paper_name
stringlengths
11
170
text
stringlengths
8.07k
307k
summary
stringlengths
152
6.16k
paper_id
stringlengths
43
43
Neural networks with motivation
1 INTRODUCTION . Motivation is a cognitive process that propels an individual ’ s behavior towards or away from a particular object , perceived event , or outcome ( Zhang et al. , 2009 ) . Mathematically , motivation can be viewed as subjective modulation of the perceived reward value before the reward is received . Th...
This paper presents a computational model of motivation for Q learning and relates it to biological models of motivation. Motivation is presented to the agent as a component of its inputs, and is encoded in a vectorised reward function where each component of the reward is weighted. This approach is explored in three d...
SP:faca1e6eda4ad3b91ab99995e420398c01cc0e42
Neural networks with motivation
1 INTRODUCTION . Motivation is a cognitive process that propels an individual ’ s behavior towards or away from a particular object , perceived event , or outcome ( Zhang et al. , 2009 ) . Mathematically , motivation can be viewed as subjective modulation of the perceived reward value before the reward is received . Th...
The authors investigate mechanisms underlying action selection in artificial agents and mice. To achieve this goal, they use RL to train neural networks to choose actions that maximize their temporally discounted sum of future rewards. Importantly, these rewards depend on a motivation factor that is itself a function o...
SP:faca1e6eda4ad3b91ab99995e420398c01cc0e42
Expected Information Maximization: Using the I-Projection for Mixture Density Estimation
1 INTRODUCTION . Learning the density of highly multi-modal distributions is a challenging machine learning problem relevant to many fields such as modelling human behavior ( Pentland & Liu , 1999 ) . Most common methods rely on maximizing the likelihood of the data . It is well known that the maximum likelihood soluti...
The paper presents an algorithm to match two distributions with latent variables, named expected information maximization (EIM). Specifically, EIM is based on the I-Projection, which basically is equivalent to minimizing the reverse KL divergence (i.e. min KL[p_model || p_data]); to handle latent variables, an upper-bo...
SP:5ca4c62eae1c6a5a870524715c3be44c40383f98
Expected Information Maximization: Using the I-Projection for Mixture Density Estimation
1 INTRODUCTION . Learning the density of highly multi-modal distributions is a challenging machine learning problem relevant to many fields such as modelling human behavior ( Pentland & Liu , 1999 ) . Most common methods rely on maximizing the likelihood of the data . It is well known that the maximum likelihood soluti...
This paper propose EIM an analog to EM but to perform the I-projection (i.e. reverse-KL) instead of the usual M-projection for EM. The motivation is that the reverse-KL is mode-seeking in contrast to the forward-KL which is mode-covering. The authors argue that in the case that the model is mis-specified, I-projection ...
SP:5ca4c62eae1c6a5a870524715c3be44c40383f98
The Usual Suspects? Reassessing Blame for VAE Posterior Collapse
1 INTRODUCTION . The variational autoencoder ( VAE ) ( Kingma & Welling , 2014 ; Rezende et al. , 2014 ) represents a powerful generative model of data points that are assumed to possess some complex yet unknown latent structure . This assumption is instantiated via the marginalized distribution pθ ( x ) = ∫ pθ ( x|z )...
The paper theoretically investigates the role of “local optima” of the variational objective in ignoring latent variables (leading to posterior collapse) in variational autoencoders. The paper first discusses various potential causes for posterior collapse before diving deeper into a particular cause: local optima. The...
SP:311d2ebcdc0f71789d6c46d23451657519495119
The Usual Suspects? Reassessing Blame for VAE Posterior Collapse
1 INTRODUCTION . The variational autoencoder ( VAE ) ( Kingma & Welling , 2014 ; Rezende et al. , 2014 ) represents a powerful generative model of data points that are assumed to possess some complex yet unknown latent structure . This assumption is instantiated via the marginalized distribution pθ ( x ) = ∫ pθ ( x|z )...
This paper is clearly written and well structured. After categorizing difference causes of posterior collapse, the authors present a theoretical analysis of one such cause extending beyond the linear case covered in existing work. The authors then extended further to the deep VAE setting and showed that issues with the...
SP:311d2ebcdc0f71789d6c46d23451657519495119
Why Does Hierarchy (Sometimes) Work So Well in Reinforcement Learning?
1 INTRODUCTION . Many real-world tasks may be decomposed into natural hierarchical structures . To navigate a large building , one first needs to learn how to walk and turn before combining these behaviors to achieve robust navigation ; to wash dishes , one first needs to learn basic object grasping and handling before...
This is an interesting paper, as it tries to understand the role of hierarchical methods (such as Options, higher level controllers etc) in RL. The core contribution of the paper is understand and evaluate the claimed benefits often proposed by hierarchical methods, and finds that the core benefit in fact comes from ex...
SP:c8c5809f731c2f0c6bf01e24bc4d9eb7cf924ccd
Why Does Hierarchy (Sometimes) Work So Well in Reinforcement Learning?
1 INTRODUCTION . Many real-world tasks may be decomposed into natural hierarchical structures . To navigate a large building , one first needs to learn how to walk and turn before combining these behaviors to achieve robust navigation ; to wash dishes , one first needs to learn basic object grasping and handling before...
This paper evaluates the benefits of using hierarchical RL (HRL) methods compared to regular shallow RL methods for fully observed MDPs. The goal of the work is to isolate and evaluate the benefits of using HRL on different control tasks (AntMaze, AntPush, AntBlock, AntBlockMaze). They find that the major benefit of HR...
SP:c8c5809f731c2f0c6bf01e24bc4d9eb7cf924ccd
Robust Learning with Jacobian Regularization
1 INTRODUCTION . Stability analysis lies at the heart of many scientific and engineering disciplines . In an unstable system , infinitesimal perturbations amplify and have substantial impacts on the performance of the system . It is especially critical to perform a thorough stability analysis on complex engineered syst...
Stability is one of the important aspects of machine learning. This paper views Jacobian regularization as a scheme to improve the stability, and studies the behavior of Jacobian regularization under random input perturbations, adversarial input perturbations, train/test distribution shift, and simply as a regularizat...
SP:385a392e6d055abd65a737f3c5be58105778ac11
Robust Learning with Jacobian Regularization
1 INTRODUCTION . Stability analysis lies at the heart of many scientific and engineering disciplines . In an unstable system , infinitesimal perturbations amplify and have substantial impacts on the performance of the system . It is especially critical to perform a thorough stability analysis on complex engineered syst...
The main contribution of this paper is that it proposed an estimator of Jacobian regularization term for neural networks to reduce the computational cost reduced by orders of magnitude, and the estimator is mathematically proved unbiased. In details, the time consumed for the application of Jacobian regularizer and the...
SP:385a392e6d055abd65a737f3c5be58105778ac11
Multichannel Generative Language Models
1 INTRODUCTION . A natural way to consider two parallel sentences in different languages is that each language is expressing the same underlying meaning under a different viewpoint . Each language can be thought of as a transformation that maps an underlying concept into a view that we collectively agree is determined ...
This paper proposes a multichannel generative language model (MGLM), which models the joint distribution p(channel_1, ..., channel_k) over k channels. MGLM can be used for both conditional generation (e.g., machine translation) and unconditional sampling. In the experiments, MGLM uses the Multi30k dataset where multipl...
SP:da1e92e9459d9f305f206e309faa8e9bbf8e6afa
Multichannel Generative Language Models
1 INTRODUCTION . A natural way to consider two parallel sentences in different languages is that each language is expressing the same underlying meaning under a different viewpoint . Each language can be thought of as a transformation that maps an underlying concept into a view that we collectively agree is determined ...
This work is an extension of KERMIT (Chan et al., 2019) to multiple languages and the proposed model is called “multichannel generative language models”. KERMIT is an extension of “Insertion Transformer” (Stern et. al, 2019), a non-autoregressive model that can jointly determine which word and which place the translate...
SP:da1e92e9459d9f305f206e309faa8e9bbf8e6afa
Convergence of Gradient Methods on Bilinear Zero-Sum Games
1 INTRODUCTION . Min-max optimization has received significant attention recently due to the popularity of generative adversarial networks ( GANs ) ( Goodfellow et al. , 2014 ) , adversarial training ( Madry et al. , 2018 ) and reinforcement learning ( Du et al. , 2017 ; Dai et al. , 2018 ) , just to name some examples...
