paper_name stringlengths 11 170 | text stringlengths 8.07k 307k | summary stringlengths 152 6.16k | paper_id stringlengths 43 43 |
|---|---|---|---|
Deformable Kernels: Adapting Effective Receptive Fields for Object Deformation | 1 INTRODUCTION . The rich diversity of object appearance in images arises from variations in object semantics and deformation . Semantics describe the high-level abstraction of what we perceive , and deformation defines the geometric transformation tied to specific data ( Gibson , 1950 ) . Humans are remarkably adept a... | This work presents the idea of deformable kernels (DKs). As opposed to rigid kernels in standard convolutional networks, DKs allow each of their grid locations to be moved around in a larger kernel field. The offset by which a DK grid cell is moved is computed conditioned on the input to the network. To motivate the id... | SP:47bcfdf057ce7e9fc379ea2e58b3c5e00b2b61a1 |
Deformable Kernels: Adapting Effective Receptive Fields for Object Deformation | 1 INTRODUCTION . The rich diversity of object appearance in images arises from variations in object semantics and deformation . Semantics describe the high-level abstraction of what we perceive , and deformation defines the geometric transformation tied to specific data ( Gibson , 1950 ) . Humans are remarkably adept a... | This paper introduce a simple algorithm called deformable kernels. It learns to generate a collection of coordinate offset Δk for each of the convolutional kernel element. Then during convolution, the kernel is treated as a 2D regular grid and sampled (interpolated) according to the generated coordinate offset before a... | SP:47bcfdf057ce7e9fc379ea2e58b3c5e00b2b61a1 |
Enhancing Adversarial Defense by k-Winners-Take-All | 1 INTRODUCTION . In the tremendous success of deep learning techniques , there is a grain of salt . It has become wellknown that deep neural networks can be easily fooled by adversarial examples ( Szegedy et al. , 2014 ) . Those deliberately crafted input samples can mislead the networks to produce an output drasticall... | The authors propose using k-winner take all (k-WTA) activation functions to prevent white box adversarial attacks. A k-WTA activation functions outputs the k highest activations in a layer while setting all other activations to zero. The reasoning given by the authors is that k-WTA activation functions have many discon... | SP:1442b188a5bac1931c3023a5529fa366f22bd8e6 |
Enhancing Adversarial Defense by k-Winners-Take-All | 1 INTRODUCTION . In the tremendous success of deep learning techniques , there is a grain of salt . It has become wellknown that deep neural networks can be easily fooled by adversarial examples ( Szegedy et al. , 2014 ) . Those deliberately crafted input samples can mislead the networks to produce an output drasticall... | This paper addresses the important question of improving the robustness of deep neural networks against adversarial attacks. The authors propose a surprisingly simple measure to improve adversarial robustness, namely replacing typical activation functions such as ReLU with a k-winners-take-all (k-WTA) functions, whereb... | SP:1442b188a5bac1931c3023a5529fa366f22bd8e6 |
Efficient meta reinforcement learning via meta goal generation | Meta reinforcement learning ( meta-RL ) is able to accelerate the acquisition of new tasks by learning from past experience . Current meta-RL methods usually learn to adapt to new tasks by directly optimizing the parameters of policies over primitive actions . However , for complex tasks which requires sophisticated co... | This paper studies the problem of leveraging past experience to quickly solve new control tasks. The starting point (and perhaps the main contribution) is the observation that some tasks have similar high-level goals, while differing in how those goals are achieved. To that end, the paper introduces an meta-RL algorith... | SP:7dd326afe8e4e148955d98fb30d561b9e6be5ba9 |
Efficient meta reinforcement learning via meta goal generation | Meta reinforcement learning ( meta-RL ) is able to accelerate the acquisition of new tasks by learning from past experience . Current meta-RL methods usually learn to adapt to new tasks by directly optimizing the parameters of policies over primitive actions . However , for complex tasks which requires sophisticated co... | In this paper, the authors focus on the problem of meta-reinforcement learning (meta-RL). Specifically, the authors consider the setting of meta-RL for goal reaching tasks where each task corresponds to an unknown goal. Existing meta-RL algorithms directly train for a policy that output low level actions, which might b... | SP:7dd326afe8e4e148955d98fb30d561b9e6be5ba9 |
Extreme Triplet Learning: Effectively Optimizing Easy Positives and Hard Negatives | 1 INTRODUCTION . Deep metric learning optimizes an embedding function that maps semantically similar images to relatively nearby locations and maps semantically dissimilar images to distant locations . A number of approaches have been proposed for this problem ( Schroff et al. , 2015a ; Sohn , 2016 ; MovshovitzAttias e... | The authors propose a new loss function as well as an adjoining visualization for improved performance of hard negative / easy positive mining for deep triplet metric learning. The authors note that under the NCA loss, if one selects an easy positive / hard negative and computes the gradient with respect to this exampl... | SP:b46afcbaf09319053c4ae21b7ee68a34c78dc28f |
Extreme Triplet Learning: Effectively Optimizing Easy Positives and Hard Negatives | 1 INTRODUCTION . Deep metric learning optimizes an embedding function that maps semantically similar images to relatively nearby locations and maps semantically dissimilar images to distant locations . A number of approaches have been proposed for this problem ( Schroff et al. , 2015a ; Sohn , 2016 ; MovshovitzAttias e... | This paper uses the triplet scatter plot as a way to describe triplet selection strategies. The authors explain previously observed bad behavior for hard-negative triplet mining showing that it tends to make all points close to each other. The authors propose a simple modification to the desired gradients and derive a ... | SP:b46afcbaf09319053c4ae21b7ee68a34c78dc28f |
Effects of Linguistic Labels on Learned Visual Representations in Convolutional Neural Networks: Labels matter! | 1 INTRODUCTION . A critical distinction between human category learning and machine category learning is that only humans have a language . A language means that human learning is not limited to a one-to-one correspondence between a visual input and a category label . Indeed , the users of a language are known to activ... | This paper assesses the effects of training an image classifier with different label types: 1-hot coarse-grained labels (10 classes), 1-hot fine grained labels (30 labels which are all subcategories of the 10 coarse-grained categories), word vector representations of the 30 fine-grained labels. They also compare the re... | SP:2fa9b2601acf885062d3c9d158f6518a9213f398 |
Effects of Linguistic Labels on Learned Visual Representations in Convolutional Neural Networks: Labels matter! | 1 INTRODUCTION . A critical distinction between human category learning and machine category learning is that only humans have a language . A language means that human learning is not limited to a one-to-one correspondence between a visual input and a category label . Indeed , the users of a language are known to activ... | This paper demonstrates the importance of labels at various levels (no label, basic level label, and superordinate level) as well as in combination to determine the importance of semantic information in classification problems. They train an identical CNN architecture either as an autoencoder (no labels), with the basi... | SP:2fa9b2601acf885062d3c9d158f6518a9213f398 |
ProtoAttend: Attention-Based Prototypical Learning | 1 INTRODUCTION . Deep neural networks have been pushing the frontiers of artificial intelligence ( AI ) by yielding excellent performance in numerous tasks , from understanding images ( He et al. , 2016 ) to text ( Conneau et al. , 2016 ) . Yet , high performance is not always a sufficient factor - as some realworld de... | The aim of this work is to make deep learning classifiers more interpretable by "projecting" each input sample into a small collection of prototype examples (with some weighting over those) and then basing the decision on a combination of the latent representations of the chosen prototypes. In this way, the chosen cate... | SP:164da37f418e9f6ce0470a329ed41e35f9ac1260 |
ProtoAttend: Attention-Based Prototypical Learning | 1 INTRODUCTION . Deep neural networks have been pushing the frontiers of artificial intelligence ( AI ) by yielding excellent performance in numerous tasks , from understanding images ( He et al. , 2016 ) to text ( Conneau et al. , 2016 ) . Yet , high performance is not always a sufficient factor - as some realworld de... | This paper presents a sample-based self-explaining method for image classification. The basic idea is adopt the attention mechanism to learn the relation between the latent representation of the query sample and training samples, and identify the training samples with higher similarity as the prototype. The classificat... | SP:164da37f418e9f6ce0470a329ed41e35f9ac1260 |
