paper_name stringlengths 11 170 | text stringlengths 8.07k 307k | summary stringlengths 152 6.16k | paper_id stringlengths 43 43 |
|---|---|---|---|
Behaviour Suite for Reinforcement Learning | 1 Introduction . The reinforcement learning ( RL ) problem describes an agent interacting with an environment with the goal of maximizing cumulative reward through time ( Sutton & Barto , 2017 ) . Unlike other branches of control , the dynamics of the environment are not fully known to the agent , but can be learned th... | This paper presents the « Behavior Suite for Reinforcement Learning » (bsuite), which is a set of RL tasks (called « experiments ») meant to evaluate an algorithm’s ability to solve various key challenges in RL. Importantly, these experiments are designed to run fast enough that one can benchmark a new algorithm within... | SP:0754d0dac07aed6b6672a6b0393087d90a5fe535 |
Behaviour Suite for Reinforcement Learning | 1 Introduction . The reinforcement learning ( RL ) problem describes an agent interacting with an environment with the goal of maximizing cumulative reward through time ( Sutton & Barto , 2017 ) . Unlike other branches of control , the dynamics of the environment are not fully known to the agent , but can be learned th... | In this paper, the authors propose a set of benchmarks for evaluating different aspects of reinforcement learning algorithms such as generalisation, exploration, and memory. The aim is to provide a set of simple environments to better understand the RL algorithms and also to provide a set of scores that summarise the p... | SP:0754d0dac07aed6b6672a6b0393087d90a5fe535 |
Matrix Multilayer Perceptron | 1 INTRODUCTION . For certain applications , it is desirable to construct a conditional covariance matrix as a function of the input ( the explanatory variable ) . The problem arises , for instance , in spatial ( and spatio-temporal ) statistics , in relation to the heteroscedastic multivariate regression ( e.g. , Poura... | This manuscript proposes a novel formulation of the MLP to address predicting a symmetric positive definite (SPD) matrix from an input vector or matrix. While the field has had methods for years to estimate SPD matrices (such as the covariance matrix estimate in the reparameterization trick), this manuscript proposes ... | SP:b0c94a1ef77cbdab8af5402718ed06964148dd95 |
Matrix Multilayer Perceptron | 1 INTRODUCTION . For certain applications , it is desirable to construct a conditional covariance matrix as a function of the input ( the explanatory variable ) . The problem arises , for instance , in spatial ( and spatio-temporal ) statistics , in relation to the heteroscedastic multivariate regression ( e.g. , Poura... | This paper explores the problem of deep heteroskedastic multivariate regression where the goal is to regress over symmetric positive definite matrices; that is, the deep learning model should take as input data points, and produce a conditional covariance matrix as the output. The key challenge in this setting is how t... | SP:b0c94a1ef77cbdab8af5402718ed06964148dd95 |
On the expected running time of nonconvex optimization with early stopping | 1 INTRODUCTION . This work considers the minimization of a differentiable and possible nonconvex objective function : min x∈Rd f ( x ) . ( 1 ) A generally accepted success criteria for algorithms that use only first-order information is that an approximate stationary point is generated . These are points x ∈ Rd at whic... | In this paper, the authors consider stochastic optimization in the setting where a validation function is used to guide the termination of the algorithm. In more details, the algorithm terminates if the gradient of the validation function at an iterate is smaller than a threshold. In this framework, the authors conside... | SP:228fd66964ccbf61d40a38bd12db78cad1401136 |
On the expected running time of nonconvex optimization with early stopping | 1 INTRODUCTION . This work considers the minimization of a differentiable and possible nonconvex objective function : min x∈Rd f ( x ) . ( 1 ) A generally accepted success criteria for algorithms that use only first-order information is that an approximate stationary point is generated . These are points x ∈ Rd at whic... | This paper proposes an optimization approach in which the optimizer computes the gradient on a given function yet uses another to decide a stopping time. Conceptually those functions are empirical errors on train and validation folds in the most common setting, although the authors seem to use other settings later in t... | SP:228fd66964ccbf61d40a38bd12db78cad1401136 |
Barcodes as summary of objective functions' topology | 1 INTRODUCTION . The learning via finding minima of objective functions is the principal strategy underlying majority of learning algorithms . For example , in Neural Network training , the objective function ’ s input is model parameters ( weights ) and the objective function ’ s output is the loss on training dataset... | This work is focused on topological characterization of target surfaces of optimization objectives (i.e. loss functions) by computing so called barcodes, which are lists of pairs of local minima and their connected saddle points. The authors claim that the barcodes constitute a representation of target objectives that ... | SP:dc8557f06ebb81345d2edeb98716e8327dcb30d8 |
Barcodes as summary of objective functions' topology | 1 INTRODUCTION . The learning via finding minima of objective functions is the principal strategy underlying majority of learning algorithms . For example , in Neural Network training , the objective function ’ s input is model parameters ( weights ) and the objective function ’ s output is the loss on training dataset... | This paper introduces the notion of barcodes as a topological invariant of loss surfaces that encodes the "depth" of local minima by associating to each minimum the lowest index-one saddle. An algorithm is presented for the computation of barcodes, and some small-scale experiments are conducted. For very small neural n... | SP:dc8557f06ebb81345d2edeb98716e8327dcb30d8 |
Differentiable Hebbian Consolidation for Continual Learning | 1 INTRODUCTION . A key aspect of human intelligence is the ability to continually adapt and learn in dynamic environments , a characteristic which is challenging to embed into artificial intelligence . Recent advances in machine learning ( ML ) have shown tremendous improvements in various problems , by learning to sol... | This paper addresses the continual learning setting, and aims to mitigate catastrophic forgetting, with results on Permuted MNIST, Split MNIST, Vision Datasets Mixture, and their own class-imbalanced version of the Permuted MNIST dataset. The authors propose to augment differentiable plastic weights - a general neural ... | SP:ecdae30f9692bf6d23cd5a571dabb82fe782c244 |
Differentiable Hebbian Consolidation for Continual Learning | 1 INTRODUCTION . A key aspect of human intelligence is the ability to continually adapt and learn in dynamic environments , a characteristic which is challenging to embed into artificial intelligence . Recent advances in machine learning ( ML ) have shown tremendous improvements in various problems , by learning to sol... | The authors introduce DIFFERENTIABLE HEBBIAN CONSOLIDATION,a new framework for continual learning that can be implemented in the usual differentiable programming setups. This framework is motivated in terms of complementary learning system (CLS) theory which features an episodic memory module. The method is shown to be... | SP:ecdae30f9692bf6d23cd5a571dabb82fe782c244 |
ROBUST DISCRIMINATIVE REPRESENTATION LEARNING VIA GRADIENT RESCALING: AN EMPHASIS REGULARISATION PERSPECTIVE | 1 INTRODUCTION . DNNs have been successfully applied in diverse applications ( Socher et al. , 2011 ; Krizhevsky et al. , 2012 ; LeCun et al. , 2015 ) . However , their success is heavily reliant on the quality of training data , especially accurate semantic labels for learning supervision . Unfortunately , on the one ... | The paper proposes a method for noise robustness based on scaling gradients of examples. By choosing the proper scaling parameters (alpha and beta), the method recovers standard losses such as CCE, MAE, and GCE, while also recovering other losses. The method is strongly related to reweighting training examples, where a... | SP:1e1a6d0bb0dc9352227b3cade1e3b88096a544b5 |
ROBUST DISCRIMINATIVE REPRESENTATION LEARNING VIA GRADIENT RESCALING: AN EMPHASIS REGULARISATION PERSPECTIVE | 1 INTRODUCTION . DNNs have been successfully applied in diverse applications ( Socher et al. , 2011 ; Krizhevsky et al. , 2012 ; LeCun et al. , 2015 ) . However , their success is heavily reliant on the quality of training data , especially accurate semantic labels for learning supervision . Unfortunately , on the one ... | This paper presents Gradient Rescaling (GR) for robust learning to combat label noise. They propose to treat each data sample with different significance scores: some samples are important to learning, and some examples are insignificant (or even detrimental) to learning. So they desire to weight each samples according... | SP:1e1a6d0bb0dc9352227b3cade1e3b88096a544b5 |
