paper_name
stringlengths
11
170
text
stringlengths
8.07k
307k
summary
stringlengths
152
6.16k
paper_id
stringlengths
43
43
Thinking While Moving: Deep Reinforcement Learning with Concurrent Control
1 INTRODUCTION . In recent years , Deep Reinforcement Learning ( DRL ) methods have achieved tremendous success on a variety of diverse environments , including video games ( Mnih et al. , 2015 ) , zero-sum games ( Silver et al. , 2016 ) , robotic grasping ( Kalashnikov et al. , 2018 ) , and in-hand manipulation tasks ...
The paper tackles the problem of making decisions for the next action while still engaged in doing the previous actions. Such a delay could either be part of the design (like a robot deciding the next action before its actors and motors have come to full rest after the current action) or an artefact of the delays inher...
SP:e62b3d784092d90177912d72a507837c113535f2
Thinking While Moving: Deep Reinforcement Learning with Concurrent Control
1 INTRODUCTION . In recent years , Deep Reinforcement Learning ( DRL ) methods have achieved tremendous success on a variety of diverse environments , including video games ( Mnih et al. , 2015 ) , zero-sum games ( Silver et al. , 2016 ) , robotic grasping ( Kalashnikov et al. , 2018 ) , and in-hand manipulation tasks ...
This paper considers the theoretically interesting and practically important problem of concurrent deep reinforcement learning (DRL), i.e., DRL in which the agent has to decide the next action while performing the previous one. This introduces several significant challenges, including delays/latency and interruption of...
SP:e62b3d784092d90177912d72a507837c113535f2
Simplified Action Decoder for Deep Multi-Agent Reinforcement Learning
1 INTRODUCTION . Humans are highly social creatures and spend vast amounts of time coordinating , collaborating and communicating with others . In contrast to these , at least partially , cooperative settings most progress on AI in games has been in zero-sum games where agents compete against each other , typically ren...
The paper examines the problem of epsilon-greedy exploration in cooperative multi-agent reinforcement learning. Such exploration makes actions less informative to other agents because of the noise added to greedy optimal actions. The suggested solution is to consider two actions: one action is epsilon-greedy and it is ...
SP:a5e28d7de6baafd260cd39c1812059854ac452c0
Simplified Action Decoder for Deep Multi-Agent Reinforcement Learning
1 INTRODUCTION . Humans are highly social creatures and spend vast amounts of time coordinating , collaborating and communicating with others . In contrast to these , at least partially , cooperative settings most progress on AI in games has been in zero-sum games where agents compete against each other , typically ren...
This paper introduces a novel exploitation of the centralized training for decentralized execution regime for Dec-POMDPs with publicly/commonly known actions. In particular, the paper augments independent value-based reinforcement learning by allowing each agent to announce the action it would have taken, had it acted...
SP:a5e28d7de6baafd260cd39c1812059854ac452c0
Learning to Retrieve Reasoning Paths over Wikipedia Graph for Question Answering
1 INTRODUCTION . Open-domain Question Answering ( QA ) is the task of answering a question given a large collection of text documents ( e.g. , Wikipedia ) . Most state-of-the-art approaches for open-domain QA ( Chen et al. , 2017 ; Wang et al. , 2018a ; Lee et al. , 2018 ; Yang et al. , 2019 ) leverage non-parameterize...
This paper introduces a graph-based recurrent retrieval model for retrieving evidence documents in a multi-hop reasoning question answering task. The main idea is that (1) the graph formed by Wikipedia links between passages can be used as constraint for constructing reasoning chains, and (2) the joint encoding of the ...
SP:2db8b4f6637f4d7171a686f3daf386b0df1067a8
Learning to Retrieve Reasoning Paths over Wikipedia Graph for Question Answering
1 INTRODUCTION . Open-domain Question Answering ( QA ) is the task of answering a question given a large collection of text documents ( e.g. , Wikipedia ) . Most state-of-the-art approaches for open-domain QA ( Chen et al. , 2017 ; Wang et al. , 2018a ; Lee et al. , 2018 ; Yang et al. , 2019 ) leverage non-parameterize...
This paper proposes a method to find a sequence of reasoning paragraphs in Wikipedia to answer queries requiring multi-hop reasoning. They make the key observation that answering multi-hop queries might require retrieving evidence that have very less lexical overlap with the question. Given a query, the proposed method...
SP:2db8b4f6637f4d7171a686f3daf386b0df1067a8
Learning by shaking: Computing policy gradients by physical forward-propagation
1 INTRODUCTION . Traditional reinforcement learning crucially relies on reward ( Sutton & Barto , 2018 ) . However , reward binds the agent to a certain task for which the reward represents success . Aligned with the recent surge of interest in unsupervised methods in reinforcement learning ( Baranes & Oudeyer , 2013 ;...
This paper investigates the use of random perturbations applied to a robotic policy to learn a local gradient useful for policy optimization. The method aims to learn a policy directly on a real physical robotic system, bypassing both simulation models and model-free RL. Training pairs are gathered by perturbations of ...
SP:bf5f51cf38c536816105c691b7f1eee6d004fc05
Learning by shaking: Computing policy gradients by physical forward-propagation
1 INTRODUCTION . Traditional reinforcement learning crucially relies on reward ( Sutton & Barto , 2018 ) . However , reward binds the agent to a certain task for which the reward represents success . Aligned with the recent surge of interest in unsupervised methods in reinforcement learning ( Baranes & Oudeyer , 2013 ;...
This paper presents a method for control by estimating the gradient of trajectories w.r.t. the policy parameters by fitting a GP to a set of noisy trajectories executing the same controller. This is opposed to the majority of current RL methods that either learn a forward model or learn a policy. They argue that learni...
SP:bf5f51cf38c536816105c691b7f1eee6d004fc05
Perceptual Regularization: Visualizing and Learning Generalizable Representations
A deployable machine learning model relies on a good representation . A desirable criteria for a good representation is to be interpretable by humans . We propose a technique termed perceptual regularization that enables visualization of the latent representation . By visualizing the learned representation , we are als...
Claims: The authors present perceptual regularization as a method for learning a visualization of deep representations for promoting interpretability and understanding of vulnerability to adversarial attacks. Second, they show their method can explain negative transfer to new tasks. Finally, they show that the represen...
SP:836cc391b0123c787ff4871813ed9f739045330f
Perceptual Regularization: Visualizing and Learning Generalizable Representations
A deployable machine learning model relies on a good representation . A desirable criteria for a good representation is to be interpretable by humans . We propose a technique termed perceptual regularization that enables visualization of the latent representation . By visualizing the learned representation , we are als...
This paper discusses the use of a reconstruction network which is fed an internal representation of a deep network and is trained to reconstruct the input. The classification and reconstruction networks are trained simultaneously with a combined loss function. Various experiments are carried out to test whether this ap...
SP:836cc391b0123c787ff4871813ed9f739045330f
High Fidelity Speech Synthesis with Adversarial Networks
1 INTRODUCTION . The Text-to-Speech ( TTS ) task consists in the conversion of text into speech audio . In recent years , the TTS field has seen remarkable progress , sparked by the development of neural autoregressive models for raw audio waveforms such as WaveNet ( van den Oord et al. , 2016 ) , SampleRNN ( Mehri et ...
I want thank the authors for solving this long-standing GAN challenge in raw waveform synthesis. With all due respect, previous GAN trials for audio synthesis are inspiring, but their audio qualities are far away from the state-of-the-art results. Although the speech fidelity of GAN-TTS is still worse than WaveNet and ...
SP:11684e08a288b62b3952baf4c8bf385be83b858d
High Fidelity Speech Synthesis with Adversarial Networks
1 INTRODUCTION . The Text-to-Speech ( TTS ) task consists in the conversion of text into speech audio . In recent years , the TTS field has seen remarkable progress , sparked by the development of neural autoregressive models for raw audio waveforms such as WaveNet ( van den Oord et al. , 2016 ) , SampleRNN ( Mehri et ...
This paper proposes to enable GAN based TTS in the time domain with the careful designs of the (non-autoregressive) generator and discriminator. There have been various trials of GAN-TTS but not so many success and I'm glad to hear that the proposed method seems to enable GAN-TTS with fast inference thanks to the non-a...
