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
42,515
2307.06962v1
2,307.06962
1
Copy Is All You Need
The dominant text generation models compose the output by sequentially selecting words from a fixed vocabulary. In this paper, we formulate text generation as progressively copying text segments (e.g., words or phrases) from an existing text collection. We compute the contextualized representations of meaningful text segments and index them using efficient vector search toolkits. The task of text generation is then decomposed into a series of copy-and-paste operations: at each time step, we seek suitable text spans from the text collection rather than selecting from a standalone vocabulary. Experiments on the standard language modeling benchmark (WikiText-103) show that our approach achieves better generation quality according to both automatic and human evaluations. Besides, its inference efficiency is comparable to token-level autoregressive models thanks to the reduction of decoding steps. We also show that our approach allows for effective domain adaptation by simply switching to domain-specific text collection without extra training. Finally, we observe that our approach attains additional performance gains by simply scaling up to larger text collections, again without further training.Our source codes are publicly available at https://github.com/gmftbyGMFTBY/Copyisallyouneed.
2023-07-13
{'id': '2307.06962v1', 'title': 'Copy Is All You Need', 'abstract': 'The dominant text generation models compose the output by sequentially\nselecting words from a fixed vocabulary. In this paper, we formulate text\ngeneration as progressively copying text segments (e.g., words or phrases) from\nan existing text collection. We compute the contextualized representations of\nmeaningful text segments and index them using efficient vector search toolkits.\nThe task of text generation is then decomposed into a series of copy-and-paste\noperations: at each time step, we seek suitable text spans from the text\ncollection rather than selecting from a standalone vocabulary. Experiments on\nthe standard language modeling benchmark (WikiText-103) show that our approach\nachieves better generation quality according to both automatic and human\nevaluations. Besides, its inference efficiency is comparable to token-level\nautoregressive models thanks to the reduction of decoding steps. We also show\nthat our approach allows for effective domain adaptation by simply switching to\ndomain-specific text collection without extra training. Finally, we observe\nthat our approach attains additional performance gains by simply scaling up to\nlarger text collections, again without further training.\\footnote{Our source\ncodes are publicly available at\n\\url{https://github.com/gmftbyGMFTBY/Copyisallyouneed}.}', 'authors': ['Tian Lan', 'Deng Cai', 'Yan Wang', 'Heyan Huang', 'Xian-Ling Mao'], 'published': '2023-07-13 05:03:26+00:00', 'categories': ['cs.CL', 'cs.AI'], 'url': 'http://arxiv.org/abs/2307.06962v1'}
42,718
2212.10701v2
2,212.10701
2
A Non-Asymptotic Analysis of Oversmoothing in Graph Neural Networks
Oversmoothing is a central challenge of building more powerful Graph Neural Networks (GNNs). While previous works have only demonstrated that oversmoothing is inevitable when the number of graph convolutions tends to infinity, in this paper, we precisely characterize the mechanism behind the phenomenon via a non-asymptotic analysis. Specifically, we distinguish between two different effects when applying graph convolutions -- an undesirable mixing effect that homogenizes node representations in different classes, and a desirable denoising effect that homogenizes node representations in the same class. By quantifying these two effects on random graphs sampled from the Contextual Stochastic Block Model (CSBM), we show that oversmoothing happens once the mixing effect starts to dominate the denoising effect, and the number of layers required for this transition is $O(N/(N))$ for sufficiently dense graphs with $N$ nodes. We also extend our analysis to study the effects of Personalized PageRank (PPR), or equivalently, the effects of initial residual connections on oversmoothing. Our results suggest that while PPR mitigates oversmoothing at deeper layers, PPR-based architectures still achieve their best performance at a shallow depth and are outperformed by the graph convolution approach on certain graphs. Finally, we support our theoretical results with numerical experiments, which further suggest that the oversmoothing phenomenon observed in practice can be magnified by the difficulty of optimizing deep GNN models.
2022-12-21
{'id': '2212.10701v2', 'title': 'A Non-Asymptotic Analysis of Oversmoothing in Graph Neural Networks', 'abstract': 'Oversmoothing is a central challenge of building more powerful Graph Neural\nNetworks (GNNs). While previous works have only demonstrated that oversmoothing\nis inevitable when the number of graph convolutions tends to infinity, in this\npaper, we precisely characterize the mechanism behind the phenomenon via a\nnon-asymptotic analysis. Specifically, we distinguish between two different\neffects when applying graph convolutions -- an undesirable mixing effect that\nhomogenizes node representations in different classes, and a desirable\ndenoising effect that homogenizes node representations in the same class. By\nquantifying these two effects on random graphs sampled from the Contextual\nStochastic Block Model (CSBM), we show that oversmoothing happens once the\nmixing effect starts to dominate the denoising effect, and the number of layers\nrequired for this transition is $O(\\log N/\\log (\\log N))$ for sufficiently\ndense graphs with $N$ nodes. We also extend our analysis to study the effects\nof Personalized PageRank (PPR), or equivalently, the effects of initial\nresidual connections on oversmoothing. Our results suggest that while PPR\nmitigates oversmoothing at deeper layers, PPR-based architectures still achieve\ntheir best performance at a shallow depth and are outperformed by the graph\nconvolution approach on certain graphs. Finally, we support our theoretical\nresults with numerical experiments, which further suggest that the\noversmoothing phenomenon observed in practice can be magnified by the\ndifficulty of optimizing deep GNN models.', 'authors': ['Xinyi Wu', 'Zhengdao Chen', 'William Wang', 'Ali Jadbabaie'], 'published': '2022-12-21 00:33:59+00:00', 'categories': ['cs.LG', 'cs.SI', 'stat.ML'], 'url': 'http://arxiv.org/abs/2212.10701v2'}
42,720
2305.14715v1
2,305.14715
1
Leveraging Future Relationship Reasoning for Vehicle Trajectory Prediction
Understanding the interaction between multiple agents is crucial for realistic vehicle trajectory prediction. Existing methods have attempted to infer the interaction from the observed past trajectories of agents using pooling, attention, or graph-based methods, which rely on a deterministic approach. However, these methods can fail under complex road structures, as they cannot predict various interactions that may occur in the future. In this paper, we propose a novel approach that uses lane information to predict a stochastic future relationship among agents. To obtain a coarse future motion of agents, our method first predicts the probability of lane-level waypoint occupancy of vehicles. We then utilize the temporal probability of passing adjacent lanes for each agent pair, assuming that agents passing adjacent lanes will highly interact. We also model the interaction using a probabilistic distribution, which allows for multiple possible future interactions. The distribution is learned from the posterior distribution of interaction obtained from ground truth future trajectories. We validate our method on popular trajectory prediction datasets: nuScenes and Argoverse. The results show that the proposed method brings remarkable performance gain in prediction accuracy, and achieves state-of-the-art performance in long-term prediction benchmark dataset.
2023-05-24
{'id': '2305.14715v1', 'title': 'Leveraging Future Relationship Reasoning for Vehicle Trajectory Prediction', 'abstract': 'Understanding the interaction between multiple agents is crucial for\nrealistic vehicle trajectory prediction. Existing methods have attempted to\ninfer the interaction from the observed past trajectories of agents using\npooling, attention, or graph-based methods, which rely on a deterministic\napproach. However, these methods can fail under complex road structures, as\nthey cannot predict various interactions that may occur in the future. In this\npaper, we propose a novel approach that uses lane information to predict a\nstochastic future relationship among agents. To obtain a coarse future motion\nof agents, our method first predicts the probability of lane-level waypoint\noccupancy of vehicles. We then utilize the temporal probability of passing\nadjacent lanes for each agent pair, assuming that agents passing adjacent lanes\nwill highly interact. We also model the interaction using a probabilistic\ndistribution, which allows for multiple possible future interactions. The\ndistribution is learned from the posterior distribution of interaction obtained\nfrom ground truth future trajectories. We validate our method on popular\ntrajectory prediction datasets: nuScenes and Argoverse. The results show that\nthe proposed method brings remarkable performance gain in prediction accuracy,\nand achieves state-of-the-art performance in long-term prediction benchmark\ndataset.', 'authors': ['Daehee Park', 'Hobin Ryu', 'Yunseo Yang', 'Jegyeong Cho', 'Jiwon Kim', 'Kuk-Jin Yoon'], 'published': '2023-05-24 04:33:28+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2305.14715v1'}
42,487
2005.08140v5
2,005.0814
5
Global inducing point variational posteriors for Bayesian neural networks and deep Gaussian processes
We consider the optimal approximate posterior over the top-layer weights in a Bayesian neural network for regression, and show that it exhibits strong dependencies on the lower-layer weights. We adapt this result to develop a correlated approximate posterior over the weights at all layers in a Bayesian neural network. We extend this approach to deep Gaussian processes, unifying inference in the two model classes. Our approximate posterior uses learned "global" inducing points, which are defined only at the input layer and propagated through the network to obtain inducing inputs at subsequent layers. By contrast, standard, "local", inducing point methods from the deep Gaussian process literature optimise a separate set of inducing inputs at every layer, and thus do not model correlations across layers. Our method gives state-of-the-art performance for a variational Bayesian method, without data augmentation or tempering, on CIFAR-10 of 86.7%, which is comparable to SGMCMC without tempering but with data augmentation (88% in Wenzel et al. 2020).
2020-05-17
{'id': '2005.08140v5', 'title': 'Global inducing point variational posteriors for Bayesian neural networks and deep Gaussian processes', 'abstract': 'We consider the optimal approximate posterior over the top-layer weights in a\nBayesian neural network for regression, and show that it exhibits strong\ndependencies on the lower-layer weights. We adapt this result to develop a\ncorrelated approximate posterior over the weights at all layers in a Bayesian\nneural network. We extend this approach to deep Gaussian processes, unifying\ninference in the two model classes. Our approximate posterior uses learned\n"global" inducing points, which are defined only at the input layer and\npropagated through the network to obtain inducing inputs at subsequent layers.\nBy contrast, standard, "local", inducing point methods from the deep Gaussian\nprocess literature optimise a separate set of inducing inputs at every layer,\nand thus do not model correlations across layers. Our method gives\nstate-of-the-art performance for a variational Bayesian method, without data\naugmentation or tempering, on CIFAR-10 of 86.7%, which is comparable to SGMCMC\nwithout tempering but with data augmentation (88% in Wenzel et al. 2020).', 'authors': ['Sebastian W. Ober', 'Laurence Aitchison'], 'published': '2020-05-17 01:10:37+00:00', 'categories': ['stat.ML', 'cs.LG'], 'url': 'http://arxiv.org/abs/2005.08140v5'}
42,489
2102.04653v2
2,102.04653
2
Proximal Gradient Descent-Ascent: Variable Convergence under KŁ Geometry
The gradient descent-ascent (GDA) algorithm has been widely applied to solve minimax optimization problems. In order to achieve convergent policy parameters for minimax optimization, it is important that GDA generates convergent variable sequences rather than convergent sequences of function values or gradient norms. However, the variable convergence of GDA has been proved only under convexity geometries, and there lacks understanding for general nonconvex minimax optimization. This paper fills such a gap by studying the convergence of a more general proximal-GDA for regularized nonconvex-strongly-concave minimax optimization. Specifically, we show that proximal-GDA admits a novel Lyapunov function, which monotonically decreases in the minimax optimization process and drives the variable sequence to a critical point. By leveraging this Lyapunov function and the K geometry that parameterizes the local geometries of general nonconvex functions, we formally establish the variable convergence of proximal-GDA to a critical point $x^*$, i.e., $x_tx^*, y_ty^*(x^*)$. Furthermore, over the full spectrum of the K-parameterized geometry, we show that proximal-GDA achieves different types of convergence rates ranging from sublinear convergence up to finite-step convergence, depending on the geometry associated with the K parameter. This is the first theoretical result on the variable convergence for nonconvex minimax optimization.
2021-02-09
{'id': '2102.04653v2', 'title': 'Proximal Gradient Descent-Ascent: Variable Convergence under KŁ Geometry', 'abstract': 'The gradient descent-ascent (GDA) algorithm has been widely applied to solve\nminimax optimization problems. In order to achieve convergent policy parameters\nfor minimax optimization, it is important that GDA generates convergent\nvariable sequences rather than convergent sequences of function values or\ngradient norms. However, the variable convergence of GDA has been proved only\nunder convexity geometries, and there lacks understanding for general nonconvex\nminimax optimization. This paper fills such a gap by studying the convergence\nof a more general proximal-GDA for regularized nonconvex-strongly-concave\nminimax optimization. Specifically, we show that proximal-GDA admits a novel\nLyapunov function, which monotonically decreases in the minimax optimization\nprocess and drives the variable sequence to a critical point. By leveraging\nthis Lyapunov function and the K{\\L} geometry that parameterizes the local\ngeometries of general nonconvex functions, we formally establish the variable\nconvergence of proximal-GDA to a critical point $x^*$, i.e., $x_t\\to x^*,\ny_t\\to y^*(x^*)$. Furthermore, over the full spectrum of the\nK{\\L}-parameterized geometry, we show that proximal-GDA achieves different\ntypes of convergence rates ranging from sublinear convergence up to finite-step\nconvergence, depending on the geometry associated with the K{\\L} parameter.\nThis is the first theoretical result on the variable convergence for nonconvex\nminimax optimization.', 'authors': ['Ziyi Chen', 'Yi Zhou', 'Tengyu Xu', 'Yingbin Liang'], 'published': '2021-02-09 05:35:53+00:00', 'categories': ['math.OC', 'cs.LG'], 'url': 'http://arxiv.org/abs/2102.04653v2'}
42,491
2006.10369v4
2,006.10369
4
Deep Encoder, Shallow Decoder: Reevaluating Non-autoregressive Machine Translation
Much recent effort has been invested in non-autoregressive neural machine translation, which appears to be an efficient alternative to state-of-the-art autoregressive machine translation on modern GPUs. In contrast to the latter, where generation is sequential, the former allows generation to be parallelized across target token positions. Some of the latest non-autoregressive models have achieved impressive translation quality-speed tradeoffs compared to autoregressive baselines. In this work, we reexamine this tradeoff and argue that autoregressive baselines can be substantially sped up without loss in accuracy. Specifically, we study autoregressive models with encoders and decoders of varied depths. Our extensive experiments show that given a sufficiently deep encoder, a single-layer autoregressive decoder can substantially outperform strong non-autoregressive models with comparable inference speed. We show that the speed disadvantage for autoregressive baselines compared to non-autoregressive methods has been overestimated in three aspects: suboptimal layer allocation, insufficient speed measurement, and lack of knowledge distillation. Our results establish a new protocol for future research toward fast, accurate machine translation. Our code is available at https://github.com/jungokasai/deep-shallow.
2020-06-18
{'id': '2006.10369v4', 'title': 'Deep Encoder, Shallow Decoder: Reevaluating Non-autoregressive Machine Translation', 'abstract': 'Much recent effort has been invested in non-autoregressive neural machine\ntranslation, which appears to be an efficient alternative to state-of-the-art\nautoregressive machine translation on modern GPUs. In contrast to the latter,\nwhere generation is sequential, the former allows generation to be parallelized\nacross target token positions. Some of the latest non-autoregressive models\nhave achieved impressive translation quality-speed tradeoffs compared to\nautoregressive baselines. In this work, we reexamine this tradeoff and argue\nthat autoregressive baselines can be substantially sped up without loss in\naccuracy. Specifically, we study autoregressive models with encoders and\ndecoders of varied depths. Our extensive experiments show that given a\nsufficiently deep encoder, a single-layer autoregressive decoder can\nsubstantially outperform strong non-autoregressive models with comparable\ninference speed. We show that the speed disadvantage for autoregressive\nbaselines compared to non-autoregressive methods has been overestimated in\nthree aspects: suboptimal layer allocation, insufficient speed measurement, and\nlack of knowledge distillation. Our results establish a new protocol for future\nresearch toward fast, accurate machine translation. Our code is available at\nhttps://github.com/jungokasai/deep-shallow.', 'authors': ['Jungo Kasai', 'Nikolaos Pappas', 'Hao Peng', 'James Cross', 'Noah A. Smith'], 'published': '2020-06-18 09:06:49+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2006.10369v4'}
42,493
2008.10766v1
2,008.10766
1
Channel-Directed Gradients for Optimization of Convolutional Neural Networks
We introduce optimization methods for convolutional neural networks that can be used to improve existing gradient-based optimization in terms of generalization error. The method requires only simple processing of existing stochastic gradients, can be used in conjunction with any optimizer, and has only a linear overhead (in the number of parameters) compared to computation of the stochastic gradient. The method works by computing the gradient of the loss function with respect to output-channel directed re-weighted L2 or Sobolev metrics, which has the effect of smoothing components of the gradient across a certain direction of the parameter tensor. We show that defining the gradients along the output channel direction leads to a performance boost, while other directions can be detrimental. We present the continuum theory of such gradients, its discretization, and application to deep networks. Experiments on benchmark datasets, several networks and baseline optimizers show that optimizers can be improved in generalization error by simply computing the stochastic gradient with respect to output-channel directed metrics.
2020-08-25
{'id': '2008.10766v1', 'title': 'Channel-Directed Gradients for Optimization of Convolutional Neural Networks', 'abstract': 'We introduce optimization methods for convolutional neural networks that can\nbe used to improve existing gradient-based optimization in terms of\ngeneralization error. The method requires only simple processing of existing\nstochastic gradients, can be used in conjunction with any optimizer, and has\nonly a linear overhead (in the number of parameters) compared to computation of\nthe stochastic gradient. The method works by computing the gradient of the loss\nfunction with respect to output-channel directed re-weighted L2 or Sobolev\nmetrics, which has the effect of smoothing components of the gradient across a\ncertain direction of the parameter tensor. We show that defining the gradients\nalong the output channel direction leads to a performance boost, while other\ndirections can be detrimental. We present the continuum theory of such\ngradients, its discretization, and application to deep networks. Experiments on\nbenchmark datasets, several networks and baseline optimizers show that\noptimizers can be improved in generalization error by simply computing the\nstochastic gradient with respect to output-channel directed metrics.', 'authors': ['Dong Lao', 'Peihao Zhu', 'Peter Wonka', 'Ganesh Sundaramoorthi'], 'published': '2020-08-25 00:44:09+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CV', 'stat.ML'], 'url': 'http://arxiv.org/abs/2008.10766v1'}
42,501
2106.07207v1
2,106.07207
1
Straight to the Gradient: Learning to Use Novel Tokens for Neural Text Generation
Advanced large-scale neural language models have led to significant success in many language generation tasks. However, the most commonly used training objective, Maximum Likelihood Estimation (MLE), has been shown problematic, where the trained model prefers using dull and repetitive phrases. In this work, we introduce ScaleGrad, a modification straight to the gradient of the loss function, to remedy the degeneration issue of the standard MLE objective. By directly maneuvering the gradient information, ScaleGrad makes the model learn to use novel tokens. Empirical results show the effectiveness of our method not only in open-ended generation, but also in directed generation tasks. With the simplicity in architecture, our method can serve as a general training objective that is applicable to most of the neural text generation tasks.
2021-06-14
{'id': '2106.07207v1', 'title': 'Straight to the Gradient: Learning to Use Novel Tokens for Neural Text Generation', 'abstract': 'Advanced large-scale neural language models have led to significant success\nin many language generation tasks. However, the most commonly used training\nobjective, Maximum Likelihood Estimation (MLE), has been shown problematic,\nwhere the trained model prefers using dull and repetitive phrases. In this\nwork, we introduce ScaleGrad, a modification straight to the gradient of the\nloss function, to remedy the degeneration issue of the standard MLE objective.\nBy directly maneuvering the gradient information, ScaleGrad makes the model\nlearn to use novel tokens. Empirical results show the effectiveness of our\nmethod not only in open-ended generation, but also in directed generation\ntasks. With the simplicity in architecture, our method can serve as a general\ntraining objective that is applicable to most of the neural text generation\ntasks.', 'authors': ['Xiang Lin', 'Simeng Han', 'Shafiq Joty'], 'published': '2021-06-14 07:46:30+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2106.07207v1'}
42,503
2006.07034v2
2,006.07034
2
Benchmarking Unsupervised Object Representations for Video Sequences
Perceiving the world in terms of objects and tracking them through time is a crucial prerequisite for reasoning and scene understanding. Recently, several methods have been proposed for unsupervised learning of object-centric representations. However, since these models were evaluated on different downstream tasks, it remains unclear how they compare in terms of basic perceptual abilities such as detection, figure-ground segmentation and tracking of objects. To close this gap, we design a benchmark with four data sets of varying complexity and seven additional test sets featuring challenging tracking scenarios relevant for natural videos. Using this benchmark, we compare the perceptual abilities of four object-centric approaches: ViMON, a video-extension of MONet, based on recurrent spatial attention, OP3, which exploits clustering via spatial mixture models, as well as TBA and SCALOR, which use explicit factorization via spatial transformers. Our results suggest that the architectures with unconstrained latent representations learn more powerful representations in terms of object detection, segmentation and tracking than the spatial transformer based architectures. We also observe that none of the methods are able to gracefully handle the most challenging tracking scenarios despite their synthetic nature, suggesting that our benchmark may provide fruitful guidance towards learning more robust object-centric video representations.
2020-06-12
{'id': '2006.07034v2', 'title': 'Benchmarking Unsupervised Object Representations for Video Sequences', 'abstract': 'Perceiving the world in terms of objects and tracking them through time is a\ncrucial prerequisite for reasoning and scene understanding. Recently, several\nmethods have been proposed for unsupervised learning of object-centric\nrepresentations. However, since these models were evaluated on different\ndownstream tasks, it remains unclear how they compare in terms of basic\nperceptual abilities such as detection, figure-ground segmentation and tracking\nof objects. To close this gap, we design a benchmark with four data sets of\nvarying complexity and seven additional test sets featuring challenging\ntracking scenarios relevant for natural videos. Using this benchmark, we\ncompare the perceptual abilities of four object-centric approaches: ViMON, a\nvideo-extension of MONet, based on recurrent spatial attention, OP3, which\nexploits clustering via spatial mixture models, as well as TBA and SCALOR,\nwhich use explicit factorization via spatial transformers. Our results suggest\nthat the architectures with unconstrained latent representations learn more\npowerful representations in terms of object detection, segmentation and\ntracking than the spatial transformer based architectures. We also observe that\nnone of the methods are able to gracefully handle the most challenging tracking\nscenarios despite their synthetic nature, suggesting that our benchmark may\nprovide fruitful guidance towards learning more robust object-centric video\nrepresentations.', 'authors': ['Marissa A. Weis', 'Kashyap Chitta', 'Yash Sharma', 'Wieland Brendel', 'Matthias Bethge', 'Andreas Geiger', 'Alexander S. Ecker'], 'published': '2020-06-12 09:37:24+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2006.07034v2'}
42,494
2306.12981v1
2,306.12981
1
Achieving Sample and Computational Efficient Reinforcement Learning by Action Space Reduction via Grouping
Reinforcement learning often needs to deal with the exponential growth of states and actions when exploring optimal control in high-dimensional spaces (often known as the curse of dimensionality). In this work, we address this issue by learning the inherent structure of action-wise similar MDP to appropriately balance the performance degradation versus sample/computational complexity. In particular, we partition the action spaces into multiple groups based on the similarity in transition distribution and reward function, and build a linear decomposition model to capture the difference between the intra-group transition kernel and the intra-group rewards. Both our theoretical analysis and experiments reveal a surprising and counter-intuitive result: while a more refined grouping strategy can reduce the approximation error caused by treating actions in the same group as identical, it also leads to increased estimation error when the size of samples or the computation resources is limited. This finding highlights the grouping strategy as a new degree of freedom that can be optimized to minimize the overall performance loss. To address this issue, we formulate a general optimization problem for determining the optimal grouping strategy, which strikes a balance between performance loss and sample/computational complexity. We further propose a computationally efficient method for selecting a nearly-optimal grouping strategy, which maintains its computational complexity independent of the size of the action space.
2023-06-22
{'id': '2306.12981v1', 'title': 'Achieving Sample and Computational Efficient Reinforcement Learning by Action Space Reduction via Grouping', 'abstract': 'Reinforcement learning often needs to deal with the exponential growth of\nstates and actions when exploring optimal control in high-dimensional spaces\n(often known as the curse of dimensionality). In this work, we address this\nissue by learning the inherent structure of action-wise similar MDP to\nappropriately balance the performance degradation versus sample/computational\ncomplexity. In particular, we partition the action spaces into multiple groups\nbased on the similarity in transition distribution and reward function, and\nbuild a linear decomposition model to capture the difference between the\nintra-group transition kernel and the intra-group rewards. Both our theoretical\nanalysis and experiments reveal a \\emph{surprising and counter-intuitive\nresult}: while a more refined grouping strategy can reduce the approximation\nerror caused by treating actions in the same group as identical, it also leads\nto increased estimation error when the size of samples or the computation\nresources is limited. This finding highlights the grouping strategy as a new\ndegree of freedom that can be optimized to minimize the overall performance\nloss. To address this issue, we formulate a general optimization problem for\ndetermining the optimal grouping strategy, which strikes a balance between\nperformance loss and sample/computational complexity. We further propose a\ncomputationally efficient method for selecting a nearly-optimal grouping\nstrategy, which maintains its computational complexity independent of the size\nof the action space.', 'authors': ['Yining Li', 'Peizhong Ju', 'Ness Shroff'], 'published': '2023-06-22 15:40:10+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2306.12981v1'}
42,496
2311.13774v1
2,311.13774
1
Learning Hierarchical Polynomials with Three-Layer Neural Networks
We study the problem of learning hierarchical polynomials over the standard Gaussian distribution with three-layer neural networks. We specifically consider target functions of the form $h = g p$ where $p : R^d R$ is a degree $k$ polynomial and $g: R R$ is a degree $q$ polynomial. This function class generalizes the single-index model, which corresponds to $k=1$, and is a natural class of functions possessing an underlying hierarchical structure. Our main result shows that for a large subclass of degree $k$ polynomials $p$, a three-layer neural network trained via layerwise gradient descent on the square loss learns the target $h$ up to vanishing test error in $O(d^k)$ samples and polynomial time. This is a strict improvement over kernel methods, which require $(d^kq)$ samples, as well as existing guarantees for two-layer networks, which require the target function to be low-rank. Our result also generalizes prior works on three-layer neural networks, which were restricted to the case of $p$ being a quadratic. When $p$ is indeed a quadratic, we achieve the information-theoretically optimal sample complexity $O(d^2)$, which is an improvement over prior work~nichani2023provable requiring a sample size of $(d^4)$. Our proof proceeds by showing that during the initial stage of training the network performs feature learning to recover the feature $p$ with $O(d^k)$ samples. This work demonstrates the ability of three-layer neural networks to learn complex features and as a result, learn a broad class of hierarchical functions.
2023-11-23
{'id': '2311.13774v1', 'title': 'Learning Hierarchical Polynomials with Three-Layer Neural Networks', 'abstract': 'We study the problem of learning hierarchical polynomials over the standard\nGaussian distribution with three-layer neural networks. We specifically\nconsider target functions of the form $h = g \\circ p$ where $p : \\mathbb{R}^d\n\\rightarrow \\mathbb{R}$ is a degree $k$ polynomial and $g: \\mathbb{R}\n\\rightarrow \\mathbb{R}$ is a degree $q$ polynomial. This function class\ngeneralizes the single-index model, which corresponds to $k=1$, and is a\nnatural class of functions possessing an underlying hierarchical structure. Our\nmain result shows that for a large subclass of degree $k$ polynomials $p$, a\nthree-layer neural network trained via layerwise gradient descent on the square\nloss learns the target $h$ up to vanishing test error in\n$\\widetilde{\\mathcal{O}}(d^k)$ samples and polynomial time. This is a strict\nimprovement over kernel methods, which require $\\widetilde \\Theta(d^{kq})$\nsamples, as well as existing guarantees for two-layer networks, which require\nthe target function to be low-rank. Our result also generalizes prior works on\nthree-layer neural networks, which were restricted to the case of $p$ being a\nquadratic. When $p$ is indeed a quadratic, we achieve the\ninformation-theoretically optimal sample complexity\n$\\widetilde{\\mathcal{O}}(d^2)$, which is an improvement over prior\nwork~\\citep{nichani2023provable} requiring a sample size of\n$\\widetilde\\Theta(d^4)$. Our proof proceeds by showing that during the initial\nstage of training the network performs feature learning to recover the feature\n$p$ with $\\widetilde{\\mathcal{O}}(d^k)$ samples. This work demonstrates the\nability of three-layer neural networks to learn complex features and as a\nresult, learn a broad class of hierarchical functions.', 'authors': ['Zihao Wang', 'Eshaan Nichani', 'Jason D. Lee'], 'published': '2023-11-23 02:19:32+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2311.13774v1'}
42,508
2403.09953v1
2,403.09953
1
Online GNN Evaluation Under Test-time Graph Distribution Shifts
Evaluating the performance of a well-trained GNN model on real-world graphs is a pivotal step for reliable GNN online deployment and serving. Due to a lack of test node labels and unknown potential training-test graph data distribution shifts, conventional model evaluation encounters limitations in calculating performance metrics (e.g., test error) and measuring graph data-level discrepancies, particularly when the training graph used for developing GNNs remains unobserved during test time. In this paper, we study a new research problem, online GNN evaluation, which aims to provide valuable insights into the well-trained GNNs's ability to effectively generalize to real-world unlabeled graphs under the test-time graph distribution shifts. Concretely, we develop an effective learning behavior discrepancy score, dubbed LeBeD, to estimate the test-time generalization errors of well-trained GNN models. Through a novel GNN re-training strategy with a parameter-free optimality criterion, the proposed LeBeD comprehensively integrates learning behavior discrepancies from both node prediction and structure reconstruction perspectives. This enables the effective evaluation of the well-trained GNNs' ability to capture test node semantics and structural representations, making it an expressive metric for estimating the generalization error in online GNN evaluation. Extensive experiments on real-world test graphs under diverse graph distribution shifts could verify the effectiveness of the proposed method, revealing its strong correlation with ground-truth test errors on various well-trained GNN models.
2024-03-15
{'id': '2403.09953v1', 'title': 'Online GNN Evaluation Under Test-time Graph Distribution Shifts', 'abstract': "Evaluating the performance of a well-trained GNN model on real-world graphs\nis a pivotal step for reliable GNN online deployment and serving. Due to a lack\nof test node labels and unknown potential training-test graph data distribution\nshifts, conventional model evaluation encounters limitations in calculating\nperformance metrics (e.g., test error) and measuring graph data-level\ndiscrepancies, particularly when the training graph used for developing GNNs\nremains unobserved during test time. In this paper, we study a new research\nproblem, online GNN evaluation, which aims to provide valuable insights into\nthe well-trained GNNs's ability to effectively generalize to real-world\nunlabeled graphs under the test-time graph distribution shifts. Concretely, we\ndevelop an effective learning behavior discrepancy score, dubbed LeBeD, to\nestimate the test-time generalization errors of well-trained GNN models.\nThrough a novel GNN re-training strategy with a parameter-free optimality\ncriterion, the proposed LeBeD comprehensively integrates learning behavior\ndiscrepancies from both node prediction and structure reconstruction\nperspectives. This enables the effective evaluation of the well-trained GNNs'\nability to capture test node semantics and structural representations, making\nit an expressive metric for estimating the generalization error in online GNN\nevaluation. Extensive experiments on real-world test graphs under diverse graph\ndistribution shifts could verify the effectiveness of the proposed method,\nrevealing its strong correlation with ground-truth test errors on various\nwell-trained GNN models.", 'authors': ['Xin Zheng', 'Dongjin Song', 'Qingsong Wen', 'Bo Du', 'Shirui Pan'], 'published': '2024-03-15 01:28:08+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2403.09953v1'}
42,513
2310.19796v3
2,310.19796
3
Re-evaluating Retrosynthesis Algorithms with Syntheseus
Automated Synthesis Planning has recently re-emerged as a research area at the intersection of chemistry and machine learning. Despite the appearance of steady progress, we argue that imperfect benchmarks and inconsistent comparisons mask systematic shortcomings of existing techniques, and unnecessarily hamper progress. To remedy this, we present a synthesis planning library with an extensive benchmarking framework, called syntheseus, which promotes best practice by default, enabling consistent meaningful evaluation of single-step models and multi-step planning algorithms. We demonstrate the capabilities of syntheseus by re-evaluating several previous retrosynthesis algorithms, and find that the ranking of state-of-the-art models changes in controlled evaluation experiments. We end with guidance for future works in this area, and call the community to engage in the discussion on how to improve benchmarks for synthesis planning.
