paper_name
stringlengths
11
170
text
stringlengths
8.07k
307k
summary
stringlengths
152
6.16k
paper_id
stringlengths
43
43
Testing For Typicality with Respect to an Ensemble of Learned Distributions
Good methods of performing anomaly detection on high-dimensional data sets are needed , since algorithms which are trained on data are only expected to perform well on data that is similar to the training data . There are theoretical results on the ability to detect if a population of data is likely to come from a know...
This paper analyzes and extends a recently proposed goodness-of-fit test based on typicality [Nalisnick et al., ArXiv 2019]. Firstly, the authors give bounds on the type-II error of this test, showing it can be characterized as a function of KLD[q || p_true] where p is the true data generating process and q is an alte...
SP:1b41929d9d98d6c198b688533d8066bd53ad085c
Testing For Typicality with Respect to an Ensemble of Learned Distributions
Good methods of performing anomaly detection on high-dimensional data sets are needed , since algorithms which are trained on data are only expected to perform well on data that is similar to the training data . There are theoretical results on the ability to detect if a population of data is likely to come from a know...
I machine learning, we often have training data representative of an underlying distribution, and we want to test whether additional data come from the same distribution as the training data (e.g. for outlier/anomaly detection, or model checking). One way to do this is to learn a model of the underlying distribution, a...
SP:1b41929d9d98d6c198b688533d8066bd53ad085c
Skip Connections Matter: On the Transferability of Adversarial Examples Generated with ResNets
1 INTRODUCTION . In deep neural networks ( DNNs ) , a skip connection builds a short-cut from a shallow layer to a deep layer by connecting the input of a convolutional block ( also known as the residual module ) directly to its output . While different layers of a neural network learn different “ levels ” of features ...
This paper proposes a modification to standard Projected Gradient Descent to improve transferability of adversarial examples, when the source model is a ResNet-like model containing skip connections. The method, Skip Gradient Method (SGM) modifies the backwards pass to scale down the gradient computed in each residual ...
SP:8cc8c5179965778ba6b0c6e9a38eeac3d903f579
Skip Connections Matter: On the Transferability of Adversarial Examples Generated with ResNets
1 INTRODUCTION . In deep neural networks ( DNNs ) , a skip connection builds a short-cut from a shallow layer to a deep layer by connecting the input of a convolutional block ( also known as the residual module ) directly to its output . While different layers of a neural network learn different “ levels ” of features ...
The paper is about adversarial attacks and highlights a security weakness of skip connections in ResNet-like CNNs, namely: skip connections make it easier to obtain adversarial examples. This observation leads to new approach to adversarial attacks, named Skip Gradient Method (SGM), which weights the residual gradient ...
SP:8cc8c5179965778ba6b0c6e9a38eeac3d903f579
How Does Learning Rate Decay Help Modern Neural Networks?
1 INTRODUCTION . Modern neural networks are deep , wide , and nonconvex . They are powerful tools for representation learning and serve as core components of deep learning systems . They are top-performing models in language translation ( Sutskever et al. , 2014 ) , visual recognition ( He et al. , 2016 ) , and decisio...
The paper investigates the role of learning rate decay in neural network training. While there are prevalent ideas of how/why learning rate decay help both optimization and generalization of neural networks, this work proposes interpretation based on pattern complexity. The mechanism the paper proposes is that initial ...
SP:b5d287a76a010838b352f4ec537fd83ef1f064cc
How Does Learning Rate Decay Help Modern Neural Networks?
1 INTRODUCTION . Modern neural networks are deep , wide , and nonconvex . They are powerful tools for representation learning and serve as core components of deep learning systems . They are top-performing models in language translation ( Sutskever et al. , 2014 ) , visual recognition ( He et al. , 2016 ) , and decisio...
This paper investigates the way decaying the learning rate helps the training of neural networks. First the paper discusses about other existing hypothesis such as the “Gradient Descent Hypothesis” by Lecun et al 1991 and SGD explanation by Kleinberg et al 2018. Then the paper tries to find contradicting examples again...
SP:b5d287a76a010838b352f4ec537fd83ef1f064cc
XLDA: Cross-Lingual Data Augmentation for Natural Language Inference and Question Answering
1 INTRODUCTION . Recent work on pretraining natural language processing systems ( Devlin et al. , 2018 ; Radford et al. , 2018 ; Howard & Ruder , 2018 ; Peters et al. , 2018 ; McCann et al. , 2017 ) has led to improvements across a wide variety of natural language tasks ( Wang et al. , 2018 ; Rajpurkar et al. , 2016 ; ...
The paper proposes a cross-lingual data augmentation method to improve the language inference and question answering tasks. The core idea is to replace a port of the input text (such as one of the sentence in a sentence pair in the language inference tasks) with its translation in another language. The authors empirica...
SP:a2a80f52b722eafc6b8d751ecfeb8f85dacfe0b8
XLDA: Cross-Lingual Data Augmentation for Natural Language Inference and Question Answering
1 INTRODUCTION . Recent work on pretraining natural language processing systems ( Devlin et al. , 2018 ; Radford et al. , 2018 ; Howard & Ruder , 2018 ; Peters et al. , 2018 ; McCann et al. , 2017 ) has led to improvements across a wide variety of natural language tasks ( Wang et al. , 2018 ; Rajpurkar et al. , 2016 ; ...
The paper provides an analysis of a cross-lingual data augmentation technique dubbed XLDA, which consists of replacing parts of an input text with its translation in another language. Building on the mBERT approach, the authors show that at fine-tuning time it is beneficial to augment the training set of XNLI with cros...
SP:a2a80f52b722eafc6b8d751ecfeb8f85dacfe0b8
Embodied Multimodal Multitask Learning
1 INTRODUCTION . Humans learn language by interacting with a dynamic perceptual environment , grounding words into visual entities and motor actions ( Smith and Gasser , 2005 ; Barsalou , 2008 ) . In recent years , there has been an increased focus on training embodied agents capable of visually-grounded language learn...
The paper describes a Dual-Attention model using Gated- and Spatial-Attention for disentanglement of attributes in feature representations for visually-grounded multitask learning. It has been shown that these models are capable of learning navigational instructions and answering questions. However, they addressed two ...
SP:8671654fe46f948a79f905bd815939dc284ca873
Embodied Multimodal Multitask Learning
1 INTRODUCTION . Humans learn language by interacting with a dynamic perceptual environment , grounding words into visual entities and motor actions ( Smith and Gasser , 2005 ; Barsalou , 2008 ) . In recent years , there has been an increased focus on training embodied agents capable of visually-grounded language learn...
The paper explores multi-task learning in embodied environments and proposes a Dual-Attention Model that disentangles the knowledge of words and visual attributes in the intermediate representations. It addresses two tasks, namely Semantic Goal Navigation (SGN) and Embodied Question Answering (EQA), using a simple synt...
SP:8671654fe46f948a79f905bd815939dc284ca873
Kaleidoscope: An Efficient, Learnable Representation For All Structured Linear Maps
1 INTRODUCTION . Structured linear maps are fundamental and ubiquitous in modern machine learning . Their efficiency in speed ( fast algorithms ) and space ( few parameters ) can reduce computation and memory usage . The class of structured linear maps includes fixed specialized transforms such as the discrete Fourier ...
This paper introduces a structured drop-in replacement for linear layers in a neural network, referred to as Kaleidoscope matrices. The class of such matrices are proven to be highly expressive and includes a very general class of sparse matrices, including convolution, Fastfood, and permutation matrices. Experiments a...
SP:27d5ff5c5032974b4cc0c6af29e414d496d99dfd
Kaleidoscope: An Efficient, Learnable Representation For All Structured Linear Maps
1 INTRODUCTION . Structured linear maps are fundamental and ubiquitous in modern machine learning . Their efficiency in speed ( fast algorithms ) and space ( few parameters ) can reduce computation and memory usage . The class of structured linear maps includes fixed specialized transforms such as the discrete Fourier ...
The authors introduce kaleidoscope matrices (K-matrices) and propose to use them as a substitute for structured matrices arising in ML applications (e.g. circulant matrix used for the convolution operation). The authors prove that K-matrices are expressive enough to capture any structured matrix with near-optimal space...
