paper_name stringlengths 11 170 | text stringlengths 8.07k 307k | summary stringlengths 152 6.16k | paper_id stringlengths 43 43 |
|---|---|---|---|
Exploring the Limits of Large Scale Pre-training | 1 INTRODUCTION . Recent impressive progress on transfer and few-shot learning ( Brown et al. , 2020 ; Goyal et al. , 2021 ; Kolesnikov et al. , 2019 ; Pham et al. , 2020 ; Dosovitskiy et al. , 2020 ; Dumoulin et al. , 2021 ; Radford et al. , 2021 ) suggests an emerging direction that scaling up models and training them... | The paper presents an empirical study (respectively meta-study) of large-scale supervised pre-training for image recognition tasks. By analysing lots of experiments with varying model sizes, dataset sizes and training durations, the paper reaches the conclusion that simply scaling them up for a generic pre-training tas... | SP:95f32813140f9e12b9d0d6f3ecad90c1ad00b0a0 |
Understanding Graph Learning with Local Intrinsic Dimensionality | 1 INTRODUCTION . Graphs are widely used to model real-life problems owing to their flexible structure and ability to carry different types of information . Graph learning has thus become essential for a wide range of applications in biomedicine ( Zitnik et al. , 2018 ) , physics ( Battaglia et al. , 2016 ) and traffic ... | In this work, the authors investigate the Local Intrinsic Dimensionality (LID), especially the feature (FLID), structure LID (SLID), and Representation LID (RLID) of a graph. Through experimental analysis, the authors demonstrate that the FLID and SLID are well correlated with the graph complexity, and real-world graph... | SP:ab4f7885ce56867b46ac82f3ded3daa83f556c62 |
Understanding Graph Learning with Local Intrinsic Dimensionality | 1 INTRODUCTION . Graphs are widely used to model real-life problems owing to their flexible structure and ability to carry different types of information . Graph learning has thus become essential for a wide range of applications in biomedicine ( Zitnik et al. , 2018 ) , physics ( Battaglia et al. , 2016 ) and traffic ... | The paper characterizes the intrinsic dimensionality of node features, graph structures, and representations learned by GNNs via the so-called Local Intrinsic Dimensionality (LID) measure, intending that it can benefit the community in understanding the difficulty of an underlying graph learning task. In addition, esti... | SP:ab4f7885ce56867b46ac82f3ded3daa83f556c62 |
Pseudo Numerical Methods for Diffusion Models on Manifolds | 1 INTRODUCTION . Denoising Diffusion Probabilistic Models ( DDPMs ) ( Sohl-Dickstein et al. , 2015 ; Ho et al. , 2020 ) is a class of generative models which model the data distribution through an iterative denoising process reversing a multi-step noising process . DDPMs have been applied successfully to a variety of a... | The paper proposes a new efficient method for denoising diffusion probabilistic models (DDPM) (generative models that optimize for the closest solution on a manifold) based on the observation that this can be seen a solving a set of differential equations on a manifold. This allows efficient pseudo numerical methods to... | SP:9be34b13f59e6e33820e863d7ed33f0479fc368e |
Pseudo Numerical Methods for Diffusion Models on Manifolds | 1 INTRODUCTION . Denoising Diffusion Probabilistic Models ( DDPMs ) ( Sohl-Dickstein et al. , 2015 ; Ho et al. , 2020 ) is a class of generative models which model the data distribution through an iterative denoising process reversing a multi-step noising process . DDPMs have been applied successfully to a variety of a... | Highlighting the high computational complexity for sampling from Denoising Diffusion Probabilistic Models (DDPMs) (e.g. wrt GANs), authors build on the connection between diffusion processes and ODEs to propose efficient (pseudo-)numerical methods so as to sample data from the data manifold. The main idea is to combine... | SP:9be34b13f59e6e33820e863d7ed33f0479fc368e |
DKM: Differentiable k-Means Clustering Layer for Neural Network Compression | Deep neural network ( DNN ) model compression for efficient on-device inference becomes increasingly important to reduce memory requirements and keep user data on-device . To this end , we propose a novel differentiable k-means clustering layer ( DKM ) and its application to train-time weight-clustering for DNN model c... | The paper is concerned with reducing the size of deep neural networks using weight sharing. The paper proposes a new building block that performs a soft k-means algorithm where each weight is assigned a convex combination of the cluster centers. At test-time, the weights are assigned to their closest cluster center suc... | SP:67efe87f8db28e0aa68246cc5b34ddc028188df7 |
DKM: Differentiable k-Means Clustering Layer for Neural Network Compression | Deep neural network ( DNN ) model compression for efficient on-device inference becomes increasingly important to reduce memory requirements and keep user data on-device . To this end , we propose a novel differentiable k-means clustering layer ( DKM ) and its application to train-time weight-clustering for DNN model c... | This paper proposes a novel differentiable k-means clustering layer (DKM) for deep neural network model compression. The DKM utilizes attention mechanism to align the weight-to-cluster assignment with the training loss function. Overall, the idea is novel but the paper is not prepared enough. | SP:67efe87f8db28e0aa68246cc5b34ddc028188df7 |
Faster Reinforcement Learning with Value Target Lower Bounding | 1 INTRODUCTION . In temporal difference ( TD ) learning , the value function is adjusted toward its Bellman target , which is the reward of the current step plus the discounted value of the next state . This forms the basis of many state of the art reinforcement learning ( RL ) algorithms such as DQN ( Mnih et al. , 20... | This paper proposed value target lower bounding as a simple modification to the Bellman operator that intends to improve convergence speed. It proves that using such a lower bound in the Bellman backup does not change the fixed point in the tabular setting. The paper then proposes two instantiations of particular value... | SP:7f894264c42d9e9670233250810e71c20d2f7fcf |
Faster Reinforcement Learning with Value Target Lower Bounding | 1 INTRODUCTION . In temporal difference ( TD ) learning , the value function is adjusted toward its Bellman target , which is the reward of the current step plus the discounted value of the next state . This forms the basis of many state of the art reinforcement learning ( RL ) algorithms such as DQN ( Mnih et al. , 20... | This paper proposes a new RL algorithm based on a modified Bellman backup equation. The main idea is to estimate the value of a state in multiple ways (using a Q function and using Monte Carlo returns) and then to take the maximum over these estimates. The paper shows that, if all estimators are lower bounds on the tru... | SP:7f894264c42d9e9670233250810e71c20d2f7fcf |
Semi-supervised learning of partial differential operators and dynamical flows | 1 INTRODUCTION . The evolution of classical and quantum physical dynamical systems in space and time is generically modeled by non-linear partial differential equations . Such are , for instance , Einstein equations of General Relativity , Maxwell equations of Electromagnetism , Schrödinger equation of Quantum Mechani... | The paper combines the FNO architecture with hypernetworks to learn the solution operator (flow map) of Markovian PDE systems. A hypernetwork whose input is the time domain (\R_+) is trained to output the weights of a FNO which then acts on an initial condition to produce the PDE solution at the time input to the hyp... | SP:ade2bc3c672043e09debf5ff560deb6eb9c16c1d |
Semi-supervised learning of partial differential operators and dynamical flows | 1 INTRODUCTION . The evolution of classical and quantum physical dynamical systems in space and time is generically modeled by non-linear partial differential equations . Such are , for instance , Einstein equations of General Relativity , Maxwell equations of Electromagnetism , Schrödinger equation of Quantum Mechani... | The paper presents a model and a loss function for approximation of the solution map of certain partial differential equations. The idea is to use HyperNetworks: the mapping of the initial condition to time t done by model f($\theta$, x), and the parameters are predicted by another neural network. The models utilize t... | SP:ade2bc3c672043e09debf5ff560deb6eb9c16c1d |
How to Robustify Black-Box ML Models? A Zeroth-Order Optimization Perspective | 1 INTRODUCTION . ML models , DNNs in particular , have achieved remarkable success owing to their superior predictive performance . However , they often lack robustness . For example , imperceptible but carefully-crafted input perturbations can fool the decision of a well-trained ML model . These input perturbations re... | The authors formulate the problem of black-box defense and propose a novel black-box defense approach called the Zero Order AutoEncoder-based Denoised Smoothing (ZO-AE-DS). Black-box defense corresponds to situations in which the defense model information cannot be obtained due to privacy protection in real scenarios. ... | SP:57cc0c93dd03b67e5edf378ed41bd492bd6da2b2 |
