance in a high-dimensional simulated driving task show that CriticSMC significantly reduces collision rates at a low computational cost while maintaining realism and diversity of driving behaviors across vehicles and environment scenarios.
2022-05-30
{'id': '2205.15460v2', 'title': 'Critic Sequential Monte Carlo', 'abstract': 'We introduce CriticSMC, a new algorithm for planning as inference built from\na composition of sequential Monte Carlo with learned Soft-Q function heuristic\nfactors. These heuristic factors, obtained from parametric approximations of\nthe marginal likelihood ahead, more effectively guide SMC towards the desired\ntarget distribution, which is particularly helpful for planning in environments\nwith hard constraints placed sparsely in time. Compared with previous work, we\nmodify the placement of such heuristic factors, which allows us to cheaply\npropose and evaluate large numbers of putative action particles, greatly\nincreasing inference and planning efficiency. CriticSMC is compatible with\ninformative priors, whose density function need not be known, and can be used\nas a model-free control algorithm. Our experiments on collision avoidance in a\nhigh-dimensional simulated driving task show that CriticSMC significantly\nreduces collision rates at a low computational cost while maintaining realism\nand diversity of driving behaviors across vehicles and environment scenarios.', 'authors': ['Vasileios Lioutas', 'Jonathan Wilder Lavington', 'Justice Sefas', 'Matthew Niedoba', 'Yunpeng Liu', 'Berend Zwartsenberg', 'Setareh Dabiri', 'Frank Wood', 'Adam Scibior'], 'published': '2022-05-30 23:14:24+00:00', 'categories': ['stat.ML', 'cs.LG'], 'url': 'http://arxiv.org/abs/2205.15460v2'}
42,371
2201.05242v2
2,201.05242
2
Neural Circuit Architectural Priors for Embodied Control
Artificial neural networks for motor control usually adopt generic architectures like fully connected MLPs. While general, these tabula rasa architectures rely on large amounts of experience to learn, are not easily transferable to new bodies, and have internal dynamics that are difficult to interpret. In nature, animals are born with highly structured connectivity in their nervous systems shaped by evolution; this innate circuitry acts synergistically with learning mechanisms to provide inductive biases that enable most animals to function well soon after birth and learn efficiently. Convolutional networks inspired by visual circuitry have encoded useful biases for vision. However, it is unknown the extent to which ANN architectures inspired by neural circuitry can yield useful biases for other AI domains. In this work, we ask what advantages biologically inspired ANN architecture can provide in the domain of motor control. Specifically, we translate C. elegans locomotion circuits into an ANN model controlling a simulated Swimmer agent. On a locomotion task, our architecture achieves good initial performance and asymptotic performance comparable with MLPs, while dramatically improving data efficiency and requiring orders of magnitude fewer parameters. Our architecture is interpretable and transfers to new body designs. An ablation analysis shows that constrained excitation/inhibition is crucial for learning, while weight initialization contributes to good initial performance. Our work demonstrates several advantages of biologically inspired ANN architecture and encourages future work in more complex embodied control.
