Datasets:

dim

/

scitldr

like 0

A disentangled representation of a data set should be capable of recovering the underlying factors that generated it. One question that arises is whether using Euclidean space for latent variable models can produce a disentangled representation when the underlying generating factors have a certain geometrical structure. Take for example the images of a car seen from different angles. The angle has a periodic structure but a 1-dimensional representation would fail to capture this topology. How can we address this problem? The submissions presented for the first stage of the NeurIPS2019 Disentanglement Challenge consist of a Diffusion Variational Autoencoder ($\Delta$VAE) with a hyperspherical latent space which can for example recover periodic true factors. The training of the $\Delta$VAE is enhanced by incorporating a modified version of the Evidence Lower Bound (ELBO) for tailoring the encoding capacity of the posterior approximate.

Description of submission to NeurIPS2019 Disentanglement Challenge based on hyperspherical variational autoencoders

Anomaly detection, finding patterns that substantially deviate from those seen previously, is one of the fundamental problems of artificial intelligence. Recently, classification-based methods were shown to achieve superior results on this task. In this work, we present a unifying view and propose an open-set method to relax current generalization assumptions. Furthermore, we extend the applicability of transformation-based methods to non-image data using random affine transformations. Our method is shown to obtain state-of-the-art accuracy and is applicable to broad data types. The strong performance of our method is extensively validated on multiple datasets from different domains.

An anomaly detection that: uses random-transformation classification for generalizing to non-image data.

Recent improvements in large-scale language models have driven progress on automatic generation of syntactically and semantically consistent text for many real-world applications. Many of these advances leverage the availability of large corpora. While training on such corpora encourages the model to understand long-range dependencies in text, it can also result in the models internalizing the social biases present in the corpora. This paper aims to quantify and reduce biases exhibited by language models. Given a conditioning context (e.g. a writing prompt) and a language model, we analyze if (and how) the sentiment of the generated text is affected by changes in values of sensitive attributes (e.g. country names, occupations, genders, etc.) in the conditioning context, a.k.a. counterfactual evaluation. We quantify these biases by adapting individual and group fairness metrics from the fair machine learning literature. Extensive evaluation on two different corpora (news articles and Wikipedia) shows that state-of-the-art Transformer-based language models exhibit biases learned from data. We propose embedding-similarity and sentiment-similarity regularization methods that improve both individual and group fairness metrics without sacrificing perplexity and semantic similarity---a positive step toward development and deployment of fairer language models for real-world applications.

We reduce sentiment biases based on counterfactual evaluation of text generation using language models.

Topic modeling of text documents is one of the most important tasks in representation learning. In this work, we propose iTM-VAE, which is a Bayesian nonparametric (BNP) topic model with variational auto-encoders. On one hand, as a BNP topic model, iTM-VAE potentially has infinite topics and can adapt the topic number to data automatically. On the other hand, different with the other BNP topic models, the inference of iTM-VAE is modeled by neural networks, which has rich representation capacity and can be computed in a simple feed-forward manner. Two variants of iTM-VAE are also proposed in this paper, where iTM-VAE-Prod models the generative process in products-of-experts fashion for better performance and iTM-VAE-G places a prior over the concentration parameter such that the model can adapt a suitable concentration parameter to data automatically. Experimental results on 20News and Reuters RCV1-V2 datasets show that the proposed models outperform the state-of-the-arts in terms of perplexity, topic coherence and document retrieval tasks. Moreover, the ability of adjusting the concentration parameter to data is also confirmed by experiments.

A Bayesian Nonparametric Topic Model with Variational Auto-Encoders which achieves the state-of-the-arts on public benchmarks in terms of perplexity, topic coherence and retrieval tasks.

Knowledge Distillation (KD) is a widely used technique in recent deep learning research to obtain small and simple models whose performance is on a par with their large and complex counterparts. Standard Knowledge Distillation tends to be time-consuming because of the training time spent to obtain a teacher model that would then provide guidance for the student model. It might be possible to cut short the time by training a teacher model on the fly, but it is not trivial to have such a high-capacity teacher that gives quality guidance to student models this way. To improve this, we present a novel framework of Knowledge Distillation exploiting dark knowledge from the whole training set. In this framework, we propose a simple and effective implementation named Distillation by Utilizing Peer Samples (DUPS) in one generation. We verify our algorithm on numerous experiments. Compared with standard training on modern architectures, DUPS achieves an average improvement of 1%-2% on various tasks with nearly zero extra cost. Considering some typical Knowledge Distillation methods which are much more time-consuming, we also get comparable or even better performance using DUPS.