Adaptive Adversarial Imitation Learning | 1 INTRODUCTION . Humans and animals can learn complex behaviors via imitation . Inspired by these learning mechanisms , Imitation Learning ( IL ) has long been a popular method for training autonomous agents from human-provided demonstrations . However , human and animal imitation differs markedly from commonly used ap... | The submission considers the problem of imitation learning when the dynamics of the expert are not known to the agent and the dynamics of the agent may change frequently. It is however assumed that the agent has access to a parameterized simulator that can simulate the expert dynamics. The parameters for the simulator ... | SP:e7d072333891bebe16584ee8276b874cb28fffda |
Adaptive Adversarial Imitation Learning | 1 INTRODUCTION . Humans and animals can learn complex behaviors via imitation . Inspired by these learning mechanisms , Imitation Learning ( IL ) has long been a popular method for training autonomous agents from human-provided demonstrations . However , human and animal imitation differs markedly from commonly used ap... | This paper proposes an algorithm for imitation of expert demonstrations, in situations where the imitator is acting under a different environment (different dynamics, for instance) than the one used to collect expert demonstrations. The algorithm builds on GAIL with the following modifications – the discriminator is ma... | SP:e7d072333891bebe16584ee8276b874cb28fffda |
Mixup Inference: Better Exploiting Mixup to Defend Adversarial Attacks | 1 INTRODUCTION . Deep neural networks ( DNNs ) have achieved state-of-the-art performance on various tasks ( Goodfellow et al. , 2016 ) . However , counter-intuitive adversarial examples generally exist in different domains , including computer vision ( Szegedy et al. , 2014 ) , natural language processing ( Jin et al.... | This paper proposes a novel use of mixup, which is originally a data augmentation method incorporating two training samples and their corresponding labels. The authors utilize mixup not for training but for inference (MI; Mixup Inference). Experimental results on Cifar 10, and Cifar 100 show that MI can boost the class... | SP:0c6c9db564f0029c12c1a1e16373970eeeb800d4 |
Mixup Inference: Better Exploiting Mixup to Defend Adversarial Attacks | 1 INTRODUCTION . Deep neural networks ( DNNs ) have achieved state-of-the-art performance on various tasks ( Goodfellow et al. , 2016 ) . However , counter-intuitive adversarial examples generally exist in different domains , including computer vision ( Szegedy et al. , 2014 ) , natural language processing ( Jin et al.... | This paper introduces a novel method for an adversarial attack named mixup inference (MI). Most of the work focuses on embedding mixup mechanism in the training phase, but MI uses the mixup in the inference phase. MI method has two main effects for the adversarial attack: one is perturbation shrinkage, and the other o... | SP:0c6c9db564f0029c12c1a1e16373970eeeb800d4 |
Neural Oblivious Decision Ensembles for Deep Learning on Tabular Data | 1 INTRODUCTION . The recent rise of deep neural networks ( DNN ) resulted in a substantial breakthrough for a large number of machine learning tasks in computer vision , natural language processing , speech recognition , reinforcement learning ( Goodfellow et al. , 2016 ) . Both gradient-based optimization via backprop... | The paper tries to ask if there is a good neural net architecture that works as effectively as gradient boosting decision trees on tabular data. The authors propose an architecture (NODE) that satisfies this conditions. NODE is an architecture consisting of differentiable oblivious decision trees that can be trained en... | SP:7f3dfc4a045d780299123b42cc712b3d7171e8eb |
Neural Oblivious Decision Ensembles for Deep Learning on Tabular Data | 1 INTRODUCTION . The recent rise of deep neural networks ( DNN ) resulted in a substantial breakthrough for a large number of machine learning tasks in computer vision , natural language processing , speech recognition , reinforcement learning ( Goodfellow et al. , 2016 ) . Both gradient-based optimization via backprop... | This paper introduces a new method to make ensembles of decision trees differentiable, and trainable with (stochastic) gradient descent. The proposed technique relies on the concept of "oblivious decision trees", which are a kind of decision trees that use the same classifier (i.e. a feature and threshold) for all the ... | SP:7f3dfc4a045d780299123b42cc712b3d7171e8eb |
Exploring Model-based Planning with Policy Networks | 1 INTRODUCTION . A model-based reinforcement learning ( MBRL ) agent learns its internal model of the world , i.e . the dynamics , from repeated interactions with the environment . With the learnt dynamics , a MBRL agent can for example perform online planning , interact with imaginary data , or optimize the controller... | This work provides a novel model-based reinforcement learning algorithm for continuous domains (Mujoco) dubbed POPLIN. The presented algorithm is similar in vein to the state-of-the-art PETS algorithm, a planning algorithm that uses state-unconditioned action proposal distributions to identify good action sequences wit... | SP:1d0977845884e1768b8a853e0c13fa71619f8164 |
Exploring Model-based Planning with Policy Networks | 1 INTRODUCTION . A model-based reinforcement learning ( MBRL ) agent learns its internal model of the world , i.e . the dynamics , from repeated interactions with the environment . With the learnt dynamics , a MBRL agent can for example perform online planning , interact with imaginary data , or optimize the controller... | This paper presents POPLIN, a novel model-based reinforcement learning algorithm, which trains a policy network to improve model-prediction control. The paper studies extensively how to utilize the policy, by planning in action space or planning in parameter space and how to train the policy, by behavioral cloning, by ... | SP:1d0977845884e1768b8a853e0c13fa71619f8164 |
Multi-agent Reinforcement Learning for Networked System Control | This paper considers multi-agent reinforcement learning ( MARL ) in networked system control . Specifically , each agent learns a decentralized control policy based on local observations and messages from connected neighbors . We formulate such a networked MARL ( NMARL ) problem as a spatiotemporal Markov decision proc... | This paper is concerned with network multi-agent RL (N-MARL), where agents need to update their policy based on messages obtained only from neighboring nodes. This is done under sensible restrictions on the state transition distribution, which can be claimed to hold true in realistic networked settings. The authors arg... | SP:28c833ad9939bcc4e355254536b610da50731d76 |
Multi-agent Reinforcement Learning for Networked System Control | This paper considers multi-agent reinforcement learning ( MARL ) in networked system control . Specifically , each agent learns a decentralized control policy based on local observations and messages from connected neighbors . We formulate such a networked MARL ( NMARL ) problem as a spatiotemporal Markov decision proc... | The authors use decentralized MARL for networked system control. Each agent might control a traffic light (exp 1) or a car in traffic (exp 2). Some features of their approach are a spatial Markov assumption (only neighborhood matters), a spatial discount factor, and NeurComm: a general message passing scheme between ag... | SP:28c833ad9939bcc4e355254536b610da50731d76 |
Intrinsic Motivation for Encouraging Synergistic Behavior | 1 INTRODUCTION . Consider a multi-agent environment such as a team of robots working together to play soccer . It is critical for a joint policy within such an environment to produce synergistic behavior , allowing multiple agents to work together to achieve a goal which they could not achieve individually . How should... | The paper focuses on using intrinsic motivation to improve the exploration process of reinforcement learning agents in tasks with sparse-reward and that require multi-agent to achieve. The authors proposed to encourage the agents toward the actions which changed the world in the ways that "would not be achieved if the ... | SP:bd79f443ec2da0a34a77be823acfc81ba45d8a18 |
Intrinsic Motivation for Encouraging Synergistic Behavior | 1 INTRODUCTION . Consider a multi-agent environment such as a team of robots working together to play soccer . It is critical for a joint policy within such an environment to produce synergistic behavior , allowing multiple agents to work together to achieve a goal which they could not achieve individually . How should... | The paper proposes a novel algorithm for encouraging synergistic behavior in multi-agent setups with an intrinsic reward that promotes the agents to work together to achieve states that they cannot achieve individually without cooperation. The paper focuses on a two-agent environment where an approximate forward dynami... | SP:bd79f443ec2da0a34a77be823acfc81ba45d8a18 |
R-TRANSFORMER: RECURRENT NEURAL NETWORK ENHANCED TRANSFORMER | 1 INTRODUCTION . Recurrent Neural Networks ( RNNs ) especially its variants such as Long Short-Term Memory ( LSTM ) and Gated Recurrent Unit ( GRU ) have achieved great success in a wide range of sequence learning tasks including language modeling , speech recognition , recommendation , etc ( Mikolov et al. , 2010 ; Su... | The paper introduces the R-Transformer architecture which adds a local RNN layer before each attention layer in Transformer. The authors claim state-of-the-art performance but only test on tiny tasks where Transformer models have not been heavily optimized and omit the main problem with RNNs - namely their speed. It is... | SP:71d504ec722cacab616fca85dd2937b93e71caaf |