MixUp as Directional Adversarial Training | 1 INTRODUCTION . Deep learning applications often require complex networks with a large number of parameters ( He et al. , 2016 ; Zagoruyko & Komodakis , 2016 ; Devlin et al. , 2018 ) . Although neural networks perform so well that their ability to generalize is an area of study in itself ( Zhang et al. , 2017a ; Arpit... | This paper proposes a novel data augmentation method, untied MixUp (UMixUp), which is a general case of both MixUp and Directional Adversarial Traning (DAT). DAT is referred to in this paper as a scheme that only input feature vectors are mixed, while MixUp also incorporates their corresponding labels. The authors prov... | SP:d74f95781e1b0f164644b7e30247791eae6afc79 |
MixUp as Directional Adversarial Training | 1 INTRODUCTION . Deep learning applications often require complex networks with a large number of parameters ( He et al. , 2016 ; Zagoruyko & Komodakis , 2016 ; Devlin et al. , 2018 ) . Although neural networks perform so well that their ability to generalize is an area of study in itself ( Zhang et al. , 2017a ; Arpit... | This paper introduces directional adversarial training (DAT) and UMixUP, which are extension methods of MixUp. DAT and UMixUp use the same method of MixUp for generating samples but use different label mixing ratios where DAT retains the sample's original label. In contrast, UMixUp uses a function of the input mixing r... | SP:d74f95781e1b0f164644b7e30247791eae6afc79 |
Neural Network Branching for Neural Network Verification | 1 INTRODUCTION . Despite their outstanding performances on various tasks , neural networks are found to be vulnerable to adversarial examples ( Goodfellow et al. , 2015 ; Szegedy et al. , 2013 ) . The brittleness of neural networks can have costly consequences in areas such as autonomous driving , finance and healthcar... | This paper deals with complete formal verification of Neural Network, based on the Branch and Bound framework. The authors focus on branching strategies, which have been shown to be a critical design decision in order to obtain good performance. The tactic employed here is to learn a Graph Neural Network (which allows ... | SP:a527aca3ea9653acb0d0a07eada4483414fd82e3 |
Neural Network Branching for Neural Network Verification | 1 INTRODUCTION . Despite their outstanding performances on various tasks , neural networks are found to be vulnerable to adversarial examples ( Goodfellow et al. , 2015 ; Szegedy et al. , 2013 ) . The brittleness of neural networks can have costly consequences in areas such as autonomous driving , finance and healthcar... | The paper proposes learning a branching heuristic to be used inside a branch-and-bound algorithm used for solving integer programming problems corresponding to neural network verification. The heuristic is parameterized as a neural network and trained to imitate an existing heuristic called Strong Branching which is co... | SP:a527aca3ea9653acb0d0a07eada4483414fd82e3 |
Representing Unordered Data Using Multiset Automata and Complex Numbers | 1 INTRODUCTION . Neural networks which operate on set-structured input have been gaining interest for their ability to handle unordered and variable-sized inputs ( Vinyals et al. , 2015 ; Wagstaff et al. , 2019 ) . They have been applied to various tasks , such as processing graph nodes ( Murphy et al. , 2018 ) , hyper... | This work presents an encoding approach for unordered set input to neural networks. The authors base their approach on weighted finite automata, where in order to absorb unordered sets, they enforce multiplicative commutativity on transition matrices by approximating them as complex diagonal matrices. The authors furth... | SP:bbde4c0910f4ef9ef433f0349f7ff3edd569b63f |
Representing Unordered Data Using Multiset Automata and Complex Numbers | 1 INTRODUCTION . Neural networks which operate on set-structured input have been gaining interest for their ability to handle unordered and variable-sized inputs ( Vinyals et al. , 2015 ; Wagstaff et al. , 2019 ) . They have been applied to various tasks , such as processing graph nodes ( Murphy et al. , 2018 ) , hyper... | This paper proposed a complex weights based multiset automata designed to represent unordered data. The main idea of multiset automata is that the transition matrices of the automata is pairwise commutative. To achieve this property, the authors proposed to restrict the transition matrices to be diagonal and shows that... | SP:bbde4c0910f4ef9ef433f0349f7ff3edd569b63f |
Defective Convolutional Layers Learn Robust CNNs | 1 INTRODUCTION . Deep learning ( LeCun et al. , 2015 ) , especially deep Convolutional Neural Network ( CNN ) ( LeCun et al. , 1998 ) , has led to state-of-the-art results spanning many machine learning fields ( He et al. , 2016 ; Ren et al. , 2015 ) . Despite the great success in numerous applications , recent studies... | The paper deals with robustness against adversarial attacks. It proposes to blank out large parts of the early convolution layers in a CNN, in an attempt to shift the focus from "texture" to "shape" features. This does seem to improve robustness against adversarial examples, with only a small decrease in general classi... | SP:942e9e4be427dd59ec333c2a3073288c4c418cdc |
Defective Convolutional Layers Learn Robust CNNs | 1 INTRODUCTION . Deep learning ( LeCun et al. , 2015 ) , especially deep Convolutional Neural Network ( CNN ) ( LeCun et al. , 1998 ) , has led to state-of-the-art results spanning many machine learning fields ( He et al. , 2016 ; Ren et al. , 2015 ) . Despite the great success in numerous applications , recent studies... | The paper proposes defective convolutional layers as a measure of defense against adversarial attacks on deep neural networks. This layer sets the outputs of a randomly sampled but *fixed* set of neurons in the convolutional layers to zero during training and testing. The authors claim that defective convolutional laye... | SP:942e9e4be427dd59ec333c2a3073288c4c418cdc |
ASGen: Answer-containing Sentence Generation to Pre-Train Question Generator for Scale-up Data in Question Answering | 1 INTRODUCTION . Machine reading comprehension ( MRC ) , which finds an answer to a given question from given paragraphs called context , is an essential task in natural language processing . With the use of high-quality human-annotated datasets for this task , such as SQuAD-v1.1 ( Rajpurkar et al. , 2016 ) , SQuAD-v2.... | This paper proposes a pretraining technique for question generation, where an answer candidate is chosen beforehand, and the objective is to predict the answer containing sentence given a paragraph excluding this sentence and the target answer candidate. The intuition of this method is that question generation requires... | SP:ff1a7f2310f3d3c647ede8e418dcc104b9da3e2b |
ASGen: Answer-containing Sentence Generation to Pre-Train Question Generator for Scale-up Data in Question Answering | 1 INTRODUCTION . Machine reading comprehension ( MRC ) , which finds an answer to a given question from given paragraphs called context , is an essential task in natural language processing . With the use of high-quality human-annotated datasets for this task , such as SQuAD-v1.1 ( Rajpurkar et al. , 2016 ) , SQuAD-v2.... | The paper in the field of machine reading comprehension. The authors address the issue of generating labeled data of question-answer tuples, without the need of manual annotation. Specifically, the authors propose a method that dynamically generates K answers given a paragraph in order to generate diverse questions an... | SP:ff1a7f2310f3d3c647ede8e418dcc104b9da3e2b |
Unsupervised Hierarchical Graph Representation Learning with Variational Bayes | 1 INTRODUCTION . Graph representation learning has attracted a surge of interest recently , inspired by the widespread success of representation learning in the image and language domains through the use of deep neural networks for parameterization . A substantial number of graph neural network ( GNN ) architectures ( ... | This work proposes an unsupervised hierarchical graph representation learning method, named BayesPool. The method learns a coarsening sequence of graphs together with the corresponding node representations. The coarsening sequence is learned using the method in Loukas (2019). The node representations are learned using ... | SP:ab4fbcfc2199b778ff071e8ccff33efc8c37e351 |
Unsupervised Hierarchical Graph Representation Learning with Variational Bayes | 1 INTRODUCTION . Graph representation learning has attracted a surge of interest recently , inspired by the widespread success of representation learning in the image and language domains through the use of deep neural networks for parameterization . A substantial number of graph neural network ( GNN ) architectures ( ... | The authors propose in this paper a new unsupervised graph representation learning method. The method leverages recent advances in graph coarsening, mainly Loukas' method. The key idea of the method consists in using a reconstruction target that is not the classical one in an auto-encoder setting. More precisely, the e... | SP:ab4fbcfc2199b778ff071e8ccff33efc8c37e351 |
Boosting Network: Learn by Growing Filters and Layers via SplitLBI | Network structures are important to learning good representations of many tasks in computer vision and machine learning communities . These structures are either manually designed , or searched by Neural Architecture Search ( NAS ) in previous works , which however requires either expert-level efforts , or prohibitive ... | This paper focuses on topic of searching for the optimal architecture for the deep network. Building on the split linearized bregman iteration strategy, the authors propose two practical algorithms to boost network, namely GT-filters Alg and GT-layers Alg. The proposed algorithms can simultaneously grow and train a net... | SP:f5783ca08d51aa886277c95f86438981e3e74810 |