2023-10-30
{'id': '2310.19796v3', 'title': 'Re-evaluating Retrosynthesis Algorithms with Syntheseus', 'abstract': 'Automated Synthesis Planning has recently re-emerged as a research area at\nthe intersection of chemistry and machine learning. Despite the appearance of\nsteady progress, we argue that imperfect benchmarks and inconsistent\ncomparisons mask systematic shortcomings of existing techniques, and\nunnecessarily hamper progress. To remedy this, we present a synthesis planning\nlibrary with an extensive benchmarking framework, called syntheseus, which\npromotes best practice by default, enabling consistent meaningful evaluation of\nsingle-step models and multi-step planning algorithms. We demonstrate the\ncapabilities of syntheseus by re-evaluating several previous retrosynthesis\nalgorithms, and find that the ranking of state-of-the-art models changes in\ncontrolled evaluation experiments. We end with guidance for future works in\nthis area, and call the community to engage in the discussion on how to improve\nbenchmarks for synthesis planning.', 'authors': ['Krzysztof Maziarz', 'Austin Tripp', 'Guoqing Liu', 'Megan Stanley', 'Shufang Xie', 'Piotr Gaiński', 'Philipp Seidl', 'Marwin Segler'], 'published': '2023-10-30 17:59:04+00:00', 'categories': ['cs.LG', 'cs.AI', 'q-bio.QM'], 'url': 'http://arxiv.org/abs/2310.19796v3'}
42,520
2210.01422v5
2,210.01422
5
Time-Varying Propensity Score to Bridge the Gap between the Past and Present
Real-world deployment of machine learning models is challenging because data evolves over time. While no model can work when data evolves in an arbitrary fashion, if there is some pattern to these changes, we might be able to design methods to address it. This paper addresses situations when data evolves gradually. We introduce a time-varying propensity score that can detect gradual shifts in the distribution of data which allows us to selectively sample past data to update the model -- not just similar data from the past like that of a standard propensity score but also data that evolved in a similar fashion in the past. The time-varying propensity score is quite general: we demonstrate different ways of implementing it and evaluate it on a variety of problems ranging from supervised learning (e.g., image classification problems) where data undergoes a sequence of gradual shifts, to reinforcement learning tasks (e.g., robotic manipulation and continuous control) where data shifts as the policy or the task changes.
2022-10-04
{'id': '2210.01422v5', 'title': 'Time-Varying Propensity Score to Bridge the Gap between the Past and Present', 'abstract': 'Real-world deployment of machine learning models is challenging because data\nevolves over time. While no model can work when data evolves in an arbitrary\nfashion, if there is some pattern to these changes, we might be able to design\nmethods to address it. This paper addresses situations when data evolves\ngradually. We introduce a time-varying propensity score that can detect gradual\nshifts in the distribution of data which allows us to selectively sample past\ndata to update the model -- not just similar data from the past like that of a\nstandard propensity score but also data that evolved in a similar fashion in\nthe past. The time-varying propensity score is quite general: we demonstrate\ndifferent ways of implementing it and evaluate it on a variety of problems\nranging from supervised learning (e.g., image classification problems) where\ndata undergoes a sequence of gradual shifts, to reinforcement learning tasks\n(e.g., robotic manipulation and continuous control) where data shifts as the\npolicy or the task changes.', 'authors': ['Rasool Fakoor', 'Jonas Mueller', 'Zachary C. Lipton', 'Pratik Chaudhari', 'Alexander J. Smola'], 'published': '2022-10-04 07:21:49+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2210.01422v5'}
42,526
2310.04966v2
2,310.04966
2
Improved Active Learning via Dependent Leverage Score Sampling
We show how to obtain improved active learning methods in the agnostic (adversarial noise) setting by combining marginal leverage score sampling with non-independent sampling strategies that promote spatial coverage. In particular, we propose an easily implemented method based on the pivotal sampling algorithm, which we test on problems motivated by learning-based methods for parametric PDEs and uncertainty quantification. In comparison to independent sampling, our method reduces the number of samples needed to reach a given target accuracy by up to $50\%$. We support our findings with two theoretical results. First, we show that any non-independent leverage score sampling method that obeys a weak one-sided $_$ independence condition (which includes pivotal sampling) can actively learn $d$ dimensional linear functions with $O(dd)$ samples, matching independent sampling. This result extends recent work on matrix Chernoff bounds under $_$ independence, and may be of interest for analyzing other sampling strategies beyond pivotal sampling. Second, we show that, for the important case of polynomial regression, our pivotal method obtains an improved bound on $O(d)$ samples.
2023-10-08
{'id': '2310.04966v2', 'title': 'Improved Active Learning via Dependent Leverage Score Sampling', 'abstract': 'We show how to obtain improved active learning methods in the agnostic\n(adversarial noise) setting by combining marginal leverage score sampling with\nnon-independent sampling strategies that promote spatial coverage. In\nparticular, we propose an easily implemented method based on the \\emph{pivotal\nsampling algorithm}, which we test on problems motivated by learning-based\nmethods for parametric PDEs and uncertainty quantification. In comparison to\nindependent sampling, our method reduces the number of samples needed to reach\na given target accuracy by up to $50\\%$. We support our findings with two\ntheoretical results. First, we show that any non-independent leverage score\nsampling method that obeys a weak \\emph{one-sided $\\ell_{\\infty}$ independence\ncondition} (which includes pivotal sampling) can actively learn $d$ dimensional\nlinear functions with $O(d\\log d)$ samples, matching independent sampling. This\nresult extends recent work on matrix Chernoff bounds under $\\ell_{\\infty}$\nindependence, and may be of interest for analyzing other sampling strategies\nbeyond pivotal sampling. Second, we show that, for the important case of\npolynomial regression, our pivotal method obtains an improved bound on $O(d)$\nsamples.', 'authors': ['Atsushi Shimizu', 'Xiaoou Cheng', 'Christopher Musco', 'Jonathan Weare'], 'published': '2023-10-08 01:51:30+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.04966v2'}
42,498
2306.14565v4
2,306.14565
4
Mitigating Hallucination in Large Multi-Modal Models via Robust Instruction Tuning
Despite the promising progress in multi-modal tasks, current large multi-modal models (LMMs) are prone to hallucinating inconsistent descriptions with respect to the associated image and human instructions. This paper addresses this issue by introducing the first large and diverse visual instruction tuning dataset, named Large-scale Robust Visual (LRV)-Instruction. Our dataset comprises 400k visual instructions generated by GPT4, covering 16 vision-and-language tasks with open-ended instructions and answers. Unlike existing studies that primarily focus on positive instruction samples, we design LRV-Instruction to include both positive and negative instructions for more robust visual instruction tuning. Our negative instructions are designed at three semantic levels: (i) Nonexistent Object Manipulation, (ii) Existent Object Manipulation and (iii) Knowledge Manipulation. To efficiently measure the hallucination generated by LMMs, we propose GPT4-Assisted Visual Instruction Evaluation (GAVIE), a stable approach to evaluate visual instruction tuning like human experts. GAVIE does not require human-annotated groundtruth answers and can adapt to diverse instruction formats. We conduct comprehensive experiments to investigate the hallucination of LMMs. Our results demonstrate existing LMMs exhibit significant hallucinations when presented with our negative instructions, particularly Existent Object and Knowledge Manipulation instructions. Moreover, we successfully mitigate hallucination by finetuning MiniGPT4 and mPLUG-Owl on LRV-Instruction while improving performance on several public datasets compared to state-of-the-art methods. Additionally, we observed that a balanced ratio of positive and negative instances in the training data leads to a more robust model. Code and data are available at https://github.com/FuxiaoLiu/LRV-Instruction.
2023-06-26
{'id': '2306.14565v4', 'title': 'Mitigating Hallucination in Large Multi-Modal Models via Robust Instruction Tuning', 'abstract': 'Despite the promising progress in multi-modal tasks, current large\nmulti-modal models (LMMs) are prone to hallucinating inconsistent descriptions\nwith respect to the associated image and human instructions. This paper\naddresses this issue by introducing the first large and diverse visual\ninstruction tuning dataset, named Large-scale Robust Visual (LRV)-Instruction.\nOur dataset comprises 400k visual instructions generated by GPT4, covering 16\nvision-and-language tasks with open-ended instructions and answers. Unlike\nexisting studies that primarily focus on positive instruction samples, we\ndesign LRV-Instruction to include both positive and negative instructions for\nmore robust visual instruction tuning. Our negative instructions are designed\nat three semantic levels: (i) Nonexistent Object Manipulation, (ii) Existent\nObject Manipulation and (iii) Knowledge Manipulation. To efficiently measure\nthe hallucination generated by LMMs, we propose GPT4-Assisted Visual\nInstruction Evaluation (GAVIE), a stable approach to evaluate visual\ninstruction tuning like human experts. GAVIE does not require human-annotated\ngroundtruth answers and can adapt to diverse instruction formats. We conduct\ncomprehensive experiments to investigate the hallucination of LMMs. Our results\ndemonstrate existing LMMs exhibit significant hallucinations when presented\nwith our negative instructions, particularly Existent Object and Knowledge\nManipulation instructions. Moreover, we successfully mitigate hallucination by\nfinetuning MiniGPT4 and mPLUG-Owl on LRV-Instruction while improving\nperformance on several public datasets compared to state-of-the-art methods.\nAdditionally, we observed that a balanced ratio of positive and negative\ninstances in the training data leads to a more robust model. Code and data are\navailable at https://github.com/FuxiaoLiu/LRV-Instruction.', 'authors': ['Fuxiao Liu', 'Kevin Lin', 'Linjie Li', 'Jianfeng Wang', 'Yaser Yacoob', 'Lijuan Wang'], 'published': '2023-06-26 10:26:33+00:00', 'categories': ['cs.CV', 'cs.AI', 'cs.CE', 'cs.CL', 'cs.MM'], 'url': 'http://arxiv.org/abs/2306.14565v4'}
42,499
2310.03695v1
2,310.03695
1
Multimarginal generative modeling with stochastic interpolants
Given a set of $K$ probability densities, we consider the multimarginal generative modeling problem of learning a joint distribution that recovers these densities as marginals. The structure of this joint distribution should identify multi-way correspondences among the prescribed marginals. We formalize an approach to this task within a generalization of the stochastic interpolant framework, leading to efficient learning algorithms built upon dynamical transport of measure. Our generative models are defined by velocity and score fields that can be characterized as the minimizers of simple quadratic objectives, and they are defined on a simplex that generalizes the time variable in the usual dynamical transport framework. The resulting transport on the simplex is influenced by all marginals, and we show that multi-way correspondences can be extracted. The identification of such correspondences has applications to style transfer, algorithmic fairness, and data decorruption. In addition, the multimarginal perspective enables an efficient algorithm for reducing the dynamical transport cost in the ordinary two-marginal setting. We demonstrate these capacities with several numerical examples.
2023-10-05
{'id': '2310.03695v1', 'title': 'Multimarginal generative modeling with stochastic interpolants', 'abstract': 'Given a set of $K$ probability densities, we consider the multimarginal\ngenerative modeling problem of learning a joint distribution that recovers\nthese densities as marginals. The structure of this joint distribution should\nidentify multi-way correspondences among the prescribed marginals. We formalize\nan approach to this task within a generalization of the stochastic interpolant\nframework, leading to efficient learning algorithms built upon dynamical\ntransport of measure. Our generative models are defined by velocity and score\nfields that can be characterized as the minimizers of simple quadratic\nobjectives, and they are defined on a simplex that generalizes the time\nvariable in the usual dynamical transport framework. The resulting transport on\nthe simplex is influenced by all marginals, and we show that multi-way\ncorrespondences can be extracted. The identification of such correspondences\nhas applications to style transfer, algorithmic fairness, and data\ndecorruption. In addition, the multimarginal perspective enables an efficient\nalgorithm for reducing the dynamical transport cost in the ordinary\ntwo-marginal setting. We demonstrate these capacities with several numerical\nexamples.', 'authors': ['Michael S. Albergo', 'Nicholas M. Boffi', 'Michael Lindsey', 'Eric Vanden-Eijnden'], 'published': '2023-10-05 17:12:38+00:00', 'categories': ['cs.LG', 'math.PR'], 'url': 'http://arxiv.org/abs/2310.03695v1'}
42,522
2211.07245v2
2,211.07245
2
Assessing Uncertainty in Similarity Scoring: Performance & Fairness in Face Recognition
The ROC curve is the major tool for assessing not only the performance but also the fairness properties of a similarity scoring function. In order to draw reliable conclusions based on empirical ROC analysis, accurately evaluating the uncertainty level related to statistical versions of the ROC curves of interest is absolutely necessary, especially for applications with considerable societal impact such as Face Recognition. In this article, we prove asymptotic guarantees for empirical ROC curves of similarity functions as well as for by-product metrics useful to assess fairness. We also explain that, because the false acceptance/rejection rates are of the form of U-statistics in the case of similarity scoring, the naive bootstrap approach may jeopardize the assessment procedure. A dedicated recentering technique must be used instead. Beyond the theoretical analysis carried out, various experiments using real face image datasets provide strong empirical evidence of the practical relevance of the methods promoted here, when applied to several ROC-based measures such as popular fairness metrics.
2022-11-14
{'id': '2211.07245v2', 'title': 'Assessing Uncertainty in Similarity Scoring: Performance & Fairness in Face Recognition', 'abstract': 'The ROC curve is the major tool for assessing not only the performance but\nalso the fairness properties of a similarity scoring function. In order to draw\nreliable conclusions based on empirical ROC analysis, accurately evaluating the\nuncertainty level related to statistical versions of the ROC curves of interest\nis absolutely necessary, especially for applications with considerable societal\nimpact such as Face Recognition. In this article, we prove asymptotic\nguarantees for empirical ROC curves of similarity functions as well as for\nby-product metrics useful to assess fairness. We also explain that, because the\nfalse acceptance/rejection rates are of the form of U-statistics in the case of\nsimilarity scoring, the naive bootstrap approach may jeopardize the assessment\nprocedure. A dedicated recentering technique must be used instead. Beyond the\ntheoretical analysis carried out, various experiments using real face image\ndatasets provide strong empirical evidence of the practical relevance of the\nmethods promoted here, when applied to several ROC-based measures such as\npopular fairness metrics.', 'authors': ['Jean-Rémy Conti', 'Stéphan Clémençon'], 'published': '2022-11-14 10:02:16+00:00', 'categories': ['cs.CV', 'cs.AI', 'cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2211.07245v2'}
42,524
2311.00287v2
2,311.00287
2
Knowledge-Infused Prompting: Assessing and Advancing Clinical Text Data Generation with Large Language Models
Clinical natural language processing requires methods that can address domain-specific challenges, such as complex medical terminology and clinical contexts. Recently, large language models (LLMs) have shown promise in this domain. Yet, their direct deployment can lead to privacy issues and are constrained by resources. To address this challenge, we delve into synthetic clinical text generation using LLMs for clinical NLP tasks. We propose an innovative, resource-efficient approach, ClinGen, which infuses knowledge into the process. Our model involves clinical knowledge extraction and context-informed LLM prompting. Both clinical topics and writing styles are drawn from external domain-specific knowledge graphs and LLMs to guide data generation. Our extensive empirical study across 7 clinical NLP tasks and 16 datasets reveals that ClinGen consistently enhances performance across various tasks, effectively aligning the distribution of real datasets and significantly enriching the diversity of generated training instances. Our code is available at https://github.com/ritaranx/ClinGen.
2023-11-01
{'id': '2311.00287v2', 'title': 'Knowledge-Infused Prompting: Assessing and Advancing Clinical Text Data Generation with Large Language Models', 'abstract': 'Clinical natural language processing requires methods that can address\ndomain-specific challenges, such as complex medical terminology and clinical\ncontexts. Recently, large language models (LLMs) have shown promise in this\ndomain. Yet, their direct deployment can lead to privacy issues and are\nconstrained by resources. To address this challenge, we delve into synthetic\nclinical text generation using LLMs for clinical NLP tasks. We propose an\ninnovative, resource-efficient approach, ClinGen, which infuses knowledge into\nthe process. Our model involves clinical knowledge extraction and\ncontext-informed LLM prompting. Both clinical topics and writing styles are\ndrawn from external domain-specific knowledge graphs and LLMs to guide data\ngeneration. Our extensive empirical study across 7 clinical NLP tasks and 16\ndatasets reveals that ClinGen consistently enhances performance across various\ntasks, effectively aligning the distribution of real datasets and significantly\nenriching the diversity of generated training instances. Our code is available\nat \\url{https://github.com/ritaranx/ClinGen}.', 'authors': ['Ran Xu', 'Hejie Cui', 'Yue Yu', 'Xuan Kan', 'Wenqi Shi', 'Yuchen Zhuang', 'Wei Jin', 'Joyce Ho', 'Carl Yang'], 'published': '2023-11-01 04:37:28+00:00', 'categories': ['cs.CL', 'cs.AI', 'cs.LG', 'q-bio.QM'], 'url': 'http://arxiv.org/abs/2311.00287v2'}
42,525
2310.11971v3
2,310.11971
3
Improving Generalization of Alignment with Human Preferences through Group Invariant Learning
The success of AI assistants based on language models (LLMs) hinges crucially on Reinforcement Learning from Human Feedback (RLHF), which enables the generation of responses more aligned with human preferences. As universal AI assistants, there's a growing expectation for them to perform consistently across various domains. However, previous work shows that Reinforcement Learning (RL) often exploits shortcuts to attain high rewards and overlooks challenging samples. This focus on quick reward gains undermines both the stability in training and the model's ability to generalize to new, unseen data. In this work, we propose a novel approach that can learn a consistent policy via RL across various data groups or domains. Given the challenges associated with acquiring group annotations, our method automatically classifies data into different groups, deliberately maximizing performance variance. Then, we optimize the policy to perform well on challenging groups. Lastly, leveraging the established groups, our approach adaptively adjusts the exploration space, allocating more learning capacity to more challenging data and preventing the model from over-optimizing on simpler data. Experimental results indicate that our approach significantly enhances training stability and model generalization.
2023-10-18
{'id': '2310.11971v3', 'title': 'Improving Generalization of Alignment with Human Preferences through Group Invariant Learning', 'abstract': "The success of AI assistants based on language models (LLMs) hinges crucially\non Reinforcement Learning from Human Feedback (RLHF), which enables the\ngeneration of responses more aligned with human preferences. As universal AI\nassistants, there's a growing expectation for them to perform consistently\nacross various domains. However, previous work shows that Reinforcement\nLearning (RL) often exploits shortcuts to attain high rewards and overlooks\nchallenging samples. This focus on quick reward gains undermines both the\nstability in training and the model's ability to generalize to new, unseen\ndata. In this work, we propose a novel approach that can learn a consistent\npolicy via RL across various data groups or domains. Given the challenges\nassociated with acquiring group annotations, our method automatically\nclassifies data into different groups, deliberately maximizing performance\nvariance. Then, we optimize the policy to perform well on challenging groups.\nLastly, leveraging the established groups, our approach adaptively adjusts the\nexploration space, allocating more learning capacity to more challenging data\nand preventing the model from over-optimizing on simpler data. Experimental\nresults indicate that our approach significantly enhances training stability\nand model generalization.", 'authors': ['Rui Zheng', 'Wei Shen', 'Yuan Hua', 'Wenbin Lai', 'Shihan Dou', 'Yuhao Zhou', 'Zhiheng Xi', 'Xiao Wang', 'Haoran Huang', 'Tao Gui', 'Qi Zhang', 'Xuanjing Huang'], 'published': '2023-10-18 13:54:15+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2310.11971v3'}
42,527
2310.01380v2
2,310.0138
2
Pessimistic Nonlinear Least-Squares Value Iteration for Offline Reinforcement Learning
Offline reinforcement learning (RL), where the agent aims to learn the optimal policy based on the data collected by a behavior policy, has attracted increasing attention in recent years. While offline RL with linear function approximation has been extensively studied with optimal results achieved under certain assumptions, many works shift their interest to offline RL with non-linear function approximation. However, limited works on offline RL with non-linear function approximation have instance-dependent regret guarantees. In this paper, we propose an oracle-efficient algorithm, dubbed Pessimistic Nonlinear Least-Square Value Iteration (PNLSVI), for offline RL with non-linear function approximation. Our algorithmic design comprises three innovative components: (1) a variance-based weighted regression scheme that can be applied to a wide range of function classes, (2) a subroutine for variance estimation, and (3) a planning phase that utilizes a pessimistic value iteration approach. Our algorithm enjoys a regret bound that has a tight dependency on the function class complexity and achieves minimax optimal instance-dependent regret when specialized to linear function approximation. Our work extends the previous instance-dependent results within simpler function classes, such as linear and differentiable function to a more general framework.
2023-10-02
{'id': '2310.01380v2', 'title': 'Pessimistic Nonlinear Least-Squares Value Iteration for Offline Reinforcement Learning', 'abstract': 'Offline reinforcement learning (RL), where the agent aims to learn the\noptimal policy based on the data collected by a behavior policy, has attracted\nincreasing attention in recent years. While offline RL with linear function\napproximation has been extensively studied with optimal results achieved under\ncertain assumptions, many works shift their interest to offline RL with\nnon-linear function approximation. However, limited works on offline RL with\nnon-linear function approximation have instance-dependent regret guarantees. In\nthis paper, we propose an oracle-efficient algorithm, dubbed Pessimistic\nNonlinear Least-Square Value Iteration (PNLSVI), for offline RL with non-linear\nfunction approximation. Our algorithmic design comprises three innovative\ncomponents: (1) a variance-based weighted regression scheme that can be applied\nto a wide range of function classes, (2) a subroutine for variance estimation,\nand (3) a planning phase that utilizes a pessimistic value iteration approach.\nOur algorithm enjoys a regret bound that has a tight dependency on the function\nclass complexity and achieves minimax optimal instance-dependent regret when\nspecialized to linear function approximation. Our work extends the previous\ninstance-dependent results within simpler function classes, such as linear and\ndifferentiable function to a more general framework.', 'authors': ['Qiwei Di', 'Heyang Zhao', 'Jiafan He', 'Quanquan Gu'], 'published': '2023-10-02 17:42:01+00:00', 'categories': ['cs.LG', 'math.OC', 'stat.ML'], 'url': 'http://arxiv.org/abs/2310.01380v2'}
42,502
2302.10796v2
2,302.10796
2
Provably Efficient Exploration in Quantum Reinforcement Learning with Logarithmic Worst-Case Regret
While quantum reinforcement learning (RL) has attracted a surge of attention recently, its theoretical understanding is limited. In particular, it remains elusive how to design provably efficient quantum RL algorithms that can address the exploration-exploitation trade-off. To this end, we propose a novel UCRL-style algorithm that takes advantage of quantum computing for tabular Markov decision processes (MDPs) with $S$ states, $A$ actions, and horizon $H$, and establish an $O(poly(S, A, H, T))$ worst-case regret for it, where $T$ is the number of episodes. Furthermore, we extend our results to quantum RL with linear function approximation, which is capable of handling problems with large state spaces. Specifically, we develop a quantum algorithm based on value target regression (VTR) for linear mixture MDPs with $d$-dimensional linear representation and prove that it enjoys $O(poly(d, H, T))$ regret. Our algorithms are variants of UCRL/UCRL-VTR algorithms in classical RL, which also leverage a novel combination of lazy updating mechanisms and quantum estimation subroutines. This is the key to breaking the $(T)$-regret barrier in classical RL. To the best of our knowledge, this is the first work studying the online exploration in quantum RL with provable logarithmic worst-case regret.
2023-02-21
{'id': '2302.10796v2', 'title': 'Provably Efficient Exploration in Quantum Reinforcement Learning with Logarithmic Worst-Case Regret', 'abstract': 'While quantum reinforcement learning (RL) has attracted a surge of attention\nrecently, its theoretical understanding is limited. In particular, it remains\nelusive how to design provably efficient quantum RL algorithms that can address\nthe exploration-exploitation trade-off. To this end, we propose a novel\nUCRL-style algorithm that takes advantage of quantum computing for tabular\nMarkov decision processes (MDPs) with $S$ states, $A$ actions, and horizon $H$,\nand establish an $\\mathcal{O}(\\mathrm{poly}(S, A, H, \\log T))$ worst-case\nregret for it, where $T$ is the number of episodes. Furthermore, we extend our\nresults to quantum RL with linear function approximation, which is capable of\nhandling problems with large state spaces. Specifically, we develop a quantum\nalgorithm based on value target regression (VTR) for linear mixture MDPs with\n$d$-dimensional linear representation and prove that it enjoys\n$\\mathcal{O}(\\mathrm{poly}(d, H, \\log T))$ regret. Our algorithms are variants\nof UCRL/UCRL-VTR algorithms in classical RL, which also leverage a novel\ncombination of lazy updating mechanisms and quantum estimation subroutines.\nThis is the key to breaking the $\\Omega(\\sqrt{T})$-regret barrier in classical\nRL. To the best of our knowledge, this is the first work studying the online\nexploration in quantum RL with provable logarithmic worst-case regret.', 'authors': ['Han Zhong', 'Jiachen Hu', 'Yecheng Xue', 'Tongyang Li', 'Liwei Wang'], 'published': '2023-02-21 16:23:11+00:00', 'categories': ['quant-ph', 'cs.AI', 'cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2302.10796v2'}
42,507
2402.11928v1
2,402.11928
1
Separating common from salient patterns with Contrastive Representation Learning
Contrastive Analysis is a sub-field of Representation Learning that aims at separating common factors of variation between two datasets, a background (i.e., healthy subjects) and a target (i.e., diseased subjects), from the salient factors of variation, only present in the target dataset. Despite their relevance, current models based on Variational Auto-Encoders have shown poor performance in learning semantically-expressive representations. On the other hand, Contrastive Representation Learning has shown tremendous performance leaps in various applications (classification, clustering, etc.). In this work, we propose to leverage the ability of Contrastive Learning to learn semantically expressive representations well adapted for Contrastive Analysis. We reformulate it under the lens of the InfoMax Principle and identify two Mutual Information terms to maximize and one to minimize. We decompose the first two terms into an Alignment and a Uniformity term, as commonly done in Contrastive Learning. Then, we motivate a novel Mutual Information minimization strategy to prevent information leakage between common and salient distributions. We validate our method, called SepCLR, on three visual datasets and three medical datasets, specifically conceived to assess the pattern separation capability in Contrastive Analysis. Code available at https://github.com/neurospin-projects/2024_rlouiset_sep_clr.
2024-02-19
{'id': '2402.11928v1', 'title': 'Separating common from salient patterns with Contrastive Representation Learning', 'abstract': 'Contrastive Analysis is a sub-field of Representation Learning that aims at\nseparating common factors of variation between two datasets, a background\n(i.e., healthy subjects) and a target (i.e., diseased subjects), from the\nsalient factors of variation, only present in the target dataset. Despite their\nrelevance, current models based on Variational Auto-Encoders have shown poor\nperformance in learning semantically-expressive representations. On the other\nhand, Contrastive Representation Learning has shown tremendous performance\nleaps in various applications (classification, clustering, etc.). In this work,\nwe propose to leverage the ability of Contrastive Learning to learn\nsemantically expressive representations well adapted for Contrastive Analysis.\nWe reformulate it under the lens of the InfoMax Principle and identify two\nMutual Information terms to maximize and one to minimize. We decompose the\nfirst two terms into an Alignment and a Uniformity term, as commonly done in\nContrastive Learning. Then, we motivate a novel Mutual Information minimization\nstrategy to prevent information leakage between common and salient\ndistributions. We validate our method, called SepCLR, on three visual datasets\nand three medical datasets, specifically conceived to assess the pattern\nseparation capability in Contrastive Analysis. Code available at\nhttps://github.com/neurospin-projects/2024_rlouiset_sep_clr.', 'authors': ['Robin Louiset', 'Edouard Duchesnay', 'Antoine Grigis', 'Pietro Gori'], 'published': '2024-02-19 08:17:13+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2402.11928v1'}
42,510
2402.14505v3
2,402.14505
3
Towards Seamless Adaptation of Pre-trained Models for Visual Place Recognition
Recent studies show that vision models pre-trained in generic visual learning tasks with large-scale data can provide useful feature representations for a wide range of visual perception problems. However, few attempts have been made to exploit pre-trained foundation models in visual place recognition (VPR). Due to the inherent difference in training objectives and data between the tasks of model pre-training and VPR, how to bridge the gap and fully unleash the capability of pre-trained models for VPR is still a key issue to address. To this end, we propose a novel method to realize seamless adaptation of pre-trained models for VPR. Specifically, to obtain both global and local features that focus on salient landmarks for discriminating places, we design a hybrid adaptation method to achieve both global and local adaptation efficiently, in which only lightweight adapters are tuned without adjusting the pre-trained model. Besides, to guide effective adaptation, we propose a mutual nearest neighbor local feature loss, which ensures proper dense local features are produced for local matching and avoids time-consuming spatial verification in re-ranking. Experimental results show that our method outperforms the state-of-the-art methods with less training data and training time, and uses about only 3% retrieval runtime of the two-stage VPR methods with RANSAC-based spatial verification. It ranks 1st on the MSLS challenge leaderboard (at the time of submission). The code is released at https://github.com/Lu-Feng/SelaVPR.
2024-02-22
{'id': '2402.14505v3', 'title': 'Towards Seamless Adaptation of Pre-trained Models for Visual Place Recognition', 'abstract': 'Recent studies show that vision models pre-trained in generic visual learning\ntasks with large-scale data can provide useful feature representations for a\nwide range of visual perception problems. However, few attempts have been made\nto exploit pre-trained foundation models in visual place recognition (VPR). Due\nto the inherent difference in training objectives and data between the tasks of\nmodel pre-training and VPR, how to bridge the gap and fully unleash the\ncapability of pre-trained models for VPR is still a key issue to address. To\nthis end, we propose a novel method to realize seamless adaptation of\npre-trained models for VPR. Specifically, to obtain both global and local\nfeatures that focus on salient landmarks for discriminating places, we design a\nhybrid adaptation method to achieve both global and local adaptation\nefficiently, in which only lightweight adapters are tuned without adjusting the\npre-trained model. Besides, to guide effective adaptation, we propose a mutual\nnearest neighbor local feature loss, which ensures proper dense local features\nare produced for local matching and avoids time-consuming spatial verification\nin re-ranking. Experimental results show that our method outperforms the\nstate-of-the-art methods with less training data and training time, and uses\nabout only 3% retrieval runtime of the two-stage VPR methods with RANSAC-based\nspatial verification. It ranks 1st on the MSLS challenge leaderboard (at the\ntime of submission). The code is released at\nhttps://github.com/Lu-Feng/SelaVPR.', 'authors': ['Feng Lu', 'Lijun Zhang', 'Xiangyuan Lan', 'Shuting Dong', 'Yaowei Wang', 'Chun Yuan'], 'published': '2024-02-22 12:55:01+00:00', 'categories': ['cs.CV', 'cs.AI'], 'url': 'http://arxiv.org/abs/2402.14505v3'}
42,517
2309.12252v3
2,309.12252
3
Parallelizing non-linear sequential models over the sequence length
Sequential models, such as Recurrent Neural Networks and Neural Ordinary Differential Equations, have long suffered from slow training due to their inherent sequential nature. For many years this bottleneck has persisted, as many thought sequential models could not be parallelized. We challenge this long-held belief with our parallel algorithm that accelerates GPU evaluation of sequential models by up to 3 orders of magnitude faster without compromising output accuracy. The algorithm does not need any special structure in the sequential models' architecture, making it applicable to a wide range of architectures. Using our method, training sequential models can be more than 10 times faster than the common sequential method without any meaningful difference in the training results. Leveraging this accelerated training, we discovered the efficacy of the Gated Recurrent Unit in a long time series classification problem with 17k time samples. By overcoming the training bottleneck, our work serves as the first step to unlock the potential of non-linear sequential models for long sequence problems.
