paper_name stringlengths 11 170 | text stringlengths 8.07k 307k | summary stringlengths 152 6.16k | paper_id stringlengths 43 43 |
|---|---|---|---|
Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift | 1 INTRODUCTION . Time-series forecasting plays a significant role in various daily problems such as health care , economics , and traffic engineering ( Che et al. , 2018 ; Bauer et al. , 2016 ; Zhang et al. , 2017 ) . Recently , time-series forecasting models have achieved comparative performance on these problems by o... | The authors propose a "reversible instance normalization" as an input pre and post processing procedure to improve the forecasting of any given base model - targeted at addressing the distribution shift that is common in time series data - e.g., time series are typically non-stationary. This works by normalizing each ... | SP:549b4b2a88e8b13656ee6bd9425fe1d2be77b334 |
Optimal ANN-SNN Conversion for High-accuracy and Ultra-low-latency Spiking Neural Networks | Spiking Neural Networks ( SNNs ) have gained great attraction due to their distinctive properties of low power consumption and fast inference on neuromorphic hardware . As the most effective method to get deep SNNs , ANN-SNN conversion has achieved comparable performance as ANNs on large-scale datasets . Despite this ,... | This paper proposes a Quantization neural network to spiking neural network (QNN2SNN) conversion method. The authors first analyze the conversion error between ANN and SNN. Then they construct the ann with quantized activation so that the error can be eliminated. Both theoretical and empirical results are presented in ... | SP:4f9c3ed91f44326e3bddf14223779d7f5fa07954 |
Optimal ANN-SNN Conversion for High-accuracy and Ultra-low-latency Spiking Neural Networks | Spiking Neural Networks ( SNNs ) have gained great attraction due to their distinctive properties of low power consumption and fast inference on neuromorphic hardware . As the most effective method to get deep SNNs , ANN-SNN conversion has achieved comparable performance as ANNs on large-scale datasets . Despite this ,... | This paper proposes a quantization clip-bottom-shift activation function to replace the ReLU activation function in ANNs, so as to better approximate the activation function of SNNs. The authors also prove that the expected error of ANN-SNN conversion can be reduced to 0 by using this method. The reported results on CI... | SP:4f9c3ed91f44326e3bddf14223779d7f5fa07954 |
Context-invariant, multi-variate time series representations | 1 INTRODUCTION . Many modern applications in the physical and virtual world are equipped with sensors that measure the state of the application , its sub-components , and the environment . Examples can be found in the Internet-of-Things ( IoT ) or in the DevOps/AIOPs space like monitoring wind turbines or cloudbased ap... | The paper presents a model for learning representations of time series. It addresses, in particular, the context of multivariate time series having exogenous and endogenous dimensions, and where the goal is to learn good representations independently from the context (exogenous dimensions). This work is an extension of... | SP:44ab69acefac1dba66c4849ea3eea5e7d3285710 |
Context-invariant, multi-variate time series representations | 1 INTRODUCTION . Many modern applications in the physical and virtual world are equipped with sensors that measure the state of the application , its sub-components , and the environment . Examples can be found in the Internet-of-Things ( IoT ) or in the DevOps/AIOPs space like monitoring wind turbines or cloudbased ap... | This article deals with representation of multivariate signals in the case where we can distinguish endogenous & exogenous channels. This article mainly relies on (Franceschi et al.,2019), whose proposal consists in representing signals using Dilated-TCNN and to apply triplet loss on the sequences as in word2vec. Then ... | SP:44ab69acefac1dba66c4849ea3eea5e7d3285710 |
Spike-inspired rank coding for fast and accurate recurrent neural networks | 1 INTRODUCTION . Neuromorphic computing is the study and use of computational mechanisms of biological neural networks in mathematical models , software simulations , and hardware emulations , both as a tool for neuroscience , and as a possible path towards improved machine intelligence ( Indiveri , 2021 ) . In fact , ... | The authors introduce a new way to train RNNs using rank order coding (ROC). With ROC the label is given by the first readout unit to reach a threshold. As soon as this happens, the processing is stopped, and BPTT is used from that particular time step, using the predictions at that particular time step and the ground ... | SP:7fa47de279e72f0782efd67722919321badcb022 |
Spike-inspired rank coding for fast and accurate recurrent neural networks | 1 INTRODUCTION . Neuromorphic computing is the study and use of computational mechanisms of biological neural networks in mathematical models , software simulations , and hardware emulations , both as a tool for neuroscience , and as a possible path towards improved machine intelligence ( Indiveri , 2021 ) . In fact , ... | The authors propose a method for fast and efficient classification of sequential data. The guiding principle is that for some data modalities it is not necessary to see the whole sequence in order to make a fairly certain classification. Their model reduces inference time by learning a rank code that is inspired by spi... | SP:7fa47de279e72f0782efd67722919321badcb022 |
Physics Informed Convex Artificial Neural Networks (PICANNs) for Optimal Transport based Density Estimation | 1 INTRODUCTION . Optimal Mass Transport ( OMT ) is a well-studied problem with a variety of applications in a diverse set of fields , ranging from physics to computer vision and in particular statistics and data science . In this paper we propose a new framework for the estimation of the solution to the L2-optimal tran... | This paper introduces a new method for density estimation that is based on the idea of optimal mass transport (OMP), i.e., finding some function that transports a probability density into another probability density while minimizing the transportation cost. To do so, the authors leverage Brenier's theorem which guarant... | SP:3a7dfa4251b64bfa4e0fef36a4dcecce97e3031c |
Physics Informed Convex Artificial Neural Networks (PICANNs) for Optimal Transport based Density Estimation | 1 INTRODUCTION . Optimal Mass Transport ( OMT ) is a well-studied problem with a variety of applications in a diverse set of fields , ranging from physics to computer vision and in particular statistics and data science . In this paper we propose a new framework for the estimation of the solution to the L2-optimal tran... | The paper presented an input convex neural networks (ICNN)-based methods to approximate the optimal transport maps. The learning objective is based on the Monge-Ampere equation, i.e. zero loss corresponds to the solution to the Monge-Ampere equation. Physics Informed Neural Networks (PINNs) are adopted to minimise the ... | SP:3a7dfa4251b64bfa4e0fef36a4dcecce97e3031c |
Churn Reduction via Distillation | In real-world systems , models are frequently updated as more data becomes available , and in addition to achieving high accuracy , the goal is to also maintain a low difference in predictions compared to the base model ( i.e . predictive “ churn ” ) . If model retraining results in vastly different behavior , then it ... | In real-life applications of predictive models, predictions often form a step of the process. Changes in the predictive model often need to be validated end to end using methods such as A/B tests before they can be used in production. Thus for certain class of problems, there is a need to control the churn or the diffe... | SP:08c1cbc07686b1f7b7fa01246b72929858c3846a |
Churn Reduction via Distillation | In real-world systems , models are frequently updated as more data becomes available , and in addition to achieving high accuracy , the goal is to also maintain a low difference in predictions compared to the base model ( i.e . predictive “ churn ” ) . If model retraining results in vastly different behavior , then it ... | This paper explores the relationship between the low-churn problem and distillation. The authors show that there is an equivalence between those two methods and that distillation performs particularly well on low-churn dataset tasks. The authors propose a novel churn reduction algorithm based on distillation which invo... | SP:08c1cbc07686b1f7b7fa01246b72929858c3846a |
Brain insights improve RNNs' accuracy and robustness for hierarchical control of continually learned autonomous motor motifs | 1 INTRODUCTION AND RELATION TO OTHER WORKS . Animals have the remarkable ability to efficiently learn and compose elaborate continuous behaviors , which often relies on the flexible chaining of ‘ motifs ’ - reproducible bouts of behavior - in response to hierarchical commands ( Zimnik & Churchland , 2021 ; Geddes et al... | The authors seek to use a small RNN to generate extended, time-varying outputs given a cue indicating which output to generate. They focus on the problem of generating a sequence of outputs, cued in series, without the end state of one interfering with the subsequently cued trajectory. They show that small vanilla RNNs... | SP:c55a51a4154e8f02f8b43cfed0a4216ba197c3a0 |
