paper_name stringlengths 11 170 | text stringlengths 8.07k 307k | summary stringlengths 152 6.16k | paper_id stringlengths 43 43 |
|---|---|---|---|
Swoosh! Rattle! Thump! - Actions that Sound | 1 INTRODUCTION . Imagine the opening of a champagne bottle ! Most vivid imaginations not only capture the celebratory visuals but also the distinctive ‘ pop ’ sound created by the act . Our world is rich and feeds all of our five senses – vision , touch , smell , sound and taste . Of these , the sense of vision , touch... | This paper studies the role of audio in object and action perception, as well as how auditory information can help learning forward and inverse dynamics models. To do this, the authors built a 'tilt-bot', which tilts a box and the object within to collect data (sound & vision) of object interactions. The authors then t... | SP:8214a2ec3d58d4fef82265c8f99e1cbb830873aa |
Swoosh! Rattle! Thump! - Actions that Sound | 1 INTRODUCTION . Imagine the opening of a champagne bottle ! Most vivid imaginations not only capture the celebratory visuals but also the distinctive ‘ pop ’ sound created by the act . Our world is rich and feeds all of our five senses – vision , touch , smell , sound and taste . Of these , the sense of vision , touch... | This paper presents audio-visual object classification and motion prediction work on a novel dataset of 60 different objects rolling around in a bin tilted to and fro by a robot, with video and 4-channel audio recordings of the object impacts. The data is rather novel, is large enough to do ML (around 17 hours of eve... | SP:8214a2ec3d58d4fef82265c8f99e1cbb830873aa |
MelNet: A Generative Model for Audio in the Frequency Domain | Capturing high-level structure in audio waveforms is challenging because a single second of audio spans tens of thousands of timesteps . While long-range dependencies are difficult to model directly in the time domain , we show that they can be more tractably modelled in two-dimensional time-frequency representations s... | In this paper the authors present a new generative model for audio in the frequency domain to capture better the global structure of the signal. For this, they use an autoregressive procedure combined with a multiscale generative model for two-dimensional time-frequency visual representation (STFT spectrogram). The pr... | SP:fa64a906a5800aa62bed00e9c2b29b9fcffd0412 |
MelNet: A Generative Model for Audio in the Frequency Domain | Capturing high-level structure in audio waveforms is challenging because a single second of audio spans tens of thousands of timesteps . While long-range dependencies are difficult to model directly in the time domain , we show that they can be more tractably modelled in two-dimensional time-frequency representations s... | The authors introduce MelNet, an autoregressive model of Mel-frequency scaled spectrograms. They convert audio into high resolution spectrograms to reduce the audio artifacts introduced by inverting spectrograms (here they use gradient-based inversion over Griffin-Lim). To improve modeling of long term dependencies, th... | SP:fa64a906a5800aa62bed00e9c2b29b9fcffd0412 |
Deep Network Classification by Scattering and Homotopy Dictionary Learning | 1 INTRODUCTION . Deep convolutional networks have spectacular applications to classification and regression ( LeCun et al. , 2015 ) , but they are black boxes that are hard to analyze mathematically because of their architecture complexity . Scattering transforms are simplified convolutional neural networks with wavele... | The paper proposes an interpretable architecture for image classification based on a scattering transform and sparse dictionary learning approach. The scattering transform acts as a pre-trained interpretable feature extractor that does not require data. A sparse dictionary on top of this representation (the scattering ... | SP:0f0e048d70d90c7b55524e88954e71efb168cee9 |
Deep Network Classification by Scattering and Homotopy Dictionary Learning | 1 INTRODUCTION . Deep convolutional networks have spectacular applications to classification and regression ( LeCun et al. , 2015 ) , but they are black boxes that are hard to analyze mathematically because of their architecture complexity . Scattering transforms are simplified convolutional neural networks with wavele... | The paper proposes a network architecture composed of three interpretable components followed by a simple MLP classifier. It first applies a scattering transform followed by a learned linear projection (to reduce dimensionality). A sparse representation of these coefficients is then obtained using dictionary learning. ... | SP:0f0e048d70d90c7b55524e88954e71efb168cee9 |
Dream to Control: Learning Behaviors by Latent Imagination | 1 INTRODUCTION Intelligent agents can achieve goals in complex environments even though they never encounter the exact same situation twice . This ability requires building representations of the world from past experience that enable generalization to novel situations . World models offer an explicit way to represent ... | This paper introduced a latent space model for reinforcement learning in vision-based control tasks. It first learns a latent dynamics model, in which the transition model and the reward model can be learned on the latent state representations. Using the learned latent state representations, it used an actor-critic mod... | SP:c457a63633d74f3637f83a95fc2f29bdd01b6411 |
Dream to Control: Learning Behaviors by Latent Imagination | 1 INTRODUCTION Intelligent agents can achieve goals in complex environments even though they never encounter the exact same situation twice . This ability requires building representations of the world from past experience that enable generalization to novel situations . World models offer an explicit way to represent ... | The paper proposes Dreamer, a model-based RL method for high-dimensional inputs such as images. The main novelty in Dreamer is to learn a policy function from latent representation-and-transition models in an end-to-end manner. Specifically, Dreamer is an actor-critic method that learns an optimal policy by backpropaga... | SP:c457a63633d74f3637f83a95fc2f29bdd01b6411 |
ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators | 1 INTRODUCTION . Current state-of-the-art representation learning methods for language can be viewed as learning denoising autoencoders ( Vincent et al. , 2008 ) . They select a small subset of the unlabeled input sequence ( typically 15 % ) , mask the identities of those tokens ( e.g. , BERT ; Devlin et al . ( 2019 ) ... | Authors offer an alternative for masked LM pretraining that's more sample-efficient called replaced token detection. Their method basically replaces certain input tokens with alternatives which are sampled from a generator and train a discriminative model to determine whether its generated or real. The work shows empir... | SP:761efdd848e9b8f43b17473ad774449ae002eeb3 |
ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators | 1 INTRODUCTION . Current state-of-the-art representation learning methods for language can be viewed as learning denoising autoencoders ( Vincent et al. , 2008 ) . They select a small subset of the unlabeled input sequence ( typically 15 % ) , mask the identities of those tokens ( e.g. , BERT ; Devlin et al . ( 2019 ) ... | The paper proposed a novel sample-efficient pretraining task. One inefficiency of BERT is that only 15% tokens are used for training in each example. The paper introduced a generator+discriminator framework to optimize the utility of training examples. The generator task is the MLM which predicts the masked word. The a... | SP:761efdd848e9b8f43b17473ad774449ae002eeb3 |
MEMO: A Deep Network for Flexible Combination of Episodic Memories | 1 INTRODUCTION . During our every day life we need to make several judgments that require connecting facts which were not experienced together , but acquired across experiences at different points in time . For instance , imagine walking your daughter to a coding summer camp and encountering another little girl with a ... | This paper proposes two main changes to the End2End Memory Network (EMN) architecture: a separation between facts and the items that comprise these facts in the external memory, policy to learn the number of memory-hops to reason. The paper also introduces a new Paired Associative Inference (PAI) task inspired by neuro... | SP:a7c5bc5a6764e8188597507fdde1cc3ad514d2ba |
MEMO: A Deep Network for Flexible Combination of Episodic Memories | 1 INTRODUCTION . During our every day life we need to make several judgments that require connecting facts which were not experienced together , but acquired across experiences at different points in time . For instance , imagine walking your daughter to a coding summer camp and encountering another little girl with a ... | This paper presents a new task (paired associate inference), drawn from cognitive psychology, which requires linking many pieces of information together to make inferences with long range dependencies. Experimental results show that standard memory architectures fail on these tasks. To redress this, the paper proposes ... | SP:a7c5bc5a6764e8188597507fdde1cc3ad514d2ba |