SP:11684e08a288b62b3952baf4c8bf385be83b858d
Evaluating The Search Phase of Neural Architecture Search
1 INTRODUCTION . By automating the design of a neural network for the task at hand , Neural Architecture Search ( NAS ) has tremendous potential to impact the practicality of deep learning ( Zoph & Le , 2017 ; Liu et al. , 2018b ; a ; Tan et al. , 2018 ; Baker et al. , 2016 ) , and has already obtained state-of-the-art...
This works studies the evaluation of search strategies for neural architecture search. It points out existing problems of the current evaluation scheme: (1) only compares the final result without testing the robustness under different random seeds; (2) lacking fair comparison with random baseline under different random...
SP:7460e4f19c4708b96f43a922a2e527f91da3afca
Evaluating The Search Phase of Neural Architecture Search
1 INTRODUCTION . By automating the design of a neural network for the task at hand , Neural Architecture Search ( NAS ) has tremendous potential to impact the practicality of deep learning ( Zoph & Le , 2017 ; Liu et al. , 2018b ; a ; Tan et al. , 2018 ; Baker et al. , 2016 ) , and has already obtained state-of-the-art...
This paper studies the effectiveness of several Neural Architecture Search (NAS) methods comparing it with that of random policy search. The paper concludes that none of these methods for a CNN (trained using CIFAR-10) and RNN model (trained using PTB) are statistically significantly better than the random search. The ...
SP:7460e4f19c4708b96f43a922a2e527f91da3afca
Unsupervised Disentanglement of Pose, Appearance and Background from Images and Videos
Unsupervised landmark learning is the task of learning semantic keypoint-like representations without the use of expensive input keypoint-level annotations . A popular approach is to factorize an image into a pose and appearance data stream , then to reconstruct the image from the factorized components . The pose repre...
The paper presents an unsupervised approach for learning landmarks in images or videos with single objects by separating the representation of the image into foreground and background and factorizing the representation of the foreground into pose and appearance. It builds upon previous work [Jakab 2018, Lorenz 2019] wh...
SP:794d8a1cf74a54e0d0b73832c44d05074ba9a246
Unsupervised Disentanglement of Pose, Appearance and Background from Images and Videos
Unsupervised landmark learning is the task of learning semantic keypoint-like representations without the use of expensive input keypoint-level annotations . A popular approach is to factorize an image into a pose and appearance data stream , then to reconstruct the image from the factorized components . The pose repre...
The paper presents an unsupervised method to get disentanglement of pose, appearance, background from both images domain and video domain. 5 sub-network are used to model pose, appearance, foreground, background, and decoders. Their methods let the network focus more on the foreground to regress the landmark and impro...
SP:794d8a1cf74a54e0d0b73832c44d05074ba9a246
Leveraging Adversarial Examples to Obtain Robust Second-Order Representations
1 INTRODUCTION . In recent years , artificial intelligence systems achieved state-of-the-art performances in image classification tasks ( Russakovsky et al. , 2015 ) ( Krizhevsky et al. , 2012 ) ( He et al. , 2016 ) . Specifically , classification algorithms surpassed top-5 human error rate of 5.1 % on ImageNet ( Russa...
This paper presents a method to represent input points by their relative positions to decisions boundaries using an adversarial attack instead of the more usual output of a deep model (before the last linear classification layer). The motivation of the method is to obtain representations that are more robust to input d...
SP:c610bfdc080b0c939039a6083d71d5f03d3a3f32
Leveraging Adversarial Examples to Obtain Robust Second-Order Representations
1 INTRODUCTION . In recent years , artificial intelligence systems achieved state-of-the-art performances in image classification tasks ( Russakovsky et al. , 2015 ) ( Krizhevsky et al. , 2012 ) ( He et al. , 2016 ) . Specifically , classification algorithms surpassed top-5 human error rate of 5.1 % on ImageNet ( Russa...
The authors propose an approach to reduce the sensitivity of neural networks to visual distortions. To do so, they modify the representation of a data point within a network, using its relative position (to other points) in representation space rather than its absolute position (which is measured as distance of a poin...
SP:c610bfdc080b0c939039a6083d71d5f03d3a3f32
Conditional Invertible Neural Networks for Guided Image Generation
1 INTRODUCTION . Generative adversarial networks ( GANs ) produce ever larger and more realistic samples ( Karras et al. , 2017 ; Brock et al. , 2019 ) . Hence they have become the primary choice for a majority of image generation tasks . As such , their conditional variants ( cGANs ) would appear to be the natural too...
The paper presents an invertible generative network, for conditional image generation. The model is an extension of Real NVP with a conditioning component. Experiments are performed for image generation on two tasks: class conditional generation on MNIST and image colorization conditioned on a grey scale image (lumina...
SP:594bd533a35136573306102a080e81fefeb0c37b
Conditional Invertible Neural Networks for Guided Image Generation
1 INTRODUCTION . Generative adversarial networks ( GANs ) produce ever larger and more realistic samples ( Karras et al. , 2017 ; Brock et al. , 2019 ) . Hence they have become the primary choice for a majority of image generation tasks . As such , their conditional variants ( cGANs ) would appear to be the natural too...
This paper proposes conditional Invertible Neural Networks (cINN), which introduces conditioning to conventional flow-based generative models. Conditioning is injected into the model via a conditional affine coupling block, which concatenates conditioning with the input to the scaling and shifting sub-networks in the c...
SP:594bd533a35136573306102a080e81fefeb0c37b
Deep Coordination Graphs
1 INTRODUCTION . One of the central challenges in cooperative multi-agent reinforcement learning ( MARL , Oliehoek & Amato , 2016 ) is coping with the size of the joint action space , which grows exponentially in the number of agents . For example , this paper evaluates tasks where eight agents each have six actions to...
Teh authors propose to learn value functions that are a sum of utility (single-agent) and payoff (2-agent) components. The weights between all function are shared (common RNN). This stands in contrast to VDNs (only uses utility functions or centralized value functions. The authors evaluate on predator-prey, where they ...
SP:3ae91f3ba2414da1c21b536b52a3e25745ff75cf
Deep Coordination Graphs
1 INTRODUCTION . One of the central challenges in cooperative multi-agent reinforcement learning ( MARL , Oliehoek & Amato , 2016 ) is coping with the size of the joint action space , which grows exponentially in the number of agents . For example , this paper evaluates tasks where eight agents each have six actions to...
This paper proposes a pairwise communication between agents using a shared neural network. The idea is to define the joint action-value function as the sum of individual agent's values + pair-wise payoff between agents, which is based on the prior work [Castellini et al.]. In particular, this paper proposes to share th...
SP:3ae91f3ba2414da1c21b536b52a3e25745ff75cf
Probability Calibration for Knowledge Graph Embedding Models
1 INTRODUCTION . Knowledge graph embedding models are neural architectures that learn vector representations ( i.e . embeddings ) of nodes and edges of a knowledge graph . Such knowledge graph embeddings have applications in knowledge graph completion , knowledge discovery , entity resolution , and link-based clusterin...
This is the first work that studies probability calibration for knowledge graph embedding models. In the case where ground-truth negatives are available the authors directly use off-the-shelf established calibration techniques (Platt scaling, isotonic regression). When ground-truth negatives are not available they prop...
SP:4fe65143f73a8de8303ad2fbd71e998ae9317152
Probability Calibration for Knowledge Graph Embedding Models
1 INTRODUCTION . Knowledge graph embedding models are neural architectures that learn vector representations ( i.e . embeddings ) of nodes and edges of a knowledge graph . Such knowledge graph embeddings have applications in knowledge graph completion , knowledge discovery , entity resolution , and link-based clusterin...
The paper studies probability calibration for three different knowledge graph embedding methods, with a focus on TransE evaluated on the task of knowledge graph triple classification. It studies Brier and log loss performance of Platt scaling and isotonic regression probability calibration on WN11 for TransE and claims...
SP:4fe65143f73a8de8303ad2fbd71e998ae9317152
Putting Machine Translation in Context with the Noisy Channel Model
1 INTRODUCTION . There have been many recent demonstrations that neural language models based on Transformers ( Vaswani et al. , 2017 ; Dai et al. , 2019 ) are capable of learning to generate remarkably coherent documents with few ( Zellers et al. , 2019 ) or no ( Radford et al. , 2019 ) conditioning variables . Althou...