Brain insights improve RNNs' accuracy and robustness for hierarchical control of continually learned autonomous motor motifs | 1 INTRODUCTION AND RELATION TO OTHER WORKS . Animals have the remarkable ability to efficiently learn and compose elaborate continuous behaviors , which often relies on the flexible chaining of ‘ motifs ’ - reproducible bouts of behavior - in response to hierarchical commands ( Zimnik & Churchland , 2021 ; Geddes et al... | This paper considers the problem of RNNs learning arbitrary sequences of “motifs” (continuous time functions, discretized) after having learned the individual motifs separately. Inspired by some previous work, it proposes an architectural approach to deal with interference by introducing a motif-dependent low-rank pert... | SP:c55a51a4154e8f02f8b43cfed0a4216ba197c3a0 |
Multi-batch Reinforcement Learning via Sample Transfer and Imitation Learning | 1 INTRODUCTION . Reinforcement learning aims to learn an optimal control policy through interactions with the environment ( Sutton & Barto , 2018 ) . Deep reinforcement learning ( Deep RL or DRL ) combines neural networks with reinforcement learning and further enables RL agents to deal with more complex environments .... | This paper studies the problem of multi-task offline RL and propose 2 methods, BAIL+ and MBAIL, each is an algorithmic variant based on the BAIL algorithm. The overall goal is to improve performance on multiple tasks. This setting concerns with a number of tasks coming from a MDP distribution where transition function ... | SP:c06abfb514f76fecbab62bd5730df2012b3e896a |
Multi-batch Reinforcement Learning via Sample Transfer and Imitation Learning | 1 INTRODUCTION . Reinforcement learning aims to learn an optimal control policy through interactions with the environment ( Sutton & Barto , 2018 ) . Deep reinforcement learning ( Deep RL or DRL ) combines neural networks with reinforcement learning and further enables RL agents to deal with more complex environments .... | (Note that the rest of the review will use Offline RL and Batch RL interchangeably) Building on top of BAIL (Chen et al. 2019), this paper provides two algorithms for the multi-task offline RL setting. The BAIL+ algorithm assumes that we know the identity of the target task and that all tasks share the same transitio... | SP:c06abfb514f76fecbab62bd5730df2012b3e896a |
SUMNAS: Supernet with Unbiased Meta-Features for Neural Architecture Search | 1 INTRODUCTION . Recent Neural Architecture Search ( NAS ) algorithms often train an over-parameterized network called a supernet to obtain supreme sub-models by sharing parameters rapidly . In this process , the shared parameters are generally trained to reach a state where better architecture discovery is possible th... | The paper proposed an improved training strategy for oneshot NAS supernetworks. The key idea is to view the training of each subnetwork as a "task", and then to apply an MAML/Reptile-style meta-learning scheme to ensure efficient cross-task adaptivity. Experiments on NAS-Bench-201 and ImageNet show improved calibration... | SP:d45b7399209937f6c1af318573d6106beff29dcf |
SUMNAS: Supernet with Unbiased Meta-Features for Neural Architecture Search | 1 INTRODUCTION . Recent Neural Architecture Search ( NAS ) algorithms often train an over-parameterized network called a supernet to obtain supreme sub-models by sharing parameters rapidly . In this process , the shared parameters are generally trained to reach a state where better architecture discovery is possible th... | This work targets better ranking performance after supernet training in NAS. The authors leverage meta-learning to make the shared supernet weights adaptive to randomly sampled subnetworks. Experiments are conducted on NAS-Bench-201 and MobileNet space. | SP:d45b7399209937f6c1af318573d6106beff29dcf |
Proof Artifact Co-Training for Theorem Proving with Language Models | 1 INTRODUCTION . Deep learning-driven automated theorem proving in large libraries of formalized mathematics ( henceforth “ neural theorem proving ” ) has been the focus of increased attention in recent years . Labeled data for imitation learning of theorem proving is scarce—formalization is notoriously labor-intensive... | To deal with data scarcity in context of learning large Transformer language models for theorem proving, this work proposes a methodology (called PACT) for extracting auxiliary task data for joint training alongside the tactic prediction objective. The methodology has been applied to Lean proof assistant. PACT signific... | SP:4a83aeec9e7a387683678c056e71a70d238511ed |
Proof Artifact Co-Training for Theorem Proving with Language Models | 1 INTRODUCTION . Deep learning-driven automated theorem proving in large libraries of formalized mathematics ( henceforth “ neural theorem proving ” ) has been the focus of increased attention in recent years . Labeled data for imitation learning of theorem proving is scarce—formalization is notoriously labor-intensive... | This paper addresses the general setup of using transformer models for interactive theorem proving (ITP) tasks. The ITP engine considered here is Lean. The contribution of the paper is a data augmentation method. This is achieved by mining low level artifacts from a given dataset of Lean proofs. This includes data extr... | SP:4a83aeec9e7a387683678c056e71a70d238511ed |
SketchODE: Learning neural sketch representation in continuous time | 1 INTRODUCTION Drawing-based communications such as sketches , writing , and diagrams come naturally to humans and have been used in some form since ancient times . Modeling such data is becoming an increasingly important and topical challenge area for machine learning systems aiming to interpret and simulate human cre... | The paper describes a way of representing online handwriting data, typically seen as a discrete sequence, in a continuous space, using neural ODEs. The learned representation allows sampling from and interpolation in the latent space. The results in the paper are compared (mostly qualitatively) to autoencoder-learned r... | SP:ccef448cb87c0cac9153d0e5b3b266a6b794c277 |
SketchODE: Learning neural sketch representation in continuous time | 1 INTRODUCTION Drawing-based communications such as sketches , writing , and diagrams come naturally to humans and have been used in some form since ancient times . Modeling such data is becoming an increasingly important and topical challenge area for machine learning systems aiming to interpret and simulate human cre... | This paper proposes a new model, called SketchODE, for learning representations of sketches using neural ODEs. Specifically, the authors parameterize hand drawn strokes as solutions of ODEs and build an auto encoder like framework for learning the vector fields of these ODEs from data. The decoder is modelled with a ... | SP:ccef448cb87c0cac9153d0e5b3b266a6b794c277 |
Autonomous Learning of Object-Centric Abstractions for High-Level Planning | 1 INTRODUCTION . Model-based methods are a promising approach to improving sample efficiency in reinforcement learning ( RL ) . However , they require the agent to either learn a highly detailed model—which is infeasible for sufficiently complex problems ( Ho et al. , 2019 ) —or to build a compact , high-level model th... | This paper introduces a method for learning symbolic, object-centric abstractions from object-factored environment observations for long-term planning tasks. It extends the symbolic representation learning framework by Konidaris et al. (2018) by factoring the state into objects, learning object type abstractions, and “... | SP:8177c43096a6a35c7f64ccf714ae038b1ea4db7a |
Autonomous Learning of Object-Centric Abstractions for High-Level Planning | 1 INTRODUCTION . Model-based methods are a promising approach to improving sample efficiency in reinforcement learning ( RL ) . However , they require the agent to either learn a highly detailed model—which is infeasible for sufficiently complex problems ( Ho et al. , 2019 ) —or to build a compact , high-level model th... | This paper proposes a method for learning an object-centric symbolic representation of an environment that allows for planning. It extends the framework introduced in Konidaris et al. (2018), which learns a symbolic representation of an environment that can be expressed using a PDDL for planning. Importantly, the metho... | SP:8177c43096a6a35c7f64ccf714ae038b1ea4db7a |
ED2: An Environment Dynamics Decomposition Framework for World Model Construction | 1 INTRODUCTION . Reinforcement Learning ( RL ) is a general learning framework for solving sequential decision-making problems and has made significant progress in many fields ( Mnih et al. , 2015 ; Silver et al. , 2016 ; Vinyals et al. , 2019 ; Schrittwieser et al. , 2019 ) . In general , RL methods can be divided int... | This paper presents a model-based RL algorithm that provides decomposed dynamics models when the state and action spaces are defined as a multi-dimensional Cartesian space. We can divide the method into two parts: * The one for partitioning the action coordinates, and * The NN architecture, one for each partition and ... | SP:e3e6dc2285271426a37bc08e8989e03fe1891ea9 |
ED2: An Environment Dynamics Decomposition Framework for World Model Construction | 1 INTRODUCTION . Reinforcement Learning ( RL ) is a general learning framework for solving sequential decision-making problems and has made significant progress in many fields ( Mnih et al. , 2015 ; Silver et al. , 2016 ; Vinyals et al. , 2019 ; Schrittwieser et al. , 2019 ) . In general , RL methods can be divided int... | The authors propose an environment dynamics decomposition (ED2) framework to decompose the environmental dynamics into multiple sub-dynamics. Empirical results show that ED2 improves the performance of several model-based reinforcement learning (MBRL) algorithms. The paper is well-written and easy to read, and the moti... | SP:e3e6dc2285271426a37bc08e8989e03fe1891ea9 |
