paper_name stringlengths 11 170 | text stringlengths 8.07k 307k | summary stringlengths 152 6.16k | paper_id stringlengths 43 43 |
|---|---|---|---|
Input Complexity and Out-of-distribution Detection with Likelihood-based Generative Models | 1 INTRODUCTION . Assessing whether input data is novel or significantly different than the one used in training is critical for real-world machine learning applications . Such data are known as out-of-distribution ( OOD ) inputs , and detecting them should facilitate safe and reliable model operation . This is particul... | This paper analyzes the peculiar case that deep generative models often assign a higher likelihood to other datasets than they were trained on. The running hypothesis here is, that input complexity plays a central role. Measuring a proxy for input complexity shows that it is tightly anticorrelated with likelihood and t... | SP:0caffdb2e566a7b43ac5fdfd8f70fca2ba66d2a8 |
Input Complexity and Out-of-distribution Detection with Likelihood-based Generative Models | 1 INTRODUCTION . Assessing whether input data is novel or significantly different than the one used in training is critical for real-world machine learning applications . Such data are known as out-of-distribution ( OOD ) inputs , and detecting them should facilitate safe and reliable model operation . This is particul... | The paper discusses the inductive biases in generative models that tend to assign higher likelihoods to "less complex" images. In particular, a likelihood based generative model (like Glow or PixelCNN++) trained on a particular dataset has significantly higher likelihood for data that have lower compression ratios (e.g... | SP:0caffdb2e566a7b43ac5fdfd8f70fca2ba66d2a8 |
Adversarial Training and Provable Defenses: Bridging the Gap | 1 INTRODUCTION . The discovery of adversarial examples in deep learning ( Szegedy et al. , 2013 ; Biggio et al. , 2013 ) has increased the importance of creating new training methods which produce accurate and robust neural networks with provable guarantees . Existing work : adversarial and provable defenses Adversaria... | the paper introduces a novel protocol for training neural networks that aims at leveraging the empirical benefits of adversarial training while allowing to certify the robustness of the network using the convex relation approach introduced by Wong & Kolter. The key ingredient is a novel algorithm for layer-wise adversa... | SP:5f2bba1f11065ba65a0b3cfa79db380aa5bba2d8 |
Adversarial Training and Provable Defenses: Bridging the Gap | 1 INTRODUCTION . The discovery of adversarial examples in deep learning ( Szegedy et al. , 2013 ; Biggio et al. , 2013 ) has increased the importance of creating new training methods which produce accurate and robust neural networks with provable guarantees . Existing work : adversarial and provable defenses Adversaria... | This paper was very clearly written and easy to follow. Kudos to the authors. In particular, the experimental evaluation section was exceptionally clear. Thanks to the authors for making the paper so easy to review. The “Main Contributions” section was excellent as well as it allows the reader to quickly understand wha... | SP:5f2bba1f11065ba65a0b3cfa79db380aa5bba2d8 |
N-BEATS: Neural basis expansion analysis for interpretable time series forecasting | We focus on solving the univariate times series point forecasting problem using deep learning . We propose a deep neural architecture based on backward and forward residual links and a very deep stack of fully-connected layers . The architecture has a number of desirable properties , being interpretable , applicable wi... | The paper proposes a DL architecture that achieves better performance on time series prediction. The proposed architecture is relatively straightforward and composes residual blocks. While the paper does achieve superior results, a lot of the text is devoted to comparing to prior work and arguing that DL approaches can... | SP:22aa0e4ed150eb964537313068e43b8b32e26a51 |
N-BEATS: Neural basis expansion analysis for interpretable time series forecasting | We focus on solving the univariate times series point forecasting problem using deep learning . We propose a deep neural architecture based on backward and forward residual links and a very deep stack of fully-connected layers . The architecture has a number of desirable properties , being interpretable , applicable wi... | The paper investigates a pure deep learning architecture for Univariate time series analysis by simply ensembling feed-forward networks, along with the residual stacking mechanism for fluid learning. Each of the generic block consists of 4 FC layers followed by the use of forward and backward predictor to have forecast... | SP:22aa0e4ed150eb964537313068e43b8b32e26a51 |
Meta Reinforcement Learning with Autonomous Inference of Subtask Dependencies | 1 INTRODUCTION . Recently , reinforcement learning ( RL ) systems have achieved super-human performance on many complex tasks ( Mnih et al. , 2015 ; Silver et al. , 2016 ; Van Seijen et al. , 2017 ) . However , these works mostly have been focused on a single known task where the agent can be trained for a long time ( ... | This paper proposes a new meta-reinforcement learning algorithm, MSGI, which focuses on the problem of adapting to unseen hierarchical tasks through interaction with the environment where the external reward is sparse. The authors make use of subtask graph inference to infer the latent subtask representation of a task ... | SP:70dd6ee712270dc93c4d1e1c20c31748e072adf2 |
Meta Reinforcement Learning with Autonomous Inference of Subtask Dependencies | 1 INTRODUCTION . Recently , reinforcement learning ( RL ) systems have achieved super-human performance on many complex tasks ( Mnih et al. , 2015 ; Silver et al. , 2016 ; Van Seijen et al. , 2017 ) . However , these works mostly have been focused on a single known task where the agent can be trained for a long time ( ... | The authors propose a novel meta-rl problem where hierarchical tasks are characterized by a graph describing all sub-tasks and their dependencies. They propose a meta-rl approach to meta-train a policy that quickly infers the subtask graph from new task data. The approach is compared to relevant baselines from both the... | SP:70dd6ee712270dc93c4d1e1c20c31748e072adf2 |
SGD with Hardness Weighted Sampling for Distributionally Robust Deep Learning | 1 INTRODUCTION . In standard deep learning pipelines , a neural network h with parameters θ is trained by minimizing the mean of a per-example loss L over a training dataset { ( xi , yi ) } n i=1 , where xi are the inputs and yi are the labels . This corresponds to Empirical Risk Minimization ( ERM ) , defined as the n... | This paper proposes a method for Distributionally Robust Optimization (DRO). DRO has recently been proposed (Namkoong and Duchi 2016 and others) as a robust learning framework compared to Empirical Risk Minimization (ERM). My analysis of this work in short is that the problem that this paper addresses is interesting a... | SP:2335481648e1f24c1da00afb90176e0c5ab0ea2b |
SGD with Hardness Weighted Sampling for Distributionally Robust Deep Learning | 1 INTRODUCTION . In standard deep learning pipelines , a neural network h with parameters θ is trained by minimizing the mean of a per-example loss L over a training dataset { ( xi , yi ) } n i=1 , where xi are the inputs and yi are the labels . This corresponds to Empirical Risk Minimization ( ERM ) , defined as the n... | This paper studies the Distributionally Robust Optimization (DRO), in the sense that the weights assigned to the training data can change, but the training data itself remains unchanged. They demonstrate that SGD with hardness weighted sampling is a principled and efficient optimization method for DRO in machine learni... | SP:2335481648e1f24c1da00afb90176e0c5ab0ea2b |
Enhancing Attention with Explicit Phrasal Alignments | The attention mechanism is an indispensable component of any state-of-the-art neural machine translation system . However , existing attention methods are often token-based and ignore the importance of phrasal alignments , which are the backbone of phrase-based statistical machine translation . We propose a novel phras... | This paper proposes an extension of the attention module that explicitly incorporates phrase information. Using convolution, attention scores are obtained independently for each n-gram type, and then combined. Transformer models with the proposed phrase attention are evaluated on multiple translation tasks, as well as ... | SP:240313e3722a62fac70228911f264ae16190fef6 |
Enhancing Attention with Explicit Phrasal Alignments | The attention mechanism is an indispensable component of any state-of-the-art neural machine translation system . However , existing attention methods are often token-based and ignore the importance of phrasal alignments , which are the backbone of phrase-based statistical machine translation . We propose a novel phras... | This work aims to incorporate phrase representation into attention mechanism. The proposed method is straightforward (which is good): a convolution window with size n is used to calculate representation for an n-gram, which then replaces the token representation in a standard attention model. The paper implements the m... | SP:240313e3722a62fac70228911f264ae16190fef6 |
