paper_name stringlengths 11 170 | text stringlengths 8.07k 307k | summary stringlengths 152 6.16k | paper_id stringlengths 43 43 |
|---|---|---|---|
Noisy Machines: Understanding noisy neural networks and enhancing robustness to analog hardware errors using distillation | The success of deep learning has brought forth a wave of interest in computer hardware design to better meet the high demands of neural network inference . In particular , analog computing hardware has been heavily motivated specifically for accelerating neural networks , based on either electronic , optical or photoni... | The article on "Noisy Machines" addresses the issue of implementing deep neural network inference on a noisy hardware computing substrate, e.g. analog accelerators. This is an important topic because analog devices allow fast and energy efficient inference, which is crucial for inference at the edge. Because of their a... | SP:cc73a630ce68477bde408cc08a92a4f98eb2c597 |
Noisy Machines: Understanding noisy neural networks and enhancing robustness to analog hardware errors using distillation | The success of deep learning has brought forth a wave of interest in computer hardware design to better meet the high demands of neural network inference . In particular , analog computing hardware has been heavily motivated specifically for accelerating neural networks , based on either electronic , optical or photoni... | The manuscript illustrates how a noisy neural network can reduce the learning capacity. To mitigate this loss, the authors propose a method that combines the method of "noise injection and "knowledge distillation". However, from a conceptual point of view, their contribution (i.e. (10) in Section 5,) is unclear to me.... | SP:cc73a630ce68477bde408cc08a92a4f98eb2c597 |
Neural ODEs for Image Segmentation with Level Sets | 1 INTRODUCTION . Image segmentation is the task of delineating pixels belonging to semantic labels . The ability to automatically segment objects is important because accurate labeling is expensive and hard ( Vittayakorn & Hays , 2011 ; Zhang et al. , 2018 ) . Automatic image segmentation can have large impact in many ... | This paper proposes to utilize Neural ODEs (NODEs) and the Level Set Method (LSM) for the task of image segmentation. The argument is that the NODE can be used to learn the force function in an LSM and solve the contour evolution process. The authors propose two architectures and demonstrate promising performance on a... | SP:16cb7d0da739f1e6a72efb9b18399d2d8b69f540 |
Neural ODEs for Image Segmentation with Level Sets | 1 INTRODUCTION . Image segmentation is the task of delineating pixels belonging to semantic labels . The ability to automatically segment objects is important because accurate labeling is expensive and hard ( Vittayakorn & Hays , 2011 ; Zhang et al. , 2018 ) . Automatic image segmentation can have large impact in many ... | This paper proposes to apply the Neural ODE framework (Chen et al 2018) for image segmentation. The method relies on contour delineation through Level Sets. Since contour estimation requires to solve an ODE, this naturally allows to apply the work presented in (Chen et al 2018). The method is here applied in two segmen... | SP:16cb7d0da739f1e6a72efb9b18399d2d8b69f540 |
Unified Probabilistic Deep Continual Learning through Generative Replay and Open Set Recognition | 1 INTRODUCTION . Most machine learning systems make the closed world assumption and are predominantly trained according to the isolated learning paradigm , where data is available at all times and is independently and identically distributed . However , in the context of continual learning , where tasks and data arrive... | This paper tackles the problem of catastrophic forgetting when data is organized in a large number of batches of data (tasks) that are sequentially made available. To avoid catastrophic forgetting, the authors learn a VAE that generates the training data (both inputs and labels) and retrain it using samples from the ne... | SP:34f3abe09b1ca5c5bbbf1a2e28b489fee010098e |
Unified Probabilistic Deep Continual Learning through Generative Replay and Open Set Recognition | 1 INTRODUCTION . Most machine learning systems make the closed world assumption and are predominantly trained according to the isolated learning paradigm , where data is available at all times and is independently and identically distributed . However , in the context of continual learning , where tasks and data arrive... | This paper combines replay and openMax approach to help continual learning. The results shows robustness on different dataset include image and audio in the continual learning condition, where the new come data has a different distribution but the model still able to maintain reasonable quality for the previously and ... | SP:34f3abe09b1ca5c5bbbf1a2e28b489fee010098e |
Asymptotic learning curves of kernel methods: empirical data v.s. Teacher-Student paradigm | 1 INTRODUCTION . In supervised learning machines learn from a finite collection of n training data , and their generalization error is then evaluated on unseen data drawn from the same distribution . How many data are needed to learn a task is characterized by the learning curve relating generalization error to n. In v... | This paper experimentally investigates how fast the generalization error decreases when some specific kernel functions are used in real datasets. This paper conducted numerical experiments on several datasets to investigate the decreasing rate of the generalization error, and the rate is determined for such datasets. T... | SP:655be2d7f8ffe68416e0c3a5b4218ffe45a37bfc |
Asymptotic learning curves of kernel methods: empirical data v.s. Teacher-Student paradigm | 1 INTRODUCTION . In supervised learning machines learn from a finite collection of n training data , and their generalization error is then evaluated on unseen data drawn from the same distribution . How many data are needed to learn a task is characterized by the learning curve relating generalization error to n. In v... | This paper studies, empirically and theoretically, the learning rates of (shift-invariant) kernel learners in a misspecified setting. In the well-specified setting, the rate of kernel learners is at least $n^{-1/2}$, and in a misspecified setting assuming only Lipschitz targets, the rate is $n^{-1/d}$. Neither seems to... | SP:655be2d7f8ffe68416e0c3a5b4218ffe45a37bfc |
On the "steerability" of generative adversarial networks | 1 INTRODUCTION . The quality of deep generative models has increased dramatically over the past few years . When introduced in 2014 , Generative Adversarial Networks ( GANs ) could only synthesize MNIST digits and low-resolution grayscale faces ( Goodfellow et al. , 2014 ) . The most recent models , however , produce d... | This work explores the extent to which the natural image manifold is captured by generative adversarial networks (GANs) by performing walks in the latent space of pretrained models. To perform these walks, a transformation vector is learned by minimizing the distance between transformed images and the corresponding ima... | SP:ad4e8d1e16eeeff006f8568bd6bf2c0862621526 |
On the "steerability" of generative adversarial networks | 1 INTRODUCTION . The quality of deep generative models has increased dramatically over the past few years . When introduced in 2014 , Generative Adversarial Networks ( GANs ) could only synthesize MNIST digits and low-resolution grayscale faces ( Goodfellow et al. , 2014 ) . The most recent models , however , produce d... | This paper propose to study the generalization properties of GANs through interpolation. They first propose to learn a linear (and non-linear) interpolation in the latent space for a specific type of image transformation for example zoom, translation, rotation, luminance, etc... They show that linear interpolation in G... | SP:ad4e8d1e16eeeff006f8568bd6bf2c0862621526 |
On Mutual Information Maximization for Representation Learning | 1 INTRODUCTION . Unsupervised representation learning is a fundamental problem in machine learning . Intuitively , one aims to learn a function g which maps the data into some , usually lower-dimensional , space where one can solve some ( generally a priori unknown ) target supervised tasks more efficiently , i.e . wit... | The paper addresses a question on whether mutual information (MI) based models for representation learning succeed primarily thanks to the MI maximization. The motivation of the work comes from the fact that although MI is known to be problematic in treatment, it has been successfully applied in a number of recent work... | SP:1572adb9d8abb9edd60d11e7cdd0e48cfdf5bd4b |
On Mutual Information Maximization for Representation Learning | 1 INTRODUCTION . Unsupervised representation learning is a fundamental problem in machine learning . Intuitively , one aims to learn a function g which maps the data into some , usually lower-dimensional , space where one can solve some ( generally a priori unknown ) target supervised tasks more efficiently , i.e . wit... | This paper gives a nice interpretation why recent works that are based on variational lower bounds of mutual information can demonstrate promising empirical results, where they argue that the success depends on "the inductive biasin both the choice of feature extractor architectures and the parametrization of theemploy... | SP:1572adb9d8abb9edd60d11e7cdd0e48cfdf5bd4b |
