paper_name stringlengths 11 170 | text stringlengths 8.07k 307k | summary stringlengths 152 6.16k | paper_id stringlengths 43 43 |
|---|---|---|---|
Bootstrapping Semantic Segmentation with Regional Contrast | 1 INTRODUCTION . Semantic segmentation is an essential part of applications such as scene understanding and autonomous driving , whose goal is to assign a semantic label to each pixel in an image . Significant progress has been achieved by use of large datasets with high quality human annotations . However , labelling ... | This work presents a new semi-supervised learning framework based on a pixel-level contrastive learning with active query/key sampling. For a given query feature with a class c, the positive key sample is generated as a mean vector of all samples with the class c, while the negative key sample is sampled by considering... | SP:0f708eb86de2495c91df25d36a9dec0cb03c8c62 |
Bootstrapping Semantic Segmentation with Regional Contrast | 1 INTRODUCTION . Semantic segmentation is an essential part of applications such as scene understanding and autonomous driving , whose goal is to assign a semantic label to each pixel in an image . Significant progress has been achieved by use of large datasets with high quality human annotations . However , labelling ... | The paper extends semantic segmentation models with a contrastive loss on "representations" at different locations in the image. A separate "representation head" is built from the intermediate features (between the "encoder" and "decoder") of the fully-convolutional segmentation network (DeepLabV3+). The "regional cont... | SP:0f708eb86de2495c91df25d36a9dec0cb03c8c62 |
Uncertainty-based out-of-distribution detection requires suitable function space priors | 1 INTRODUCTION . One of the challenges that the modern machine learning community is striving to tackle is the detection of unseen inputs for which predictions should not be trusted . This problem is also known as out-of-distribution ( OOD ) detection . The challenging nature of this task is partly rooted in the fact t... | This paper carries out an analysis that motivates that the use of the Bayesian predictive distribution and its uncertainty is not appropriate for detecting out of distribution data. The paper focuses on the case of Bayesian Neural networks and its infinite-wide generalization. Namely, a Gaussian process. The paper has ... | SP:9c11e1076ba18ba379ab89048cb69d7e5039351f |
Uncertainty-based out-of-distribution detection requires suitable function space priors | 1 INTRODUCTION . One of the challenges that the modern machine learning community is striving to tackle is the detection of unseen inputs for which predictions should not be trusted . This problem is also known as out-of-distribution ( OOD ) detection . The challenging nature of this task is partly rooted in the fact t... | The manuscript challenges the widely believed assumption that Bayesian neural networks (BNNs) are well suit for out-of-distribution (OOD) detection, by showing empirical results obtained using infinite-width (allowing the exact inference, because a network can be equivalently represented by a GP with the included kerne... | SP:9c11e1076ba18ba379ab89048cb69d7e5039351f |
Uncertainty-based out-of-distribution detection requires suitable function space priors | 1 INTRODUCTION . One of the challenges that the modern machine learning community is striving to tackle is the detection of unseen inputs for which predictions should not be trusted . This problem is also known as out-of-distribution ( OOD ) detection . The challenging nature of this task is partly rooted in the fact t... | In this work, the authors challenge the assumption that underlies many recent works that Bayesian neural networks should be well-suited to out-of-distribution detection. In order to do this, the authors focus on a function-space view by examining the properties of infinite-width BNNs. They use this analysis to argue th... | SP:9c11e1076ba18ba379ab89048cb69d7e5039351f |
Improving Hyperparameter Optimization by Planning Ahead | 1 INTRODUCTION . Hyperparameter optimization ( HPO ) is a ubiquitous problem within the research community and an integral aspect of tuning machine learning algorithms to ensure generalization beyond the training data . HPO is often posed as a sequential decision-making process , however , it can be seen as a special u... | The authors establish the equivalency of hyperparameter optimization (HPO) and model-based reinforcement learning (MRL). On one hand, hyperparameter optimization is seen as an optimization of a sequence of actions (hyperparameter candidates) that improves a reward function. On the other hand, planning replaces the acqu... | SP:a05319d4a1fdaf55457e30366eb5a9632346471d |
Improving Hyperparameter Optimization by Planning Ahead | 1 INTRODUCTION . Hyperparameter optimization ( HPO ) is a ubiquitous problem within the research community and an integral aspect of tuning machine learning algorithms to ensure generalization beyond the training data . HPO is often posed as a sequential decision-making process , however , it can be seen as a special u... | This paper proposes a method for hyperparameter optimization (HO) drawing inspiration from model-based RL (MBRL). They recast the HO problem in a markov decision process (MDP) formulation. This formulation permits the use of MBRL methods to solve the HO problem. Concretely, having reframed HO using this MDP formulati... | SP:a05319d4a1fdaf55457e30366eb5a9632346471d |
Improving Hyperparameter Optimization by Planning Ahead | 1 INTRODUCTION . Hyperparameter optimization ( HPO ) is a ubiquitous problem within the research community and an integral aspect of tuning machine learning algorithms to ensure generalization beyond the training data . HPO is often posed as a sequential decision-making process , however , it can be seen as a special u... | The paper presents a novel transfer learning approach to hyperparameter optimization (HPO) by formulating the problem in a model-based reinforcement learning (MbRL) framework. In this setting, the transition function serves as a surrogate model for learning the validation loss of any black-box ML model which is trained... | SP:a05319d4a1fdaf55457e30366eb5a9632346471d |
Fast Differentiable Matrix Square Root | 1 INTRODUCTION . Consider a positive semi-definite matrix A . The principle square root A 1 2 and the inverse square root A− 1 2 ( often derived by calculating the inverse of A 1 2 ) are mathematically of practical interests , mainly because some desired spectral properties can be obtained by such transformations . An ... | The work proposes a fast method to solve the matrix square root. The proposed approach is differentiable and is faster than SVD and NS iteration. Thus, the proposed method is very suitable for optimizing deep neural networks that involves matrix square root computation. Its forward pass uses matrix taylor polynomial or... | SP:7925e57e7c96f6f86bb0058759369a89e35330f4 |
Fast Differentiable Matrix Square Root | 1 INTRODUCTION . Consider a positive semi-definite matrix A . The principle square root A 1 2 and the inverse square root A− 1 2 ( often derived by calculating the inverse of A 1 2 ) are mathematically of practical interests , mainly because some desired spectral properties can be obtained by such transformations . An ... | The paper addresses the problem of computing the matrix square root of a positive semidefinite matrix and computing the gradient of the matrix square root, so they can be used as a forward pass and a backward pass in a deep learning framework. The paper has two separate contributions. In the forward pass, two approxi... | SP:7925e57e7c96f6f86bb0058759369a89e35330f4 |
Fast Differentiable Matrix Square Root | 1 INTRODUCTION . Consider a positive semi-definite matrix A . The principle square root A 1 2 and the inverse square root A− 1 2 ( often derived by calculating the inverse of A 1 2 ) are mathematically of practical interests , mainly because some desired spectral properties can be obtained by such transformations . An ... | This paper aims to solve an important problem: how to efficiently compute matrix square root in both forward and backward propagations. To this end, the authors proposed to exploit Matrix Taylor Polynomial and Matrix Pade Approximation to compute matrix square root in forward propagations, while approximate Lyapunov eq... | SP:7925e57e7c96f6f86bb0058759369a89e35330f4 |
Neuro-Symbolic Forward Reasoning | 1 INTRODUCTION . Right from the time of Aristotle , reasoning has been in the center of the study of human behavior ( Miller , 1984 ) . Reasoning can be defined as the process of deriving conclusions and predictions from available data . The long-lasting goal of artificial intelligence has been to develop rational agen... | The paper proposes a method for differentiable reasoning using soft first-order logic. The key idea is to combine forward reasoning with object-based deep learning. In particular, after the perception process from a pre-trained object detector, the proposed Neuro-Symbolic Forward Reasoner (NSFR) converts the object-bas... | SP:13162e80e8792affe9dc33538a6d5e0802a437ec |