FoveaBox: Beyound Anchor-based Object Detection | We present FoveaBox , an accurate , flexible , and completely anchor-free framework for object detection . While almost all state-of-the-art object detectors utilize predefined anchors to enumerate possible locations , scales and aspect ratios for the search of the objects , their performance and generalization ability... | This paper introduces an anchor-free object detection framework that aims at simultaneously predicting the object position and the corresponding boundary. To achieve this, the proposed FoveaBox detector predicts category-sensitive semantic maps for the object existing possibility, and produces category-agnostic boundi... | SP:4e5cd6138722d9bfe662959519d7edadbbd2ea52 |
FoveaBox: Beyound Anchor-based Object Detection | We present FoveaBox , an accurate , flexible , and completely anchor-free framework for object detection . While almost all state-of-the-art object detectors utilize predefined anchors to enumerate possible locations , scales and aspect ratios for the search of the objects , their performance and generalization ability... | The paper introduces foveabox, a method that performs "keypoint" like object detection -- instead of "anchor" based detection (to be discussed later). The idea is simple: predict class labels for pixels that fall within (a reduced version) the GT boxes of the instance; and predict bounding box offsets for those positiv... | SP:4e5cd6138722d9bfe662959519d7edadbbd2ea52 |
Recurrent Neural Networks are Universal Filters | ∀ > 0 , ∃ RNN-based filter , s.t . lim sup k→∞ ∥∥x̂k|k − E [ xk|Yk ] ∥∥ < , where x̂k|k is RNN-based filter ’ s estimate of state xk at step k conditioned on the observation history and E [ xk|Yk ] is the conditional mean of xk , known as the optimal estimate of the state in minimum mean square error sense . As an inte... | This paper shows that RNN (of infinite horizon) can be universal approximators for any stochastic dynamics system. It does that by using the Bayesian framework of filtering, which shows that if sufficient statistics given the observations so far can be computed at each stage of filtering, then the expected hidden stat... | SP:86b5887f47a2acc510a7fb242214a6bece82f682 |
Recurrent Neural Networks are Universal Filters | ∀ > 0 , ∃ RNN-based filter , s.t . lim sup k→∞ ∥∥x̂k|k − E [ xk|Yk ] ∥∥ < , where x̂k|k is RNN-based filter ’ s estimate of state xk at step k conditioned on the observation history and E [ xk|Yk ] is the conditional mean of xk , known as the optimal estimate of the state in minimum mean square error sense . As an inte... | The paper attempts to establish the asymptotic accuracy of the "RNN" (but not the RNN models that are well-known in the literature - see the below comments) as a universal functional approximator. It considers a general state space model and uses feedforward neural nets (FNN) to learn the filtering and forecast distrib... | SP:86b5887f47a2acc510a7fb242214a6bece82f682 |
Scaling Up Neural Architecture Search with Big Single-Stage Models | Neural architecture search ( NAS ) methods have shown promising results discovering models that are both accurate and fast . For NAS , training a one-shot model has became a popular strategy to approximate the quality of multiple architectures ( child models ) using a single set of shared weights . To avoid performance... | The authors propose a search for neural architectures with different resource requirements by training a single model only. Furthermore, models found at the end of the search require no additional post-processing and are ready for deployment. A weight-sharing technique is applied to make this happen. The authors argue ... | SP:915803c7eee23272e85bb7933618a240f57d6c36 |
Scaling Up Neural Architecture Search with Big Single-Stage Models | Neural architecture search ( NAS ) methods have shown promising results discovering models that are both accurate and fast . For NAS , training a one-shot model has became a popular strategy to approximate the quality of multiple architectures ( child models ) using a single set of shared weights . To avoid performance... | This paper presents a method for architecture search in deep neural networks in order to identify scaled-down networks that can operate on resource limited hardware. The approach taken in this paper is different from other approaches, which train a single big model then fine tune smaller models for specific hardware, o... | SP:915803c7eee23272e85bb7933618a240f57d6c36 |
Modelling the influence of data structure on learning in neural networks | 1 INTRODUCTION AND RELATED WORK . A major impediment for understanding the effectiveness of deep neural networks is our lack of mathematical models for the data sets on which neural networks are trained . This lack of tractable models prevents us from analysing the impact of data sets on the training of neural networks... | This paper studies influences of data structures on neural network learning. The data structures discussed in this paper are structured inputs (concentrating on a low-dimensional manifold) versus unstructured ones, as well as the teacher task (labels are obtained as a function of high-dimensional inputs) versus the lat... | SP:cb85203d43368ce095c1f37015bf7af061914a5c |
Modelling the influence of data structure on learning in neural networks | 1 INTRODUCTION AND RELATED WORK . A major impediment for understanding the effectiveness of deep neural networks is our lack of mathematical models for the data sets on which neural networks are trained . This lack of tractable models prevents us from analysing the impact of data sets on the training of neural networks... | The authors consider the general problem of "structure" in datasets--particularly, what are the features of datasets that govern the learning dynamics of neural networks trained to classify that data. They approach this problem by looking at combinations of [iid gaussian, structure] inputs and [teacher, latent] tasks ... | SP:cb85203d43368ce095c1f37015bf7af061914a5c |
Interpretable Complex-Valued Neural Networks for Privacy Protection | 1 INTRODUCTION . Deep neural networks ( DNNs ) have shown superb capabilities to process massive volume of data , and local devices such as mobile phones , medical equipment , Internet of Things ( IoT ) devices have become major data entry points in recent years . Although on-device machine learning has exhibited vario... | In this paper, the authors propose to hide information in phase of the input features. They proposed that if each layer of processing layer, sitting outside of the local unit, is phase preserving, then they can recover the phase back. They propose a modification to the most popular layers in DNN to satisfy that propert... | SP:545da4fe9052319abda88a3ca5ddccaabd2ff2b4 |
Interpretable Complex-Valued Neural Networks for Privacy Protection | 1 INTRODUCTION . Deep neural networks ( DNNs ) have shown superb capabilities to process massive volume of data , and local devices such as mobile phones , medical equipment , Internet of Things ( IoT ) devices have become major data entry points in recent years . Although on-device machine learning has exhibited vario... | This paper proposes a novel way to outsource a part of the information processing in a deep learning model to an untrusted remote location while revealing only little information about the input or final output of the computation. To this end the result of an on-chip encoder (e.g. the first N layers of a ConvNet) is en... | SP:545da4fe9052319abda88a3ca5ddccaabd2ff2b4 |
LEARNING TO LEARN WITH BETTER CONVERGENCE | 1 INTRODUCTION . Optimization is an important problem for almost all tasks in machine learning . For optimizing non-convex problems that arise in machine learning , such as minimizing training loss of a neural network , SGD is still the de-facto algorithm . However , success of SGD often hinges on careful selection of ... | This paper presents several improvements over the existing learning to learn models including Andrychowicz et al. (2016) and Lv et al. (2017). Specifically, this paper analyzes the issues in the original learning to learn paradigm (L2L), including instability during training and bias term issues in the RNN. It proposes... | SP:49577cf8ac5482e49fde9d222b889e5799c53c11 |
LEARNING TO LEARN WITH BETTER CONVERGENCE | 1 INTRODUCTION . Optimization is an important problem for almost all tasks in machine learning . For optimizing non-convex problems that arise in machine learning , such as minimizing training loss of a neural network , SGD is still the de-facto algorithm . However , success of SGD often hinges on careful selection of ... | In this paper, the authors build on the 'learning to learn' work, that aims to leverage deep learning models with optimization algorithms, most commonly with recurrent networks. The goal is to utilize meta-learners that can adapt the optimization update strategy based on data/experience. The authors aim to tackle pro... | SP:49577cf8ac5482e49fde9d222b889e5799c53c11 |
A multi-task U-net for segmentation with lazy labels | 1 INTRODUCTION . Image segmentation has been an active research field in the past decades . Deep learning approaches play an increasingly important role and have become state-of-the-art in various segmentation tasks ( Huang et al. , 2018 ; Khoreva et al. , 2017 ; Tsutsui et al. , 2018 ; Ghosh et al. , 2018 ; Litjens et... | This paper addresses the problem of learning segmentation models in the presence of weak and strong labels of different types. This is an important problem that arises while learning in data-scarce settings. The presented approach optimizes jointly over the various types of labels by treating each one as a task. They e... | SP:b5f024a484f19add02eb52a485c2b4806ff5528e |