R-TRANSFORMER: RECURRENT NEURAL NETWORK ENHANCED TRANSFORMER | 1 INTRODUCTION . Recurrent Neural Networks ( RNNs ) especially its variants such as Long Short-Term Memory ( LSTM ) and Gated Recurrent Unit ( GRU ) have achieved great success in a wide range of sequence learning tasks including language modeling , speech recognition , recommendation , etc ( Mikolov et al. , 2010 ; Su... | This paper proposes a new architecture, R-Transformer, that blends the Transformer networks and the recurrent networks, so as to better capture both the long- and short-term features. By injecting a local RNN layer at every level of the network, the authors hoped to enhance the Transformer's ability to model locality s... | SP:71d504ec722cacab616fca85dd2937b93e71caaf |
A Data-Efficient Mutual Information Neural Estimator for Statistical Dependency Testing | 1 INTRODUCTION . Mutual Information ( MI ) is an important , theoretically grounded measure of similarity between random variables . MI captures general , non-linear , statistical dependencies between random variables . MI estimators that estimate MI from samples are important tools widely used in not only subjects suc... | The paper proposes a neural-network-based estimation of mutual information, following the earlier line of work in [A]. The main focus has been to develop an estimator that can reliably work with small dataset sizes. They first reduce the sample complexity of estimating mutual information by decoupling the network learn... | SP:e020226557ee78b133893665ef4f30c9cb81ff9f |
A Data-Efficient Mutual Information Neural Estimator for Statistical Dependency Testing | 1 INTRODUCTION . Mutual Information ( MI ) is an important , theoretically grounded measure of similarity between random variables . MI captures general , non-linear , statistical dependencies between random variables . MI estimators that estimate MI from samples are important tools widely used in not only subjects suc... | This manuscript studies mutual-information estimation, in particular variational lower bounds, and focuses on reducing their sample complexity. The first contribution is based on adapting the MINE energy-based MI estimator family to out-of-sample testing. MINE involves fitting a very flexible parametric form of the dis... | SP:e020226557ee78b133893665ef4f30c9cb81ff9f |
Variational Hyper RNN for Sequence Modeling | In this work , we propose a novel probabilistic sequence model that excels at capturing high variability in time series data , both across sequences and within an individual sequence . Our method uses temporal latent variables to capture information about the underlying data pattern and dynamically decodes the latent i... | This paper proposes a variational hyper recurrent neural network which is a combination of the variational RNN and the hypernetwork. The hypernetwork is an RNN whose output modifies the parameters of the variational RNN dynamically at runtime. Overall, this seems like an extension of the idea of using a hypernetwork wi... | SP:eb9803ef7698cade762d39290f842a7b3bf897d0 |
Variational Hyper RNN for Sequence Modeling | In this work , we propose a novel probabilistic sequence model that excels at capturing high variability in time series data , both across sequences and within an individual sequence . Our method uses temporal latent variables to capture information about the underlying data pattern and dynamically decodes the latent i... | In this paper the authors propose an architecture based on variational autoencoders and hyper-networks. The basic idea is that the weights of the underlying RNN/autoencoder are not fixed, but are coming from another RNN/feed-forward network which captures the underlying dynamics and adjusts the weights accordingly. The... | SP:eb9803ef7698cade762d39290f842a7b3bf897d0 |
Frequency Principle: Fourier Analysis Sheds Light on Deep Neural Networks | 1 INTRODUCTION . Understanding the training process of Deep Neural Networks ( DNNs ) is a fundamental problem in the area of deep learning . We find a common behavior of the gradient-based training process of DNNs , that is , a Frequency Principle ( F-Principle ) : DNNs often fit target functions from low to high frequ... | This paper proposes to analyze the loss of neural networks in the Fourier domain. Since this is computationally expensive for larger-dimensional datasets, the analysis instead first projects the data onto the principal component of the data, and then using a Gaussian kernel estimation (which has nice properties in the ... | SP:74aafec80535022cbdd83067763fd7bced294ace |
Frequency Principle: Fourier Analysis Sheds Light on Deep Neural Networks | 1 INTRODUCTION . Understanding the training process of Deep Neural Networks ( DNNs ) is a fundamental problem in the area of deep learning . We find a common behavior of the gradient-based training process of DNNs , that is , a Frequency Principle ( F-Principle ) : DNNs often fit target functions from low to high frequ... | The paper studies the training process of NNs through the lens of Fourier analysis. The authors argue that during the training process, NNs will first learn low frequencies part of the function first and then the high frequency part. To verify this claim empirically, the author propose two methods: 1. examine the conve... | SP:74aafec80535022cbdd83067763fd7bced294ace |
Consistency-Based Semi-Supervised Active Learning: Towards Minimizing Labeling Budget | Active learning ( AL ) integrates data labeling and model training to minimize the labeling cost by prioritizing the selection of high value data that can best improve model performance . Readily-available unlabeled data are used for selection mechanisms , but are not used for model training in most conventional pool-b... | This paper proposes a semi-supervised active learning method to reduce the labeling cost. In the proposed method, a selection criterion to better integrate AL selection mechanism in SSL training framework is designed. The simple metric that aims to measure the inconsistency across a certain number of meaningful perturb... | SP:b2a573333b5b1b89b68f307c2b5de571fc84a481 |
Consistency-Based Semi-Supervised Active Learning: Towards Minimizing Labeling Budget | Active learning ( AL ) integrates data labeling and model training to minimize the labeling cost by prioritizing the selection of high value data that can best improve model performance . Readily-available unlabeled data are used for selection mechanisms , but are not used for model training in most conventional pool-b... | This paper proposes a new combination method for active learning and semi-supervised learning, where the objective is to make predictions that are robust to perturbations (for SSL) and select points for labeling with labels that differ under perturbations. This technique achieves 2x label efficiency over SSL with unifo... | SP:b2a573333b5b1b89b68f307c2b5de571fc84a481 |
LambdaNet: Probabilistic Type Inference using Graph Neural Networks | As gradual typing becomes increasingly popular in languages like Python and TypeScript , there is a growing need to infer type annotations automatically . While type annotations help with tasks like code completion and static error catching , these annotations can not be fully determined by compilers and are tedious to... | This paper proposed to use Graph Neural Networks (GNN) to do type inference for dynamically typed languages. The key technique is to construct a type dependency graph and infer the type on top of it. The type dependency graph contains edges specifying hard constraints derived from the static analysis, as well as soft r... | SP:b4c82616d2410a07ecce89da0e5dc9428f9209ae |
LambdaNet: Probabilistic Type Inference using Graph Neural Networks | As gradual typing becomes increasingly popular in languages like Python and TypeScript , there is a growing need to infer type annotations automatically . While type annotations help with tasks like code completion and static error catching , these annotations can not be fully determined by compilers and are tedious to... | A method to predict likely type of program variables in TypeScript is presented. It consists of a translation of a program's type constraints and defined objects into a (hyper)graph, and a specialised neural message passing architecture to learn from the generated graphs. Experiments show that the method substantially ... | SP:b4c82616d2410a07ecce89da0e5dc9428f9209ae |
Are Transformers universal approximators of sequence-to-sequence functions? | 1 INTRODUCTION . Self-attention based Transformer networks ( Vaswani et al. , 2017 ) have been at the center of the recent progress on various natural language processing ( NLP ) tasks , including machine translation ( Vaswani et al. , 2017 ) , language modeling ( Radford et al. , 2018 ; 2019 ) , and question answering... | C1. Transformers (without positional encodings and without layer normalization), with 2 attention heads of dimension 1 and feed-forward layers (FFN) with 4 hidden nodes, are universal approximators of continuous permutation-equivariant functions f of compact support, relative to any Lp metric (1 <= p < \infty). (Thm. 2... | SP:3489d6d9dde3bec6f8d50f309d28572c393eac61 |