This paper studies the convergence of multiple methods (Gradient, extragradient, optimistic and momentum) on a bilinear minmax game. More precisely, this paper uses spectral condition to study the difference between simultaneous (Jacobi) and alternating (Gau\ss-Seidel) updates. The analysis is based on Schur theorem an...
SP:69704bad659d8cc6e35dc5b7f372bf2e39805f4f
Convergence of Gradient Methods on Bilinear Zero-Sum Games
1 INTRODUCTION . Min-max optimization has received significant attention recently due to the popularity of generative adversarial networks ( GANs ) ( Goodfellow et al. , 2014 ) , adversarial training ( Madry et al. , 2018 ) and reinforcement learning ( Du et al. , 2017 ; Dai et al. , 2018 ) , just to name some examples...
The paper presents exact conditions for the convergence of several gradient based methods for solving bilinear games. In particular, the methods under study are Gradient Descent(GD), Extragradient (EG), Optimizatic Gradient descent (OGD) and Momentum methods. For these methods, the authors provide convergence rates (wi...
SP:69704bad659d8cc6e35dc5b7f372bf2e39805f4f
Style-based Encoder Pre-training for Multi-modal Image Synthesis
1 INTRODUCTION . Image-to-Image ( I2I ) translation is the task of transforming images from one domain to another ( e.g. , semantic maps→ scenes , sketches→ photo-realistic images , etc. ) . Many problems in computer vision and graphics can be cast as I2I translation , such as photo-realistic image synthesis ( Chen & K...
In this paper, the authors tackle the problem of multi-modal image-to-image translation by pre-training a style-based encoder. The style-based encoder is trained with a triplet loss that encourages similarity between images with similar styles and dissimilarity between images with different styles. The output of the en...
SP:0a523e5c8790b62fef099d7c5bec61bb18a2703c
Style-based Encoder Pre-training for Multi-modal Image Synthesis
1 INTRODUCTION . Image-to-Image ( I2I ) translation is the task of transforming images from one domain to another ( e.g. , semantic maps→ scenes , sketches→ photo-realistic images , etc. ) . Many problems in computer vision and graphics can be cast as I2I translation , such as photo-realistic image synthesis ( Chen & K...
The authors propose to use a non-end-to-end approach to the problem of multi-modal I2I. Firstly, a metric learning problem is solved to embed images into space, taking into account the pairwise style discrepancy (style is defined, e.g., based on VGG Gramians). As the notion of style is universal for similar datasets, t...
SP:0a523e5c8790b62fef099d7c5bec61bb18a2703c
Prediction, Consistency, Curvature: Representation Learning for Locally-Linear Control
1 INTRODUCTION . Decomposing the problem of decision-making in an unknown environment into estimating dynamics followed by planning provides a powerful framework for building intelligent agents . This decomposition confers several notable benefits . First , it enables the handling of sparse-reward environments by lever...
This paper considers learning low-dimensional representations from high-dimensional observations for control purposes. The authors extend the E2C framework by introducing the new PCC-Loss function. This new loss function aims to reflect the prediction in the observation space, the consistency between latent and observa...
SP:8ec794421e38087b73f7d7fb4fbf373728ea39c7
Prediction, Consistency, Curvature: Representation Learning for Locally-Linear Control
1 INTRODUCTION . Decomposing the problem of decision-making in an unknown environment into estimating dynamics followed by planning provides a powerful framework for building intelligent agents . This decomposition confers several notable benefits . First , it enables the handling of sparse-reward environments by lever...
This work proposes a regularization strategy for learning optimal policy for a dynamic control problem in a latent low-dimensional domain. The work is based on LCE approach, but with in-depth analysis on how to choose/design the regularization for the \hat{P} operator, which consists of an encoder, a decoder, and dynam...
SP:8ec794421e38087b73f7d7fb4fbf373728ea39c7
IsoNN: Isomorphic Neural Network for Graph Representation Learning and Classification
1 INTRODUCTION . The graph structure is attracting increasing interests because of its great representation power on various types of data . Researchers have done many analyses based on different types of graphs , such as social networks , brain networks and biological networks . In this paper , we will focus on the bi...
This paper proposes a neural network architecture to classify graph structure. A graph is specified using its adjacency matrix, and the authors prose to extract features by identifying temples, implemented as small kernels on sub matrices of the adjacency matrix. The main problem is how to handle isomorphism: there is ...
SP:2656017dbf3c1e8b659857d3a44fdbb91e186237
IsoNN: Isomorphic Neural Network for Graph Representation Learning and Classification
1 INTRODUCTION . The graph structure is attracting increasing interests because of its great representation power on various types of data . Researchers have done many analyses based on different types of graphs , such as social networks , brain networks and biological networks . In this paper , we will focus on the bi...
This paper proposes a new neural network architecture for dealing with graphs dealing with the lack of order of the nodes. The first step called the graph isomorphic layer compute features invariant to the order of nodes by extracting sub-graphs and cosidering all possible permutation of these subgraphs. There is no tr...
SP:2656017dbf3c1e8b659857d3a44fdbb91e186237
Toward Understanding The Effect of Loss Function on The Performance of Knowledge Graph Embedding
1 INTRODUCTION . Knowledge is considered as commonsense facts and other information accumulated from different sources . A Knowledge Graph ( KG ) is collection of facts and is usually represented as a set of triples ( h , r , t ) where h , t are entities and r is a relation , e.g . ( iphone , hyponym , smartphone ) . E...
This paper list several limitations of translational-based Knowledge Graph embedding methods, TransE which have been identified by prior works and have theoretically/empirically shown that all limitations can be addressed by altering the loss function and shifting to Complex domain. The authors propose four variants of...
SP:86076eabb48ef1fe9d51b54945bf81ed44bcacd7
Toward Understanding The Effect of Loss Function on The Performance of Knowledge Graph Embedding
1 INTRODUCTION . Knowledge is considered as commonsense facts and other information accumulated from different sources . A Knowledge Graph ( KG ) is collection of facts and is usually represented as a set of triples ( h , r , t ) where h , t are entities and r is a relation , e.g . ( iphone , hyponym , smartphone ) . E...
In this paper, the authors investigate the main limitations of TransE in the light of loss function. The authors claim that their contributions consist of two parts: 1) proving that the proper selection of loss functions is vital in KGE; 2) proposing a model called TransComplEx. The results show that the proper selecti...
SP:86076eabb48ef1fe9d51b54945bf81ed44bcacd7
Minimally distorted Adversarial Examples with a Fast Adaptive Boundary Attack
1 Introduction . The finding of the vulnerability of neural networks-based classifiers to adversarial examples , that is small perturbations of the input able to modify the decision of the models , started a fast development of a variety of attack algorithms . The high effectiveness of adversarial attacks reveals the f...
The authors propose a new gradient-based method (FAB) for constructing adversarial perturbations for deep neural networks. At a high level, the method repeatedly estimates the decision boundary based on the linearization of the classifier at a given point and projects to the closest "misclassified" example based on tha...
SP:3d3842a5e0816084c5a2406f1b0143d0215b9559
Minimally distorted Adversarial Examples with a Fast Adaptive Boundary Attack
1 Introduction . The finding of the vulnerability of neural networks-based classifiers to adversarial examples , that is small perturbations of the input able to modify the decision of the models , started a fast development of a variety of attack algorithms . The high effectiveness of adversarial attacks reveals the f...
Authors extend deepFool by adding extra steps and constraints to find closer points to the source image as the adversarial image. They both project onto the decision boundary. Deepfool does and adhoc clipping to keep the pixel values in (0,1) but the new proposed method respects the constraints during the steps. Also d...
SP:3d3842a5e0816084c5a2406f1b0143d0215b9559
INFERENCE, PREDICTION, AND ENTROPY RATE OF CONTINUOUS-TIME, DISCRETE-EVENT PROCESSES
The inference of models , prediction of future symbols , and entropy rate estimation of discrete-time , discrete-event processes is well-worn ground . However , many time series are better conceptualized as continuous-time , discrete-event processes . Here , we provide new methods for inferring models , predicting futu...
The authors present a model for time series which are represented as discrete events in continuous time and describe methods for doing parameter inference, future event prediction and entropy rate estimation for such processes. Their model is based on models for Bayesian Structure prediction where they add the temporal...