We present a novel framework of Knowledge Distillation utilizing peer samples as the teacher

We develop a metalearning approach for learning hierarchically structured poli- cies, improving sample efficiency on unseen tasks through the use of shared primitives—policies that are executed for large numbers of timesteps. Specifi- cally, a set of primitives are shared within a distribution of tasks, and are switched between by task-specific policies. We provide a concrete metric for measuring the strength of such hierarchies, leading to an optimization problem for quickly reaching high reward on unseen tasks. We then present an algorithm to solve this problem end-to-end through the use of any off-the-shelf reinforcement learning method, by repeatedly sampling new tasks and resetting task-specific policies. We successfully discover meaningful motor primitives for the directional movement of four-legged robots, solely by interacting with distributions of mazes. We also demonstrate the transferability of primitives to solve long-timescale sparse-reward obstacle courses, and we enable 3D humanoid robots to robustly walk and crawl with the same policy.

learn hierarchal sub-policies through end-to-end training over a distribution of tasks

This paper proposes a new model for document embedding. Existing approaches either require complex inference or use recurrent neural networks that are difficult to parallelize. We take a different route and use recent advances in language modeling to develop a convolutional neural network embedding model. This allows us to train deeper architectures that are fully parallelizable. Stacking layers together increases the receptive filed allowing each successive layer to model increasingly longer range semantic dependences within the document. Empirically we demonstrate superior results on two publicly available benchmarks. Full code will be released with the final version of this paper.

Convolutional neural network model for unsupervised document embedding.

We prove bounds on the generalization error of convolutional networks. The bounds are in terms of the training loss, the number of parameters, the Lipschitz constant of the loss and the distance from the weights to the initial weights. They are independent of the number of pixels in the input, and the height and width of hidden feature maps. We present experiments with CIFAR-10, along with varying hyperparameters of a deep convolutional network, comparing our bounds with practical generalization gaps.

We prove generalization bounds for convolutional neural networks that take account of weight-tying

MobileNets family of computer vision neural networks have fueled tremendous progress in the design and organization of resource-efficient architectures in recent years. New applications with stringent real-time requirements in highly constrained devices require further compression of MobileNets-like already computeefficient networks. Model quantization is a widely used technique to compress and accelerate neural network inference and prior works have quantized MobileNets to 4 − 6 bits albeit with a modest to significant drop in accuracy. While quantization to sub-byte values (i.e. precision ≤ 8 bits) has been valuable, even further quantization of MobileNets to binary or ternary values is necessary to realize significant energy savings and possibly runtime speedups on specialized hardware, such as ASICs and FPGAs. Under the key observation that convolutional filters at each layer of a deep neural network may respond differently to ternary quantization, we propose a novel quantization method that generates per-layer hybrid filter banks consisting of full-precision and ternary weight filters for MobileNets. The layer-wise hybrid filter banks essentially combine the strengths of full-precision and ternary weight filters to derive a compact, energy-efficient architecture for MobileNets. Using this proposed quantization method, we quantized a substantial portion of weight filters of MobileNets to ternary values resulting in 27.98% savings in energy, and a 51.07% reduction in the model size, while achieving comparable accuracy and no degradation in throughput on specialized hardware in comparison to the baseline full-precision MobileNets.

2x savings in model size, 28% energy reduction for MobileNets on ImageNet at no loss in accuracy using hybrid layers composed of conventional full-precision filters and ternary filters

Performing controlled experiments on noisy data is essential in thoroughly understanding deep learning across a spectrum of noise levels. Due to the lack of suitable datasets, previous research have only examined deep learning on controlled synthetic noise, and real-world noise has never been systematically studied in a controlled setting. To this end, this paper establishes a benchmark of real-world noisy labels at 10 controlled noise levels. As real-world noise possesses unique properties, to understand the difference, we conduct a large-scale study across a variety of noise levels and types, architectures, methods, and training settings. Our study shows that: (1) Deep Neural Networks (DNNs) generalize much better on real-world noise. (2) DNNs may not learn patterns first on real-world noisy data. (3) When networks are fine-tuned, ImageNet architectures generalize well on noisy data. (4) Real-world noise appears to be less harmful, yet it is more difficult for robust DNN methods to improve. (5) Robust learning methods that work well on synthetic noise may not work as well on real-world noise, and vice versa. We hope our benchmark, as well as our findings, will facilitate deep learning research on noisy data.

We establish a benchmark of controlled real noise and reveal several interesting findings about real-world noisy data.

Designing RNA molecules has garnered recent interest in medicine, synthetic biology, biotechnology and bioinformatics since many functional RNA molecules were shown to be involved in regulatory processes for transcription, epigenetics and translation. Since an RNA's function depends on its structural properties, the RNA Design problem is to find an RNA sequence which satisfies given structural constraints. Here, we propose a new algorithm for the RNA Design problem, dubbed LEARNA. LEARNA uses deep reinforcement learning to train a policy network to sequentially design an entire RNA sequence given a specified target structure. By meta-learning across 65000 different RNA Design tasks for one hour on 20 CPU cores, our extension Meta-LEARNA constructs an RNA Design policy that can be applied out of the box to solve novel RNA Design tasks. Methodologically, for what we believe to be the first time, we jointly optimize over a rich space of architectures for the policy network, the hyperparameters of the training procedure and the formulation of the decision process. Comprehensive empirical results on two widely-used RNA Design benchmarks, as well as a third one that we introduce, show that our approach achieves new state-of-the-art performance on the former while also being orders of magnitudes faster in reaching the previous state-of-the-art performance. In an ablation study, we analyze the importance of our method's different components.