Deep Multi-View Learning via Task-Optimal CCA | 1 INTRODUCTION . Parallel modalities of data are increasingly common in a variety of applications , including images and text , audio and video , parallel texts of different languages , and a variety of medical imaging and omics modalities for each patient . Each view provides essential information for classification a... | CCA is a generative model that learns a shared subspace based on two (or multi) views of the data. Being generative, it might not have strong discriminative power for some downstream classification tasks. Previous approaches to infuse discriminative power into the shared subspace estimated by CCA are linear. So, this ... | SP:1bc27efda9dd80ce3cfabd6a1e16a904c85f37fc |
Deep Multi-View Learning via Task-Optimal CCA | 1 INTRODUCTION . Parallel modalities of data are increasingly common in a variety of applications , including images and text , audio and video , parallel texts of different languages , and a variety of medical imaging and omics modalities for each patient . Each view provides essential information for classification a... | This paper addresses the problem of jointly performing CCA with task labeling. The problem is timely and important as it is challenging to perform CCA jointly with the task classification (see below) and hence previous work typically perform this in a pipeline - that is, first projecting the data using a pre-trained CC... | SP:1bc27efda9dd80ce3cfabd6a1e16a904c85f37fc |
Coherent Gradients: An Approach to Understanding Generalization in Gradient Descent-based Optimization | 1 INTRODUCTION AND OVERVIEW . Neural networks used in practice often have sufficient effective capacity to learn arbitrary maps from their inputs to their outputs . This is typically demonstrated by training a classification network that achieves good test accuracy on a real dataset S , on a modified version of S ( cal... | This paper posits that similar input examples will have similar gradients, leading to a gradient "coherence" phenomenon. A simple argument then suggests that the loss should decrease much more rapidly when gradients cohere than when they do not. This hypothesis and analysis is supported with clever experiments that con... | SP:d905f831fd48700ebbaf8286fa71f77b45aa685f |
Coherent Gradients: An Approach to Understanding Generalization in Gradient Descent-based Optimization | 1 INTRODUCTION AND OVERVIEW . Neural networks used in practice often have sufficient effective capacity to learn arbitrary maps from their inputs to their outputs . This is typically demonstrated by training a classification network that achieves good test accuracy on a real dataset S , on a modified version of S ( cal... | The surprising generalization properties of neural networks trained with stochastic gradient descent are still poorly understood. The present work suggests that they can be explained at least partly by the fact that patterns shared across many data points will lead to gradients pointing in similar directions, thus rein... | SP:d905f831fd48700ebbaf8286fa71f77b45aa685f |
Walking the Tightrope: An Investigation of the Convolutional Autoencoder Bottleneck | In this paper , we present an in-depth investigation of the convolutional autoencoder ( CAE ) bottleneck . Autoencoders ( AE ) , and especially their convolutional variants , play a vital role in the current deep learning toolbox . Researchers and practitioners employ CAEs for a variety of tasks , ranging from outlier ... | The authors evaluate convolutional autoencoders (CAE) by varying the size (width & height) and depth of the bottleneck layer on three datasets and compare test and training performance. They furthermore evaluate the quality of the bottleneck activations for linear classification. The authors also investigate the belief... | SP:761207caf0d1b23f060e3957a6309bc6d76819a6 |
Walking the Tightrope: An Investigation of the Convolutional Autoencoder Bottleneck | In this paper , we present an in-depth investigation of the convolutional autoencoder ( CAE ) bottleneck . Autoencoders ( AE ) , and especially their convolutional variants , play a vital role in the current deep learning toolbox . Researchers and practitioners employ CAEs for a variety of tasks , ranging from outlier ... | This paper studies some of the properties of fully convolutional autoencoders (CAE) as a function of the shape and total size of the bottleneck. They train and test CAEs with bottlenecks consisting of different ratios of spatial resolution versus number of channels, as well as different total number of neurons. The aut... | SP:761207caf0d1b23f060e3957a6309bc6d76819a6 |
On the Global Convergence of Training Deep Linear ResNets | 1 INTRODUCTION . Despite the remarkable power of deep neural networks ( DNNs ) trained using stochastic gradient descent ( SGD ) in many machine learning applications , theoretical understanding of the properties of this algorithm , or even plain gradient descent ( GD ) , remains limited . Many key properties of the le... | This paper deals with the global convergence of deep linear ResNets. The author show that under some initialization conditions for the first and the last layer (that are not optimized !) GD and SGD does converge to a global minimum of the min squared error. The closed related work seems to be Bartlett et al. 2019 that ... | SP:4363825dfbd8c5b5a616ea5b0f67a751dcbe7eaf |
On the Global Convergence of Training Deep Linear ResNets | 1 INTRODUCTION . Despite the remarkable power of deep neural networks ( DNNs ) trained using stochastic gradient descent ( SGD ) in many machine learning applications , theoretical understanding of the properties of this algorithm , or even plain gradient descent ( GD ) , remains limited . Many key properties of the le... | In this paper, the authors study the convergence of (stochastic) gradient descent in training deep linear residual networks, where linear transformation at input and output layers are fixed and matrices in other layers are trained. They first establish a global convergence of GD/SGD under some conditions on the fixed l... | SP:4363825dfbd8c5b5a616ea5b0f67a751dcbe7eaf |
HiLLoC: lossless image compression with hierarchical latent variable models | 1 INTRODUCTION . Bits back coding ( Wallace , 1990 ; Hinton & van Camp , 1993 ) is a method for performing lossless compression using a latent variable model . In an ideal implementation , the method can achieve an expected message length equal to the variational free energy , often referred to as the negative evidence... | This paper focuses on lossless source compression with bits back coding for hierarchical fully convolutional VAEs. The focus/contribution is three-fold: 1. Improve the compression rate performance by adapting the discretization of latent space required for the entropy coder ANS. The newly proposed discretization scheme... | SP:cbff5688c7be72a90c6e2ff6e3629c6feac3717c |
HiLLoC: lossless image compression with hierarchical latent variable models | 1 INTRODUCTION . Bits back coding ( Wallace , 1990 ; Hinton & van Camp , 1993 ) is a method for performing lossless compression using a latent variable model . In an ideal implementation , the method can achieve an expected message length equal to the variational free energy , often referred to as the negative evidence... | This paper proposes a method for lossless image compression consisting of a VAE and using a bits-back version of ANS. The results are very impressive on a ImageNet (but maybe not so impressive on the other benchmarks). The authors also discuss how to speed up inference and present some frightening runtime numbers for t... | SP:cbff5688c7be72a90c6e2ff6e3629c6feac3717c |
Adversarially Robust Neural Networks via Optimal Control: Bridging Robustness with Lyapunov Stability | Deep neural networks are known to be vulnerable to adversarial perturbations . In this paper , we bridge adversarial robustness of neural nets with Lyapunov stability of dynamical systems . From this viewpoint , training neural nets is equivalent to finding an optimal control of the discrete dynamical system , which al... | The goal of this paper is to train neural networks (NNs) in a way to be robust to adversarial attacks. The authors formulate training a NN as finding an optimal controller for a discrete dynamical system. This formulation allows them to use an optimal control algorithm, called method of successive approximations (MSA),... | SP:3dd9ae7b88b3e6848ee1fbd11c274d7a395d3167 |
Adversarially Robust Neural Networks via Optimal Control: Bridging Robustness with Lyapunov Stability | Deep neural networks are known to be vulnerable to adversarial perturbations . In this paper , we bridge adversarial robustness of neural nets with Lyapunov stability of dynamical systems . From this viewpoint , training neural nets is equivalent to finding an optimal control of the discrete dynamical system , which al... | Neural Networks are vulnerable to adversarial perturbations. This paper proposes a method that based on optimal control theory that uses semidefinite-programming. This is a quite popular topic in Adversarial training recently, there has been a few works in that line. There are almost no experiments in this paper. There... | SP:3dd9ae7b88b3e6848ee1fbd11c274d7a395d3167 |
Imitation Learning via Off-Policy Distribution Matching | 1 INTRODUCTION . Reinforcement learning ( RL ) is typically framed as learning a behavior policy based on reward feedback from trial-and-error experience . Accordingly , many successful demonstrations of RL often rely on carefully handcrafted rewards with various bonuses and penalties designed to encourage intended beh... | This paper presents an algorithm for adversarial imitation that uses off-policy data in a principled manner, unlike prior work. The core idea is to express the KL-divergence between the policy's state-action marginal and the expert's state action marginal using the Donsker-Varadhan representation and then applying the ... | SP:b63d45fa7937d0efe9d4d471ca75e52114393ea7 |