2023-09-21
{'id': '2309.12252v3', 'title': 'Parallelizing non-linear sequential models over the sequence length', 'abstract': "Sequential models, such as Recurrent Neural Networks and Neural Ordinary\nDifferential Equations, have long suffered from slow training due to their\ninherent sequential nature. For many years this bottleneck has persisted, as\nmany thought sequential models could not be parallelized. We challenge this\nlong-held belief with our parallel algorithm that accelerates GPU evaluation of\nsequential models by up to 3 orders of magnitude faster without compromising\noutput accuracy. The algorithm does not need any special structure in the\nsequential models' architecture, making it applicable to a wide range of\narchitectures. Using our method, training sequential models can be more than 10\ntimes faster than the common sequential method without any meaningful\ndifference in the training results. Leveraging this accelerated training, we\ndiscovered the efficacy of the Gated Recurrent Unit in a long time series\nclassification problem with 17k time samples. By overcoming the training\nbottleneck, our work serves as the first step to unlock the potential of\nnon-linear sequential models for long sequence problems.", 'authors': ['Yi Heng Lim', 'Qi Zhu', 'Joshua Selfridge', 'Muhammad Firmansyah Kasim'], 'published': '2023-09-21 16:52:34+00:00', 'categories': ['cs.LG', 'cs.DC', 'physics.comp-ph'], 'url': 'http://arxiv.org/abs/2309.12252v3'}
42,519
2310.19698v2
2,310.19698
2
When Do Prompting and Prefix-Tuning Work? A Theory of Capabilities and Limitations
Context-based fine-tuning methods, including prompting, in-context learning, soft prompting (also known as prompt tuning), and prefix-tuning, have gained popularity due to their ability to often match the performance of full fine-tuning with a fraction of the parameters. Despite their empirical successes, there is little theoretical understanding of how these techniques influence the internal computation of the model and their expressiveness limitations. We show that despite the continuous embedding space being more expressive than the discrete token space, soft-prompting and prefix-tuning are potentially less expressive than full fine-tuning, even with the same number of learnable parameters. Concretely, context-based fine-tuning cannot change the relative attention pattern over the content and can only bias the outputs of an attention layer in a fixed direction. This suggests that while techniques like prompting, in-context learning, soft prompting, and prefix-tuning can effectively elicit skills present in the pretrained model, they may not be able to learn novel tasks that require new attention patterns.
2023-10-30
{'id': '2310.19698v2', 'title': 'When Do Prompting and Prefix-Tuning Work? A Theory of Capabilities and Limitations', 'abstract': 'Context-based fine-tuning methods, including prompting, in-context learning,\nsoft prompting (also known as prompt tuning), and prefix-tuning, have gained\npopularity due to their ability to often match the performance of full\nfine-tuning with a fraction of the parameters. Despite their empirical\nsuccesses, there is little theoretical understanding of how these techniques\ninfluence the internal computation of the model and their expressiveness\nlimitations. We show that despite the continuous embedding space being more\nexpressive than the discrete token space, soft-prompting and prefix-tuning are\npotentially less expressive than full fine-tuning, even with the same number of\nlearnable parameters. Concretely, context-based fine-tuning cannot change the\nrelative attention pattern over the content and can only bias the outputs of an\nattention layer in a fixed direction. This suggests that while techniques like\nprompting, in-context learning, soft prompting, and prefix-tuning can\neffectively elicit skills present in the pretrained model, they may not be able\nto learn novel tasks that require new attention patterns.', 'authors': ['Aleksandar Petrov', 'Philip H. S. Torr', 'Adel Bibi'], 'published': '2023-10-30 16:19:34+00:00', 'categories': ['cs.LG', 'cs.CL'], 'url': 'http://arxiv.org/abs/2310.19698v2'}
42,533
2311.14904v1
2,311.14904
1
LLM-Assisted Code Cleaning For Training Accurate Code Generators
Natural language to code generation is an important application area of LLMs and has received wide attention from the community. The majority of relevant studies have exclusively concentrated on increasing the quantity and functional correctness of training sets while disregarding other stylistic elements of programs. More recently, data quality has garnered a lot of interest and multiple works have showcased its importance for improving performance. In this work, we investigate data quality for code and find that making the code more structured and readable leads to improved code generation performance of the system. We build a novel data-cleaning pipeline that uses these principles to transform existing programs by 1.) renaming variables, 2.) modularizing and decomposing complex code into smaller helper sub-functions, and 3.) inserting natural-language based plans via LLM based transformations. We evaluate our approach on two challenging algorithmic code generation benchmarks and find that fine-tuning CodeLLaMa-7B on our transformed modularized programs improves the performance by up to 30% compared to fine-tuning on the original dataset. Additionally, we demonstrate improved performance from using a smaller amount of higher-quality data, finding that a model fine-tuned on the entire original dataset is outperformed by a model trained on 15% of our cleaned dataset. Even in comparison to closed-source models, our models outperform the much larger AlphaCoder models.
2023-11-25
{'id': '2311.14904v1', 'title': 'LLM-Assisted Code Cleaning For Training Accurate Code Generators', 'abstract': 'Natural language to code generation is an important application area of LLMs\nand has received wide attention from the community. The majority of relevant\nstudies have exclusively concentrated on increasing the quantity and functional\ncorrectness of training sets while disregarding other stylistic elements of\nprograms. More recently, data quality has garnered a lot of interest and\nmultiple works have showcased its importance for improving performance. In this\nwork, we investigate data quality for code and find that making the code more\nstructured and readable leads to improved code generation performance of the\nsystem. We build a novel data-cleaning pipeline that uses these principles to\ntransform existing programs by 1.) renaming variables, 2.) modularizing and\ndecomposing complex code into smaller helper sub-functions, and 3.) inserting\nnatural-language based plans via LLM based transformations. We evaluate our\napproach on two challenging algorithmic code generation benchmarks and find\nthat fine-tuning CodeLLaMa-7B on our transformed modularized programs improves\nthe performance by up to 30% compared to fine-tuning on the original dataset.\nAdditionally, we demonstrate improved performance from using a smaller amount\nof higher-quality data, finding that a model fine-tuned on the entire original\ndataset is outperformed by a model trained on 15% of our cleaned dataset. Even\nin comparison to closed-source models, our models outperform the much larger\nAlphaCoder models.', 'authors': ['Naman Jain', 'Tianjun Zhang', 'Wei-Lin Chiang', 'Joseph E. Gonzalez', 'Koushik Sen', 'Ion Stoica'], 'published': '2023-11-25 02:45:50+00:00', 'categories': ['cs.LG', 'cs.SE'], 'url': 'http://arxiv.org/abs/2311.14904v1'}
42,504
2110.03372v2
2,110.03372
2
Unifying Likelihood-free Inference with Black-box Optimization and Beyond
Black-box optimization formulations for biological sequence design have drawn recent attention due to their promising potential impact on the pharmaceutical industry. In this work, we propose to unify two seemingly distinct worlds: likelihood-free inference and black-box optimization, under one probabilistic framework. In tandem, we provide a recipe for constructing various sequence design methods based on this framework. We show how previous optimization approaches can be "reinvented" in our framework, and further propose new probabilistic black-box optimization algorithms. Extensive experiments on sequence design application illustrate the benefits of the proposed methodology.
2021-10-06
{'id': '2110.03372v2', 'title': 'Unifying Likelihood-free Inference with Black-box Optimization and Beyond', 'abstract': 'Black-box optimization formulations for biological sequence design have drawn\nrecent attention due to their promising potential impact on the pharmaceutical\nindustry. In this work, we propose to unify two seemingly distinct worlds:\nlikelihood-free inference and black-box optimization, under one probabilistic\nframework. In tandem, we provide a recipe for constructing various sequence\ndesign methods based on this framework. We show how previous optimization\napproaches can be "reinvented" in our framework, and further propose new\nprobabilistic black-box optimization algorithms. Extensive experiments on\nsequence design application illustrate the benefits of the proposed\nmethodology.', 'authors': ['Dinghuai Zhang', 'Jie Fu', 'Yoshua Bengio', 'Aaron Courville'], 'published': '2021-10-06 02:41:50+00:00', 'categories': ['cs.LG', 'cs.AI', 'q-bio.BM', 'stat.ME', 'stat.ML'], 'url': 'http://arxiv.org/abs/2110.03372v2'}
42,563
2306.17499v1
2,306.17499
1
The Implicit Bias of Minima Stability in Multivariate Shallow ReLU Networks
We study the type of solutions to which stochastic gradient descent converges when used to train a single hidden-layer multivariate ReLU network with the quadratic loss. Our results are based on a dynamical stability analysis. In the univariate case, it was shown that linearly stable minima correspond to network functions (predictors), whose second derivative has a bounded weighted $L^1$ norm. Notably, the bound gets smaller as the step size increases, implying that training with a large step size leads to `smoother' predictors. Here we generalize this result to the multivariate case, showing that a similar result applies to the Laplacian of the predictor. We demonstrate the tightness of our bound on the MNIST dataset, and show that it accurately captures the behavior of the solutions as a function of the step size. Additionally, we prove a depth separation result on the approximation power of ReLU networks corresponding to stable minima of the loss. Specifically, although shallow ReLU networks are universal approximators, we prove that stable shallow networks are not. Namely, there is a function that cannot be well-approximated by stable single hidden-layer ReLU networks trained with a non-vanishing step size. This is while the same function can be realized as a stable two hidden-layer ReLU network. Finally, we prove that if a function is sufficiently smooth (in a Sobolev sense) then it can be approximated arbitrarily well using shallow ReLU networks that correspond to stable solutions of gradient descent.
2023-06-30
{'id': '2306.17499v1', 'title': 'The Implicit Bias of Minima Stability in Multivariate Shallow ReLU Networks', 'abstract': "We study the type of solutions to which stochastic gradient descent converges\nwhen used to train a single hidden-layer multivariate ReLU network with the\nquadratic loss. Our results are based on a dynamical stability analysis. In the\nunivariate case, it was shown that linearly stable minima correspond to network\nfunctions (predictors), whose second derivative has a bounded weighted $L^1$\nnorm. Notably, the bound gets smaller as the step size increases, implying that\ntraining with a large step size leads to `smoother' predictors. Here we\ngeneralize this result to the multivariate case, showing that a similar result\napplies to the Laplacian of the predictor. We demonstrate the tightness of our\nbound on the MNIST dataset, and show that it accurately captures the behavior\nof the solutions as a function of the step size. Additionally, we prove a depth\nseparation result on the approximation power of ReLU networks corresponding to\nstable minima of the loss. Specifically, although shallow ReLU networks are\nuniversal approximators, we prove that stable shallow networks are not. Namely,\nthere is a function that cannot be well-approximated by stable single\nhidden-layer ReLU networks trained with a non-vanishing step size. This is\nwhile the same function can be realized as a stable two hidden-layer ReLU\nnetwork. Finally, we prove that if a function is sufficiently smooth (in a\nSobolev sense) then it can be approximated arbitrarily well using shallow ReLU\nnetworks that correspond to stable solutions of gradient descent.", 'authors': ['Mor Shpigel Nacson', 'Rotem Mulayoff', 'Greg Ongie', 'Tomer Michaeli', 'Daniel Soudry'], 'published': '2023-06-30 09:17:39+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2306.17499v1'}
42,505
2003.02234v1
2,003.02234
1
Contrastive estimation reveals topic posterior information to linear models
Contrastive learning is an approach to representation learning that utilizes naturally occurring similar and dissimilar pairs of data points to find useful embeddings of data. In the context of document classification under topic modeling assumptions, we prove that contrastive learning is capable of recovering a representation of documents that reveals their underlying topic posterior information to linear models. We apply this procedure in a semi-supervised setup and demonstrate empirically that linear classifiers with these representations perform well in document classification tasks with very few training examples.
2020-03-04
{'id': '2003.02234v1', 'title': 'Contrastive estimation reveals topic posterior information to linear models', 'abstract': 'Contrastive learning is an approach to representation learning that utilizes\nnaturally occurring similar and dissimilar pairs of data points to find useful\nembeddings of data. In the context of document classification under topic\nmodeling assumptions, we prove that contrastive learning is capable of\nrecovering a representation of documents that reveals their underlying topic\nposterior information to linear models. We apply this procedure in a\nsemi-supervised setup and demonstrate empirically that linear classifiers with\nthese representations perform well in document classification tasks with very\nfew training examples.', 'authors': ['Christopher Tosh', 'Akshay Krishnamurthy', 'Daniel Hsu'], 'published': '2020-03-04 18:20:55+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2003.02234v1'}
42,521
2010.04495v1
2,010.04495
1
Linear Mode Connectivity in Multitask and Continual Learning
Continual (sequential) training and multitask (simultaneous) training are often attempting to solve the same overall objective: to find a solution that performs well on all considered tasks. The main difference is in the training regimes, where continual learning can only have access to one task at a time, which for neural networks typically leads to catastrophic forgetting. That is, the solution found for a subsequent task does not perform well on the previous ones anymore. However, the relationship between the different minima that the two training regimes arrive at is not well understood. What sets them apart? Is there a local structure that could explain the difference in performance achieved by the two different schemes? Motivated by recent work showing that different minima of the same task are typically connected by very simple curves of low error, we investigate whether multitask and continual solutions are similarly connected. We empirically find that indeed such connectivity can be reliably achieved and, more interestingly, it can be done by a linear path, conditioned on having the same initialization for both. We thoroughly analyze this observation and discuss its significance for the continual learning process. Furthermore, we exploit this finding to propose an effective algorithm that constrains the sequentially learned minima to behave as the multitask solution. We show that our method outperforms several state of the art continual learning algorithms on various vision benchmarks.
2020-10-09
{'id': '2010.04495v1', 'title': 'Linear Mode Connectivity in Multitask and Continual Learning', 'abstract': 'Continual (sequential) training and multitask (simultaneous) training are\noften attempting to solve the same overall objective: to find a solution that\nperforms well on all considered tasks. The main difference is in the training\nregimes, where continual learning can only have access to one task at a time,\nwhich for neural networks typically leads to catastrophic forgetting. That is,\nthe solution found for a subsequent task does not perform well on the previous\nones anymore. However, the relationship between the different minima that the\ntwo training regimes arrive at is not well understood. What sets them apart? Is\nthere a local structure that could explain the difference in performance\nachieved by the two different schemes? Motivated by recent work showing that\ndifferent minima of the same task are typically connected by very simple curves\nof low error, we investigate whether multitask and continual solutions are\nsimilarly connected. We empirically find that indeed such connectivity can be\nreliably achieved and, more interestingly, it can be done by a linear path,\nconditioned on having the same initialization for both. We thoroughly analyze\nthis observation and discuss its significance for the continual learning\nprocess. Furthermore, we exploit this finding to propose an effective algorithm\nthat constrains the sequentially learned minima to behave as the multitask\nsolution. We show that our method outperforms several state of the art\ncontinual learning algorithms on various vision benchmarks.', 'authors': ['Seyed Iman Mirzadeh', 'Mehrdad Farajtabar', 'Dilan Gorur', 'Razvan Pascanu', 'Hassan Ghasemzadeh'], 'published': '2020-10-09 10:53:25+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CV'], 'url': 'http://arxiv.org/abs/2010.04495v1'}
42,557
2102.11535v4
2,102.11535
4
Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective
Neural Architecture Search (NAS) has been explosively studied to automate the discovery of top-performer neural networks. Current works require heavy training of supernet or intensive architecture evaluations, thus suffering from heavy resource consumption and often incurring search bias due to truncated training or approximations. Can we select the best neural architectures without involving any training and eliminate a drastic portion of the search cost? We provide an affirmative answer, by proposing a novel framework called training-free neural architecture search (TE-NAS). TE-NAS ranks architectures by analyzing the spectrum of the neural tangent kernel (NTK) and the number of linear regions in the input space. Both are motivated by recent theory advances in deep networks and can be computed without any training and any label. We show that: (1) these two measurements imply the trainability and expressivity of a neural network; (2) they strongly correlate with the network's test accuracy. Further on, we design a pruning-based NAS mechanism to achieve a more flexible and superior trade-off between the trainability and expressivity during the search. In NAS-Bench-201 and DARTS search spaces, TE-NAS completes high-quality search but only costs 0.5 and 4 GPU hours with one 1080Ti on CIFAR-10 and ImageNet, respectively. We hope our work inspires more attempts in bridging the theoretical findings of deep networks and practical impacts in real NAS applications. Code is available at: https://github.com/VITA-Group/TENAS.
2021-02-23
{'id': '2102.11535v4', 'title': 'Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective', 'abstract': "Neural Architecture Search (NAS) has been explosively studied to automate the\ndiscovery of top-performer neural networks. Current works require heavy\ntraining of supernet or intensive architecture evaluations, thus suffering from\nheavy resource consumption and often incurring search bias due to truncated\ntraining or approximations. Can we select the best neural architectures without\ninvolving any training and eliminate a drastic portion of the search cost? We\nprovide an affirmative answer, by proposing a novel framework called\ntraining-free neural architecture search (TE-NAS). TE-NAS ranks architectures\nby analyzing the spectrum of the neural tangent kernel (NTK) and the number of\nlinear regions in the input space. Both are motivated by recent theory advances\nin deep networks and can be computed without any training and any label. We\nshow that: (1) these two measurements imply the trainability and expressivity\nof a neural network; (2) they strongly correlate with the network's test\naccuracy. Further on, we design a pruning-based NAS mechanism to achieve a more\nflexible and superior trade-off between the trainability and expressivity\nduring the search. In NAS-Bench-201 and DARTS search spaces, TE-NAS completes\nhigh-quality search but only costs 0.5 and 4 GPU hours with one 1080Ti on\nCIFAR-10 and ImageNet, respectively. We hope our work inspires more attempts in\nbridging the theoretical findings of deep networks and practical impacts in\nreal NAS applications. Code is available at:\nhttps://github.com/VITA-Group/TENAS.", 'authors': ['Wuyang Chen', 'Xinyu Gong', 'Zhangyang Wang'], 'published': '2021-02-23 07:50:44+00:00', 'categories': ['cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2102.11535v4'}
42,558
2105.03193v1
2,105.03193
1
Network Pruning That Matters: A Case Study on Retraining Variants
Network pruning is an effective method to reduce the computational expense of over-parameterized neural networks for deployment on low-resource systems. Recent state-of-the-art techniques for retraining pruned networks such as weight rewinding and learning rate rewinding have been shown to outperform the traditional fine-tuning technique in recovering the lost accuracy (Renda et al., 2020), but so far it is unclear what accounts for such performance. In this work, we conduct extensive experiments to verify and analyze the uncanny effectiveness of learning rate rewinding. We find that the reason behind the success of learning rate rewinding is the usage of a large learning rate. Similar phenomenon can be observed in other learning rate schedules that involve large learning rates, e.g., the 1-cycle learning rate schedule (Smith et al., 2019). By leveraging the right learning rate schedule in retraining, we demonstrate a counter-intuitive phenomenon in that randomly pruned networks could even achieve better performance than methodically pruned networks (fine-tuned with the conventional approach). Our results emphasize the cruciality of the learning rate schedule in pruned network retraining - a detail often overlooked by practitioners during the implementation of network pruning. One-sentence Summary: We study the effective of different retraining mechanisms while doing pruning
2021-05-07
{'id': '2105.03193v1', 'title': 'Network Pruning That Matters: A Case Study on Retraining Variants', 'abstract': 'Network pruning is an effective method to reduce the computational expense of\nover-parameterized neural networks for deployment on low-resource systems.\nRecent state-of-the-art techniques for retraining pruned networks such as\nweight rewinding and learning rate rewinding have been shown to outperform the\ntraditional fine-tuning technique in recovering the lost accuracy (Renda et\nal., 2020), but so far it is unclear what accounts for such performance. In\nthis work, we conduct extensive experiments to verify and analyze the uncanny\neffectiveness of learning rate rewinding. We find that the reason behind the\nsuccess of learning rate rewinding is the usage of a large learning rate.\nSimilar phenomenon can be observed in other learning rate schedules that\ninvolve large learning rates, e.g., the 1-cycle learning rate schedule (Smith\net al., 2019). By leveraging the right learning rate schedule in retraining, we\ndemonstrate a counter-intuitive phenomenon in that randomly pruned networks\ncould even achieve better performance than methodically pruned networks\n(fine-tuned with the conventional approach). Our results emphasize the\ncruciality of the learning rate schedule in pruned network retraining - a\ndetail often overlooked by practitioners during the implementation of network\npruning. One-sentence Summary: We study the effective of different retraining\nmechanisms while doing pruning', 'authors': ['Duong H. Le', 'Binh-Son Hua'], 'published': '2021-05-07 12:03:24+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2105.03193v1'}
42,560
2205.15379v1
2,205.15379
1
Truly Deterministic Policy Optimization
In this paper, we present a policy gradient method that avoids exploratory noise injection and performs policy search over the deterministic landscape. By avoiding noise injection all sources of estimation variance can be eliminated in systems with deterministic dynamics (up to the initial state distribution). Since deterministic policy regularization is impossible using traditional non-metric measures such as the KL divergence, we derive a Wasserstein-based quadratic model for our purposes. We state conditions on the system model under which it is possible to establish a monotonic policy improvement guarantee, propose a surrogate function for policy gradient estimation, and show that it is possible to compute exact advantage estimates if both the state transition model and the policy are deterministic. Finally, we describe two novel robotic control environments -- one with non-local rewards in the frequency domain and the other with a long horizon (8000 time-steps) -- for which our policy gradient method (TDPO) significantly outperforms existing methods (PPO, TRPO, DDPG, and TD3). Our implementation with all the experimental settings is available at https://github.com/ehsansaleh/code_tdpo
2022-05-30
{'id': '2205.15379v1', 'title': 'Truly Deterministic Policy Optimization', 'abstract': 'In this paper, we present a policy gradient method that avoids exploratory\nnoise injection and performs policy search over the deterministic landscape. By\navoiding noise injection all sources of estimation variance can be eliminated\nin systems with deterministic dynamics (up to the initial state distribution).\nSince deterministic policy regularization is impossible using traditional\nnon-metric measures such as the KL divergence, we derive a Wasserstein-based\nquadratic model for our purposes. We state conditions on the system model under\nwhich it is possible to establish a monotonic policy improvement guarantee,\npropose a surrogate function for policy gradient estimation, and show that it\nis possible to compute exact advantage estimates if both the state transition\nmodel and the policy are deterministic. Finally, we describe two novel robotic\ncontrol environments -- one with non-local rewards in the frequency domain and\nthe other with a long horizon (8000 time-steps) -- for which our policy\ngradient method (TDPO) significantly outperforms existing methods (PPO, TRPO,\nDDPG, and TD3). Our implementation with all the experimental settings is\navailable at https://github.com/ehsansaleh/code_tdpo', 'authors': ['Ehsan Saleh', 'Saba Ghaffari', 'Timothy Bretl', 'Matthew West'], 'published': '2022-05-30 18:49:33+00:00', 'categories': ['cs.AI', 'cs.LG', 'cs.RO', 'cs.SY', 'eess.SY'], 'url': 'http://arxiv.org/abs/2205.15379v1'}
42,518
2310.18168v5
2,310.18168
5
Personas as a Way to Model Truthfulness in Language Models
Large language models (LLMs) are trained on vast amounts of text from the internet, which contains both factual and misleading information about the world. While unintuitive from a classic view of LMs, recent work has shown that the truth value of a statement can be elicited from the model's representations. This paper presents an explanation for why LMs appear to know the truth despite not being trained with truth labels. We hypothesize that the pretraining data is generated by groups of (un)truthful agents whose outputs share common features, and they form a (un)truthful persona. By training on this data, LMs can infer and represent the persona in its activation space. This allows the model to separate truth from falsehoods and controls the truthfulness of its generation. We show evidence for the persona hypothesis via two observations: (1) we can probe whether a model's answer will be truthful before it is generated; (2) finetuning a model on a set of facts improves its truthfulness on unseen topics. Next, using arithmetics as a synthetic environment, we show that structures of the pretraining data are crucial for the model to infer the truthful persona. Overall, our findings suggest that models can exploit hierarchical structures in the data to learn abstract concepts like truthfulness.
2023-10-27
{'id': '2310.18168v5', 'title': 'Personas as a Way to Model Truthfulness in Language Models', 'abstract': "Large language models (LLMs) are trained on vast amounts of text from the\ninternet, which contains both factual and misleading information about the\nworld. While unintuitive from a classic view of LMs, recent work has shown that\nthe truth value of a statement can be elicited from the model's\nrepresentations. This paper presents an explanation for why LMs appear to know\nthe truth despite not being trained with truth labels. We hypothesize that the\npretraining data is generated by groups of (un)truthful agents whose outputs\nshare common features, and they form a (un)truthful persona. By training on\nthis data, LMs can infer and represent the persona in its activation space.\nThis allows the model to separate truth from falsehoods and controls the\ntruthfulness of its generation. We show evidence for the persona hypothesis via\ntwo observations: (1) we can probe whether a model's answer will be truthful\nbefore it is generated; (2) finetuning a model on a set of facts improves its\ntruthfulness on unseen topics. Next, using arithmetics as a synthetic\nenvironment, we show that structures of the pretraining data are crucial for\nthe model to infer the truthful persona. Overall, our findings suggest that\nmodels can exploit hierarchical structures in the data to learn abstract\nconcepts like truthfulness.", 'authors': ['Nitish Joshi', 'Javier Rando', 'Abulhair Saparov', 'Najoung Kim', 'He He'], 'published': '2023-10-27 14:27:43+00:00', 'categories': ['cs.CL', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2310.18168v5'}
42,523
2305.18702v2
2,305.18702
2
Adversarial Adaptive Sampling: Unify PINN and Optimal Transport for the Approximation of PDEs
Solving partial differential equations (PDEs) is a central task in scientific computing. Recently, neural network approximation of PDEs has received increasing attention due to its flexible meshless discretization and its potential for high-dimensional problems. One fundamental numerical difficulty is that random samples in the training set introduce statistical errors into the discretization of loss functional which may become the dominant error in the final approximation, and therefore overshadow the modeling capability of the neural network. In this work, we propose a new minmax formulation to optimize simultaneously the approximate solution, given by a neural network model, and the random samples in the training set, provided by a deep generative model. The key idea is to use a deep generative model to adjust random samples in the training set such that the residual induced by the approximate PDE solution can maintain a smooth profile when it is being minimized. Such an idea is achieved by implicitly embedding the Wasserstein distance between the residual-induced distribution and the uniform distribution into the loss, which is then minimized together with the residual. A nearly uniform residual profile means that its variance is small for any normalized weight function such that the Monte Carlo approximation error of the loss functional is reduced significantly for a certain sample size. The adversarial adaptive sampling (AAS) approach proposed in this work is the first attempt to formulate two essential components, minimizing the residual and seeking the optimal training set, into one minmax objective functional for the neural network approximation of PDEs.
2023-05-30
{'id': '2305.18702v2', 'title': 'Adversarial Adaptive Sampling: Unify PINN and Optimal Transport for the Approximation of PDEs', 'abstract': 'Solving partial differential equations (PDEs) is a central task in scientific\ncomputing. Recently, neural network approximation of PDEs has received\nincreasing attention due to its flexible meshless discretization and its\npotential for high-dimensional problems. One fundamental numerical difficulty\nis that random samples in the training set introduce statistical errors into\nthe discretization of loss functional which may become the dominant error in\nthe final approximation, and therefore overshadow the modeling capability of\nthe neural network. In this work, we propose a new minmax formulation to\noptimize simultaneously the approximate solution, given by a neural network\nmodel, and the random samples in the training set, provided by a deep\ngenerative model. The key idea is to use a deep generative model to adjust\nrandom samples in the training set such that the residual induced by the\napproximate PDE solution can maintain a smooth profile when it is being\nminimized. Such an idea is achieved by implicitly embedding the Wasserstein\ndistance between the residual-induced distribution and the uniform distribution\ninto the loss, which is then minimized together with the residual. A nearly\nuniform residual profile means that its variance is small for any normalized\nweight function such that the Monte Carlo approximation error of the loss\nfunctional is reduced significantly for a certain sample size. The adversarial\nadaptive sampling (AAS) approach proposed in this work is the first attempt to\nformulate two essential components, minimizing the residual and seeking the\noptimal training set, into one minmax objective functional for the neural\nnetwork approximation of PDEs.', 'authors': ['Kejun Tang', 'Jiayu Zhai', 'Xiaoliang Wan', 'Chao Yang'], 'published': '2023-05-30 02:59:18+00:00', 'categories': ['stat.ML', 'cs.LG', 'cs.NA', 'math.NA'], 'url': 'http://arxiv.org/abs/2305.18702v2'}
42,537
2303.10137v2
2,303.10137
2
A Recipe for Watermarking Diffusion Models
Diffusion models (DMs) have demonstrated advantageous potential on generative tasks. Widespread interest exists in incorporating DMs into downstream applications, such as producing or editing photorealistic images. However, practical deployment and unprecedented power of DMs raise legal issues, including copyright protection and monitoring of generated content. In this regard, watermarking has been a proven solution for copyright protection and content monitoring, but it is underexplored in the DMs literature. Specifically, DMs generate samples from longer tracks and may have newly designed multimodal structures, necessitating the modification of conventional watermarking pipelines. To this end, we conduct comprehensive analyses and derive a recipe for efficiently watermarking state-of-the-art DMs (e.g., Stable Diffusion), via training from scratch or finetuning. Our recipe is straightforward but involves empirically ablated implementation details, providing a foundation for future research on watermarking DMs. The code is available at https://github.com/yunqing-me/WatermarkDM.
2023-03-17
{'id': '2303.10137v2', 'title': 'A Recipe for Watermarking Diffusion Models', 'abstract': 'Diffusion models (DMs) have demonstrated advantageous potential on generative\ntasks. Widespread interest exists in incorporating DMs into downstream\napplications, such as producing or editing photorealistic images. However,\npractical deployment and unprecedented power of DMs raise legal issues,\nincluding copyright protection and monitoring of generated content. In this\nregard, watermarking has been a proven solution for copyright protection and\ncontent monitoring, but it is underexplored in the DMs literature.\nSpecifically, DMs generate samples from longer tracks and may have newly\ndesigned multimodal structures, necessitating the modification of conventional\nwatermarking pipelines. To this end, we conduct comprehensive analyses and\nderive a recipe for efficiently watermarking state-of-the-art DMs (e.g., Stable\nDiffusion), via training from scratch or finetuning. Our recipe is\nstraightforward but involves empirically ablated implementation details,\nproviding a foundation for future research on watermarking DMs. The code is\navailable at https://github.com/yunqing-me/WatermarkDM.', 'authors': ['Yunqing Zhao', 'Tianyu Pang', 'Chao Du', 'Xiao Yang', 'Ngai-Man Cheung', 'Min Lin'], 'published': '2023-03-17 17:25:10+00:00', 'categories': ['cs.CV', 'cs.CR', 'cs.LG'], 'url': 'http://arxiv.org/abs/2303.10137v2'}
42,543
2306.07124v2
2,306.07124
2
Diverse Projection Ensembles for Distributional Reinforcement Learning
In contrast to classical reinforcement learning (RL), distributional RL algorithms aim to learn the distribution of returns rather than their expected value. Since the nature of the return distribution is generally unknown a priori or arbitrarily complex, a common approach finds approximations within a set of representable, parametric distributions. Typically, this involves a projection of the unconstrained distribution onto the set of simplified distributions. We argue that this projection step entails a strong inductive bias when coupled with neural networks and gradient descent, thereby profoundly impacting the generalization behavior of learned models. In order to facilitate reliable uncertainty estimation through diversity, we study the combination of several different projections and representations in a distributional ensemble. We establish theoretical properties of such projection ensembles and derive an algorithm that uses ensemble disagreement, measured by the average 1-Wasserstein distance, as a bonus for deep exploration. We evaluate our algorithm on the behavior suite benchmark and VizDoom and find that diverse projection ensembles lead to significant performance improvements over existing methods on a variety of tasks with the most pronounced gains in directed exploration problems.