Deep Relational Factorization Machines | 1 INTRODUCTION . Many supervised learning tasks need to model data with numerous categorical features , which is usually converted into a set of binary features using one-hot encoding . However , when the original categorical features have high cardinalities , such data becomes high-dimensional and sparse . The difficu... | In this paper, the authors propose generalize the FM to consider both interaction between features and interaction between samples. For the interaction between features, the authors propose to use graph convolution to capture high-order feature interactions. Moreover, the authors construct a graph on the instances base... | SP:b054b02760d839fe09152fbdc75e3090d147345b |
Deep Relational Factorization Machines | 1 INTRODUCTION . Many supervised learning tasks need to model data with numerous categorical features , which is usually converted into a set of binary features using one-hot encoding . However , when the original categorical features have high cardinalities , such data becomes high-dimensional and sparse . The difficu... | This paper proposes to combine the graph neural networks and factorization machines. First, the authors propose a relational feature interaction component (RFO) tp deal with the categorical features. This component first uses the factorization machine to project the features to h^FI(x), then it uses an aggregation oper... | SP:b054b02760d839fe09152fbdc75e3090d147345b |
Pareto Optimality in No-Harm Fairness | 1 INTRODUCTION . Machine learning algorithms play an important role in decision making in society . When these algorithms are used to make high-impact decisions such as hiring , credit-lending , predicting mortality for intensive care unit patients , or classifying benign/malign skin lesions , it is paramount to guaran... | This paper introduces a new kind of algorithmic fairness framework where the focus is on first finding a fair classifier that does "no harm" and then in a subsequent step potentially allow doing harm in order to achieve even fairer outcomes. Fairness is here understood as risk disparity: how different are the risks ach... | SP:c3d608213089ac61f4887e18c5c1e58363c78a09 |
Pareto Optimality in No-Harm Fairness | 1 INTRODUCTION . Machine learning algorithms play an important role in decision making in society . When these algorithms are used to make high-impact decisions such as hiring , credit-lending , predicting mortality for intensive care unit patients , or classifying benign/malign skin lesions , it is paramount to guaran... | This paper considers the notion of "no-harm" group fairness, i.e. trying to reduce the risk gap between minority and majority groups without excessive reduction in performance on the majority groups. Authors formalize the problem by defining a Pareto fair classifier, i.e. one that minimizes the risk gaps between groups... | SP:c3d608213089ac61f4887e18c5c1e58363c78a09 |
The problem with DDPG: understanding failures in deterministic environments with sparse rewards | 1 INTRODUCTION . The Deep Deterministic Policy Gradient ( DDPG ) algorithm ( Lillicrap et al . ( 2015 ) ) is one of the earliest deep Reinforcement Learning ( RL ) algorithms designed to operate on potentially large continuous state and action spaces with a deterministic policy , and it is still one of the most widely ... | Overview: This paper describes a shortfall with the DDPG algorithm on a continuous state action space with sparse rewards. To first prove the existence of this shortfall, the authors demonstrate its theoretical possibility by reviewing the behavior of DDPG actor critic equations and the “two-regimes” proofs in the appe... | SP:b9eff5f0e2d89e5074e564fcbe7b0183c8c4818b |
The problem with DDPG: understanding failures in deterministic environments with sparse rewards | 1 INTRODUCTION . The Deep Deterministic Policy Gradient ( DDPG ) algorithm ( Lillicrap et al . ( 2015 ) ) is one of the earliest deep Reinforcement Learning ( RL ) algorithms designed to operate on potentially large continuous state and action spaces with a deterministic policy , and it is still one of the most widely ... | The paper investigates why DDPG can sometimes fail in environments with sparse rewards. It presents a simple environment that helps the reader build intuition and supports the paper's empirical investigation. First, the paper shows that DDPG fails on the simple environment in ~6% of cases, despite the solution being tr... | SP:b9eff5f0e2d89e5074e564fcbe7b0183c8c4818b |
Provenance detection through learning transformation-resilient watermarking | 1 INTRODUCTION . Generative models have contributed to impressive advancements in content generation and representation learning in both digital image and audio domains ( Brock et al . ( 2018 ) ; Kalchbrenner et al . ( 2018 ) ; Mehri et al . ( 2016 ) ; Zhu et al . ( 2017 ) ; Prenger et al . ( 2019 ) ; Donahue & Simonya... | The authors introduce ReSWAT, a method for transformation-resilient watermarking of images via adversarial training. The high level idea is to learn a watermark/detector pair (W,D). W can be any transformation (in this paper, an l-infty bounded perturbation) that imputes an imperceptible distortion to a given input, wh... | SP:902b1484ef76a82c7a43a9eac6e65c5e08f8345a |
Provenance detection through learning transformation-resilient watermarking | 1 INTRODUCTION . Generative models have contributed to impressive advancements in content generation and representation learning in both digital image and audio domains ( Brock et al . ( 2018 ) ; Kalchbrenner et al . ( 2018 ) ; Mehri et al . ( 2016 ) ; Zhu et al . ( 2017 ) ; Prenger et al . ( 2019 ) ; Donahue & Simonya... | This paper is about a novel method to add watermarks to images and audio that is highly robust to several transformations that is closely related to gan methods. The idea is that the watermark signal is learned concurrently to the detector network, which share similarities to a generator and detector networks. Five sta... | SP:902b1484ef76a82c7a43a9eac6e65c5e08f8345a |
Anomaly Detection Based on Unsupervised Disentangled Representation Learning in Combination with Manifold Learning | 1 INTRODUCTION . Detecting anomalies in data flow of modern intelligent systems is an important but challenging problem . Formally speaking , anomaly detection problems can be statistically viewed as identifying outliers having low probabilities from the modelling of data distribution p ( x ) . Practically , since stat... | In a paper a new way to compute anomality score (for a test point) is suggested. A paper is purely experimental, based on existing techniques to dimension reduction (beta-VAE and t-SNE). Given trained beta-VAE, latent vectors, obtained for training set, are feed into t-SNE algorithm. The overall anomality score for a t... | SP:1b32bd6b0a6c8672c109415f6fcbbad4c13c40f4 |
Anomaly Detection Based on Unsupervised Disentangled Representation Learning in Combination with Manifold Learning | 1 INTRODUCTION . Detecting anomalies in data flow of modern intelligent systems is an important but challenging problem . Formally speaking , anomaly detection problems can be statistically viewed as identifying outliers having low probabilities from the modelling of data distribution p ( x ) . Practically , since stat... | This paper presents a novel deep anomaly detection model. It combines two existing models: B-VAE and t-SNE. The B-VAE is trained unsupervised and learns an encoder and decoder which provide both an embedding and a reconstruction. Using t-SNE to reduce its dimensionality, the embedding is projected into a 2 dimensional ... | SP:1b32bd6b0a6c8672c109415f6fcbbad4c13c40f4 |
Differentiable Architecture Compression | 1 INTRODUCTION . In machine learning , resources at inference time are often much more constrained than at training time . For example , while neural networks for computer vision and natural language processing ( NLP ) are routinely trained using GPUs , trained networks are often deployed on embedded systems or mobile ... | The authors proposed a new gradient-based architecture search method that tries to find more efficient alternatives starting from the pre-trained model. The approach is similar to DARTS (Liu et al., 2019) with a budget constraint, such as size and throughput. One major difference is to modify the update of architectura... | SP:0ae436583d8ace9acd5810d146933893f229ba9b |
Differentiable Architecture Compression | 1 INTRODUCTION . In machine learning , resources at inference time are often much more constrained than at training time . For example , while neural networks for computer vision and natural language processing ( NLP ) are routinely trained using GPUs , trained networks are often deployed on embedded systems or mobile ... | This paper proposes a new architecture search method called "DARC" that utilizes a differentiable objective function. Since a naive formulation of architecture search is reduced to a combinatorial optimization which is not differentiable, the optimization requires much computational cost. To overcome this difficulty, t... | SP:0ae436583d8ace9acd5810d146933893f229ba9b |
Adaptive Learned Bloom Filter (Ada-BF): Efficient Utilization of the Classifier | 1 INTRODUCTION . Bloom filter ( BF ) is a widely used data structure for low-memory and high-speed approximate membership testing ( Bloom , 1970 ) . Bloom filters compress a given set S into bit arrays , where we can approximately test whether a given element ( or query ) x belongs to a set S , i.e. , x ∈ S or otherwis... | The paper proposed an adaptive learned bloom filter. Rather than setting a threshold of prediction score, the paper partitions the score into several intervals; for query insider each interval, the paper either uses a group of independent hash functions to hash the query in one unified bloom filter or introduce an inde... | SP:1a02536d11f939c007732cb7b8619107170e8c64 |