Boosting Network: Learn by Growing Filters and Layers via SplitLBI | Network structures are important to learning good representations of many tasks in computer vision and machine learning communities . These structures are either manually designed , or searched by Neural Architecture Search ( NAS ) in previous works , which however requires either expert-level efforts , or prohibitive ... | This paper proposes an architecture search method for deep convolutional neural network models that progressively increases the number of filters per layer as well as the number of layers, and the authors refer to this general approach as boosting networks. The algorithm for increasing the number of filters is based on... | SP:f5783ca08d51aa886277c95f86438981e3e74810 |
A Theory of Usable Information under Computational Constraints | We propose a new framework for reasoning about information in complex systems . Our foundation is based on a variational extension of Shannon ’ s information theory that takes into account the modeling power and computational constraints of the observer . The resulting predictive V-information encompasses mutual inform... | The paper introduces a framework for quantifying information about one random variable, given another random variable (“side information”) and, importantly, a function class of allowed transformations that can be applied to the latter. This matches the typical scenario in machine learning, where observations (playing t... | SP:03b7bce7c88de2434b54fc0483d8905aa04203e9 |
A Theory of Usable Information under Computational Constraints | We propose a new framework for reasoning about information in complex systems . Our foundation is based on a variational extension of Shannon ’ s information theory that takes into account the modeling power and computational constraints of the observer . The resulting predictive V-information encompasses mutual inform... | The paper presents a generalization of classical definitions of entropy and mutual information that can capture computational constraints. Intuitively, information theoretic results assume infinite computational resources, so they may not correspond to how we treat "information" in practice. One example is public-key e... | SP:03b7bce7c88de2434b54fc0483d8905aa04203e9 |
Deep exploration by novelty-pursuit with maximum state entropy | 1 INTRODUCTION . Efficient exploration is important to learn a ( near- ) optimal policy for reinforcement learning ( RL ) in huge state space ( Sutton & Barto , 1998 ) . Dithering strategies like epsilon-greedy , Gaussian action noise , and Boltzmann exploration are inefficient and require exponential interactions to e... | The paper addresses the challenge of intrinsically-driven exploration in tasks with sparse or delayed rewards. First, the authors try to bridge the gap between the objectives of intrinsically-motivated goal generation and maximum state entropy exploration. Then, they propose a new exploration method, called novelty-pur... | SP:f3287af29c0148119a9b84df7681dbccb4884ef7 |
Deep exploration by novelty-pursuit with maximum state entropy | 1 INTRODUCTION . Efficient exploration is important to learn a ( near- ) optimal policy for reinforcement learning ( RL ) in huge state space ( Sutton & Barto , 1998 ) . Dithering strategies like epsilon-greedy , Gaussian action noise , and Boltzmann exploration are inefficient and require exponential interactions to e... | This paper proposes novelty-pursuit for exploration in large state space. In theory, novelty-pursuit is motivated by connecting intrinsically motivated goal exploration process (IMGEP) and the maximum state entropy exploration (MSEE), showing that exploring least visited state can increase state distribution entropy mo... | SP:f3287af29c0148119a9b84df7681dbccb4884ef7 |
Abstractive Dialog Summarization with Semantic Scaffolds | 1 INTRODUCTION . Summarization aims to condense a piece of text to a shorter version , retaining the critical information . On dialogs , summarization has various promising applications in the real world . For instance , the automatic doctor-patient interaction summary can save doctors ’ massive amount of time used for... | Authors proposed an enhanced Pointer-Generator model called SPNet. The key difference between SPNet and PG are the separate handling or using of speaker role, semantic slot and domain labels. Authors also proposed a new metrics called Critical Information Completeness (CIC) to address ROUGE's weakness in assessing if k... | SP:33a9dffdcc2a5fc2a30a5a2e9b8cb65cd1010bed |
Abstractive Dialog Summarization with Semantic Scaffolds | 1 INTRODUCTION . Summarization aims to condense a piece of text to a shorter version , retaining the critical information . On dialogs , summarization has various promising applications in the real world . For instance , the automatic doctor-patient interaction summary can save doctors ’ massive amount of time used for... | The authors propose a new abstractive dialog summarization dataset and task based on the MultiWOZ dataset. Unlike previous work which targets very short descriptions of dialog transcripts (e.g. 'industrial designer presentation'), this paper looks to generate long descriptions of the entire dialog using the prompts in ... | SP:33a9dffdcc2a5fc2a30a5a2e9b8cb65cd1010bed |
On Concept-Based Explanations in Deep Neural Networks | 1 INTRODUCTION . Deep neural networks ( DNNs ) have shown great success in numerous tasks ( Goodfellow et al. , 2016 ) , from understanding images ( Zoph et al. , 2017 ) to answering questions ( Devlin et al. , 2018 ) . Yet , in many scenarios their lack of explainability serves as a bottleneck against their real-world... | The paper proposes metrics for evaluating concept based explanations in terms of ‘completeness’ -- characterized by (1) whether the set of presented concepts if sufficient to retain the predictive performance of the original model and (2) how is performance affected when all information useful to a complete set of conc... | SP:af31abf3d1d705bca5b2d35e7689d502c1520e99 |
On Concept-Based Explanations in Deep Neural Networks | 1 INTRODUCTION . Deep neural networks ( DNNs ) have shown great success in numerous tasks ( Goodfellow et al. , 2016 ) , from understanding images ( Zoph et al. , 2017 ) to answering questions ( Devlin et al. , 2018 ) . Yet , in many scenarios their lack of explainability serves as a bottleneck against their real-world... | The authors build on the work by Ghorbani et al. in concept-based interpretability methods by taking into account the "completeness" of the concepts. This basically tests whether the models accuracy holds if the input is projected onto the span of the discovered concepts. They propose "ConceptSHAP", based on Shapley va... | SP:af31abf3d1d705bca5b2d35e7689d502c1520e99 |
PROVABLY BENEFITS OF DEEP HIERARCHICAL RL | 1 INTRODUCTION . Reinforcement learning ( RL ) is a powerful tool to solve sequential decision making problems in various domains , including computer games ( Mnih et al. , 2013 ) , Go ( Silver et al. , 2016 ) , robotics ( Schulman et al. , 2015 ) . A particular feature in these successful applications of RL is that th... | This paper proposes a new kind of episodic finite MDPs called "deep hierarchical MDP" (hMDP). An L-layer hMDP can be *roughly* thought of as L episodic finite MDPs stacked together. A variant of UCRL2 [JOA10] is proposed to solve these hMDPs and some results from its regret analysis are provided. | SP:46d31f575928c68f60302520901feabe823e0dd4 |
PROVABLY BENEFITS OF DEEP HIERARCHICAL RL | 1 INTRODUCTION . Reinforcement learning ( RL ) is a powerful tool to solve sequential decision making problems in various domains , including computer games ( Mnih et al. , 2013 ) , Go ( Silver et al. , 2016 ) , robotics ( Schulman et al. , 2015 ) . A particular feature in these successful applications of RL is that th... | This paper studies the theoretical aspects of HRL. It provides theoretical analysis for the complexity of Deep HRL. The idea is to exploit a given action hierarchy, and known state decomposition, the fact that the high-level state space shares similar low-level structures. The final result is an exponential improvement... | SP:46d31f575928c68f60302520901feabe823e0dd4 |
Neural Machine Translation with Universal Visual Representation | 1 INTRODUCTION . Visual information has been introduced for neural machine translation in some previous studies ( NMT ) ( Specia et al. , 2016 ; Elliott et al. , 2017 ; Barrault et al. , 2018 ; Ive et al. , 2019 ) though the contribution of images is still an open question ( Elliott , 2018 ; Caglayan et al. , 2019 ) . ... | This paper uses visual representation learned over monolingual corpora with image annotations, which overcomes the lack of large-scale bilingual sentence-image pairs for multimodal NMT. Their approach enables visual information to be integrated into large-scale text-only NMT. Experiments on four widely used translation... | SP:55f583b190d59af8aaa7bda3c9e44bf5ed7ea96c |
Neural Machine Translation with Universal Visual Representation | 1 INTRODUCTION . Visual information has been introduced for neural machine translation in some previous studies ( NMT ) ( Specia et al. , 2016 ; Elliott et al. , 2017 ; Barrault et al. , 2018 ; Ive et al. , 2019 ) though the contribution of images is still an open question ( Elliott , 2018 ; Caglayan et al. , 2019 ) . ... | The authors propose to augment NMT with a grounded inventory of images. The intuition is clear and the premise is very tempting. The key architectural choice is to allow the transformer to use language embeddings to attend into a topic-image lookup table. The proportion is learned to balance how much signal comes fr... | SP:55f583b190d59af8aaa7bda3c9e44bf5ed7ea96c |
