paper_name
stringlengths
11
170
text
stringlengths
8.07k
307k
summary
stringlengths
152
6.16k
paper_id
stringlengths
43
43
Deep Auto-Deferring Policy for Combinatorial Optimization
Designing efficient algorithms for combinatorial optimization appears ubiquitously in various scientific fields . Recently , deep reinforcement learning ( DRL ) frameworks have gained considerable attention as a new approach : they can automatically train a good solver while relying less on sophisticated domain knowled...
This paper aims at solving graph-based combinatorial optimization problems using a new paradigm for Deep Reinforcement Learning. In contrast with standard Markov Decision Processes used for combinatorial DRL, the authors advocate the use of “deferred” MDPs capturing more complex actions (which can choose a subset of no...
SP:265254f820e23f0ccc344a78330bf0361b4e0499
Deep Auto-Deferring Policy for Combinatorial Optimization
Designing efficient algorithms for combinatorial optimization appears ubiquitously in various scientific fields . Recently , deep reinforcement learning ( DRL ) frameworks have gained considerable attention as a new approach : they can automatically train a good solver while relying less on sophisticated domain knowled...
The paper proposes a Deep RL approach called Auto-Deferring Policy (ADP) to learning a policy for constructing solutions for the Maximum Independent Set (MIS) problem. Rather than constructing a solution one variable per episode step, the policy can make decisions about multiple variables per step, as well as defer dec...
SP:265254f820e23f0ccc344a78330bf0361b4e0499
Continuous Control with Contexts, Provably
( √ T ) regret bound in the online setting where T is the num- ber of environments the agent played . This also implies after playing Õ ( 1/ 2 ) environments , the agent is able to transfer the learned knowledge to obtain an -suboptimal policy for an unseen environment . To our knowledge , this is first provably effic...
This paper considers the problem of changing environments for LQR. The authors model this through the use of a decoder that maps an incoming context (C,D) to the LQR matrices (A,B). They provide an algorithm for this setting based on a UCB strategy, prove sample complexity and regret bounds, and experimental results.
SP:9ef265599f0065faadcb80c57cd845c76ef70ecf
Continuous Control with Contexts, Provably
( √ T ) regret bound in the online setting where T is the num- ber of environments the agent played . This also implies after playing Õ ( 1/ 2 ) environments , the agent is able to transfer the learned knowledge to obtain an -suboptimal policy for an unseen environment . To our knowledge , this is first provably effic...
In order to generalize the RL agent to unseen environment, in this work the authors studied the theoretical learning problem of building a decoder on top of linear continuous control using linear quadratic regulator (LQR). They presented a simple, UCB-based algorithm that refines the estimates of the encoder while doin...
SP:9ef265599f0065faadcb80c57cd845c76ef70ecf
Why Convolutional Networks Learn Oriented Bandpass Filters: A Hypothesis
1 INTRODUCTION . 1.1 MOTIVATION . Convolutional networks ( ConvNets ) in conjunction with deep learning have shown state-of-the-art performance in application to computer vision , ranging across both classification , e.g. , ( Krizhevsky et al. , 2012 ; Tran et al. , 2015 ; Ge et al. , 2019 ) and regression , e.g. , ( S...
This short, interesting paper provides a theoretical analysis to explain why we may expect to see bandpass oriented filters arise as a result of the convolutional structure of deep networks. The explanation boils down to the fact that the eigenfunctions of convolutions correspond to bandpass filters. This can explain s...
SP:8d751df2c357c435f1e46f81d4e28b1661aa6aad
Why Convolutional Networks Learn Oriented Bandpass Filters: A Hypothesis
1 INTRODUCTION . 1.1 MOTIVATION . Convolutional networks ( ConvNets ) in conjunction with deep learning have shown state-of-the-art performance in application to computer vision , ranging across both classification , e.g. , ( Krizhevsky et al. , 2012 ; Tran et al. , 2015 ; Ge et al. , 2019 ) and regression , e.g. , ( S...
This paper proposed a hypothesis on why neural network learns oriented bandpass filters. While most existing work attribute this phenomenon to image structures, this paper suggests that it might be a property of convolution. In particular, it shows Fourier basis are eigenfunctions of convolution, and band pass filters ...
SP:8d751df2c357c435f1e46f81d4e28b1661aa6aad
Relation-based Generalized Zero-shot Classification with the Domain Discriminator on the shared representation
1 INTRODUCTION . The recent high performance of deep neural networks on image classification and object recognition depends greatly on whether one can obtain sufficiently labeled images of classes to predict . Nevertheless , it is difficult to do this in the real world because the number of existing classes is enormous...
This paper proposes a relation-based ZSL model which can effectively alleviate the domain bias problem. To this end, first, the paper claims that a good relation-based ZSL model should consider two requirements -- modality invariance and class separability. And the paper designed Modality-invariant and Class-separable ...
SP:14da0ec3ce7769bf3e3371097ae805909a1b636d
Relation-based Generalized Zero-shot Classification with the Domain Discriminator on the shared representation
1 INTRODUCTION . The recent high performance of deep neural networks on image classification and object recognition depends greatly on whether one can obtain sufficiently labeled images of classes to predict . Nevertheless , it is difficult to do this in the real world because the number of existing classes is enormous...
The main topic of this paper is generalized zero-shot learning. This paper modifies traditional VAE method with attribute matching prior to release the hidden features from original regularization. This paper also proposes a domain discriminator to enhance class-separability of learned features to avoid unseen classes ...
SP:14da0ec3ce7769bf3e3371097ae805909a1b636d
Low-dimensional statistical manifold embedding of directed graphs
We propose a novel node embedding of directed graphs to statistical manifolds , which is based on a global minimization of pairwise relative entropy and graph geodesics in a non-linear way . Each node is encoded with a probability density function over a measurable space . Furthermore , we analyze the connection betwee...
This paper proposed another graph embedding method. It focuses on directed graphs, and it embedded the graph nodes into exponential power distributions, which include the Gaussian distribution as a special case. The method is implemented by optimizing with respect to the free distributions on a statistical manifold so ...
SP:7419b8bf021dbef66e8483d178ef9838bc790a48
Low-dimensional statistical manifold embedding of directed graphs
We propose a novel node embedding of directed graphs to statistical manifolds , which is based on a global minimization of pairwise relative entropy and graph geodesics in a non-linear way . Each node is encoded with a probability density function over a measurable space . Furthermore , we analyze the connection betwee...
This paper proposes an unsupervised method for learning node embeddings of directed graphs into statistical manifolds. Each node in the graph is mapped to a distribution in the space of k-variate power distributions, endowed with the KL divergence as asymetric similarity. The authors propose an optimization method base...
SP:7419b8bf021dbef66e8483d178ef9838bc790a48
Neural Subgraph Isomorphism Counting
1 INTRODUCTION . Graphs are general data structures widely used in many applications , including social network analysis , molecular structure analysis , natural language processing and knowledge graph modeling , etc . Learning with graphs has recently drawn much attention as neural network approaches to representation...
This paper studied how to leverage the power of graph neural networks for counting subgraph isomorphism. The motivation is that the current subgraph isomorphism detection is NP-complete problem and a proposed approach based on GNN could approximately solve the counting problem in polynomial time. Then they relaxed orig...
SP:4ae6b28a86ddef3fcb212ba132c2058109fbc938
Neural Subgraph Isomorphism Counting
1 INTRODUCTION . Graphs are general data structures widely used in many applications , including social network analysis , molecular structure analysis , natural language processing and knowledge graph modeling , etc . Learning with graphs has recently drawn much attention as neural network approaches to representation...
This paper proposes a method called Dynamic Intermedium Attention Memory Network (DIAMNet) to learn the subgraph isomorphism counting for a given pattern graph P and target graph G. This requires global information unlike usual GNN cases such as node classification, link prediction, community detection. First, input gr...
SP:4ae6b28a86ddef3fcb212ba132c2058109fbc938
Progressive Memory Banks for Incremental Domain Adaptation
1 INTRODUCTION . Domain adaptation aims to transfer knowledge from one domain ( called the source domain ) to another ( called the target domain ) in a machine learning system.2 If the data of the target domain are not large enough , using data from the source domain typically helps to improve model performance in the ...