Adaptive Learned Bloom Filter (Ada-BF): Efficient Utilization of the Classifier | 1 INTRODUCTION . Bloom filter ( BF ) is a widely used data structure for low-memory and high-speed approximate membership testing ( Bloom , 1970 ) . Bloom filters compress a given set S into bit arrays , where we can approximately test whether a given element ( or query ) x belongs to a set S , i.e. , x ∈ S or otherwis... | This paper extends the Bloom filter learning by using the complete spectrum of the scores regions. It uses multiple thresholds and then varies the number of hash functions among different scores regions to obtain better trade-off. Detailed theoretical analysis provides guaranteed superiority over learned Bloom filter u... | SP:1a02536d11f939c007732cb7b8619107170e8c64 |
Deep Ensembles: A Loss Landscape Perspective | 1 INTRODUCTION . Consider a typical classification problem , where xn ∈ RD denotes the D-dimensional features and yn ∈ [ 1 , . . . , K ] denotes the class label . Assume we have a parametric model p ( y|x , θ ) for the conditional distribution where θ denotes weights and biases of a neural network , and p ( θ ) is a pr... | This paper is trying to answer the question why ensembles of deep neural networks trained with random initialization work so well in practice in improving accuracy. Their proposed hypothesis is that networks trained from different initializations, although all converge to a low-loss/high accuracy optimum, explore diffe... | SP:e13275073d8298a924331305623d86a4b41c670e |
Deep Ensembles: A Loss Landscape Perspective | 1 INTRODUCTION . Consider a typical classification problem , where xn ∈ RD denotes the D-dimensional features and yn ∈ [ 1 , . . . , K ] denotes the class label . Assume we have a parametric model p ( y|x , θ ) for the conditional distribution where θ denotes weights and biases of a neural network , and p ( θ ) is a pr... | This paper analyzes ensembling methods in deep learning from the perspective of the loss landscapes. The authors empirically show that popular methods for learning Bayesian neural networks produce samples with limited diversity in the function space compared to modes of the loss found using different random initializat... | SP:e13275073d8298a924331305623d86a4b41c670e |
Semi-Supervised Boosting via Self Labelling | 1 INTRODUCTION . The rise of the Internet has made it easy to collect massive amounts of data to perform machine learning tasks . However , providing quality labels to each of the samples collected within these large datasets is a long and expensive process . There is a rich literature aiming to alleviate this issue , ... | In this paper, the authors present an approach for semi-supervised learning which combines noisy labels with boosting. In a first step, the labeled instances are used to train a set of classifiers, and these are used to create noisy labels for the unlabeled instances. Then, an EM procedure is used to estimate the noise... | SP:d5d2c965b30b18749ef11e08271d76ff9c556329 |
Semi-Supervised Boosting via Self Labelling | 1 INTRODUCTION . The rise of the Internet has made it easy to collect massive amounts of data to perform machine learning tasks . However , providing quality labels to each of the samples collected within these large datasets is a long and expensive process . There is a rich literature aiming to alleviate this issue , ... | The authors propose a new semi-supervised boosting approach. The approach takes a set of supervised learning algorithms to simulate "crowd-source" labels of the unlabeled data, which are then used to generate a noisy label per unlabeled instance. The noise level is then estimated with an agreement-based scheme, and fed... | SP:d5d2c965b30b18749ef11e08271d76ff9c556329 |
Learning with Social Influence through Interior Policy Differentiation | 1 INTRODUCTION . The paradigm of Reinforcement Learning ( RL ) , inspired by cognition and animal studies ( Thorndike , 2017 ; Schultz et al. , 1997 ) , can be described as learning by interacting with the environment to maximize a cumulative reward ( Sutton et al. , 1998 ) . From the perspective of ecology , biodivers... | This paper proposes a new method for learning diverse policies in RL environments, with the ultimate goal of increasing reward. The paper develops a novel method, called interior policy differentiation (IPD), that constrains trained policy to be sufficiently different from one another. They test on 3 Mujoco domains, sh... | SP:6b1f56de94f5edc349fed07546f5964151b51d8e |
Learning with Social Influence through Interior Policy Differentiation | 1 INTRODUCTION . The paradigm of Reinforcement Learning ( RL ) , inspired by cognition and animal studies ( Thorndike , 2017 ; Schultz et al. , 1997 ) , can be described as learning by interacting with the environment to maximize a cumulative reward ( Sutton et al. , 1998 ) . From the perspective of ecology , biodivers... | The paper presents a new algorithm for maximizing the diversity of different policies learned for a given task. The diversity is quantified using a metric, where in this case the total variation is used. A policy is different from a set of other policy if its minimum distance to all the other policies is high. The auth... | SP:6b1f56de94f5edc349fed07546f5964151b51d8e |
Deep Graph Matching Consensus | 1 INTRODUCTION . Graph matching refers to the problem of establishing meaningful structural correspondences of nodes between two or more graphs by taking both node similarities and pairwise edge similarities into account ( Wang et al. , 2019b ) . Since graphs are natural representations for encoding relational data , t... | This paper proposes a two-stage GNN-based architecture to establish correspondences between two graphs. The first step is to learn node embeddings using a GNN to obtain soft node correspondences between two graphs. The second step is to iteratively refine them using the constraints of matching consensus in local neighb... | SP:e8af90f522657cb1cc069da98c22ae60d04b8879 |
Deep Graph Matching Consensus | 1 INTRODUCTION . Graph matching refers to the problem of establishing meaningful structural correspondences of nodes between two or more graphs by taking both node similarities and pairwise edge similarities into account ( Wang et al. , 2019b ) . Since graphs are natural representations for encoding relational data , t... | The authors proposed a message passing neural network-based graph matching methods. The overall framework can be viewed as a graph siamese network, where two set of points are passing through the same graph neural network, and then two new embeddings are generated. Using the two embedding the similarity between points ... | SP:e8af90f522657cb1cc069da98c22ae60d04b8879 |
Pitfalls of In-Domain Uncertainty Estimation and Ensembling in Deep Learning | video / code / blog post 1 INTRODUCTION . Deep neural networks ( DNNs ) have become one of the most popular families of machine learning models . The predictive performance of DNNs for classification is often measured in terms of accuracy . However , DNNs have been shown to yield inaccurate and unreliable probability e... | The paper provide an extensive review of current advances in uncertainty estimation in neural networks with the analysis of drawbacks of currently used uncertainty metrics and comparison on scale the recent method to estimate uncertainty. The paper covers a lot of uncertainty metrics and a wide range of methods. The pa... | SP:454c98d15b785ccd0128dbf7d8209adbda1fd2e8 |
Pitfalls of In-Domain Uncertainty Estimation and Ensembling in Deep Learning | video / code / blog post 1 INTRODUCTION . Deep neural networks ( DNNs ) have become one of the most popular families of machine learning models . The predictive performance of DNNs for classification is often measured in terms of accuracy . However , DNNs have been shown to yield inaccurate and unreliable probability e... | This paper mainly concerns the quality of in-domain uncertainty for image classification. After exploring common standards for uncertainty quantification, the authors point out pitfalls of existing metrics by investigating different ensembling techniques and introduce a novel metric called deep ensemble equivalent (DEE... | SP:454c98d15b785ccd0128dbf7d8209adbda1fd2e8 |
Deep Reasoning Networks: Thinking Fast and Slow, for Pattern De-mixing | 1 INTRODUCTION . Deep learning has achieved tremendous success in areas such as vision , speech recognition , language translation , and autonomous driving . Nevertheless , certain limitations of deep learning are generally recognized , in particular , limitations due to the fact that deep learning approaches heavily d... | This work proposes a framework for solving de-mixing problems. The hard constraints from human inputs about a specific problem are relaxed into continuous constraints (the "slow" reasoning part), and a reconstruction loss measures the fitness of the inferred labels with the observations (the "fast" pattern recognition ... | SP:f68087fb27b761dc1b71889ab84723526b621a6c |
Deep Reasoning Networks: Thinking Fast and Slow, for Pattern De-mixing | 1 INTRODUCTION . Deep learning has achieved tremendous success in areas such as vision , speech recognition , language translation , and autonomous driving . Nevertheless , certain limitations of deep learning are generally recognized , in particular , limitations due to the fact that deep learning approaches heavily d... | This paper proposes a new encoder-decoder framework that combines prior knowledge-based regularization and constrained reconstruction for unsupervised and weakly-supervised classification in structure rich scenarios. This framework injects prior knowledge in the form of relaxed constraints that act as regularization du... | SP:f68087fb27b761dc1b71889ab84723526b621a6c |