Imitation Learning via Off-Policy Distribution Matching | 1 INTRODUCTION . Reinforcement learning ( RL ) is typically framed as learning a behavior policy based on reward feedback from trial-and-error experience . Accordingly , many successful demonstrations of RL often rely on carefully handcrafted rewards with various bonuses and penalties designed to encourage intended beh... | This paper provides a novel off policy objective to solve imitation learning. It resolves the limitation of the famous GAIL algorithm that we need on-policy samples to interact with the environment. The new algorithm is simple but efficient, and can handle off-policy settings. The derivation of equation (12) is nice an... | SP:b63d45fa7937d0efe9d4d471ca75e52114393ea7 |
Contextual Inverse Reinforcement Learning | 1 INTRODUCTION . We study sequential decision-making in a Contextual Markov Decision Process ( CMDP , Hallak et al . ( 2015 ) ) , where the reward , while unknown to the agent , depends on a static parameter referred to as the context . For a concrete example , consider the dynamic treatment regime ( Chakraborty & Murp... | This paper introduces a formulation for the contextual inverse reinforcement learning (COIRL) problem and proposed three algorithms for solving the proposed problem. Theoretical analysis of scalability and sample complexity are conducted for cases where both the feature function and the context-to-reward mapping functi... | SP:caa7fcf551f4ee75b6c06f05581bc5ef298fedbe |
Contextual Inverse Reinforcement Learning | 1 INTRODUCTION . We study sequential decision-making in a Contextual Markov Decision Process ( CMDP , Hallak et al . ( 2015 ) ) , where the reward , while unknown to the agent , depends on a static parameter referred to as the context . For a concrete example , consider the dynamic treatment regime ( Chakraborty & Murp... | This work focuses on the problem of 'contextual' inverse reinforcement learning, where the reward is a function of the current state of the MDP, and a set of context features, which remain constant within each episode. The primary contribution of this work is the formulation of inverse reinforcement learning (for rest... | SP:caa7fcf551f4ee75b6c06f05581bc5ef298fedbe |
MULTI-STAGE INFLUENCE FUNCTION | Multi-stage training and knowledge transfer from a large-scale pretraining task to various finetuning tasks have revolutionized natural language processing ( NLP ) and computer vision ( CV ) , with state-of-the-art performances constantly being improved . In this paper , we develop a multi-stage influence function scor... | The authors derive the influence function of models that are first pre-trained and then fine-tuned. This extends influence functions beyond the standard supervised setting that they have been primarily considered in. To do so, the authors make two methodological contributions: 1) working through the calculus for the pr... | SP:5406872be7f8a36576284f9a18ecb76d658bf25c |
MULTI-STAGE INFLUENCE FUNCTION | Multi-stage training and knowledge transfer from a large-scale pretraining task to various finetuning tasks have revolutionized natural language processing ( NLP ) and computer vision ( CV ) , with state-of-the-art performances constantly being improved . In this paper , we develop a multi-stage influence function scor... | This is an analysis paper of pretraining with the tool “influence function”. First, the authors calculate the influence score for the models with/without pretraining, and then propose some implementation details (i.e., use CG to estimate the inversed Hessian). To calculate the influence function of a model with pretrai... | SP:5406872be7f8a36576284f9a18ecb76d658bf25c |
Difference-Seeking Generative Adversarial Network--Unseen Sample Generation | 1 INTRODUCTION . Unseen data1are not samples from the distribution of the training data and are difficult to collect . It has been demonstrated that unseen samples can be applied to several applications . Dai et al . ( 2017 ) proposed how to create complement data , and theoretically showed that complement data , consi... | This paper proposed DSGAN which learns to generate unseen data from seen data distribution p_d and its somehow “broad” version p_{\hat d} (E.g., p_d convolved with Gaussian). The “unseen data” is the one that appears in p_{\hat d} but not in p_d. DSGAN is trained to generate such data. In particular, it uses samples fr... | SP:1879f9692f92bb2249772e14f27839d9e426f9b3 |
Difference-Seeking Generative Adversarial Network--Unseen Sample Generation | 1 INTRODUCTION . Unseen data1are not samples from the distribution of the training data and are difficult to collect . It has been demonstrated that unseen samples can be applied to several applications . Dai et al . ( 2017 ) proposed how to create complement data , and theoretically showed that complement data , consi... | This paper provides an interesting application of GAN which can generate the outlier distribution of training data which forces generator to learn the distribution of the low probability density area of given data. To show the effectiveness of the method, the author intuitively shows how it works on 2-D points data as ... | SP:1879f9692f92bb2249772e14f27839d9e426f9b3 |
Generative Cleaning Networks with Quantized Nonlinear Transform for Deep Neural Network Defense | 1 INTRODUCTION . Recent research has shown that deep neural networks are sensitive to adversarial attacks ( Szegedy et al. , 2013 ) . Very small changes of the input image can fool the state-of-art classifier with very high success probabilities . During the past few years , a number of methods have been proposed to co... | This paper proposes a new method to defend a neural network agains adversarial attacks (both white-box and black-box attacks). By jointly training a Generative Cleaning Network with quantized nonlinear transform, and a Detector Network, the proposed cleans the incoming attacked image and correctly classifies its true l... | SP:b386db96a13357984b61761342ae1cc876fe6d3a |
Generative Cleaning Networks with Quantized Nonlinear Transform for Deep Neural Network Defense | 1 INTRODUCTION . Recent research has shown that deep neural networks are sensitive to adversarial attacks ( Szegedy et al. , 2013 ) . Very small changes of the input image can fool the state-of-art classifier with very high success probabilities . During the past few years , a number of methods have been proposed to co... | This paper developed a method for defending deep neural networks against adversarial attacks based on generative cleaning networks with quantized nonlinear transform. The network is claimed to recover the original image while cleaning up the residual attack noise. The authors developed a detector network, which serves ... | SP:b386db96a13357984b61761342ae1cc876fe6d3a |
Optimistic Adaptive Acceleration for Optimization | 1 INTRODUCTION . Nowadays deep learning has been very successful in numerous applications , from robotics ( e.g. , Levine et al . ( 2017 ) ) , computer vision ( e.g. , He et al . ( 2016 ) ; Goodfellow et al . ( 2014 ) ) , reinforcement learning ( e.g. , Mnih et al . ( 2013 ) ) , to natural language processing ( e.g. , ... | This paper studies an optimistic variant of AMSGrad algorithm, where an estimate of the future gradient is incorporated into the optimization problem. The main claim is that when we have good enough (distance from the ground truth is small) estimate of the unknown gradient, the proposed algorithm will enjoy lower regre... | SP:35b6bf3da512cae6ad93e1422b1e272474f9a8cb |
Optimistic Adaptive Acceleration for Optimization | 1 INTRODUCTION . Nowadays deep learning has been very successful in numerous applications , from robotics ( e.g. , Levine et al . ( 2017 ) ) , computer vision ( e.g. , He et al . ( 2016 ) ; Goodfellow et al . ( 2014 ) ) , reinforcement learning ( e.g. , Mnih et al . ( 2013 ) ) , to natural language processing ( e.g. , ... | This paper proposes an online optimization method called Optimistic-AMSGrad, which combines two existing methods: (i) AMSGrad (Reddi et al 2018) and (ii) optimistic online learning where the prediction step is done with the extrapolation algorithm by Scieur et al 2016. The authors do a good job of presenting the method... | SP:35b6bf3da512cae6ad93e1422b1e272474f9a8cb |
Making the Shoe Fit: Architectures, Initializations, and Tuning for Learning with Privacy | 1 INTRODUCTION . Machine learning ( ML ) can be usefully applied to the analysis of sensitive data , e.g. , in the domain of healthcare ( Kononenko , 2001 ) . However , ML models may unintentionally reveal sensitive aspects of their training data , e.g. , due to overfitting ( Shokri et al. , 2017 ; Song & Shmatikov , 2... | This paper presents experimental evidence that learning with privacy requires approaches that are not identical to those used when learning without privacy. These approaches include re-considering different model choices (i.e., its structure and activation functions), its initialization, and its optimization procedure.... | SP:8ff1115adfd50e2c1512534ec8b90f91e0c0c331 |
