paper_name stringlengths 11 170 | text stringlengths 8.07k 307k | summary stringlengths 152 6.16k | paper_id stringlengths 43 43 |
|---|---|---|---|
Modal Uncertainty Estimation via Discrete Latent Representations | 1 INTRODUCTION . Making predictions in the real world has to face with various uncertainties . One of the arguably most common uncertainties is due to partial or corrupted observations , as such it is often insufficient for making a unique and deterministic prediction . For example , when inspecting where a single CT s... | This paper introduces a novel conditional generative model for high dimensional data with multimodal output distributions. The proposed method, called modal uncertainty estimation (MUE), is a conditional VAE but with discrete latent representations. This discrete latent space allows the model to better handle multimoda... | SP:dd1ac7776d55534c5458d43d1fe39af30386343d |
Practical Evaluation of Out-of-Distribution Detection Methods for Image Classification | We reconsider the evaluation of OOD detection methods for image recognition . Although many studies have been conducted so far to build better OOD detection methods , most of them follow Hendrycks and Gimpel ’ s work for the method of experimental evaluation . While the unified evaluation method is necessary for a fair... | The paper empirically analyzes the evaluation framework of the current OOD detection systems for the image recognition task, specifically the evaluation described in [1] using Max-softmax and calibrated confidence. They motivate the paper by the necessity of having better evaluation for OOD detection to be reflective o... | SP:07471c50632db15eedbbc63f360a391140c1e094 |
Group Equivariant Generative Adversarial Networks | 1 INTRODUCTION . Generative visual modeling is an area of active research , time and again finding diverse and creative applications . A prevailing approach is the generative adversarial network ( GAN ) , wherein density estimation is implicitly approximated by a min-max game between two neural networks ( Goodfellow et... | The submission concerns an application of group convolutions (Cohen & Welling, 2016) to the image synthesis setting, where images are produced by the generator of a GAN. The two GAN components are augmented mainly by a straightforward replacement of "regular" convolutions by group convolutions, in addition to some othe... | SP:74ef7a70748db738244d9e402bbc4a9b43002896 |
Integrating linguistic knowledge into DNNs: Application to online grooming detection | 1 INTRODUCTION . Online grooming ( OG ) is a communicative process of entrapment in which an adult lures a minor into taking part in sexual activities online and , at times , offline ( Lorenzo-Dus et al. , 2016 ; Chiang & Grant , 2019 ) . Our aim is to detect instances of OG . This is achieved through binary classifica... | This work proposes the approach of integrating priors into a DNN in the form of Linguistic sub-models that capture characteristics of OG. The authors use the example of the PAN-12 dataset for sexual predators to use information about linguistics behaviour for the grooming phases. The work then goes to highlight the aug... | SP:f7611cb09eeb69912df93a040cf1ea98f59fd309 |
GraphCGAN: Convolutional Graph Neural Network with Generative Adversarial Networks | 1 INTRODUCTION . Graph-based semi-supervised learning ( SSL ) aims to classify nodes in graph , where only small amounts of nodes are labeled due to the expensive and time-consuming label collection process . To solve such task , various graph neural networks ( GNNs ) have been proposed using the idea of convolutional ... | The paper presents a method to combine graph convolutional neural networks (GCNs) with generative adversarial networks (GANs). The authors focus on the problem of semi-supervised learning on graphs and propose an end-to-end framework in which the generative model is followed by direct convolutions on the graph nodes. E... | SP:98871703cab28ed757a6ea54eea0407621624d62 |
Multi-Time Attention Networks for Irregularly Sampled Time Series | 1 INTRODUCTION . Irregularly sampled time series occur in application domains including healthcare , climate science , ecology , astronomy , biology and others . It is well understood that irregular sampling poses a significant challenge to machine learning models , which typically assume fully-observed , fixed-size fe... | This paper proposes a novel approach to learn an embedding of continuous time values and use an attention mechanism to produce a fixed-length representation of a time series containing a variable number of observations. In particular, it proposes an mTAN network to leverage the mTAN module in an encoder-decoder framewo... | SP:1996387f48b0d87ffe78a2c08a08faeb618c2213 |
Capturing Label Characteristics in VAEs | 1 INTRODUCTION . Learning the characteristic factors of perceptual observations has long been desired for effective machine intelligence ( Brooks , 1991 ; Bengio et al. , 2013 ; Hinton & Salakhutdinov , 2006 ; Tenenbaum , 1998 ) . In particular , the ability to learn meaningful factors—capturing human-understandable ch... | The paper proposes to re-think the fashion of using label information in the VAE framework. The authors propose to disentangle information about the label (or, more generally, the context) in a "hard-coded" manner, namely, by using a separate set of variables for the label (context). The paper is written in a lucid man... | SP:3b9ce25cba7d3b62e4927a76feccea0106d9b338 |
Regioned Episodic Reinforcement Learning | Goal-oriented reinforcement learning algorithms are often good at exploration , not exploitation , while episodic algorithms excel at exploitation , not exploration . As a result , neither of these approaches alone can lead to a sample-efficient algorithm in complex environments with high dimensional state space and de... | This paper presents a new algorithm called Regioned Episodic Reinforcement Learning (RERL), which combines ideas from episodic memory, with automatic sub-goal creation or “goal-oriented” RL. The method works by dividing the state space into regions, where a different goal identifies each region. Then, using an episodic... | SP:c80e745edb60717dcaa312fb3c01723bdb72f81d |
Graph Representation Learning for Multi-Task Settings: a Meta-Learning Approach | 1 INTRODUCTION Original Embeddings Transferred Embeddings NC GC- > NC LP- > NC0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 Ac cu ra cy - 13.21 % - 14.52 % Node Classification ( b ) GC NC- > GC LP- > GC0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 Ac cu ra cy - 21.29 % - 10.82 % Graph Classification ( c ) LP NC- > LP GC-... | The manuscript proposes SAME, a model based on GNN and meta-learning for learning multi-task node embeddings. Unlike multi-task learning setting, SAME aims at learning to quickly adapt to multiple tasks. Two model variants iSAME and eSAME are proposed base on different settings in inner/outer loop of parameter update. ... | SP:4b9cb72dcc70459c938b5ba8aaec2ea8fa253e1b |
Semi-Supervised Audio Representation Learning for Modeling Beehive Strengths | Honey bees are critical to our ecosystem and food security as a pollinator , contributing 35 % of our global agriculture yield ( Klein et al. , 2007 ) . In spite of their importance , beekeeping is exclusively dependent on human labor and experiencederived heuristics , while requiring frequent human checkups to ensure ... | The paper presents a semi-supervised model to predict the vitality of beehives. The inputs of the model are data from sensors (audio on one hand and environmental on the other hand such as temperature, humidity ...). The objective is to predict simultaneously 3 values of interest: the frames state of beehives, the pot... | SP:a7713950962f783173dbcf3ecd14289782380561 |
Learning Two-Time-Scale Representations For Large Scale Recommendations | 1 INTRODUCTION A hypothetical user ’ s interaction with recommendation systems gives us diminishing returns in terms of its information value in understanding the user . For an active user who has lots of historical interactions , she is typically well understood by the recommender , and each new interaction gives rela... | The paper considers the sequential recommendation problem. The proposed method essentially combines the following two ideas: (i) two-stage learning: using conventional CF to pretrain user/item embeddings, and feed them (fixed, unlearned) into the 2nd stage learning. (ii) two-time-scale: using 2 RNNs to model active use... | SP:47cbc46d73c5ad9d50744a7ff9fd6797eff273c4 |
You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling | 1 INTRODUCTION . The Transformer model ( Vaswani et al. , 2017 ) is incredibly effective across a diverse set of natural language processing ( NLP ) applications including machine translation ( Vaswani et al. , 2017 ) , language inference ( Devlin et al. , 2018 ) and paraphrasing ( Raffel et al. , 2019 ) . Transformer-... | The paper proposes to replace the weighted average of the values in standard self-attention with the average of values sampled in a way that the expectation is close to the result of self-attention. In particular, the authors associate with each query-key pair, a Bernoulli random variable with expected value close to t... | SP:2749a34e8528dfd4fcc733f9b9f175fcacbcb223 |