Reducing Transformer Depth on Demand with Structured Dropout | 1 INTRODUCTION . Transformer architectures ( Vaswani et al. , 2017 ) have become the dominant architecture in natural language processing , with state-of-the-art performance across a variety of tasks , including machine translation ( Vaswani et al. , 2017 ; Ott et al. , 2018 ) , language modeling ( Dai et al. , 2019 ; ... | This work explored the effect of LayerDrop training in efficient pruning at inference time. The authors showed that it is possible to have comparable performance from sub-networks of smaller depth selected from one large network without additional finetuning. More encouraging is that the sub-networks are able to perfor... | SP:278819106a3ae8c15442e56994fb175a0cad70dd |
Reducing Transformer Depth on Demand with Structured Dropout | 1 INTRODUCTION . Transformer architectures ( Vaswani et al. , 2017 ) have become the dominant architecture in natural language processing , with state-of-the-art performance across a variety of tasks , including machine translation ( Vaswani et al. , 2017 ; Ott et al. , 2018 ) , language modeling ( Dai et al. , 2019 ; ... | The paper proposes a method, LayerDrop, for pruning layers in Transformer based models. The goal is to explore the stochastic depth of transformer models during training in order to do efficient layer pruning at inference time. The key idea is simple and easy to understand: randomly dropping transformer layers during t... | SP:278819106a3ae8c15442e56994fb175a0cad70dd |
Influence-Based Multi-Agent Exploration | 1 INTRODUCTION . Reinforcement learning algorithms aim to learn a policy that maximizes the accumulative reward from an environment . Many advances of deep reinforcement learning rely on a dense shaped reward function , such as distance to the goal ( Mirowski et al. , 2016 ; Wu et al. , 2018 ) , scores in games ( Mnih ... | This paper studies the problem of designing effective exploration strategies in multi-agent domains. The key idea is to define one agent's exploration in terms of its interactions with other agents. This leads to two auxiliary exploration objectives, which measure how one agent's actions affect the dynamics and value o... | SP:4677ba60a6346626bfba72170b2d0c68cf9ed6be |
Influence-Based Multi-Agent Exploration | 1 INTRODUCTION . Reinforcement learning algorithms aim to learn a policy that maximizes the accumulative reward from an environment . Many advances of deep reinforcement learning rely on a dense shaped reward function , such as distance to the goal ( Mirowski et al. , 2016 ; Wu et al. , 2018 ) , scores in games ( Mnih ... | This paper proposes methods for incentivizing exploration in multi-agent RL. There are two approaches that are proposed, both framed as influence maximization (of either the state transitions or the decisions of the other agents). The scaling to multiple agents is done via decomposing to pairwise interactions. This i... | SP:4677ba60a6346626bfba72170b2d0c68cf9ed6be |
Bandlimiting Neural Networks Against Adversarial Attacks | 1 INTRODUCTION . Although deep neural networks ( DNN ) have shown to be powerful in many machine learning tasks , Szegedy et al . ( 2013 ) found that they are vulnerable to adversarial samples . Adversarial samples are subtly altered inputs that can fool the trained model to produce erroneous outputs . They are more co... | The paper is dedicated to studying adversarial attack and defense problems from the perspective of Fourier analysis. They demonstrate that the adversarial vulnerability of neural networks can be attributed to non-zero high-frequency components. Then, the author proposes a simple post average approach to smooth out the ... | SP:561cb64e5320799677a5a7108830aaec9d33a963 |
Bandlimiting Neural Networks Against Adversarial Attacks | 1 INTRODUCTION . Although deep neural networks ( DNN ) have shown to be powerful in many machine learning tasks , Szegedy et al . ( 2013 ) found that they are vulnerable to adversarial samples . Adversarial samples are subtly altered inputs that can fool the trained model to produce erroneous outputs . They are more co... | The paper proposes an approach for improving robustness of already trained artificial neural networks with relu activation functions. The main motivation comes from signal processing where robustness is typically obtained via averaging moduli of Fourier coefficients over some frequency band (e.g., mel-frequency coeffic... | SP:561cb64e5320799677a5a7108830aaec9d33a963 |
Minimizing FLOPs to Learn Efficient Sparse Representations | 1 INTRODUCTION . Learning semantic representations using deep neural networks ( DNN ) is now a fundamental facet of applications ranging from visual search ( Jing et al. , 2015 ; Hadi Kiapour et al. , 2015 ) , semantic text matching ( Neculoiu et al. , 2016 ) , oneshot classification ( Koch et al. , 2015 ) , clustering... | This paper focuses on learning a representation that facilitates efficient content-based retrieval. Although the representations that are learned from deep neural networks can contain rich information, it is computationally expensive to use those representations to perform a search for the best match. In particular, co... | SP:e86910b99dbc07691e2882c85c87d150de40a0ff |
Minimizing FLOPs to Learn Efficient Sparse Representations | 1 INTRODUCTION . Learning semantic representations using deep neural networks ( DNN ) is now a fundamental facet of applications ranging from visual search ( Jing et al. , 2015 ; Hadi Kiapour et al. , 2015 ) , semantic text matching ( Neculoiu et al. , 2016 ) , oneshot classification ( Koch et al. , 2015 ) , clustering... | This paper proposes to learn sparse representation in neural networks for retrieval in large database of vectors. Such sparse representation, when the fraction of non-zeros is high, can be computed using sparse matrix multiplication, or variants of inverted index scoring and lead to potentially lower FLOPs needed. This... | SP:e86910b99dbc07691e2882c85c87d150de40a0ff |
Likelihood Contribution based Multi-scale Architecture for Generative Flows | 1 INTRODUCTION . Deep Generative Modeling aims to learn the embedded distributions and representations in input ( especially unlabelled ) data , requiring no/minimal human labelling effort . Learning without knowledge of labels ( unsupervised learning ) is of increasing importance because of the abundance of unlabelled... | This paper propose a heuristic algorithm for deciding which random variables to be Gaussianized early in flow-based generative models. The proposed algorithm involves first training a flow without multi-scale training, for example, 32*32*c - 32*32*c - 32*32*c. Then, it computes the logdet term for each variable at eac... | SP:962c445bf9fa4cc39b00aa1a57073320ba145865 |
Likelihood Contribution based Multi-scale Architecture for Generative Flows | 1 INTRODUCTION . Deep Generative Modeling aims to learn the embedded distributions and representations in input ( especially unlabelled ) data , requiring no/minimal human labelling effort . Learning without knowledge of labels ( unsupervised learning ) is of increasing importance because of the abundance of unlabelled... | This paper presents a new multi-scale architecture for flow-based generative models. Unlike prior work on multi-scale flow architectures which use fixed dimension-splitting heuristics, the proposed approach learns which dimensions to process further. The features are chosen for further processing based on a heuristic m... | SP:962c445bf9fa4cc39b00aa1a57073320ba145865 |
BOSH: An Efficient Meta Algorithm for Decision-based Attacks | 1 INTRODUCTION . It has been shown that machine learning models , including deep neural networks , are vulnerable to adversarial examples ( Goodfellow et al. , 2014 ; Szegedy et al. , 2013 ; Chen et al. , 2017a ) . Therefore , evaluating the robustness of a given model becomes crucial for security sensitive application... | In this paper, the authors study the adversarial example generation problem, in the difficult case where the attacked model is a black box. Since the model is unknown, the approaches based on the minimization of a loss function with a gradient based optimizer do not apply. The current alternatives, known as decision-ba... | SP:fb717cacd65d17e3d1971170a82b902ee94d4dfc |
BOSH: An Efficient Meta Algorithm for Decision-based Attacks | 1 INTRODUCTION . It has been shown that machine learning models , including deep neural networks , are vulnerable to adversarial examples ( Goodfellow et al. , 2014 ; Szegedy et al. , 2013 ; Chen et al. , 2017a ) . Therefore , evaluating the robustness of a given model becomes crucial for security sensitive application... | This paper proposes a meta-algorithm for the so-called "decision-based attack" problem, where a model that can be accessed only via label queries for a given input is attacked by a minimal perturbation to the input that changes the predicted label. The algorithm, BOSH, augments any iterative algorithm for this problem ... | SP:fb717cacd65d17e3d1971170a82b902ee94d4dfc |