This paper proposes an extensible attention mechanism applied on the previous hidden state of an RNN and resulting in supplementary input for the next RNN step. For each added domain, new pairs of attentions key and values can be added to provide more capacity for the model. This method is applied in the context of inc...
SP:fc2d2ef400ef225020f6aa37cba475fbbbf44fba
Progressive Memory Banks for Incremental Domain Adaptation
1 INTRODUCTION . Domain adaptation aims to transfer knowledge from one domain ( called the source domain ) to another ( called the target domain ) in a machine learning system.2 If the data of the target domain are not large enough , using data from the source domain typically helps to improve model performance in the ...
This work proposes to use an augmented RNN model to address the incremental domain adaptation problem. In particular, it designs the progressive memory bank approach which expands the memory capacity by adding parameters every time a new task comes in. The RNN retrieves knowledge from the memory bank via key-value atte...
SP:fc2d2ef400ef225020f6aa37cba475fbbbf44fba
BERT-AL: BERT for Arbitrarily Long Document Understanding
1 INTRODUCTION . In recent years , neural networks are proposed to solve various NLP tasks . Especially , pretrained language models ( Peters et al. , 2018 ; Radford et al. , 2018 ; Devlin et al. , 2019 ; Yang et al. , 2019 ) attract lots of attentions , which take advantage of the two-stages training process : pretrai...
The paper proposes a methodology to overcome the problem of processing long sequences with a pre-trained Transformer model, which suffers from high computational costs due to the complexity being quadratic in the length of the sequence. The authors also point out that BERT needs to be retrained from scratch if sequence...
SP:fa6f35939da0aa66e90846ba9077b67c91f33eb4
BERT-AL: BERT for Arbitrarily Long Document Understanding
1 INTRODUCTION . In recent years , neural networks are proposed to solve various NLP tasks . Especially , pretrained language models ( Peters et al. , 2018 ; Radford et al. , 2018 ; Devlin et al. , 2019 ; Yang et al. , 2019 ) attract lots of attentions , which take advantage of the two-stages training process : pretrai...
This paper proposed another variant of BERT, called BERT-AL, which can deal with arbitrarily long inputs. The authors constructed the proposed method by combining the segment-wise BERT with the multi-channel LSTM. The authors validated the proposed method on the text summarization task and achieved higher performance t...
SP:fa6f35939da0aa66e90846ba9077b67c91f33eb4
Neural Markov Logic Networks
We introduce Neural Markov Logic Networks ( NMLNs ) , a statistical relational learning system that borrows ideas from Markov logic . Like Markov Logic Networks ( MLNs ) , NMLNs are an exponential-family model for modelling distributions over possible worlds , but unlike MLNs , they do not rely on explicitly specified ...
This paper presents Neural Markov Logic Networks (NMLN), which is a generalization of Markov Logic Networks (MLN). Unlike MLN which relies on pre-specified first-order logic (FOL) rules, NMLN learns potential functions parameterized by neural networks on fragments of the graph. The potential function can possibly take...
SP:cf446212c3be85660aea33c9604c240473a42105
Neural Markov Logic Networks
We introduce Neural Markov Logic Networks ( NMLNs ) , a statistical relational learning system that borrows ideas from Markov logic . Like Markov Logic Networks ( MLNs ) , NMLNs are an exponential-family model for modelling distributions over possible worlds , but unlike MLNs , they do not rely on explicitly specified ...
The following paper provides an extension to Markov Logic Networks(MLNs), by removing their dependency on pre-defined first-order logic rules. This is handled via neural networks which are able to capture the statistical relations, so-called Neural Markov Logic Networks(NMLNs). As this is an implicit representation fr...
SP:cf446212c3be85660aea33c9604c240473a42105
Meta-Learning with Warped Gradient Descent
1 INTRODUCTION . Learning ( how ) to learn implies inferring a learning strategy from some set of past experiences via a meta-learner that a task-learner can leverage when learning a new task . One approach is to directly parameterise an update rule via the memory of a recurrent neural network ( Andrychowicz et al. , 2...
The current paper deals with meta-learning and essentially proposes a generalization of MAML (a popular gradient-based meta-learning algorithm) that mostly builds upon two main recent advances in meta-learning: 1) an architectural one (see e.g. T-Nets), which consists in optimizing the parameters of additional layers d...
SP:66a3dbe015b07649121b664fbb28c7e198b0282d
Meta-Learning with Warped Gradient Descent
1 INTRODUCTION . Learning ( how ) to learn implies inferring a learning strategy from some set of past experiences via a meta-learner that a task-learner can leverage when learning a new task . One approach is to directly parameterise an update rule via the memory of a recurrent neural network ( Andrychowicz et al. , 2...
The authors propose warped gradient descent (WarpGrad) an optimisation framework for facilitating gradient-based meta-learning. WarpGrad interleaves within the learner meta-learned warp-layers that implicitly precondition the gradients of the task-specific parameters during backpropagation. In contrast to the linear pr...
SP:66a3dbe015b07649121b664fbb28c7e198b0282d
Policy path programming
1 INTRODUCTION . Reinforcement learning algorithms can leverage internal models of environment dynamics to facilitate the development of good control policies ( Sutton & Barto , 2018 ) . Dynamic programming methods iteratively implement one-step , full-width backups in order to propagate reward information across a sta...
This work proposes a policy iteration algorithm that implements full-depth, full-width backups in contrast to one-step, full-width methods. The authors go over existing algorithms and talks a bit how their proposal conceptually differs in how it performs said backups. They provide a bit of intuition to help explain the...
SP:65cf439b128b8b7e2596b065df7c9294b408ed19
Policy path programming
1 INTRODUCTION . Reinforcement learning algorithms can leverage internal models of environment dynamics to facilitate the development of good control policies ( Sutton & Barto , 2018 ) . Dynamic programming methods iteratively implement one-step , full-width backups in order to propagate reward information across a sta...
The paper considers the problem of finding the optimal policy in the Markovian decision Processes, where a KL policy regularizer is added to the objective function. Instead of the closed form solution which leads to the KL-regularized Bellman equation the paper proposes to use an incremental gradient ascent algorithm....
SP:65cf439b128b8b7e2596b065df7c9294b408ed19
Fully Polynomial-Time Randomized Approximation Schemes for Global Optimization of High-Dimensional Folded Concave Penalized Generalized Linear Models
1 INTRODUCTION . This paper concerns global optimization of a folded concave penalized formulation of highdimensional learning generalized linear models , which belongs to statistical/machine learning problems such that the number of dimensions ( or number of fitting parameters ) p is ( much ) larger than the number of...
The paper studies the problem of global optimization of high-dimensional sparse estimators regularized by PCP (folded concave penalty). The main result is showing that under certain conditions, with high probability, the desired global solution is an oracle stationary point satisfying the so-called S^3ONC conditions. I...
SP:4ee46486b69550aea4f1a3a6beb0caf6ea8c635e
Fully Polynomial-Time Randomized Approximation Schemes for Global Optimization of High-Dimensional Folded Concave Penalized Generalized Linear Models
1 INTRODUCTION . This paper concerns global optimization of a folded concave penalized formulation of highdimensional learning generalized linear models , which belongs to statistical/machine learning problems such that the number of dimensions ( or number of fitting parameters ) p is ( much ) larger than the number of...
The authors study the high-dimensional sparse estimation problems, which is one of the fundamental topics in both machine learning and optimization communities. In the literature, the folded concave penalty (FCP) methods have been shown to enjoy the strong oracle property for high-dimensional sparse estimation. While ...
SP:4ee46486b69550aea4f1a3a6beb0caf6ea8c635e
AtomNAS: Fine-Grained End-to-End Neural Architecture Search
1 INTRODUCTION . Human-designed neural networks are already surpassed by machine-designed ones . Neural Architecture Search ( NAS ) has become the mainstream approach to discover efficient and powerful network structures ( Zoph & Le ( 2017 ) ; Pham et al . ( 2018 ) ; Tan et al . ( 2019 ) ; Liu et al . ( 2019a ) ) . Alt...
[Summary] This paper proposes a channel-wise neural architecture search (NAS) approach. The NAS search algorithm is similar to previous one-shot NAS, but the search space is channel-wise: each channel has it’s own kernel size, which is quite novel and interesting. Results are strong in terms of FLOPS and parameters.