Neuro-Symbolic Forward Reasoning | 1 INTRODUCTION . Right from the time of Aristotle , reasoning has been in the center of the study of human behavior ( Miller , 1984 ) . Reasoning can be defined as the process of deriving conclusions and predictions from available data . The long-lasting goal of artificial intelligence has been to develop rational agen... | The authors propose a model for doing forwards reasoning. There are three parts: The model decomposes a scene into object, converts that to a grounded symbolic representation and then performs differentiable reasoning. The authors show classification performance on two datasets. While the authors propose a Neuro-Symbo... | SP:13162e80e8792affe9dc33538a6d5e0802a437ec |
Neuro-Symbolic Forward Reasoning | 1 INTRODUCTION . Right from the time of Aristotle , reasoning has been in the center of the study of human behavior ( Miller , 1984 ) . Reasoning can be defined as the process of deriving conclusions and predictions from available data . The long-lasting goal of artificial intelligence has been to develop rational agen... | The paper tackles the problem of how to incorporate symbolic reasoning into a differentiable, deep learning architecture. The authors present an architecture comprising a pre-trained, slot-based encoder, and a differentiable clausal reasoner. They apply it to visual, relational reasoning problems, evaluating the archit... | SP:13162e80e8792affe9dc33538a6d5e0802a437ec |
Distributionally Robust Fair Principal Components via Geodesic Descents | 1 INTRODUCTION . Machine learning models are ubiquitous in our daily lives and supporting the decision-making process in diverse domains . With their flourishing applications , there also surface numerous concerns regarding the fairness of the models ’ outputs ( Mehrabi et al. , 2021 ) . Indeed , these models are prone... | This paper studies the formulation, reformulation and algorithm for distributionally robust fairness-aware PCA. The reformulation exploits techniques from distributionally robust optimization, and the algorithm is based on reimannian sub-gradient descent. The theory is tested on UCI datasets. | SP:3a0592645fae2a4732b188588648b0b31f948233 |
Distributionally Robust Fair Principal Components via Geodesic Descents | 1 INTRODUCTION . Machine learning models are ubiquitous in our daily lives and supporting the decision-making process in diverse domains . With their flourishing applications , there also surface numerous concerns regarding the fairness of the models ’ outputs ( Mehrabi et al. , 2021 ) . Indeed , these models are prone... | This paper aims to improve fairness and distributional robustness in dimensionality reduction techniques. More specifically, it proposes a regularization term to principal compoenent analysis (PCA) such that the expected reconstruction errors for each groups (conditioned on a binary sensitive variable) are similar, and... | SP:3a0592645fae2a4732b188588648b0b31f948233 |
Distributionally Robust Fair Principal Components via Geodesic Descents | 1 INTRODUCTION . Machine learning models are ubiquitous in our daily lives and supporting the decision-making process in diverse domains . With their flourishing applications , there also surface numerous concerns regarding the fairness of the models ’ outputs ( Mehrabi et al. , 2021 ) . Indeed , these models are prone... | This paper considers the fair PCA with distribution ally robust optimization algorithm. The problem is of great interest. The key idea of the proposed method is to consider a penalized loss function with the penalty on the fairness criterion. The distributionally robust optimization and its reformulation is developed f... | SP:3a0592645fae2a4732b188588648b0b31f948233 |
Provable Hierarchy-Based Meta-Reinforcement Learning | 1 INTRODUCTION . Reinforcement learning ( RL ) has demonstrated tremendous successes in many domains ( Schulman et al. , 2015 ; Vinyals et al. , 2019 ; Schrittwieser et al. , 2020 ) , learning near-optimal policies despite limited supervision . Nevertheless , RL remains difficult to apply to problems requiring temporal... | This paper proposes a novel formulation to analyze the provable benefits of hierarchical RL algorithms. Based on this new formulation of hierarchical structures, they propose new algorithms to learn the latent hierarchy and apply the extracted hierarchy on the downstream tasks. Under several assumptions, they prove th... | SP:1138b5cfe53a5ddba9aeeb3c2b495ddf4b638708 |
Provable Hierarchy-Based Meta-Reinforcement Learning | 1 INTRODUCTION . Reinforcement learning ( RL ) has demonstrated tremendous successes in many domains ( Schulman et al. , 2015 ; Vinyals et al. , 2019 ; Schrittwieser et al. , 2020 ) , learning near-optimal policies despite limited supervision . Nevertheless , RL remains difficult to apply to problems requiring temporal... | This paper presents a theoretical analysis of hierarchical reinforcement learning in a meta-RL setting. This work is focused on a tabular case. To quantify the importance of the state-action pair, $\alpha$-importance is introduced. Subsequently, for exit coverage, optimistic imagination is introduced. The proposed alg... | SP:1138b5cfe53a5ddba9aeeb3c2b495ddf4b638708 |
Provable Hierarchy-Based Meta-Reinforcement Learning | 1 INTRODUCTION . Reinforcement learning ( RL ) has demonstrated tremendous successes in many domains ( Schulman et al. , 2015 ; Vinyals et al. , 2019 ; Schrittwieser et al. , 2020 ) , learning near-optimal policies despite limited supervision . Nevertheless , RL remains difficult to apply to problems requiring temporal... | The authors develop a theoretical framework to discover the latent hierarchical structures shared across meta-training RL tasks, and propose a tractable hierarchy-learning algorithm with provable guarantees. The paper seems to be theoretically solid, and is well organized. I appreciate the proposed notions, such as lat... | SP:1138b5cfe53a5ddba9aeeb3c2b495ddf4b638708 |
Improved Image Generation via Sparsity | 1 INTRODUCTION . The use of Generative Adversarial Networks ( GANs ) for image synthesis is one of the most fascinating outcomes of the emerging deep learning era , e.g. , Goodfellow et al . ( 2014 ) ; Radford et al . ( 2015 ) ; Zhu et al . ( 2017 ) ; Ledig et al . ( 2017 ) ; Karras et al . ( 2018 ) . GAN is a machine-... | The paper proposes viewing CNN's through the lens of sparse coding. In practice it mostly boils down to encouraging sparsity in the last layer activations. To this end the paper proposes three different mechanisms that promote sparsity. This is shown to improve the FID scores somewhat in one dataset with a range of dif... | SP:fcb3dd25c73be8cf2a9bd4109c6e93530a4fc626 |
Improved Image Generation via Sparsity | 1 INTRODUCTION . The use of Generative Adversarial Networks ( GANs ) for image synthesis is one of the most fascinating outcomes of the emerging deep learning era , e.g. , Goodfellow et al . ( 2014 ) ; Radford et al . ( 2015 ) ; Zhu et al . ( 2017 ) ; Ledig et al . ( 2017 ) ; Karras et al . ( 2018 ) . GAN is a machine-... | This paper proposes a CNN-based (specifically GAN-based) solution to tackle image synthesis using sparse coding concept. The proposed method utilises the generator from a generative adversarial network to synthesise a sparse representation, i.e. the sparse code, for image synthesis. This method proposes to split origin... | SP:fcb3dd25c73be8cf2a9bd4109c6e93530a4fc626 |
Improved Image Generation via Sparsity | 1 INTRODUCTION . The use of Generative Adversarial Networks ( GANs ) for image synthesis is one of the most fascinating outcomes of the emerging deep learning era , e.g. , Goodfellow et al . ( 2014 ) ; Radford et al . ( 2015 ) ; Zhu et al . ( 2017 ) ; Ledig et al . ( 2017 ) ; Karras et al . ( 2018 ) . GAN is a machine-... | The paper describes image generation as a sparse coding reconstruction process. The authors formulate the sparse coding as a core block of a convolution image generator and test several different approaches for enforcing sparsity in the trained network. The experiments show that the method can improve FID scores of com... | SP:fcb3dd25c73be8cf2a9bd4109c6e93530a4fc626 |
SAINT: Improved Neural Networks for Tabular Data via Row Attention and Contrastive Pre-Training | 1 INTRODUCTION . While machine learning for image and language processing has seen major advances over the last decade , many critical industries , including financial services , health care , and logistics , rely heavily on data in structured format . Tabular data is unique in several ways that have prevented it from ... | The paper provides (main contributions) a new deep learning architecture (SAINT) for tabular data that performs attention over both samples and features. For datasets with missing labels, the authors also analyse a new contrastive self-supervised pre-training approach. The paper also introduces a new embedding method ... | SP:9b39210369f5124989f3d154c8293b00bed5c19b |
SAINT: Improved Neural Networks for Tabular Data via Row Attention and Contrastive Pre-Training | 1 INTRODUCTION . While machine learning for image and language processing has seen major advances over the last decade , many critical industries , including financial services , health care , and logistics , rely heavily on data in structured format . Tabular data is unique in several ways that have prevented it from ... | This paper proposes SAINT, a neural network model for handling tabular data with both continuous and discrete values. SAINT uses both self-attention among variables and inter-sample attention among different samples. Using both InfoNCE and denoising objectives for pre-training strategies, SAINT was able to outrank all ... | SP:9b39210369f5124989f3d154c8293b00bed5c19b |