Learning Numeral Embedding | 1 INTRODUCTION . Word embeddings , the distributed vector representations of words , have become the essential building block for deep learning approaches to natural language processing ( NLP ) . The quality of pretrained word embeddings has been shown to significantly impact the performance of neural approaches to a v... | The paper talks about a recently highlighted problem in word embeddings which is their incapability to represent numerals, especially the out-of-vocabulary numerals. For addressing the problem, they propose a method that induces a finite set of prototype numerals using either self-organizing map or Gaussian Mixture mod... | SP:4c1ba325175a1a289d7467dc269d20eabc67383c |
Learning Numeral Embedding | 1 INTRODUCTION . Word embeddings , the distributed vector representations of words , have become the essential building block for deep learning approaches to natural language processing ( NLP ) . The quality of pretrained word embeddings has been shown to significantly impact the performance of neural approaches to a v... | The paper proposes a novel method for embedding numerals which can be learned by using neural word embedding learning techniques. The paper motivates the work by reviewing the difficulty of embedding components to represent numerals: OOV in most cases. Their main contribution is the introduction of a method composes nu... | SP:4c1ba325175a1a289d7467dc269d20eabc67383c |
Imitation Learning of Robot Policies using Language, Vision and Motion | 1 INTRODUCTION . A significant challenge when designing robots to operate in the real world lies in the generation of control policies that can adapt to changing environments . Programming such policies is a labor and time-consuming process which requires substantial technical expertise . Imitation learning ( Schaal , ... | This work uses imitations learning (from synthetic data) to train a deep model which takes a natural language instruction, and a visual representation of a robot's environment, and outputs a trajectory for the robot to follow which executes this instruction. The work focuses on a robotic pick-and-place task, where the... | SP:1db20b6170874b3c477e7429d5d6e853680b6e5b |
Imitation Learning of Robot Policies using Language, Vision and Motion | 1 INTRODUCTION . A significant challenge when designing robots to operate in the real world lies in the generation of control policies that can adapt to changing environments . Programming such policies is a labor and time-consuming process which requires substantial technical expertise . Imitation learning ( Schaal , ... | The paper addresses the problem of using multiple modalities for learning from demonstration. Approaches that take in task or joint space data to learn a policy for replicating that task are numerous. Doing the same with multiple modalities involved, in particular vision, language and motion, has only been recently con... | SP:1db20b6170874b3c477e7429d5d6e853680b6e5b |
A Bilingual Generative Transformer for Semantic Sentence Embedding | 1 INTRODUCTION . Learning useful representations of language has been a source of recent success in natural language processing ( NLP ) . Much work has been done on learning representations for words ( Mikolov et al. , 2013 ; Pennington et al. , 2014 ) and sentences ( Kiros et al. , 2015 ; Conneau et al. , 2017 ) . Mor... | This paper presents a bilingual generative model for sentence embedding based variational probabilistic framework. By separating a common latent variable from language-specific latent variables, the model is able to capture what's in common between parallel bilingual sentences and language-specific semantics. Experimen... | SP:f676894db5781369ec25d27ccf44e51c12d081ea |
A Bilingual Generative Transformer for Semantic Sentence Embedding | 1 INTRODUCTION . Learning useful representations of language has been a source of recent success in natural language processing ( NLP ) . Much work has been done on learning representations for words ( Mikolov et al. , 2013 ; Pennington et al. , 2014 ) and sentences ( Kiros et al. , 2015 ; Conneau et al. , 2017 ) . Mor... | This paper addresses the problem of constructing a sentence embedding using a generative transformer model which encodes semantic aspects and language-specific aspect separately. They use transformers to encode and decode sentence embedding, and the objective reconstructs input with a latent variables (language variabl... | SP:f676894db5781369ec25d27ccf44e51c12d081ea |
Unifying Question Answering, Text Classification, and Regression via Span Extraction | 1 INTRODUCTION . Pre-trained natural language processing ( NLP ) systems ( Radford et al. , 2019 ; Devlin et al. , 2018 ; Radford et al. , 2018 ; Howard & Ruder , 2018 ; Peters et al. , 2018 ; McCann et al. , 2017 ; Liu et al. , 2019b ) have been shown to transfer remarkably well on downstream tasks including text clas... | This paper asks whether it works to remove task-specific heads and treat classification and regression problems as span extraction, by formatting problems in such a way that a single span extraction model can be used. This is a reasonable question to ask, and the authors performed a very large number of experiments at... | SP:345e244321aa18121d73d55e9e572eb904f48e9e |
Unifying Question Answering, Text Classification, and Regression via Span Extraction | 1 INTRODUCTION . Pre-trained natural language processing ( NLP ) systems ( Radford et al. , 2019 ; Devlin et al. , 2018 ; Radford et al. , 2018 ; Howard & Ruder , 2018 ; Peters et al. , 2018 ; McCann et al. , 2017 ; Liu et al. , 2019b ) have been shown to transfer remarkably well on downstream tasks including text clas... | This paper introduces a method for converting sentence pair classification tasks and sentence regression tasks into span into span extraction tasks, by listing all the possible classes (entailment, contradiction, neural) or the discretized scores (0.0, 0.25 ...) and concatenating them with the source text. With this fo... | SP:345e244321aa18121d73d55e9e572eb904f48e9e |
Denoising Improves Latent Space Geometry in Text Autoencoders | 1 INTRODUCTION . Autoencoder based generative models have recently become popular tools for advancing controllable text generation such as style or sentiment transfer ( Bowman et al. , 2016 ; Hu et al. , 2017 ; Shen et al. , 2017 ; Zhao et al. , 2018 ) . By mapping sentences to vectors in the latent space , these model... | The paper "Denoising Improves Latent Space Geometry in Text Autoencoders" tackles the problem of text autoencoding in a space which respects text similarities. It is an interesting problem for which various attempts have been proposed, while still facing difficulties for encoding in smooth spaces. The paper proposes a ... | SP:891db9f5c3c7f981f4b9e37e36436a471f65117a |
Denoising Improves Latent Space Geometry in Text Autoencoders | 1 INTRODUCTION . Autoencoder based generative models have recently become popular tools for advancing controllable text generation such as style or sentiment transfer ( Bowman et al. , 2016 ; Hu et al. , 2017 ; Shen et al. , 2017 ; Zhao et al. , 2018 ) . By mapping sentences to vectors in the latent space , these model... | This paper presented a denoising adversarial autoencoder for sentence embeddings. The idea is that by introducing perturbations (word omissions, etc) the embeddings are more meaningful and less "memorized". Evaluations include measuring sentence perplexity in generation/reconstruction, tense changing via vector arithme... | SP:891db9f5c3c7f981f4b9e37e36436a471f65117a |
Learning Reusable Options for Multi-Task Reinforcement Learning | 1 INTRODUCTION . Reinforcement learning ( RL ) techniques have experienced much of their success in simulated environments , such as video games ( Mnih et al. , 2015 ) or board games ( Silver et al. , 2016 ; Tesauro , 1995 ) . One of the main reasons why RL has worked so well in these applications is that we are able s... | This paper proposes a new option discovery method for multi-task RL to reuse the option learned in previous tasks for better generalization. The authors utilize demonstrations collected beforehand and train an option learning framework offline by minimizing the expected number of terminations while encouraging diverse ... | SP:fe137babff80e9e5f48e44f36a86a71d095d6264 |
Learning Reusable Options for Multi-Task Reinforcement Learning | 1 INTRODUCTION . Reinforcement learning ( RL ) techniques have experienced much of their success in simulated environments , such as video games ( Mnih et al. , 2015 ) or board games ( Silver et al. , 2016 ; Tesauro , 1995 ) . One of the main reasons why RL has worked so well in these applications is that we are able s... | The authors propose to learn reusable options to make use of prior information and claim to do so with minimal information from the user (such as # of options needed to solve the task, which options etc). The claim is that the agent is first able to learn a near-optimal policy for a small # of problems and then is able... | SP:fe137babff80e9e5f48e44f36a86a71d095d6264 |