Making the Shoe Fit: Architectures, Initializations, and Tuning for Learning with Privacy | 1 INTRODUCTION . Machine learning ( ML ) can be usefully applied to the analysis of sensitive data , e.g. , in the domain of healthcare ( Kononenko , 2001 ) . However , ML models may unintentionally reveal sensitive aspects of their training data , e.g. , due to overfitting ( Shokri et al. , 2017 ; Song & Shmatikov , 2... | The paper methodically analyses the settings and choices used when training neural networks (specifically CNNs) via the DP-SGD algorithm and suggests changes to the standard procedures that empirically lead to higher accuracies despite the added noise. The main statement of the paper is quite simple: optimize hyperpara... | SP:8ff1115adfd50e2c1512534ec8b90f91e0c0c331 |
A Closer Look at the Optimization Landscapes of Generative Adversarial Networks | 1 INTRODUCTION . Deep neural networks have exhibited remarkable success in many applications ( Krizhevsky et al. , 2012 ) . This success has motivated many studies of their non-convex loss landscape ( Choromanska et al. , 2015 ; Kawaguchi , 2016 ; Li et al. , 2018b ) , which , in turn , has led to many improvements , s... | This paper proposes visualization techniques for the optimization landscape in GANs. The primary tool presented in this paper is a quantity called path-angle, which looks at the angle between the game vector field and the linear path between a point away from a stationary point and a point near a stationary point. The ... | SP:276b6721d8cad9d323908d8677706c8ad1668c95 |
A Closer Look at the Optimization Landscapes of Generative Adversarial Networks | 1 INTRODUCTION . Deep neural networks have exhibited remarkable success in many applications ( Krizhevsky et al. , 2012 ) . This success has motivated many studies of their non-convex loss landscape ( Choromanska et al. , 2015 ; Kawaguchi , 2016 ; Li et al. , 2018b ) , which , in turn , has led to many improvements , s... | This paper tries to provide a deeper understanding of the training dynamics of GANs in practice via characterizing and visualizing the rotation and attraction phenomena nearby a locally stable stationary point (LSSP) and questions the necessity to access a differential/local Nash equilibrium (LNE). In particular, this ... | SP:276b6721d8cad9d323908d8677706c8ad1668c95 |
Adversarially Robust Representations with Smooth Encoders | 1 INTRODUCTION . Representation learning is a fundamental problem in Machine learning and holds the promise to enable data-efficient learning and transfer to new tasks . Researchers working in domains like Computer Vision ( Krizhevsky et al. , 2012 ) and Natural Language Processing ( Devlin et al. , 2018 ) have already... | This paper studies the vulnerability of representations learned by variational auto-encoders (VAE). It first show that the learned representation of VAE is susceptible to small changes, similar to the adversarial examples in supervised learning setting. Then propose a regularization method, called smooth encoder, to im... | SP:881f632bbadf0cac11ec1e466f02b26762f67073 |
Adversarially Robust Representations with Smooth Encoders | 1 INTRODUCTION . Representation learning is a fundamental problem in Machine learning and holds the promise to enable data-efficient learning and transfer to new tasks . Researchers working in domains like Computer Vision ( Krizhevsky et al. , 2012 ) and Natural Language Processing ( Devlin et al. , 2018 ) have already... | This paper analyzes the shortcoming of VAE objective, and propose a regularization method based on a selection mechanism that creates a fictive data point by explicitly perturbing an observed true data point. It is lead to Wasserstein distance between representations. Experiments are made on three datasets; ColorMNIST,... | SP:881f632bbadf0cac11ec1e466f02b26762f67073 |
On Robustness of Neural Ordinary Differential Equations | 1 INTRODUCTION Neural ordinary differential equations ( Chen et al. , 2018 ) form a family of models that approximate nonlinear mappings by using continuous-time ODEs . Due to their desirable properties , such as invertibility and parameter efficiency , neural ODEs have attracted increasing attention recently ( Dupont ... | The paper is concerned with neural ODE-based networks, specifically their robustness. While ODEs are a classical subject in mathematics with many applications in the sciences and beyond, neural ODEs are a recently proposed family of models for nonlinear mappings in the context of machine learning systems. There they sh... | SP:12411220098647e9bc26769218f2f64d82867493 |
On Robustness of Neural Ordinary Differential Equations | 1 INTRODUCTION Neural ordinary differential equations ( Chen et al. , 2018 ) form a family of models that approximate nonlinear mappings by using continuous-time ODEs . Due to their desirable properties , such as invertibility and parameter efficiency , neural ODEs have attracted increasing attention recently ( Dupont ... | This paper investigates the robustness of Neural Ordinary differential equations (ODEs) against corrupted and adversarial examples. The crux of the analysis is based on the separation property of ODE integral curves. The insights from empirical robustness evaluation show that controlling the difference between neighbor... | SP:12411220098647e9bc26769218f2f64d82867493 |
Granger Causal Structure Reconstruction from Heterogeneous Multivariate Time Series | Granger causal structure reconstruction is an emerging topic that can uncover causal relationship behind multivariate time series data . In many real-world systems , it is common to encounter a large amount of multivariate time series data collected from heterogeneous individuals with sharing commonalities , however th... | This paper proposes a new way of finding the Granger temporal-causal network based on attention mechanism on the predictions obtained by individual time series. It describes a surprisingly complex procedure for computing the attention vector based on combining Granger-inspired attentions with attentions obtained during... | SP:674372d2a8bfd6460e61cf6d39f85a9128cdf131 |
Granger Causal Structure Reconstruction from Heterogeneous Multivariate Time Series | Granger causal structure reconstruction is an emerging topic that can uncover causal relationship behind multivariate time series data . In many real-world systems , it is common to encounter a large amount of multivariate time series data collected from heterogeneous individuals with sharing commonalities , however th... | The paper proposes a novel way of reconstructing Granger causal structures using a differentiable neural network architecture that contains attention modules that are proportional to the Granger causality of the input layers. Furthermore, the architecture blends individual-specific induced causal structures and cross-p... | SP:674372d2a8bfd6460e61cf6d39f85a9128cdf131 |
Deep Randomized Least Squares Value Iteration | 1 INTRODUCTION . In Reinforcement Learning ( RL ) , an agent seeks to maximize the cumulative rewards obtained from interactions with an unknown environment ( Sutton et al. , 1998 ) . Since the agent can learn only by its interactions with the environment , it faces the exploration-exploitation dilemma : Should it take... | The paper proposes to extend the popular linear-control algorithm, RLSVI, to utilize learned representations. This is done by adapting a work from bandit literature that utilizes BLR with representations that are learned via a DNN. The proposed solution is then compared to DQN with fixed epsilon as the exploration stra... | SP:0bd5556d83764d1445a07e46ea5fcd074789b6e0 |
Deep Randomized Least Squares Value Iteration | 1 INTRODUCTION . In Reinforcement Learning ( RL ) , an agent seeks to maximize the cumulative rewards obtained from interactions with an unknown environment ( Sutton et al. , 1998 ) . Since the agent can learn only by its interactions with the environment , it faces the exploration-exploitation dilemma : Should it take... | This paper introduces a deep learning-based adaptation for the RLVSI algorithm, where the agent uses the representation learned by the deep neural network-based RL agent (DQN). They use the last layer of DQN as a state representation for RLSVI. In order to work with the changing representations of the deep agent, they... | SP:0bd5556d83764d1445a07e46ea5fcd074789b6e0 |
Noise Regularization for Conditional Density Estimation | 1 INTRODUCTION . While regression analysis aims to describe the conditional mean E [ y|x ] of a response y given inputs x , many problems such as risk management and planning under uncertainty require gaining insight about deviations from the mean and their associated likelihood . The stochastic dependency of y on x ca... | The paper considers the problem of parametric conditional density estimation, i.e. given a set of points {(x_n, y_n)} drawn from a distribution $p(x,y)$, the task is to estimate the conditional distribution p(x|y). The paper considers parametric estimation where in given a parametrized family of distributions f_{theta}... | SP:86bd95d8a233760200cafb7cb72ac48a7d50b7d1 |
Noise Regularization for Conditional Density Estimation | 1 INTRODUCTION . While regression analysis aims to describe the conditional mean E [ y|x ] of a response y given inputs x , many problems such as risk management and planning under uncertainty require gaining insight about deviations from the mean and their associated likelihood . The stochastic dependency of y on x ca... | The paper presents a regularization technique for conditional density estimation. The method is simple: adding noise to the data points, and training on the noisy data points. The paper also further gives an interpretation of the method, as a form of smoothing the curvature of the density function. It further proves th... | SP:86bd95d8a233760200cafb7cb72ac48a7d50b7d1 |
