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
42,611
2301.12386v4
2,301.12386
4
Plugin estimators for selective classification with out-of-distribution detection
Real-world classifiers can benefit from the option of abstaining from predicting on samples where they have low confidence. Such abstention is particularly useful on samples which are close to the learned decision boundary, or which are outliers with respect to the training sample. These settings have been the subject of extensive but disjoint study in the selective classification (SC) and out-of-distribution (OOD) detection literature. Recent work on selective classification with OOD detection (SCOD) has argued for the unified study of these problems; however, the formal underpinnings of this problem are still nascent, and existing techniques are heuristic in nature. In this paper, we propose new plugin estimators for SCOD that are theoretically grounded, effective, and generalise existing approaches from the SC and OOD detection literature. In the course of our analysis, we formally explicate how na\"ive use of existing SC and OOD detection baselines may be inadequate for SCOD. We empirically demonstrate that our approaches yields competitive SC and OOD detection performance compared to baselines from both literatures.
2023-01-29
{'id': '2301.12386v4', 'title': 'Plugin estimators for selective classification with out-of-distribution detection', 'abstract': 'Real-world classifiers can benefit from the option of abstaining from\npredicting on samples where they have low confidence. Such abstention is\nparticularly useful on samples which are close to the learned decision\nboundary, or which are outliers with respect to the training sample. These\nsettings have been the subject of extensive but disjoint study in the selective\nclassification (SC) and out-of-distribution (OOD) detection literature. Recent\nwork on selective classification with OOD detection (SCOD) has argued for the\nunified study of these problems; however, the formal underpinnings of this\nproblem are still nascent, and existing techniques are heuristic in nature. In\nthis paper, we propose new plugin estimators for SCOD that are theoretically\ngrounded, effective, and generalise existing approaches from the SC and OOD\ndetection literature. In the course of our analysis, we formally explicate how\nna\\"{i}ve use of existing SC and OOD detection baselines may be inadequate for\nSCOD. We empirically demonstrate that our approaches yields competitive SC and\nOOD detection performance compared to baselines from both literatures.', 'authors': ['Harikrishna Narasimhan', 'Aditya Krishna Menon', 'Wittawat Jitkrittum', 'Sanjiv Kumar'], 'published': '2023-01-29 07:45:17+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2301.12386v4'}
42,615
2305.02299v4
2,305.02299
4
Dynamic Sparse Training with Structured Sparsity
Dynamic Sparse Training (DST) methods achieve state-of-the-art results in sparse neural network training, matching the generalization of dense models while enabling sparse training and inference. Although the resulting models are highly sparse and theoretically less computationally expensive, achieving speedups with unstructured sparsity on real-world hardware is challenging. In this work, we propose a sparse-to-sparse DST method, Structured RigL (SRigL), to learn a variant of fine-grained structured N:M sparsity by imposing a constant fan-in constraint. Using our empirical analysis of existing DST methods at high sparsity, we additionally employ a neuron ablation method which enables SRigL to achieve state-of-the-art sparse-to-sparse structured DST performance on a variety of Neural Network (NN) architectures. Using a 90% sparse linear layer, we demonstrate a real-world acceleration of 3.4x/2.5x on CPU for online inference and 1.7x/13.0x on GPU for inference with a batch size of 256 when compared to equivalent dense/unstructured (CSR) sparse layers, respectively.
2023-05-03
{'id': '2305.02299v4', 'title': 'Dynamic Sparse Training with Structured Sparsity', 'abstract': 'Dynamic Sparse Training (DST) methods achieve state-of-the-art results in\nsparse neural network training, matching the generalization of dense models\nwhile enabling sparse training and inference. Although the resulting models are\nhighly sparse and theoretically less computationally expensive, achieving\nspeedups with unstructured sparsity on real-world hardware is challenging. In\nthis work, we propose a sparse-to-sparse DST method, Structured RigL (SRigL),\nto learn a variant of fine-grained structured N:M sparsity by imposing a\nconstant fan-in constraint. Using our empirical analysis of existing DST\nmethods at high sparsity, we additionally employ a neuron ablation method which\nenables SRigL to achieve state-of-the-art sparse-to-sparse structured DST\nperformance on a variety of Neural Network (NN) architectures. Using a 90%\nsparse linear layer, we demonstrate a real-world acceleration of 3.4x/2.5x on\nCPU for online inference and 1.7x/13.0x on GPU for inference with a batch size\nof 256 when compared to equivalent dense/unstructured (CSR) sparse layers,\nrespectively.', 'authors': ['Mike Lasby', 'Anna Golubeva', 'Utku Evci', 'Mihai Nica', 'Yani Ioannou'], 'published': '2023-05-03 17:48:55+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2305.02299v4'}
42,624
2310.04686v1
2,310.04686
1
Tight Rates in Supervised Outlier Transfer Learning
A critical barrier to learning an accurate decision rule for outlier detection is the scarcity of outlier data. As such, practitioners often turn to the use of similar but imperfect outlier data from which they might transfer information to the target outlier detection task. Despite the recent empirical success of transfer learning approaches in outlier detection, a fundamental understanding of when and how knowledge can be transferred from a source to a target outlier detection task remains elusive. In this work, we adopt the traditional framework of Neyman-Pearson classification -- which formalizes supervised outlier detection -- with the added assumption that one has access to some related but imperfect outlier data. Our main results are as follows: We first determine the information-theoretic limits of the problem under a measure of discrepancy that extends some existing notions from traditional balanced classification; interestingly, unlike in balanced classification, seemingly very dissimilar sources can provide much information about a target, thus resulting in fast transfer. We then show that, in principle, these information-theoretic limits are achievable by adaptive procedures, i.e., procedures with no a priori information on the discrepancy between source and target outlier distributions.
2023-10-07
{'id': '2310.04686v1', 'title': 'Tight Rates in Supervised Outlier Transfer Learning', 'abstract': 'A critical barrier to learning an accurate decision rule for outlier\ndetection is the scarcity of outlier data. As such, practitioners often turn to\nthe use of similar but imperfect outlier data from which they might transfer\ninformation to the target outlier detection task. Despite the recent empirical\nsuccess of transfer learning approaches in outlier detection, a fundamental\nunderstanding of when and how knowledge can be transferred from a source to a\ntarget outlier detection task remains elusive. In this work, we adopt the\ntraditional framework of Neyman-Pearson classification -- which formalizes\nsupervised outlier detection -- with the added assumption that one has access\nto some related but imperfect outlier data. Our main results are as follows:\n We first determine the information-theoretic limits of the problem under a\nmeasure of discrepancy that extends some existing notions from traditional\nbalanced classification; interestingly, unlike in balanced classification,\nseemingly very dissimilar sources can provide much information about a target,\nthus resulting in fast transfer.\n We then show that, in principle, these information-theoretic limits are\nachievable by adaptive procedures, i.e., procedures with no a priori\ninformation on the discrepancy between source and target outlier distributions.', 'authors': ['Mohammadreza M. Kalan', 'Samory Kpotufe'], 'published': '2023-10-07 04:14:13+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.04686v1'}
42,633
2310.00112v2
2,310.00112
2
Reinforcement Learning for Node Selection in Branch-and-Bound
A big challenge in branch and bound lies in identifying the optimal node within the search tree from which to proceed. Current state-of-the-art selectors utilize either hand-crafted ensembles that automatically switch between naive sub-node selectors, or learned node selectors that rely on individual node data. We propose a novel simulation technique that uses reinforcement learning (RL) while considering the entire tree state, rather than just isolated nodes. To achieve this, we train a graph neural network that produces a probability distribution based on the path from the model's root to its "to-be-selected" leaves. Modelling node-selection as a probability distribution allows us to train the model using state-of-the-art RL techniques that capture both intrinsic node-quality and node-evaluation costs. Our method induces a high quality node selection policy on a set of varied and complex problem sets, despite only being trained on specially designed, synthetic travelling salesmen problem (TSP) instances. Using such a fixed pretrained policy shows significant improvements on several benchmarks in optimality gap reductions and per-node efficiency under strict time constraints.
2023-09-29
{'id': '2310.00112v2', 'title': 'Reinforcement Learning for Node Selection in Branch-and-Bound', 'abstract': 'A big challenge in branch and bound lies in identifying the optimal node\nwithin the search tree from which to proceed. Current state-of-the-art\nselectors utilize either hand-crafted ensembles that automatically switch\nbetween naive sub-node selectors, or learned node selectors that rely on\nindividual node data. We propose a novel simulation technique that uses\nreinforcement learning (RL) while considering the entire tree state, rather\nthan just isolated nodes. To achieve this, we train a graph neural network that\nproduces a probability distribution based on the path from the model\'s root to\nits "to-be-selected" leaves. Modelling node-selection as a probability\ndistribution allows us to train the model using state-of-the-art RL techniques\nthat capture both intrinsic node-quality and node-evaluation costs. Our method\ninduces a high quality node selection policy on a set of varied and complex\nproblem sets, despite only being trained on specially designed, synthetic\ntravelling salesmen problem (TSP) instances. Using such a fixed pretrained\npolicy shows significant improvements on several benchmarks in optimality gap\nreductions and per-node efficiency under strict time constraints.', 'authors': ['Alexander Mattick', 'Christopher Mutschler'], 'published': '2023-09-29 19:55:56+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.00112v2'}
42,594
2211.10691v2
2,211.10691
2
Two Facets of SDE Under an Information-Theoretic Lens: Generalization of SGD via Training Trajectories and via Terminal States
Stochastic differential equations (SDEs) have been shown recently to characterize well the dynamics of training machine learning models with SGD. When the generalization error of the SDE approximation closely aligns with that of SGD in expectation, it provides two opportunities for understanding better the generalization behaviour of SGD through its SDE approximation. Firstly, viewing SGD as full-batch gradient descent with Gaussian gradient noise allows us to obtain trajectory-based generalization bound using the information-theoretic bound from Xu and Raginsky [2017]. Secondly, assuming mild conditions, we estimate the steady-state weight distribution of SDE and use information-theoretic bounds from Xu and Raginsky [2017] and Negrea et al. [2019] to establish terminal-state-based generalization bounds. Our proposed bounds have some advantages, notably the trajectory-based bound outperforms results in Wang and Mao [2022], and the terminal-state-based bound exhibits a fast decay rate comparable to stability-based bounds.
2022-11-19
{'id': '2211.10691v2', 'title': 'Two Facets of SDE Under an Information-Theoretic Lens: Generalization of SGD via Training Trajectories and via Terminal States', 'abstract': 'Stochastic differential equations (SDEs) have been shown recently to\ncharacterize well the dynamics of training machine learning models with SGD.\nWhen the generalization error of the SDE approximation closely aligns with that\nof SGD in expectation, it provides two opportunities for understanding better\nthe generalization behaviour of SGD through its SDE approximation. Firstly,\nviewing SGD as full-batch gradient descent with Gaussian gradient noise allows\nus to obtain trajectory-based generalization bound using the\ninformation-theoretic bound from Xu and Raginsky [2017]. Secondly, assuming\nmild conditions, we estimate the steady-state weight distribution of SDE and\nuse information-theoretic bounds from Xu and Raginsky [2017] and Negrea et al.\n[2019] to establish terminal-state-based generalization bounds. Our proposed\nbounds have some advantages, notably the trajectory-based bound outperforms\nresults in Wang and Mao [2022], and the terminal-state-based bound exhibits a\nfast decay rate comparable to stability-based bounds.', 'authors': ['Ziqiao Wang', 'Yongyi Mao'], 'published': '2022-11-19 13:15:39+00:00', 'categories': ['cs.LG', 'cs.IT', 'math.IT'], 'url': 'http://arxiv.org/abs/2211.10691v2'}
42,597
2310.05327v2
2,310.05327
2
Provable Compositional Generalization for Object-Centric Learning
Learning representations that generalize to novel compositions of known concepts is crucial for bridging the gap between human and machine perception. One prominent effort is learning object-centric representations, which are widely conjectured to enable compositional generalization. Yet, it remains unclear when this conjecture will be true, as a principled theoretical or empirical understanding of compositional generalization is lacking. In this work, we investigate when compositional generalization is guaranteed for object-centric representations through the lens of identifiability theory. We show that autoencoders that satisfy structural assumptions on the decoder and enforce encoder-decoder consistency will learn object-centric representations that provably generalize compositionally. We validate our theoretical result and highlight the practical relevance of our assumptions through experiments on synthetic image data.
2023-10-09
{'id': '2310.05327v2', 'title': 'Provable Compositional Generalization for Object-Centric Learning', 'abstract': 'Learning representations that generalize to novel compositions of known\nconcepts is crucial for bridging the gap between human and machine perception.\nOne prominent effort is learning object-centric representations, which are\nwidely conjectured to enable compositional generalization. Yet, it remains\nunclear when this conjecture will be true, as a principled theoretical or\nempirical understanding of compositional generalization is lacking. In this\nwork, we investigate when compositional generalization is guaranteed for\nobject-centric representations through the lens of identifiability theory. We\nshow that autoencoders that satisfy structural assumptions on the decoder and\nenforce encoder-decoder consistency will learn object-centric representations\nthat provably generalize compositionally. We validate our theoretical result\nand highlight the practical relevance of our assumptions through experiments on\nsynthetic image data.', 'authors': ['Thaddäus Wiedemer', 'Jack Brady', 'Alexander Panfilov', 'Attila Juhos', 'Matthias Bethge', 'Wieland Brendel'], 'published': '2023-10-09 01:18:07+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.05327v2'}
42,599
2306.02204v2
2,306.02204
2
Cycle Consistency Driven Object Discovery
Developing deep learning models that effectively learn object-centric representations, akin to human cognition, remains a challenging task. Existing approaches facilitate object discovery by representing objects as fixed-size vectors, called ``slots'' or ``object files''. While these approaches have shown promise in certain scenarios, they still exhibit certain limitations. First, they rely on architectural priors which can be unreliable and usually require meticulous engineering to identify the correct objects. Second, there has been a notable gap in investigating the practical utility of these representations in downstream tasks. To address the first limitation, we introduce a method that explicitly optimizes the constraint that each object in a scene should be associated with a distinct slot. We formalize this constraint by introducing consistency objectives which are cyclic in nature. By integrating these consistency objectives into various existing slot-based object-centric methods, we showcase substantial improvements in object-discovery performance. These enhancements consistently hold true across both synthetic and real-world scenes, underscoring the effectiveness and adaptability of the proposed approach. To tackle the second limitation, we apply the learned object-centric representations from the proposed method to two downstream reinforcement learning tasks, demonstrating considerable performance enhancements compared to conventional slot-based and monolithic representation learning methods. Our results suggest that the proposed approach not only improves object discovery, but also provides richer features for downstream tasks.
2023-06-03
{'id': '2306.02204v2', 'title': 'Cycle Consistency Driven Object Discovery', 'abstract': "Developing deep learning models that effectively learn object-centric\nrepresentations, akin to human cognition, remains a challenging task. Existing\napproaches facilitate object discovery by representing objects as fixed-size\nvectors, called ``slots'' or ``object files''. While these approaches have\nshown promise in certain scenarios, they still exhibit certain limitations.\nFirst, they rely on architectural priors which can be unreliable and usually\nrequire meticulous engineering to identify the correct objects. Second, there\nhas been a notable gap in investigating the practical utility of these\nrepresentations in downstream tasks. To address the first limitation, we\nintroduce a method that explicitly optimizes the constraint that each object in\na scene should be associated with a distinct slot. We formalize this constraint\nby introducing consistency objectives which are cyclic in nature. By\nintegrating these consistency objectives into various existing slot-based\nobject-centric methods, we showcase substantial improvements in\nobject-discovery performance. These enhancements consistently hold true across\nboth synthetic and real-world scenes, underscoring the effectiveness and\nadaptability of the proposed approach. To tackle the second limitation, we\napply the learned object-centric representations from the proposed method to\ntwo downstream reinforcement learning tasks, demonstrating considerable\nperformance enhancements compared to conventional slot-based and monolithic\nrepresentation learning methods. Our results suggest that the proposed approach\nnot only improves object discovery, but also provides richer features for\ndownstream tasks.", 'authors': ['Aniket Didolkar', 'Anirudh Goyal', 'Yoshua Bengio'], 'published': '2023-06-03 21:49:06+00:00', 'categories': ['cs.CV', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2306.02204v2'}
42,603
2403.10063v2
2,403.10063
2
Unified Projection-Free Algorithms for Adversarial DR-Submodular Optimization
This paper introduces unified projection-free Frank-Wolfe type algorithms for adversarial continuous DR-submodular optimization, spanning scenarios such as full information and (semi-)bandit feedback, monotone and non-monotone functions, different constraints, and types of stochastic queries. For every problem considered in the non-monotone setting, the proposed algorithms are either the first with proven sub-linear $$-regret bounds or have better $$-regret bounds than the state of the art, where $$ is a corresponding approximation bound in the offline setting. In the monotone setting, the proposed approach gives state-of-the-art sub-linear $$-regret bounds among projection-free algorithms in 7 of the 8 considered cases while matching the result of the remaining case. Additionally, this paper addresses semi-bandit and bandit feedback for adversarial DR-submodular optimization, advancing the understanding of this optimization area.
2024-03-15
{'id': '2403.10063v2', 'title': 'Unified Projection-Free Algorithms for Adversarial DR-Submodular Optimization', 'abstract': 'This paper introduces unified projection-free Frank-Wolfe type algorithms for\nadversarial continuous DR-submodular optimization, spanning scenarios such as\nfull information and (semi-)bandit feedback, monotone and non-monotone\nfunctions, different constraints, and types of stochastic queries. For every\nproblem considered in the non-monotone setting, the proposed algorithms are\neither the first with proven sub-linear $\\alpha$-regret bounds or have better\n$\\alpha$-regret bounds than the state of the art, where $\\alpha$ is a\ncorresponding approximation bound in the offline setting. In the monotone\nsetting, the proposed approach gives state-of-the-art sub-linear\n$\\alpha$-regret bounds among projection-free algorithms in 7 of the 8\nconsidered cases while matching the result of the remaining case. Additionally,\nthis paper addresses semi-bandit and bandit feedback for adversarial\nDR-submodular optimization, advancing the understanding of this optimization\narea.', 'authors': ['Mohammad Pedramfar', 'Yididiya Y. Nadew', 'Christopher J. Quinn', 'Vaneet Aggarwal'], 'published': '2024-03-15 07:05:44+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CC', 'math.OC'], 'url': 'http://arxiv.org/abs/2403.10063v2'}
42,604
2309.17400v2
2,309.174
2
Directly Fine-Tuning Diffusion Models on Differentiable Rewards
We present Direct Reward Fine-Tuning (DRaFT), a simple and effective method for fine-tuning diffusion models to maximize differentiable reward functions, such as scores from human preference models. We first show that it is possible to backpropagate the reward function gradient through the full sampling procedure, and that doing so achieves strong performance on a variety of rewards, outperforming reinforcement learning-based approaches. We then propose more efficient variants of DRaFT: DRaFT-K, which truncates backpropagation to only the last K steps of sampling, and DRaFT-LV, which obtains lower-variance gradient estimates for the case when K=1. We show that our methods work well for a variety of reward functions and can be used to substantially improve the aesthetic quality of images generated by Stable Diffusion 1.4. Finally, we draw connections between our approach and prior work, providing a unifying perspective on the design space of gradient-based fine-tuning algorithms.
2023-09-29
{'id': '2309.17400v2', 'title': 'Directly Fine-Tuning Diffusion Models on Differentiable Rewards', 'abstract': 'We present Direct Reward Fine-Tuning (DRaFT), a simple and effective method\nfor fine-tuning diffusion models to maximize differentiable reward functions,\nsuch as scores from human preference models. We first show that it is possible\nto backpropagate the reward function gradient through the full sampling\nprocedure, and that doing so achieves strong performance on a variety of\nrewards, outperforming reinforcement learning-based approaches. We then propose\nmore efficient variants of DRaFT: DRaFT-K, which truncates backpropagation to\nonly the last K steps of sampling, and DRaFT-LV, which obtains lower-variance\ngradient estimates for the case when K=1. We show that our methods work well\nfor a variety of reward functions and can be used to substantially improve the\naesthetic quality of images generated by Stable Diffusion 1.4. Finally, we draw\nconnections between our approach and prior work, providing a unifying\nperspective on the design space of gradient-based fine-tuning algorithms.', 'authors': ['Kevin Clark', 'Paul Vicol', 'Kevin Swersky', 'David J Fleet'], 'published': '2023-09-29 17:01:02+00:00', 'categories': ['cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2309.17400v2'}
42,607
2310.02535v1
2,310.02535
1
Linear programming using diagonal linear networks
Linear programming has played a crucial role in shaping decision-making, resource allocation, and cost reduction in various domains. In this paper, we investigate the application of overparametrized neural networks and their implicit bias in solving linear programming problems. Specifically, our findings reveal that training diagonal linear networks with gradient descent, while optimizing the squared $L_2$-norm of the slack variable, leads to solutions for entropically regularized linear programming problems. Remarkably, the strength of this regularization depends on the initialization used in the gradient descent process. We analyze the convergence of both discrete-time and continuous-time dynamics and demonstrate that both exhibit a linear rate of convergence, requiring only mild assumptions on the constraint matrix. For the first time, we introduce a comprehensive framework for solving linear programming problems using diagonal neural networks. We underscore the significance of our discoveries by applying them to address challenges in basis pursuit and optimal transport problems.
2023-10-04
{'id': '2310.02535v1', 'title': 'Linear programming using diagonal linear networks', 'abstract': 'Linear programming has played a crucial role in shaping decision-making,\nresource allocation, and cost reduction in various domains. In this paper, we\ninvestigate the application of overparametrized neural networks and their\nimplicit bias in solving linear programming problems. Specifically, our\nfindings reveal that training diagonal linear networks with gradient descent,\nwhile optimizing the squared $L_2$-norm of the slack variable, leads to\nsolutions for entropically regularized linear programming problems. Remarkably,\nthe strength of this regularization depends on the initialization used in the\ngradient descent process. We analyze the convergence of both discrete-time and\ncontinuous-time dynamics and demonstrate that both exhibit a linear rate of\nconvergence, requiring only mild assumptions on the constraint matrix. For the\nfirst time, we introduce a comprehensive framework for solving linear\nprogramming problems using diagonal neural networks. We underscore the\nsignificance of our discoveries by applying them to address challenges in basis\npursuit and optimal transport problems.', 'authors': ['Haoyue Wang', 'Promit Ghosal', 'Rahul Mazumder'], 'published': '2023-10-04 02:24:28+00:00', 'categories': ['math.OC'], 'url': 'http://arxiv.org/abs/2310.02535v1'}
42,601
2402.16158v2
2,402.16158
2
Distribution-Free Fair Federated Learning with Small Samples
As federated learning gains increasing importance in real-world applications due to its capacity for decentralized data training, addressing fairness concerns across demographic groups becomes critically important. However, most existing machine learning algorithms for ensuring fairness are designed for centralized data environments and generally require large-sample and distributional assumptions, underscoring the urgent need for fairness techniques adapted for decentralized and heterogeneous systems with finite-sample and distribution-free guarantees. To address this issue, this paper introduces FedFaiREE, a post-processing algorithm developed specifically for distribution-free fair learning in decentralized settings with small samples. Our approach accounts for unique challenges in decentralized environments, such as client heterogeneity, communication costs, and small sample sizes. We provide rigorous theoretical guarantees for both fairness and accuracy, and our experimental results further provide robust empirical validation for our proposed method.
2024-02-25
{'id': '2402.16158v2', 'title': 'Distribution-Free Fair Federated Learning with Small Samples', 'abstract': 'As federated learning gains increasing importance in real-world applications\ndue to its capacity for decentralized data training, addressing fairness\nconcerns across demographic groups becomes critically important. However, most\nexisting machine learning algorithms for ensuring fairness are designed for\ncentralized data environments and generally require large-sample and\ndistributional assumptions, underscoring the urgent need for fairness\ntechniques adapted for decentralized and heterogeneous systems with\nfinite-sample and distribution-free guarantees. To address this issue, this\npaper introduces FedFaiREE, a post-processing algorithm developed specifically\nfor distribution-free fair learning in decentralized settings with small\nsamples. Our approach accounts for unique challenges in decentralized\nenvironments, such as client heterogeneity, communication costs, and small\nsample sizes. We provide rigorous theoretical guarantees for both fairness and\naccuracy, and our experimental results further provide robust empirical\nvalidation for our proposed method.', 'authors': ['Qichuan Yin', 'Zexian Wang', 'Junzhou Huang', 'Huaxiu Yao', 'Linjun Zhang'], 'published': '2024-02-25 17:37:53+00:00', 'categories': ['stat.ML', 'cs.CY', 'cs.LG'], 'url': 'http://arxiv.org/abs/2402.16158v2'}
42,609
2305.16791v4
2,305.16791
4
On the Generalization and Approximation Capacities of Neural Controlled Differential Equations
Neural Controlled Differential Equations (NCDEs) are a state-of-the-art tool for supervised learning with irregularly sampled time series (Kidger, 2020). However, no theoretical analysis of their performance has been provided yet, and it remains unclear in particular how the irregularity of the time series affects their predictions. By merging the rich theory of controlled differential equations (CDE) and Lipschitz-based measures of the complexity of deep neural nets, we take a first step towards the theoretical understanding of NCDE. Our first result is a generalization bound for this class of predictors that depends on the regularity of the time series data. In a second time, we leverage the continuity of the flow of CDEs to provide a detailed analysis of both the sampling-induced bias and the approximation bias. Regarding this last result, we show how classical approximation results on neural nets may transfer to NCDEs. Our theoretical results are validated through a series of experiments.
2023-05-26
{'id': '2305.16791v4', 'title': 'On the Generalization and Approximation Capacities of Neural Controlled Differential Equations', 'abstract': 'Neural Controlled Differential Equations (NCDEs) are a state-of-the-art tool\nfor supervised learning with irregularly sampled time series (Kidger, 2020).\nHowever, no theoretical analysis of their performance has been provided yet,\nand it remains unclear in particular how the irregularity of the time series\naffects their predictions. By merging the rich theory of controlled\ndifferential equations (CDE) and Lipschitz-based measures of the complexity of\ndeep neural nets, we take a first step towards the theoretical understanding of\nNCDE. Our first result is a generalization bound for this class of predictors\nthat depends on the regularity of the time series data. In a second time, we\nleverage the continuity of the flow of CDEs to provide a detailed analysis of\nboth the sampling-induced bias and the approximation bias. Regarding this last\nresult, we show how classical approximation results on neural nets may transfer\nto NCDEs. Our theoretical results are validated through a series of\nexperiments.', 'authors': ['Linus Bleistein', 'Agathe Guilloux'], 'published': '2023-05-26 10:02:32+00:00', 'categories': ['stat.ML', 'cs.LG'], 'url': 'http://arxiv.org/abs/2305.16791v4'}
42,613
2309.05516v5
2,309.05516
5
Optimize Weight Rounding via Signed Gradient Descent for the Quantization of LLMs
Large Language Models (LLMs) have demonstrated exceptional proficiency in language-related tasks, but their deployment poses significant challenges due to substantial memory and storage requirements. Weight-only quantization has emerged as a promising solution, significantly reducing memory and storage needs without sacrificing too much performance. In this study, we introduce SignRound, a method that leverages signed gradient descent (SignSGD) to optimize rounding values and weight clipping in just 200 steps. SignRound integrates the advantages of Quantization-Aware Training (QAT) and Post-Training Quantization (PTQ), delivering exceptional results across 2 to 4 bits while minimizing tuning costs and avoiding additional inference overhead. For example, SignRound achieved absolute average accuracy improvements ranging from 6.91% to 33.22% at 2bits, as measured by the average zero-shot accuracy across 11 tasks. It also demonstrates strong generalization in recent models, achieving near-lossless 4-bit quantization in most scenarios. The source code is publicly available at https://github.com/intel/auto-round.
2023-09-11
{'id': '2309.05516v5', 'title': 'Optimize Weight Rounding via Signed Gradient Descent for the Quantization of LLMs', 'abstract': 'Large Language Models (LLMs) have demonstrated exceptional proficiency in\nlanguage-related tasks, but their deployment poses significant challenges due\nto substantial memory and storage requirements. Weight-only quantization has\nemerged as a promising solution, significantly reducing memory and storage\nneeds without sacrificing too much performance. In this study, we introduce\nSignRound, a method that leverages signed gradient descent (SignSGD) to\noptimize rounding values and weight clipping in just 200 steps. SignRound\nintegrates the advantages of Quantization-Aware Training (QAT) and\nPost-Training Quantization (PTQ), delivering exceptional results across 2 to 4\nbits while minimizing tuning costs and avoiding additional inference overhead.\nFor example, SignRound achieved absolute average accuracy improvements ranging\nfrom 6.91% to 33.22% at 2bits, as measured by the average zero-shot accuracy\nacross 11 tasks. It also demonstrates strong generalization in recent models,\nachieving near-lossless 4-bit quantization in most scenarios. The source code\nis publicly available at https://github.com/intel/auto-round.', 'authors': ['Wenhua Cheng', 'Weiwei Zhang', 'Haihao Shen', 'Yiyang Cai', 'Xin He', 'Kaokao Lv', 'Yi Liu'], 'published': '2023-09-11 14:58:23+00:00', 'categories': ['cs.CL', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2309.05516v5'}
42,617
2507.00025v1
2,507.00025
1
Generalizing to New Dynamical Systems via Frequency Domain Adaptation
Learning the underlying dynamics from data with deep neural networks has shown remarkable potential in modeling various complex physical dynamics. However, current approaches are constrained in their ability to make reliable predictions in a specific domain and struggle with generalizing to unseen systems that are governed by the same general dynamics but differ in environmental characteristics. In this work, we formulate a parameter-efficient method, Fourier Neural Simulator for Dynamical Adaptation (FNSDA), that can readily generalize to new dynamics via adaptation in the Fourier space. Specifically, FNSDA identifies the shareable dynamics based on the known environments using an automatic partition in Fourier modes and learns to adjust the modes specific for each new environment by conditioning on low-dimensional latent systematic parameters for efficient generalization. We evaluate our approach on four representative families of dynamic systems, and the results show that FNSDA can achieve superior or competitive generalization performance compared to existing methods with a significantly reduced parameter cost. Our code is available at https://github.com/WonderSeven/FNSDA.