D2p-fed:Differentially Private Federated Learning with Efficient Communication | 1 INTRODUCTION . Federated learning ( FL ) is a popular machine learning paradigm that allows a central server to train models over decentralized data sources . In federated learning , each client performs training locally on their data source and only updates the model change to the server , which then updates the glo... | The paper proposes the discrete Gaussian based differentially private federated learning algorithm to achieve both differential privacy and communication efficiency in federated learning. In particular, it adds discrete Gaussian noise into client updates and uses secure aggregation to prevent the server from observing ... | SP:f17c1ecc9bb74a6c267c54a8863d0fcd336f4fdf |
Trusted Multi-View Classification | 1 INTRODUCTION . Multi-view data , typically associated with multiple modalities or multiple types of features , often exists in real-world scenarios . State-of-the-art multi-view learning methods achieve tremendous success across a wide range of real-world applications . However , this success typically relies on comp... | This paper proposes a reliable multi-view classification mechanism equipped with uncertainty, called Trusted Multi-View Classification. The goal is to dynamically assess the quality of different views for different samples to provide reliable uncertainty estimation. The idea is clear and well-motivated. The authors per... | SP:20a4cfac4c8e66208f4a4bd6b2ceeb3c8cabac3a |
Secure Byzantine-Robust Machine Learning | 1 INTRODUCTION . Recent years have witnessed fast growth of successful machine learning applications based on data collected from decentralized user devices . Unfortunately , however , currently most of the important machine learning models on a societal level do not have their utility , control , and privacy aligned w... | This work proposes a method to robustly (<.5 adversarial workers) aggregate model updates using two non-colluding servers. The proposed method scales well with the number of workers and is compatible with local DP and different robust aggregation protocols. Especially the scalability is a big improvement compared to pr... | SP:697a56b8f9152e50ee683f5a1b59bc272b01c4db |
Deep Reinforcement Learning For Wireless Scheduling with Multiclass Services | In this paper , we investigate the problem of scheduling and resource allocation over a time varying set of clients with heterogeneous demands . This problem appears when service providers need to serve traffic generated by users with different classes of requirements . We thus have to allocate bandwidth resources over... | Basically, it seems that the proposed method is interesting and meaningful. The scheduling problem in this paper is based on the analogy to a server having a water pitcher, and the deep reinforcement learning approach for the scheduling problem has been designed. However, the scheduling problem in wireless networks is ... | SP:9caede157f5546829e12c95bd290a760c1aa2dce |
Multi-modal Self-Supervision from Generalized Data Transformations | 1 INTRODUCTION . Recent works such as PIRL ( Misra & van der Maaten , 2020 ) , MoCo ( He et al. , 2019 ) and SimCLR ( Tian et al. , 2019 ) have shown that it is possible to pre-train state-of-the-art image representations without the use of any manually-provided labels . Furthermore , many of these approaches use varia... | The paper introduces a general framework dubbed Generalized Data Transformations (GDT) for self supervised learning. The framework is used to perform video-audio self supervised learning and analyze what kind of transformations the representations should be invariant to or on the contrary variant to thanks to a contras... | SP:858bb0278078b780b1fe163c7a7a084fd142f186 |
Overparameterisation and worst-case generalisation: friend or foe? | 1 INTRODUCTION . Overparameterised neural networks have demonstrated the remarkable ability to perfectly fit training samples , while still generalising to unseen test samples ( Zhang et al. , 2017 ; Neyshabur et al. , 2019 ; Nakkiran et al. , 2020 ) . However , several recent works have revealed that overparameterised... | The paper builds upon prior work that shows that overparameterized networks learned by ERM can have poor worst-case performance over pre-defined groups. Specifically, the paper demonstrates that this result is not necessarily due to overparameterized learning poor representations for rare subgroups, but rather mis-cali... | SP:3e9c01477200929c84f6725472107beab75a573e |
High-Capacity Expert Binary Networks | 1 INTRODUCTION . A promising , hardware-aware , direction for designing efficient deep learning models case is that of network binarization , in which filter and activation values are restricted to two states only : ±1 ( Rastegari et al. , 2016 ; Courbariaux et al. , 2016 ) . This comes with two important advantages : ... | This paper proposes some techniques to improve the accuracy of binary networks without adding much computational overhead. To improve model capacity, the author proposes mixture-of-experts convolution with a winner-takes-all gating mechanisms. To deal with the limited representation power of binary activations, the pap... | SP:a0c493b218741a8b49a12458bf78c88dc3aa596a |
Neural CDEs for Long Time Series via the Log-ODE Method | 1 INTRODUCTION . Neural controlled differential equations ( Neural CDEs ) ( Kidger et al. , 2020 ) are the continuous-time analogue to a recurrent neural network ( RNN ) , and provide a natural method for modelling temporal dynamics with neural networks . Neural CDEs are similar to neural ordinary differential equation... | .** The authors describe how to apply a log signature to temporal datasets. This operation reduces dimensionality along the time axis at the price of adding some dimensionality to the spatial dimension. Then they train a neural controlled differential equation (Neural CDE) on the transformed dataset and show that their... | SP:b8f49fdda704b0206febd3c09d1f475047919099 |
Counterfactual Fairness through Data Preprocessing | 1 INTRODUCTION . The rapid popularization of machine learning methods and the growing availability of personal data have enabled decision-makers from various fields such as graduate admission ( Waters & Miikkulainen , 2014 ) , hiring ( Ajunwa et al. , 2016 ) , credit scoring ( Thomas , 2009 ) , and criminal justice ( B... | The paper addresses the problem of preprocessing the data in a way that the predictions of a learning task will be counterfactually fair. The counterfactual fairness definition is borrowed from that of (Kusner et al., 2017). The authors propose ortogonaliza tion and marginal distribution mapping so as to achieve counte... | SP:e3e7028a84d8a272b7714e91bc08e67af40152c1 |
ZCal: Machine learning methods for calibrating radio interferometric data | 1 INTRODUCTION . Modern-day astronomy is at an unprecedented stage , with a deluge of data from different telescopes . In contrast to conventional methods , today astronomical discoveries are data-driven . The upcoming Square Kilometer Array ( SKA ) is expected to produce terabytes of data every hour ( The SKA telescop... | The paper presents a study of using machine learning methods to calibrate a radio telescope using information from sensor data on, e.g., atmospheric conditions. The authors consider tree- and neighbourhood-based methods for predicting amplitudes and phases for seven antennas. The results show that the methods perform q... | SP:4f59251101a0aad11518673e5571dceb4fcff65e |
Hierarchical Reinforcement Learning by Discovering Intrinsic Options | 1 INTRODUCTION . Imagine a wheeled robot learning to kick a soccer ball into a goal with sparse reward supervision . In order to succeed , it must discover how to first navigate in its environment , then touch the ball , and finally kick it into the goal , only receiving a positive reward at the end for completing the ... | The paper develops a hierarchical reinforcement learning algorithm and analyzes its behaviour in four robotic manipulation and navigation tasks. The approach is based on a two-level hierarchy, *scheduler* at the top and *worker* at the bottom. This is similar to other approaches in the literature and the algorithm uses... | SP:62750e67412021ffe9ef18e104833255aa6ed606 |
Towards Practical Second Order Optimization for Deep Learning | 1 Introduction . Second order methods are among the most powerful algorithms in mathematical optimization . Algorithms in this family often use a preconditioning matrix to transform the gradient before applying each step . Classically , the preconditioner is the matrix of second-order derivatives ( i.e. , the Hessian )... | This work addresses practical challenges in applying full matrix pre-conditioner methods (such as Shampoo) on problems involving large datasets and architectures trained using a distributed setup. In particular, this work presents a practical extension for the Shampoo algorithm by (1) using only a left or right precon... | SP:8bdbbc8a8bc54620675393fd822f56fb9ec53ffc |
Towards Understanding Fast Adversarial Training | 1 INTRODUCTION . Adversarial examples are carefully crafted versions of the original data that successfully mislead a classifier ( Szegedy et al. , 2013 ) , while realizing minimal change in appearance when viewed by most humans . Although deep neural networks have achieved impressive success on a variety of challengin... | The authors claimed in this paper that as the most empirically successful approach to defending adversarial examples, PGD-based adversarial training, is computationally inefficient. Fast adversarial training could mitigate this issue by training a model using FGSM attacks initialized with large randomized perturbations... | SP:f30f2cd322e3995e29563d5f6045e0f427c267af |