How to Robustify Black-Box ML Models? A Zeroth-Order Optimization Perspective | 1 INTRODUCTION . ML models , DNNs in particular , have achieved remarkable success owing to their superior predictive performance . However , they often lack robustness . For example , imperceptible but carefully-crafted input perturbations can fool the decision of a well-trained ML model . These input perturbations re... | This work provides an algorithm to ensure robust training of an ML model with just black-box knowledge of it i.e., input and output access. The algorithm relies on using Denoised Smoothing with zeroth-order optimization where the gradients are estimated using random perturbations (finite-differencing). They avoid the c... | SP:57cc0c93dd03b67e5edf378ed41bd492bd6da2b2 |
Boosting the Certified Robustness of L-infinity Distance Nets | 1 INTRODUCTION . Modern neural networks , while achieving high accuracy on various tasks , are found to be vulnerable to small , adversarially-chosen perturbations of the inputs ( Szegedy et al. , 2013 ; Biggio et al. , 2013 ) . Given an image x correctly classified by a neural network , there often exists a small adve... | This paper proposed a simple modification of $\ell_\infty$ net training, which boosts the accuracy for certified robustness under $\ell_\infty$ attack. It provides a trainable scale on the output of the network and uses a clipped hinge loss. The paper also proves the expressive ability of $\ell_\infty$ nets for classif... | SP:f3620e9c72efa8512c624f0e055aa229b1af949e |
Boosting the Certified Robustness of L-infinity Distance Nets | 1 INTRODUCTION . Modern neural networks , while achieving high accuracy on various tasks , are found to be vulnerable to small , adversarially-chosen perturbations of the inputs ( Szegedy et al. , 2013 ; Biggio et al. , 2013 ) . Given an image x correctly classified by a neural network , there often exists a small adve... | This paper is a follow-up paper of Zhang et al. (2021). In Zhang et al. (2021), the authors proposed a new network architecture, l_infty distance net. By construction, the network is 1-Lipschitz w.r.t. l_infty distance. However, the training procedure therein is problematic. This paper resolves the issue by a new loss ... | SP:f3620e9c72efa8512c624f0e055aa229b1af949e |
Wisdom of Committees: An Overlooked Approach To Faster and More Accurate Models | 1 INTRODUCTION Optimizing the efficiency of neural networks is important for real-world applications as they can only use limited computational resources and often have requirements on response time . There has been considerable work in this direction ( Howard et al. , 2017 ; Zhang et al. , 2018 ; Tan & Le , 2019 ) , b... | The work provides an empirical accuracy-efficiency comparison of model ensembles and cascades of shallow models against single deeper models. The main finding, which supports previous results, is rather interesting: compositions of shallow models tend to provide better efficiency-accuracy trade-off than single deep mod... | SP:a0537bc2883ff413f0ffafc44a53a65be9fc7738 |
Wisdom of Committees: An Overlooked Approach To Faster and More Accurate Models | 1 INTRODUCTION Optimizing the efficiency of neural networks is important for real-world applications as they can only use limited computational resources and often have requirements on response time . There has been considerable work in this direction ( Howard et al. , 2017 ; Zhang et al. , 2018 ; Tan & Le , 2019 ) , b... | This paper investigates the effectiveness of model cascades in computation/accuracy tradeoff improvement. A straightforward procedure is used, where all combination-permutations of a handful of models are evaluated in a cascade, and exit thresholds are determined by choosing best computation work within accuracy degra... | SP:a0537bc2883ff413f0ffafc44a53a65be9fc7738 |
MoReL: Multi-omics Relational Learning | 1 INTRODUCTION . Multi-view learning tries to fully leverage the information from multiple sources ( i.e . different types of omics data in molecular biology ) and represents them in a shared embedding space , which is beneficial for many downstream tasks with a limited number of training samples . In biomedical applic... | The paper proposed a deep Bayesian model for heterogeneous multi-omics data integration. The Gromov-Wasserstein (FGW) regularization between latent representations of heterogeneous views is used to align nodes/features in every pair of views. The experimental results have demonstrated improvement in inferring meaningfu... | SP:2d073cd15d16bdf07f9e934c28192a1b36a27b38 |
MoReL: Multi-omics Relational Learning | 1 INTRODUCTION . Multi-view learning tries to fully leverage the information from multiple sources ( i.e . different types of omics data in molecular biology ) and represents them in a shared embedding space , which is beneficial for many downstream tasks with a limited number of training samples . In biomedical applic... | The authors propose a Bayesian framework to learn relations among multi-omic datasets. The main unique advantage over existing methods is the proposed method is able to learn without a priori dependency structure and it allows a certain degree of missingness and mismatching. Experiments on two biomedical multi-omics da... | SP:2d073cd15d16bdf07f9e934c28192a1b36a27b38 |
Tactics on Refining Decision Boundary for Improving Certification-based Robust Training | In certification-based robust training , existing methods utilize relaxation based methods to bound the worst case performance of neural networks given certain perturbation . However , these certification based methods treat all the examples equally regardless of their vulnerability and true adversarial distribution , ... | This paper proposes bound-based weighted loss and epsilon auto-tuning to improve the performance of certifiable training. The insights of the improvements are mainly borrowed from well-developed adversarial training while they are customized for certifiable training considering bound margins provided bound propagation ... | SP:d7b4a400a8376f863b898e6ab363485bb95f0cde |
Tactics on Refining Decision Boundary for Improving Certification-based Robust Training | In certification-based robust training , existing methods utilize relaxation based methods to bound the worst case performance of neural networks given certain perturbation . However , these certification based methods treat all the examples equally regardless of their vulnerability and true adversarial distribution , ... | This paper proposes two ideas for improving the performance of certified training. The first idea is to use assign weight for each input based on the margin to the decision boundary. The second idea is to use automatic scheduling of perturbation radius during training. They show that using these two ideas leads to impr... | SP:d7b4a400a8376f863b898e6ab363485bb95f0cde |
Sparsity Winning Twice: Better Robust Generalization from More Efficient Training | 1 INTRODUCTION Deep neural networks ( DNNs ) are notoriously vulnerable to maliciously crafted adversarial attacks . To conquer this fragility , numerous adversarial defense mechanisms are proposed to establish robust neural networks ( Schmidt et al. , 2018 ; Sun et al. , 2019 ; Nakkiran , 2019 ; Raghunathan et al. , 2... | Recent studies demonstrate adversarial training suffers from severe overfitting besides getting very expensive. This paper proposes to handle the two problems organically altogether, with the tool of sparse training. The authors show that injecting appropriate sparsity forms in training could substantially shrink the... | SP:2d4b408a083d8ccd887b847c98ff1faed9d90d30 |
Sparsity Winning Twice: Better Robust Generalization from More Efficient Training | 1 INTRODUCTION Deep neural networks ( DNNs ) are notoriously vulnerable to maliciously crafted adversarial attacks . To conquer this fragility , numerous adversarial defense mechanisms are proposed to establish robust neural networks ( Schmidt et al. , 2018 ; Sun et al. , 2019 ; Nakkiran , 2019 ; Raghunathan et al. , 2... | This paper deals with the problem of training a neural network so that it generalizes well over data unseen at training time. Namely, they address the particular case where a network is trained over an adversarial scheme. This paper proposes two methods for learning a sparse architecture called robust and flying bird. ... | SP:2d4b408a083d8ccd887b847c98ff1faed9d90d30 |
High Probability Bounds for a Class of Nonconvex Algorithms with AdaGrad Stepsize | √ T ) with high probability without the knowledge of smoothness and variance . We use a particular version of Freedman ’ s concentration bound for martingale difference sequences [ Kakade & Tewari , 2008 ] which enables us to achieve the best-known dependence of log ( 1/δ ) on the probability margin δ . We present our ... | This paper proposed a new analysis for AdaGrad method in smooth and non-convex optimization, to get high probability convergence toward stationary points. Based on some assumptions (Eqs. (2), (6) and (7)), i.e., Lipschitz, bounded variance of gradient estimates, and bounded stochastic gradient, the authors analyzed Al... | SP:e384abbadce76420670e40b9597cc5511369d422 |
High Probability Bounds for a Class of Nonconvex Algorithms with AdaGrad Stepsize | √ T ) with high probability without the knowledge of smoothness and variance . We use a particular version of Freedman ’ s concentration bound for martingale difference sequences [ Kakade & Tewari , 2008 ] which enables us to achieve the best-known dependence of log ( 1/δ ) on the probability margin δ . We present our ... | This paper provides an analysis of adaptive learning rate scheme in stochastic non-convex settings. Under assumptions of smoothness and bounded stochastic gradients (a light tailed condition), it is shown that the sum of the gradient norms of the iterates of the non-convex algorithm is small. In order to accomplish thi... | SP:e384abbadce76420670e40b9597cc5511369d422 |