Are Transformers universal approximators of sequence-to-sequence functions? | 1 INTRODUCTION . Self-attention based Transformer networks ( Vaswani et al. , 2017 ) have been at the center of the recent progress on various natural language processing ( NLP ) tasks , including machine translation ( Vaswani et al. , 2017 ) , language modeling ( Radford et al. , 2018 ; 2019 ) , and question answering... | This paper tries to analyse the Transformer, widely applied building block of a neural network component, to improve understanding of the internals of the model. The analysis starts showing that the transformer blocks generate permutation equivalent maps and then shows that the transformer can approximate any permutati... | SP:3489d6d9dde3bec6f8d50f309d28572c393eac61 |
Imagining the Latent Space of a Variational Auto-Encoders | 1 INTRODUCTION . Variational auto-encoders ( VAEs ) have made a significant impact since their introduction by Kingma and Welling ( 2014 ) . However , one of their perceived problems is their reconstruction performance . This has spawned a wave of research into trying to improve the reconstruction performance ( Zhao et... | The paper proposes a new method for improving generative properties of VAE model. The idea is to train VAE in two stages: at first, train the vanilla VAE, then at the second stage freeze the encoder part and train the decoder part as a GAN generator with an additional regularizer which encourages cycle consistency in t... | SP:214f7d764cebce811e175531d0b2e7f0c8dc18c3 |
Imagining the Latent Space of a Variational Auto-Encoders | 1 INTRODUCTION . Variational auto-encoders ( VAEs ) have made a significant impact since their introduction by Kingma and Welling ( 2014 ) . However , one of their perceived problems is their reconstruction performance . This has spawned a wave of research into trying to improve the reconstruction performance ( Zhao et... | This paper proposes a hybrid VAE-GAN model, called the latent space renderer-GAN (LSR-GAN), with the goal to “imagine” the latent space of a VAE, and to improve the decoding and sampling quality of a VAE. First, a VAE-like model is trained, after which the encoder weights are frozen, and the decoder is trained as the g... | SP:214f7d764cebce811e175531d0b2e7f0c8dc18c3 |
DASGrad: Double Adaptive Stochastic Gradient | 1 INTRODUCTION AND MOTIVATION . Stochastic gradient descent ( SGD ) is a widely used optimization method , and currently through backpropagation this algorithm has propelled the success of many deep learning applications . The Deep Learning community has particularly adopted variants of adaptive moment methods for SGD ... | In this paper, the doubly adaptive stochastic gradient method (DASGrad) is introduced via augmenting adaptive moment methods with adaptive (as opposed to uniform) probabilities for data sampling. The convergence of the proposed method is analyzed in terms of regret bound and is compared to similar results for ADAM. The... | SP:de54df40429714a5d975e1b0ef2c4b529ba5f6e3 |
DASGrad: Double Adaptive Stochastic Gradient | 1 INTRODUCTION AND MOTIVATION . Stochastic gradient descent ( SGD ) is a widely used optimization method , and currently through backpropagation this algorithm has propelled the success of many deep learning applications . The Deep Learning community has particularly adopted variants of adaptive moment methods for SGD ... | In the paper, the authors propose a general form that covers most stochastic gradient methods so far, e.g. stochastic gradient descent, Adam, or adaptive probabilities methods. Then, they also provide a convergence analysis of convex problems. In the experiments, they compared the proposed DASGRAD method with Adam, A... | SP:de54df40429714a5d975e1b0ef2c4b529ba5f6e3 |
Random Matrix Theory Proves that Deep Learning Representations of GAN-data Behave as Gaussian Mixtures | 1 INTRODUCTION . The performance of machine learning methods depends strongly on the choice of the data representation ( or features ) on which they are applied . This data representation should ideally contain relevant information about the learning task in order to achieve learning with simple models and small amount... | The authors generalize Gaussian random vectors to a broader class of "concentrated" vectors which they use as their primary tool for analysis of the latent representations learned by GANs. They show that the spectral behavior (i.e. spectra and leading eigenspaces) of the Gram matrix computed over GAN representations is... | SP:b05dc7c8eb4108fcf028d86de6dc4db5e70fed49 |
Random Matrix Theory Proves that Deep Learning Representations of GAN-data Behave as Gaussian Mixtures | 1 INTRODUCTION . The performance of machine learning methods depends strongly on the choice of the data representation ( or features ) on which they are applied . This data representation should ideally contain relevant information about the learning task in order to achieve learning with simple models and small amount... | In this paper, the authors claim to establish a Lipschitz bound on neural networks, initialized randomly, and trained until convergence. They also claim to establish probabilistic concentration of the resolvent of the Gram matrix from a mixture of k distributions with varying means and covariances. The authors also p... | SP:b05dc7c8eb4108fcf028d86de6dc4db5e70fed49 |
Weakly-Supervised Trajectory Segmentation for Learning Reusable Skills | Learning useful and reusable skill , or sub-task primitives , is a long-standing problem in sensorimotor control . This is challenging because it ’ s hard to define what constitutes a useful skill . Instead of direct manual supervision which is tedious and prone to bias , in this work , our goal is to extract reusable ... | This paper tackles the problem of learning to label individual timesteps of sequential data, when given labels only for the sequence as a whole. The authors take an approach derived from the multiple-instance learning (MIL) literature that involves pooling the per-timestep predictions into a sequence-level prediction a... | SP:b16e4a52ca4ae6105ae7506b5446d31901b71d3f |
Weakly-Supervised Trajectory Segmentation for Learning Reusable Skills | Learning useful and reusable skill , or sub-task primitives , is a long-standing problem in sensorimotor control . This is challenging because it ’ s hard to define what constitutes a useful skill . Instead of direct manual supervision which is tedious and prone to bias , in this work , our goal is to extract reusable ... | The paper presents a weakly supervised method for segmentation of trajectories into sub-skills inspired by multi-instance learning (MIL) in image classification by Andrews et al. (2002). This is done via training a classifier to label each observation per time-step with the probability of skills corresponding to that o... | SP:b16e4a52ca4ae6105ae7506b5446d31901b71d3f |
SINGLE PATH ONE-SHOT NEURAL ARCHITECTURE SEARCH WITH UNIFORM SAMPLING | 1 INTRODUCTION . Deep learning automates feature engineering and solves the weight optimization problem . Neural Architecture Search ( NAS ) aims to automate architecture engineering by solving one more problem , architecture design . Early NAS approaches ( Zoph et al. , 2018 ; Zhong et al. , 2018a ; b ; Liu et al. , 2... | This paper presents a new one-shot NAS approach. Parameter updating and structure updating are optimized separately. In the process of parameter optimization, different from the previous methods, the network samples the structure according to the uniform probability distribution. This sampling method can avoid the coup... | SP:745b6d9c4d1e7c3d0b52c91984356c66e3c7ba5b |
SINGLE PATH ONE-SHOT NEURAL ARCHITECTURE SEARCH WITH UNIFORM SAMPLING | 1 INTRODUCTION . Deep learning automates feature engineering and solves the weight optimization problem . Neural Architecture Search ( NAS ) aims to automate architecture engineering by solving one more problem , architecture design . Early NAS approaches ( Zoph et al. , 2018 ; Zhong et al. , 2018a ; b ; Liu et al. , 2... | Authors revise the one-shot NAS algorithm in this work. One-shot NAS that employs a supernet to share the weights between subnets is an efficient NAS algorithm. Authors develop a new training paradigm to train the supernet sufficiently. Specifically, they uniformly sample a single path from supernet at each iteration t... | SP:745b6d9c4d1e7c3d0b52c91984356c66e3c7ba5b |
Storage Efficient and Dynamic Flexible Runtime Channel Pruning via Deep Reinforcement Learning | 1 INTRODUCTION . In recent years , convolutional neural networks ( CNNs ) have been proven to be effective in a wide range of computer vision tasks , such as image classification ( Krizhevsky et al. , 2012 ; Simonyan & Zisserman , 2015 ; He et al. , 2016 ) , objection detection ( He et al. , 2017 ; Zhou et al. , 2019 ;... | This paper proposes to learn static and dynamic channel pruning policies for convolutional neural networks. Static pruning depends only on the training dataset and is computed once before the model is deployed. Dynamic pruning is input-dependent. The policies are obtained with deep reinforcement learning on the trainin... | SP:b451949bef9c16fe4dea78ef337d7f7dcbae3f90 |
Storage Efficient and Dynamic Flexible Runtime Channel Pruning via Deep Reinforcement Learning | 1 INTRODUCTION . In recent years , convolutional neural networks ( CNNs ) have been proven to be effective in a wide range of computer vision tasks , such as image classification ( Krizhevsky et al. , 2012 ; Simonyan & Zisserman , 2015 ; He et al. , 2016 ) , objection detection ( He et al. , 2017 ; Zhou et al. , 2019 ;... | This work introduces a Reinforcement Learning based framework that simultaneously learns both a static and dynamic pruning strategy. The combination allows the static pruner to decrease the required storage while the dynamic pruning can optimize the required compute using input-dependent pruned weights. The RL agent ca... | SP:b451949bef9c16fe4dea78ef337d7f7dcbae3f90 |