ALFA: Adversarial Feature Augmentation for Enhanced Image Recognition | 1 INTRODUCTION . Neural networks often fall vulnerable when presented adversarial examples injected with imperceptible perturbations , and suffer significant performance drop when facing such attacks ( Szegedy et al. , 2013 ; Goodfellow et al. , 2015b ) . Such susceptibility has motivated abundant studies on adversaria... | Overview of paper: this work tackles the task of adversarial augmentation for better generalization. Instead of augmentation the pixels space, which is expensive and potentially harder, they augment the intermediate feature representation. As the choice of the particular layer for application of the perturbations affec... | SP:b5daf21a7a1df819b39afd967085b64a55d14fb4 |
Using Deep Reinforcement Learning to Train and Evaluate Instructional Sequencing Policies for an Intelligent Tutoring System | 1 INTRODUCTION . An Intelligent Tutoring System ( ITS ) aims at teaching a set of skills to users by individualizing instructions . Giving instruction to users requires many sequential decisions , such as what to teach , what activities to present , what problems to include , and what help to give . Our aim is to take ... | The paper describes variants of an intelligent tutoring system (ITS) developed using a newer (but previously published) variant of Knowledge Tracing (HOT-DINA) for assessing student proficiency and an RL algorithm (PPO) for making decisions on items and content areas to try next. An empirical simulation calibrated to ... | SP:bba6a0856c8f3bb5a7ef8a768c38b999e6438df9 |
Nonconvex Continual Learning with Episodic Memory | 1 INTRODUCTION . Learning new tasks without forgetting previously learned tasks is a key aspect of artificial intelligence to be as versatile as humans . Unlike the conventional deep learning that observes tasks from an i.i.d . distribution , continual learning train sequentially a model on a non-stationary stream of d... | This paper analyses the convergence of episodic memory-based continual learning methods by looking at it as a nonconvex optimisation problem. They analyse the convergence rates for the case where all memory from past tasks is stored, and then consider the case where there is only a subset of past data, leading to overf... | SP:f4fc140928d2b4901d76664e62569545c70d8a5e |
Control-Aware Representations for Model-based Reinforcement Learning | 1 INTRODUCTION . Control of non-linear dynamical systems is a key problem in control theory . Many methods have been developed with different levels of success in different classes of such problems . The majority of these methods assume that a model of the system is known and its underlying state is low-dimensional and... | This paper aims to address an important question in reinforcement learning: policy learning from high-dimensional sensory observations. The authors propose an algorithm for Learning Controllable Embedding (LCE) based on policy iteration in the latent space. The authors provide a theorem to show how the policy performan... | SP:84f9003af6de793a1fd9c75c2cf9bb9dc495d56e |
Search Data Structure Learning | 1 INTRODUCTION . In many applications , the machines need to perform many searches in a gigantic database where the number of relevant documents is minuscule , e.g . ten in a billion . It is like searching for some needles in a haystack . In those cases , considering every document is extremely inefficient . For produc... | In this paper, the authors proposed Search Data Structure Learning (SDSL), which they claim to be a generalization of the standard Search Data Structure. They also present a new metric called Sequential Search Work Ratio (SSWR) to evaluate the quality and efficiency of the search. They introduced a new loss called F-be... | SP:892315ac5e3431d1be76ae8dbeb2121ea22b4ed8 |
Rapid Neural Architecture Search by Learning to Generate Graphs from Datasets | 1 INTRODUCTION . The rapid progress in the design of neural architectures has largely contributed to the success of deep learning on many applications ( Krizhevsky et al. , 2012 ; Cho et al. , 2014 ; He et al. , 2016 ; Szegedy et al . ; Vaswani et al. , 2017 ; Zhang et al. , 2018 ) . However , due to the vast search sp... | The authors address neural architecture search (NAS) scenarios. In particular, a framework, MetaD2A, is proposed, which yields a neural architecture for a new dataset. In a nutshell, the framework learns a "dataset-to-neural-network-architecture" transformation using a database of datasets and architectures. Each datas... | SP:d366dee57fb1f10beeef03e52f8a93ee6ff39f33 |
To Learn Effective Features: Understanding the Task-Specific Adaptation of MAML | 1 INTRODUCTION . Few-shot learning , aiming to learn from few labelled examples , is a great challenge for modern machine learning systems . Meta learning , an effective way for tracking this challenge , enables the model to learn general knowledge across a distribution of tasks . Various ideas of meta learning have be... | In this paper, the authors investigate the inner-loop optimization mechanism of meta-learning algorithms. The analysis shows the effectiveness of the multi-step adaptation and (1) the key of meta-learning is how to design a well-differentiated classifier. They then propose Random Decision Planes (RDP) and Meta Contrast... | SP:74dc640c4b7e724036bc4f772059fab7e9e33007 |
Interpretable Meta-Reinforcement Learning with Actor-Critic Method | 1 INTRODUCTION . Reinforcement learning problems have been studied for a long time and there are many impressive works that achieved human-level control in real world tasks ( Mnih et al. , 2013 ; Silver et al. , 2017 ; Vinyals et al. , 2019 ; Schrittwieser et al. , 2019 ) . These agents are trained separately on each t... | Authors introduce a new meta-RL algorithm based on SAC. It uses a context variable $c$ that they condition the Q-function on and the adaptation mechanism which is based on the values of the value function (ie. $\mathbb{E_a} Q(\dot, a)$) instead of the true returns. Authors claim their method reduces variance and bias o... | SP:5b537c8e2d4559f2980b079e46f23eeb8b6f30ad |
UPDeT: Universal Multi-agent RL via Policy Decoupling with Transformers | 1 INTRODUCTION . Reinforcement Learning ( RL ) provides a framework for decision-making problems in an interactive environment , with applications including robotics control ( Hester et al . ( 2010 ) ) , video gaming ( Mnih et al . ( 2015 ) ) , auto-driving ( Bojarski et al . ( 2016 ) ) , person search ( Chang et al . ... | 1. In this paper the authors proposed a transferrable framework for multi-agent RL, which enables the learned policies easily generalize to more challenging scenarios. This seems to be a good contribution to the community of multi-agent RL. It bears a potential to handle large-scale tasks with only limited training dat... | SP:e7976ca1bd206e20cbff3147a2b607ff6d658b2a |
Gradient Descent Ascent for Min-Max Problems on Riemannian Manifolds | 1 INTRODUCTION . In the paper , we study a class of useful non-convex minimax ( a.k.a . min-max ) problems on the Riemannian manifoldM with the definition as : min x∈M max y∈Y f ( x , y ) , ( 1 ) where the function f ( x , y ) is µ-strongly concave in y but possibly nonconvex in x . Here Y ⊆ Rd is a convex and closed s... | In this paper, the authors present and analyze a class of gradient-descent algorithms for solving min-max problems when the first (minimization) variable is constrained to live on a Riemaniann manifold. In the case when i) a retraction and an isometric transport are available on the manifold; and ii) the objective is s... | SP:63859002bed6542b5fe469aecb01e3070572885c |
Fooling a Complete Neural Network Verifier | 1 INTRODUCTION . In their seminal work , Szegedy et al . found that for a given neural network and input example one can always find a very small adversarial input perturbation that results in an incorrect output ( Szegedy et al. , 2014 ) . This striking discovery motivated a substantial amount of research . In this ar... | The authors show that certain complete neural network verifiers can be mislead by carefully crafted neural networks that exploit round-off errors, which when large magnitude values overwhelm low magnitude values. Such a construction can be obfuscated by taking advantage of the compounding effect when there are many lay... | SP:2308aac0572e5a7bca7552cfaf89617012da87b4 |
FAST GRAPH ATTENTION NETWORKS USING EFFECTIVE RESISTANCE BASED GRAPH SPARSIFICATION | 1 INTRODUCTION . Graphs are efficient representations of pairwise relations , with many real-world applications including product co-purchasing network ( ( McAuley et al. , 2015 ) ) , co-author network ( ( Hamilton et al. , 2017b ) ) , etc . Graph neural networks ( GNN ) have become popular as a tool for inference from... | This paper proposes a paradigm which speeds up the training/inference time of GATs while not compromising too much performance. The method adopts a layerwise sampling procedure. In particular. The authors propose to sample a sub-portion of edges for each layer based on their effective resistance. Such sampling keeps th... | SP:12c875bb1a25581a9f1e4eebfb1e1519d47ee6c7 |