We learn to solve the RNA Design problem with reinforcement learning using meta learning and autoML approaches.

Pruning is a popular technique for compressing a neural network: a large pre-trained network is fine-tuned while connections are successively removed. However, the value of pruning has largely evaded scrutiny. In this extended abstract, we examine residual networks obtained through Fisher-pruning and make two interesting observations. First, when time-constrained, it is better to train a simple, smaller network from scratch than prune a large network. Second, it is the architectures obtained through the pruning process --- not the learnt weights --- that prove valuable. Such architectures are powerful when trained from scratch. Furthermore, these architectures are easy to approximate without any further pruning: we can prune once and obtain a family of new, scalable network architectures for different memory requirements.

Training small networks beats pruning, but pruning finds good small networks to train that are easy to copy.

Supervised learning problems---particularly those involving social data---are often subjective. That is, human readers, looking at the same data, might come to legitimate but completely different conclusions based on their personal experiences. Yet in machine learning settings feedback from multiple human annotators is often reduced to a single ``ground truth'' label, thus hiding the true, potentially rich and diverse interpretations of the data found across the social spectrum. We explore the rewards and challenges of discovering and learning representative distributions of the labeling opinions of a large human population. A major, critical cost to this approach is the number of humans needed to provide enough labels not only to obtain representative samples but also to train a machine to predict representative distributions on unlabeled data. We propose aggregating label distributions over, not just individuals, but also data items, in order to maximize the costs of humans in the loop. We test different aggregation approaches on state-of-the-art deep learning models. Our results suggest that careful label aggregation methods can greatly reduce the number of samples needed to obtain representative distributions.

We study the problem of learning to predict the underlying diversity of beliefs present in supervised learning domains.

Recent advancements in deep learning techniques such as Convolutional Neural Networks(CNN) and Generative Adversarial Networks(GAN) have achieved breakthroughs in the problem of semantic image inpainting, the task of reconstructing missing pixels in given images. While much more effective than conventional approaches, deep learning models require large datasets and great computational resources for training, and inpainting quality varies considerably when training data vary in size and diversity. To address these problems, we present in this paper a inpainting strategy of \textit{Comparative Sample Augmentation}, which enhances the quality of training set by filtering out irrelevant images and constructing additional images using information about the surrounding regions of the images to be inpainted. Experiments on multiple datasets demonstrate that our method extends the applicability of deep inpainting models to training sets with varying sizes, while maintaining inpainting quality as measured by qualitative and quantitative metrics for a large class of deep models, with little need for model-specific consideration.

We introduced a strategy which enables inpainting models on datasets of various sizes

Generative adversarial networks (GANs) are a family of generative models that do not minimize a single training criterion. Unlike other generative models, the data distribution is learned via a game between a generator (the generative model) and a discriminator (a teacher providing training signal) that each minimize their own cost. GANs are designed to reach a Nash equilibrium at which each player cannot reduce their cost without changing the other players’ parameters. One useful approach for the theory of GANs is to show that a divergence between the training distribution and the model distribution obtains its minimum value at equilibrium. Several recent research directions have been motivated by the idea that this divergence is the primary guide for the learning process and that every step of learning should decrease the divergence. We show that this view is overly restrictive. During GAN training, the discriminator provides learning signal in situations where the gradients of the divergences between distributions would not be useful. We provide empirical counterexamples to the view of GAN training as divergence minimization. Specifically, we demonstrate that GANs are able to learn distributions in situations where the divergence minimization point of view predicts they would fail. We also show that gradient penalties motivated from the divergence minimization perspective are equally helpful when applied in other contexts in which the divergence minimization perspective does not predict they would be helpful. This contributes to a growing body of evidence that GAN training may be more usefully viewed as approaching Nash equilibria via trajectories that do not necessarily minimize a specific divergence at each step.

We find evidence that divergence minimization may not be an accurate characterization of GAN training.