SAINT: Improved Neural Networks for Tabular Data via Row Attention and Contrastive Pre-Training | 1 INTRODUCTION . While machine learning for image and language processing has seen major advances over the last decade , many critical industries , including financial services , health care , and logistics , rely heavily on data in structured format . Tabular data is unique in several ways that have prevented it from ... | The paper presents a new deep-learning network architecture for tabular data classification and regression problems, called SAINT. While deep-learning provided significant improvements in many domains, tabular data problems are still dominated by classical algorithms like XGBoost and Catboost. The new architecture is b... | SP:9b39210369f5124989f3d154c8293b00bed5c19b |
Should We Be Pre-training? An Argument for End-task Aware Training as an Alternative | In most settings of practical concern , machine learning practitioners know in advance what end-task they wish to boost with auxiliary tasks . However , widely used methods for leveraging auxiliary data like pre-training and its continuedpretraining variant are end-task agnostic : they rarely , if ever , exploit knowle... | This paper considers the common setup of pretraining and finetuning paradigm and argues that an end-task aware setting, either by multitasking or an additional meta learning that learns the weights between various auxiliary tasks and the end task, is superior than end-task agnostic pretraining. The paper provides form... | SP:e585363319b3741d3342cb271b359d1541c11e99 |
Should We Be Pre-training? An Argument for End-task Aware Training as an Alternative | In most settings of practical concern , machine learning practitioners know in advance what end-task they wish to boost with auxiliary tasks . However , widely used methods for leveraging auxiliary data like pre-training and its continuedpretraining variant are end-task agnostic : they rarely , if ever , exploit knowle... | In the paper the authors propose TARTAN, methods to enable end-task aware pre-training. MT-TARTAN simply combines the pre-training objectives and the end-task objective as multi-task learning. META-TARTAN learns a set of weights for the pre-training objective and the end-task objective, using meta-learning. MT-TARTAN a... | SP:e585363319b3741d3342cb271b359d1541c11e99 |
Should We Be Pre-training? An Argument for End-task Aware Training as an Alternative | In most settings of practical concern , machine learning practitioners know in advance what end-task they wish to boost with auxiliary tasks . However , widely used methods for leveraging auxiliary data like pre-training and its continuedpretraining variant are end-task agnostic : they rarely , if ever , exploit knowle... | The paper makes the argument that generic pre-training (on auxiliary tasks) is inferior to task-specific pre-training. The authors argue that the final task should be learned together with the auxiliary tasks in a multi-task setup, which they call MT-TARTAN. They also propose a meta-learning algorithm META-TARTAN that ... | SP:e585363319b3741d3342cb271b359d1541c11e99 |
Learned Simulators for Turbulence | 1 INTRODUCTION . Turbulent fluid dynamics are ubiquitous throughout the natural world , and many scientific and engineering disciplines depend on high quality simulations of turbulence . Examples range from design problems in aeronautics ( Rhie , 1983 ) and medicine ( Sallam & Hwang , 1984 ) to scientific problems of u... | The authors evaluate some different architectures of neural nets for learning turbulence. They showcase the results for four different systems and in terms of several metrics. The main conclusion is that a learned simulator can indeed be good even for turbulence and possibly improved with additional mechanisms such as ... | SP:d55e3120b000495878f24bc589dd386152dad943 |
Learned Simulators for Turbulence | 1 INTRODUCTION . Turbulent fluid dynamics are ubiquitous throughout the natural world , and many scientific and engineering disciplines depend on high quality simulations of turbulence . Examples range from design problems in aeronautics ( Rhie , 1983 ) and medicine ( Sallam & Hwang , 1984 ) to scientific problems of u... | This paper present a neural network-based model by training learned simulators at low spatial and temporal resolutions to capture turbulent dynamics generated at high resolution. It also shows that the proposed model can simulate turbulent dynamics more accurately than classical numerical solvers at the same low resol... | SP:d55e3120b000495878f24bc589dd386152dad943 |
Learned Simulators for Turbulence | 1 INTRODUCTION . Turbulent fluid dynamics are ubiquitous throughout the natural world , and many scientific and engineering disciplines depend on high quality simulations of turbulence . Examples range from design problems in aeronautics ( Rhie , 1983 ) and medicine ( Sallam & Hwang , 1984 ) to scientific problems of u... | This paper proposes a method to predict turbulent dynamics of coarse simulations. To achieve this, a modified ResNet architecture is proposed, which the authors call Dilated ResNet (dResNet). This dResNet is combined with an encoder/decoder model inspired by graph networks. Unfortunately, the details of this architectu... | SP:d55e3120b000495878f24bc589dd386152dad943 |
The NTK Adversary: An Approach to Adversarial Attacks without any Model Access | 1 INTRODUCTION . Adversarial examples , a term coined by Szegedy et al . ( 2014 ) , are inputs of a machine learning model that are carefully crafted so that the model is mistaken on them . Their existence , together with the empirically verified susceptibility of many classifiers to them ( Szegedy et al. , 2014 ; Carl... | This paper proposes the use of Neural Tangent Kernel (NTK) to generate transferrable adversarial examples without access to the target model. The main contributions are: 1. The derivation of adversarial perturbation under NTK setting. 2. Illustration of different attack scenarios using NTK-attack (e.g. no training n... | SP:9d07d9b76c1e7dc32a1ac08188d28d04d5c260b2 |
The NTK Adversary: An Approach to Adversarial Attacks without any Model Access | 1 INTRODUCTION . Adversarial examples , a term coined by Szegedy et al . ( 2014 ) , are inputs of a machine learning model that are carefully crafted so that the model is mistaken on them . Their existence , together with the empirically verified susceptibility of many classifiers to them ( Szegedy et al. , 2014 ; Carl... | In this paper the authors introduce and study a new approach to generate Adversarial Attacks for image classification models. The core of their approach lies in leveraging Neural Tangent Kernel (NTK) formulations of Deep Neural Networks. By using the analytic and empirical NTK versions of common network architectures t... | SP:9d07d9b76c1e7dc32a1ac08188d28d04d5c260b2 |
The NTK Adversary: An Approach to Adversarial Attacks without any Model Access | 1 INTRODUCTION . Adversarial examples , a term coined by Szegedy et al . ( 2014 ) , are inputs of a machine learning model that are carefully crafted so that the model is mistaken on them . Their existence , together with the empirically verified susceptibility of many classifiers to them ( Szegedy et al. , 2014 ; Carl... | This paper investigates to use NTK as a proxy to generate adversarial examples for neural nets. I like the paper in general. The derivations are clear and the experiments are clearly presented. However, I think this paper suffers significant problems from unclear formulation of the threat model. | SP:9d07d9b76c1e7dc32a1ac08188d28d04d5c260b2 |
Sample Efficient Stochastic Policy Extragradient Algorithm for Zero-Sum Markov Game | Two-player zero-sum Markov game is a fundamental problem in reinforcement learning and game theory . Although many algorithms have been proposed for solving zero-sum Markov games in the existing literature , many of them either require a full knowledge of the environment or are not sample-efficient . In this paper , we... | The authors consider two-player zero-sum Markov games. The goal is to develop decentralized, model-free, symmetric algorithm. With entropy regularization and new estimators, the authors show the stochastic policy extragradient algorithms have these properties and the sample complexity improves the sate-of-the-art. | SP:2bdd583398e2aac0b1816c19c4e88757fc24cbd9 |
Sample Efficient Stochastic Policy Extragradient Algorithm for Zero-Sum Markov Game | Two-player zero-sum Markov game is a fundamental problem in reinforcement learning and game theory . Although many algorithms have been proposed for solving zero-sum Markov games in the existing literature , many of them either require a full knowledge of the environment or are not sample-efficient . In this paper , we... | ## Summary This paper studies a decentralized stochastic policy extra-gradient algorithm for solving two-player zero-sum Markov game. In comparison to the standard policy extra-gradient algorithm, this algorithm uses a set of stochastic estimators to estimate the value functions involved in the stochastic updates. Im... | SP:2bdd583398e2aac0b1816c19c4e88757fc24cbd9 |
Sample Efficient Stochastic Policy Extragradient Algorithm for Zero-Sum Markov Game | Two-player zero-sum Markov game is a fundamental problem in reinforcement learning and game theory . Although many algorithms have been proposed for solving zero-sum Markov games in the existing literature , many of them either require a full knowledge of the environment or are not sample-efficient . In this paper , we... | This paper focuses on the two-player zero-sum Markov Game setting and proposes a stochastic version of the policy extragradient (PE). The PE algorithm, which solves an entropy-regularized minimax matrix game problem by predictive update (PU) [Cen et al., 2021], has been well studied under the deterministic case in the ... | SP:2bdd583398e2aac0b1816c19c4e88757fc24cbd9 |