A multi-task U-net for segmentation with lazy labels | 1 INTRODUCTION . Image segmentation has been an active research field in the past decades . Deep learning approaches play an increasingly important role and have become state-of-the-art in various segmentation tasks ( Huang et al. , 2018 ; Khoreva et al. , 2017 ; Tsutsui et al. , 2018 ; Ghosh et al. , 2018 ; Litjens et... | This paper proposes a method for semantic segmentation using "lazy" segmentation labels. Lazy labels are defined as coarse labels of the segmented objects. The proposed method is a UNET trained in a multitask fashion whit 3 tasks: object detection, object separation, and object segmentation. The method is trained on 2 ... | SP:b5f024a484f19add02eb52a485c2b4806ff5528e |
Self-Attentional Credit Assignment for Transfer in Reinforcement Learning | 1 INTRODUCTION . To some , intelligence is measured as the capability of transferring knowledge to unprecedented situations . While the notion of intellect itself is hard to define , the ability to reuse learned information is a desirable trait for learning agents . The coffee test ( Goertzel et al. , 2012 ) , presente... | This paper proposes to consider the problem of transfer in the context of sequential decision-making -- in particular reinforcement learning -- from the view-point of learning transferable credit assignment capability. They hypothesize that by learning how to assign credit, structural invariants can be learned which th... | SP:8fa7e40de8cef28991e918a6ff372dfac981df6c |
Self-Attentional Credit Assignment for Transfer in Reinforcement Learning | 1 INTRODUCTION . To some , intelligence is measured as the capability of transferring knowledge to unprecedented situations . While the notion of intellect itself is hard to define , the ability to reuse learned information is a desirable trait for learning agents . The coffee test ( Goertzel et al. , 2012 ) , presente... | This work focuses on credit assignment using a self-attention module for transfer RL problems. Specifically, the reward signals are assigned backward to previous states according to the attention weights. This can be helpful especially when the reward signal is sparse. Experiments on the newly proposed Triggers environ... | SP:8fa7e40de8cef28991e918a6ff372dfac981df6c |
Role of two learning rates in convergence of model-agnostic meta-learning | Model-agnostic meta-learning ( MAML ) is known as a powerful meta-learning method . However , MAML is notorious for being hard to train because of the existence of two learning rates . Therefore , in this paper , we derived a sufficinent condition of inner learning rate α and meta-learning rate β for a simplified MAML ... | The authors study a method to help tuning the two learning rates used in the MAML training algorithm. First, they derive a necessary condition for the convergence of the gradient descent in the single task setting. The condition relies on the eigenvalues of the Hessian of the task loss. This condition is reminiscent of... | SP:5f32c357c2f143c81d31176fe38fae9826e138b3 |
Role of two learning rates in convergence of model-agnostic meta-learning | Model-agnostic meta-learning ( MAML ) is known as a powerful meta-learning method . However , MAML is notorious for being hard to train because of the existence of two learning rates . Therefore , in this paper , we derived a sufficinent condition of inner learning rate α and meta-learning rate β for a simplified MAML ... | This paper studies the two learning rates \alpha and \beta used in Model-Agnostic Meta-Learning (MAML) algorithm by [Finn et al., 2017]. MAML is known to be difficult to train, and part of the reason why is the need to tune the two learning rates. Under simplifications, the paper derives some necessary conditions on \a... | SP:5f32c357c2f143c81d31176fe38fae9826e138b3 |
GPNET: MONOCULAR 3D VEHICLE DETECTION BASED ON LIGHTWEIGHT WHEEL GROUNDING POINT DETECTION NETWORK | 1 INTRODUCTION . 3D location and orientation detection is a basic but challenging problem in computer vision , which focuses on the prediction accuracy of visible and invisible points . It has been applied in many ways , including human action recognition , human-computer interaction , recently popular object detection... | This paper introduces a method to detect cars from a single image. The method imposes several handcrafted constraints specific to the dataset in order to achieve higher improvement and efficiency. These constraints are quite strong and they not generalize to new situations (eg. a car in the sky, a car upside down a car... | SP:535c313f14cbdf5a973f7e61e059f9d27d07a3c2 |
GPNET: MONOCULAR 3D VEHICLE DETECTION BASED ON LIGHTWEIGHT WHEEL GROUNDING POINT DETECTION NETWORK | 1 INTRODUCTION . 3D location and orientation detection is a basic but challenging problem in computer vision , which focuses on the prediction accuracy of visible and invisible points . It has been applied in many ways , including human action recognition , human-computer interaction , recently popular object detection... | I find it very hard to review this paper. The idea of using keypoints to carry pose estimation is is more than 15 years old, and for the car examples reported in this paper, I'm wondering why not just you SURF or SIFT - these would certainly have been reasonable baselines. The convnets cited in this paper are mostly ta... | SP:535c313f14cbdf5a973f7e61e059f9d27d07a3c2 |
Reanalysis of Variance Reduced Temporal Difference Learning | 1 INTRODUCTION . In reinforcement learning ( RL ) , policy evaluation aims to obtain the expected long-term reward of a given policy and plays an important role in identifying the optimal policy that achieves the maximal cumulative reward over time Bertsekas and Tsitsiklis ( 1995 ) ; Dayan and Watkins ( 1992 ) ; Rummer... | The paper is on temporal difference learning, specifically variance reduction of it. As per the claims of the paper, a previous method from (Korda and La, 2015) had technical errors, which the paper corrects and provides a better analysis of variance reduction. In the end, the paper focuses on the variance of the gradi... | SP:9ed2924f9f5890e7e790b6f16702efdd93319421 |
Reanalysis of Variance Reduced Temporal Difference Learning | 1 INTRODUCTION . In reinforcement learning ( RL ) , policy evaluation aims to obtain the expected long-term reward of a given policy and plays an important role in identifying the optimal policy that achieves the maximal cumulative reward over time Bertsekas and Tsitsiklis ( 1995 ) ; Dayan and Watkins ( 1992 ) ; Rummer... | In this paper, the authors study the variance reduced TD (VRTD) algorithm, by Korda and Prashanth (2015) (KP15), for policy evaluation in RL. They first highlight technical errors in the analysis of KP15, and then provide new convergence analysis for this algorithm. The new analysis is based on a new technique to boun... | SP:9ed2924f9f5890e7e790b6f16702efdd93319421 |
On The Difficulty of Warm-Starting Neural Network Training | We design a new algorithm for batch active learning with deep neural network models . Our algorithm , Batch Active learning by Diverse Gradient Embeddings ( BADGE ) , samples groups of points that are disparate and high-magnitude when represented in a hallucinated gradient space , a strategy designed to incorporate bot... | This paper examines the problem of warm-starting the training of neural networks. In particular, a generalization gap arises when the network is trained on the full training set from the start versus being warm-started, where the network is initially (partially) trained on a subset of the training set, then switched to... | SP:076e7698bb0dbd1917516c5c8445d79bf7fead7f |
On The Difficulty of Warm-Starting Neural Network Training | We design a new algorithm for batch active learning with deep neural network models . Our algorithm , Batch Active learning by Diverse Gradient Embeddings ( BADGE ) , samples groups of points that are disparate and high-magnitude when represented in a hallucinated gradient space , a strategy designed to incorporate bot... | The author proposes different approaches to the problem of "warm-started" neural networks. Models trained from scratch on the whole dataset have better performance than "warm-started" models, which are trained with weights initialized from training using part of the available data. The authors change hyperparameters ... | SP:076e7698bb0dbd1917516c5c8445d79bf7fead7f |
Learning robust visual representations using data augmentation invariance | Keywords : deep neural networks ; visual cortex ; invariance ; data augmentation 1 INTRODUCTION . Deep artificial neural networks ( DNNs ) have borrowed much inspiration from neuroscience and are , at the same time , the current best model class for predicting neural responses across the visual system in the brain ( Ki... | Motivated by biological visual systems, this paper investigates whether the representations of convolutional networks for visual recognition are invariant to identity preserving transformations. The results show that empirically they are not, and they further propose a data-augmentation approach to learn this invarianc... | SP:93c014b25cbb01bc2f2808750e351efa1210ed80 |