Measuring Mutual Information (MI) between high-dimensional, continuous, random variables from observed samples has wide theoretical and practical applications. Recent works have developed accurate MI estimators through provably low-bias approximations and tight variational lower bounds assuming abundant supply of samples, but require an unrealistic number of samples to guarantee statistical significance of the estimation. In this work, we focus on improving data efficiency and propose a Data-Efficient MINE Estimator (DEMINE) that can provide a tight lower confident interval of MI under limited data, through adding cross-validation to the MINE lower bound (Belghazi et al., 2018). Hyperparameter search is employed and a novel meta-learning approach with task augmentation is developed to increase robustness to hyperparamters, reduce overfitting and improve accuracy. With improved data-efficiency, our DEMINE estimator enables statistical testing of dependency at practical dataset sizes. We demonstrate the effectiveness of DEMINE on synthetic benchmarks and a real world fMRI dataset, with application of inter-subject correlation analysis.

A new & practical statistical test of dependency using neural networks, benchmarked on synthetic and a real fMRI datasets.

Language and vision are processed as two different modal in current work for image captioning. However, recent work on Super Characters method shows the effectiveness of two-dimensional word embedding, which converts text classification problem into image classification problem. In this paper, we propose the SuperCaptioning method, which borrows the idea of two-dimensional word embedding from Super Characters method, and processes the information of language and vision together in one single CNN model. The experimental results on Flickr30k data shows the proposed method gives high quality image captions. An interactive demo is ready to show at the workshop.

Image captioning using two-dimensional word embedding.

Determining the optimal order in which data examples are presented to Deep Neural Networks during training is a non-trivial problem. However, choosing a non-trivial scheduling method may drastically improve convergence. In this paper, we propose a Self-Paced Learning (SPL)-fused Deep Metric Learning (DML) framework, which we call Learning Embeddings for Adaptive Pace (LEAP). Our method parameterizes mini-batches dynamically based on the \textit{easiness} and \textit{true diverseness} of the sample within a salient feature representation space. In LEAP, we train an \textit{embedding} Convolutional Neural Network (CNN) to learn an expressive representation space by adaptive density discrimination using the Magnet Loss. The \textit{student} CNN classifier dynamically selects samples to form a mini-batch based on the \textit{easiness} from cross-entropy losses and \textit{true diverseness} of examples from the representation space sculpted by the \textit{embedding} CNN. We evaluate LEAP using deep CNN architectures for the task of supervised image classification on MNIST, FashionMNIST, CIFAR-10, CIFAR-100, and SVHN. We show that the LEAP framework converges faster with respect to the number of mini-batch updates required to achieve a comparable or better test performance on each of the datasets.

LEAP combines the strength of adaptive sampling with that of mini-batch online learning and adaptive representation learning to formulate a representative self-paced strategy in an end-to-end DNN training protocol.

Conventional deep reinforcement learning typically determines an appropriate primitive action at each timestep, which requires enormous amount of time and effort for learning an effective policy, especially in large and complex environments. To deal with the issue fundamentally, we incorporate macro actions, defined as sequences of primitive actions, into the primitive action space to form an augmented action space. The problem lies in how to find an appropriate macro action to augment the primitive action space. The agent using a proper augmented action space is able to jump to a farther state and thus speed up the exploration process as well as facilitate the learning procedure. In previous researches, macro actions are developed by mining the most frequently used action sequences or repeating previous actions. However, the most frequently used action sequences are extracted from a past policy, which may only reinforce the original behavior of that policy. On the other hand, repeating actions may limit the diversity of behaviors of the agent. Instead, we propose to construct macro actions by a genetic algorithm, which eliminates the dependency of the macro action derivation procedure from the past policies of the agent. Our approach appends a macro action to the primitive action space once at a time and evaluates whether the augmented action space leads to promising performance or not. We perform extensive experiments and show that the constructed macro actions are able to speed up the learning process for a variety of deep reinforcement learning methods. Our experimental results also demonstrate that the macro actions suggested by our approach are transferable among deep reinforcement learning methods and similar environments. We further provide a comprehensive set of ablation analysis to validate our methodology.

We propose to construct macro actions by a genetic algorithm, which eliminates the dependency of the macro action derivation procedure from the past policies of the agent.

A key problem in neuroscience and life sciences more generally is that the data generation process is often best thought of as a hierarchy of dynamic systems. One example of this is in-vivo calcium imaging data, where observed calcium transients are driven by a combination of electro-chemical kinetics where hypothesized trajectories around manifolds determining the frequency of these transients. A recent approach using sequential variational auto-encoders demonstrated it was possible to learn the latent dynamic structure of reaching behaviour from spiking data modelled as a Poisson process. Here we extend this approach using a ladder method to infer the spiking events driving calcium transients along with the deeper latent dynamic system. We show strong performance of this approach on a benchmark synthetic dataset against a number of alternatives.

We propose an extension to LFADS capable of inferring spike trains to reconstruct calcium fluorescence traces using hierarchical VAEs.