Chunked Autoregressive GAN for Conditional Waveform Synthesis | Conditional waveform synthesis models learn a distribution of audio waveforms given conditioning such as text , mel-spectrograms , or MIDI . These systems employ deep generative models that model the waveform via either sequential ( autoregressive ) or parallel ( non-autoregressive ) sampling . Generative adversarial n... | This paper proposes a conditional waveform synthesis (CWS) model called Chunked Autoregressive GAN (CARGAN). By combining the advantages of the AR and non-AR generative models, CARGAN achieves better pitch accuracy with faster training speed and less memory usage without much decline in generation speed. First, it sho... | SP:2c3b558b60fa3fb1d888ad5c6677a499b13e82d4 |
Chunked Autoregressive GAN for Conditional Waveform Synthesis | Conditional waveform synthesis models learn a distribution of audio waveforms given conditioning such as text , mel-spectrograms , or MIDI . These systems employ deep generative models that model the waveform via either sequential ( autoregressive ) or parallel ( non-autoregressive ) sampling . Generative adversarial n... | The paper introduces a Chunked Autoregressive GAN (CARGAN) method for conditional synthesis which is autoregressive over chunks of audio but uses Hifi-GAN like parallel generation within a chunk. The method is motivated by the periodicity and pitch errors shown by existing parallel (non-autoregressive) GAN generators. ... | SP:2c3b558b60fa3fb1d888ad5c6677a499b13e82d4 |
Chunked Autoregressive GAN for Conditional Waveform Synthesis | Conditional waveform synthesis models learn a distribution of audio waveforms given conditioning such as text , mel-spectrograms , or MIDI . These systems employ deep generative models that model the waveform via either sequential ( autoregressive ) or parallel ( non-autoregressive ) sampling . Generative adversarial n... | The authors present a chunk-wise auto-regressive generative model for audio with adversarial loss. In particular, the authors note the limitations of purely convolutional adversarial audio generation model for text to speech as those fail to provide a consistent pitch for an extended duration. The authors provide a num... | SP:2c3b558b60fa3fb1d888ad5c6677a499b13e82d4 |
Self-supervised regression learning using domain knowledge: Applications to improving self-supervised image denoising | 1 INTRODUCTION . Deep regression neural network ( NN ) -based methods that can accurately predict real- or complexvalued output have been rapidly gaining popularity in a wide range of computational imaging and computer vision applications including image denoising ( Vincent et al. , 2010 ; Xie et al. , 2012 ; Zhang et ... | This paper proposes a method for unsupervised image denoising. It shows that a better designed operator based on domain knowledge can benefit unsupervised image denoising task. The provided experimental results show the proposed method outperforms existing unsupervised denoising ones and achieves similar performance to... | SP:fc1c3797c7b073cfa8db2e40d53aa0b2f12829de |
Self-supervised regression learning using domain knowledge: Applications to improving self-supervised image denoising | 1 INTRODUCTION . Deep regression neural network ( NN ) -based methods that can accurately predict real- or complexvalued output have been rapidly gaining popularity in a wide range of computational imaging and computer vision applications including image denoising ( Vincent et al. , 2010 ; Xie et al. , 2012 ; Zhang et ... | This paper proposed a self-supervised denoing method using domain knowledge. The proposed method seems somewhat like knowledge distillation, where a pre-trained denoiser or handcrafted designed noise models g can provide better initial results than original noisy images, and then a better denoiser f can be trained base... | SP:fc1c3797c7b073cfa8db2e40d53aa0b2f12829de |
Self-supervised regression learning using domain knowledge: Applications to improving self-supervised image denoising | 1 INTRODUCTION . Deep regression neural network ( NN ) -based methods that can accurately predict real- or complexvalued output have been rapidly gaining popularity in a wide range of computational imaging and computer vision applications including image denoising ( Vincent et al. , 2010 ; Xie et al. , 2012 ; Zhang et ... | This paper present an self-supervised deep learning method for image denoising, which is a generalization of the concept proposed in noise2self. The basic idea is to approximate the unbiased estimation of the supervised loss function, which utilized the noise independence of the input and the output of the network. Som... | SP:fc1c3797c7b073cfa8db2e40d53aa0b2f12829de |
Lottery Ticket Structured Node Pruning for Tabular Datasets | In this paper we presented two pruning approaches on tabular neural networks based on the lottery ticket hypothesis that went beyond masking nodes by resizing the models accordingly . We showed top performing models in 6 of 8 datasets tested in terms of F1/RMSE . We also showed in 6 of 8 datasets a total reduction of o... | This paper investigates model pruning and the Lottery Ticket Hypothesis in the context of tabular datasets and model training. The authors apply a set of pruning techniques to the tabular neural networks from FastAI and examine whether the LTH still holds on the tabular datasets and their corresponding models. Differen... | SP:3ac384440a8866ac421210c154c904bfbdbb6276 |
Lottery Ticket Structured Node Pruning for Tabular Datasets | In this paper we presented two pruning approaches on tabular neural networks based on the lottery ticket hypothesis that went beyond masking nodes by resizing the models accordingly . We showed top performing models in 6 of 8 datasets tested in terms of F1/RMSE . We also showed in 6 of 8 datasets a total reduction of o... | This paper proposes two approaches to use sparse neural networks on tabular data. The topic is interesting and sometimes overlooked in the literature. The proposed approaches are based on node (likely referring here to neuron) pruning. Overall, I believe that the proposed method description, experimental evaluation, an... | SP:3ac384440a8866ac421210c154c904bfbdbb6276 |
Lottery Ticket Structured Node Pruning for Tabular Datasets | In this paper we presented two pruning approaches on tabular neural networks based on the lottery ticket hypothesis that went beyond masking nodes by resizing the models accordingly . We showed top performing models in 6 of 8 datasets tested in terms of F1/RMSE . We also showed in 6 of 8 datasets a total reduction of o... | This paper proposes two structured pruning strategies for compressing a (very specific) tabular neural network: iterative pruning, and one-shot pruning. Both methods work by removing the node with smallest $L_{-\infty}$ norm, i.e., the one with smallest minimum weight of each node. After removing a designated amount of... | SP:3ac384440a8866ac421210c154c904bfbdbb6276 |
On the benefits of deep RL in accelerated MRI sampling | 1 INTRODUCTION . Magnetic resonance imaging ( MRI ) is a non-invasive , non-ionizing medical imaging method which has been widely adopted in clinical settings due to its with unmatched quality in soft tissue contrast . However , MRI suffers from long scanning times , which limits patient comfort , imaging quality as we... | The authors report on the use and current limitations of using deep reinforcement learning for accelerated sampling in MR imaging. They focus on two specific works that use deep RL for sampling: Pineda and Bakker. They mainly compare these methods against a previously described non-RL technique: Stochastic Learning-bas... | SP:9c872eda015a2b603701fd5cc263f0ceb244baa7 |
On the benefits of deep RL in accelerated MRI sampling | 1 INTRODUCTION . Magnetic resonance imaging ( MRI ) is a non-invasive , non-ionizing medical imaging method which has been widely adopted in clinical settings due to its with unmatched quality in soft tissue contrast . However , MRI suffers from long scanning times , which limits patient comfort , imaging quality as we... | The paper first surveys several recent Fast MRI Recon methods, including RL-based and greedy-policy-based ones, that reconstruct high quality images from highly undersampled data from the perspective of optimizing the sampling of k-space data and then concerns the RL-based methods. Further it considers a typical Fast M... | SP:9c872eda015a2b603701fd5cc263f0ceb244baa7 |
On the benefits of deep RL in accelerated MRI sampling | 1 INTRODUCTION . Magnetic resonance imaging ( MRI ) is a non-invasive , non-ionizing medical imaging method which has been widely adopted in clinical settings due to its with unmatched quality in soft tissue contrast . However , MRI suffers from long scanning times , which limits patient comfort , imaging quality as we... | The paper investigates the effectiveness of deep RL methods for accelerating the acquisition of MRI scans. It compares two recent methods by Bakker et al. as well as Pineda et al. to non-RL baselines. The results indicate that deep RL provides little benefits over the LBCS baseline. | SP:9c872eda015a2b603701fd5cc263f0ceb244baa7 |