This paper presents a simple approach for document-level machine translation. The idea is to use a language model on the target side and a reverse translation model to choose the best document-level translation. This is theoretically justified by Bayes’ rule and the assumption that the sentences are conditionally indep...
SP:50fc4ed275dfba23494957478389b0a9ba30838e
Putting Machine Translation in Context with the Noisy Channel Model
1 INTRODUCTION . There have been many recent demonstrations that neural language models based on Transformers ( Vaswani et al. , 2017 ; Dai et al. , 2019 ) are capable of learning to generate remarkably coherent documents with few ( Zellers et al. , 2019 ) or no ( Radford et al. , 2019 ) conditioning variables . Althou...
The paper describes a noisy channel approach for document-level translation, which does not rely on parallel documents to train. The approach relies on a sentence-level translation model (from target-to-source languages) and a document level language model (on target language), each is trained separately. For decoding,...
SP:50fc4ed275dfba23494957478389b0a9ba30838e
Are there any 'object detectors' in the hidden layers of CNNs trained to identify objects or scenes?
1 INTRODUCTION . There have been recent attempts to understand how neural networks ( NNs ) work by analyzing hidden units one-at-a-time using various measures such as localist selectivity ( Bowers et al. , 2014 ) , class-conditional mean activity selectivity ( CCMAS ) ( Morcos et al. , 2018 ) , precision ( Zhou et al. ...
The paper makes an empirical claim that CNNs for object recognition do not contain hidden neuron which is highly selective to each class, mainly based on three aspects: (a) metrics related to the maximum informedness, (b) jitterplots of activation data, and (c) a user study assessing whether generated images maximizing...
SP:fa9f5943cd6501a9e72e66ca071ba4b15923bfa1
Are there any 'object detectors' in the hidden layers of CNNs trained to identify objects or scenes?
1 INTRODUCTION . There have been recent attempts to understand how neural networks ( NNs ) work by analyzing hidden units one-at-a-time using various measures such as localist selectivity ( Bowers et al. , 2014 ) , class-conditional mean activity selectivity ( CCMAS ) ( Morcos et al. , 2018 ) , precision ( Zhou et al. ...
The paper empirically studies the category selectivity of individual cells in hidden units of CNNs. It is a sort of "meta-study" and comparison of different metrics proposed to identify cells with a preference for a specific target category. The claimed finding is that there are no cells that are "sufficiently" selecti...
SP:fa9f5943cd6501a9e72e66ca071ba4b15923bfa1
Learning to Make Generalizable and Diverse Predictions for Retrosynthesis
1 INTRODUCTION . This paper proposes a novel approach for one-step retrosynthesis . This task is crucial for material and drug manufacturing ( Corey & Wipke , 1969 ; Corey , 1991 ) and aims to predict which reactants are needed to generate a given target molecule as the main product . For instance , Figure 1 demonstrat...
This paper is very well-written and combines the state-of-the-art NLP model and the domain knowledge in retrosynthetic reaction predictions. The authors propose pre-training models to help improve the model's generation to rare reactions. In addition, a discrete latent variable model is used in the model to encourage ...
SP:1eb66a8cc370cb0a84c08a12a7a8842e785f5d6f
Learning to Make Generalizable and Diverse Predictions for Retrosynthesis
1 INTRODUCTION . This paper proposes a novel approach for one-step retrosynthesis . This task is crucial for material and drug manufacturing ( Corey & Wipke , 1969 ; Corey , 1991 ) and aims to predict which reactants are needed to generate a given target molecule as the main product . For instance , Figure 1 demonstrat...
Given a target compound, the authors suggest a method to predict likely chemical reactants to produce the target. The authors provide a transformer based model to predict the reactants. Existing methods do not generalize well for rare reactions and the training data has only one reactant set for each target even though...
SP:1eb66a8cc370cb0a84c08a12a7a8842e785f5d6f
Masked Based Unsupervised Content Transfer
We consider the problem of translating , in an unsupervised manner , between two domains where one contains some additional information compared to the other . The proposed method disentangles the common and separate parts of these domains and , through the generation of a mask , focuses the attention of the underlying...
This paper proposes a method for unpaired image-to-image translation, where the target domain explicitly contains some additional information than the source domain. The authors use auto-encoders to separate the common and specific representations and to generate masks, which seems to be related to [1]. The authors emp...
SP:0d7b31f997fb88fad3c01a6d7e74e1f8227b02b0
Masked Based Unsupervised Content Transfer
We consider the problem of translating , in an unsupervised manner , between two domains where one contains some additional information compared to the other . The proposed method disentangles the common and separate parts of these domains and , through the generation of a mask , focuses the attention of the underlying...
This work proposed a mask based approach for instance-level unsupervised content transfer, which is an extension of the disentanglement work in (Press et al., 2019) and the attention guided translation (Chen et al., 2018, Mejjati et al., 2018). Unlike the disentanglement work, the introduced mask allows the adaptation ...
SP:0d7b31f997fb88fad3c01a6d7e74e1f8227b02b0
Semi-Supervised Learning with Normalizing Flows
1 INTRODUCTION . In many domains unlabeled data is plentiful , while labeled data may be scarce . Semi-supervised learning framework leverages both labeled and unlabeled data reducing the need for expensive manual annotation . Recently , consistency-based methods have shown outstanding performance in semisupervised ima...
The paper describes how to use normalising flows for Semi Supervised Learning (SSL). Briefly, the method consists in finding a (bijective) map for transforming a mixture of Gaussians into a density approximating the empirical data-distribution -- as usual for flow methods, the parameters are found through likelihood ma...
SP:b109c13a5d110eb9254eafc226881e44e393120d
Semi-Supervised Learning with Normalizing Flows
1 INTRODUCTION . In many domains unlabeled data is plentiful , while labeled data may be scarce . Semi-supervised learning framework leverages both labeled and unlabeled data reducing the need for expensive manual annotation . Recently , consistency-based methods have shown outstanding performance in semisupervised ima...
The paper describes a normalising flow with the prior distribution represented by a Gaussian mixture model (GMM). The method, FlowGMM, maps each class of the dataset to a Gaussian distribution in the latent space by optimising the joint likelihood of both labelled and unlabelled data, thus making the method useful for ...
SP:b109c13a5d110eb9254eafc226881e44e393120d
Unsupervised Distillation of Syntactic Information from Contextualized Word Representations
Contextualized word representations , such as ELMo and BERT , were shown to perform well on various semantic and structural ( syntactic ) task . In this work , we tackle the task of unsupervised disentanglement between semantics and structure in neural language representations : we aim to learn a transformation of the ...
This paper aims to disentangle semantics and syntax in contextualized word representations. The main idea is to learn a transformation of the contexualized representations that will make two word representations to be more similar if they appear in the same syntactic context, but less similar if they appear in differen...
SP:909aed1f85884ac86edbb664e8617ef711a47eb0
Unsupervised Distillation of Syntactic Information from Contextualized Word Representations
Contextualized word representations , such as ELMo and BERT , were shown to perform well on various semantic and structural ( syntactic ) task . In this work , we tackle the task of unsupervised disentanglement between semantics and structure in neural language representations : we aim to learn a transformation of the ...
The authors state a clear hypothesis: it is possible to extract syntactic information from contextualized word vectors in an unsupervised manner. The method of creating syntactically equivalent (but semantically different) sentences is indeed interesting on its own. Experiments do support the main hypothesis -- the dis...
SP:909aed1f85884ac86edbb664e8617ef711a47eb0
Neural Clustering Processes
1 INTRODUCTION . Mixture models ( or equivalently , probabilistic clustering models ) are a staple of statistical modelling in which a discrete latent variable is introduced for each observation , indicating its mixture component identity . Popular inference methods in these models fall into two main classes . When exp...
In this paper, the authors consider the neural amortized inference for clustering processes, in which the number of cluster can be automatically adapted based on the observed samples. The proposed algorithm largely follows the standard variational auto-encoder. The major contribution of the paper is the design of the p...