SP:ebf224ba53e628c06a426f49cf8f1761f3da7fc8
AtomNAS: Fine-Grained End-to-End Neural Architecture Search
1 INTRODUCTION . Human-designed neural networks are already surpassed by machine-designed ones . Neural Architecture Search ( NAS ) has become the mainstream approach to discover efficient and powerful network structures ( Zoph & Le ( 2017 ) ; Pham et al . ( 2018 ) ; Tan et al . ( 2019 ) ; Liu et al . ( 2019a ) ) . Alt...
This basic idea of this paper is to decompose the common building blocks of large network into atomic blocks, which equips NAS with more fine-grained search space. What's more, the authors propose a resource-aware search to reduce the computation and dynamically shrinkage the model to accelerate the learning. Retrainin...
SP:ebf224ba53e628c06a426f49cf8f1761f3da7fc8
Sparse Weight Activation Training
Training convolutional neural networks ( CNNs ) is time consuming . Prior work has explored how to reduce the computational demands of training by eliminating gradients with relatively small magnitude . We show that eliminating small magnitude components has limited impact on the direction of high-dimensional vectors ....
This paper studies training neural networks with sparse weights and sparse activations (SWAT training). By using sparse weights in forward passes as well as sparse weights and activations in backward passes, SWAT can reduce the computation overhead and also reduce the training memory footprint. The primary contribution...
SP:2a6531194b14410a59ae43cda18bc29bd633a0ec
Sparse Weight Activation Training
Training convolutional neural networks ( CNNs ) is time consuming . Prior work has explored how to reduce the computational demands of training by eliminating gradients with relatively small magnitude . We show that eliminating small magnitude components has limited impact on the direction of high-dimensional vectors ....
This paper proposes SWAT as a training algorithm for sparse networks on different architectures. The paper claims being able to reach a level of sparsity with no drop in accuracy. The goal is to minimize the computations during training time. To this end, SWAT sets to zero the vectors where necessary. Different from o...
SP:2a6531194b14410a59ae43cda18bc29bd633a0ec
Deep Learning of Determinantal Point Processes via Proper Spectral Sub-gradient
1 INTRODUCTION . Diversity is desired in multiple machine learning and computer vision tasks ( e.g. , image hashing ( Chen et al. , 2017 ; Carreira-Perpinán & Raziperchikolaei , 2016 ) , descriptor learning ( Zhang et al. , 2017 ) , metric learning ( Mishchuk et al. , 2017 ) and video summarization ( Sharghi et al. , ...
Determinantal Point Processes (DPPs) are statistical models that allow efficient sampling of diverse solutions - a problem that is hard with most machine learning modeling frameworks. However, learning diverse features via DPPs with deep learning frameworks is challenging due the instability in computing the gradient ...
SP:bf0cf7080618b0578eab60961d20e7a99a50c080
Deep Learning of Determinantal Point Processes via Proper Spectral Sub-gradient
1 INTRODUCTION . Diversity is desired in multiple machine learning and computer vision tasks ( e.g. , image hashing ( Chen et al. , 2017 ; Carreira-Perpinán & Raziperchikolaei , 2016 ) , descriptor learning ( Zhang et al. , 2017 ) , metric learning ( Mishchuk et al. , 2017 ) and video summarization ( Sharghi et al. , ...
The authors present an approach to optimize determinantal point processes directly (by gradient descent, instead of . via approximations), so that diversity could be modeled in objective functions for deep learning systems. The approach taken is to express the DPP term as an L-ensemble in the spectral domain over the g...
SP:bf0cf7080618b0578eab60961d20e7a99a50c080
Effective and Robust Detection of Adversarial Examples via Benford-Fourier Coefficients
1 INTRODUCTION . Deep neural networks ( DNNs ) have achieved a remarkable success in many important applications , such as image classification , face recognition , object detection , etc . In the meanwhile , DNNs have been shown to be very vulnerable to adversarial examples . However , many real-world scenarios have v...
This paper proposes an approach to adversarial detection. The approach first computes a representation of the activation layers using the Benford-Fourier coefficients. One then generates a range of noisy instances, and trains an SVM using those noisy instances as supervised labels (e.g., noisy instances are adversari...
SP:5a0fa18c836d63dae05055f3197e479099d7abb5
Effective and Robust Detection of Adversarial Examples via Benford-Fourier Coefficients
1 INTRODUCTION . Deep neural networks ( DNNs ) have achieved a remarkable success in many important applications , such as image classification , face recognition , object detection , etc . In the meanwhile , DNNs have been shown to be very vulnerable to adversarial examples . However , many real-world scenarios have v...
This paper presents a new discriminator metric for adversarial attack's detection by deriving the different properties of l-th neuron network layer on different adv/benign samples. This method can achieve good AUC score comparing to other start-of-art detection methods and also achieve good robustness under correspondi...
SP:5a0fa18c836d63dae05055f3197e479099d7abb5
Identifying through Flows for Recovering Latent Representations
Identifiability , or recovery of the true latent representations from which the observed data originates , is de facto a fundamental goal of representation learning . Yet , most deep generative models do not address the question of identifiability , and thus fail to deliver on the promise of the recovery of the true la...
This paper is about learning an identifiable generative model, iFlow, that builds upon a recent result on nonlinear ICA. The key idea is providing side information to identify the latent representation, i.e., essentially a prior conditioned on extra information such as labels and restricting the mapping to flows for be...
SP:7ab75453e99ff53d62b1fdeeb02602cc8ecf94a6
Identifying through Flows for Recovering Latent Representations
Identifiability , or recovery of the true latent representations from which the observed data originates , is de facto a fundamental goal of representation learning . Yet , most deep generative models do not address the question of identifiability , and thus fail to deliver on the promise of the recovery of the true la...
The paper tackles the identifiability problem in generative modeling, i.e., recovering the true latent representations from which the observed data originates. The paper argues that identifiable variational autoencoder (iVAE) suffers from intractability issue which leads to suboptimal solutions. The paper instead propo...
SP:7ab75453e99ff53d62b1fdeeb02602cc8ecf94a6
Individualised Dose-Response Estimation using Generative Adversarial Nets
1 INTRODUCTION . Most of the methods developed in the causal inference literature focus on learning the effects of binary or categorical treatments ( Bertsimas et al. , 2017 ; Alaa et al. , 2017 ; Alaa & van der Schaar , 2017 ; Athey & Imbens , 2016 ; Wager & Athey , 2018 ; Yoon et al. , 2018 ) . These treatments , tho...
The paper introduces Dose Response Generative Adversarial Network (DRGAN) that is aimed at generating entire dose-response curve from observational data with single dose treatments. This work is an extension of GANITE (Yoon et al., 2018) for the case of real-valued treatments (i.e., dosage). The proposed model consists...
SP:5630dcc454f4e040626c29557dc5c67cbb289dda
Individualised Dose-Response Estimation using Generative Adversarial Nets
1 INTRODUCTION . Most of the methods developed in the causal inference literature focus on learning the effects of binary or categorical treatments ( Bertsimas et al. , 2017 ; Alaa et al. , 2017 ; Alaa & van der Schaar , 2017 ; Athey & Imbens , 2016 ; Wager & Athey , 2018 ; Yoon et al. , 2018 ) . These treatments , tho...
The paper proposes a generative adversarial net model for heterogeneous dose-response causal effect estimation. The idea is to generate counterfactual dose-response curves using the generator that can fool the discriminator that tries to distinguish between factual data and counterfactual data. Factual data along with ...
SP:5630dcc454f4e040626c29557dc5c67cbb289dda
Unbiased Contrastive Divergence Algorithm for Training Energy-Based Latent Variable Models
1 INTRODUCTION . Energy-based latent variable models cover a broad class of generative models that are frequently used to characterize sophisticated distributions of high-dimensional data . Popular examples of this class include the restricted Boltzmann machines ( RBM , Smolensky , 1986 ; Hinton , 2012 ) , deep belief ...
Based on recent progress in unbiased MCMC sampling the paper proposes an unbiased contrastive divergence (UCD) algorithm for training energy based models. Specifically they developed an unbiased version of the gibbs sampling contrastive divergence algorithm for training restricted Boltzman machines. The authors demonst...