Learning robust visual representations using data augmentation invariance | Keywords : deep neural networks ; visual cortex ; invariance ; data augmentation 1 INTRODUCTION . Deep artificial neural networks ( DNNs ) have borrowed much inspiration from neuroscience and are , at the same time , the current best model class for predicting neural responses across the visual system in the brain ( Ki... | The paper proposes to explicitly improve the robustness of image-classification models to invariant transformations, via a secondary multi-task objective. The idea is that the secondary objective makes intermediate-layer representations invariant to transformations of the image that should lead to the same classificati... | SP:93c014b25cbb01bc2f2808750e351efa1210ed80 |
AutoQ: Automated Kernel-Wise Neural Network Quantization | 1 INTRODUCTION . Although convolutional neural networks ( CNNs ) have been the dominant approach ( Sandler et al. , 2018 ) to solving a wide variety of problems such as computer vision and recommendation systems , it is challenging to deploy CNNs to mobile devices having only limited hardware resources and tight power ... | This paper proposes a new method for quantizing neural network weights and activations that uses deep reinforcement learning to select the appropriate bitwidth for individual kernels in each layer. The algorithm uses a reward function that weights accuracy of the quantized model with latency, energy, and FPGA area, and... | SP:85bec3a997bfbed2252e9549d7a9856d35890241 |
AutoQ: Automated Kernel-Wise Neural Network Quantization | 1 INTRODUCTION . Although convolutional neural networks ( CNNs ) have been the dominant approach ( Sandler et al. , 2018 ) to solving a wide variety of problems such as computer vision and recommendation systems , it is challenging to deploy CNNs to mobile devices having only limited hardware resources and tight power ... | The paper proposed a method for network quantization. Similar with the work of "HAQ: Hardware-Aware Automated Quantization with Mixed Precision"(CVPR 2019), the proposed method is based on reinforcement learning. The contribution of the work is on the kernel-wise quantization, i.e., assigning different bitwidth to diff... | SP:85bec3a997bfbed2252e9549d7a9856d35890241 |
Stablizing Adversarial Invariance Induction by Discriminator Matching | 1 INTRODUCTION . Extensive studies have demonstrated that deep neural networks ( DNNs ) can uncover complicated variations in data to provide powerful representations that are useful for classification tasks ( Hinton et al. , 2006 ; Krizhevsky et al. , 2012 ) . However , in some scenarios , the learned representation s... | The paper points out that the practical behavior of AII assumes the optimality of the attribute classifier, which is rarely held in practice. And claims that the paper analyzes the practical behavior of AII both theoretically and empirically, indicating that AII has theoretical difficulty as it maximizes variational up... | SP:b700b2e67fe44a82672de943640710d3b2141405 |
Stablizing Adversarial Invariance Induction by Discriminator Matching | 1 INTRODUCTION . Extensive studies have demonstrated that deep neural networks ( DNNs ) can uncover complicated variations in data to provide powerful representations that are useful for classification tasks ( Hinton et al. , 2006 ; Krizhevsky et al. , 2012 ) . However , in some scenarios , the learned representation s... | The paper studies the problem of representation learning under invariance constraints (i.e., the representation should be invariant wrt some attributes). The authors first review the adversarial invariance induction (AII) approach and they point out its limitations and then they propose a novel variant, which introduce... | SP:b700b2e67fe44a82672de943640710d3b2141405 |
CLEVRER: Collision Events for Video Representation and Reasoning | 1 INTRODUCTION . The ability to recognize objects and reason about their behaviors in physical events from videos lies at the core of human cognitive development ( Spelke , 2000 ) . Humans , even young infants , group segments into objects based on motion , and use concepts of object permanence , solidity , and continu... | This paper studies the temporal and causal structures in videos. Specifically, the authors first introduce a new dataset called CLEVRER drawing motivation from CLEVR, a well-known visual reasoning dataset. They further evaluate a set of state-of-the-art methods on the newly introduced dataset to confirm their initial b... | SP:5f7434c2a4e815f70674cd3684bb31de7401a648 |
CLEVRER: Collision Events for Video Representation and Reasoning | 1 INTRODUCTION . The ability to recognize objects and reason about their behaviors in physical events from videos lies at the core of human cognitive development ( Spelke , 2000 ) . Humans , even young infants , group segments into objects based on motion , and use concepts of object permanence , solidity , and continu... | Authors propose a new Dataset CLEVRER, a simulated video dataset involving interaction between objects. It is discussed, the existing state-of-the-art models for visual question answering, doesn’t capture the causal structure between the objects and their claim is supported by their experiments. Authors also proposed ... | SP:5f7434c2a4e815f70674cd3684bb31de7401a648 |
Cross Domain Imitation Learning | 1 INTRODUCTION . Humans possess an astonishing ability to recognize latent structural similarities between behaviors in related but distinct domains , and learn new skills from cross domain demonstrations alone . Not only are we capable of learning from third person observations that have no obvious correspondence to o... | This paper proposes Generative Adversarial MDP Alignment (GAMA) for imitation learning. Given a set of paired MDPs, GAMA learns a state mapping f and an action mapping g such that one MDP can be reduced to another. For a new test MDP pair (x,y) where expert demonstrations are available for y, GAMA can use f to map a st... | SP:0f655731e050572011b5ac4f394d725e1b35fc56 |
Cross Domain Imitation Learning | 1 INTRODUCTION . Humans possess an astonishing ability to recognize latent structural similarities between behaviors in related but distinct domains , and learn new skills from cross domain demonstrations alone . Not only are we capable of learning from third person observations that have no obvious correspondence to o... | The paper proposes a learning approach for zero-shot imitation learning in an RL setting across domains with different embodiments and viewpoint mismatch. The proposed approach involves two steps, alignment and adaptation. In contrast to previous work, the alignment between domains, represented as MDPs, is learned from... | SP:0f655731e050572011b5ac4f394d725e1b35fc56 |
Provable Representation Learning for Imitation Learning via Bi-level Optimization | 1 INTRODUCTION . Humans can often learn from experts quickly and with a few demonstrations and we would like our artificial agents to do the same . However , even for simple imitation learning tasks , the current state-of-the-art methods require thousand of demonstrations . Humans do not learn new skills from scratch .... | The paper tackles the representation learning problem where the aim is to learn a generic representation that is useful for a variety of downstream tasks. A two-level optimization framework is proposed: an inner optimization over the specific problem-at-hand, and an outer optimization over other similar problems. The p... | SP:e108dc35910955986e9f58d77965d34863c5374b |
Provable Representation Learning for Imitation Learning via Bi-level Optimization | 1 INTRODUCTION . Humans can often learn from experts quickly and with a few demonstrations and we would like our artificial agents to do the same . However , even for simple imitation learning tasks , the current state-of-the-art methods require thousand of demonstrations . Humans do not learn new skills from scratch .... | This paper theoretically explores reinforcement/imitation learning via representation learning. The key theoretical question being investigated is the relationship between representation learning in a multi-task/meta learning setup and its dependence to the sample/task complexity. The paper sets up the problem in bilev... | SP:e108dc35910955986e9f58d77965d34863c5374b |
The Dual Information Bottleneck | 1 INTRODUCTION . 1.1 THE INFORMATION BOTTLENECK METHOD . The Information Bottleneck ( IB ) method ( Tishby et al. , 1999 ) , is an information-theoretic framework for describing efficient representations of a large scale “ input ” random variable X ( input patterns ) , for predicting an “ output ” variable Y ( desired ... | The information bottleneck (IB) is an information theoretic principle for optimizing a mapping (encoding, e.g. clustering) of an input, to trade off two kinds of mutual information: minimize mutual information between the original input and the mapped version (to compress the input), and maximize mutual information bet... | SP:998b957bfb35d2c7c7aa08e433fb4aed556e2c44 |
The Dual Information Bottleneck | 1 INTRODUCTION . 1.1 THE INFORMATION BOTTLENECK METHOD . The Information Bottleneck ( IB ) method ( Tishby et al. , 1999 ) , is an information-theoretic framework for describing efficient representations of a large scale “ input ” random variable X ( input patterns ) , for predicting an “ output ” variable Y ( desired ... | This paper proposes a new "dual" variant of the Information Bottleneck framework. The IB framework has been the subject of many papers in past years, with a focus on understanding the inner workings of deep learning. The framework poses machine learning as optimizing an internal representation to tradeoff between retai... | SP:998b957bfb35d2c7c7aa08e433fb4aed556e2c44 |