In spite of the recent success of neural machine translation (NMT) in standard benchmarks, the lack of large parallel corpora poses a major practical problem for many language pairs. There have been several proposals to alleviate this issue with, for instance, triangulation and semi-supervised learning techniques, but they still require a strong cross-lingual signal. In this work, we completely remove the need of parallel data and propose a novel method to train an NMT system in a completely unsupervised manner, relying on nothing but monolingual corpora. Our model builds upon the recent work on unsupervised embedding mappings, and consists of a slightly modified attentional encoder-decoder model that can be trained on monolingual corpora alone using a combination of denoising and backtranslation. Despite the simplicity of the approach, our system obtains 15.56 and 10.21 BLEU points in WMT 2014 French-to-English and German-to-English translation. The model can also profit from small parallel corpora, and attains 21.81 and 15.24 points when combined with 100,000 parallel sentences, respectively. Our implementation is released as an open source project.

We introduce the first successful method to train neural machine translation in an unsupervised manner, using nothing but monolingual corpora

We describe a new training methodology for generative adversarial networks. The key idea is to grow both the generator and discriminator progressively: starting from a low resolution, we add new layers that model increasingly fine details as training progresses. This both speeds the training up and greatly stabilizes it, allowing us to produce images of unprecedented quality, e.g., CelebA images at 1024^2. We also propose a simple way to increase the variation in generated images, and achieve a record inception score of 8.80 in unsupervised CIFAR10. Additionally, we describe several implementation details that are important for discouraging unhealthy competition between the generator and discriminator. Finally, we suggest a new metric for evaluating GAN results, both in terms of image quality and variation. As an additional contribution, we construct a higher-quality version of the CelebA dataset.

We train generative adversarial networks in a progressive fashion, enabling us to generate high-resolution images with high quality.

Designing a convolution for a spherical neural network requires a delicate tradeoff between efficiency and rotation equivariance. DeepSphere, a method based on a graph representation of the discretized sphere, strikes a controllable balance between these two desiderata. This contribution is twofold. First, we study both theoretically and empirically how equivariance is affected by the underlying graph with respect to the number of pixels and neighbors. Second, we evaluate DeepSphere on relevant problems. Experiments show state-of-the-art performance and demonstrates the efficiency and flexibility of this formulation. Perhaps surprisingly, comparison with previous work suggests that anisotropic filters might be an unnecessary price to pay.

A graph-based spherical CNN that strikes an interesting balance of trade-offs for a wide variety of applications.

The notion of the stationary equilibrium ensemble has played a central role in statistical mechanics. In machine learning as well, training serves as generalized equilibration that drives the probability distribution of model parameters toward stationarity. Here, we derive stationary fluctuation-dissipation relations that link measurable quantities and hyperparameters in the stochastic gradient descent algorithm. These relations hold exactly for any stationary state and can in particular be used to adaptively set training schedule. We can further use the relations to efficiently extract information pertaining to a loss-function landscape such as the magnitudes of its Hessian and anharmonicity. Our claims are empirically verified.

We prove fluctuation-dissipation relations for SGD, which can be used to (i) adaptively set learning rates and (ii) probe loss surfaces.

Recurrent neural networks (RNNs) are difficult to train on sequence processing tasks, not only because input noise may be amplified through feedback, but also because any inaccuracy in the weights has similar consequences as input noise. We describe a method for denoising the hidden state during training to achieve more robust representations thereby improving generalization performance. Attractor dynamics are incorporated into the hidden state to `clean up' representations at each step of a sequence. The attractor dynamics are trained through an auxillary denoising loss to recover previously experienced hidden states from noisy versions of those states. This state-denoised recurrent neural network (SDRNN) performs multiple steps of internal processing for each external sequence step. On a range of tasks, we show that the SDRNN outperforms a generic RNN as well as a variant of the SDRNN with attractor dynamics on the hidden state but without the auxillary loss. We argue that attractor dynamics---and corresponding connectivity constraints---are an essential component of the deep learning arsenal and should be invoked not only for recurrent networks but also for improving deep feedforward nets and intertask transfer.

We propose a mechanism for denoising the internal state of an RNN to improve generalization performance.

We consider reinforcement learning in input-driven environments, where an exogenous, stochastic input process affects the dynamics of the system. Input processes arise in many applications, including queuing systems, robotics control with disturbances, and object tracking. Since the state dynamics and rewards depend on the input process, the state alone provides limited information for the expected future returns. Therefore, policy gradient methods with standard state-dependent baselines suffer high variance during training. We derive a bias-free, input-dependent baseline to reduce this variance, and analytically show its benefits over state-dependent baselines. We then propose a meta-learning approach to overcome the complexity of learning a baseline that depends on a long sequence of inputs. Our experimental results show that across environments from queuing systems, computer networks, and MuJoCo robotic locomotion, input-dependent baselines consistently improve training stability and result in better eventual policies.

For environments dictated partially by external input processes, we derive an input-dependent baseline that provably reduces the variance for policy gradient methods and improves the policy performance in a wide range of RL tasks.