SP:51a88b77450225e0f80f9fa25510fb4ea64463b2
INFERENCE, PREDICTION, AND ENTROPY RATE OF CONTINUOUS-TIME, DISCRETE-EVENT PROCESSES
The inference of models , prediction of future symbols , and entropy rate estimation of discrete-time , discrete-event processes is well-worn ground . However , many time series are better conceptualized as continuous-time , discrete-event processes . Here , we provide new methods for inferring models , predicting futu...
The paper focuses on the problem of modeling, predicting and estimating entropy information over continuous-time discrete event processes. Specifically, the paper leverages unifilar HSMM's for model inference and then uses the inferred states to make future predictions. The authors also use inferred model with previous...
SP:51a88b77450225e0f80f9fa25510fb4ea64463b2
Graph inference learning for semi-supervised classification
1 INTRODUCTION . Graph , which comprises a set of vertices/nodes together with connected edges , is a formal structural representation of non-regular data . Due to the strong representation ability , it accommodates many potential applications , e.g. , social network ( Orsini et al. , 2017 ) , world wide data ( Page et...
This paper proposes to leverage the between-node-path information into the inference of conventional graph neural network methods. Specifically, the proposed method treats the nodes in training set as a reference corpus and, when infering the label of a specific node, makes this node "attend" to the reference corpus, w...
SP:06bbc70edab65f046adb46bc364c3b91f5880845
Graph inference learning for semi-supervised classification
1 INTRODUCTION . Graph , which comprises a set of vertices/nodes together with connected edges , is a formal structural representation of non-regular data . Due to the strong representation ability , it accommodates many potential applications , e.g. , social network ( Orsini et al. , 2017 ) , world wide data ( Page et...
This paper presents a semi-supervised classification method for classifying unlabeled nodes in graph data. The authors propose a Graph Inference Learning (GIL) framework to learn node labels on graph topology. The node labeling is based of three aspects: 1) node representation to measure the similarity between the cent...
SP:06bbc70edab65f046adb46bc364c3b91f5880845
Ridge Regression: Structure, Cross-Validation, and Sketching
We study the following three fundamental problems about ridge regression : ( 1 ) what is the structure of the estimator ? ( 2 ) how to correctly use cross-validation to choose the regularization parameter ? and ( 3 ) how to accelerate computation without losing too much accuracy ? We consider the three problems in a un...
This paper deals with 3 theoretical properties of ridge regression. First, it proves that the ridge regression estimator is equivalent to a specific representation which is useful as for instance it can be used to derive the training error of the ridge estimator. Second, it provides a bias correction mechanism for ridg...
SP:bbcb77fc764f7e90ef6126d97d8195734fcdafe8
Ridge Regression: Structure, Cross-Validation, and Sketching
We study the following three fundamental problems about ridge regression : ( 1 ) what is the structure of the estimator ? ( 2 ) how to correctly use cross-validation to choose the regularization parameter ? and ( 3 ) how to accelerate computation without losing too much accuracy ? We consider the three problems in a un...
This paper presents a theoretical study of ridge regression, focusing on the practical problems of correcting for the bias of the cross-validation based estimate of the optimal regularisation parameter, and quantification of the asymptotic risk of sketching algorithms for ridge regression, both in the p / n -> gamma in...
SP:bbcb77fc764f7e90ef6126d97d8195734fcdafe8
Incorporating Horizontal Connections in Convolution by Spatial Shuffling
INCORPORATING HORIZONTAL CONNECTIONS IN CONVOLUTION BY SPATIAL SHUFFLING Anonymous authors Paper under double-blind review Convolutional Neural Networks ( CNNs ) are composed of multiple convolution layers and show elegant performance in vision tasks . The design of the regular convolution is based on the Receptive Fie...
The authors extended the regular convolution and proposed spatially shuffled convolution to use the information outside of its RF, which is inspired by the idea that horizontal connections are believed to be important for visual processing in the visual cortex in biological brain. The authors proposed ss convolution fo...
SP:d5ccf8fdd029c2a99dac0441385f280ed3fc03fb
Incorporating Horizontal Connections in Convolution by Spatial Shuffling
INCORPORATING HORIZONTAL CONNECTIONS IN CONVOLUTION BY SPATIAL SHUFFLING Anonymous authors Paper under double-blind review Convolutional Neural Networks ( CNNs ) are composed of multiple convolution layers and show elegant performance in vision tasks . The design of the regular convolution is based on the Receptive Fie...
In this paper, the authors proposed a shuffle strategy for convolution layers in convolutional neural networks (CNNs). Specifically, the authors argued that the receptive field (RF) of each convolutional filter should be not constrained in the small patch. Instead, it should also cover other locations beyond the local ...
SP:d5ccf8fdd029c2a99dac0441385f280ed3fc03fb
Guiding Program Synthesis by Learning to Generate Examples
1 INTRODUCTION . Over the years , program synthesis has been applied to a wide variety of different tasks including string , number or date transformations ( Gulwani , 2011 ; Singh & Gulwani , 2012 ; 2016 ; Ellis et al. , 2019 ; Menon et al. , 2013 ; Ellis & Gulwani , 2017 ) , layout and graphic program generation ( Bi...
A method for a refinement loop for program synthesizers operating on input/ouput specifications is presented. The core idea is to generate several candidate solutions, execute them on several inputs, and then use a learned component to judge which of the resulting input/output pairs are most likely to be correct. This ...
SP:aec7ce88f21b38c205522c88b3a3253e24754182
Guiding Program Synthesis by Learning to Generate Examples
1 INTRODUCTION . Over the years , program synthesis has been applied to a wide variety of different tasks including string , number or date transformations ( Gulwani , 2011 ; Singh & Gulwani , 2012 ; 2016 ; Ellis et al. , 2019 ; Menon et al. , 2013 ; Ellis & Gulwani , 2017 ) , layout and graphic program generation ( Bi...
This paper handles the challenge of generating generalizable programs from input-output specifications when the size of the specification can be quite limited and therefore ambiguous. When proposed candidate programs lead to divergent outputs on a new input, the paper proposes to use a learned neural oracle that can ev...
SP:aec7ce88f21b38c205522c88b3a3253e24754182
Meta-Learning by Hallucinating Useful Examples
1 INTRODUCTION . Modern deep learning models rely heavily on large amounts of annotated examples ( Deng et al. , 2009 ) . Their data-hungry nature limits their applicability to real-world scenarios , where the cost of annotating examples is prohibitive , or they involve rare concepts ( Zhu et al. , 2014 ; Fink , 2011 )...
In this paper, the authors address few-shot learning via a precise collaborative hallucinator. In particular, they follow the framework of (Wang et al., 2018), and introduce two kinds of training regularization. The soft precision-inducing loss follows the spirit of adversarial learning, by using knowledge distillation...
SP:ca085e8e2675fe579df4187290b7b7dc37b8a729
Meta-Learning by Hallucinating Useful Examples
1 INTRODUCTION . Modern deep learning models rely heavily on large amounts of annotated examples ( Deng et al. , 2009 ) . Their data-hungry nature limits their applicability to real-world scenarios , where the cost of annotating examples is prohibitive , or they involve rare concepts ( Zhu et al. , 2014 ; Fink , 2011 )...
This paper describes a method that builds upon the work of Wang et al. It meta-learns to hallucinate additional samples for few-shot learning for classification tasks. Their two main insights of this paper are to propose a soft-precision term which compares the classifiers' predictions for all classes other than the gr...
SP:ca085e8e2675fe579df4187290b7b7dc37b8a729
DyNet: Dynamic Convolution for Accelerating Convolution Neural Networks
1 INTRODUCTION . Convolutional neural networks ( CNNs ) have achieved state-of-the-art performance in many computer vision tasks ( Krizhevsky et al. , 2012 ; Szegedy et al. , 2013 ) , and the neural architectures of CNNs are evolving over the years ( Krizhevsky et al. , 2012 ; Simonyan & Zisserman , 2014 ; Szegedy et a...
The authors propose to use dynamic convolutional kernels as a means to reduce the computation cost in static CNNs while maintaining their performance. The dynamic kernels are obtained by a linear combination of static kernels where the weights of the linear combination are input-dependent (they are obtained similarly t...