Cut-and-Paste Neural Rendering | Cut-and-paste methods take an object from one image and insert it into another . Doing so often results in unrealistic looking images because the inserted object ’ s shading is inconsistent with the target scene ’ s shading . Existing reshading methods require a geometric and physical model of the inserted object , whi... | This paper proposes cut-and-paste neural rendering that allows to insert objects into a target scene in a plausible manner, i.e., in terms of shading plausibility. At the core of the approach is a deep image prior that allows to match the shading and albedo fields based on shading and albedo consistency losses. A norma... | SP:5207e34f58574e18c30192be6e2312863129fccd |
Optimizing Transformers with Approximate Computing for Faster, Smaller and more Accurate NLP Models | 1 INTRODUCTION . Transformer networks with hundreds of billions of parameters , such as T5 ( Raffel et al . ( 2019 ) ) , Megatron ( Shoeybi et al . ( 2019 ) ) , BERT ( Devlin et al . ( 2019 ) ) , GPT-2 ( Radford et al . ( 2019 ) ) and GPT-3 ( Brown et al . ( 2020 ) ) , have achieved state-of-the-art performance in seve... | This paper presents a method for improving a fine-turned Transformer in terms of a specific metric such as size, speed, or accuracy. The candidates of removed elements are considered hierarchically with some heuristics and are evaluated in terms of training and validation loss to determine whether they should actually ... | SP:0e68a02aff6bc3918d91083d6b48a3d625ebdc5d |
Multi-Level Generative Models for Partial Label Learning with Non-random Label Noise | 1 INTRODUCTION . Partial label ( PL ) learning is a weakly supervised learning problem with ambiguous labels ( Hüllermeier & Beringer , 2006 ; Zeng et al. , 2013 ) , where each training instance is assigned a set of candidate labels , among which only one is the true label . Since it is typically difficult and costly ... | This submission proposes a new method of learning from data with partially observed labels. In this problem, every instance has a label candidate set, which contains the true label. This submission introduces adversarial learning to improve the disambiguation of inexact labels. Particularly, there are two adversarial l... | SP:c0f80cb8844c1d9e6490f25a0b8feaa27557086c |
A Discriminative Gaussian Mixture Model with Sparsity | 1 INTRODUCTION . In probabilistic classification , a discriminative model is an approach that assigns a class label c to an input sample x by estimating the posterior probability P ( c | x ) . The posterior probability P ( c | x ) should correctly be modeled because it is not only related to classification accuracy , b... | The paper proposes a sparse classifier via discriminative GMM. This model is trained based on sparse Bayesian learning. The sparsity constraint removes redundant Gaussian components which results in reducing the number of parameters and improving the generalization. This framework can potentially be embedded into the... | SP:c9bda3b4e9859b304a8a3d1bc30ae0c8618a509d |
Meta-Learning of Structured Task Distributions in Humans and Machines | 1 INTRODUCTION . While machine learning has supported tremendous progress in artificial intelligence , a major weakness – especially in comparison to humans – has been its relative inability to learn structured representations , such as compositional grammar rules , causal graphs , discrete symbolic objects , etc . ( L... | This work is an exploration of model behaviour upon meta-learning tasks with compositional structure. The authors discover that, unlike humans, machine learning models do not readily pick up on the underlying compositional generative structure of a set of tasks, and hence cannot match the performance of humans. Convers... | SP:f0574c6588c9dc844b3e651e490092f058b7eb3c |
Transformers with Competitive Ensembles of Independent Mechanisms | 1 INTRODUCTION . A major theme throughout the history of deep learning has been the introduction of inductive biases in neural architectures , more recently with a focus on the ability to dynamically keep distinct types of information separated . While an MLP architecture has one large hidden representation at each lay... | This paper proposes an independent mechanism that divides hidden representations and parameters into multiple independent mechanisms. The authors claim that the mechanism benefits the computation of sparse tensors; it does learn better inductive biases than a sizeable monolithic model. This idea is particularly similar... | SP:21f106f8f8fa276557c2d46d25ab456370502f75 |
Environment Predictive Coding for Embodied Agents | 1 INTRODUCTION . In visual navigation tasks , an intelligent embodied agent must move around a 3D environment using its stream of egocentric observations to sense objects and obstacles , typically without the benefit of a pre-computed map . Significant recent progress on this problem can be attributed to the availabili... | The paper proposes a self-supervised approach for learning environment-level representations for embodied agents. The idea is that agents collect images and their corresponding poses during a walk-through phase. The images are clustered into multiple "zones". The zones are divided into seen and unseen zones. Using cont... | SP:eeab784f22aaf84838d021cc4c93a8707389d002 |
KETG: A Knowledge Enhanced Text Generation Framework | 1 INTRODUCTION . Recent pre-trained language models such as GPT-2 can capture clear semantic and syntactic features ( Radford , 2018 ) , performing well in machine translation and abstract generation tasks ( Li et al. , 2016 ; Wang et al. , 2016 ) . However , the application of language models in text generation still ... | This paper proposes to use a rhetoric knowledge graph for rhetorical text generation. One of its key contributions is to construct a rhetoric knowledge graph by leveraging SOTA NER and relation classification models. To generate a rhetorical text, the new method starts with sending a keyword to the knowledge graph to ... | SP:a0417f78d102a7c5ae83d98abe990dc03e3405ec |
Learning to Make Decisions via Submodular Regularization | 1 INTRODUCTION . In real-world automated decision making tasks we seek the optimal set of actions that jointly achieve the maximal utility . Many of such tasks — either deterministic/non-adaptive or stochastic/adaptive — can be viewed as combinatorial optimization problems over a large number of actions . As an example... | This paper combines combines submodular surrogates for sequential decision making with imitation learning. Specifically, it proposes to learn an acquisition function g by imitating an expert which is assumed to be following a greedy policy wrt a general submodular surrogate f. This is accomplished by regularizing g to ... | SP:364842bf9376198df47a7323185d72cc73380d4d |
CLOPS: Continual Learning of Physiological Signals | 1 INTRODUCTION . Many deep learning algorithms operate under the assumption that instances are independent and identically-distributed ( i.i.d. ) . The violation of this assumption can be detrimental to the training behaviour and performance of an algorithm . The assumption of independence can be violated , for example... | The authors propose a learning methodology designed to offset detriments to algorithm performance that arise when instances are not i.i.d (independent and identically distributed), focusing on cases in continual learning (CL) given physiological signals. They designed a replay-based learning method that handles an inst... | SP:61d83ed48f892bcb7d0488c9b918132b2623eea1 |
SALD: Sign Agnostic Learning with Derivatives | 1 INTRODUCTION . Recently , neural networks ( NN ) have been used for representing and reconstructing 3D surfaces . Current NN-based 3D learning approaches differ in two aspects : the choice of surface representation , and the supervision method . Common representations of surfaces include using NN as parametric charts... | This paper presents SALD, a new type of implicit shape representation that, in addition to predicting the signed distance function, aligns the gradients of the distance function with that of the neural distance field. The resulting algorithm, for example, has improved approximation power and better preserves the sharp ... | SP:ceacad438130adfb746240e36dd32d14794b4291 |
Sequential Density Ratio Estimation for Simultaneous Optimization of Speed and Accuracy | 1 INTRODUCTION . The sequential probability ratio test , or SPRT , was originally invented by Abraham Wald , and an equivalent approach was also independently developed and used by Alan Turing in the 1940s ( Good , 1979 ; Simpson , 2010 ; Wald , 1945 ) . SPRT calculates the log-likelihood ratio ( LLR ) of two competing... | This work introduces SPRT-TANDEM an algorithm to train a sequential probability ratio test (SPRT) as a neural network. This network is then used to discriminate between two hypotheses as fast as possible (seeing the smallest number of observations in a sequence) while maintaining a certain level of accuracy. The main c... | SP:3120ae529b5b2964470ad055d1f13989f192c961 |
NOVAS: Non-convex Optimization via Adaptive Stochastic Search for End-to-end Learning and Control | 1 INTRODUCTION . Deep learning has experienced a drastic increase in the diversity of neural network architectures , both in terms of proposed structure , as well as in the repertoire of operations that define the interdependencies of its elements . With respect to the latter , a significant amount of attention has bee... | This paper aims to present a method that allows efficient learning in neural networks architecture that present optimization blocks. These blocks have the form of x_{i+1} = \arg \min_x F(x, x_i, \theta), and can be thought of as a neural network layer. The addition of this block results in a complex optimization proble... | SP:b64d32119a136b5957e85e52c3ab32c27d3c2f3f |