2023-06-12
{'id': '2306.07124v2', 'title': 'Diverse Projection Ensembles for Distributional Reinforcement Learning', 'abstract': 'In contrast to classical reinforcement learning (RL), distributional RL\nalgorithms aim to learn the distribution of returns rather than their expected\nvalue. Since the nature of the return distribution is generally unknown a\npriori or arbitrarily complex, a common approach finds approximations within a\nset of representable, parametric distributions. Typically, this involves a\nprojection of the unconstrained distribution onto the set of simplified\ndistributions. We argue that this projection step entails a strong inductive\nbias when coupled with neural networks and gradient descent, thereby profoundly\nimpacting the generalization behavior of learned models. In order to facilitate\nreliable uncertainty estimation through diversity, we study the combination of\nseveral different projections and representations in a distributional ensemble.\nWe establish theoretical properties of such projection ensembles and derive an\nalgorithm that uses ensemble disagreement, measured by the average\n1-Wasserstein distance, as a bonus for deep exploration. We evaluate our\nalgorithm on the behavior suite benchmark and VizDoom and find that diverse\nprojection ensembles lead to significant performance improvements over existing\nmethods on a variety of tasks with the most pronounced gains in directed\nexploration problems.', 'authors': ['Moritz A. Zanger', 'Wendelin Böhmer', 'Matthijs T. J. Spaan'], 'published': '2023-06-12 13:59:48+00:00', 'categories': ['cs.LG', 'cs.AI', 'stat.ML'], 'url': 'http://arxiv.org/abs/2306.07124v2'}
42,554
2403.17921v1
2,403.17921
1
The Need for Speed: Pruning Transformers with One Recipe
We introduce the $O$ne-shot $P$runing $T$echnique for $I$nterchangeable $N$etworks ($OPTIN$) framework as a tool to increase the efficiency of pre-trained transformer architectures $without requiring re-training$. Recent works have explored improving transformer efficiency, however often incur computationally expensive re-training procedures or depend on architecture-specific characteristics, thus impeding practical wide-scale adoption. To address these shortcomings, the OPTIN framework leverages intermediate feature distillation, capturing the long-range dependencies of model parameters (coined $trajectory$), to produce state-of-the-art results on natural language, image classification, transfer learning, and semantic segmentation tasks $without re-training$. Given a FLOP constraint, the OPTIN framework will compress the network while maintaining competitive accuracy performance and improved throughput. Particularly, we show a $2$% accuracy degradation from NLP baselines and a $0.5$% improvement from state-of-the-art methods on image classification at competitive FLOPs reductions. We further demonstrate the generalization of tasks and architecture with comparative performance using Mask2Former for semantic segmentation and cnn-style networks. OPTIN presents one of the first one-shot efficient frameworks for compressing transformer architectures that generalizes well across different class domains, in particular: natural language and image-related tasks, without $re-training$.
2024-03-26
{'id': '2403.17921v1', 'title': 'The Need for Speed: Pruning Transformers with One Recipe', 'abstract': 'We introduce the $\\textbf{O}$ne-shot $\\textbf{P}$runing $\\textbf{T}$echnique\nfor $\\textbf{I}$nterchangeable $\\textbf{N}$etworks ($\\textbf{OPTIN}$) framework\nas a tool to increase the efficiency of pre-trained transformer architectures\n$\\textit{without requiring re-training}$. Recent works have explored improving\ntransformer efficiency, however often incur computationally expensive\nre-training procedures or depend on architecture-specific characteristics, thus\nimpeding practical wide-scale adoption. To address these shortcomings, the\nOPTIN framework leverages intermediate feature distillation, capturing the\nlong-range dependencies of model parameters (coined $\\textit{trajectory}$), to\nproduce state-of-the-art results on natural language, image classification,\ntransfer learning, and semantic segmentation tasks $\\textit{without\nre-training}$. Given a FLOP constraint, the OPTIN framework will compress the\nnetwork while maintaining competitive accuracy performance and improved\nthroughput. Particularly, we show a $\\leq 2$% accuracy degradation from NLP\nbaselines and a $0.5$% improvement from state-of-the-art methods on image\nclassification at competitive FLOPs reductions. We further demonstrate the\ngeneralization of tasks and architecture with comparative performance using\nMask2Former for semantic segmentation and cnn-style networks. OPTIN presents\none of the first one-shot efficient frameworks for compressing transformer\narchitectures that generalizes well across different class domains, in\nparticular: natural language and image-related tasks, without\n$\\textit{re-training}$.', 'authors': ['Samir Khaki', 'Konstantinos N. Plataniotis'], 'published': '2024-03-26 17:55:58+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2403.17921v1'}
42,567
2310.16802v2
2,310.16802
2
From Molecules to Materials: Pre-training Large Generalizable Models for Atomic Property Prediction
Foundation models have been transformational in machine learning fields such as natural language processing and computer vision. Similar success in atomic property prediction has been limited due to the challenges of training effective models across multiple chemical domains. To address this, we introduce Joint Multi-domain Pre-training (JMP), a supervised pre-training strategy that simultaneously trains on multiple datasets from different chemical domains, treating each dataset as a unique pre-training task within a multi-task framework. Our combined training dataset consists of $$120M systems from OC20, OC22, ANI-1x, and Transition-1x. We evaluate performance and generalization by fine-tuning over a diverse set of downstream tasks and datasets including: QM9, rMD17, MatBench, QMOF, SPICE, and MD22. JMP demonstrates an average improvement of 59% over training from scratch, and matches or sets state-of-the-art on 34 out of 40 tasks. Our work highlights the potential of pre-training strategies that utilize diverse data to advance property prediction across chemical domains, especially for low-data tasks. Please visit https://nima.sh/jmp for further information.
2023-10-25
{'id': '2310.16802v2', 'title': 'From Molecules to Materials: Pre-training Large Generalizable Models for Atomic Property Prediction', 'abstract': 'Foundation models have been transformational in machine learning fields such\nas natural language processing and computer vision. Similar success in atomic\nproperty prediction has been limited due to the challenges of training\neffective models across multiple chemical domains. To address this, we\nintroduce Joint Multi-domain Pre-training (JMP), a supervised pre-training\nstrategy that simultaneously trains on multiple datasets from different\nchemical domains, treating each dataset as a unique pre-training task within a\nmulti-task framework. Our combined training dataset consists of $\\sim$120M\nsystems from OC20, OC22, ANI-1x, and Transition-1x. We evaluate performance and\ngeneralization by fine-tuning over a diverse set of downstream tasks and\ndatasets including: QM9, rMD17, MatBench, QMOF, SPICE, and MD22. JMP\ndemonstrates an average improvement of 59% over training from scratch, and\nmatches or sets state-of-the-art on 34 out of 40 tasks. Our work highlights the\npotential of pre-training strategies that utilize diverse data to advance\nproperty prediction across chemical domains, especially for low-data tasks.\nPlease visit https://nima.sh/jmp for further information.', 'authors': ['Nima Shoghi', 'Adeesh Kolluru', 'John R. Kitchin', 'Zachary W. Ulissi', 'C. Lawrence Zitnick', 'Brandon M. Wood'], 'published': '2023-10-25 17:32:23+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.16802v2'}
42,529
2302.14853v2
2,302.14853
2
An Efficient Tester-Learner for Halfspaces
We give the first efficient algorithm for learning halfspaces in the testable learning model recently defined by Rubinfeld and Vasilyan (2023). In this model, a learner certifies that the accuracy of its output hypothesis is near optimal whenever the training set passes an associated test, and training sets drawn from some target distribution -- e.g., the Gaussian -- must pass the test. This model is more challenging than distribution-specific agnostic or Massart noise models where the learner is allowed to fail arbitrarily if the distributional assumption does not hold. We consider the setting where the target distribution is Gaussian (or more generally any strongly log-concave distribution) in $d$ dimensions and the noise model is either Massart or adversarial (agnostic). For Massart noise, our tester-learner runs in polynomial time and outputs a hypothesis with (information-theoretically optimal) error $opt + $ for any strongly log-concave target distribution. For adversarial noise, our tester-learner obtains error $O(opt) + $ in polynomial time when the target distribution is Gaussian; for strongly log-concave distributions, we obtain $O(opt) + $ in quasipolynomial time. Prior work on testable learning ignores the labels in the training set and checks that the empirical moments of the covariates are close to the moments of the base distribution. Here we develop new tests of independent interest that make critical use of the labels and combine them with the moment-matching approach of Gollakota et al. (2023). This enables us to simulate a variant of the algorithm of Diakonikolas et al. (2020) for learning noisy halfspaces using nonconvex SGD but in the testable learning setting.
2023-02-28
{'id': '2302.14853v2', 'title': 'An Efficient Tester-Learner for Halfspaces', 'abstract': 'We give the first efficient algorithm for learning halfspaces in the testable\nlearning model recently defined by Rubinfeld and Vasilyan (2023). In this\nmodel, a learner certifies that the accuracy of its output hypothesis is near\noptimal whenever the training set passes an associated test, and training sets\ndrawn from some target distribution -- e.g., the Gaussian -- must pass the\ntest. This model is more challenging than distribution-specific agnostic or\nMassart noise models where the learner is allowed to fail arbitrarily if the\ndistributional assumption does not hold.\n We consider the setting where the target distribution is Gaussian (or more\ngenerally any strongly log-concave distribution) in $d$ dimensions and the\nnoise model is either Massart or adversarial (agnostic). For Massart noise, our\ntester-learner runs in polynomial time and outputs a hypothesis with\n(information-theoretically optimal) error $\\mathsf{opt} + \\epsilon$ for any\nstrongly log-concave target distribution. For adversarial noise, our\ntester-learner obtains error $O(\\mathsf{opt}) + \\epsilon$ in polynomial time\nwhen the target distribution is Gaussian; for strongly log-concave\ndistributions, we obtain $\\tilde{O}(\\mathsf{opt}) + \\epsilon$ in\nquasipolynomial time.\n Prior work on testable learning ignores the labels in the training set and\nchecks that the empirical moments of the covariates are close to the moments of\nthe base distribution. Here we develop new tests of independent interest that\nmake critical use of the labels and combine them with the moment-matching\napproach of Gollakota et al. (2023). This enables us to simulate a variant of\nthe algorithm of Diakonikolas et al. (2020) for learning noisy halfspaces using\nnonconvex SGD but in the testable learning setting.', 'authors': ['Aravind Gollakota', 'Adam R. Klivans', 'Konstantinos Stavropoulos', 'Arsen Vasilyan'], 'published': '2023-02-28 18:51:55+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2302.14853v2'}
42,531
2410.12166v1
2,410.12166
1
Reclaiming the Source of Programmatic Policies: Programmatic versus Latent Spaces
Recent works have introduced LEAPS and HPRL, systems that learn latent spaces of domain-specific languages, which are used to define programmatic policies for partially observable Markov decision processes (POMDPs). These systems induce a latent space while optimizing losses such as the behavior loss, which aim to achieve locality in program behavior, meaning that vectors close in the latent space should correspond to similarly behaving programs. In this paper, we show that the programmatic space, induced by the domain-specific language and requiring no training, presents values for the behavior loss similar to those observed in latent spaces presented in previous work. Moreover, algorithms searching in the programmatic space significantly outperform those in LEAPS and HPRL. To explain our results, we measured the "friendliness" of the two spaces to local search algorithms. We discovered that algorithms are more likely to stop at local maxima when searching in the latent space than when searching in the programmatic space. This implies that the optimization topology of the programmatic space, induced by the reward function in conjunction with the neighborhood function, is more conducive to search than that of the latent space. This result provides an explanation for the superior performance in the programmatic space.
2024-10-16
{'id': '2410.12166v1', 'title': 'Reclaiming the Source of Programmatic Policies: Programmatic versus Latent Spaces', 'abstract': 'Recent works have introduced LEAPS and HPRL, systems that learn latent spaces\nof domain-specific languages, which are used to define programmatic policies\nfor partially observable Markov decision processes (POMDPs). These systems\ninduce a latent space while optimizing losses such as the behavior loss, which\naim to achieve locality in program behavior, meaning that vectors close in the\nlatent space should correspond to similarly behaving programs. In this paper,\nwe show that the programmatic space, induced by the domain-specific language\nand requiring no training, presents values for the behavior loss similar to\nthose observed in latent spaces presented in previous work. Moreover,\nalgorithms searching in the programmatic space significantly outperform those\nin LEAPS and HPRL. To explain our results, we measured the "friendliness" of\nthe two spaces to local search algorithms. We discovered that algorithms are\nmore likely to stop at local maxima when searching in the latent space than\nwhen searching in the programmatic space. This implies that the optimization\ntopology of the programmatic space, induced by the reward function in\nconjunction with the neighborhood function, is more conducive to search than\nthat of the latent space. This result provides an explanation for the superior\nperformance in the programmatic space.', 'authors': ['Tales H. Carvalho', 'Kenneth Tjhia', 'Levi H. S. Lelis'], 'published': '2024-10-16 02:10:04+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2410.12166v1'}
42,532
2310.04418v2
2,310.04418
2
Functional Interpolation for Relative Positions Improves Long Context Transformers
Preventing the performance decay of Transformers on inputs longer than those used for training has been an important challenge in extending the context length of these models. Though the Transformer architecture has fundamentally no limits on the input sequence lengths it can process, the choice of position encoding used during training can limit the performance of these models on longer inputs. We propose a novel functional relative position encoding with progressive interpolation, FIRE, to improve Transformer generalization to longer contexts. We theoretically prove that this can represent some of the popular relative position encodings, such as T5's RPE, Alibi, and Kerple. We next empirically show that FIRE models have better generalization to longer contexts on both zero-shot language modeling and long text benchmarks.
2023-10-06
{'id': '2310.04418v2', 'title': 'Functional Interpolation for Relative Positions Improves Long Context Transformers', 'abstract': "Preventing the performance decay of Transformers on inputs longer than those\nused for training has been an important challenge in extending the context\nlength of these models. Though the Transformer architecture has fundamentally\nno limits on the input sequence lengths it can process, the choice of position\nencoding used during training can limit the performance of these models on\nlonger inputs. We propose a novel functional relative position encoding with\nprogressive interpolation, FIRE, to improve Transformer generalization to\nlonger contexts. We theoretically prove that this can represent some of the\npopular relative position encodings, such as T5's RPE, Alibi, and Kerple. We\nnext empirically show that FIRE models have better generalization to longer\ncontexts on both zero-shot language modeling and long text benchmarks.", 'authors': ['Shanda Li', 'Chong You', 'Guru Guruganesh', 'Joshua Ainslie', 'Santiago Ontanon', 'Manzil Zaheer', 'Sumit Sanghai', 'Yiming Yang', 'Sanjiv Kumar', 'Srinadh Bhojanapalli'], 'published': '2023-10-06 17:59:11+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.04418v2'}
42,535
2405.14853v1
2,405.14853
1
Privileged Sensing Scaffolds Reinforcement Learning
We need to look at our shoelaces as we first learn to tie them but having mastered this skill, can do it from touch alone. We call this phenomenon "sensory scaffolding": observation streams that are not needed by a master might yet aid a novice learner. We consider such sensory scaffolding setups for training artificial agents. For example, a robot arm may need to be deployed with just a low-cost, robust, general-purpose camera; yet its performance may improve by having privileged training-time-only access to informative albeit expensive and unwieldy motion capture rigs or fragile tactile sensors. For these settings, we propose "Scaffolder", a reinforcement learning approach which effectively exploits privileged sensing in critics, world models, reward estimators, and other such auxiliary components that are only used at training time, to improve the target policy. For evaluating sensory scaffolding agents, we design a new "S3" suite of ten diverse simulated robotic tasks that explore a wide range of practical sensor setups. Agents must use privileged camera sensing to train blind hurdlers, privileged active visual perception to help robot arms overcome visual occlusions, privileged touch sensors to train robot hands, and more. Scaffolder easily outperforms relevant prior baselines and frequently performs comparably even to policies that have test-time access to the privileged sensors. Website: https://penn-pal-lab.github.io/scaffolder/
2024-05-23
{'id': '2405.14853v1', 'title': 'Privileged Sensing Scaffolds Reinforcement Learning', 'abstract': 'We need to look at our shoelaces as we first learn to tie them but having\nmastered this skill, can do it from touch alone. We call this phenomenon\n"sensory scaffolding": observation streams that are not needed by a master\nmight yet aid a novice learner. We consider such sensory scaffolding setups for\ntraining artificial agents. For example, a robot arm may need to be deployed\nwith just a low-cost, robust, general-purpose camera; yet its performance may\nimprove by having privileged training-time-only access to informative albeit\nexpensive and unwieldy motion capture rigs or fragile tactile sensors. For\nthese settings, we propose "Scaffolder", a reinforcement learning approach\nwhich effectively exploits privileged sensing in critics, world models, reward\nestimators, and other such auxiliary components that are only used at training\ntime, to improve the target policy. For evaluating sensory scaffolding agents,\nwe design a new "S3" suite of ten diverse simulated robotic tasks that explore\na wide range of practical sensor setups. Agents must use privileged camera\nsensing to train blind hurdlers, privileged active visual perception to help\nrobot arms overcome visual occlusions, privileged touch sensors to train robot\nhands, and more. Scaffolder easily outperforms relevant prior baselines and\nfrequently performs comparably even to policies that have test-time access to\nthe privileged sensors. Website: https://penn-pal-lab.github.io/scaffolder/', 'authors': ['Edward S. Hu', 'James Springer', 'Oleh Rybkin', 'Dinesh Jayaraman'], 'published': '2024-05-23 17:57:14+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.RO'], 'url': 'http://arxiv.org/abs/2405.14853v1'}
42,539
2312.06960v1
2,312.0696
1
Remote Sensing Vision-Language Foundation Models without Annotations via Ground Remote Alignment
We introduce a method to train vision-language models for remote-sensing images without using any textual annotations. Our key insight is to use co-located internet imagery taken on the ground as an intermediary for connecting remote-sensing images and language. Specifically, we train an image encoder for remote sensing images to align with the image encoder of CLIP using a large amount of paired internet and satellite images. Our unsupervised approach enables the training of a first-of-its-kind large-scale vision language model (VLM) for remote sensing images at two different resolutions. We show that these VLMs enable zero-shot, open-vocabulary image classification, retrieval, segmentation and visual question answering for satellite images. On each of these tasks, our VLM trained without textual annotations outperforms existing VLMs trained with supervision, with gains of up to 20% for classification and 80% for segmentation.
2023-12-12
{'id': '2312.06960v1', 'title': 'Remote Sensing Vision-Language Foundation Models without Annotations via Ground Remote Alignment', 'abstract': 'We introduce a method to train vision-language models for remote-sensing\nimages without using any textual annotations. Our key insight is to use\nco-located internet imagery taken on the ground as an intermediary for\nconnecting remote-sensing images and language. Specifically, we train an image\nencoder for remote sensing images to align with the image encoder of CLIP using\na large amount of paired internet and satellite images. Our unsupervised\napproach enables the training of a first-of-its-kind large-scale vision\nlanguage model (VLM) for remote sensing images at two different resolutions. We\nshow that these VLMs enable zero-shot, open-vocabulary image classification,\nretrieval, segmentation and visual question answering for satellite images. On\neach of these tasks, our VLM trained without textual annotations outperforms\nexisting VLMs trained with supervision, with gains of up to 20% for\nclassification and 80% for segmentation.', 'authors': ['Utkarsh Mall', 'Cheng Perng Phoo', 'Meilin Kelsey Liu', 'Carl Vondrick', 'Bharath Hariharan', 'Kavita Bala'], 'published': '2023-12-12 03:39:07+00:00', 'categories': ['cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2312.06960v1'}
42,544
2310.06301v1
2,310.06301
1
Dynamical versus Bayesian Phase Transitions in a Toy Model of Superposition
We investigate phase transitions in a Toy Model of Superposition (TMS) using Singular Learning Theory (SLT). We derive a closed formula for the theoretical loss and, in the case of two hidden dimensions, discover that regular $k$-gons are critical points. We present supporting theory indicating that the local learning coefficient (a geometric invariant) of these $k$-gons determines phase transitions in the Bayesian posterior as a function of training sample size. We then show empirically that the same $k$-gon critical points also determine the behavior of SGD training. The picture that emerges adds evidence to the conjecture that the SGD learning trajectory is subject to a sequential learning mechanism. Specifically, we find that the learning process in TMS, be it through SGD or Bayesian learning, can be characterized by a journey through parameter space from regions of high loss and low complexity to regions of low loss and high complexity.
2023-10-10
{'id': '2310.06301v1', 'title': 'Dynamical versus Bayesian Phase Transitions in a Toy Model of Superposition', 'abstract': 'We investigate phase transitions in a Toy Model of Superposition (TMS) using\nSingular Learning Theory (SLT). We derive a closed formula for the theoretical\nloss and, in the case of two hidden dimensions, discover that regular $k$-gons\nare critical points. We present supporting theory indicating that the local\nlearning coefficient (a geometric invariant) of these $k$-gons determines phase\ntransitions in the Bayesian posterior as a function of training sample size. We\nthen show empirically that the same $k$-gon critical points also determine the\nbehavior of SGD training. The picture that emerges adds evidence to the\nconjecture that the SGD learning trajectory is subject to a sequential learning\nmechanism. Specifically, we find that the learning process in TMS, be it\nthrough SGD or Bayesian learning, can be characterized by a journey through\nparameter space from regions of high loss and low complexity to regions of low\nloss and high complexity.', 'authors': ['Zhongtian Chen', 'Edmund Lau', 'Jake Mendel', 'Susan Wei', 'Daniel Murfet'], 'published': '2023-10-10 04:26:04+00:00', 'categories': ['cs.LG', 'cs.AI', '62F15, 68T07'], 'url': 'http://arxiv.org/abs/2310.06301v1'}
42,548
2405.03864v1
2,405.03864
1
Learning Planning Abstractions from Language
This paper presents a framework for learning state and action abstractions in sequential decision-making domains. Our framework, planning abstraction from language (PARL), utilizes language-annotated demonstrations to automatically discover a symbolic and abstract action space and induce a latent state abstraction based on it. PARL consists of three stages: 1) recovering object-level and action concepts, 2) learning state abstractions, abstract action feasibility, and transition models, and 3) applying low-level policies for abstract actions. During inference, given the task description, PARL first makes abstract action plans using the latent transition and feasibility functions, then refines the high-level plan using low-level policies. PARL generalizes across scenarios involving novel object instances and environments, unseen concept compositions, and tasks that require longer planning horizons than settings it is trained on.
2024-05-06
{'id': '2405.03864v1', 'title': 'Learning Planning Abstractions from Language', 'abstract': 'This paper presents a framework for learning state and action abstractions in\nsequential decision-making domains. Our framework, planning abstraction from\nlanguage (PARL), utilizes language-annotated demonstrations to automatically\ndiscover a symbolic and abstract action space and induce a latent state\nabstraction based on it. PARL consists of three stages: 1) recovering\nobject-level and action concepts, 2) learning state abstractions, abstract\naction feasibility, and transition models, and 3) applying low-level policies\nfor abstract actions. During inference, given the task description, PARL first\nmakes abstract action plans using the latent transition and feasibility\nfunctions, then refines the high-level plan using low-level policies. PARL\ngeneralizes across scenarios involving novel object instances and environments,\nunseen concept compositions, and tasks that require longer planning horizons\nthan settings it is trained on.', 'authors': ['Weiyu Liu', 'Geng Chen', 'Joy Hsu', 'Jiayuan Mao', 'Jiajun Wu'], 'published': '2024-05-06 21:24:22+00:00', 'categories': ['cs.RO', 'cs.AI'], 'url': 'http://arxiv.org/abs/2405.03864v1'}
42,536
2305.14718v5
2,305.14718
5
Leftover Lunch: Advantage-based Offline Reinforcement Learning for Language Models
Reinforcement Learning with Human Feedback (RLHF) is the most prominent method for Language Model (LM) alignment. However, RLHF is an unstable and data-hungry process that continually requires new high-quality LM-generated data for finetuning. We introduce Advantage-Leftover Lunch RL (A-LoL), a new class of offline policy gradient algorithms that enable RL training on any pre-existing data. By assuming the entire LM output sequence as a single action, A-LoL allows incorporating sequence-level classifiers or human-designed scoring functions as rewards. Subsequently, by using LM's value estimate, A-LoL only trains on positive advantage (leftover) data points, making it resilient to noise. Overall, A-LoL is an easy-to-implement, sample-efficient, and stable LM training recipe. We demonstrate the effectiveness of A-LoL and its variants with a set of four different language generation tasks. We compare against both online RL (PPO) and recent preference-based (DPO, PRO) and reward-based (GOLD) offline RL baselines. On the commonly-used RLHF benchmark, Helpful and Harmless Assistant (HHA), LMs trained with A-LoL methods achieve the highest diversity while also being rated more safe and helpful than the baselines according to humans. Additionally, in the remaining three tasks, A-LoL could optimize multiple distinct reward functions even when using noisy or suboptimal training data. We also release our experimental code. https://github.com/abaheti95/LoL-RL
2023-05-24
{'id': '2305.14718v5', 'title': 'Leftover Lunch: Advantage-based Offline Reinforcement Learning for Language Models', 'abstract': "Reinforcement Learning with Human Feedback (RLHF) is the most prominent\nmethod for Language Model (LM) alignment. However, RLHF is an unstable and\ndata-hungry process that continually requires new high-quality LM-generated\ndata for finetuning. We introduce Advantage-Leftover Lunch RL (A-LoL), a new\nclass of offline policy gradient algorithms that enable RL training on any\npre-existing data. By assuming the entire LM output sequence as a single\naction, A-LoL allows incorporating sequence-level classifiers or human-designed\nscoring functions as rewards. Subsequently, by using LM's value estimate, A-LoL\nonly trains on positive advantage (leftover) data points, making it resilient\nto noise. Overall, A-LoL is an easy-to-implement, sample-efficient, and stable\nLM training recipe.\n We demonstrate the effectiveness of A-LoL and its variants with a set of four\ndifferent language generation tasks. We compare against both online RL (PPO)\nand recent preference-based (DPO, PRO) and reward-based (GOLD) offline RL\nbaselines. On the commonly-used RLHF benchmark, Helpful and Harmless Assistant\n(HHA), LMs trained with A-LoL methods achieve the highest diversity while also\nbeing rated more safe and helpful than the baselines according to humans.\nAdditionally, in the remaining three tasks, A-LoL could optimize multiple\ndistinct reward functions even when using noisy or suboptimal training data.\n We also release our experimental code. https://github.com/abaheti95/LoL-RL", 'authors': ['Ashutosh Baheti', 'Ximing Lu', 'Faeze Brahman', 'Ronan Le Bras', 'Maarten Sap', 'Mark Riedl'], 'published': '2023-05-24 04:42:17+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2305.14718v5'}
42,545
2502.14160v1
2,502.1416
1
Efficient Inverse Multiagent Learning
In this paper, we study inverse game theory (resp. inverse multiagent learning) in which the goal is to find parameters of a game's payoff functions for which the expected (resp. sampled) behavior is an equilibrium. We formulate these problems as generative-adversarial (i.e., min-max) optimization problems, for which we develop polynomial-time algorithms to solve, the former of which relies on an exact first-order oracle, and the latter, a stochastic one. We extend our approach to solve inverse multiagent simulacral learning in polynomial time and number of samples. In these problems, we seek a simulacrum, meaning parameters and an associated equilibrium that replicate the given observations in expectation. We find that our approach outperforms the widely-used ARIMA method in predicting prices in Spanish electricity markets based on time-series data.
2025-02-20
{'id': '2502.14160v1', 'title': 'Efficient Inverse Multiagent Learning', 'abstract': "In this paper, we study inverse game theory (resp. inverse multiagent\nlearning) in which the goal is to find parameters of a game's payoff functions\nfor which the expected (resp. sampled) behavior is an equilibrium. We formulate\nthese problems as generative-adversarial (i.e., min-max) optimization problems,\nfor which we develop polynomial-time algorithms to solve, the former of which\nrelies on an exact first-order oracle, and the latter, a stochastic one. We\nextend our approach to solve inverse multiagent simulacral learning in\npolynomial time and number of samples. In these problems, we seek a simulacrum,\nmeaning parameters and an associated equilibrium that replicate the given\nobservations in expectation. We find that our approach outperforms the\nwidely-used ARIMA method in predicting prices in Spanish electricity markets\nbased on time-series data.", 'authors': ['Denizalp Goktas', 'Amy Greenwald', 'Sadie Zhao', 'Alec Koppel', 'Sumitra Ganesh'], 'published': '2025-02-20 00:07:06+00:00', 'categories': ['cs.GT', 'cs.AI', 'cs.LG', 'econ.TH'], 'url': 'http://arxiv.org/abs/2502.14160v1'}
42,549
2310.10047v1
2,310.10047
1
Improving Large Language Model Fine-tuning for Solving Math Problems
Despite their success in many natural language tasks, solving math problems remains a significant challenge for large language models (LLMs). A large gap exists between LLMs' pass-at-one and pass-at-N performance in solving math problems, suggesting LLMs might be close to finding correct solutions, motivating our exploration of fine-tuning methods to unlock LLMs' performance. Using the challenging MATH dataset, we investigate three fine-tuning strategies: (1) solution fine-tuning, where we fine-tune to generate a detailed solution for a given math problem; (2) solution-cluster re-ranking, where the LLM is fine-tuned as a solution verifier/evaluator to choose among generated candidate solution clusters; (3) multi-task sequential fine-tuning, which integrates both solution generation and evaluation tasks together efficiently to enhance the LLM performance. With these methods, we present a thorough empirical study on a series of PaLM 2 models and find: (1) The quality and style of the step-by-step solutions used for fine-tuning can make a significant impact on the model performance; (2) While solution re-ranking and majority voting are both effective for improving the model performance when used separately, they can also be used together for an even greater performance boost; (3) Multi-task fine-tuning that sequentially separates the solution generation and evaluation tasks can offer improved performance compared with the solution fine-tuning baseline. Guided by these insights, we design a fine-tuning recipe that yields approximately 58.8% accuracy on the MATH dataset with fine-tuned PaLM 2-L models, an 11.2% accuracy improvement over the few-shot performance of pre-trained PaLM 2-L model with majority voting.