Learning to Reason: Distilling Hierarchy via Self-Supervision and Reinforcement Learning | 1 INTRODUCTION . Reinforcement learning ( RL ) aims to compute the optimal control policy while an agent interacts with the environment . Recent advances in deep learning enable RL frameworks to utilize deep neural networks to efficiently represent and learn a policy having a flexible and expressive structure . As a re... | This paper presents a framework for learning hierarchical policies using a latent variable conditioned policy operating at the low level, with model based planning at the high level. Unlike prior work which does hierarchical reinforcement learning, the key technical contribution of this work is that they use planning w... | SP:95408bc20d6c07d7f4d7239faf8f08969e2f8722 |
Learning to Reason: Distilling Hierarchy via Self-Supervision and Reinforcement Learning | 1 INTRODUCTION . Reinforcement learning ( RL ) aims to compute the optimal control policy while an agent interacts with the environment . Recent advances in deep learning enable RL frameworks to utilize deep neural networks to efficiently represent and learn a policy having a flexible and expressive structure . As a re... | This paper proposes a latent variable model to perform imitation learning. The authors propose the model in the control-as-inference framework and introduce two additional latent variables: one that represents a latent state (z) and another that represents a latent action (h). For the generative model, the authors use ... | SP:95408bc20d6c07d7f4d7239faf8f08969e2f8722 |
Improving Sequential Latent Variable Models with Autoregressive Flows | We propose an approach for sequence modeling based on autoregressive normalizing flows . Each autoregressive transform , acting across time , serves as a moving reference frame for modeling higher-level dynamics . This technique provides a simple , general-purpose method for improving sequence modeling , with connectio... | The paper proposes to combine the video modeling approaches based on autoregressive flows (e.g. Kumar’19) with amortized variational inference (e.g. Denton’18), wherein an autoregressive latent variable model optimized with variational inference is extended with an autoregressive flow that further transforms the output... | SP:31950e78fd806be2e0f7cc4728251fdc1a3c437d |
Improving Sequential Latent Variable Models with Autoregressive Flows | We propose an approach for sequence modeling based on autoregressive normalizing flows . Each autoregressive transform , acting across time , serves as a moving reference frame for modeling higher-level dynamics . This technique provides a simple , general-purpose method for improving sequence modeling , with connectio... | This paper proposes to model temporal sequences using autoregressive flows across time steps, that allow to model more explicitly temporal changes of the input, i.e. how the input x_t has changed w.r.t x_{<t}. As also stated by the authors, this is a generalization of other work that instead of modelling the input at e... | SP:31950e78fd806be2e0f7cc4728251fdc1a3c437d |
Estimating Gradients for Discrete Random Variables by Sampling without Replacement | 1 INTRODUCTION . Put replacement in your basement ! We derive the unordered set estimator1 : an unbiased ( gradient ) estimator for expectations over discrete random variables based on ( unordered sets of ) samples without replacement . In particular , we consider the problem of estimating ( the gradient of ) the expec... | In this paper, an unbiased estimator for expectations over discrete random variables is developed based on a sampling-without-replacement strategy. The proposed estimator is shown to be a Rao-Blackwellization of three existing unbiased estimators with guaranteed reduction in estimation variance. The connections of the ... | SP:ce4cfc10fe405005267e62712d939275d2847128 |
Estimating Gradients for Discrete Random Variables by Sampling without Replacement | 1 INTRODUCTION . Put replacement in your basement ! We derive the unordered set estimator1 : an unbiased ( gradient ) estimator for expectations over discrete random variables based on ( unordered sets of ) samples without replacement . In particular , we consider the problem of estimating ( the gradient of ) the expec... | This paper introduces an gradient estimator for loss functions that are expectations over discrete random variables. The basic idea is that an estimator over a discrete distribution can be Rao-Blackwellized by conditioning on the event that the discrete realization was produced by being the first sample drawn from an u... | SP:ce4cfc10fe405005267e62712d939275d2847128 |
CURSOR-BASED ADAPTIVE QUANTIZATION FOR DEEP NEURAL NETWORK | 1 INTRODUCTION . Deep learning ( DL ) has achieved great successes in varied fields such as gaming , natural language processing , speech recognition , computer vision and so on . However , its huge computational burden and large memory consumption still intimidate many potential applications , especially for mobile de... | The authors developed a novel quantization technique that yields layer-wise different mixed-precision quantization. To do so, they alternatively update the pre-trained weights and the quantizer, which they call cursor. The following two features distinguish this paper: using two precision values around the cursor's val... | SP:e677ee557b7b802ce2588bfc05b16054913f9662 |
CURSOR-BASED ADAPTIVE QUANTIZATION FOR DEEP NEURAL NETWORK | 1 INTRODUCTION . Deep learning ( DL ) has achieved great successes in varied fields such as gaming , natural language processing , speech recognition , computer vision and so on . However , its huge computational burden and large memory consumption still intimidate many potential applications , especially for mobile de... | This paper is about using quantization to compress the DNN models. The main idea is to use NAS to obtain the mixed precision model. More specifically, it adaptively chooses the number of quantization bit for each layer using NAS by minimizing the cross-entropy loss and the total number of bits (or model size) used to c... | SP:e677ee557b7b802ce2588bfc05b16054913f9662 |
EgoMap: Projective mapping and structured egocentric memory for Deep RL | 1 INTRODUCTION . A critical part of intelligence is navigation , memory and planning . An animal that is able to store and recall pertinent information about their environment is likely to exceed the performance of an animal whose behavior is purely reactive . Many control problems in partially observed 3D environments... | The paper proposes a novel architecture for spatially structured memory. The main idea is to incorporate inductive bias/invariance derived from projective geometry arguments. The experiments seem to clearly show that this new architecture improves previous approaches to tasks which require spatial reasoning and memory,... | SP:0b1459e58145faa54216d433648da41f64c39a23 |
EgoMap: Projective mapping and structured egocentric memory for Deep RL | 1 INTRODUCTION . A critical part of intelligence is navigation , memory and planning . An animal that is able to store and recall pertinent information about their environment is likely to exceed the performance of an animal whose behavior is purely reactive . Many control problems in partially observed 3D environments... | This paper studies how to build semantic spatial maps for the purpose of navigation in 3D environments. The paper presents a differentiable policy network that pastes together semantic map predictions into a spatial map. Information is read out from this map using a global read operation (that looks at the entire map) ... | SP:0b1459e58145faa54216d433648da41f64c39a23 |
Model Inversion Networks for Model-Based Optimization | 1 INTRODUCTION . Data-driven optimization problems arise in a range of domains : from protein design ( Brookes et al. , 2019 ) to automated aircraft design ( Hoburg & Abbeel , 2012 ) , from the design of robots ( Liao et al. , 2019 ) to the design of neural net architectures ( Zoph & Le , 2017 ) and learning from logge... | This paper tackles the problem of solving a black-box optimization problem where only some samples have been observed. This task requires a good model that can be both expressive and generalizable. Instead of learning only a single forward model of x -> y, this paper proposes to additionally use a mapping from y -> x. ... | SP:dab57601f3910855870d72fb2729f4ce011f11a7 |
Model Inversion Networks for Model-Based Optimization | 1 INTRODUCTION . Data-driven optimization problems arise in a range of domains : from protein design ( Brookes et al. , 2019 ) to automated aircraft design ( Hoburg & Abbeel , 2012 ) , from the design of robots ( Liao et al. , 2019 ) to the design of neural net architectures ( Zoph & Le , 2017 ) and learning from logge... | The paper prposes to learn an inverse network to predict x given a target y for optimisation, instead of the traditional way of optimisation (e.g. using Bayesian optimisation for the complex cases considered in the paper). However, unfortunately, this paper is too close in concept, and in my understanding lower in the ... | SP:dab57601f3910855870d72fb2729f4ce011f11a7 |