Enhancing Transformation-Based Defenses Against Adversarial Attacks with a Distribution Classifier | 1 INTRODUCTION . There has been widespread use of convolutional neural networks ( CNN ) in many critical real-life applications such as facial recognition ( Parkhi et al. , 2015 ) and self-driving cars ( Jung et al. , 2016 ) . However , it has been found that CNNs could misclassify the input image when the image has be... | This paper presents a novel defense method to make the classification more robust. The motivation is based on the observation: the distribution of the soft-max for the cleaned image and its transformed images for one class is similar to the distribution of the soft-max for the adversarial image and its transformed imag... | SP:ad9f47b416f144d43c2c4f66599529e1bb49bca6 |
Enhancing Transformation-Based Defenses Against Adversarial Attacks with a Distribution Classifier | 1 INTRODUCTION . There has been widespread use of convolutional neural networks ( CNN ) in many critical real-life applications such as facial recognition ( Parkhi et al. , 2015 ) and self-driving cars ( Jung et al. , 2016 ) . However , it has been found that CNNs could misclassify the input image when the image has be... | The authors analyze the use of image transformation as a defense against adversarial examples, where a challenge is to prevent the deterioration of performance on clean images. To do this, they show that the softmax distributions for clean and adversarial images share similar "features", and therefore one can apply a t... | SP:ad9f47b416f144d43c2c4f66599529e1bb49bca6 |
Unsupervised Progressive Learning and the STAM Architecture | 1 INTRODUCTION . We start by posing a challenging problem , referred to as Unsupervised Progressive Learning ( UPL ) ( see Figure 1 ) . In the UPL problem , the agent observes a sequence ( or stream ) of unlabeled data vectors { xt } t∈N with xt ∈ Rn . Each vector xt is associated with a class k ( xt ) and the vectors ... | This paper proposes a method called Self-Taught Associative Memory (STAM) for Unsupervised Progressive Learning (UPL) , i.e., learning salient representation from streams of mostly unlabeled data with occasional class labels, where the number of class increases over time. The motivation of this paper is quite interesti... | SP:beaa3dfef4bdf3d8fea64d4cf86911f45edd2873 |
Unsupervised Progressive Learning and the STAM Architecture | 1 INTRODUCTION . We start by posing a challenging problem , referred to as Unsupervised Progressive Learning ( UPL ) ( see Figure 1 ) . In the UPL problem , the agent observes a sequence ( or stream ) of unlabeled data vectors { xt } t∈N with xt ∈ Rn . Each vector xt is associated with a class k ( xt ) and the vectors ... | This paper sets up a new problem based on a continuous stream of potentially partially labelled data, which the authors call the Unsupervised Progressive Learning problem. The paper also introduces a new model designed to approach this problem, called the STAM architecture, with many concepts applied in a novel way. Th... | SP:beaa3dfef4bdf3d8fea64d4cf86911f45edd2873 |
Credible Sample Elicitation by Deep Learning, for Deep Learning | 1 INTRODUCTION . The availability of a large quantity of credible samples is crucial for building high-fidelity machine learning models . This is particularly true for deep learning systems that are data-hungry . Arguably , the most scalable way to collect a large amount of training samples is to crowdsource from a dec... | This paper proposes a sample elicitation framework to tackle the problem of eliciting credible samples from agents for complex distributions. The authors suggest that deep neural frameworks can be applied in this framework for sample elicitation through the derivations. The authors also show the connection between the ... | SP:3504773d062b05d1f7c358dfdc0da2ad78f5bc5e |
Credible Sample Elicitation by Deep Learning, for Deep Learning | 1 INTRODUCTION . The availability of a large quantity of credible samples is crucial for building high-fidelity machine learning models . This is particularly true for deep learning systems that are data-hungry . Arguably , the most scalable way to collect a large amount of training samples is to crowdsource from a dec... | This paper studies the sample elicitation problem where agents are asked to report samples. The goal is then to evaluate the quality of these reported samples by means of a scoring function S. Following previous related works, the authors use the equivalence between maximizing the expected proper score and minimizing s... | SP:3504773d062b05d1f7c358dfdc0da2ad78f5bc5e |
Detecting Change in Seasonal Pattern via Autoencoder and Temporal Regularization | 1 INTRODUCTION . Time series data are sequences of measurements over time describing the behavior of systems . Time series analysis has become increasingly important in monitoring systems health and performance . As the system behavior changes over time due to external events and/or internal modifications , the problem... | The paper raises an alarm that state-of-the art change-point detection methods in the ML literature do not handle important practical aspects arising in time-series modeling, namely seasonality. Indeed, methods designed to detect changing distribution under an i.i.d. setting can fail dramatically when the assumption is... | SP:c8a8e6b90a56186572e90e21755e73effbeaea14 |
Detecting Change in Seasonal Pattern via Autoencoder and Temporal Regularization | 1 INTRODUCTION . Time series data are sequences of measurements over time describing the behavior of systems . Time series analysis has become increasingly important in monitoring systems health and performance . As the system behavior changes over time due to external events and/or internal modifications , the problem... | This paper proposed a new model for change point detection, using autoencoders with temporal regularization, in order to impose temporal smoothness in the latent codes. To motivate this new model, the authors also provided a toy example to show how the abnormality in a time series is removed in the reconstructed signal... | SP:c8a8e6b90a56186572e90e21755e73effbeaea14 |
Adversarial Robustness as a Prior for Learned Representations | 1 INTRODUCTION . Beyond achieving remarkably high accuracy on a variety of tasks ( Krizhevsky et al. , 2012 ; He et al. , 2015 ; Collobert & Weston , 2008 ) , a major appeal of deep learning is the ability to learn effective representations of data . Specifically , deep neural networks can be thought of as linear class... | The paper presents a study about the representations learned by neural networks trained using robust optimization — a type of optimization that requires the model to be robust to small perturbations in the data. Specifically, the paper presents results of ResNet-50 trained on ImageNet with standard optimization and rob... | SP:33d9ed48ce72f65860ffb34e77ed8b79b95b4869 |
Adversarial Robustness as a Prior for Learned Representations | 1 INTRODUCTION . Beyond achieving remarkably high accuracy on a variety of tasks ( Krizhevsky et al. , 2012 ; He et al. , 2015 ; Collobert & Weston , 2008 ) , a major appeal of deep learning is the ability to learn effective representations of data . Specifically , deep neural networks can be thought of as linear class... | The paper shows that the learnt representations of robustly trained models align more closely with features that the human perceive as meaningful. They propose that robust optimization can be viewed as inducing a human prior over learnt features. Extensive experiments demonstrate that robust representations are approxi... | SP:33d9ed48ce72f65860ffb34e77ed8b79b95b4869 |
Large Batch Optimization for Deep Learning: Training BERT in 76 minutes | 1 INTRODUCTION . With the advent of large scale datasets , training large deep neural networks , even using computationally efficient optimization methods like Stochastic gradient descent ( SGD ) , has become particularly challenging . For instance , training state-of-the-art deep learning models like BERT and ResNet-5... | This paper developed a novel layerwise adaptation strategy, LAMB, that allows training BERT model with large mini-batches (32k vs baseline 512). This significantly speeds up the status quo in training BERT model, and effectively reduces the training time from original 3 days to only 76 minutes. In addition to demonstra... | SP:14dca47a505818106502978bddde5a7f294bfeeb |
Large Batch Optimization for Deep Learning: Training BERT in 76 minutes | 1 INTRODUCTION . With the advent of large scale datasets , training large deep neural networks , even using computationally efficient optimization methods like Stochastic gradient descent ( SGD ) , has become particularly challenging . For instance , training state-of-the-art deep learning models like BERT and ResNet-5... | This paper proposes a learning rate adaptation mechanism, called LAMB, for large-batch distributed training. The goal is to stabilize the training as the batch size increases. The idea is simple and straightforward -- there should be a layerwise learning rate adjusted by normalizing the layer weights and gradients at e... | SP:14dca47a505818106502978bddde5a7f294bfeeb |
Soft Token Matching for Interpretable Low-Resource Classification | 1 INTRODUCTION . Gathering and labeling data is a task that can be expensive in terms of time , human effort and resources . When we can not rely on already available datasets , training a model with acceptable performance on few data points annotated by few annotators , becomes critical in many practical applications ... | The authors propose PARCUS ("Pattern Representations on Continuous Spaces"), a model which computes a soft-matching probability for all words in an input sequence with so-called prototypes in order to predict a label for the input. Furthermore, for training, PARCUS makes use of rationales. Those are indicators of input... | SP:4f8f34e95732b3f87b23878289062d359cda110f |