Strength of Minibatch Noise in SGD | 1 INTRODUCTION . Stochastic gradient descent ( SGD ) is the simple and efficient optimization algorithm behind the success of deep learning ( Allen-Zhu et al. , 2019 ; Xing et al. , 2018 ; Zhang et al. , 2018 ; Wang et al. , 2020 ; He and Tao , 2020 ; Liu et al. , 2021 ; Simsekli et al. , 2019 ; Wu et al. , 2020 ) . Mi... | This paper investigates the importance of noise in mini-batch SGD. The main contribution of this paper is to derive an analytic solution to the shape and strength of SGD minibatch noise for linear regression with random noise in the label, linear regression with additional L2 regularization and non-linear regression gi... | SP:c355f51df75e7fecabc14774158f1af55436389d |
Strength of Minibatch Noise in SGD | 1 INTRODUCTION . Stochastic gradient descent ( SGD ) is the simple and efficient optimization algorithm behind the success of deep learning ( Allen-Zhu et al. , 2019 ; Xing et al. , 2018 ; Zhang et al. , 2018 ; Wang et al. , 2020 ; He and Tao , 2020 ; Liu et al. , 2021 ; Simsekli et al. , 2019 ; Wu et al. , 2020 ) . Mi... | This paper studies the properties of the gradient noise in mini-batch SGD using discrete-time analysis for a fixed learning rate. Their analysis is more general than prior work and considers SGD with momentum, a learning rate matrix (that subsumes preconditioning methods), and regularization. They provide closed form e... | SP:c355f51df75e7fecabc14774158f1af55436389d |
Strength of Minibatch Noise in SGD | 1 INTRODUCTION . Stochastic gradient descent ( SGD ) is the simple and efficient optimization algorithm behind the success of deep learning ( Allen-Zhu et al. , 2019 ; Xing et al. , 2018 ; Zhang et al. , 2018 ; Wang et al. , 2020 ; He and Tao , 2020 ; Liu et al. , 2021 ; Simsekli et al. , 2019 ; Wu et al. , 2020 ) . Mi... | This paper studies the minibatch noise for discrete time SGD and discusses its approximation on different applications. The novelty of the paper stands on the derivation of minibatch noise covatiance of discrete time SGD. For special cases with label noise and L2 regularization, this work derives the exact solution of ... | SP:c355f51df75e7fecabc14774158f1af55436389d |
Evaluating Deep Graph Neural Networks | 1 INTRODUCTION . The recent success of Graph Neural Networks ( GNNs ) ( Zhang et al. , 2020 ) has boosted researches on knowledge discovery and data mining on graph data . Designed for graph-structured data , GNNs provide a universal way to tackle node-level , edge-level , and graph-level tasks , including social netwo... | This work performs a systematic study to analyze the main issues of the difficulty in training deep GNNs by disentangling the effects of embedding propagation (EP) and embedding transformation (ET). They find that the large $D_t$ is the root cause for the failure of deep GNNs. Node-adaptive combination mechanism and re... | SP:6bec4007a1821af4b17d5d0afef0ea9d48e74f52 |
Evaluating Deep Graph Neural Networks | 1 INTRODUCTION . The recent success of Graph Neural Networks ( GNNs ) ( Zhang et al. , 2020 ) has boosted researches on knowledge discovery and data mining on graph data . Designed for graph-structured data , GNNs provide a universal way to tackle node-level , edge-level , and graph-level tasks , including social netwo... | The paper considers the application of GNNs to semi-supervised node classification tasks. The problem arises in big data problems such as modeling citation databases. In this line of work it is assumed that the data follows a graph structure, i.e., that locally connected nodes are likely to have the same label. Each n... | SP:6bec4007a1821af4b17d5d0afef0ea9d48e74f52 |
Evaluating Deep Graph Neural Networks | 1 INTRODUCTION . The recent success of Graph Neural Networks ( GNNs ) ( Zhang et al. , 2020 ) has boosted researches on knowledge discovery and data mining on graph data . Designed for graph-structured data , GNNs provide a universal way to tackle node-level , edge-level , and graph-level tasks , including social netwo... | In this work, the authors performed an experimental evaluation on several GNNs in order to understand what aspects of the current architecture designs that leads to the compromised performance of deep GNNs. They claimed to find the root causes: large propagation depth leads to the over-smoothing issue and large transfo... | SP:6bec4007a1821af4b17d5d0afef0ea9d48e74f52 |
A Survey on Evidential Deep Learning For Single-Pass Uncertainty Estimation | 1 INTRODUCTION Many existing methods for uncertainty estimation leverage the concept of Bayesian Model Averaging , that approaches such as Monte Carlo ( MC ) Dropout ( Gal & Ghahramani , 2016 ) , Bayes-by-backprop ( Blundell et al. , 2015 ) or ensembling ( Lakshminarayanan et al. , 2017 ) can be grouped under ( Wilson ... | This paper surveys a collection of existing works that the author frames as evidential deep learning. For the classification case, evidential deep learning tries to train a network to output the parameters of a Dirichlet distribution, hoping the one could directly obtain the data uncertainty and model uncertainty from ... | SP:b7d4d7bb36f7033094e318b8ac64c283de026190 |
A Survey on Evidential Deep Learning For Single-Pass Uncertainty Estimation | 1 INTRODUCTION Many existing methods for uncertainty estimation leverage the concept of Bayesian Model Averaging , that approaches such as Monte Carlo ( MC ) Dropout ( Gal & Ghahramani , 2016 ) , Bayes-by-backprop ( Blundell et al. , 2015 ) or ensembling ( Lakshminarayanan et al. , 2017 ) can be grouped under ( Wilson ... | The paper is a survey of methods in evidential deep learning. It gives a brief motivation for this set of methods, explains a general framework, and describes previous works for classification and regressions tasks in varying amount of details. As a survey paper, it doesn't introduce novel ideas, but gives a useful ove... | SP:b7d4d7bb36f7033094e318b8ac64c283de026190 |
A Survey on Evidential Deep Learning For Single-Pass Uncertainty Estimation | 1 INTRODUCTION Many existing methods for uncertainty estimation leverage the concept of Bayesian Model Averaging , that approaches such as Monte Carlo ( MC ) Dropout ( Gal & Ghahramani , 2016 ) , Bayes-by-backprop ( Blundell et al. , 2015 ) or ensembling ( Lakshminarayanan et al. , 2017 ) can be grouped under ( Wilson ... | The paper presents a survey of evidential deep learning, a family of machine learning methods that suggest accounting for various forms of uncertainties via a hierarchical predictive model whose hyperprior parameters are a function of input observations. The paper classifies evidential deep learning approaches into cat... | SP:b7d4d7bb36f7033094e318b8ac64c283de026190 |
Hermitry Ratio: Evaluating the validity of perturbation methods for explainable deep learning | 1 INTRODUCTION . In recent years there has been an explosion of explanation methods for artificial intelligence ( XAI ) algorithms , largely due to the fact that AI is used in pretty much every facet of our modern lives . The need for trust and understanding in AI algorithms has never been more important . A popular wa... | This paper aims to evaluate whether perturbation methods for explainable deep learning generates out-of-the-distribution (OOD) samples which can weaken the explanations provided by these methods. Basically, we would not know if the changes in prediction are due to the removal of some important image feature or because ... | SP:f43d886e94cad012c564827d43004ef57c0ce020 |
Hermitry Ratio: Evaluating the validity of perturbation methods for explainable deep learning | 1 INTRODUCTION . In recent years there has been an explosion of explanation methods for artificial intelligence ( XAI ) algorithms , largely due to the fact that AI is used in pretty much every facet of our modern lives . The need for trust and understanding in AI algorithms has never been more important . A popular wa... | The focus of this paper is on perturbation methods as a means of extracting explainable information about a learner. The authors question the premise underlying perturbation methods as existent in the current literature, and hence introduce a concept they term the hermitry ratio. This ratio is supposed to "to indicat... | SP:f43d886e94cad012c564827d43004ef57c0ce020 |
Hermitry Ratio: Evaluating the validity of perturbation methods for explainable deep learning | 1 INTRODUCTION . In recent years there has been an explosion of explanation methods for artificial intelligence ( XAI ) algorithms , largely due to the fact that AI is used in pretty much every facet of our modern lives . The need for trust and understanding in AI algorithms has never been more important . A popular wa... | The paper measures a statistic how much image perturbation methods generate outliers. In principle they measure the amount of samples whose distances are above a treshold which is determined by a quantile on in-manifold data. This is a relevant question for the research community. The idea of hermitry ratio can be foun... | SP:f43d886e94cad012c564827d43004ef57c0ce020 |
Dataset transformations trade-offs to adapt machine learning methods across domains | Machine learning-based methods have been proved to be quite successful in different domains . However , applying the same methods across domains is not a trivial task . In the literature , the most common approach is to convert a dataset into the same format as the original domain to employ the same architecture that w... | This paper attempts to study the effect of different transformations of input datasets, and the performance of the resulting models thereof. In particular, authors consider time-series data which is then represented in (1) time-series format, (2) vectorized form, (3) tensorized form. For each of the representation, a d... | SP:6b4c9e9813f8b9da3bcf4bbb72f92a202ee86722 |