2023-10-16
{'id': '2310.10047v1', 'title': 'Improving Large Language Model Fine-tuning for Solving Math Problems', 'abstract': "Despite their success in many natural language tasks, solving math problems\nremains a significant challenge for large language models (LLMs). A large gap\nexists between LLMs' pass-at-one and pass-at-N performance in solving math\nproblems, suggesting LLMs might be close to finding correct solutions,\nmotivating our exploration of fine-tuning methods to unlock LLMs' performance.\nUsing the challenging MATH dataset, we investigate three fine-tuning\nstrategies: (1) solution fine-tuning, where we fine-tune to generate a detailed\nsolution for a given math problem; (2) solution-cluster re-ranking, where the\nLLM is fine-tuned as a solution verifier/evaluator to choose among generated\ncandidate solution clusters; (3) multi-task sequential fine-tuning, which\nintegrates both solution generation and evaluation tasks together efficiently\nto enhance the LLM performance. With these methods, we present a thorough\nempirical study on a series of PaLM 2 models and find: (1) The quality and\nstyle of the step-by-step solutions used for fine-tuning can make a significant\nimpact on the model performance; (2) While solution re-ranking and majority\nvoting are both effective for improving the model performance when used\nseparately, they can also be used together for an even greater performance\nboost; (3) Multi-task fine-tuning that sequentially separates the solution\ngeneration and evaluation tasks can offer improved performance compared with\nthe solution fine-tuning baseline. Guided by these insights, we design a\nfine-tuning recipe that yields approximately 58.8% accuracy on the MATH dataset\nwith fine-tuned PaLM 2-L models, an 11.2% accuracy improvement over the\nfew-shot performance of pre-trained PaLM 2-L model with majority voting.", 'authors': ['Yixin Liu', 'Avi Singh', 'C. Daniel Freeman', 'John D. Co-Reyes', 'Peter J. Liu'], 'published': '2023-10-16 04:11:19+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2310.10047v1'}
42,551
2401.14469v1
2,401.14469
1
Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels
Recent advances in depthwise-separable convolutional neural networks (DS-CNNs) have led to novel architectures, that surpass the performance of classical CNNs, by a considerable scalability and accuracy margin. This paper reveals another striking property of DS-CNN architectures: discernible and explainable patterns emerge in their trained depthwise convolutional kernels in all layers. Through an extensive analysis of millions of trained filters, with different sizes and from various models, we employed unsupervised clustering with autoencoders, to categorize these filters. Astonishingly, the patterns converged into a few main clusters, each resembling the difference of Gaussian (DoG) functions, and their first and second-order derivatives. Notably, we were able to classify over 95\% and 90\% of the filters from state-of-the-art ConvNextV2 and ConvNeXt models, respectively. This finding is not merely a technological curiosity; it echoes the foundational models neuroscientists have long proposed for the vision systems of mammals. Our results thus deepen our understanding of the emergent properties of trained DS-CNNs and provide a bridge between artificial and biological visual processing systems. More broadly, they pave the way for more interpretable and biologically-inspired neural network designs in the future.
2024-01-25
{'id': '2401.14469v1', 'title': 'Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels', 'abstract': 'Recent advances in depthwise-separable convolutional neural networks\n(DS-CNNs) have led to novel architectures, that surpass the performance of\nclassical CNNs, by a considerable scalability and accuracy margin. This paper\nreveals another striking property of DS-CNN architectures: discernible and\nexplainable patterns emerge in their trained depthwise convolutional kernels in\nall layers. Through an extensive analysis of millions of trained filters, with\ndifferent sizes and from various models, we employed unsupervised clustering\nwith autoencoders, to categorize these filters. Astonishingly, the patterns\nconverged into a few main clusters, each resembling the difference of Gaussian\n(DoG) functions, and their first and second-order derivatives. Notably, we were\nable to classify over 95\\% and 90\\% of the filters from state-of-the-art\nConvNextV2 and ConvNeXt models, respectively. This finding is not merely a\ntechnological curiosity; it echoes the foundational models neuroscientists have\nlong proposed for the vision systems of mammals. Our results thus deepen our\nunderstanding of the emergent properties of trained DS-CNNs and provide a\nbridge between artificial and biological visual processing systems. More\nbroadly, they pave the way for more interpretable and biologically-inspired\nneural network designs in the future.', 'authors': ['Zahra Babaiee', 'Peyman M. Kiasari', 'Daniela Rus', 'Radu Grosu'], 'published': '2024-01-25 19:05:53+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CV', 'cs.NE'], 'url': 'http://arxiv.org/abs/2401.14469v1'}
42,553
2310.02557v3
2,310.02557
3
Generalization in diffusion models arises from geometry-adaptive harmonic representations
Deep neural networks (DNNs) trained for image denoising are able to generate high-quality samples with score-based reverse diffusion algorithms. These impressive capabilities seem to imply an escape from the curse of dimensionality, but recent reports of memorization of the training set raise the question of whether these networks are learning the "true" continuous density of the data. Here, we show that two DNNs trained on non-overlapping subsets of a dataset learn nearly the same score function, and thus the same density, when the number of training images is large enough. In this regime of strong generalization, diffusion-generated images are distinct from the training set, and are of high visual quality, suggesting that the inductive biases of the DNNs are well-aligned with the data density. We analyze the learned denoising functions and show that the inductive biases give rise to a shrinkage operation in a basis adapted to the underlying image. Examination of these bases reveals oscillating harmonic structures along contours and in homogeneous regions. We demonstrate that trained denoisers are inductively biased towards these geometry-adaptive harmonic bases since they arise not only when the network is trained on photographic images, but also when it is trained on image classes supported on low-dimensional manifolds for which the harmonic basis is suboptimal. Finally, we show that when trained on regular image classes for which the optimal basis is known to be geometry-adaptive and harmonic, the denoising performance of the networks is near-optimal.
2023-10-04
{'id': '2310.02557v3', 'title': 'Generalization in diffusion models arises from geometry-adaptive harmonic representations', 'abstract': 'Deep neural networks (DNNs) trained for image denoising are able to generate\nhigh-quality samples with score-based reverse diffusion algorithms. These\nimpressive capabilities seem to imply an escape from the curse of\ndimensionality, but recent reports of memorization of the training set raise\nthe question of whether these networks are learning the "true" continuous\ndensity of the data. Here, we show that two DNNs trained on non-overlapping\nsubsets of a dataset learn nearly the same score function, and thus the same\ndensity, when the number of training images is large enough. In this regime of\nstrong generalization, diffusion-generated images are distinct from the\ntraining set, and are of high visual quality, suggesting that the inductive\nbiases of the DNNs are well-aligned with the data density. We analyze the\nlearned denoising functions and show that the inductive biases give rise to a\nshrinkage operation in a basis adapted to the underlying image. Examination of\nthese bases reveals oscillating harmonic structures along contours and in\nhomogeneous regions. We demonstrate that trained denoisers are inductively\nbiased towards these geometry-adaptive harmonic bases since they arise not only\nwhen the network is trained on photographic images, but also when it is trained\non image classes supported on low-dimensional manifolds for which the harmonic\nbasis is suboptimal. Finally, we show that when trained on regular image\nclasses for which the optimal basis is known to be geometry-adaptive and\nharmonic, the denoising performance of the networks is near-optimal.', 'authors': ['Zahra Kadkhodaie', 'Florentin Guth', 'Eero P. Simoncelli', 'Stéphane Mallat'], 'published': '2023-10-04 03:30:32+00:00', 'categories': ['cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2310.02557v3'}
42,562
2310.08796v1
2,310.08796
1
End-to-end Story Plot Generator
Story plots, while short, carry most of the essential information of a full story that may contain tens of thousands of words. We study the problem of automatic generation of story plots, which includes story premise, character descriptions, plot outlines, etc. To generate a single engaging plot, existing plot generators (e.g., DOC (Yang et al., 2022a)) require hundreds to thousands of calls to LLMs (e.g., OpenAI API) in the planning stage of the story plot, which is costly and takes at least several minutes. Moreover, the hard-wired nature of the method makes the pipeline non-differentiable, blocking fast specialization and personalization of the plot generator. In this paper, we propose three models, $OpenPlot$, $E2EPlot$ and $RLPlot$, to address these challenges. $OpenPlot$ replaces expensive OpenAI API calls with LLaMA2 (Touvron et al., 2023) calls via careful prompt designs, which leads to inexpensive generation of high-quality training datasets of story plots. We then train an end-to-end story plot generator, $E2EPlot$, by supervised fine-tuning (SFT) using approximately 13000 story plots generated by $OpenPlot$. $E2EPlot$ generates story plots of comparable quality to $OpenPlot$, and is > 10$$ faster (1k tokens in only 30 seconds on average). Finally, we obtain $RLPlot$ that is further fine-tuned with RLHF on several different reward models for different aspects of story quality, which yields 60.0$\%$ winning rate against $E2EPlot$ along the aspect of suspense and surprise.
2023-10-13
{'id': '2310.08796v1', 'title': 'End-to-end Story Plot Generator', 'abstract': 'Story plots, while short, carry most of the essential information of a full\nstory that may contain tens of thousands of words. We study the problem of\nautomatic generation of story plots, which includes story premise, character\ndescriptions, plot outlines, etc. To generate a single engaging plot, existing\nplot generators (e.g., DOC (Yang et al., 2022a)) require hundreds to thousands\nof calls to LLMs (e.g., OpenAI API) in the planning stage of the story plot,\nwhich is costly and takes at least several minutes. Moreover, the hard-wired\nnature of the method makes the pipeline non-differentiable, blocking fast\nspecialization and personalization of the plot generator. In this paper, we\npropose three models, $\\texttt{OpenPlot}$, $\\texttt{E2EPlot}$ and\n$\\texttt{RLPlot}$, to address these challenges. $\\texttt{OpenPlot}$ replaces\nexpensive OpenAI API calls with LLaMA2 (Touvron et al., 2023) calls via careful\nprompt designs, which leads to inexpensive generation of high-quality training\ndatasets of story plots. We then train an end-to-end story plot generator,\n$\\texttt{E2EPlot}$, by supervised fine-tuning (SFT) using approximately 13000\nstory plots generated by $\\texttt{OpenPlot}$. $\\texttt{E2EPlot}$ generates\nstory plots of comparable quality to $\\texttt{OpenPlot}$, and is > 10$\\times$\nfaster (1k tokens in only 30 seconds on average). Finally, we obtain\n$\\texttt{RLPlot}$ that is further fine-tuned with RLHF on several different\nreward models for different aspects of story quality, which yields 60.0$\\%$\nwinning rate against $\\texttt{E2EPlot}$ along the aspect of suspense and\nsurprise.', 'authors': ['Hanlin Zhu', 'Andrew Cohen', 'Danqing Wang', 'Kevin Yang', 'Xiaomeng Yang', 'Jiantao Jiao', 'Yuandong Tian'], 'published': '2023-10-13 00:49:59+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2310.08796v1'}
42,538
2310.05473v1
2,310.05473
1
Sentence-level Prompts Benefit Composed Image Retrieval
Composed image retrieval (CIR) is the task of retrieving specific images by using a query that involves both a reference image and a relative caption. Most existing CIR models adopt the late-fusion strategy to combine visual and language features. Besides, several approaches have also been suggested to generate a pseudo-word token from the reference image, which is further integrated into the relative caption for CIR. However, these pseudo-word-based prompting methods have limitations when target image encompasses complex changes on reference image, e.g., object removal and attribute modification. In this work, we demonstrate that learning an appropriate sentence-level prompt for the relative caption (SPRC) is sufficient for achieving effective composed image retrieval. Instead of relying on pseudo-word-based prompts, we propose to leverage pretrained V-L models, e.g., BLIP-2, to generate sentence-level prompts. By concatenating the learned sentence-level prompt with the relative caption, one can readily use existing text-based image retrieval models to enhance CIR performance. Furthermore, we introduce both image-text contrastive loss and text prompt alignment loss to enforce the learning of suitable sentence-level prompts. Experiments show that our proposed method performs favorably against the state-of-the-art CIR methods on the Fashion-IQ and CIRR datasets. The source code and pretrained model are publicly available at https://github.com/chunmeifeng/SPRC
2023-10-09
{'id': '2310.05473v1', 'title': 'Sentence-level Prompts Benefit Composed Image Retrieval', 'abstract': 'Composed image retrieval (CIR) is the task of retrieving specific images by\nusing a query that involves both a reference image and a relative caption. Most\nexisting CIR models adopt the late-fusion strategy to combine visual and\nlanguage features. Besides, several approaches have also been suggested to\ngenerate a pseudo-word token from the reference image, which is further\nintegrated into the relative caption for CIR. However, these pseudo-word-based\nprompting methods have limitations when target image encompasses complex\nchanges on reference image, e.g., object removal and attribute modification. In\nthis work, we demonstrate that learning an appropriate sentence-level prompt\nfor the relative caption (SPRC) is sufficient for achieving effective composed\nimage retrieval. Instead of relying on pseudo-word-based prompts, we propose to\nleverage pretrained V-L models, e.g., BLIP-2, to generate sentence-level\nprompts. By concatenating the learned sentence-level prompt with the relative\ncaption, one can readily use existing text-based image retrieval models to\nenhance CIR performance. Furthermore, we introduce both image-text contrastive\nloss and text prompt alignment loss to enforce the learning of suitable\nsentence-level prompts. Experiments show that our proposed method performs\nfavorably against the state-of-the-art CIR methods on the Fashion-IQ and CIRR\ndatasets. The source code and pretrained model are publicly available at\nhttps://github.com/chunmeifeng/SPRC', 'authors': ['Yang Bai', 'Xinxing Xu', 'Yong Liu', 'Salman Khan', 'Fahad Khan', 'Wangmeng Zuo', 'Rick Siow Mong Goh', 'Chun-Mei Feng'], 'published': '2023-10-09 07:31:44+00:00', 'categories': ['cs.CV', 'cs.AI'], 'url': 'http://arxiv.org/abs/2310.05473v1'}
42,540
2310.11346v3
2,310.11346
3
Towards Generalizable Multi-Camera 3D Object Detection via Perspective Debiasing
Detecting objects in 3D space using multiple cameras, known as Multi-Camera 3D Object Detection (MC3D-Det), has gained prominence with the advent of bird's-eye view (BEV) approaches. However, these methods often struggle when faced with unfamiliar testing environments due to the lack of diverse training data encompassing various viewpoints and environments. To address this, we propose a novel method that aligns 3D detection with 2D camera plane results, ensuring consistent and accurate detections. Our framework, anchored in perspective debiasing, helps the learning of features resilient to domain shifts. In our approach, we render diverse view maps from BEV features and rectify the perspective bias of these maps, leveraging implicit foreground volumes to bridge the camera and BEV planes. This two-step process promotes the learning of perspective- and context-independent features, crucial for accurate object detection across varying viewpoints, camera parameters, and environmental conditions. Notably, our model-agnostic approach preserves the original network structure without incurring additional inference costs, facilitating seamless integration across various models and simplifying deployment. Furthermore, we also show our approach achieves satisfactory results in real data when trained only with virtual datasets, eliminating the need for real scene annotations. Experimental results on both Domain Generalization (DG) and Unsupervised Domain Adaptation (UDA) clearly demonstrate its effectiveness. The codes are available at https://github.com/EnVision-Research/Generalizable-BEV.
2023-10-17
{'id': '2310.11346v3', 'title': 'Towards Generalizable Multi-Camera 3D Object Detection via Perspective Debiasing', 'abstract': "Detecting objects in 3D space using multiple cameras, known as Multi-Camera\n3D Object Detection (MC3D-Det), has gained prominence with the advent of\nbird's-eye view (BEV) approaches. However, these methods often struggle when\nfaced with unfamiliar testing environments due to the lack of diverse training\ndata encompassing various viewpoints and environments. To address this, we\npropose a novel method that aligns 3D detection with 2D camera plane results,\nensuring consistent and accurate detections. Our framework, anchored in\nperspective debiasing, helps the learning of features resilient to domain\nshifts. In our approach, we render diverse view maps from BEV features and\nrectify the perspective bias of these maps, leveraging implicit foreground\nvolumes to bridge the camera and BEV planes. This two-step process promotes the\nlearning of perspective- and context-independent features, crucial for accurate\nobject detection across varying viewpoints, camera parameters, and\nenvironmental conditions. Notably, our model-agnostic approach preserves the\noriginal network structure without incurring additional inference costs,\nfacilitating seamless integration across various models and simplifying\ndeployment. Furthermore, we also show our approach achieves satisfactory\nresults in real data when trained only with virtual datasets, eliminating the\nneed for real scene annotations. Experimental results on both Domain\nGeneralization (DG) and Unsupervised Domain Adaptation (UDA) clearly\ndemonstrate its effectiveness. The codes are available at\nhttps://github.com/EnVision-Research/Generalizable-BEV.", 'authors': ['Hao Lu', 'Yunpeng Zhang', 'Qing Lian', 'Dalong Du', 'Yingcong Chen'], 'published': '2023-10-17 15:31:28+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2310.11346v3'}
42,541
2310.02611v2
2,310.02611
2
Analyzing and Improving Optimal-Transport-based Adversarial Networks
Optimal Transport (OT) problem aims to find a transport plan that bridges two distributions while minimizing a given cost function. OT theory has been widely utilized in generative modeling. In the beginning, OT distance has been used as a measure for assessing the distance between data and generated distributions. Recently, OT transport map between data and prior distributions has been utilized as a generative model. These OT-based generative models share a similar adversarial training objective. In this paper, we begin by unifying these OT-based adversarial methods within a single framework. Then, we elucidate the role of each component in training dynamics through a comprehensive analysis of this unified framework. Moreover, we suggest a simple but novel method that improves the previously best-performing OT-based model. Intuitively, our approach conducts a gradual refinement of the generated distribution, progressively aligning it with the data distribution. Our approach achieves a FID score of 2.51 on CIFAR-10 and 5.99 on CelebA-HQ-256, outperforming unified OT-based adversarial approaches.
2023-10-04
{'id': '2310.02611v2', 'title': 'Analyzing and Improving Optimal-Transport-based Adversarial Networks', 'abstract': 'Optimal Transport (OT) problem aims to find a transport plan that bridges two\ndistributions while minimizing a given cost function. OT theory has been widely\nutilized in generative modeling. In the beginning, OT distance has been used as\na measure for assessing the distance between data and generated distributions.\nRecently, OT transport map between data and prior distributions has been\nutilized as a generative model. These OT-based generative models share a\nsimilar adversarial training objective. In this paper, we begin by unifying\nthese OT-based adversarial methods within a single framework. Then, we\nelucidate the role of each component in training dynamics through a\ncomprehensive analysis of this unified framework. Moreover, we suggest a simple\nbut novel method that improves the previously best-performing OT-based model.\nIntuitively, our approach conducts a gradual refinement of the generated\ndistribution, progressively aligning it with the data distribution. Our\napproach achieves a FID score of 2.51 on CIFAR-10 and 5.99 on CelebA-HQ-256,\noutperforming unified OT-based adversarial approaches.', 'authors': ['Jaemoo Choi', 'Jaewoong Choi', 'Myungjoo Kang'], 'published': '2023-10-04 06:52:03+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2310.02611v2'}
42,546
2310.20673v2
2,310.20673
2
Balancing Act: Constraining Disparate Impact in Sparse Models
Model pruning is a popular approach to enable the deployment of large deep learning models on edge devices with restricted computational or storage capacities. Although sparse models achieve performance comparable to that of their dense counterparts at the level of the entire dataset, they exhibit high accuracy drops for some data sub-groups. Existing methods to mitigate this disparate impact induced by pruning (i) rely on surrogate metrics that address the problem indirectly and have limited interpretability; or (ii) scale poorly with the number of protected sub-groups in terms of computational cost. We propose a constrained optimization approach that directly addresses the disparate impact of pruning: our formulation bounds the accuracy change between the dense and sparse models, for each sub-group. This choice of constraints provides an interpretable success criterion to determine if a pruned model achieves acceptable disparity levels. Experimental results demonstrate that our technique scales reliably to problems involving large models and hundreds of protected sub-groups.
2023-10-31
{'id': '2310.20673v2', 'title': 'Balancing Act: Constraining Disparate Impact in Sparse Models', 'abstract': 'Model pruning is a popular approach to enable the deployment of large deep\nlearning models on edge devices with restricted computational or storage\ncapacities. Although sparse models achieve performance comparable to that of\ntheir dense counterparts at the level of the entire dataset, they exhibit high\naccuracy drops for some data sub-groups. Existing methods to mitigate this\ndisparate impact induced by pruning (i) rely on surrogate metrics that address\nthe problem indirectly and have limited interpretability; or (ii) scale poorly\nwith the number of protected sub-groups in terms of computational cost. We\npropose a constrained optimization approach that directly addresses the\ndisparate impact of pruning: our formulation bounds the accuracy change between\nthe dense and sparse models, for each sub-group. This choice of constraints\nprovides an interpretable success criterion to determine if a pruned model\nachieves acceptable disparity levels. Experimental results demonstrate that our\ntechnique scales reliably to problems involving large models and hundreds of\nprotected sub-groups.', 'authors': ['Meraj Hashemizadeh', 'Juan Ramirez', 'Rohan Sukumaran', 'Golnoosh Farnadi', 'Simon Lacoste-Julien', 'Jose Gallego-Posada'], 'published': '2023-10-31 17:37:35+00:00', 'categories': ['cs.LG', 'cs.CY'], 'url': 'http://arxiv.org/abs/2310.20673v2'}
42,556
2401.03350v2
2,401.0335
2
Accurate and Scalable Estimation of Epistemic Uncertainty for Graph Neural Networks
While graph neural networks (GNNs) are widely used for node and graph representation learning tasks, the reliability of GNN uncertainty estimates under distribution shifts remains relatively under-explored. Indeed, while post-hoc calibration strategies can be used to improve in-distribution calibration, they need not also improve calibration under distribution shift. However, techniques which produce GNNs with better intrinsic uncertainty estimates are particularly valuable, as they can always be combined with post-hoc strategies later. Therefore, in this work, we propose G-$$UQ, a novel training framework designed to improve intrinsic GNN uncertainty estimates. Our framework adapts the principle of stochastic data centering to graph data through novel graph anchoring strategies, and is able to support partially stochastic GNNs. While, the prevalent wisdom is that fully stochastic networks are necessary to obtain reliable estimates, we find that the functional diversity induced by our anchoring strategies when sampling hypotheses renders this unnecessary and allows us to support G-$$UQ on pretrained models. Indeed, through extensive evaluation under covariate, concept and graph size shifts, we show that G-$$UQ leads to better calibrated GNNs for node and graph classification. Further, it also improves performance on the uncertainty-based tasks of out-of-distribution detection and generalization gap estimation. Overall, our work provides insights into uncertainty estimation for GNNs, and demonstrates the utility of G-$$UQ in obtaining reliable estimates.
2024-01-07
{'id': '2401.03350v2', 'title': 'Accurate and Scalable Estimation of Epistemic Uncertainty for Graph Neural Networks', 'abstract': 'While graph neural networks (GNNs) are widely used for node and graph\nrepresentation learning tasks, the reliability of GNN uncertainty estimates\nunder distribution shifts remains relatively under-explored. Indeed, while\npost-hoc calibration strategies can be used to improve in-distribution\ncalibration, they need not also improve calibration under distribution shift.\nHowever, techniques which produce GNNs with better intrinsic uncertainty\nestimates are particularly valuable, as they can always be combined with\npost-hoc strategies later. Therefore, in this work, we propose G-$\\Delta$UQ, a\nnovel training framework designed to improve intrinsic GNN uncertainty\nestimates. Our framework adapts the principle of stochastic data centering to\ngraph data through novel graph anchoring strategies, and is able to support\npartially stochastic GNNs. While, the prevalent wisdom is that fully stochastic\nnetworks are necessary to obtain reliable estimates, we find that the\nfunctional diversity induced by our anchoring strategies when sampling\nhypotheses renders this unnecessary and allows us to support G-$\\Delta$UQ on\npretrained models. Indeed, through extensive evaluation under covariate,\nconcept and graph size shifts, we show that G-$\\Delta$UQ leads to better\ncalibrated GNNs for node and graph classification. Further, it also improves\nperformance on the uncertainty-based tasks of out-of-distribution detection and\ngeneralization gap estimation. Overall, our work provides insights into\nuncertainty estimation for GNNs, and demonstrates the utility of G-$\\Delta$UQ\nin obtaining reliable estimates.', 'authors': ['Puja Trivedi', 'Mark Heimann', 'Rushil Anirudh', 'Danai Koutra', 'Jayaraman J. Thiagarajan'], 'published': '2024-01-07 00:58:33+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2401.03350v2'}
42,572
2308.01316v1
2,308.01316
1
Patched Denoising Diffusion Models For High-Resolution Image Synthesis
We propose an effective denoising diffusion model for generating high-resolution images (e.g., 1024$$512), trained on small-size image patches (e.g., 64$$64). We name our algorithm Patch-DM, in which a new feature collage strategy is designed to avoid the boundary artifact when synthesizing large-size images. Feature collage systematically crops and combines partial features of the neighboring patches to predict the features of a shifted image patch, allowing the seamless generation of the entire image due to the overlap in the patch feature space. Patch-DM produces high-quality image synthesis results on our newly collected dataset of nature images (1024$$512), as well as on standard benchmarks of smaller sizes (256$$256), including LSUN-Bedroom, LSUN-Church, and FFHQ. We compare our method with previous patch-based generation methods and achieve state-of-the-art FID scores on all four datasets. Further, Patch-DM also reduces memory complexity compared to the classic diffusion models.
2023-08-02
{'id': '2308.01316v1', 'title': 'Patched Denoising Diffusion Models For High-Resolution Image Synthesis', 'abstract': 'We propose an effective denoising diffusion model for generating\nhigh-resolution images (e.g., 1024$\\times$512), trained on small-size image\npatches (e.g., 64$\\times$64). We name our algorithm Patch-DM, in which a new\nfeature collage strategy is designed to avoid the boundary artifact when\nsynthesizing large-size images. Feature collage systematically crops and\ncombines partial features of the neighboring patches to predict the features of\na shifted image patch, allowing the seamless generation of the entire image due\nto the overlap in the patch feature space. Patch-DM produces high-quality image\nsynthesis results on our newly collected dataset of nature images\n(1024$\\times$512), as well as on standard benchmarks of smaller sizes\n(256$\\times$256), including LSUN-Bedroom, LSUN-Church, and FFHQ. We compare our\nmethod with previous patch-based generation methods and achieve\nstate-of-the-art FID scores on all four datasets. Further, Patch-DM also\nreduces memory complexity compared to the classic diffusion models.', 'authors': ['Zheng Ding', 'Mengqi Zhang', 'Jiajun Wu', 'Zhuowen Tu'], 'published': '2023-08-02 17:58:01+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2308.01316v1'}
42,542
2310.12808v2
2,310.12808
2
Model Merging by Uncertainty-Based Gradient Matching
Models trained on different datasets can be merged by a weighted-averaging of their parameters, but why does it work and when can it fail? Here, we connect the inaccuracy of weighted-averaging to mismatches in the gradients and propose a new uncertainty-based scheme to improve the performance by reducing the mismatch. The connection also reveals implicit assumptions in other schemes such as averaging, task arithmetic, and Fisher-weighted averaging. Our new method gives consistent improvements for large language models and vision transformers, both in terms of performance and robustness to hyperparameters. Code available here.
2023-10-19
{'id': '2310.12808v2', 'title': 'Model Merging by Uncertainty-Based Gradient Matching', 'abstract': 'Models trained on different datasets can be merged by a weighted-averaging of\ntheir parameters, but why does it work and when can it fail? Here, we connect\nthe inaccuracy of weighted-averaging to mismatches in the gradients and propose\na new uncertainty-based scheme to improve the performance by reducing the\nmismatch. The connection also reveals implicit assumptions in other schemes\nsuch as averaging, task arithmetic, and Fisher-weighted averaging. Our new\nmethod gives consistent improvements for large language models and vision\ntransformers, both in terms of performance and robustness to hyperparameters.\nCode available here.', 'authors': ['Nico Daheim', 'Thomas Möllenhoff', 'Edoardo Maria Ponti', 'Iryna Gurevych', 'Mohammad Emtiyaz Khan'], 'published': '2023-10-19 15:02:45+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CL'], 'url': 'http://arxiv.org/abs/2310.12808v2'}
42,547
2306.13103v1
2,306.13103
1
Evaluating the Robustness of Text-to-image Diffusion Models against Real-world Attacks
Text-to-image (T2I) diffusion models (DMs) have shown promise in generating high-quality images from textual descriptions. The real-world applications of these models require particular attention to their safety and fidelity, but this has not been sufficiently explored. One fundamental question is whether existing T2I DMs are robust against variations over input texts. To answer it, this work provides the first robustness evaluation of T2I DMs against real-world attacks. Unlike prior studies that focus on malicious attacks involving apocryphal alterations to the input texts, we consider an attack space spanned by realistic errors (e.g., typo, glyph, phonetic) that humans can make, to ensure semantic consistency. Given the inherent randomness of the generation process, we develop novel distribution-based attack objectives to mislead T2I DMs. We perform attacks in a black-box manner without any knowledge of the model. Extensive experiments demonstrate the effectiveness of our method for attacking popular T2I DMs and simultaneously reveal their non-trivial robustness issues. Moreover, we provide an in-depth analysis of our method to show that it is not designed to attack the text encoder in T2I DMs solely.
2023-06-16
{'id': '2306.13103v1', 'title': 'Evaluating the Robustness of Text-to-image Diffusion Models against Real-world Attacks', 'abstract': 'Text-to-image (T2I) diffusion models (DMs) have shown promise in generating\nhigh-quality images from textual descriptions. The real-world applications of\nthese models require particular attention to their safety and fidelity, but\nthis has not been sufficiently explored. One fundamental question is whether\nexisting T2I DMs are robust against variations over input texts. To answer it,\nthis work provides the first robustness evaluation of T2I DMs against\nreal-world attacks. Unlike prior studies that focus on malicious attacks\ninvolving apocryphal alterations to the input texts, we consider an attack\nspace spanned by realistic errors (e.g., typo, glyph, phonetic) that humans can\nmake, to ensure semantic consistency. Given the inherent randomness of the\ngeneration process, we develop novel distribution-based attack objectives to\nmislead T2I DMs. We perform attacks in a black-box manner without any knowledge\nof the model. Extensive experiments demonstrate the effectiveness of our method\nfor attacking popular T2I DMs and simultaneously reveal their non-trivial\nrobustness issues. Moreover, we provide an in-depth analysis of our method to\nshow that it is not designed to attack the text encoder in T2I DMs solely.', 'authors': ['Hongcheng Gao', 'Hao Zhang', 'Yinpeng Dong', 'Zhijie Deng'], 'published': '2023-06-16 00:43:35+00:00', 'categories': ['cs.CR', 'cs.AI', 'cs.CV', 'cs.LG'], 'url': 'http://arxiv.org/abs/2306.13103v1'}
42,550
2310.06743v2
2,310.06743
2
Geographic Location Encoding with Spherical Harmonics and Sinusoidal Representation Networks
Learning representations of geographical space is vital for any machine learning model that integrates geolocated data, spanning application domains such as remote sensing, ecology, or epidemiology. Recent work embeds coordinates using sine and cosine projections based on Double Fourier Sphere (DFS) features. These embeddings assume a rectangular data domain even on global data, which can lead to artifacts, especially at the poles. At the same time, little attention has been paid to the exact design of the neural network architectures with which these functional embeddings are combined. This work proposes a novel location encoder for globally distributed geographic data that combines spherical harmonic basis functions, natively defined on spherical surfaces, with sinusoidal representation networks (SirenNets) that can be interpreted as learned Double Fourier Sphere embedding. We systematically evaluate positional embeddings and neural network architectures across various benchmarks and synthetic evaluation datasets. In contrast to previous approaches that require the combination of both positional encoding and neural networks to learn meaningful representations, we show that both spherical harmonics and sinusoidal representation networks are competitive on their own but set state-of-the-art performances across tasks when combined. The model code and experiments are available at https://github.com/marccoru/locationencoder.
2023-10-10
{'id': '2310.06743v2', 'title': 'Geographic Location Encoding with Spherical Harmonics and Sinusoidal Representation Networks', 'abstract': 'Learning representations of geographical space is vital for any machine\nlearning model that integrates geolocated data, spanning application domains\nsuch as remote sensing, ecology, or epidemiology. Recent work embeds\ncoordinates using sine and cosine projections based on Double Fourier Sphere\n(DFS) features. These embeddings assume a rectangular data domain even on\nglobal data, which can lead to artifacts, especially at the poles. At the same\ntime, little attention has been paid to the exact design of the neural network\narchitectures with which these functional embeddings are combined. This work\nproposes a novel location encoder for globally distributed geographic data that\ncombines spherical harmonic basis functions, natively defined on spherical\nsurfaces, with sinusoidal representation networks (SirenNets) that can be\ninterpreted as learned Double Fourier Sphere embedding. We systematically\nevaluate positional embeddings and neural network architectures across various\nbenchmarks and synthetic evaluation datasets. In contrast to previous\napproaches that require the combination of both positional encoding and neural\nnetworks to learn meaningful representations, we show that both spherical\nharmonics and sinusoidal representation networks are competitive on their own\nbut set state-of-the-art performances across tasks when combined. The model\ncode and experiments are available at\nhttps://github.com/marccoru/locationencoder.', 'authors': ['Marc Rußwurm', 'Konstantin Klemmer', 'Esther Rolf', 'Robin Zbinden', 'Devis Tuia'], 'published': '2023-10-10 16:12:17+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2310.06743v2'}
42,566
2311.14645v2
2,311.14645
2
A General Framework for User-Guided Bayesian Optimization
The optimization of expensive-to-evaluate black-box functions is prevalent in various scientific disciplines. Bayesian optimization is an automatic, general and sample-efficient method to solve these problems with minimal knowledge of the underlying function dynamics. However, the ability of Bayesian optimization to incorporate prior knowledge or beliefs about the function at hand in order to accelerate the optimization is limited, which reduces its appeal for knowledgeable practitioners with tight budgets. To allow domain experts to customize the optimization routine, we propose ColaBO, the first Bayesian-principled framework for incorporating prior beliefs beyond the typical kernel structure, such as the likely location of the optimizer or the optimal value. The generality of ColaBO makes it applicable across different Monte Carlo acquisition functions and types of user beliefs. We empirically demonstrate ColaBO's ability to substantially accelerate optimization when the prior information is accurate, and to retain approximately default performance when it is misleading.