A Unified framework for randomized smoothing based certified defenses | √ d ) , where d is the dimensionality of the data . This theoretical finding sug- gests that certifying ` ∞-normed robustness by randomized smoothing may not be scalable to high-dimensional data . The veracity of our framework and analysis is verified by extensive evaluations on CIFAR10 and ImageNet . 1 INTRODUCTION . ... | This work examines the recently proposed randomized smoothing method for certifying the robustness of neural networks. The authors explain a theoretical framework for analyzing randomized smoothing as a certification method, propose two alternative definitions of robustness (D_MR and D_inf), and prove that using Gaussi... | SP:cefb35a0bba2e8d3b11b1c81bde283b4e0699da6 |
A Unified framework for randomized smoothing based certified defenses | √ d ) , where d is the dimensionality of the data . This theoretical finding sug- gests that certifying ` ∞-normed robustness by randomized smoothing may not be scalable to high-dimensional data . The veracity of our framework and analysis is verified by extensive evaluations on CIFAR10 and ImageNet . 1 INTRODUCTION . ... | The authors propose a new definition for robustness of random functions. This definition is ideal for analyzing the certified robustness under randomized smoothing techniques. They analyze and show that the Gaussian smoothing is near optimal for \ell_2 smoothing as the mean maximum error is only off by a factor of log ... | SP:cefb35a0bba2e8d3b11b1c81bde283b4e0699da6 |
Latent Variables on Spheres for Sampling and Inference | 1 INTRODUCTION . Deep generative models , such as Variational Auto-Encoder ( VAE ) ( Kingma & Welling , 2013 ; Rezende et al. , 2014 ) and Generative Adversarial Network ( GAN ) ( Goodfellow et al. , 2014 ) , play more and more important role in machine learning and computer vision . However , the problem of variationa... | This paper considers the L2 normalization of samples “z” from a given prior p(z) in Generative Adversarial Netowks (GAN) and autoencoders. The L2 normalization corresponds to projecting samples onto the surface of a unit-hypersphere. Hence, to attempt to justify this normalization, the authors rely on some already esta... | SP:8b572d1f037184bb002765442e1ab35f57a1f084 |
Latent Variables on Spheres for Sampling and Inference | 1 INTRODUCTION . Deep generative models , such as Variational Auto-Encoder ( VAE ) ( Kingma & Welling , 2013 ; Rezende et al. , 2014 ) and Generative Adversarial Network ( GAN ) ( Goodfellow et al. , 2014 ) , play more and more important role in machine learning and computer vision . However , the problem of variationa... | This paper proposes a novel autoencoder algorithm, named Spherical AutoEncoder (SAE). In this paper, the authors argue that the sphere structure has good properties in high-dimensional. To leverage the properties, proposed algorithm centerizes latent variables and projects them onto unit sphere. To show the empirical p... | SP:8b572d1f037184bb002765442e1ab35f57a1f084 |
Feature Partitioning for Efficient Multi-Task Architectures | 1 INTRODUCTION . Multi-task learning allows models to leverage similarities across tasks and avoid overfitting to the particular features of any one task ( Caruana , 1997 ; Zamir et al. , 2018 ) . This can result in better generalization and more robust feature representations . While this makes multi-task learning app... | This paper studies multi-task learning (MTL) from the deep learning perspective where a number of layers are shared between tasks followed by specific heads for each task. One of the main challenges in this problem is to decide the best configuration among a large number of possible ones (e.g., the number of layers , n... | SP:237b129348ea81633989e45c3db9b6e8ef6fdfa2 |
Feature Partitioning for Efficient Multi-Task Architectures | 1 INTRODUCTION . Multi-task learning allows models to leverage similarities across tasks and avoid overfitting to the particular features of any one task ( Caruana , 1997 ; Zamir et al. , 2018 ) . This can result in better generalization and more robust feature representations . While this makes multi-task learning app... | This paper proposes a framework for learning multi-task convolutional neural networks. For each layer of the network, the proposed algorithm assigns a subset of the layer's channels to each of the tasks. This is in contrast to existing methods that assign whole layers to tasks. There are two key ideas here: (1) instead... | SP:237b129348ea81633989e45c3db9b6e8ef6fdfa2 |
Learning Boolean Circuits with Neural Networks | 1 INTRODUCTION AND MOTIVATION . It is well known ( e.g . Livni et al . ( 2014 ) ) that while deep neural-networks can express any function that can be run efficiently on a computer , in the general case , training them is computationally hard . Despite this theoretic pessimism , in practice , deep neural networks are s... | The paper proposes a layer-wise method for training the weights of a binary-tree-structured neural network such that it correctly reproduces certain classes of Boolean functions defined by binary-tree-structured Boolean circuits. Specifically, this paper shows analytically that if a circuit satisfies a property termed ... | SP:fcee5370a61cbfb74d07727f29d83623f2f452e5 |
Learning Boolean Circuits with Neural Networks | 1 INTRODUCTION AND MOTIVATION . It is well known ( e.g . Livni et al . ( 2014 ) ) that while deep neural-networks can express any function that can be run efficiently on a computer , in the general case , training them is computationally hard . Despite this theoretic pessimism , in practice , deep neural networks are s... | This paper aims to study the correlation between the neural network's input and output by abstracting the network as a binary tree Boolean circuit problem. The paper is well-written, motivations are clearly presented, and literature reviews are well placed. The contributions are mainly theoretical, and the experimental... | SP:fcee5370a61cbfb74d07727f29d83623f2f452e5 |
Stiffness: A New Perspective on Generalization in Neural Networks | 1 INTRODUCTION . Neural networks are a class of highly expressive function approximators that proved to be successful in approximating solutions to complex tasks across many domains such as vision , natural language understanding , and game-play . They have long been recognized as universal function approximators ( Hor... | This paper introduces “stiffness”, a new metric to characterize generalization in neural networks. Stiffness is a pretty simple concept and is relatively straightforward to compute. The authors evaluate this metric on standard datasets using two relatively small neural networks. On the whole, the paper is written clear... | SP:201070333d2ca3ad49d9f4783d190e2c1772afe9 |
Stiffness: A New Perspective on Generalization in Neural Networks | 1 INTRODUCTION . Neural networks are a class of highly expressive function approximators that proved to be successful in approximating solutions to complex tasks across many domains such as vision , natural language understanding , and game-play . They have long been recognized as universal function approximators ( Hor... | This submission introduces a metric, termed stiffness, to evaluate the generalization capability of neural networks. The metric is novel and straightforward, it measures how stiff a network is by looking at how a small gradient step on one example affects the loss on another example. The authors study several configura... | SP:201070333d2ca3ad49d9f4783d190e2c1772afe9 |
Nesterov Accelerated Gradient and Scale Invariance for Adversarial Attacks | 1 INTRODUCTION . Deep learning models have been shown to be vulnerable to adversarial examples ( Goodfellow et al. , 2014 ; Szegedy et al. , 2014 ) , which are generated by applying human-imperceptible perturbations on benign input to result in the misclassification . In addition , adversarial examples have an intrigui... | This paper studies how to generate transferable adversarial examples for black-box attacks. Two methods have been proposed, namely Nesterov Iterative Fast Gradient Sign Method (NI-FGSM) and Scale-Invariant attack Method (SIM). The first method adopts Nesterov optimizer instead of momentum optimizer to generate adversa... | SP:19318b52fa22d612f81c72457f9876d9abe7d701 |
Nesterov Accelerated Gradient and Scale Invariance for Adversarial Attacks | 1 INTRODUCTION . Deep learning models have been shown to be vulnerable to adversarial examples ( Goodfellow et al. , 2014 ; Szegedy et al. , 2014 ) , which are generated by applying human-imperceptible perturbations on benign input to result in the misclassification . In addition , adversarial examples have an intrigui... | In this paper, the authors apply the Nesterov Accelerated Gradient method to the adversarial attack task and achieve better transferability of the adversarial examples. Furthermore, the authors introduce a scale transformation method to provide the augmentation on the model, which also boosts the transferability of the... | SP:19318b52fa22d612f81c72457f9876d9abe7d701 |
At Your Fingertips: Automatic Piano Fingering Detection | 1 INTRODUCTION . Learning to play the piano is a hard task taking years to master . One of the challenging aspects when learning a new piece is the fingering choice in which to play each note . While beginner booklets contain many fingering suggestions , advanced pieces often contain none or a select few . Automatic pr... | The paper is a nice piece of works which clearly articulates the objective and the subsequent discussion. The focus of the paper--i.e. disclose the difficulties of piano fingering data annotation and the proposal of automating this process by automatically extracting fingerings from public videos and MIDI files, using ... | SP:42d41dec3695a319b32a212d33682ae15535f27c |