Deep networks have shown great performance in classification tasks. However, the parameters learned by the classifier networks usually discard stylistic information of the input, in favour of information strictly relevant to classification. We introduce a network that has the capacity to do both classification and reconstruction by adding a "style memory" to the output layer of the network. We also show how to train such a neural network as a deep multi-layer autoencoder, jointly minimizing both classification and reconstruction losses. The generative capacity of our network demonstrates that the combination of style-memory neurons with the classifier neurons yield good reconstructions of the inputs when the classification is correct. We further investigate the nature of the style memory, and how it relates to composing digits and letters.

Augmenting the top layer of a classifier network with a style memory enables it to be generative.

Routing models, a form of conditional computation where examples are routed through a subset of components in a larger network, have shown promising results in recent works. Surprisingly, routing models to date have lacked important properties, such as architectural diversity and large numbers of routing decisions. Both architectural diversity and routing depth can increase the representational power of a routing network. In this work, we address both of these deficiencies. We discuss the significance of architectural diversity in routing models, and explain the tradeoffs between capacity and optimization when increasing routing depth. In our experiments, we find that adding architectural diversity to routing models significantly improves performance, cutting the error rates of a strong baseline by 35% on an Omniglot setup. However, when scaling up routing depth, we find that modern routing techniques struggle with optimization. We conclude by discussing both the positive and negative results, and suggest directions for future research.

Per-example routing models benefit from architectural diversity, but still struggle to scale to a large number of routing decisions.

Across numerous applications, forecasting relies on numerical solvers for partial differential equations (PDEs). Although the use of deep-learning techniques has been proposed, the uses have been restricted by the fact the training data are obtained using PDE solvers. Thereby, the uses were limited to domains, where the PDE solver was applicable, but no further. We present methods for training on small domains, while applying the trained models on larger domains, with consistency constraints ensuring the solutions are physically meaningful even at the boundary of the small domains. We demonstrate the results on an air-pollution forecasting model for Dublin, Ireland.

We present RNNs for training surrogate models of PDEs, wherein consistency constraints ensure the solutions are physically meaningful, even when the training uses much smaller domains than the trained model is applied to.

We address the issue of limit cycling behavior in training Generative Adversarial Networks and propose the use of Optimistic Mirror Decent (OMD) for training Wasserstein GANs. Recent theoretical results have shown that optimistic mirror decent (OMD) can enjoy faster regret rates in the context of zero-sum games. WGANs is exactly a context of solving a zero-sum game with simultaneous no-regret dynamics. Moreover, we show that optimistic mirror decent addresses the limit cycling problem in training WGANs. We formally show that in the case of bi-linear zero-sum games the last iterate of OMD dynamics converges to an equilibrium, in contrast to GD dynamics which are bound to cycle. We also portray the huge qualitative difference between GD and OMD dynamics with toy examples, even when GD is modified with many adaptations proposed in the recent literature, such as gradient penalty or momentum. We apply OMD WGAN training to a bioinformatics problem of generating DNA sequences. We observe that models trained with OMD achieve consistently smaller KL divergence with respect to the true underlying distribution, than models trained with GD variants. Finally, we introduce a new algorithm, Optimistic Adam, which is an optimistic variant of Adam. We apply it to WGAN training on CIFAR10 and observe improved performance in terms of inception score as compared to Adam.

We propose the use of optimistic mirror decent to address cycling problems in the training of GANs. We also introduce the Optimistic Adam algorithm

Learning good representations of users and items is crucially important to recommendation with implicit feedback. Matrix factorization is the basic idea to derive the representations of users and items by decomposing the given interaction matrix. However, existing matrix factorization based approaches share the limitation in that the interaction between user embedding and item embedding is only weakly enforced by fitting the given individual rating value, which may lose potentially useful information. In this paper, we propose a novel Augmented Generalized Matrix Factorization (AGMF) approach that is able to incorporate the historical interaction information of users and items for learning effective representations of users and items. Despite the simplicity of our proposed approach, extensive experiments on four public implicit feedback datasets demonstrate that our approach outperforms state-of-the-art counterparts. Furthermore, the ablation study demonstrates that by using multi-hot encoding to enrich user embedding and item embedding for Generalized Matrix Factorization, better performance, faster convergence, and lower training loss can be achieved.

A simple extension of generalized matrix factorization can outperform state-of-the-art approaches for recommendation.