Federated Adversarial Domain Adaptation | Federated learning improves data privacy and efficiency in machine learning performed over networks of distributed devices , such as mobile phones , IoT and wearable devices , etc . Yet models trained with federated learning can still fail to generalize to new devices due to the problem of domain shift . Domain shift o... | This paper introduces an unsupervised federated domain adaptation (UFDA) problem and proposes a new model called Federated Adversarial Domain Adaptation (FADA) to transfer the knowledge learned from distributed source domains to an unlabeled target domain. This paper uses a dynamic attention mechanism by leveraging the... | SP:eaae68b4effc3ed45f560b74377bdcb524220f81 |
Federated Adversarial Domain Adaptation | Federated learning improves data privacy and efficiency in machine learning performed over networks of distributed devices , such as mobile phones , IoT and wearable devices , etc . Yet models trained with federated learning can still fail to generalize to new devices due to the problem of domain shift . Domain shift o... | The authors present a novel algorithm for dealing with domain adaptation in the setting of federated learning (classification, specifically). That is, they tackle the issue of learning a model on a new domain when access to the data points used in training the source models is not possible due to privacy constraints. T... | SP:eaae68b4effc3ed45f560b74377bdcb524220f81 |
Learning Compact Reward for Image Captioning | Adversarial learning has shown its advances in generating natural and diverse descriptions in image captioning . However , the learned reward of existing adversarial methods is vague and ill-defined due to the reward ambiguity problem . In this paper , we propose a refined Adversarial Inverse Reinforcement Learning ( r... | The authors propose using a recent method for adversarial inverse reinforcement learning (AIRL) for the task for generating high-quality image captions. Leveraging the GAN framework, a discriminator is trained to distinguish real captions from those produced by the generator, while the generator is optimized with polic... | SP:c6f58fd967bd81a9b46c030dc5b2d1d16670e057 |
Learning Compact Reward for Image Captioning | Adversarial learning has shown its advances in generating natural and diverse descriptions in image captioning . However , the learned reward of existing adversarial methods is vague and ill-defined due to the reward ambiguity problem . In this paper , we propose a refined Adversarial Inverse Reinforcement Learning ( r... | This paper proposes a refined Adversarial Inverse Reinforcement Learning (rAIRL) to remedy the reward ambiguity by decoupling the reward for each word in a sentence, while the existing methods that utilize reinforcement learning to optimize evaluation score handle only sentence-level rewards. Furthermore, a conditional... | SP:c6f58fd967bd81a9b46c030dc5b2d1d16670e057 |
On Understanding Knowledge Graph Representation | 1 INTRODUCTION . Knowledge graphs are large repositories of binary relations between words ( or entities ) in the form of fact triples ( subject , relation , object ) . Many models have been developed for learning representations of entities and relations in knowledge graphs , such that known facts can be recalled and ... | This paper proposes to provide a detailed study on the explainability of link prediction (LP) models by utilizing a recent interpretation of word embeddings. More specifically, the authors categorize the relations in KG into three categories (R, S, C) using the correlation between the semantic relation between two word... | SP:81ec2c343c6a792251a62e680683c75ef0662ad6 |
On Understanding Knowledge Graph Representation | 1 INTRODUCTION . Knowledge graphs are large repositories of binary relations between words ( or entities ) in the form of fact triples ( subject , relation , object ) . Many models have been developed for learning representations of entities and relations in knowledge graphs , such that known facts can be recalled and ... | The paper attempts to understand the latent structure underlying knowledge graph embedding methods. The work can be seen as an extension of understanding of PMI-based word embedding methods. They categorize knowledge graph relations into three categories based on their relation conditions: Relatedness (R), Specialisati... | SP:81ec2c343c6a792251a62e680683c75ef0662ad6 |
A Coordinate-Free Construction of Scalable Natural Gradient | 1 INTRODUCTION . Most neural networks are trained using stochastic gradient descent ( SGD ) ( Bottou & Bousquet , 2007 ) , or variants thereof which adapt step sizes for individual dimensions ( Duchi et al. , 2011 ; Kingma & Ba , 2015 ) . One well-known deficiency of SGD is that the updates are sensitive to the paramet... | This paper analyzes the invariance properties of the K-FAC algorithm by reconstructing the algorithm in a coordinate-free way where the neural network is viewed as a series of affine mappings alternating with nonlinear activation functions. It converts the original metric into an approximate metric, whose coordinate re... | SP:cf57db4d77bafc3f7741422fe54e9eb2d1ab7051 |
A Coordinate-Free Construction of Scalable Natural Gradient | 1 INTRODUCTION . Most neural networks are trained using stochastic gradient descent ( SGD ) ( Bottou & Bousquet , 2007 ) , or variants thereof which adapt step sizes for individual dimensions ( Duchi et al. , 2011 ; Kingma & Ba , 2015 ) . One well-known deficiency of SGD is that the updates are sensitive to the paramet... | Natural gradient (NG) has been proven efficient in statistical learning, and one of its attractive properties is being invariant under smooth transformations of the parameter space. Computing NG is often difficult as one has to derive the Fisher matrix and its inverse. K-FAC offers an approximate method for approximati... | SP:cf57db4d77bafc3f7741422fe54e9eb2d1ab7051 |
Efficient Exploration via State Marginal Matching | 1 INTRODUCTION . Reinforcement learning ( RL ) algorithms must be equipped with exploration mechanisms to effectively solve tasks with limited reward signals . These tasks arise in many real-world applications where providing human supervision is expensive . The inability of current RL algorithms to adequately explore ... | The paper proposes to frame exploration in reinforcement learning as a distribution matching problem. More specifically, the proposed method (SMM) aims to minimize the reverse KL between the state distribution induced by the policy and a desired state distribution. The desired state distribution can be used to guide ex... | SP:b4874a9e8d55c290a6fe83bb7d88be93e31d9842 |
Efficient Exploration via State Marginal Matching | 1 INTRODUCTION . Reinforcement learning ( RL ) algorithms must be equipped with exploration mechanisms to effectively solve tasks with limited reward signals . These tasks arise in many real-world applications where providing human supervision is expensive . The inability of current RL algorithms to adequately explore ... | This paper proposes to optimize the state marginal distribution to match a target distribution for the purposes of exploration. This target distribution could be uniform or could encode prior knowledge about downstream tasks. This matching can be done by iteratively fitting a density model on the historical data from t... | SP:b4874a9e8d55c290a6fe83bb7d88be93e31d9842 |
Generalization bounds for deep convolutional neural networks | We prove bounds on the generalization error of convolutional networks . The bounds are in terms of the training loss , the number of parameters , the Lipschitz constant of the loss and the distance from the weights to the initial weights . They are independent of the number of pixels in the input , and the height and w... | The paper describes new norm-based generalization bounds that were specifically adapted to convolutional neural networks. Since convolutional neural networks do not explicitly depend on the input dimension, these bounds share the same property. Further additional improvement over Bartlett et al. ‘17 bound, is that this... | SP:25502178fec50290fe1b6f1fa099df8e5d5d6751 |
Generalization bounds for deep convolutional neural networks | We prove bounds on the generalization error of convolutional networks . The bounds are in terms of the training loss , the number of parameters , the Lipschitz constant of the loss and the distance from the weights to the initial weights . They are independent of the number of pixels in the input , and the height and w... | This paper studied the generalization power of CNNs and showed several upper bounds of generalization errors. Their results have two characteristics. First, the bounds are in terms of the quantity that is independent of the input dimension (size-free). Second, the upper bounds involve the distance between initial and l... | SP:25502178fec50290fe1b6f1fa099df8e5d5d6751 |