On Learning with Fairness Trade-Offs | 1 INTRODUCTION . Reducing bias in an algorithm requires a number of steps ; first , specify the fairness definitions ( and thus fairness metrics ) that apply , second , encode them with a penalty so as to measure the discrepancy between outcomes and the perfect fairness scenario , third , choose the trade-off between t... | This paper establishes the guarantee for the generalization of fairness-aware learning in binary classification under PAC-learning and a more practical asymptotic framework. Through the derived theorem, authors conclude that low sample size and class balance lead to the poor generalization of fairness-aware learning, a... | SP:3e69b216b8568e1ae0f25c73d45219ac3bef8a43 |
On Learning with Fairness Trade-Offs | 1 INTRODUCTION . Reducing bias in an algorithm requires a number of steps ; first , specify the fairness definitions ( and thus fairness metrics ) that apply , second , encode them with a penalty so as to measure the discrepancy between outcomes and the perfect fairness scenario , third , choose the trade-off between t... | The paper provides an analysis of fairness via regularizing a loss function with a suitably selected fairness measure / metric. In particular, the paper analysis this under a PAC learning setting and a sample limit (asymptotic) setting. A majority of the paper focuses on the PAC setting, where Rademacher complexity bou... | SP:3e69b216b8568e1ae0f25c73d45219ac3bef8a43 |
On the Pitfalls of Analyzing Individual Neurons in Language Models | 1 INTRODUCTION . Many studies attempt to interpret language models by predicting different linguistic properties from word representations , an approach called probing classifiers ( Adi et al. , 2017 ; Conneau et al. , 2018 , inter alia ) . A growing body of work focuses on individual neurons within the representation ... | The paper revisits two methods, Linear and Gaussian, as described in the literature, to probe language models based on individual neurons. The paper suggests that these methods contain two limitations - 1. by ranking the neurons on a linguistic task, the methods conflate between ranking quality and the probe's classifi... | SP:71274cc432cc95a2d80418365896096b875be5c1 |
On the Pitfalls of Analyzing Individual Neurons in Language Models | 1 INTRODUCTION . Many studies attempt to interpret language models by predicting different linguistic properties from word representations , an approach called probing classifiers ( Adi et al. , 2017 ; Conneau et al. , 2018 , inter alia ) . A growing body of work focuses on individual neurons within the representation ... | This paper responds to several recent works focused on identifying important individual neurons for particular classifying tasks. They consider 2 existing methods which rely on an external probe to rank the neurons in a network. They also introduce a method that does not rely on a probe, instead ranking neurons accordi... | SP:71274cc432cc95a2d80418365896096b875be5c1 |
Rethinking Temperature in Graph Contrastive Learning | 1 INTRODUCTION . Self-supervised learning provides a good learning paradigm without high-cost label information for computer vision ( Chen et al . ( 2020 ) ; Chen & He ( 2021 ) ; Grill et al . ( 2020 ) ) , natural language processing ( Wu et al . ( 2019 ) ; Gao et al . ( 2021 ) ) , and speech recognition ( Ravanelli et... | This paper investigates a crucial problem, i.e., how to generate good node representations under the InfoNCE contrastive loss in graphs. By studying the feature of loss function’s gradients, it finds that the dynamic temperature change is beneficial to learn uniform node representations. The proposed method GLATE is si... | SP:b3951670f71d8af5e59144d68c7d50405a001227 |
Rethinking Temperature in Graph Contrastive Learning | 1 INTRODUCTION . Self-supervised learning provides a good learning paradigm without high-cost label information for computer vision ( Chen et al . ( 2020 ) ; Chen & He ( 2021 ) ; Grill et al . ( 2020 ) ) , natural language processing ( Wu et al . ( 2019 ) ; Gao et al . ( 2021 ) ) , and speech recognition ( Ravanelli et... | The authors explore the role of the temperature in the loss function for graph contrastive learning. They argue that global uniformity and local separation are both necessary to the learning quality and this can be controlled by the temperature. Thus, they develop a simple but effective algorithm GLATE to dynamically a... | SP:b3951670f71d8af5e59144d68c7d50405a001227 |
Resolving label uncertainty with implicit generative models | 1 INTRODUCTION . We consider the problem of joint inference of latent label variables ℓ8 in a collection of data samples indexed by 8 consisting of observations ( features ) G8 and corresponding prior beliefs about their latent label variables ? 8 ( ℓ ) . Two illustrative examples are shown in Fig . 1 . In the first ex... | The paper presents a weakly supervised learning strategy, which exploits instance labels in a form of label prior distributions for training classifiers. The main idea of this work is to build an implicit generative model from a probabilistic label prediction network, which is then trained with an ELBO loss. The paper ... | SP:aa1054c1782ad4562ed402bb56f70f9287d197ce |
Resolving label uncertainty with implicit generative models | 1 INTRODUCTION . We consider the problem of joint inference of latent label variables ℓ8 in a collection of data samples indexed by 8 consisting of observations ( features ) G8 and corresponding prior beliefs about their latent label variables ? 8 ( ℓ ) . Two illustrative examples are shown in Fig . 1 . In the first ex... | In this paper, the authors present implicit generative models in a free energy criterion with a combination of both the training of neural networks for label prediction and modeling of a label prior. They also discuss multiple sources of label priors to handle label uncertainty for coarse and imprecise data input. The ... | SP:aa1054c1782ad4562ed402bb56f70f9287d197ce |
Neural Capacitance: A New Perspective of Neural Network Selection via Edge Dynamics | Efficient model selection for identifying a suitable pre-trained neural network to a downstream task is a fundamental yet challenging task in deep learning . Current practice requires expensive computational costs in model training for performance prediction . In this paper , we propose a novel framework for neural net... | Accurately predicting the performance at early stage is important for efficient model selection without incurring too much computation. The paper proposes a neural capacitance metric as a predictive measure to capture the performance of a model on the downstream task using only a handful of early training results. The ... | SP:d18240d9eca55834f96c0223924e425a38685368 |
Neural Capacitance: A New Perspective of Neural Network Selection via Edge Dynamics | Efficient model selection for identifying a suitable pre-trained neural network to a downstream task is a fundamental yet challenging task in deep learning . Current practice requires expensive computational costs in model training for performance prediction . In this paper , we propose a novel framework for neural net... | This paper proposes a framework to select the neural networks for downstream tasks. To identify the better generalization model, the authors propose a new metric (Neural Capacitance, NCP) to predict precise learning curves. And the authors provide the theoretical explanations for NCP. Then the authors have verified th... | SP:d18240d9eca55834f96c0223924e425a38685368 |
Know Your Action Set: Learning Action Relations for Reinforcement Learning | Intelligent agents can solve tasks in a variety of ways depending on the action set at their disposal . For instance , while using a toolkit for repair , the choice of tool ( the action ) closely depends on what other tools are available . Yet , such dependence on other available actions is ignored in conventional rein... | This paper tackles an RL problem setting in which the actions available to an agent vary from episode to episode, and the optimal action in some states depends on the other actions that are available. The authors’ approach to this setting is to use a graph neural network to process all available actions, both to summar... | SP:80376e14141e0c667c4e1c1568ca9d545a1c5fbd |
Know Your Action Set: Learning Action Relations for Reinforcement Learning | Intelligent agents can solve tasks in a variety of ways depending on the action set at their disposal . For instance , while using a toolkit for repair , the choice of tool ( the action ) closely depends on what other tools are available . Yet , such dependence on other available actions is ignored in conventional rein... | This paper introduces a novel policy architecture (AGILE) for RL agents that learn action interdependence from a varying action space. A graph attention network is used to calculate the action utility and to summarize the action set input. Authors argue that this architecture allows the RL agents’ to learn action relat... | SP:80376e14141e0c667c4e1c1568ca9d545a1c5fbd |
MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer | 1 INTRODUCTION . Self-attention-based models , especially vision transformers ( ViTs ; Figure 1a ; Dosovitskiy et al. , 2021 ) , are an alternative to convolutional neural networks ( CNNs ) to learn visual representations . Briefly , ViT divides an image into a sequence of non-overlapping patches and then learns interp... | This work proposes a hybrid backbone which combines existing MV2 block and new MobileViT block for efficient classification, detection and segmentation. The core idea with MobileViT block is to use both convolution and self-attention block to aggregate local and global feature, respectively. Experiments on IN-1K, COCO ... | SP:9b2ac905ecf71d7207dab06d2367d7259ac4968b |
MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer | 1 INTRODUCTION . Self-attention-based models , especially vision transformers ( ViTs ; Figure 1a ; Dosovitskiy et al. , 2021 ) , are an alternative to convolutional neural networks ( CNNs ) to learn visual representations . Briefly , ViT divides an image into a sequence of non-overlapping patches and then learns interp... | This paper introduces a lightweight and general-purpose vision transformer, termed MobileViT, for mobile devices. It attempts to build a model which can combine the strengths of CNNs and ViTs to build a light-weight and low latency network for mobile vision tasks. MobileViT achieves top-1 accuracy of 78.4% on ImageNet-... | SP:9b2ac905ecf71d7207dab06d2367d7259ac4968b |
Learned Index with Dynamic $\epsilon$ | 1 INTRODUCTION . Data indexing ( Graefe & Kuno , 2011 ; Wang et al. , 2018 ; Luo & Carey , 2020 ; Zhou et al. , 2020 ) , which stores keys and corresponding payloads with designed structures , supports efficient query operations over data and benefits various data retrieval applications . Recently , Machine Learning ( ... | This paper considers learning based methods for constructing index structures, a fundamental data structure. Recent work has given learning based index structures which guarantee a maximum error of $\epsilon$ in the predicted index. At a high level, this is done by breaking up the data set into several segments, and ... | SP:123d38f124b95d956b18e6906a2e26171fb94224 |
Learned Index with Dynamic $\epsilon$ | 1 INTRODUCTION . Data indexing ( Graefe & Kuno , 2011 ; Wang et al. , 2018 ; Luo & Carey , 2020 ; Zhou et al. , 2020 ) , which stores keys and corresponding payloads with designed structures , supports efficient query operations over data and benefits various data retrieval applications . Recently , Machine Learning ( ... | This paper mainly studies the index structure problem. Existing learned index methods use a fixed value for all the learned segments. In this paper, the author contributes a deeper understanding of how the impacts the index performance, and enlightens the exploration of fine-grained trade-off adjustments by considerin... | SP:123d38f124b95d956b18e6906a2e26171fb94224 |
Generalizing MLPs With Dropouts, Batch Normalization, and Skip Connections | 1 INTRODUCTION . 1.1 MLP AND ITS TRAINING OBJECTIVE . A multilayer perceptron ( MLP ) is one of the simplest and the oldest artificial neural network architectures . Its origin dates back to 1958 ( Rosenblatt , 1958 ) . Its basic building blocks are linear regression and nonlinear activation functions . One can stack t... | The paper proposes a network architecture for MLP networks. The authors build upon work by Chen et al. (2019) to achieve whitening by batchnorm followed by dropout (IC Layer). Further they use skip-connections (He et al., 2016) in their architecture. The authors tested their approach in an ablation study for gender and... | SP:a5347c01c781befd65ffc3899fc313627891fd65 |
Generalizing MLPs With Dropouts, Batch Normalization, and Skip Connections | 1 INTRODUCTION . 1.1 MLP AND ITS TRAINING OBJECTIVE . A multilayer perceptron ( MLP ) is one of the simplest and the oldest artificial neural network architectures . Its origin dates back to 1958 ( Rosenblatt , 1958 ) . Its basic building blocks are linear regression and nonlinear activation functions . One can stack t... | This paper proposes a general architecture for MLPs. The main ingredients are skip connections and whitening. These methods have been proven to be effective in CNNs. The authors show that MLPs do benefit from them as well. The proposed skip connection (SO,SI - skip-out, skip-in) is added before the last ReLu (R) of eac... | SP:a5347c01c781befd65ffc3899fc313627891fd65 |
Interpretable Semantic Role Relation Table for Supporting Facts Recognition of Reading Comprehension | 1 INTRODUCTION . There has been an increasing interest in the explainability of Machine Reading Comprehension ( MRC ) in recent years . For enhancing the explainability in MRC , some researchers ( Qiu et al. , 2019 ; Tu et al. , 2020 ; Fang et al. , 2020 ) utilize Graph Networks . The relational inductive bias encoded ... | The paper presents a model for the supporting sentence prediction task in machine reading comprehension. To enhance interpretability of the model, the paper proposes to use the semantic relations among a question, target sentence, and sentences in an article as features of the model. The semantic relations are obtained... | SP:46f8bfb644d14e8ff9a9053057e27f96c4c32a97 |
Interpretable Semantic Role Relation Table for Supporting Facts Recognition of Reading Comprehension | 1 INTRODUCTION . There has been an increasing interest in the explainability of Machine Reading Comprehension ( MRC ) in recent years . For enhancing the explainability in MRC , some researchers ( Qiu et al. , 2019 ; Tu et al. , 2020 ; Fang et al. , 2020 ) utilize Graph Networks . The relational inductive bias encoded ... | The authors propose a method for adding interpretable semantic features for a Machine Reading Comprehension model based on semantic relations across words and sentences. The goal is to use semantic role labeling annotation to produce features used on a neural model for a language downstream task. The main contributi... | SP:46f8bfb644d14e8ff9a9053057e27f96c4c32a97 |
ViDT: An Efficient and Effective Fully Transformer-based Object Detector | Transformers are transforming the landscape of computer vision , especially for recognition tasks . Detection transformers are the first fully end-to-end learning systems for object detection , while vision transformers are the first fully transformer-based architecture for image classification . In this paper , we int... | This paper proposes ViDT, a high-performance Detection Transformer with an impressive accuracy-speed trade-off. A lot of experiments as well as ablation studies are conducted to prove the effectiveness of the proposed detector. The design principle of ViDT can also generalize and inspire future detector design. Moreove... | SP:c5fa77eef56d8878f58c99c5c50902a3f739e2fb |
ViDT: An Efficient and Effective Fully Transformer-based Object Detector | Transformers are transforming the landscape of computer vision , especially for recognition tasks . Detection transformers are the first fully end-to-end learning systems for object detection , while vision transformers are the first fully transformer-based architecture for image classification . In this paper , we int... | The paper builds upon the recent advances in transformer based image classification methods (ViT variants) and detection methods (DETR variants). It argues that naively replacing the conv feature backbone in DETR with a ViT based one, is problematic due to (i) the quadratic complexity of the self attention module in Vi... | SP:c5fa77eef56d8878f58c99c5c50902a3f739e2fb |
Graphon based Clustering and Testing of Networks: Algorithms and Theory | 1 INTRODUCTION . Machine learning on graphs has evolved considerably over the past two decades . The traditional view towards network analysis is limited to modelling interactions among entities of interest , for instance social networks or world wide web , and learning algorithms based on graph theory have been common... | Motivated by the sort-and-smooth graphon estimator of Chan and Airoldi, 2014, this paper proposes two new clustering algorithms (graph distance based spectral clustering and similarity based semidefinite programming) for multiple graphs observed without vertex correspondence. The idea is to use the graphon approximatio... | SP:789952dc98861a18ec301641a66f64eabc8a6e13 |
Graphon based Clustering and Testing of Networks: Algorithms and Theory | 1 INTRODUCTION . Machine learning on graphs has evolved considerably over the past two decades . The traditional view towards network analysis is limited to modelling interactions among entities of interest , for instance social networks or world wide web , and learning algorithms based on graph theory have been common... | The paper proposes a graph distance based on graphons that can operate on the small sample size regime; this is possible by exploiting a large number of nodes. The proposed paper is meant to address two shortcomings that the authors have found in the literature: - a lack of study of current graph processing methods be... | SP:789952dc98861a18ec301641a66f64eabc8a6e13 |
Large-Scale Representation Learning on Graphs via Bootstrapping | 1 INTRODUCTION . Graphs provide a powerful abstraction for complex datasets that arise in a variety of applications such as social networks , transportation networks , and biological sciences ( Hamilton et al. , 2017 ; DerrowPinion et al. , 2021 ; Zitnik & Leskovec , 2017 ; Chanussot et al. , 2021 ) . Despite recent ad... | The paper proposes a self-supervised graph representation learning algorithm named BGRL. BGRL employs similar architecture & training mechanism with BYOL and necessary adjustments (e.g., the augmentation methods, dropping the prejector module) to adapt to the characteristics of graph datasets. Given the nature of BYOL-... | SP:cd488af7b4ffc290fcd06ca9577156fa67ec3286 |