SP:71511ac32bc7e86a19296e8b049f38bd43e31d29
Unbiased Contrastive Divergence Algorithm for Training Energy-Based Latent Variable Models
1 INTRODUCTION . Energy-based latent variable models cover a broad class of generative models that are frequently used to characterize sophisticated distributions of high-dimensional data . Popular examples of this class include the restricted Boltzmann machines ( RBM , Smolensky , 1986 ; Hinton , 2012 ) , deep belief ...
The paper proposes an algorithmic improvement that significantly simplifies training of energy-based models, such as the Restricted Boltzmann Machine. The key issue in training such models is computing the gradient of the log partition function, which can be framed as computing the expected value of f(x) = dE(x; theta)...
SP:71511ac32bc7e86a19296e8b049f38bd43e31d29
Lean Images for Geo-Localization
1 INTRODUCTION . Imagine you are brought blindfolded to a street corner of a city you know well . Now , you remove the blindfold . Can you tell where you are ? This is the geo-localization task . In computer vision , this amounts to estimating the position ( and sometimes the orientation ) of a camera given its current...
This paper evaluates the performances of Deep Learning for geo-localization tasks in a world of images without textures ("lean images"). More exactly, the lean images are images rendered from a 3D model of a city. made of the depth and/or the buildings' edges and/or the buildings' faces. For the purpose of the evaluati...
SP:b8b826f478d505b38f97ea351b08290bbcd0029d
Lean Images for Geo-Localization
1 INTRODUCTION . Imagine you are brought blindfolded to a street corner of a city you know well . Now , you remove the blindfold . Can you tell where you are ? This is the geo-localization task . In computer vision , this amounts to estimating the position ( and sometimes the orientation ) of a camera given its current...
The paper evaluates a geo-localization task based on "lean" images only, obtained by projection of 3d models without texture information. Multiple levels of granularity of the lean images (edges/edges+faces/edges+faces+depth) are compared for the learning, both in a "memorization" setting and in a "generalization to un...
SP:b8b826f478d505b38f97ea351b08290bbcd0029d
Ranking Policy Gradient
1 INTRODUCTION . One of the major challenges in reinforcement learning ( RL ) is the high sample complexity ( Kakade et al. , 2003 ) , which is the number of samples must be collected to conduct successful learning . There are different reasons leading to poor sample efficiency of RL ( Yu , 2018 ) . Because policy grad...
This paper presents a new view on policy gradient methods from the perspective of ranking. The end goal in policy learning is to achieve the right ranking of actions at a state (in the case when deterministic policies are optimal), and the paper proposes a method of doing this inspired from the work on learning to rank...
SP:1d0c0485f95d6cf21922e0ac949333e58cc87bdf
Ranking Policy Gradient
1 INTRODUCTION . One of the major challenges in reinforcement learning ( RL ) is the high sample complexity ( Kakade et al. , 2003 ) , which is the number of samples must be collected to conduct successful learning . There are different reasons leading to poor sample efficiency of RL ( Yu , 2018 ) . Because policy grad...
This work first establishes the connection of maximizing the lower bound of accumulated reward and supervised learning on near-optimal policies. Then it proposes a general framework for policy learning: during the exploration stage, the agent will collect near-optimal trajectories while in the exploitation stage, the a...
SP:1d0c0485f95d6cf21922e0ac949333e58cc87bdf
Modeling question asking using neural program generation
1 INTRODUCTION . People can ask rich , creative questions to learn efficiently about their environment . Question asking is central to human learning yet it is a tremendous challenge for computational models . There is always an infinite set of possible questions that one can ask , leading to challenges both in represe...
The paper uses a deep neural network architecture (CNN + Transformer) to model logical translations of questions in the form of programs. The experimental setup uses the "battleship" game scenario, which is an interesting domain for questions because of the inherent partial observability present in the game. The paper ...
SP:d9608f07e3a5f6d8775080df89a261e054c1a2f8
Modeling question asking using neural program generation
1 INTRODUCTION . People can ask rich , creative questions to learn efficiently about their environment . Question asking is central to human learning yet it is a tremendous challenge for computational models . There is always an infinite set of possible questions that one can ask , leading to challenges both in represe...
The authors explore different ways to generate questions about the current state of a “Battleship” game. To do this, they introduce a neural network architecture with a convolutional encoder and a Transformer-based decoder and consider both supervised and reinforcement-learned training approaches. They evaluate the int...
SP:d9608f07e3a5f6d8775080df89a261e054c1a2f8
Meta-Graph: Few shot Link Prediction via Meta Learning
1 INTRODUCTION . Given a graph representing known relationships between a set of nodes , the goal of link prediction is to learn from the graph and infer novel or previously unknown relationships ( Liben-Nowell & Kleinberg , 2003 ) . For instance , in a social network we may use link prediction to power a friendship re...
Overview: In this paper, a meta-learning approach is proposed to perform link prediction across multi-graphs with scarce data. To do so, each graph is treated as a link prediction "task". Different from the tasks in conventional meta-learning, the graphs here are generally non i.i.d. Based on the variational graph auto...
SP:ab5deaa838db01236f0d3db05b39e2f10e39184b
Meta-Graph: Few shot Link Prediction via Meta Learning
1 INTRODUCTION . Given a graph representing known relationships between a set of nodes , the goal of link prediction is to learn from the graph and infer novel or previously unknown relationships ( Liben-Nowell & Kleinberg , 2003 ) . For instance , in a social network we may use link prediction to power a friendship re...
This paper proposes to provide a novel gradient-based meta-learning framework (Meta-Graph) for a few shot link prediction task. More specifically, they generate an effective parameter initialization for a local link prediction model for any unseen graph by leveraging higher-order gradients and introducing graph signatu...
SP:ab5deaa838db01236f0d3db05b39e2f10e39184b
Deep geometric matrix completion: Are we doing it right?
1 INTRODUCTION . Matrix completion deals with the recovery of missing values of a matrix from a subset of its entries , Find X s.t . X S = M S. ( 1 ) HereX stands for the unknown matrix , M ∈ Rm×n for the ground truth matrix , S is a binary mask representing the input support , and denotes the Hadamard product . Since ...
This paper proposes a novel approach for the loss function of matrix completion when geometric information is available. The proposed method consists of two ideas: (1) spectral regularization (i.e., Dirichlet energy) with a re-parameterizing basis and (2) multiresolution of spectral loss (i.e., zoomout loss). In additi...
SP:41e70121223317279d2026514d8e786cf22a4aaf
Deep geometric matrix completion: Are we doing it right?
1 INTRODUCTION . Matrix completion deals with the recovery of missing values of a matrix from a subset of its entries , Find X s.t . X S = M S. ( 1 ) HereX stands for the unknown matrix , M ∈ Rm×n for the ground truth matrix , S is a binary mask representing the input support , and denotes the Hadamard product . Since ...
This paper aims to solve the matrix completion problem by incorporating geometric information. The proposed approach involves using graphs encoding relations between rows (and columns), applying spectral decomposition to these graphs, and using a multi-resolution spectral geometric loss to reconstruct the functional ma...
SP:41e70121223317279d2026514d8e786cf22a4aaf
Flexible and Efficient Long-Range Planning Through Curious Exploration
1 INTRODUCTION . Many complex behaviors such as cleaning a kitchen , organizing a drawer , or cooking a meal require plans that are a combination of low-level geometric manipulation and high-level action sequencing . For example , boiling water requires sequencing high-level actions such as fetching a pot , pouring wat...
The paper introduces Curious Sample Planner (CSP) a long-horizon motion planning method that combines task and motion planning with deep reinforcement learning in order to solve simulated robotic tasks with sparse rewards. The CSP algorithm considers two different hierarchies of actions: primitive actions, which contro...
SP:5db89a98e14e6cb553431c750c85944f0b766b6b
Flexible and Efficient Long-Range Planning Through Curious Exploration
1 INTRODUCTION . Many complex behaviors such as cleaning a kitchen , organizing a drawer , or cooking a meal require plans that are a combination of low-level geometric manipulation and high-level action sequencing . For example , boiling water requires sequencing high-level actions such as fetching a pot , pouring wat...