SP:13d9cb4549113d278e5b7e6fe8185002137e0671
Neural Clustering Processes
1 INTRODUCTION . Mixture models ( or equivalently , probabilistic clustering models ) are a staple of statistical modelling in which a discrete latent variable is introduced for each observation , indicating its mixture component identity . Popular inference methods in these models fall into two main classes . When exp...
This paper introduces a novel deep learning architecture for efficient amortized Bayesian inference over mixture models. Unlike previous approaches to amortized clustering, the proposed method allows us to treat local discrete labels of data points and infer the unbounded number of mixture components, making it more fl...
SP:13d9cb4549113d278e5b7e6fe8185002137e0671
Learning Compact Embedding Layers via Differentiable Product Quantization
Embedding layers are commonly used to map discrete symbols into continuous embedding vectors that reflect their semantic meanings . Despite their effectiveness , the number of parameters in an embedding layer increases linearly with the number of symbols and poses a critical challenge on memory and storage constraints ...
This paper works on methods for compressed embedding layers for low memory inference, where the compressed embedding are learned together with the task-specific models in a differentiable end-to-end fashion. The methods in the paper build on a slight variant of the K-way D-dimensional discrete code representation propo...
SP:5bf8fd6f74c8aa558dd854cdd70d1afa31a2989b
Learning Compact Embedding Layers via Differentiable Product Quantization
Embedding layers are commonly used to map discrete symbols into continuous embedding vectors that reflect their semantic meanings . Despite their effectiveness , the number of parameters in an embedding layer increases linearly with the number of symbols and poses a critical challenge on memory and storage constraints ...
This paper considers the problem of having compact yet expressive KD code for NLP tasks. The authors claim that the proposed differentiable product quantization framework has better compression but similar performance compared to existing KD codes.The authors present two instances of the DPQ framework: DPQ-SX using sof...
SP:5bf8fd6f74c8aa558dd854cdd70d1afa31a2989b
Counterfactual Regularization for Model-Based Reinforcement Learning
1 INTRODUCTION . Recent advancements in the use of variational inference and generative neural network architectures have made it possible to build accurate transition models for high-dimensional sequential decisionmaking environments . Networks have been trained to accurately predict future states conditioned on actio...
The paper presents regularization techniques for model based reinforcement learning which attempt to build counterfactual reasoning into the model. In particular, they present auxiliary loss terms which can be used in "what if" scenarios where the actual state is unknown. Given certain assumptions, they show that this ...
SP:bb4ad6af4f43dab9af6411b2d72bacebbb9f29ab
Counterfactual Regularization for Model-Based Reinforcement Learning
1 INTRODUCTION . Recent advancements in the use of variational inference and generative neural network architectures have made it possible to build accurate transition models for high-dimensional sequential decisionmaking environments . Networks have been trained to accurately predict future states conditioned on actio...
This paper proposes a method called "counterfactual regularization" whereby the dynamics/transition model is encourage to not have degeneracies where the actions don't influence the state transitions. Concretely, this is done by, for every state, computing the maximum deviation of Transition model under a different ac...
SP:bb4ad6af4f43dab9af6411b2d72bacebbb9f29ab
Efficient Wrapper Feature Selection using Autoencoder and Model Based Elimination
1 INTRODUCTION . Feature selection is a preprocessing technique that ranks the significance of features to eliminate features that are insignificant to the task at hand . As examined by Yu and Liu ( 2003 ) , it is a powerful tool to alleviate the curse of dimensionality , reduce training time and increase the accuracy ...
In this paper, the authors present an iterative approach for feature selection which selects features based both on the relevance and redundancy of each feature. The relevance of each feature is determined using a mild variant of the Feature Quality Index; essentially, the relevance is computed as the loss in model per...
SP:860fadd8179582b24954784b0020a94a322e4675
Efficient Wrapper Feature Selection using Autoencoder and Model Based Elimination
1 INTRODUCTION . Feature selection is a preprocessing technique that ranks the significance of features to eliminate features that are insignificant to the task at hand . As examined by Yu and Liu ( 2003 ) , it is a powerful tool to alleviate the curse of dimensionality , reduce training time and increase the accuracy ...
The article "Efficient Wrapper Feature Selection using Autoencoder and Model Based Elimination" considers the problem of feature selection for a broad class of machine learning models. The authors argue that it is important to consider the relevance of features for the considered supervised ML problem and redundancy of...
SP:860fadd8179582b24954784b0020a94a322e4675
Stabilizing Off-Policy Reinforcement Learning with Conservative Policy Gradients
1 INTRODUCTION . Reinforcement Learning ( RL ) is a dynamical learning paradigm , in which the algorithm ( also known as the ‘ agent ’ ) learns through sequential interaction with the environment . On each round , the agent performs an action , which transitions it into a new state , and is provided with a state-depend...
This paper proposes an approach called conservative policy gradients to stabilize the training of deep policy gradient methods. At fixed intervals, the current policy and a separate target policy are evaluated with a number of rollouts. The target policy is then updated to match the current policy only if the current p...
SP:c49cea9b8b1ad72568948d3d184b2d95d4cc00e3
Stabilizing Off-Policy Reinforcement Learning with Conservative Policy Gradients
1 INTRODUCTION . Reinforcement Learning ( RL ) is a dynamical learning paradigm , in which the algorithm ( also known as the ‘ agent ’ ) learns through sequential interaction with the environment . On each round , the agent performs an action , which transitions it into a new state , and is provided with a state-depend...
This paper proposes a simple method for stabilizing the off-policy deep reinforcement learning algorithm, which updates the target network only when the online network performs better than the target network in order to ensure the stability guarantees. More specifically, at every T time steps, they execute both the onl...
SP:c49cea9b8b1ad72568948d3d184b2d95d4cc00e3
Thieves on Sesame Street! Model Extraction of BERT-based APIs
1 INTRODUCTION . Machine learning models represent valuable intellectual property : the process of gathering training data , iterating over model design , and tuning hyperparameters costs considerable money and effort . As such , these models are often only indirectly accessible through web APIs that allow users to que...
The authors explore how well model extraction works on recent BERT-based NLP models. The question is: how easy is it for an adversary model to learn to imitate the victim model, only from novel inputs and the corresponding outputs? Importantly, the adversary is supposed to not have access to the original training set. ...
SP:051dd7bc82951cead504d0861e2048c99d20dbe8
Thieves on Sesame Street! Model Extraction of BERT-based APIs
1 INTRODUCTION . Machine learning models represent valuable intellectual property : the process of gathering training data , iterating over model design , and tuning hyperparameters costs considerable money and effort . As such , these models are often only indirectly accessible through web APIs that allow users to que...
This paper studies the effectiveness of model extraction techniques on large pretrained language models like BERT. The core hypothesis of the paper is that using pretrained language models, and pretrained contextualized embeddings, has made it easier to reconstruct models using model stealing/extraction methods. Furthe...
SP:051dd7bc82951cead504d0861e2048c99d20dbe8
Continuous Convolutional Neural Network forNonuniform Time Series
1 INTRODUCTION . Convolutional neural network ( CNN ) , together with recurrent neural network ( RNN ) , is among the most popular deep learning architectures to process time series data . However , both CNN and RNN rest on the assumption that both the input and output data are sampled uniformly . However , many time-s...
A method that was proposed by authors deals with a problem of non-uniform data in time series. One of ways to deal with this problem is interpolate input signal between data points. In signal processing a standard way to interpolate is to apply a convolution with a kernel. This operation, by itself, is a non-trivial, s...
SP:dc098f0332d6cda0884d293608bf992133b9aa20
Continuous Convolutional Neural Network forNonuniform Time Series
1 INTRODUCTION . Convolutional neural network ( CNN ) , together with recurrent neural network ( RNN ) , is among the most popular deep learning architectures to process time series data . However , both CNN and RNN rest on the assumption that both the input and output data are sampled uniformly . However , many time-s...
The paper proposes a continuous CNN model to accommodate the nonuniform time series data. The model learns the interpolation and the convolution kernel functions in an end-to-end manner, so that it can capture the signal patterns and be flexible. A layer has three networks, which learn a kernel function to represent th...