FrugalMCT: Efficient Online ML API Selection for Multi-Label Classification Tasks | 1 INTRODUCTION . Many machine learning users are starting to adopt machine learning as a service ( MLaaS ) APIs to obtain high-quality predictions . One of the most common tasks these APIs target is multi-label classification . For example , one can use Google ’ s computer vision API ( Goo ) to tag an image with a wide... | Given several mutli-label machine learning APIs, this paper study how to select those APIs under a budget constrant while striving to improve the overall accuracy. The author first formulate the budget API select problem as an integer linear programming problem, then relax the integer contraint and solving the relax pr... | SP:31412e46449bccbfe0b74080a1c15df64b2363d5 |
FrugalMCT: Efficient Online ML API Selection for Multi-Label Classification Tasks | 1 INTRODUCTION . Many machine learning users are starting to adopt machine learning as a service ( MLaaS ) APIs to obtain high-quality predictions . One of the most common tasks these APIs target is multi-label classification . For example , one can use Google ’ s computer vision API ( Goo ) to tag an image with a wide... | This paper addresses the practical task to use the combination of ML APIs for multi-label classification. Different from the related work FrugalML which ignores the correlation between ML APIs, the proposed FrugalMCT allows selecting and combining the different ML APIs based on a budget. Sufficient theoretical and empi... | SP:31412e46449bccbfe0b74080a1c15df64b2363d5 |
Reinforcement Learning for Adaptive Mesh Refinement | 1 INTRODUCTION . Numerical simulation of PDEs via the finite element method ( FEM ) ( Brenner & Scott , 2007 ) plays an integral role in computational science and engineering ( Reddy & Gartling , 2010 ; Monk et al. , 2003 ) . Given a fixed set of basis functions , the resolution of the finite element mesh determines th... | This paper proposes an application of reinforcement learning for adaptive mesh refinement in large-scale finite element simulations of complex physical systems. The authors suggest to formulate the mesh refinement problem as a MDP and propose different policy architectures for scalable application of reinforcement lear... | SP:50912255573295ef5e76ec95e6e83b9ee0b3534e |
Reinforcement Learning for Adaptive Mesh Refinement | 1 INTRODUCTION . Numerical simulation of PDEs via the finite element method ( FEM ) ( Brenner & Scott , 2007 ) plays an integral role in computational science and engineering ( Reddy & Gartling , 2010 ; Monk et al. , 2003 ) . Given a fixed set of basis functions , the resolution of the finite element mesh determines th... | For various complicated problems governed by PDEs (e.g. solid/fluid interactions, aerodynamics, elasticity, backscattering, etc.) the computational cost can become prohibitive even for one inquiry, let alone parametric study. In the same time, mesh refinement is crucial to achieve acceptable accuracy. To mitigate such ... | SP:50912255573295ef5e76ec95e6e83b9ee0b3534e |
Information Gain Propagation: a New Way to Graph Active Learning with Soft Labels | 1 INTRODUCTION . Graph Neural Networks ( GNNs ) have recently achieved remarkable success in various graph-based tasks , ranging from traffic networks , biology , to social networks ( Zhang et al. , 2020 ; Wang et al. , 2019 ; Li et al. , 2019 ; Do et al. , 2019 ) . Despite their effectiveness and popularity , GNNs typ... | The paper proposes an active learning method for GNNs that is based on an information gain maximization where ethe information gain is obtained by querying a data point and looking at the influence of the queried node on the neighborhood relative to their previous information. They also claim the setting of relaxing t... | SP:3907616cf8748efca1c63a16cfb9335e1380aec8 |
Information Gain Propagation: a New Way to Graph Active Learning with Soft Labels | 1 INTRODUCTION . Graph Neural Networks ( GNNs ) have recently achieved remarkable success in various graph-based tasks , ranging from traffic networks , biology , to social networks ( Zhang et al. , 2020 ; Wang et al. , 2019 ; Li et al. , 2019 ; Do et al. , 2019 ) . Despite their effectiveness and popularity , GNNs typ... | The paper proposes a new method active learning (AL) on graphs. Unlike other AL approaches, the proposed approach provides soft labels via *relaxed queries* to the *domain experts*. Main Contributions: 1). The paper proposes a new innovative approach for graph active learning with soft labels. The key idea is to ... | SP:3907616cf8748efca1c63a16cfb9335e1380aec8 |
Style Equalization: Unsupervised Learning of Controllable Generative Sequence Models | 1 INTRODUCTION . The goal of controllable generative sequence models is to generate sequences containing target content in a target style . With the capability to select speaker voices , multi-speaker text-to-speech models have been successfully adopted in many voice assistants ( Gibiansky et al. , 2017 ; Ping et al. ,... | In this paper, the authors argue that the typical training algorithms for controllable sequence generative models suffers from the 'training-inference mismatch'. Therefore, to address such a problem, they introduce a style transformation module that is called 'style equalization'. Such a module is designed to enable tr... | SP:fd11668b6b0d122ede44d1dcec6f33e3f4e20e0c |
Style Equalization: Unsupervised Learning of Controllable Generative Sequence Models | 1 INTRODUCTION . The goal of controllable generative sequence models is to generate sequences containing target content in a target style . With the capability to select speaker voices , multi-speaker text-to-speech models have been successfully adopted in many voice assistants ( Gibiansky et al. , 2017 ; Ping et al. ,... | - To enhance the quality of style-controlled generation, especially in an unsupervised manner and non-parallel setting, this paper proposes a "style equalization" mechanism to prevent the content leakage problem. In the style equalization module, the style of a sample is transformed to be the same as the style of groun... | SP:fd11668b6b0d122ede44d1dcec6f33e3f4e20e0c |
ANCER: Anisotropic Certification via Sample-wise Volume Maximization | 1 INTRODUCTION . The well-studied fact that Deep Neural Networks ( DNNs ) are vulnerable to additive imperceptible noise perturbations has led to a growing interest in developing robust classifiers ( Goodfellow et al. , 2015 ; Szegedy et al. , 2014 ) . A recent promising approach to achieve state-of-the-art provable ro... | In this paper, the authors discuss the extension of $\ell_{p}$-randomized smoothing to anisotropic counterparts. In particular, they consider the extension of $\ell_{2}$-certificates from (hyper)spheres to (hyper)ellipsoids by sampling anisotropic rather than isotropic Gaussian noise, as well as the extension of $\ell_... | SP:b6fef1ef35ccf967c2df9d1704ba38df1af3e879 |
ANCER: Anisotropic Certification via Sample-wise Volume Maximization | 1 INTRODUCTION . The well-studied fact that Deep Neural Networks ( DNNs ) are vulnerable to additive imperceptible noise perturbations has led to a growing interest in developing robust classifiers ( Goodfellow et al. , 2015 ; Szegedy et al. , 2014 ) . A recent promising approach to achieve state-of-the-art provable ro... | The paper proposes the anisotropic version of randomized smoothing. Evaluation metrics based on the volume of the certified region are proposed, allowing comparisons with the certified regions provided from isotropic randomized smoothing. Experimental results show the usefulness of introducing anisotropic randomized sm... | SP:b6fef1ef35ccf967c2df9d1704ba38df1af3e879 |
Temporal Alignment Prediction for Supervised Representation Learning and Few-Shot Sequence Classification | 1 INTRODUCTION . Distance between sequences plays a crucial role in sequence classification ( Sakoe & Chiba , 1978 ) , retrieval ( Su et al. , 2019 ) , clustering ( Garcı́a-Garcı́a et al. , 2008 ) , etc . Measuring distance between sequences is difficult , since different sequences may have different sampling rates , e... | This paper focuses on temporal sequence alignment, i.e. the task of finding an optimal alignment between sequences of different lengths. This task has been addressed by various traditional methods (e.g. DTW) that involve dynamic programming or other optimisations techniques that cannot be easily embedded in end-to-end ... | SP:35fdcb687dd1aff475c0c3cee2b899579de19b46 |
Temporal Alignment Prediction for Supervised Representation Learning and Few-Shot Sequence Classification | 1 INTRODUCTION . Distance between sequences plays a crucial role in sequence classification ( Sakoe & Chiba , 1978 ) , retrieval ( Su et al. , 2019 ) , clustering ( Garcı́a-Garcı́a et al. , 2008 ) , etc . Measuring distance between sequences is difficult , since different sequences may have different sampling rates , e... | One-shot sequence alignment with CNN. Comparisons of sequence data are important for many tasks including action recognition and retrieval. Previous approaches include generating fixed-size feature vectors (e.g., RNN) or temporally aligning two sequences (e.g., DTW). Instead, the proposed method directly predict the al... | SP:35fdcb687dd1aff475c0c3cee2b899579de19b46 |