2025-06-17
{'id': '2507.00025v1', 'title': 'Generalizing to New Dynamical Systems via Frequency Domain Adaptation', 'abstract': 'Learning the underlying dynamics from data with deep neural networks has\nshown remarkable potential in modeling various complex physical dynamics.\nHowever, current approaches are constrained in their ability to make reliable\npredictions in a specific domain and struggle with generalizing to unseen\nsystems that are governed by the same general dynamics but differ in\nenvironmental characteristics. In this work, we formulate a parameter-efficient\nmethod, Fourier Neural Simulator for Dynamical Adaptation (FNSDA), that can\nreadily generalize to new dynamics via adaptation in the Fourier space.\nSpecifically, FNSDA identifies the shareable dynamics based on the known\nenvironments using an automatic partition in Fourier modes and learns to adjust\nthe modes specific for each new environment by conditioning on low-dimensional\nlatent systematic parameters for efficient generalization. We evaluate our\napproach on four representative families of dynamic systems, and the results\nshow that FNSDA can achieve superior or competitive generalization performance\ncompared to existing methods with a significantly reduced parameter cost. Our\ncode is available at https://github.com/WonderSeven/FNSDA.', 'authors': ['Tiexin Qin', 'Hong Yan', 'Haoliang Li'], 'published': '2025-06-17 09:11:34+00:00', 'categories': ['cs.LG', 'cs.AI', 'stat.ML'], 'url': 'http://arxiv.org/abs/2507.00025v1'}
42,620
2311.05613v1
2,311.05613
1
Window Attention is Bugged: How not to Interpolate Position Embeddings
Window attention, position embeddings, and high resolution finetuning are core concepts in the modern transformer era of computer vision. However, we find that naively combining these near ubiquitous components can have a detrimental effect on performance. The issue is simple: interpolating position embeddings while using window attention is wrong. We study two state-of-the-art methods that have these three components, namely Hiera and ViTDet, and find that both do indeed suffer from this bug. To fix it, we introduce a simple absolute window position embedding strategy, which solves the bug outright in Hiera and allows us to increase both speed and performance of the model in ViTDet. We finally combine the two to obtain HieraDet, which achieves 61.7 box mAP on COCO, making it state-of-the-art for models that only use ImageNet-1k pretraining. This all stems from what is essentially a 3 line bug fix, which we name "absolute win".
2023-11-09
{'id': '2311.05613v1', 'title': 'Window Attention is Bugged: How not to Interpolate Position Embeddings', 'abstract': 'Window attention, position embeddings, and high resolution finetuning are\ncore concepts in the modern transformer era of computer vision. However, we\nfind that naively combining these near ubiquitous components can have a\ndetrimental effect on performance. The issue is simple: interpolating position\nembeddings while using window attention is wrong. We study two state-of-the-art\nmethods that have these three components, namely Hiera and ViTDet, and find\nthat both do indeed suffer from this bug. To fix it, we introduce a simple\nabsolute window position embedding strategy, which solves the bug outright in\nHiera and allows us to increase both speed and performance of the model in\nViTDet. We finally combine the two to obtain HieraDet, which achieves 61.7 box\nmAP on COCO, making it state-of-the-art for models that only use ImageNet-1k\npretraining. This all stems from what is essentially a 3 line bug fix, which we\nname "absolute win".', 'authors': ['Daniel Bolya', 'Chaitanya Ryali', 'Judy Hoffman', 'Christoph Feichtenhofer'], 'published': '2023-11-09 18:59:58+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2311.05613v1'}
42,629
2312.05807v1
2,312.05807
1
Federated Learning Empowered by Generative Content
Federated learning (FL) enables leveraging distributed private data for model training in a privacy-preserving way. However, data heterogeneity significantly limits the performance of current FL methods. In this paper, we propose a novel FL framework termed FedGC, designed to mitigate data heterogeneity issues by diversifying private data with generative content. FedGC is a simple-to-implement framework as it only introduces a one-shot step of data generation. In data generation, we summarize three crucial and worth-exploring aspects (budget allocation, prompt design, and generation guidance) and propose three solution candidates for each aspect. Specifically, to achieve a better trade-off between data diversity and fidelity for generation guidance, we propose to generate data based on the guidance of prompts and real data simultaneously. The generated data is then merged with private data to facilitate local model training. Such generative data increases the diversity of private data to prevent each client from fitting the potentially biased private data, alleviating the issue of data heterogeneity. We conduct a systematic empirical study on FedGC, covering diverse baselines, datasets, scenarios, and modalities. Interesting findings include (1) FedGC consistently and significantly enhances the performance of FL methods, even when notable disparities exist between generative and private data; (2) FedGC achieves both better performance and privacy-preservation. We wish this work can inspire future works to further explore the potential of enhancing FL with generative content.
2023-12-10
{'id': '2312.05807v1', 'title': 'Federated Learning Empowered by Generative Content', 'abstract': 'Federated learning (FL) enables leveraging distributed private data for model\ntraining in a privacy-preserving way. However, data heterogeneity significantly\nlimits the performance of current FL methods. In this paper, we propose a novel\nFL framework termed FedGC, designed to mitigate data heterogeneity issues by\ndiversifying private data with generative content. FedGC is a\nsimple-to-implement framework as it only introduces a one-shot step of data\ngeneration. In data generation, we summarize three crucial and worth-exploring\naspects (budget allocation, prompt design, and generation guidance) and propose\nthree solution candidates for each aspect. Specifically, to achieve a better\ntrade-off between data diversity and fidelity for generation guidance, we\npropose to generate data based on the guidance of prompts and real data\nsimultaneously. The generated data is then merged with private data to\nfacilitate local model training. Such generative data increases the diversity\nof private data to prevent each client from fitting the potentially biased\nprivate data, alleviating the issue of data heterogeneity. We conduct a\nsystematic empirical study on FedGC, covering diverse baselines, datasets,\nscenarios, and modalities. Interesting findings include (1) FedGC consistently\nand significantly enhances the performance of FL methods, even when notable\ndisparities exist between generative and private data; (2) FedGC achieves both\nbetter performance and privacy-preservation. We wish this work can inspire\nfuture works to further explore the potential of enhancing FL with generative\ncontent.', 'authors': ['Rui Ye', 'Xinyu Zhu', 'Jingyi Chai', 'Siheng Chen', 'Yanfeng Wang'], 'published': '2023-12-10 07:38:56+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2312.05807v1'}
42,608
2102.01862v2
2,102.01862
2
Towards Robust Neural Networks via Close-loop Control
Despite their success in massive engineering applications, deep neural networks are vulnerable to various perturbations due to their black-box nature. Recent study has shown that a deep neural network can misclassify the data even if the input data is perturbed by an imperceptible amount. In this paper, we address the robustness issue of neural networks by a novel close-loop control method from the perspective of dynamic systems. Instead of modifying the parameters in a fixed neural network architecture, a close-loop control process is added to generate control signals adaptively for the perturbed or corrupted data. We connect the robustness of neural networks with optimal control using the geometrical information of underlying data to design the control objective. The detailed analysis shows how the embedding manifolds of state trajectory affect error estimation of the proposed method. Our approach can simultaneously maintain the performance on clean data and improve the robustness against many types of data perturbations. It can also further improve the performance of robustly trained neural networks against different perturbations. To the best of our knowledge, this is the first work that improves the robustness of neural networks with close-loop control.
2021-02-03
{'id': '2102.01862v2', 'title': 'Towards Robust Neural Networks via Close-loop Control', 'abstract': 'Despite their success in massive engineering applications, deep neural\nnetworks are vulnerable to various perturbations due to their black-box nature.\nRecent study has shown that a deep neural network can misclassify the data even\nif the input data is perturbed by an imperceptible amount. In this paper, we\naddress the robustness issue of neural networks by a novel close-loop control\nmethod from the perspective of dynamic systems. Instead of modifying the\nparameters in a fixed neural network architecture, a close-loop control process\nis added to generate control signals adaptively for the perturbed or corrupted\ndata. We connect the robustness of neural networks with optimal control using\nthe geometrical information of underlying data to design the control objective.\nThe detailed analysis shows how the embedding manifolds of state trajectory\naffect error estimation of the proposed method. Our approach can simultaneously\nmaintain the performance on clean data and improve the robustness against many\ntypes of data perturbations. It can also further improve the performance of\nrobustly trained neural networks against different perturbations. To the best\nof our knowledge, this is the first work that improves the robustness of neural\nnetworks with close-loop control.', 'authors': ['Zhuotong Chen', 'Qianxiao Li', 'Zheng Zhang'], 'published': '2021-02-03 03:50:35+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2102.01862v2'}
42,612
2009.05359v5
2,009.05359
5
Activation Relaxation: A Local Dynamical Approximation to Backpropagation in the Brain
The backpropagation of error algorithm (backprop) has been instrumental in the recent success of deep learning. However, a key question remains as to whether backprop can be formulated in a manner suitable for implementation in neural circuitry. The primary challenge is to ensure that any candidate formulation uses only local information, rather than relying on global signals as in standard backprop. Recently several algorithms for approximating backprop using only local signals have been proposed. However, these algorithms typically impose other requirements which challenge biological plausibility: for example, requiring complex and precise connectivity schemes, or multiple sequential backwards phases with information being stored across phases. Here, we propose a novel algorithm, Activation Relaxation (AR), which is motivated by constructing the backpropagation gradient as the equilibrium point of a dynamical system. Our algorithm converges rapidly and robustly to the correct backpropagation gradients, requires only a single type of computational unit, utilises only a single parallel backwards relaxation phase, and can operate on arbitrary computation graphs. We illustrate these properties by training deep neural networks on visual classification tasks, and describe simplifications to the algorithm which remove further obstacles to neurobiological implementation (for example, the weight-transport problem, and the use of nonlinear derivatives), while preserving performance.
2020-09-11
{'id': '2009.05359v5', 'title': 'Activation Relaxation: A Local Dynamical Approximation to Backpropagation in the Brain', 'abstract': 'The backpropagation of error algorithm (backprop) has been instrumental in\nthe recent success of deep learning. However, a key question remains as to\nwhether backprop can be formulated in a manner suitable for implementation in\nneural circuitry. The primary challenge is to ensure that any candidate\nformulation uses only local information, rather than relying on global signals\nas in standard backprop. Recently several algorithms for approximating backprop\nusing only local signals have been proposed. However, these algorithms\ntypically impose other requirements which challenge biological plausibility:\nfor example, requiring complex and precise connectivity schemes, or multiple\nsequential backwards phases with information being stored across phases. Here,\nwe propose a novel algorithm, Activation Relaxation (AR), which is motivated by\nconstructing the backpropagation gradient as the equilibrium point of a\ndynamical system. Our algorithm converges rapidly and robustly to the correct\nbackpropagation gradients, requires only a single type of computational unit,\nutilises only a single parallel backwards relaxation phase, and can operate on\narbitrary computation graphs. We illustrate these properties by training deep\nneural networks on visual classification tasks, and describe simplifications to\nthe algorithm which remove further obstacles to neurobiological implementation\n(for example, the weight-transport problem, and the use of nonlinear\nderivatives), while preserving performance.', 'authors': ['Beren Millidge', 'Alexander Tschantz', 'Anil K Seth', 'Christopher L Buckley'], 'published': '2020-09-11 11:56:34+00:00', 'categories': ['cs.NE', 'cs.AI', 'cs.LG', 'q-bio.NC'], 'url': 'http://arxiv.org/abs/2009.05359v5'}
42,616
2104.00795v1
2,104.00795
1
No Cost Likelihood Manipulation at Test Time for Making Better Mistakes in Deep Networks
There has been increasing interest in building deep hierarchy-aware classifiers that aim to quantify and reduce the severity of mistakes, and not just reduce the number of errors. The idea is to exploit the label hierarchy (e.g., the WordNet ontology) and consider graph distances as a proxy for mistake severity. Surprisingly, on examining mistake-severity distributions of the top-1 prediction, we find that current state-of-the-art hierarchy-aware deep classifiers do not always show practical improvement over the standard cross-entropy baseline in making better mistakes. The reason for the reduction in average mistake-severity can be attributed to the increase in low-severity mistakes, which may also explain the noticeable drop in their accuracy. To this end, we use the classical Conditional Risk Minimization (CRM) framework for hierarchy-aware classification. Given a cost matrix and a reliable estimate of likelihoods (obtained from a trained network), CRM simply amends mistakes at inference time; it needs no extra hyperparameters and requires adding just a few lines of code to the standard cross-entropy baseline. It significantly outperforms the state-of-the-art and consistently obtains large reductions in the average hierarchical distance of top-$k$ predictions across datasets, with very little loss in accuracy. CRM, because of its simplicity, can be used with any off-the-shelf trained model that provides reliable likelihood estimates.
2021-04-01
{'id': '2104.00795v1', 'title': 'No Cost Likelihood Manipulation at Test Time for Making Better Mistakes in Deep Networks', 'abstract': 'There has been increasing interest in building deep hierarchy-aware\nclassifiers that aim to quantify and reduce the severity of mistakes, and not\njust reduce the number of errors. The idea is to exploit the label hierarchy\n(e.g., the WordNet ontology) and consider graph distances as a proxy for\nmistake severity. Surprisingly, on examining mistake-severity distributions of\nthe top-1 prediction, we find that current state-of-the-art hierarchy-aware\ndeep classifiers do not always show practical improvement over the standard\ncross-entropy baseline in making better mistakes. The reason for the reduction\nin average mistake-severity can be attributed to the increase in low-severity\nmistakes, which may also explain the noticeable drop in their accuracy. To this\nend, we use the classical Conditional Risk Minimization (CRM) framework for\nhierarchy-aware classification. Given a cost matrix and a reliable estimate of\nlikelihoods (obtained from a trained network), CRM simply amends mistakes at\ninference time; it needs no extra hyperparameters and requires adding just a\nfew lines of code to the standard cross-entropy baseline. It significantly\noutperforms the state-of-the-art and consistently obtains large reductions in\nthe average hierarchical distance of top-$k$ predictions across datasets, with\nvery little loss in accuracy. CRM, because of its simplicity, can be used with\nany off-the-shelf trained model that provides reliable likelihood estimates.', 'authors': ['Shyamgopal Karthik', 'Ameya Prabhu', 'Puneet K. Dokania', 'Vineet Gandhi'], 'published': '2021-04-01 22:40:25+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2104.00795v1'}
42,619
2104.03869v1
2,104.03869
1
Probing BERT in Hyperbolic Spaces
Recently, a variety of probing tasks are proposed to discover linguistic properties learned in contextualized word embeddings. Many of these works implicitly assume these embeddings lay in certain metric spaces, typically the Euclidean space. This work considers a family of geometrically special spaces, the hyperbolic spaces, that exhibit better inductive biases for hierarchical structures and may better reveal linguistic hierarchies encoded in contextualized representations. We introduce a Poincare probe, a structural probe projecting these embeddings into a Poincare subspace with explicitly defined hierarchies. We focus on two probing objectives: (a) dependency trees where the hierarchy is defined as head-dependent structures; (b) lexical sentiments where the hierarchy is defined as the polarity of words (positivity and negativity). We argue that a key desideratum of a probe is its sensitivity to the existence of linguistic structures. We apply our probes on BERT, a typical contextualized embedding model. In a syntactic subspace, our probe better recovers tree structures than Euclidean probes, revealing the possibility that the geometry of BERT syntax may not necessarily be Euclidean. In a sentiment subspace, we reveal two possible meta-embeddings for positive and negative sentiments and show how lexically-controlled contextualization would change the geometric localization of embeddings. We demonstrate the findings with our Poincare probe via extensive experiments and visualization. Our results can be reproduced at https://github.com/FranxYao/PoincareProbe.
2021-04-08
{'id': '2104.03869v1', 'title': 'Probing BERT in Hyperbolic Spaces', 'abstract': 'Recently, a variety of probing tasks are proposed to discover linguistic\nproperties learned in contextualized word embeddings. Many of these works\nimplicitly assume these embeddings lay in certain metric spaces, typically the\nEuclidean space. This work considers a family of geometrically special spaces,\nthe hyperbolic spaces, that exhibit better inductive biases for hierarchical\nstructures and may better reveal linguistic hierarchies encoded in\ncontextualized representations. We introduce a Poincare probe, a structural\nprobe projecting these embeddings into a Poincare subspace with explicitly\ndefined hierarchies. We focus on two probing objectives: (a) dependency trees\nwhere the hierarchy is defined as head-dependent structures; (b) lexical\nsentiments where the hierarchy is defined as the polarity of words (positivity\nand negativity). We argue that a key desideratum of a probe is its sensitivity\nto the existence of linguistic structures. We apply our probes on BERT, a\ntypical contextualized embedding model. In a syntactic subspace, our probe\nbetter recovers tree structures than Euclidean probes, revealing the\npossibility that the geometry of BERT syntax may not necessarily be Euclidean.\nIn a sentiment subspace, we reveal two possible meta-embeddings for positive\nand negative sentiments and show how lexically-controlled contextualization\nwould change the geometric localization of embeddings. We demonstrate the\nfindings with our Poincare probe via extensive experiments and visualization.\nOur results can be reproduced at https://github.com/FranxYao/PoincareProbe.', 'authors': ['Boli Chen', 'Yao Fu', 'Guangwei Xu', 'Pengjun Xie', 'Chuanqi Tan', 'Mosha Chen', 'Liping Jing'], 'published': '2021-04-08 16:24:53+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2104.03869v1'}
42,621
2102.10307v1
2,102.10307
1
Large-width functional asymptotics for deep Gaussian neural networks
In this paper, we consider fully connected feed-forward deep neural networks where weights and biases are independent and identically distributed according to Gaussian distributions. Extending previous results (Matthews et al., 2018a;b; Yang, 2019) we adopt a function-space perspective, i.e. we look at neural networks as infinite-dimensional random elements on the input space $R^I$. Under suitable assumptions on the activation function we show that: i) a network defines a continuous Gaussian process on the input space $R^I$; ii) a network with re-scaled weights converges weakly to a continuous Gaussian process in the large-width limit; iii) the limiting Gaussian process has almost surely locally $$-H\"older continuous paths, for $0 < <1$. Our results contribute to recent theoretical studies on the interplay between infinitely wide deep neural networks and Gaussian processes by establishing weak convergence in function-space with respect to a stronger metric.
2021-02-20
{'id': '2102.10307v1', 'title': 'Large-width functional asymptotics for deep Gaussian neural networks', 'abstract': 'In this paper, we consider fully connected feed-forward deep neural networks\nwhere weights and biases are independent and identically distributed according\nto Gaussian distributions. Extending previous results (Matthews et al.,\n2018a;b; Yang, 2019) we adopt a function-space perspective, i.e. we look at\nneural networks as infinite-dimensional random elements on the input space\n$\\mathbb{R}^I$. Under suitable assumptions on the activation function we show\nthat: i) a network defines a continuous Gaussian process on the input space\n$\\mathbb{R}^I$; ii) a network with re-scaled weights converges weakly to a\ncontinuous Gaussian process in the large-width limit; iii) the limiting\nGaussian process has almost surely locally $\\gamma$-H\\"older continuous paths,\nfor $0 < \\gamma <1$. Our results contribute to recent theoretical studies on\nthe interplay between infinitely wide deep neural networks and Gaussian\nprocesses by establishing weak convergence in function-space with respect to a\nstronger metric.', 'authors': ['Daniele Bracale', 'Stefano Favaro', 'Sandra Fortini', 'Stefano Peluchetti'], 'published': '2021-02-20 10:14:37+00:00', 'categories': ['math.PR', 'cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2102.10307v1'}
42,623
1905.11623v2
1,905.11623
2
Solving NP-Hard Problems on Graphs with Extended AlphaGo Zero
There have been increasing challenges to solve combinatorial optimization problems by machine learning. Khalil et al. proposed an end-to-end reinforcement learning framework, S2V-DQN, which automatically learns graph embeddings to construct solutions to a wide range of problems. To improve the generalization ability of their Q-learning method, we propose a novel learning strategy based on AlphaGo Zero which is a Go engine that achieved a superhuman level without the domain knowledge of the game. Our framework is redesigned for combinatorial problems, where the final reward might take any real number instead of a binary response, win/lose. In experiments conducted for five kinds of NP-hard problems including MinimumVertexCover and MaxCut, our method is shown to generalize better to various graphs than S2V-DQN. Furthermore, our method can be combined with recently-developed graph neural network (GNN) models such as the Graph Isomorphism Network, resulting in even better performance. This experiment also gives an interesting insight into a suitable choice of GNN models for each task.
2019-05-28
{'id': '1905.11623v2', 'title': 'Solving NP-Hard Problems on Graphs with Extended AlphaGo Zero', 'abstract': 'There have been increasing challenges to solve combinatorial optimization\nproblems by machine learning. Khalil et al. proposed an end-to-end\nreinforcement learning framework, S2V-DQN, which automatically learns graph\nembeddings to construct solutions to a wide range of problems. To improve the\ngeneralization ability of their Q-learning method, we propose a novel learning\nstrategy based on AlphaGo Zero which is a Go engine that achieved a superhuman\nlevel without the domain knowledge of the game. Our framework is redesigned for\ncombinatorial problems, where the final reward might take any real number\ninstead of a binary response, win/lose. In experiments conducted for five kinds\nof NP-hard problems including {\\sc MinimumVertexCover} and {\\sc MaxCut}, our\nmethod is shown to generalize better to various graphs than S2V-DQN.\nFurthermore, our method can be combined with recently-developed graph neural\nnetwork (GNN) models such as the \\emph{Graph Isomorphism Network}, resulting in\neven better performance. This experiment also gives an interesting insight into\na suitable choice of GNN models for each task.', 'authors': ['Kenshin Abe', 'Zijian Xu', 'Issei Sato', 'Masashi Sugiyama'], 'published': '2019-05-28 06:04:25+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/1905.11623v2'}
42,614
2305.14122v2
2,305.14122
2
Transferring Learning Trajectories of Neural Networks
Training deep neural networks (DNNs) is computationally expensive, which is problematic especially when performing duplicated or similar training runs in model ensemble or fine-tuning pre-trained models, for example. Once we have trained one DNN on some dataset, we have its learning trajectory (i.e., a sequence of intermediate parameters during training) which may potentially contain useful information for learning the dataset. However, there has been no attempt to utilize such information of a given learning trajectory for another training. In this paper, we formulate the problem of "transferring" a given learning trajectory from one initial parameter to another one (learning transfer problem) and derive the first algorithm to approximately solve it by matching gradients successively along the trajectory via permutation symmetry. We empirically show that the transferred parameters achieve non-trivial accuracy before any direct training, and can be trained significantly faster than training from scratch.
2023-05-23
{'id': '2305.14122v2', 'title': 'Transferring Learning Trajectories of Neural Networks', 'abstract': 'Training deep neural networks (DNNs) is computationally expensive, which is\nproblematic especially when performing duplicated or similar training runs in\nmodel ensemble or fine-tuning pre-trained models, for example. Once we have\ntrained one DNN on some dataset, we have its learning trajectory (i.e., a\nsequence of intermediate parameters during training) which may potentially\ncontain useful information for learning the dataset. However, there has been no\nattempt to utilize such information of a given learning trajectory for another\ntraining. In this paper, we formulate the problem of "transferring" a given\nlearning trajectory from one initial parameter to another one (learning\ntransfer problem) and derive the first algorithm to approximately solve it by\nmatching gradients successively along the trajectory via permutation symmetry.\nWe empirically show that the transferred parameters achieve non-trivial\naccuracy before any direct training, and can be trained significantly faster\nthan training from scratch.', 'authors': ['Daiki Chijiwa'], 'published': '2023-05-23 14:46:32+00:00', 'categories': ['cs.LG', 'cs.AI', 'stat.ML'], 'url': 'http://arxiv.org/abs/2305.14122v2'}
42,618
2310.00935v3
2,310.00935
3
Resolving Knowledge Conflicts in Large Language Models
Large language models (LLMs) often encounter knowledge conflicts, scenarios where discrepancy arises between the internal parametric knowledge of LLMs and non-parametric information provided in the prompt context. In this work we ask what are the desiderata for LLMs when a knowledge conflict arises and whether existing LLMs fulfill them. We posit that LLMs should 1) identify knowledge conflicts, 2) pinpoint conflicting information segments, and 3) provide distinct answers or viewpoints in conflicting scenarios. To this end, we introduce an evaluation framework for simulating contextual knowledge conflicts and quantitatively evaluating to what extent LLMs achieve these goals. It includes diverse and complex situations of knowledge conflict, knowledge from diverse entities and domains, two synthetic conflict creation methods, and settings with progressively increasing difficulty to reflect realistic knowledge conflicts. Extensive experiments with the framework reveal that while LLMs perform well in identifying the existence of knowledge conflicts, they struggle to determine the specific conflicting knowledge and produce a response with distinct answers amidst conflicting information. To address these challenges, we propose new instruction-based approaches that augment LLMs to better achieve the three goals. Further analysis shows that abilities to tackle knowledge conflicts are greatly impacted by factors such as knowledge domain, while generating robust responses to knowledge conflict scenarios remains an open research question.
2023-10-02
{'id': '2310.00935v3', 'title': 'Resolving Knowledge Conflicts in Large Language Models', 'abstract': 'Large language models (LLMs) often encounter knowledge conflicts, scenarios\nwhere discrepancy arises between the internal parametric knowledge of LLMs and\nnon-parametric information provided in the prompt context. In this work we ask\nwhat are the desiderata for LLMs when a knowledge conflict arises and whether\nexisting LLMs fulfill them. We posit that LLMs should 1) identify knowledge\nconflicts, 2) pinpoint conflicting information segments, and 3) provide\ndistinct answers or viewpoints in conflicting scenarios. To this end, we\nintroduce an evaluation framework for simulating contextual knowledge conflicts\nand quantitatively evaluating to what extent LLMs achieve these goals. It\nincludes diverse and complex situations of knowledge conflict, knowledge from\ndiverse entities and domains, two synthetic conflict creation methods, and\nsettings with progressively increasing difficulty to reflect realistic\nknowledge conflicts. Extensive experiments with the framework reveal that while\nLLMs perform well in identifying the existence of knowledge conflicts, they\nstruggle to determine the specific conflicting knowledge and produce a response\nwith distinct answers amidst conflicting information. To address these\nchallenges, we propose new instruction-based approaches that augment LLMs to\nbetter achieve the three goals. Further analysis shows that abilities to tackle\nknowledge conflicts are greatly impacted by factors such as knowledge domain,\nwhile generating robust responses to knowledge conflict scenarios remains an\nopen research question.', 'authors': ['Yike Wang', 'Shangbin Feng', 'Heng Wang', 'Weijia Shi', 'Vidhisha Balachandran', 'Tianxing He', 'Yulia Tsvetkov'], 'published': '2023-10-02 06:57:45+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2310.00935v3'}
42,622
2310.10616v1
2,310.10616
1
How Do Transformers Learn In-Context Beyond Simple Functions? A Case Study on Learning with Representations
While large language models based on the transformer architecture have demonstrated remarkable in-context learning (ICL) capabilities, understandings of such capabilities are still in an early stage, where existing theory and mechanistic understanding focus mostly on simple scenarios such as learning simple function classes. This paper takes initial steps on understanding ICL in more complex scenarios, by studying learning with representations. Concretely, we construct synthetic in-context learning problems with a compositional structure, where the label depends on the input through a possibly complex but fixed representation function, composed with a linear function that differs in each instance. By construction, the optimal ICL algorithm first transforms the inputs by the representation function, and then performs linear ICL on top of the transformed dataset. We show theoretically the existence of transformers that approximately implement such algorithms with mild depth and size. Empirically, we find trained transformers consistently achieve near-optimal ICL performance in this setting, and exhibit the desired dissection where lower layers transforms the dataset and upper layers perform linear ICL. Through extensive probing and a new pasting experiment, we further reveal several mechanisms within the trained transformers, such as concrete copying behaviors on both the inputs and the representations, linear ICL capability of the upper layers alone, and a post-ICL representation selection mechanism in a harder mixture setting. These observed mechanisms align well with our theory and may shed light on how transformers perform ICL in more realistic scenarios.
2023-10-16
{'id': '2310.10616v1', 'title': 'How Do Transformers Learn In-Context Beyond Simple Functions? A Case Study on Learning with Representations', 'abstract': 'While large language models based on the transformer architecture have\ndemonstrated remarkable in-context learning (ICL) capabilities, understandings\nof such capabilities are still in an early stage, where existing theory and\nmechanistic understanding focus mostly on simple scenarios such as learning\nsimple function classes. This paper takes initial steps on understanding ICL in\nmore complex scenarios, by studying learning with representations. Concretely,\nwe construct synthetic in-context learning problems with a compositional\nstructure, where the label depends on the input through a possibly complex but\nfixed representation function, composed with a linear function that differs in\neach instance. By construction, the optimal ICL algorithm first transforms the\ninputs by the representation function, and then performs linear ICL on top of\nthe transformed dataset. We show theoretically the existence of transformers\nthat approximately implement such algorithms with mild depth and size.\nEmpirically, we find trained transformers consistently achieve near-optimal ICL\nperformance in this setting, and exhibit the desired dissection where lower\nlayers transforms the dataset and upper layers perform linear ICL. Through\nextensive probing and a new pasting experiment, we further reveal several\nmechanisms within the trained transformers, such as concrete copying behaviors\non both the inputs and the representations, linear ICL capability of the upper\nlayers alone, and a post-ICL representation selection mechanism in a harder\nmixture setting. These observed mechanisms align well with our theory and may\nshed light on how transformers perform ICL in more realistic scenarios.', 'authors': ['Tianyu Guo', 'Wei Hu', 'Song Mei', 'Huan Wang', 'Caiming Xiong', 'Silvio Savarese', 'Yu Bai'], 'published': '2023-10-16 17:40:49+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.10616v1'}
42,625
2310.04416v2
2,310.04416
2
Alice Benchmarks: Connecting Real World Re-Identification with the Synthetic
For object re-identification (re-ID), learning from synthetic data has become a promising strategy to cheaply acquire large-scale annotated datasets and effective models, with few privacy concerns. Many interesting research problems arise from this strategy, e.g., how to reduce the domain gap between synthetic source and real-world target. To facilitate developing more new approaches in learning from synthetic data, we introduce the Alice benchmarks, large-scale datasets providing benchmarks as well as evaluation protocols to the research community. Within the Alice benchmarks, two object re-ID tasks are offered: person and vehicle re-ID. We collected and annotated two challenging real-world target datasets: AlicePerson and AliceVehicle, captured under various illuminations, image resolutions, etc. As an important feature of our real target, the clusterability of its training set is not manually guaranteed to make it closer to a real domain adaptation test scenario. Correspondingly, we reuse existing PersonX and VehicleX as synthetic source domains. The primary goal is to train models from synthetic data that can work effectively in the real world. In this paper, we detail the settings of Alice benchmarks, provide an analysis of existing commonly-used domain adaptation methods, and discuss some interesting future directions. An online server has been set up for the community to evaluate methods conveniently and fairly. Datasets and the online server details are available at https://sites.google.com/view/alice-benchmarks.
2023-10-06
{'id': '2310.04416v2', 'title': 'Alice Benchmarks: Connecting Real World Re-Identification with the Synthetic', 'abstract': 'For object re-identification (re-ID), learning from synthetic data has become\na promising strategy to cheaply acquire large-scale annotated datasets and\neffective models, with few privacy concerns. Many interesting research problems\narise from this strategy, e.g., how to reduce the domain gap between synthetic\nsource and real-world target. To facilitate developing more new approaches in\nlearning from synthetic data, we introduce the Alice benchmarks, large-scale\ndatasets providing benchmarks as well as evaluation protocols to the research\ncommunity. Within the Alice benchmarks, two object re-ID tasks are offered:\nperson and vehicle re-ID. We collected and annotated two challenging real-world\ntarget datasets: AlicePerson and AliceVehicle, captured under various\nilluminations, image resolutions, etc. As an important feature of our real\ntarget, the clusterability of its training set is not manually guaranteed to\nmake it closer to a real domain adaptation test scenario. Correspondingly, we\nreuse existing PersonX and VehicleX as synthetic source domains. The primary\ngoal is to train models from synthetic data that can work effectively in the\nreal world. In this paper, we detail the settings of Alice benchmarks, provide\nan analysis of existing commonly-used domain adaptation methods, and discuss\nsome interesting future directions. An online server has been set up for the\ncommunity to evaluate methods conveniently and fairly. Datasets and the online\nserver details are available at https://sites.google.com/view/alice-benchmarks.', 'authors': ['Xiaoxiao Sun', 'Yue Yao', 'Shengjin Wang', 'Hongdong Li', 'Liang Zheng'], 'published': '2023-10-06 17:58:26+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2310.04416v2'}
42,628
2310.17722v2
2,310.17722
2
Large Language Models as Generalizable Policies for Embodied Tasks
We show that large language models (LLMs) can be adapted to be generalizable policies for embodied visual tasks. Our approach, called Large LAnguage model Reinforcement Learning Policy (LLaRP), adapts a pre-trained frozen LLM to take as input text instructions and visual egocentric observations and output actions directly in the environment. Using reinforcement learning, we train LLaRP to see and act solely through environmental interactions. We show that LLaRP is robust to complex paraphrasings of task instructions and can generalize to new tasks that require novel optimal behavior. In particular, on 1,000 unseen tasks it achieves 42% success rate, 1.7x the success rate of other common learned baselines or zero-shot applications of LLMs. Finally, to aid the community in studying language conditioned, massively multi-task, embodied AI problems we release a novel benchmark, Language Rearrangement, consisting of 150,000 training and 1,000 testing tasks for language-conditioned rearrangement. Video examples of LLaRP in unseen Language Rearrangement instructions are at https://llm-rl.github.io.