SP:dc098f0332d6cda0884d293608bf992133b9aa20
DUAL ADVERSARIAL MODEL FOR GENERATING 3D POINT CLOUD
1 INTRODUCTION . Point clouds are widely used in many fields , such as vision , 3D object detection and so on . In particular , 3D point clouds can represent geometric details of the object , and suitable for simple geometric transformation . However , there exist significant disadvantages in the point cloud data . For...
The paper focuses on designing a generative framework for 3-D point data clouds. These point clouds correspond to objects shapes in 3-dimensions. According to the paper, previous approaches for generating such 3-D point clouds involved autoencoder and GANs used separately. The authors propose a framework combining both...
SP:a77c90c209b32b8b5fce5f1ce55752a5b028cb7e
DUAL ADVERSARIAL MODEL FOR GENERATING 3D POINT CLOUD
1 INTRODUCTION . Point clouds are widely used in many fields , such as vision , 3D object detection and so on . In particular , 3D point clouds can represent geometric details of the object , and suitable for simple geometric transformation . However , there exist significant disadvantages in the point cloud data . For...
This paper proposed a new dual generation model for learning representation by combining GAN and Autoencoder for 3D point cloud data. The idea is to add a regularization term over the GAN loss, and the regularization term measures the distance between the output of the encoder of Autoencoder and that of GAN by two poin...
SP:a77c90c209b32b8b5fce5f1ce55752a5b028cb7e
Kernel and Rich Regimes in Overparametrized Models
1 INTRODUCTION . A string of recent papers study neural networks trained with gradient descent in the “ kernel regime. ” They observe that , in a certain regime , networks trained with gradient descent behave as kernel methods ( Jacot et al. , 2018 ; Daniely et al. , 2016 ; Daniely , 2017 ) . This allows one to prove c...
This paper analyzes an inductive bias of the gradient flow for diagonal two-or higher-homogeneous models and characterizes a limit point depending on the initialization scale of parameters. Concretely, the paper shows that the gradient flow converges to an interpolator attaining minimum L1- (or L2-norm) when the scale ...
SP:b030108c8ae5c6ad968b76072d0ab08af9de4cd9
Kernel and Rich Regimes in Overparametrized Models
1 INTRODUCTION . A string of recent papers study neural networks trained with gradient descent in the “ kernel regime. ” They observe that , in a certain regime , networks trained with gradient descent behave as kernel methods ( Jacot et al. , 2018 ; Daniely et al. , 2016 ; Daniely , 2017 ) . This allows one to prove c...
I really appreciated this paper. It discusses a very complex question ("Are we learning in a kernel regime, or in a rich regime where features are identified") by looking at perhaps the simplest model the authors could think of, and then study in detail the model. And how simple it turns out to be: just a linear regres...
SP:b030108c8ae5c6ad968b76072d0ab08af9de4cd9
StructBERT: Incorporating Language Structures into Pre-training for Deep Language Understanding
1 INTRODUCTION . A pre-trained language model ( LM ) is a key component in many natural language understanding ( NLU ) tasks such as semantic textual similarity ( Cer et al. , 2017 ) , question answering ( Rajpurkar et al. , 2016 ) and sentiment classification ( Socher et al. , 2013 ) . In order to obtain reliable lang...
This paper proposes to use additional structures within and between sentences for pre-training BERT. The basic idea is to shuffle either some n-grams within sentences or the sentences in texts, then train the model to predict the correct orders. Experiments in this work show that, with this additional training objectiv...
SP:bae39837d7167b44abee412dd258c7640d315471
StructBERT: Incorporating Language Structures into Pre-training for Deep Language Understanding
1 INTRODUCTION . A pre-trained language model ( LM ) is a key component in many natural language understanding ( NLU ) tasks such as semantic textual similarity ( Cer et al. , 2017 ) , question answering ( Rajpurkar et al. , 2016 ) and sentiment classification ( Socher et al. , 2013 ) . In order to obtain reliable lang...
This paper proposed a new pre-trained language model based on BERT, called StructBERT. The key contributions are the two new pre-train objectives, (1) word structural objective, where the goal is to reconstruct the right order of intentionally shuffled word tokens, and (2) sentence structural objective, a three-class s...
SP:bae39837d7167b44abee412dd258c7640d315471
Intensity-Free Learning of Temporal Point Processes
1 INTRODUCTION . Visits to hospitals , purchases in e-commerce systems , financial transactions , posts in social media — various forms of human activity can be represented as discrete events happening at irregular intervals . The framework of temporal point processes is a natural choice for modeling such data . By com...
The authors propose a new paradigm for learning models for point processes which circumvents the need to explicitly model the conditional intensity. This utilizes recent work on Normalizing flows and upends a long-standing paradigm. The paper is a true tour-de-force and the authors make a very convincing case for why i...
SP:681fe0a9929f5e6473029e70d5073913183e38a3
Intensity-Free Learning of Temporal Point Processes
1 INTRODUCTION . Visits to hospitals , purchases in e-commerce systems , financial transactions , posts in social media — various forms of human activity can be represented as discrete events happening at irregular intervals . The framework of temporal point processes is a natural choice for modeling such data . By com...
This paper describes a simple yet effective technique for learning temporal point processes using a mixture of log-normal densities whose parameters are estimated with neural networks that also adds conditional information. The method is shown to perform better than more recent techniques for density estimation such as...
SP:681fe0a9929f5e6473029e70d5073913183e38a3
Optimal Attacks on Reinforcement Learning Policies
1 INTRODUCTION . Advances in Deep Reinforcement Learning ( RL ) have made it possible to train end-to-end policies achieving superhuman performance on a large variety of tasks , such as playing Atari games Mnih et al . ( 2013 ; 2016 ) , playing Go Silver et al . ( 2016 ; 2017 ) , as well as controlling systems with con...
The authors of this paper propose a novel adversarial attack for deep reinforcement learning. Different from the classical attacks, e.g., FGSM, they explicitly minimize the reward collect by the agent in a form of Markov decision process. Experiment results demonstrate that the proposed approach can damage the well-per...
SP:94d8eb827399e58de2a0aed1e5d3a1d629d7fcf7
Optimal Attacks on Reinforcement Learning Policies
1 INTRODUCTION . Advances in Deep Reinforcement Learning ( RL ) have made it possible to train end-to-end policies achieving superhuman performance on a large variety of tasks , such as playing Atari games Mnih et al . ( 2013 ; 2016 ) , playing Go Silver et al . ( 2016 ; 2017 ) , as well as controlling systems with con...
The paper investigates adversarial attacks on learned (fixed) policies. In particular, they devise optimal attacks in the sense that e.g. the agent’s objective should be minimised by the attacker. It is assumed that the attacker can manipulate the observed state of the agent at each time step during testing in a restri...
SP:94d8eb827399e58de2a0aed1e5d3a1d629d7fcf7
Unifying Graph Convolutional Neural Networks and Label Propagation
Label Propagation ( LPA ) and Graph Convolutional Neural Networks ( GCN ) are both message passing algorithms on graphs . Both solve the task of node classification but LPA propagates node label information across the edges of the graph , while GCN propagates and transforms node feature information . However , while co...
The paper proposed an unified model for Label Propagation (LPA) and Graph Convolutional Neural Networks (GCN). It is shown how it is possible to infer the relationship between LPA and GCN in terms of label or feature smoothing (how label/feature does propagate over the neighbors) and label or feature influence over the...
SP:d9f17344cd266b16a70c37d891b2c64a6d454908
Unifying Graph Convolutional Neural Networks and Label Propagation
Label Propagation ( LPA ) and Graph Convolutional Neural Networks ( GCN ) are both message passing algorithms on graphs . Both solve the task of node classification but LPA propagates node label information across the edges of the graph , while GCN propagates and transforms node feature information . However , while co...
This paper introduces a unified model which combines label propagation algorithm (LPA) and graph convolutional networks (GCNs) for node classification. The motivation of this combination is supported by two analysis on the feature/label smoothing and feature/label influence. The proposed GCN-LPA framework utilizeds LPA...
SP:d9f17344cd266b16a70c37d891b2c64a6d454908
Learning to Link
1 INTRODUCTION . Overview . Clustering is an important component of modern data analysis . For example , we might cluster emails as a pre-processing step for spam detection , or we might cluster individuals in a social network in order to suggest new connections . There are a myriad of different clustering algorithms ,...