This paper tackles the problem of enabling robots to learn long-horizon, sparse-reward tasks. The proposed approach, the Curious Sample Planner (CSP), builds on insights in task and motion planning (TAMP), which is a standard approach for tackling these kinds of tasks. TAMP constructs a plan in the space of macro-actio...
SP:5db89a98e14e6cb553431c750c85944f0b766b6b
Iterative Target Augmentation for Effective Conditional Generation
1 INTRODUCTION . Deep architectures are becoming increasingly adept at generating complex objects such as images , text , molecules , or programs . Many useful generation problems can be seen as translation tasks , where the goal is to take a source ( precursor ) object such as a molecule and turn it into a target sati...
This paper proposes a training scheme to enhance the optimization process where the outputs are required to meet certain constraints. The authors propose to insert an additional target augmentation phase after the regular training. For each datapoint, the algorithm samples candidate outputs until it find a valid output...
SP:0a5525d1cbaa4856f6556e1426a1236b73cfc8d3
Iterative Target Augmentation for Effective Conditional Generation
1 INTRODUCTION . Deep architectures are becoming increasingly adept at generating complex objects such as images , text , molecules , or programs . Many useful generation problems can be seen as translation tasks , where the goal is to take a source ( precursor ) object such as a molecule and turn it into a target sati...
This paper proposes a data augmentation strategy for a class of problems that the amount of labelled data is limited while the evaluation procedure is easier. Specifically, they are able to incorporate some of the model’s output into training data to guide the training procedure. The idea is quite simple and effective ...
SP:0a5525d1cbaa4856f6556e1426a1236b73cfc8d3
FasterSeg: Searching for Faster Real-time Semantic Segmentation
1 INTRODUCTION . Semantic segmentation predicts pixel-level annotations of different semantic categories for an image . Despite its performance breakthrough thanks to the prosperity of convolutional neural networks ( CNNs ) ( Long et al. , 2015 ) , as a dense structured prediction task , segmentation models commonly su...
This paper presents an automatically designed semantic segmentation network utilising neural architecture search. The proposed method is discovered from a search space integrating multi-resolution branches, that has been recently found to be vital in manually designed segmentation models. To calibrate the balance betwe...
SP:4c52d1c81fa793c07fb6ea1dc85b107acb709254
FasterSeg: Searching for Faster Real-time Semantic Segmentation
1 INTRODUCTION . Semantic segmentation predicts pixel-level annotations of different semantic categories for an image . Despite its performance breakthrough thanks to the prosperity of convolutional neural networks ( CNNs ) ( Long et al. , 2015 ) , as a dense structured prediction task , segmentation models commonly su...
This paper proposes a neural architecture search (NAS) algorithm which automatically finds a efficient network architecture, FasterSeg, for real time semantic segmentation. In designing a NAS algorithm the author takes cue from recent architectural advances introduced for faster segmentation as well as improved accurac...
SP:4c52d1c81fa793c07fb6ea1dc85b107acb709254
Balancing Cost and Benefit with Tied-Multi Transformers
1 INTRODUCTION . Neural networks for sequence-to-sequence modeling typically consist of an encoder and a decoder coupled via an attention mechanism . Whereas the very first deep models used stacked recurrent neural networks ( RNN ) ( Sutskever et al. , 2014 ; Cho et al. , 2014 ; Bahdanau et al. , 2015 ) in the encoder ...
In this paper, the authors propose a new variant of Transformer called Tied-multi Transformer. Given such a model with an N-layer encoder and an M-layer decoder, it is trained with M*N loss functions, where each combination of the nth-layer of the encoder and the mth-layer of the decoder is used to train an NMT model. ...
SP:4a2116053dafd3977fbf2e9d2879fce7c3965448
Balancing Cost and Benefit with Tied-Multi Transformers
1 INTRODUCTION . Neural networks for sequence-to-sequence modeling typically consist of an encoder and a decoder coupled via an attention mechanism . Whereas the very first deep models used stacked recurrent neural networks ( RNN ) ( Sutskever et al. , 2014 ; Cho et al. , 2014 ; Bahdanau et al. , 2015 ) in the encoder ...
This work proposes a way to reduce the latencies incurred in inference for neural machine translation. Basic idea is to train a model with softmax attached to each output of decoder layers, and computes a loss by aggregating the cross entropy losses over the softmaxes. During inference, it could either use one of the s...
SP:4a2116053dafd3977fbf2e9d2879fce7c3965448
Skew-Fit: State-Covering Self-Supervised Reinforcement Learning
1 INTRODUCTION Reinforcement learning ( RL ) provides an appealing formalism for automated learning of behavioral skills , but separately learning every potentially useful skill becomes prohibitively time consuming , both in terms of the experience required for the agent and the effort required for the user to design r...
The paper proposes an exploratory objective that can maximize state coverage in RL. They show that a formal objective for maximizing state coverage is equivalent to maximizing the entropy of a goal distribution. The core idea is to propose a method to maximize entropy of a goal distribution, or a state distribution sin...
SP:fe8c2ff8b3d5980861f2b5f81057d361903d985b
Skew-Fit: State-Covering Self-Supervised Reinforcement Learning
1 INTRODUCTION Reinforcement learning ( RL ) provides an appealing formalism for automated learning of behavioral skills , but separately learning every potentially useful skill becomes prohibitively time consuming , both in terms of the experience required for the agent and the effort required for the user to design r...
This paper introduced a very interesting idea to facilitate exploration in goal-conditioned reinforcement learning. The key idea is to learn a generative model of goal distribution to match the weighted empirical distribution, where the rare states receive larger weights. This encourages the model to generate more dive...
SP:fe8c2ff8b3d5980861f2b5f81057d361903d985b
Learn to Explain Efficiently via Neural Logic Inductive Learning
1 INTRODUCTION The recent years have witnessed the growing success of deep learning models in a wide range of applications . However , these models are also criticized for the lack of interpretability in its behavior and decision making process ( Lipton , 2016 ; Mittelstadt et al. , 2019 ) , and for being data-hungry ....
This paper presents a model for effectively hierarchically ‘searching’ through the space of (continuously relaxed) FOL formulas that explain the underlying dataset. The model presented employs a three-level architecture to produce logic entailment formulas, skolemized with a set of skolem functions, i.e. ‘operators’. T...
SP:d909573bd57a8f14009fd399d88cd2eeddd0aab5
Learn to Explain Efficiently via Neural Logic Inductive Learning
1 INTRODUCTION The recent years have witnessed the growing success of deep learning models in a wide range of applications . However , these models are also criticized for the lack of interpretability in its behavior and decision making process ( Lipton , 2016 ; Mittelstadt et al. , 2019 ) , and for being data-hungry ....
This paper proposes a novel architecture of integrating neural models with logic inference capabilities to achieve the goal of scalable predictions with explanatory decisions, which are of significant importance in the real deployment. In general, the article is well-written and nicely-structured with clear motivations...
SP:d909573bd57a8f14009fd399d88cd2eeddd0aab5
Weight-space symmetry in neural network loss landscapes revisited
Neural network training depends on the structure of the underlying loss landscape , i.e . local minima , saddle points , flat plateaus , and loss barriers . In relation to the structure of the landscape , we study the permutation symmetry of neurons in each layer of a deep neural network , which gives rise not only to ...
The paper presented a method for studying the landscape of the loss function w.r.t. parameters in a neural network from the perspective of weight-space symmetry. The detailed method includes constructing/optimising a low-loss path in between two parameter vectors (incoming connections from the previous layer) of two ne...
SP:d30f90a1fcfc6cb120ed92d1f36b3e4312948201
Weight-space symmetry in neural network loss landscapes revisited
Neural network training depends on the structure of the underlying loss landscape , i.e . local minima , saddle points , flat plateaus , and loss barriers . In relation to the structure of the landscape , we study the permutation symmetry of neurons in each layer of a deep neural network , which gives rise not only to ...
This paper studies the permutation symmetry of deep neural networks. It was known that by reordering neurons and their connections in each layer, the input -> output map the neural network represents can be preserved. This corresponded to a set of unconnected equivalent points in the weight space. The authors study the...