2023-10-26
{'id': '2310.17722v2', 'title': 'Large Language Models as Generalizable Policies for Embodied Tasks', 'abstract': 'We show that large language models (LLMs) can be adapted to be generalizable\npolicies for embodied visual tasks. Our approach, called Large LAnguage model\nReinforcement Learning Policy (LLaRP), adapts a pre-trained frozen LLM to take\nas input text instructions and visual egocentric observations and output\nactions directly in the environment. Using reinforcement learning, we train\nLLaRP to see and act solely through environmental interactions. We show that\nLLaRP is robust to complex paraphrasings of task instructions and can\ngeneralize to new tasks that require novel optimal behavior. In particular, on\n1,000 unseen tasks it achieves 42% success rate, 1.7x the success rate of other\ncommon learned baselines or zero-shot applications of LLMs. Finally, to aid the\ncommunity in studying language conditioned, massively multi-task, embodied AI\nproblems we release a novel benchmark, Language Rearrangement, consisting of\n150,000 training and 1,000 testing tasks for language-conditioned\nrearrangement. Video examples of LLaRP in unseen Language Rearrangement\ninstructions are at https://llm-rl.github.io.', 'authors': ['Andrew Szot', 'Max Schwarzer', 'Harsh Agrawal', 'Bogdan Mazoure', 'Walter Talbott', 'Katherine Metcalf', 'Natalie Mackraz', 'Devon Hjelm', 'Alexander Toshev'], 'published': '2023-10-26 18:32:05+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CL'], 'url': 'http://arxiv.org/abs/2310.17722v2'}
42,632
2305.15357v5
2,305.15357
5
Solving Diffusion ODEs with Optimal Boundary Conditions for Better Image Super-Resolution
Diffusion models, as a kind of powerful generative model, have given impressive results on image super-resolution (SR) tasks. However, due to the randomness introduced in the reverse process of diffusion models, the performances of diffusion-based SR models are fluctuating at every time of sampling, especially for samplers with few resampled steps. This inherent randomness of diffusion models results in ineffectiveness and instability, making it challenging for users to guarantee the quality of SR results. However, our work takes this randomness as an opportunity: fully analyzing and leveraging it leads to the construction of an effective plug-and-play sampling method that owns the potential to benefit a series of diffusion-based SR methods. More in detail, we propose to steadily sample high-quality SR images from pre-trained diffusion-based SR models by solving diffusion ordinary differential equations (diffusion ODEs) with optimal boundary conditions (BCs) and analyze the characteristics between the choices of BCs and their corresponding SR results. Our analysis shows the route to obtain an approximately optimal BC via an efficient exploration in the whole space. The quality of SR results sampled by the proposed method with fewer steps outperforms the quality of results sampled by current methods with randomness from the same pre-trained diffusion-based SR model, which means that our sampling method "boosts" current diffusion-based SR models without any additional training.
2023-05-24
{'id': '2305.15357v5', 'title': 'Solving Diffusion ODEs with Optimal Boundary Conditions for Better Image Super-Resolution', 'abstract': 'Diffusion models, as a kind of powerful generative model, have given\nimpressive results on image super-resolution (SR) tasks. However, due to the\nrandomness introduced in the reverse process of diffusion models, the\nperformances of diffusion-based SR models are fluctuating at every time of\nsampling, especially for samplers with few resampled steps. This inherent\nrandomness of diffusion models results in ineffectiveness and instability,\nmaking it challenging for users to guarantee the quality of SR results.\nHowever, our work takes this randomness as an opportunity: fully analyzing and\nleveraging it leads to the construction of an effective plug-and-play sampling\nmethod that owns the potential to benefit a series of diffusion-based SR\nmethods. More in detail, we propose to steadily sample high-quality SR images\nfrom pre-trained diffusion-based SR models by solving diffusion ordinary\ndifferential equations (diffusion ODEs) with optimal boundary conditions (BCs)\nand analyze the characteristics between the choices of BCs and their\ncorresponding SR results. Our analysis shows the route to obtain an\napproximately optimal BC via an efficient exploration in the whole space. The\nquality of SR results sampled by the proposed method with fewer steps\noutperforms the quality of results sampled by current methods with randomness\nfrom the same pre-trained diffusion-based SR model, which means that our\nsampling method "boosts" current diffusion-based SR models without any\nadditional training.', 'authors': ['Yiyang Ma', 'Huan Yang', 'Wenhan Yang', 'Jianlong Fu', 'Jiaying Liu'], 'published': '2023-05-24 17:09:54+00:00', 'categories': ['eess.IV', 'cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2305.15357v5'}
42,626
2003.05033v5
2,003.05033
5
Generalized Energy Based Models
We introduce the Generalized Energy Based Model (GEBM) for generative modelling. These models combine two trained components: a base distribution (generally an implicit model), which can learn the support of data with low intrinsic dimension in a high dimensional space; and an energy function, to refine the probability mass on the learned support. Both the energy function and base jointly constitute the final model, unlike GANs, which retain only the base distribution (the "generator"). GEBMs are trained by alternating between learning the energy and the base. We show that both training stages are well-defined: the energy is learned by maximising a generalized likelihood, and the resulting energy-based loss provides informative gradients for learning the base. Samples from the posterior on the latent space of the trained model can be obtained via MCMC, thus finding regions in this space that produce better quality samples. Empirically, the GEBM samples on image-generation tasks are of much better quality than those from the learned generator alone, indicating that all else being equal, the GEBM will outperform a GAN of the same complexity. When using normalizing flows as base measures, GEBMs succeed on density modelling tasks, returning comparable performance to direct maximum likelihood of the same networks.
2020-03-10
{'id': '2003.05033v5', 'title': 'Generalized Energy Based Models', 'abstract': 'We introduce the Generalized Energy Based Model (GEBM) for generative\nmodelling. These models combine two trained components: a base distribution\n(generally an implicit model), which can learn the support of data with low\nintrinsic dimension in a high dimensional space; and an energy function, to\nrefine the probability mass on the learned support. Both the energy function\nand base jointly constitute the final model, unlike GANs, which retain only the\nbase distribution (the "generator"). GEBMs are trained by alternating between\nlearning the energy and the base. We show that both training stages are\nwell-defined: the energy is learned by maximising a generalized likelihood, and\nthe resulting energy-based loss provides informative gradients for learning the\nbase. Samples from the posterior on the latent space of the trained model can\nbe obtained via MCMC, thus finding regions in this space that produce better\nquality samples. Empirically, the GEBM samples on image-generation tasks are of\nmuch better quality than those from the learned generator alone, indicating\nthat all else being equal, the GEBM will outperform a GAN of the same\ncomplexity. When using normalizing flows as base measures, GEBMs succeed on\ndensity modelling tasks, returning comparable performance to direct maximum\nlikelihood of the same networks.', 'authors': ['Michael Arbel', 'Liang Zhou', 'Arthur Gretton'], 'published': '2020-03-10 23:22:09+00:00', 'categories': ['stat.ML', 'cs.LG'], 'url': 'http://arxiv.org/abs/2003.05033v5'}
42,645
2103.09716v5
2,103.09716
5
Quantitative Performance Assessment of CNN Units via Topological Entropy Calculation
Identifying the status of individual network units is critical for understanding the mechanism of convolutional neural networks (CNNs). However, it is still challenging to reliably give a general indication of unit status, especially for units in different network models. To this end, we propose a novel method for quantitatively clarifying the status of single unit in CNN using algebraic topological tools. Unit status is indicated via the calculation of a defined topological-based entropy, called feature entropy, which measures the degree of chaos of the global spatial pattern hidden in the unit for a category. In this way, feature entropy could provide an accurate indication of status for units in different networks with diverse situations like weight-rescaling operation. Further, we show that feature entropy decreases as the layer goes deeper and shares almost simultaneous trend with loss during training. We show that by investigating the feature entropy of units on only training data, it could give discrimination between networks with different generalization ability from the view of the effectiveness of feature representations.
2021-03-17
{'id': '2103.09716v5', 'title': 'Quantitative Performance Assessment of CNN Units via Topological Entropy Calculation', 'abstract': 'Identifying the status of individual network units is critical for\nunderstanding the mechanism of convolutional neural networks (CNNs). However,\nit is still challenging to reliably give a general indication of unit status,\nespecially for units in different network models. To this end, we propose a\nnovel method for quantitatively clarifying the status of single unit in CNN\nusing algebraic topological tools. Unit status is indicated via the calculation\nof a defined topological-based entropy, called feature entropy, which measures\nthe degree of chaos of the global spatial pattern hidden in the unit for a\ncategory. In this way, feature entropy could provide an accurate indication of\nstatus for units in different networks with diverse situations like\nweight-rescaling operation. Further, we show that feature entropy decreases as\nthe layer goes deeper and shares almost simultaneous trend with loss during\ntraining. We show that by investigating the feature entropy of units on only\ntraining data, it could give discrimination between networks with different\ngeneralization ability from the view of the effectiveness of feature\nrepresentations.', 'authors': ['Yang Zhao', 'Hao Zhang'], 'published': '2021-03-17 15:18:18+00:00', 'categories': ['cs.CV', 'cs.LG', 'I.2.10'], 'url': 'http://arxiv.org/abs/2103.09716v5'}
42,647
2110.04121v2
2,110.04121
2
On the Limitations of Multimodal VAEs
Multimodal variational autoencoders (VAEs) have shown promise as efficient generative models for weakly-supervised data. Yet, despite their advantage of weak supervision, they exhibit a gap in generative quality compared to unimodal VAEs, which are completely unsupervised. In an attempt to explain this gap, we uncover a fundamental limitation that applies to a large family of mixture-based multimodal VAEs. We prove that the sub-sampling of modalities enforces an undesirable upper bound on the multimodal ELBO and thereby limits the generative quality of the respective models. Empirically, we showcase the generative quality gap on both synthetic and real data and present the tradeoffs between different variants of multimodal VAEs. We find that none of the existing approaches fulfills all desired criteria of an effective multimodal generative model when applied on more complex datasets than those used in previous benchmarks. In summary, we identify, formalize, and validate fundamental limitations of VAE-based approaches for modeling weakly-supervised data and discuss implications for real-world applications.
2021-10-08
{'id': '2110.04121v2', 'title': 'On the Limitations of Multimodal VAEs', 'abstract': 'Multimodal variational autoencoders (VAEs) have shown promise as efficient\ngenerative models for weakly-supervised data. Yet, despite their advantage of\nweak supervision, they exhibit a gap in generative quality compared to unimodal\nVAEs, which are completely unsupervised. In an attempt to explain this gap, we\nuncover a fundamental limitation that applies to a large family of\nmixture-based multimodal VAEs. We prove that the sub-sampling of modalities\nenforces an undesirable upper bound on the multimodal ELBO and thereby limits\nthe generative quality of the respective models. Empirically, we showcase the\ngenerative quality gap on both synthetic and real data and present the\ntradeoffs between different variants of multimodal VAEs. We find that none of\nthe existing approaches fulfills all desired criteria of an effective\nmultimodal generative model when applied on more complex datasets than those\nused in previous benchmarks. In summary, we identify, formalize, and validate\nfundamental limitations of VAE-based approaches for modeling weakly-supervised\ndata and discuss implications for real-world applications.', 'authors': ['Imant Daunhawer', 'Thomas M. Sutter', 'Kieran Chin-Cheong', 'Emanuele Palumbo', 'Julia E. Vogt'], 'published': '2021-10-08 13:28:28+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2110.04121v2'}
42,650
2205.10186v3
2,205.10186
3
Bayesian Active Learning with Fully Bayesian Gaussian Processes
The bias-variance trade-off is a well-known problem in machine learning that only gets more pronounced the less available data there is. In active learning, where labeled data is scarce or difficult to obtain, neglecting this trade-off can cause inefficient and non-optimal querying, leading to unnecessary data labeling. In this paper, we focus on active learning with Gaussian Processes (GPs). For the GP, the bias-variance trade-off is made by optimization of the two hyperparameters: the length scale and noise-term. Considering that the optimal mode of the joint posterior of the hyperparameters is equivalent to the optimal bias-variance trade-off, we approximate this joint posterior and utilize it to design two new acquisition functions. The first one is a Bayesian variant of Query-by-Committee (B-QBC), and the second is an extension that explicitly minimizes the predictive variance through a Query by Mixture of Gaussian Processes (QB-MGP) formulation. Across six simulators, we empirically show that B-QBC, on average, achieves the best marginal likelihood, whereas QB-MGP achieves the best predictive performance. We show that incorporating the bias-variance trade-off in the acquisition functions mitigates unnecessary and expensive data labeling.
2022-05-20
{'id': '2205.10186v3', 'title': 'Bayesian Active Learning with Fully Bayesian Gaussian Processes', 'abstract': 'The bias-variance trade-off is a well-known problem in machine learning that\nonly gets more pronounced the less available data there is. In active learning,\nwhere labeled data is scarce or difficult to obtain, neglecting this trade-off\ncan cause inefficient and non-optimal querying, leading to unnecessary data\nlabeling. In this paper, we focus on active learning with Gaussian Processes\n(GPs). For the GP, the bias-variance trade-off is made by optimization of the\ntwo hyperparameters: the length scale and noise-term. Considering that the\noptimal mode of the joint posterior of the hyperparameters is equivalent to the\noptimal bias-variance trade-off, we approximate this joint posterior and\nutilize it to design two new acquisition functions. The first one is a Bayesian\nvariant of Query-by-Committee (B-QBC), and the second is an extension that\nexplicitly minimizes the predictive variance through a Query by Mixture of\nGaussian Processes (QB-MGP) formulation. Across six simulators, we empirically\nshow that B-QBC, on average, achieves the best marginal likelihood, whereas\nQB-MGP achieves the best predictive performance. We show that incorporating the\nbias-variance trade-off in the acquisition functions mitigates unnecessary and\nexpensive data labeling.', 'authors': ['Christoffer Riis', 'Francisco Antunes', 'Frederik Boe Hüttel', 'Carlos Lima Azevedo', 'Francisco Câmara Pereira'], 'published': '2022-05-20 13:52:04+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2205.10186v3'}
42,651
2110.04744v2
2,110.04744
2
Long Expressive Memory for Sequence Modeling
We propose a novel method called Long Expressive Memory (LEM) for learning long-term sequential dependencies. LEM is gradient-based, it can efficiently process sequential tasks with very long-term dependencies, and it is sufficiently expressive to be able to learn complicated input-output maps. To derive LEM, we consider a system of multiscale ordinary differential equations, as well as a suitable time-discretization of this system. For LEM, we derive rigorous bounds to show the mitigation of the exploding and vanishing gradients problem, a well-known challenge for gradient-based recurrent sequential learning methods. We also prove that LEM can approximate a large class of dynamical systems to high accuracy. Our empirical results, ranging from image and time-series classification through dynamical systems prediction to speech recognition and language modeling, demonstrate that LEM outperforms state-of-the-art recurrent neural networks, gated recurrent units, and long short-term memory models.
2021-10-10
{'id': '2110.04744v2', 'title': 'Long Expressive Memory for Sequence Modeling', 'abstract': 'We propose a novel method called Long Expressive Memory (LEM) for learning\nlong-term sequential dependencies. LEM is gradient-based, it can efficiently\nprocess sequential tasks with very long-term dependencies, and it is\nsufficiently expressive to be able to learn complicated input-output maps. To\nderive LEM, we consider a system of multiscale ordinary differential equations,\nas well as a suitable time-discretization of this system. For LEM, we derive\nrigorous bounds to show the mitigation of the exploding and vanishing gradients\nproblem, a well-known challenge for gradient-based recurrent sequential\nlearning methods. We also prove that LEM can approximate a large class of\ndynamical systems to high accuracy. Our empirical results, ranging from image\nand time-series classification through dynamical systems prediction to speech\nrecognition and language modeling, demonstrate that LEM outperforms\nstate-of-the-art recurrent neural networks, gated recurrent units, and long\nshort-term memory models.', 'authors': ['T. Konstantin Rusch', 'Siddhartha Mishra', 'N. Benjamin Erichson', 'Michael W. Mahoney'], 'published': '2021-10-10 09:43:02+00:00', 'categories': ['cs.LG', 'math.DS', 'stat.ML'], 'url': 'http://arxiv.org/abs/2110.04744v2'}
42,658
2112.03518v2
2,112.03518
2
Genetic Algorithm for Constrained Molecular Inverse Design
A genetic algorithm is suitable for exploring large search spaces as it finds an approximate solution. Because of this advantage, genetic algorithm is effective in exploring vast and unknown space such as molecular search space. Though the algorithm is suitable for searching vast chemical space, it is difficult to optimize pharmacological properties while maintaining molecular substructure. To solve this issue, we introduce a genetic algorithm featuring a constrained molecular inverse design. The proposed algorithm successfully produces valid molecules for crossover and mutation. Furthermore, it optimizes specific properties while adhering to structural constraints using a two-phase optimization. Experiments prove that our algorithm effectively finds molecules that satisfy specific properties while maintaining structural constraints.
2021-12-07
{'id': '2112.03518v2', 'title': 'Genetic Algorithm for Constrained Molecular Inverse Design', 'abstract': 'A genetic algorithm is suitable for exploring large search spaces as it finds\nan approximate solution. Because of this advantage, genetic algorithm is\neffective in exploring vast and unknown space such as molecular search space.\nThough the algorithm is suitable for searching vast chemical space, it is\ndifficult to optimize pharmacological properties while maintaining molecular\nsubstructure. To solve this issue, we introduce a genetic algorithm featuring a\nconstrained molecular inverse design. The proposed algorithm successfully\nproduces valid molecules for crossover and mutation. Furthermore, it optimizes\nspecific properties while adhering to structural constraints using a two-phase\noptimization. Experiments prove that our algorithm effectively finds molecules\nthat satisfy specific properties while maintaining structural constraints.', 'authors': ['Yurim Lee', 'Gydam Choi', 'Minsung Yoon', 'Cheongwon Kim'], 'published': '2021-12-07 05:58:44+00:00', 'categories': ['cs.NE', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2112.03518v2'}
42,627
2304.01910v4
2,304.0191
4
On the Variance of Neural Network Training with respect to Test Sets and Distributions
Typical neural network trainings have substantial variance in test-set performance between repeated runs, impeding hyperparameter comparison and training reproducibility. In this work we present the following results towards understanding this variation. (1) Despite having significant variance on their test-sets, we demonstrate that standard CIFAR-10 and ImageNet trainings have little variance in performance on the underlying test-distributions from which their test-sets are sampled. (2) We show that these trainings make approximately independent errors on their test-sets. That is, the event that a trained network makes an error on one particular example does not affect its chances of making errors on other examples, relative to their average rates over repeated runs of training with the same hyperparameters. (3) We prove that the variance of neural network trainings on their test-sets is a downstream consequence of the class-calibration property discovered by Jiang et al. (2021). Our analysis yields a simple formula which accurately predicts variance for the binary classification case. (4) We conduct preliminary studies of data augmentation, learning rate, finetuning instability and distribution-shift through the lens of variance between runs.
2023-04-04
{'id': '2304.01910v4', 'title': 'On the Variance of Neural Network Training with respect to Test Sets and Distributions', 'abstract': 'Typical neural network trainings have substantial variance in test-set\nperformance between repeated runs, impeding hyperparameter comparison and\ntraining reproducibility. In this work we present the following results towards\nunderstanding this variation. (1) Despite having significant variance on their\ntest-sets, we demonstrate that standard CIFAR-10 and ImageNet trainings have\nlittle variance in performance on the underlying test-distributions from which\ntheir test-sets are sampled. (2) We show that these trainings make\napproximately independent errors on their test-sets. That is, the event that a\ntrained network makes an error on one particular example does not affect its\nchances of making errors on other examples, relative to their average rates\nover repeated runs of training with the same hyperparameters. (3) We prove that\nthe variance of neural network trainings on their test-sets is a downstream\nconsequence of the class-calibration property discovered by Jiang et al.\n(2021). Our analysis yields a simple formula which accurately predicts variance\nfor the binary classification case. (4) We conduct preliminary studies of data\naugmentation, learning rate, finetuning instability and distribution-shift\nthrough the lens of variance between runs.', 'authors': ['Keller Jordan'], 'published': '2023-04-04 16:09:55+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2304.01910v4'}
42,630
2402.03921v2
2,402.03921
2
Large Language Models to Enhance Bayesian Optimization
Bayesian optimization (BO) is a powerful approach for optimizing complex and expensive-to-evaluate black-box functions. Its importance is underscored in many applications, notably including hyperparameter tuning, but its efficacy depends on efficiently balancing exploration and exploitation. While there has been substantial progress in BO methods, striking this balance remains a delicate process. In this light, we present LLAMBO, a novel approach that integrates the capabilities of Large Language Models (LLM) within BO. At a high level, we frame the BO problem in natural language, enabling LLMs to iteratively propose and evaluate promising solutions conditioned on historical evaluations. More specifically, we explore how combining contextual understanding, few-shot learning proficiency, and domain knowledge of LLMs can improve model-based BO. Our findings illustrate that LLAMBO is effective at zero-shot warmstarting, and enhances surrogate modeling and candidate sampling, especially in the early stages of search when observations are sparse. Our approach is performed in context and does not require LLM finetuning. Additionally, it is modular by design, allowing individual components to be integrated into existing BO frameworks, or function cohesively as an end-to-end method. We empirically validate LLAMBO's efficacy on the problem of hyperparameter tuning, highlighting strong empirical performance across a range of diverse benchmarks, proprietary, and synthetic tasks.
2024-02-06
{'id': '2402.03921v2', 'title': 'Large Language Models to Enhance Bayesian Optimization', 'abstract': "Bayesian optimization (BO) is a powerful approach for optimizing complex and\nexpensive-to-evaluate black-box functions. Its importance is underscored in\nmany applications, notably including hyperparameter tuning, but its efficacy\ndepends on efficiently balancing exploration and exploitation. While there has\nbeen substantial progress in BO methods, striking this balance remains a\ndelicate process. In this light, we present LLAMBO, a novel approach that\nintegrates the capabilities of Large Language Models (LLM) within BO. At a high\nlevel, we frame the BO problem in natural language, enabling LLMs to\niteratively propose and evaluate promising solutions conditioned on historical\nevaluations. More specifically, we explore how combining contextual\nunderstanding, few-shot learning proficiency, and domain knowledge of LLMs can\nimprove model-based BO. Our findings illustrate that LLAMBO is effective at\nzero-shot warmstarting, and enhances surrogate modeling and candidate sampling,\nespecially in the early stages of search when observations are sparse. Our\napproach is performed in context and does not require LLM finetuning.\nAdditionally, it is modular by design, allowing individual components to be\nintegrated into existing BO frameworks, or function cohesively as an end-to-end\nmethod. We empirically validate LLAMBO's efficacy on the problem of\nhyperparameter tuning, highlighting strong empirical performance across a range\nof diverse benchmarks, proprietary, and synthetic tasks.", 'authors': ['Tennison Liu', 'Nicolás Astorga', 'Nabeel Seedat', 'Mihaela van der Schaar'], 'published': '2024-02-06 11:44:06+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2402.03921v2'}
42,634
2501.06999v1
2,501.06999
1
Likelihood Training of Cascaded Diffusion Models via Hierarchical Volume-preserving Maps
Cascaded models are multi-scale generative models with a marked capacity for producing perceptually impressive samples at high resolutions. In this work, we show that they can also be excellent likelihood models, so long as we overcome a fundamental difficulty with probabilistic multi-scale models: the intractability of the likelihood function. Chiefly, in cascaded models each intermediary scale introduces extraneous variables that cannot be tractably marginalized out for likelihood evaluation. This issue vanishes by modeling the diffusion process on latent spaces induced by a class of transformations we call hierarchical volume-preserving maps, which decompose spatially structured data in a hierarchical fashion without introducing local distortions in the latent space. We demonstrate that two such maps are well-known in the literature for multiscale modeling: Laplacian pyramids and wavelet transforms. Not only do such reparameterizations allow the likelihood function to be directly expressed as a joint likelihood over the scales, we show that the Laplacian pyramid and wavelet transform also produces significant improvements to the state-of-the-art on a selection of benchmarks in likelihood modeling, including density estimation, lossless compression, and out-of-distribution detection. Investigating the theoretical basis of our empirical gains we uncover deep connections to score matching under the Earth Mover's Distance (EMD), which is a well-known surrogate for perceptual similarity. Code can be found at https://github.com/lihenryhfl/pcdmthis https url.
2025-01-13
{'id': '2501.06999v1', 'title': 'Likelihood Training of Cascaded Diffusion Models via Hierarchical Volume-preserving Maps', 'abstract': "Cascaded models are multi-scale generative models with a marked capacity for\nproducing perceptually impressive samples at high resolutions. In this work, we\nshow that they can also be excellent likelihood models, so long as we overcome\na fundamental difficulty with probabilistic multi-scale models: the\nintractability of the likelihood function. Chiefly, in cascaded models each\nintermediary scale introduces extraneous variables that cannot be tractably\nmarginalized out for likelihood evaluation. This issue vanishes by modeling the\ndiffusion process on latent spaces induced by a class of transformations we\ncall hierarchical volume-preserving maps, which decompose spatially structured\ndata in a hierarchical fashion without introducing local distortions in the\nlatent space. We demonstrate that two such maps are well-known in the\nliterature for multiscale modeling: Laplacian pyramids and wavelet transforms.\nNot only do such reparameterizations allow the likelihood function to be\ndirectly expressed as a joint likelihood over the scales, we show that the\nLaplacian pyramid and wavelet transform also produces significant improvements\nto the state-of-the-art on a selection of benchmarks in likelihood modeling,\nincluding density estimation, lossless compression, and out-of-distribution\ndetection. Investigating the theoretical basis of our empirical gains we\nuncover deep connections to score matching under the Earth Mover's Distance\n(EMD), which is a well-known surrogate for perceptual similarity. Code can be\nfound at \\href{https://github.com/lihenryhfl/pcdm}{this https url}.", 'authors': ['Henry Li', 'Ronen Basri', 'Yuval Kluger'], 'published': '2025-01-13 01:20:23+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2501.06999v1'}
42,636
2401.09671v2
2,401.09671
2
Towards Identifiable Unsupervised Domain Translation: A Diversified Distribution Matching Approach
Unsupervised domain translation (UDT) aims to find functions that convert samples from one domain (e.g., sketches) to another domain (e.g., photos) without changing the high-level semantic meaning (also referred to as ``content''). The translation functions are often sought by probability distribution matching of the transformed source domain and target domain. CycleGAN stands as arguably the most representative approach among this line of work. However, it was noticed in the literature that CycleGAN and variants could fail to identify the desired translation functions and produce content-misaligned translations. This limitation arises due to the presence of multiple translation functions -- referred to as ``measure-preserving automorphism" (MPA) -- in the solution space of the learning criteria. Despite awareness of such identifiability issues, solutions have remained elusive. This study delves into the core identifiability inquiry and introduces an MPA elimination theory. Our analysis shows that MPA is unlikely to exist, if multiple pairs of diverse cross-domain conditional distributions are matched by the learning function. Our theory leads to a UDT learner using distribution matching over auxiliary variable-induced subsets of the domains -- other than over the entire data domains as in the classical approaches. The proposed framework is the first to rigorously establish translation identifiability under reasonable UDT settings, to our best knowledge. Experiments corroborate with our theoretical claims.
2024-01-18
{'id': '2401.09671v2', 'title': 'Towards Identifiable Unsupervised Domain Translation: A Diversified Distribution Matching Approach', 'abstract': 'Unsupervised domain translation (UDT) aims to find functions that convert\nsamples from one domain (e.g., sketches) to another domain (e.g., photos)\nwithout changing the high-level semantic meaning (also referred to as\n``content\'\'). The translation functions are often sought by probability\ndistribution matching of the transformed source domain and target domain.\nCycleGAN stands as arguably the most representative approach among this line of\nwork. However, it was noticed in the literature that CycleGAN and variants\ncould fail to identify the desired translation functions and produce\ncontent-misaligned translations. This limitation arises due to the presence of\nmultiple translation functions -- referred to as ``measure-preserving\nautomorphism" (MPA) -- in the solution space of the learning criteria. Despite\nawareness of such identifiability issues, solutions have remained elusive. This\nstudy delves into the core identifiability inquiry and introduces an MPA\nelimination theory. Our analysis shows that MPA is unlikely to exist, if\nmultiple pairs of diverse cross-domain conditional distributions are matched by\nthe learning function. Our theory leads to a UDT learner using distribution\nmatching over auxiliary variable-induced subsets of the domains -- other than\nover the entire data domains as in the classical approaches. The proposed\nframework is the first to rigorously establish translation identifiability\nunder reasonable UDT settings, to our best knowledge. Experiments corroborate\nwith our theoretical claims.', 'authors': ['Sagar Shrestha', 'Xiao Fu'], 'published': '2024-01-18 01:07:00+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CV'], 'url': 'http://arxiv.org/abs/2401.09671v2'}
42,641
2310.03716v2
2,310.03716
2
A Long Way to Go: Investigating Length Correlations in RLHF
Great success has been reported using Reinforcement Learning from Human Feedback (RLHF) to align large language models, with open preference datasets enabling wider experimentation, particularly for "helpfulness" in tasks like dialogue and web question answering. Alongside these improvements, however, RLHF also often drives models to produce longer outputs. This paper demonstrates, on three diverse settings, that optimizing for response length is, much more than previously thought, a significant factor behind RLHF. Studying the strategies RL optimization uses to maximize reward, we find improvements in reward to largely be driven by increasing response length, instead of other features. Indeed, we find that even a purely length-based reward reproduces most downstream RLHF improvements over supervised fine-tuned models. Testing a comprehensive set of length-countering interventions, we identify the dominant source of these biases to be reward models, which, by studying training dynamics, we find are non-robust and easily influenced by length biases in preference data.