ZARTS: On Zero-order Optimization for Neural Architecture Search | Differentiable architecture search ( DARTS ) has been a popular one-shot paradigm for NAS due to its high efficiency . It introduces trainable architecture parameters to represent the importance of candidate operations and proposes first/secondorder approximation to estimate their gradients , making it possible to solv... | This paper presents ZARTS, a zero-order optimization method for DARTS, to search without enforcing the approximation of the network weights. It conducts in-depth analysis on the first/second order approximation in DARTS, and points out that such approximation leads to bias and instability. Then the work proposes three ... | SP:072d9da072c1ef224e3cfb6f67ef7c0e78e456af |
ZARTS: On Zero-order Optimization for Neural Architecture Search | Differentiable architecture search ( DARTS ) has been a popular one-shot paradigm for NAS due to its high efficiency . It introduces trainable architecture parameters to represent the importance of candidate operations and proposes first/secondorder approximation to estimate their gradients , making it possible to solv... | This paper (ZARTS) proposes to apply gradient-estimation-based zero-order optimization methods to tackle neural architecture search (NAS). Two major contributions are made by this paper: (1) it is the first to borrow the methods of zero-order optimization to solve NAS problem; (2) it shows that zero-order optimization... | SP:072d9da072c1ef224e3cfb6f67ef7c0e78e456af |
Model-Agnostic Meta-Attack: Towards Reliable Evaluation of Adversarial Robustness | 1 INTRODUCTION . Deep neural networks are vulnerable to maliciously crafted adversarial examples ( Biggio et al. , 2013 ; Szegedy et al. , 2014 ; Goodfellow et al. , 2015 ) , which aim to induce erroneous model predictions by adding small perturbations to the normal inputs . Due to the threats , a multitude of defense ... | This paper proposes an automatic approach for attacking classifiers, by approximating a possible adaptive attack that can take place on a newly-published defense. The methodology (MAMA) applies meta machine learning techniques, by training it on different defenses to let it grasp what might be a good signal to follow w... | SP:c44e42aed2bb38106bdc33fd5e7b3dfb4e9c5584 |
Model-Agnostic Meta-Attack: Towards Reliable Evaluation of Adversarial Robustness | 1 INTRODUCTION . Deep neural networks are vulnerable to maliciously crafted adversarial examples ( Biggio et al. , 2013 ; Szegedy et al. , 2014 ; Goodfellow et al. , 2015 ) , which aim to induce erroneous model predictions by adding small perturbations to the normal inputs . Due to the threats , a multitude of defense ... | This paper proposes a model-agnostic meta-attack and achieves promising adversarial attack performance compared with state-of-the-art adversarial attack algorithms. The proposed algorithm overcomes many issues, such as the vanishing gradient problem, training instability problem, generalization to other defense models.... | SP:c44e42aed2bb38106bdc33fd5e7b3dfb4e9c5584 |
GradSign: Model Performance Inference with Theoretical Insights | A key challenge in neural architecture search ( NAS ) is quickly inferring the predictive performance of a broad spectrum of networks to discover statistically accurate and computationally efficient ones . We refer to this task as model performance inference ( MPI ) . The current practice for efficient MPI is gradient-... | Model performance inference is a key challenge in neural architecture search. This paper introduces GradSign, an accurate, simple, and flexible metric for model performance inference. GradSign approximately analyzes the optimization landscape of different networks at the granularity of individual training samples using... | SP:0eb1faf1d1b44c5c80745824ae303ee086eae8a4 |
GradSign: Model Performance Inference with Theoretical Insights | A key challenge in neural architecture search ( NAS ) is quickly inferring the predictive performance of a broad spectrum of networks to discover statistically accurate and computationally efficient ones . We refer to this task as model performance inference ( MPI ) . The current practice for efficient MPI is gradient-... | This work discusses the problem of neural architecture search. While existing gradient methods are based on heuristics, this work proposes a metric called Gradsign for model performance inference, which provides some theoretical guarantees and performs well in practise. Authors compare gradsign to a number of state o... | SP:0eb1faf1d1b44c5c80745824ae303ee086eae8a4 |
On the approximation properties of recurrent encoder-decoder architectures | 1 Introduction . Encoder-decoder is an increasingly popular architecture for sequence to sequence modelling problems ( Sutskever et al. , 2014 ; Chiu et al. , 2018 ; Venugopalan et al. , 2015 ) . The core of this architecture is to first encode the input sequence into a vector using the encoder and then map the vector ... | This paper provides theoretical studies for why encoder-decoder can be seen as a generalization of RNNs in time-inhomogenous sequence modeling. The authors put an impressive amount of effort into mathematically defining approximation properties of RNN encoder-decoders beginning from a universal approximation result. ... | SP:b33082267afa3cdcb3a8e2a049b1a4bc8f3d9d5d |
On the approximation properties of recurrent encoder-decoder architectures | 1 Introduction . Encoder-decoder is an increasingly popular architecture for sequence to sequence modelling problems ( Sutskever et al. , 2014 ; Chiu et al. , 2018 ; Venugopalan et al. , 2015 ) . The core of this architecture is to first encode the input sequence into a vector using the encoder and then map the vector ... | This paper provides theoretical insight for approximation properties of RNN encoder-decoder architecture in linear setting. More specifically, they study supervised learning problem of temporal modelling where a first RNN encodes a given sequence into a coding vector and a second RNN is responsible to decode said vecto... | SP:b33082267afa3cdcb3a8e2a049b1a4bc8f3d9d5d |
Debiasing Pretrained Text Encoders by Paying Attention to Paying Attention | 1 INTRODUCTION . Natural Language Processing ( NLP ) is increasingly penetrating real-world operations such as recruitment ( Hansen et al. , 2015 ) , legal systems ( Dale , 2019 ) , healthcare ( Velupillai et al. , 2018 ) and Web Search ( Nalisnick et al. , 2016 ) . Part of this success is attributed to the underlying ... | This paper proposes a new debiasing method for contextualized word embeddings, specifically for attention-based text encoders. At a very high level, the proposed method tries to calibrate the attention scores of words from different groups, e.g. to reduce gender bias, the method forces the model (text encoder) to alloc... | SP:e7c0d655d20b3a6de09dd2ea2d150b149ed60845 |
Debiasing Pretrained Text Encoders by Paying Attention to Paying Attention | 1 INTRODUCTION . Natural Language Processing ( NLP ) is increasingly penetrating real-world operations such as recruitment ( Hansen et al. , 2015 ) , legal systems ( Dale , 2019 ) , healthcare ( Velupillai et al. , 2018 ) and Web Search ( Nalisnick et al. , 2016 ) . Part of this success is attributed to the underlying ... | This paper addresses potential biases introduced by attention models by re-weighing the attention weights. First, the paper provides a few examples that demonstrate attention weights correlated with social stereotype (e.g., doctor attending to he and nurse attending to she). Then, it proposes to reduce this type of bia... | SP:e7c0d655d20b3a6de09dd2ea2d150b149ed60845 |
Information Bottleneck: Exact Analysis of (Quantized) Neural Networks | 1 INTRODUCTION . Improving our theoretical understanding of why over-parameterized deep neural networks generalize well is arguably one of main problems in current machine learning research ( Poggio et al. , 2020 ) . Tishby & Zaslavsky ( 2015 ) suggested to analyze deep neural networks based on their Information Bottle... | **Update after authors' response:** The authors have managed to clarify some issues and make a couple of small improvements to the manuscript. I am tempted to raise my score to a 7, but to me personally the paper does not quite pass the threshold for an 8 (which is the next possible rating on the conference scale). I a... | SP:5626c1fc910929420a1453636be1da17572c3872 |
Information Bottleneck: Exact Analysis of (Quantized) Neural Networks | 1 INTRODUCTION . Improving our theoretical understanding of why over-parameterized deep neural networks generalize well is arguably one of main problems in current machine learning research ( Poggio et al. , 2020 ) . Tishby & Zaslavsky ( 2015 ) suggested to analyze deep neural networks based on their Information Bottle... | This paper considers the important problem of mutual information estimation in neural networks, a problem at the root of a debate on the usefulness of the information-bottleneck approach for the analysis of information flow in neural networks. There exist many approximation schemes to get estimates of the inter-layers ... | SP:5626c1fc910929420a1453636be1da17572c3872 |
A Simple Approach to Adversarial Robustness in Few-shot Image Classification | 1 INTRODUCTION . Few-shot learning presents the challenge of generalizing to unseen tasks with limited data . The problem is aimed at learning quickly from few examples of data , which is generally considered the hallmark of human intelligence . This is an important practical problem due to the scarce availability of f... | This paper aims to address the problem of adversarial attack for low shot image classification. This work is motivated by the challenging scenario where there is a need of significant amount of data to train an adversarial robust classifier and there is not much data under few-shot setting. This work proposed and demon... | SP:b27d0bb34999cb1d197f68cca7e4c01c433ed6e8 |