We propose an unsupervised method for building dynamic representations of sequential data, particularly of observed interactions. The method simultaneously acquires representations of input data and its dynamics. It is based on a hierarchical generative model composed of two levels. In the first level, a model learns representations to generate observed data. In the second level, representational states encode the dynamics of the lower one. The model is designed as a Bayesian network with switching variables represented in the higher level, and which generates transition models. The method actively explores the latent space guided by its knowledge and the uncertainty about it. That is achieved by updating the latent variables from prediction error signals backpropagated to the latent space. So, no encoder or inference models are used since the generators also serve as their inverse transformations. The method is evaluated in two scenarios, with static images and with videos. The results show that the adaptation over time leads to better performance than with similar architectures without temporal dependencies, e.g., variational autoencoders. With videos, it is shown that the system extracts the dynamics of the data in states that highly correlate with the ground truth of the actions observed.

A method that build representations of sequential data and its dynamics through generative models with an active process

Activation is a nonlinearity function that plays a predominant role in the convergence and performance of deep neural networks. While Rectified Linear Unit (ReLU) is the most successful activation function, its derivatives have shown superior performance on benchmark datasets. In this work, we explore the polynomials as activation functions (order ≥ 2) that can approximate continuous real valued function within a given interval. Leveraging this property, the main idea is to learn the nonlinearity, accepting that the ensuing function may not be monotonic. While having the ability to learn more suitable nonlinearity, we cannot ignore the fact that it is a challenge to achieve stable performance due to exploding gradients - which is prominent with the increase in order. To handle this issue, we introduce dynamic input scaling, output scaling, and lower learning rate for the polynomial weights. Moreover, lower learning rate will control the abrupt fluctuations of the polynomials between weight updates. In experiments on three public datasets, our proposed method matches the performance of prior activation functions, thus providing insight into a network’s nonlinearity preference.

We propose polynomial as activation functions.

We introduce CBF, an exploration method that works in the absence of rewards or end of episode signal. CBF is based on intrinsic reward derived from the error of a dynamics model operating in feature space. It was inspired by (Pathak et al., 2017), is easy to implement, and can achieve results such as passing four levels of Super Mario Bros, navigating VizDoom mazes and passing two levels of SpaceInvaders. We investigated the effect of combining the method with several auxiliary tasks, but find inconsistent improvements over the CBF baseline.

A simple intrinsic motivation method using forward dynamics model error in feature space of the policy.

This paper is concerned with the robustness of VAEs to adversarial attacks. We highlight that conventional VAEs are brittle under attack but that methods recently introduced for disentanglement such as β-TCVAE (Chen et al., 2018) improve robustness, as demonstrated through a variety of previously proposed adversarial attacks (Tabacof et al. (2016); Gondim-Ribeiro et al. (2018); Kos et al.(2018)). This motivated us to develop Seatbelt-VAE, a new hierarchical disentangled VAE that is designed to be significantly more robust to adversarial attacks than existing approaches, while retaining high quality reconstructions.

We show that disentangled VAEs are more robust than vanilla VAEs to adversarial attacks that aim to trick them into decoding the adversarial input to a chosen target. We then develop an even more robust hierarchical disentangled VAE, Seatbelt-VAE.

The backpropagation algorithm is the de-facto standard for credit assignment in artificial neural networks due to its empirical results. Since its conception, variants of the backpropagation algorithm have emerged. More specifically, variants that leverage function changes in the backpropagation equations to satisfy their specific requirements. Feedback Alignment is one such example, which replaces the weight transpose matrix in the backpropagation equations with a random matrix in search of a more biologically plausible credit assignment algorithm. In this work, we show that function changes in the backpropagation procedure is equivalent to adding an implicit learning rate to an artificial neural network. Furthermore, we learn activation function derivatives in the backpropagation equations to demonstrate early convergence in these artificial neural networks. Our work reports competitive performances with early convergence on MNIST and CIFAR10 on sufficiently large deep neural network architectures.

We demonstrate that function changes in the backpropagation is equivalent to an implicit learning rate

Unsupervised text style transfer is the task of re-writing text of a given style into a target style without using a parallel corpus of source style and target style sentences for training. Style transfer systems are evaluated on their ability to generate sentences that 1) possess the target style, 2) are fluent and natural sounding, and 3) preserve the non-stylistic parts (content) of the source sentence. We train a reinforcement learning (RL) based unsupervised style transfer system that incorporates rewards for the above measures, and describe novel rewards shaping methods for the same. Our approach does not attempt to disentangle style and content, and leverages the power of massively pre-trained language models as well as the Transformer. Our system significantly outperforms existing state-of-art systems based on human as well as automatic evaluations on target style, fluency and content preservation as well as on overall success of style transfer, on a variety of datasets.