Large-Scale Representation Learning on Graphs via Bootstrapping | 1 INTRODUCTION . Graphs provide a powerful abstraction for complex datasets that arise in a variety of applications such as social networks , transportation networks , and biological sciences ( Hamilton et al. , 2017 ; DerrowPinion et al. , 2021 ; Zitnik & Leskovec , 2017 ; Chanussot et al. , 2021 ) . Despite recent ad... | This paper proposes BRGL, a method for graph representation learning based on ‘bootstrapping’ (in the same sense BYOL is). Different from the prior art, which mainly relies on contrastive learning, BRGL naturally scales linearly with graph size. An extensive suite of experiments shows that BRGL scales better than prev... | SP:cd488af7b4ffc290fcd06ca9577156fa67ec3286 |
Learning Versatile Neural Architectures by Propagating Network Codes | 1 INTRODUCTION . Designing a single neural network architecture that adapts to multiple different tasks is challenging . This is because different tasks , such as image segmentation in Cityscapes ( Cordts et al. , 2016 ) and video recognition in HMDB51 ( Kuehne et al. , 2011 ) , have different data distributions and re... | This paper tackles neural architecture search (NAS), addressing the problem of finding architectures suitable for a multitude of vision-related tasks, ranging from object detection to semantic segmentation. To the best of my knowledge, this is the first attempt in NAS for computer vision models that explicitly design t... | SP:36f276695fe54a7290e68a69eac687c0557bf454 |
Learning Versatile Neural Architectures by Propagating Network Codes | 1 INTRODUCTION . Designing a single neural network architecture that adapts to multiple different tasks is challenging . This is because different tasks , such as image segmentation in Cityscapes ( Cordts et al. , 2016 ) and video recognition in HMDB51 ( Kuehne et al. , 2011 ) , have different data distributions and re... | This paper organized a multi-task NAS benchmark including 4 widely used datasets and more than 20,000 models. It also proposed a searching architecture named Network Coding Propagation (NCP) to effectively find the optical model for specific tasks. Experimental results indicate that the proposed model can be applied to... | SP:36f276695fe54a7290e68a69eac687c0557bf454 |
Joint Shapley values: a measure of joint feature importance | The Shapley value is one of the most widely used measures of feature importance partly as it measures a feature ’ s average effect on a model ’ s prediction . We introduce joint Shapley values , which directly extend Shapley ’ s axioms and intuitions : joint Shapley values measure a set of features ’ average contributi... | This work introduces "joint Shapley values", which directly extend Shapley’s axioms and intuitions: joint Shapley values measure a set of features’ average effect on a model’s prediction. This work naturally extends Shapley's axioms from a single feature, to sets of features. In a nutshell: joint Shapley values measure... | SP:86972f8c4420d86b88dfcf2aaf13b88d53b98de5 |
Joint Shapley values: a measure of joint feature importance | The Shapley value is one of the most widely used measures of feature importance partly as it measures a feature ’ s average effect on a model ’ s prediction . We introduce joint Shapley values , which directly extend Shapley ’ s axioms and intuitions : joint Shapley values measure a set of features ’ average contributi... | The paper proposes an extension of the Shapley values, namely, the joint Shapley values which measure a set of features' average effect of a models prediction. The uniqueness of the joint Shapley values is proved. Moreover, training details and tuning parameters are provided in the accompanying code. | SP:86972f8c4420d86b88dfcf2aaf13b88d53b98de5 |
DMSANET: DUAL MULTI SCALE ATTENTION NETWORK | 1 INTRODUCTION . The local receptive field of the human eye has led to the construction of convolutional neural networks which has powered much of the recent advances in computer vision . Multi scale architecture used in the famous InceptionNet ( Szegedy et al. , 2016 ) aggregates multi-scale information from different... | In this paper, the authors proposed a new attention module, named Dual Multi-Scale Attention. Three commonly used methods are combined: multi-scale, channel-attention, and position attention. By introducing the new module (which is a simple combination without any novelty) to ResNet, it SIGNIFICANTLY improves the perfo... | SP:ee608c38bc5cba62161a824d237c680050ba3484 |
DMSANET: DUAL MULTI SCALE ATTENTION NETWORK | 1 INTRODUCTION . The local receptive field of the human eye has led to the construction of convolutional neural networks which has powered much of the recent advances in computer vision . Multi scale architecture used in the famous InceptionNet ( Szegedy et al. , 2016 ) aggregates multi-scale information from different... | In this paper, the authors propose a new attention module that shows better performance and lesser computation than most existing attention modules. Based on this module, the so-called Dual Multi Scale Attention Network is proposed. Several experiments are conducted to verify the performance on image classification, ob... | SP:ee608c38bc5cba62161a824d237c680050ba3484 |
Rethinking Goal-Conditioned Supervised Learning and Its Connection to Offline RL | 1 INTRODUCTION . Reinforcement learning ( RL ) enables automatic skill learning and has achieved great success in various tasks ( Mnih et al. , 2015 ; Lillicrap et al. , 2016 ; Haarnoja et al. , 2018 ; Vinyals et al. , 2019 ) . Recently , goal-conditioned RL is gaining attention from the community , as it encourages ag... | The proposed method proposes a method for goal-conditioned RL that can be interpreted as a weighted version of prior work.These weights resemble a combination of discounting and advantage weighting. The paper provides theory arguing that these weights cause the proposed method to optimize a tighter lower bound than pri... | SP:b35f74a2fd21a865da1621b2c59ead20e912e3ad |
Rethinking Goal-Conditioned Supervised Learning and Its Connection to Offline RL | 1 INTRODUCTION . Reinforcement learning ( RL ) enables automatic skill learning and has achieved great success in various tasks ( Mnih et al. , 2015 ; Lillicrap et al. , 2016 ; Haarnoja et al. , 2018 ; Vinyals et al. , 2019 ) . Recently , goal-conditioned RL is gaining attention from the community , as it encourages ag... | This paper proposes an extension of GCSL where the goal-conditioned BC loss is weighted by a variable that correlates with the number of steps necessary to achieve the desired goal. The advantage of this approach is that sub-optimal trajectories to a particular goal will get downweighted to the benefit of more direct ... | SP:b35f74a2fd21a865da1621b2c59ead20e912e3ad |
Interpreting Reinforcement Policies through Local Behaviors | 1 INTRODUCTION . Deep reinforcement learning has seen stupendous success over the last decade with superhuman performance in games such as Go ( Silver et al. , 2016 ) , Chess ( Silver et al. , 2018 ) as well as Atari benchmarks ( Mnih et al. , 2015 ) . With increasing superior capabilities of automated ( learning ) sys... | This work proposes to explain an RL policy by clustering states into meta-states and presenting strategic state(s) for each meta-state. This clustering is performed based on policy rollouts, balancing likelihood of paths within a meta-state and number of paths from states within the meta-state. The authors present exam... | SP:dde554a1db9bb15202e77adaf6edd9868865605b |
Interpreting Reinforcement Policies through Local Behaviors | 1 INTRODUCTION . Deep reinforcement learning has seen stupendous success over the last decade with superhuman performance in games such as Go ( Silver et al. , 2016 ) , Chess ( Silver et al. , 2018 ) as well as Atari benchmarks ( Mnih et al. , 2015 ) . With increasing superior capabilities of automated ( learning ) sys... | The paper proposes an approach to interpret a black-box control policy of a reinforcement learning (RL) problem such that its interpretations can be understood by a human user. For a given expert policy (i.e., a fitted deep neural network), the proposed approach uses transition probabilities induced by the expert polic... | SP:dde554a1db9bb15202e77adaf6edd9868865605b |
Analytically Tractable Bayesian Deep Q-Learning | 1 INTRODUCTION . Reinforcement learning ( RL ) has gained increasing interest since the demonstration it was able to reach human performance on video game benchmarks using deep Q-learning ( DQN ) ( Mnih et al. , 2015 ; Van Hasselt et al. , 2016 ) . Deep RL methods typically require an explicit definition of an explorat... | This paper proposed a new Bayesian Deep RL algorithm that learns deep Q networks (DQN) without using gradient descent. The updates of DQN's parameters follows from the tractable approximate Gaussian inference (TAGI) algorithm. Then it studies the empirical performance of proposed algorithm with DQN learned by gradient ... | SP:879e4af8012f31b4bc12387f8573889b7d341bd6 |