At Your Fingertips: Automatic Piano Fingering Detection | 1 INTRODUCTION . Learning to play the piano is a hard task taking years to master . One of the challenging aspects when learning a new piece is the fingering choice in which to play each note . While beginner booklets contain many fingering suggestions , advanced pieces often contain none or a select few . Automatic pr... | In this paper, the authors proposed an automatic piano fingering algorithm, that accepts YouTube videos and corresponding MIDI files and outputs fingering prediction for each note. The claimed contribution is two-fold: First, they proposed the algorithm, and second, they claim that the algorithm can be used to automati... | SP:42d41dec3695a319b32a212d33682ae15535f27c |
Representation Quality Explain Adversarial Attacks | 1 INTRODUCTION . Adversarial samples are slightly perturbed inputs that can make neural networks misclassify . They are carefully crafted by searching for variations in the input that , for example , could decrease the soft-labels of the correct class . Since they were discovered some years ago ( Szegedy , 2014 ) , the... | This paper aims at revealing the relationship between the quality of deep representations and the attack susceptibility of deep classification models. To this end, they propose the zero-shot test to investigate the "quality" of learned representations for unknown classes. Specifically, they leverage two kinds of qualit... | SP:f78edf237bfd944156163801e210e08fd16f8625 |
Representation Quality Explain Adversarial Attacks | 1 INTRODUCTION . Adversarial samples are slightly perturbed inputs that can make neural networks misclassify . They are carefully crafted by searching for variations in the input that , for example , could decrease the soft-labels of the correct class . Since they were discovered some years ago ( Szegedy , 2014 ) , the... | This paper proposes to evaluate the robustness of the neural networks by extrapolating to the unseen classes. However, the authors only include evaluation for non-robust trained models, without considering the robust trained model, such as Madry et al. [1]. The conclusion is not convincing that the authors studied the ... | SP:f78edf237bfd944156163801e210e08fd16f8625 |
Peer Loss Functions: Learning from Noisy Labels without Knowing Noise Rates | 1 INTRODUCTION . The quality of supervised learning models depends on the training data { ( xn , yn ) } Nn=1 . In practice , label noise can arise due to a host of reasons . For instance , the observed labels ỹns may represent human observations of a ground truth label . In this case , human annotators may observe the... | The paper studies the label noise problem with the motivation of without estimating the flip rate or transition matrix. This is an interesting direction for dealing with label noise. Most of the previous studies need either estimate the transition matrix or put restrictions on it, e.g., to be symmetric. A very related ... | SP:3072194753b9f5af63f5d4c9a06b6d67f39e6b0b |
Peer Loss Functions: Learning from Noisy Labels without Knowing Noise Rates | 1 INTRODUCTION . The quality of supervised learning models depends on the training data { ( xn , yn ) } Nn=1 . In practice , label noise can arise due to a host of reasons . For instance , the observed labels ỹns may represent human observations of a ground truth label . In this case , human annotators may observe the... | This paper studies the problem of learning classifiers from noisy data without specifying the noise rates. Inspired by the literature of peer prediction, the authors propose peer loss. First, a scoring function is introduced, minimizing which we can elicit the Bayes optimal classifier f*. Then the authors use the setti... | SP:3072194753b9f5af63f5d4c9a06b6d67f39e6b0b |
Goten: GPU-Outsourcing Trusted Execution of Neural Network Training and Prediction | 1 INTRODUCTION . While deep neural networks ( DNN ) can produce predictive models with unparalleled performance , its training phase requires enormous data as input . A single data owner may not possess enough data to train a good DNN . Multiple data owners , say , financial institutions , may want to collaborate in tr... | The paper builds a privacy-preserving training framework within a Trusted Execution Environment (TEE) such as Intel SGX. The work is heavily inspired from Slalom, which does privacy-preserving inference in TEEs. The main drawbacks of Slalom when extending to training are (1) weight quantization needs to be dynamics as ... | SP:da2ce3fdc90fc70d3f51e3e26fb844b4b1759af5 |
Goten: GPU-Outsourcing Trusted Execution of Neural Network Training and Prediction | 1 INTRODUCTION . While deep neural networks ( DNN ) can produce predictive models with unparalleled performance , its training phase requires enormous data as input . A single data owner may not possess enough data to train a good DNN . Multiple data owners , say , financial institutions , may want to collaborate in tr... | The paper proposes a method for privacy-preserving training and evaluation of DNNs. The method is based on a combination of hardware support from a trusted execution enclave (Intel SGX) and an algorithm for offloading intensive computation to unsecure GPU devices and communicating with the trusted environment without l... | SP:da2ce3fdc90fc70d3f51e3e26fb844b4b1759af5 |
Bootstrapping the Expressivity with Model-based Planning | 1 INTRODUCTION . Model-based deep reinforcement learning ( RL ) algorithms offer a lot of potentials in achieving significantly better sample efficiency than the model-free algorithms for continuous control tasks . We can largely categorize the model-based deep RL algorithms into two types : 1. model-based policy optim... | This paper presents a mainly theoretical argument comparing the expressivity of model-free and model-based RL methods contrary to analysis in the past which usually relies on sample complexity. They construct a family of MDPs, where the true dynamics belong to a simple function class (in terms of the number of linear p... | SP:e95b1fa1f8d1b66ef0fdfb6c7aa56983f83e2277 |
Bootstrapping the Expressivity with Model-based Planning | 1 INTRODUCTION . Model-based deep reinforcement learning ( RL ) algorithms offer a lot of potentials in achieving significantly better sample efficiency than the model-free algorithms for continuous control tasks . We can largely categorize the model-based deep RL algorithms into two types : 1. model-based policy optim... | The paper highlights an interesting issue regarding approximability of function approximators (neural networks). The paper provides cases where the action value function is difficult to approximate and is much more difficult than the dynamics of a model. The author conducts some experiments to claim that even with a la... | SP:e95b1fa1f8d1b66ef0fdfb6c7aa56983f83e2277 |
Improved Training Speed, Accuracy, and Data Utilization via Loss Function Optimization | 1 INTRODUCTION . Much of the power of modern neural networks originates from their complexity , i.e. , number of parameters , hyperparameters , and topology . This complexity is often beyond human ability to optimize , and automated methods are needed . An entire field of metalearning has emerged recently to address th... | This paper proposes a very interesting idea of loss function optimization. At first sight, loss function is the goal of optimization and can not be optimized directly. However, the true goal of optimization is the final accuracy (for classification). So lots of loss functions can be designed and combined to form a larg... | SP:a3f8fc0a93ecd80f88cf808e8cd228588010b1b0 |
Improved Training Speed, Accuracy, and Data Utilization via Loss Function Optimization | 1 INTRODUCTION . Much of the power of modern neural networks originates from their complexity , i.e. , number of parameters , hyperparameters , and topology . This complexity is often beyond human ability to optimize , and automated methods are needed . An entire field of metalearning has emerged recently to address th... | The authors propose using evolutionary computation (EC) to perform meta learning over the set of symbolic expressions for loss functions. It's a compelling idea that is well-motivated. They find that applying their EC method to mnist yields an interesting loss function that they name the 'Baikal loss.' Much of the pape... | SP:a3f8fc0a93ecd80f88cf808e8cd228588010b1b0 |
Learning Heuristics for Quantified Boolean Formulas through Reinforcement Learning | 1 INTRODUCTION . One of the most intriguing questions for artificial intelligence is : can ( deep ) learning be effectively used for symbolic reasoning ? The benefits of combining deductive reasoning with inductive learning for automated reasoning and in formal methods for system design have been noted ( e.g. , see Ses... | This paper investigates the problem of predicting the truth of quantified boolean formulae using deep reinforcement learning. In this setting, the problem is formulated as a reinforcement learning task, in which the learner is interacting with a solver (CADET), and its goal is to find a sequence of actions (each associ... | SP:ff7ab4e497018b2fa801bd05e7e14d59265babed |