2023-10-05
{'id': '2310.03716v2', 'title': 'A Long Way to Go: Investigating Length Correlations in RLHF', 'abstract': 'Great success has been reported using Reinforcement Learning from Human\nFeedback (RLHF) to align large language models, with open preference datasets\nenabling wider experimentation, particularly for "helpfulness" in tasks like\ndialogue and web question answering. Alongside these improvements, however,\nRLHF also often drives models to produce longer outputs. This paper\ndemonstrates, on three diverse settings, that optimizing for response length\nis, much more than previously thought, a significant factor behind RLHF.\nStudying the strategies RL optimization uses to maximize reward, we find\nimprovements in reward to largely be driven by increasing response length,\ninstead of other features. Indeed, we find that even a purely length-based\nreward reproduces most downstream RLHF improvements over supervised fine-tuned\nmodels. Testing a comprehensive set of length-countering interventions, we\nidentify the dominant source of these biases to be reward models, which, by\nstudying training dynamics, we find are non-robust and easily influenced by\nlength biases in preference data.', 'authors': ['Prasann Singhal', 'Tanya Goyal', 'Jiacheng Xu', 'Greg Durrett'], 'published': '2023-10-05 17:38:28+00:00', 'categories': ['cs.CL', 'cs.LG'], 'url': 'http://arxiv.org/abs/2310.03716v2'}
42,631
2403.02484v1
2,403.02484
1
Encodings for Prediction-based Neural Architecture Search
Predictor-based methods have substantially enhanced Neural Architecture Search (NAS) optimization. The efficacy of these predictors is largely influenced by the method of encoding neural network architectures. While traditional encodings used an adjacency matrix describing the graph structure of a neural network, novel encodings embrace a variety of approaches from unsupervised pretraining of latent representations to vectors of zero-cost proxies. In this paper, we categorize and investigate neural encodings from three main types: structural, learned, and score-based. Furthermore, we extend these encodings and introduce unified encodings, that extend NAS predictors to multiple search spaces. Our analysis draws from experiments conducted on over 1.5 million neural network architectures on NAS spaces such as NASBench-101 (NB101), NB201, NB301, Network Design Spaces (NDS), and TransNASBench-101. Building on our study, we present our predictor FLAN: Flow Attention for NAS. FLAN integrates critical insights on predictor design, transfer learning, and unified encodings to enable more than an order of magnitude cost reduction for training NAS accuracy predictors. Our implementation and encodings for all neural networks are open-sourced at https://github.com/abdelfattah-lab/flan_nashttps://github.com/abdelfattah-lab/flan\_nas.
2024-03-04
{'id': '2403.02484v1', 'title': 'Encodings for Prediction-based Neural Architecture Search', 'abstract': 'Predictor-based methods have substantially enhanced Neural Architecture\nSearch (NAS) optimization. The efficacy of these predictors is largely\ninfluenced by the method of encoding neural network architectures. While\ntraditional encodings used an adjacency matrix describing the graph structure\nof a neural network, novel encodings embrace a variety of approaches from\nunsupervised pretraining of latent representations to vectors of zero-cost\nproxies. In this paper, we categorize and investigate neural encodings from\nthree main types: structural, learned, and score-based. Furthermore, we extend\nthese encodings and introduce \\textit{unified encodings}, that extend NAS\npredictors to multiple search spaces. Our analysis draws from experiments\nconducted on over 1.5 million neural network architectures on NAS spaces such\nas NASBench-101 (NB101), NB201, NB301, Network Design Spaces (NDS), and\nTransNASBench-101. Building on our study, we present our predictor\n\\textbf{FLAN}: \\textbf{Fl}ow \\textbf{A}ttention for \\textbf{N}AS. FLAN\nintegrates critical insights on predictor design, transfer learning, and\n\\textit{unified encodings} to enable more than an order of magnitude cost\nreduction for training NAS accuracy predictors. Our implementation and\nencodings for all neural networks are open-sourced at\n\\href{https://github.com/abdelfattah-lab/flan_nas}{https://github.com/abdelfattah-lab/flan\\_nas}.', 'authors': ['Yash Akhauri', 'Mohamed S. Abdelfattah'], 'published': '2024-03-04 21:05:52+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CV', 'cs.NE'], 'url': 'http://arxiv.org/abs/2403.02484v1'}
42,635
2209.09371v4
2,209.09371
4
Topological data analysis on noisy quantum computers
Topological data analysis (TDA) is a powerful technique for extracting complex and valuable shape-related summaries of high-dimensional data. However, the computational demands of classical algorithms for computing TDA are exorbitant, and quickly become impractical for high-order characteristics. Quantum computers offer the potential of achieving significant speedup for certain computational problems. Indeed, TDA has been purported to be one such problem, yet, quantum computing algorithms proposed for the problem, such as the original Quantum TDA (QTDA) formulation by Lloyd, Garnerone and Zanardi, require fault-tolerance qualifications that are currently unavailable. In this study, we present NISQ-TDA, a fully implemented end-to-end quantum machine learning algorithm needing only a short circuit-depth, that is applicable to high-dimensional classical data, and with provable asymptotic speedup for certain classes of problems. The algorithm neither suffers from the data-loading problem nor does it need to store the input data on the quantum computer explicitly. The algorithm was successfully executed on quantum computing devices, as well as on noisy quantum simulators, applied to small datasets. Preliminary empirical results suggest that the algorithm is robust to noise.
2022-09-19
{'id': '2209.09371v4', 'title': 'Topological data analysis on noisy quantum computers', 'abstract': 'Topological data analysis (TDA) is a powerful technique for extracting\ncomplex and valuable shape-related summaries of high-dimensional data. However,\nthe computational demands of classical algorithms for computing TDA are\nexorbitant, and quickly become impractical for high-order characteristics.\nQuantum computers offer the potential of achieving significant speedup for\ncertain computational problems. Indeed, TDA has been purported to be one such\nproblem, yet, quantum computing algorithms proposed for the problem, such as\nthe original Quantum TDA (QTDA) formulation by Lloyd, Garnerone and Zanardi,\nrequire fault-tolerance qualifications that are currently unavailable. In this\nstudy, we present NISQ-TDA, a fully implemented end-to-end quantum machine\nlearning algorithm needing only a short circuit-depth, that is applicable to\nhigh-dimensional classical data, and with provable asymptotic speedup for\ncertain classes of problems. The algorithm neither suffers from the\ndata-loading problem nor does it need to store the input data on the quantum\ncomputer explicitly. The algorithm was successfully executed on quantum\ncomputing devices, as well as on noisy quantum simulators, applied to small\ndatasets. Preliminary empirical results suggest that the algorithm is robust to\nnoise.', 'authors': ['Ismail Yunus Akhalwaya', 'Shashanka Ubaru', 'Kenneth L. Clarkson', 'Mark S. Squillante', 'Vishnu Jejjala', 'Yang-Hui He', 'Kugendran Naidoo', 'Vasileios Kalantzis', 'Lior Horesh'], 'published': '2022-09-19 22:45:00+00:00', 'categories': ['quant-ph', 'cs.LG', 'cs.NA', 'math.NA'], 'url': 'http://arxiv.org/abs/2209.09371v4'}
42,640
2309.16540v3
2,309.1654
3
Unsupervised Pretraining for Fact Verification by Language Model Distillation
Fact verification aims to verify a claim using evidence from a trustworthy knowledge base. To address this challenge, algorithms must produce features for every claim that are both semantically meaningful, and compact enough to find a semantic alignment with the source information. In contrast to previous work, which tackled the alignment problem by learning over annotated corpora of claims and their corresponding labels, we propose SFAVEL (Self-supervised Fact Verification via Language Model Distillation), a novel unsupervised pretraining framework that leverages pre-trained language models to distil self-supervised features into high-quality claim-fact alignments without the need for annotations. This is enabled by a novel contrastive loss function that encourages features to attain high-quality claim and evidence alignments whilst preserving the semantic relationships across the corpora. Notably, we present results that achieve a new state-of-the-art on FB15k-237 (+5.3% Hits@1) and FEVER (+8% accuracy) with linear evaluation.
2023-09-28
{'id': '2309.16540v3', 'title': 'Unsupervised Pretraining for Fact Verification by Language Model Distillation', 'abstract': 'Fact verification aims to verify a claim using evidence from a trustworthy\nknowledge base. To address this challenge, algorithms must produce features for\nevery claim that are both semantically meaningful, and compact enough to find a\nsemantic alignment with the source information. In contrast to previous work,\nwhich tackled the alignment problem by learning over annotated corpora of\nclaims and their corresponding labels, we propose SFAVEL (Self-supervised Fact\nVerification via Language Model Distillation), a novel unsupervised pretraining\nframework that leverages pre-trained language models to distil self-supervised\nfeatures into high-quality claim-fact alignments without the need for\nannotations. This is enabled by a novel contrastive loss function that\nencourages features to attain high-quality claim and evidence alignments whilst\npreserving the semantic relationships across the corpora. Notably, we present\nresults that achieve a new state-of-the-art on FB15k-237 (+5.3% Hits@1) and\nFEVER (+8% accuracy) with linear evaluation.', 'authors': ['Adrián Bazaga', 'Pietro Liò', 'Gos Micklem'], 'published': '2023-09-28 15:53:44+00:00', 'categories': ['cs.CL', 'cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2309.16540v3'}
42,648
2402.03131v1
2,402.03131
1
Constrained Decoding for Cross-lingual Label Projection
Zero-shot cross-lingual transfer utilizing multilingual LLMs has become a popular learning paradigm for low-resource languages with no labeled training data. However, for NLP tasks that involve fine-grained predictions on words and phrases, the performance of zero-shot cross-lingual transfer learning lags far behind supervised fine-tuning methods. Therefore, it is common to exploit translation and label projection to further improve the performance by (1) translating training data that is available in a high-resource language (e.g., English) together with the gold labels into low-resource languages, and/or (2) translating test data in low-resource languages to a high-source language to run inference on, then projecting the predicted span-level labels back onto the original test data. However, state-of-the-art marker-based label projection methods suffer from translation quality degradation due to the extra label markers injected in the input to the translation model. In this work, we explore a new direction that leverages constrained decoding for label projection to overcome the aforementioned issues. Our new method not only can preserve the quality of translated texts but also has the versatility of being applicable to both translating training and translating test data strategies. This versatility is crucial as our experiments reveal that translating test data can lead to a considerable boost in performance compared to translating only training data. We evaluate on two cross-lingual transfer tasks, namely Named Entity Recognition and Event Argument Extraction, spanning 20 languages. The results demonstrate that our approach outperforms the state-of-the-art marker-based method by a large margin and also shows better performance than other label projection methods that rely on external word alignment.
2024-02-05
{'id': '2402.03131v1', 'title': 'Constrained Decoding for Cross-lingual Label Projection', 'abstract': 'Zero-shot cross-lingual transfer utilizing multilingual LLMs has become a\npopular learning paradigm for low-resource languages with no labeled training\ndata. However, for NLP tasks that involve fine-grained predictions on words and\nphrases, the performance of zero-shot cross-lingual transfer learning lags far\nbehind supervised fine-tuning methods. Therefore, it is common to exploit\ntranslation and label projection to further improve the performance by (1)\ntranslating training data that is available in a high-resource language (e.g.,\nEnglish) together with the gold labels into low-resource languages, and/or (2)\ntranslating test data in low-resource languages to a high-source language to\nrun inference on, then projecting the predicted span-level labels back onto the\noriginal test data. However, state-of-the-art marker-based label projection\nmethods suffer from translation quality degradation due to the extra label\nmarkers injected in the input to the translation model. In this work, we\nexplore a new direction that leverages constrained decoding for label\nprojection to overcome the aforementioned issues. Our new method not only can\npreserve the quality of translated texts but also has the versatility of being\napplicable to both translating training and translating test data strategies.\nThis versatility is crucial as our experiments reveal that translating test\ndata can lead to a considerable boost in performance compared to translating\nonly training data. We evaluate on two cross-lingual transfer tasks, namely\nNamed Entity Recognition and Event Argument Extraction, spanning 20 languages.\nThe results demonstrate that our approach outperforms the state-of-the-art\nmarker-based method by a large margin and also shows better performance than\nother label projection methods that rely on external word alignment.', 'authors': ['Duong Minh Le', 'Yang Chen', 'Alan Ritter', 'Wei Xu'], 'published': '2024-02-05 15:57:32+00:00', 'categories': ['cs.CL', 'cs.LG'], 'url': 'http://arxiv.org/abs/2402.03131v1'}
42,656
2309.16108v4
2,309.16108
4
Channel Vision Transformers: An Image Is Worth 1 x 16 x 16 Words
Vision Transformer (ViT) has emerged as a powerful architecture in the realm of modern computer vision. However, its application in certain imaging fields, such as microscopy and satellite imaging, presents unique challenges. In these domains, images often contain multiple channels, each carrying semantically distinct and independent information. Furthermore, the model must demonstrate robustness to sparsity in input channels, as they may not be densely available during training or testing. In this paper, we propose a modification to the ViT architecture that enhances reasoning across the input channels and introduce Hierarchical Channel Sampling (HCS) as an additional regularization technique to ensure robustness when only partial channels are presented during test time. Our proposed model, ChannelViT, constructs patch tokens independently from each input channel and utilizes a learnable channel embedding that is added to the patch tokens, similar to positional embeddings. We evaluate the performance of ChannelViT on ImageNet, JUMP-CP (microscopy cell imaging), and So2Sat (satellite imaging). Our results show that ChannelViT outperforms ViT on classification tasks and generalizes well, even when a subset of input channels is used during testing. Across our experiments, HCS proves to be a powerful regularizer, independent of the architecture employed, suggesting itself as a straightforward technique for robust ViT training. Lastly, we find that ChannelViT generalizes effectively even when there is limited access to all channels during training, highlighting its potential for multi-channel imaging under real-world conditions with sparse sensors. Our code is available at https://github.com/insitro/ChannelViT.
2023-09-28
{'id': '2309.16108v4', 'title': 'Channel Vision Transformers: An Image Is Worth 1 x 16 x 16 Words', 'abstract': 'Vision Transformer (ViT) has emerged as a powerful architecture in the realm\nof modern computer vision. However, its application in certain imaging fields,\nsuch as microscopy and satellite imaging, presents unique challenges. In these\ndomains, images often contain multiple channels, each carrying semantically\ndistinct and independent information. Furthermore, the model must demonstrate\nrobustness to sparsity in input channels, as they may not be densely available\nduring training or testing. In this paper, we propose a modification to the ViT\narchitecture that enhances reasoning across the input channels and introduce\nHierarchical Channel Sampling (HCS) as an additional regularization technique\nto ensure robustness when only partial channels are presented during test time.\nOur proposed model, ChannelViT, constructs patch tokens independently from each\ninput channel and utilizes a learnable channel embedding that is added to the\npatch tokens, similar to positional embeddings. We evaluate the performance of\nChannelViT on ImageNet, JUMP-CP (microscopy cell imaging), and So2Sat\n(satellite imaging). Our results show that ChannelViT outperforms ViT on\nclassification tasks and generalizes well, even when a subset of input channels\nis used during testing. Across our experiments, HCS proves to be a powerful\nregularizer, independent of the architecture employed, suggesting itself as a\nstraightforward technique for robust ViT training. Lastly, we find that\nChannelViT generalizes effectively even when there is limited access to all\nchannels during training, highlighting its potential for multi-channel imaging\nunder real-world conditions with sparse sensors. Our code is available at\nhttps://github.com/insitro/ChannelViT.', 'authors': ['Yujia Bao', 'Srinivasan Sivanandan', 'Theofanis Karaletsos'], 'published': '2023-09-28 02:20:59+00:00', 'categories': ['cs.CV', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2309.16108v4'}
42,693
2205.12532v2
2,205.12532
2
Skill Machines: Temporal Logic Skill Composition in Reinforcement Learning
It is desirable for an agent to be able to solve a rich variety of problems that can be specified through language in the same environment. A popular approach towards obtaining such agents is to reuse skills learned in prior tasks to generalise compositionally to new ones. However, this is a challenging problem due to the curse of dimensionality induced by the combinatorially large number of ways high-level goals can be combined both logically and temporally in language. To address this problem, we propose a framework where an agent first learns a sufficient set of skill primitives to achieve all high-level goals in its environment. The agent can then flexibly compose them both logically and temporally to provably achieve temporal logic specifications in any regular language, such as regular fragments of linear temporal logic. This provides the agent with the ability to map from complex temporal logic task specifications to near-optimal behaviours zero-shot. We demonstrate this experimentally in a tabular setting, as well as in a high-dimensional video game and continuous control environment. Finally, we also demonstrate that the performance of skill machines can be improved with regular off-policy reinforcement learning algorithms when optimal behaviours are desired.
2022-05-25
{'id': '2205.12532v2', 'title': 'Skill Machines: Temporal Logic Skill Composition in Reinforcement Learning', 'abstract': 'It is desirable for an agent to be able to solve a rich variety of problems\nthat can be specified through language in the same environment. A popular\napproach towards obtaining such agents is to reuse skills learned in prior\ntasks to generalise compositionally to new ones. However, this is a challenging\nproblem due to the curse of dimensionality induced by the combinatorially large\nnumber of ways high-level goals can be combined both logically and temporally\nin language. To address this problem, we propose a framework where an agent\nfirst learns a sufficient set of skill primitives to achieve all high-level\ngoals in its environment. The agent can then flexibly compose them both\nlogically and temporally to provably achieve temporal logic specifications in\nany regular language, such as regular fragments of linear temporal logic. This\nprovides the agent with the ability to map from complex temporal logic task\nspecifications to near-optimal behaviours zero-shot. We demonstrate this\nexperimentally in a tabular setting, as well as in a high-dimensional video\ngame and continuous control environment. Finally, we also demonstrate that the\nperformance of skill machines can be improved with regular off-policy\nreinforcement learning algorithms when optimal behaviours are desired.', 'authors': ['Geraud Nangue Tasse', 'Devon Jarvis', 'Steven James', 'Benjamin Rosman'], 'published': '2022-05-25 07:05:24+00:00', 'categories': ['cs.LG', 'cs.LO'], 'url': 'http://arxiv.org/abs/2205.12532v2'}
42,637
2310.00808v2
2,310.00808
2
Completing Visual Objects via Bridging Generation and Segmentation
This paper presents a novel approach to object completion, with the primary goal of reconstructing a complete object from its partially visible components. Our method, named MaskComp, delineates the completion process through iterative stages of generation and segmentation. In each iteration, the object mask is provided as an additional condition to boost image generation, and, in return, the generated images can lead to a more accurate mask by fusing the segmentation of images. We demonstrate that the combination of one generation and one segmentation stage effectively functions as a mask denoiser. Through alternation between the generation and segmentation stages, the partial object mask is progressively refined, providing precise shape guidance and yielding superior object completion results. Our experiments demonstrate the superiority of MaskComp over existing approaches, e.g., ControlNet and Stable Diffusion, establishing it as an effective solution for object completion.
2023-10-01
{'id': '2310.00808v2', 'title': 'Completing Visual Objects via Bridging Generation and Segmentation', 'abstract': 'This paper presents a novel approach to object completion, with the primary\ngoal of reconstructing a complete object from its partially visible components.\nOur method, named MaskComp, delineates the completion process through iterative\nstages of generation and segmentation. In each iteration, the object mask is\nprovided as an additional condition to boost image generation, and, in return,\nthe generated images can lead to a more accurate mask by fusing the\nsegmentation of images. We demonstrate that the combination of one generation\nand one segmentation stage effectively functions as a mask denoiser. Through\nalternation between the generation and segmentation stages, the partial object\nmask is progressively refined, providing precise shape guidance and yielding\nsuperior object completion results. Our experiments demonstrate the superiority\nof MaskComp over existing approaches, e.g., ControlNet and Stable Diffusion,\nestablishing it as an effective solution for object completion.', 'authors': ['Xiang Li', 'Yinpeng Chen', 'Chung-Ching Lin', 'Hao Chen', 'Kai Hu', 'Rita Singh', 'Bhiksha Raj', 'Lijuan Wang', 'Zicheng Liu'], 'published': '2023-10-01 22:25:40+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2310.00808v2'}
42,639
2407.11451v1
2,407.11451
1
Isometric Representation Learning for Disentangled Latent Space of Diffusion Models
The latent space of diffusion model mostly still remains unexplored, despite its great success and potential in the field of generative modeling. In fact, the latent space of existing diffusion models are entangled, with a distorted mapping from its latent space to image space. To tackle this problem, we present Isometric Diffusion, equipping a diffusion model with a geometric regularizer to guide the model to learn a geometrically sound latent space of the training data manifold. This approach allows diffusion models to learn a more disentangled latent space, which enables smoother interpolation, more accurate inversion, and more precise control over attributes directly in the latent space. Our extensive experiments consisting of image interpolations, image inversions, and linear editing show the effectiveness of our method.
2024-07-16
{'id': '2407.11451v1', 'title': 'Isometric Representation Learning for Disentangled Latent Space of Diffusion Models', 'abstract': 'The latent space of diffusion model mostly still remains unexplored, despite\nits great success and potential in the field of generative modeling. In fact,\nthe latent space of existing diffusion models are entangled, with a distorted\nmapping from its latent space to image space. To tackle this problem, we\npresent Isometric Diffusion, equipping a diffusion model with a geometric\nregularizer to guide the model to learn a geometrically sound latent space of\nthe training data manifold. This approach allows diffusion models to learn a\nmore disentangled latent space, which enables smoother interpolation, more\naccurate inversion, and more precise control over attributes directly in the\nlatent space. Our extensive experiments consisting of image interpolations,\nimage inversions, and linear editing show the effectiveness of our method.', 'authors': ['Jaehoon Hahm', 'Junho Lee', 'Sunghyun Kim', 'Joonseok Lee'], 'published': '2024-07-16 07:36:01+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2407.11451v1'}
42,642
2402.02653v1
2,402.02653
1
Learning with Mixture of Prototypes for Out-of-Distribution Detection
Out-of-distribution (OOD) detection aims to detect testing samples far away from the in-distribution (ID) training data, which is crucial for the safe deployment of machine learning models in the real world. Distance-based OOD detection methods have emerged with enhanced deep representation learning. They identify unseen OOD samples by measuring their distances from ID class centroids or prototypes. However, existing approaches learn the representation relying on oversimplified data assumptions, e.g, modeling ID data of each class with one centroid class prototype or using loss functions not designed for OOD detection, which overlook the natural diversities within the data. Naively enforcing data samples of each class to be compact around only one prototype leads to inadequate modeling of realistic data and limited performance. To tackle these issues, we propose PrototypicAl Learning with a Mixture of prototypes (PALM) which models each class with multiple prototypes to capture the sample diversities, and learns more faithful and compact samples embeddings to enhance OOD detection. Our method automatically identifies and dynamically updates prototypes, assigning each sample to a subset of prototypes via reciprocal neighbor soft assignment weights. PALM optimizes a maximum likelihood estimation (MLE) loss to encourage the sample embeddings to be compact around the associated prototypes, as well as a contrastive loss on all prototypes to enhance intra-class compactness and inter-class discrimination at the prototype level. Moreover, the automatic estimation of prototypes enables our approach to be extended to the challenging OOD detection task with unlabelled ID data. Extensive experiments demonstrate the superiority of PALM, achieving state-of-the-art average AUROC performance of 93.82 on the challenging CIFAR-100 benchmark. Code is available at https://github.com/jeff024/PALM.
2024-02-05
{'id': '2402.02653v1', 'title': 'Learning with Mixture of Prototypes for Out-of-Distribution Detection', 'abstract': 'Out-of-distribution (OOD) detection aims to detect testing samples far away\nfrom the in-distribution (ID) training data, which is crucial for the safe\ndeployment of machine learning models in the real world. Distance-based OOD\ndetection methods have emerged with enhanced deep representation learning. They\nidentify unseen OOD samples by measuring their distances from ID class\ncentroids or prototypes. However, existing approaches learn the representation\nrelying on oversimplified data assumptions, e.g, modeling ID data of each class\nwith one centroid class prototype or using loss functions not designed for OOD\ndetection, which overlook the natural diversities within the data. Naively\nenforcing data samples of each class to be compact around only one prototype\nleads to inadequate modeling of realistic data and limited performance. To\ntackle these issues, we propose PrototypicAl Learning with a Mixture of\nprototypes (PALM) which models each class with multiple prototypes to capture\nthe sample diversities, and learns more faithful and compact samples embeddings\nto enhance OOD detection. Our method automatically identifies and dynamically\nupdates prototypes, assigning each sample to a subset of prototypes via\nreciprocal neighbor soft assignment weights. PALM optimizes a maximum\nlikelihood estimation (MLE) loss to encourage the sample embeddings to be\ncompact around the associated prototypes, as well as a contrastive loss on all\nprototypes to enhance intra-class compactness and inter-class discrimination at\nthe prototype level. Moreover, the automatic estimation of prototypes enables\nour approach to be extended to the challenging OOD detection task with\nunlabelled ID data. Extensive experiments demonstrate the superiority of PALM,\nachieving state-of-the-art average AUROC performance of 93.82 on the\nchallenging CIFAR-100 benchmark. Code is available at\nhttps://github.com/jeff024/PALM.', 'authors': ['Haodong Lu', 'Dong Gong', 'Shuo Wang', 'Jason Xue', 'Lina Yao', 'Kristen Moore'], 'published': '2024-02-05 00:52:50+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2402.02653v1'}
42,644
2312.04819v2
2,312.04819
2
Attention-Guided Contrastive Role Representations for Multi-Agent Reinforcement Learning
Real-world multi-agent tasks usually involve dynamic team composition with the emergence of roles, which should also be a key to efficient cooperation in multi-agent reinforcement learning (MARL). Drawing inspiration from the correlation between roles and agent's behavior patterns, we propose a novel framework of **A**ttention-guided **CO**ntrastive **R**ole representation learning for **M**ARL (**ACORM**) to promote behavior heterogeneity, knowledge transfer, and skillful coordination across agents. First, we introduce mutual information maximization to formalize role representation learning, derive a contrastive learning objective, and concisely approximate the distribution of negative pairs. Second, we leverage an attention mechanism to prompt the global state to attend to learned role representations in value decomposition, implicitly guiding agent coordination in a skillful role space to yield more expressive credit assignment. Experiments on challenging StarCraft II micromanagement and Google research football tasks demonstrate the state-of-the-art performance of our method and its advantages over existing approaches. Our code is available at [https://github.com/NJU-RL/ACORM](https://github.com/NJU-RL/ACORM).
2023-12-08
{'id': '2312.04819v2', 'title': 'Attention-Guided Contrastive Role Representations for Multi-Agent Reinforcement Learning', 'abstract': "Real-world multi-agent tasks usually involve dynamic team composition with\nthe emergence of roles, which should also be a key to efficient cooperation in\nmulti-agent reinforcement learning (MARL). Drawing inspiration from the\ncorrelation between roles and agent's behavior patterns, we propose a novel\nframework of **A**ttention-guided **CO**ntrastive **R**ole representation\nlearning for **M**ARL (**ACORM**) to promote behavior heterogeneity, knowledge\ntransfer, and skillful coordination across agents. First, we introduce mutual\ninformation maximization to formalize role representation learning, derive a\ncontrastive learning objective, and concisely approximate the distribution of\nnegative pairs. Second, we leverage an attention mechanism to prompt the global\nstate to attend to learned role representations in value decomposition,\nimplicitly guiding agent coordination in a skillful role space to yield more\nexpressive credit assignment. Experiments on challenging StarCraft II\nmicromanagement and Google research football tasks demonstrate the\nstate-of-the-art performance of our method and its advantages over existing\napproaches. Our code is available at\n[https://github.com/NJU-RL/ACORM](https://github.com/NJU-RL/ACORM).", 'authors': ['Zican Hu', 'Zongzhang Zhang', 'Huaxiong Li', 'Chunlin Chen', 'Hongyu Ding', 'Zhi Wang'], 'published': '2023-12-08 03:40:38+00:00', 'categories': ['cs.MA'], 'url': 'http://arxiv.org/abs/2312.04819v2'}
42,655
2310.07106v1
2,310.07106
1
The Temporal Structure of Language Processing in the Human Brain Corresponds to The Layered Hierarchy of Deep Language Models
Deep Language Models (DLMs) provide a novel computational paradigm for understanding the mechanisms of natural language processing in the human brain. Unlike traditional psycholinguistic models, DLMs use layered sequences of continuous numerical vectors to represent words and context, allowing a plethora of emerging applications such as human-like text generation. In this paper we show evidence that the layered hierarchy of DLMs may be used to model the temporal dynamics of language comprehension in the brain by demonstrating a strong correlation between DLM layer depth and the time at which layers are most predictive of the human brain. Our ability to temporally resolve individual layers benefits from our use of electrocorticography (ECoG) data, which has a much higher temporal resolution than noninvasive methods like fMRI. Using ECoG, we record neural activity from participants listening to a 30-minute narrative while also feeding the same narrative to a high-performing DLM (GPT2-XL). We then extract contextual embeddings from the different layers of the DLM and use linear encoding models to predict neural activity. We first focus on the Inferior Frontal Gyrus (IFG, or Broca's area) and then extend our model to track the increasing temporal receptive window along the linguistic processing hierarchy from auditory to syntactic and semantic areas. Our results reveal a connection between human language processing and DLMs, with the DLM's layer-by-layer accumulation of contextual information mirroring the timing of neural activity in high-order language areas.
2023-10-11
{'id': '2310.07106v1', 'title': 'The Temporal Structure of Language Processing in the Human Brain Corresponds to The Layered Hierarchy of Deep Language Models', 'abstract': "Deep Language Models (DLMs) provide a novel computational paradigm for\nunderstanding the mechanisms of natural language processing in the human brain.\nUnlike traditional psycholinguistic models, DLMs use layered sequences of\ncontinuous numerical vectors to represent words and context, allowing a\nplethora of emerging applications such as human-like text generation. In this\npaper we show evidence that the layered hierarchy of DLMs may be used to model\nthe temporal dynamics of language comprehension in the brain by demonstrating a\nstrong correlation between DLM layer depth and the time at which layers are\nmost predictive of the human brain. Our ability to temporally resolve\nindividual layers benefits from our use of electrocorticography (ECoG) data,\nwhich has a much higher temporal resolution than noninvasive methods like fMRI.\nUsing ECoG, we record neural activity from participants listening to a\n30-minute narrative while also feeding the same narrative to a high-performing\nDLM (GPT2-XL). We then extract contextual embeddings from the different layers\nof the DLM and use linear encoding models to predict neural activity. We first\nfocus on the Inferior Frontal Gyrus (IFG, or Broca's area) and then extend our\nmodel to track the increasing temporal receptive window along the linguistic\nprocessing hierarchy from auditory to syntactic and semantic areas. Our results\nreveal a connection between human language processing and DLMs, with the DLM's\nlayer-by-layer accumulation of contextual information mirroring the timing of\nneural activity in high-order language areas.", 'authors': ['Ariel Goldstein', 'Eric Ham', 'Mariano Schain', 'Samuel Nastase', 'Zaid Zada', 'Avigail Dabush', 'Bobbi Aubrey', 'Harshvardhan Gazula', 'Amir Feder', 'Werner K Doyle', 'Sasha Devore', 'Patricia Dugan', 'Daniel Friedman', 'Roi Reichart', 'Michael Brenner', 'Avinatan Hassidim', 'Orrin Devinsky', 'Adeen Flinker', 'Omer Levy', 'Uri Hasson'], 'published': '2023-10-11 01:03:42+00:00', 'categories': ['cs.CL', 'cs.AI', 'cs.LG', 'q-bio.NC'], 'url': 'http://arxiv.org/abs/2310.07106v1'}
42,638
2310.08566v2
2,310.08566
2
Transformers as Decision Makers: Provable In-Context Reinforcement Learning via Supervised Pretraining
Large transformer models pretrained on offline reinforcement learning datasets have demonstrated remarkable in-context reinforcement learning (ICRL) capabilities, where they can make good decisions when prompted with interaction trajectories from unseen environments. However, when and how transformers can be trained to perform ICRL have not been theoretically well-understood. In particular, it is unclear which reinforcement-learning algorithms transformers can perform in context, and how distribution mismatch in offline training data affects the learned algorithms. This paper provides a theoretical framework that analyzes supervised pretraining for ICRL. This includes two recently proposed training methods -- algorithm distillation and decision-pretrained transformers. First, assuming model realizability, we prove the supervised-pretrained transformer will imitate the conditional expectation of the expert algorithm given the observed trajectory. The generalization error will scale with model capacity and a distribution divergence factor between the expert and offline algorithms. Second, we show transformers with ReLU attention can efficiently approximate near-optimal online reinforcement learning algorithms like LinUCB and Thompson sampling for stochastic linear bandits, and UCB-VI for tabular Markov decision processes. This provides the first quantitative analysis of the ICRL capabilities of transformers pretrained from offline trajectories.