SP:28a2ee0012e23223b2c3501a94a5e72e0c718c66
DyNet: Dynamic Convolution for Accelerating Convolution Neural Networks
1 INTRODUCTION . Convolutional neural networks ( CNNs ) have achieved state-of-the-art performance in many computer vision tasks ( Krizhevsky et al. , 2012 ; Szegedy et al. , 2013 ) , and the neural architectures of CNNs are evolving over the years ( Krizhevsky et al. , 2012 ; Simonyan & Zisserman , 2014 ; Szegedy et a...
This paper proposed dynamic convolution (DyNet) to accelerating convolution networks. The new method is tested on the ImageNet dataset with three different backbones. It reduces the computation flops by a large margin while keeps similar classification accuracy. The additional segmentation experiment on the Cityscapes ...
SP:28a2ee0012e23223b2c3501a94a5e72e0c718c66
Star-Convexity in Non-Negative Matrix Factorization
1 INTRODUCTION . Non-negative matrix factorization ( NMF ) is a ubiquitous technique for data analysis where one attempts to factorize a measurement matrix X into the product of non-negative matrices U , V ( Lee and Seung , 1999 ) . This simple problem has applications in recommender systems ( Luo et al. , 2014 ) , sci...
The paper derives results for nonnegative-matrix factorization along the lines of recent results on SGD for DNNs, showing that the loss is star-convex towards randomized planted solutions. The star-convexity property is also shown to hold to some degree on real world datasets. The paper argues that these results explai...
SP:9e712c6f60b19d9309721eea514589755b4ce648
Star-Convexity in Non-Negative Matrix Factorization
1 INTRODUCTION . Non-negative matrix factorization ( NMF ) is a ubiquitous technique for data analysis where one attempts to factorize a measurement matrix X into the product of non-negative matrices U , V ( Lee and Seung , 1999 ) . This simple problem has applications in recommender systems ( Luo et al. , 2014 ) , sci...
This paper studies loss landscape of Non-negative matrix factorization (NMF) when the matrix is very large. It shows that with high probability, the landscape is quasi-convex under some conditions. This suggests that the optimization problem would become easier as the size of the matrix becomes very large. Implications...
SP:9e712c6f60b19d9309721eea514589755b4ce648
SemanticAdv: Generating Adversarial Examples via Attribute-Conditional Image Editing
1 INTRODUCTION . Deep neural networks ( DNNs ) have demonstrated great successes in advancing the state-of-the-art performance of discriminative tasks ( Krizhevsky et al. , 2012 ; Goodfellow et al. , 2016 ; He et al. , 2016 ; Collobert & Weston , 2008 ; Deng et al. , 2013 ; Silver et al. , 2016 ) . However , recent res...
The authors describe a method for adversarially modifying a given (test) example that 1) still retains the correct label on the example, but 2) causes a model to make an incorrect prediction on it. The novelty of their proposed method is that their adversarial modifications are along a provided semantic axis (e.g., cha...
SP:37c8908c43beda4efc9db25216225f0106fe009c
SemanticAdv: Generating Adversarial Examples via Attribute-Conditional Image Editing
1 INTRODUCTION . Deep neural networks ( DNNs ) have demonstrated great successes in advancing the state-of-the-art performance of discriminative tasks ( Krizhevsky et al. , 2012 ; Goodfellow et al. , 2016 ; He et al. , 2016 ; Collobert & Weston , 2008 ; Deng et al. , 2013 ; Silver et al. , 2016 ) . However , recent res...
This paper proposes to generate "unrestricted adversarial examples" via attribute-conditional image editing. Their method, SemanticAdv, leverages disentangled semantic factors and interpolates feature-map with higher freedom than attribute-space. Their adversarial optimization objectives combine both attack effectivene...
SP:37c8908c43beda4efc9db25216225f0106fe009c
Annealed Denoising score matching: learning Energy based model in high-dimensional spaces
1 INTRODUCTION AND MOTIVATION . Treating data as stochastic samples from a probability distribution and developing models that can learn such distributions is at the core for solving a large variety of application problems , such as error correction/denoising ( Vincent et al. , 2010 ) , outlier/novelty detection ( Zhai...
The paper proposes to learn an energy based generative model using an ‘annealed’ denoising score matching objective. The main contribution of the paper is to show that denoising score matching can be trained on a range of noise scales concurrently using a small modification to the loss. Compared to approximate likeliho...
SP:e84523133b0c393a7d673a3faef8cd2d6368830a
Annealed Denoising score matching: learning Energy based model in high-dimensional spaces
1 INTRODUCTION AND MOTIVATION . Treating data as stochastic samples from a probability distribution and developing models that can learn such distributions is at the core for solving a large variety of application problems , such as error correction/denoising ( Vincent et al. , 2010 ) , outlier/novelty detection ( Zhai...
This paper presents a method of learning of energy based models using denoising score matching. This technique has been used before but only with limited success. The authors hypothesize that this is due to the fact that the matching was only performed over a single noise scale. The main idea of this work is to employ ...
SP:e84523133b0c393a7d673a3faef8cd2d6368830a
Gradientless Descent: High-Dimensional Zeroth-Order Optimization
1 INTRODUCTION . We consider the problem of zeroth-order optimization ( also known as gradient-free optimization , or bandit optimization ) , where our goal is to minimize an objective function f : Rn → R with as few evaluations of f ( x ) as possible . For many practical and interesting objective functions , gradients...
This paper proposes stable GradientLess Descent (GLD) algorithms that do not rely on gradient estimate. Based on the low-rank assumption on P_A, the iteration complexity is poly-logarithmically dependent on dimensionality. The theoretical analysis of the main results is based on a geometric perspective, which is intere...
SP:e958fbb0b004f454b79944ca72958254087147d4
Gradientless Descent: High-Dimensional Zeroth-Order Optimization
1 INTRODUCTION . We consider the problem of zeroth-order optimization ( also known as gradient-free optimization , or bandit optimization ) , where our goal is to minimize an objective function f : Rn → R with as few evaluations of f ( x ) as possible . For many practical and interesting objective functions , gradients...
The paper proposes a novel zeroth-order algorithm for high-dimensional optimization. In particular, the algorithm as an instance of direct search algorithms where no attempt is made to estimate the gradient of the function during the optimization process. The authors study the optimization of monotone transformations o...
SP:e958fbb0b004f454b79944ca72958254087147d4
Analyzing Privacy Loss in Updates of Natural Language Models
1 INTRODUCTION . Over the last few years , deep learning has made sufficient progress to be integrated into intelligent , user-facing systems , which means that machine learning models are now part of the regular software development lifecycle . As part of this move towards concrete products , models are regularly re-t...
This paper looks at privacy concerns regarding data for a specific model before and after a single update. It discusses the privacy concerns thoroughly and look at language modeling as a representative task. They find that there are plenty of cases namely when the composition of the sequences involve low frequency word...
SP:9d2476df24b81661dc5ad76b13c8fd5fd1653381
Analyzing Privacy Loss in Updates of Natural Language Models
1 INTRODUCTION . Over the last few years , deep learning has made sufficient progress to be integrated into intelligent , user-facing systems , which means that machine learning models are now part of the regular software development lifecycle . As part of this move towards concrete products , models are regularly re-t...
This paper studies the privacy issue of widely used neural language models in the current literature. The authors consider the privacy implication phenomena of two model snapshots before and after an update. The updating setting considered in this paper is kind of interesting. However, the contribution of the current p...
SP:9d2476df24b81661dc5ad76b13c8fd5fd1653381
Attraction-Repulsion Actor-Critic for Continuous Control Reinforcement Learning
In reinforcement learning , robotic control tasks are often useful for understanding how agents perform in environments with deceptive rewards where the agent can easily become trapped into suboptimal solutions . One way to avoid these local optima is to use a population of agents to ensure coverage of the policy space...
The paper proposes an ensemble method for reinforcement learning in which the policy updates are modulated with a loss which encourages diversity among all experienced policies. It is a combination of SAC, normalizing flow policies, and an approach to diversity considered by Hong et al. (2018). The work seems rather in...
SP:044d99499c4a9cb383f5e39a28fc7ccb700040d1
Attraction-Repulsion Actor-Critic for Continuous Control Reinforcement Learning
In reinforcement learning , robotic control tasks are often useful for understanding how agents perform in environments with deceptive rewards where the agent can easily become trapped into suboptimal solutions . One way to avoid these local optima is to use a population of agents to ensure coverage of the policy space...