Analytically Tractable Bayesian Deep Q-Learning | 1 INTRODUCTION . Reinforcement learning ( RL ) has gained increasing interest since the demonstration it was able to reach human performance on video game benchmarks using deep Q-learning ( DQN ) ( Mnih et al. , 2015 ; Van Hasselt et al. , 2016 ) . Deep RL methods typically require an explicit definition of an explorat... | The paper describes how to adapt the Q-learning algorithm so that it is compatible with Bayesian deep learning, as opposed to the standard (semi-)gradient-based deep learning implementation. It explains how to propagate uncertainties about network parameters through the neural network computations and ultimately to the... | SP:879e4af8012f31b4bc12387f8573889b7d341bd6 |
Revisiting the Lottery Ticket Hypothesis: A Ramanujan Graph Perspective | 1 INTRODUCTION AND RELATED WORK . Neural Network ( NN ) and its recent advancements have made a significant contribution to solve various machine learning applications . The power of an over-parameterized NN lies in its capability to learn simple patterns and memorize the noise in the data ( Neyshabur et al. , 2018 ) .... | This paper revisits the Lottery Ticket Hypotheis (LTH), where aggressively pruning a neural network in a smart way allows it to retain its convergence properties, at a lesser computational cost. The authors advance an explanation relating this LTH to expanders, graphs that have the property that information circulates... | SP:18de06d24b04d5bda2a3f7d6c3cd4f28076cfee5 |
Revisiting the Lottery Ticket Hypothesis: A Ramanujan Graph Perspective | 1 INTRODUCTION AND RELATED WORK . Neural Network ( NN ) and its recent advancements have made a significant contribution to solve various machine learning applications . The power of an over-parameterized NN lies in its capability to learn simple patterns and memorize the noise in the data ( Neyshabur et al. , 2018 ) .... | In this paper, the authors study the properties of the subnetworks in lottery ticket hypothesis (LTH) from the perspective of the spectral graph theory. They argue that the pruned network in LTH remains a Ramanujan graph. The performance of the subnetworks begins to degrade with the loss of Ramanujan graph property. To... | SP:18de06d24b04d5bda2a3f7d6c3cd4f28076cfee5 |
Associated Learning: an Alternative to End-to-End Backpropagation that Works on CNN, RNN, and Transformer | 1 INTRODUCTION . Backpropagation ( BP ) is the keystone of modern deep learning . Although BP is the standard way to learn network parameters , it is far from ideal . Some of the most discussed issues of BP are optimization difficulties ( e.g. , vanishing gradients and exploding gradients ( Hochreiter et al. , 2001 ) )... | This paper is a continuation of an original associated learning paper by Kao&Chen 2021. It attempts to propoose new learning approach associated learning as an alternative way to back-propagation. On top of the original paper, it discovers more interesting properties and extend AL to CNN, LSTM and transformers (though ... | SP:590a081dcfa6a61b6228700092654cf8647ffecd |
Associated Learning: an Alternative to End-to-End Backpropagation that Works on CNN, RNN, and Transformer | 1 INTRODUCTION . Backpropagation ( BP ) is the keystone of modern deep learning . Although BP is the standard way to learn network parameters , it is far from ideal . Some of the most discussed issues of BP are optimization difficulties ( e.g. , vanishing gradients and exploding gradients ( Hochreiter et al. , 2001 ) )... | This paper proposes associated learning (AL) for CNN, RNN, and transformer. Different from back-propagation (BP), AL decomposes BP’s global end-to-end training strategy into several small local optimization targets such that each sub-networks has an isolated gradient flow. To achieve this, the paper proposes to map inp... | SP:590a081dcfa6a61b6228700092654cf8647ffecd |
Diverse and Consistent Multi-view Networks for Semi-supervised Regression | 1 INTRODUCTION . Deep neural networks have achieved tremendous success across several domains , ranging from computer vision , natural language processing , to audio analysis ( LeCun et al. , 2015 ) . However , to train neural networks that perform well typically requires a large amount of labeled data . In many cases ... | This paper proposed a method for **semi-supervised multi-view regression** based on **data augmentation** and an **undirected graphical model**. The author derived a "**consistency**" term for unlabeled data and a "**diversity**" term for labeled data based on their log-likelihood and combined them linearly with two hy... | SP:da67860e703f8b08a84c6ba79fe3a32b33a187ad |
Diverse and Consistent Multi-view Networks for Semi-supervised Regression | 1 INTRODUCTION . Deep neural networks have achieved tremendous success across several domains , ranging from computer vision , natural language processing , to audio analysis ( LeCun et al. , 2015 ) . However , to train neural networks that perform well typically requires a large amount of labeled data . In many cases ... | This paper concerns the semi-supervised problem. Different from SOTA deep semi-supervised methods, the proposed DiCom employs a diversity measure on the labeled multi-view data, and combines diversity with consistency based on underlying probabilistic graphical assumptions. Experiments verify the effectiveness. | SP:da67860e703f8b08a84c6ba79fe3a32b33a187ad |
Learning to Abstain in the Presence of Uninformative Data | 1 INTRODUCTION . Despite the success of machine learning in computer vision ( Deng et al. , 2009 ; Krizhevsky et al. , 2009 ; He et al. , 2016a ; Huang et al. , 2017 ) and natural language processing ( Vaswani et al. , 2017 ; Devlin et al. , 2018 ) , the power of ML is yet to make significant impact in other areas . On... | The paper is concerned with selective classification in a stylised 'realisably noisy' data model, wherein the support of the input distribution is partitioned into two chunks, the "informative" $\Omega_I$ and the "uninformative" $\Omega_U,$ such that - The labels are completely noisy ($\mathrm{Bern}(1/2)$) if the inpu... | SP:4313c3bed7eaa75d90537a6d993bb6207f4adae2 |
Learning to Abstain in the Presence of Uninformative Data | 1 INTRODUCTION . Despite the success of machine learning in computer vision ( Deng et al. , 2009 ; Krizhevsky et al. , 2009 ; He et al. , 2016a ; Huang et al. , 2017 ) and natural language processing ( Vaswani et al. , 2017 ; Devlin et al. , 2018 ) , the power of ML is yet to make significant impact in other areas . On... | This paper considers supervised learning with abstention -- where the learning method can decide to make prediction in some region of the feature space, and declare the rest of the feature space unpredictable. The setting is related to but different from the classical selective prediction, and to prediction uncertainty... | SP:4313c3bed7eaa75d90537a6d993bb6207f4adae2 |
Learn Together, Stop Apart: a Novel Approach to Ensemble Pruning | 1 INTRODUCTION . There are still many areas where classical machine learning algorithms prevail over deep neural networks despite the dramatic growth of their usage in artificial intelligence research . One of such classical algorithms is Gradient Boosting ( GB ) ( Friedman ( 2001 ) ) . It allows to obtain high-quality... | This paper proposes to tune the number of models in a boosting ensemble in an instance-wise fashion. The idea is to first cluster the samples using a decision tree and then to tune the size of the ensemble independently for each cluster, instead of doing it globally for all instances. An efficient two-level cross-valid... | SP:5af5a69509d6176336791a840c17703dee176a1d |
Learn Together, Stop Apart: a Novel Approach to Ensemble Pruning | 1 INTRODUCTION . There are still many areas where classical machine learning algorithms prevail over deep neural networks despite the dramatic growth of their usage in artificial intelligence research . One of such classical algorithms is Gradient Boosting ( GB ) ( Friedman ( 2001 ) ) . It allows to obtain high-quality... | The paper proposes a novel method to set the optimal ensemble size in gradient boosting. In particular, the authors propose an adaptive strategy that sets distinct ensemble sizes for different regions of the input space. For that, they propose dividing the input space in coherent regions (whose instances are similar bo... | SP:5af5a69509d6176336791a840c17703dee176a1d |
State-Action Joint Regularized Implicit Policy for Offline Reinforcement Learning | Offline reinforcement learning enables learning from a fixed dataset , without further interactions with the environment . The lack of environmental interactions makes the policy training vulnerable to state-action pairs far from the training dataset and prone to missing rewarding actions . For training more effective ... | This paper proposed a regularized policy learning algorithm for offline reinforcement learning. The implicit policy is trained by a GAN-like framework, and the regularization loss constrains the distance between learned policy and behavior policy. Experiments and ablation study on the D4RL dataset validate the proposed... | SP:558e16a6d21c668b92792c51293357eaecb4dc9d |
State-Action Joint Regularized Implicit Policy for Offline Reinforcement Learning | Offline reinforcement learning enables learning from a fixed dataset , without further interactions with the environment . The lack of environmental interactions makes the policy training vulnerable to state-action pairs far from the training dataset and prone to missing rewarding actions . For training more effective ... | One of the existing Offline RL algorithms is to constrain the learned policy, such as constraining the learned policy to be consistent with the behavior policy itself or the action distribution based on state conditions, or adopting a Gaussian policy. However, in any given state s, the potential action value function i... | SP:558e16a6d21c668b92792c51293357eaecb4dc9d |