2023-10-12
{'id': '2310.08566v2', 'title': 'Transformers as Decision Makers: Provable In-Context Reinforcement Learning via Supervised Pretraining', 'abstract': 'Large transformer models pretrained on offline reinforcement learning\ndatasets have demonstrated remarkable in-context reinforcement learning (ICRL)\ncapabilities, where they can make good decisions when prompted with interaction\ntrajectories from unseen environments. However, when and how transformers can\nbe trained to perform ICRL have not been theoretically well-understood. In\nparticular, it is unclear which reinforcement-learning algorithms transformers\ncan perform in context, and how distribution mismatch in offline training data\naffects the learned algorithms. This paper provides a theoretical framework\nthat analyzes supervised pretraining for ICRL. This includes two recently\nproposed training methods -- algorithm distillation and decision-pretrained\ntransformers. First, assuming model realizability, we prove the\nsupervised-pretrained transformer will imitate the conditional expectation of\nthe expert algorithm given the observed trajectory. The generalization error\nwill scale with model capacity and a distribution divergence factor between the\nexpert and offline algorithms. Second, we show transformers with ReLU attention\ncan efficiently approximate near-optimal online reinforcement learning\nalgorithms like LinUCB and Thompson sampling for stochastic linear bandits, and\nUCB-VI for tabular Markov decision processes. This provides the first\nquantitative analysis of the ICRL capabilities of transformers pretrained from\noffline trajectories.', 'authors': ['Licong Lin', 'Yu Bai', 'Song Mei'], 'published': '2023-10-12 17:55:02+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CL', 'math.ST', 'stat.ML', 'stat.TH'], 'url': 'http://arxiv.org/abs/2310.08566v2'}
42,652
2312.12226v2
2,312.12226
2
On the Parameterization of Second-Order Optimization Effective Towards the Infinite Width
Second-order optimization has been developed to accelerate the training of deep neural networks and it is being applied to increasingly larger-scale models. In this study, towards training on further larger scales, we identify a specific parameterization for second-order optimization that promotes feature learning in a stable manner even if the network width increases significantly. Inspired by a maximal update parameterization, we consider a one-step update of the gradient and reveal the appropriate scales of hyperparameters including random initialization, learning rates, and damping terms. Our approach covers two major second-order optimization algorithms, K-FAC and Shampoo, and we demonstrate that our parameterization achieves higher generalization performance in feature learning. In particular, it enables us to transfer the hyperparameters across models with different widths.
2023-12-19
{'id': '2312.12226v2', 'title': 'On the Parameterization of Second-Order Optimization Effective Towards the Infinite Width', 'abstract': 'Second-order optimization has been developed to accelerate the training of\ndeep neural networks and it is being applied to increasingly larger-scale\nmodels. In this study, towards training on further larger scales, we identify a\nspecific parameterization for second-order optimization that promotes feature\nlearning in a stable manner even if the network width increases significantly.\nInspired by a maximal update parameterization, we consider a one-step update of\nthe gradient and reveal the appropriate scales of hyperparameters including\nrandom initialization, learning rates, and damping terms. Our approach covers\ntwo major second-order optimization algorithms, K-FAC and Shampoo, and we\ndemonstrate that our parameterization achieves higher generalization\nperformance in feature learning. In particular, it enables us to transfer the\nhyperparameters across models with different widths.', 'authors': ['Satoki Ishikawa', 'Ryo Karakida'], 'published': '2023-12-19 15:12:39+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2312.12226v2'}
42,657
2310.07246v2
2,310.07246
2
Vec-Tok Speech: speech vectorization and tokenization for neural speech generation
Language models (LMs) have recently flourished in natural language processing and computer vision, generating high-fidelity texts or images in various tasks. In contrast, the current speech generative models are still struggling regarding speech quality and task generalization. This paper presents Vec-Tok Speech, an extensible framework that resembles multiple speech generation tasks, generating expressive and high-fidelity speech. Specifically, we propose a novel speech codec based on speech vectors and semantic tokens. Speech vectors contain acoustic details contributing to high-fidelity speech reconstruction, while semantic tokens focus on the linguistic content of speech, facilitating language modeling. Based on the proposed speech codec, Vec-Tok Speech leverages an LM to undertake the core of speech generation. Moreover, Byte-Pair Encoding (BPE) is introduced to reduce the token length and bit rate for lower exposure bias and longer context coverage, improving the performance of LMs. Vec-Tok Speech can be used for intra- and cross-lingual zero-shot voice conversion (VC), zero-shot speaking style transfer text-to-speech (TTS), speech-to-speech translation (S2ST), speech denoising, and speaker de-identification and anonymization. Experiments show that Vec-Tok Speech, built on 50k hours of speech, performs better than other SOTA models. Code will be available at https://github.com/BakerBunker/VecTok .
2023-10-11
{'id': '2310.07246v2', 'title': 'Vec-Tok Speech: speech vectorization and tokenization for neural speech generation', 'abstract': 'Language models (LMs) have recently flourished in natural language processing\nand computer vision, generating high-fidelity texts or images in various tasks.\nIn contrast, the current speech generative models are still struggling\nregarding speech quality and task generalization. This paper presents Vec-Tok\nSpeech, an extensible framework that resembles multiple speech generation\ntasks, generating expressive and high-fidelity speech. Specifically, we propose\na novel speech codec based on speech vectors and semantic tokens. Speech\nvectors contain acoustic details contributing to high-fidelity speech\nreconstruction, while semantic tokens focus on the linguistic content of\nspeech, facilitating language modeling. Based on the proposed speech codec,\nVec-Tok Speech leverages an LM to undertake the core of speech generation.\nMoreover, Byte-Pair Encoding (BPE) is introduced to reduce the token length and\nbit rate for lower exposure bias and longer context coverage, improving the\nperformance of LMs. Vec-Tok Speech can be used for intra- and cross-lingual\nzero-shot voice conversion (VC), zero-shot speaking style transfer\ntext-to-speech (TTS), speech-to-speech translation (S2ST), speech denoising,\nand speaker de-identification and anonymization. Experiments show that Vec-Tok\nSpeech, built on 50k hours of speech, performs better than other SOTA models.\nCode will be available at https://github.com/BakerBunker/VecTok .', 'authors': ['Xinfa Zhu', 'Yuanjun Lv', 'Yi Lei', 'Tao Li', 'Wendi He', 'Hongbin Zhou', 'Heng Lu', 'Lei Xie'], 'published': '2023-10-11 07:23:27+00:00', 'categories': ['cs.SD', 'eess.AS'], 'url': 'http://arxiv.org/abs/2310.07246v2'}
42,665
2311.01410v2
2,311.0141
2
The Blessing of Randomness: SDE Beats ODE in General Diffusion-based Image Editing
We present a unified probabilistic formulation for diffusion-based image editing, where a latent variable is edited in a task-specific manner and generally deviates from the corresponding marginal distribution induced by the original stochastic or ordinary differential equation (SDE or ODE). Instead, it defines a corresponding SDE or ODE for editing. In the formulation, we prove that the Kullback-Leibler divergence between the marginal distributions of the two SDEs gradually decreases while that for the ODEs remains as the time approaches zero, which shows the promise of SDE in image editing. Inspired by it, we provide the SDE counterparts for widely used ODE baselines in various tasks including inpainting and image-to-image translation, where SDE shows a consistent and substantial improvement. Moreover, we propose SDE-Drag -- a simple yet effective method built upon the SDE formulation for point-based content dragging. We build a challenging benchmark (termed DragBench) with open-set natural, art, and AI-generated images for evaluation. A user study on DragBench indicates that SDE-Drag significantly outperforms our ODE baseline, existing diffusion-based methods, and the renowned DragGAN. Our results demonstrate the superiority and versatility of SDE in image editing and push the boundary of diffusion-based editing methods.
2023-11-02
{'id': '2311.01410v2', 'title': 'The Blessing of Randomness: SDE Beats ODE in General Diffusion-based Image Editing', 'abstract': 'We present a unified probabilistic formulation for diffusion-based image\nediting, where a latent variable is edited in a task-specific manner and\ngenerally deviates from the corresponding marginal distribution induced by the\noriginal stochastic or ordinary differential equation (SDE or ODE). Instead, it\ndefines a corresponding SDE or ODE for editing. In the formulation, we prove\nthat the Kullback-Leibler divergence between the marginal distributions of the\ntwo SDEs gradually decreases while that for the ODEs remains as the time\napproaches zero, which shows the promise of SDE in image editing. Inspired by\nit, we provide the SDE counterparts for widely used ODE baselines in various\ntasks including inpainting and image-to-image translation, where SDE shows a\nconsistent and substantial improvement. Moreover, we propose SDE-Drag -- a\nsimple yet effective method built upon the SDE formulation for point-based\ncontent dragging. We build a challenging benchmark (termed DragBench) with\nopen-set natural, art, and AI-generated images for evaluation. A user study on\nDragBench indicates that SDE-Drag significantly outperforms our ODE baseline,\nexisting diffusion-based methods, and the renowned DragGAN. Our results\ndemonstrate the superiority and versatility of SDE in image editing and push\nthe boundary of diffusion-based editing methods.', 'authors': ['Shen Nie', 'Hanzhong Allan Guo', 'Cheng Lu', 'Yuhao Zhou', 'Chenyu Zheng', 'Chongxuan Li'], 'published': '2023-11-02 17:23:14+00:00', 'categories': ['cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2311.01410v2'}
42,681
2310.01362v3
2,310.01362
3
Robot Fleet Learning via Policy Merging
Fleets of robots ingest massive amounts of heterogeneous streaming data silos generated by interacting with their environments, far more than what can be stored or transmitted with ease. At the same time, teams of robots should co-acquire diverse skills through their heterogeneous experiences in varied settings. How can we enable such fleet-level learning without having to transmit or centralize fleet-scale data? In this paper, we investigate policy merging (PoMe) from such distributed heterogeneous datasets as a potential solution. To efficiently merge policies in the fleet setting, we propose FLEET-MERGE, an instantiation of distributed learning that accounts for the permutation invariance that arises when parameterizing the control policies with recurrent neural networks. We show that FLEET-MERGE consolidates the behavior of policies trained on 50 tasks in the Meta-World environment, with good performance on nearly all training tasks at test time. Moreover, we introduce a novel robotic tool-use benchmark, FLEET-TOOLS, for fleet policy learning in compositional and contact-rich robot manipulation tasks, to validate the efficacy of FLEET-MERGE on the benchmark.
2023-10-02
{'id': '2310.01362v3', 'title': 'Robot Fleet Learning via Policy Merging', 'abstract': 'Fleets of robots ingest massive amounts of heterogeneous streaming data silos\ngenerated by interacting with their environments, far more than what can be\nstored or transmitted with ease. At the same time, teams of robots should\nco-acquire diverse skills through their heterogeneous experiences in varied\nsettings. How can we enable such fleet-level learning without having to\ntransmit or centralize fleet-scale data? In this paper, we investigate policy\nmerging (PoMe) from such distributed heterogeneous datasets as a potential\nsolution. To efficiently merge policies in the fleet setting, we propose\nFLEET-MERGE, an instantiation of distributed learning that accounts for the\npermutation invariance that arises when parameterizing the control policies\nwith recurrent neural networks. We show that FLEET-MERGE consolidates the\nbehavior of policies trained on 50 tasks in the Meta-World environment, with\ngood performance on nearly all training tasks at test time. Moreover, we\nintroduce a novel robotic tool-use benchmark, FLEET-TOOLS, for fleet policy\nlearning in compositional and contact-rich robot manipulation tasks, to\nvalidate the efficacy of FLEET-MERGE on the benchmark.', 'authors': ['Lirui Wang', 'Kaiqing Zhang', 'Allan Zhou', 'Max Simchowitz', 'Russ Tedrake'], 'published': '2023-10-02 17:23:51+00:00', 'categories': ['cs.RO', 'cs.LG'], 'url': 'http://arxiv.org/abs/2310.01362v3'}
42,688
2106.02190v6
2,106.0219
6
Spatial Graph Attention and Curiosity-driven Policy for Antiviral Drug Discovery
We developed Distilled Graph Attention Policy Network (DGAPN), a reinforcement learning model to generate novel graph-structured chemical representations that optimize user-defined objectives by efficiently navigating a physically constrained domain. The framework is examined on the task of generating molecules that are designed to bind, noncovalently, to functional sites of SARS-CoV-2 proteins. We present a spatial Graph Attention (sGAT) mechanism that leverages self-attention over both node and edge attributes as well as encoding the spatial structure -- this capability is of considerable interest in synthetic biology and drug discovery. An attentional policy network is introduced to learn the decision rules for a dynamic, fragment-based chemical environment, and state-of-the-art policy gradient techniques are employed to train the network with stability. Exploration is driven by the stochasticity of the action space design and the innovation reward bonuses learned and proposed by random network distillation. In experiments, our framework achieved outstanding results compared to state-of-the-art algorithms, while reducing the complexity of paths to chemical synthesis.
2021-06-04
{'id': '2106.02190v6', 'title': 'Spatial Graph Attention and Curiosity-driven Policy for Antiviral Drug Discovery', 'abstract': 'We developed Distilled Graph Attention Policy Network (DGAPN), a\nreinforcement learning model to generate novel graph-structured chemical\nrepresentations that optimize user-defined objectives by efficiently navigating\na physically constrained domain. The framework is examined on the task of\ngenerating molecules that are designed to bind, noncovalently, to functional\nsites of SARS-CoV-2 proteins. We present a spatial Graph Attention (sGAT)\nmechanism that leverages self-attention over both node and edge attributes as\nwell as encoding the spatial structure -- this capability is of considerable\ninterest in synthetic biology and drug discovery. An attentional policy network\nis introduced to learn the decision rules for a dynamic, fragment-based\nchemical environment, and state-of-the-art policy gradient techniques are\nemployed to train the network with stability. Exploration is driven by the\nstochasticity of the action space design and the innovation reward bonuses\nlearned and proposed by random network distillation. In experiments, our\nframework achieved outstanding results compared to state-of-the-art algorithms,\nwhile reducing the complexity of paths to chemical synthesis.', 'authors': ['Yulun Wu', 'Mikaela Cashman', 'Nicholas Choma', 'Érica T. Prates', 'Verónica G. Melesse Vergara', 'Manesh Shah', 'Andrew Chen', 'Austin Clyde', 'Thomas S. Brettin', 'Wibe A. de Jong', 'Neeraj Kumar', 'Martha S. Head', 'Rick L. Stevens', 'Peter Nugent', 'Daniel A. Jacobson', 'James B. Brown'], 'published': '2021-06-04 00:36:47+00:00', 'categories': ['cs.LG', 'cs.AI', 'q-bio.BM'], 'url': 'http://arxiv.org/abs/2106.02190v6'}
42,643
2310.04971v2
2,310.04971
2
Understanding the Robustness of Multi-modal Contrastive Learning to Distribution Shift
Recently, multimodal contrastive learning (MMCL) approaches, such as CLIP, have achieved a remarkable success in learning representations that are robust against distribution shift and generalize to new domains. Despite the empirical success, the mechanism behind learning such generalizable representations is not understood. In this work, we rigorously analyze this problem and uncover two mechanisms behind MMCL's robustness: intra-class contrasting, which allows the model to learn features with a high variance, and inter-class feature sharing, where annotated details in one class help learning other classes better. Both mechanisms prevent spurious features that are over-represented in the training data to overshadow the generalizable core features. This yields superior zero-shot classification accuracy under distribution shift. Furthermore, we theoretically demonstrate the benefits of using rich captions on robustness and explore the effect of annotating different types of details in the captions. We validate our theoretical findings through experiments, including a well-designed synthetic experiment and an experiment involving training CLIP models on MSCOCO/Conceptual Captions and evaluating them on shifted ImageNets.
2023-10-08
{'id': '2310.04971v2', 'title': 'Understanding the Robustness of Multi-modal Contrastive Learning to Distribution Shift', 'abstract': "Recently, multimodal contrastive learning (MMCL) approaches, such as CLIP,\nhave achieved a remarkable success in learning representations that are robust\nagainst distribution shift and generalize to new domains. Despite the empirical\nsuccess, the mechanism behind learning such generalizable representations is\nnot understood. In this work, we rigorously analyze this problem and uncover\ntwo mechanisms behind MMCL's robustness: \\emph{intra-class contrasting}, which\nallows the model to learn features with a high variance, and \\emph{inter-class\nfeature sharing}, where annotated details in one class help learning other\nclasses better. Both mechanisms prevent spurious features that are\nover-represented in the training data to overshadow the generalizable core\nfeatures. This yields superior zero-shot classification accuracy under\ndistribution shift. Furthermore, we theoretically demonstrate the benefits of\nusing rich captions on robustness and explore the effect of annotating\ndifferent types of details in the captions. We validate our theoretical\nfindings through experiments, including a well-designed synthetic experiment\nand an experiment involving training CLIP models on MSCOCO/Conceptual Captions\nand evaluating them on shifted ImageNets.", 'authors': ['Yihao Xue', 'Siddharth Joshi', 'Dang Nguyen', 'Baharan Mirzasoleiman'], 'published': '2023-10-08 02:25:52+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.04971v2'}
42,660
2312.09234v3
2,312.09234
3
Let's do the time-warp-attend: Learning topological invariants of dynamical systems
Dynamical systems across the sciences, from electrical circuits to ecological networks, undergo qualitative and often catastrophic changes in behavior, called bifurcations, when their underlying parameters cross a threshold. Existing methods predict oncoming catastrophes in individual systems but are primarily time-series-based and struggle both to categorize qualitative dynamical regimes across diverse systems and to generalize to real data. To address this challenge, we propose a data-driven, physically-informed deep-learning framework for classifying dynamical regimes and characterizing bifurcation boundaries based on the extraction of topologically invariant features. We focus on the paradigmatic case of the supercritical Hopf bifurcation, which is used to model periodic dynamics across a wide range of applications. Our convolutional attention method is trained with data augmentations that encourage the learning of topological invariants which can be used to detect bifurcation boundaries in unseen systems and to design models of biological systems like oscillatory gene regulatory networks. We further demonstrate our method's use in analyzing real data by recovering distinct proliferation and differentiation dynamics along pancreatic endocrinogenesis trajectory in gene expression space based on single-cell data. Our method provides valuable insights into the qualitative, long-term behavior of a wide range of dynamical systems, and can detect bifurcations or catastrophic transitions in large-scale physical and biological systems.
2023-12-14
{'id': '2312.09234v3', 'title': "Let's do the time-warp-attend: Learning topological invariants of dynamical systems", 'abstract': "Dynamical systems across the sciences, from electrical circuits to ecological\nnetworks, undergo qualitative and often catastrophic changes in behavior,\ncalled bifurcations, when their underlying parameters cross a threshold.\nExisting methods predict oncoming catastrophes in individual systems but are\nprimarily time-series-based and struggle both to categorize qualitative\ndynamical regimes across diverse systems and to generalize to real data. To\naddress this challenge, we propose a data-driven, physically-informed\ndeep-learning framework for classifying dynamical regimes and characterizing\nbifurcation boundaries based on the extraction of topologically invariant\nfeatures. We focus on the paradigmatic case of the supercritical Hopf\nbifurcation, which is used to model periodic dynamics across a wide range of\napplications. Our convolutional attention method is trained with data\naugmentations that encourage the learning of topological invariants which can\nbe used to detect bifurcation boundaries in unseen systems and to design models\nof biological systems like oscillatory gene regulatory networks. We further\ndemonstrate our method's use in analyzing real data by recovering distinct\nproliferation and differentiation dynamics along pancreatic endocrinogenesis\ntrajectory in gene expression space based on single-cell data. Our method\nprovides valuable insights into the qualitative, long-term behavior of a wide\nrange of dynamical systems, and can detect bifurcations or catastrophic\ntransitions in large-scale physical and biological systems.", 'authors': ['Noa Moriel', 'Matthew Ricci', 'Mor Nitzan'], 'published': '2023-12-14 18:57:16+00:00', 'categories': ['cs.LG', 'math.DS', 'stat.ML'], 'url': 'http://arxiv.org/abs/2312.09234v3'}
42,668
2405.20439v1
2,405.20439
1
Sharpness-Aware Minimization Enhances Feature Quality via Balanced Learning
Sharpness-Aware Minimization (SAM) has emerged as a promising alternative optimizer to stochastic gradient descent (SGD). The originally-proposed motivation behind SAM was to bias neural networks towards flatter minima that are believed to generalize better. However, recent studies have shown conflicting evidence on the relationship between flatness and generalization, suggesting that flatness does fully explain SAM's success. Sidestepping this debate, we identify an orthogonal effect of SAM that is beneficial out-of-distribution: we argue that SAM implicitly balances the quality of diverse features. SAM achieves this effect by adaptively suppressing well-learned features which gives remaining features opportunity to be learned. We show that this mechanism is beneficial in datasets that contain redundant or spurious features where SGD falls for the simplicity bias and would not otherwise learn all available features. Our insights are supported by experiments on real data: we demonstrate that SAM improves the quality of features in datasets containing redundant or spurious features, including CelebA, Waterbirds, CIFAR-MNIST, and DomainBed.
2024-05-30
{'id': '2405.20439v1', 'title': 'Sharpness-Aware Minimization Enhances Feature Quality via Balanced Learning', 'abstract': "Sharpness-Aware Minimization (SAM) has emerged as a promising alternative\noptimizer to stochastic gradient descent (SGD). The originally-proposed\nmotivation behind SAM was to bias neural networks towards flatter minima that\nare believed to generalize better. However, recent studies have shown\nconflicting evidence on the relationship between flatness and generalization,\nsuggesting that flatness does fully explain SAM's success. Sidestepping this\ndebate, we identify an orthogonal effect of SAM that is beneficial\nout-of-distribution: we argue that SAM implicitly balances the quality of\ndiverse features. SAM achieves this effect by adaptively suppressing\nwell-learned features which gives remaining features opportunity to be learned.\nWe show that this mechanism is beneficial in datasets that contain redundant or\nspurious features where SGD falls for the simplicity bias and would not\notherwise learn all available features. Our insights are supported by\nexperiments on real data: we demonstrate that SAM improves the quality of\nfeatures in datasets containing redundant or spurious features, including\nCelebA, Waterbirds, CIFAR-MNIST, and DomainBed.", 'authors': ['Jacob Mitchell Springer', 'Vaishnavh Nagarajan', 'Aditi Raghunathan'], 'published': '2024-05-30 19:32:56+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2405.20439v1'}
42,675
2402.08530v2
2,402.0853
2
A Distributional Analogue to the Successor Representation
This paper contributes a new approach for distributional reinforcement learning which elucidates a clean separation of transition structure and reward in the learning process. Analogous to how the successor representation (SR) describes the expected consequences of behaving according to a given policy, our distributional successor measure (SM) describes the distributional consequences of this behaviour. We formulate the distributional SM as a distribution over distributions and provide theory connecting it with distributional and model-based reinforcement learning. Moreover, we propose an algorithm that learns the distributional SM from data by minimizing a two-level maximum mean discrepancy. Key to our method are a number of algorithmic techniques that are independently valuable for learning generative models of state. As an illustration of the usefulness of the distributional SM, we show that it enables zero-shot risk-sensitive policy evaluation in a way that was not previously possible.
2024-02-13
{'id': '2402.08530v2', 'title': 'A Distributional Analogue to the Successor Representation', 'abstract': 'This paper contributes a new approach for distributional reinforcement\nlearning which elucidates a clean separation of transition structure and reward\nin the learning process. Analogous to how the successor representation (SR)\ndescribes the expected consequences of behaving according to a given policy,\nour distributional successor measure (SM) describes the distributional\nconsequences of this behaviour. We formulate the distributional SM as a\ndistribution over distributions and provide theory connecting it with\ndistributional and model-based reinforcement learning. Moreover, we propose an\nalgorithm that learns the distributional SM from data by minimizing a two-level\nmaximum mean discrepancy. Key to our method are a number of algorithmic\ntechniques that are independently valuable for learning generative models of\nstate. As an illustration of the usefulness of the distributional SM, we show\nthat it enables zero-shot risk-sensitive policy evaluation in a way that was\nnot previously possible.', 'authors': ['Harley Wiltzer', 'Jesse Farebrother', 'Arthur Gretton', 'Yunhao Tang', 'André Barreto', 'Will Dabney', 'Marc G. Bellemare', 'Mark Rowland'], 'published': '2024-02-13 15:35:24+00:00', 'categories': ['cs.LG', 'cs.AI', 'stat.ML'], 'url': 'http://arxiv.org/abs/2402.08530v2'}
42,679
2401.10148v2
2,401.10148
2
Explicitly Disentangled Representations in Object-Centric Learning
Extracting structured representations from raw visual data is an important and long-standing challenge in machine learning. Recently, techniques for unsupervised learning of object-centric representations have raised growing interest. In this context, enhancing the robustness of the latent features can improve the efficiency and effectiveness of the training of downstream tasks. A promising step in this direction is to disentangle the factors that cause variation in the data. Previously, Invariant Slot Attention disentangled position, scale, and orientation from the remaining features. Extending this approach, we focus on separating the shape and texture components. In particular, we propose a novel architecture that biases object-centric models toward disentangling shape and texture components into two non-overlapping subsets of the latent space dimensions. These subsets are known a priori, hence before the training process. Experiments on a range of object-centric benchmarks reveal that our approach achieves the desired disentanglement while also numerically improving baseline performance in most cases. In addition, we show that our method can generate novel textures for a specific object or transfer textures between objects with distinct shapes.
2024-01-18
{'id': '2401.10148v2', 'title': 'Explicitly Disentangled Representations in Object-Centric Learning', 'abstract': 'Extracting structured representations from raw visual data is an important\nand long-standing challenge in machine learning. Recently, techniques for\nunsupervised learning of object-centric representations have raised growing\ninterest. In this context, enhancing the robustness of the latent features can\nimprove the efficiency and effectiveness of the training of downstream tasks. A\npromising step in this direction is to disentangle the factors that cause\nvariation in the data. Previously, Invariant Slot Attention disentangled\nposition, scale, and orientation from the remaining features. Extending this\napproach, we focus on separating the shape and texture components. In\nparticular, we propose a novel architecture that biases object-centric models\ntoward disentangling shape and texture components into two non-overlapping\nsubsets of the latent space dimensions. These subsets are known a priori, hence\nbefore the training process. Experiments on a range of object-centric\nbenchmarks reveal that our approach achieves the desired disentanglement while\nalso numerically improving baseline performance in most cases. In addition, we\nshow that our method can generate novel textures for a specific object or\ntransfer textures between objects with distinct shapes.', 'authors': ['Riccardo Majellaro', 'Jonathan Collu', 'Aske Plaat', 'Thomas M. Moerland'], 'published': '2024-01-18 17:22:11+00:00', 'categories': ['cs.CV', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2401.10148v2'}
42,680
2304.06911v2
2,304.06911
2
3D Feature Prediction for Masked-AutoEncoder-Based Point Cloud Pretraining
Masked autoencoders (MAE) have recently been introduced to 3D self-supervised pretraining for point clouds due to their great success in NLP and computer vision. Unlike MAEs used in the image domain, where the pretext task is to restore features at the masked pixels, such as colors, the existing 3D MAE works reconstruct the missing geometry only, i.e, the location of the masked points. In contrast to previous studies, we advocate that point location recovery is inessential and restoring intrinsic point features is much superior. To this end, we propose to ignore point position reconstruction and recover high-order features at masked points including surface normals and surface variations, through a novel attention-based decoder which is independent of the encoder design. We validate the effectiveness of our pretext task and decoder design using different encoder structures for 3D training and demonstrate the advantages of our pretrained networks on various point cloud analysis tasks.
2023-04-14
{'id': '2304.06911v2', 'title': '3D Feature Prediction for Masked-AutoEncoder-Based Point Cloud Pretraining', 'abstract': 'Masked autoencoders (MAE) have recently been introduced to 3D self-supervised\npretraining for point clouds due to their great success in NLP and computer\nvision. Unlike MAEs used in the image domain, where the pretext task is to\nrestore features at the masked pixels, such as colors, the existing 3D MAE\nworks reconstruct the missing geometry only, i.e, the location of the masked\npoints. In contrast to previous studies, we advocate that point location\nrecovery is inessential and restoring intrinsic point features is much\nsuperior. To this end, we propose to ignore point position reconstruction and\nrecover high-order features at masked points including surface normals and\nsurface variations, through a novel attention-based decoder which is\nindependent of the encoder design. We validate the effectiveness of our pretext\ntask and decoder design using different encoder structures for 3D training and\ndemonstrate the advantages of our pretrained networks on various point cloud\nanalysis tasks.', 'authors': ['Siming Yan', 'Yuqi Yang', 'Yuxiao Guo', 'Hao Pan', 'Peng-shuai Wang', 'Xin Tong', 'Yang Liu', 'Qixing Huang'], 'published': '2023-04-14 03:25:24+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2304.06911v2'}
42,646
2506.12553v1
2,506.12553
1
Beyond Laplace and Gaussian: Exploring the Generalized Gaussian Mechanism for Private Machine Learning
Differential privacy (DP) is obtained by randomizing a data analysis algorithm, which necessarily introduces a tradeoff between its utility and privacy. Many DP mechanisms are built upon one of two underlying tools: Laplace and Gaussian additive noise mechanisms. We expand the search space of algorithms by investigating the Generalized Gaussian mechanism, which samples the additive noise term $x$ with probability proportional to $e^-| x |^ $ for some $1$. The Laplace and Gaussian mechanisms are special cases of GG for $=1$ and $=2$, respectively. In this work, we prove that all members of the GG family satisfy differential privacy, and provide an extension of an existing numerical accountant (the PRV accountant) for these mechanisms. We show that privacy accounting for the GG Mechanism and its variants is dimension independent, which substantially improves computational costs of privacy accounting. We apply the GG mechanism to two canonical tools for private machine learning, PATE and DP-SGD; we show empirically that $$ has a weak relationship with test-accuracy, and that generally $=2$ (Gaussian) is nearly optimal. This provides justification for the widespread adoption of the Gaussian mechanism in DP learning, and can be interpreted as a negative result, that optimizing over $$ does not lead to meaningful improvements in performance.