Few-Shot One-Class Classification via Meta-Learning | 1 INTRODUCTION . The anomaly detection ( AD ) task ( Chandola et al. , 2009 ; Aggarwal , 2015 ) consists in differentiating between normal and abnormal data samples . AD applications are common in various domains that involve different data types , including medical diagnosis ( Prastawa et al. , 2004 ) , cybersecurity ... | This paper tackles an interesting problem, one-class classification or anomaly detection, using a meta-learning approach. The main contribution is to introduce a parameter such that the inner-loop of the meta-learning algorithm better reflects the imbalance which occurs during meta-testing. Results are shown comparing... | SP:5260bc0d3c1b956f31d8921a51bbc776843cd6ef |
Few-Shot One-Class Classification via Meta-Learning | 1 INTRODUCTION . The anomaly detection ( AD ) task ( Chandola et al. , 2009 ; Aggarwal , 2015 ) consists in differentiating between normal and abnormal data samples . AD applications are common in various domains that involve different data types , including medical diagnosis ( Prastawa et al. , 2004 ) , cybersecurity ... | One of promising approach to tackle the few-shot problems is to use meta-learning so that the learner can quickly generalize to an unseen task. One-class classification requires only a set of positive examples to discriminate negative examples from positive examples. The current paper addresses a method of meta-trainin... | SP:5260bc0d3c1b956f31d8921a51bbc776843cd6ef |
Automated curriculum generation through setter-solver interactions | 1 INTRODUCTION . Reinforcement learning ( RL ) algorithms use correlations between policies and environmental rewards to reinforce and improve agent performance . But such correlation-based learning may struggle in dynamic environments with constantly changing settings or goals , because policies that correlate with re... | This paper tackles the task of automatically inducing a curriculum for agents learning through reinforcement. Specifically, they use two agents — a setter agent that sets goals, and a solver agent that solves the goals provided by the setter. While this has been explored before, the difficulty lies in training both ag... | SP:cd63a80ffd1039df8b4b470f26353da3ce0022ec |
Automated curriculum generation through setter-solver interactions | 1 INTRODUCTION . Reinforcement learning ( RL ) algorithms use correlations between policies and environmental rewards to reinforce and improve agent performance . But such correlation-based learning may struggle in dynamic environments with constantly changing settings or goals , because policies that correlate with re... | This paper proposes an autocurricula scheme to train a goal-conditional agent in a dynamic and sparse-rewarding environment. The main idea is to train a setter model to sample goals for next-step training, where the setter can make the decision either based on the training history or the environmental observation (cond... | SP:cd63a80ffd1039df8b4b470f26353da3ce0022ec |
RTFM: Generalising to New Environment Dynamics via Reading | 1 INTRODUCTION . Reinforcement learning ( RL ) has been successful in a variety of areas such as continuous control ( Lillicrap et al. , 2015 ) , dialogue systems ( Li et al. , 2016 ) , and game-playing ( Mnih et al. , 2013 ) . However , RL adoption in real-world problems is limited due to poor sample efficiency and fa... | This work proposes a new environment, Read to Fight Monsters (RTFM), and correspondingly a new algorithm, txt2\pi, for solving this problem. The RTFM requires the agent to read a description of the rules (x beats y, etc) and a description of the goal (to eliminate y), and perform the task correctly to win the game. The... | SP:41de1f1971e860acdbb74dcd266fd308c035b47b |
RTFM: Generalising to New Environment Dynamics via Reading | 1 INTRODUCTION . Reinforcement learning ( RL ) has been successful in a variety of areas such as continuous control ( Lillicrap et al. , 2015 ) , dialogue systems ( Li et al. , 2016 ) , and game-playing ( Mnih et al. , 2013 ) . However , RL adoption in real-world problems is limited due to poor sample efficiency and fa... | This paper constructs a new game that requires combining visual reasoning with text understanding to win. The authors propose a new model txt2π, based on a new layer called FiLM², which combines text and visual features in a way that allows visual features to be encoded with knowledge of the text features (as in the F... | SP:41de1f1971e860acdbb74dcd266fd308c035b47b |
Learning to Discretize: Solving 1D Scalar Conservation Laws via Deep Reinforcement Learning | Conservation laws are considered to be fundamental laws of nature . It has broad application in many fields including physics , chemistry , biology , geology , and engineering . Solving the differential equations associated with conservation laws is a major branch in computational mathematics . Recent success of machin... | In this paper, the author maps the problem of time series PDE into a naive reinforcement learning problem. Under the MDP assumption, the author sets the initial state of the particles as the current state, the flux at all spaces as the possible actions, and map the state-action pair deterministically to the next state ... | SP:5eed765bdae8974a4dc216b49631d9709767e29e |
Learning to Discretize: Solving 1D Scalar Conservation Laws via Deep Reinforcement Learning | Conservation laws are considered to be fundamental laws of nature . It has broad application in many fields including physics , chemistry , biology , geology , and engineering . Solving the differential equations associated with conservation laws is a major branch in computational mathematics . Recent success of machin... | This paper proposes to use reinforcement learning for constructing discretziation stencils of numerical schemes. More specifically, the method focuses on the widely used WENO schemes, which are an established class of finite difference schemes. Within this context, the method aims for training models to infer the weigh... | SP:5eed765bdae8974a4dc216b49631d9709767e29e |
Adversarial Lipschitz Regularization | 1 INTRODUCTION . In recent years , Generative adversarial networks ( GANs ) ( Goodfellow et al. , 2014 ) have been becoming the state-of-the-art in several generative modeling tasks , ranging from image generation ( Karras et al. , 2018 ) to imitation learning ( Ho and Ermon , 2016 ) . They are based on an idea of a tw... | Virtual Adversarial Training (Miyato et al., 2017) can be viewed as a form of Lipschitz regularization. Inspired by this, the paper proposes a Lipschitz regularization technique that tries to ensure that the function being regularized doesn’t change a lot in virtual adversarial directions. This method is shown to be ef... | SP:d3e5ddd5bff36693dda6d3fb3fc19ab47706ec74 |
Adversarial Lipschitz Regularization | 1 INTRODUCTION . In recent years , Generative adversarial networks ( GANs ) ( Goodfellow et al. , 2014 ) have been becoming the state-of-the-art in several generative modeling tasks , ranging from image generation ( Karras et al. , 2018 ) to imitation learning ( Ho and Ermon , 2016 ) . They are based on an idea of a tw... | It is an interesting idea about how to enforce the Lipsthitz constrain in WGAN by using virtual adversarial training. The connection between virtual adversarial and this paper method - ALR is quite simple and clear. In the experiments, the FID score in the table is not complete which can not clearly compare the ability... | SP:d3e5ddd5bff36693dda6d3fb3fc19ab47706ec74 |
Attention Interpretability Across NLP Tasks | 1 INTRODUCTION . Attention is a way of obtaining a weighted sum of the vector representations of a layer in a neural network model ( Bahdanau et al. , 2015 ) . It is used in diverse tasks ranging from machine translation ( Luong et al. , 2015 ) , language modeling ( Liu & Lapata , 2018 ) to image captioning ( Xu et al.... | This paper investigates the degree to which we might view attention weights as explanatory across NLP tasks and architectures. Notably, the authors distinguish between single and "pair" sequence tasks, the latter including NLI, and generation tasks (e.g., translation). The argument here is that attention weights do not... | SP:50780fb6b72c0da68cb960a12530c54a831222de |
Attention Interpretability Across NLP Tasks | 1 INTRODUCTION . Attention is a way of obtaining a weighted sum of the vector representations of a layer in a neural network model ( Bahdanau et al. , 2015 ) . It is used in diverse tasks ranging from machine translation ( Luong et al. , 2015 ) , language modeling ( Liu & Lapata , 2018 ) to image captioning ( Xu et al.... | I use (unqualified) “self-attention” to refer to attention of tokens in a sequence to other tokens in the same sequence, as described by [some corrected version of] Eq (1) and the paragraph following it (citing Bahdanau et al. 2015). This contrasts with “Transformer self-attention” and “cross-sequence attention”. | SP:50780fb6b72c0da68cb960a12530c54a831222de |
Spectral Nonlocal Block for Neural Network | 1 INTRODUCTION . Capturing the long-range spatial-temporal dependencies is crucial for the Deep Convolutional Neural Networks ( CNNs ) to extract discriminate features in vision tasks such as image and video classification . However , the traditional convolution operator only focuses on processing local neighborhood at... | In this paper, authors propose a spectral nonlocal block. First, they re-interpret the nonlocal blocks in a graph view and then use Chebyshev approximation to obtain the spectral nonlocal block which is quite simple by adding a ZW_1 term. Furthermore, they analyze the steady-state to build up a deeper nonlocal structur... | SP:eb76b7126106346a97624a40277bb28c57f0629b |