SP:d30f90a1fcfc6cb120ed92d1f36b3e4312948201
INTERNAL-CONSISTENCY CONSTRAINTS FOR EMERGENT COMMUNICATION
1 INTRODUCTION Emergent communication is the study of how linguistic protocols evolve when agents are tasked to cooperate . For example , agents engaged in a simple object retrieval task learn to communicate with one another in order to get the items they want ( Lazaridou et al. , 2018 ) . To date , work of this type h...
The paper analyzes if enforcing internal-consistency for speaker-listener setup can (i) improve the ability of the agents to refer to unseen referents (ii) generalize for different communicative roles. The paper evaluates a transformer and arecurrent model modified with various sharing strategies on a single-turn refer...
SP:05e114ff99351e68ad8c16f0655335d183262296
INTERNAL-CONSISTENCY CONSTRAINTS FOR EMERGENT COMMUNICATION
1 INTRODUCTION Emergent communication is the study of how linguistic protocols evolve when agents are tasked to cooperate . For example , agents engaged in a simple object retrieval task learn to communicate with one another in order to get the items they want ( Lazaridou et al. , 2018 ) . To date , work of this type h...
This paper investigates the question of internal consistency in emergent communication. In other words, the paper aims to answer the question ‘how is emergent communication improved if we enforce the constraint that an agent must speak in the same way that it listens?’ The paper explores three methods of enforcing inte...
SP:05e114ff99351e68ad8c16f0655335d183262296
Black-Box Adversarial Attack with Transferable Model-based Embedding
1 INTRODUCTION . The wide adoption of neural network models in modern applications has caused major security concerns , as such models are known to be vulnerable to adversarial examples that can fool neural networks to make wrong predictions ( Szegedy et al. , 2014 ) . Methods to attack neural networks can be divided i...
This paper proposed a new method for black-box adversarial attacks which tries to learn a low-dimensional embedding using a pretrained model and then performs efficient search within the embedding space to attack the target network. The proposed method can produce perturbation with semantic patterns are easily transfe...
SP:c5f7b60e4be5b38d3b29fc4e0395aca077e23eae
Black-Box Adversarial Attack with Transferable Model-based Embedding
1 INTRODUCTION . The wide adoption of neural network models in modern applications has caused major security concerns , as such models are known to be vulnerable to adversarial examples that can fool neural networks to make wrong predictions ( Szegedy et al. , 2014 ) . Methods to attack neural networks can be divided i...
Review: The paper proposes a new framework (TREMBA) for black-box adversarial attack. The method utilizes a pretrained source network to learn a low dimensional embedding, it then searches efficiently within the embedding space (using NES) and produces an adversarial perturbation that can attack an unknown target netwo...
SP:c5f7b60e4be5b38d3b29fc4e0395aca077e23eae
Time2Vec: Learning a Vector Representation of Time
1 INTRODUCTION . In building machine learning models , “ time ” is often an important feature . Examples include predicting daily sales for a company based on the date ( and other available features ) , predicting the time for a patient ’ s next health event based on their medical history , and predicting the song a pe...
This paper proposes a simple representation of time (Time2Vec) for modelling sequential data. The idea is to apply multiple sine functions to the time with trainable period and offset and concatenate them together, which is similar to positional encoding [Vaswani et al.] except that the periods and offsets are learned....
SP:a6d659731143b2b4b7c0b3940db926b80c2cf59f
Time2Vec: Learning a Vector Representation of Time
1 INTRODUCTION . In building machine learning models , “ time ” is often an important feature . Examples include predicting daily sales for a company based on the date ( and other available features ) , predicting the time for a patient ’ s next health event based on their medical history , and predicting the song a pe...
This paper introduces a particular learnable vector representation of time which is applicable across problems without the use of a hand-crafted time representation. Their representation makes use of a feed-forward layer with sine activations which operates on time data. As it is a vector representation, it combines we...
SP:a6d659731143b2b4b7c0b3940db926b80c2cf59f
Regularization Matters in Policy Optimization
Deep Reinforcement Learning ( Deep RL ) has been receiving increasingly more attention thanks to its encouraging performance on a variety of control tasks . Yet , conventional regularization techniques in training neural networks ( e.g. , L2 regularization , dropout ) have been largely ignored in RL methods , possibly ...
The paper provides an empirical study of regularization in policy optimization methods in multiple continuous control tasks. The paper focuses on the effect of conventional regularization on performance in training environments, not generalization ability to different (but similar) testing environments. Their findings ...
SP:2b7f7b8d90f9d6bf2d58fe973a988322c0a4a557
Regularization Matters in Policy Optimization
Deep Reinforcement Learning ( Deep RL ) has been receiving increasingly more attention thanks to its encouraging performance on a variety of control tasks . Yet , conventional regularization techniques in training neural networks ( e.g. , L2 regularization , dropout ) have been largely ignored in RL methods , possibly ...
This paper investigates the use of conventional regularizers for neural networks in the reinforcement learning setting. Contrary to the standard practice of foregoing regularizers in deep RL, the paper finds that their addition can improve the performance of policy gradient algorithms on a standard suite of continuous ...
SP:2b7f7b8d90f9d6bf2d58fe973a988322c0a4a557
Layer Flexible Adaptive Computation Time for Recurrent Neural Networks
1 INTRODUCTION . Recurrent neural networks ( RNN ) are widely used in supervised machine learning tasks for their superior performance in sequence data , such as machine translation Auli et al . ( 2013 ) ; Liu et al . ( 2014 ) , speech recognition Graves et al . ( 2013 ) ; Hannun et al . ( 2014 ) , image description ge...
The authors propose Layer Flexible Adaptive Computation Time, an RNN-esque sequence model with varying depth at each time step. The idea is that the model can adaptively choose how much computational effort to spend on each example. The authors evaluate the model empirically on a financial dataset and Wikipedia languag...
SP:fd5ef5edf4f2aba1e3570d05ed55f80a620b50e3
Layer Flexible Adaptive Computation Time for Recurrent Neural Networks
1 INTRODUCTION . Recurrent neural networks ( RNN ) are widely used in supervised machine learning tasks for their superior performance in sequence data , such as machine translation Auli et al . ( 2013 ) ; Liu et al . ( 2014 ) , speech recognition Graves et al . ( 2013 ) ; Hannun et al . ( 2014 ) , image description ge...
This paper proposes a layer-flexible adaptive computation time model which enables learning with a different number of layers at each time step. It proposed a set of mechanisms to make the variable layer possible. It uses attention to re-arrange the hidden states in different layers into a different number of hidden s...
SP:fd5ef5edf4f2aba1e3570d05ed55f80a620b50e3
Causal Discovery with Reinforcement Learning
1 INTRODUCTION . Discovering and understanding causal mechanisms underlying natural phenomena are important to many disciplines of sciences . An effective approach is to conduct controlled randomized experiments , which however is expensive or even impossible in certain fields such as social sciences ( Bollen , 1989 ) ...
In this paper, the authors propose an RL-based structure searching method for causal discovery. The authors reformulate the score-based causal discovery problem into an RL-format, which includes the reward function re-design, hyper-parameter choose, and graph generation. To my knowledge, it’s the first time that the RL...
SP:bc4cb6a48758c5a9d81fa13e2b144f151dd6a85b
Causal Discovery with Reinforcement Learning
1 INTRODUCTION . Discovering and understanding causal mechanisms underlying natural phenomena are important to many disciplines of sciences . An effective approach is to conduct controlled randomized experiments , which however is expensive or even impossible in certain fields such as social sciences ( Bollen , 1989 ) ...
This work addresses the task of causal discovery. The proposed contribution is to apply prior work which uses reinforcement learning for combinatorial optimization to structure learning. Specifically, the proposed optimization problem seeks to maximize a penalized score criterion subject to the acyclicity constraint pr...
SP:bc4cb6a48758c5a9d81fa13e2b144f151dd6a85b
Optimizing Loss Landscape Connectivity via Neuron Alignment
The loss landscapes of deep neural networks are poorly understood due to their high nonconvexity . Empirically , the local minima of these loss functions can be connected by a learned curve in model space , along which the loss remains nearly constant . Yet , current path finding algorithms do not consider the influenc...
Given two parameters theta_1 and theta_2 of the same architecture, the authors propose to learn a minimal loss curve between theta_1 and P theta_2, where P is a permutation matrix yielding another equivalent parameterization of the same network. The authors show that either by initializing P with neuron alignment or by...