Dataset transformations trade-offs to adapt machine learning methods across domains | Machine learning-based methods have been proved to be quite successful in different domains . However , applying the same methods across domains is not a trivial task . In the literature , the most common approach is to convert a dataset into the same format as the original domain to employ the same architecture that w... | - Draft attempts to understand the effect of converting dataset format while applying an ML technique. It argues that converting the target data into a format on which the specific ML technique was shown to be successful is suboptimal. Via simple experiments it shows that the data sample format conversion need not be a... | SP:6b4c9e9813f8b9da3bcf4bbb72f92a202ee86722 |
Dataset transformations trade-offs to adapt machine learning methods across domains | Machine learning-based methods have been proved to be quite successful in different domains . However , applying the same methods across domains is not a trivial task . In the literature , the most common approach is to convert a dataset into the same format as the original domain to employ the same architecture that w... | This paper aims to study the effects of different transformation methods when processing a dataset for supervised deep learning tasks. The authors conducted one experiment to show that different data arranging methods leads to different vulnerability for adversarial attacks. In the experimental procedure, an optimal tr... | SP:6b4c9e9813f8b9da3bcf4bbb72f92a202ee86722 |
Group-based Interleaved Pipeline Parallelism for Large-scale DNN Training | 1 INTRODUCTION . Several recent lines of research ( Liu et al. , 2019 ; Yang et al. , 2019 ; Lan et al. , 2019 ; Raffel et al. , 2019 ; Brown et al. , 2020 ; Lin et al. , 2021 ) on various application domains have collectively demonstrated that larger DNN models can yield better performance . This creates an emerging t... | The paper presents a novel pipeline parallelism technique for training large DNN models named as WPipe. The method aims to improve upon existing pipeline methods such as Pipedream-2BW by having a better memory efficiency and more fresh weight updates. Key idea is to divide the model partitions into two parts groups and... | SP:1a2313cdfc69d8feae1ca20ffce030c5412020f6 |
Group-based Interleaved Pipeline Parallelism for Large-scale DNN Training | 1 INTRODUCTION . Several recent lines of research ( Liu et al. , 2019 ; Yang et al. , 2019 ; Lan et al. , 2019 ; Raffel et al. , 2019 ; Brown et al. , 2020 ; Lin et al. , 2021 ) on various application domains have collectively demonstrated that larger DNN models can yield better performance . This creates an emerging t... | This paper introduces a novel pipeline training strategy WPipe. WPipe divides model partitions into two groups and updates each group alternatively, which eliminates half of the delayed gradients and memory redundancy compared to Pipedream-2BW. The experimental results show that WPipe can achieve higher throughputs a... | SP:1a2313cdfc69d8feae1ca20ffce030c5412020f6 |
Group-based Interleaved Pipeline Parallelism for Large-scale DNN Training | 1 INTRODUCTION . Several recent lines of research ( Liu et al. , 2019 ; Yang et al. , 2019 ; Lan et al. , 2019 ; Raffel et al. , 2019 ; Brown et al. , 2020 ; Lin et al. , 2021 ) on various application domains have collectively demonstrated that larger DNN models can yield better performance . This creates an emerging t... | The paper addresses the problem of very large scale deep neural network training through model parallelism. A new model parallel pipeline called WPipe is proposed, that builds on previously existing schemes while targeting (1) lower memory redundancy and (2) fresher weight updates. Experimental evaluation shows that WP... | SP:1a2313cdfc69d8feae1ca20ffce030c5412020f6 |
Frequency-aware SGD for Efficient Embedding Learning with Provable Benefits | 1 INTRODUCTION . Embedding learning describes a problem of learning dense real-valued vector representation for categorical data , often referred to as token ( Pennington et al. , 2014 ; Mikolov et al. , 2013a ; b ) . Good quality embeddings can capture rich semantic information of tokens , and thus serve as the corner... | This paper proposes a frequency-aware learning algorithm for embedding learning in recommender systems. The idea of incoporating frequency information into the learning algorithm is very interesting to me. The proposed method is very simple and easy to implement with provable benefits. | SP:66e64c8417f109860030ce79e151384f9dd33308 |
Frequency-aware SGD for Efficient Embedding Learning with Provable Benefits | 1 INTRODUCTION . Embedding learning describes a problem of learning dense real-valued vector representation for categorical data , often referred to as token ( Pennington et al. , 2014 ; Mikolov et al. , 2013a ; b ) . Good quality embeddings can capture rich semantic information of tokens , and thus serve as the corner... | This paper proposes a counter-based learning-rate scheduler for SGD. This algorithm is designed based on the long tail distribution in recommendations and languages. The proposed algorithm enjoys a theoretic guarantee unlike other adaptive learning rate methods, e.g., Adams. Simultaneously, the authors demonstrated tha... | SP:66e64c8417f109860030ce79e151384f9dd33308 |
Frequency-aware SGD for Efficient Embedding Learning with Provable Benefits | 1 INTRODUCTION . Embedding learning describes a problem of learning dense real-valued vector representation for categorical data , often referred to as token ( Pennington et al. , 2014 ; Mikolov et al. , 2013a ; b ) . Good quality embeddings can capture rich semantic information of tokens , and thus serve as the corner... | This paper proposes two optimization algorithms for recommendation where the token distributions are highly imbalanced. The frequency information is integrated into the optimization algorithms for fast convergence and better performance. The proposed algorithms are easy to understand and implement, and the theoretical ... | SP:66e64c8417f109860030ce79e151384f9dd33308 |
Conditional Image Generation by Conditioning Variational Auto-Encoders | 1 INTRODUCTION . A major challenge with applying variational auto-encoders ( VAEs ) to high-dimensional data is the typically slow training times . For example , training a state-of-the-art VAE ( Vahdat & Kautz , 2020 ; Child , 2020 ) on the 256× 256 FFHQ dataset ( Karras et al. , 2019 ) takes on the order of 1 GPU-yea... | In this paper, the authors focus on training conditional variational autoencoders. They propose an architecture and training objective which leverages pretraining an initial unconditional VAE. The approach effectively infers the latent variables of the original unconditional VAE given the new conditioning input. They d... | SP:61aced7e3c84fd2a9f9eef68bcc2f5cbf35fd304 |
Conditional Image Generation by Conditioning Variational Auto-Encoders | 1 INTRODUCTION . A major challenge with applying variational auto-encoders ( VAEs ) to high-dimensional data is the typically slow training times . For example , training a state-of-the-art VAE ( Vahdat & Kautz , 2020 ; Child , 2020 ) on the 256× 256 FFHQ dataset ( Karras et al. , 2019 ) takes on the order of 1 GPU-yea... | The paper proposes a method of leveraging state-of-the-art VAE decoders (foundational models) to create new conditional VAEs by training a new encoder for prior of the latent space. The authors demonstrate quite convincingly that the new conditional VAE can generate images with more variation than a state-of-the-art G... | SP:61aced7e3c84fd2a9f9eef68bcc2f5cbf35fd304 |
Conditional Image Generation by Conditioning Variational Auto-Encoders | 1 INTRODUCTION . A major challenge with applying variational auto-encoders ( VAEs ) to high-dimensional data is the typically slow training times . For example , training a state-of-the-art VAE ( Vahdat & Kautz , 2020 ; Child , 2020 ) on the 256× 256 FFHQ dataset ( Karras et al. , 2019 ) takes on the order of 1 GPU-yea... | This paper proposes a method, IPA, that converts an unconditional VAE to a conditional one by reusing the pretrained weights and training a partial encoder. Experiments on the image completion task show favorable results compared to GAN-based approach. The authors also explored an application in Bayesian optimal experi... | SP:61aced7e3c84fd2a9f9eef68bcc2f5cbf35fd304 |
Graph Condensation for Graph Neural Networks | Given the prevalence of large-scale graphs in real-world applications , the storage and time for training neural models have raised increasing concerns . To alleviate the concerns , we propose and study the problem of graph condensation for graph neural networks ( GNNs ) . Specifically , we aim to condense the large , ... | The paper proposes and addresses the problem of graph condensation. In a nutshell, provided a large graph G, the scope of the paper is to propose a solution able to generate a smaller synthetic graph G’, which effectively allows to train Graph Neural Networks (GNNs) able to achieve similar performance as if they were t... | SP:2a46ee245f3b6c6f881d99b2721733493ec1603d |