Objective Evaluation of Deep Visual Interpretations on Time Series Data | 1 INTRODUCTION . Due to its high performance on complex multi-modal data , deep learning ( DL ) becomes increasingly popular in many real-world applications that process time series data ( Fawaz et al. , 2019b ) . While we fundamentally rely on their classification accuracy in many safety-relevant applications ( Berken... | This study deals with the issue of interpretability in the analysis of time series data using deep learning models. Six metrics for evaluating interpretation methods are introduced and nine existing interpretation methods are evaluated. One of these metrics is proposed by the author. One of the metrics is based on an e... | SP:175791387c30347cc7f014645095f3a2c82a6c32 |
Objective Evaluation of Deep Visual Interpretations on Time Series Data | 1 INTRODUCTION . Due to its high performance on complex multi-modal data , deep learning ( DL ) becomes increasingly popular in many real-world applications that process time series data ( Fawaz et al. , 2019b ) . While we fundamentally rely on their classification accuracy in many safety-relevant applications ( Berken... | This paper studies deep neural network (DNN) explainability methods in the context of time-series data. Several metrics exist for evaluating the validity of DNN explainability methods on computer vision tasks. However, it is not clear whether these metrics are reliable when applied to DNN explainability methods on time... | SP:175791387c30347cc7f014645095f3a2c82a6c32 |
Unifying Top-down and Bottom-up for Recurrent Visual Attention | 1 INTRODUCTION . Recurrent visual attention model Mnih et al . ( 2014 ) ( abbreviated as RAM ) leverages reinforcement learning Sutton & Barto ( 1998 ) and recurrent neural networks Schuster et al . ( 1997 ) ; LeCun et al . ( 2015 ) to recognize the objects of interests in a sequential manner . Specifically , RAM model... | This submission proposes to offer a better initialization by using image pyramids and Q-learning (top-down manner) for the original recurrent visual attention (RAM) model (bottom-up manner). Two new constraints are also proposed for better exploration for RAM. The proposed method has been tested on several image classi... | SP:1eae7d672cd3f9a61f41af1cd46082370ee5b666 |
Unifying Top-down and Bottom-up for Recurrent Visual Attention | 1 INTRODUCTION . Recurrent visual attention model Mnih et al . ( 2014 ) ( abbreviated as RAM ) leverages reinforcement learning Sutton & Barto ( 1998 ) and recurrent neural networks Schuster et al . ( 1997 ) ; LeCun et al . ( 2015 ) to recognize the objects of interests in a sequential manner . Specifically , RAM model... | This paper extends the recurrent attention model(RAM) with another extra top-down attention. Specifically, they exploit image pyramids and Q-learning to select regions-of-interest first in the top-down attention mechanism, and then follow RAM to use policy gradient to find the patch in the bottom-up attention. Meanwhil... | SP:1eae7d672cd3f9a61f41af1cd46082370ee5b666 |
Adaptive Behavior Cloning Regularization for Stable Offline-to-Online Reinforcement Learning | 1 INTRODUCTION . Offline or batch reinforcement learning ( RL ) deals with the training of RL agents from fixed datasets generated by possibly unknown behavior policies , without any interactions with the environment . This is important in problems like robotics , autonomous driving , and healthcare where data collecti... | This paper proposes a new offline RL with online fine-tuning method. The authors first pretrain the policy using recent offline RL method TD3+BC with offline data and then collect on-policy data to further improve the pretrained policy. To prevent the policy from either degrading performance or failing to improve at th... | SP:c3b9fafd3676ec970b5d2a7dbb9d92b1ffda5959 |
Adaptive Behavior Cloning Regularization for Stable Offline-to-Online Reinforcement Learning | 1 INTRODUCTION . Offline or batch reinforcement learning ( RL ) deals with the training of RL agents from fixed datasets generated by possibly unknown behavior policies , without any interactions with the environment . This is important in problems like robotics , autonomous driving , and healthcare where data collecti... | This paper studies the fine-tuning problem from offline to online RL. While the naive approach to fine-tune offline policy suffers from a sudden distributional shift by online samples and too much behavior constraint in offline algorithms. The proposed method leverage (1) the adaptive coefficient tuning in TD3+BC loss,... | SP:c3b9fafd3676ec970b5d2a7dbb9d92b1ffda5959 |
Two Instances of Interpretable Neural Network for Universal Approximations | 1 INTRODUCTION . Artificial neural networks ( NN ) have recently seen successful applications in many fields . Modern deep neural network ( DNN ) architecture , usually trained through the backpropagation mechanism , has been called a black-box because of its lack of interpretability . To tackle this issue , various st... | This paper proposes two new neural networks (NN) construction schemes that aim at better interpretability, in the sense that (1) the NN should always memorize the training data; (2) the NN can roughly tell if a new test data has any similarity to any data in the training sample. The authors also provide approximation e... | SP:2e6e3e928f4a398cb89208e3af72d60022951b0f |
Two Instances of Interpretable Neural Network for Universal Approximations | 1 INTRODUCTION . Artificial neural networks ( NN ) have recently seen successful applications in many fields . Modern deep neural network ( DNN ) architecture , usually trained through the backpropagation mechanism , has been called a black-box because of its lack of interpretability . To tackle this issue , various st... | Summary: 1. This paper proposes two neural networks by construction, i.e., Triangularly-constructed Neural Network (TNN) and Semi-Quantized Activation Neural Network (SQANN). 2. These two neural networks are universal approximators, which is proven by construction. 3. These two neural networks are resistant to catas... | SP:2e6e3e928f4a398cb89208e3af72d60022951b0f |
ZenDet: Revisiting Efficient Object Detection Backbones from Zero-Shot Neural Architecture Search | In object detection models , the detection backbone consumes more than half of the overall inference cost . Recent researches attempt to reduce this cost by optimizing the backbone architecture with the help of Neural Architecture Search ( NAS ) . However , existing NAS methods for object detection require hundreds to ... | In this paper, the authors propose a zero-shot NAS to search backbone for detection task. Specifically, this method uses the differential entropy of output features as a metric to measure the performance of architecture on detection task. With the differential entropy , this method does not need training network parame... | SP:e2de45a266957eb22a06444dba4e4d58be9292be |
ZenDet: Revisiting Efficient Object Detection Backbones from Zero-Shot Neural Architecture Search | In object detection models , the detection backbone consumes more than half of the overall inference cost . Recent researches attempt to reduce this cost by optimizing the backbone architecture with the help of Neural Architecture Search ( NAS ) . However , existing NAS methods for object detection require hundreds to ... | This paper proposes a zero-shot neural architecture search approach for backbone design in object detection. The idea is to compute an entropy-based multi-scale Zen-score and use the score as an objective for evolutionary architecture search. The paper achieves better results comparing with previous zero-shot NAS appro... | SP:e2de45a266957eb22a06444dba4e4d58be9292be |
The Importance of the Current Input in Sequence Modeling | 1 INTRODUCTION . Deep learning models constitute the current state of the art in most artificial intelligence applications , from computer vision to robotics or medicine . When dealing with sequential data , Recurrent Neural Networks ( RNNs ) , specially those architectures with gating mechanisms such as the LSTM ( Hoc... | This paper proposes a simple improvement to recurrent network architectures for language modeling. The idea is to insert a single additional layer into a network with one or more recurrent layers, just before the output layer (Eq. 6). This is termed a "dual connection" and it combines the output of the last recurrent l... | SP:9be76bd104a9119c6aa5480f7cb70aba77519907 |
The Importance of the Current Input in Sequence Modeling | 1 INTRODUCTION . Deep learning models constitute the current state of the art in most artificial intelligence applications , from computer vision to robotics or medicine . When dealing with sequential data , Recurrent Neural Networks ( RNNs ) , specially those architectures with gating mechanisms such as the LSTM ( Hoc... | This work revisits the LSTM architecture. They propose to modify a recurrent architecture by adding a direct connection between the input and the output of the recurrent module, called "dual". They also consider a double-layer LSTM, where the output of the recurrent module is obtained by the concatenated application of... | SP:9be76bd104a9119c6aa5480f7cb70aba77519907 |