2025-06-14
{'id': '2506.12553v1', 'title': 'Beyond Laplace and Gaussian: Exploring the Generalized Gaussian Mechanism for Private Machine Learning', 'abstract': 'Differential privacy (DP) is obtained by randomizing a data analysis\nalgorithm, which necessarily introduces a tradeoff between its utility and\nprivacy. Many DP mechanisms are built upon one of two underlying tools: Laplace\nand Gaussian additive noise mechanisms. We expand the search space of\nalgorithms by investigating the Generalized Gaussian mechanism, which samples\nthe additive noise term $x$ with probability proportional to $e^{-\\frac{| x\n|}{\\sigma}^{\\beta} }$ for some $\\beta \\geq 1$. The Laplace and Gaussian\nmechanisms are special cases of GG for $\\beta=1$ and $\\beta=2$, respectively.\n In this work, we prove that all members of the GG family satisfy differential\nprivacy, and provide an extension of an existing numerical accountant (the PRV\naccountant) for these mechanisms. We show that privacy accounting for the GG\nMechanism and its variants is dimension independent, which substantially\nimproves computational costs of privacy accounting.\n We apply the GG mechanism to two canonical tools for private machine\nlearning, PATE and DP-SGD; we show empirically that $\\beta$ has a weak\nrelationship with test-accuracy, and that generally $\\beta=2$ (Gaussian) is\nnearly optimal. This provides justification for the widespread adoption of the\nGaussian mechanism in DP learning, and can be interpreted as a negative result,\nthat optimizing over $\\beta$ does not lead to meaningful improvements in\nperformance.', 'authors': ['Roy Rinberg', 'Ilia Shumailov', 'Vikrant Singhal', 'Rachel Cummings', 'Nicolas Papernot'], 'published': '2025-06-14 15:49:25+00:00', 'categories': ['cs.LG', 'cs.CR', 'stat.ML'], 'url': 'http://arxiv.org/abs/2506.12553v1'}
42,649
2309.05196v3
2,309.05196
3
Does Writing with Language Models Reduce Content Diversity?
Large language models (LLMs) have led to a surge in collaborative writing with model assistance. As different users incorporate suggestions from the same model, there is a risk of decreased diversity in the produced content, potentially limiting diverse perspectives in public discourse. In this work, we measure the impact of co-writing on diversity via a controlled experiment, where users write argumentative essays in three setups -- using a base LLM (GPT3), a feedback-tuned LLM (InstructGPT), and writing without model help. We develop a set of diversity metrics and find that writing with InstructGPT (but not the GPT3) results in a statistically significant reduction in diversity. Specifically, it increases the similarity between the writings of different authors and reduces the overall lexical and content diversity. We additionally find that this effect is mainly attributable to InstructGPT contributing less diverse text to co-written essays. In contrast, the user-contributed text remains unaffected by model collaboration. This suggests that the recent improvement in generation quality from adapting models to human feedback might come at the cost of more homogeneous and less diverse content.
2023-09-11
{'id': '2309.05196v3', 'title': 'Does Writing with Language Models Reduce Content Diversity?', 'abstract': 'Large language models (LLMs) have led to a surge in collaborative writing\nwith model assistance. As different users incorporate suggestions from the same\nmodel, there is a risk of decreased diversity in the produced content,\npotentially limiting diverse perspectives in public discourse. In this work, we\nmeasure the impact of co-writing on diversity via a controlled experiment,\nwhere users write argumentative essays in three setups -- using a base LLM\n(GPT3), a feedback-tuned LLM (InstructGPT), and writing without model help. We\ndevelop a set of diversity metrics and find that writing with InstructGPT (but\nnot the GPT3) results in a statistically significant reduction in diversity.\nSpecifically, it increases the similarity between the writings of different\nauthors and reduces the overall lexical and content diversity. We additionally\nfind that this effect is mainly attributable to InstructGPT contributing less\ndiverse text to co-written essays. In contrast, the user-contributed text\nremains unaffected by model collaboration. This suggests that the recent\nimprovement in generation quality from adapting models to human feedback might\ncome at the cost of more homogeneous and less diverse content.', 'authors': ['Vishakh Padmakumar', 'He He'], 'published': '2023-09-11 02:16:47+00:00', 'categories': ['cs.CL', 'cs.CY', 'cs.HC', 'cs.LG'], 'url': 'http://arxiv.org/abs/2309.05196v3'}
42,653
2310.15386v2
2,310.15386
2
Course Correcting Koopman Representations
Koopman representations aim to learn features of nonlinear dynamical systems (NLDS) which lead to linear dynamics in the latent space. Theoretically, such features can be used to simplify many problems in modeling and control of NLDS. In this work we study autoencoder formulations of this problem, and different ways they can be used to model dynamics, specifically for future state prediction over long horizons. We discover several limitations of predicting future states in the latent space and propose an inference-time mechanism, which we refer to as Periodic Reencoding, for faithfully capturing long term dynamics. We justify this method both analytically and empirically via experiments in low and high dimensional NLDS.
2023-10-23
{'id': '2310.15386v2', 'title': 'Course Correcting Koopman Representations', 'abstract': 'Koopman representations aim to learn features of nonlinear dynamical systems\n(NLDS) which lead to linear dynamics in the latent space. Theoretically, such\nfeatures can be used to simplify many problems in modeling and control of NLDS.\nIn this work we study autoencoder formulations of this problem, and different\nways they can be used to model dynamics, specifically for future state\nprediction over long horizons. We discover several limitations of predicting\nfuture states in the latent space and propose an inference-time mechanism,\nwhich we refer to as Periodic Reencoding, for faithfully capturing long term\ndynamics. We justify this method both analytically and empirically via\nexperiments in low and high dimensional NLDS.', 'authors': ['Mahan Fathi', 'Clement Gehring', 'Jonathan Pilault', 'David Kanaa', 'Pierre-Luc Bacon', 'Ross Goroshin'], 'published': '2023-10-23 22:36:31+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.RO', 'cs.SY', 'eess.SY'], 'url': 'http://arxiv.org/abs/2310.15386v2'}
42,654
2310.02832v2
2,310.02832
2
Out-of-Distribution Detection by Leveraging Between-Layer Transformation Smoothness
Effective out-of-distribution (OOD) detection is crucial for reliable machine learning models, yet most current methods are limited in practical use due to requirements like access to training data or intervention in training. We present a novel method for detecting OOD data in Transformers based on transformation smoothness between intermediate layers of a network (BLOOD), which is applicable to pre-trained models without access to training data. BLOOD utilizes the tendency of between-layer representation transformations of in-distribution (ID) data to be smoother than the corresponding transformations of OOD data, a property that we also demonstrate empirically. We evaluate BLOOD on several text classification tasks with Transformer networks and demonstrate that it outperforms methods with comparable resource requirements. Our analysis also suggests that when learning simpler tasks, OOD data transformations maintain their original sharpness, whereas sharpness increases with more complex tasks.
2023-10-04
{'id': '2310.02832v2', 'title': 'Out-of-Distribution Detection by Leveraging Between-Layer Transformation Smoothness', 'abstract': 'Effective out-of-distribution (OOD) detection is crucial for reliable machine\nlearning models, yet most current methods are limited in practical use due to\nrequirements like access to training data or intervention in training. We\npresent a novel method for detecting OOD data in Transformers based on\ntransformation smoothness between intermediate layers of a network (BLOOD),\nwhich is applicable to pre-trained models without access to training data.\nBLOOD utilizes the tendency of between-layer representation transformations of\nin-distribution (ID) data to be smoother than the corresponding transformations\nof OOD data, a property that we also demonstrate empirically. We evaluate BLOOD\non several text classification tasks with Transformer networks and demonstrate\nthat it outperforms methods with comparable resource requirements. Our analysis\nalso suggests that when learning simpler tasks, OOD data transformations\nmaintain their original sharpness, whereas sharpness increases with more\ncomplex tasks.', 'authors': ['Fran Jelenić', 'Josip Jukić', 'Martin Tutek', 'Mate Puljiz', 'Jan Šnajder'], 'published': '2023-10-04 13:59:45+00:00', 'categories': ['cs.LG', 'cs.CL'], 'url': 'http://arxiv.org/abs/2310.02832v2'}
42,659
2401.08815v1
2,401.08815
1
Adversarial Supervision Makes Layout-to-Image Diffusion Models Thrive
Despite the recent advances in large-scale diffusion models, little progress has been made on the layout-to-image (L2I) synthesis task. Current L2I models either suffer from poor editability via text or weak alignment between the generated image and the input layout. This limits their usability in practice. To mitigate this, we propose to integrate adversarial supervision into the conventional training pipeline of L2I diffusion models (ALDM). Specifically, we employ a segmentation-based discriminator which provides explicit feedback to the diffusion generator on the pixel-level alignment between the denoised image and the input layout. To encourage consistent adherence to the input layout over the sampling steps, we further introduce the multistep unrolling strategy. Instead of looking at a single timestep, we unroll a few steps recursively to imitate the inference process, and ask the discriminator to assess the alignment of denoised images with the layout over a certain time window. Our experiments show that ALDM enables layout faithfulness of the generated images, while allowing broad editability via text prompts. Moreover, we showcase its usefulness for practical applications: by synthesizing target distribution samples via text control, we improve domain generalization of semantic segmentation models by a large margin (~12 mIoU points).
2024-01-16
{'id': '2401.08815v1', 'title': 'Adversarial Supervision Makes Layout-to-Image Diffusion Models Thrive', 'abstract': 'Despite the recent advances in large-scale diffusion models, little progress\nhas been made on the layout-to-image (L2I) synthesis task. Current L2I models\neither suffer from poor editability via text or weak alignment between the\ngenerated image and the input layout. This limits their usability in practice.\nTo mitigate this, we propose to integrate adversarial supervision into the\nconventional training pipeline of L2I diffusion models (ALDM). Specifically, we\nemploy a segmentation-based discriminator which provides explicit feedback to\nthe diffusion generator on the pixel-level alignment between the denoised image\nand the input layout. To encourage consistent adherence to the input layout\nover the sampling steps, we further introduce the multistep unrolling strategy.\nInstead of looking at a single timestep, we unroll a few steps recursively to\nimitate the inference process, and ask the discriminator to assess the\nalignment of denoised images with the layout over a certain time window. Our\nexperiments show that ALDM enables layout faithfulness of the generated images,\nwhile allowing broad editability via text prompts. Moreover, we showcase its\nusefulness for practical applications: by synthesizing target distribution\nsamples via text control, we improve domain generalization of semantic\nsegmentation models by a large margin (~12 mIoU points).', 'authors': ['Yumeng Li', 'Margret Keuper', 'Dan Zhang', 'Anna Khoreva'], 'published': '2024-01-16 20:31:46+00:00', 'categories': ['cs.CV', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2401.08815v1'}
42,661
2311.15647v1
2,311.15647
1
Bandits Meet Mechanism Design to Combat Clickbait in Online Recommendation
We study a strategic variant of the multi-armed bandit problem, which we coin the strategic click-bandit. This model is motivated by applications in online recommendation where the choice of recommended items depends on both the click-through rates and the post-click rewards. Like in classical bandits, rewards follow a fixed unknown distribution. However, we assume that the click-rate of each arm is chosen strategically by the arm (e.g., a host on Airbnb) in order to maximize the number of times it gets clicked. The algorithm designer does not know the post-click rewards nor the arms' actions (i.e., strategically chosen click-rates) in advance, and must learn both values over time. To solve this problem, we design an incentive-aware learning algorithm, UCB-S, which achieves two goals simultaneously: (a) incentivizing desirable arm behavior under uncertainty; (b) minimizing regret by learning unknown parameters. We characterize all approximate Nash equilibria among arms under UCB-S and show a $O (KT)$ regret bound uniformly in every equilibrium. We also show that incentive-unaware algorithms generally fail to achieve low regret in the strategic click-bandit. Finally, we support our theoretical results by simulations of strategic arm behavior which confirm the effectiveness and robustness of our proposed incentive design.
2023-11-27
{'id': '2311.15647v1', 'title': 'Bandits Meet Mechanism Design to Combat Clickbait in Online Recommendation', 'abstract': "We study a strategic variant of the multi-armed bandit problem, which we coin\nthe strategic click-bandit. This model is motivated by applications in online\nrecommendation where the choice of recommended items depends on both the\nclick-through rates and the post-click rewards. Like in classical bandits,\nrewards follow a fixed unknown distribution. However, we assume that the\nclick-rate of each arm is chosen strategically by the arm (e.g., a host on\nAirbnb) in order to maximize the number of times it gets clicked. The algorithm\ndesigner does not know the post-click rewards nor the arms' actions (i.e.,\nstrategically chosen click-rates) in advance, and must learn both values over\ntime. To solve this problem, we design an incentive-aware learning algorithm,\nUCB-S, which achieves two goals simultaneously: (a) incentivizing desirable arm\nbehavior under uncertainty; (b) minimizing regret by learning unknown\nparameters. We characterize all approximate Nash equilibria among arms under\nUCB-S and show a $\\tilde{\\mathcal{O}} (\\sqrt{KT})$ regret bound uniformly in\nevery equilibrium. We also show that incentive-unaware algorithms generally\nfail to achieve low regret in the strategic click-bandit. Finally, we support\nour theoretical results by simulations of strategic arm behavior which confirm\nthe effectiveness and robustness of our proposed incentive design.", 'authors': ['Thomas Kleine Buening', 'Aadirupa Saha', 'Christos Dimitrakakis', 'Haifeng Xu'], 'published': '2023-11-27 09:19:01+00:00', 'categories': ['cs.LG', 'cs.GT'], 'url': 'http://arxiv.org/abs/2311.15647v1'}
42,671
2401.14544v1
2,401.14544
1
Bayesian Optimization through Gaussian Cox Process Models for Spatio-temporal Data
Bayesian optimization (BO) has established itself as a leading strategy for efficiently optimizing expensive-to-evaluate functions. Existing BO methods mostly rely on Gaussian process (GP) surrogate models and are not applicable to (doubly-stochastic) Gaussian Cox processes, where the observation process is modulated by a latent intensity function modeled as a GP. In this paper, we propose a novel maximum a posteriori inference of Gaussian Cox processes. It leverages the Laplace approximation and change of kernel technique to transform the problem into a new reproducing kernel Hilbert space, where it becomes more tractable computationally. It enables us to obtain both a functional posterior of the latent intensity function and the covariance of the posterior, thus extending existing works that often focus on specific link functions or estimating the posterior mean. Using the result, we propose a BO framework based on the Gaussian Cox process model and further develop a Nystr\"om approximation for efficient computation. Extensive evaluations on various synthetic and real-world datasets demonstrate significant improvement over state-of-the-art inference solutions for Gaussian Cox processes, as well as effective BO with a wide range of acquisition functions designed through the underlying Gaussian Cox process model.
2024-01-25
{'id': '2401.14544v1', 'title': 'Bayesian Optimization through Gaussian Cox Process Models for Spatio-temporal Data', 'abstract': 'Bayesian optimization (BO) has established itself as a leading strategy for\nefficiently optimizing expensive-to-evaluate functions. Existing BO methods\nmostly rely on Gaussian process (GP) surrogate models and are not applicable to\n(doubly-stochastic) Gaussian Cox processes, where the observation process is\nmodulated by a latent intensity function modeled as a GP. In this paper, we\npropose a novel maximum a posteriori inference of Gaussian Cox processes. It\nleverages the Laplace approximation and change of kernel technique to transform\nthe problem into a new reproducing kernel Hilbert space, where it becomes more\ntractable computationally. It enables us to obtain both a functional posterior\nof the latent intensity function and the covariance of the posterior, thus\nextending existing works that often focus on specific link functions or\nestimating the posterior mean. Using the result, we propose a BO framework\nbased on the Gaussian Cox process model and further develop a Nystr\\"om\napproximation for efficient computation. Extensive evaluations on various\nsynthetic and real-world datasets demonstrate significant improvement over\nstate-of-the-art inference solutions for Gaussian Cox processes, as well as\neffective BO with a wide range of acquisition functions designed through the\nunderlying Gaussian Cox process model.', 'authors': ['Yongsheng Mei', 'Mahdi Imani', 'Tian Lan'], 'published': '2024-01-25 22:26:15+00:00', 'categories': ['cs.LG', 'math.FA', 'math.PR'], 'url': 'http://arxiv.org/abs/2401.14544v1'}
42,677
2310.02902v1
2,310.02902
1
Searching for High-Value Molecules Using Reinforcement Learning and Transformers
Reinforcement learning (RL) over text representations can be effective for finding high-value policies that can search over graphs. However, RL requires careful structuring of the search space and algorithm design to be effective in this challenge. Through extensive experiments, we explore how different design choices for text grammar and algorithmic choices for training can affect an RL policy's ability to generate molecules with desired properties. We arrive at a new RL-based molecular design algorithm (ChemRLformer) and perform a thorough analysis using 25 molecule design tasks, including computationally complex protein docking simulations. From this analysis, we discover unique insights in this problem space and show that ChemRLformer achieves state-of-the-art performance while being more straightforward than prior work by demystifying which design choices are actually helpful for text-based molecule design.
2023-10-04
{'id': '2310.02902v1', 'title': 'Searching for High-Value Molecules Using Reinforcement Learning and Transformers', 'abstract': "Reinforcement learning (RL) over text representations can be effective for\nfinding high-value policies that can search over graphs. However, RL requires\ncareful structuring of the search space and algorithm design to be effective in\nthis challenge. Through extensive experiments, we explore how different design\nchoices for text grammar and algorithmic choices for training can affect an RL\npolicy's ability to generate molecules with desired properties. We arrive at a\nnew RL-based molecular design algorithm (ChemRLformer) and perform a thorough\nanalysis using 25 molecule design tasks, including computationally complex\nprotein docking simulations. From this analysis, we discover unique insights in\nthis problem space and show that ChemRLformer achieves state-of-the-art\nperformance while being more straightforward than prior work by demystifying\nwhich design choices are actually helpful for text-based molecule design.", 'authors': ['Raj Ghugare', 'Santiago Miret', 'Adriana Hugessen', 'Mariano Phielipp', 'Glen Berseth'], 'published': '2023-10-04 15:40:07+00:00', 'categories': ['cs.LG', 'cond-mat.mtrl-sci', 'cs.AI'], 'url': 'http://arxiv.org/abs/2310.02902v1'}
42,683
2305.15508v4
2,305.15508
4
How to Fix a Broken Confidence Estimator: Evaluating Post-hoc Methods for Selective Classification with Deep Neural Networks
This paper addresses the problem of selective classification for deep neural networks, where a model is allowed to abstain from low-confidence predictions to avoid potential errors. We focus on so-called post-hoc methods, which replace the confidence estimator of a given classifier without modifying or retraining it, thus being practically appealing. Considering neural networks with softmax outputs, our goal is to identify the best confidence estimator that can be computed directly from the unnormalized logits. This problem is motivated by the intriguing observation in recent work that many classifiers appear to have a "broken" confidence estimator, in the sense that their selective classification performance is much worse than what could be expected by their corresponding accuracies. We perform an extensive experimental study of many existing and proposed confidence estimators applied to 84 pretrained ImageNet classifiers available from popular repositories. Our results show that a simple $p$-norm normalization of the logits, followed by taking the maximum logit as the confidence estimator, can lead to considerable gains in selective classification performance, completely fixing the pathological behavior observed in many classifiers. As a consequence, the selective classification performance of any classifier becomes almost entirely determined by its corresponding accuracy. Moreover, these results are shown to be consistent under distribution shift. Our code is available at https://github.com/lfpc/FixSelectiveClassification.
2023-05-24
{'id': '2305.15508v4', 'title': 'How to Fix a Broken Confidence Estimator: Evaluating Post-hoc Methods for Selective Classification with Deep Neural Networks', 'abstract': 'This paper addresses the problem of selective classification for deep neural\nnetworks, where a model is allowed to abstain from low-confidence predictions\nto avoid potential errors. We focus on so-called post-hoc methods, which\nreplace the confidence estimator of a given classifier without modifying or\nretraining it, thus being practically appealing. Considering neural networks\nwith softmax outputs, our goal is to identify the best confidence estimator\nthat can be computed directly from the unnormalized logits. This problem is\nmotivated by the intriguing observation in recent work that many classifiers\nappear to have a "broken" confidence estimator, in the sense that their\nselective classification performance is much worse than what could be expected\nby their corresponding accuracies. We perform an extensive experimental study\nof many existing and proposed confidence estimators applied to 84 pretrained\nImageNet classifiers available from popular repositories. Our results show that\na simple $p$-norm normalization of the logits, followed by taking the maximum\nlogit as the confidence estimator, can lead to considerable gains in selective\nclassification performance, completely fixing the pathological behavior\nobserved in many classifiers. As a consequence, the selective classification\nperformance of any classifier becomes almost entirely determined by its\ncorresponding accuracy. Moreover, these results are shown to be consistent\nunder distribution shift. Our code is available at\nhttps://github.com/lfpc/FixSelectiveClassification.', 'authors': ['Luís Felipe P. Cattelan', 'Danilo Silva'], 'published': '2023-05-24 18:56:55+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2305.15508v4'}
42,698
2212.08774v2
2,212.08774
2
Annotation by Clicks: A Point-Supervised Contrastive Variance Method for Medical Semantic Segmentation
Medical image segmentation methods typically rely on numerous dense annotated images for model training, which are notoriously expensive and time-consuming to collect. To alleviate this burden, weakly supervised techniques have been exploited to train segmentation models with less expensive annotations. In this paper, we propose a novel point-supervised contrastive variance method (PSCV) for medical image semantic segmentation, which only requires one pixel-point from each organ category to be annotated. The proposed method trains the base segmentation network by using a novel contrastive variance (CV) loss to exploit the unlabeled pixels and a partial cross-entropy loss on the labeled pixels. The CV loss function is designed to exploit the statistical spatial distribution properties of organs in medical images and their variance distribution map representations to enforce discriminative predictions over the unlabeled pixels. Experimental results on two standard medical image datasets demonstrate that the proposed method outperforms the state-of-the-art weakly supervised methods on point-supervised medical image semantic segmentation tasks.
2022-12-17
{'id': '2212.08774v2', 'title': 'Annotation by Clicks: A Point-Supervised Contrastive Variance Method for Medical Semantic Segmentation', 'abstract': 'Medical image segmentation methods typically rely on numerous dense annotated\nimages for model training, which are notoriously expensive and time-consuming\nto collect. To alleviate this burden, weakly supervised techniques have been\nexploited to train segmentation models with less expensive annotations. In this\npaper, we propose a novel point-supervised contrastive variance method (PSCV)\nfor medical image semantic segmentation, which only requires one pixel-point\nfrom each organ category to be annotated. The proposed method trains the base\nsegmentation network by using a novel contrastive variance (CV) loss to exploit\nthe unlabeled pixels and a partial cross-entropy loss on the labeled pixels.\nThe CV loss function is designed to exploit the statistical spatial\ndistribution properties of organs in medical images and their variance\ndistribution map representations to enforce discriminative predictions over the\nunlabeled pixels. Experimental results on two standard medical image datasets\ndemonstrate that the proposed method outperforms the state-of-the-art weakly\nsupervised methods on point-supervised medical image semantic segmentation\ntasks.', 'authors': ['Qing En', 'Yuhong Guo'], 'published': '2022-12-17 01:07:21+00:00', 'categories': ['cs.CV', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2212.08774v2'}
42,704
2403.02690v1
2,403.0269
1
Dirichlet-based Per-Sample Weighting by Transition Matrix for Noisy Label Learning
For learning with noisy labels, the transition matrix, which explicitly models the relation between noisy label distribution and clean label distribution, has been utilized to achieve the statistical consistency of either the classifier or the risk. Previous researches have focused more on how to estimate this transition matrix well, rather than how to utilize it. We propose good utilization of the transition matrix is crucial and suggest a new utilization method based on resampling, coined RENT. Specifically, we first demonstrate current utilizations can have potential limitations for implementation. As an extension to Reweighting, we suggest the Dirichlet distribution-based per-sample Weight Sampling (DWS) framework, and compare reweighting and resampling under DWS framework. With the analyses from DWS, we propose RENT, a REsampling method with Noise Transition matrix. Empirically, RENT consistently outperforms existing transition matrix utilization methods, which includes reweighting, on various benchmark datasets. Our code is available at https://github.com/BaeHeeSun/RENT.
2024-03-05
{'id': '2403.02690v1', 'title': 'Dirichlet-based Per-Sample Weighting by Transition Matrix for Noisy Label Learning', 'abstract': 'For learning with noisy labels, the transition matrix, which explicitly\nmodels the relation between noisy label distribution and clean label\ndistribution, has been utilized to achieve the statistical consistency of\neither the classifier or the risk. Previous researches have focused more on how\nto estimate this transition matrix well, rather than how to utilize it. We\npropose good utilization of the transition matrix is crucial and suggest a new\nutilization method based on resampling, coined RENT. Specifically, we first\ndemonstrate current utilizations can have potential limitations for\nimplementation. As an extension to Reweighting, we suggest the Dirichlet\ndistribution-based per-sample Weight Sampling (DWS) framework, and compare\nreweighting and resampling under DWS framework. With the analyses from DWS, we\npropose RENT, a REsampling method with Noise Transition matrix. Empirically,\nRENT consistently outperforms existing transition matrix utilization methods,\nwhich includes reweighting, on various benchmark datasets. Our code is\navailable at \\url{https://github.com/BaeHeeSun/RENT}.', 'authors': ['HeeSun Bae', 'Seungjae Shin', 'Byeonghu Na', 'Il-Chul Moon'], 'published': '2024-03-05 06:20:49+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2403.02690v1'}
42,662
2403.10231v2
2,403.10231
2
Less is More: One-shot Subgraph Reasoning on Large-scale Knowledge Graphs
To deduce new facts on a knowledge graph (KG), a link predictor learns from the graph structure and collects local evidence to find the answer to a given query. However, existing methods suffer from a severe scalability problem due to the utilization of the whole KG for prediction, which hinders their promise on large scale KGs and cannot be directly addressed by vanilla sampling methods. In this work, we propose the one-shot-subgraph link prediction to achieve efficient and adaptive prediction. The design principle is that, instead of directly acting on the whole KG, the prediction procedure is decoupled into two steps, i.e., (i) extracting only one subgraph according to the query and (ii) predicting on this single, query dependent subgraph. We reveal that the non-parametric and computation-efficient heuristics Personalized PageRank (PPR) can effectively identify the potential answers and supporting evidence. With efficient subgraph-based prediction, we further introduce the automated searching of the optimal configurations in both data and model spaces. Empirically, we achieve promoted efficiency and leading performances on five large-scale benchmarks. The code is publicly available at: https://github.com/tmlr-group/one-shot-subgraph.
2024-03-15
{'id': '2403.10231v2', 'title': 'Less is More: One-shot Subgraph Reasoning on Large-scale Knowledge Graphs', 'abstract': 'To deduce new facts on a knowledge graph (KG), a link predictor learns from\nthe graph structure and collects local evidence to find the answer to a given\nquery. However, existing methods suffer from a severe scalability problem due\nto the utilization of the whole KG for prediction, which hinders their promise\non large scale KGs and cannot be directly addressed by vanilla sampling\nmethods. In this work, we propose the one-shot-subgraph link prediction to\nachieve efficient and adaptive prediction. The design principle is that,\ninstead of directly acting on the whole KG, the prediction procedure is\ndecoupled into two steps, i.e., (i) extracting only one subgraph according to\nthe query and (ii) predicting on this single, query dependent subgraph. We\nreveal that the non-parametric and computation-efficient heuristics\nPersonalized PageRank (PPR) can effectively identify the potential answers and\nsupporting evidence. With efficient subgraph-based prediction, we further\nintroduce the automated searching of the optimal configurations in both data\nand model spaces. Empirically, we achieve promoted efficiency and leading\nperformances on five large-scale benchmarks. The code is publicly available at:\nhttps://github.com/tmlr-group/one-shot-subgraph.', 'authors': ['Zhanke Zhou', 'Yongqi Zhang', 'Jiangchao Yao', 'Quanming Yao', 'Bo Han'], 'published': '2024-03-15 12:00:12+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.SI'], 'url': 'http://arxiv.org/abs/2403.10231v2'}
42,669
2311.18177v2
2,311.18177
2
An Effective Universal Polynomial Basis for Spectral Graph Neural Networks
Spectral Graph Neural Networks (GNNs), also referred to as graph filters have gained increasing prevalence for heterophily graphs. Optimal graph filters rely on Laplacian eigendecomposition for Fourier transform. In an attempt to avert the prohibitive computations, numerous polynomial filters by leveraging distinct polynomials have been proposed to approximate the desired graph filters. However, polynomials in the majority of polynomial filters are predefined and remain fixed across all graphs, failing to accommodate the diverse heterophily degrees across different graphs. To tackle this issue, we first investigate the correlation between polynomial bases of desired graph filters and the degrees of graph heterophily via a thorough theoretical analysis. Afterward, we develop an adaptive heterophily basis by incorporating graph heterophily degrees. Subsequently, we integrate this heterophily basis with the homophily basis, creating a universal polynomial basis UniBasis. In consequence, we devise a general polynomial filter UniFilter. Comprehensive experiments on both real-world and synthetic datasets with varying heterophily degrees significantly support the superiority of UniFilter, demonstrating the effectiveness and generality of UniBasis, as well as its promising capability as a new method for graph analysis.
2023-11-30
{'id': '2311.18177v2', 'title': 'An Effective Universal Polynomial Basis for Spectral Graph Neural Networks', 'abstract': 'Spectral Graph Neural Networks (GNNs), also referred to as graph filters have\ngained increasing prevalence for heterophily graphs. Optimal graph filters rely\non Laplacian eigendecomposition for Fourier transform. In an attempt to avert\nthe prohibitive computations, numerous polynomial filters by leveraging\ndistinct polynomials have been proposed to approximate the desired graph\nfilters. However, polynomials in the majority of polynomial filters are\npredefined and remain fixed across all graphs, failing to accommodate the\ndiverse heterophily degrees across different graphs. To tackle this issue, we\nfirst investigate the correlation between polynomial bases of desired graph\nfilters and the degrees of graph heterophily via a thorough theoretical\nanalysis. Afterward, we develop an adaptive heterophily basis by incorporating\ngraph heterophily degrees. Subsequently, we integrate this heterophily basis\nwith the homophily basis, creating a universal polynomial basis UniBasis. In\nconsequence, we devise a general polynomial filter UniFilter. Comprehensive\nexperiments on both real-world and synthetic datasets with varying heterophily\ndegrees significantly support the superiority of UniFilter, demonstrating the\neffectiveness and generality of UniBasis, as well as its promising capability\nas a new method for graph analysis.', 'authors': ['Keke Huang', 'Pietro Liò'], 'published': '2023-11-30 01:48:42+00:00', 'categories': ['cs.LG', 'cs.SI', 'eess.SP'], 'url': 'http://arxiv.org/abs/2311.18177v2'}
42,670
2309.16397v3
2,309.16397
3
Uncertainty-Aware Decision Transformer for Stochastic Driving Environments
Offline Reinforcement Learning (RL) enables policy learning without active interactions, making it especially appealing for self-driving tasks. Recent successes of Transformers inspire casting offline RL as sequence modeling, which, however, fails in stochastic environments with incorrect assumptions that identical actions can consistently achieve the same goal. In this paper, we introduce an UNcertainty-awaRE deciSion Transformer (UNREST) for planning in stochastic driving environments without introducing additional transition or complex generative models. Specifically, UNREST estimates uncertainties by conditional mutual information between transitions and returns. Discovering 'uncertainty accumulation' and 'temporal locality' properties of driving environments, we replace the global returns in decision transformers with truncated returns less affected by environments to learn from actual outcomes of actions rather than environment transitions. We also dynamically evaluate uncertainty at inference for cautious planning. Extensive experiments demonstrate UNREST's superior performance in various driving scenarios and the power of our uncertainty estimation strategy.