RL in environments with deceptive rewards can produce sub-optimal policies. To remedy this, the paper proposes a method for population-based exploration. Multiple actors, each parameterized with policies based on Normalizing Flows (radial contractions), are optimized over iterations using the off-policy SAC algorithm. ...
SP:044d99499c4a9cb383f5e39a28fc7ccb700040d1
From Variational to Deterministic Autoencoders
1 INTRODUCTION . Generative models lie at the core of machine learning . By capturing the mechanisms behind the data generation process , one can reason about data probabilistically , access and traverse the lowdimensional manifold the data is assumed to live on , and ultimately generate new data . It is therefore not ...
This paper propose an extension to deterministic autoencoders. Motivated from VAEs, the authors propose RAEs, which replace the noise injection in the encoders of VAEs with an explicit regularization term on the latent representations. As a result, the model becomes a deterministic autoencoder with a L_2 regularization...
SP:e4f5ca770474ba98dc7643522ea6435f0586c292
From Variational to Deterministic Autoencoders
1 INTRODUCTION . Generative models lie at the core of machine learning . By capturing the mechanisms behind the data generation process , one can reason about data probabilistically , access and traverse the lowdimensional manifold the data is assumed to live on , and ultimately generate new data . It is therefore not ...
The paper studies (the more conventional) deterministic auto-encoders, as they are easier to train than VAE. To then try to maintain the model's capability of approximating the data distribution and to draw/synthesize new unseen samples, the paper both looks at imposing additional regularization terms towards a smooth ...
SP:e4f5ca770474ba98dc7643522ea6435f0586c292
Predictive Coding for Boosting Deep Reinforcement Learning with Sparse Rewards
1 INTRODUCTION . Recent progress in deep reinforcement learning ( DRL ) has enabled robots to learn and execute complex tasks , ranging from game playing ( Jaderberg et al. , 2018 ; OpenAI , 2019 ) , robotic manipulations ( Andrychowicz et al. , 2017 ; Haarnoja et al. , 2018 ) , to navigation ( Zhang et al. , 2017 ) . ...
This paper proposes using the features learned through Contrastive Predictive Coding as a means for reward shaping. Specifically, they propose to cluster the embedding using the clusters to provide feedback to the agent by applying a positive reward when the agent enters the goal cluster. In more complex domains they...
SP:7cd001a35175d8565c046093dcf070ba7fa988d6
Predictive Coding for Boosting Deep Reinforcement Learning with Sparse Rewards
1 INTRODUCTION . Recent progress in deep reinforcement learning ( DRL ) has enabled robots to learn and execute complex tasks , ranging from game playing ( Jaderberg et al. , 2018 ; OpenAI , 2019 ) , robotic manipulations ( Andrychowicz et al. , 2017 ; Haarnoja et al. , 2018 ) , to navigation ( Zhang et al. , 2017 ) . ...
The paper proposes a reward shaping method which aim to tackle sparse reward tasks. The paper first trains a representation using contrastive predictive coding and then uses the learned representation to provide feedback to the control agent. The main difference from the previous work (i.e. CPC) is that the paper uses ...
SP:7cd001a35175d8565c046093dcf070ba7fa988d6
Skew-Explore: Learn faster in continuous spaces with sparse rewards
1 INTRODUCTION Reinforcement Learning ( RL ) is based on performing exploratory actions in a trial-and-error manner and reinforcing those actions that result in superior reward outcomes . Exploration plays an important role in solving a given sequential decision-making problem . A RL agent can not improve its behaviour...
This paper studies the problem of exploration in reinforcement learning. The key idea is to learn a goal-conditioned agent and do exploration by selecting goals at the frontier of previously visited states. This frontier is estimated using an extension of prior work (Pong 2019). The method is evaluated on two continuo...
SP:1e4d48aca131f5ff12775ba51dd1176397038d59
Skew-Explore: Learn faster in continuous spaces with sparse rewards
1 INTRODUCTION Reinforcement Learning ( RL ) is based on performing exploratory actions in a trial-and-error manner and reinforcing those actions that result in superior reward outcomes . Exploration plays an important role in solving a given sequential decision-making problem . A RL agent can not improve its behaviour...
This paper proposes a new exploration algorithm by proposing a new way of generating intrinsic rewards. Specifically, the authors propose to maintain a "novelty frontier" which consists of states that have low-likelihood under some likelihood model trained on their replay buffer. The authors propose to sample from the ...
SP:1e4d48aca131f5ff12775ba51dd1176397038d59
Improving Generalization in Meta Reinforcement Learning using Learned Objectives
1 INTRODUCTION . The process of evolution has equipped humans with incredibly general learning algorithms . They enable us to solve a wide range of problems , even in the absence of a large number of related prior experiences . The algorithms that give rise to these capabilities are the result of distilling the collect...
This paper presents a novel meta reinforcement learning algorithm capable of meta-generalizing to unseen tasks. They make use of a learned objective function used in combination with DDPG style update. Results are presented on different combinations of meta-training and meta-testing on lunar, half cheetah, and hopper e...
SP:9043128647ca5b26b38c11af6fddf166e012a390
Improving Generalization in Meta Reinforcement Learning using Learned Objectives
1 INTRODUCTION . The process of evolution has equipped humans with incredibly general learning algorithms . They enable us to solve a wide range of problems , even in the absence of a large number of related prior experiences . The algorithms that give rise to these capabilities are the result of distilling the collect...
The paper proposes a meta reinforcement learning algorithm called MetaGenRL, which meta-learns learning rules to generalize to different environments. The paper poses an important observation where learning rules in reinforcement learning to train the agents are results of human engineering and design, instead, the pap...
SP:9043128647ca5b26b38c11af6fddf166e012a390
On summarized validation curves and generalization
The validation curve is widely used for model selection and hyper-parameter search with the curve usually summarized over all the training tasks . However , this summarization tends to lose the intricacies of the per-task curves and it isn ’ t able to reflect if all the tasks are at their validation optimum even if the...
The paper examines the common practice of performing model selection by choosing the model that maximizes validation accuracy. In a setting where there are multiple tasks, the average validation error hides performance on individual tasks, which may be relevant. The paper casts multi-class image classification as a mul...
SP:f48d609519e10cdf6de5dd0341edd5544d96402c
On summarized validation curves and generalization
The validation curve is widely used for model selection and hyper-parameter search with the curve usually summarized over all the training tasks . However , this summarization tends to lose the intricacies of the per-task curves and it isn ’ t able to reflect if all the tasks are at their validation optimum even if the...
Model validation curve typically aggregates accuracies of all labels. This paper investigates the fine-grained per-label model validation curve. It shows that the optimal epoch varies by label. The paper proposes a visualization method to detect if there is a disparity between the per-label curves and the summarized va...
SP:f48d609519e10cdf6de5dd0341edd5544d96402c
A Non-asymptotic comparison of SVRG and SGD: tradeoffs between compute and speed
1 INTRODUCTION . Many large-scale machine learning problems , especially in deep learning , are formulated as minimizing the sum of loss functions on millions of training examples ( Krizhevsky et al. , 2012 ; Devlin et al. , 2018 ) . Computing exact gradient over the entire training set is intractable for these problem...
This paper compares SGD and SVRG (as a representative variance reduced method) to explore tradeoffs. Although the computational complexity vs overall convergence performance tradeoff is well-known at this point, an interesting new perspective is the comparison in regions of interpolation (where SGD gradient variance wi...
SP:67c44f33dff59e4d218f753fdbc6296da62cdf62
A Non-asymptotic comparison of SVRG and SGD: tradeoffs between compute and speed
1 INTRODUCTION . Many large-scale machine learning problems , especially in deep learning , are formulated as minimizing the sum of loss functions on millions of training examples ( Krizhevsky et al. , 2012 ; Devlin et al. , 2018 ) . Computing exact gradient over the entire training set is intractable for these problem...
This paper examines the tradeoffs between applying SVRG and SGD for training neural networks by providing an analysis of noisy least squares regression problems as well as experiments on simple MLPs and CNNs on MNIST and CIFAR-10. The theory analyzes a linear model where both the input $x$ and label noise $\epsilon$ fo...
SP:67c44f33dff59e4d218f753fdbc6296da62cdf62
SMiRL: Surprise Minimizing RL in Entropic Environments
1 INTRODUCTION . The general struggle for existence of animate beings is not a struggle for raw materials , nor for energy , but a struggle for negative entropy . ( Ludwig Boltzmann , 1886 ) All living organisms carve out environmental niches within which they can maintain relative predictability amidst the ever-increa...