SP:27d5ff5c5032974b4cc0c6af29e414d496d99dfd
Unsupervised Model Selection for Variational Disentangled Representation Learning
1 INTRODUCTION . Happy families are all alike ; every unhappy family is unhappy in its own way . — Leo Tolstoy , Anna Karenina Despite the success of deep learning in the recent years ( Hu et al. , 2018 ; Espeholt et al. , 2018 ; Silver et al. , 2018 ; Lample et al. , 2018 ; Hessel et al. , 2017 ; Oord et al. , 2016 ) ...
The paper proposes a metric for unsupervised model (and hyperparameter) selection for VAE-based models. The essential basis for the metric is to rank the models based on how much disentanglement they provide. This method relies on a key observation from this paper [A] viz., disentangled representations by any VAE-based...
SP:4c48dff5afc7fefe00e4c7e92e319ae4a68165cd
Unsupervised Model Selection for Variational Disentangled Representation Learning
1 INTRODUCTION . Happy families are all alike ; every unhappy family is unhappy in its own way . — Leo Tolstoy , Anna Karenina Despite the success of deep learning in the recent years ( Hu et al. , 2018 ; Espeholt et al. , 2018 ; Silver et al. , 2018 ; Lample et al. , 2018 ; Hessel et al. , 2017 ; Oord et al. , 2016 ) ...
This paper addresses the problem of unsupervised model selection for disentangled representation learning. Based on the understanding of “why VAEs disentangle” [Burgess et al. 2017, Locatello et al. 2018, Mathieu et al. 2019, Rolinek et al. 2019], the authors adopt the assumption that disentangled representations are a...
SP:4c48dff5afc7fefe00e4c7e92e319ae4a68165cd
On Solving Minimax Optimization Locally: A Follow-the-Ridge Approach
1 INTRODUCTION We consider differentiable sequential games with two players : a leader who can commit to an action , and a follower who responds after observing the leader ’ s action . Particularly , we focus on the zero-sum case of this problem which is also known as minimax optimization , i.e. , min x∈Rn max y∈Rm f (...
This paper designs a set of dynamics for learning in games called follow-the-ridge with the goal of finding local stackelberg equilibria. The main theoretical results show that the only stable attractors of the dynamics are stackelberg equilibria. Moreover, the authors give a deterministic convergence rate for the vani...
SP:27b73923af173446aa087b192767dedd7119231b
On Solving Minimax Optimization Locally: A Follow-the-Ridge Approach
1 INTRODUCTION We consider differentiable sequential games with two players : a leader who can commit to an action , and a follower who responds after observing the leader ’ s action . Particularly , we focus on the zero-sum case of this problem which is also known as minimax optimization , i.e. , min x∈Rn max y∈Rm f (...
The present work proposes a new algorithm, "Follow the Ridge" (FR) that uses second order gradient information to iteratively find local minimax points, or Stackelberg equilibria in two player continuous games. The authors show rigorously that the only stable fixed points of their algorithm are local minimax points and...
SP:27b73923af173446aa087b192767dedd7119231b
EXACT ANALYSIS OF CURVATURE CORRECTED LEARNING DYNAMICS IN DEEP LINEAR NETWORKS
1 INTRODUCTION . Difficulty in training deep neural networks arises from the fact that the network ’ s input-output map fθ ( · ) is nonlinearly related to its parameters θ . This causes non-convex loss landscape with proliferation of saddle-points and poorly-conditioned curvature where gradient-based first order optimi...
Authors analyse curvature corrected optimization methods in the context of deep learning. They build their analysis on Saxe et.al.s work. They show that curvature corrected methods preserve properties of SGD. They also show the disadvantages of layer restricted approximations. They show the importance of time scales in...
SP:72f151a2ffa8c63b2d7740ba2d2074ca6125c3ba
EXACT ANALYSIS OF CURVATURE CORRECTED LEARNING DYNAMICS IN DEEP LINEAR NETWORKS
1 INTRODUCTION . Difficulty in training deep neural networks arises from the fact that the network ’ s input-output map fθ ( · ) is nonlinearly related to its parameters θ . This causes non-convex loss landscape with proliferation of saddle-points and poorly-conditioned curvature where gradient-based first order optimi...
In this manuscript, the authors analyze the dynamics of training deep linear neural networks under a generalized family of natural gradient methods that apply curvature corrections. They first show that the learning trajectory (direction of singular mode dynamics) in natural gradient descent follows the same path as gr...
SP:72f151a2ffa8c63b2d7740ba2d2074ca6125c3ba
Combining graph and sequence information to learn protein representations
1 INTRODUCTION . Proteins can be described by their primary , secondary , tertiary , and quaternary structure or even as nodes in protein-protein interaction networks ( Creighton , 1993 ) . Some proteins with similar sequences play similar roles ; others with high levels of sequence similarity can play different roles ...
This paper introduces a method to incorporate both sequence information and graph information to learn the protein representations. The idea is very straightforward. Basically, it used the embedding from OhmNet [Marinka et al, 2017] for the graph information and used the sequence information from UniRep [Ethan et al, 2...
SP:3d4981a5b80d3f1f2b1249fa1a310988dfc81a91
Combining graph and sequence information to learn protein representations
1 INTRODUCTION . Proteins can be described by their primary , secondary , tertiary , and quaternary structure or even as nodes in protein-protein interaction networks ( Creighton , 1993 ) . Some proteins with similar sequences play similar roles ; others with high levels of sequence similarity can play different roles ...
In this study, the authors develop a method to predict the function of proteins from their structure as well as the network of proteins with which they interact in a given tissue. The method consists in training a linear classifier on the output of two existing embedding methods, UniRep/SeqVec and OhmNet, respectively ...
SP:3d4981a5b80d3f1f2b1249fa1a310988dfc81a91
Multi-source Multi-view Transfer Learning in Neural Topic Modeling with Pretrained Topic and Word Embeddings
1 INTRODUCTION . Probabilistic topic models , such as LDA ( Blei et al. , 2003 ) , Replicated Softmax ( RSM ) ( Salakhutdinov & Hinton , 2009 ) and Document Neural Autoregressive Distribution Estimator ( DocNADE ) ( Larochelle & Lauly , 2012 ) are often used to extract topics from text collections and learn latent docu...
On the basis of existing topic modelling approaches, the authors apply a transfer learning approach to incorporate additional knowledge to topic models, using both word embeddings and topic models. The underlying idea is that topic models contain a global view that differs on a thematic level, while word embeddings con...
SP:979cb5eda94e85ac70c6652abb1580295f39c46b
Multi-source Multi-view Transfer Learning in Neural Topic Modeling with Pretrained Topic and Word Embeddings
1 INTRODUCTION . Probabilistic topic models , such as LDA ( Blei et al. , 2003 ) , Replicated Softmax ( RSM ) ( Salakhutdinov & Hinton , 2009 ) and Document Neural Autoregressive Distribution Estimator ( DocNADE ) ( Larochelle & Lauly , 2012 ) are often used to extract topics from text collections and learn latent docu...
The paper proposes a multi-source and multi-view transfer learning for neural topic modelling with the pre-trained topic and word embedding. The method is based on NEURAL AUTOREGRESSIVE TOPIC MODELs --- DocNADE (Larochelle&Lauly,2012). DocNADE learns topics using language modelling framework. DocNADEe (Gupta et al., 20...
SP:979cb5eda94e85ac70c6652abb1580295f39c46b
Convolutional Tensor-Train LSTM for Long-Term Video Prediction
1 INTRODUCTION . Understanding dynamics of videos and performing long-term predictions of the future is a highly challenging problem . It entails learning complex representation of real-world environment without external supervision . This arises in a wide range of applications , including autonomous driving , robot co...
This paper proposes a method that saves memory and computation in the task of video prediction by low-rank tensor representations via tensor decomposition. The method is able to outperform standard convolutional lstm and other methods by using less parameters when testing it in the Moving MNIST and KTH datasets. The au...
SP:a289020322570c222c7bfdd2c6da0bd2cac95381
Convolutional Tensor-Train LSTM for Long-Term Video Prediction
1 INTRODUCTION . Understanding dynamics of videos and performing long-term predictions of the future is a highly challenging problem . It entails learning complex representation of real-world environment without external supervision . This arises in a wide range of applications , including autonomous driving , robot co...
This paper proposed a convolutional tensor-train (CTT) format based high-order and convolutional LSTM approach for long-term video prediction. This paper is well-motivated. Video data usually have high dimensional input, and the proposed method aims to explicitly take into account more than one hidden representation of...