SP:8a9d38f92030daeccbe4a9d4489d116d3316d519
Optimizing Loss Landscape Connectivity via Neuron Alignment
The loss landscapes of deep neural networks are poorly understood due to their high nonconvexity . Empirically , the local minima of these loss functions can be connected by a learned curve in model space , along which the loss remains nearly constant . Yet , current path finding algorithms do not consider the influenc...
The paper investigates the connection between symmetries in the neural network architectures and the loss landscape of the corresponding neural networks. In the previous works, there was shown that the two local minima of a neural network can be connected by a curve with the low validation/train loss along the curve. D...
SP:8a9d38f92030daeccbe4a9d4489d116d3316d519
Coresets for Accelerating Incremental Gradient Methods
1 INTRODUCTION . Mathematical optimization lies at the core of training large-scale machine learning systems , and is now widely used over massive data sets with great practical success , assuming sufficient data resources are available . Achieving this success , however , also requires large amounts of ( often GPU ) c...
This paper proposes a novel extension to SGD/incremental gradient methods called CRAIG. The algorithm selects a subset of datapoints to approximate the training loss at the beginning of each epoch in order to reduce the total amount of time necessary to solve the empirical risk minimization problem. In particular, the ...
SP:e0f8626408b0f7931b4778c5696f8df856a49f3a
Coresets for Accelerating Incremental Gradient Methods
1 INTRODUCTION . Mathematical optimization lies at the core of training large-scale machine learning systems , and is now widely used over massive data sets with great practical success , assuming sufficient data resources are available . Achieving this success , however , also requires large amounts of ( often GPU ) c...
The paper proposes a theoretically founded method to generate subsets of a dataset, together with corresponding sample weights in a way that the average gradient of the subset is at most epsilon far from the average gradient of the full dataset. Given such a subset, the authors provide theoretical guarantees for conver...
SP:e0f8626408b0f7931b4778c5696f8df856a49f3a
Watch the Unobserved: A Simple Approach to Parallelizing Monte Carlo Tree Search
1 INTRODUCTION . Recently , Monte Carlo Tree Search ( MCTS ) algorithms such as UCT ( Kocsis et al. , 2006 ) have achieved great success in solving many challenging artificial intelligence ( AI ) benchmarks , including video games ( Guo et al. , 2016 ) and Go ( Silver et al. , 2016 ) . However , they rely on a large nu...
The paper introduces a new algorithm for parallelizing monte carlo tree search (MCTS). MCTS is hard to parallelize as we have to keep track of the statistics of the node of the tree, which are typically not up-to-date in a parallel execution. The paper introduces a new algorithm that updates the visitation counts befor...
SP:f428ec6777a6b4119612bb27b43fe6128ef23c58
Watch the Unobserved: A Simple Approach to Parallelizing Monte Carlo Tree Search
1 INTRODUCTION . Recently , Monte Carlo Tree Search ( MCTS ) algorithms such as UCT ( Kocsis et al. , 2006 ) have achieved great success in solving many challenging artificial intelligence ( AI ) benchmarks , including video games ( Guo et al. , 2016 ) and Go ( Silver et al. , 2016 ) . However , they rely on a large nu...
This paper introduces a new algorithm for parallelizing Monte-Carlo Tree Search (MCTS). Specifically, when expanding a new node in the search tree, the algorithm updates the parent nodes’ statistics of the visit counts but not their values; it is only when the expansion and simulation steps are complete that the values...
SP:f428ec6777a6b4119612bb27b43fe6128ef23c58
Hamiltonian Generative Networks
1 INTRODUCTION Any system capable of a wide range of intelligent behaviours within a dynamic environment requires a good predictive model of the environment ’ s dynamics . This is true for intelligence in both biological ( Friston , 2009 ; 2010 ; Clark , 2013 ) and artificial ( Hafner et al. , 2019 ; Battaglia et al. ,...
The authors present a method for learning Hamiltonian functions that govern a dynamical directly from observational data. The basic approach uses three networks: 1) an inference network (I'm not clear why this is not just called an encoder), that maps past observations to a latent p,q space in a VAE-like fashion; 2) a...
SP:aba5a1845cf2e38d592d640278f9ed0134a3d96d
Hamiltonian Generative Networks
1 INTRODUCTION Any system capable of a wide range of intelligent behaviours within a dynamic environment requires a good predictive model of the environment ’ s dynamics . This is true for intelligence in both biological ( Friston , 2009 ; 2010 ; Clark , 2013 ) and artificial ( Hafner et al. , 2019 ; Battaglia et al. ,...
The paper introduces a novel way of learning Hamiltonian dynamics with a generative network. The Hamiltonian generative network (HGN) learns the dynamics directly from data by embedding observations in a latent space, which is then transformed into a phase space describing the system's initial (abstract) position and m...
SP:aba5a1845cf2e38d592d640278f9ed0134a3d96d
Why Gradient Clipping Accelerates Training: A Theoretical Justification for Adaptivity
1 INTRODUCTION . We study optimization algorithms for neural network training and aim to resolve the mystery of why adaptive methods converge fast . Specifically , we study gradient-based methods for minimizing a differentiable nonconvex function f : Rd → R , where f ( x ) can potentially be stochastic , i.e. , f ( x )...
In this paper, the authors relax the generally used Lipschitz smoothness condition in optimization, to a more general smoothness condition that may depend on norm of the gradient. The authors proved that, with this relaxed condition, under such cases, both GD and clipped GD can converge within O(1/\epsilon^2) time, but...
SP:f97c429a41ec4715cb88acf6dde1c4d4ffa5957d
Why Gradient Clipping Accelerates Training: A Theoretical Justification for Adaptivity
1 INTRODUCTION . We study optimization algorithms for neural network training and aim to resolve the mystery of why adaptive methods converge fast . Specifically , we study gradient-based methods for minimizing a differentiable nonconvex function f : Rd → R , where f ( x ) can potentially be stochastic , i.e. , f ( x )...
This paper applies new assumption on smoothness that assume the norm of Hessian is bounded by a scalar plus norm of gradient. The traditional smoothness is only bounded with a scalar, the proposed assumption is more relaxed because now the norm of Hessian can grow with the norm of gradient. Under this assumption, the a...
SP:f97c429a41ec4715cb88acf6dde1c4d4ffa5957d
Perturbations are not Enough: Generating Adversarial Examples with Spatial Distortions
Deep neural network image classifiers are reported to be susceptible to adversarial evasion attacks , which use carefully crafted images created to mislead a classifier . Recently , various kinds of adversarial attack methods have been proposed , most of which focus on adding small perturbations to input images . Despi...
This paper proposes a new adversarial attack method by combining spatial transformations with perturbation-based noises. The proposed method uses two networks to generate the parameters of spatial transformation and the perturbation noise. The whole architecture is trained by a variant of GAN-loss to make the adversari...
SP:6feb5e58a91fa2eaab6a915a3897235ab02ebd82
Perturbations are not Enough: Generating Adversarial Examples with Spatial Distortions
Deep neural network image classifiers are reported to be susceptible to adversarial evasion attacks , which use carefully crafted images created to mislead a classifier . Recently , various kinds of adversarial attack methods have been proposed , most of which focus on adding small perturbations to input images . Despi...
This paper builds upon the work of AdvGAN and proposes to add spatial transformations on top of it. The resulting attacking framework is demonstrated to outperform AdvGAN on attacking several defense approaches, such as Defense-GAN, AdvCritic and adversarial training. Compared to previous approaches on generating spati...
SP:6feb5e58a91fa2eaab6a915a3897235ab02ebd82
Learning De-biased Representations with Biased Representations
1 INTRODUCTION . Most machine learning algorithms are trained and evaluated by randomly splitting a single source of data into training and test sets . Although this is a standard protocol , it is blind to a critical problem : the existence of dataset bias ( Torralba & Efros , 2011 ) . For instance , many frog images a...
This manuscript discusses the problem of bias shortcut employed by many machine learning algorithms (due to dataset problems or underlying effects of any algorithmic bias within an application). The authors argue that models tend to underutilize their capacities to extract non-bias signals when bias shortcuts provide e...