This paper proposes a novel form of surprise-minimizing intrinsic reward signal that leads to interesting behavior in the absence of an external reward signal. The proposed approach encourages an agent to visit states with high probability / density under a parametric marginal state distribution that is learned as the ...
SP:6022b52e1e160bd034df1a7c71c6ca163bcf4dc0
SMiRL: Surprise Minimizing RL in Entropic Environments
1 INTRODUCTION . The general struggle for existence of animate beings is not a struggle for raw materials , nor for energy , but a struggle for negative entropy . ( Ludwig Boltzmann , 1886 ) All living organisms carve out environmental niches within which they can maintain relative predictability amidst the ever-increa...
This paper proposes Surprise Minimizing RL (SMiRL), a conceptual framework for training a reinforcement learning agent to seek out states with high likelihood under a density model trained on visited states. They qualitatively and quantitatively explore various aspects of the behaviour of these agents and argue that th...
SP:6022b52e1e160bd034df1a7c71c6ca163bcf4dc0
Projected Canonical Decomposition for Knowledge Base Completion
1 INTRODUCTION . The problems of representation learning and link prediction in multi-relational data can be formulated as a binary tensor completion problem , where the tensor is obtained by stacking the adjacency matrices of every relations between entities . This tensor can then be intrepreted as a `` knowledge base...
In this paper, a tensor decomposition method is studied for link prediction problems. The model is based on Tucker decomposition but the core tensor is decomposed as CP decomposition so that it can be seen as an interpolation between Tucker and CP. The performance is evaluated with several NLP data sets (e.g., subject-...
SP:8bdeb36997d6699e48511d9abac87df8c14bd087
Projected Canonical Decomposition for Knowledge Base Completion
1 INTRODUCTION . The problems of representation learning and link prediction in multi-relational data can be formulated as a binary tensor completion problem , where the tensor is obtained by stacking the adjacency matrices of every relations between entities . This tensor can then be intrepreted as a `` knowledge base...
The paper introduces a novel tensor decomposition that is reminiscent of canonical decomposition (CP) with low-rank factors, based on the observation that the core tensor in Tucker decomposition can be decomposed, resulting in a model interpolating between CP and Tucker. The authors argue that a straight application of...
SP:8bdeb36997d6699e48511d9abac87df8c14bd087
Robust And Interpretable Blind Image Denoising Via Bias-Free Convolutional Neural Networks
1 INTRODUCTION AND CONTRIBUTIONS . The problem of denoising consists of recovering a signal from measurements corrupted by noise , and is a canonical application of statistical estimation that has been studied since the 1950 ’ s . Achieving high-quality denoising results requires ( at least implicitly ) quantifying and...
This paper proposed to remove all bias terms in denoising networks to avoid overfitting when different noise levels exist. With analysis, the paper concludes that the dimensions of subspaces of image features are adaptively changing according to the noise level. An interesting result is that the MSE is proportional to ...
SP:62a75399aa97a61432385cf1dffabb674741a18a
Robust And Interpretable Blind Image Denoising Via Bias-Free Convolutional Neural Networks
1 INTRODUCTION AND CONTRIBUTIONS . The problem of denoising consists of recovering a signal from measurements corrupted by noise , and is a canonical application of statistical estimation that has been studied since the 1950 ’ s . Achieving high-quality denoising results requires ( at least implicitly ) quantifying and...
This paper looks at how deep convolutional neural networks for image denoising can generalize across various noise levels. First, they argue that state-of-the-art denoising networks perform poorly outside of the training noise range. The authors empirically show that as denoising performance degrades on unseen noise le...
SP:62a75399aa97a61432385cf1dffabb674741a18a
Pay Attention to Features, Transfer Learn Faster CNNs
1 Introduction . Despite recent successes of CNNs achieving state-of-the-art performance in vision applications ( Tan & Le , 2019 ; Cai & Vasconcelos , 2018 ; Zhao et al. , 2018 ; Ren et al. , 2015 ) , there are two major shortcomings limiting their deployments in real life . First , training CNNs from random initializ...
This paper proposes a method called attentive feature distillation and selection (AFDS) to improve the performance of transfer learning for CNNs. The authors argue that the regularization should constrain the proximity of feature maps, instead of pre-trained model weights. Specifically, the authors proposes two modifi...
SP:35407fdffbf982a97312ef16673be781d593ff22
Pay Attention to Features, Transfer Learn Faster CNNs
1 Introduction . Despite recent successes of CNNs achieving state-of-the-art performance in vision applications ( Tan & Le , 2019 ; Cai & Vasconcelos , 2018 ; Zhao et al. , 2018 ; Ren et al. , 2015 ) , there are two major shortcomings limiting their deployments in real life . First , training CNNs from random initializ...
The paper presents an improvement to the task of transfer learning by being deliberate about which channels from the base model are most relevant to the new task at hand. It does this by apply attentive feature selection (AFS) to select channels or features that align well with the down stream task and attentive featur...
SP:35407fdffbf982a97312ef16673be781d593ff22
Multi-objective Neural Architecture Search via Predictive Network Performance Optimization
1 INTRODUCTION . Recently Neural Architecture Search ( NAS ) has aroused a surge of interest by its potentials of freeing the researchers from tedious and time-consuming architecture tuning for each new task and dataset . Specifically , NAS has already shown some competitive results comparing with hand-crafted architec...
This paper proposed BOGCN-NAS that encodes current architecture with Graph convolutional network (GCN) and uses the feature extracted from GCN as the input to perform a Bayesian regression (predicting bias and variance, See Eqn. 5-6). They use Bayesian Optimization to pick the most promising next model with Expected Im...
SP:d510a4587befa21d3f6b151d437e9d5272ce03a2
Multi-objective Neural Architecture Search via Predictive Network Performance Optimization
1 INTRODUCTION . Recently Neural Architecture Search ( NAS ) has aroused a surge of interest by its potentials of freeing the researchers from tedious and time-consuming architecture tuning for each new task and dataset . Specifically , NAS has already shown some competitive results comparing with hand-crafted architec...
This paper provide a NAS algorithm using Bayesian Optimization with Graph Convolutional Network predictor. The method apply GCN as a surrogate model to adaptively discover and incorporate nodes structure to approximate the performance of the architecture. The method further considers an efficient multi-objective search...
SP:d510a4587befa21d3f6b151d437e9d5272ce03a2
Learning Temporal Coherence via Self-Supervision for GAN-based Video Generation
1 INTRODUCTION . Generative adversarial models ( GANs ) have been extremely successful at learning complex distributions such as natural images ( Zhu et al. , 2017 ; Isola et al. , 2017 ) . However , for sequence generation , directly applying GANs without carefully engineered constraints typically results in strong ar...
Augments the loss of video generation systems with a discriminator that considers multiple frames (as opposed to single frames independently) and a new objective termed ping-pong loss which is introduced in order to deal with “artifacts” that appear in video generation. The paper also proposes a few automatic metrics w...
SP:f719db5d0209fd670518cf1e28a66dfcd9de0a8c
Learning Temporal Coherence via Self-Supervision for GAN-based Video Generation
1 INTRODUCTION . Generative adversarial models ( GANs ) have been extremely successful at learning complex distributions such as natural images ( Zhu et al. , 2017 ; Isola et al. , 2017 ) . However , for sequence generation , directly applying GANs without carefully engineered constraints typically results in strong ar...
The paper presents a novel method for training video-to-video translation (vid2vid) models. The authors introduce a spatio-temporal adversarial discriminator for GAN training, that shows significant benefits over prior methods, in particular, parallel (as opposed to joint) spatial and temporal discriminators. In additi...
SP:f719db5d0209fd670518cf1e28a66dfcd9de0a8c
Towards Stable and Efficient Training of Verifiably Robust Neural Networks
1 INTRODUCTION . The success of deep neural networks ( DNNs ) has motivated their deployment in some safety-critical environments , such as autonomous driving and facial recognition systems . Applications in these areas make understanding the robustness and security of deep neural networks urgently needed , especially ...
This paper proposes a new method for training certifiably robust models that achieves better results than the previous SOTA results by IBP, with a moderate increase in training time. It uses a CROWN-based bound in the warm up phase of IBP, which serves as a better initialization for the later phase of IBP and lead to i...