2022-01-13
{'id': '2201.05242v2', 'title': 'Neural Circuit Architectural Priors for Embodied Control', 'abstract': 'Artificial neural networks for motor control usually adopt generic\narchitectures like fully connected MLPs. While general, these tabula rasa\narchitectures rely on large amounts of experience to learn, are not easily\ntransferable to new bodies, and have internal dynamics that are difficult to\ninterpret. In nature, animals are born with highly structured connectivity in\ntheir nervous systems shaped by evolution; this innate circuitry acts\nsynergistically with learning mechanisms to provide inductive biases that\nenable most animals to function well soon after birth and learn efficiently.\nConvolutional networks inspired by visual circuitry have encoded useful biases\nfor vision. However, it is unknown the extent to which ANN architectures\ninspired by neural circuitry can yield useful biases for other AI domains. In\nthis work, we ask what advantages biologically inspired ANN architecture can\nprovide in the domain of motor control. Specifically, we translate C. elegans\nlocomotion circuits into an ANN model controlling a simulated Swimmer agent. On\na locomotion task, our architecture achieves good initial performance and\nasymptotic performance comparable with MLPs, while dramatically improving data\nefficiency and requiring orders of magnitude fewer parameters. Our architecture\nis interpretable and transfers to new body designs. An ablation analysis shows\nthat constrained excitation/inhibition is crucial for learning, while weight\ninitialization contributes to good initial performance. Our work demonstrates\nseveral advantages of biologically inspired ANN architecture and encourages\nfuture work in more complex embodied control.', 'authors': ['Nikhil X. Bhattasali', 'Anthony M. Zador', 'Tatiana A. Engel'], 'published': '2022-01-13 23:22:16+00:00', 'categories': ['cs.LG', 'cs.NE', 'cs.RO', 'q-bio.NC'], 'url': 'http://arxiv.org/abs/2201.05242v2'}
42,375
2202.12162v2
2,202.12162
2
Measuring CLEVRness: Blackbox testing of Visual Reasoning Models
How can we measure the reasoning capabilities of intelligence systems? Visual question answering provides a convenient framework for testing the model's abilities by interrogating the model through questions about the scene. However, despite scores of various visual QA datasets and architectures, which sometimes yield even a super-human performance, the question of whether those architectures can actually reason remains open to debate. To answer this, we extend the visual question answering framework and propose the following behavioral test in the form of a two-player game. We consider black-box neural models of CLEVR. These models are trained on a diagnostic dataset benchmarking reasoning. Next, we train an adversarial player that re-configures the scene to fool the CLEVR model. We show that CLEVR models, which otherwise could perform at a human level, can easily be fooled by our agent. Our results put in doubt whether data-driven approaches can do reasoning without exploiting the numerous biases that are often present in those datasets. Finally, we also propose a controlled experiment measuring the efficiency of such models to learn and perform reasoning.
2022-02-24
{'id': '2202.12162v2', 'title': 'Measuring CLEVRness: Blackbox testing of Visual Reasoning Models', 'abstract': "How can we measure the reasoning capabilities of intelligence systems? Visual\nquestion answering provides a convenient framework for testing the model's\nabilities by interrogating the model through questions about the scene.\nHowever, despite scores of various visual QA datasets and architectures, which\nsometimes yield even a super-human performance, the question of whether those\narchitectures can actually reason remains open to debate. To answer this, we\nextend the visual question answering framework and propose the following\nbehavioral test in the form of a two-player game. We consider black-box neural\nmodels of CLEVR. These models are trained on a diagnostic dataset benchmarking\nreasoning. Next, we train an adversarial player that re-configures the scene to\nfool the CLEVR model. We show that CLEVR models, which otherwise could perform\nat a human level, can easily be fooled by our agent. Our results put in doubt\nwhether data-driven approaches can do reasoning without exploiting the numerous\nbiases that are often present in those datasets. Finally, we also propose a\ncontrolled experiment measuring the efficiency of such models to learn and\nperform reasoning.", 'authors': ['Spyridon Mouselinos', 'Henryk Michalewski', 'Mateusz Malinowski'], 'published': '2022-02-24 15:59:29+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CV'], 'url': 'http://arxiv.org/abs/2202.12162v2'}
42,393
2310.18531v1
2,310.18531
1
Feature Selection in the Contrastive Analysis Setting
Contrastive analysis (CA) refers to the exploration of variations uniquely enriched in a target dataset as compared to a corresponding background dataset generated from sources of variation that are irrelevant to a given task. For example, a biomedical data analyst may wish to find a small set of genes to use as a proxy for variations in genomic data only present among patients with a given disease (target) as opposed to healthy control subjects (background). However, as of yet the problem of feature selection in the CA setting has received little attention from the machine learning community. In this work we present contrastive feature selection (CFS), a method for performing feature selection in the CA setting. We motivate our approach with a novel information-theoretic analysis of representation learning in the CA setting, and we empirically validate CFS on a semi-synthetic dataset and four real-world biomedical datasets. We find that our method consistently outperforms previously proposed state-of-the-art supervised and fully unsupervised feature selection methods not designed for the CA setting. An open-source implementation of our method is available at https://github.com/suinleelab/CFS.