2023-09-28
{'id': '2309.16397v3', 'title': 'Uncertainty-Aware Decision Transformer for Stochastic Driving Environments', 'abstract': "Offline Reinforcement Learning (RL) enables policy learning without active\ninteractions, making it especially appealing for self-driving tasks. Recent\nsuccesses of Transformers inspire casting offline RL as sequence modeling,\nwhich, however, fails in stochastic environments with incorrect assumptions\nthat identical actions can consistently achieve the same goal. In this paper,\nwe introduce an UNcertainty-awaRE deciSion Transformer (UNREST) for planning in\nstochastic driving environments without introducing additional transition or\ncomplex generative models. Specifically, UNREST estimates uncertainties by\nconditional mutual information between transitions and returns. Discovering\n'uncertainty accumulation' and 'temporal locality' properties of driving\nenvironments, we replace the global returns in decision transformers with\ntruncated returns less affected by environments to learn from actual outcomes\nof actions rather than environment transitions. We also dynamically evaluate\nuncertainty at inference for cautious planning. Extensive experiments\ndemonstrate UNREST's superior performance in various driving scenarios and the\npower of our uncertainty estimation strategy.", 'authors': ['Zenan Li', 'Fan Nie', 'Qiao Sun', 'Fang Da', 'Hang Zhao'], 'published': '2023-09-28 12:44:51+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2309.16397v3'}
42,672
2310.03789v3
2,310.03789
3
Grokking as a First Order Phase Transition in Two Layer Networks
A key property of deep neural networks (DNNs) is their ability to learn new features during training. This intriguing aspect of deep learning stands out most clearly in recently reported Grokking phenomena. While mainly reflected as a sudden increase in test accuracy, Grokking is also believed to be a beyond lazy-learning/Gaussian Process (GP) phenomenon involving feature learning. Here we apply a recent development in the theory of feature learning, the adaptive kernel approach, to two teacher-student models with cubic-polynomial and modular addition teachers. We provide analytical predictions on feature learning and Grokking properties of these models and demonstrate a mapping between Grokking and the theory of phase transitions. We show that after Grokking, the state of the DNN is analogous to the mixed phase following a first-order phase transition. In this mixed phase, the DNN generates useful internal representations of the teacher that are sharply distinct from those before the transition.
2023-10-05
{'id': '2310.03789v3', 'title': 'Grokking as a First Order Phase Transition in Two Layer Networks', 'abstract': 'A key property of deep neural networks (DNNs) is their ability to learn new\nfeatures during training. This intriguing aspect of deep learning stands out\nmost clearly in recently reported Grokking phenomena. While mainly reflected as\na sudden increase in test accuracy, Grokking is also believed to be a beyond\nlazy-learning/Gaussian Process (GP) phenomenon involving feature learning. Here\nwe apply a recent development in the theory of feature learning, the adaptive\nkernel approach, to two teacher-student models with cubic-polynomial and\nmodular addition teachers. We provide analytical predictions on feature\nlearning and Grokking properties of these models and demonstrate a mapping\nbetween Grokking and the theory of phase transitions. We show that after\nGrokking, the state of the DNN is analogous to the mixed phase following a\nfirst-order phase transition. In this mixed phase, the DNN generates useful\ninternal representations of the teacher that are sharply distinct from those\nbefore the transition.', 'authors': ['Noa Rubin', 'Inbar Seroussi', 'Zohar Ringel'], 'published': '2023-10-05 18:00:01+00:00', 'categories': ['stat.ML', 'cond-mat.dis-nn', 'cs.LG'], 'url': 'http://arxiv.org/abs/2310.03789v3'}
42,674
2308.15321v6
2,308.15321
6
Elucidating the Exposure Bias in Diffusion Models
Diffusion models have demonstrated impressive generative capabilities, but their exposure bias problem, described as the input mismatch between training and sampling, lacks in-depth exploration. In this paper, we systematically investigate the exposure bias problem in diffusion models by first analytically modelling the sampling distribution, based on which we then attribute the prediction error at each sampling step as the root cause of the exposure bias issue. Furthermore, we discuss potential solutions to this issue and propose an intuitive metric for it. Along with the elucidation of exposure bias, we propose a simple, yet effective, training-free method called Epsilon Scaling to alleviate the exposure bias. We show that Epsilon Scaling explicitly moves the sampling trajectory closer to the vector field learned in the training phase by scaling down the network output, mitigating the input mismatch between training and sampling. Experiments on various diffusion frameworks (ADM, DDIM, EDM, LDM, DiT, PFGM++) verify the effectiveness of our method. Remarkably, our ADM-ES, as a state-of-the-art stochastic sampler, obtains 2.17 FID on CIFAR-10 under 100-step unconditional generation. The code is available at https://github.com/forever208/ADM-ES and https://github.com/forever208/EDM-ES.
2023-08-29
{'id': '2308.15321v6', 'title': 'Elucidating the Exposure Bias in Diffusion Models', 'abstract': 'Diffusion models have demonstrated impressive generative capabilities, but\ntheir \\textit{exposure bias} problem, described as the input mismatch between\ntraining and sampling, lacks in-depth exploration. In this paper, we\nsystematically investigate the exposure bias problem in diffusion models by\nfirst analytically modelling the sampling distribution, based on which we then\nattribute the prediction error at each sampling step as the root cause of the\nexposure bias issue. Furthermore, we discuss potential solutions to this issue\nand propose an intuitive metric for it. Along with the elucidation of exposure\nbias, we propose a simple, yet effective, training-free method called Epsilon\nScaling to alleviate the exposure bias. We show that Epsilon Scaling explicitly\nmoves the sampling trajectory closer to the vector field learned in the\ntraining phase by scaling down the network output, mitigating the input\nmismatch between training and sampling. Experiments on various diffusion\nframeworks (ADM, DDIM, EDM, LDM, DiT, PFGM++) verify the effectiveness of our\nmethod. Remarkably, our ADM-ES, as a state-of-the-art stochastic sampler,\nobtains 2.17 FID on CIFAR-10 under 100-step unconditional generation. The code\nis available at \\url{https://github.com/forever208/ADM-ES} and\n\\url{https://github.com/forever208/EDM-ES}.', 'authors': ['Mang Ning', 'Mingxiao Li', 'Jianlin Su', 'Albert Ali Salah', 'Itir Onal Ertugrul'], 'published': '2023-08-29 14:16:09+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CV'], 'url': 'http://arxiv.org/abs/2308.15321v6'}
42,687
2311.12447v1
2,311.12447
1
Designing Long-term Group Fair Policies in Dynamical Systems
Neglecting the effect that decisions have on individuals (and thus, on the underlying data distribution) when designing algorithmic decision-making policies may increase inequalities and unfairness in the long term - even if fairness considerations were taken in the policy design process. In this paper, we propose a novel framework for achieving long-term group fairness in dynamical systems, in which current decisions may affect an individual's features in the next step, and thus, future decisions. Specifically, our framework allows us to identify a time-independent policy that converges, if deployed, to the targeted fair stationary state of the system in the long term, independently of the initial data distribution. We model the system dynamics with a time-homogeneous Markov chain and optimize the policy leveraging the Markov chain convergence theorem to ensure unique convergence. We provide examples of different targeted fair states of the system, encompassing a range of long-term goals for society and policymakers. Furthermore, we show how our approach facilitates the evaluation of different long-term targets by examining their impact on the group-conditional population distribution in the long term and how it evolves until convergence.
2023-11-21
{'id': '2311.12447v1', 'title': 'Designing Long-term Group Fair Policies in Dynamical Systems', 'abstract': "Neglecting the effect that decisions have on individuals (and thus, on the\nunderlying data distribution) when designing algorithmic decision-making\npolicies may increase inequalities and unfairness in the long term - even if\nfairness considerations were taken in the policy design process. In this paper,\nwe propose a novel framework for achieving long-term group fairness in\ndynamical systems, in which current decisions may affect an individual's\nfeatures in the next step, and thus, future decisions. Specifically, our\nframework allows us to identify a time-independent policy that converges, if\ndeployed, to the targeted fair stationary state of the system in the long term,\nindependently of the initial data distribution. We model the system dynamics\nwith a time-homogeneous Markov chain and optimize the policy leveraging the\nMarkov chain convergence theorem to ensure unique convergence. We provide\nexamples of different targeted fair states of the system, encompassing a range\nof long-term goals for society and policymakers. Furthermore, we show how our\napproach facilitates the evaluation of different long-term targets by examining\ntheir impact on the group-conditional population distribution in the long term\nand how it evolves until convergence.", 'authors': ['Miriam Rateike', 'Isabel Valera', 'Patrick Forré'], 'published': '2023-11-21 08:58:50+00:00', 'categories': ['cs.AI'], 'url': 'http://arxiv.org/abs/2311.12447v1'}
42,663
2403.11120v2
2,403.1112
2
Unifying Feature and Cost Aggregation with Transformers for Semantic and Visual Correspondence
This paper introduces a Transformer-based integrative feature and cost aggregation network designed for dense matching tasks. In the context of dense matching, many works benefit from one of two forms of aggregation: feature aggregation, which pertains to the alignment of similar features, or cost aggregation, a procedure aimed at instilling coherence in the flow estimates across neighboring pixels. In this work, we first show that feature aggregation and cost aggregation exhibit distinct characteristics and reveal the potential for substantial benefits stemming from the judicious use of both aggregation processes. We then introduce a simple yet effective architecture that harnesses self- and cross-attention mechanisms to show that our approach unifies feature aggregation and cost aggregation and effectively harnesses the strengths of both techniques. Within the proposed attention layers, the features and cost volume both complement each other, and the attention layers are interleaved through a coarse-to-fine design to further promote accurate correspondence estimation. Finally at inference, our network produces multi-scale predictions, computes their confidence scores, and selects the most confident flow for final prediction. Our framework is evaluated on standard benchmarks for semantic matching, and also applied to geometric matching, where we show that our approach achieves significant improvements compared to existing methods.
2024-03-17
{'id': '2403.11120v2', 'title': 'Unifying Feature and Cost Aggregation with Transformers for Semantic and Visual Correspondence', 'abstract': 'This paper introduces a Transformer-based integrative feature and cost\naggregation network designed for dense matching tasks. In the context of dense\nmatching, many works benefit from one of two forms of aggregation: feature\naggregation, which pertains to the alignment of similar features, or cost\naggregation, a procedure aimed at instilling coherence in the flow estimates\nacross neighboring pixels. In this work, we first show that feature aggregation\nand cost aggregation exhibit distinct characteristics and reveal the potential\nfor substantial benefits stemming from the judicious use of both aggregation\nprocesses. We then introduce a simple yet effective architecture that harnesses\nself- and cross-attention mechanisms to show that our approach unifies feature\naggregation and cost aggregation and effectively harnesses the strengths of\nboth techniques. Within the proposed attention layers, the features and cost\nvolume both complement each other, and the attention layers are interleaved\nthrough a coarse-to-fine design to further promote accurate correspondence\nestimation. Finally at inference, our network produces multi-scale predictions,\ncomputes their confidence scores, and selects the most confident flow for final\nprediction. Our framework is evaluated on standard benchmarks for semantic\nmatching, and also applied to geometric matching, where we show that our\napproach achieves significant improvements compared to existing methods.', 'authors': ['Sunghwan Hong', 'Seokju Cho', 'Seungryong Kim', 'Stephen Lin'], 'published': '2024-03-17 07:02:55+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2403.11120v2'}
42,678
2402.03445v2
2,402.03445
2
Denoising Diffusion via Image-Based Rendering
Generating 3D scenes is a challenging open problem, which requires synthesizing plausible content that is fully consistent in 3D space. While recent methods such as neural radiance fields excel at view synthesis and 3D reconstruction, they cannot synthesize plausible details in unobserved regions since they lack a generative capability. Conversely, existing generative methods are typically not capable of reconstructing detailed, large-scale scenes in the wild, as they use limited-capacity 3D scene representations, require aligned camera poses, or rely on additional regularizers. In this work, we introduce the first diffusion model able to perform fast, detailed reconstruction and generation of real-world 3D scenes. To achieve this, we make three contributions. First, we introduce a new neural scene representation, IB-planes, that can efficiently and accurately represent large 3D scenes, dynamically allocating more capacity as needed to capture details visible in each image. Second, we propose a denoising-diffusion framework to learn a prior over this novel 3D scene representation, using only 2D images without the need for any additional supervision signal such as masks or depths. This supports 3D reconstruction and generation in a unified architecture. Third, we develop a principled approach to avoid trivial 3D solutions when integrating the image-based rendering with the diffusion model, by dropping out representations of some images. We evaluate the model on several challenging datasets of real and synthetic images, and demonstrate superior results on generation, novel view synthesis and 3D reconstruction.
2024-02-05
{'id': '2402.03445v2', 'title': 'Denoising Diffusion via Image-Based Rendering', 'abstract': 'Generating 3D scenes is a challenging open problem, which requires\nsynthesizing plausible content that is fully consistent in 3D space. While\nrecent methods such as neural radiance fields excel at view synthesis and 3D\nreconstruction, they cannot synthesize plausible details in unobserved regions\nsince they lack a generative capability. Conversely, existing generative\nmethods are typically not capable of reconstructing detailed, large-scale\nscenes in the wild, as they use limited-capacity 3D scene representations,\nrequire aligned camera poses, or rely on additional regularizers. In this work,\nwe introduce the first diffusion model able to perform fast, detailed\nreconstruction and generation of real-world 3D scenes. To achieve this, we make\nthree contributions. First, we introduce a new neural scene representation,\nIB-planes, that can efficiently and accurately represent large 3D scenes,\ndynamically allocating more capacity as needed to capture details visible in\neach image. Second, we propose a denoising-diffusion framework to learn a prior\nover this novel 3D scene representation, using only 2D images without the need\nfor any additional supervision signal such as masks or depths. This supports 3D\nreconstruction and generation in a unified architecture. Third, we develop a\nprincipled approach to avoid trivial 3D solutions when integrating the\nimage-based rendering with the diffusion model, by dropping out representations\nof some images. We evaluate the model on several challenging datasets of real\nand synthetic images, and demonstrate superior results on generation, novel\nview synthesis and 3D reconstruction.', 'authors': ['Titas Anciukevičius', 'Fabian Manhardt', 'Federico Tombari', 'Paul Henderson'], 'published': '2024-02-05 19:00:45+00:00', 'categories': ['cs.CV', 'cs.GR', 'cs.LG'], 'url': 'http://arxiv.org/abs/2402.03445v2'}
42,694
2306.08494v2
2,306.08494
2
Langevin Monte Carlo for strongly log-concave distributions: Randomized midpoint revisited
We revisit the problem of sampling from a target distribution that has a smooth strongly log-concave density everywhere in $R^p$. In this context, if no additional density information is available, the randomized midpoint discretization for the kinetic Langevin diffusion is known to be the most scalable method in high dimensions with large condition numbers. Our main result is a nonasymptotic and easy to compute upper bound on the Wasserstein-2 error of this method. To provide a more thorough explanation of our method for establishing the computable upper bound, we conduct an analysis of the midpoint discretization for the vanilla Langevin process. This analysis helps to clarify the underlying principles and provides valuable insights that we use to establish an improved upper bound for the kinetic Langevin process with the midpoint discretization. Furthermore, by applying these techniques we establish new guarantees for the kinetic Langevin process with Euler discretization, which have a better dependence on the condition number than existing upper bounds.
2023-06-14
{'id': '2306.08494v2', 'title': 'Langevin Monte Carlo for strongly log-concave distributions: Randomized midpoint revisited', 'abstract': 'We revisit the problem of sampling from a target distribution that has a\nsmooth strongly log-concave density everywhere in $\\mathbb R^p$. In this\ncontext, if no additional density information is available, the randomized\nmidpoint discretization for the kinetic Langevin diffusion is known to be the\nmost scalable method in high dimensions with large condition numbers. Our main\nresult is a nonasymptotic and easy to compute upper bound on the Wasserstein-2\nerror of this method. To provide a more thorough explanation of our method for\nestablishing the computable upper bound, we conduct an analysis of the midpoint\ndiscretization for the vanilla Langevin process. This analysis helps to clarify\nthe underlying principles and provides valuable insights that we use to\nestablish an improved upper bound for the kinetic Langevin process with the\nmidpoint discretization. Furthermore, by applying these techniques we establish\nnew guarantees for the kinetic Langevin process with Euler discretization,\nwhich have a better dependence on the condition number than existing upper\nbounds.', 'authors': ['Lu Yu', 'Avetik Karagulyan', 'Arnak Dalalyan'], 'published': '2023-06-14 13:18:09+00:00', 'categories': ['math.ST', 'cs.LG', 'math.PR', 'stat.TH'], 'url': 'http://arxiv.org/abs/2306.08494v2'}
42,697
2404.18033v1
2,404.18033
1
Exposing Text-Image Inconsistency Using Diffusion Models
In the battle against widespread online misinformation, a growing problem is text-image inconsistency, where images are misleadingly paired with texts with different intent or meaning. Existing classification-based methods for text-image inconsistency can identify contextual inconsistencies but fail to provide explainable justifications for their decisions that humans can understand. Although more nuanced, human evaluation is impractical at scale and susceptible to errors. To address these limitations, this study introduces D-TIIL (Diffusion-based Text-Image Inconsistency Localization), which employs text-to-image diffusion models to localize semantic inconsistencies in text and image pairs. These models, trained on large-scale datasets act as ``omniscient" agents that filter out irrelevant information and incorporate background knowledge to identify inconsistencies. In addition, D-TIIL uses text embeddings and modified image regions to visualize these inconsistencies. To evaluate D-TIIL's efficacy, we introduce a new TIIL dataset containing 14K consistent and inconsistent text-image pairs. Unlike existing datasets, TIIL enables assessment at the level of individual words and image regions and is carefully designed to represent various inconsistencies. D-TIIL offers a scalable and evidence-based approach to identifying and localizing text-image inconsistency, providing a robust framework for future research combating misinformation.
2024-04-28
{'id': '2404.18033v1', 'title': 'Exposing Text-Image Inconsistency Using Diffusion Models', 'abstract': 'In the battle against widespread online misinformation, a growing problem is\ntext-image inconsistency, where images are misleadingly paired with texts with\ndifferent intent or meaning. Existing classification-based methods for\ntext-image inconsistency can identify contextual inconsistencies but fail to\nprovide explainable justifications for their decisions that humans can\nunderstand. Although more nuanced, human evaluation is impractical at scale and\nsusceptible to errors. To address these limitations, this study introduces\nD-TIIL (Diffusion-based Text-Image Inconsistency Localization), which employs\ntext-to-image diffusion models to localize semantic inconsistencies in text and\nimage pairs. These models, trained on large-scale datasets act as ``omniscient"\nagents that filter out irrelevant information and incorporate background\nknowledge to identify inconsistencies. In addition, D-TIIL uses text embeddings\nand modified image regions to visualize these inconsistencies. To evaluate\nD-TIIL\'s efficacy, we introduce a new TIIL dataset containing 14K consistent\nand inconsistent text-image pairs. Unlike existing datasets, TIIL enables\nassessment at the level of individual words and image regions and is carefully\ndesigned to represent various inconsistencies. D-TIIL offers a scalable and\nevidence-based approach to identifying and localizing text-image inconsistency,\nproviding a robust framework for future research combating misinformation.', 'authors': ['Mingzhen Huang', 'Shan Jia', 'Zhou Zhou', 'Yan Ju', 'Jialing Cai', 'Siwei Lyu'], 'published': '2024-04-28 00:29:24+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2404.18033v1'}
42,664
2111.06636v3
2,111.06636
3
Closed-Loop Data Transcription to an LDR via Minimaxing Rate Reduction
This work proposes a new computational framework for learning a structured generative model for real-world datasets. In particular, we propose to learn a closed-loop transcription between a multi-class multi-dimensional data distribution and a linear discriminative representation (LDR) in the feature space that consists of multiple independent multi-dimensional linear subspaces. In particular, we argue that the optimal encoding and decoding mappings sought can be formulated as the equilibrium point of a two-player minimax game between the encoder and decoder. A natural utility function for this game is the so-called rate reduction, a simple information-theoretic measure for distances between mixtures of subspace-like Gaussians in the feature space. Our formulation draws inspiration from closed-loop error feedback from control systems and avoids expensive evaluating and minimizing approximated distances between arbitrary distributions in either the data space or the feature space. To a large extent, this new formulation unifies the concepts and benefits of Auto-Encoding and GAN and naturally extends them to the settings of learning a both discriminative and generative representation for multi-class and multi-dimensional real-world data. Our extensive experiments on many benchmark imagery datasets demonstrate tremendous potential of this new closed-loop formulation: under fair comparison, visual quality of the learned decoder and classification performance of the encoder is competitive and often better than existing methods based on GAN, VAE, or a combination of both. Unlike existing generative models, the so learned features of the multiple classes are structured: different classes are explicitly mapped onto corresponding independent principal subspaces in the feature space. Source code can be found at https://github.com/Delay-Xili/LDR.
2021-11-12
{'id': '2111.06636v3', 'title': 'Closed-Loop Data Transcription to an LDR via Minimaxing Rate Reduction', 'abstract': 'This work proposes a new computational framework for learning a structured\ngenerative model for real-world datasets. In particular, we propose to learn a\nclosed-loop transcription between a multi-class multi-dimensional data\ndistribution and a linear discriminative representation (LDR) in the feature\nspace that consists of multiple independent multi-dimensional linear subspaces.\nIn particular, we argue that the optimal encoding and decoding mappings sought\ncan be formulated as the equilibrium point of a two-player minimax game between\nthe encoder and decoder. A natural utility function for this game is the\nso-called rate reduction, a simple information-theoretic measure for distances\nbetween mixtures of subspace-like Gaussians in the feature space. Our\nformulation draws inspiration from closed-loop error feedback from control\nsystems and avoids expensive evaluating and minimizing approximated distances\nbetween arbitrary distributions in either the data space or the feature space.\nTo a large extent, this new formulation unifies the concepts and benefits of\nAuto-Encoding and GAN and naturally extends them to the settings of learning a\nboth discriminative and generative representation for multi-class and\nmulti-dimensional real-world data. Our extensive experiments on many benchmark\nimagery datasets demonstrate tremendous potential of this new closed-loop\nformulation: under fair comparison, visual quality of the learned decoder and\nclassification performance of the encoder is competitive and often better than\nexisting methods based on GAN, VAE, or a combination of both. Unlike existing\ngenerative models, the so learned features of the multiple classes are\nstructured: different classes are explicitly mapped onto corresponding\nindependent principal subspaces in the feature space. Source code can be found\nat https://github.com/Delay-Xili/LDR.', 'authors': ['Xili Dai', 'Shengbang Tong', 'Mingyang Li', 'Ziyang Wu', 'Michael Psenka', 'Kwan Ho Ryan Chan', 'Pengyuan Zhai', 'Yaodong Yu', 'Xiaojun Yuan', 'Heung Yeung Shum', 'Yi Ma'], 'published': '2021-11-12 10:06:08+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2111.06636v3'}
42,666
2110.03128v2
2,110.03128
2
On the Generalization of Models Trained with SGD: Information-Theoretic Bounds and Implications
This paper follows up on a recent work of Neu et al. (2021) and presents some new information-theoretic upper bounds for the generalization error of machine learning models, such as neural networks, trained with SGD. We apply these bounds to analyzing the generalization behaviour of linear and two-layer ReLU networks. Experimental study of these bounds provide some insights on the SGD training of neural networks. They also point to a new and simple regularization scheme which we show performs comparably to the current state of the art.
2021-10-07
{'id': '2110.03128v2', 'title': 'On the Generalization of Models Trained with SGD: Information-Theoretic Bounds and Implications', 'abstract': 'This paper follows up on a recent work of Neu et al. (2021) and presents some\nnew information-theoretic upper bounds for the generalization error of machine\nlearning models, such as neural networks, trained with SGD. We apply these\nbounds to analyzing the generalization behaviour of linear and two-layer ReLU\nnetworks. Experimental study of these bounds provide some insights on the SGD\ntraining of neural networks. They also point to a new and simple regularization\nscheme which we show performs comparably to the current state of the art.', 'authors': ['Ziqiao Wang', 'Yongyi Mao'], 'published': '2021-10-07 00:53:33+00:00', 'categories': ['cs.LG', 'cs.IT', 'math.IT'], 'url': 'http://arxiv.org/abs/2110.03128v2'}
42,667
2110.02870v3
2,110.0287
3
Capturing Structural Locality in Non-parametric Language Models
Structural locality is a ubiquitous feature of real-world datasets, wherein data points are organized into local hierarchies. Some examples include topical clusters in text or project hierarchies in source code repositories. In this paper, we explore utilizing this structural locality within non-parametric language models, which generate sequences that reference retrieved examples from an external source. We propose a simple yet effective approach for adding locality information into such models by adding learned parameters that improve the likelihood of retrieving examples from local neighborhoods. Experiments on two different domains, Java source code and Wikipedia text, demonstrate that locality features improve model efficacy over models without access to these features, with interesting differences. We also perform an analysis of how and where locality features contribute to improved performance and why the traditionally used contextual similarity metrics alone are not enough to grasp the locality structure.
2021-10-06
{'id': '2110.02870v3', 'title': 'Capturing Structural Locality in Non-parametric Language Models', 'abstract': 'Structural locality is a ubiquitous feature of real-world datasets, wherein\ndata points are organized into local hierarchies. Some examples include topical\nclusters in text or project hierarchies in source code repositories. In this\npaper, we explore utilizing this structural locality within non-parametric\nlanguage models, which generate sequences that reference retrieved examples\nfrom an external source. We propose a simple yet effective approach for adding\nlocality information into such models by adding learned parameters that improve\nthe likelihood of retrieving examples from local neighborhoods. Experiments on\ntwo different domains, Java source code and Wikipedia text, demonstrate that\nlocality features improve model efficacy over models without access to these\nfeatures, with interesting differences. We also perform an analysis of how and\nwhere locality features contribute to improved performance and why the\ntraditionally used contextual similarity metrics alone are not enough to grasp\nthe locality structure.', 'authors': ['Frank F. Xu', 'Junxian He', 'Graham Neubig', 'Vincent J. Hellendoorn'], 'published': '2021-10-06 15:53:38+00:00', 'categories': ['cs.CL', 'cs.SE'], 'url': 'http://arxiv.org/abs/2110.02870v3'}
42,673
2112.02612v3
2,112.02612
3
Training Structured Neural Networks Through Manifold Identification and Variance Reduction
This paper proposes an algorithm (RMDA) for training neural networks (NNs) with a regularization term for promoting desired structures. RMDA does not incur computation additional to proximal SGD with momentum, and achieves variance reduction without requiring the objective function to be of the finite-sum form. Through the tool of manifold identification from nonlinear optimization, we prove that after a finite number of iterations, all iterates of RMDA possess a desired structure identical to that induced by the regularizer at the stationary point of asymptotic convergence, even in the presence of engineering tricks like data augmentation and dropout that complicate the training process. Experiments on training NNs with structured sparsity confirm that variance reduction is necessary for such an identification, and show that RMDA thus significantly outperforms existing methods for this task. For unstructured sparsity, RMDA also outperforms a state-of-the-art pruning method, validating the benefits of training structured NNs through regularization.
2021-12-05
{'id': '2112.02612v3', 'title': 'Training Structured Neural Networks Through Manifold Identification and Variance Reduction', 'abstract': 'This paper proposes an algorithm (RMDA) for training neural networks (NNs)\nwith a regularization term for promoting desired structures. RMDA does not\nincur computation additional to proximal SGD with momentum, and achieves\nvariance reduction without requiring the objective function to be of the\nfinite-sum form. Through the tool of manifold identification from nonlinear\noptimization, we prove that after a finite number of iterations, all iterates\nof RMDA possess a desired structure identical to that induced by the\nregularizer at the stationary point of asymptotic convergence, even in the\npresence of engineering tricks like data augmentation and dropout that\ncomplicate the training process. Experiments on training NNs with structured\nsparsity confirm that variance reduction is necessary for such an\nidentification, and show that RMDA thus significantly outperforms existing\nmethods for this task. For unstructured sparsity, RMDA also outperforms a\nstate-of-the-art pruning method, validating the benefits of training structured\nNNs through regularization.', 'authors': ['Zih-Syuan Huang', 'Ching-pei Lee'], 'published': '2021-12-05 16:23:53+00:00', 'categories': ['cs.LG', 'math.OC', 'stat.ML'], 'url': 'http://arxiv.org/abs/2112.02612v3'}
42,676
2106.06946v3
2,106.06946
3
Boosting Randomized Smoothing with Variance Reduced Classifiers
Randomized Smoothing (RS) is a promising method for obtaining robustness certificates by evaluating a base model under noise. In this work, we: (i) theoretically motivate why ensembles are a particularly suitable choice as base models for RS, and (ii) empirically confirm this choice, obtaining state-of-the-art results in multiple settings. The key insight of our work is that the reduced variance of ensembles over the perturbations introduced in RS leads to significantly more consistent classifications for a given input. This, in turn, leads to substantially increased certifiable radii for samples close to the decision boundary. Additionally, we introduce key optimizations which enable an up to 55-fold decrease in sample complexity of RS for predetermined radii, thus drastically reducing its computational overhead. Experimentally, we show that ensembles of only 3 to 10 classifiers consistently improve on their strongest constituting model with respect to their average certified radius (ACR) by 5% to 21% on both CIFAR10 and ImageNet, achieving a new state-of-the-art ACR of 0.86 and 1.11, respectively. We release all code and models required to reproduce our results at https://github.com/eth-sri/smoothing-ensembles.
2021-06-13
{'id': '2106.06946v3', 'title': 'Boosting Randomized Smoothing with Variance Reduced Classifiers', 'abstract': 'Randomized Smoothing (RS) is a promising method for obtaining robustness\ncertificates by evaluating a base model under noise. In this work, we: (i)\ntheoretically motivate why ensembles are a particularly suitable choice as base\nmodels for RS, and (ii) empirically confirm this choice, obtaining\nstate-of-the-art results in multiple settings. The key insight of our work is\nthat the reduced variance of ensembles over the perturbations introduced in RS\nleads to significantly more consistent classifications for a given input. This,\nin turn, leads to substantially increased certifiable radii for samples close\nto the decision boundary. Additionally, we introduce key optimizations which\nenable an up to 55-fold decrease in sample complexity of RS for predetermined\nradii, thus drastically reducing its computational overhead. Experimentally, we\nshow that ensembles of only 3 to 10 classifiers consistently improve on their\nstrongest constituting model with respect to their average certified radius\n(ACR) by 5% to 21% on both CIFAR10 and ImageNet, achieving a new\nstate-of-the-art ACR of 0.86 and 1.11, respectively. We release all code and\nmodels required to reproduce our results at\nhttps://github.com/eth-sri/smoothing-ensembles.', 'authors': ['Miklós Z. Horváth', 'Mark Niklas Müller', 'Marc Fischer', 'Martin Vechev'], 'published': '2021-06-13 08:40:27+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CV'], 'url': 'http://arxiv.org/abs/2106.06946v3'}
42,690
2203.12748v1
2,203.12748
1
Is Fairness Only Metric Deep? Evaluating and Addressing Subgroup Gaps in Deep Metric Learning
Deep metric learning (DML) enables learning with less supervision through its emphasis on the similarity structure of representations. There has been much work on improving generalization of DML in settings like zero-shot retrieval, but little is known about its implications for fairness. In this paper, we are the first to evaluate state-of-the-art DML methods trained on imbalanced data, and to show the negative impact these representations have on minority subgroup performance when used for downstream tasks. In this work, we first define fairness in DML through an analysis of three properties of the representation space -- inter-class alignment, intra-class alignment, and uniformity -- and propose finDML, the fairness in non-balanced DML benchmark to characterize representation fairness. Utilizing finDML, we find bias in DML representations to propagate to common downstream classification tasks. Surprisingly, this bias is propagated even when training data in the downstream task is re-balanced. To address this problem, we present Partial Attribute De-correlation (PARADE) to de-correlate feature representations from sensitive attributes and reduce performance gaps between subgroups in both embedding space and downstream metrics.