SP:5c78aac08d907ff07205fe28bf9fa4385c58f40d
Towards Stable and Efficient Training of Verifiably Robust Neural Networks
1 INTRODUCTION . The success of deep neural networks ( DNNs ) has motivated their deployment in some safety-critical environments , such as autonomous driving and facial recognition systems . Applications in these areas make understanding the robustness and security of deep neural networks urgently needed , especially ...
This work proposes CROWN-IBP - novel and efficient certified defense method against adversarial attacks, by combining linear relaxation methods which tend to have tighter bounds with the more efficient interval-based methods. With an attempt to augment the IBP method with its lower computation complexity with the tight...
SP:5c78aac08d907ff07205fe28bf9fa4385c58f40d
Smooth Kernels Improve Adversarial Robustness and Perceptually-Aligned Gradients
Recent research has shown that CNNs are often overly sensitive to high-frequency textural patterns . Inspired by the intuition that humans tend to be more sensitive to lower-frequency ( larger-scale ) patterns , we design a regularization scheme that penalizes large differences between adjacent components within each c...
Paper summary: This paper argues that reducing the reliance of neural networks on high-frequency components of images could help robustness against adversarial examples. To attain this goal, the authors propose a new regularization scheme that encourages convolutional kernels to be smoother. The authors augment standa...
SP:687a3382a219565eb3eb85b707017eb582439565
Smooth Kernels Improve Adversarial Robustness and Perceptually-Aligned Gradients
Recent research has shown that CNNs are often overly sensitive to high-frequency textural patterns . Inspired by the intuition that humans tend to be more sensitive to lower-frequency ( larger-scale ) patterns , we design a regularization scheme that penalizes large differences between adjacent components within each c...
The authors propose a method for learning smoother convolutional kernels with the goal of improving robustness and human alignment. Specifically, they propose a regularizer penalizing large changes between consecutive pixels of the kernel with the intuition of penalizing the use of high-frequency input components. They...
SP:687a3382a219565eb3eb85b707017eb582439565
Discovering Topics With Neural Topic Models Built From PLSA Loss
1 INTRODUCTION . Nowadays , with the digital era , electronic text corpora are ubiquitous . These corpora can be company emails , news groups articles , online journal articles , Wikipedia articles , video metadata ( titles , descriptions , tags ) . These corpora can be very large , thus requiring automatic analysis me...
This paper proposes a neural topic model that aim to discover topics by minimizing a version of the PLSA loss. According to PLSA, a document is presented as a mixture of topics, while a topic is a probability distribution over words, with documents and words assumed independent given topics. Thanks to this assumption, ...
SP:b9b8e3efa69342c90b91dcb29bda1e2f8127581e
Discovering Topics With Neural Topic Models Built From PLSA Loss
1 INTRODUCTION . Nowadays , with the digital era , electronic text corpora are ubiquitous . These corpora can be company emails , news groups articles , online journal articles , Wikipedia articles , video metadata ( titles , descriptions , tags ) . These corpora can be very large , thus requiring automatic analysis me...
I am unimpressed with the quality of writing and presentation, to begin with. There are numerous grammatical errors and typos that make the paper a very difficult read. The presentation also follows an inequitable pattern where the backgrounds and related works are overemphasized and the actual contribution of the pape...
SP:b9b8e3efa69342c90b91dcb29bda1e2f8127581e
GDP: Generalized Device Placement for Dataflow Graphs
1 INTRODUCTION . Neural networks have demonstrated remarkable scalability–improved performance can usually be achieved by training a larger model on a larger dataset ( Hestness et al. , 2017 ; Shazeer et al. , 2017 ; Jozefowicz et al. , 2016 ; Mahajan et al. , 2018 ; Radford et al. ) . Training such large models effici...
In this paper the authors propose an end-to-end policy for graph placement and partitioning of computational graphs produced "under-the-hood" by platforms like Tensorflow. As the sizes of the neural networks increase, using distributed deep learning is becoming more and more necessary. Primitives like the one suggested...
SP:a396624adb04f88f4ba9d10a7968be1926b5d226
GDP: Generalized Device Placement for Dataflow Graphs
1 INTRODUCTION . Neural networks have demonstrated remarkable scalability–improved performance can usually be achieved by training a larger model on a larger dataset ( Hestness et al. , 2017 ; Shazeer et al. , 2017 ; Jozefowicz et al. , 2016 ; Mahajan et al. , 2018 ; Radford et al. ) . Training such large models effici...
This work proposes to use a combination of graph neural networks (GNNs) and proximal policy optimization (PPO) to train policies for generalized device placement in dataflow graphs. Essentially, (1) a GNN is used to learn representations of a dataflow graph (in an inductive manner), (2) a transformer is used to output ...
SP:a396624adb04f88f4ba9d10a7968be1926b5d226
CEB Improves Model Robustness
1 INTRODUCTION . We aim to make models that make meaningful predictions beyond the data they were trained on . Generally we want our models to be robust . Broadly , robustness is the ability of a model to continue making valid predictions as the distribution the model is tested on moves away from the empirical training...
The paper modifies existing classifier architectures and training objective, in order to minimize "conditional entropy bottleneck" (CEB) objective, in attempts to force the representation to maximize the information bottleneck objective. Consequently, the paper claims that this CEB model improves general test accuracy ...
SP:caca11294236433df3e4a14e0ae263ef332372c9
CEB Improves Model Robustness
1 INTRODUCTION . We aim to make models that make meaningful predictions beyond the data they were trained on . Generally we want our models to be robust . Broadly , robustness is the ability of a model to continue making valid predictions as the distribution the model is tested on moves away from the empirical training...
This paper studied the effectiveness of Conditional Entropy Bottleneck (CEB) on improving model robustness. Three tasks are considered to demonstrate its effectiveness; generalization performance over clean test images, adversarially perturbed images, and images corrupted by various synthetic noises. The experiment res...
SP:caca11294236433df3e4a14e0ae263ef332372c9
Inductive representation learning on temporal graphs
1 INTRODUCTION . The technique of learning lower-dimensional vector embeddings on graphs have been widely applied to graph analysis tasks ( Perozzi et al. , 2014 ; Tang et al. , 2015 ; Wang et al. , 2016 ) and deployed in industrial systems ( Ying et al. , 2018 ; Wang et al. , 2018a ) . Most of the graph representation...
This paper addresses the problem of representation learning for temporal graphs. That is, graphs where the topology can evolve over time. The contribution is a temporal graph attention (TGAT) layer aims to exploit learned temporal dynamics of graph evolution in tasks such as node classification and link prediction. Thi...
SP:50073cbe6ab4b44b3c68f141542c1e81df0c5f61
Inductive representation learning on temporal graphs
1 INTRODUCTION . The technique of learning lower-dimensional vector embeddings on graphs have been widely applied to graph analysis tasks ( Perozzi et al. , 2014 ; Tang et al. , 2015 ; Wang et al. , 2016 ) and deployed in industrial systems ( Ying et al. , 2018 ; Wang et al. , 2018a ) . Most of the graph representation...
This paper proposed the temporal graph attention layer which aggregates in-hop features with self-attention and incorporates temporal information with Fourier based relative positional encoding. This idea is novel in GCN field. Experimental results demonstrate that the TGAT which adds temporal encoding outperforms the ...
SP:50073cbe6ab4b44b3c68f141542c1e81df0c5f61
Graph Neural Networks for Reasoning 2-Quantified Boolean Formulas
1 INTRODUCTION . As deep learning makes astonishing achievements in the domain of image ( He et al. , 2016 ) and audio ( Hannun et al. , 2014 ) processing , natural languages ( Vaswani et al. , 2017 ) , and discrete heuristics decisions in games ( Silver et al. , 2017 ) , there is a profound interest in applying the re...
This paper explores how graph neural networks can be applied to test satisfiability of 2QBF logical formulas. They show that a straightforward extension of a GNN-based SAT solver to 2QBF fails to outperform random chance, and argue that this is because proving either satisfiability or unsatisfiability of 2QBF requires ...
SP:8361d709b85b1c717e2cf742dab0145fae667660
Graph Neural Networks for Reasoning 2-Quantified Boolean Formulas
1 INTRODUCTION . As deep learning makes astonishing achievements in the domain of image ( He et al. , 2016 ) and audio ( Hannun et al. , 2014 ) processing , natural languages ( Vaswani et al. , 2017 ) , and discrete heuristics decisions in games ( Silver et al. , 2017 ) , there is a profound interest in applying the re...