SP:686bfd24ecc46724adfe6b4fba91dc2012b188c0
Learning De-biased Representations with Biased Representations
1 INTRODUCTION . Most machine learning algorithms are trained and evaluated by randomly splitting a single source of data into training and test sets . Although this is a standard protocol , it is blind to a critical problem : the existence of dataset bias ( Torralba & Efros , 2011 ) . For instance , many frog images a...
The paper describes a methodology for reducing model dependance on bias by specifying a model family of biases (i.e. conv nets with only 1x1 convs to model color biases), and then forcing independence between feature representations of the bias model and the a full model (i.e. conv nets with 3x3 convs to also model edg...
SP:686bfd24ecc46724adfe6b4fba91dc2012b188c0
On Empirical Comparisons of Optimizers for Deep Learning
1 INTRODUCTION . The optimization algorithm chosen by a deep learning practitioner determines the training speed and the final predictive performance of their model . To date , there is no theory that adequately explains how to make this choice . Instead , our community relies on empirical studies ( Wilson et al. , 201...
This paper presents experimental data supporting the claim that the under aggressive hyper-parameter tuning different optimizers are essentially ranked by inclusion --- if the hyper-parameters of method A can simulate any setting of the hyper-parameters of method B then under aggressive hyper-parameter tuning A will do...
SP:18de04936d8fc216f861fd000e156957311ea391
On Empirical Comparisons of Optimizers for Deep Learning
1 INTRODUCTION . The optimization algorithm chosen by a deep learning practitioner determines the training speed and the final predictive performance of their model . To date , there is no theory that adequately explains how to make this choice . Instead , our community relies on empirical studies ( Wilson et al. , 201...
The paper provides an empirical comparison of a set of first-order optimization methods for deep learning models. Those optimizers include stochastic gradient descent, momentum method, RMSProp, Adam, Nesterov, and Nadam, which arguably covers all popular variants used in the literature. Although it is not the first em...
SP:18de04936d8fc216f861fd000e156957311ea391
Stochastic Prototype Embeddings
1 INTRODUCTION . Supervised deep-embedding methods map instances from an input space to a latent embedding space in which same-label pairs are near and different-label pairs are far . The embedding thus captures semantic relationships without discarding inter-class structure . In contrast , consider a standard neural n...
The paper proposes stochastic prototype embeddings (SPE) for few-shot learning. The method is an extension of Prototypical Networks (PN, [1]) with Gaussian embeddings. The idea is to take representation uncertainty into account when classifying objects which makes the model more robust to input and label noise. The aut...
SP:0b3653c04024ddae217d91fdf4bec5f3a1250453
Stochastic Prototype Embeddings
1 INTRODUCTION . Supervised deep-embedding methods map instances from an input space to a latent embedding space in which same-label pairs are near and different-label pairs are far . The embedding thus captures semantic relationships without discarding inter-class structure . In contrast , consider a standard neural n...
By extending prototypical networks, this paper proposes a probabilistic model, i.e., stochastic prototype embedding, that treats embeddings as random variables. The model is very straightforward and easy to understand. The authors make a few assumptions to simplify the problem. For example, the distance between every i...
SP:0b3653c04024ddae217d91fdf4bec5f3a1250453
Don't Use Large Mini-batches, Use Local SGD
1 INTRODUCTION . Fast and efficient training of large scale deep-learning models relies on distributed hardware and on distributed optimization algorithms . For efficient use of system resources , these algorithms crucially must ( i ) enable parallelization while being communication efficient , and ( ii ) exhibit good ...
This paper proposes a variant of local SGD, post-local SGD, for distributed training of deep neural networks. It targets to mitigate the generalization gap caused by large batch training. The idea is straightforward and easy to understand-- start the training with standard mini-batch SGD and later switch to local SGD. ...
SP:2194acbf643b821c3c8e83481ff8af8a43e6fa83
Don't Use Large Mini-batches, Use Local SGD
1 INTRODUCTION . Fast and efficient training of large scale deep-learning models relies on distributed hardware and on distributed optimization algorithms . For efficient use of system resources , these algorithms crucially must ( i ) enable parallelization while being communication efficient , and ( ii ) exhibit good ...
This paper proposes a new distributed computation technique for SGD training of deep neural networks. The proposed method is a modification of the local SGD which updates models distributed to several workers in a parallel way and synchronize the model parameters at every few epochs. The local SGD shows a nice performa...
SP:2194acbf643b821c3c8e83481ff8af8a43e6fa83
Count-guided Weakly Supervised Localization Based on Density Map
1 INTRODUCTION . Deep convolutional neural networks ( CNN ) have significantly pushed the frontier of image processing . Trained with enormous amount of data , CNNs have surpassed human performance in many object detecting tasks . However , humans are still ahead in many aspects . One of them is the segmentation based ...
This article proposes a method for object counting which can be trained with weak supervision. Object counting methods are often trained with point annotations, i.e., one click-point per object. In this article, a weaker way of annotation is used: count-based annotation, i.e., the number of objects of each class presen...
SP:608ec90d9c9d08fc825ae49a17cff75d816c95a7
Count-guided Weakly Supervised Localization Based on Density Map
1 INTRODUCTION . Deep convolutional neural networks ( CNN ) have significantly pushed the frontier of image processing . Trained with enormous amount of data , CNNs have surpassed human performance in many object detecting tasks . However , humans are still ahead in many aspects . One of them is the segmentation based ...
The main contribution of the paper is the extension of techniques for weakly supervised localization, i.e. given ground truth counts of objects in a given image, one can do training to generate hidden layer density maps that allow for feature detection and localization of objects. The main contribution of the paper s...
SP:608ec90d9c9d08fc825ae49a17cff75d816c95a7
Are Few-shot Learning Benchmarks Too Simple ?
We argue that the widely used Omniglot and miniImageNet benchmarks are too simple because their class semantics do not vary across episodes , which defeats their intended purpose of evaluating few-shot classification methods . The class semantics of Omniglot is invariably “ characters ” and the class semantics of miniI...
The paper is concerned with few-shot classification, both its benchmarks and method used to tackle it. The scope of the few-shot classification problem can be set relatively widely, depending on what data is available at what stage. In general few-shot classification is an important ability of intelligent systems and a...
SP:200f79d7a66aa5ffdaf5f302c564b1567d3e1589
Are Few-shot Learning Benchmarks Too Simple ?
We argue that the widely used Omniglot and miniImageNet benchmarks are too simple because their class semantics do not vary across episodes , which defeats their intended purpose of evaluating few-shot classification methods . The class semantics of Omniglot is invariably “ characters ” and the class semantics of miniI...
The authors argue that the popular benchmark datasets, Omniglot and miniImageNet, are too simple to evaluate supervised few-shot classification methods due to their insufficient variety of class semantics. To validate this, the authors proposed clustering-based meta-learning method, called Centroid Network. Although it...
SP:200f79d7a66aa5ffdaf5f302c564b1567d3e1589
Graph Neural Networks Exponentially Lose Expressive Power for Node Classification
1 INTRODUCTION . Motivated by the success of Deep Learning ( DL ) , several attempts have been made to apply DL models to non-Euclidean data , particularly , graph-structured data such as chemical compounds , social networks , and polygons . Recently , Graph Neural Networks ( graph NNs ) ( Duvenaud et al. , 2015 ; Li e...
In the paper, the authors carry out theoretically analysis on the expressive power for GNN. The analysis focused on the limiting case when the depth of layers goes to infinite. The authors prove that if the weights of the GNN satisfy certain condition based on the graph Laplacian, then the transformed features contain ...
SP:7028036b485c4e14aa50cc8a9f788f11b8807dd2
Graph Neural Networks Exponentially Lose Expressive Power for Node Classification
1 INTRODUCTION . Motivated by the success of Deep Learning ( DL ) , several attempts have been made to apply DL models to non-Euclidean data , particularly , graph-structured data such as chemical compounds , social networks , and polygons . Recently , Graph Neural Networks ( graph NNs ) ( Duvenaud et al. , 2015 ; Li e...
The paper studies why graph NNs lose the expressive power as additional layers are added. A dynamical system perspective is adopted and used to show that under certain conditions on the weights, the expressiveness of the network deteriorates. This is since the network's output eventually only carries information about ...
SP:7028036b485c4e14aa50cc8a9f788f11b8807dd2