Graph Condensation for Graph Neural Networks | Given the prevalence of large-scale graphs in real-world applications , the storage and time for training neural models have raised increasing concerns . To alleviate the concerns , we propose and study the problem of graph condensation for graph neural networks ( GNNs ) . Specifically , we aim to condense the large , ... | To alleviate the storage and time consumption for training GNN models on large graphs, the paper studies graph condensation, which draws inspirations from data distillation/condensation. It first constructs a much smaller synthetic graph, and then train GNN models on this small graph. Empirical results show that, graph... | SP:2a46ee245f3b6c6f881d99b2721733493ec1603d |
Graph Condensation for Graph Neural Networks | Given the prevalence of large-scale graphs in real-world applications , the storage and time for training neural models have raised increasing concerns . To alleviate the concerns , we propose and study the problem of graph condensation for graph neural networks ( GNNs ) . Specifically , we aim to condense the large , ... | This paper introduces GCond, a graph condensation framework designed to compress graph datasets and reduce storage and time requirements while training GNNs on large-scale graphs. GCond makes use of gradient matching and graph sampling to reduce the graph size. Experiments demonstrate that GCond can maintain a high deg... | SP:2a46ee245f3b6c6f881d99b2721733493ec1603d |
On-Policy Model Errors in Reinforcement Learning | 1 INTRODUCTION . Model-free reinforcement learning ( RL ) has made great advancements in diverse domains such as single- and multi-agent game playing ( Mnih et al. , 2015 ; Silver et al. , 2016 ; Vinyals et al. , 2019 ) , robotics ( Kalashnikov et al. , 2018 ) and neural architecture search ( Zoph & Le , 2016 ) . All o... | The paper considers model-based reinforcement learning (MBRL) of the MBPO flavour, where model-free RL methods are accelerated using rollouts of a learned model. THis paper proposes to alleviate (off-policy) model bias per rollout using the (on-policy) data and calls this approach on-policy correction (OPC). With thi... | SP:f2753053e21f32ea10baf41029bb8ea6709b9a77 |
On-Policy Model Errors in Reinforcement Learning | 1 INTRODUCTION . Model-free reinforcement learning ( RL ) has made great advancements in diverse domains such as single- and multi-agent game playing ( Mnih et al. , 2015 ; Silver et al. , 2016 ; Vinyals et al. , 2019 ) , robotics ( Kalashnikov et al. , 2018 ) and neural architecture search ( Zoph & Le , 2016 ) . All o... | This paper addresses on-policy errors in model-based Reinforcement Learning. The authors decompose the policy improvement bound in terms that depend on the off-policy and on-policy model errors. They first present two major techniques, one purely based on the replay buffer, which asymptotically obtains zero on-policy e... | SP:f2753053e21f32ea10baf41029bb8ea6709b9a77 |
On-Policy Model Errors in Reinforcement Learning | 1 INTRODUCTION . Model-free reinforcement learning ( RL ) has made great advancements in diverse domains such as single- and multi-agent game playing ( Mnih et al. , 2015 ; Silver et al. , 2016 ; Vinyals et al. , 2019 ) , robotics ( Kalashnikov et al. , 2018 ) and neural architecture search ( Zoph & Le , 2016 ) . All o... | The paper proposes a way to combine observed data with a learned model to mitigate the cost of each choice separately - the observed data is over-fitted to a specific policy while the learned model has approximation errors. The authors mix the two options using probabilistic corrections based on the learned model appli... | SP:f2753053e21f32ea10baf41029bb8ea6709b9a77 |
Scalable Robust Federated Learning with Provable Security Guarantees | Federated averaging , the most popular aggregation approach in federated learning , is known to be vulnerable to failures and adversarial updates from clients that wish to disrupt training . While median aggregation remains one of the most popular alternatives to improve training robustness , relying on secure multi-pa... | The paper considers federated learning with two non-colluding honest-but-curious servers and honest-but-curious clients, where a subset of clients can be faulty. It proposes a protocol to securely compute approximate median. The main idea is to use secure median algorithm from [Tueno et al. 2019] along with a bucketing... | SP:b4a4baf0c0d01d14bf76b349461039244245ad07 |
Scalable Robust Federated Learning with Provable Security Guarantees | Federated averaging , the most popular aggregation approach in federated learning , is known to be vulnerable to failures and adversarial updates from clients that wish to disrupt training . While median aggregation remains one of the most popular alternatives to improve training robustness , relying on secure multi-pa... | This paper proposes a fast, secure, private and scalable approximate median aggregation approach for federated learning with two semi-honest non-colluding servers. The convergence analysis of the proposed approach for the IID case has been also provided under certain assumptions, which shows that the proposed approach ... | SP:b4a4baf0c0d01d14bf76b349461039244245ad07 |
Scalable Robust Federated Learning with Provable Security Guarantees | Federated averaging , the most popular aggregation approach in federated learning , is known to be vulnerable to failures and adversarial updates from clients that wish to disrupt training . While median aggregation remains one of the most popular alternatives to improve training robustness , relying on secure multi-pa... | This paper aims to build privacy-preserving and robust federated learning systems. The proposed method includes two key ideas. The first idea is to propose an approximate, crypto-friendly median aggregation rule, which aims to achieve robustness. The second idea is to use cryto methods to implement this aggregation rul... | SP:b4a4baf0c0d01d14bf76b349461039244245ad07 |
Directional Domain Generalization | 1 INTRODUCTION . Modern machine learning techniques have achieved unprecedented success over the past decades in various areas . However , one fundamental limitation of most existing techniques is that a model trained on one data set can not generalize well on another data set if it is sampled from a different distribu... | Summary: The paper proposes a new domain generalization paradigm called evolving domain generalization, where the source domains and the target domain are not random but have an evolving pattern. The paper provides theoretical results on the evolving domain generalization, which bounds the target domain test error wi... | SP:5f85d2331ff99d5bb36e2e0b6506861d88927423 |
Directional Domain Generalization | 1 INTRODUCTION . Modern machine learning techniques have achieved unprecedented success over the past decades in various areas . However , one fundamental limitation of most existing techniques is that a model trained on one data set can not generalize well on another data set if it is sampled from a different distribu... | This paper tackles the issue of training a model to generalize to data from unseen domains when there is a natural progression between domain at training and test time (eg. temporal progression). With the assumption that there is a fixed underlying transformation "g" which underlies the evolution of the data distributi... | SP:5f85d2331ff99d5bb36e2e0b6506861d88927423 |
Directional Domain Generalization | 1 INTRODUCTION . Modern machine learning techniques have achieved unprecedented success over the past decades in various areas . However , one fundamental limitation of most existing techniques is that a model trained on one data set can not generalize well on another data set if it is sampled from a different distribu... | In domain generalization (DG), we normally assume that the target domain is static rather than dynamic. However, we might meet many dynamic target domains in the real world (e.g., domains change over time), which will cause significant prediction errors on such dynamic target domains. To avoid this issue, this paper co... | SP:5f85d2331ff99d5bb36e2e0b6506861d88927423 |
Graph-less Neural Networks: Teaching Old MLPs New Tricks Via Distillation | 1 INTRODUCTION . Graph Neural Networks ( GNNs ) have recently become very popular for graph machine learning ( GML ) research and have shown great results on node classification tasks ( Kipf & Welling , 2016 ; Hamilton et al. , 2017 ; Veličković et al. , 2017 ) like product prediction on co-purchasing graphs and pape... | This paper applies KD in the context of the graph. It aims to distill the teacher output of a GNN model into a simple MLP model. Empirically, they show that this simple KD design is able to improve the student MLP model by a large margin and can match the results coming from a teacher GNN model. Besides, it shows empat... | SP:0ac9fb146e42a3eab0bc3f13a71c74445a49b730 |
Graph-less Neural Networks: Teaching Old MLPs New Tricks Via Distillation | 1 INTRODUCTION . Graph Neural Networks ( GNNs ) have recently become very popular for graph machine learning ( GML ) research and have shown great results on node classification tasks ( Kipf & Welling , 2016 ; Hamilton et al. , 2017 ; Veličković et al. , 2017 ) like product prediction on co-purchasing graphs and pape... | To overcome inference latency of GNN models, this paper proposes a GLNN framework that trains a student MLP with supervision from a teacher GNN model. GLNNs significantly improves performances of pure MLPs and are comparable with traditional GNNs in many cases. The paper explores different real-world settings including... | SP:0ac9fb146e42a3eab0bc3f13a71c74445a49b730 |