A Simple Approach to Adversarial Robustness in Few-shot Image Classification | 1 INTRODUCTION . Few-shot learning presents the challenge of generalizing to unseen tasks with limited data . The problem is aimed at learning quickly from few examples of data , which is generally considered the hallmark of human intelligence . This is an important practical problem due to the scarce availability of f... | This paper 1) proposes a simple transfer learning approach to enable train adversarially robust few-shot classifiers for few-shot image classification, and 2) present a method for novel classification task based on calibrating the centroid of the few-shot category towards the base classes. Results show good performance... | SP:b27d0bb34999cb1d197f68cca7e4c01c433ed6e8 |
NAS-Bench-360: Benchmarking Diverse Tasks for Neural Architecture Search | 1 INTRODUCTION . Neural architecture search ( NAS ) aims to automate the design of deep neural networks , ensuring performance on par with hand-crafted architectures while reducing human labor devoted to tedious architecture tuning ( Elsken et al. , 2019 ) . With the growing number of application areas of ML , and thus... | This paper proposes a new benchmark for NAS methods, which is called NAS-Bench-360. Unlike the existing benchmark datasets for NAS, the proposed benchmark contains ten diverse tasks derived from various fields of research. This paper has tested several standard NAS methods on the proposed benchmark and confirmed that t... | SP:5893f3dba5c2341a1e9dad1002d7ac226417c026 |
NAS-Bench-360: Benchmarking Diverse Tasks for Neural Architecture Search | 1 INTRODUCTION . Neural architecture search ( NAS ) aims to automate the design of deep neural networks , ensuring performance on par with hand-crafted architectures while reducing human labor devoted to tedious architecture tuning ( Elsken et al. , 2019 ) . With the growing number of application areas of ML , and thus... | This paper proposes a benchmark to test the performance of NAS algorithms and search spaces on a diverse set of tasks. The benchmark consists of 10 different datasets across different modalities. On these tasks, a variety of NAS algorithm as well as search spaces are allowed a fixed amount of compute resources (in term... | SP:5893f3dba5c2341a1e9dad1002d7ac226417c026 |
Can Stochastic Gradient Langevin Dynamics Provide Differential Privacy for Deep Learning? | 1 INTRODUCTION Machine learning and , specifically , deep learning models show state-of-the-art results in various fields such as computer vision , natural language processing , and signal processing ( e.g. , Carion et al . ( 2020 ) ; Devlin et al . ( 2019 ) ; Balevi & Andrews ( 2021 ) ) . Training these models require... | This paper shows that even when the posterior is as private as targeted in the beginning, sampling from posterior with SGLS might not be as private as targeted. The authors prove the theorem on Bayesian linear regression problem. They prove that for n big enough sampling from the posterior is (\epsilon, \delta) differe... | SP:e11a3ee0dce61bc8647a345d8947b9d36e2323f8 |
Can Stochastic Gradient Langevin Dynamics Provide Differential Privacy for Deep Learning? | 1 INTRODUCTION Machine learning and , specifically , deep learning models show state-of-the-art results in various fields such as computer vision , natural language processing , and signal processing ( e.g. , Carion et al . ( 2020 ) ; Devlin et al . ( 2019 ) ; Balevi & Andrews ( 2021 ) ) . Training these models require... | This paper studies the privacy guarantee of Bayesian learning using Stochastic Gradient Langevin Dynamics (SGLD). Since the SGLD updates are stochastic, it is often thought the solution can be suitable for privacy-preserving of the data that is used to train the algorithm. Using a counter-example, this paper shows that... | SP:e11a3ee0dce61bc8647a345d8947b9d36e2323f8 |
Better state exploration using action sequence equivalence | 1 INTRODUCTION . Despite the rapidly improving performance of Reinforcement Learning ( RL ) agents on a variety of tasks ( Mnih et al. , 2015 ; Silver et al. , 2016 ) , they remain largely sample-inefficient learners compared to humans ( Toromanoff et al. , 2019 ) . Contributing to this is the vast amount of prior know... | In the context of reinforcement learning, authors propose an exploration strategy based on environment-specific prior knowledge of action equivalence. An example of such equivalence is rotating 180° twice in a grid world, as the agent comes back to the same original state: the action sequence forms an identity in this ... | SP:1f35871b8ec295dd84e991fdc57a45024bc07607 |
Better state exploration using action sequence equivalence | 1 INTRODUCTION . Despite the rapidly improving performance of Reinforcement Learning ( RL ) agents on a variety of tasks ( Mnih et al. , 2015 ; Silver et al. , 2016 ) , they remain largely sample-inefficient learners compared to humans ( Toromanoff et al. , 2019 ) . Contributing to this is the vast amount of prior know... | The paper proposes a method that, from a simple encoding of sequences of actions that have equivalent outcome in an MDP, allows to compute a local policy for local high-quality exploration (it replaces the random action of $\varepsilon$-greedy with an action that maximizes the entropy of future visited states). Full al... | SP:1f35871b8ec295dd84e991fdc57a45024bc07607 |
Personalized Heterogeneous Federated Learning with Gradient Similarity | 1 INTRODUCTION . With the popularity of smartphones , personal computers , and other devices , the data stored on them has increased dramatically . These data are related to each other but exist independently on different devices . Also , their owners are often not willing to share their private data , which prompted F... | This paper proposed a personalized federated learning algorithm which takes into account the similarity of gradient of different users to update the model. More formally, the authors define $\tilde{S}(i,j)$ as a measure of similarity between the gradients of two user $i$ and $j$, and then update the model of user $i$by... | SP:edda5940fcfe72533d9925b2d73f5ed4c411e4bb |
Personalized Heterogeneous Federated Learning with Gradient Similarity | 1 INTRODUCTION . With the popularity of smartphones , personal computers , and other devices , the data stored on them has increased dramatically . These data are related to each other but exist independently on different devices . Also , their owners are often not willing to share their private data , which prompted F... | This paper proposes two methods for personalized federated learning, one synchronous and one asynchronous. The general approach taken in both cases is to adapt the weights when averaging information from different clients, so that clients with more similar gradients are given more weight in the update for each client. ... | SP:edda5940fcfe72533d9925b2d73f5ed4c411e4bb |
Continual Backprop: Stochastic Gradient Descent with Persistent Randomness | 1 INTRODUCTION . In the last decade , deep learning methods have been successful and become the state-of-the-art in many machine learning problems and applications , including supervised classification , reinforcement learning ( Silver et al. , 2016 ) , computer vision ( Krizhevsky et al. , 2012 ) , and natural languag... | This paper demonstrates and proposes a solution for a new problem in continual learning which is the inverse of catastrophic forgetting. Compared to prior work, they study problems where the data distribution changes much more rapidly. They demonstrate that backpropagation based optimization loses its ability to adapt ... | SP:2416c3d070cc9b54e096cc57687749731f3b9193 |
Continual Backprop: Stochastic Gradient Descent with Persistent Randomness | 1 INTRODUCTION . In the last decade , deep learning methods have been successful and become the state-of-the-art in many machine learning problems and applications , including supervised classification , reinforcement learning ( Silver et al. , 2016 ) , computer vision ( Krizhevsky et al. , 2012 ) , and natural languag... | This paper investigates the problem of fast adaptation in a non-stationary online continual learning(CL) setting. It argues that keeping weight randomnization is important to fast adaptation in CL. However, current CL methods only performs weight randomization in the beginning of the algorithm; the weights loss randomn... | SP:2416c3d070cc9b54e096cc57687749731f3b9193 |
Omni-Dimensional Dynamic Convolution | 1 INTRODUCTION . In the past decade , we have witnessed the tremendous success of deep Convolutional Neural Networks ( CNNs ) in many computer vision applications ( Krizhevsky et al. , 2012 ; Girshick et al. , 2014 ; Long et al. , 2015 ; He et al. , 2017 ) . The most common way of constructing a deep CNN is to stack a ... | This work mainly focuses on designing a new dynamic network for large-scale image recognition problems. Specifically, the author discussed the weakness of the existing dynamic convolution operations, and based on the analysis the author proposed a novel framework with the name ODConv. Extensive experiments confirm the ... | SP:8467ec8e80c64d6648e1053b1f7cb593de940132 |