This paper proposed a data-driven method of selecting a linkage-based clustering algorithm from a large space. The space of algorithms is parameterized by two sets of parameters which indicate the convex combinations of metrics and merge functions. They analyze the sample complexity for small generalization error. An e...
SP:72719789a5653ae47a06265616e77c12fe9da7d4
Learning to Link
1 INTRODUCTION . Overview . Clustering is an important component of modern data analysis . For example , we might cluster emails as a pre-processing step for spam detection , or we might cluster individuals in a social network in order to suggest new connections . There are a myriad of different clustering algorithms ,...
This paper studies the problem of learning both the distance metric and a linkage rule from clustering examples. Suppose we have L metrics d_1, …, d_L and L’ linkage rules for hierarchical agglomerative clustering, D_1, …, D_L’ where each rule is a 2-point-based merge function (i.e. computes the distance between some t...
SP:72719789a5653ae47a06265616e77c12fe9da7d4
Improving Adversarial Robustness Requires Revisiting Misclassified Examples
1 INTRODUCTION . Despite their great success in applications such as computer vision ( He et al. , 2016 ) , speech recognition ( Wang et al. , 2017 ) and natural language processing ( Devlin et al. , 2018 ; Zeng et al. , 2019 ) , deep neural networks ( DNNs ) are extremely vulnerable to adversarial examples crafted by ...
Neural Networks (NN) have been shown to be susceptible to various adversarial attacks i.e. if we perturb the "x" just a little, the output prediction changes. So, there has been much research devoted to how we can make NNs robust to such attacks. Typically, the adversarial examples that are used to train adversarially ...
SP:141200456d0b5c1970dcb11c11acefca7717cf1b
Improving Adversarial Robustness Requires Revisiting Misclassified Examples
1 INTRODUCTION . Despite their great success in applications such as computer vision ( He et al. , 2016 ) , speech recognition ( Wang et al. , 2017 ) and natural language processing ( Devlin et al. , 2018 ; Zeng et al. , 2019 ) , deep neural networks ( DNNs ) are extremely vulnerable to adversarial examples crafted by ...
The paper improves adversarial training by introducing two modifications to the loss function: (i) a "boosted" version of the cross-entropy loss that involves a term similar to a large-margin loss, and (ii) weighting the adversarial loss differently depending on how correctly classified an example is. When put together...
SP:141200456d0b5c1970dcb11c11acefca7717cf1b
Learning Surrogate Losses
1 INTRODUCTION . In reality , a large set of loss functions can not be directly minimized by gradient-based methods because they are either piece-wise continuous , non-differentiable , or non-decomposable ( Zhang et al. , 2018 ) . For example , binary classification models are often evaluated using the miss-classificat...
This paper proposes a method of learning loss functions in addition to the learning of predictors. Since it's not easy to optimize loss functions that evaluate the accuracy, surrogate loss functions have been widely employed. The design of the surrogate loss is problem-dependent, and handcraft is required. This paper t...
SP:6f36e4a7b604bf894e65be9f88cc0f353a8bd1bf
Learning Surrogate Losses
1 INTRODUCTION . In reality , a large set of loss functions can not be directly minimized by gradient-based methods because they are either piece-wise continuous , non-differentiable , or non-decomposable ( Zhang et al. , 2018 ) . For example , binary classification models are often evaluated using the miss-classificat...
In this paper, the authors propose to learn surrogate loss functions for non-differentiable and non-decomposable loss. An alternative minimization method is used for training the surrogate network and prediction model. Learning surrogate loss functions for different tasks is somewhat novel, although there are some prio...
SP:6f36e4a7b604bf894e65be9f88cc0f353a8bd1bf
Emergent Tool Use From Multi-Agent Autocurricula
1 INTRODUCTION . Creating intelligent artificial agents that can solve a wide variety of complex human-relevant tasks has been a long-standing challenge in the artificial intelligence community . Of particular relevance to humans will be agents that can sense and interact with objects in a physical world . One approach...
Authors in introduce a new competitive/cooperative physics-based environment in which different teams of agents compete in a visual concealment and search task with visibility-based team-based rewards (although There are no explicit incentives for agents to interact with objects in the environment). They show that, com...
SP:f42b9d3416085bf309889db41c1981a32b7d8b21
Emergent Tool Use From Multi-Agent Autocurricula
1 INTRODUCTION . Creating intelligent artificial agents that can solve a wide variety of complex human-relevant tasks has been a long-standing challenge in the artificial intelligence community . Of particular relevance to humans will be agents that can sense and interact with objects in a physical world . One approach...
The authors report on an empirical study of emergent behavior of multiple RL agents learning to play hide-and-seek (a sparse reward task). The main point of this paper is that RL agents learning at scale (large number of samples, batch-size 64000). can learn to solve tasks with strategies that are human-interpretable (...
SP:f42b9d3416085bf309889db41c1981a32b7d8b21
A Mutual Information Maximization Perspective of Language Representation Learning
1 INTRODUCTION . Advances in representation learning have driven progress in natural language processing . Performance on many downstream tasks have improved considerably , achieving parity with human baselines in benchmark leaderboards such as SQuAD ( Rajpurkar et al. , 2016 ; 2018 ) and GLUE ( Wang et al. , 2019 ) . ...
The paper proposes to make a clear connection between the InfoNCE learning objective (which is a lower bound of the mutual information) and multiple language models like BERT and XLN. Then based on the observation that classical LM can be seen as instances of InfoNCE, they propose a new (InfoWord) model relying on the ...
SP:6102c0d1cbc3ee0f3e9260c40e53c09a34d3d94d
A Mutual Information Maximization Perspective of Language Representation Learning
1 INTRODUCTION . Advances in representation learning have driven progress in natural language processing . Performance on many downstream tasks have improved considerably , achieving parity with human baselines in benchmark leaderboards such as SQuAD ( Rajpurkar et al. , 2016 ; 2018 ) and GLUE ( Wang et al. , 2019 ) . ...
This paper first gives a concise yet precise summary of maximizing one of variational lower bounds of mutual information, InfoNCE, then it provides an alternative view to explain case by case why word embedding Skip-gram, BERT, XLNet work in practice can be viewed by InfoNCE framework, thus we have a good understand fo...
SP:6102c0d1cbc3ee0f3e9260c40e53c09a34d3d94d
Unsupervised Few Shot Learning via Self-supervised Training
1 INTRODUCTION . Few-shot learning , which aims to accomplish a learning task by using very few training examples , is receiving increasing attention in the machine learning community . The challenge of few-shot learning lies on that traditional techniques such as fine-tuning would normally incur overfitting ( Wang et ...
This paper aims to conduct few-shot learning on unlabeled data (instead of on training tasks with few-shot labeled data per task). The proposed algorithm is a trivial combination of existing clustering method and a few-shot learning method, i.e., the clustering provides pseudo labels, from which a series of few-shot tr...
SP:443b64bd0352a4283857b9c0e416436b6936732d
Unsupervised Few Shot Learning via Self-supervised Training
1 INTRODUCTION . Few-shot learning , which aims to accomplish a learning task by using very few training examples , is receiving increasing attention in the machine learning community . The challenge of few-shot learning lies on that traditional techniques such as fine-tuning would normally incur overfitting ( Wang et ...
This paper considers the problem of learning an image representation for few-shot learning without using image labels during training. This is a well-motivated problem since (as the paper points out) learning such a representation using "episodes" of low-shot learning problems as examples may require a large amount of ...
SP:443b64bd0352a4283857b9c0e416436b6936732d
Intrinsically Motivated Discovery of Diverse Patterns in Self-Organizing Systems
1 INTRODUCTION . Self-organization of patterns that emerge from local rules is a pervasive phenomena in natural and artificial dynamical systems ( Ball , 1999 ) . It ranges from the formation of snow flakes , spots and rays on animal ’ s skin , to spiral galaxies . Understanding these processes has boosted progress in ...
The paper uses the continuous Game of Life as a testing ground for algorithms that discover diverse behaviors. The problem is interesting, under-explored, and rich. The combines a variety of interesting ideas including compositional pattern producing networks (CPPNs) to learn structured primitives. Although the author...