Data-Driven Approach to Encoding and Decoding 3-D Crystal Structures | Generative models have achieved impressive results in many domains including image and text generation . In the natural sciences , generative models have lead to rapid progress in automated drug discovery . Many of the current methods focus on either 1-D or 2-D representations of typically small , drug-like molecules .... | PAPER SUMMARY: This paper addresses the problem of encoding and decoding 3D chemical structures, with the ultimate goal of generating 3D crystal structures. The authors propose an auto-encoder framework for encoding the 3D locations of atoms in the crystal to a latent representation and then decoding that representatio... | SP:ab6f2c56bfe6fb6c479345d113e647ba9acf2035 |
Data-Driven Approach to Encoding and Decoding 3-D Crystal Structures | Generative models have achieved impressive results in many domains including image and text generation . In the natural sciences , generative models have lead to rapid progress in automated drug discovery . Many of the current methods focus on either 1-D or 2-D representations of typically small , drug-like molecules .... | The authors describe a method to encode and decode the position of atoms in 3-D molecules. An encoder-decoder architecture is used to create a representation of a molecule and to reconstruct the molecule from its representation. Then a second Neural Network segments the output and assigns an atomic number. Prior work o... | SP:ab6f2c56bfe6fb6c479345d113e647ba9acf2035 |
NPTC-net: Narrow-Band Parallel Transport Convolutional Neural Network on Point Clouds | 1 INTRODUCTION . Convolution is one of the most widely used operators in applied mathematics , computer science and engineering . It is also the most important building block of Convolutional Neural Netowrks ( CNNs ) which are the main driven force in the recent success of deep learning LeCun et al . ( 2015 ) ; Goodfel... | The authors propose NPTC a new convolutional operator for 3D point clouds embedded on a 2D manifold based no parallel transport defined by a narrow-band approximation. The method combines voxelization within a local neighborhood in 3D (narrow band) and geometric convolution. Experimental results are presneted for a ran... | SP:7506acf48272f5455ce8192cda8ccd390ef2c618 |
NPTC-net: Narrow-Band Parallel Transport Convolutional Neural Network on Point Clouds | 1 INTRODUCTION . Convolution is one of the most widely used operators in applied mathematics , computer science and engineering . It is also the most important building block of Convolutional Neural Netowrks ( CNNs ) which are the main driven force in the recent success of deep learning LeCun et al . ( 2015 ) ; Goodfel... | This paper proposes Narrow-Band Parallel Transport Convolution (NPTC) for point cloud data. The general idea is to use gradients of some distance function to define the vector field. The authors use voxelization to approximate the point cloud in a narrow-band covering the point cloud so that distance function can be ca... | SP:7506acf48272f5455ce8192cda8ccd390ef2c618 |
Adversarial AutoAugment | 1 INTRODUCTION . Massive amount of data have promoted the great success of deep learning in academia and industry . The performance of deep neural networks ( DNNs ) would be improved substantially when more supervised data is available or better data augmentation method is adapted . Data augmentation such as rotation ,... | This paper proposes a technique called Adversarial AutoAugment which dynamically learns good data augmentation policies during training. An adversarial approach is used: a target network tries to achieve good classification performance on a training set, while a policy network attempts to foil the target network by dev... | SP:08957836ae755ff67421cbefd5e48859bd403a83 |
Adversarial AutoAugment | 1 INTRODUCTION . Massive amount of data have promoted the great success of deep learning in academia and industry . The performance of deep neural networks ( DNNs ) would be improved substantially when more supervised data is available or better data augmentation method is adapted . Data augmentation such as rotation ,... | This paper describes a method to learn data augmentation policies using an adversarial loss. It builds on the AutoAugment method. In AutoAugment, an augmentation policy generator is trained by reinforcement learning. At each iteration, a classifier network is trained from scratch based on the current augmentation polic... | SP:08957836ae755ff67421cbefd5e48859bd403a83 |
DiffTaichi: Differentiable Programming for Physical Simulation | 1 INTRODUCTION . Differentiable physical simulators are effective components in machine learning systems . For example , de Avila Belbute-Peres et al . ( 2018a ) and Hu et al . ( 2019b ) have shown that controller optimization with differentiable simulators converges one to four orders of magnitude faster than model-fr... | This paper introduces DiffSim, a programming language for high-performance differentiable physics simulations. The paper demonstrates 10 different simulations with controller optimization. It shows that the proposed language is easier to use and faster than the other alternatives, such as CUDA and TensorFlow. At the en... | SP:81d02cae96de58b1b7c74e7522cc82f61b782417 |
DiffTaichi: Differentiable Programming for Physical Simulation | 1 INTRODUCTION . Differentiable physical simulators are effective components in machine learning systems . For example , de Avila Belbute-Peres et al . ( 2018a ) and Hu et al . ( 2019b ) have shown that controller optimization with differentiable simulators converges one to four orders of magnitude faster than model-fr... | This paper describes DiffSim, a differentiable programming system for learning with physical simulation. The system (built on the Taichi system) allows users to specify a forward simulation in a Python-like syntax, after which the program is compiled and iteratively run in both forward-mode and gradients computed for s... | SP:81d02cae96de58b1b7c74e7522cc82f61b782417 |
Fast Neural Network Adaptation via Parameter Remapping and Architecture Search | 1 INTRODUCTION . Deep convolutional neural networks have achieved great successes in computer vision tasks such as image classification ( Krizhevsky et al. , 2012 ; He et al. , 2016 ; Howard et al. , 2017 ) , semantic segmentation ( Long et al. , 2015 ; Ronneberger et al. , 2015 ; Chen et al. , 2017b ) and object detec... | This paper provides a new technique to adapt a source neural network performed well on classification task to image segmentation and objective detection tasks via the author called parameter-remapping trick. The parameter remapping uses weights from the source neural network to the two-stages: architecture adaption pha... | SP:235671d38b4e5b8b3d978b285ffc60f22f30889b |
Fast Neural Network Adaptation via Parameter Remapping and Architecture Search | 1 INTRODUCTION . Deep convolutional neural networks have achieved great successes in computer vision tasks such as image classification ( Krizhevsky et al. , 2012 ; He et al. , 2016 ; Howard et al. , 2017 ) , semantic segmentation ( Long et al. , 2015 ; Ronneberger et al. , 2015 ; Chen et al. , 2017b ) and object detec... | The paper proposes a method called FNA (fast network adaptation), which takes a pretrained image classification network, and produces a network for the task of object detection/semantic segmentation. The process consists of three phases: Network Expansion, Architecture Adaptation and Parameters Adaptation, and uses the... | SP:235671d38b4e5b8b3d978b285ffc60f22f30889b |
Discrepancy Ratio: Evaluating Model Performance When Even Experts Disagree on the Truth | 1 INTRODUCTION . The canonical supervised machine learning paradigm assumes the presence of both inputs and their corresponding , unambiguous outputs . In practice , the vast majority of datasets exhibit some degree of label noise . Acknowledgement of this fact has resulted in the development of algorithms tailored to ... | This paper proposes a novel metric "discrepancy ratio" for evaluating the performance of a model where the ground truth for each data point comes from many expert-yet-imperfect annotators. The authors suggested that this problem has remained largely unexplored. The proposed metric is intuitive and is easy-to-use. The a... | SP:f819e7a08540c2ebf35d975caf628971376e5c82 |
Discrepancy Ratio: Evaluating Model Performance When Even Experts Disagree on the Truth | 1 INTRODUCTION . The canonical supervised machine learning paradigm assumes the presence of both inputs and their corresponding , unambiguous outputs . In practice , the vast majority of datasets exhibit some degree of label noise . Acknowledgement of this fact has resulted in the development of algorithms tailored to ... | This paper proposed to evaluate model performance when the ground truth labels were not available and noisy labels provided by multiple uncertain experts were provided instead. The proposed evaluation metric, called discrepancy ratio, is defined as the ratio between the average model-annotator discrepancy and the avera... | SP:f819e7a08540c2ebf35d975caf628971376e5c82 |
Mixed-curvature Variational Autoencoders | Euclidean geometry has historically been the typical “ workhorse ” for machine learning applications due to its power and simplicity . However , it has recently been shown that geometric spaces with constant non-zero curvature improve representations and performance on a variety of data types and downstream tasks . Con... | This paper is about developing VAEs in non-Euclidean spaces. Fairly recently, ML researchers have developed non-Euclidean embeddings, initially in hyperbolic space (constant negative curvature), and then in product spaces that have varying curvatures. These ideas were developed for embeddings, and recent attempts have ... | SP:44aca1ce39be826e389afbff70936eb0ef774f8f |