SP:a289020322570c222c7bfdd2c6da0bd2cac95381
Order Learning and Its Application to Age Estimation
1 INTRODUCTION . To measure the quality of something , we often compare it with other things of a similar kind . Before assigning 4 stars to a film , a critic would have thought , “ It is better than 3-star films but worse than 5-stars. ” This ranking through pairwise comparisons is done in various decision processes (...
The paper proposes a method for learning partial orders based on learning fuzzy pairwise comparisons (smaller/greater/approximatively equal to), and the retaining of a set of representants in each chain in order to allow consistent result by consistency maximization. The method is applied to the problem of estimating t...
SP:1c4adc8ff01ce8ca27baf0d7e438634fa84e26e7
Order Learning and Its Application to Age Estimation
1 INTRODUCTION . To measure the quality of something , we often compare it with other things of a similar kind . Before assigning 4 stars to a film , a critic would have thought , “ It is better than 3-star films but worse than 5-stars. ” This ranking through pairwise comparisons is done in various decision processes (...
This paper presents an order learning method and applies it to age estimation from facial images. It designs a pairwise comparator that categorizes ordering relationship between two instances into ternary classes of greater than, similar, and smaller than. Instead of directly estimating the class of each instance, it l...
SP:1c4adc8ff01ce8ca27baf0d7e438634fa84e26e7
Equivariant neural networks and equivarification
1 INTRODUCTION . One key issue in deep neural network training is the difficulty of tuning parameters , especially when the network size grows larger and larger Han et al . ( 2015 ) . In order to reduce the complexity of the network , many techniques have been proposed by analyzing the structural characteristics of dat...
In this work, the authors employ concepts from group theory to turn an arbitrary feed forward neural network into an equivariant one, i.e. a network whose output transforms in a way that is consistent with the transformation of the input. To this end, the authors first introduce the basic concepts of group theory requi...
SP:a6047a76fb92417053518625713c7174eab16680
Equivariant neural networks and equivarification
1 INTRODUCTION . One key issue in deep neural network training is the difficulty of tuning parameters , especially when the network size grows larger and larger Han et al . ( 2015 ) . In order to reduce the complexity of the network , many techniques have been proposed by analyzing the structural characteristics of dat...
The paper adds an interesting new perspective to equivariant neural nets. However, the actual construction looks equivalent to steerable neural nets to me (see the papers by Cohen and Welling). The generalization of steerable nets has been published under the name "gauge equivariant neural nets", it would be very inter...
SP:a6047a76fb92417053518625713c7174eab16680
On Variational Learning of Controllable Representations for Text without Supervision
1 INTRODUCTION . High-dimensional data , such as images and text , are often causally generated through the interaction of many complex factors , such as lighting and pose in images or style and content in texts . Recently , VAEs and other unsupervised generative models have found successes in modelling the manifold of...
The paper "On Variational Learning of Controllable Representations for Text without Supervision" tackles the problem of latent vacancy of text representation via variational text auto-encoders. Based on the observation that a single factor of the sentence encoding gathers most of relevant information for classifying th...
SP:ac74f77f4d4c8c6c2ca9304bb1050aa22a87df63
On Variational Learning of Controllable Representations for Text without Supervision
1 INTRODUCTION . High-dimensional data , such as images and text , are often causally generated through the interaction of many complex factors , such as lighting and pose in images or style and content in texts . Recently , VAEs and other unsupervised generative models have found successes in modelling the manifold of...
This paper presents a method for controlled text generation by using a new loss function (standard VAE loss with auxiliary losses added on). The method is tested on style transfer datasets: Yelp and Amazon. The central hypothesis is that when manipulating latent codes of a VAE, you can end up in low-density regions of ...
SP:ac74f77f4d4c8c6c2ca9304bb1050aa22a87df63
Aggregating explanation methods for neural networks stabilizes explanations
Despite a growing literature on explaining neural networks , no consensus has been reached on how to explain a neural network decision or how to evaluate an explanation . Our contributions in this paper are twofold . First , we investigate schemes to combine explanation methods and reduce model uncertainty to obtain a ...
The paper has two main messages: 1- Averaging over the explanation (saliency map in the case of image data) of different methods results in a smaller error than an expected error of a single explanation method. 2- Introducing a new saliency map evaluation method by seeking to mitigate the effect of high spatial correla...
SP:eb5d45ef0112f93ade7aa89d9a5132062590f9e1
Aggregating explanation methods for neural networks stabilizes explanations
Despite a growing literature on explaining neural networks , no consensus has been reached on how to explain a neural network decision or how to evaluate an explanation . Our contributions in this paper are twofold . First , we investigate schemes to combine explanation methods and reduce model uncertainty to obtain a ...
This paper, inspired by the established technique of model ensembling, proposes two methods (AGG-Mean and AGG-Var) for aggregating different model explanations into a single unified explanation. The authors mathematically prove that the derived explanation is guaranteed to be more truthful than the average performance ...
SP:eb5d45ef0112f93ade7aa89d9a5132062590f9e1
Do Deep Neural Networks for Segmentation Understand Insideness?
1 INTRODUCTION . Image segmentation is necessary for complete image understanding . A key component of image segmentation is to determine whether a pixel is inside or outside a region , ie . the “ insideness ” problem ( Ullman , 1984 ; 1996 ) . Deep Neural Networks ( DNNs ) have been tremendously successful in image se...
This submission introduces a new concept, termed insideness, to study semantic segmentation in deep learning era. The authors raise many interesting questions, such as (1) Does deep neural networks (DNN) understand insideness? (2) What representations do DNNs use to address the long-range relationships of insideness? (...
SP:6aab83c6e2805838fee314ae400ce5a8bb08f8f3
Do Deep Neural Networks for Segmentation Understand Insideness?
1 INTRODUCTION . Image segmentation is necessary for complete image understanding . A key component of image segmentation is to determine whether a pixel is inside or outside a region , ie . the “ insideness ” problem ( Ullman , 1984 ; 1996 ) . Deep Neural Networks ( DNNs ) have been tremendously successful in image se...
This paper investigates the problem of modeling insideness using neural networks. To this end, the authors carefully designed both feedforward and recurrent neural networks, which are, in principle, able to learn the insideness in its global optima. For evaluation, these methods are trained to predict the insideness in...
SP:6aab83c6e2805838fee314ae400ce5a8bb08f8f3
Understanding Knowledge Distillation in Non-autoregressive Machine Translation
1 INTRODUCTION . Traditional neural machine translation ( NMT ) systems ( Bahdanau et al. , 2015 ; Gehring et al. , 2017 ; Vaswani et al. , 2017 ) generate sequences in an autoregressive fashion ; each target token is predicted step-by-step by conditioning on the previous generated tokens in a monotonic ( e.g . left-to...
In this paper, the authors investigate non-autoregressive translation (NAT). They specifically look into how using different auto-regressive translation (AT) models for knowledge distillation impacts the quality of NAT models. The paper is well organised and the experiments are sound and interesting, shedding light on ...
SP:92370bd193d2a808b6803cb3a22ab3d690f1e13d
Understanding Knowledge Distillation in Non-autoregressive Machine Translation
1 INTRODUCTION . Traditional neural machine translation ( NMT ) systems ( Bahdanau et al. , 2015 ; Gehring et al. , 2017 ; Vaswani et al. , 2017 ) generate sequences in an autoregressive fashion ; each target token is predicted step-by-step by conditioning on the previous generated tokens in a monotonic ( e.g . left-to...
The paper analyses recent distillation techniques for non-autoregressive machine translation models (NAT). These models use a autoregressive teacher (AT), which typically perform better. However, AT models can not be parallelized that easily such as the NAT models. The distillation has the effect of removing modes from...
SP:92370bd193d2a808b6803cb3a22ab3d690f1e13d
ShardNet: One Filter Set to Rule Them All
Deep CNNs have achieved state-of-the-art performance on numerous machine learning and computer vision tasks in recent years , but as they have become increasingly deep , the number of parameters they use has also increased , making them hard to deploy in memory-constrained environments and difficult to interpret . Mach...
In this paper, the authors propose to use the *same* convolutional layer in every layer of a DNN. The network effectively is converted into repeatedly applying the same convolutional filter at multiple scales. The idea is motivated by wavelet decompositions and related work. The authors show that by repeatedly applying...