2022-03-23
{'id': '2203.12748v1', 'title': 'Is Fairness Only Metric Deep? Evaluating and Addressing Subgroup Gaps in Deep Metric Learning', 'abstract': 'Deep metric learning (DML) enables learning with less supervision through its\nemphasis on the similarity structure of representations. There has been much\nwork on improving generalization of DML in settings like zero-shot retrieval,\nbut little is known about its implications for fairness. In this paper, we are\nthe first to evaluate state-of-the-art DML methods trained on imbalanced data,\nand to show the negative impact these representations have on minority subgroup\nperformance when used for downstream tasks. In this work, we first define\nfairness in DML through an analysis of three properties of the representation\nspace -- inter-class alignment, intra-class alignment, and uniformity -- and\npropose finDML, the fairness in non-balanced DML benchmark to characterize\nrepresentation fairness. Utilizing finDML, we find bias in DML representations\nto propagate to common downstream classification tasks. Surprisingly, this bias\nis propagated even when training data in the downstream task is re-balanced. To\naddress this problem, we present Partial Attribute De-correlation (PARADE) to\nde-correlate feature representations from sensitive attributes and reduce\nperformance gaps between subgroups in both embedding space and downstream\nmetrics.', 'authors': ['Natalie Dullerud', 'Karsten Roth', 'Kimia Hamidieh', 'Nicolas Papernot', 'Marzyeh Ghassemi'], 'published': '2022-03-23 22:20:29+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CY', 'stat.ML'], 'url': 'http://arxiv.org/abs/2203.12748v1'}
42,682
2310.06721v3
2,310.06721
3
Tweedie Moment Projected Diffusions For Inverse Problems
Diffusion generative models unlock new possibilities for inverse problems as they allow for the incorporation of strong empirical priors in scientific inference. Recently, diffusion models are repurposed for solving inverse problems using Gaussian approximations to conditional densities of the reverse process via Tweedie's formula to parameterise the mean, complemented with various heuristics. To address various challenges arising from these approximations, we leverage higher order information using Tweedie's formula and obtain a statistically principled approximation. We further provide a theoretical guarantee specifically for posterior sampling which can lead to a better theoretical understanding of diffusion-based conditional sampling. Finally, we illustrate the empirical effectiveness of our approach for general linear inverse problems on toy synthetic examples as well as image restoration. We show that our method (i) removes any time-dependent step-size hyperparameters required by earlier methods, (ii) brings stability and better sample quality across multiple noise levels, (iii) is the only method that works in a stable way with variance exploding (VE) forward processes as opposed to earlier works.
2023-10-10
{'id': '2310.06721v3', 'title': 'Tweedie Moment Projected Diffusions For Inverse Problems', 'abstract': "Diffusion generative models unlock new possibilities for inverse problems as\nthey allow for the incorporation of strong empirical priors in scientific\ninference. Recently, diffusion models are repurposed for solving inverse\nproblems using Gaussian approximations to conditional densities of the reverse\nprocess via Tweedie's formula to parameterise the mean, complemented with\nvarious heuristics. To address various challenges arising from these\napproximations, we leverage higher order information using Tweedie's formula\nand obtain a statistically principled approximation. We further provide a\ntheoretical guarantee specifically for posterior sampling which can lead to a\nbetter theoretical understanding of diffusion-based conditional sampling.\nFinally, we illustrate the empirical effectiveness of our approach for general\nlinear inverse problems on toy synthetic examples as well as image restoration.\nWe show that our method (i) removes any time-dependent step-size\nhyperparameters required by earlier methods, (ii) brings stability and better\nsample quality across multiple noise levels, (iii) is the only method that\nworks in a stable way with variance exploding (VE) forward processes as opposed\nto earlier works.", 'authors': ['Benjamin Boys', 'Mark Girolami', 'Jakiw Pidstrigach', 'Sebastian Reich', 'Alan Mosca', 'O. Deniz Akyildiz'], 'published': '2023-10-10 15:48:55+00:00', 'categories': ['stat.CO'], 'url': 'http://arxiv.org/abs/2310.06721v3'}
42,684
2310.20082v2
2,310.20082
2
Efficient Subgraph GNNs by Learning Effective Selection Policies
Subgraph GNNs are provably expressive neural architectures that learn graph representations from sets of subgraphs. Unfortunately, their applicability is hampered by the computational complexity associated with performing message passing on many subgraphs. In this paper, we consider the problem of learning to select a small subset of the large set of possible subgraphs in a data-driven fashion. We first motivate the problem by proving that there are families of WL-indistinguishable graphs for which there exist efficient subgraph selection policies: small subsets of subgraphs that can already identify all the graphs within the family. We then propose a new approach, called Policy-Learn, that learns how to select subgraphs in an iterative manner. We prove that, unlike popular random policies and prior work addressing the same problem, our architecture is able to learn the efficient policies mentioned above. Our experimental results demonstrate that Policy-Learn outperforms existing baselines across a wide range of datasets.
2023-10-30
{'id': '2310.20082v2', 'title': 'Efficient Subgraph GNNs by Learning Effective Selection Policies', 'abstract': 'Subgraph GNNs are provably expressive neural architectures that learn graph\nrepresentations from sets of subgraphs. Unfortunately, their applicability is\nhampered by the computational complexity associated with performing message\npassing on many subgraphs. In this paper, we consider the problem of learning\nto select a small subset of the large set of possible subgraphs in a\ndata-driven fashion. We first motivate the problem by proving that there are\nfamilies of WL-indistinguishable graphs for which there exist efficient\nsubgraph selection policies: small subsets of subgraphs that can already\nidentify all the graphs within the family. We then propose a new approach,\ncalled Policy-Learn, that learns how to select subgraphs in an iterative\nmanner. We prove that, unlike popular random policies and prior work addressing\nthe same problem, our architecture is able to learn the efficient policies\nmentioned above. Our experimental results demonstrate that Policy-Learn\noutperforms existing baselines across a wide range of datasets.', 'authors': ['Beatrice Bevilacqua', 'Moshe Eliasof', 'Eli Meirom', 'Bruno Ribeiro', 'Haggai Maron'], 'published': '2023-10-30 23:41:05+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.20082v2'}
42,685
2402.03590v1
2,402.0359
1
Assessing the Impact of Distribution Shift on Reinforcement Learning Performance
Research in machine learning is making progress in fixing its own reproducibility crisis. Reinforcement learning (RL), in particular, faces its own set of unique challenges. Comparison of point estimates, and plots that show successful convergence to the optimal policy during training, may obfuscate overfitting or dependence on the experimental setup. Although researchers in RL have proposed reliability metrics that account for uncertainty to better understand each algorithm's strengths and weaknesses, the recommendations of past work do not assume the presence of out-of-distribution observations. We propose a set of evaluation methods that measure the robustness of RL algorithms under distribution shifts. The tools presented here argue for the need to account for performance over time while the agent is acting in its environment. In particular, we recommend time series analysis as a method of observational RL evaluation. We also show that the unique properties of RL and simulated dynamic environments allow us to make stronger assumptions to justify the measurement of causal impact in our evaluations. We then apply these tools to single-agent and multi-agent environments to show the impact of introducing distribution shifts during test time. We present this methodology as a first step toward rigorous RL evaluation in the presence of distribution shifts.
2024-02-05
{'id': '2402.03590v1', 'title': 'Assessing the Impact of Distribution Shift on Reinforcement Learning Performance', 'abstract': "Research in machine learning is making progress in fixing its own\nreproducibility crisis. Reinforcement learning (RL), in particular, faces its\nown set of unique challenges. Comparison of point estimates, and plots that\nshow successful convergence to the optimal policy during training, may\nobfuscate overfitting or dependence on the experimental setup. Although\nresearchers in RL have proposed reliability metrics that account for\nuncertainty to better understand each algorithm's strengths and weaknesses, the\nrecommendations of past work do not assume the presence of out-of-distribution\nobservations. We propose a set of evaluation methods that measure the\nrobustness of RL algorithms under distribution shifts. The tools presented here\nargue for the need to account for performance over time while the agent is\nacting in its environment. In particular, we recommend time series analysis as\na method of observational RL evaluation. We also show that the unique\nproperties of RL and simulated dynamic environments allow us to make stronger\nassumptions to justify the measurement of causal impact in our evaluations. We\nthen apply these tools to single-agent and multi-agent environments to show the\nimpact of introducing distribution shifts during test time. We present this\nmethodology as a first step toward rigorous RL evaluation in the presence of\ndistribution shifts.", 'authors': ['Ted Fujimoto', 'Joshua Suetterlein', 'Samrat Chatterjee', 'Auroop Ganguly'], 'published': '2024-02-05 23:50:55+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.MA'], 'url': 'http://arxiv.org/abs/2402.03590v1'}
42,686
2305.12205v2
2,305.12205
2
Vocabulary for Universal Approximation: A Linguistic Perspective of Mapping Compositions
In recent years, deep learning-based sequence modelings, such as language models, have received much attention and success, which pushes researchers to explore the possibility of transforming non-sequential problems into a sequential form. Following this thought, deep neural networks can be represented as composite functions of a sequence of mappings, linear or nonlinear, where each composition can be viewed as a word. However, the weights of linear mappings are undetermined and hence require an infinite number of words. In this article, we investigate the finite case and constructively prove the existence of a finite vocabulary $V=\_i: R^d R^d | i=1,...,n\$ with $n=O(d^2)$ for the universal approximation. That is, for any continuous mapping $f: R^d R^d$, compact domain $$ and $>0$, there is a sequence of mappings $_i_1, ..., _i_m V, m Z_+$, such that the composition $_i_m ... _i_1 $ approximates $f$ on $$ with an error less than $$. Our results demonstrate an unusual approximation power of mapping compositions and motivate a novel compositional model for regular languages.
2023-05-20
{'id': '2305.12205v2', 'title': 'Vocabulary for Universal Approximation: A Linguistic Perspective of Mapping Compositions', 'abstract': 'In recent years, deep learning-based sequence modelings, such as language\nmodels, have received much attention and success, which pushes researchers to\nexplore the possibility of transforming non-sequential problems into a\nsequential form. Following this thought, deep neural networks can be\nrepresented as composite functions of a sequence of mappings, linear or\nnonlinear, where each composition can be viewed as a \\emph{word}. However, the\nweights of linear mappings are undetermined and hence require an infinite\nnumber of words. In this article, we investigate the finite case and\nconstructively prove the existence of a finite \\emph{vocabulary} $V=\\{\\phi_i:\n\\mathbb{R}^d \\to \\mathbb{R}^d | i=1,...,n\\}$ with $n=O(d^2)$ for the universal\napproximation. That is, for any continuous mapping $f: \\mathbb{R}^d \\to\n\\mathbb{R}^d$, compact domain $\\Omega$ and $\\varepsilon>0$, there is a sequence\nof mappings $\\phi_{i_1}, ..., \\phi_{i_m} \\in V, m \\in \\mathbb{Z}_+$, such that\nthe composition $\\phi_{i_m} \\circ ... \\circ \\phi_{i_1} $ approximates $f$ on\n$\\Omega$ with an error less than $\\varepsilon$. Our results demonstrate an\nunusual approximation power of mapping compositions and motivate a novel\ncompositional model for regular languages.', 'authors': ['Yongqiang Cai'], 'published': '2023-05-20 14:50:34+00:00', 'categories': ['cs.LG', 'cs.NA', 'math.DS', 'math.NA'], 'url': 'http://arxiv.org/abs/2305.12205v2'}
42,689
2305.19190v4
2,305.1919
4
Inverse Approximation Theory for Nonlinear Recurrent Neural Networks
We prove an inverse approximation theorem for the approximation of nonlinear sequence-to-sequence relationships using recurrent neural networks (RNNs). This is a so-called Bernstein-type result in approximation theory, which deduces properties of a target function under the assumption that it can be effectively approximated by a hypothesis space. In particular, we show that nonlinear sequence relationships that can be stably approximated by nonlinear RNNs must have an exponential decaying memory structure - a notion that can be made precise. This extends the previously identified curse of memory in linear RNNs into the general nonlinear setting, and quantifies the essential limitations of the RNN architecture for learning sequential relationships with long-term memory. Based on the analysis, we propose a principled reparameterization method to overcome the limitations. Our theoretical results are confirmed by numerical experiments. The code has been released in https://github.com/radarFudan/Curse-of-memory
2023-05-30
{'id': '2305.19190v4', 'title': 'Inverse Approximation Theory for Nonlinear Recurrent Neural Networks', 'abstract': 'We prove an inverse approximation theorem for the approximation of nonlinear\nsequence-to-sequence relationships using recurrent neural networks (RNNs). This\nis a so-called Bernstein-type result in approximation theory, which deduces\nproperties of a target function under the assumption that it can be effectively\napproximated by a hypothesis space. In particular, we show that nonlinear\nsequence relationships that can be stably approximated by nonlinear RNNs must\nhave an exponential decaying memory structure - a notion that can be made\nprecise. This extends the previously identified curse of memory in linear RNNs\ninto the general nonlinear setting, and quantifies the essential limitations of\nthe RNN architecture for learning sequential relationships with long-term\nmemory. Based on the analysis, we propose a principled reparameterization\nmethod to overcome the limitations. Our theoretical results are confirmed by\nnumerical experiments. The code has been released in\nhttps://github.com/radarFudan/Curse-of-memory', 'authors': ['Shida Wang', 'Zhong Li', 'Qianxiao Li'], 'published': '2023-05-30 16:34:28+00:00', 'categories': ['cs.LG', 'cs.AI', 'math.DS'], 'url': 'http://arxiv.org/abs/2305.19190v4'}
42,691
2405.18792v1
2,405.18792
1
Kernel Metric Learning for In-Sample Off-Policy Evaluation of Deterministic RL Policies
We consider off-policy evaluation (OPE) of deterministic target policies for reinforcement learning (RL) in environments with continuous action spaces. While it is common to use importance sampling for OPE, it suffers from high variance when the behavior policy deviates significantly from the target policy. In order to address this issue, some recent works on OPE proposed in-sample learning with importance resampling. Yet, these approaches are not applicable to deterministic target policies for continuous action spaces. To address this limitation, we propose to relax the deterministic target policy using a kernel and learn the kernel metrics that minimize the overall mean squared error of the estimated temporal difference update vector of an action value function, where the action value function is used for policy evaluation. We derive the bias and variance of the estimation error due to this relaxation and provide analytic solutions for the optimal kernel metric. In empirical studies using various test domains, we show that the OPE with in-sample learning using the kernel with optimized metric achieves significantly improved accuracy than other baselines.
2024-05-29
{'id': '2405.18792v1', 'title': 'Kernel Metric Learning for In-Sample Off-Policy Evaluation of Deterministic RL Policies', 'abstract': 'We consider off-policy evaluation (OPE) of deterministic target policies for\nreinforcement learning (RL) in environments with continuous action spaces.\nWhile it is common to use importance sampling for OPE, it suffers from high\nvariance when the behavior policy deviates significantly from the target\npolicy. In order to address this issue, some recent works on OPE proposed\nin-sample learning with importance resampling. Yet, these approaches are not\napplicable to deterministic target policies for continuous action spaces. To\naddress this limitation, we propose to relax the deterministic target policy\nusing a kernel and learn the kernel metrics that minimize the overall mean\nsquared error of the estimated temporal difference update vector of an action\nvalue function, where the action value function is used for policy evaluation.\nWe derive the bias and variance of the estimation error due to this relaxation\nand provide analytic solutions for the optimal kernel metric. In empirical\nstudies using various test domains, we show that the OPE with in-sample\nlearning using the kernel with optimized metric achieves significantly improved\naccuracy than other baselines.', 'authors': ['Haanvid Lee', 'Tri Wahyu Guntara', 'Jongmin Lee', 'Yung-Kyun Noh', 'Kee-Eung Kim'], 'published': '2024-05-29 06:17:33+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2405.18792v1'}
42,695
2310.03111v1
2,310.03111
1
Multi-modal Gaussian Process Variational Autoencoders for Neural and Behavioral Data
Characterizing the relationship between neural population activity and behavioral data is a central goal of neuroscience. While latent variable models (LVMs) are successful in describing high-dimensional time-series data, they are typically only designed for a single type of data, making it difficult to identify structure shared across different experimental data modalities. Here, we address this shortcoming by proposing an unsupervised LVM which extracts temporally evolving shared and independent latents for distinct, simultaneously recorded experimental modalities. We do this by combining Gaussian Process Factor Analysis (GPFA), an interpretable LVM for neural spiking data with temporally smooth latent space, with Gaussian Process Variational Autoencoders (GP-VAEs), which similarly use a GP prior to characterize correlations in a latent space, but admit rich expressivity due to a deep neural network mapping to observations. We achieve interpretability in our model by partitioning latent variability into components that are either shared between or independent to each modality. We parameterize the latents of our model in the Fourier domain, and show improved latent identification using this approach over standard GP-VAE methods. We validate our model on simulated multi-modal data consisting of Poisson spike counts and MNIST images that scale and rotate smoothly over time. We show that the multi-modal GP-VAE (MM-GPVAE) is able to not only identify the shared and independent latent structure across modalities accurately, but provides good reconstructions of both images and neural rates on held-out trials. Finally, we demonstrate our framework on two real world multi-modal experimental settings: Drosophila whole-brain calcium imaging alongside tracked limb positions, and Manduca sexta spike train measurements from ten wing muscles as the animal tracks a visual stimulus.
2023-10-04
{'id': '2310.03111v1', 'title': 'Multi-modal Gaussian Process Variational Autoencoders for Neural and Behavioral Data', 'abstract': 'Characterizing the relationship between neural population activity and\nbehavioral data is a central goal of neuroscience. While latent variable models\n(LVMs) are successful in describing high-dimensional time-series data, they are\ntypically only designed for a single type of data, making it difficult to\nidentify structure shared across different experimental data modalities. Here,\nwe address this shortcoming by proposing an unsupervised LVM which extracts\ntemporally evolving shared and independent latents for distinct, simultaneously\nrecorded experimental modalities. We do this by combining Gaussian Process\nFactor Analysis (GPFA), an interpretable LVM for neural spiking data with\ntemporally smooth latent space, with Gaussian Process Variational Autoencoders\n(GP-VAEs), which similarly use a GP prior to characterize correlations in a\nlatent space, but admit rich expressivity due to a deep neural network mapping\nto observations. We achieve interpretability in our model by partitioning\nlatent variability into components that are either shared between or\nindependent to each modality. We parameterize the latents of our model in the\nFourier domain, and show improved latent identification using this approach\nover standard GP-VAE methods. We validate our model on simulated multi-modal\ndata consisting of Poisson spike counts and MNIST images that scale and rotate\nsmoothly over time. We show that the multi-modal GP-VAE (MM-GPVAE) is able to\nnot only identify the shared and independent latent structure across modalities\naccurately, but provides good reconstructions of both images and neural rates\non held-out trials. Finally, we demonstrate our framework on two real world\nmulti-modal experimental settings: Drosophila whole-brain calcium imaging\nalongside tracked limb positions, and Manduca sexta spike train measurements\nfrom ten wing muscles as the animal tracks a visual stimulus.', 'authors': ['Rabia Gondur', 'Usama Bin Sikandar', 'Evan Schaffer', 'Mikio Christian Aoi', 'Stephen L Keeley'], 'published': '2023-10-04 19:04:55+00:00', 'categories': ['cs.LG', 'q-bio.NC'], 'url': 'http://arxiv.org/abs/2310.03111v1'}
42,696
2310.03368v4
2,310.03368
4
Evaluating Hallucinations in Chinese Large Language Models
In this paper, we establish a benchmark named HalluQA (Chinese Hallucination Question-Answering) to measure the hallucination phenomenon in Chinese large language models. HalluQA contains 450 meticulously designed adversarial questions, spanning multiple domains, and takes into account Chinese historical culture, customs, and social phenomena. During the construction of HalluQA, we consider two types of hallucinations: imitative falsehoods and factual errors, and we construct adversarial samples based on GLM-130B and ChatGPT. For evaluation, we design an automated evaluation method using GPT-4 to judge whether a model output is hallucinated. We conduct extensive experiments on 24 large language models, including ERNIE-Bot, Baichuan2, ChatGLM, Qwen, SparkDesk and etc. Out of the 24 models, 18 achieved non-hallucination rates lower than 50%. This indicates that HalluQA is highly challenging. We analyze the primary types of hallucinations in different types of models and their causes. Additionally, we discuss which types of hallucinations should be prioritized for different types of models.
2023-10-05
{'id': '2310.03368v4', 'title': 'Evaluating Hallucinations in Chinese Large Language Models', 'abstract': 'In this paper, we establish a benchmark named HalluQA (Chinese Hallucination\nQuestion-Answering) to measure the hallucination phenomenon in Chinese large\nlanguage models. HalluQA contains 450 meticulously designed adversarial\nquestions, spanning multiple domains, and takes into account Chinese historical\nculture, customs, and social phenomena. During the construction of HalluQA, we\nconsider two types of hallucinations: imitative falsehoods and factual errors,\nand we construct adversarial samples based on GLM-130B and ChatGPT. For\nevaluation, we design an automated evaluation method using GPT-4 to judge\nwhether a model output is hallucinated. We conduct extensive experiments on 24\nlarge language models, including ERNIE-Bot, Baichuan2, ChatGLM, Qwen, SparkDesk\nand etc. Out of the 24 models, 18 achieved non-hallucination rates lower than\n50%. This indicates that HalluQA is highly challenging. We analyze the primary\ntypes of hallucinations in different types of models and their causes.\nAdditionally, we discuss which types of hallucinations should be prioritized\nfor different types of models.', 'authors': ['Qinyuan Cheng', 'Tianxiang Sun', 'Wenwei Zhang', 'Siyin Wang', 'Xiangyang Liu', 'Mozhi Zhang', 'Junliang He', 'Mianqiu Huang', 'Zhangyue Yin', 'Kai Chen', 'Xipeng Qiu'], 'published': '2023-10-05 07:57:09+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2310.03368v4'}
42,699
2311.00416v1
2,311.00416
1
Efficient Human-AI Coordination via Preparatory Language-based Convention
Developing intelligent agents capable of seamless coordination with humans is a critical step towards achieving artificial general intelligence. Existing methods for human-AI coordination typically train an agent to coordinate with a diverse set of policies or with human models fitted from real human data. However, the massively diverse styles of human behavior present obstacles for AI systems with constrained capacity, while high quality human data may not be readily available in real-world scenarios. In this study, we observe that prior to coordination, humans engage in communication to establish conventions that specify individual roles and actions, making their coordination proceed in an orderly manner. Building upon this observation, we propose employing the large language model (LLM) to develop an action plan (or equivalently, a convention) that effectively guides both human and AI. By inputting task requirements, human preferences, the number of agents, and other pertinent information into the LLM, it can generate a comprehensive convention that facilitates a clear understanding of tasks and responsibilities for all parties involved. Furthermore, we demonstrate that decomposing the convention formulation problem into sub-problems with multiple new sessions being sequentially employed and human feedback, will yield a more efficient coordination convention. Experimental evaluations conducted in the Overcooked-AI environment, utilizing a human proxy model, highlight the superior performance of our proposed method compared to existing learning-based approaches. When coordinating with real humans, our method achieves better alignment with human preferences and an average performance improvement of 15% compared to the state-of-the-art.
2023-11-01
{'id': '2311.00416v1', 'title': 'Efficient Human-AI Coordination via Preparatory Language-based Convention', 'abstract': 'Developing intelligent agents capable of seamless coordination with humans is\na critical step towards achieving artificial general intelligence. Existing\nmethods for human-AI coordination typically train an agent to coordinate with a\ndiverse set of policies or with human models fitted from real human data.\nHowever, the massively diverse styles of human behavior present obstacles for\nAI systems with constrained capacity, while high quality human data may not be\nreadily available in real-world scenarios. In this study, we observe that prior\nto coordination, humans engage in communication to establish conventions that\nspecify individual roles and actions, making their coordination proceed in an\norderly manner. Building upon this observation, we propose employing the large\nlanguage model (LLM) to develop an action plan (or equivalently, a convention)\nthat effectively guides both human and AI. By inputting task requirements,\nhuman preferences, the number of agents, and other pertinent information into\nthe LLM, it can generate a comprehensive convention that facilitates a clear\nunderstanding of tasks and responsibilities for all parties involved.\nFurthermore, we demonstrate that decomposing the convention formulation problem\ninto sub-problems with multiple new sessions being sequentially employed and\nhuman feedback, will yield a more efficient coordination convention.\nExperimental evaluations conducted in the Overcooked-AI environment, utilizing\na human proxy model, highlight the superior performance of our proposed method\ncompared to existing learning-based approaches. When coordinating with real\nhumans, our method achieves better alignment with human preferences and an\naverage performance improvement of 15% compared to the state-of-the-art.', 'authors': ['Cong Guan', 'Lichao Zhang', 'Chunpeng Fan', 'Yichen Li', 'Feng Chen', 'Lihe Li', 'Yunjia Tian', 'Lei Yuan', 'Yang Yu'], 'published': '2023-11-01 10:18:23+00:00', 'categories': ['cs.LG', 'cs.CL'], 'url': 'http://arxiv.org/abs/2311.00416v1'}
42,701
2405.07414v2
2,405.07414
2
Binning as a Pretext Task: Improving Self-Supervised Learning in Tabular Domains
The ability of deep networks to learn superior representations hinges on leveraging the proper inductive biases, considering the inherent properties of datasets. In tabular domains, it is critical to effectively handle heterogeneous features (both categorical and numerical) in a unified manner and to grasp irregular functions like piecewise constant functions. To address the challenges in the self-supervised learning framework, we propose a novel pretext task based on the classical binning method. The idea is straightforward: reconstructing the bin indices (either orders or classes) rather than the original values. This pretext task provides the encoder with an inductive bias to capture the irregular dependencies, mapping from continuous inputs to discretized bins, and mitigates the feature heterogeneity by setting all features to have category-type targets. Our empirical investigations ascertain several advantages of binning: capturing the irregular function, compatibility with encoder architecture and additional modifications, standardizing all features into equal sets, grouping similar values within a feature, and providing ordering information. Comprehensive evaluations across diverse tabular datasets corroborate that our method consistently improves tabular representation learning performance for a wide range of downstream tasks. The codes are available in https://github.com/kyungeun-lee/tabularbinning.
2024-05-13
{'id': '2405.07414v2', 'title': 'Binning as a Pretext Task: Improving Self-Supervised Learning in Tabular Domains', 'abstract': 'The ability of deep networks to learn superior representations hinges on\nleveraging the proper inductive biases, considering the inherent properties of\ndatasets. In tabular domains, it is critical to effectively handle\nheterogeneous features (both categorical and numerical) in a unified manner and\nto grasp irregular functions like piecewise constant functions. To address the\nchallenges in the self-supervised learning framework, we propose a novel\npretext task based on the classical binning method. The idea is\nstraightforward: reconstructing the bin indices (either orders or classes)\nrather than the original values. This pretext task provides the encoder with an\ninductive bias to capture the irregular dependencies, mapping from continuous\ninputs to discretized bins, and mitigates the feature heterogeneity by setting\nall features to have category-type targets. Our empirical investigations\nascertain several advantages of binning: capturing the irregular function,\ncompatibility with encoder architecture and additional modifications,\nstandardizing all features into equal sets, grouping similar values within a\nfeature, and providing ordering information. Comprehensive evaluations across\ndiverse tabular datasets corroborate that our method consistently improves\ntabular representation learning performance for a wide range of downstream\ntasks. The codes are available in\nhttps://github.com/kyungeun-lee/tabularbinning.', 'authors': ['Kyungeun Lee', 'Ye Seul Sim', 'Hye-Seung Cho', 'Moonjung Eo', 'Suhee Yoon', 'Sanghyu Yoon', 'Woohyung Lim'], 'published': '2024-05-13 01:23:14+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2405.07414v2'}
42,702
2402.03348v2
2,402.03348
2
Respect the model: Fine-grained and Robust Explanation with Sharing Ratio Decomposition
The truthfulness of existing explanation methods in authentically elucidating the underlying model's decision-making process has been questioned. Existing methods have deviated from faithfully representing the model, thus susceptible to adversarial attacks. To address this, we propose a novel eXplainable AI (XAI) method called SRD (Sharing Ratio Decomposition), which sincerely reflects the model's inference process, resulting in significantly enhanced robustness in our explanations. Different from the conventional emphasis on the neuronal level, we adopt a vector perspective to consider the intricate nonlinear interactions between filters. We also introduce an interesting observation termed Activation-Pattern-Only Prediction (APOP), letting us emphasize the importance of inactive neurons and redefine relevance encapsulating all relevant information including both active and inactive neurons. Our method, SRD, allows for the recursive decomposition of a Pointwise Feature Vector (PFV), providing a high-resolution Effective Receptive Field (ERF) at any layer.
2024-01-25
{'id': '2402.03348v2', 'title': 'Respect the model: Fine-grained and Robust Explanation with Sharing Ratio Decomposition', 'abstract': "The truthfulness of existing explanation methods in authentically elucidating\nthe underlying model's decision-making process has been questioned. Existing\nmethods have deviated from faithfully representing the model, thus susceptible\nto adversarial attacks. To address this, we propose a novel eXplainable AI\n(XAI) method called SRD (Sharing Ratio Decomposition), which sincerely reflects\nthe model's inference process, resulting in significantly enhanced robustness\nin our explanations. Different from the conventional emphasis on the neuronal\nlevel, we adopt a vector perspective to consider the intricate nonlinear\ninteractions between filters. We also introduce an interesting observation\ntermed Activation-Pattern-Only Prediction (APOP), letting us emphasize the\nimportance of inactive neurons and redefine relevance encapsulating all\nrelevant information including both active and inactive neurons. Our method,\nSRD, allows for the recursive decomposition of a Pointwise Feature Vector\n(PFV), providing a high-resolution Effective Receptive Field (ERF) at any\nlayer.", 'authors': ['Sangyu Han', 'Yearim Kim', 'Nojun Kwak'], 'published': '2024-01-25 07:20:23+00:00', 'categories': ['cs.CV', 'cs.AI'], 'url': 'http://arxiv.org/abs/2402.03348v2'}