Spectral Nonlocal Block for Neural Network | 1 INTRODUCTION . Capturing the long-range spatial-temporal dependencies is crucial for the Deep Convolutional Neural Networks ( CNNs ) to extract discriminate features in vision tasks such as image and video classification . However , the traditional convolution operator only focuses on processing local neighborhood at... | The paper proposes a spectral non-local block, which is a generalized method of the non-local block and non-local stage in the literature. The proposed spectral non-local block can be plugged into a neural network to improve its effectiveness. The paper also provides theoretical analyses of the stability of the propose... | SP:eb76b7126106346a97624a40277bb28c57f0629b |
ReClor: A Reading Comprehension Dataset Requiring Logical Reasoning | 1 INTRODUCTION . Machine reading comprehension ( MRC ) is a fundamental task in Natural Language Processing , which requires models to understand a body of text and answer a particular question related to the context . With success of unsupervised representation learning in NLP , language pre-training based models such... | This paper presents a machine reading comprehension dataset called ReClor. It is different from existing datasets in that ReClor targets logical reasoning. The authors identified biased data points and separated the testing dataset into biased and non-biased sets. Experimental results show that state-of-the-art models ... | SP:26734ce3b17851304312b1211ff74e054046d1d3 |
ReClor: A Reading Comprehension Dataset Requiring Logical Reasoning | 1 INTRODUCTION . Machine reading comprehension ( MRC ) is a fundamental task in Natural Language Processing , which requires models to understand a body of text and answer a particular question related to the context . With success of unsupervised representation learning in NLP , language pre-training based models such... | This paper presents a new reading comprehension dataset for logical reasoning. It is a multi-choice problem where questions are mainly from GMAT and LSAT, containing 4139 data points. The analyses of the data demonstrate that questions require diverse types of reasoning such as finding necessary/sufficient assumptions,... | SP:26734ce3b17851304312b1211ff74e054046d1d3 |
Training binary neural networks with real-to-binary convolutions | 1 INTRODUCTION . Following the introduction of the BinaryNeuralNet ( BNN ) algorithm ( Courbariaux et al. , 2016 ) , binary neural networks emerged as one of the most promising approaches for obtaining highly efficient neural networks that can be deployed on devices with limited computational resources . Binary convolu... | This paper is on building binary network. The steps for building binary network takes several components: traditional strategy to binary/optimize a model (like data augmentation, binary initialization using 2-stage optimization, etc), real-to-binary attention matching that tries to match the output of real values and ... | SP:11e711f93423bcab7a9bad9c9bfd969519b09eb2 |
Training binary neural networks with real-to-binary convolutions | 1 INTRODUCTION . Following the introduction of the BinaryNeuralNet ( BNN ) algorithm ( Courbariaux et al. , 2016 ) , binary neural networks emerged as one of the most promising approaches for obtaining highly efficient neural networks that can be deployed on devices with limited computational resources . Binary convolu... | This paper studies the problem of training binary neural networks. The authors first provide a strong baseline by assembling a group of training techniques that appeared in recent work that achieves state-of-the-art performance. Then the authors proposed two methods to further boost the performance gain. The first meth... | SP:11e711f93423bcab7a9bad9c9bfd969519b09eb2 |
Global Relational Models of Source Code | Models of code can learn distributed representations of a program ’ s syntax and semantics to predict many non-trivial properties of a program . Recent state-ofthe-art models leverage highly structured representations of programs , such as trees , graphs and paths therein ( e.g. , data-flow relations ) , which are prec... | The paper proposes improvements on existing probabilistic models for code that predicts and repairs variable misuses. This is a variant of the task, proposed by Vasic et al. The task takes a dataset of python functions, introduces errors in these functions and makes a classifier that would identify what errors were int... | SP:c7c4415e10a9426b0cffb18491d42922700dce85 |
Global Relational Models of Source Code | Models of code can learn distributed representations of a program ’ s syntax and semantics to predict many non-trivial properties of a program . Recent state-ofthe-art models leverage highly structured representations of programs , such as trees , graphs and paths therein ( e.g. , data-flow relations ) , which are prec... | In this paper, the authors proposed a new method to model the source code for the bug repairing task. Traditional methods use either a global sequence based model or a local graph based model. The authors proposed a new sandwich model like [RNN GNN RNN]. The experiments show that such simple combination of models signi... | SP:c7c4415e10a9426b0cffb18491d42922700dce85 |
Reinforcement Learning with Structured Hierarchical Grammar Representations of Actions | 1 INTRODUCTION . Reinforcement Learning ( RL ) has made great progress in recent years , successfully being applied to settings such as board games ( Silver et al. , 2017 ) , video games ( Mnih et al. , 2015 ) and robot tasks ( OpenAI et al. , 2018 ) . Some of this advance is due to the use of deep learning techniques ... | The authors propose a method for learning macro-actions in a multi-step manner, where Sequitur, a grammar calculator, is leveraged together with an entropy-minimisation based strategy to find relevant macro-actions. The authors propose a system to bootstrap the weights of these macro-actions when increasing the policy'... | SP:23a43ab91a13463f9b5185d5bb0ab328ea6eb0c7 |
Reinforcement Learning with Structured Hierarchical Grammar Representations of Actions | 1 INTRODUCTION . Reinforcement Learning ( RL ) has made great progress in recent years , successfully being applied to settings such as board games ( Silver et al. , 2017 ) , video games ( Mnih et al. , 2015 ) and robot tasks ( OpenAI et al. , 2018 ) . Some of this advance is due to the use of deep learning techniques ... | This paper introduced a way to combine actions into meta-actions through action grammar. The authors trained agents that executes both primitive actions and meta-actions, resulting in better performance on Atari games. Specifically, meta-actions are generated after a period of training from collected greedy action sequ... | SP:23a43ab91a13463f9b5185d5bb0ab328ea6eb0c7 |
GraphSAINT: Graph Sampling Based Inductive Learning Method | 1 INTRODUCTION . Recently , representation learning on graphs has attracted much attention , since it greatly facilitates tasks such as classification and clustering ( Wu et al. , 2019 ; Cai et al. , 2017 ) . Current works on Graph Convolutional Networks ( GCNs ) ( Hamilton et al. , 2017 ; Chen et al. , 2018b ; Gao et ... | This paper proposed a new sampling method to train GCN in the mini-batch manner. In particular, unlike existing methods which samples the mini-batch in the node-wise way, GraphSAINT proposed to sample a mini-batch in the graph-wise way. As a result, GraphSAINT uses the same graph across different GCN layers, while most... | SP:7227922e5ec088fabf0fa9c0584ee4f5c1f3887a |
GraphSAINT: Graph Sampling Based Inductive Learning Method | 1 INTRODUCTION . Recently , representation learning on graphs has attracted much attention , since it greatly facilitates tasks such as classification and clustering ( Wu et al. , 2019 ; Cai et al. , 2017 ) . Current works on Graph Convolutional Networks ( GCNs ) ( Hamilton et al. , 2017 ; Chen et al. , 2018b ; Gao et ... | Scaling GCNs to large graphs is important for real applications. Instead of sampling the nodes or edges across GCN layers, this paper proposes to sample the training graph to improve training efficiency and accuracy. It is a smart idea to construct a complete GCN from the sampled subgraphs. Convincing experiments can... | SP:7227922e5ec088fabf0fa9c0584ee4f5c1f3887a |
When Covariate-shifted Data Augmentation Increases Test Error And How to Fix It | 1 INTRODUCTION . Adversarial training improves the robustness of neural networks to perturbations , commonly referred to as adversarial examples ( Goodfellow et al. , 2015 ; Szegedy et al. , 2014 ; Biggio et al. , 2013 ) . However , adversarial training also causes an undesirable increase in the error on the unperturbe... | This paper studies data augmentation in the regime where labels for the augmented datapoints are known. Special emphasis is put on the study of overparametrised linear models with minimum Euclidean norm of the regression weights (a.k.a. ridgeless regression). The results of this study are then used to motivate their “X... | SP:37e258666bfb1bbd89749be3543e3511bf3a81f7 |
When Covariate-shifted Data Augmentation Increases Test Error And How to Fix It | 1 INTRODUCTION . Adversarial training improves the robustness of neural networks to perturbations , commonly referred to as adversarial examples ( Goodfellow et al. , 2015 ; Szegedy et al. , 2014 ; Biggio et al. , 2013 ) . However , adversarial training also causes an undesirable increase in the error on the unperturbe... | This paper provides some theory into the question of whether data augmentation can hurt test-set performance. To paraphrase the theory, data augmentation can hurt when it causes the model to learn a spurious local details instead of global structure, even if the augmented data comes from the same (predictive) distribut... | SP:37e258666bfb1bbd89749be3543e3511bf3a81f7 |