SP:3ed52251f2462bc5bb61c384ecafc1cce376ef26
ShardNet: One Filter Set to Rule Them All
Deep CNNs have achieved state-of-the-art performance on numerous machine learning and computer vision tasks in recent years , but as they have become increasingly deep , the number of parameters they use has also increased , making them hard to deploy in memory-constrained environments and difficult to interpret . Mach...
This paper proposes to modify a standard CNN by requiring all of its layers to share the same filter set, essentially allowing it to be expressed as an iterative (or recurrent) network. This also has the effect of forcing the same number of feature channels to be used throughout the network. For ResNet-like architect...
SP:3ed52251f2462bc5bb61c384ecafc1cce376ef26
Conservative Uncertainty Estimation By Fitting Prior Networks
1 INTRODUCTION . Deep learning has achieved huge success in many applications . In particular , increasingly often , it is used as a component in decision-making systems . In order to have confidence in decisions made by such systems , it is necessary to obtain good uncertainty estimates , which quantify how certain th...
This paper introduces a new method for uncertainty estimation which utilizes randomly initialized networks. Essentially, instead of training a single predictor that outputs means and uncertainty estimates together, authors propose to have two separate models: one that outputs means, and one that outputs uncertainties. ...
SP:9a703a4562558d32a372047cd46cfe57b3695d38
Conservative Uncertainty Estimation By Fitting Prior Networks
1 INTRODUCTION . Deep learning has achieved huge success in many applications . In particular , increasingly often , it is used as a component in decision-making systems . In order to have confidence in decisions made by such systems , it is necessary to obtain good uncertainty estimates , which quantify how certain th...
This work introduces a simple technique to obtain uncertainty estimates for deep neural networks. This is achieved by having a set of random networks (i.e. neural networks where their parameters are randomly initialized) and then computing an uncertainty value based on the difference in the predictions between those ra...
SP:9a703a4562558d32a372047cd46cfe57b3695d38
Cross-Iteration Batch Normalization
1 INTRODUCTION . Batch Normalization ( BN ) ( Ioffe & Szegedy , 2015 ) has played a significant role in the success of deep neural networks . It was introduced to address the issue of internal covariate shift , where the distribution of network activations changes during training iterations due to the updates of networ...
Paper summary: This paper proposes a new normalization technique specially designed for settings with small mini-batch sizes (where previous methods like BatchNorm are known to suffer). The approach aggregates mean/variance statistics from previous iterations, weighted based on the Taylor expansion, to get a better est...
SP:1707af8ace423f653ad0355d3a363fa1af8c7daf
Cross-Iteration Batch Normalization
1 INTRODUCTION . Batch Normalization ( BN ) ( Ioffe & Szegedy , 2015 ) has played a significant role in the success of deep neural networks . It was introduced to address the issue of internal covariate shift , where the distribution of network activations changes during training iterations due to the updates of networ...
This paper proposes a novel Cross-Iteration Batch Normalization (CBN) to address the limitation of BN in the case of small mini-batch sizes. Different from existing methods, CBN exploits the statistics cross different iterations to obtain more accurate estimates of the data statistics. Specifically, the proposed CBN us...
SP:1707af8ace423f653ad0355d3a363fa1af8c7daf
Characterize and Transfer Attention in Graph Neural Networks
1 INTRODUCTION . The modeling of graphs has become an active research topic in deep learning ( Bronstein et al. , 2017 ) . Dozens of neural network models have been developed for exploiting the structural information of graphs ( Scarselli et al. , 2009 ; Bruna et al. , 2014 ; Henaff et al. , 2015 ; Duvenaud et al. , 20...
This paper presents an empirical study of the attention mechanism in the graph attention networks (GAT). The study reveals that the attention patterns largely depend on the dataset, on some datasets they are sharp, but on others the attention patterns are almost uniform and not so different from the uniform aggregatio...
SP:a037146bb5c073f2764346596ec1f13c7391d894
Characterize and Transfer Attention in Graph Neural Networks
1 INTRODUCTION . The modeling of graphs has become an active research topic in deep learning ( Bronstein et al. , 2017 ) . Dozens of neural network models have been developed for exploiting the structural information of graphs ( Scarselli et al. , 2009 ; Bruna et al. , 2014 ; Henaff et al. , 2015 ; Duvenaud et al. , 20...
This paper carries out several kinds of analysis on the GAT networks of Velickovic (2018), which augment GNN updates with multihead self attention. Three standard attention types are compared, on several different datasets, and differences between uniform attention and learned attention are reported. An experiment is c...
SP:a037146bb5c073f2764346596ec1f13c7391d894
Conditional generation of molecules from disentangled representations
Though machine learning approaches have shown great success in estimating properties of small molecules , the inverse problem of generating molecules with desired properties remains challenging . This difficulty is in part because the set of molecules which have a given property is structurally very diverse . Treating ...
The authors introduce a variational autoencoder for conditional generation of molecules. The model is borrowed from text-based style transfer, applied here on sequence (SMILES) representation of molecules rather than viewing molecules as graphs (as more recent approaches). From a modeling point of view, the main new pa...
SP:e4482ea19c071040799a23293a00fef8305126d5
Conditional generation of molecules from disentangled representations
Though machine learning approaches have shown great success in estimating properties of small molecules , the inverse problem of generating molecules with desired properties remains challenging . This difficulty is in part because the set of molecules which have a given property is structurally very diverse . Treating ...
This paper proposes a VAE-based conditional molecular graph generation model. For that purpose, the disentanglement approach is adopted in this paper: learn to separate property information from the structure representation of a molecular graph. The authors use the supervised VAE objective since the KL regularizer in t...
SP:e4482ea19c071040799a23293a00fef8305126d5
Episodic Reinforcement Learning with Associative Memory
1 INTRODUCTION . Deep reinforcement learning ( RL ) has achieved remarkable performance on extensive complex domains ( Mnih et al. , 2015 ; Lillicrap et al. , 2016 ; Silver et al. , 2016 ; Schulman et al. , 2017 ) . Deep RL research largely focuses on parametric methods , which usually depend on a parametrized value fu...
The paper proposes a new method for organizing episodic memory in with deep Q-networks. It organizes the memory as a graph in which nodes are internal representations of observed states and edges link state transitions. Additionally, nodes from different episodes are merged into a single node if they represent the same...
SP:961781ea8113343d82568b49c26f0889d5632aba
Episodic Reinforcement Learning with Associative Memory
1 INTRODUCTION . Deep reinforcement learning ( RL ) has achieved remarkable performance on extensive complex domains ( Mnih et al. , 2015 ; Lillicrap et al. , 2016 ; Silver et al. , 2016 ; Schulman et al. , 2017 ) . Deep RL research largely focuses on parametric methods , which usually depend on a parametrized value fu...
This paper proposes Episode Reinforcement Learning with Associative Memory (ERLAM), which maintains a graph based on the state transitions (i.e. nodes correspond to states, and edges correspond to transitions) and propagates the values through the edges in the graph in the reverse order of each trajectory. The learned ...
SP:961781ea8113343d82568b49c26f0889d5632aba
Variational Information Bottleneck for Unsupervised Clustering: Deep Gaussian Mixture Embedding
In this paper , we develop an unsupervised generative clustering framework that combines variational information bottleneck and the Gaussian Mixture Model . Specifically , in our approach we use the variational information bottleneck method and model the latent space as a mixture of Gaussians . We derive a bound on the...
This paper considers the autoencoder model combining the usual information bottleneck and the Gaussian mixture model (GMM). Using an approximation to deal with GMMs, the authors derive a bound on the cost function generalizing the ELBO. The performance of the proposed method is tested on three benchmark datasets and co...
SP:24d509d2318c32e01958afb57b50ec166fdb872f
Variational Information Bottleneck for Unsupervised Clustering: Deep Gaussian Mixture Embedding
In this paper , we develop an unsupervised generative clustering framework that combines variational information bottleneck and the Gaussian Mixture Model . Specifically , in our approach we use the variational information bottleneck method and model the latent space as a mixture of Gaussians . We derive a bound on the...
The author(s) posit a Mixture of Gaussian's prior for a compressed latent space representation of high-dimensional data (e.g. images and documents). They propose fitting this model using the Variational Information Bottleneck paradigm and explicate its derivation and tie it to the variational objective used by similar ...