Neural Point Process for Forecasting Spatiotemporal Events | 1 . Introduction . Accurate modeling of spatiotemporal event dynamics is fundamentally important for disaster response ( Veen and Schoenberg , 2008 ) , logistic optimization ( Safikhani et al. , 2018 ) and social media analysis ( Liang et al. , 2019 ) . Compared to other sequence data such as texts or time series , spa... | This work studies the DNN-based spatiotemporal point process model. It points out the drawback of most existing DNN-based point process models: incapability to incorporate the spatio information. Although in statistics, the spatiotemporal point process is capable of capturing events in continuous space and time, such m... | SP:401998f890d05e3c22e89754ed6b64403e1a6ead |
Discovering Parametric Activation Functions | 1 INTRODUCTION . The rectified linear unit ( ReLU ( x ) = max { x , 0 } ) is the most commonly used activation function in modern deep learning architectures ( Nair & Hinton , 2010 ) . When introduced , it offered substantial improvements over the previously popular tanh and sigmoid activation functions . Because ReLU ... | The authors propose to search for activation functions with regularized evolution, an evolutionary algorithm proposed by Real et al. Various mutations are proposed that allow to investigate a larger search space than prior work. In particular, a mutation is added which adds trainable parameters to the activation functi... | SP:510133bddf8cd65c97348e4a8161009fc1d791e0 |
Efficient Competitive Self-Play Policy Optimization | 1 INTRODUCTION . Reinforcement learning ( RL ) from self-play has drawn tremendous attention over the past few years . Empirical successes have been observed in several challenging tasks , including Go ( Silver et al. , 2016 ; 2017 ; 2018 ) , simulated hide-and-seek ( Baker et al. , 2020 ) , simulated sumo wrestling ( ... | The paper “Efficient Competitive Self-Play Policy Optimization” introduces a new self-play scheme for solving zero-sum two-player games. It is suggested to train a population of N agents in parallel, where each agent is matched against the comparatively strongest opponent in the next round of training. As baselines, th... | SP:d23a1168bdf9f77e67f24b5062525cefd213a43e |
Defending against black-box adversarial attacks with gradient-free trained sign activation neural networks | While machine learning models today can achieve high accuracies on classification tasks , they can be deceived by minor imperceptible distortions to the data . These are known as adversarial attacks and can be lethal in the black-box setting which does not require knowledge of the target model type or its parameters . ... | The paper proposes an architecture (ensemble of networks) aiming at being robust against black-box attacks, based on the idea that crafting an adversarial example able to fool enough individual networks such that the majority vote changes is a more difficult task. The paper presents ways of training such ensembles and... | SP:be01b10daaf670341722afb0c2d8570156ba7b53 |
Flow Neural Network for Traffic Flow Modelling in IP Networks | 1 INTRODUCTION . Deep Learning ( DL ) has gained substantial popularity in light of its applicability to real-world tasks across computer vision , natural language processing ( Goodfellow et al. , 2016 ) , protein structure prediction ( Senior et al. , 2020 ) and challenging games such as Go ( Silver et al. , 2017 ) . ... | The goal of this study is the 1-step prediction of flow rate in flow networks. They first define a “spatial-temporal induction effect (STI)” and claim it to be the universal property of flow networks. Their main contribution is their proposed “flow neural network” which is based on the STI effect and a combination of G... | SP:c582c4634f7e343732bab5e9cc7024efbf6d88d0 |
ARMCMC: Online Model Parameters full probability Estimation in Bayesian Paradigm | 1 INTRODUCTION . Bayesian methods are powerful tools to not only obtain a numerical estimate of a parameter but also to give a measure of confidence ( Kuśmierczyk et al. , 2019 ; Bishop , 2006 ; Joho et al. , 2013 ) . In particular , Bayesian inferences calculate the probability distribution of parameters rather than ... | The paper introduces a new Markov chain Monte-Carlo (MCMC) algorithm to obtain and track the posterior distribution over unknown parameters in a non-linear system. Despite its simple elegance, i.e., the introduction of a data-driven _temporal forgetting factor_ into the usual Metropolis-Hastings algorithm, the approach... | SP:d9610d460905f545ccdd7524b9efc049ecdc0f25 |
MiCE: Mixture of Contrastive Experts for Unsupervised Image Clustering | 1 INTRODUCTION . Unsupervised clustering is a fundamental task that aims to partition data into distinct groups of similar ones without explicit human labels . Deep clustering methods ( Xie et al. , 2016 ; Wu et al. , 2019 ) exploit the representations learned by neural networks and have made large progress on high-dim... | Authors present “mixture of experts” type of method to solve a clustering with unsupervised learning problem. Method is called as Mixture of Contrastive Experts (MiCE) which uses contrastive learning as a base module and combines it with latent mixture models. Authors develop a scalable algorithm for MiCE and empirical... | SP:c3995e4d2f6dcf282fa8312606a43471c82f629f |
PDE-regularized Neural Networks for Image Classification | 1 INTRODUCTION . It had been discovered that interpreting neural networks as differential equations is possible by several independent research groups ( Weinan , 2017 ; Ruthotto & Haber , 2019 ; Lu et al. , 2018 ; Ciccone et al. , 2018 ; Chen et al. , 2018 ; Gholami et al. , 2019 ) . Among them , the seminal neural ord... | The paper proposed the method of neural PDE as an improvement of neural ODE. In specific, neural PDE considers both the layer and the hidden dimension as continuous variables of the PDE. The new part of neural PDE compared to neural ODE is essentially solving PDE inverse problems (learning PDE from data) in the computa... | SP:6357b56f8b11f6eb3ccd152460b4aff5ab9ff6d4 |
Combining Ensembles and Data Augmentation Can Harm Your Calibration | 1 INTRODUCTION . Many success stories in deep learning ( Krizhevsky et al. , 2012 ; Sutskever et al. , 2014 ) are in restricted settings where predictions are only made for inputs similar to the training distribution . In real-world scenarios , neural networks can face truly novel data points during inference , and in ... | This work analyses the interaction between data-augmentation strategies such as MixUp and model ensembles with regards to calibration performance. The authors note how strategies such as mixup and label smoothing, which reduce a single model's over-confidence, lead to degradation in calibration performance when such mo... | SP:8079cb72ef8db9b5ab9275770ade605746840832 |
Do Deeper Convolutional Networks Perform Better? | 1 INTRODUCTION . Traditional statistical learning theory argues that over-parameterized models will overfit training data and thus generalize poorly to unseen data ( Hastie et al. , 2001 ) . This is explained through the bias-variance tradeoff ; as model complexity increases , so will variance , and thus more complex m... | This paper mainly answers a fundamental question: what is the role of depth in convolutional networks? Specifically, the authors present an empirical analysis of the impact of the depth on the generalization in CNNs. Experiments on CIFAR10 and ImageNet32 demonstrate that the test performance beyond a critical depth. My... | SP:4f6e5411e0d5a017100c74a3842fed4ff323d883 |
Machine Learning Algorithms for Data Labeling: An Empirical Evaluation | 1 INTRODUCTION . Supervised learning is the most commonly used machine learning paradigms . There are problems with supervised learning and machine learning in general . The first problem is that machine learning requires huge amounts of data . Secondly , supervised learning needs labels in the data . In a case study p... | This paper aims to evaluate the performance of seven automated labeling algorithms in terms of accuracy. The authors conducted a set of experiments on six datasets from different domains under two typical settings where 10% and 50%of labels in the datasets are available. Experimental results show that the algorithms la... | SP:975e5116fe8c4160a6e0c875044d95ee569208a9 |
IALE: Imitating Active Learner Ensembles | 1 INTRODUCTION . The high performance of deep learning on various tasks from computer vision ( Voulodimos et al. , 2018 ) to natural language processing ( NLP ) ( Barrault et al. , 2019 ) also comes with disadvantages . One of their main drawbacks is the large amount of labeled training data they require . Obtaining su... | In this work, an imitation learning (AL) approach is proposed to imitate multiple active learning algorithms, in order to take their advantages to learn a better active learning algorithm. The main idea is to treat the active learning algorithms as experts and utilize the DAGGER algorithm for imitation learning. The pr... | SP:4063187f00775058a7d47814b0062648d88f0b8d |