Improving SAT Solver Heuristics with Graph Networks and Reinforcement Learning | 1 INTRODUCTION . Boolean satisfiability ( SAT ) is an important problem for both industry and academia impacting various fields , including circuit design , computer security , artificial intelligence , automatic theorem proving , and combinatorial optimization . As a result , modern SAT solvers are well-crafted , soph... | This paper investigates the problem of learning new branching heuristics in SAT solvers. The idea is very simple: take MiniSat, remove the usual VSIDS heuristic, and replace it with a variable selection policy that has been trained from a deep reinforcement learning algorithm. The architecture advocated in the present ... | SP:720aa05838e9926dafd1161847b197b8f2f8a64a |
Improving SAT Solver Heuristics with Graph Networks and Reinforcement Learning | 1 INTRODUCTION . Boolean satisfiability ( SAT ) is an important problem for both industry and academia impacting various fields , including circuit design , computer security , artificial intelligence , automatic theorem proving , and combinatorial optimization . As a result , modern SAT solvers are well-crafted , soph... | The paper proposes learning a branching heuristic to be used inside the SAT solver MiniSat using reinforcement learning. The state is represented as a graph representation of the Boolean formula as in previous works, and the policy is parameterized as a graph neural network. At each step of an episode the policy select... | SP:720aa05838e9926dafd1161847b197b8f2f8a64a |
Towards Understanding the Regularization of Adversarial Robustness on Neural Networks | 1 INTRODUCTION . Despite the remarkable performance ( Krizhevsky et al. , 2012 ) of Deep Neural Networks ( NNs ) , they are found to be rather fragile and easily fooled by adversarial examples ( Szegedy et al. , 2014 ) . More intriguingly , these adversarial examples are generated by adding imperceptible noise to norma... | The paper presents new theory to develop understanding about why adversarially robust neural networks show lower test performance compared to their standard counterparts despite being more robust to perturbations in the data. The main hypothesis is that the degradation in performance in adversarially robust networks is... | SP:4ae68ed1b9175b904a6f026277e9ff8bb288797b |
Towards Understanding the Regularization of Adversarial Robustness on Neural Networks | 1 INTRODUCTION . Despite the remarkable performance ( Krizhevsky et al. , 2012 ) of Deep Neural Networks ( NNs ) , they are found to be rather fragile and easily fooled by adversarial examples ( Szegedy et al. , 2014 ) . More intriguingly , these adversarial examples are generated by adding imperceptible noise to norma... | This paper focuses on analyzing the regularization of adversarial robustness (AR) on neural networks (NNs). They establish a generalization error (GE) bound characterizing the regularization of AR, and identify two quantities: margin distributions and singular values of NNs' weight matrices. With empirical studies, the... | SP:4ae68ed1b9175b904a6f026277e9ff8bb288797b |
TWO-STEP UNCERTAINTY NETWORK FOR TASKDRIVEN SENSOR PLACEMENT | 1 INTRODUCTION . Sensor placement is widely studied in the areas of environment monitoring ( Hu et al. , 2018 ; Nguyen et al. , 2015 ) , structural health monitoring ( Ostachowicz et al. , 2019 ) , security screening ( Masoudi et al. , 2016 ) , and adaptive computed tomography ( Ouadah et al. , 2017 ) . The optimal sen... | This paper describes a sensor placement strategy based on information gain on an unknown quantity of interest, which already exists in the active learning literature. As is well-known in the literature, this is equivalent to minimizing the expected remaining entropy. What the authors have done differently is to conside... | SP:e16dc7a0c8ab7f163f6b8f06926aeec03161280d |
TWO-STEP UNCERTAINTY NETWORK FOR TASKDRIVEN SENSOR PLACEMENT | 1 INTRODUCTION . Sensor placement is widely studied in the areas of environment monitoring ( Hu et al. , 2018 ; Nguyen et al. , 2015 ) , structural health monitoring ( Ostachowicz et al. , 2019 ) , security screening ( Masoudi et al. , 2016 ) , and adaptive computed tomography ( Ouadah et al. , 2017 ) . The optimal sen... | This paper addresses the issue of how to optimize sensor placement. The authors propose a framework for sensor placement called Two-step Uncertainty Network (TUN) based on the idea of information gain maximization. More concretely, the proposed method encodes an arbitrary number of measurements, models the conditional ... | SP:e16dc7a0c8ab7f163f6b8f06926aeec03161280d |
Continual learning with hypernetworks | 1 INTRODUCTION . We assume that a neural network f ( x , Θ ) with trainable weights Θ is given data from a set of tasks { ( X ( 1 ) , Y ( 1 ) ) , . . . , ( X ( T ) , Y ( T ) ) } , with input samples X ( t ) = { x ( t , i ) } nti=1 and output samples Y ( t ) = { y ( t , i ) } nti=1 , where nt ≡ |X ( t ) | . A standard t... | Paper proposes a method for CL. The method is based on hypernetworks. These networks are a metamodel, which produce the parameters (from a task-conditioned embedding) which will be used in the main network. Preventing forgetting in the main network is now, replaced by preventing forgetting in the hypernetwork. This is ... | SP:99fd9fac1678bb46d41967f397f237561a3890d3 |
Continual learning with hypernetworks | 1 INTRODUCTION . We assume that a neural network f ( x , Θ ) with trainable weights Θ is given data from a set of tasks { ( X ( 1 ) , Y ( 1 ) ) , . . . , ( X ( T ) , Y ( T ) ) } , with input samples X ( t ) = { x ( t , i ) } nti=1 and output samples Y ( t ) = { y ( t , i ) } nti=1 , where nt ≡ |X ( t ) | . A standard t... | This paper proposes to use hypernetwork to prevent catastrophic forgetting. In deep learning, the information of the samples are converted to parameters during the training process, however, future training process could interfere with the information from the previous tasks. One of the method to prevent forgetting is ... | SP:99fd9fac1678bb46d41967f397f237561a3890d3 |
IMPACT: Importance Weighted Asynchronous Architectures with Clipped Target Networks | 1 INTRODUCTION . Proximal Policy Optimization ( Schulman et al. , 2017 ) is one of the most sample-efficient on-policy algorithms . However , it relies on a synchronous architecture for collecting experiences , which is closely tied to its trust region optimization objective . Other architectures such as IMPALA can ach... | This paper introduces IMPACT which is a distributed RL algorithm that shortens training time of RL systems while maintaining/ improving the sample efficiency. It is built on top of the famous PPO algorithm (https://arxiv.org/abs/1707.06347). The authors break down the novel component of their model into three categorie... | SP:139a4db0a387a52eebf873a8f37f974492aa0d2f |
IMPACT: Importance Weighted Asynchronous Architectures with Clipped Target Networks | 1 INTRODUCTION . Proximal Policy Optimization ( Schulman et al. , 2017 ) is one of the most sample-efficient on-policy algorithms . However , it relies on a synchronous architecture for collecting experiences , which is closely tied to its trust region optimization objective . Other architectures such as IMPALA can ach... | Reinforcement learning (RL) training speed is broadly evaluated on two dimensions: sample efficiency (the number of environment interactions required) and wall-clock time. Improved wall-clock training time has been achieved through distributed actors and learners, but often at the expense of sample efficiency. IMPAC... | SP:139a4db0a387a52eebf873a8f37f974492aa0d2f |
Overcoming Catastrophic Forgetting via Hessian-free Curvature Estimates | Learning neural networks with gradient descent over a long sequence of tasks is problematic as their fine-tuning to new tasks overwrites the network weights that are important for previous tasks . This leads to a poor performance on old tasks – a phenomenon framed as catastrophic forgetting . While early approaches use... | The paper focuses on alleviating the problem of "catastrophic forgetting", exhibited by neural networks learned with gradient-based algorithms over long sequence of tasks. In such learning scenarios, tuning of parameters over the new tasks lead to degradation of performance over the old tasks as the parameters importan... | SP:a4b0890cdeb53d7ea32798703955b14baeb60715 |
Overcoming Catastrophic Forgetting via Hessian-free Curvature Estimates | Learning neural networks with gradient descent over a long sequence of tasks is problematic as their fine-tuning to new tasks overwrites the network weights that are important for previous tasks . This leads to a poor performance on old tasks – a phenomenon framed as catastrophic forgetting . While early approaches use... | This paper proposes a method for tackling catastrophic forgetting. Similar to previous methods such as EWC (Kirkpatrick et al., 2017), they penalize parameter updates that align with the Fisher information matrix of the previous tasks. This will prevent the model from changing the previously useful parameters. They try... | SP:a4b0890cdeb53d7ea32798703955b14baeb60715 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.