SP:24d509d2318c32e01958afb57b50ec166fdb872f
Match prediction from group comparison data using neural networks
1 INTRODUCTION . The most elementary form of comparisons is pairwise : we often compare a pair of items and make judgments as to which one is of higher utility or simply preferable over the other . With a large amount of such comparison data , one can consider various interesting tasks . One may wish to predict future ...
This paper proposed a novel architecture to tackle the match prediction problem. There are two/three modules in the architecture, the R/P modules and the G module. R/P modules take the current utility estimates of the individuals in a given group comparison as input and produce the current R/P estimates for the indivi...
SP:61e38c36fc69f1cc6f7867971e75e06f7248283b
Match prediction from group comparison data using neural networks
1 INTRODUCTION . The most elementary form of comparisons is pairwise : we often compare a pair of items and make judgments as to which one is of higher utility or simply preferable over the other . With a large amount of such comparison data , one can consider various interesting tasks . One may wish to predict future ...
This paper attempts to solve match prediction problem, i.e., whether a group is preferred over the other. The key challenge is "consistency" since it's hard to find the universal pattern over tasks. Instead, this paper propose to learn reward and penalty modules and both vary when the underlying model changes. Experime...
SP:61e38c36fc69f1cc6f7867971e75e06f7248283b
THE EFFECT OF ADVERSARIAL TRAINING: A THEORETICAL CHARACTERIZATION
1 INTRODUCTION . Despite the impressive performance of deep neural networks on various learning tasks , the widely existing adversarial examples ( Goodfellow et al. , 2014 ; Szegedy et al. , 2017 ) has thwarted its application in the safety-sensitive scenarios ( Kurakin et al. , 2016 ; Chen et al. , 2015 ) . A well tra...
The aim of this paper is to provide a theoretical analysis of adversarial training under the linear classification setting. The main result states that, under many technical assumptions, adversarial training using gradient descent may converge to the hard margin SVM classifier with a fast rate. Here "fast" is not the ...
SP:3c637265c7844256d8e73afb0d1ac811db505c73
THE EFFECT OF ADVERSARIAL TRAINING: A THEORETICAL CHARACTERIZATION
1 INTRODUCTION . Despite the impressive performance of deep neural networks on various learning tasks , the widely existing adversarial examples ( Goodfellow et al. , 2014 ; Szegedy et al. , 2017 ) has thwarted its application in the safety-sensitive scenarios ( Kurakin et al. , 2016 ; Chen et al. , 2015 ) . A well tra...
This paper provides some analyses of the difference between adversarial training and standard training for linear classification problem. In particular, it proves that when the data is \eps linearly separable, adversarial training converges faster than standard trading. It also argues that when the data is not \eps lin...
SP:3c637265c7844256d8e73afb0d1ac811db505c73
Channel Equilibrium Networks
1 INTRODUCTION . Normalization is an important technique for a wide range of tasks such as image classification ( Ioffe & Szegedy , 2015 ) , object detection ( He et al. , 2017a ; Wu & He , 2018 ) , and image generation ( Miyato et al. , 2018 ) . In recent years , a lot of work improved normalization methods , such as ...
The authors point out that CNNs can develop collapsed channels that limit their capacity. They propose to remedy this with batch decorrelation (BD), which focuses on ensuring that channels play an equal role in the feature map and are less likely to collapse. The claim is supported with experiments on CIFAR10, ImageNet...
SP:8939f2377046904b82dd2b219dc2df9b008078b4
Channel Equilibrium Networks
1 INTRODUCTION . Normalization is an important technique for a wide range of tasks such as image classification ( Ioffe & Szegedy , 2015 ) , object detection ( He et al. , 2017a ; Wu & He , 2018 ) , and image generation ( Miyato et al. , 2018 ) . In recent years , a lot of work improved normalization methods , such as ...
This paper studies the channel-collapsed problem in CNNs using 'BN+ReLU' . The Channel Equilibrium block which consists of batch decorrelation branch and adaptive instance inverse branch are proposed to reduce the channel-level sparsity. Experiments on ImageNet and COCO demonstrate that the proposed CE block can achiev...
SP:8939f2377046904b82dd2b219dc2df9b008078b4
Physics-Aware Flow Data Completion Using Neural Inpainting
1 INTRODUCTION . Realistically modeling and predicting fluid phenomena is important to a large number of applications , which may range from optimizing objects ’ aerodynamic properties to creating special effects in movies . Fluids are commonly modelled by numerically solving the Navier-Stokes equations , however compu...
In this paper the authors adopt prior work in image inpainting to the problem of 2d fluid velocity field inpainting by extending the network architecture and using additional loss functions. Specifically, the U-net network is extended with a DenseBlock in the middle, and a separate branch of the network is added which ...
SP:71b7633050462a3cfc3d64e81ae3f6cec758f068
Physics-Aware Flow Data Completion Using Neural Inpainting
1 INTRODUCTION . Realistically modeling and predicting fluid phenomena is important to a large number of applications , which may range from optimizing objects ’ aerodynamic properties to creating special effects in movies . Fluids are commonly modelled by numerically solving the Navier-Stokes equations , however compu...
I am not an expert in recent Navier-Stokes approaches, but note that there is a lot of recent work in physics aware modeling. Specifically the sections on e.g. loss seem to have a lot of prior work. It’s difficult for me to judge the exact amount of novelty in this paper with respect to the physics. With respects to t...
SP:71b7633050462a3cfc3d64e81ae3f6cec758f068
Knowledge Consistency between Neural Networks and Beyond
1 INTRODUCTION . Deep neural networks ( DNNs ) have shown promise in many tasks of artificial intelligence . However , there is still lack of mathematical tools to diagnose representations in intermediate layers of a DNN , e.g . discovering flaws in representations or identifying reliable and unreliable features . Trad...
The goal of this paper is to analyze knowledge consistency between pretrained deep neural nets. In order to do so the paper trains neural networks to predict a hidden layer of one DNN using a hidden layer of another DNN. The model is interesting in that it is multi layer but it also allows decomposing its prediction as...
SP:3312a33df68b98ddffec4706f031fc2173181d05
Knowledge Consistency between Neural Networks and Beyond
1 INTRODUCTION . Deep neural networks ( DNNs ) have shown promise in many tasks of artificial intelligence . However , there is still lack of mathematical tools to diagnose representations in intermediate layers of a DNN , e.g . discovering flaws in representations or identifying reliable and unreliable features . Trad...
This paper presents a method to disentangle intermediate features between two different deep neural networks. More specifically, given two networks, the proposed approach aims to find consistent and inconsistent feature components for a certain layer in each network. If one network is more powerful to the other (e.g., ...
SP:3312a33df68b98ddffec4706f031fc2173181d05
Generative Adversarial Nets for Multiple Text Corpora
1 Introduction . Generative adversarial nets ( GAN ) ( Goodfellow et al. , 2014 ) belong to a class of generative models which are trainable and can generate artificial data examples similar to the existing ones . In a GAN model , there are two sub-models simultaneously trained : a generative model G from which artific...
The paper proposes extensions of Generative Adversarial Networks to modeling multiple text corpora. Concretely, the paper looks at two problems: 1) given independently pretrained word embeddings from K corpora, finding a common word embdding, 2) extracting document representations from a discriminator of a GAN trained ...
SP:f44604decdd75946cc41dbdc8f25039e141276fe
Generative Adversarial Nets for Multiple Text Corpora
1 Introduction . Generative adversarial nets ( GAN ) ( Goodfellow et al. , 2014 ) belong to a class of generative models which are trainable and can generate artificial data examples similar to the existing ones . In a GAN model , there are two sub-models simultaneously trained : a generative model G from which artific...
The paper proposes to use Generative Adversarial Networks (GANs) in the context of natural language processing and introduces two models for generating document embeddings. The first model, weGAN, aggregates multiple sets of single-corpus word2vec embeddings into one set of cross-corpus word representations; document e...
SP:f44604decdd75946cc41dbdc8f25039e141276fe
Constrained Markov Decision Processes via Backward Value Functions
1 INTRODUCTION . Reinforcement Learning ( RL ) provides a sound decision-theoretic framework to optimize the behavior of learning agents in an interactive setting ( Sutton & Barto , 2018 ) . Recently , the field of RL has found success in many high-dimensional domains , like video games , Go , robot locomotion and navi...
The paper approaches the CMDP problem, in which one wishes to learn a max return policy subject to trajectory-based constraints. The paper proposes a technique based on the introduced concept of "backward value functions". These functions satisfy a sort of Bellman equation. The paper proposes a safe policy improveme...