Neural networks with late-phase weights | 1 INTRODUCTION . Neural networks trained with SGD generalize remarkably well on a wide range of problems . A classic technique to further improve generalization is to ensemble many such models ( Lakshminarayanan et al. , 2017 ) . At test time , the predictions made by each model are combined , usually through a simple ... | This work suggests a variant of ensembling that is more compute-efficient. Specifically, it involves forking an ensemble only in the late stage of training, and forming this ensemble via a "low-dimentional" family. That is, instead of maintaining independent networks, maintain only "low-rank"-style perturbations of the... | SP:651166f4bdf2eb56689f790d3c697a43be974521 |
Multi-Agent Collaboration via Reward Attribution Decomposition | 1 INTRODUCTION . In recent years , multi-agent deep reinforcement learning ( MARL ) has drawn increasing interest from the research community . MARL algorithms have shown super-human level performance in various games like Dota 2 ( Berner et al. , 2019 ) , Quake 3 Arena ( Jaderberg et al. , 2019 ) , and StarCraft ( Sam... | To address the ad hoc team play, the authors propose a residual term of Q function, which additionally considers the states of nearby agents. A novel MARA loss is introduced to the residual term as a regularization to achieve the reward assignment implicitly. The proposed CollaQ could be easily built on QMIX and traine... | SP:6adf73371c97da34bca974dbffb5b7dd211b9e44 |
Statistical inference for individual fairness | 1 INTRODUCTION . The problem of bias in machine learning systems is at the forefront of contemporary ML research . Numerous media outlets have scrutinized machine learning systems deployed in practice for violations of basic societal equality principles ( Angwin et al. , 2016 ; Dastin , 2018 ; Vigdor , 2019 ) . In resp... | The paper introduces a framework to statistically test whether a given model is individually fair or not. In particular, given a model, a distance metric over individuals, and a data point z, the authors propose an algorithm that finds a new data point z' such that z' is similar to z but their corresponding losses are ... | SP:85843d0456fb7791c3edfc1f81dec00be5abc41f |
Revisiting the Train Loss: an Efficient Performance Estimator for Neural Architecture Search | 1 INTRODUCTION . Reliably estimating the generalisation performance of a proposed architecture is crucial to the success of Neural Architecture Search ( NAS ) but has always been a major bottleneck in NAS algorithms ( Elsken et al. , 2018 ) . The traditional approach of training each architecture for a large number of ... | This paper proposes a simple model-free method to estimate the generalization performance of deep neural architectures based on their early training losses. The proposed method uses the sum of training losses during training to estimate the performance and is motivated by recent empirical and theoretical results. The e... | SP:e7bd23e8d01a469909890d06581882da634a3e0f |
Augmenting Physical Models with Deep Networks for Complex Dynamics Forecasting | 1 INTRODUCTION . Modeling and forecasting complex dynamical systems is a major challenge in domains such as environment and climate ( Rolnick et al. , 2019 ) , health science ( Choi et al. , 2016 ) , and in many industrial applications ( Toubeau et al. , 2018 ) . Model Based ( MB ) approaches typically rely on partial ... | This paper outlines a method for forecasting and parameter estimation when you have a partial physics model (possibly with unknown parameters) and time series data. This is a hybrid approach where the data-driven (deep learning) approach only learns the parts not accounted for by the physical model. A key feature is be... | SP:ddf5fcf80d3a1d2c18cf4432d29c0eda32dbbef3 |
Winning the L2RPN Challenge: Power Grid Management via Semi-Markov Afterstate Actor-Critic | 1 INTRODUCTION . The power grid , an interconnected network for delivering electricity from producers to consumers , has become an essential component of modern society . For a safe and reliable transmission of electricity , it is constantly monitored and managed by human experts in the control room . Therefore , there... | This paper proposes an effective method for managing power grid topology to increase efficiency. They use Transformer attention over a Graph Neural Network as the basic architecture, then propose a hierarchical technique in which the upper level learns to output goal network topologies, which are then implemented by a ... | SP:839dcc82412b1e77aa5e3f267ef421dae1bc0cfc |
A spherical analysis of Adam with Batch Normalization | A SPHERICAL ANALYSIS OF ADAM WITH BATCH NORMALIZATION . Anonymous authors Paper under double-blind review Batch Normalization ( BN ) is a prominent deep learning technique . In spite of its apparent simplicity , its implications over optimization are yet to be fully understood . While previous studies mostly focus on t... | This work studies optimization dynamics for neural network models that are scaling invariant with respect to parameters. A general formulation of optimization algorithms is considered, covering many widely used algorithms like SGD and Adam. The projected dynamics (to the unit sphere) is studied, and the effective learn... | SP:2d1b5b2da4802fb7f229112fb841bc194ba47204 |
Sliced Kernelized Stein Discrepancy | 1 INTRODUCTION . Discrepancy measures for quantifying differences between two probability distributions play key roles in statistics and machine learning . Among many existing discrepancy measures , Stein discrepancy ( SD ) is unique in that it only requires samples from one distribution and the score function ( i.e . ... | This paper tries to solve the curse-of-dimensionality problem of KSD and corresponding mode-collapse problem of SVGD by projecting both the input and output of test function onto 1D slices. By doing so, the paper proposes the new discrepancies called SSD and maxSKSD, and a new variant of SVGD called S-SVGD. Experiments... | SP:23124b43054b8f3b0cf5860a1fa0728f7edf8e63 |
Synthesising Realistic Calcium Traces of Neuronal Populations Using GAN | 1 INTRODUCTION . The ability to record accurate neuronal activities from behaving animals is essential for the study of information processing in the brain . Electrophysiological recording , which measures the rate of change in voltage by microelectrodes inserted in the cell membrane of a neuron , has high temporal res... | The paper proposes to use a GAN framework to generate the realistic neuronal calcium signals, enabling to scale-up the neuronal population activity data. The solution is based on WAVEGAN architecture with Wasserstein distance to train on calcium fluorescent signals. The experiments are performed in comparison to artifi... | SP:3f164a85f782ec9beeb00b19638f98d0cb6a6265 |
Episodic Memory for Learning Subjective-Timescale Models | 1 INTRODUCTION . An agent endowed with a model of its environment has the ability to predict the consequences of its actions and perform planning into the future before deciding on its next move . Models can allow agents to simulate the possible action-conditioned futures from their current state , even if the state wa... | Most model-based RL algorithms learn dynamics models that predicts the next timestep. However, because of model-bias, frequency of timesteps, and objective timescales, the dynamics models can accumulate errors and limited by timescales. The authors propose subjective-timescale model (STM) that instead of predicting the... | SP:3e360ec6c3c576d09fc38169789f9df9dada9bea |
Efficient randomized smoothing by denoising with learned score function | 1 INTRODUCTION . The deep image classifiers are susceptible to deliberate noises as known as adversarial attacks ( Szegedy et al. , 2013 ; Goodfellow et al. , 2014 ; Carlini & Wagner , 2017 ) . Even though many works proposed heuristics that can annul or mitigate adversarial attacks , most of them were broken by strong... | This paper presents a denoising-based method for randomized smoothing that converts a base classifier into a smoothed one with p-robustness to adversarial examples. It considers a practical setting where the retraining/finetuning of the base classifier is largely inapplicable (e.g. the commercial classification service... | SP:074d113e06bfa79b8a5314560ef0b6669278abd5 |
Random Feature Attention | 1 INTRODUCTION . Transformer architectures ( Vaswani et al. , 2017 ) have achieved tremendous success on a variety of sequence modeling tasks ( Ott et al. , 2018 ; Radford et al. , 2018 ; Parmar et al. , 2018 ; Devlin et al. , 2019 ; Parisotto et al. , 2020 , inter alia ) . Under the hood , the key component is attenti... | The paper presents a linear time and space attention mechanism based on random features to approximate the softmax. The paper is clearly written and easy to follow. The results are convincing: not chasing SOTA, but comparing to sensible baselines, namely [Baevski & Auli 2019] for language modeling on Wikitext-103, and ... | SP:e79752ff486049e2e9ec9f588aa918ca2399a5e2 |
Directed Acyclic Graph Neural Networks | 1 INTRODUCTION . Graph-structured data is ubiquitous across various disciplines ( Gilmer et al. , 2017 ; Zitnik et al. , 2018 ; Sanchez-Gonzalez et al. , 2020 ) . Graph neural networks ( GNNs ) use both the graph structure and node features to produce a vectorial representation , which can be used for classification , ... | This paper introduces a model, Directed Acyclic Graph Neural Network (DAGNN), which processes information according to the flow defined by partial order. DAGNN can be regarded as a special case of previous GNN models, but specific to directed acyclic graph structures. The authors prove that the model satisfies the prop... | SP:3d2faa84203e50f95080e9d2de9660affe58e157 |