Soft Token Matching for Interpretable Low-Resource Classification | 1 INTRODUCTION . Gathering and labeling data is a task that can be expensive in terms of time , human effort and resources . When we can not rely on already available datasets , training a model with acceptable performance on few data points annotated by few annotators , becomes critical in many practical applications ... | This paper considers the problem of text classification, especially the settings in which the number of labeled sentences is very small. However, authors assume, annotations of rationales behind the label, i.e. highlighting tokens in a sentence which are important in deciding its label. As per my understanding, this is... | SP:4f8f34e95732b3f87b23878289062d359cda110f |
Dual-module Inference for Efficient Recurrent Neural Networks | 1 INTRODUCTION . Recurrent Neural Networks ( RNNs ) play a critical role in many natural language processing ( NLP ) tasks , such as machine translation ( Bahdanau et al. , 2014 ; Wu et al. , 2016 ) , speech recognition ( Graves et al. , 2013 ; He et al. , 2019 ) , and speech synthesis ( Wang et al. , 2017 ) , owing to... | This paper attempts to compress the networks so as to accelerate the running procedure as well as save the storage. The authors propose a dual-module that is composed of a little module and big module. The big module use the full original data and parameters whereas the little module use small data and parameters by ra... | SP:bf0b8ec1ea69eb1b54e6502182b66ab3a8321c42 |
Dual-module Inference for Efficient Recurrent Neural Networks | 1 INTRODUCTION . Recurrent Neural Networks ( RNNs ) play a critical role in many natural language processing ( NLP ) tasks , such as machine translation ( Bahdanau et al. , 2014 ; Wu et al. , 2016 ) , speech recognition ( Graves et al. , 2013 ; He et al. , 2019 ) , and speech synthesis ( Wang et al. , 2017 ) , owing to... | This manuscript proposes an approach to reduce memory access and computation in Recurrent Neural Networks. Specifically, they train a second "little" neural network to approximate a pre-trained "big" network and use simple rules to switch between the little and the big network. The approach can provide some speedups ... | SP:bf0b8ec1ea69eb1b54e6502182b66ab3a8321c42 |
Scoring-Aggregating-Planning: Learning task-agnostic priors from interactions and sparse rewards for zero-shot generalization | 1 INTRODUCTION . While deep Reinforcement Learning ( RL ) methods have shown impressive performance on video games ( Mnih et al. , 2015 ) and robotics tasks ( Schulman et al. , 2015 ; Lillicrap et al. , 2015 ) , they solve each problem tabula rasa . Hence , it will be hard for them to generalize to new tasks without re... | The paper proposes a framework (Scoring-Aggregating-Planning (SAP)) for learning task-agnostic priors that allow generalization to new tasks without finetuning. The motivation for this is very clear - humans can perform much better than machines in zero-shot conditions because humans have learned priors about objects, ... | SP:b6ca7f80548c640f173512386883ce7e305dd96c |
Scoring-Aggregating-Planning: Learning task-agnostic priors from interactions and sparse rewards for zero-shot generalization | 1 INTRODUCTION . While deep Reinforcement Learning ( RL ) methods have shown impressive performance on video games ( Mnih et al. , 2015 ) and robotics tasks ( Schulman et al. , 2015 ; Lillicrap et al. , 2015 ) , they solve each problem tabula rasa . Hence , it will be hard for them to generalize to new tasks without re... | The paper describes a method that aims to learn task-agnostic priors for zero-shot generalization. The main idea is to employ the following modeling approach on top of the model-based RL framework: a local convolution network is used to compute a score for each local state action pair, and then another network is used ... | SP:b6ca7f80548c640f173512386883ce7e305dd96c |
Why Not to Use Zero Imputation? Correcting Sparsity Bias in Training Neural Networks | Handling missing data is one of the most fundamental problems in machine learning . Among many approaches , the simplest and most intuitive way is zero imputation , which treats the value of a missing entry simply as zero . However , many studies have experimentally confirmed that zero imputation results in suboptimal ... | This paper provides a novel solution to the variable sparsity problem, where the output of neural networks biased with respect to the number of missing inputs. The authors proposed a sparsity normalization algorithm to process the input vectors to encounter the bias. In experiments, the authors evaluated the proposed s... | SP:9fdc1a88425fd5d103163f7bbcbafc7ca7fe81be |
Why Not to Use Zero Imputation? Correcting Sparsity Bias in Training Neural Networks | Handling missing data is one of the most fundamental problems in machine learning . Among many approaches , the simplest and most intuitive way is zero imputation , which treats the value of a missing entry simply as zero . However , many studies have experimentally confirmed that zero imputation results in suboptimal ... | This paper studies a very interesting phenomena in machine learning called VSP, that is the output of the model is highly affected via the level of missing values in its input. The authors demonstrate the existence of such phenomena empirically, analyze the root cause for it theoretically, and propose a simple yet eff... | SP:9fdc1a88425fd5d103163f7bbcbafc7ca7fe81be |
LEARNING TO IMPUTE: A GENERAL FRAMEWORK FOR SEMI-SUPERVISED LEARNING | 1 INTRODUCTION . Semi-supervised learning ( SSL ) ( Chapelle et al. , 2009 ) is one of the approaches to learn not only from labeled samples but also unlabeled ones . Under certain assumptions such as presence of smooth prediction functions that map data to labels , of low-dimensional manifolds that the high-dimensiona... | This paper proposes a semi-supervised approach to impute the labels of unlabeled samples such that a network achieves better generalization when it is trained on these labels. The proposed strategy can be easily used to improve the state-of-the-art semi-supervised methods. It mainly uses a validation data set to evalua... | SP:79ddf8eda1c2247a1fc928cd7f4ca3d1d95b6adc |
LEARNING TO IMPUTE: A GENERAL FRAMEWORK FOR SEMI-SUPERVISED LEARNING | 1 INTRODUCTION . Semi-supervised learning ( SSL ) ( Chapelle et al. , 2009 ) is one of the approaches to learn not only from labeled samples but also unlabeled ones . Under certain assumptions such as presence of smooth prediction functions that map data to labels , of low-dimensional manifolds that the high-dimensiona... | This paper uses a meta-learning approach to solve semi-supervised learning. The main idea is to simulate an SGD step on the loss of the meta-validation data and see how the model will perform if the pseudo-labels of unlabelled data are perturbed. Experiments on classification and regression problems show that the propo... | SP:79ddf8eda1c2247a1fc928cd7f4ca3d1d95b6adc |
Revisiting Self-Training for Neural Sequence Generation | 1 INTRODUCTION . Deep neural networks often require large amounts of labeled data to achieve good performance . However , acquiring labels is a costly process , which motivates research on methods that can effectively utilize unlabeled data to improve performance . Towards this goal , semi-supervised learning ( Chapell... | The paper introduces an interesting study that tries to explain why conditional text generation models with autoregressive decoders benefit from self-training on pseudo labels created from the same model. The paper introduces and verifies two hypotheses: 1) Decoding strategy: Since beam search is a biased estimator sam... | SP:c57202d97644413a7a1586156e0ea2d88950cc80 |
Revisiting Self-Training for Neural Sequence Generation | 1 INTRODUCTION . Deep neural networks often require large amounts of labeled data to achieve good performance . However , acquiring labels is a costly process , which motivates research on methods that can effectively utilize unlabeled data to improve performance . Towards this goal , semi-supervised learning ( Chapell... | This paper presents a self-training approach for improving sequence-to-sequence tasks. As a preliminary experiment, this study randomly sampled 100k sentences from WMT 2014 English-German dataset (WMT100K, hereafter), trained a baseline (Transformer) model on WMT100K, and applied self-training methods on the remaining ... | SP:c57202d97644413a7a1586156e0ea2d88950cc80 |
Weakly Supervised Clustering by Exploiting Unique Class Count | A weakly supervised learning based clustering framework is proposed in this paper . As the core of this framework , we introduce a novel multiple instance learning task based on a bag level label called unique class count ( ucc ) , which is the number of unique classes among all instances inside the bag . In this task ... | This paper proposes a MIL clustering method. The proposed MIL setup is called "unique class count (ucc)", this is, for a bag os samples ucc is the number of clusters in the bag. The method learns the features of the samples using two losses: an autoender loss and the ucc loss. Once trained on a dataset the method can p... | SP:f70ab30f1b31fa2dcf450582b6798c4da8841687 |