SP:a078647f423f16068679fd5621f3600f1d96f7bf
Constrained Markov Decision Processes via Backward Value Functions
1 INTRODUCTION . Reinforcement Learning ( RL ) provides a sound decision-theoretic framework to optimize the behavior of learning agents in an interactive setting ( Sutton & Barto , 2018 ) . Recently , the field of RL has found success in many high-dimensional domains , like video games , Go , robot locomotion and navi...
This paper presents a new approach for solving Constrained MDPs. Because the cost constraint is cumulative, the best action depends on the cumulative cost so far. They address this issue by learning a backward value function of the estimated cumulative cost so far. Their theoretical results show that the same propertie...
SP:a078647f423f16068679fd5621f3600f1d96f7bf
Data Augmentation in Training CNNs: Injecting Noise to Images
1 INTRODUCTION . Convolutional Neural Networks ( CNNs ) find an ever-growing field of application throughout image and sound processing tasks , since the success of AlexNet ( Krizhevsky et al. , 2012 ) in the 2012 ImageNet competition . Yet , training these networks still keeps the need of an ” artistic ” touch : even ...
The paper studies the effect of various data augmentation methods on image classification tasks. The Authors propose the Structural Similarity (SSIM) as a measure of the magnitude of the various types of data augmentation noise they consider. The Authors argue that SSIM is superior to PSNR as a measure of the intensity...
SP:99e7452e7b7c5a1071af9370aa61acad39f99833
Data Augmentation in Training CNNs: Injecting Noise to Images
1 INTRODUCTION . Convolutional Neural Networks ( CNNs ) find an ever-growing field of application throughout image and sound processing tasks , since the success of AlexNet ( Krizhevsky et al. , 2012 ) in the 2012 ImageNet competition . Yet , training these networks still keeps the need of an ” artistic ” touch : even ...
This paper aims at analyzing the effect of injecting noise to images as data augmentation in training CNN for the image classification task. Based on the SSIM metric (which is shown to be a better metric than PSNR), different noise level on a set of different kinds of noise are explored. Experimental results on two sub...
SP:99e7452e7b7c5a1071af9370aa61acad39f99833
SCELMo: Source Code Embeddings from Language Models
1 INTRODUCTION . Learning rich representations for source code is an open problem that has the potential to enable software engineering and development tools . Some work on machine learning for source code has used hand engineered features ( Long & Rinard , 2016 , e.g . ) , but designing and implementing such features ...
The paper proposes to use ELMO embeddings to improve the precision on the first step of the DeepBugs tasks defined by Pradel and Sen (2018). This first step is an artificial problem created by taking real programs (with and without bugs, but assuming almost all of them are correct) and introducing bugs of certain type ...
SP:88d2b8efd477ec41d0bb7720d3a1ce366e1c3060
SCELMo: Source Code Embeddings from Language Models
1 INTRODUCTION . Learning rich representations for source code is an open problem that has the potential to enable software engineering and development tools . Some work on machine learning for source code has used hand engineered features ( Long & Rinard , 2016 , e.g . ) , but designing and implementing such features ...
This paper leverage recent advances of ELMo in context embedding and apply it in the source code embedding. With the help of ELMo, source embedding can take the three benefits: (1)  Surrounding names provide indirect information about possible values the variable could take; (2) an variable’s value evolves through the ...
SP:88d2b8efd477ec41d0bb7720d3a1ce366e1c3060
Starfire: Regularization-Free Adversarially-Robust Structured Sparse Training
This paper studies structured sparse training of CNNs with a gradual pruning technique that leads to fixed , sparse weight matrices after a set number of epochs . We simplify the structure of the enforced sparsity so that it reduces overhead caused by regularization . The proposed training methodology explores several ...
This paper introduces a strategy to prune a convolutional neural network during training. To speed up training, the proposed method prunes the weights with the smallest magnitude during only a small number of epochs at the beginning of training, later on continuing training with a fixed sparsity pattern. Several granul...
SP:f16157d7cd025cddd1b7f8024983737cfed9e8d4
Starfire: Regularization-Free Adversarially-Robust Structured Sparse Training
This paper studies structured sparse training of CNNs with a gradual pruning technique that leads to fixed , sparse weight matrices after a set number of epochs . We simplify the structure of the enforced sparsity so that it reduces overhead caused by regularization . The proposed training methodology explores several ...
This paper explores a series of incremental variations of existing pruning techniques for compressing Resnet-50 for ImageNet. Specifically, it proposes concentrating all pruning during an early "era" of training (the first 20-50 epochs out of 100 total). It also explores hybrids between sparse pruning and structured pr...
SP:f16157d7cd025cddd1b7f8024983737cfed9e8d4
On the Pareto Efficiency of Quantized CNN
1 INTRODUCTION . Recent success of convolutional neural networks ( CNNs ) in computer vision applications such as image classification and semantic segmentation , have fueled many important applications in energyconstrained devices , e.g. , virtual reality headsets , drones , and robots . As a result , improving the en...
This paper studies the accuracy vs model-size trade-off of quantized CNNs under different channel width multipliers. The authors demonstrated that while all-to-all convolution works well under low bit settings, depthwise conv needs a different sweet spot. The authors then proceed to use the insight to design quantized ...
SP:79a050f3b4f6466e5bee5533a7b018b2f200cb01
On the Pareto Efficiency of Quantized CNN
1 INTRODUCTION . Recent success of convolutional neural networks ( CNNs ) in computer vision applications such as image classification and semantic segmentation , have fueled many important applications in energyconstrained devices , e.g. , virtual reality headsets , drones , and robots . As a result , improving the en...
The author studies the quantization strategy of CNNs in terms of Pareto Efficiency. Through a series of experiments with three standard CNN models (ResNet, VGG11, MobileNetV2), the authors demonstrated that lower precision value can be better than high precision values in term of Pareto efficiency under the iso-model s...
SP:79a050f3b4f6466e5bee5533a7b018b2f200cb01
Local Label Propagation for Large-Scale Semi-Supervised Learning
1 INTRODUCTION . Deep neural networks ( DNNs ) have achieved impressive performance on tasks across a variety of domains , including vision ( Krizhevsky et al. , 2012 ; Simonyan & Zisserman , 2014 ; He et al. , 2016a ; 2017 ) , speech recognition ( Hinton et al. , 2012 ; Hannun et al. , 2014 ; Deng et al. , 2013 ; Noda...
The authors propose a local label propagation approach for large-scale semi-supervised learning. The approach learns a representation that tries to minimize a combination of the cross-entropy loss on the labeled data and a negative inner-product-based likelihood between the propagated pseudo-label and other examples wi...
SP:624274b6944826b6f9597298b290ae50566d6e5c
Local Label Propagation for Large-Scale Semi-Supervised Learning
1 INTRODUCTION . Deep neural networks ( DNNs ) have achieved impressive performance on tasks across a variety of domains , including vision ( Krizhevsky et al. , 2012 ; Simonyan & Zisserman , 2014 ; He et al. , 2016a ; 2017 ) , speech recognition ( Hinton et al. , 2012 ; Hannun et al. , 2014 ; Deng et al. , 2013 ; Noda...
The paper introduces an approach for semi-supervised learning based on local label propagation. The idea is to leverage the geometric structure in the embedding space, such that data near to each other in the embedding space should have the same labels. The labels of the K-nearest labeled examples are weighted to form ...
SP:624274b6944826b6f9597298b290ae50566d6e5c
Certifying Distributional Robustness using Lipschitz Regularisation
1 INTRODUCTION . Regularised risk minimisation has been the workhorse of learning nonlinear hypotheses such as deep neural networks and kernel machines . Recently , distributional robust risk ( DRR ) minimization has emerged as a promising instance with marked efficacy and flexibility . Instead of perturbing the observ...
Through the lens of Distributional Robust Risk (DRR), this work draws a link between adversarial robustness and Lipschitz constant regularisation. The authors first provide an upper bound of the DRR (with a Wasserstein ball as the ambiguity set) in terms of the true risk and the Lipschitz constant of the loss function ...
SP:dbf67fa98a71f8c3b7b62e9b5695ced62bcb730d
Certifying Distributional Robustness using Lipschitz Regularisation
1 INTRODUCTION . Regularised risk minimisation has been the workhorse of learning nonlinear hypotheses such as deep neural networks and kernel machines . Recently , distributional robust risk ( DRR ) minimization has emerged as a promising instance with marked efficacy and flexibility . Instead of perturbing the observ...