Mirror Descent View For Neural Network Quantization | 1 INTRODUCTION . Despite the success of deep neural networks in various domains , their excessive computational and memory requirements limit their practical usability for real-time applications or in resource-limited devices . Quantization is a prominent technique for network compression , where the objective is to le... | This paper proposes a Mirror Descent (MD) framework for the quantization of neural networks, which, different with previous quantization methods, enables us to derive valid mirror maps and the respective MD updates. Moreover, the authors also provide a stable implementation of MD by storing an additional set of auxilia... | SP:96afb20c4d7fe41c083a0217c9cb8d1f21a73a15 |
Mirror Descent View For Neural Network Quantization | 1 INTRODUCTION . Despite the success of deep neural networks in various domains , their excessive computational and memory requirements limit their practical usability for real-time applications or in resource-limited devices . Quantization is a prominent technique for network compression , where the objective is to le... | This paper proposes a neural network (NN) quantization based on Mirror Descent (MD) framework. The core of the proposal is the construction of the mirror map from the unconstrained auxiliary variables to the quantized space. Building on that core, the authors derive some mapping functions from the corresponding project... | SP:96afb20c4d7fe41c083a0217c9cb8d1f21a73a15 |
Natural Image Manipulation for Autoregressive Models Using Fisher Scores | 1 INTRODUCTION . Over the last few decades , unsupervised learning has been a rapidly growing field , with the development of more complex and better probabilistic density models . Autoregressive generative models ( Salimans et al. , 2017 ; Oord et al. , 2016a ; Menick & Kalchbrenner , 2018 ) are one of the most powerf... | Motivated by the observation that powerful deep autoregressive models such as PixelCNNs lack the ability to produce semantically meaningful latent embeddings and generate visually appealing interpolated images by latent representation manipulations, this paper proposes using Fisher scores projected to a reasonably low-... | SP:662edd2fd9437de887821ebf7de06415eba13fae |
Natural Image Manipulation for Autoregressive Models Using Fisher Scores | 1 INTRODUCTION . Over the last few decades , unsupervised learning has been a rapidly growing field , with the development of more complex and better probabilistic density models . Autoregressive generative models ( Salimans et al. , 2017 ; Oord et al. , 2016a ; Menick & Kalchbrenner , 2018 ) are one of the most powerf... | This paper focuses on the problem of interpolating between data points using neural autoregressive models. The core idea is that it is possible to use (a smaller-dimensional projection of) the Fisher score of the density function defined by the autoregressive model to represent data points in embedding space, and a neu... | SP:662edd2fd9437de887821ebf7de06415eba13fae |
Meta-RCNN: Meta Learning for Few-Shot Object Detection | 1 INTRODUCTION . Object detection is the task of identifying various objects in a given image , and localizing them with a bounding box . It is a widely studied problem in computer vision , and following the success deep convolutional neural networks ( DCNN ) in image classification ( Karpathy et al. , 2014 ; Krizhevsk... | The paper proposes a method for few-shot object detection (FSOD), a variant of few-shot learning (FSL) where using a support set of few training images for novel categories (usually 1 or 5) not only the correct category labels are predicted on the query images, but also the object instances from the novel categories ar... | SP:9e36913574414b98fd6c4a66061cf0216dcc536b |
Meta-RCNN: Meta Learning for Few-Shot Object Detection | 1 INTRODUCTION . Object detection is the task of identifying various objects in a given image , and localizing them with a bounding box . It is a widely studied problem in computer vision , and following the success deep convolutional neural networks ( DCNN ) in image classification ( Karpathy et al. , 2014 ; Krizhevsk... | This paper is about the task of object detection in the setting of few-shots dataset. The problem is addressed in the learning scheme of meta-learning paradigm: the proposed meta-rcnn trains the popular faster-rcnn on several tasks of few shots object detection while the RPN and the object classification networks are m... | SP:9e36913574414b98fd6c4a66061cf0216dcc536b |
Angular Visual Hardness | Although convolutional neural networks ( CNNs ) are inspired by the mechanisms behind human visual systems , they diverge on many measures such as ambiguity or hardness . In this paper , we make a surprising discovery : there exists a ( nearly ) universal score function for CNNs whose correlation is statistically signi... | This paper defined Angular Visual Hardness (AVH) based on the angle between image feature embedding and the weights of the target class. The authors compared the correlation between AVH and human selection frequency with model confidence and feature norm. The results show that both AVH and model confidence have correla... | SP:ab1e96008989209a4c5423f723bfae327416e78a |
Angular Visual Hardness | Although convolutional neural networks ( CNNs ) are inspired by the mechanisms behind human visual systems , they diverge on many measures such as ambiguity or hardness . In this paper , we make a surprising discovery : there exists a ( nearly ) universal score function for CNNs whose correlation is statistically signi... | This paper is trying to bridge the gap between CNN and the human visual system by proposing a metric (angular visual distance) and validate that this metric is correlated to the human visual hardness and this metric has a stronger relation compared to the softmax score which has been viewed as a metric measuring the... | SP:ab1e96008989209a4c5423f723bfae327416e78a |
Retrieving Signals in the Frequency Domain with Deep Complex Extractors | 1 INTRODUCTION . Complex-valued neural networks have been studied since long before the emergence of modern deep learning techniques ( Georgiou & Koutsougeras , 1992 ; Zemel et al. , 1995 ; Kim & Adalı , 2003 ; Hirose , 2003 ; Nitta , 2004 ) . Nevertheless , deep complex-valued models have only started to gain momentum... | This paper proposes a new method for source separation, by using deep learning UNets, complex-valued representations and the Fourier domain. Concretely, their contribution is : i) a complex-valued convolutional version of the Feature-Wise Linear Modulation, able to optimise the parameters needed to create multiple sepa... | SP:32ff67eb5376e6c8c52c5adc601c520abc9a648c |
Retrieving Signals in the Frequency Domain with Deep Complex Extractors | 1 INTRODUCTION . Complex-valued neural networks have been studied since long before the emergence of modern deep learning techniques ( Georgiou & Koutsougeras , 1992 ; Zemel et al. , 1995 ; Kim & Adalı , 2003 ; Hirose , 2003 ; Nitta , 2004 ) . Nevertheless , deep complex-valued models have only started to gain momentum... | This work researches the deep complex-valued neural networks. Specifically, it proposes a new signal extraction mechanism that operates in frequency domain and applies to address the speech separation issue. Also, a function is proposed to explicitly consider both the magnitude and phase information of a signal. Relate... | SP:32ff67eb5376e6c8c52c5adc601c520abc9a648c |
Neural Text Generation With Unlikelihood Training | 1 INTRODUCTION . Neural text generation is a vital tool in a wide range of natural language applications . However , the standard approach – training a sequence to sequence model , e.g . Transformer ( Vaswani et al. , 2017 ) , to maximize log-likelihood and approximately decoding the most likely sequence – is known to ... | This paper proposes training losses, unlikelihood objective, for mitigating the repetition problem of the text generated by recent neural language models. The problem is well-motivated by evidence from the existing literature. Specifically, the paper argues that the main cause of the degenerated output is the maximum l... | SP:4140a212888e058dc1f0bfaa5233f54e9d87fcee |
Neural Text Generation With Unlikelihood Training | 1 INTRODUCTION . Neural text generation is a vital tool in a wide range of natural language applications . However , the standard approach – training a sequence to sequence model , e.g . Transformer ( Vaswani et al. , 2017 ) , to maximize log-likelihood and approximately decoding the most likely sequence – is known to ... | The main contribution of this paper lies in the proposed unlikelihood training objective for open-ended text generation. The key idea is to enforce the unlikely generations to be assigned lower probability by the model. Both token and sequence-level unlikelihood training objectives are provided. Impressively, the autho... | SP:4140a212888e058dc1f0bfaa5233f54e9d87fcee |
Solving Packing Problems by Conditional Query Learning | 1 INTRODUCTION . How to pack boxes with the smallest bin size ? With the development of globalization and ECommerce , this question becomes more and more important . Boxes are packed to various bins such as shipping container and boxcar . Many of these boxes are made by paper or plastic ; packing boxes in a more effici... | This paper aims at solving geometric bin packing (2D or 3D) problems using a deep reinforcement learning framework. Namely, the framework is based on the actor-critic paradigm, and uses a conditional query learning model for performing composite actions (selections, rotations) in geometric bin packing. Experiments are ... | SP:3af65f4601748c89802e82f7e312d169ab8f54f2 |