SP:f0962d6f8e5bc3a7dc5488d4d0bd4731fdd63756
Intrinsically Motivated Discovery of Diverse Patterns in Self-Organizing Systems
1 INTRODUCTION . Self-organization of patterns that emerge from local rules is a pervasive phenomena in natural and artificial dynamical systems ( Ball , 1999 ) . It ranges from the formation of snow flakes , spots and rays on animal ’ s skin , to spiral galaxies . Understanding these processes has boosted progress in ...
The focus of the presented paper is on formulating the automated discovery of self-organized patterns in high-dimensional dynamic systems. The introduced framework uses cellular automata (game of life) as a testbed for experimentation and evaluation and existing machine learning algorithms (POP-IMGEPs). The goal of the...
SP:f0962d6f8e5bc3a7dc5488d4d0bd4731fdd63756
Few-shot Learning by Focusing on Differences
1 INTRODUCTION . Progress in artificial intelligence ( AI ) has been rapid . AI agents have been outperforming humans in an increasing variety of tasks , such as in recognizing images on ImageNet ( He et al. , 2016 ) and in the ancient game of Go ( Silver et al. , 2016 ) . However , challenges remain – systems that out...
In this paper the authors propose a metric based model for few-shot learning. The goal of the proposed technique is to incorporate a prior that highlight better the dissimilarity between closely related class prototype. Thus, the proposed paper is related to prototypical neural network (use of prototype to represent a ...
SP:8861e607941d5f65eae84cb2f2ac04066254b18a
Few-shot Learning by Focusing on Differences
1 INTRODUCTION . Progress in artificial intelligence ( AI ) has been rapid . AI agents have been outperforming humans in an increasing variety of tasks , such as in recognizing images on ImageNet ( He et al. , 2016 ) and in the ancient game of Go ( Silver et al. , 2016 ) . However , challenges remain – systems that out...
The stated contributions of the paper are: (1) a method for performing few-shot learning and (2) an approach for building harder few-shot learning datasets from existing datasets. The authors describe a model for creating a task-aware embedding for different novel sets (for different image classification settings) usin...
SP:8861e607941d5f65eae84cb2f2ac04066254b18a
Meta-Learning without Memorization
1 INTRODUCTION . The ability to learn new concepts and skills with small amounts of data is a critical aspect of intelligence that many machine learning systems lack . Meta-learning ( Schmidhuber , 1987 ) has emerged as a promising approach for enabling systems to quickly learn new tasks by building upon experience fro...
This paper analyses a pitfall of current meta-learning algorithms, where the task can be inferred from the meta-training data alone, leaving the task-training data unused. Such a meta-learner would generalise well on the meta-training tasks, but will fail to generalise on new tasks at test time. This kind of overfittin...
SP:3350fe8ea74f715832130fe2c8a5309b721aa24b
Meta-Learning without Memorization
1 INTRODUCTION . The ability to learn new concepts and skills with small amounts of data is a critical aspect of intelligence that many machine learning systems lack . Meta-learning ( Schmidhuber , 1987 ) has emerged as a promising approach for enabling systems to quickly learn new tasks by building upon experience fro...
This paper illustrates, identifies, and formally defines a memorization problem in meta-learning -- the model can simply memorize meta-training tasks and ignore meta-training train sets. The paper proposes to optimize the mutual information between testing predictions and the training data (given input and meta model),...
SP:3350fe8ea74f715832130fe2c8a5309b721aa24b
Biologically Plausible Neural Networks via Evolutionary Dynamics and Dopaminergic Plasticity
1 INTRODUCTION . In his Turing award lecture , neural networks pioneer Geoff Hinton opined that “ evolution can ’ t get gradients because a lot of what determines the relationship between the genotype and the phenotype is outside your control ” ( Hinton , 2019 ) . We beg to differ . Evolution does have what amounts to ...
In this paper the authors propose a method for training neural networks using evolutionary methods. The aim of developing this method is to provide a biological alternative to back-propagation. The authors prove that their method converges and with high probability succeeds in learning linear classification problems. A...
SP:a9c2860abb6a9df585aecea0dfb9a833458f184f
Biologically Plausible Neural Networks via Evolutionary Dynamics and Dopaminergic Plasticity
1 INTRODUCTION . In his Turing award lecture , neural networks pioneer Geoff Hinton opined that “ evolution can ’ t get gradients because a lot of what determines the relationship between the genotype and the phenotype is outside your control ” ( Hinton , 2019 ) . We beg to differ . Evolution does have what amounts to ...
This paper argues that Artificial Neural Network (ANN) lack in biological plausibility because of the back-propagation process. Therefore, the authors provide an alternative approach, named neural net evolution (NNE) that follows evolutionary theory. This approach uses a large number of genotypes (in the form of vector...
SP:a9c2860abb6a9df585aecea0dfb9a833458f184f
Geometry-aware Generation of Adversarial and Cooperative Point Clouds
1 INTRODUCTION . The existence of adversarial examples shows the vulnerability of machine learning models , and triggers a great amount of research attention paid to either attacking and defense studies for safetycritical issues , or robustness analysis on machine learning models themselves . In existing literature , a...
This paper proposes a novel loss function to account for imperceptible, geometry-aware deformations of point clouds. The loss is used in two cases: generating adversarial point clouds to attack representative models of point set classifiers, and generating cooperative point clouds to improve classification confidence o...
SP:823860fad022b8abde97b597f8b3881453489dc1
Geometry-aware Generation of Adversarial and Cooperative Point Clouds
1 INTRODUCTION . The existence of adversarial examples shows the vulnerability of machine learning models , and triggers a great amount of research attention paid to either attacking and defense studies for safetycritical issues , or robustness analysis on machine learning models themselves . In existing literature , a...
This paper describes a new targeted adversarial attack against 3D point cloud object classifiers that is robust to several countermeasures. The attack finds a point of the target class that is close to the original point cloud in terms of a more complicated metric that combines the Hausdorff distance, the Chamfer dista...
SP:823860fad022b8abde97b597f8b3881453489dc1
GraphNVP: an Invertible Flow-based Model for Generating Molecular Graphs
1 INTRODUCTION . Generation of molecules with certain desirable properties is a crucial problem in computational drug discovery . Recently , deep learning approaches are being actively studied for generating promising candidate molecules quickly . Earlier models ( Kusner et al. , 2017 ; Gómez-Bombarelli et al. , 2018 )...
This paper presents a new reversible flow-based graph generative model wherein the whole graph i.e., representative attributes such as node features and adjacency tensor is modeled using seperate streams of invertible flow model. This allows training of generative model using exact likelihood maximization over the unde...
SP:f447505525e50eabf0dd4dd7a09ce763f16a233c
GraphNVP: an Invertible Flow-based Model for Generating Molecular Graphs
1 INTRODUCTION . Generation of molecules with certain desirable properties is a crucial problem in computational drug discovery . Recently , deep learning approaches are being actively studied for generating promising candidate molecules quickly . Earlier models ( Kusner et al. , 2017 ; Gómez-Bombarelli et al. , 2018 )...
In this paper, a GraphNVP framework for molecular graph generation is proposed. The main difference from the previously proposed models is the use of the invertible normalizing flow idea for the generative model, which doesn’t require a separate decoder for sampling. This architecture is implemented with coupling layer...
SP:f447505525e50eabf0dd4dd7a09ce763f16a233c
A Mean-Field Theory for Kernel Alignment with Random Features in Generative Adverserial Networks
1 INTRODUCTION . A fundamental and long-standing problem in unsupervised learning systems is to capture the underlying distribution of data . While deep generative models such as Boltzmann machines Salakhutdinov & Hinton ( 2009 ) and auto-encoding variational Bayes Kingma & Welling ( 2013 ) accomplish this task to some...
This paper proposes to learn a kernel for training MMD-GAN by optimizing over the probability distribution that defines the kernel by means of random features. This is unlike the usual setting of MMD-GAN where the kernel is parametrized by composing a fixed top-kernel with a discriminator network that is optimized duri...
SP:e0dd4a62106a2c2fc6a248c601ddb8422e148864
A Mean-Field Theory for Kernel Alignment with Random Features in Generative Adverserial Networks
1 INTRODUCTION . A fundamental and long-standing problem in unsupervised learning systems is to capture the underlying distribution of data . While deep generative models such as Boltzmann machines Salakhutdinov & Hinton ( 2009 ) and auto-encoding variational Bayes Kingma & Welling ( 2013 ) accomplish this task to some...