Omni-Dimensional Dynamic Convolution | 1 INTRODUCTION . In the past decade , we have witnessed the tremendous success of deep Convolutional Neural Networks ( CNNs ) in many computer vision applications ( Krizhevsky et al. , 2012 ; Girshick et al. , 2014 ; Long et al. , 2015 ; He et al. , 2017 ) . The most common way of constructing a deep CNN is to stack a ... | The authors present ODConv, a type of dynamic convolutional operation. ODConv combines two prior ideas, i.e. (1) filter recalibration with attention in SENet and (2) additive kernels in CondConv/DyConv, and also generalizes to all remaining dimensions of convolutional filters. The authors propose to use ODConv as drop-... | SP:8467ec8e80c64d6648e1053b1f7cb593de940132 |
The Manifold Hypothesis for Gradient-Based Explanations | 1 INTRODUCTION . A large number of algorithms aim to provide post-hoc explanations for the output of neural networks ( Simonyan et al. , 2014 ; Bach et al. , 2015 ; Shrikumar et al. , 2017 ; Ancona et al. , 2018 ; Lim et al. , 2021 ) . Many of them are , directly or indirectly , based on the gradient with respect to th... | The paper argues that the main reason (or a good reason) for the "meaningfulness" of a gradient with data manifold. The authors perform a set of controlled experiments with different feature attribution methods. Finally, they theoretically show that alignment of the gradient with data manifold has nothing to do with ge... | SP:c3276f7bbc7faa158569f67c2cd806e4154e0048 |
The Manifold Hypothesis for Gradient-Based Explanations | 1 INTRODUCTION . A large number of algorithms aim to provide post-hoc explanations for the output of neural networks ( Simonyan et al. , 2014 ; Bach et al. , 2015 ; Shrikumar et al. , 2017 ; Ancona et al. , 2018 ; Lim et al. , 2021 ) . Many of them are , directly or indirectly , based on the gradient with respect to th... | The paper constructs a synthetic classification task with a known manifold structure by training the classifier with data from a variational autoencoder with a low-dimensional latent space. The paper argues that the components of image gradients that lie in the tangent space of the data manifold are semantically meanin... | SP:c3276f7bbc7faa158569f67c2cd806e4154e0048 |
Deep learning via message passing algorithms based on belief propagation | 1 INTRODUCTION . Belief Propagation is a method for computing marginals and entropies in probabilistic inference problems ( Bethe , 1935 ; Peierls , 1936 ; Gallager , 1962 ; Pearl , 1982 ) . These include optimization problems as well once they are written as zero temperature limit of a Gibbs distribution that uses the... | This paper introduces a belief-propagation message-passing training algorithm for multi-layer neural networks. This algorithm is adapted to mini-batch training and biases distributions toward high entropy solutions. Empirical results show that neural networks with discrete weights and activations trained with this algo... | SP:a736d2fa98e58e22b69e55daf8b678d1583cc7e8 |
Deep learning via message passing algorithms based on belief propagation | 1 INTRODUCTION . Belief Propagation is a method for computing marginals and entropies in probabilistic inference problems ( Bethe , 1935 ; Peierls , 1936 ; Gallager , 1962 ; Pearl , 1982 ) . These include optimization problems as well once they are written as zero temperature limit of a Gibbs distribution that uses the... | This manuscript provides an interesting try on alterative training algorithms for deep neural networks, based on (approximate) message-passing algorithms based on the well-known belief propagation (BP) algorithm. In particular, the binary neural network is considered and four algorithms (BP, three variants of BP, i.e.... | SP:a736d2fa98e58e22b69e55daf8b678d1583cc7e8 |
Demystifying Batch Normalization in ReLU Networks: Equivalent Convex Optimization Models and Implicit Regularization | 1 INTRODUCTION . Deep neural networks have achieved dramatic progress in the past decade . This dramatic progress largely hinged on improvements in terms of optimization techniques . One of the most prominent recent optimization techniques is Batch Normalization ( BN ) ( Ioffe & Szegedy , 2015 ) . BN is an operation th... | The claim of this paper is casting training neural networks with batch normalization to a convex program solvable in polynomial time. The convex reduction sparks an implicit regularization of batch normalization. Taking inspiration from the convex program and the implicit regularization, the authors improve BN. | SP:b392cdc4ce546566457a48e95bfbaea6cad5b44b |
Demystifying Batch Normalization in ReLU Networks: Equivalent Convex Optimization Models and Implicit Regularization | 1 INTRODUCTION . Deep neural networks have achieved dramatic progress in the past decade . This dramatic progress largely hinged on improvements in terms of optimization techniques . One of the most prominent recent optimization techniques is Batch Normalization ( BN ) ( Ioffe & Szegedy , 2015 ) . BN is an operation th... | The paper studies batch normalization in deep neural networks. For a two-layer network with scalar output and batch normalization, a dual of the problem is derived. It is then shown that in the high-dimensional regime, the dual can be further simplified so that an optimal solution can be computed in closed form. In the... | SP:b392cdc4ce546566457a48e95bfbaea6cad5b44b |
Blur Is an Ensemble: Spatial Smoothings to Improve Accuracy, Uncertainty, and Robustness | Bayesian neural networks ( BNNs ) have shown success in the areas of uncertainty1 estimation and robustness . However , a crucial challenge prohibits their use in2 practice . Bayesian NNs require a large number of predictions to produce reliable3 results , leading to a significant increase in computational cost . To al... | This paper proposed a spatial `smooth` layer including a feature range bounding layer `prob` and `blur` the intermediate feature map in a CNN. 'Smooth' improves the accuracy and uncertainty of both deterministic CNN and a Bayesian NN approximated by MC-dropout. Authors tried to justify how `smooth` improves the optimiz... | SP:58d0e331b89085a01a2c56ec63efb1126f616846 |
Blur Is an Ensemble: Spatial Smoothings to Improve Accuracy, Uncertainty, and Robustness | Bayesian neural networks ( BNNs ) have shown success in the areas of uncertainty1 estimation and robustness . However , a crucial challenge prohibits their use in2 practice . Bayesian NNs require a large number of predictions to produce reliable3 results , leading to a significant increase in computational cost . To al... | The motivation of this work is on the computational cost of using BNNs in practice, where applications might require a large number of BNNs in an ensemble formation for achieving good performance. The work in this manuscript aims to reduce the computational cost ensemble. The manuscript's insight for solving the compu... | SP:58d0e331b89085a01a2c56ec63efb1126f616846 |
Safe Deep RL in 3D Environments using Human Feedback | Agents should avoid unsafe behaviour during both training and deployment . This typically requires a simulator and a procedural specification of unsafe behaviour . Unfortunately , a simulator is not always available , and procedurally specifying constraints can be difficult or impossible for many real-world tasks . A r... | ### Contributions * This paper proposes a safe model-based deep RL approach where * No simulator is needed. The dynamic model is learned from data. * No constraint is specified. * This work is an extension of reward query synthesis via trajectory optimization (ReQueST). * ReQueSt (previous work) * Algorit... | SP:d8fe4568447b255f04befefa320abc6d2f32ccdc |
Safe Deep RL in 3D Environments using Human Feedback | Agents should avoid unsafe behaviour during both training and deployment . This typically requires a simulator and a procedural specification of unsafe behaviour . Unfortunately , a simulator is not always available , and procedurally specifying constraints can be difficult or impossible for many real-world tasks . A r... | The paper proposes an extension to ReQueST, which learn learns a neural simulator of the environment from safe human trajectories and then learns a reward model from human feedback. This work extends ReQueST by dense reward sketches on imagines trajectories and evaluates the idea on a visually-complex 3D environment. T... | SP:d8fe4568447b255f04befefa320abc6d2f32ccdc |
Communicate Then Adapt: An Effective Decentralized Adaptive Method for Deep Training | 1 INTRODUCTION . Decentralized SGD ( Lopes & Sayed , 2008 ; Nedic & Ozdaglar , 2009 ; Chen & Sayed , 2012 ; Lian et al. , 2017 ; Assran et al. , 2019 ) is an emerging training approach for deep learning known for its much less communication overhead . In contrast to parallel SGD in which a global averaging across all c... | This paper proposes a decentralized adaptive method for distributed deep learning, termed DAG-Adam. Convergence results are provided for smooth non-convex objectives under a bounded gradient assumption. Numerical experiments are conducted on Image Classification (CIFAR10, ImageNet-1k) and Language Modelling (fine-tunin... | SP:20b3645f16dca8252b31b43661e565d180529a4e |
Communicate Then Adapt: An Effective Decentralized Adaptive Method for Deep Training | 1 INTRODUCTION . Decentralized SGD ( Lopes & Sayed , 2008 ; Nedic & Ozdaglar , 2009 ; Chen & Sayed , 2012 ; Lian et al. , 2017 ; Assran et al. , 2019 ) is an emerging training approach for deep learning known for its much less communication overhead . In contrast to parallel SGD in which a global averaging across all c... | This paper developed a new decentralized adaptive gradient descent method to address the data heterogeneity problem. The motivation is clear and the experimental results show improvement over existing methods. However, the theoretical analysis is not solid. | SP:20b3645f16dca8252b31b43661e565d180529a4e |