Solving Packing Problems by Conditional Query Learning | 1 INTRODUCTION . How to pack boxes with the smallest bin size ? With the development of globalization and ECommerce , this question becomes more and more important . Boxes are packed to various bins such as shipping container and boxcar . Many of these boxes are made by paper or plastic ; packing boxes in a more effici... | This paper proposes an end-to-end deep reinforcement learning-based algorithm for the 2D and 3D bin packing problems. Its main contribution is conditional query learning (CQL) which allows effective decision over mutually conditioned action spaces through policy expressed as a sequence of conditional distributions. Eff... | SP:3af65f4601748c89802e82f7e312d169ab8f54f2 |
Learning to Anneal and Prune Proximity Graphs for Similarity Search | This paper studies similarity search , which is a crucial enabler of many feature vector–based applications . The problem of similarity search has been extensively studied in the machine learning community . Recent advances of proximity graphs have achieved outstanding performance through exploiting the navigability of... | This paper suggests an approach for learning how to sparsify similarity search graphs. Graph-based methods currently attain state of the art performance for similarity search, and reducing their number of edges may speed them up even further. The paper suggests a learning framework that uses sample queries in order to ... | SP:158dd8882013a9a5efa7fd4579ad3900ca76a4b5 |
Learning to Anneal and Prune Proximity Graphs for Similarity Search | This paper studies similarity search , which is a crucial enabler of many feature vector–based applications . The problem of similarity search has been extensively studied in the machine learning community . Recent advances of proximity graphs have achieved outstanding performance through exploiting the navigability of... | This paper studies the problem of improving proximity graph for nearest neighbor search. It formulates the task of pruning the graph as a problem of learning annealable proximity graph. A hard pruning processes is used after the learning process, and the results shows that the proposed method can reduce 50% of the edge... | SP:158dd8882013a9a5efa7fd4579ad3900ca76a4b5 |
Understanding and Improving Information Transfer in Multi-Task Learning | 1 INTRODUCTION . Multi-task learning has recently emerged as a powerful paradigm in deep learning to obtain language ( Devlin et al . ( 2018 ) ; Liu et al . ( 2019a ; b ) ) and visual representations ( Kokkinos ( 2017 ) ) from large-scale data . By leveraging supervised data from related tasks , multi-task learning app... | This paper studies how to improve the multi-task learning from both theoretical and experimental viewpoints. More specifically, they study an architecture where there is a shared model for all of the tasks and a separate module specific to each task. They show that data similarity of the tasks, measured by task covaria... | SP:23726c6ff50e4ff1beb7f21e31a9f6286a656b1e |
Understanding and Improving Information Transfer in Multi-Task Learning | 1 INTRODUCTION . Multi-task learning has recently emerged as a powerful paradigm in deep learning to obtain language ( Devlin et al . ( 2018 ) ; Liu et al . ( 2019a ; b ) ) and visual representations ( Kokkinos ( 2017 ) ) from large-scale data . By leveraging supervised data from related tasks , multi-task learning app... | This paper analyzed the principles for a successful transfer in the hard-parameter sharing multitask learning model. They analyzed three key factors of multi-task learning on linear model and relu linear model: model capacity (output dimension after common transformation), task covariance (similarity between tasks) and... | SP:23726c6ff50e4ff1beb7f21e31a9f6286a656b1e |
FreeLB: Enhanced Adversarial Training for Natural Language Understanding | 1 INTRODUCTION . Adversarial training is a method for creating robust neural networks . During adversarial training , mini-batches of training samples are contaminated with adversarial perturbations ( alterations that are small and yet cause misclassification ) , and then used to update network parameters until the res... | - This paper modifies and extends the recent “free” training strategies in adversarial training for representation learning for natural language. The proposed “Free” Large-Batch Adversarial Training is well motived, in comparison with plain PGD-based adversarial training and the existing methods like FreeAT and YOPO, ... | SP:027dfebf9732ce68cb3985ef873b00d65e6e7205 |
FreeLB: Enhanced Adversarial Training for Natural Language Understanding | 1 INTRODUCTION . Adversarial training is a method for creating robust neural networks . During adversarial training , mini-batches of training samples are contaminated with adversarial perturbations ( alterations that are small and yet cause misclassification ) , and then used to update network parameters until the res... | In this paper, the authors present a new adversarial training algorithm and apply it to the fintuning stage large scale language models BERT and RoBERTa. They find that with FreeLB applied to finetuning, both BERT and RoBERTa see small boosts in performance on GLUE, ARC, and CommonsenseQA. The gains they see on GLUE ar... | SP:027dfebf9732ce68cb3985ef873b00d65e6e7205 |
An Inductive Bias for Distances: Neural Nets that Respect the Triangle Inequality | 1 INTRODUCTION . Many machine learning tasks involve a distance measure over the input domain . A good measure can make a once hard task easy , even trivial . In many cases—including graph distances , certain clustering algorithms , and general value functions in reinforcement learning ( RL ) —it is either known that d... | This paper proposes a modeling approach for norm and metric learning that ensures triangle inequalities are satisfied by the very design of the architecture. The main idea is that convexity together with homogeneity imply subadditivity, so starting from an input-convex architecture and using activations that preserve h... | SP:1dae5dd9635962d35767ee1a5a4da01170e18029 |
An Inductive Bias for Distances: Neural Nets that Respect the Triangle Inequality | 1 INTRODUCTION . Many machine learning tasks involve a distance measure over the input domain . A good measure can make a once hard task easy , even trivial . In many cases—including graph distances , certain clustering algorithms , and general value functions in reinforcement learning ( RL ) —it is either known that d... | This manuscript proposes a general framework to learn non-Euclidean distances from data using neural networks. The authors provide a combination of theoretical and experimental results in support of the use of several neural architectures to learn such distances. In particular, the develop “deep norms” and “wide norms”... | SP:1dae5dd9635962d35767ee1a5a4da01170e18029 |
Neural Networks for Principal Component Analysis: A New Loss Function Provably Yields Ordered Exact Eigenvectors | 1 INTRODUCTION . Ranking among the most widely-used and valuable statistical tools , Principal Component Analysis ( PCA ) represents a given set of data within a new orthogonal coordinate system in which the data are uncorrelated and the variance of the data along each orthogonal axis is successively ordered from the h... | This paper proposes a new loss function to compute the exact ordered eigenvectors of a dataset. The loss is motivated from the idea of computing the eigenvectors sequentially. However doing so would be computationally expensive, and the authors show that the loss function they propose (sum of sequential losses) has the... | SP:3a670c06bf87ba895ed91ed2280d88881defa412 |
Neural Networks for Principal Component Analysis: A New Loss Function Provably Yields Ordered Exact Eigenvectors | 1 INTRODUCTION . Ranking among the most widely-used and valuable statistical tools , Principal Component Analysis ( PCA ) represents a given set of data within a new orthogonal coordinate system in which the data are uncorrelated and the variance of the data along each orthogonal axis is successively ordered from the h... | This paper proposes a new loss function for performing principal component analysis (PCA) using linear autoencoders (LAEs). With this new loss function, the decoder weights of LAEs can eventually converge to the exact ordered unnormalized eigenvectors of the sample covariance matrix. The main contribution is to add the... | SP:3a670c06bf87ba895ed91ed2280d88881defa412 |
Hierarchical Foresight: Self-Supervised Learning of Long-Horizon Tasks via Visual Subgoal Generation | 1 INTRODUCTION . Developing robotic systems that can complete long horizon visual control tasks , while generalizing to novel scenes and objectives , remains an unsolved and challenging problem . Generalization to unseen objects and scenes requires robots to be trained across diverse environments , meaning that detaile... | This paper introduces a hierarchical extension to existing work in vision-based model predictive control. Here, a hierarchical model is optimised to find sub-goals that minimise the planning cost (bottleneck states), so as to allow for improved planning to goal states expressed in higher dimensional state spaces. As ex... | SP:7c442073ca3d80b472665b8bd9ec3534ef010950 |