Learning Heuristics for Quantified Boolean Formulas through Reinforcement Learning | 1 INTRODUCTION . One of the most intriguing questions for artificial intelligence is : can ( deep ) learning be effectively used for symbolic reasoning ? The benefits of combining deductive reasoning with inductive learning for automated reasoning and in formal methods for system design have been noted ( e.g. , see Ses... | This will be an uncharacteristically short review. The work poses an interesting idea: why not mix heuristics and learning. It reads as if the paper was written a while ago and the intro was not updated, since there is a lot of related work using the same concept. Please cite existing work in the introduction, it refle... | SP:ff7ab4e497018b2fa801bd05e7e14d59265babed |
Language GANs Falling Short | 1 INTRODUCTION . Generating fluent natural language is a central aim in Natural Language Processing ( NLP ) . Transformer architectures ( Vaswani et al. , 2017 ) with hundreds of millions or billions of parameters regularly reestablish state-of-the-art on held-out validation perplexities , however , the generated sampl... | This paper concerns the limitation of the quality-only evaluation metric for text generation models. Instead, a desirable evaluation metric should not only measure the sample quality, but also the sample diversity, to prevent the mode collapse problem in gan-based models generation. The author presents an interesting, ... | SP:7dffea29b6080871ab1737a0032361627fb8f5aa |
Language GANs Falling Short | 1 INTRODUCTION . Generating fluent natural language is a central aim in Natural Language Processing ( NLP ) . Transformer architectures ( Vaswani et al. , 2017 ) with hundreds of millions or billions of parameters regularly reestablish state-of-the-art on held-out validation perplexities , however , the generated sampl... | Recently many language GAN papers have been published to overcome the so called exposure bias, and demonstrated improvements in natural language generation in terms of sample quality, some works propose to assess the generation in terms of diversity, however, quality and diversity are two conflicting measures that are... | SP:7dffea29b6080871ab1737a0032361627fb8f5aa |
Asynchronous Stochastic Subgradient Methods for General Nonsmooth Nonconvex Optimization | Asynchronous distributed methods are a popular way to reduce the communication and synchronization costs of large-scale optimization . Yet , for all their success , little is known about their convergence guarantees in the challenging case of general non-smooth , non-convex objectives , beyond cases where closed-form p... | This paper analyzes convergence of asynchronous methods on general non-smooth and non-convex functions (typically arising from deep leaning). Stochastic sub-gradient asynchronous methods are of particular challenging when coupled with complicated hardware behavior of modern NUMA architecture. To validate the analysis, ... | SP:f9ed35b4b0c9410d718d872f80a50e182455e04d |
Asynchronous Stochastic Subgradient Methods for General Nonsmooth Nonconvex Optimization | Asynchronous distributed methods are a popular way to reduce the communication and synchronization costs of large-scale optimization . Yet , for all their success , little is known about their convergence guarantees in the challenging case of general non-smooth , non-convex objectives , beyond cases where closed-form p... | This paper proposes a model to study asynchronous stochastic subgradient methods for minimizing a nonsmooth nonconvex function. Studying stochastic subgradient methods in the nonsmooth nonconvex setting is already very challenging. Throwing asynchronous updates into the mix further complicates the analysis. This is ove... | SP:f9ed35b4b0c9410d718d872f80a50e182455e04d |
Oblique Decision Trees from Derivatives of ReLU Networks | 1 INTRODUCTION . Decision trees ( Breiman et al. , 1984 ) employ a series of simple decision nodes , arranged in a tree , to transparently capture how the predicted outcome is reached . Functionally , such tree-based models , including random forest ( Breiman , 2001 ) , realize piece-wise constant functions . Beyond th... | This paper leverages the piecewise linearity of predictions in ReLU neural networks to encode and learn piecewise constant predictors akin to oblique decision trees (trees with splits made on linear combinations of features instead of axis-aligned splits). The core observation is that the Jacobian of a ReLU network is ... | SP:4758a1780d55349f71362bc930fba49a08895733 |
Oblique Decision Trees from Derivatives of ReLU Networks | 1 INTRODUCTION . Decision trees ( Breiman et al. , 1984 ) employ a series of simple decision nodes , arranged in a tree , to transparently capture how the predicted outcome is reached . Functionally , such tree-based models , including random forest ( Breiman , 2001 ) , realize piece-wise constant functions . Beyond th... | This paper proposes locally constant network (LCN), which is implemented via the gradient of piece-wise linear networks such as ReLU networks. The authors built the equivalence between LCN and decision trees, and also demonstrated that LCN with M neurons has the same representation capability as decision trees with 2^M... | SP:4758a1780d55349f71362bc930fba49a08895733 |
InfoGraph: Unsupervised and Semi-supervised Graph-Level Representation Learning via Mutual Information Maximization | 1 INTRODUCTION . Graphs have proven to be an effective way to represent very diverse types of data including social networks Newman & Girvan ( 2004 ) , biological reaction networksPavlopoulos et al . ( 2011 ) , proteinprotein interactions Krogan et al . ( 2006 ) , the quantum mechanical properties of individual molecul... | In this paper, the authors propose a graph-level representation, which extends the existing node-level representation learning mechanism. Besides, both unsupervised and semi-supervised learning are leveraged for InfoGraph and InfoGraph*, receptively. The authors naturally apply Deep Graph Infomax, a contrastive represe... | SP:b3180fc8a3ed68988595e881da81393fac04847c |
InfoGraph: Unsupervised and Semi-supervised Graph-Level Representation Learning via Mutual Information Maximization | 1 INTRODUCTION . Graphs have proven to be an effective way to represent very diverse types of data including social networks Newman & Girvan ( 2004 ) , biological reaction networksPavlopoulos et al . ( 2011 ) , proteinprotein interactions Krogan et al . ( 2006 ) , the quantum mechanical properties of individual molecul... | The paper presents a new graph representation learning method for the whole graph under both unsupervised and semi-supervised setting. Different from existing ones using graph kernel, or graph2vec, the proposed InfoGraph is able to extract graph-level representation with fixed-length features that are generalized well.... | SP:b3180fc8a3ed68988595e881da81393fac04847c |
Theory and Evaluation Metrics for Learning Disentangled Representations | 1 INTRODUCTION . Disentanglement learning holds the key for understanding the world from observations , transferring knowledge across different tasks and domains , generating novel designs , and learning compositional concepts ( Bengio et al. , 2013 ; Higgins et al. , 2017b ; Lake et al. , 2017 ; Peters et al. , 2017 ;... | This paper defines precise semantics of disentanglement representations and presents evaluation metrics to evaluate such representations. The authors provides information-theoretic characterization disentangled representations along three dimensions: informativeness, separability, and interpretability; and propose metr... | SP:746ebd0772b2236009e7ff32cf8669f8bc8b6615 |
Theory and Evaluation Metrics for Learning Disentangled Representations | 1 INTRODUCTION . Disentanglement learning holds the key for understanding the world from observations , transferring knowledge across different tasks and domains , generating novel designs , and learning compositional concepts ( Bengio et al. , 2013 ; Higgins et al. , 2017b ; Lake et al. , 2017 ; Peters et al. , 2017 ;... | The paper presents a new set of metrics for evaluating disentangled representations in both supervised and unsupervised settings. Disentangled representations are evaluated along three dimensions: informativeness, separability, and interpretability. While previous work offers metrics for similar dimensions (e.g., (East... | SP:746ebd0772b2236009e7ff32cf8669f8bc8b6615 |
Linguistic Embeddings as a Common-Sense Knowledge Repository: Challenges and Opportunities | Many applications of linguistic embedding models rely on their value as pretrained inputs for end-to-end tasks such as dialog modeling , machine translation , or question answering . This position paper presents an alternate paradigm : Rather than using learned embeddings as input features , we instead treat them as a ... | This paper provides an overview of methods to use embeddings for texts as common-sense knowledge. It mentions many aspects where embeddings can be used as common-sense knowledge. However, the paper lacks both novelty and in-depth analysis. Most methods proposed are basically computing the cosine distances between lang... | SP:640d652c3d0bdfd7a5b3e557d2cb390882183268 |
Linguistic Embeddings as a Common-Sense Knowledge Repository: Challenges and Opportunities | Many applications of linguistic embedding models rely on their value as pretrained inputs for end-to-end tasks such as dialog modeling , machine translation , or question answering . This position paper presents an alternate paradigm : Rather than using learned embeddings as input features , we instead treat them as a ... | - This paper claims to provide a “alternate” view of pretrained embedding as commonsense repository. But this is not a new view at all. It is well known embedding can encode commonsense knowledge, which is a reason it helps a wide variety of recent commonsense related tasks (and there has been considerable analysis on ... | SP:640d652c3d0bdfd7a5b3e557d2cb390882183268 |