Weakly Supervised Clustering by Exploiting Unique Class Count | A weakly supervised learning based clustering framework is proposed in this paper . As the core of this framework , we introduce a novel multiple instance learning task based on a bag level label called unique class count ( ucc ) , which is the number of unique classes among all instances inside the bag . In this task ... | This paper proposes a new type of weakly supervised clustering / multiple instance learning (MIL) problem in which bags of instances (data points) are labeled with a "unique class count (UCC)*, rather than any bag-level or instance-level labels. For example, a histopathology slide (the bag), consisting of many individ... | SP:f70ab30f1b31fa2dcf450582b6798c4da8841687 |
Closed loop deep Bayesian inversion: Uncertainty driven acquisition for fast MRI | This work proposes a closed loop , uncertainty-driven adaptive sampling framework ( CLUDAS ) for accelerating magnetic resonance imaging ( MRI ) via deep Bayesian inversion . By closed loop , we mean that our samples adapt in real-time to the incoming data . To our knowledge , we demonstrate the first generative advers... | The paper describes a method for accelerating MRI scans by proposing lines in k-space to acquire next. The proposals are based on posterior uncertainty estimates obtained from GAN-based reconstructions from parts of the k-space acquired thus far. The authors address an interesting and important problem of speeding up M... | SP:8a2a437441032f68341e305c568f59643a1c81e8 |
Closed loop deep Bayesian inversion: Uncertainty driven acquisition for fast MRI | This work proposes a closed loop , uncertainty-driven adaptive sampling framework ( CLUDAS ) for accelerating magnetic resonance imaging ( MRI ) via deep Bayesian inversion . By closed loop , we mean that our samples adapt in real-time to the incoming data . To our knowledge , we demonstrate the first generative advers... | The paper proposes an uncertainty driven acquisition for MRI reconstruction. Contrary to most previous approaches (which try to get best reconstruction for a fixed sampling pattern) the method incorporates an adaptive, on-the-fly masking building (which is similar in spirit to Zhang at al. 2019). The measurements to ac... | SP:8a2a437441032f68341e305c568f59643a1c81e8 |
Enhancing Language Emergence through Empathy | 1 INTRODUCTION . Natural language is not as rule-based as researchers in supervised language learning would prefer . There are limitless context-dependent notions to it , and flexible language use is considered as a necessary aspect of general AI . Originally , natural language emerged through a necessity to achieve su... | This paper takes the reference-game setup of Lazaridou et al. (2018), as a means of enabling emergent communication, and adds an auxiliary task to demonstrate that this helps with language emergence. The auxiliary task is to enable the speaker to predict the hidden state of the listener, after the message has been rece... | SP:0c4124acde5770c92f5afb3a0f12d2f70eead48d |
Enhancing Language Emergence through Empathy | 1 INTRODUCTION . Natural language is not as rule-based as researchers in supervised language learning would prefer . There are limitless context-dependent notions to it , and flexible language use is considered as a necessary aspect of general AI . Originally , natural language emerged through a necessity to achieve su... | This paper aims to take insight from human language acquisition and the importance of empathic connection to learn better models for emergent language. The authors propose an approach to introduce the notion of empathy to multi-agent deep RL by extending existing approaches on referential games with an auxiliary task f... | SP:0c4124acde5770c92f5afb3a0f12d2f70eead48d |
DYNAMIC SELF-TRAINING FRAMEWORK FOR GRAPH CONVOLUTIONAL NETWORKS | 1 INTRODUCTION . Graphs or networks can be used to model any interactions between entities such as social interactions ( Facebook , Twitter ) , biological networks ( protein-protein interaction ) , and citation networks . There has been an increasing research interest in deep learning on graph structured data , e.g. , ... | This paper propose to modify the existing work [1] of self-training framework for graph convolutional networks. It tracks three limitations of [1] and propose three use a threshold-based rule to insert new pseudo-labels and dynamic change the pseudo-label set. Moreover personalized weight are assigned to each activate... | SP:6bf1569771191ea913217f527173f454d50e266c |
DYNAMIC SELF-TRAINING FRAMEWORK FOR GRAPH CONVOLUTIONAL NETWORKS | 1 INTRODUCTION . Graphs or networks can be used to model any interactions between entities such as social interactions ( Facebook , Twitter ) , biological networks ( protein-protein interaction ) , and citation networks . There has been an increasing research interest in deep learning on graph structured data , e.g. , ... | This paper proposes a generalised self-training framework to build a Graph Neural Network to label graphs. Of importance is the dynamic nature of the self-training. The authors do not change the GCN but extend the self-training portion as per the prior GCN paper by introducing Dynamic Self-Training that keeps a confid... | SP:6bf1569771191ea913217f527173f454d50e266c |
A Novel Analysis Framework of Lower Complexity Bounds for Finite-Sum Optimization | √ κn ) log ( 1/ε ) ) iterations , where κ is the condition number of the objective function . This lower bound is tighter than previous results and perfectly matches the upper bound of the existing proximal incremental first-order oracle algorithm Point-SAGA . We develop a novel construction to show the above result , ... | The authors prove lower bounds on the number of queries required for optimizing sums of convex functions. They consider more powerful queries than the usual queries that provide function evaluation/gradient pairs for chosen summands. As was done in [1] (which is cited in the submission), in this work algorithms can a... | SP:bcf9ed060b00d47720785afbdfd540a4c98715d4 |
A Novel Analysis Framework of Lower Complexity Bounds for Finite-Sum Optimization | √ κn ) log ( 1/ε ) ) iterations , where κ is the condition number of the objective function . This lower bound is tighter than previous results and perfectly matches the upper bound of the existing proximal incremental first-order oracle algorithm Point-SAGA . We develop a novel construction to show the above result , ... | This paper proves a better complexity lower bound for stochastic PIFO optimizers on the problem of finite-sum minimization. The paper assumes that the objective function is the sum of n individual loss functions. It further assumes that (1) the optimizer initializes at a fixed point, and (2) at each iteration, it rando... | SP:bcf9ed060b00d47720785afbdfd540a4c98715d4 |
Generalized Natural Language Grounded Navigation via Environment-agnostic Multitask Learning | Recent research efforts enable study for natural language grounded navigation in photo-realistic environments , e.g. , following natural language instructions or dialog . However , existing methods tend to overfit training data in seen environments and fail to generalize well in previously unseen environments . In orde... | This paper addresses some challenges of following natural language instructions for navigating in visual environments. The main challenge in such tasks is the scarcity of available training data, which results in generalization problems where the agent has difficulty navigating in unseen environments. Therefore, the au... | SP:3e3e429ab3ba27875731c7cecf2d00bd959973b6 |
Generalized Natural Language Grounded Navigation via Environment-agnostic Multitask Learning | Recent research efforts enable study for natural language grounded navigation in photo-realistic environments , e.g. , following natural language instructions or dialog . However , existing methods tend to overfit training data in seen environments and fail to generalize well in previously unseen environments . In orde... | This paper aims to apply the model of Wang 2019 to the new NDH task of Thomason '19. Both of these datasets are built on the same room-to-room environment and both are for natural language instruction following. Thomason's work extends the R2R paradigm to include a dialogue history which is collapsed into a single in... | SP:3e3e429ab3ba27875731c7cecf2d00bd959973b6 |
Dual Graph Representation Learning | Graph representation learning embeds nodes in large graphs as low-dimensional vectors and benefit to many downstream applications . Most embedding frameworks , however , are inherently transductive and unable to generalize to unseen nodes or learn representations across different graphs . Inductive approaches , such as... | This paper proposed a dual graph representation method to learn the representation of nodes in a graph. In particular, it learns the embedding of paired nodes simultaneously for multiple times, and use the mean values as the final representation. The experimental result demonstrates some improvement over existing metho... | SP:a0ba8e10e93f74cf923317f94b7dcd7f880d04c3 |