Hierarchical Foresight: Self-Supervised Learning of Long-Horizon Tasks via Visual Subgoal Generation | 1 INTRODUCTION . Developing robotic systems that can complete long horizon visual control tasks , while generalizing to novel scenes and objectives , remains an unsolved and challenging problem . Generalization to unseen objects and scenes requires robots to be trained across diverse environments , meaning that detaile... | This paper proposes a method, hierarchical visual foresight (HVF) that learns to break down the long horizon tasks into short horizon segments. It first generates the subgoals conditioned on the main goal. These subgoals are optimized to have meaningful states and used for planning. The experiments on Maze navigation, ... | SP:7c442073ca3d80b472665b8bd9ec3534ef010950 |
Do Image Classifiers Generalize Across Time? | 1 INTRODUCTION . Convolutional neural networks ( CNNs ) still exhibit many troubling failure modes . At one extreme , ` p-adversarial examples cause large drops in accuracy for state-of-the-art models while relying only on visually imperceptible changes to the input image ( Goodfellow et al. , 2014 ; Biggio and Roli , ... | This paper presents new datasets based on ImageNet and Youtube-BB to assert networks performance consistency across time. Compared to previous work, it uses human labeler to further validate the dataset and discard frames that are deemed too different from the reference one. It provides results on image classification ... | SP:ef3afd5d34fbb7c8310a1dc9d6e49c2f37db07e6 |
Do Image Classifiers Generalize Across Time? | 1 INTRODUCTION . Convolutional neural networks ( CNNs ) still exhibit many troubling failure modes . At one extreme , ` p-adversarial examples cause large drops in accuracy for state-of-the-art models while relying only on visually imperceptible changes to the input image ( Goodfellow et al. , 2014 ; Biggio and Roli , ... | In this paper, the authors curated two datasets: ImageNet-Vid and Youtube-BB in order to create human-reviewed perceptibly similar sets (Imagenet-Vid-Robust and YTBB-Robust). The obtained datasets are evaluated over 45 different models pre-trained on ImageNet in order to see their drop in accuracy on natural perturbati... | SP:ef3afd5d34fbb7c8310a1dc9d6e49c2f37db07e6 |
Pipelined Training with Stale Weights of Deep Convolutional Neural Networks | The growth in the complexity of Convolutional Neural Networks ( CNNs ) is increasing interest in partitioning a network across multiple accelerators during training and pipelining the backpropagation computations over the accelerators . Existing approaches avoid or limit the use of stale weights through techniques such... | This paper investigates the impact of stale weights on the statistical efficiency and performance in a pipelined backpropagation scheme that maximizes accelerator utilization while keeping the memory overhead modest. The paper proposes to combine pipelined and non-pipelined training in a hybrid scheme to address the is... | SP:3b0d0ac062a7bc618741cff17c7d507b0b0a7489 |
Pipelined Training with Stale Weights of Deep Convolutional Neural Networks | The growth in the complexity of Convolutional Neural Networks ( CNNs ) is increasing interest in partitioning a network across multiple accelerators during training and pipelining the backpropagation computations over the accelerators . Existing approaches avoid or limit the use of stale weights through techniques such... | This paper proposes a new pipelined training approach to speedup the training for neural networks. The approach separates forward and backpropagation processes into multiple stages, cache the activation and gradients between stages, processes stages simultaneously, and then uses the stored activations to compute gradie... | SP:3b0d0ac062a7bc618741cff17c7d507b0b0a7489 |
Efficient and Information-Preserving Future Frame Prediction and Beyond | 1 INTRODUCTION . Deep learning has enjoyed tremendous success in recent years due to its ability to capture complex dependencies and non-linearities in large datasets ( Krizhevsky et al . ( 2012 ) ; He et al . ( 2016 ) ; Gomez et al . ( 2017 ) ) . Excellent performance has been achieved on a wide range of supervised ma... | This paper introduces Conditionally Reversible Network (CrevNet) that consists of the invertible autoencoder and a reversible predictive module (RPM). The two-way autoencoder is an invertible network that preserves the volume with no information loss while reducing memory consumption by using bijective downsampling. Th... | SP:00b48a43a5037915e21ddb2f0941cdd26a69d44d |
Efficient and Information-Preserving Future Frame Prediction and Beyond | 1 INTRODUCTION . Deep learning has enjoyed tremendous success in recent years due to its ability to capture complex dependencies and non-linearities in large datasets ( Krizhevsky et al . ( 2012 ) ; He et al . ( 2016 ) ; Gomez et al . ( 2017 ) ) . Excellent performance has been achieved on a wide range of supervised ma... | In this paper, the authors propose a new method of self-supervised feature learning from videos based on learning future frame prediction. The idea is similar as BERT like NLP tasks, but for videos, the computational cost and memory cost could be very large. To solve this problem efficiently, the authors adopt several ... | SP:00b48a43a5037915e21ddb2f0941cdd26a69d44d |
Global Concavity and Optimization in a Class of Dynamic Discrete Choice Models | Discrete choice models with unobserved heterogeneity are commonly used Econometric models for dynamic Economic behavior which have been adopted in practice to predict behavior of individuals and firms from schooling and job choices to strategic decisions in market competition . These models feature optimizing agents wh... | This paper considers reinforcement learning for discrete choice models with unobserved heterogeneity, which is useful for analyzing dynamic Economic behavior. Random choice-specific shocks in reward is accommodated, which are only observed by the agent but not recorded in the data. Existing optimization approaches rel... | SP:5a0e35b51548e82135b965e7b692e8a0af1289f8 |
Global Concavity and Optimization in a Class of Dynamic Discrete Choice Models | Discrete choice models with unobserved heterogeneity are commonly used Econometric models for dynamic Economic behavior which have been adopted in practice to predict behavior of individuals and firms from schooling and job choices to strategic decisions in market competition . These models feature optimizing agents wh... | This paper deals with a certain class of models, known as discrete choice models. These models are popular in econometrics, and aim at modelling the complex behavioural patterns of individuals or firms. Entities in these models are typically modelled as rational agents, that behave optimally for reaching their goal of ... | SP:5a0e35b51548e82135b965e7b692e8a0af1289f8 |
On the implicit minimization of alternative loss functions when training deep networks | 1 INTRODUCTION . In the last few years , deep learning has succeeded in establishing state of the art performances in a wide variety of tasks in fields like computer vision , natural language processing and bioinformatics ( LeCun et al. , 2015 ) . Understanding when and how these networks generalize better is important... | This paper want to show that minimizing cross-entropy loss will simultaneously minimize Hinge loss with different margins, cross-entropy loss with different temperatures and a newly introduced Gcdf loss with different standard deviations. The main contribution is a new gcdf loss based on Gaussian-perturbed parameters. ... | SP:3145e0027567692ea5c3ca4ef8d0d94b40f8f27f |
On the implicit minimization of alternative loss functions when training deep networks | 1 INTRODUCTION . In the last few years , deep learning has succeeded in establishing state of the art performances in a wide variety of tasks in fields like computer vision , natural language processing and bioinformatics ( LeCun et al. , 2015 ) . Understanding when and how these networks generalize better is important... | This paper makes a step towards understanding of the implicit bias of optimization algorithms in deep learning. The authors consider alternative loss functions for deep networks: (1) the temperature-scaled cross-entropy loss with different values of the temperature; (2) the hinge-loss with different values of the margi... | SP:3145e0027567692ea5c3ca4ef8d0d94b40f8f27f |
The Curious Case of Neural Text Degeneration | Context : In a shocking finding , scientist discovered a herd of unicorns living in a remote , previously unexplored valley , in the Andes Mountains . Even more surprising to the researchers was the fact that the unicorns spoke perfect English . Beam Search , b=32 : `` The study , published in the Proceedings of the Na... | This paper is motivated by an observation that maximization-based decoding approaches such as beam search can lead to incoherent and repetitive sentences when open-ended long-form text generation based on neural language model such as GPT-2 is performed. To solve the problem, this paper proposes a sampling method calle... | SP:08889d3b0659e76092dbb9a9fd2825701cebda44 |