This paper uses results from distributional robustness to provide bounds of p-norm-constrained adversarial risk which depend on the Lipschitz constant of the underlying classifier. The bulk of the paper focuses on sample-efficient mechanisms to approximate the Lipschitz constants of kernel methods so that a constraint ...
SP:dbf67fa98a71f8c3b7b62e9b5695ced62bcb730d
Graph Residual Flow for Molecular Graph Generation
1 INTRODUCTION . We propose a deep generative model for molecular graphs based on invertible functions . We especially focus on introducing an invertible function that is tuned for the use in graph structured data , which allows for flexible mappings with less number of parameters than previous invertible models for gr...
The paper introduces an invertible deep generative model architecture for modeling molecular graphs. The model is based on graph residual flows (GRF), which is a graph variant of normalizing flows. The GRF model is a refinement of the GraphNVP generative model, which is also invertible, but which does not seem to work...
SP:9ad896111e20da136d179dcd72aad658eba76d93
Graph Residual Flow for Molecular Graph Generation
1 INTRODUCTION . We propose a deep generative model for molecular graphs based on invertible functions . We especially focus on introducing an invertible function that is tuned for the use in graph structured data , which allows for flexible mappings with less number of parameters than previous invertible models for gr...
GraphNVP is the first paper to introduce the concept of "invertible flow", that is to construct the invertible mapping from latent vector z to the graph G. By constructing the mapping from G to z, GraphNVP first changes the discrete feature vector into continuous variables, then update this matrix representation by sca...
SP:9ad896111e20da136d179dcd72aad658eba76d93
A Boolean Task Algebra for Reinforcement Learning
1 INTRODUCTION . Reinforcement learning ( RL ) has achieved recent success in a number of difficult , high-dimensional environments ( Mnih et al. , 2015 ; Levine et al. , 2016 ; Lillicrap et al. , 2016 ; Silver et al. , 2017 ) . However , these methods generally require millions of samples from the environment to learn...
The paper proposes a method of combining value functions for a certain class of tasks, including shortest path problems, to solve composed tasks. By expressing tasks as a Boolean algebra, they can be combined using the negation, conjunction and disjunction operations. Analogous operations are available for the optimal ...
SP:61c1ba5a02194732b56c6491b40e80d2d0846851
A Boolean Task Algebra for Reinforcement Learning
1 INTRODUCTION . Reinforcement learning ( RL ) has achieved recent success in a number of difficult , high-dimensional environments ( Mnih et al. , 2015 ; Levine et al. , 2016 ; Lillicrap et al. , 2016 ; Silver et al. , 2017 ) . However , these methods generally require millions of samples from the environment to learn...
This paper introduces a framework for composing tasks by treating tasks as a Boolean algebra. The paper assumes an undiscounted MDP with a 0-1 reward and a fixed absorbing set G, and considers a family of tasks defined by different reward functions. Each task defers only by the value of the reward function at the absor...
SP:61c1ba5a02194732b56c6491b40e80d2d0846851
Q-learning with UCB Exploration is Sample Efficient for Infinite-Horizon MDP
A fundamental question in reinforcement learning is whether model-free algorithms are sample efficient . Recently , Jin et al . ( 2018 ) proposed a Q-learning algorithm with UCB exploration policy , and proved it has nearly optimal regret bound for finite-horizon episodic MDP . In this paper , we adapt Q-learning with ...
In this paper, the authors extend the UCB Q-learning algorithm by Jin et al. (2018) to infinite horizon discounted MDPs, and prove a PAC bound of \tilde{O}(SA/\epsilon^2 (1-\gamma)^7) for the resulting algorithm. This bound improves the one for delayed Q-learning by Strehl et al. (2006) and matches the lower-bound in t...
SP:a3911fe147060f3b790ea85cfaf18034add4368c
Q-learning with UCB Exploration is Sample Efficient for Infinite-Horizon MDP
A fundamental question in reinforcement learning is whether model-free algorithms are sample efficient . Recently , Jin et al . ( 2018 ) proposed a Q-learning algorithm with UCB exploration policy , and proved it has nearly optimal regret bound for finite-horizon episodic MDP . In this paper , we adapt Q-learning with ...
This paper extends Jin et al. (2018)'s idea to infinite horizon and improves the best known sample complexity to $\tilde{O}(\frac{SA}{\epsilon^2 (1-\gamma)^7})$. The derivation is similar to Jin's paper except a very careful selection on the pseudo-horizon length $H$, where $H$ is given in finite horizon and work as th...
SP:a3911fe147060f3b790ea85cfaf18034add4368c
Enforcing Physical Constraints in Neural Neural Networks through Differentiable PDE Layer
1 INTRODUCTION . Convolutional Neural Network ( CNN ) based deep learning architectures have achieved huge success in many tasks across computer vision , but their use in the physical sciences have only recently been explored . Many parallels exist between physical science problems and those in computer vision . For in...
The paper describes a way to efficiently enforce physical constraints expressed by linear PDEs on the output of a neural network. The idea is to have, as a last layer of the network, a projection onto the constrained solution space, and to back-propagate through it. That projection layer is made efficient for high-dime...
SP:bb2885554d98533633a54e0a84ec5c08ba87db2d
Enforcing Physical Constraints in Neural Neural Networks through Differentiable PDE Layer
1 INTRODUCTION . Convolutional Neural Network ( CNN ) based deep learning architectures have achieved huge success in many tasks across computer vision , but their use in the physical sciences have only recently been explored . Many parallels exist between physical science problems and those in computer vision . For in...
This work develops a differentiable spectral projection layer to enforce spatial PDE constraints using spectral methods, to achieve the introduction of the physical constraints in the end-to-end network without damaging the intrinsic property of the network. Analysis of computational cost shows the proposed layer is ch...
SP:bb2885554d98533633a54e0a84ec5c08ba87db2d
DBA: Distributed Backdoor Attacks against Federated Learning
1 INTRODUCTION . Federated learning ( FL ) has been recently proposed to address the problems for training machine learning models without direct access to diverse training data , especially for privacy-sensitive tasks ( Smith et al. , 2017 ; McMahan et al. , 2017 ; Zhao et al. , 2018 ) . Utilizing local training data ...
This paper studies backdoor attacks under federated learning setting. To inject a certain backdoor pattern, existing work generate poisoning samples by blending the same pattern with different input samples. Even for federated learning where the adversary can control multiple parties, such as [1], all parties still use...
SP:5a1d5dd1a128cc32d3e9c71f309cb7031fcffcdb
DBA: Distributed Backdoor Attacks against Federated Learning
1 INTRODUCTION . Federated learning ( FL ) has been recently proposed to address the problems for training machine learning models without direct access to diverse training data , especially for privacy-sensitive tasks ( Smith et al. , 2017 ; McMahan et al. , 2017 ; Zhao et al. , 2018 ) . Utilizing local training data ...
The authors introduce the idea of distributed backdoor attacks in the FL framework, in which the dishonest participants in FL add local triggers to their training data to influence the global model to classify triggered images in a desired way. They show empirically that the learned models then are more likely to be su...
SP:5a1d5dd1a128cc32d3e9c71f309cb7031fcffcdb
Tensor Decompositions for Temporal Knowledge Base Completion
1 INTRODUCTION . Link prediction in relational data has been the subject of interest , given the widespread availability of such data and the breadth of its use in bioinformatics ( Zitnik et al. , 2018 ) , recommender systems ( Koren et al. , 2009 ) or Knowledge Base completion ( Nickel et al. , 2016a ) . Relational da...
In this paper, the authors study an important problem, i.e., time-aware link prediction in a knowledge base. Specifically, the authors focus on predicting the missing link in a quadruple, i.e., (subject, predicate, ?, timestamp). In particular, the authors design a new tensor (order 4) factorization based method with p...
SP:e666899b4e1cfe12cb58b2dc76e6ec923c0e5059
Tensor Decompositions for Temporal Knowledge Base Completion
1 INTRODUCTION . Link prediction in relational data has been the subject of interest , given the widespread availability of such data and the breadth of its use in bioinformatics ( Zitnik et al. , 2018 ) , recommender systems ( Koren et al. , 2009 ) or Knowledge Base completion ( Nickel et al. , 2016a ) . Relational da...