Dual Graph Representation Learning | Graph representation learning embeds nodes in large graphs as low-dimensional vectors and benefit to many downstream applications . Most embedding frameworks , however , are inherently transductive and unable to generalize to unseen nodes or learn representations across different graphs . Inductive approaches , such as... | This paper extends GraphSAGE in several dimensions: 1) applying attention when aggregating neighbors (already used by GAT and many other approached); 2) Ensembling node embedding by applying DualENC multiple times on positive pairs selected by random walk (this is doing aggregation of neighborhood again); and 3) adding... | SP:a0ba8e10e93f74cf923317f94b7dcd7f880d04c3 |
Energy-based models for atomic-resolution protein conformations | 1 INTRODUCTION . Methods for the rational design of proteins make use of complex energy functions that approximate the physical forces that determine protein conformations ( Cornell et al. , 1995 ; Jorgensen et al. , 1996 ; MacKerell Jr et al. , 1998 ) , incorporating knowledge about statistical patterns in databases o... | The paper proposes an Energy-Based-Model (EBM) for scoring the possible configurations of amino acid side chain conformations in protein structures with known amino acid backbone structure. The energy of the side-chain conformation (the chi-angle) for a given amino acid in the structure is calculated as a function of a... | SP:6ebcd4fc6279bf7662a6691dae25f1bf4616432d |
Energy-based models for atomic-resolution protein conformations | 1 INTRODUCTION . Methods for the rational design of proteins make use of complex energy functions that approximate the physical forces that determine protein conformations ( Cornell et al. , 1995 ; Jorgensen et al. , 1996 ; MacKerell Jr et al. , 1998 ) , incorporating knowledge about statistical patterns in databases o... | The authors propose a predictive model based on the energy based model that uses a Transformer architecture for the energy function. It accepts as input an atom and its neighboring atoms and computes an energy for their configuration. The input features include representations of physical properties (atom identity, ato... | SP:6ebcd4fc6279bf7662a6691dae25f1bf4616432d |
Frequency-based Search-control in Dyna | 1 INTRODUCTION . Model-based reinforcement learning ( MBRL ) ( Lin , 1992 ; Sutton , 1991b ; Daw , 2012 ; Sutton & Barto , 2018 ) methods have successfully been applied to many benchmark domains ( Gu et al. , 2016 ; Ha , David and Schmidhuber , Jürgen , 2018 ; Kaiser et al. , 2020 ) . The Dyna architecture , introduce... | This paper basically built upon [1]. The authors propose to do sampling in the high-frequency domain to increase the sample efficiency. They first argue that the high-frequency part of the function is hard to approximate (i.e., needs more sample points) in section 3.1. They argue that the gradient and Hessian can be u... | SP:b71f0c38b5308ce902baeff4d457745c50034894 |
Frequency-based Search-control in Dyna | 1 INTRODUCTION . Model-based reinforcement learning ( MBRL ) ( Lin , 1992 ; Sutton , 1991b ; Daw , 2012 ; Sutton & Barto , 2018 ) methods have successfully been applied to many benchmark domains ( Gu et al. , 2016 ; Ha , David and Schmidhuber , Jürgen , 2018 ; Kaiser et al. , 2020 ) . The Dyna architecture , introduce... | This paper proposes a new way to select states from which do do transitions in dyna algorithm (which trains policy from model experience as if it was a real experience). It proposes to look for states where frequency of value function as a function of a real valued state is large, because these are the states where the... | SP:b71f0c38b5308ce902baeff4d457745c50034894 |
The Benefits of Over-parameterization at Initialization in Deep ReLU Networks | 1 INTRODUCTION . Deep rectifier ( ReLU ) networks are popular in deep learning due to their ease of training and stateof-the-art generalization . This success of deep rectifier networks can be partly attributed to good initialization strategies ( for example Glorot & Bengio ( 2010 ) ; He et al . ( 2015 ) ) . Essentiall... | This paper studies initialization techniques for deep ReLU networks from a theoretical standpoint and derives finite layer width concentration bounds to show that with the He initialization scheme, deep ReLU networks preserve the norm of the input sample during a forward pass and the norm of the gradient with respect t... | SP:7c36047790a8d3e229748fea4d9ff7572a97fd0a |
The Benefits of Over-parameterization at Initialization in Deep ReLU Networks | 1 INTRODUCTION . Deep rectifier ( ReLU ) networks are popular in deep learning due to their ease of training and stateof-the-art generalization . This success of deep rectifier networks can be partly attributed to good initialization strategies ( for example Glorot & Bengio ( 2010 ) ; He et al . ( 2015 ) ) . Essentiall... | This work considers random parameter initialization in neural networks (In particular the initialization presented in He et al.) and develops non-asymptotic bounds for the norms and gradients of neural networks during initialization. The authors show that the norms of the outputs and gradients (for gradients, under a d... | SP:7c36047790a8d3e229748fea4d9ff7572a97fd0a |
AdvCodec: Towards A Unified Framework for Adversarial Text Generation | 1 INTRODUCTION . Recent studies have demonstrated that deep neural networks ( DNNs ) are vulnerable to carefully crafted adversarial examples ( Goodfellow et al. , 2015 ; Papernot et al. , 2016 ; Eykholt et al. , 2017 ; Moosavi-Dezfooli et al. , 2016 ) . While there are a lot of successful attacks proposed in the conti... | This paper proposes a new attack framework AdvCodec for adversarial text generation. The main idea is to use a tree-based autoencoder to embed text data into the continuous vector space and then optimize to find the adversarial perturbation in the vector space. The authors consider two types of attacks: concat attack a... | SP:2034dffa26a8e68c466d835ada625fe635a71b66 |
AdvCodec: Towards A Unified Framework for Adversarial Text Generation | 1 INTRODUCTION . Recent studies have demonstrated that deep neural networks ( DNNs ) are vulnerable to carefully crafted adversarial examples ( Goodfellow et al. , 2015 ; Papernot et al. , 2016 ; Eykholt et al. , 2017 ; Moosavi-Dezfooli et al. , 2016 ) . While there are a lot of successful attacks proposed in the conti... | Motivated by recent development of attack/defense methods addressing the vulnerability of deep CNN classifiers for images, this paper proposes an attack framework for adversarial text generation, in which an autoencoder is employed to map discrete text to a high-dimensional continuous latent space, standard iterative o... | SP:2034dffa26a8e68c466d835ada625fe635a71b66 |
SesameBERT: Attention for Anywhere | 1 INTRODUCTION . In recent years , unsupervised pretrained models have dominated the field of natural language processing ( NLP ) . The construction of a framework for such a model involves two steps : pretraining and fine-tuning . During pretraining , an encoder neural network model is trained using large-scale unlabe... | The paper proposes adding two mechanisms to the BERT architecture for NLU. The first is based on integrating information from all layers of the encoder via a method called Squeeze and Excitation. The second uses Gaussian blurring to encourage information sharing among neighboring words. The proposed method improves mod... | SP:b5433e6f4dc436a4a15554124a790aa794d5dc0d |
SesameBERT: Attention for Anywhere | 1 INTRODUCTION . In recent years , unsupervised pretrained models have dominated the field of natural language processing ( NLP ) . The construction of a framework for such a model involves two steps : pretraining and fine-tuning . During pretraining , an encoder neural network model is trained using large-scale unlabe... | This paper proposes a novel BERT based neural architecture, SESAME-BERT, which consists of “Squeeze and Excitation” method and Gaussian blurring. “Squeeze and Excitation” method extracts features from BERT by calculating a weighted sum of layers in BERT to feed the feature vectors to a downstream classifier. To capture... | SP:b5433e6f4dc436a4a15554124a790aa794d5dc0d |
Filter redistribution templates for iteration-lessconvolutional model reduction | 1 INTRODUCTION . Convolutional neural networks are built by stacking layers of neurons following the principle explained by Fukushima ’ s Neocognitron model ( Fukushima , 1980 ) . The Neocognitron design made neural networks invariant to shift in feature locations by arranging cells locally connected in a hierarchical ... | The search/design space of neural network architectures is vast, so researchers tend to use simple heuristics to guide their designs; alternatively neural architecture search methods may minimise heuristics in order to remove bias within the search. The authors propose applying a few simple heuristics in the form of "t... | SP:bfdade05a90180a4b30d9593598b9d7c2a9e0533 |
Filter redistribution templates for iteration-lessconvolutional model reduction | 1 INTRODUCTION . Convolutional neural networks are built by stacking layers of neurons following the principle explained by Fukushima ’ s Neocognitron model ( Fukushima , 1980 ) . The Neocognitron design made neural networks invariant to shift in feature locations by arranging cells locally connected in a hierarchical ... | This paper presents a simple methodological study on the effect of the distribution of convolutional filters on the accuracy of deep convolutional networks on the CIFAR 10 and CIFAR 100 data sets. There are five different kind of distributions studied: constant number of filters, monotonically increasing and decreasin... | SP:bfdade05a90180a4b30d9593598b9d7c2a9e0533 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.