Adversarial Robustness as a Prior for Learned Representations | 1 INTRODUCTION . Beyond achieving remarkably high accuracy on a variety of tasks ( Krizhevsky et al. , 2012 ; He et al. , 2015 ; Collobert & Weston , 2008 ) , a major appeal of deep learning is the ability to learn effective representations of data . Specifically , deep neural networks can be thought of as linear class... | The paper looks at favorable properties of feature representations of an adversarially robust model. In particular, the authors look at a model trained with PGD training with an $\ell_p$ adversary. In terms of favourable properties, the authors look at representation inversion and feature manipulation and with experim... | SP:33d9ed48ce72f65860ffb34e77ed8b79b95b4869 |
Adversarial Robustness as a Prior for Learned Representations | 1 INTRODUCTION . Beyond achieving remarkably high accuracy on a variety of tasks ( Krizhevsky et al. , 2012 ; He et al. , 2015 ; Collobert & Weston , 2008 ) , a major appeal of deep learning is the ability to learn effective representations of data . Specifically , deep neural networks can be thought of as linear class... | This paper empirically demonstrates that robust optimization encourages deep neural networks to learn a high-level encoding of inputs. Specifically, this paper first utilize $\ell_2$- norm adversarial training to train robust neural networks. Then, this paper leverage two visualization techniques, i.e., *representation... | SP:33d9ed48ce72f65860ffb34e77ed8b79b95b4869 |
Two Regimes of Generalization for Non-Linear Metric Learning | 1 INTRODUCTION . Metric Learning , Bellet et al . ( 2015 ) , is the problem of finding a metric ρ on the space of features , such that ρ reflects some semantic properties of a given task . Generally , the input can be thought of as a set of labeled pairs { ( ( xi , x′i ) , yi ) } ni=1 , where xi , x′i ∈ Rd are the feat... | In this paper, the authors look at the Rademacher complexity of the family of Euclidean metrics learned on a data set via an $L$ layer network where the activations functions are Lipschitz. The idea behind the proof is to use the bounds for $\epsilon$ net for the embedding network from Barlett, Foster, and Telgarsky 20... | SP:935ed98ae6074d793f311b963ea3944a8e2c6293 |
Two Regimes of Generalization for Non-Linear Metric Learning | 1 INTRODUCTION . Metric Learning , Bellet et al . ( 2015 ) , is the problem of finding a metric ρ on the space of features , such that ρ reflects some semantic properties of a given task . Generally , the input can be thought of as a set of labeled pairs { ( ( xi , x′i ) , yi ) } ni=1 , where xi , x′i ∈ Rd are the feat... | The paper provides two new generalization bounds for non-linear metric learning with deep neural networks, by extending results of Bartlett et al. 2017 to the metric learning setting. The two bounds have been called the 'sparse' and 'non-sparse' bounds and differ in the norm used for the last layer. Experiments are per... | SP:935ed98ae6074d793f311b963ea3944a8e2c6293 |
Understanding the robustness-accuracy tradeoff by rethinking robust fairness | 1 INTRODUCTION . 1.1 BACKGROUND . Deep neural networks ( DNNs ) have been proven to be vulnerable to the adversarial attacks , as demonstrated in ( Szegedy ; Goodfellow et al . ; Kurakin et al . ; Carlini & Wagner ) . By adding crafted imperceptible perturbations to the input , attackers can easily fool the model to gi... | This paper identifies how adversarial training (AT) algorithms for robustness may negatively affect the notion of robust fairness and proposes two methods ME-AT and ME-TRADES, which combine existing AT methods with a maximum entropy (ME) term, to improve the accuracy-robustness tradeoff and robustness fairness. Althoug... | SP:e85499f1a98415c4392e71b1b5702cd89a35bd12 |
Understanding the robustness-accuracy tradeoff by rethinking robust fairness | 1 INTRODUCTION . 1.1 BACKGROUND . Deep neural networks ( DNNs ) have been proven to be vulnerable to the adversarial attacks , as demonstrated in ( Szegedy ; Goodfellow et al . ; Kurakin et al . ; Carlini & Wagner ) . By adding crafted imperceptible perturbations to the input , attackers can easily fool the model to gi... | This paper investigates inter-class similarity and intra-class variance, and corroborates that AT will cause an increase in inter-class similarity, which could be the root of both the robustness-accuracy tradeoff and robust fairness phenomena. The authors first considers Label Smoothing (LS) as the regularizer, and con... | SP:e85499f1a98415c4392e71b1b5702cd89a35bd12 |
Transformed CNNs: recasting pre-trained convolutional layers with self-attention | INTRODUCTION Since the success of AlexNet in 2012 ( Krizhevsky et al. , 2017 ) , the field of Computer Vision has been dominated by Convolutional Neural Networks ( CNNs ) ( LeCun et al. , 1998 ; 1989 ) . Their local receptive fields give them a strong inductive bias to exploit the spatial structure of natural images ( ... | This work proposes an approach to bridge CNNs and vision transformers for image recognition. The idea is to replace the last convolutional stage of a ResNet by a self-attention layer which is initialized from the weights of the convolutional stage. The approach is shown to improve the performance of the CNN, especially... | SP:be490ce3cc1184daa229cab9c9dbfede83f79469 |
Transformed CNNs: recasting pre-trained convolutional layers with self-attention | INTRODUCTION Since the success of AlexNet in 2012 ( Krizhevsky et al. , 2017 ) , the field of Computer Vision has been dominated by Convolutional Neural Networks ( CNNs ) ( LeCun et al. , 1998 ; 1989 ) . Their local receptive fields give them a strong inductive bias to exploit the spatial structure of natural images ( ... | The paper explores a hybrid type of model architectures that combines the recently popular transformer based model architectures with already well established convolutional neural networks. Specifically, the proposed hybrid model follows a two-stage training strategy where first a CNN model is trained and the pre-train... | SP:be490ce3cc1184daa229cab9c9dbfede83f79469 |
AS-MLP: An Axial Shifted MLP Architecture for Vision | 1 INTRODUCTION . In the past decade , Convolutional Neural Networks ( CNNs ) ( Krizhevsky et al. , 2012 ; He et al. , 2016 ) have received widespread attention and have become the de-facto standard for computer vision . Furthermore , with the in-depth exploration and research on self-attention , transformer-based archi... | The paper proposes a new architecture for computer vision that is inspired by (a) the Swin Transformer, (b) MLP-Mixer (and colleagues) and (c) CNN-like local context via shifts (like Shift, TSM, ViP, S2-MLP). The architecture is based on the Swin Transformer, removes the windowed-attention, and then adds "local shifts"... | SP:ad02657d8b6e822881aac3b489431e1ee0afeddc |
AS-MLP: An Axial Shifted MLP Architecture for Vision | 1 INTRODUCTION . In the past decade , Convolutional Neural Networks ( CNNs ) ( Krizhevsky et al. , 2012 ; He et al. , 2016 ) have received widespread attention and have become the de-facto standard for computer vision . Furthermore , with the in-depth exploration and research on self-attention , transformer-based archi... | This paper proposes to use the shift operation (Wu et al. CVPR 2018) in an axial manner for MLP-mixer architectures. The proposed method performs much better than previous MLP-based methods on ImageNet-1K and on par with Swin-transformer. | SP:ad02657d8b6e822881aac3b489431e1ee0afeddc |
Unsupervised Federated Learning is Possible | 1 INTRODUCTION . Federated Learning ( FL ) has received significant attention from both academic and industrial perspectives in that it can bring together separate data sources and allow multiple clients to train a central model in a collaborative but private manner ( McMahan et al. , 2017 ; Kairouz et al. , 2019 ; Yan... | This work presents a novel federated learning scheme to address the problem of learning from only unlabeled data. The main idea is clean and interesting, which constructs a global (server) model by aggregating the surrogate clients’ tasks from observing only unlabeled data for the classification tasks. The unlabeled da... | SP:a69fd07798a781ba3b1f06d26c6c4554abcfd30f |
Unsupervised Federated Learning is Possible | 1 INTRODUCTION . Federated Learning ( FL ) has received significant attention from both academic and industrial perspectives in that it can bring together separate data sources and allow multiple clients to train a central model in a collaborative but private manner ( McMahan et al. , 2017 ; Kairouz et al. , 2019 ; Yan... | The authors are proposing an approach for unsupervised Federated Learning. The authors propose to use Emperical risk minimizing for unsupervised learning. The authors assign labels to each of the class and uses a prior for the classes so that each of the client can learn without any labels. The authors have shown theo... | SP:a69fd07798a781ba3b1f06d26c6c4554abcfd30f |
Does Entity Abstraction Help Generative Transformers Reason? | 1 INTRODUCTION . Transformer language models ( TLMs ; Vaswani et al . 2017 ) have enabled rapid progress in natural language processing ( NLP ) . When pre-trained on large corpora ( such as the web ) to predict the next tokens or a set of masked tokens from an input sequence , TLMs can capture linguistic knowledge ( Pe... | This paper investigate incorporating entity abstraction to transformer language models for text reasoning tasks. The paper proposes different methods to inject entity abstraction information into transformer LMs and experiments on a synthetic dataset show that the proposed method helps compositional generalization. How... | SP:944ebd7c4c81d2266fe950501917c948f6925bb2 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.