The Curious Case of Neural Text Degeneration | Context : In a shocking finding , scientist discovered a herd of unicorns living in a remote , previously unexplored valley , in the Andes Mountains . Even more surprising to the researchers was the fact that the unicorns spoke perfect English . Beam Search , b=32 : `` The study , published in the Proceedings of the Na... | This paper studies an important problem, i.e., how to find a good decoding strategy for open-ended text generation. To this end, the authors provide a deep analysis of the most common decoding methods, and propose Nucleus Sampling, a very simple yet effective method to generate higher-quality text. Compared with top-k ... | SP:08889d3b0659e76092dbb9a9fd2825701cebda44 |
Implementation Matters in Deep RL: A Case Study on PPO and TRPO | 1 INTRODUCTION . Deep reinforcement learning ( RL ) algorithms have fueled many of the most publicized achievements in modern machine learning ( Silver et al. , 2017 ; OpenAI , 2018 ; Abbeel & Schulman , 2016 ; Mnih et al. , 2013 ) . However , despite these accomplishments , deep RL methods still are not nearly as reli... | This paper calls to attention the importance of specifying all performance altering implementation details that are current inherent in the state-of-the-art deep policy gradient community. Specifically, this paper builds very closely on the work started by Henderson et al. 2017, building a conversation around the impo... | SP:133403fbbb8b1195da7a017675d19d3b7b270811 |
Implementation Matters in Deep RL: A Case Study on PPO and TRPO | 1 INTRODUCTION . Deep reinforcement learning ( RL ) algorithms have fueled many of the most publicized achievements in modern machine learning ( Silver et al. , 2017 ; OpenAI , 2018 ; Abbeel & Schulman , 2016 ; Mnih et al. , 2013 ) . However , despite these accomplishments , deep RL methods still are not nearly as reli... | This paper investigates the impact of implementation "details", with existing implementations of TRPO and PPO as examples. The main takeaway is that the performance gains observed in PPO (compared to TRPO) are actually caused by differences in implementation, and not by the differences between the two learning algorith... | SP:133403fbbb8b1195da7a017675d19d3b7b270811 |
Effective Use of Variational Embedding Capacity in Expressive End-to-End Speech Synthesis | Recent work has explored sequence-to-sequence latent variable models for expressive speech synthesis ( supporting control and transfer of prosody and style ) , but has not presented a coherent framework for understanding the trade-offs between the competing methods . In this paper , we propose embedding capacity ( the ... | 1. Summary: This paper proposes Capacitron, a conditional variational latent variable model for TTS which allow for controllable latent variable capacity. They optimize the Lagrangian dual of the ELBO and restrict the capacity of the rate-term through a learnable, non-negative multiplier. They demonstrate the effective... | SP:af656384c8eec0891912cc1893a5d827bc6efb78 |
Effective Use of Variational Embedding Capacity in Expressive End-to-End Speech Synthesis | Recent work has explored sequence-to-sequence latent variable models for expressive speech synthesis ( supporting control and transfer of prosody and style ) , but has not presented a coherent framework for understanding the trade-offs between the competing methods . In this paper , we propose embedding capacity ( the ... | In this work authors present a regularized, variational autoencoder method for speech synthesis. To endow the latent space with more capacity, the authors employ a modified variational autoencoder objective, which uses a learnable Lagrange multiplier to impose a capacity limit on KL divergence between latent posterior ... | SP:af656384c8eec0891912cc1893a5d827bc6efb78 |
Hyperbolic Discounting and Learning Over Multiple Horizons | 1 INTRODUCTION . The standard treatment of the reinforcement learning ( RL ) problem is the Markov Decision Process ( MDP ) which includes a discount factor 0 ≤ γ ≤ 1 that exponentially reduces the present value of future rewards ( Bellman , 1957 ; Sutton & Barto , 1998 ) . A reward rt received in t-time steps is deval... | The paper investigates hyperbolic discounting as a more biologically plausible alternative to exponential discounting in reinforcement learning. First, it formulates a notion of hazard in MDPs as constant exponential discounting and shows that hyperbolic discounting is consistent with uncertainty over the hazard rate. ... | SP:87d54568606a9c1f752dae1420484a7b02a7ab1f |
Hyperbolic Discounting and Learning Over Multiple Horizons | 1 INTRODUCTION . The standard treatment of the reinforcement learning ( RL ) problem is the Markov Decision Process ( MDP ) which includes a discount factor 0 ≤ γ ≤ 1 that exponentially reduces the present value of future rewards ( Bellman , 1957 ; Sutton & Barto , 1998 ) . A reward rt received in t-time steps is deval... | This paper argues that hyperbolic and other non-exponential discounting mechanisms have been more utilized by humans and animals for value preferences than exponential discounting as widely used in RL literature. The authors claim that hyperbolic discounting mechanisms are especially preferred in the setting of maintai... | SP:87d54568606a9c1f752dae1420484a7b02a7ab1f |
Information-Theoretic Local Minima Characterization and Regularization | 1 INTRODUCTION . Recently , there has been a surge in the interest of acquiring a theoretical understanding over deep neural network ’ s behavior . Breakthroughs have been made in characterizing the optimization process , showing that learning algorithms such as stochastic gradient descent ( SGD ) tend to end up in one... | This paper contributes to the deep learning generalization theory, mainly from the theoretical perspective with experimental verifications. The key proposition is given by the unnumbered simple equation in the middle of page 4 (please number it), where \mathcal{I} is the Fisher information matrix. According to the auth... | SP:39c5ad94a057196b513d4a96d3478ddf73add838 |
Information-Theoretic Local Minima Characterization and Regularization | 1 INTRODUCTION . Recently , there has been a surge in the interest of acquiring a theoretical understanding over deep neural network ’ s behavior . Breakthroughs have been made in characterizing the optimization process , showing that learning algorithms such as stochastic gradient descent ( SGD ) tend to end up in one... | This paper provides a metric to characterize local minima of deep network loss landscapes based on the Fisher information matrix of the model parameterized by the deep network. The authors connect the Fisher information to the curvature of the loss landscape (the loss considered is the negative loss likelihood) and obt... | SP:39c5ad94a057196b513d4a96d3478ddf73add838 |
Dynamic Model Pruning with Feedback | 1 INTRODUCTION . Highly overparametrized deep neural networks show impressive results on machine learning tasks . However , with the increase in model size comes also the demand for memory and computer power at inference stage—two resources that are scarcely available on low-end devices . Pruning techniques have been s... | This work proposes a simple pruning method that dynamically sparsifies the network during training. This is achieved by performing at fixed intervals magnitude based pruning for either individual weights or entire neurons. While similar methods have been explored before, this work proposes a slight twist; instead of up... | SP:cf0aed09560d12961f718e915b72a2c5403c4e4a |
Dynamic Model Pruning with Feedback | 1 INTRODUCTION . Highly overparametrized deep neural networks show impressive results on machine learning tasks . However , with the increase in model size comes also the demand for memory and computer power at inference stage—two resources that are scarcely available on low-end devices . Pruning techniques have been s... | In this paper, the authors proposed a novel model compression method that uses error feedbacks to dynamically allocates sparsity patterns during training. The authors provided a systematic overview of a good number of existing model compression algorithms depending on the relative order of pruning and training processe... | SP:cf0aed09560d12961f718e915b72a2c5403c4e4a |
Contextual Text Style Transfer | 1 INTRODUCTION . Text style transfer has recently been applied to many applications with remarkable success ( e.g. , sentiment manipulation , formalized writing ) . Early work relied on parallel corpora with a sequenceto-sequence learning framework ( Bahdanau et al. , 2015 ; Jhamtani et al. , 2017 ) . However , collect... | The authors propose the task of contextual text style transfer: transferring the style of one text into another (i.e., informal to formal, or offensive to non-offensive), when the text is present within some larger, provided context. The authors propose a model (CAST) which takes advantage of the additional context to ... | SP:27e5d87807bde38fd23e80517608417aaaf724f3 |
Contextual Text Style Transfer | 1 INTRODUCTION . Text style transfer has recently been applied to many applications with remarkable success ( e.g. , sentiment manipulation , formalized writing ) . Early work relied on parallel corpora with a sequenceto-sequence learning framework ( Bahdanau et al. , 2015 ; Jhamtani et al. , 2017 ) . However , collect... | The paper proposes a new task for text style transfer, based on the idea that the the surrounding context of a sentence is important, whereas previous such tasks have only looked at sentences in isolation. Two new crowd-sourced datasets are created, and a combination of now fairly standard neural components is shown to... | SP:27e5d87807bde38fd23e80517608417aaaf724f3 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.