2023-11-24
{'id': '2311.14645v2', 'title': 'A General Framework for User-Guided Bayesian Optimization', 'abstract': "The optimization of expensive-to-evaluate black-box functions is prevalent in\nvarious scientific disciplines. Bayesian optimization is an automatic, general\nand sample-efficient method to solve these problems with minimal knowledge of\nthe underlying function dynamics. However, the ability of Bayesian optimization\nto incorporate prior knowledge or beliefs about the function at hand in order\nto accelerate the optimization is limited, which reduces its appeal for\nknowledgeable practitioners with tight budgets. To allow domain experts to\ncustomize the optimization routine, we propose ColaBO, the first\nBayesian-principled framework for incorporating prior beliefs beyond the\ntypical kernel structure, such as the likely location of the optimizer or the\noptimal value. The generality of ColaBO makes it applicable across different\nMonte Carlo acquisition functions and types of user beliefs. We empirically\ndemonstrate ColaBO's ability to substantially accelerate optimization when the\nprior information is accurate, and to retain approximately default performance\nwhen it is misleading.", 'authors': ['Carl Hvarfner', 'Frank Hutter', 'Luigi Nardi'], 'published': '2023-11-24 18:27:26+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2311.14645v2'}
42,574
2305.19838v4
2,305.19838
4
Relaxing the Additivity Constraints in Decentralized No-Regret High-Dimensional Bayesian Optimization
Bayesian Optimization (BO) is typically used to optimize an unknown function $f$ that is noisy and costly to evaluate, by exploiting an acquisition function that must be maximized at each optimization step. Even if provably asymptotically optimal BO algorithms are efficient at optimizing low-dimensional functions, scaling them to high-dimensional spaces remains an open problem, often tackled by assuming an additive structure for $f$. By doing so, BO algorithms typically introduce additional restrictive assumptions on the additive structure that reduce their applicability domain. This paper contains two main contributions: (i) we relax the restrictive assumptions on the additive structure of $f$ without weakening the maximization guarantees of the acquisition function, and (ii) we address the over-exploration problem for decentralized BO algorithms. To these ends, we propose DuMBO, an asymptotically optimal decentralized BO algorithm that achieves very competitive performance against state-of-the-art BO algorithms, especially when the additive structure of $f$ comprises high-dimensional factors.
2023-05-31
{'id': '2305.19838v4', 'title': 'Relaxing the Additivity Constraints in Decentralized No-Regret High-Dimensional Bayesian Optimization', 'abstract': 'Bayesian Optimization (BO) is typically used to optimize an unknown function\n$f$ that is noisy and costly to evaluate, by exploiting an acquisition function\nthat must be maximized at each optimization step. Even if provably\nasymptotically optimal BO algorithms are efficient at optimizing\nlow-dimensional functions, scaling them to high-dimensional spaces remains an\nopen problem, often tackled by assuming an additive structure for $f$. By doing\nso, BO algorithms typically introduce additional restrictive assumptions on the\nadditive structure that reduce their applicability domain. This paper contains\ntwo main contributions: (i) we relax the restrictive assumptions on the\nadditive structure of $f$ without weakening the maximization guarantees of the\nacquisition function, and (ii) we address the over-exploration problem for\ndecentralized BO algorithms. To these ends, we propose DuMBO, an asymptotically\noptimal decentralized BO algorithm that achieves very competitive performance\nagainst state-of-the-art BO algorithms, especially when the additive structure\nof $f$ comprises high-dimensional factors.', 'authors': ['Anthony Bardou', 'Patrick Thiran', 'Thomas Begin'], 'published': '2023-05-31 13:26:49+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2305.19838v4'}
42,581
2310.02207v3
2,310.02207
3
Language Models Represent Space and Time
The capabilities of large language models (LLMs) have sparked debate over whether such systems just learn an enormous collection of superficial statistics or a set of more coherent and grounded representations that reflect the real world. We find evidence for the latter by analyzing the learned representations of three spatial datasets (world, US, NYC places) and three temporal datasets (historical figures, artworks, news headlines) in the Llama-2 family of models. We discover that LLMs learn linear representations of space and time across multiple scales. These representations are robust to prompting variations and unified across different entity types (e.g. cities and landmarks). In addition, we identify individual "space neurons" and "time neurons" that reliably encode spatial and temporal coordinates. While further investigation is needed, our results suggest modern LLMs learn rich spatiotemporal representations of the real world and possess basic ingredients of a world model.
2023-10-03
{'id': '2310.02207v3', 'title': 'Language Models Represent Space and Time', 'abstract': 'The capabilities of large language models (LLMs) have sparked debate over\nwhether such systems just learn an enormous collection of superficial\nstatistics or a set of more coherent and grounded representations that reflect\nthe real world. We find evidence for the latter by analyzing the learned\nrepresentations of three spatial datasets (world, US, NYC places) and three\ntemporal datasets (historical figures, artworks, news headlines) in the Llama-2\nfamily of models. We discover that LLMs learn linear representations of space\nand time across multiple scales. These representations are robust to prompting\nvariations and unified across different entity types (e.g. cities and\nlandmarks). In addition, we identify individual "space neurons" and "time\nneurons" that reliably encode spatial and temporal coordinates. While further\ninvestigation is needed, our results suggest modern LLMs learn rich\nspatiotemporal representations of the real world and possess basic ingredients\nof a world model.', 'authors': ['Wes Gurnee', 'Max Tegmark'], 'published': '2023-10-03 17:06:52+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CL'], 'url': 'http://arxiv.org/abs/2310.02207v3'}
42,552
2311.02805v2
2,311.02805
2
Tailoring Self-Rationalizers with Multi-Reward Distillation
Large language models (LMs) are capable of generating free-text rationales to aid question answering. However, prior work 1) suggests that useful self-rationalization is emergent only at significant scales (e.g., 175B parameter GPT-3); and 2) focuses largely on downstream performance, ignoring the semantics of the rationales themselves, e.g., are they faithful, true, and helpful for humans? In this work, we enable small-scale LMs (approx. 200x smaller than GPT-3) to generate rationales that not only improve downstream task performance, but are also more plausible, consistent, and diverse, assessed both by automatic and human evaluation. Our method, MaRio (Multi-rewArd RatIOnalization), is a multi-reward conditioned self-rationalization algorithm that optimizes multiple distinct properties like plausibility, diversity and consistency. Results on five difficult question-answering datasets StrategyQA, QuaRel, OpenBookQA, NumerSense and QASC show that not only does MaRio improve task accuracy, but it also improves the self-rationalization quality of small LMs across the aforementioned axes better than a supervised fine-tuning (SFT) baseline. Extensive human evaluations confirm that MaRio rationales are preferred vs. SFT rationales, as well as qualitative improvements in plausibility and consistency.
2023-11-06
{'id': '2311.02805v2', 'title': 'Tailoring Self-Rationalizers with Multi-Reward Distillation', 'abstract': 'Large language models (LMs) are capable of generating free-text rationales to\naid question answering. However, prior work 1) suggests that useful\nself-rationalization is emergent only at significant scales (e.g., 175B\nparameter GPT-3); and 2) focuses largely on downstream performance, ignoring\nthe semantics of the rationales themselves, e.g., are they faithful, true, and\nhelpful for humans? In this work, we enable small-scale LMs (approx. 200x\nsmaller than GPT-3) to generate rationales that not only improve downstream\ntask performance, but are also more plausible, consistent, and diverse,\nassessed both by automatic and human evaluation. Our method, MaRio\n(Multi-rewArd RatIOnalization), is a multi-reward conditioned\nself-rationalization algorithm that optimizes multiple distinct properties like\nplausibility, diversity and consistency. Results on five difficult\nquestion-answering datasets StrategyQA, QuaRel, OpenBookQA, NumerSense and QASC\nshow that not only does MaRio improve task accuracy, but it also improves the\nself-rationalization quality of small LMs across the aforementioned axes better\nthan a supervised fine-tuning (SFT) baseline. Extensive human evaluations\nconfirm that MaRio rationales are preferred vs. SFT rationales, as well as\nqualitative improvements in plausibility and consistency.', 'authors': ['Sahana Ramnath', 'Brihi Joshi', 'Skyler Hallinan', 'Ximing Lu', 'Liunian Harold Li', 'Aaron Chan', 'Jack Hessel', 'Yejin Choi', 'Xiang Ren'], 'published': '2023-11-06 00:20:11+00:00', 'categories': ['cs.CL'], 'url': 'http://arxiv.org/abs/2311.02805v2'}
42,555
2307.02485v2
2,307.02485
2
Building Cooperative Embodied Agents Modularly with Large Language Models
In this work, we address challenging multi-agent cooperation problems with decentralized control, raw sensory observations, costly communication, and multi-objective tasks instantiated in various embodied environments. While previous research either presupposes a cost-free communication channel or relies on a centralized controller with shared observations, we harness the commonsense knowledge, reasoning ability, language comprehension, and text generation prowess of LLMs and seamlessly incorporate them into a cognitive-inspired modular framework that integrates with perception, memory, and execution. Thus building a Cooperative Embodied Language Agent CoELA, who can plan, communicate, and cooperate with others to accomplish long-horizon tasks efficiently. Our experiments on C-WAH and TDW-MAT demonstrate that CoELA driven by GPT-4 can surpass strong planning-based methods and exhibit emergent effective communication. Though current Open LMs like LLAMA-2 still underperform, we fine-tune a CoELA with data collected with our agents and show how they can achieve promising performance. We also conducted a user study for human-agent interaction and discovered that CoELA communicating in natural language can earn more trust and cooperate more effectively with humans. Our research underscores the potential of LLMs for future research in multi-agent cooperation. Videos can be found on the project website https://vis-www.cs.umass.edu/Co-LLM-Agents/.
2023-07-05
{'id': '2307.02485v2', 'title': 'Building Cooperative Embodied Agents Modularly with Large Language Models', 'abstract': 'In this work, we address challenging multi-agent cooperation problems with\ndecentralized control, raw sensory observations, costly communication, and\nmulti-objective tasks instantiated in various embodied environments. While\nprevious research either presupposes a cost-free communication channel or\nrelies on a centralized controller with shared observations, we harness the\ncommonsense knowledge, reasoning ability, language comprehension, and text\ngeneration prowess of LLMs and seamlessly incorporate them into a\ncognitive-inspired modular framework that integrates with perception, memory,\nand execution. Thus building a Cooperative Embodied Language Agent CoELA, who\ncan plan, communicate, and cooperate with others to accomplish long-horizon\ntasks efficiently. Our experiments on C-WAH and TDW-MAT demonstrate that CoELA\ndriven by GPT-4 can surpass strong planning-based methods and exhibit emergent\neffective communication. Though current Open LMs like LLAMA-2 still\nunderperform, we fine-tune a CoELA with data collected with our agents and show\nhow they can achieve promising performance. We also conducted a user study for\nhuman-agent interaction and discovered that CoELA communicating in natural\nlanguage can earn more trust and cooperate more effectively with humans. Our\nresearch underscores the potential of LLMs for future research in multi-agent\ncooperation. Videos can be found on the project website\nhttps://vis-www.cs.umass.edu/Co-LLM-Agents/.', 'authors': ['Hongxin Zhang', 'Weihua Du', 'Jiaming Shan', 'Qinhong Zhou', 'Yilun Du', 'Joshua B. Tenenbaum', 'Tianmin Shu', 'Chuang Gan'], 'published': '2023-07-05 17:59:27+00:00', 'categories': ['cs.AI', 'cs.CL', 'cs.CV', 'cs.RO'], 'url': 'http://arxiv.org/abs/2307.02485v2'}
42,559
2406.19579v1
2,406.19579
1
Private Zeroth-Order Nonsmooth Nonconvex Optimization
We introduce a new zeroth-order algorithm for private stochastic optimization on nonconvex and nonsmooth objectives. Given a dataset of size $M$, our algorithm ensures $(,^2/2)$-R\'enyi differential privacy and finds a $(,)$-stationary point so long as $M=(d^3 + d^3/2^2)$. This matches the optimal complexity of its non-private zeroth-order analog. Notably, although the objective is not smooth, we have privacy ``for free'' whenever $d$.
2024-06-27
{'id': '2406.19579v1', 'title': 'Private Zeroth-Order Nonsmooth Nonconvex Optimization', 'abstract': "We introduce a new zeroth-order algorithm for private stochastic optimization\non nonconvex and nonsmooth objectives. Given a dataset of size $M$, our\nalgorithm ensures $(\\alpha,\\alpha\\rho^2/2)$-R\\'enyi differential privacy and\nfinds a $(\\delta,\\epsilon)$-stationary point so long as\n$M=\\tilde\\Omega\\left(\\frac{d}{\\delta\\epsilon^3} +\n\\frac{d^{3/2}}{\\rho\\delta\\epsilon^2}\\right)$. This matches the optimal\ncomplexity of its non-private zeroth-order analog. Notably, although the\nobjective is not smooth, we have privacy ``for free'' whenever $\\rho \\ge\n\\sqrt{d}\\epsilon$.", 'authors': ['Qinzi Zhang', 'Hoang Tran', 'Ashok Cutkosky'], 'published': '2024-06-27 23:57:01+00:00', 'categories': ['math.OC', 'cs.CR', 'cs.LG'], 'url': 'http://arxiv.org/abs/2406.19579v1'}
42,561
2305.15371v2
2,305.15371
2
Stochastic Unrolled Federated Learning
Algorithm unrolling has emerged as a learning-based optimization paradigm that unfolds truncated iterative algorithms in trainable neural-network optimizers. We introduce Stochastic UnRolled Federated learning (SURF), a method that expands algorithm unrolling to federated learning in order to expedite its convergence. Our proposed method tackles two challenges of this expansion, namely the need to feed whole datasets to the unrolled optimizers to find a descent direction and the decentralized nature of federated learning. We circumvent the former challenge by feeding stochastic mini-batches to each unrolled layer and imposing descent constraints to guarantee its convergence. We address the latter challenge by unfolding the distributed gradient descent (DGD) algorithm in a graph neural network (GNN)-based unrolled architecture, which preserves the decentralized nature of training in federated learning. We theoretically prove that our proposed unrolled optimizer converges to a near-optimal region infinitely often. Through extensive numerical experiments, we also demonstrate the effectiveness of the proposed framework in collaborative training of image classifiers.
2023-05-24
{'id': '2305.15371v2', 'title': 'Stochastic Unrolled Federated Learning', 'abstract': 'Algorithm unrolling has emerged as a learning-based optimization paradigm\nthat unfolds truncated iterative algorithms in trainable neural-network\noptimizers. We introduce Stochastic UnRolled Federated learning (SURF), a\nmethod that expands algorithm unrolling to federated learning in order to\nexpedite its convergence. Our proposed method tackles two challenges of this\nexpansion, namely the need to feed whole datasets to the unrolled optimizers to\nfind a descent direction and the decentralized nature of federated learning. We\ncircumvent the former challenge by feeding stochastic mini-batches to each\nunrolled layer and imposing descent constraints to guarantee its convergence.\nWe address the latter challenge by unfolding the distributed gradient descent\n(DGD) algorithm in a graph neural network (GNN)-based unrolled architecture,\nwhich preserves the decentralized nature of training in federated learning. We\ntheoretically prove that our proposed unrolled optimizer converges to a\nnear-optimal region infinitely often. Through extensive numerical experiments,\nwe also demonstrate the effectiveness of the proposed framework in\ncollaborative training of image classifiers.', 'authors': ['Samar Hadou', 'Navid NaderiAlizadeh', 'Alejandro Ribeiro'], 'published': '2023-05-24 17:26:22+00:00', 'categories': ['cs.LG', 'eess.SP'], 'url': 'http://arxiv.org/abs/2305.15371v2'}
42,569
2310.04564v1
2,310.04564
1
ReLU Strikes Back: Exploiting Activation Sparsity in Large Language Models
Large Language Models (LLMs) with billions of parameters have drastically transformed AI applications. However, their demanding computation during inference has raised significant challenges for deployment on resource-constrained devices. Despite recent trends favoring alternative activation functions such as GELU or SiLU, known for increased computation, this study strongly advocates for reinstating ReLU activation in LLMs. We demonstrate that using the ReLU activation function has a negligible impact on convergence and performance while significantly reducing computation and weight transfer. This reduction is particularly valuable during the memory-bound inference step, where efficiency is paramount. Exploring sparsity patterns in ReLU-based LLMs, we unveil the reutilization of activated neurons for generating new tokens and leveraging these insights, we propose practical strategies to substantially reduce LLM inference computation up to three times, using ReLU activations with minimal performance trade-offs.
2023-10-06
{'id': '2310.04564v1', 'title': 'ReLU Strikes Back: Exploiting Activation Sparsity in Large Language Models', 'abstract': 'Large Language Models (LLMs) with billions of parameters have drastically\ntransformed AI applications. However, their demanding computation during\ninference has raised significant challenges for deployment on\nresource-constrained devices. Despite recent trends favoring alternative\nactivation functions such as GELU or SiLU, known for increased computation,\nthis study strongly advocates for reinstating ReLU activation in LLMs. We\ndemonstrate that using the ReLU activation function has a negligible impact on\nconvergence and performance while significantly reducing computation and weight\ntransfer. This reduction is particularly valuable during the memory-bound\ninference step, where efficiency is paramount. Exploring sparsity patterns in\nReLU-based LLMs, we unveil the reutilization of activated neurons for\ngenerating new tokens and leveraging these insights, we propose practical\nstrategies to substantially reduce LLM inference computation up to three times,\nusing ReLU activations with minimal performance trade-offs.', 'authors': ['Iman Mirzadeh', 'Keivan Alizadeh', 'Sachin Mehta', 'Carlo C Del Mundo', 'Oncel Tuzel', 'Golnoosh Samei', 'Mohammad Rastegari', 'Mehrdad Farajtabar'], 'published': '2023-10-06 20:01:33+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2310.04564v1'}
42,575
2403.10424v2
2,403.10424
2
Structured Evaluation of Synthetic Tabular Data
Tabular data is common yet typically incomplete, small in volume, and access-restricted due to privacy concerns. Synthetic data generation offers potential solutions. Many metrics exist for evaluating the quality of synthetic tabular data; however, we lack an objective, coherent interpretation of the many metrics. To address this issue, we propose an evaluation framework with a single, mathematical objective that posits that the synthetic data should be drawn from the same distribution as the observed data. Through various structural decomposition of the objective, this framework allows us to reason for the first time the completeness of any set of metrics, as well as unifies existing metrics, including those that stem from fidelity considerations, downstream application, and model-based approaches. Moreover, the framework motivates model-free baselines and a new spectrum of metrics. We evaluate structurally informed synthesizers and synthesizers powered by deep learning. Using our structured framework, we show that synthetic data generators that explicitly represent tabular structure outperform other methods, especially on smaller datasets.
2024-03-15
{'id': '2403.10424v2', 'title': 'Structured Evaluation of Synthetic Tabular Data', 'abstract': 'Tabular data is common yet typically incomplete, small in volume, and\naccess-restricted due to privacy concerns. Synthetic data generation offers\npotential solutions. Many metrics exist for evaluating the quality of synthetic\ntabular data; however, we lack an objective, coherent interpretation of the\nmany metrics. To address this issue, we propose an evaluation framework with a\nsingle, mathematical objective that posits that the synthetic data should be\ndrawn from the same distribution as the observed data. Through various\nstructural decomposition of the objective, this framework allows us to reason\nfor the first time the completeness of any set of metrics, as well as unifies\nexisting metrics, including those that stem from fidelity considerations,\ndownstream application, and model-based approaches. Moreover, the framework\nmotivates model-free baselines and a new spectrum of metrics. We evaluate\nstructurally informed synthesizers and synthesizers powered by deep learning.\nUsing our structured framework, we show that synthetic data generators that\nexplicitly represent tabular structure outperform other methods, especially on\nsmaller datasets.', 'authors': ['Scott Cheng-Hsin Yang', 'Baxter Eaves', 'Michael Schmidt', 'Ken Swanson', 'Patrick Shafto'], 'published': '2024-03-15 15:58:37+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2403.10424v2'}
42,564
2404.06280v2
2,404.0628
2
Algorithms for Caching and MTS with reduced number of predictions
ML-augmented algorithms utilize predictions to achieve performance beyond their worst-case bounds. Producing these predictions might be a costly operation -- this motivated Im et al. '22 to introduce the study of algorithms which use predictions parsimoniously. We design parsimonious algorithms for caching and MTS with action predictions, proposed by Antoniadis et al. '20, focusing on the parameters of consistency (performance with perfect predictions) and smoothness (dependence of their performance on the prediction error). Our algorithm for caching is 1-consistent, robust, and its smoothness deteriorates with the decreasing number of available predictions. We propose an algorithm for general MTS whose consistency and smoothness both scale linearly with the decreasing number of predictions. Without the restriction on the number of available predictions, both algorithms match the earlier guarantees achieved by Antoniadis et al. '20.
2024-04-09
{'id': '2404.06280v2', 'title': 'Algorithms for Caching and MTS with reduced number of predictions', 'abstract': "ML-augmented algorithms utilize predictions to achieve performance beyond\ntheir worst-case bounds. Producing these predictions might be a costly\noperation -- this motivated Im et al. '22 to introduce the study of algorithms\nwhich use predictions parsimoniously. We design parsimonious algorithms for\ncaching and MTS with action predictions, proposed by Antoniadis et al. '20,\nfocusing on the parameters of consistency (performance with perfect\npredictions) and smoothness (dependence of their performance on the prediction\nerror). Our algorithm for caching is 1-consistent, robust, and its smoothness\ndeteriorates with the decreasing number of available predictions. We propose an\nalgorithm for general MTS whose consistency and smoothness both scale linearly\nwith the decreasing number of predictions. Without the restriction on the\nnumber of available predictions, both algorithms match the earlier guarantees\nachieved by Antoniadis et al. '20.", 'authors': ['Karim Abdel Sadek', 'Marek Elias'], 'published': '2024-04-09 13:02:40+00:00', 'categories': ['cs.LG', 'cs.DS'], 'url': 'http://arxiv.org/abs/2404.06280v2'}
42,577
2402.17205v3
2,402.17205
3
Measuring Vision-Language STEM Skills of Neural Models
We introduce a new challenge to test the STEM skills of neural models. The problems in the real world often require solutions, combining knowledge from STEM (science, technology, engineering, and math). Unlike existing datasets, our dataset requires the understanding of multimodal vision-language information of STEM. Our dataset features one of the largest and most comprehensive datasets for the challenge. It includes 448 skills and 1,073,146 questions spanning all STEM subjects. Compared to existing datasets that often focus on examining expert-level ability, our dataset includes fundamental skills and questions designed based on the K-12 curriculum. We also add state-of-the-art foundation models such as CLIP and GPT-3.5-Turbo to our benchmark. Results show that the recent model advances only help master a very limited number of lower grade-level skills (2.5% in the third grade) in our dataset. In fact, these models are still well below (averaging 54.7%) the performance of elementary students, not to mention near expert-level performance. To understand and increase the performance on our dataset, we teach the models on a training split of our dataset. Even though we observe improved performance, the model performance remains relatively low compared to average elementary students. To solve STEM problems, we will need novel algorithmic innovations from the community.
2024-02-27
{'id': '2402.17205v3', 'title': 'Measuring Vision-Language STEM Skills of Neural Models', 'abstract': 'We introduce a new challenge to test the STEM skills of neural models. The\nproblems in the real world often require solutions, combining knowledge from\nSTEM (science, technology, engineering, and math). Unlike existing datasets,\nour dataset requires the understanding of multimodal vision-language\ninformation of STEM. Our dataset features one of the largest and most\ncomprehensive datasets for the challenge. It includes 448 skills and 1,073,146\nquestions spanning all STEM subjects. Compared to existing datasets that often\nfocus on examining expert-level ability, our dataset includes fundamental\nskills and questions designed based on the K-12 curriculum. We also add\nstate-of-the-art foundation models such as CLIP and GPT-3.5-Turbo to our\nbenchmark. Results show that the recent model advances only help master a very\nlimited number of lower grade-level skills (2.5% in the third grade) in our\ndataset. In fact, these models are still well below (averaging 54.7%) the\nperformance of elementary students, not to mention near expert-level\nperformance. To understand and increase the performance on our dataset, we\nteach the models on a training split of our dataset. Even though we observe\nimproved performance, the model performance remains relatively low compared to\naverage elementary students. To solve STEM problems, we will need novel\nalgorithmic innovations from the community.', 'authors': ['Jianhao Shen', 'Ye Yuan', 'Srbuhi Mirzoyan', 'Ming Zhang', 'Chenguang Wang'], 'published': '2024-02-27 04:55:03+00:00', 'categories': ['cs.CL', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2402.17205v3'}
42,580
2310.06668v1
2,310.06668
1
Latent Diffusion Counterfactual Explanations
Counterfactual explanations have emerged as a promising method for elucidating the behavior of opaque black-box models. Recently, several works leveraged pixel-space diffusion models for counterfactual generation. To handle noisy, adversarial gradients during counterfactual generation -- causing unrealistic artifacts or mere adversarial perturbations -- they required either auxiliary adversarially robust models or computationally intensive guidance schemes. However, such requirements limit their applicability, e.g., in scenarios with restricted access to the model's training data. To address these limitations, we introduce Latent Diffusion Counterfactual Explanations (LDCE). LDCE harnesses the capabilities of recent class- or text-conditional foundation latent diffusion models to expedite counterfactual generation and focus on the important, semantic parts of the data. Furthermore, we propose a novel consensus guidance mechanism to filter out noisy, adversarial gradients that are misaligned with the diffusion model's implicit classifier. We demonstrate the versatility of LDCE across a wide spectrum of models trained on diverse datasets with different learning paradigms. Finally, we showcase how LDCE can provide insights into model errors, enhancing our understanding of black-box model behavior.
2023-10-10
{'id': '2310.06668v1', 'title': 'Latent Diffusion Counterfactual Explanations', 'abstract': "Counterfactual explanations have emerged as a promising method for\nelucidating the behavior of opaque black-box models. Recently, several works\nleveraged pixel-space diffusion models for counterfactual generation. To handle\nnoisy, adversarial gradients during counterfactual generation -- causing\nunrealistic artifacts or mere adversarial perturbations -- they required either\nauxiliary adversarially robust models or computationally intensive guidance\nschemes. However, such requirements limit their applicability, e.g., in\nscenarios with restricted access to the model's training data. To address these\nlimitations, we introduce Latent Diffusion Counterfactual Explanations (LDCE).\nLDCE harnesses the capabilities of recent class- or text-conditional foundation\nlatent diffusion models to expedite counterfactual generation and focus on the\nimportant, semantic parts of the data. Furthermore, we propose a novel\nconsensus guidance mechanism to filter out noisy, adversarial gradients that\nare misaligned with the diffusion model's implicit classifier. We demonstrate\nthe versatility of LDCE across a wide spectrum of models trained on diverse\ndatasets with different learning paradigms. Finally, we showcase how LDCE can\nprovide insights into model errors, enhancing our understanding of black-box\nmodel behavior.", 'authors': ['Karim Farid', 'Simon Schrodi', 'Max Argus', 'Thomas Brox'], 'published': '2023-10-10 14:42:34+00:00', 'categories': ['cs.LG', 'cs.CV'], 'url': 'http://arxiv.org/abs/2310.06668v1'}
42,585
2311.18207v3
2,311.18207
3
Towards Assessing and Benchmarking Risk-Return Tradeoff of Off-Policy Evaluation
Off-Policy Evaluation (OPE) aims to assess the effectiveness of counterfactual policies using only offline logged data and is often used to identify the top-k promising policies for deployment in online A/B tests. Existing evaluation metrics for OPE estimators primarily focus on the "accuracy" of OPE or that of downstream policy selection, neglecting risk-return tradeoff in the subsequent online policy deployment. To address this issue, we draw inspiration from portfolio evaluation in finance and develop a new metric, called SharpeRatio@k, which measures the risk-return tradeoff of policy portfolios formed by an OPE estimator under varying online evaluation budgets (k). We validate our metric in two example scenarios, demonstrating its ability to effectively distinguish between low-risk and high-risk estimators and to accurately identify the most efficient one. Efficiency of an estimator is characterized by its capability to form the most advantageous policy portfolios, maximizing returns while minimizing risks during online deployment, a nuance that existing metrics typically overlook. To facilitate a quick, accurate, and consistent evaluation of OPE via SharpeRatio@k, we have also integrated this metric into an open-source software, SCOPE-RL (https://github.com/hakuhodo-technologies/scope-rl). Employing SharpeRatio@k and SCOPE-RL, we conduct comprehensive benchmarking experiments on various estimators and RL tasks, focusing on their risk-return tradeoff. These experiments offer several interesting directions and suggestions for future OPE research.
2023-11-30
{'id': '2311.18207v3', 'title': 'Towards Assessing and Benchmarking Risk-Return Tradeoff of Off-Policy Evaluation', 'abstract': 'Off-Policy Evaluation (OPE) aims to assess the effectiveness of\ncounterfactual policies using only offline logged data and is often used to\nidentify the top-k promising policies for deployment in online A/B tests.\nExisting evaluation metrics for OPE estimators primarily focus on the\n"accuracy" of OPE or that of downstream policy selection, neglecting\nrisk-return tradeoff in the subsequent online policy deployment. To address\nthis issue, we draw inspiration from portfolio evaluation in finance and\ndevelop a new metric, called SharpeRatio@k, which measures the risk-return\ntradeoff of policy portfolios formed by an OPE estimator under varying online\nevaluation budgets (k). We validate our metric in two example scenarios,\ndemonstrating its ability to effectively distinguish between low-risk and\nhigh-risk estimators and to accurately identify the most efficient one.\nEfficiency of an estimator is characterized by its capability to form the most\nadvantageous policy portfolios, maximizing returns while minimizing risks\nduring online deployment, a nuance that existing metrics typically overlook. To\nfacilitate a quick, accurate, and consistent evaluation of OPE via\nSharpeRatio@k, we have also integrated this metric into an open-source\nsoftware, SCOPE-RL (https://github.com/hakuhodo-technologies/scope-rl).\nEmploying SharpeRatio@k and SCOPE-RL, we conduct comprehensive benchmarking\nexperiments on various estimators and RL tasks, focusing on their risk-return\ntradeoff. These experiments offer several interesting directions and\nsuggestions for future OPE research.', 'authors': ['Haruka Kiyohara', 'Ren Kishimoto', 'Kosuke Kawakami', 'Ken Kobayashi', 'Kazuhide Nakata', 'Yuta Saito'], 'published': '2023-11-30 02:56:49+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2311.18207v3'}
42,592
2407.00806v1
2,407.00806
1
Benchmarks for Reinforcement Learning with Biased Offline Data and Imperfect Simulators
In many reinforcement learning (RL) applications one cannot easily let the agent act in the world; this is true for autonomous vehicles, healthcare applications, and even some recommender systems, to name a few examples. Offline RL provides a way to train agents without real-world exploration, but is often faced with biases due to data distribution shifts, limited coverage, and incomplete representation of the environment. To address these issues, practical applications have tried to combine simulators with grounded offline data, using so-called hybrid methods. However, constructing a reliable simulator is in itself often challenging due to intricate system complexities as well as missing or incomplete information. In this work, we outline four principal challenges for combining offline data with imperfect simulators in RL: simulator modeling error, partial observability, state and action discrepancies, and hidden confounding. To help drive the RL community to pursue these problems, we construct ``Benchmarks for Mechanistic Offline Reinforcement Learning'' (B4MRL), which provide dataset-simulator benchmarks for the aforementioned challenges. Our results suggest the key necessity of such benchmarks for future research.