Synthesizer: Rethinking Self-Attention for Transformer Models | The dot product self-attention is known to be central and indispensable to stateof-the-art Transformer models . But is it really required ? This paper investigates the true importance and contribution of the dot product-based self-attention mechanism on the performance of Transformer models . Via extensive experiments ... | This paper challenges the common belief that self-attention with dot product is necessary to train good NLP models. Several variants of the Synthesizer model is proposed. The effectiveness of Synthesizer is surprisingly good, although not beating the dot-product attention. The authors further showed that mixing synthes... | SP:081c48c667eef561333c5b0d739e9dbebefa0f34 |
Learning Graph Normalization for Graph Neural Networks | 1 INTRODUCTION . Graph Neural Networks ( GNNs ) have shown great popularity due to their efficiency in learning on graphs for various application areas , such as natural language processing ( Yao et al. , 2019 ; Zhang et al. , 2018 ) , computer vision ( Li et al. , 2020 ; Cheng et al. , 2020 ) , point cloud ( Shi & Raj... | This paper proposes and evaluates different normalization techniques for graph neural networks. Also, the authors argue that the best normalization technique is task dependent, so they propose to use a weighted average of different normalizations that is learned during training, called AGN. In the paper they propose 4 ... | SP:e8de5995140c90ed95c915f5724c0a910a99cfb9 |
Iterative Graph Self-Distillation | 1 INTRODUCTION . Graphs are ubiquitous representations encoding relational structures across various domains . Learning low-dimensional vector representations of graphs is critical in various domains ranging from social science ( Newman & Girvan , 2004 ) to bioinformatics ( Duvenaud et al. , 2015 ; Zhou et al. , 2020 )... | Learning graph-level representations with only labels has been explored by many works. However, it's not easy to annotate every graph. This paper applies the ideas from semi-supervised classification task to improve the representation quality learned by graph neural network. Specifically the proposed solution combines ... | SP:3647115d0449f579f5ad7305103ecb553046d613 |
How Important is Importance Sampling for Deep Budgeted Training? | 1 INTRODUCTION . The availability of vast amounts of labeled data is crucial in training deep neural networks ( DNNs ) ( Mahajan et al. , 2018 ; Xie et al. , 2020 ) . Despite prompting considerable advances in many computer vision tasks ( Yao et al. , 2018 ; Sun et al. , 2019a ) , this dependence poses two challenges :... | This paper investigates the use of importance sampling in budgeted training. Four importance sampling techniques from prior works are applied within the context of fixed training budgets, and compared under different conditions of training set selection, learning rate schedule and data augmentations. Each aims to sam... | SP:1fc676213cbcfd690a3aea055066a3004f974325 |
VideoFlow: A Framework for Building Visual Analysis Pipelines | 1 INTRODUCTION . The success of computer vision techniques is spawning intelligent visual analysis systems in real applications . Rather than serving individual models , these systems are often powered by a workflow of image/video decoding , several serial or parallel algorithm processing stages , as well as sinking an... | The paper presents a tutorial to a video analysis platform software, i.e., VideoFlow, which represents a video analysis task as a computation graph, provides common functions like video decoding and database storage, integrates deep learning frameworks, e.g. Caffe/Pytorch/MXNet as built-in inference engines, and suppor... | SP:b31d37adc24ddff6ef32dc607fe3c8c29341a81d |
SOAR: Second-Order Adversarial Regularization | 1 INTRODUCTION . Adversarial training ( Szegedy et al. , 2013 ) is the standard approach for improving the robustness of deep neural networks ( DNN ) , or any other model , against adversarial examples . It is a data augmentation method that adds adversarial examples to the training set and updates the network with new... | The paper proposed a regularizer loss as an alternative to adversarial training to improve the robustness of neural networks against adversarial attacks. The new regularizer is derived from a second-order Tyler series expansion of the loss function in the model robustness optimization problem. Clear mathematical deriva... | SP:f67271e00a669e2b64580762c04eb7b88965061d |
How to Design Sample and Computationally Efficient VQA Models | 1 INTRODUCTION . Many real-world complex tasks require both perception and reasoning ( or System I and System II intelligence ( Sutton & Barto , 2018 ) ) , such as VQA . What is the best way to integrate perception and reasoning components in a single model ? Furthermore , how would such an integration lead to accurate... | The paper proposes a neuro-symbolic model for sample-efficient VQA, which turns each question into a probabilistic program which is then softly executed. The problem explored in the paper and its background and context presented clearly and it does a good job in motivating its importance and trade-offs between possible... | SP:885d09e9fb6fa10be309dcbfe259ecf35ccabb82 |
A Chaos Theory Approach to Understand Neural Network Optimization | Despite the complicated structure of modern deep neural network architectures , they are still optimized with algorithms based on Stochastic Gradient Descent ( SGD ) . However , the reason behind the effectiveness of SGD is not well understood , making its study an active research area . In this paper , we formulate de... | The authors use an insight from chaos theory to derive an efficient method of estimating the largest and smallest eigenvalues of the loss Hessian wrt the weights. To do that, they use nearby weight space positions, optimize for a bit (either gradient climbing or descending), check how quickly the points are departing f... | SP:fbb217eb911fc3b0d40b941281d08d0a399a459a |
Deep Learning meets Projective Clustering | 1 INTRODUCTION AND MOTIVATION . Deep Learning revolutionized Machine Learning by improving the accuracy by dozens of percents for fundamental tasks in Natural Language Processing ( NLP ) through learning representations of a natural language via a deep neural network ( Mikolov et al. , 2013 ; Radford et al. , 2018 ; Le... | This work proposes a new approach, based on projective clustering, for compressing the embedding layers of DNNs for natural language modeling tasks. The authors show that the trade-off between compression and model accuracy can be improved by considering a set of k subspaces rather than just a single subspace. Methods ... | SP:d8f80f84b089766124693485390dbfce0c94527c |
Box-To-Box Transformation for Modeling Joint Hierarchies | 1 INTRODUCTION . Representation learning for hierarchical relations is crucial in natural language processing because of the hierarchical nature of common knowledge , for example , < Bird ISA Animal > ( Athiwaratkun & Wilson , 2018 ; Vendrov et al. , 2016 ; Vilnis et al. , 2018 ; Nickel & Kiela , 2017 ) . The ISA relat... | The paper focuses on modeling multiple hierarchical relations on a heterogenous graph. The task “modeling joint hierarchies” is essentially trying to infer whether a given pair of entities has a hierarchical connection especially when there exists multiple hierarchical relations (2 in the paper), and missing links. Th... | SP:2f3bb20ca38e10fde160e4961d6b1796cadd465f |
Spatio-Temporal Graph Scattering Transform | 1 INTRODUCTION . Processing and learning from spatio-temporal data have received increasing attention recently . Examples include : i ) skeleton-based human action recognition based on a sequence of human poses ( Liu et al . ( 2019 ) ) , which is critical to human behavior understanding ( Borges et al . ( 2013 ) ) , an... | The authors propose wavelets for both separable and joint spatio-temporal graphs. And then the authors design a spatio-temporal graph scattering transform (ST-GST), which is a non-trainable counterpart of spatio-temporal graph convolutional networks and a nonlinear version of spatiotemporal graph wavelets. Finally, the... | SP:03895ea221824f6e57ea88ec7332efbbec207c7d |
Explicit homography estimation improves contrastive self-supervised learning | 1 INTRODUCTION . There is an ever-increasing pool of data , particularly unstructured data such as images , text , video , and audio . The vast majority of this data is unlabelled . The process of labelling is time-consuming , labour-intensive , and expensive . Such an environment makes algorithms that can leverage ful... | The authors propose a module that regresses the parameters of an affine transformation or homography as an additional objective in the contrastive self-supervised learning framework. The authors argue that the geometric information encoded in the proposed module can supplement the signal provided by a contrastive loss,... | SP:b6083b2193bf2ab0df08746ef2ec9e51b513525f |