Model-free Learning Control of Nonlinear Stochastic Systems with Stability Guarantee | 1 INTRODUCTION . Control of discrete-time nonlinear stochastic systems is an important topic in both control theory and reinforcement learning . In the past decades , the advancement of nonlinear control theory in the control community has been successfully applied in aircraft , automobiles , advanced robots and space ... | In this paper, the authors introduce an algorithm to learn a stable controller using deep NN actor-critic method. They define the stability in the mean cost criteria, which is used to constrain the critic network as a Lyapunov function. In addition, the semi-positive definiteness of the Lyapunov function is enforced b... | SP:7618656900318448b5c9fd7ce2fb33bab3384a4c |
Model-free Learning Control of Nonlinear Stochastic Systems with Stability Guarantee | 1 INTRODUCTION . Control of discrete-time nonlinear stochastic systems is an important topic in both control theory and reinforcement learning . In the past decades , the advancement of nonlinear control theory in the control community has been successfully applied in aircraft , automobiles , advanced robots and space ... | In this work the authors studied the model-free RL approach for learning a policy with stability guarantees. Leveraging the Lyapunov stochastic stability criterion, instead if minimizing the cumulative cost (plus a soft entropy), they propose optimizing an objective function with a specific Lyapunov critic, which is a ... | SP:7618656900318448b5c9fd7ce2fb33bab3384a4c |
An Explicitly Relational Neural Network Architecture | 1 INTRODUCTION . When humans face novel problems , they are able to draw effectively on past experience with other problems that are superficially very different , but that have similarities on a more abstract , structural level . This ability is essential for lifelong , continual learning , and confers on humans a deg... | This paper presents PrediNet: an architecture explicitly designed to extract representations in the form of three-place predicates (relations). They evaluate the architecture on a visual relational task called the "Relations Game" which involves comparing Tetris-like shapes according to their appearance, relative posit... | SP:2af62c222666760e3a313b875e9bcb8190fed30b |
An Explicitly Relational Neural Network Architecture | 1 INTRODUCTION . When humans face novel problems , they are able to draw effectively on past experience with other problems that are superficially very different , but that have similarities on a more abstract , structural level . This ability is essential for lifelong , continual learning , and confers on humans a deg... | The authors propose a new neural network architecture that learns to form propositional representations with an explicitly relational structure from raw pixel data. The authors testified the proposed algorithm using the Relations Game, whose aim is to label an image containing a number of objects as True or False accor... | SP:2af62c222666760e3a313b875e9bcb8190fed30b |
Multiagent Reinforcement Learning in Games with an Iterated Dominance Solution | 1 INTRODUCTION . Intelligent agents sharing a common environment are affected by the actions taken by their peers . Using reinforcement learning ( RL ) to derive agent policies becomes challenging since the environment becomes non-stationary for each agent when its peers adapt their behaviour through their learning pro... | This paper studies independent multi-agent reinforcement learning (MARL) in dominance solvable games. The main contribution of this paper is that the authors have proved the convergence to the iterated dominance solution for two RL algorithms: REINFORCE (Section 3.1, binary action case only) and Importance Weighted Mon... | SP:e9dcb99e720c5ddd920ccbeda176536132323a3d |
Multiagent Reinforcement Learning in Games with an Iterated Dominance Solution | 1 INTRODUCTION . Intelligent agents sharing a common environment are affected by the actions taken by their peers . Using reinforcement learning ( RL ) to derive agent policies becomes challenging since the environment becomes non-stationary for each agent when its peers adapt their behaviour through their learning pro... | This paper studies reinforcement learning algorithms in a specific subset of multi-agent environments that are 'dominance solvable'. This means that, given an initial set of strategies in the game, if we iteratively remove 'dominated strategies' (those whose utility is strictly less than another strategy independent of... | SP:e9dcb99e720c5ddd920ccbeda176536132323a3d |
Supervised learning with incomplete data via sparse representations | 1 Introduction . In many machine learning applications sometimes the measurements are noisy or affected by artifacts , resulting in incomplete data samples . Examples of this situation include : self-driving vehicle or robot where objects in the view field can be partially occluded ; recommendation systems built from t... | This manuscript studies supervised learning with incomplete observation of the features, assuming a low-rank structure in the full set of features. The work tackles the problem with a global cost function that optimizes the classifier on observations reconstructed with a dictionary penalized jointly for sparse coding. ... | SP:bf0dc1023654d3b515119e32a8bd9ab97a79398a |
Supervised learning with incomplete data via sparse representations | 1 Introduction . In many machine learning applications sometimes the measurements are noisy or affected by artifacts , resulting in incomplete data samples . Examples of this situation include : self-driving vehicle or robot where objects in the view field can be partially occluded ; recommendation systems built from t... | The authors propose a scheme for simultaneous dictionary learning and classification based on sparse representation of the data within the learned dictionary. Their goal is achieved via optimization of a three-part training cost function that explicitly models the accuracy and sparsity of the sparse model, simultaneous... | SP:bf0dc1023654d3b515119e32a8bd9ab97a79398a |
Gap-Aware Mitigation of Gradient Staleness | 1 INTRODUCTION . The steady growth of deep neural networks over the years has made it impractical to train them from scratch on a single worker ( i.e. , computational device ) . Distributing the computations over several workers can drastically reduce the training time . However , due to the sequential nature of the wi... | This paper studies training large machine learning models in a distributed setup. For such a setup, as the number of workers increases, employing synchronous stochastic gradient descent incurs a significant delay due to the presence of straggling workers. Using asynchronous methods should circumvent the issue of stragg... | SP:96a3d3453cca0349f176d351b653d25c94b77acb |
Gap-Aware Mitigation of Gradient Staleness | 1 INTRODUCTION . The steady growth of deep neural networks over the years has made it impractical to train them from scratch on a single worker ( i.e. , computational device ) . Distributing the computations over several workers can drastically reduce the training time . However , due to the sequential nature of the wi... | The paper introduces a new variant of asynchronous SGD, GA-ASGD, for distributed training. The goal is to mitigate the gradient staleness issue caused by asynchronously applying gradients to an old version of parameters. Prior work addresses this issue by penalizing the learning step of a worker linearly to its missed ... | SP:96a3d3453cca0349f176d351b653d25c94b77acb |
XD: Cross-lingual Knowledge Distillation for Polyglot Sentence Embeddings | Current state-of-the-art results in multilingual natural language inference ( NLI ) are based on tuning XLM ( a pre-trained polyglot language model ) separately for each language involved , resulting in multiple models . We reach significantly higher NLI results with a single model for all languages via multilingual tu... | This paper proposes two improved strategies for fine-tuning XLM (a multilingual variant of BERT) for cross-lingual NLI. First of all, it shows that fine-tuning a single model on the combination of all languages (the original English data from MultiNLI and their MT translation into the rest of languages) performs better... | SP:8208b999d9994f1bbee5f5c4806cc4c7b8b1ea8f |
XD: Cross-lingual Knowledge Distillation for Polyglot Sentence Embeddings | Current state-of-the-art results in multilingual natural language inference ( NLI ) are based on tuning XLM ( a pre-trained polyglot language model ) separately for each language involved , resulting in multiple models . We reach significantly higher NLI results with a single model for all languages via multilingual tu... | First, the authors propose to train a model for natural language inference (NLI) on multiple languages simultaneously. In particular, they translate English examples into all target languages and fine-tune a pretrained language model on all thereby obtained data at once. This is different from the previous state-of-the... | SP:8208b999d9994f1bbee5f5c4806cc4c7b8b1ea8f |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.