1-bit LAMB: Communication Efficient Large-Scale Large-Batch Training with LAMB's Convergence Speed | 1 INTRODUCTION . Training large-scale deep learning models in a distributed fashion is computation-heavy and expensive ( Brown et al. , 2020 ) . In addition to computation , communication overhead becomes a serious system challenge for such large models . A recent study of BERT pre-training with Adam demonstrates that ... | This work studies the problem of distributed training with large batches in a communication-bottlenecked setup, where regular versions of algorithms such as LAMB become a constraint. Authors propose a new algorithm, which compresses the gradient momentum before aggregation and then reconstructs the gradients to recover... | SP:348371af70bb81a998b7dcfc8de2d60ea9b506e5 |
1-bit LAMB: Communication Efficient Large-Scale Large-Batch Training with LAMB's Convergence Speed | 1 INTRODUCTION . Training large-scale deep learning models in a distributed fashion is computation-heavy and expensive ( Brown et al. , 2020 ) . In addition to computation , communication overhead becomes a serious system challenge for such large models . A recent study of BERT pre-training with Adam demonstrates that ... | The paper proposes a communication-efficient distributed LAMB optimizer with 1-bit compression. It follows previous work to first warm-up the variance, but proposes to inference the scaling factor based on reconstructed variance. Experiments show training speedup due to communication compression. The proposed 1-bit LAM... | SP:348371af70bb81a998b7dcfc8de2d60ea9b506e5 |
Contact Points Discovery for Soft-Body Manipulations with Differentiable Physics | 1 INTRODUCTION . Soft body manipulation has a wide application in cooking ( Bollini et al. , 2013 ) , fabric manipulation ( Wu et al. , 2020 ) , healthcare ( Mayer et al. , 2008 ) and manufacturing of deformable objects ( Sanchez et al. , 2018 ) . Differentiable physics has recently been shown as a powerful and effecti... | The paper proposed an algorithm to discover appropriate contact points for deformable object manipulation. A key component of the proposed algorithm is to use an optimal-transport approach that computes a transport priority score for each particle in the deformable body. This score is then used to guide a grid search p... | SP:92e58feb55f1f058d36bac600b9f8f196fe4cc43 |
Contact Points Discovery for Soft-Body Manipulations with Differentiable Physics | 1 INTRODUCTION . Soft body manipulation has a wide application in cooking ( Bollini et al. , 2013 ) , fabric manipulation ( Wu et al. , 2020 ) , healthcare ( Mayer et al. , 2008 ) and manufacturing of deformable objects ( Sanchez et al. , 2018 ) . Differentiable physics has recently been shown as a powerful and effecti... | This paper proposes a method to solve the multistage manipulation tasks for soft materials. Differentiable physics has been used in controlling and manipulating soft materials since DiffTaichi. However, optimization methods based on local gradient information can be easily trapped in local minima. This paper divides a ... | SP:92e58feb55f1f058d36bac600b9f8f196fe4cc43 |
Accuracy-Privacy Trade-off in Deep Ensemble: A Membership Inference Perspective | 1 INTRODUCTION . Ensemble learning has been shown to improve classification accuracy of neural networks in particular , and machine learning classifiers in general ( Kondratyuk et al. , 2020 ; Kuncheva & Whitaker , 2003 ; Sagi & Rokach , 2018 ) . The most commonly used approach for deep models involves averaging the ou... | This work analyzes the accuracy-privacy trade-off in ensemble learning by performing model inference attacks. The key finding of the paper is that the presence of an ensemble (that averages the predictions of individual learners) exacerbates the disparity between the confidence distribution of samples that were seen du... | SP:a492824ed04e34de0d1a54373e4cc15348c14a45 |
Accuracy-Privacy Trade-off in Deep Ensemble: A Membership Inference Perspective | 1 INTRODUCTION . Ensemble learning has been shown to improve classification accuracy of neural networks in particular , and machine learning classifiers in general ( Kondratyuk et al. , 2020 ; Kuncheva & Whitaker , 2003 ; Sagi & Rokach , 2018 ) . The most commonly used approach for deep models involves averaging the ou... | This paper provide a systemantic analysis on the accuracy-privacy trade off for deep ensmebles. They show that the effectiveness of membership inference attacks is likely to increase when ensembling improves accuracy. The authors further study the impact of various factors such as prediction confidence and agreement b... | SP:a492824ed04e34de0d1a54373e4cc15348c14a45 |
Autoregressive Quantile Flows for Predictive Uncertainty Estimation | 1 INTRODUCTION . Reasoning about uncertainty via the language of probability is important in many application domains of machine learning , including medicine ( Saria , 2018 ) , robotics ( Chua et al. , 2018 ; Buckman et al. , 2018 ) , and operations research ( Van Roy et al. , 1997 ) . Especially important is the esti... | This paper proposed a quantile regression method for uncertainty estimation based on autoregressive quantile flow. The flow model can be trained in both forward and reverse setting using different loss functions, and the quantile flow framework can be combined with other linear or non-linear transformations. The author... | SP:52701ccbe77facd26fa921a2610dad1da60e1a5f |
Autoregressive Quantile Flows for Predictive Uncertainty Estimation | 1 INTRODUCTION . Reasoning about uncertainty via the language of probability is important in many application domains of machine learning , including medicine ( Saria , 2018 ) , robotics ( Chua et al. , 2018 ; Buckman et al. , 2018 ) , and operations research ( Van Roy et al. , 1997 ) . Especially important is the esti... | This paper proposes a novel framework for training flow models named Autoregressive Quantile Flows (AQF). The proposed method utilizes a new objective by evaluating forecasts with proper scoring rules, including the continuous ranked probability score and the check score. The advantages of the proposed objective are 1)... | SP:52701ccbe77facd26fa921a2610dad1da60e1a5f |
New Perspective on the Global Convergence of Finite-Sum Optimization | 1 INTRODUCTION . In recent years , deep neural networks ( DNNs ) have shown a great success in many machine learning tasks . However , training these neural networks is challenging since the loss surface of network architecture is generally non-convex , or even non-smooth . Thus , there have been a long-standing questi... | This paper presents a new optimization method for finding global minima of nonconvex finite sum problems. In particular, the summands are functions of the form $\phi_{i}\circ h$ where $\phi_{i}$ is convex and Lipschitz smooth, while $h$ is nonconvex. Each iteration of the method consists of solving an auxiliary regular... | SP:030b2045318e6e4189685793b5eab37ffb8b1a82 |
New Perspective on the Global Convergence of Finite-Sum Optimization | 1 INTRODUCTION . In recent years , deep neural networks ( DNNs ) have shown a great success in many machine learning tasks . However , training these neural networks is challenging since the loss surface of network architecture is generally non-convex , or even non-smooth . Thus , there have been a long-standing questi... | The paper provides a new gradient-based algorithm. The algorithm is based on the observation that a loss function for a single sample can be written as composition of two functions (the logits and the actual loss function). It computes the direction by means of solving a quadratic MSE problem. They provide a convergent... | SP:030b2045318e6e4189685793b5eab37ffb8b1a82 |
Efficient representations for privacy-preserving inference | 1 INTRODUCTION . In recent years , deep neural networks have achieved state-of-the-art accuracy for tasks such as image recognition . They have been deployed in a range of sectors , powering a wide variety of applications such as recommendation systems , medical diagnosis , and content filtering . Machine Learning as a... | The paper considers the problem of privacy preserving inference on deep learning models using homomorphic encryption. HE is a special type of encryption that allows one to perform certain types of computations while the data is encrypted. However, the catch is HE based inferences can be significantly slower than the no... | SP:89a1b45eb1420f7259acaf8289fcd30523941e03 |
Efficient representations for privacy-preserving inference | 1 INTRODUCTION . In recent years , deep neural networks have achieved state-of-the-art accuracy for tasks such as image recognition . They have been deployed in a range of sectors , powering a wide variety of applications such as recommendation systems , medical diagnosis , and content filtering . Machine Learning as a... | To accelerate the privacy-preserving inference through convolution neural networks (CNNs) with homomorphic encryption (HE), the authors aim to reduce the number of homomorphic operations (HOPs) required for the algorithm to save the data needs to be transferred while preserving prediction accuracy. Using a LOLA method... | SP:89a1b45eb1420f7259acaf8289fcd30523941e03 |