This paper investigated the GNN-based solvers for the 2-Quantified Boolean Formula satisfiability problem. This paper points out that GNN has limitations in reasoning about unsatisfiability of SAT problems possibly due to the simple message-passing scheme. To extend the GNN-based SAT solvers to 2-QBF solvers, this pape...
SP:8361d709b85b1c717e2cf742dab0145fae667660
The Probabilistic Fault Tolerance of Neural Networks in the Continuous Limit
1 INTRODUCTION . Understanding the inner working of artificial neural networks ( NNs ) is currently one of the most pressing questions ( 20 ) in learning theory . As of now , neural networks are the backbone of the most successful machine learning solutions ( 37 ; 18 ) . They are deployed in safety-critical tasks in wh...
Review: This paper considers the problem of dropping neurons from a neural network. In the case where this is done randomly, this corresponds to the widely studied dropout algorithm. If the goal is to become robust to randomly dropped neurons during evaluation, then it seems sufficient to just train with dropout (the...
SP:2b8df72b380b893a55a82934afd558d75a3f42f2
The Probabilistic Fault Tolerance of Neural Networks in the Continuous Limit
1 INTRODUCTION . Understanding the inner working of artificial neural networks ( NNs ) is currently one of the most pressing questions ( 20 ) in learning theory . As of now , neural networks are the backbone of the most successful machine learning solutions ( 37 ; 18 ) . They are deployed in safety-critical tasks in wh...
This contribution studies the impact of deletions of random neurons on prediction accuracy of trained architecture, with the application to failure analysis and the specific context of neuromorphic hardware. The manuscript shows that worst-case analysis of failure modes is NP hard and contributes a theoretical analysis...
SP:2b8df72b380b893a55a82934afd558d75a3f42f2
Learning representations for binary-classification without backpropagation
1 INTRODUCTION . A key factor enabling the successes of Deep Learning is the backpropagation of error ( BP ) algorithm ( Rumelhart et al. , 1986 ) . Since it has been introduced , BP has sparked several discussions on whether physical brains are realizing BP-like learning or not ( Grossberg , 1987 ; Crick , 1989 ) . To...
This paper presents an approach towards extending the capabilities of feedback alignment algorithms, that in essence replace the error backpropagation weights with random matrices. The authors propose a particular type of network where all weights are constraint to positive values except the first layers, a monotonica...
SP:8d95af673099b1df7b837f583aa55678d67c5bd6
Learning representations for binary-classification without backpropagation
1 INTRODUCTION . A key factor enabling the successes of Deep Learning is the backpropagation of error ( BP ) algorithm ( Rumelhart et al. , 1986 ) . Since it has been introduced , BP has sparked several discussions on whether physical brains are realizing BP-like learning or not ( Grossberg , 1987 ; Crick , 1989 ) . To...
This paper examines the question of learning in neural networks with random, fixed feedback weights, a technique known as “feedback alignment”. Feedback alignment was originally discovered by Lillicrap et al. (2016; Nature Communications, 7, 13276) when they were exploring potential means of solving the “weight transpo...
SP:8d95af673099b1df7b837f583aa55678d67c5bd6
Quantized Reinforcement Learning (QuaRL)
Recent work has shown that quantization can help reduce the memory , compute , and energy demands of deep neural networks without significantly harming their quality . However , whether these prior techniques , applied traditionally to imagebased models , work with the same efficacy to the sequential decision making pr...
This paper investigates the impact of using a reduced precision (i.e., quantization) in different deep reinforcement learning (DRL) algorithms. It shows that overall, reducing the precision of the neural network in DRL algorithms from 32 bits to 16 or 8 bits doesn't have much effect on the quality of the learned policy...
SP:0cfa52672cf34ffafece1171e48d6c344645dcf3
Quantized Reinforcement Learning (QuaRL)
Recent work has shown that quantization can help reduce the memory , compute , and energy demands of deep neural networks without significantly harming their quality . However , whether these prior techniques , applied traditionally to imagebased models , work with the same efficacy to the sequential decision making pr...
Training and deployment of DRL models is expensive. Quantization has proven useful in supervised learning, however it is yet to be tested thoroughly in DRL. This paper investigates whether quantization can be applied in DRL towards better resource usage (compute, energy) without harming the model quality. Both quantiza...
SP:0cfa52672cf34ffafece1171e48d6c344645dcf3
A critical analysis of self-supervision, or what we can learn from a single image
1 INTRODUCTION . Despite tremendous progress in supervised learning , learning without external supervision remains difficult . Self-supervision has recently emerged as one of the most promising approaches to address this limitation . Self-supervision builds on the fact that convolutional neural networks ( CNNs ) trans...
The paper studies self-supervised learning from very few unlabeled images, down to the extreme case where only a single image is used for training. From the few/single image(s) available for training, a data set of the same size as some unmodified reference data set (ImageNet, Cifar-10/100) is generated through heavy d...
SP:8283eb652046558e12c67447dddebcb52ee9de94
A critical analysis of self-supervision, or what we can learn from a single image
1 INTRODUCTION . Despite tremendous progress in supervised learning , learning without external supervision remains difficult . Self-supervision has recently emerged as one of the most promising approaches to address this limitation . Self-supervision builds on the fact that convolutional neural networks ( CNNs ) trans...
This paper explores self-supervised learning in the low-data regime, comparing results to self-supervised learning on larger datasets. BiGAN, RotNet, and DeepCluster serve as the reference self-supervised methods. It argues that early layers of a convolutional neural network can be effectively learned from a single s...
SP:8283eb652046558e12c67447dddebcb52ee9de94
AutoSlim: Towards One-Shot Architecture Search for Channel Numbers
1 INTRODUCTION . The channel configuration ( a.k.a .. filter numbers or channel numbers ) of a neural network plays a critical role in its affordability on resource constrained platforms , such as mobile phones , wearables and Internet of Things ( IoT ) devices . The most common constraints ( Liu et al. , 2017b ; Huang...
In this paper, the authors propose a method to perform architecture search on the number of channels in convolutional layers. The proposed method, called AutoSlim, is a one-shot approach based on previous work of Slimmable Networks [2,3]. The authors have tested the proposed methods on a variety of architectures on Ima...
SP:5abcf6f6bd3c0079e6f942f614949a3f566afed8
AutoSlim: Towards One-Shot Architecture Search for Channel Numbers
1 INTRODUCTION . The channel configuration ( a.k.a .. filter numbers or channel numbers ) of a neural network plays a critical role in its affordability on resource constrained platforms , such as mobile phones , wearables and Internet of Things ( IoT ) devices . The most common constraints ( Liu et al. , 2017b ; Huang...
This paper proposes a simple and one-shot approach on neural architecture search for the number of channels to achieve better accuracy. Rather than training a lot of network samples, the proposed method trains a single slimmable network to approximate the network accuracy of different channel configurations. The experi...
SP:5abcf6f6bd3c0079e6f942f614949a3f566afed8
Generalized Clustering by Learning to Optimize Expected Normalized Cuts
We introduce a novel end-to-end approach for learning to cluster in the absence of labeled examples . Our clustering objective is based on optimizing normalized cuts , a criterion which measures both intra-cluster similarity as well as inter-cluster dissimilarity . We define a differentiable loss function equivalent to...
This paper presents an end-to-end approach for clustering. The proposed model is called CNC. It simultaneously learns a data embedding that preserve data affinity using Siamese networks, and clusters data in the embedding space. The model is trained by minimizing a differentiable loss function that is derived from norm...
SP:6c5368ae026fc1aaf92bdc208d90e4eec999575a
Generalized Clustering by Learning to Optimize Expected Normalized Cuts
We introduce a novel end-to-end approach for learning to cluster in the absence of labeled examples . Our clustering objective is based on optimizing normalized cuts , a criterion which measures both intra-cluster similarity as well as inter-cluster dissimilarity . We define a differentiable loss function equivalent to...
The paper suggests a differentiable objective that can be used to train a network to output cluster probabilities for a given datapoint, given a fixed number of clusters and embeddings of the data points to be clustered. In particular, this objective can be seen as a relaxation of the normalized cut objective, where in...
SP:6c5368ae026fc1aaf92bdc208d90e4eec999575a