Mixed-curvature Variational Autoencoders | Euclidean geometry has historically been the typical “ workhorse ” for machine learning applications due to its power and simplicity . However , it has recently been shown that geometric spaces with constant non-zero curvature improve representations and performance on a variety of data types and downstream tasks . Con... | This paper introduces a general formulation of the notion of a VAE with a latent space composed by a curved manifold. It follows the current trend of learning representations on curved spaces by proposing a formulation of the latent distributions of the VAE in a variety of fixed-curvature spaces, and introduces an appr... | SP:44aca1ce39be826e389afbff70936eb0ef774f8f |
Prototype Recalls for Continual Learning | 1 INTRODUCTION . Continual learning , also known as lifelong learning , is the crucial ability for humans to continually acquire and transfer new knowledge across their lifespans while retaining previously learnt experiences Hassabis et al . ( 2017 ) . This ability is also critical for artificial intelligence ( AI ) sy... | Paper proposes a method for continual learning. The method is based on the learning of a metric space where classes are represented by prototypes in this space. To prevent forgetting the method proposes to perform prototype recall, aiming to keep prototypes in the same location in embedding space (Fig 1b). The method i... | SP:58b67f1e081e61982d524768c88f3754c3470e0a |
Prototype Recalls for Continual Learning | 1 INTRODUCTION . Continual learning , also known as lifelong learning , is the crucial ability for humans to continually acquire and transfer new knowledge across their lifespans while retaining previously learnt experiences Hassabis et al . ( 2017 ) . This ability is also critical for artificial intelligence ( AI ) sy... | The proposed method addresses continual learning, by learning a mapping from the input space to an embedding space, and employing a loss that encourages clustering the embeddings by class (and task?) around some centroids called prototypes. Catastrophic forgetting is mitigated by adding a penalty term that is proportio... | SP:58b67f1e081e61982d524768c88f3754c3470e0a |
Pre-training as Batch Meta Reinforcement Learning with tiMe | 1 INTRODUCTION . Deep Reinforcement Learning algorithms still require millions of environment interactions to obtain reasonable performance , hindering their applications ( Mnih et al. , 2015 ; Lillicrap et al. , 2016 ; Vuong et al. , 2018 ; Fujimoto et al. , 2018b ; Jaderberg et al. , 2018 ; Arulkumaran et al. , 2019 ... | The paper studies batch meta learning, i.e. the problem of using a fixed experience from past tasks to learn a policy which can quickly adapt to a new related task. The proposed method combines the techniques of Fujimoto et al. (2018) for stabilizing batch off-policy learning with ideas from Rakelly et al. (2019) for l... | SP:27b73a836058b5e3cf5430f5c64fec2d1475da1b |
Pre-training as Batch Meta Reinforcement Learning with tiMe | 1 INTRODUCTION . Deep Reinforcement Learning algorithms still require millions of environment interactions to obtain reasonable performance , hindering their applications ( Mnih et al. , 2015 ; Lillicrap et al. , 2016 ; Vuong et al. , 2018 ; Fujimoto et al. , 2018b ; Jaderberg et al. , 2018 ; Arulkumaran et al. , 2019 ... | This paper studies the meta-RL problem in the off-policy, batch learning setting. Batch-RL is the setting in which a policy is learned entirely offline, that is, without interaction with the environment and given only trajectories collected by some policy. Compared to RL, Meta-RL involves the additional challenge of ta... | SP:27b73a836058b5e3cf5430f5c64fec2d1475da1b |
Graph Convolutional Reinforcement Learning | 1 INTRODUCTION . Cooperation is a widespread phenomenon in nature from viruses , bacteria , and social amoebae to insect societies , social animals , and humans ( Melis & Semmann , 2010 ) . Human exceeds all other species in terms of range and scale of cooperation . The development of human cooperation is facilitated b... | This paper introduces Graph Convolutional Reinforcement Learning (referred to as DGN). DGN is a Deep Q-Learning (DQN) agent structured as a graph neural network / graph convolutional network with multi-head dot product attention as a message aggregation function. Graphs are obtained based on spatial neighborhoods (e.g.... | SP:7dc520ce87edf76ac948de085da9855d2f32c7ab |
Graph Convolutional Reinforcement Learning | 1 INTRODUCTION . Cooperation is a widespread phenomenon in nature from viruses , bacteria , and social amoebae to insect societies , social animals , and humans ( Melis & Semmann , 2010 ) . Human exceeds all other species in terms of range and scale of cooperation . The development of human cooperation is facilitated b... | This paper proposes an algorithm allowing "cooperation" between agents in multi-agent reinforcement learning, modeling agents as nodes in a graph. Each agent having only a partial view of the environment, the proposed algorithm uses multi-head attention as a (graph) convolution kernel but otherwise remains similar to t... | SP:7dc520ce87edf76ac948de085da9855d2f32c7ab |
A Dynamic Approach to Accelerate Deep Learning Training | 1 INTRODUCTION . The use of Deep Neural Networks ( DNNs ) is becoming ubiquitous in areas like computer vision ( Krizhevsky et al. , 2012 ; Szegedy et al. , 2015 ) , speech recognition ( Hinton et al. , 2012 ) , or language translation ( Wu et al. , 2016 ) . DNNs display very remarkable pattern detection capacities and... | This proposes two techniques to replace mixed-precision arithmetic with half-precision training for a large part of the training process. In the first approach, the authors simply switch all mixed-precision operations with half-precision operations, and can achieve performances slightly lower than SOTA. In the second a... | SP:52cbd90cb8f3de333a12f8c42cad968d36b509c3 |
A Dynamic Approach to Accelerate Deep Learning Training | 1 INTRODUCTION . The use of Deep Neural Networks ( DNNs ) is becoming ubiquitous in areas like computer vision ( Krizhevsky et al. , 2012 ; Szegedy et al. , 2015 ) , speech recognition ( Hinton et al. , 2012 ) , or language translation ( Wu et al. , 2016 ) . DNNs display very remarkable pattern detection capacities and... | The author(s) propose to accelerate the training of deep neural networks while also maintain the performance of the trained model by switching between fully half-precision computation and mixed-precision computation. Compared to the commonly-used mixed-precision training strategy, the proposed method can accelerate the... | SP:52cbd90cb8f3de333a12f8c42cad968d36b509c3 |
PairNorm: Tackling Oversmoothing in GNNs | The performance of graph neural nets ( GNNs ) is known to gradually decrease with increasing number of layers . This decay is partly attributed to oversmoothing , where repeated graph convolutions eventually make node embeddings indistinguishable . We take a closer look at two different interpretations , aiming to quan... | It is known that GNNs are vulnerable to the oversmoothing problem, in which feature vectors on nodes get closer as we increase the number of (message passing type graph convolution layers). This paper proposed PairNorm, which is a normalization layer for GNNs to tackle this problem. The idea is to pull apart feature ve... | SP:d11f0eb42f1ea12686290a2936b5aee262c8d84d |
PairNorm: Tackling Oversmoothing in GNNs | The performance of graph neural nets ( GNNs ) is known to gradually decrease with increasing number of layers . This decay is partly attributed to oversmoothing , where repeated graph convolutions eventually make node embeddings indistinguishable . We take a closer look at two different interpretations , aiming to quan... | The article "PairNorm: Tackling Oversmoothing in GNNs" considers the interesting phenomenon of performance degradation of graph neural network when the depth of the network increases beyond the values of 2-4. The authors argue that one of the reasons for such behavior is so-called "oversmoothing", when intermediate rep... | SP:d11f0eb42f1ea12686290a2936b5aee262c8d84d |
CNAS: Channel-Level Neural Architecture Search | 1 INTRODUCTION . Nowadays , deep neural networks ( DNNs ) are used extensively and successfully in many fields and applications such as computer vision , speech recognition , machine translation , and automated vehicles . Designing DNNs often requires significant architecture engineering , a large amount of trial and e... | This paper aims to search a sparse but competitive architecture with using a single fixed type of operation by proposing a channel-level neural architecture search (CNAS). Different from most previous NAS works, this paper conducts NAS process on channel-level such that different cell has different topology. CNAS provi... | SP:f077ef67a022348d5d4d455cb313a691cfe63e47 |