Variational inference for diffusion modulated Cox processes | 1 INTRODUCTION . Cox processes ( Cox , 1955 ; Cox & Isham , 1980 ) , also known as doubly-stochastic Poisson processes , are a class of stochastic point processes wherein the point intensity is itself stochastic and , conditional on a realization of the intensity process , the number of points in any subset of space is... | The paper under review proposes a variational inference procedure for a specific class of Cox processes whose intensity is derived from a stochastic differential equation. The methodology relies on a restriction of candidate solutions the the subset for which the drift depends on $x_t$, $N_t$ and $t$; the drift is then... | SP:0268dac3486fd3de176b7170b12d864092ad856a |
On Position Embeddings in BERT | 1 INTRODUCTION . Position embeddings ( PEs ) are crucial in Transformer-based architectures for capturing word order ; without them , the representation is bag-of-words . Fully learnable absolute position embeddings ( APEs ) were first proposed by Gehring et al . ( 2017 ) to capture word position in Convolutional Seq2s... | The paper presents a systematic analysis of approaches used to encode position information in transformers and in particular BERT-based models. The paper investigates absolute and relative position embedding strategies that use either fixed/learnable sinusoidal or fully learnable position embeddings. These embeddings a... | SP:c653e54cd37cd4f661b12551c59344dbdfbb8329 |
Improving Calibration through the Relationship with Adversarial Robustness | 1 Introduction . The robustness of machine learning algorithms is becoming increasingly important as ML systems are being used in higher-stakes applications . In one line of research , neural networks are shown to lack adversarial robustness – small perturbations to the input can successfully fool classifiers into maki... | This paper proposes a new method (AR-AdaLS) for label smoothing to improve deep network calibration. In particular, the authors draw a connection between lack of calibration (overconfidence) and examples which are prone to adversarial attacks. They show that by generating smoothed targets based on the adversarial robus... | SP:34177dc9d2e81610d167b996c3f106327c666f94 |
MIROSTAT: A NEURAL TEXT DECODING ALGORITHM THAT DIRECTLY CONTROLS PERPLEXITY | 1 INTRODUCTION . Large-scale generative language models ( LMs ) have received recent attention due to their highquality open-ended text generation ability ( Brown et al. , 2020 ; Radford et al. , 2019 ) . Generating texts from these LMs usually relies on some form of random sampling . Pure sampling often leads to incoh... | Neural text generation models typically rely on sampling schemes for autoregressive decoding. This may range from pure sampling, top-k, top-p to temperature modulated sampling. These methods are mostly heuristic schemes and lack theoretical analysis. This paper tries to fill that gap by analyzing these schemes theoreti... | SP:e1a78b637ef015d15ae3283f6bd3299e5244d457 |
Learning Aggregation Functions | 1 INTRODUCTION . The need to aggregate representations is ubiquitous in deep learning . Some recent examples include max-over-time pooling used in convolutional networks for sequence classification ( Kim , 2014 ) , average pooling of neighbors in graph convolutional networks ( Kipf & Welling , 2017 ) , max-pooling in D... | Universal function representation guarantee requires either highly discontinuous mappings or a highly dimensional latent space. For this reason the authors propose a new parametric family of aggregation functions, called LAF (for learning aggregation functions). It can be seen as a smooth version of the class of functi... | SP:a3f2c5b8bc8bfa03ad589b322c82ac84bca605b2 |
Precondition Layer and Its Use for GANs | 1 INTRODUCTION . Generative Adversarial Nets ( GANs ) ( Goodfellow et al. , 2014 ) successfully transform samples from one distribution to another . Nevertheless , training GANs is known to be challenging , and its performance is often sensitive to hyper-parameters and datasets . Understanding the training difficulties... | This paper mainly solves the instability issue on the spectral normalization for generative adversarial networks (SN-GANs) when training with high dimensional data. To address this, the authors present a preconditioning layer (PC-layer) with two different ways (i.e., FPC and APC) to perform a low-degree polynomial prec... | SP:2434dec4e18251ecfe3d6a7838881e799aad8b4f |
Differentiable Learning of Graph-like Logical Rules from Knowledge Graphs | Logical rules inside a knowledge graph ( KG ) are essential for reasoning , logical inference , and rule mining . However , existing works can only handle simple , i.e. , chain-like and tree-like , rules and can not capture KG ’ s complex semantics , which can be better captured by graph-like rules . Besides , learning... | This paper proposes techniques that generate logical rules out of knowledge graphs; the idea is to produce more complex rules than usual by exploiting a differentiable formulation of the associated learning process. This is a relevant theme as rule learning from knowledge graphs is important in practice due to its pote... | SP:bc280e927e60317d6c2382d5507f522ba58ebe42 |
Meta-learning with negative learning rates | 1 INTRODUCTION . Deep Learning models represent the state-of-the-art in several machine learning benchmarks ( LeCun et al . ( 2015 ) ) , and their performance does not seem to stop improving when adding more data and computing resources ( Rosenfeld et al . ( 2020 ) , Kaplan et al . ( 2020 ) ) . However , they require a... | This paper studies meta-learning in the mixed linear regression setting, focusing on the effect of the within-task step-size on performance. For over-parameterized, under-parameterized, and NTK regimes they derive expressions for test-time loss that suggest that negative or close-to-zero learning rates are optimal, and... | SP:ad96575881588cd2566d2c9c589882a6db9b3874 |
On the Consistency Loss for Leveraging Augmented Data to Learn Robust and Invariant Representations | 1 INTRODUCTION . Recent advances in deep learning has delivered remarkable empirical performance over i.i.d test data , and the community continues to investigate the more challenging and realistic scenario when models are tested in robustness over non-i.i.d data ( e.g. , Ben-David et al. , 2010 ; Szegedy et al. , 2013... | In order to improve the robustness of the learned models, prior work has proposed various data augmentation techniques and different ways of incorporating them into training. This work seeks to provide a general understanding of how we should train with augmented samples in order to learn robust and invariant models fr... | SP:638a6687e5846937cea0e0be3a6e68ad743a787d |
Achieving Linear Speedup with Partial Worker Participation in Non-IID Federated Learning | mKT + 1T ) for full worker participation and a convergence rate O ( √ K√ nT + 1T ) for partial worker participation , where K is the number of local steps , T is the number of total communication rounds , m is the total worker number and n is the worker number in one communication round if for partial worker participat... | This paper provides a new analysis for the FedAvg algorithm, which assumes the data on different workers are non-IID and the objective functions are non-convex. The new analysis improved the existing bounds of FedAvg. Besides, the analysis is also extended to the non-stationary network, where the number of workers part... | SP:33cd383e425b23699614bcff904cc4e52720c29c |
Byzantine-Resilient Non-Convex Stochastic Gradient Descent | 1 INTRODUCTION . Motivated by the pervasiveness of large-scale distributed machine learning , there has recently been significant interest in providing distributed optimization algorithms with strong fault-tolerance guarantees . In this context , the strongest , most stringent fault model is that of Byzantine faults ( ... | The paper considers stochastic gradient descent convergence in a distributed setting with m workers, where up to α workers can be Byzantine, i.e. perform in an arbitrarily adversarial way. In this setting, they develop a variant of SGD which finds a second-order stationary point, prevents Byzantine workers from signifi... | SP:86adaa9dd2414906f708b26e60c86b6e854bb222 |
Improving VAEs' Robustness to Adversarial Attack | 1 INTRODUCTION . Variational autoencoders ( VAEs ) are a powerful approach to learning deep generative models and probabilistic autoencoders ( Kingma & Welling , 2014 ; Rezende et al. , 2014 ) . However , previous work has shown that they are vulnerable to adversarial attacks ( Tabacof et al. , 2016 ; Gondim-Ribeiro et... | This work builds on the vulnerability of VAEs to adversarial attacks to propose investigate how training with alternative losses may alleviate this problem, with a specific focus on disentanglement. In particular it is found that disentanglement constraints may improve the robustness to adversarial attacks, to the detr... | SP:14e55fd6a62febf4c0884964989ac6eb4ae70f63 |
Molecule Optimization by Explainable Evolution | 1 INTRODUCTION . The space of organic molecules is vast , the size of which is exceeding 1060 ( Reymond et al. , 2010 ) . Searching over this vast space for molecules of interest is a challenging task in chemistry , material science , and drug discovery , especially given that molecules are desired to meet multiple cri... | The paper tackles the problem of molecule property optimisation. To this end, the authors proposes an alternating approach consisting of an explainer model and a molecule completion model. The explainer model takes a complete molecule as input and outputs a subgraph that represents the part that contributes most to pro... | SP:cf9319c2a107d0d34ff04da0f53201f3cdff4c24 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.