This paper extends the ComplEx model (Trouillon et al., 2016) for completing temporal knowledge bases by augmenting it with timestamp embeddings. Besides, based on the assumption that these timestamp representations evolve slowly over time, the paper introduces this prior as a regularizer. Also, the paper adds a non-te...
SP:e666899b4e1cfe12cb58b2dc76e6ec923c0e5059
On Stochastic Sign Descent Methods
1 INTRODUCTION . One of the key factors behind the success of modern machine learning models is the availability of large amounts of training data ( Bottou & Le Cun , 2003 ; Krizhevsky et al. , 2012 ; Schmidhuber , 2015 ) . However , the state-of-the-art deep learning models deployed in industry typically rely on datas...
The paper presents an improved analysis of the signSGD gradient estimator. The authors propose to relax the requirements on the gradient estimator in Bernstein (2019). The only requirement imposed on the gradient is that it should have the correct sign with probability greater than 1/2. In particular this approach allo...
SP:138632d011d3fcc86cff90f9e2fa8b1929d008cb
On Stochastic Sign Descent Methods
1 INTRODUCTION . One of the key factors behind the success of modern machine learning models is the availability of large amounts of training data ( Bottou & Le Cun , 2003 ; Krizhevsky et al. , 2012 ; Schmidhuber , 2015 ) . However , the state-of-the-art deep learning models deployed in industry typically rely on datas...
This paper focuses on signSGD with the aim of improving theoretical understanding of the method. The main contribution of the paper is to identify a condition SPB (success probability bounds), which is necessary for convergence of signSGD and study its connections with the other conditions known in the literature for s...
SP:138632d011d3fcc86cff90f9e2fa8b1929d008cb
Towards Interpretable Molecular Graph Representation Learning
1 INTRODUCTION . Following the recent rise of deep learning for image and speech processing , there has been great interest in generalizing convolutional neural networks to arbitrary graph-structured data ( Gilmer et al. , 2017 ; Henaff et al. , 2015 ; Xu et al. , 2018 ) . To this end , graph neural networks ( GNN ) fa...
The authors propose a new pooling layer, LaPool, for hierarchical graph representation learning (Ying et al., 2019) by clustering nodes around centroids that are selected based on "signal intensity variation". The signal intensity variation of node x is defined as sum_{y in HOP(x, h)} ||x - y|| where HOP(x, h) is the ...
SP:c9affd2ef30c7b4e0873aeb5783105b8ea6c056b
Towards Interpretable Molecular Graph Representation Learning
1 INTRODUCTION . Following the recent rise of deep learning for image and speech processing , there has been great interest in generalizing convolutional neural networks to arbitrary graph-structured data ( Gilmer et al. , 2017 ; Henaff et al. , 2015 ; Xu et al. , 2018 ) . To this end , graph neural networks ( GNN ) fa...
The paper introduces a new pooling approach "Laplacian pooling" for graph neural networks, which the authors claim is able to better preserve information about the local structure, and to provide interpretability. Namely, the pooling approach is based on finding centroids (nodes having high signal variation compared t...
SP:c9affd2ef30c7b4e0873aeb5783105b8ea6c056b
AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty
1 INTRODUCTION . Current machine learning models depend on the ability of training data to faithfully represent the data encountered during deployment . In practice , data distributions evolve ( Lipton et al. , 2018 ) , models encounter new scenarios ( Hendrycks & Gimpel , 2017 ) , and data curation procedures may capt...
This paper proposes a method called AugMix, which is intended to improve model robustness to data distribution shift. AugMix appears fairly simple to implement. Several new images are created by augmenting an original image through chains of sequentially applied transformations (the "Aug" part of AugMix), then the augm...
SP:3e64cbaffc0f2c9cf7bb2d7716b50795f03fe1fa
AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty
1 INTRODUCTION . Current machine learning models depend on the ability of training data to faithfully represent the data encountered during deployment . In practice , data distributions evolve ( Lipton et al. , 2018 ) , models encounter new scenarios ( Hendrycks & Gimpel , 2017 ) , and data curation procedures may capt...
The paper discusses a new data augmentation method which improves the accuracy of the network for several specific shifted domain scenarios. The main goal of the paper is to increase the robustness of the deep model trained on the augmented data to generalize well beyond the data corruption like the rotation, transla...
SP:3e64cbaffc0f2c9cf7bb2d7716b50795f03fe1fa
Self-Imitation Learning via Trajectory-Conditioned Policy for Hard-Exploration Tasks
1 INTRODUCTION . Hard-exploration tasks , particularly characterized by sparse environment rewards , are traditionally challenging in reinforcement learning ( RL ) , because the agent must carefully balance the exploration and exploitation when taking a long sequence of actions to receive infrequent non-zero rewards . ...
The paper addresses the challenge of hard exploration tasks. The approach taken is to apply self-imitation to a diverse selection of trajectories from past experience -- practice re-doing the strangest things you've ever done. This is claimed to drive more efficient exploration in sparse-reward problems, leading to SOT...
SP:9c4bfe5e2bd7e16ad54d8b37b67f9d86192f9124
Self-Imitation Learning via Trajectory-Conditioned Policy for Hard-Exploration Tasks
1 INTRODUCTION . Hard-exploration tasks , particularly characterized by sparse environment rewards , are traditionally challenging in reinforcement learning ( RL ) , because the agent must carefully balance the exploration and exploitation when taking a long sequence of actions to receive infrequent non-zero rewards . ...
The authors identify and address the problem of sub-optimal and myopic behaviors of self-imitation learning in environments with sparse rewards. The authors propose DTSIL to learn a trajectory-conditioned policy to imitate diverse trajectories from the agent’s own past experience. Unlike other self-imitation learning m...
SP:9c4bfe5e2bd7e16ad54d8b37b67f9d86192f9124
Computation Reallocation for Object Detection
1 INTRODUCTION . Object detection is one of the fundamental tasks in computer vision . The backbone feature extractor is usually taken directly from classification literature ( Girshick , 2015 ; Ren et al. , 2015 ; Lin et al. , 2017a ; Lu et al. , 2019 ) . However , comparing with classification , object detection aims...
This paper works on neural architecture search for object detection. Two search directions are proposed: 1) searching the number of conv blocks at each resolution (or "stage"). 2) searching the dilations for each conv block. A greedy neighbor-based search algorithm is adopted. The results show healthy improvements amon...
SP:6b29ca414857bbf1cb0dbf01e67520b37155f3a4
Computation Reallocation for Object Detection
1 INTRODUCTION . Object detection is one of the fundamental tasks in computer vision . The backbone feature extractor is usually taken directly from classification literature ( Girshick , 2015 ; Ren et al. , 2015 ; Lin et al. , 2017a ; Lu et al. , 2019 ) . However , comparing with classification , object detection aims...
The paper attempts to apply neural architecture search (NAS) to re-arrange, or re-allocate the network backbone blocks and the convolution filters for object detection. The search space is two-fold: 1) the network is allowed to search over allocation of different number of blocks in the backbone (e.g. ResNet, MobileNet...
SP:6b29ca414857bbf1cb0dbf01e67520b37155f3a4
Concise Multi-head Attention Models
1 INTRODUCTION . Attention based architectures , such as Transformers , have been effective for sequence modelling tasks such as machine translation ( Gehring et al. , 2017 ; Vaswani et al. , 2017 ) , question answering , sentence classification ( Radford et al. , 2018 ; Devlin et al. , 2018 ) and document generation (...
This work studies the head size <--> head number tradeoff in multihead attention. It argues and formally establishes that (1) the expressivity of an attention head is determined by its dimension and (b) fixing the head dimension, one gains additional expressive power by using more heads. In response to such observation...
SP:49ef0331201083490748c1dbcd12d130cb0a68d4
Concise Multi-head Attention Models
1 INTRODUCTION . Attention based architectures , such as Transformers , have been effective for sequence modelling tasks such as machine translation ( Gehring et al. , 2017 ; Vaswani et al. , 2017 ) , question answering , sentence classification ( Radford et al. , 2018 ; Devlin et al. , 2018 ) and document generation (...
This work discusses how to set the projection size for each head (head size) in multi-head attention module, especially Transformer. Theorem 1 is interesting, which points out a lower bound for the head size. The proposed method is to decouple the dependency between the head size and the embedding size. The experiments...
SP:49ef0331201083490748c1dbcd12d130cb0a68d4