CNAS: Channel-Level Neural Architecture Search | 1 INTRODUCTION . Nowadays , deep neural networks ( DNNs ) are used extensively and successfully in many fields and applications such as computer vision , speech recognition , machine translation , and automated vehicles . Designing DNNs often requires significant architecture engineering , a large amount of trial and e... | This paper aims to propose a novel framework for neural architecture search. Although there have been many solutions in the literature, the authors try to build a NAS model that is sparse in structure while being similarly effective as conventional dense models. The method is straightforward - they select a single fix... | SP:f077ef67a022348d5d4d455cb313a691cfe63e47 |
Mode Connectivity and Sparse Neural Networks | 1 INTRODUCTION . The lottery ticket hypothesis ( Frankle & Carbin , 2019 ) conjectures that neural networks contain sparse subnetworks that are capable of training in isolation from initialization to full accuracy . The sole empirical evidence in support of the lottery ticket hypothesis is a series of experiments using... | This paper works on empirically demonstrating the connection between model connectivity and the lottery ticket hypothesis, which are individually explored in the literature. Here the model connectivity refers to the fact that SGD produces different solutions (from the randomness, such as data ordering) that are connect... | SP:b403e36027a1f260c7daead40764de7984c943ef |
Mode Connectivity and Sparse Neural Networks | 1 INTRODUCTION . The lottery ticket hypothesis ( Frankle & Carbin , 2019 ) conjectures that neural networks contain sparse subnetworks that are capable of training in isolation from initialization to full accuracy . The sole empirical evidence in support of the lottery ticket hypothesis is a series of experiments using... | This paper empirically presents a very interesting connection between two also very interesting phenomena (mode connectivity and lottery ticket hypothesis), while removing a previous limitation of the lottery ticket hypothesis on larger networks. through a good amount of experiments, the authors empirically showed thes... | SP:b403e36027a1f260c7daead40764de7984c943ef |
Compressive Transformers for Long-Range Sequence Modelling | 1 INTRODUCTION . Humans have a remarkable ability to remember information over long time horizons . When reading a book , we build up a compressed representation of the past narrative , such as the characters and events that have built up the story so far . We can do this even if they are separated by thousands of word... | This paper proposes a way to compress past hidden states for modeling long sequences. Attention is used to query the compressed representation. The authors introduce several methods for compression such as convolution, pooling etc. The outcome is a versatile model that enables long-range sequence modeling, achieving st... | SP:2237245aeb115eb318e447d63ab3a4614d8eec06 |
Compressive Transformers for Long-Range Sequence Modelling | 1 INTRODUCTION . Humans have a remarkable ability to remember information over long time horizons . When reading a book , we build up a compressed representation of the past narrative , such as the characters and events that have built up the story so far . We can do this even if they are separated by thousands of word... | This paper investigates a so-called "compressive transformer" approach. The idea is to compress distant past memories into a coarse-grained representation while keeping a fine-grained representation for close past memories. A variety of compression techniques and training strategies have been investigated in the paper... | SP:2237245aeb115eb318e447d63ab3a4614d8eec06 |
Assessing Generalization in TD methods for Deep Reinforcement Learning | 1 INTRODUCTION . Deep neural networks ( DNNs ) trained on supervised learning tasks using i.i.d . data have shown the capacity to learn quickly even from a small amount of samples ( Hardt et al. , 2016 ) . Intuitively , this is due to each sample also providing information about the estimate corresponding to other samp... | This paper performs an empirical evaluation of generalization by TD methods with neural nets as function approximators. To quantify generalization, the paper considers the change in the loss function at similar states to the one where the update rule is being applied (where “similar” is usually defined as nearby in tim... | SP:a11a05bf95d8dcd7adb929912430615c73f4b531 |
Assessing Generalization in TD methods for Deep Reinforcement Learning | 1 INTRODUCTION . Deep neural networks ( DNNs ) trained on supervised learning tasks using i.i.d . data have shown the capacity to learn quickly even from a small amount of samples ( Hardt et al. , 2016 ) . Intuitively , this is due to each sample also providing information about the estimate corresponding to other samp... | The manuscript is analyzing the "generalization" in TD(lambda) methods. It includes supervised learning from trajectories, on-policy imitation learning, and basic RL setting. Moreover, memoization performance has also been measured. Main conclusion is the fact that TD(0) performs very similar to tabular learning failin... | SP:a11a05bf95d8dcd7adb929912430615c73f4b531 |
AssembleNet: Searching for Multi-Stream Neural Connectivity in Video Architectures | 1 INTRODUCTION . Learning to represent videos is a challenging problem . Because a video contains spatio-temporal data , its representation is required to abstract both appearance and motion information . This is particularly important for tasks such as activity recognition , as understanding detailed semantic contents... | This paper is a neural architecture search paper. In particular, it applies this to finding better neural architectures for video understanding, emphasizing exploring the video temporal resolutions needed and how to combine intermediate representations capturing appearance and motion. It introduces a somewhat new algor... | SP:f6624fed0b38b3937355e3b4c9e4c1070d60dcc3 |
AssembleNet: Searching for Multi-Stream Neural Connectivity in Video Architectures | 1 INTRODUCTION . Learning to represent videos is a challenging problem . Because a video contains spatio-temporal data , its representation is required to abstract both appearance and motion information . This is particularly important for tasks such as activity recognition , as understanding detailed semantic contents... | This paper aims to adapt the standard neural architecture search scheme to search a two-input convolutional neural network for video representations. To this end, the paper formulates a direct acyclic graph with two input nodes (for RGB image and optical flow), where each node represents some pre-composed layers and ed... | SP:f6624fed0b38b3937355e3b4c9e4c1070d60dcc3 |
Learning Disentangled Representations for CounterFactual Regression | We consider the challenge of estimating treatment effects from observational data ; and point out that , in general , only some factors based on the observed covariates X contribute to selection of the treatment T , and only some to determining the outcomes Y . We model this by considering three underlying sources of {... | The authors consider the problem of estimating average treatment effects when observed X and treatment T causes Y. Observational data for X,T,Y is available and strong ignorability is assumed. Previous work (Shalit et al 2017) introduced learning a representation that is invariant in distribution across treatment an... | SP:1fcf3b2eec374cb379819564c4dbf5cfabe3ff8a |
Learning Disentangled Representations for CounterFactual Regression | We consider the challenge of estimating treatment effects from observational data ; and point out that , in general , only some factors based on the observed covariates X contribute to selection of the treatment T , and only some to determining the outcomes Y . We model this by considering three underlying sources of {... | The paper proposes an algorithm that identifies disentangled representation to find out an individual treatment effect. A very specific model that tries to find out the underlying dynamics of such a problem is proposed and is learned by minimizing a suggested objective that takes the strengths of previous approaches. T... | SP:1fcf3b2eec374cb379819564c4dbf5cfabe3ff8a |
Machine Truth Serum | 1 INTRODUCTION . Wisdom of the crowd harnesses the power of aggregated opinion of a diverse group rather than a few individuals . Though initially proposed for mainly aggregating human judgements , this idea has been successfully implemented in the context of machine learning . In particular , ensemble learning was pro... | This paper proposes two machine learning adaptations of the Bayesian truth serum approach to aggregating predictions from human experts. The first method proposed involves training two regression models for each classifier in the ensemble that predicts the proportion of other classifiers that assign the same label to a... | SP:ec3b4ae82ca6f34505dbb909d0a705804f8eb22c |
Machine Truth Serum | 1 INTRODUCTION . Wisdom of the crowd harnesses the power of aggregated opinion of a diverse group rather than a few individuals . Though initially proposed for mainly aggregating human judgements , this idea has been successfully implemented in the context of machine learning . In particular , ensemble learning was pro... | Inspired by work in ensembling human decisions, the authors propose an ensembling technique called "Machine Truth Serum" (based off "Bayesian Truth Serum"). Instead of using majority vote to ensemble the decisions of several classifiers, this paper follows the "surprisingly popular" algorithm; the ensembled decision is... | SP:ec3b4ae82ca6f34505dbb909d0a705804f8eb22c |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.