NormFormer: Improved Transformer Pretraining with Extra Normalization | 1 INTRODUCTION . The original transformer architecture ( Vaswani et al. , 2017 ) applies Layer Normalization ( Ba et al. , 2016 ) after each sublayer ’ s residual connection ( “ Post-LN ” ) in order to reduce the variance of the inputs to the following sublayer , i.e . : PostLN ( x ) = LayerNorm ( x+ Sublayer ( x ) ) ,... | This paper aims to improve pretraining Pre-LayerNorm transformers by alleviating two issues: early layers have much larger gradients than later ones, and naive residual learning can't provide optimal weighting. To this end, it proposes to add two LayerNorms after the multi-head attention and the GELU non-linear activat... | SP:45ba88126844e65868d6284c7175a9893ccaf67e |
NormFormer: Improved Transformer Pretraining with Extra Normalization | 1 INTRODUCTION . The original transformer architecture ( Vaswani et al. , 2017 ) applies Layer Normalization ( Ba et al. , 2016 ) after each sublayer ’ s residual connection ( “ Post-LN ” ) in order to reduce the variance of the inputs to the following sublayer , i.e . : PostLN ( x ) = LayerNorm ( x+ Sublayer ( x ) ) ,... | NormFormer improves on Pre-LN transformers by making the following modifications: learnable scaling parameters for each dimension of the output of each attention head prior to concatenation across heads (*Scaled Attention*); layer norm on the attention output (*Post Attn LN*); layer norm on the FFN nonlinearity output ... | SP:45ba88126844e65868d6284c7175a9893ccaf67e |
Local Calibration: Metrics and Recalibration | Probabilistic classifiers output confidence scores along with their predictions , and these confidence scores should be calibrated , i.e. , they should reflect the reliability of the prediction . Confidence scores that minimize standard metrics such as the expected calibration error ( ECE ) accurately measure the relia... | This work proposes a new metric for calibration in classification where calibration is measured over localities in the input space, and the localities are determined with a kernel over the feature space. A recalibration algorithm (LoRe) is additionally proposed, which aims to recalibrate the class predictions w.r.t. lo... | SP:4c8f0cf7f6196f586ec83d16a768742d13a16cea |
Local Calibration: Metrics and Recalibration | Probabilistic classifiers output confidence scores along with their predictions , and these confidence scores should be calibrated , i.e. , they should reflect the reliability of the prediction . Confidence scores that minimize standard metrics such as the expected calibration error ( ECE ) accurately measure the relia... | This paper proposes a new measure of calibration called Local Calibration. While conventional calibration measures are only defined with probabilistic outputs, the proposed local calibration measure further incorporates the feature space by considering the neighbouring region with a kernel. The authors also propose a c... | SP:4c8f0cf7f6196f586ec83d16a768742d13a16cea |
X-model: Improving Data Efficiency in Deep Learning with A Minimax Model | 1 INTRODUCTION . In the last decade , deep learning has become the de facto choice for numerous machine learning applications in the presence of large-scale labeled datasets . However , collecting sufficient labeled data through manual labeling , especially for deep regression tasks such as keypoint localization and ag... | The paper focuses on reducing data labeling efforts by improving data efficiency. In contrast to most existing approaches that address this problem only in the classification setup, the paper focuses on both classification and regression set ups. The proposed method primarily is built on leveraging invariance to data ... | SP:fe04f7ffacf4dfa43448503ac2fa7a5f7f14ab3d |
X-model: Improving Data Efficiency in Deep Learning with A Minimax Model | 1 INTRODUCTION . In the last decade , deep learning has become the de facto choice for numerous machine learning applications in the presence of large-scale labeled datasets . However , collecting sufficient labeled data through manual labeling , especially for deep regression tasks such as keypoint localization and ag... | The paper presents a g data-efficient approach that encourages invariance to both data and model stochasticity that works for both classification and regression tasks. Furthermore, the proposed minimax loss function can specifically enhance invariance to model stochasticity. The extensive experimental results verify th... | SP:fe04f7ffacf4dfa43448503ac2fa7a5f7f14ab3d |
A General Theory of Relativity in Reinforcement Learning | 1 INTRODUCTION . Deep reinforcement learning ( RL ) has demonstrated its great successes in recent years , including breakthrough of solving a number of challenging problems like Atari ( Mnih et al. , 2015 ) , GO ( Silver et al. , 2016 ; 2017 ) , DOTA2 ( Berner et al. , 2019 ) and StarCraft II ( Vinyals et al. , 2019 )... | The paper studies transfer in reinforcement learning (RL), beginning with a theorem that relates the performance of one policy under a particular dynamics to another policy under different dynamics. This is broken down into a “dynamics-induced gap” and a “policy-induced gap”, for which explicit expressions are given. O... | SP:56a799994baed2b0f32c40c7586cb50c8a43f855 |
A General Theory of Relativity in Reinforcement Learning | 1 INTRODUCTION . Deep reinforcement learning ( RL ) has demonstrated its great successes in recent years , including breakthrough of solving a number of challenging problems like Atari ( Mnih et al. , 2015 ) , GO ( Silver et al. , 2016 ; 2017 ) , DOTA2 ( Berner et al. , 2019 ) and StarCraft II ( Vinyals et al. , 2019 )... | This paper proposed a way to decompose the difference between values of two policies in two MDPs respectively. Such a decomposition results in two parts, the first one is the difference between values of one policy under two MDPs; the second part is the difference between values of two policies under the same MDP. Usin... | SP:56a799994baed2b0f32c40c7586cb50c8a43f855 |
Flashlight: Enabling Innovation in Tools for Machine Learning | 1 INTRODUCTION . The recent rise of deep learning-based techniques has been accompanied and sustained by the wide availability of dedicated frameworks such as TensorFlow ( Abadi et al. , 2016 ) and PyTorch ( Paszke et al. , 2019 ) . These frameworks have enabled the democratization of machine learning research by provi... | The paper proposed a minimal design API (or mostly API?) of machine learning framework called Flashlight. The key argument of the paper is that Flashlight is modular and agile. The Flashlight captured the key aspects of machine learning frameworks: Tensor and Operation, Memory Management, and Distributed. There are som... | SP:97873277c2891819393aeebbd3256b7445794e89 |
Flashlight: Enabling Innovation in Tools for Machine Learning | 1 INTRODUCTION . The recent rise of deep learning-based techniques has been accompanied and sustained by the wide availability of dedicated frameworks such as TensorFlow ( Abadi et al. , 2016 ) and PyTorch ( Paszke et al. , 2019 ) . These frameworks have enabled the democratization of machine learning research by provi... | The paper describes the design philosophy and structure of the Flashlight deep learning framework. Flashlight is modular, small, and narrowly oriented toward systems researchers. Rather than (or in addition to) high level productivity, Flashlight focuses on internal and external simplicity of ML tools. The authors eval... | SP:97873277c2891819393aeebbd3256b7445794e89 |
Constructing a Good Behavior Basis for Transfer using Generalized Policy Updates | 1 INTRODUCTION . Reinforcement learning ( RL ) studies the problem of building rational decision-making agents that maximize long term cumulative reward through trial-and-error interaction with a given environment . In recent years , RL algorithms combined with powerful function approximators such as deep neural networ... | The paper extends the successor features framework to answer the following question: which policies should we learn and store so that, when presented with a new task, we achieve the best performance possible? The paper defines the notion of independent policies (forming a kind of basis over policy space), which can the... | SP:fbacc4b906328e10a7f61a351bc02cf99aa33c4c |
Constructing a Good Behavior Basis for Transfer using Generalized Policy Updates | 1 INTRODUCTION . Reinforcement learning ( RL ) studies the problem of building rational decision-making agents that maximize long term cumulative reward through trial-and-error interaction with a given environment . In recent years , RL algorithms combined with powerful function approximators such as deep neural networ... | The paper focuses on reinforcement learning problems with known successor features and rewards expressible as their linear combination. Building on recent research, it presents a concept of independent features and independent policies and way to construct them. Theoretically it shows that the set of independent polici... | SP:fbacc4b906328e10a7f61a351bc02cf99aa33c4c |
Learning the Dynamics of Physical Systems from Sparse Observations with Finite Element Networks | 1 INTRODUCTION The laws driving the physical world are often best described by partial differential equations ( PDEs ) that relate how a magnitude of interest changes in time with its change in space . They describe how the atmosphere and oceans circulate and interact , how structures deform under load and how electrom... | The author proposes a method for forecasting in Partial Differential Equations by coupling Finite Element Method on an arbitrary grid with the learning of the dynamics from data. For this purpose a variant of message passing based graph networks is used. It is show in the paper that it's possible to incorporate priors ... | SP:a76c1a2b18015e647fa687abbb2840e2426b31f8 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.