This paper aims to improve the kernel selection issue of the MMD-based generative models. The author formulates the kernels via inverse Fourier transform and the goal is to learn the optimal N finite random Fourier features (RFF). The RFF samples are optimized by the proposed kernel alignment loss where the positive an...
SP:e0dd4a62106a2c2fc6a248c601ddb8422e148864
A Constructive Prediction of the Generalization Error Across Scales
1 INTRODUCTION . With the success and heightened adoption of neural networks for real world tasks , some questions remain poorly answered . For a given task and model architecture , how much data would one require to reach a prescribed performance level ? How big a model would be needed ? Addressing such questions is m...
This paper proposes a functional form to model the dependence of generalization error on a held-out test set on model and dataset size. The functional form is derived based on empirical observations of the generalizing error for various model and dataset sizes (sections O1, O2, and O3) and on certain necessary criteria...
SP:32d80c08e2e1a76e06e701d537264421493db122
A Constructive Prediction of the Generalization Error Across Scales
1 INTRODUCTION . With the success and heightened adoption of neural networks for real world tasks , some questions remain poorly answered . For a given task and model architecture , how much data would one require to reach a prescribed performance level ? How big a model would be needed ? Addressing such questions is m...
This work proposes a functional form for the relationship between <dataset size, model size> and generalization error, and performs an empirical study to validate it. First, it states 5 criteria that such a functional form must take, and proposes one such functional form containing 6 free coefficients that satisfy all ...
SP:32d80c08e2e1a76e06e701d537264421493db122
Rethinking the Hyperparameters for Fine-tuning
1 INTRODUCTION . Many real-world applications often have a limited number of training instances , which makes directly training deep neural networks hard and prone to overfitting . Transfer learning with the knowledge of models learned on a similar task can help to avoid overfitting . Fine-tuning is a simple and effect...
This submission studies the problem of transfer learning and fine tuning. This submission proposes four insights: Momentum hyperparameters are essential for fine-tuning; When the hyperparameters satisfy some certain relationships, the results of fine-tuning are optimal; The similarity between source and target datasets...
SP:f08e59bd838b72a61a2ddcbd9027df8bca75ccea
Rethinking the Hyperparameters for Fine-tuning
1 INTRODUCTION . Many real-world applications often have a limited number of training instances , which makes directly training deep neural networks hard and prone to overfitting . Transfer learning with the knowledge of models learned on a similar task can help to avoid overfitting . Fine-tuning is a simple and effect...
This paper studies the role of different hyperparameters in finetuning image recognition models on new target tasks. The authors run a large set of experiments and show that, perhaps non-surprisingly, hyperparameters matter. In particular, they show that momentum, which is typically ignored in finetuning, is quite impo...
SP:f08e59bd838b72a61a2ddcbd9027df8bca75ccea
Convolutional Bipartite Attractor Networks
1 INTRODUCTION . Under ordinary conditions , human visual perception is quick and accurate . Studying circumstances that give rise to slow or inaccurate perception can help reveal the underlying mechanisms of visual information processing . Recent investigations of occluded ( Tang et al. , 2018 ) and empirically challe...
The paper argues for the use of attractive networks (AN) for the tasks that involve learning from noisy data. Attractor networks are recurrent in nature and use energy minimization dynamics. As motivation, the authors point to studies that give evidence for the usefulness of recurrence for visual tasks. The experiments...
SP:e2bc61c78d53d0b72fcc5cde34368e88290371b6
Convolutional Bipartite Attractor Networks
1 INTRODUCTION . Under ordinary conditions , human visual perception is quick and accurate . Studying circumstances that give rise to slow or inaccurate perception can help reveal the underlying mechanisms of visual information processing . Recent investigations of occluded ( Tang et al. , 2018 ) and empirically challe...
This paper presents an attractor network (AN) approach for pattern interpretation and completion. The authors propose a convolutional bipartite architecture consisting of visible (input and output) and hidden layers with weight constraints and squared and energy-based losses. To prevent vanishing/exploding gradients, t...
SP:e2bc61c78d53d0b72fcc5cde34368e88290371b6
Piecewise linear activations substantially shape the loss surfaces of neural networks
1 INTRODUCTION . Neural networks have been successfully deployed in many real-world applications ( LeCun et al. , 2015 ; Witten et al. , 2016 ; Silver et al. , 2016 ; He et al. , 2016 ; Litjens et al. , 2017 ) . In spite of this , the theoretical foundations of neural networks are somewhat premature . To the many defic...
This paper studies the theoretical property of neural network's loss surface. The main contribution is to prove that the loss surface of every neural network (with arbitrary depth) with piecewise linear activations has infinite spurious local minima. Moreover, the paper further characterizes the partition of the local ...
SP:e963ad4e47263da9f64c76505e1853cbf8b012c4
Piecewise linear activations substantially shape the loss surfaces of neural networks
1 INTRODUCTION . Neural networks have been successfully deployed in many real-world applications ( LeCun et al. , 2015 ; Witten et al. , 2016 ; Silver et al. , 2016 ; He et al. , 2016 ; Litjens et al. , 2017 ) . In spite of this , the theoretical foundations of neural networks are somewhat premature . To the many defic...
This paper studies the landscape of deep neural networks with piecewise-linear activation functions. The paper showed that under very mild assumptions, the loss surface admits infinite spurious local minima. Further, it is shown that the loss surface is partitioned into many multilinear cells. If the network is two-lay...
SP:e963ad4e47263da9f64c76505e1853cbf8b012c4
Manifold Modeling in Embedded Space: A Perspective for Interpreting "Deep Image Prior"
1 INTRODUCTION . The most important piece of information for image/tensor restoration would be the “ prior ” which usually converts the optimization problems from ill-posed to well-posed , and/or gives some robustness for specific noises and outliers . Many priors were studied in computer science problems such as low-r...
In this paper, the authors present a natural image model based on the manifold of image patches. It is similar to the Deep Image Prior in that it is untrained and has a convolutional-like structure. It leads to an optimization problem with a reconstruction loss term and an auto encoding term. The authors show empiri...
SP:cf11852f87d71e71dc4e5327eef4236db46fe1d5
Manifold Modeling in Embedded Space: A Perspective for Interpreting "Deep Image Prior"
1 INTRODUCTION . The most important piece of information for image/tensor restoration would be the “ prior ” which usually converts the optimization problems from ill-posed to well-posed , and/or gives some robustness for specific noises and outliers . Many priors were studied in computer science problems such as low-r...
This paper introduces a transformation from the deep image prior (DIP) to an embedding with an autoencoder (MMES). The authors aim to use this transformation to explain ("in words") why the DIP works so well and explain why convolutions are needed in the DIP. The contributions are summarised as a) providing an interpre...
SP:cf11852f87d71e71dc4e5327eef4236db46fe1d5
Evidence-Aware Entropy Decomposition For Active Deep Learning
1 INTRODUCTION . Deep learning ( DL ) models establish dominating status among other types of supervised learning models by achieving the state-of-the-art performance in various application domains . However , such an advantage only emerges when a huge amount of labeled training data is available . This limitation slow...
This paper propose an active deep learning model. By leveraging subjective Logic, they propose to decompose the entropy of a predicted class distribution into vacuity (lack of evidence) and dissonance (conflict of strong evidence). Instead of using the predicted class distribution, they estimate the supporting evidence...
SP:45bf7ca342ad1752c7f7c056653137b9283c487f
Evidence-Aware Entropy Decomposition For Active Deep Learning
1 INTRODUCTION . Deep learning ( DL ) models establish dominating status among other types of supervised learning models by achieving the state-of-the-art performance in various application domains . However , such an advantage only emerges when a huge amount of labeled training data is available . This limitation slow...
The authors consider active deep learning. They propose decomposing predictive entropy into a) vacuity (lack of evidence) and b) dissonance (contradictory evidence). They frame this in terms of "subjective logic". In practice this is achieved by having the NN output the parameters of a Dirichlet, which allows an additi...
SP:45bf7ca342ad1752c7f7c056653137b9283c487f