2024-06-30
{'id': '2407.00806v1', 'title': 'Benchmarks for Reinforcement Learning with Biased Offline Data and Imperfect Simulators', 'abstract': "In many reinforcement learning (RL) applications one cannot easily let the\nagent act in the world; this is true for autonomous vehicles, healthcare\napplications, and even some recommender systems, to name a few examples.\nOffline RL provides a way to train agents without real-world exploration, but\nis often faced with biases due to data distribution shifts, limited coverage,\nand incomplete representation of the environment. To address these issues,\npractical applications have tried to combine simulators with grounded offline\ndata, using so-called hybrid methods. However, constructing a reliable\nsimulator is in itself often challenging due to intricate system complexities\nas well as missing or incomplete information. In this work, we outline four\nprincipal challenges for combining offline data with imperfect simulators in\nRL: simulator modeling error, partial observability, state and action\ndiscrepancies, and hidden confounding. To help drive the RL community to pursue\nthese problems, we construct ``Benchmarks for Mechanistic Offline Reinforcement\nLearning'' (B4MRL), which provide dataset-simulator benchmarks for the\naforementioned challenges. Our results suggest the key necessity of such\nbenchmarks for future research.", 'authors': ['Ori Linial', 'Guy Tennenholtz', 'Uri Shalit'], 'published': '2024-06-30 19:22:59+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2407.00806v1'}
42,568
2310.04292v3
2,310.04292
3
Towards Foundational Models for Molecular Learning on Large-Scale Multi-Task Datasets
Recently, pre-trained foundation models have enabled significant advancements in multiple fields. In molecular machine learning, however, where datasets are often hand-curated, and hence typically small, the lack of datasets with labeled features, and codebases to manage those datasets, has hindered the development of foundation models. In this work, we present seven novel datasets categorized by size into three distinct categories: ToyMix, LargeMix and UltraLarge. These datasets push the boundaries in both the scale and the diversity of supervised labels for molecular learning. They cover nearly 100 million molecules and over 3000 sparsely defined tasks, totaling more than 13 billion individual labels of both quantum and biological nature. In comparison, our datasets contain 300 times more data points than the widely used OGB-LSC PCQM4Mv2 dataset, and 13 times more than the quantum-only QM1B dataset. In addition, to support the development of foundational models based on our proposed datasets, we present the Graphium graph machine learning library which simplifies the process of building and training molecular machine learning models for multi-task and multi-level molecular datasets. Finally, we present a range of baseline results as a starting point of multi-task and multi-level training on these datasets. Empirically, we observe that performance on low-resource biological datasets show improvement by also training on large amounts of quantum data. This indicates that there may be potential in multi-task and multi-level training of a foundation model and fine-tuning it to resource-constrained downstream tasks.
2023-10-06
{'id': '2310.04292v3', 'title': 'Towards Foundational Models for Molecular Learning on Large-Scale Multi-Task Datasets', 'abstract': 'Recently, pre-trained foundation models have enabled significant advancements\nin multiple fields. In molecular machine learning, however, where datasets are\noften hand-curated, and hence typically small, the lack of datasets with\nlabeled features, and codebases to manage those datasets, has hindered the\ndevelopment of foundation models. In this work, we present seven novel datasets\ncategorized by size into three distinct categories: ToyMix, LargeMix and\nUltraLarge. These datasets push the boundaries in both the scale and the\ndiversity of supervised labels for molecular learning. They cover nearly 100\nmillion molecules and over 3000 sparsely defined tasks, totaling more than 13\nbillion individual labels of both quantum and biological nature. In comparison,\nour datasets contain 300 times more data points than the widely used OGB-LSC\nPCQM4Mv2 dataset, and 13 times more than the quantum-only QM1B dataset. In\naddition, to support the development of foundational models based on our\nproposed datasets, we present the Graphium graph machine learning library which\nsimplifies the process of building and training molecular machine learning\nmodels for multi-task and multi-level molecular datasets. Finally, we present a\nrange of baseline results as a starting point of multi-task and multi-level\ntraining on these datasets. Empirically, we observe that performance on\nlow-resource biological datasets show improvement by also training on large\namounts of quantum data. This indicates that there may be potential in\nmulti-task and multi-level training of a foundation model and fine-tuning it to\nresource-constrained downstream tasks.', 'authors': ['Dominique Beaini', 'Shenyang Huang', 'Joao Alex Cunha', 'Zhiyi Li', 'Gabriela Moisescu-Pareja', 'Oleksandr Dymov', 'Samuel Maddrell-Mander', 'Callum McLean', 'Frederik Wenkel', 'Luis Müller', 'Jama Hussein Mohamud', 'Ali Parviz', 'Michael Craig', 'Michał Koziarski', 'Jiarui Lu', 'Zhaocheng Zhu', 'Cristian Gabellini', 'Kerstin Klaser', 'Josef Dean', 'Cas Wognum', 'Maciej Sypetkowski', 'Guillaume Rabusseau', 'Reihaneh Rabbany', 'Jian Tang', 'Christopher Morris', 'Ioannis Koutis', 'Mirco Ravanelli', 'Guy Wolf', 'Prudencio Tossou', 'Hadrien Mary', 'Therence Bois', 'Andrew Fitzgibbon', 'Błażej Banaszewski', 'Chad Martin', 'Dominic Masters'], 'published': '2023-10-06 14:51:17+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.04292v3'}
42,570
2310.09675v2
2,310.09675
2
Efficient Model-Agnostic Multi-Group Equivariant Networks
Constructing model-agnostic group equivariant networks, such as equitune (Basu et al., 2023b) and its generalizations (Kim et al., 2023), can be computationally expensive for large product groups. We address this problem by providing efficient model-agnostic equivariant designs for two related problems: one where the network has multiple inputs each with potentially different groups acting on them, and another where there is a single input but the group acting on it is a large product group. For the first design, we initially consider a linear model and characterize the entire equivariant space that satisfies this constraint. This characterization gives rise to a novel fusion layer between different channels that satisfies an invariance-symmetry (IS) constraint, which we call an IS layer. We then extend this design beyond linear models, similar to equitune, consisting of equivariant and IS layers. We also show that the IS layer is a universal approximator of invariant-symmetric functions. Inspired by the first design, we use the notion of the IS property to design a second efficient model-agnostic equivariant design for large product groups acting on a single input. For the first design, we provide experiments on multi-image classification where each view is transformed independently with transformations such as rotations. We find equivariant models are robust to such transformations and perform competitively otherwise. For the second design, we consider three applications: language compositionality on the SCAN dataset to product groups; fairness in natural language generation from GPT-2 to address intersectionality; and robust zero-shot image classification with CLIP. Overall, our methods are simple and general, competitive with equitune and its variants, while also being computationally more efficient.
2023-10-14
{'id': '2310.09675v2', 'title': 'Efficient Model-Agnostic Multi-Group Equivariant Networks', 'abstract': 'Constructing model-agnostic group equivariant networks, such as equitune\n(Basu et al., 2023b) and its generalizations (Kim et al., 2023), can be\ncomputationally expensive for large product groups. We address this problem by\nproviding efficient model-agnostic equivariant designs for two related\nproblems: one where the network has multiple inputs each with potentially\ndifferent groups acting on them, and another where there is a single input but\nthe group acting on it is a large product group. For the first design, we\ninitially consider a linear model and characterize the entire equivariant space\nthat satisfies this constraint. This characterization gives rise to a novel\nfusion layer between different channels that satisfies an invariance-symmetry\n(IS) constraint, which we call an IS layer. We then extend this design beyond\nlinear models, similar to equitune, consisting of equivariant and IS layers. We\nalso show that the IS layer is a universal approximator of invariant-symmetric\nfunctions. Inspired by the first design, we use the notion of the IS property\nto design a second efficient model-agnostic equivariant design for large\nproduct groups acting on a single input. For the first design, we provide\nexperiments on multi-image classification where each view is transformed\nindependently with transformations such as rotations. We find equivariant\nmodels are robust to such transformations and perform competitively otherwise.\nFor the second design, we consider three applications: language\ncompositionality on the SCAN dataset to product groups; fairness in natural\nlanguage generation from GPT-2 to address intersectionality; and robust\nzero-shot image classification with CLIP. Overall, our methods are simple and\ngeneral, competitive with equitune and its variants, while also being\ncomputationally more efficient.', 'authors': ['Razan Baltaji', 'Sourya Basu', 'Lav R. Varshney'], 'published': '2023-10-14 22:24:26+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CL'], 'url': 'http://arxiv.org/abs/2310.09675v2'}
42,573
2403.07888v2
2,403.07888
2
Cross-modality debiasing: using language to mitigate sub-population shifts in imaging
Sub-population shift is a specific type of domain shift that highlights changes in data distribution within specific sub-groups or populations between training and testing. Sub-population shift accounts for a significant source of algorithmic bias and calls for distributional robustness. Recent studies found inherent distributional robustness in multi-modality foundation models, such as the vision-language model CLIP, yet this robustness is vulnerable through parameter fine-tuning. In this paper, we propose leveraging the connection of robustness among different modalities and reshaping the distributional robustness of one modality with another. Specifically, in the context of the distributional robustness of CLIP, we propose to leverage natural language inputs to debias the image feature representations, to improve worst-case performance on sub-populations. Our extensive empirical studies show that image representations debiased by natural language can achieve significant performance improvement and reduction of performance instability under sub-population shifts.
2024-02-02
{'id': '2403.07888v2', 'title': 'Cross-modality debiasing: using language to mitigate sub-population shifts in imaging', 'abstract': 'Sub-population shift is a specific type of domain shift that highlights\nchanges in data distribution within specific sub-groups or populations between\ntraining and testing. Sub-population shift accounts for a significant source of\nalgorithmic bias and calls for distributional robustness. Recent studies found\ninherent distributional robustness in multi-modality foundation models, such as\nthe vision-language model CLIP, yet this robustness is vulnerable through\nparameter fine-tuning. In this paper, we propose leveraging the connection of\nrobustness among different modalities and reshaping the distributional\nrobustness of one modality with another. Specifically, in the context of the\ndistributional robustness of CLIP, we propose to leverage natural language\ninputs to debias the image feature representations, to improve worst-case\nperformance on sub-populations. Our extensive empirical studies show that image\nrepresentations debiased by natural language can achieve significant\nperformance improvement and reduction of performance instability under\nsub-population shifts.', 'authors': ['Yijiang Pang', 'Bao Hoang', 'Jiayu Zhou'], 'published': '2024-02-02 18:54:48+00:00', 'categories': ['cs.CV', 'cs.AI'], 'url': 'http://arxiv.org/abs/2403.07888v2'}
42,576
2403.19963v1
2,403.19963
1
Efficient Modulation for Vision Networks
In this work, we present efficient modulation, a novel design for efficient vision networks. We revisit the modulation mechanism, which operates input through convolutional context modeling and feature projection layers, and fuses features via element-wise multiplication and an MLP block. We demonstrate that the modulation mechanism is particularly well suited for efficient networks and further tailor the modulation design by proposing the efficient modulation (EfficientMod) block, which is considered the essential building block for our networks. Benefiting from the prominent representational ability of modulation mechanism and the proposed efficient design, our network can accomplish better trade-offs between accuracy and efficiency and set new state-of-the-art performance in the zoo of efficient networks. When integrating EfficientMod with the vanilla self-attention block, we obtain the hybrid architecture which further improves the performance without loss of efficiency. We carry out comprehensive experiments to verify EfficientMod's performance. With fewer parameters, our EfficientMod-s performs 0.6 top-1 accuracy better than EfficientFormerV2-s2 and is 25% faster on GPU, and 2.9 better than MobileViTv2-1.0 at the same GPU latency. Additionally, our method presents a notable improvement in downstream tasks, outperforming EfficientFormerV2-s by 3.6 mIoU on the ADE20K benchmark. Code and checkpoints are available at https://github.com/ma-xu/EfficientMod.
2024-03-29
{'id': '2403.19963v1', 'title': 'Efficient Modulation for Vision Networks', 'abstract': "In this work, we present efficient modulation, a novel design for efficient\nvision networks. We revisit the modulation mechanism, which operates input\nthrough convolutional context modeling and feature projection layers, and fuses\nfeatures via element-wise multiplication and an MLP block. We demonstrate that\nthe modulation mechanism is particularly well suited for efficient networks and\nfurther tailor the modulation design by proposing the efficient modulation\n(EfficientMod) block, which is considered the essential building block for our\nnetworks. Benefiting from the prominent representational ability of modulation\nmechanism and the proposed efficient design, our network can accomplish better\ntrade-offs between accuracy and efficiency and set new state-of-the-art\nperformance in the zoo of efficient networks. When integrating EfficientMod\nwith the vanilla self-attention block, we obtain the hybrid architecture which\nfurther improves the performance without loss of efficiency. We carry out\ncomprehensive experiments to verify EfficientMod's performance. With fewer\nparameters, our EfficientMod-s performs 0.6 top-1 accuracy better than\nEfficientFormerV2-s2 and is 25% faster on GPU, and 2.9 better than\nMobileViTv2-1.0 at the same GPU latency. Additionally, our method presents a\nnotable improvement in downstream tasks, outperforming EfficientFormerV2-s by\n3.6 mIoU on the ADE20K benchmark. Code and checkpoints are available at\nhttps://github.com/ma-xu/EfficientMod.", 'authors': ['Xu Ma', 'Xiyang Dai', 'Jianwei Yang', 'Bin Xiao', 'Yinpeng Chen', 'Yun Fu', 'Lu Yuan'], 'published': '2024-03-29 03:48:35+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2403.19963v1'}
42,579
2310.12962v1
2,310.12962
1
An Emulator for Fine-Tuning Large Language Models using Small Language Models
Widely used language models (LMs) are typically built by scaling up a two-stage training pipeline: a pre-training stage that uses a very large, diverse dataset of text and a fine-tuning (sometimes, 'alignment') stage that uses targeted examples or other specifications of desired behaviors. While it has been hypothesized that knowledge and skills come from pre-training, and fine-tuning mostly filters this knowledge and skillset, this intuition has not been extensively tested. To aid in doing so, we introduce a novel technique for decoupling the knowledge and skills gained in these two stages, enabling a direct answer to the question, "What would happen if we combined the knowledge learned by a large model during pre-training with the knowledge learned by a small model during fine-tuning (or vice versa)?" Using an RL-based framework derived from recent developments in learning from human preferences, we introduce emulated fine-tuning (EFT), a principled and practical method for sampling from a distribution that approximates (or 'emulates') the result of pre-training and fine-tuning at different scales. Our experiments with EFT show that scaling up fine-tuning tends to improve helpfulness, while scaling up pre-training tends to improve factuality. Beyond decoupling scale, we show that EFT enables test-time adjustment of competing behavioral traits like helpfulness and harmlessness without additional training. Finally, a special case of emulated fine-tuning, which we call LM up-scaling, avoids resource-intensive fine-tuning of large pre-trained models by ensembling them with small fine-tuned models, essentially emulating the result of fine-tuning the large pre-trained model. Up-scaling consistently improves helpfulness and factuality of instruction-following models in the Llama, Llama-2, and Falcon families, without additional hyperparameters or training.
2023-10-19
{'id': '2310.12962v1', 'title': 'An Emulator for Fine-Tuning Large Language Models using Small Language Models', 'abstract': 'Widely used language models (LMs) are typically built by scaling up a\ntwo-stage training pipeline: a pre-training stage that uses a very large,\ndiverse dataset of text and a fine-tuning (sometimes, \'alignment\') stage that\nuses targeted examples or other specifications of desired behaviors. While it\nhas been hypothesized that knowledge and skills come from pre-training, and\nfine-tuning mostly filters this knowledge and skillset, this intuition has not\nbeen extensively tested. To aid in doing so, we introduce a novel technique for\ndecoupling the knowledge and skills gained in these two stages, enabling a\ndirect answer to the question, "What would happen if we combined the knowledge\nlearned by a large model during pre-training with the knowledge learned by a\nsmall model during fine-tuning (or vice versa)?" Using an RL-based framework\nderived from recent developments in learning from human preferences, we\nintroduce emulated fine-tuning (EFT), a principled and practical method for\nsampling from a distribution that approximates (or \'emulates\') the result of\npre-training and fine-tuning at different scales. Our experiments with EFT show\nthat scaling up fine-tuning tends to improve helpfulness, while scaling up\npre-training tends to improve factuality. Beyond decoupling scale, we show that\nEFT enables test-time adjustment of competing behavioral traits like\nhelpfulness and harmlessness without additional training. Finally, a special\ncase of emulated fine-tuning, which we call LM up-scaling, avoids\nresource-intensive fine-tuning of large pre-trained models by ensembling them\nwith small fine-tuned models, essentially emulating the result of fine-tuning\nthe large pre-trained model. Up-scaling consistently improves helpfulness and\nfactuality of instruction-following models in the Llama, Llama-2, and Falcon\nfamilies, without additional hyperparameters or training.', 'authors': ['Eric Mitchell', 'Rafael Rafailov', 'Archit Sharma', 'Chelsea Finn', 'Christopher D. Manning'], 'published': '2023-10-19 17:57:16+00:00', 'categories': ['cs.CL', 'cs.AI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2310.12962v1'}
42,571
2405.09999v2
2,405.09999
2
Reward Centering
We show that discounted methods for solving continuing reinforcement learning problems can perform significantly better if they center their rewards by subtracting out the rewards' empirical average. The improvement is substantial at commonly used discount factors and increases further as the discount factor approaches one. In addition, we show that if a problem's rewards are shifted by a constant, then standard methods perform much worse, whereas methods with reward centering are unaffected. Estimating the average reward is straightforward in the on-policy setting; we propose a slightly more sophisticated method for the off-policy setting. Reward centering is a general idea, so we expect almost every reinforcement-learning algorithm to benefit by the addition of reward centering.
2024-05-16
{'id': '2405.09999v2', 'title': 'Reward Centering', 'abstract': "We show that discounted methods for solving continuing reinforcement learning\nproblems can perform significantly better if they center their rewards by\nsubtracting out the rewards' empirical average. The improvement is substantial\nat commonly used discount factors and increases further as the discount factor\napproaches one. In addition, we show that if a problem's rewards are shifted by\na constant, then standard methods perform much worse, whereas methods with\nreward centering are unaffected. Estimating the average reward is\nstraightforward in the on-policy setting; we propose a slightly more\nsophisticated method for the off-policy setting. Reward centering is a general\nidea, so we expect almost every reinforcement-learning algorithm to benefit by\nthe addition of reward centering.", 'authors': ['Abhishek Naik', 'Yi Wan', 'Manan Tomar', 'Richard S. Sutton'], 'published': '2024-05-16 11:33:49+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2405.09999v2'}
42,582
2305.14632v1
2,305.14632
1
Supermodular Rank: Set Function Decomposition and Optimization
We define the supermodular rank of a function on a lattice. This is the smallest number of terms needed to decompose it into a sum of supermodular functions. The supermodular summands are defined with respect to different partial orders. We characterize the maximum possible value of the supermodular rank and describe the functions with fixed supermodular rank. We analogously define the submodular rank. We use submodular decompositions to optimize set functions. Given a bound on the submodular rank of a set function, we formulate an algorithm that splits an optimization problem into submodular subproblems. We show that this method improves the approximation ratio guarantees of several algorithms for monotone set function maximization and ratio of set functions minimization, at a computation overhead that depends on the submodular rank.
2023-05-24
{'id': '2305.14632v1', 'title': 'Supermodular Rank: Set Function Decomposition and Optimization', 'abstract': 'We define the supermodular rank of a function on a lattice. This is the\nsmallest number of terms needed to decompose it into a sum of supermodular\nfunctions. The supermodular summands are defined with respect to different\npartial orders. We characterize the maximum possible value of the supermodular\nrank and describe the functions with fixed supermodular rank. We analogously\ndefine the submodular rank. We use submodular decompositions to optimize set\nfunctions. Given a bound on the submodular rank of a set function, we formulate\nan algorithm that splits an optimization problem into submodular subproblems.\nWe show that this method improves the approximation ratio guarantees of several\nalgorithms for monotone set function maximization and ratio of set functions\nminimization, at a computation overhead that depends on the submodular rank.', 'authors': ['Rishi Sonthalia', 'Anna Seigal', 'Guido Montufar'], 'published': '2023-05-24 02:09:28+00:00', 'categories': ['math.CO', 'cs.CC', 'cs.DM', 'cs.LG', 'math.OC'], 'url': 'http://arxiv.org/abs/2305.14632v1'}
42,584
2311.13628v2
2,311.13628
2
Prompt Risk Control: A Rigorous Framework for Responsible Deployment of Large Language Models
The recent explosion in the capabilities of large language models has led to a wave of interest in how best to prompt a model to perform a given task. While it may be tempting to simply choose a prompt based on average performance on a validation set, this can lead to a deployment where unexpectedly poor responses are generated, especially for the worst-off users. To mitigate this prospect, we propose Prompt Risk Control, a lightweight framework for selecting a prompt based on rigorous upper bounds on families of informative risk measures. We offer methods for producing bounds on a diverse set of metrics, including quantities that measure worst-case responses and disparities in generation quality across the population of users. In addition, we extend the underlying statistical bounding techniques to accommodate the possibility of distribution shifts in deployment. Experiments on applications such as open-ended chat, medical question summarization, and code generation highlight how such a framework can foster responsible deployment by reducing the risk of the worst outcomes.
2023-11-22
{'id': '2311.13628v2', 'title': 'Prompt Risk Control: A Rigorous Framework for Responsible Deployment of Large Language Models', 'abstract': 'The recent explosion in the capabilities of large language models has led to\na wave of interest in how best to prompt a model to perform a given task. While\nit may be tempting to simply choose a prompt based on average performance on a\nvalidation set, this can lead to a deployment where unexpectedly poor responses\nare generated, especially for the worst-off users. To mitigate this prospect,\nwe propose Prompt Risk Control, a lightweight framework for selecting a prompt\nbased on rigorous upper bounds on families of informative risk measures. We\noffer methods for producing bounds on a diverse set of metrics, including\nquantities that measure worst-case responses and disparities in generation\nquality across the population of users. In addition, we extend the underlying\nstatistical bounding techniques to accommodate the possibility of distribution\nshifts in deployment. Experiments on applications such as open-ended chat,\nmedical question summarization, and code generation highlight how such a\nframework can foster responsible deployment by reducing the risk of the worst\noutcomes.', 'authors': ['Thomas P. Zollo', 'Todd Morrill', 'Zhun Deng', 'Jake C. Snell', 'Toniann Pitassi', 'Richard Zemel'], 'published': '2023-11-22 18:50:47+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CL'], 'url': 'http://arxiv.org/abs/2311.13628v2'}
42,587
2301.10352v2
2,301.10352
2
Capacity Analysis of Vector Symbolic Architectures
Hyperdimensional computing (HDC) is a biologically-inspired framework which represents symbols with high-dimensional vectors, and uses vector operations to manipulate them. The ensemble of a particular vector space and a prescribed set of vector operations (including one addition-like for "bundling" and one outer-product-like for "binding") form a *vector symbolic architecture* (VSA). While VSAs have been employed in numerous applications and have been studied empirically, many theoretical questions about VSAs remain open. We analyze the *representation capacities* of four common VSAs: MAP-I, MAP-B, and two VSAs based on sparse binary vectors. "Representation capacity' here refers to bounds on the dimensions of the VSA vectors required to perform certain symbolic tasks, such as testing for set membership $i S$ and estimating set intersection sizes $|X Y|$ for two sets of symbols $X$ and $Y$, to a given degree of accuracy. We also analyze the ability of a novel variant of a Hopfield network (a simple model of associative memory) to perform some of the same tasks that are typically asked of VSAs. In addition to providing new bounds on VSA capacities, our analyses establish and leverage connections between VSAs, "sketching" (dimensionality reduction) algorithms, and Bloom filters.
2023-01-24
{'id': '2301.10352v2', 'title': 'Capacity Analysis of Vector Symbolic Architectures', 'abstract': 'Hyperdimensional computing (HDC) is a biologically-inspired framework which\nrepresents symbols with high-dimensional vectors, and uses vector operations to\nmanipulate them. The ensemble of a particular vector space and a prescribed set\nof vector operations (including one addition-like for "bundling" and one\nouter-product-like for "binding") form a *vector symbolic architecture* (VSA).\nWhile VSAs have been employed in numerous applications and have been studied\nempirically, many theoretical questions about VSAs remain open. We analyze the\n*representation capacities* of four common VSAs: MAP-I, MAP-B, and two VSAs\nbased on sparse binary vectors. "Representation capacity\' here refers to bounds\non the dimensions of the VSA vectors required to perform certain symbolic\ntasks, such as testing for set membership $i \\in S$ and estimating set\nintersection sizes $|X \\cap Y|$ for two sets of symbols $X$ and $Y$, to a given\ndegree of accuracy. We also analyze the ability of a novel variant of a\nHopfield network (a simple model of associative memory) to perform some of the\nsame tasks that are typically asked of VSAs. In addition to providing new\nbounds on VSA capacities, our analyses establish and leverage connections\nbetween VSAs, "sketching" (dimensionality reduction) algorithms, and Bloom\nfilters.', 'authors': ['Kenneth L. Clarkson', 'Shashanka Ubaru', 'Elizabeth Yang'], 'published': '2023-01-24 23:43:25+00:00', 'categories': ['cs.LG', 'cs.DS'], 'url': 'http://arxiv.org/abs/2301.10352v2'}
42,593
2407.01776v1
2,407.01776
1
Federated Binary Matrix Factorization using Proximal Optimization
Identifying informative components in binary data is an essential task in many research areas, including life sciences, social sciences, and recommendation systems. Boolean matrix factorization (BMF) is a family of methods that performs this task by efficiently factorizing the data. In real-world settings, the data is often distributed across stakeholders and required to stay private, prohibiting the straightforward application of BMF. To adapt BMF to this context, we approach the problem from a federated-learning perspective, while building on a state-of-the-art continuous binary matrix factorization relaxation to BMF that enables efficient gradient-based optimization. We propose to only share the relaxed component matrices, which are aggregated centrally using a proximal operator that regularizes for binary outcomes. We show the convergence of our federated proximal gradient descent algorithm and provide differential privacy guarantees. Our extensive empirical evaluation demonstrates that our algorithm outperforms, in terms of quality and efficacy, federation schemes of state-of-the-art BMF methods on a diverse set of real-world and synthetic data.
2024-07-01
{'id': '2407.01776v1', 'title': 'Federated Binary Matrix Factorization using Proximal Optimization', 'abstract': 'Identifying informative components in binary data is an essential task in\nmany research areas, including life sciences, social sciences, and\nrecommendation systems. Boolean matrix factorization (BMF) is a family of\nmethods that performs this task by efficiently factorizing the data. In\nreal-world settings, the data is often distributed across stakeholders and\nrequired to stay private, prohibiting the straightforward application of BMF.\nTo adapt BMF to this context, we approach the problem from a federated-learning\nperspective, while building on a state-of-the-art continuous binary matrix\nfactorization relaxation to BMF that enables efficient gradient-based\noptimization. We propose to only share the relaxed component matrices, which\nare aggregated centrally using a proximal operator that regularizes for binary\noutcomes. We show the convergence of our federated proximal gradient descent\nalgorithm and provide differential privacy guarantees. Our extensive empirical\nevaluation demonstrates that our algorithm outperforms, in terms of quality and\nefficacy, federation schemes of state-of-the-art BMF methods on a diverse set\nof real-world and synthetic data.', 'authors': ['Sebastian Dalleiger', 'Jilles Vreeken', 'Michael Kamp'], 'published': '2024-07-01 20:10:24+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2407.01776v1'}
42,578
2310.01382v2
2,310.01382
2
Compressing LLMs: The Truth is Rarely Pure and Never Simple
Despite their remarkable achievements, modern Large Language Models (LLMs) face exorbitant computational and memory footprints. Recently, several works have shown significant success in training-free and data-free compression (pruning and quantization) of LLMs that achieve 50 - 60% sparsity and reduce the bit width to 3 or 4 bits per weight, with negligible degradation of perplexity over the uncompressed baseline. As recent research efforts are focused on developing increasingly sophisticated compression methods, our work takes a step back and re-evaluates the effectiveness of existing SoTA compression methods, which rely on a fairly simple and widely questioned metric, perplexity (even for dense LLMs). We introduce Knowledge-Intensive Compressed LLM BenchmarK (LLM-KICK), a collection of carefully curated tasks to redefine the evaluation protocol for compressed LLMs, which have significant alignment with their dense counterparts and perplexity fail to capture subtle change in their true capabilities. LLM-KICK unveils many favorable merits and unfortunate plights of current SoTA compression methods: all pruning methods suffer significant performance degradation, sometimes at trivial sparsity ratios (e.g., 25-30%), and fail for N:M sparsity in knowledge-intensive tasks; current quantization methods are more successful than pruning; yet, pruned LLMs even at $50$% sparsity are robust in-context retrieval and summarization systems; among others. LLM-KICK is designed to holistically access compressed LLMs' ability for language understanding, reasoning, generation, in-context retrieval, in-context summarization, etc. We hope our study can foster the development of better LLM compression methods. The reproduced codes are available at https://github.com/VITA-Group/llm-kick.
2023-10-02
{'id': '2310.01382v2', 'title': 'Compressing LLMs: The Truth is Rarely Pure and Never Simple', 'abstract': "Despite their remarkable achievements, modern Large Language Models (LLMs)\nface exorbitant computational and memory footprints. Recently, several works\nhave shown significant success in training-free and data-free compression\n(pruning and quantization) of LLMs that achieve 50 - 60% sparsity and reduce\nthe bit width to 3 or 4 bits per weight, with negligible degradation of\nperplexity over the uncompressed baseline. As recent research efforts are\nfocused on developing increasingly sophisticated compression methods, our work\ntakes a step back and re-evaluates the effectiveness of existing SoTA\ncompression methods, which rely on a fairly simple and widely questioned\nmetric, perplexity (even for dense LLMs). We introduce Knowledge-Intensive\nCompressed LLM BenchmarK (LLM-KICK), a collection of carefully curated tasks to\nredefine the evaluation protocol for compressed LLMs, which have significant\nalignment with their dense counterparts and perplexity fail to capture subtle\nchange in their true capabilities. LLM-KICK unveils many favorable merits and\nunfortunate plights of current SoTA compression methods: all pruning methods\nsuffer significant performance degradation, sometimes at trivial sparsity\nratios (e.g., 25-30%), and fail for N:M sparsity in knowledge-intensive tasks;\ncurrent quantization methods are more successful than pruning; yet, pruned LLMs\neven at $\\geq 50$% sparsity are robust in-context retrieval and summarization\nsystems; among others. LLM-KICK is designed to holistically access compressed\nLLMs' ability for language understanding, reasoning, generation, in-context\nretrieval, in-context summarization, etc. We hope our study can foster the\ndevelopment of better LLM compression methods. The reproduced codes are\navailable at https://github.com/VITA-Group/llm-kick.", 'authors': ['Ajay Jaiswal', 'Zhe Gan', 'Xianzhi Du', 'Bowen Zhang', 'Zhangyang Wang', 'Yinfei Yang'], 'published': '2023-10-02 17:42:37+00:00', 'categories': ['cs.CL', 'cs.LG'], 'url': 'http://arxiv.org/abs/2310.01382v2'}
42,583
2312.11529v4
2,312.11529
4
Efficient and Scalable Graph Generation through Iterative Local Expansion
In the realm of generative models for graphs, extensive research has been conducted. However, most existing methods struggle with large graphs due to the complexity of representing the entire joint distribution across all node pairs and capturing both global and local graph structures simultaneously. To overcome these issues, we introduce a method that generates a graph by progressively expanding a single node to a target graph. In each step, nodes and edges are added in a localized manner through denoising diffusion, building first the global structure, and then refining the local details. The local generation avoids modeling the entire joint distribution over all node pairs, achieving substantial computational savings with subquadratic runtime relative to node count while maintaining high expressivity through multiscale generation. Our experiments show that our model achieves state-of-the-art performance on well-established benchmark datasets while successfully scaling to graphs with at least 5000 nodes. Our method is also the first to successfully extrapolate to graphs outside of the training distribution, showcasing a much better generalization capability over existing methods.