2023-10-27
{'id': '2310.18531v1', 'title': 'Feature Selection in the Contrastive Analysis Setting', 'abstract': 'Contrastive analysis (CA) refers to the exploration of variations uniquely\nenriched in a target dataset as compared to a corresponding background dataset\ngenerated from sources of variation that are irrelevant to a given task. For\nexample, a biomedical data analyst may wish to find a small set of genes to use\nas a proxy for variations in genomic data only present among patients with a\ngiven disease (target) as opposed to healthy control subjects (background).\nHowever, as of yet the problem of feature selection in the CA setting has\nreceived little attention from the machine learning community. In this work we\npresent contrastive feature selection (CFS), a method for performing feature\nselection in the CA setting. We motivate our approach with a novel\ninformation-theoretic analysis of representation learning in the CA setting,\nand we empirically validate CFS on a semi-synthetic dataset and four real-world\nbiomedical datasets. We find that our method consistently outperforms\npreviously proposed state-of-the-art supervised and fully unsupervised feature\nselection methods not designed for the CA setting. An open-source\nimplementation of our method is available at https://github.com/suinleelab/CFS.', 'authors': ['Ethan Weinberger', 'Ian Covert', 'Su-In Lee'], 'published': '2023-10-27 23:16:03+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.18531v1'}
42,396
2204.05842v1
2,204.05842
1
Generative Negative Replay for Continual Learning
Learning continually is a key aspect of intelligence and a necessary ability to solve many real-life problems. One of the most effective strategies to control catastrophic forgetting, the Achilles' heel of continual learning, is storing part of the old data and replaying them interleaved with new experiences (also known as the replay approach). Generative replay, which is using generative models to provide replay patterns on demand, is particularly intriguing, however, it was shown to be effective mainly under simplified assumptions, such as simple scenarios and low-dimensional data. In this paper, we show that, while the generated data are usually not able to improve the classification accuracy for the old classes, they can be effective as negative examples (or antagonists) to better learn the new classes, especially when the learning experiences are small and contain examples of just one or few classes. The proposed approach is validated on complex class-incremental and data-incremental continual learning scenarios (CORe50 and ImageNet-1000) composed of high-dimensional data and a large number of training experiences: a setup where existing generative replay approaches usually fail.
2022-04-12
{'id': '2204.05842v1', 'title': 'Generative Negative Replay for Continual Learning', 'abstract': "Learning continually is a key aspect of intelligence and a necessary ability\nto solve many real-life problems. One of the most effective strategies to\ncontrol catastrophic forgetting, the Achilles' heel of continual learning, is\nstoring part of the old data and replaying them interleaved with new\nexperiences (also known as the replay approach). Generative replay, which is\nusing generative models to provide replay patterns on demand, is particularly\nintriguing, however, it was shown to be effective mainly under simplified\nassumptions, such as simple scenarios and low-dimensional data. In this paper,\nwe show that, while the generated data are usually not able to improve the\nclassification accuracy for the old classes, they can be effective as negative\nexamples (or antagonists) to better learn the new classes, especially when the\nlearning experiences are small and contain examples of just one or few classes.\nThe proposed approach is validated on complex class-incremental and\ndata-incremental continual learning scenarios (CORe50 and ImageNet-1000)\ncomposed of high-dimensional data and a large number of training experiences: a\nsetup where existing generative replay approaches usually fail.", 'authors': ['Gabriele Graffieti', 'Davide Maltoni', 'Lorenzo Pellegrini', 'Vincenzo Lomonaco'], 'published': '2022-04-12 14:38:00+00:00', 'categories': ['cs.LG', 'cs.CV', 'stat.ML'], 'url': 'http://arxiv.org/abs/2204.05842v1'}
42,397
2110.02642v5
2,110.02642
5
Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy
Unsupervised detection of anomaly points in time series is a challenging problem, which requires the model to derive a distinguishable criterion. Previous methods tackle the problem mainly through learning pointwise representation or pairwise association, however, neither is sufficient to reason about the intricate dynamics. Recently, Transformers have shown great power in unified modeling of pointwise representation and pairwise association, and we find that the self-attention weight distribution of each time point can embody rich association with the whole series. Our key observation is that due to the rarity of anomalies, it is extremely difficult to build nontrivial associations from abnormal points to the whole series, thereby, the anomalies' associations shall mainly concentrate on their adjacent time points. This adjacent-concentration bias implies an association-based criterion inherently distinguishable between normal and abnormal points, which we highlight through the Association Discrepancy. Technically, we propose the Anomaly Transformer with a new Anomaly-Attention mechanism to compute the association discrepancy. A minimax strategy is devised to amplify the normal-abnormal distinguishability of the association discrepancy. The Anomaly Transformer achieves state-of-the-art results on six unsupervised time series anomaly detection benchmarks of three applications: service monitoring, space & earth exploration, and water treatment.