Graph-less Neural Networks: Teaching Old MLPs New Tricks Via Distillation | 1 INTRODUCTION . Graph Neural Networks ( GNNs ) have recently become very popular for graph machine learning ( GML ) research and have shown great results on node classification tasks ( Kipf & Welling , 2016 ; Hamilton et al. , 2017 ; Veličković et al. , 2017 ) like product prediction on co-purchasing graphs and pape... | The paper proposes a Knowledge Distillation (KD) based approach for producing MLPs able to achieve comparable performance to Graph Neural Networks (GNNs) on node-wise classification tasks. The MLPs receive as input only the features of each target node v and it is trained to imitate the behavior of a pre-trained GNN on... | SP:0ac9fb146e42a3eab0bc3f13a71c74445a49b730 |
Curriculum Discovery through an Encompassing Curriculum Learning Framework | We describe a curriculum learning framework capable of discovering optimal curricula in addition to performing standard curriculum learning . We show that this framework encompasses existing curriculum learning approaches such as difficulty-based data sub-sampling , data pruning , and loss re-weighting . We employ the ... | This paper proposes a new parameterized data partitioning and weighing scheme, that partitions data into three groups {easy, medium, hard} and determines a curriculum based on relative importance of different samples. They evaluate on three datasets (full and balanced versions) and show improvements over other CL appro... | SP:a983ecdce2b0b57e8717695b192478bfc6e4b0a2 |
Curriculum Discovery through an Encompassing Curriculum Learning Framework | We describe a curriculum learning framework capable of discovering optimal curricula in addition to performing standard curriculum learning . We show that this framework encompasses existing curriculum learning approaches such as difficulty-based data sub-sampling , data pruning , and loss re-weighting . We employ the ... | The paper proposes to learn training curricula by considering 3 sigmoids (representing "easy", "mid" and "hard" difficulty levels), that would define instances' weight as a function of time. The sigmoids' parameters are fitted to 3-bucket values of entropy of multi-annotator labels (that need to pre-exist), or of the i... | SP:a983ecdce2b0b57e8717695b192478bfc6e4b0a2 |
Curriculum Discovery through an Encompassing Curriculum Learning Framework | We describe a curriculum learning framework capable of discovering optimal curricula in addition to performing standard curriculum learning . We show that this framework encompasses existing curriculum learning approaches such as difficulty-based data sub-sampling , data pruning , and loss re-weighting . We employ the ... | The paper presents a curriculum learning approach applied to NLP models. Texts are separated into easy, medium and hard subsets based on a difficulty score that is given a standard NLP model. Hyperparameter tuning is used to determining evolving weights for each batch. The approach is evaluated on three benchmark datas... | SP:a983ecdce2b0b57e8717695b192478bfc6e4b0a2 |
Evaluation Metrics for Graph Generative Models: Problems, Pitfalls, and Practical Solutions | 1 INTRODUCTION . Graph generative models have become an active research branch , making it possible to generalise structural patterns inherent to graphs from certain domains—such as chemoinformatics—and actively synthesise new graphs ( Liao et al. , 2019 ) . Next to the development of improved models , their evaluation... | In this work, the authors aim to provide a principled way to evaluate and compare graph generative models. The authors initially list desirable criteria an evaluation metric should possess and subsequently discuss the usage of maximum mean discrepancy (MMD) for model comparison. Subsequently they highlight issues with ... | SP:a1a80e849c4ed5f0c631020deabe3907f01f8cbe |
Evaluation Metrics for Graph Generative Models: Problems, Pitfalls, and Practical Solutions | 1 INTRODUCTION . Graph generative models have become an active research branch , making it possible to generalise structural patterns inherent to graphs from certain domains—such as chemoinformatics—and actively synthesise new graphs ( Liao et al. , 2019 ) . Next to the development of improved models , their evaluation... | The paper criticizes Maximum Mean Discrepancy (MMD) an evaluation metric that has been used lately to evaluate Generative Graph Models (GGMs). The main problems of the MMD are: It does not capture difference upon perturbations; It does not have a scale; It requires the selection of a kernel and parameters that could le... | SP:a1a80e849c4ed5f0c631020deabe3907f01f8cbe |
Evaluation Metrics for Graph Generative Models: Problems, Pitfalls, and Practical Solutions | 1 INTRODUCTION . Graph generative models have become an active research branch , making it possible to generalise structural patterns inherent to graphs from certain domains—such as chemoinformatics—and actively synthesise new graphs ( Liao et al. , 2019 ) . Next to the development of improved models , their evaluation... | This paper describes desiderata (expressivity, robustness, and efficiency) for metrics for comparing graph generative models and details the various ways that recent work has used maximum mean discrepancy (MMD) to evaluate graph generative models. Several limitations of MMD are described, as well as the consequences o... | SP:a1a80e849c4ed5f0c631020deabe3907f01f8cbe |
A neural network framework for learning Green's function | 1 INTRODUCTION . Green ’ s function plays an important role in the theoretical research and engineering application of many important partial differential equations ( PDEs ) , such as the Poisson equation , Helmholtz equation , Maxwell equation , and so on . For one thing , Green ’ s function can help solve PDE problem... | This paper explores a deep learning scheme to solve partial differential equations on regular domains (mostly 2D). This paper focuses on the method of Greens function for solving a class of linear PDE’s. More specifically: the Greens function corresponding to a PDE is a function that depends on (1.) the differential op... | SP:78cf8aa7ee43ec8782e73370d3a9d4a972f53a25 |
A neural network framework for learning Green's function | 1 INTRODUCTION . Green ’ s function plays an important role in the theoretical research and engineering application of many important partial differential equations ( PDEs ) , such as the Poisson equation , Helmholtz equation , Maxwell equation , and so on . For one thing , Green ’ s function can help solve PDE problem... | The paper proposed a neural network based method for solving linear PDEs. The approach is based on approximating the solution using a combination of PINN and neural network based Green's function method. The architecture is then verified against the classical Poisson and Helmholtz equation. | SP:78cf8aa7ee43ec8782e73370d3a9d4a972f53a25 |
A neural network framework for learning Green's function | 1 INTRODUCTION . Green ’ s function plays an important role in the theoretical research and engineering application of many important partial differential equations ( PDEs ) , such as the Poisson equation , Helmholtz equation , Maxwell equation , and so on . For one thing , Green ’ s function can help solve PDE problem... | This paper proposes a neural network based method for computing Green's functions of Poisson and Helmholtz equations defined on various domains. The idea is to first subtract the target Green's function with some background green's function which is assumed to be known and easy to evaluation. Then the next step is appr... | SP:78cf8aa7ee43ec8782e73370d3a9d4a972f53a25 |
Gradient Broadcast Adaptation: Defending against the backdoor attack in pre-trained models | 1 INTRODUCTION . Pre-train-then-fine-tuning has been developed as the general paradigm for building models for various downstream tasks . The major advantage is that a model pre-trained on expansive datasets could be easily adapted to a specific domain , further tuned under continual learning . For example , Devlin et ... | This paper proposes a method to defend against NLP backdoor attacks. The authors propose to calculate the global direction of gradients of loss with respect to input word embeddings and update word embeddings using the global direction. By doing so, rare words can be updated to a "normal state" and are expected to be n... | SP:212e480fbeb43ffb00707628f48058a8d8517e96 |
Gradient Broadcast Adaptation: Defending against the backdoor attack in pre-trained models | 1 INTRODUCTION . Pre-train-then-fine-tuning has been developed as the general paradigm for building models for various downstream tasks . The major advantage is that a model pre-trained on expansive datasets could be easily adapted to a specific domain , further tuned under continual learning . For example , Devlin et ... | This paper proposes a defense against backdoor attack on pre-trained large language models. The proposed defense computes the average of the gradients per input sentence to contribute to updating all tokens in the sentence. The approach is empirically shown to outperform two baselines. | SP:212e480fbeb43ffb00707628f48058a8d8517e96 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.