2023-12-14
{'id': '2312.11529v4', 'title': 'Efficient and Scalable Graph Generation through Iterative Local Expansion', 'abstract': 'In the realm of generative models for graphs, extensive research has been\nconducted. However, most existing methods struggle with large graphs due to the\ncomplexity of representing the entire joint distribution across all node pairs\nand capturing both global and local graph structures simultaneously. To\novercome these issues, we introduce a method that generates a graph by\nprogressively expanding a single node to a target graph. In each step, nodes\nand edges are added in a localized manner through denoising diffusion, building\nfirst the global structure, and then refining the local details. The local\ngeneration avoids modeling the entire joint distribution over all node pairs,\nachieving substantial computational savings with subquadratic runtime relative\nto node count while maintaining high expressivity through multiscale\ngeneration. Our experiments show that our model achieves state-of-the-art\nperformance on well-established benchmark datasets while successfully scaling\nto graphs with at least 5000 nodes. Our method is also the first to\nsuccessfully extrapolate to graphs outside of the training distribution,\nshowcasing a much better generalization capability over existing methods.', 'authors': ['Andreas Bergmeister', 'Karolis Martinkus', 'Nathanaël Perraudin', 'Roger Wattenhofer'], 'published': '2023-12-14 10:42:42+00:00', 'categories': ['cs.SI', 'cs.LG'], 'url': 'http://arxiv.org/abs/2312.11529v4'}
42,589
2302.06354v3
2,302.06354
3
Less is More: Selective Layer Finetuning with SubTuning
Finetuning a pretrained model has become a standard approach for training neural networks on novel tasks, resulting in fast convergence and improved performance. In this work, we study an alternative finetuning method, where instead of finetuning all the weights of the network, we only train a carefully chosen subset of layers, keeping the rest of the weights frozen at their initial (pretrained) values. We demonstrate that subset finetuning (or SubTuning) often achieves accuracy comparable to full finetuning of the model, and even surpasses the performance of full finetuning when training data is scarce. Therefore, SubTuning allows deploying new tasks at minimal computational cost, while enjoying the benefits of finetuning the entire model. This yields a simple and effective method for multi-task learning, where different tasks do not interfere with one another, and yet share most of the resources at inference time. We demonstrate the efficiency of SubTuning across multiple tasks, using different network architectures and pretraining methods.
2023-02-13
{'id': '2302.06354v3', 'title': 'Less is More: Selective Layer Finetuning with SubTuning', 'abstract': 'Finetuning a pretrained model has become a standard approach for training\nneural networks on novel tasks, resulting in fast convergence and improved\nperformance. In this work, we study an alternative finetuning method, where\ninstead of finetuning all the weights of the network, we only train a carefully\nchosen subset of layers, keeping the rest of the weights frozen at their\ninitial (pretrained) values. We demonstrate that \\emph{subset finetuning} (or\nSubTuning) often achieves accuracy comparable to full finetuning of the model,\nand even surpasses the performance of full finetuning when training data is\nscarce. Therefore, SubTuning allows deploying new tasks at minimal\ncomputational cost, while enjoying the benefits of finetuning the entire model.\nThis yields a simple and effective method for multi-task learning, where\ndifferent tasks do not interfere with one another, and yet share most of the\nresources at inference time. We demonstrate the efficiency of SubTuning across\nmultiple tasks, using different network architectures and pretraining methods.', 'authors': ['Gal Kaplun', 'Andrey Gurevich', 'Tal Swisa', 'Mazor David', 'Shai Shalev-Shwartz', 'Eran Malach'], 'published': '2023-02-13 13:38:46+00:00', 'categories': ['cs.LG', 'cs.AI'], 'url': 'http://arxiv.org/abs/2302.06354v3'}
42,596
2305.08960v2
2,305.0896
2
One Forward is Enough for Neural Network Training via Likelihood Ratio Method
While backpropagation (BP) is the mainstream approach for gradient computation in neural network training, its heavy reliance on the chain rule of differentiation constrains the designing flexibility of network architecture and training pipelines. We avoid the recursive computation in BP and develop a unified likelihood ratio (ULR) method for gradient estimation with just one forward propagation. Not only can ULR be extended to train a wide variety of neural network architectures, but the computation flow in BP can also be rearranged by ULR for better device adaptation. Moreover, we propose several variance reduction techniques to further accelerate the training process. Our experiments offer numerical results across diverse aspects, including various neural network training scenarios, computation flow rearrangement, and fine-tuning of pre-trained models. All findings demonstrate that ULR effectively enhances the flexibility of neural network training by permitting localized module training without compromising the global objective and significantly boosts the network robustness.
2023-05-15
{'id': '2305.08960v2', 'title': 'One Forward is Enough for Neural Network Training via Likelihood Ratio Method', 'abstract': 'While backpropagation (BP) is the mainstream approach for gradient\ncomputation in neural network training, its heavy reliance on the chain rule of\ndifferentiation constrains the designing flexibility of network architecture\nand training pipelines. We avoid the recursive computation in BP and develop a\nunified likelihood ratio (ULR) method for gradient estimation with just one\nforward propagation. Not only can ULR be extended to train a wide variety of\nneural network architectures, but the computation flow in BP can also be\nrearranged by ULR for better device adaptation. Moreover, we propose several\nvariance reduction techniques to further accelerate the training process. Our\nexperiments offer numerical results across diverse aspects, including various\nneural network training scenarios, computation flow rearrangement, and\nfine-tuning of pre-trained models. All findings demonstrate that ULR\neffectively enhances the flexibility of neural network training by permitting\nlocalized module training without compromising the global objective and\nsignificantly boosts the network robustness.', 'authors': ['Jinyang Jiang', 'Zeliang Zhang', 'Chenliang Xu', 'Zhaofei Yu', 'Yijie Peng'], 'published': '2023-05-15 19:02:46+00:00', 'categories': ['cs.LG', 'cs.NE', 'math.OC'], 'url': 'http://arxiv.org/abs/2305.08960v2'}
42,598
2310.09297v2
2,310.09297
2
A Framework for Inference Inspired by Human Memory Mechanisms
How humans and machines make sense of current inputs for relation reasoning and question-answering while putting the perceived information into context of our past memories, has been a challenging conundrum in cognitive science and artificial intelligence. Inspired by human brain's memory system and cognitive architectures, we propose a PMI framework that consists of perception, memory and inference components. Notably, the memory module comprises working and long-term memory, with the latter endowed with a higher-order structure to retain extensive and complex relational knowledge and experience. Through a differentiable competitive write access, current perceptions update working memory, which is later merged with long-term memory via outer product associations, reducing information conflicts and averting memory overflow. In the inference module, relevant information is retrieved from two separate memory origins and associatively integrated to attain a more comprehensive and precise interpretation of current perceptions. We exploratively apply our PMI to improve prevailing Transformers and CNN models on question-answering tasks like bAbI-20k and Sort-of-CLEVR datasets, as well as detecting equilateral triangles, language modeling and image classification tasks, and in each case, our PMI enhancements consistently outshine their original counterparts significantly. Visualization analyses reveal that relational memory consolidation, along with the interaction and integration of information from diverse memory sources, substantially contributes to the model effectiveness on inference tasks.
2023-10-01
{'id': '2310.09297v2', 'title': 'A Framework for Inference Inspired by Human Memory Mechanisms', 'abstract': "How humans and machines make sense of current inputs for relation reasoning\nand question-answering while putting the perceived information into context of\nour past memories, has been a challenging conundrum in cognitive science and\nartificial intelligence. Inspired by human brain's memory system and cognitive\narchitectures, we propose a PMI framework that consists of perception, memory\nand inference components. Notably, the memory module comprises working and\nlong-term memory, with the latter endowed with a higher-order structure to\nretain extensive and complex relational knowledge and experience. Through a\ndifferentiable competitive write access, current perceptions update working\nmemory, which is later merged with long-term memory via outer product\nassociations, reducing information conflicts and averting memory overflow. In\nthe inference module, relevant information is retrieved from two separate\nmemory origins and associatively integrated to attain a more comprehensive and\nprecise interpretation of current perceptions. We exploratively apply our PMI\nto improve prevailing Transformers and CNN models on question-answering tasks\nlike bAbI-20k and Sort-of-CLEVR datasets, as well as detecting equilateral\ntriangles, language modeling and image classification tasks, and in each case,\nour PMI enhancements consistently outshine their original counterparts\nsignificantly. Visualization analyses reveal that relational memory\nconsolidation, along with the interaction and integration of information from\ndiverse memory sources, substantially contributes to the model effectiveness on\ninference tasks.", 'authors': ['Xiangyu Zeng', 'Jie Lin', 'Piao Hu', 'Ruizheng Huang', 'Zhicheng Zhang'], 'published': '2023-10-01 08:12:55+00:00', 'categories': ['cs.LG', 'cs.AI', 'cs.CL'], 'url': 'http://arxiv.org/abs/2310.09297v2'}
42,586
2003.01897v2
2,003.01897
2
Optimal Regularization Can Mitigate Double Descent
Recent empirical and theoretical studies have shown that many learning algorithms -- from linear regression to neural networks -- can have test performance that is non-monotonic in quantities such the sample size and model size. This striking phenomenon, often referred to as "double descent", has raised questions of if we need to re-think our current understanding of generalization. In this work, we study whether the double-descent phenomenon can be avoided by using optimal regularization. Theoretically, we prove that for certain linear regression models with isotropic data distribution, optimally-tuned $_2$ regularization achieves monotonic test performance as we grow either the sample size or the model size. We also demonstrate empirically that optimally-tuned $_2$ regularization can mitigate double descent for more general models, including neural networks. Our results suggest that it may also be informative to study the test risk scalings of various algorithms in the context of appropriately tuned regularization.
2020-03-04
{'id': '2003.01897v2', 'title': 'Optimal Regularization Can Mitigate Double Descent', 'abstract': 'Recent empirical and theoretical studies have shown that many learning\nalgorithms -- from linear regression to neural networks -- can have test\nperformance that is non-monotonic in quantities such the sample size and model\nsize. This striking phenomenon, often referred to as "double descent", has\nraised questions of if we need to re-think our current understanding of\ngeneralization. In this work, we study whether the double-descent phenomenon\ncan be avoided by using optimal regularization. Theoretically, we prove that\nfor certain linear regression models with isotropic data distribution,\noptimally-tuned $\\ell_2$ regularization achieves monotonic test performance as\nwe grow either the sample size or the model size. We also demonstrate\nempirically that optimally-tuned $\\ell_2$ regularization can mitigate double\ndescent for more general models, including neural networks. Our results suggest\nthat it may also be informative to study the test risk scalings of various\nalgorithms in the context of appropriately tuned regularization.', 'authors': ['Preetum Nakkiran', 'Prayaag Venkat', 'Sham Kakade', 'Tengyu Ma'], 'published': '2020-03-04 05:19:09+00:00', 'categories': ['cs.LG', 'cs.NE', 'math.ST', 'stat.ML', 'stat.TH'], 'url': 'http://arxiv.org/abs/2003.01897v2'}
42,588
2008.02598v3
2,008.02598
3
Offline Meta Learning of Exploration
Consider the following instance of the Offline Meta Reinforcement Learning (OMRL) problem: given the complete training logs of $N$ conventional RL agents, trained on $N$ different tasks, design a meta-agent that can quickly maximize reward in a new, unseen task from the same task distribution. In particular, while each conventional RL agent explored and exploited its own different task, the meta-agent must identify regularities in the data that lead to effective exploration/exploitation in the unseen task. Here, we take a Bayesian RL (BRL) view, and seek to learn a Bayes-optimal policy from the offline data. Building on the recent VariBAD BRL approach, we develop an off-policy BRL method that learns to plan an exploration strategy based on an adaptive neural belief estimate. However, learning to infer such a belief from offline data brings a new identifiability issue we term MDP ambiguity. We characterize the problem, and suggest resolutions via data collection and modification procedures. Finally, we evaluate our framework on a diverse set of domains, including difficult sparse reward tasks, and demonstrate learning of effective exploration behavior that is qualitatively different from the exploration used by any RL agent in the data.
2020-08-06
{'id': '2008.02598v3', 'title': 'Offline Meta Learning of Exploration', 'abstract': 'Consider the following instance of the Offline Meta Reinforcement Learning\n(OMRL) problem: given the complete training logs of $N$ conventional RL agents,\ntrained on $N$ different tasks, design a meta-agent that can quickly maximize\nreward in a new, unseen task from the same task distribution. In particular,\nwhile each conventional RL agent explored and exploited its own different task,\nthe meta-agent must identify regularities in the data that lead to effective\nexploration/exploitation in the unseen task. Here, we take a Bayesian RL (BRL)\nview, and seek to learn a Bayes-optimal policy from the offline data. Building\non the recent VariBAD BRL approach, we develop an off-policy BRL method that\nlearns to plan an exploration strategy based on an adaptive neural belief\nestimate. However, learning to infer such a belief from offline data brings a\nnew identifiability issue we term MDP ambiguity. We characterize the problem,\nand suggest resolutions via data collection and modification procedures.\nFinally, we evaluate our framework on a diverse set of domains, including\ndifficult sparse reward tasks, and demonstrate learning of effective\nexploration behavior that is qualitatively different from the exploration used\nby any RL agent in the data.', 'authors': ['Ron Dorfman', 'Idan Shenfeld', 'Aviv Tamar'], 'published': '2020-08-06 12:09:18+00:00', 'categories': ['cs.LG', 'cs.AI', 'stat.ML'], 'url': 'http://arxiv.org/abs/2008.02598v3'}
42,591
2006.07029v1
2,006.07029
1
Rethinking Sampling in 3D Point Cloud Generative Adversarial Networks
In this paper, we examine the long-neglected yet important effects of point sampling patterns in point cloud GANs. Through extensive experiments, we show that sampling-insensitive discriminators (e.g.PointNet-Max) produce shape point clouds with point clustering artifacts while sampling-oversensitive discriminators (e.g.PointNet++, DGCNN) fail to guide valid shape generation. We propose the concept of sampling spectrum to depict the different sampling sensitivities of discriminators. We further study how different evaluation metrics weigh the sampling pattern against the geometry and propose several perceptual metrics forming a sampling spectrum of metrics. Guided by the proposed sampling spectrum, we discover a middle-point sampling-aware baseline discriminator, PointNet-Mix, which improves all existing point cloud generators by a large margin on sampling-related metrics. We point out that, though recent research has been focused on the generator design, the main bottleneck of point cloud GAN actually lies in the discriminator design. Our work provides both suggestions and tools for building future discriminators. We will release the code to facilitate future research.
2020-06-12
{'id': '2006.07029v1', 'title': 'Rethinking Sampling in 3D Point Cloud Generative Adversarial Networks', 'abstract': 'In this paper, we examine the long-neglected yet important effects of point\nsampling patterns in point cloud GANs. Through extensive experiments, we show\nthat sampling-insensitive discriminators (e.g.PointNet-Max) produce shape point\nclouds with point clustering artifacts while sampling-oversensitive\ndiscriminators (e.g.PointNet++, DGCNN) fail to guide valid shape generation. We\npropose the concept of sampling spectrum to depict the different sampling\nsensitivities of discriminators. We further study how different evaluation\nmetrics weigh the sampling pattern against the geometry and propose several\nperceptual metrics forming a sampling spectrum of metrics. Guided by the\nproposed sampling spectrum, we discover a middle-point sampling-aware baseline\ndiscriminator, PointNet-Mix, which improves all existing point cloud generators\nby a large margin on sampling-related metrics. We point out that, though recent\nresearch has been focused on the generator design, the main bottleneck of point\ncloud GAN actually lies in the discriminator design. Our work provides both\nsuggestions and tools for building future discriminators. We will release the\ncode to facilitate future research.', 'authors': ['He Wang', 'Zetian Jiang', 'Li Yi', 'Kaichun Mo', 'Hao Su', 'Leonidas J. Guibas'], 'published': '2020-06-12 09:29:24+00:00', 'categories': ['cs.CV', 'cs.LG', 'eess.IV'], 'url': 'http://arxiv.org/abs/2006.07029v1'}
42,595
2006.00555v3
2,006.00555
3
Transferring Inductive Biases through Knowledge Distillation
Having the right inductive biases can be crucial in many tasks or scenarios where data or computing resources are a limiting factor, or where training data is not perfectly representative of the conditions at test time. However, defining, designing and efficiently adapting inductive biases is not necessarily straightforward. In this paper, we explore the power of knowledge distillation for transferring the effect of inductive biases from one model to another. We consider families of models with different inductive biases, LSTMs vs. Transformers and CNNs vs. MLPs, in the context of tasks and scenarios where having the right inductive biases is critical. We study the effect of inductive biases on the solutions the models converge to and investigate how and to what extent the effect of inductive biases is transferred through knowledge distillation, in terms of not only performance but also different aspects of converged solutions.
2020-05-31
{'id': '2006.00555v3', 'title': 'Transferring Inductive Biases through Knowledge Distillation', 'abstract': 'Having the right inductive biases can be crucial in many tasks or scenarios\nwhere data or computing resources are a limiting factor, or where training data\nis not perfectly representative of the conditions at test time. However,\ndefining, designing and efficiently adapting inductive biases is not\nnecessarily straightforward. In this paper, we explore the power of knowledge\ndistillation for transferring the effect of inductive biases from one model to\nanother. We consider families of models with different inductive biases, LSTMs\nvs. Transformers and CNNs vs. MLPs, in the context of tasks and scenarios where\nhaving the right inductive biases is critical. We study the effect of inductive\nbiases on the solutions the models converge to and investigate how and to what\nextent the effect of inductive biases is transferred through knowledge\ndistillation, in terms of not only performance but also different aspects of\nconverged solutions.', 'authors': ['Samira Abnar', 'Mostafa Dehghani', 'Willem Zuidema'], 'published': '2020-05-31 16:34:08+00:00', 'categories': ['cs.LG', 'cs.AI', 'stat.ML'], 'url': 'http://arxiv.org/abs/2006.00555v3'}
42,600
2002.03206v3
2,002.03206
3
Characterizing Structural Regularities of Labeled Data in Overparameterized Models
Humans are accustomed to environments that contain both regularities and exceptions. For example, at most gas stations, one pays prior to pumping, but the occasional rural station does not accept payment in advance. Likewise, deep neural networks can generalize across instances that share common patterns or structures, yet have the capacity to memorize rare or irregular forms. We analyze how individual instances are treated by a model via a consistency score. The score characterizes the expected accuracy for a held-out instance given training sets of varying size sampled from the data distribution. We obtain empirical estimates of this score for individual instances in multiple data sets, and we show that the score identifies out-of-distribution and mislabeled examples at one end of the continuum and strongly regular examples at the other end. We identify computationally inexpensive proxies to the consistency score using statistics collected during training. We show examples of potential applications to the analysis of deep-learning systems.
2020-02-08
{'id': '2002.03206v3', 'title': 'Characterizing Structural Regularities of Labeled Data in Overparameterized Models', 'abstract': 'Humans are accustomed to environments that contain both regularities and\nexceptions. For example, at most gas stations, one pays prior to pumping, but\nthe occasional rural station does not accept payment in advance. Likewise, deep\nneural networks can generalize across instances that share common patterns or\nstructures, yet have the capacity to memorize rare or irregular forms. We\nanalyze how individual instances are treated by a model via a consistency\nscore. The score characterizes the expected accuracy for a held-out instance\ngiven training sets of varying size sampled from the data distribution. We\nobtain empirical estimates of this score for individual instances in multiple\ndata sets, and we show that the score identifies out-of-distribution and\nmislabeled examples at one end of the continuum and strongly regular examples\nat the other end. We identify computationally inexpensive proxies to the\nconsistency score using statistics collected during training. We show examples\nof potential applications to the analysis of deep-learning systems.', 'authors': ['Ziheng Jiang', 'Chiyuan Zhang', 'Kunal Talwar', 'Michael C. Mozer'], 'published': '2020-02-08 17:39:46+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2002.03206v3'}
42,602
2102.10055v1
2,102.10055
1
Effective and Efficient Vote Attack on Capsule Networks
Standard Convolutional Neural Networks (CNNs) can be easily fooled by images with small quasi-imperceptible artificial perturbations. As alternatives to CNNs, the recently proposed Capsule Networks (CapsNets) are shown to be more robust to white-box attacks than CNNs under popular attack protocols. Besides, the class-conditional reconstruction part of CapsNets is also used to detect adversarial examples. In this work, we investigate the adversarial robustness of CapsNets, especially how the inner workings of CapsNets change when the output capsules are attacked. The first observation is that adversarial examples misled CapsNets by manipulating the votes from primary capsules. Another observation is the high computational cost, when we directly apply multi-step attack methods designed for CNNs to attack CapsNets, due to the computationally expensive routing mechanism. Motivated by these two observations, we propose a novel vote attack where we attack votes of CapsNets directly. Our vote attack is not only effective but also efficient by circumventing the routing process. Furthermore, we integrate our vote attack into the detection-aware attack paradigm, which can successfully bypass the class-conditional reconstruction based detection method. Extensive experiments demonstrate the superior attack performance of our vote attack on CapsNets.
2021-02-19
{'id': '2102.10055v1', 'title': 'Effective and Efficient Vote Attack on Capsule Networks', 'abstract': 'Standard Convolutional Neural Networks (CNNs) can be easily fooled by images\nwith small quasi-imperceptible artificial perturbations. As alternatives to\nCNNs, the recently proposed Capsule Networks (CapsNets) are shown to be more\nrobust to white-box attacks than CNNs under popular attack protocols. Besides,\nthe class-conditional reconstruction part of CapsNets is also used to detect\nadversarial examples. In this work, we investigate the adversarial robustness\nof CapsNets, especially how the inner workings of CapsNets change when the\noutput capsules are attacked. The first observation is that adversarial\nexamples misled CapsNets by manipulating the votes from primary capsules.\nAnother observation is the high computational cost, when we directly apply\nmulti-step attack methods designed for CNNs to attack CapsNets, due to the\ncomputationally expensive routing mechanism. Motivated by these two\nobservations, we propose a novel vote attack where we attack votes of CapsNets\ndirectly. Our vote attack is not only effective but also efficient by\ncircumventing the routing process. Furthermore, we integrate our vote attack\ninto the detection-aware attack paradigm, which can successfully bypass the\nclass-conditional reconstruction based detection method. Extensive experiments\ndemonstrate the superior attack performance of our vote attack on CapsNets.', 'authors': ['Jindong Gu', 'Baoyuan Wu', 'Volker Tresp'], 'published': '2021-02-19 17:35:07+00:00', 'categories': ['cs.CV'], 'url': 'http://arxiv.org/abs/2102.10055v1'}
42,605
2010.02347v2
2,010.02347
2
Learning with Instance-Dependent Label Noise: A Sample Sieve Approach
Human-annotated labels are often prone to noise, and the presence of such noise will degrade the performance of the resulting deep neural network (DNN) models. Much of the literature (with several recent exceptions) of learning with noisy labels focuses on the case when the label noise is independent of features. Practically, annotations errors tend to be instance-dependent and often depend on the difficulty levels of recognizing a certain task. Applying existing results from instance-independent settings would require a significant amount of estimation of noise rates. Therefore, providing theoretically rigorous solutions for learning with instance-dependent label noise remains a challenge. In this paper, we propose CORES$^2$ (COnfidence REgularized Sample Sieve), which progressively sieves out corrupted examples. The implementation of CORES$^2$ does not require specifying noise rates and yet we are able to provide theoretical guarantees of CORES$^2$ in filtering out the corrupted examples. This high-quality sample sieve allows us to treat clean examples and the corrupted ones separately in training a DNN solution, and such a separation is shown to be advantageous in the instance-dependent noise setting. We demonstrate the performance of CORES$^2$ on CIFAR10 and CIFAR100 datasets with synthetic instance-dependent label noise and Clothing1M with real-world human noise. As of independent interests, our sample sieve provides a generic machinery for anatomizing noisy datasets and provides a flexible interface for various robust training techniques to further improve the performance. Code is available at https://github.com/UCSC-REAL/cores.
2020-10-05
{'id': '2010.02347v2', 'title': 'Learning with Instance-Dependent Label Noise: A Sample Sieve Approach', 'abstract': 'Human-annotated labels are often prone to noise, and the presence of such\nnoise will degrade the performance of the resulting deep neural network (DNN)\nmodels. Much of the literature (with several recent exceptions) of learning\nwith noisy labels focuses on the case when the label noise is independent of\nfeatures. Practically, annotations errors tend to be instance-dependent and\noften depend on the difficulty levels of recognizing a certain task. Applying\nexisting results from instance-independent settings would require a significant\namount of estimation of noise rates. Therefore, providing theoretically\nrigorous solutions for learning with instance-dependent label noise remains a\nchallenge. In this paper, we propose CORES$^{2}$ (COnfidence REgularized Sample\nSieve), which progressively sieves out corrupted examples. The implementation\nof CORES$^{2}$ does not require specifying noise rates and yet we are able to\nprovide theoretical guarantees of CORES$^{2}$ in filtering out the corrupted\nexamples. This high-quality sample sieve allows us to treat clean examples and\nthe corrupted ones separately in training a DNN solution, and such a separation\nis shown to be advantageous in the instance-dependent noise setting. We\ndemonstrate the performance of CORES$^{2}$ on CIFAR10 and CIFAR100 datasets\nwith synthetic instance-dependent label noise and Clothing1M with real-world\nhuman noise. As of independent interests, our sample sieve provides a generic\nmachinery for anatomizing noisy datasets and provides a flexible interface for\nvarious robust training techniques to further improve the performance. Code is\navailable at https://github.com/UCSC-REAL/cores.', 'authors': ['Hao Cheng', 'Zhaowei Zhu', 'Xingyu Li', 'Yifei Gong', 'Xing Sun', 'Yang Liu'], 'published': '2020-10-05 21:44:09+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2010.02347v2'}
42,590
2310.04334v1
2,310.04334
1
Saliency-Guided Hidden Associative Replay for Continual Learning
Continual Learning is a burgeoning domain in next-generation AI, focusing on training neural networks over a sequence of tasks akin to human learning. While CL provides an edge over traditional supervised learning, its central challenge remains to counteract catastrophic forgetting and ensure the retention of prior tasks during subsequent learning. Amongst various strategies to tackle this, replay based methods have emerged as preeminent, echoing biological memory mechanisms. However, these methods are memory intensive, often preserving entire data samples, an approach inconsistent with humans selective memory retention of salient experiences. While some recent works have explored the storage of only significant portions of data in episodic memory, the inherent nature of partial data necessitates innovative retrieval mechanisms. Current solutions, like inpainting, approximate full data reconstruction from partial cues, a method that diverges from genuine human memory processes. Addressing these nuances, this paper presents the Saliency Guided Hidden Associative Replay for Continual Learning. This novel framework synergizes associative memory with replay-based strategies. SHARC primarily archives salient data segments via sparse memory encoding. Importantly, by harnessing associative memory paradigms, it introduces a content focused memory retrieval mechanism, promising swift and near-perfect recall, bringing CL a step closer to authentic human memory processes. Extensive experimental results demonstrate the effectiveness of our proposed method for various continual learning tasks.
2023-10-06
{'id': '2310.04334v1', 'title': 'Saliency-Guided Hidden Associative Replay for Continual Learning', 'abstract': 'Continual Learning is a burgeoning domain in next-generation AI, focusing on\ntraining neural networks over a sequence of tasks akin to human learning. While\nCL provides an edge over traditional supervised learning, its central challenge\nremains to counteract catastrophic forgetting and ensure the retention of prior\ntasks during subsequent learning. Amongst various strategies to tackle this,\nreplay based methods have emerged as preeminent, echoing biological memory\nmechanisms. However, these methods are memory intensive, often preserving\nentire data samples, an approach inconsistent with humans selective memory\nretention of salient experiences. While some recent works have explored the\nstorage of only significant portions of data in episodic memory, the inherent\nnature of partial data necessitates innovative retrieval mechanisms. Current\nsolutions, like inpainting, approximate full data reconstruction from partial\ncues, a method that diverges from genuine human memory processes. Addressing\nthese nuances, this paper presents the Saliency Guided Hidden Associative\nReplay for Continual Learning. This novel framework synergizes associative\nmemory with replay-based strategies. SHARC primarily archives salient data\nsegments via sparse memory encoding. Importantly, by harnessing associative\nmemory paradigms, it introduces a content focused memory retrieval mechanism,\npromising swift and near-perfect recall, bringing CL a step closer to authentic\nhuman memory processes. Extensive experimental results demonstrate the\neffectiveness of our proposed method for various continual learning tasks.', 'authors': ['Guangji Bai', 'Qilong Zhao', 'Xiaoyang Jiang', 'Yifei Zhang', 'Liang Zhao'], 'published': '2023-10-06 15:54:12+00:00', 'categories': ['cs.LG'], 'url': 'http://arxiv.org/abs/2310.04334v1'}
42,606
2306.04793v1
2,306.04793
1
On the Joint Interaction of Models, Data, and Features
Learning features from data is one of the defining characteristics of deep learning, but our theoretical understanding of the role features play in deep learning is still rudimentary. To address this gap, we introduce a new tool, the interaction tensor, for empirically analyzing the interaction between data and model through features. With the interaction tensor, we make several key observations about how features are distributed in data and how models with different random seeds learn different features. Based on these observations, we propose a conceptual framework for feature learning. Under this framework, the expected accuracy for a single hypothesis and agreement for a pair of hypotheses can both be derived in closed-form. We demonstrate that the proposed framework can explain empirically observed phenomena, including the recently discovered Generalization Disagreement Equality (GDE) that allows for estimating the generalization error with only unlabeled data. Further, our theory also provides explicit construction of natural data distributions that break the GDE. Thus, we believe this work provides valuable new insight into our understanding of feature learning.
2023-06-07
{'id': '2306.04793v1', 'title': 'On the Joint Interaction of Models, Data, and Features', 'abstract': 'Learning features from data is one of the defining characteristics of deep\nlearning, but our theoretical understanding of the role features play in deep\nlearning is still rudimentary. To address this gap, we introduce a new tool,\nthe interaction tensor, for empirically analyzing the interaction between data\nand model through features. With the interaction tensor, we make several key\nobservations about how features are distributed in data and how models with\ndifferent random seeds learn different features. Based on these observations,\nwe propose a conceptual framework for feature learning. Under this framework,\nthe expected accuracy for a single hypothesis and agreement for a pair of\nhypotheses can both be derived in closed-form. We demonstrate that the proposed\nframework can explain empirically observed phenomena, including the recently\ndiscovered Generalization Disagreement Equality (GDE) that allows for\nestimating the generalization error with only unlabeled data. Further, our\ntheory also provides explicit construction of natural data distributions that\nbreak the GDE. Thus, we believe this work provides valuable new insight into\nour understanding of feature learning.', 'authors': ['Yiding Jiang', 'Christina Baek', 'J. Zico Kolter'], 'published': '2023-06-07 21:35:26+00:00', 'categories': ['cs.LG', 'stat.ML'], 'url': 'http://arxiv.org/abs/2306.04793v1'}