2021-10-06
{'id': '2110.02642v5', 'title': 'Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy', 'abstract': "Unsupervised detection of anomaly points in time series is a challenging\nproblem, which requires the model to derive a distinguishable criterion.\nPrevious methods tackle the problem mainly through learning pointwise\nrepresentation or pairwise association, however, neither is sufficient to\nreason about the intricate dynamics. Recently, Transformers have shown great\npower in unified modeling of pointwise representation and pairwise association,\nand we find that the self-attention weight distribution of each time point can\nembody rich association with the whole series. Our key observation is that due\nto the rarity of anomalies, it is extremely difficult to build nontrivial\nassociations from abnormal points to the whole series, thereby, the anomalies'\nassociations shall mainly concentrate on their adjacent time points. This\nadjacent-concentration bias implies an association-based criterion inherently\ndistinguishable between normal and abnormal points, which we highlight through\nthe \\emph{Association Discrepancy}. Technically, we propose the \\emph{Anomaly\nTransformer} with a new \\emph{Anomaly-Attention} mechanism to compute the\nassociation discrepancy. A minimax strategy is devised to amplify the\nnormal-abnormal distinguishability of the association discrepancy. The Anomaly\nTransformer achieves state-of-the-art results on six unsupervised time series\nanomaly detection benchmarks of three applications: service monitoring, space &\nearth exploration, and water treatment.", 'authors': ['Jiehui Xu', 'Haixu Wu', 'Jianmin Wang', 'Mingsheng Long'], 'published': '2021-10-06 10:33:55+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2110.02642v5'}
42,400
2107.05216v1
2,107.05216
1
A Simple Reward-free Approach to Constrained Reinforcement Learning
In constrained reinforcement learning (RL), a learning agent seeks to not only optimize the overall reward but also satisfy the additional safety, diversity, or budget constraints. Consequently, existing constrained RL solutions require several new algorithmic ingredients that are notably different from standard RL. On the other hand, reward-free RL is independently developed in the unconstrained literature, which learns the transition dynamics without using the reward information, and thus naturally capable of addressing RL with multiple objectives under the common dynamics. This paper bridges reward-free RL and constrained RL. Particularly, we propose a simple meta-algorithm such that given any reward-free RL oracle, the approachability and constrained RL problems can be directly solved with negligible overheads in sample complexity. Utilizing the existing reward-free RL solvers, our framework provides sharp sample complexity results for constrained RL in the tabular MDP setting, matching the best existing results up to a factor of horizon dependence; our framework directly extends to a setting of tabular two-player Markov games, and gives a new result for constrained RL with linear function approximation.
2021-07-12
{'id': '2107.05216v1', 'title': 'A Simple Reward-free Approach to Constrained Reinforcement Learning', 'abstract': 'In constrained reinforcement learning (RL), a learning agent seeks to not\nonly optimize the overall reward but also satisfy the additional safety,\ndiversity, or budget constraints. Consequently, existing constrained RL\nsolutions require several new algorithmic ingredients that are notably\ndifferent from standard RL. On the other hand, reward-free RL is independently\ndeveloped in the unconstrained literature, which learns the transition dynamics\nwithout using the reward information, and thus naturally capable of addressing\nRL with multiple objectives under the common dynamics. This paper bridges\nreward-free RL and constrained RL. Particularly, we propose a simple\nmeta-algorithm such that given any reward-free RL oracle, the approachability\nand constrained RL problems can be directly solved with negligible overheads in\nsample complexity. Utilizing the existing reward-free RL solvers, our framework\nprovides sharp sample complexity results for constrained RL in the tabular MDP\nsetting, matching the best existing results up to a factor of horizon\ndependence; our framework directly extends to a setting of tabular two-player\nMarkov games, and gives a new result for constrained RL with linear function\napproximation.', 'authors': ['Sobhan Miryoosefi', 'Chi Jin'], 'published': '2021-07-12 06:27:30+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2107.05216v1'}