A reinforcement learning approach to text style transfer

Despite the success of Generative Adversarial Networks (GANs) in image synthesis, there lacks enough understanding on what networks have learned inside the deep generative representations and how photo-realistic images are able to be composed from random noises. In this work, we show that highly-structured semantic hierarchy emerges from the generative representations as the variation factors for synthesizing scenes. By probing the layer-wise representations with a broad set of visual concepts at different abstraction levels, we are able to quantify the causality between the activations and the semantics occurring in the output image. Such a quantification identifies the human-understandable variation factors learned by GANs to compose scenes. The qualitative and quantitative results suggest that the generative representations learned by GAN are specialized to synthesize different hierarchical semantics: the early layers tend to determine the spatial layout and configuration, the middle layers control the categorical objects, and the later layers finally render the scene attributes as well as color scheme. Identifying such a set of manipulatable latent semantics facilitates semantic scene manipulation.

We show that highly-structured semantic hierarchy emerges in the deep generative representations as a result for synthesizing scenes.

Variational autoencoders (VAEs) defined over SMILES string and graph-based representations of molecules promise to improve the optimization of molecular properties, thereby revolutionizing the pharmaceuticals and materials industries. However, these VAEs are hindered by the non-unique nature of SMILES strings and the computational cost of graph convolutions. To efficiently pass messages along all paths through the molecular graph, we encode multiple SMILES strings of a single molecule using a set of stacked recurrent neural networks, harmonizing hidden representations of each atom between SMILES representations, and use attentional pooling to build a final fixed-length latent representation. By then decoding to a disjoint set of SMILES strings of the molecule, our All SMILES VAE learns an almost bijective mapping between molecules and latent representations near the high-probability-mass subspace of the prior. Our SMILES-derived but molecule-based latent representations significantly surpass the state-of-the-art in a variety of fully- and semi-supervised property regression and molecular property optimization tasks.

We pool messages amongst multiple SMILES strings of the same molecule to pass information along all paths through the molecular graph, producing latent representations that significantly surpass the state-of-the-art in a variety of tasks.

We propose a simple yet highly effective method that addresses the mode-collapse problem in the Conditional Generative Adversarial Network (cGAN). Although conditional distributions are multi-modal (i.e., having many modes) in practice, most cGAN approaches tend to learn an overly simplified distribution where an input is always mapped to a single output regardless of variations in latent code. To address such issue, we propose to explicitly regularize the generator to produce diverse outputs depending on latent codes. The proposed regularization is simple, general, and can be easily integrated into most conditional GAN objectives. Additionally, explicit regularization on generator allows our method to control a balance between visual quality and diversity. We demonstrate the effectiveness of our method on three conditional generation tasks: image-to-image translation, image inpainting, and future video prediction. We show that simple addition of our regularization to existing models leads to surprisingly diverse generations, substantially outperforming the previous approaches for multi-modal conditional generation specifically designed in each individual task.

We propose a simple and general approach that avoids a mode collapse problem in various conditional GANs.

The transformer is a state-of-the-art neural translation model that uses attention to iteratively refine lexical representations with information drawn from the surrounding context. Lexical features are fed into the first layer and propagated through a deep network of hidden layers. We argue that the need to represent and propagate lexical features in each layer limits the model’s capacity for learning and representing other information relevant to the task. To alleviate this bottleneck, we introduce gated shortcut connections between the embedding layer and each subsequent layer within the encoder and decoder. This enables the model to access relevant lexical content dynamically, without expending limited resources on storing it within intermediate states. We show that the proposed modification yields consistent improvements on standard WMT translation tasks and reduces the amount of lexical information passed along the hidden layers. We furthermore evaluate different ways to integrate lexical connections into the transformer architecture and present ablation experiments exploring the effect of proposed shortcuts on model behavior.

Equipping the transformer model with shortcuts to the embedding layer frees up model capacity for learning novel information.

Probability density estimation is a classical and well studied problem, but standard density estimation methods have historically lacked the power to model complex and high-dimensional image distributions. More recent generative models leverage the power of neural networks to implicitly learn and represent probability models over complex images. We describe methods to extract explicit probability density estimates from GANs, and explore the properties of these image density functions. We perform sanity check experiments to provide evidence that these probabilities are reasonable. However, we also show that density functions of natural images are difficult to interpret and thus limited in use. We study reasons for this lack of interpretability, and suggest that we can get better interpretability by doing density estimation on latent representations of images.

We examine the relationship between probability density values and image content in non-invertible GANs.

Convolutional Neural Networks (CNNs) are composed of multiple convolution layers and show elegant performance in vision tasks. The design of the regular convolution is based on the Receptive Field (RF) where the information within a specific region is processed. In the view of the regular convolution's RF, the outputs of neurons in lower layers with smaller RF are bundled to create neurons in higher layers with larger RF. As a result, the neurons in high layers are able to capture the global context even though the neurons in low layers only see the local information. However, in lower layers of the biological brain, the information outside of the RF changes the properties of neurons. In this work, we extend the regular convolution and propose spatially shuffled convolution (ss convolution). In ss convolution, the regular convolution is able to use the information outside of its RF by spatial shuffling which is a simple and lightweight operation. We perform experiments on CIFAR-10 and ImageNet-1k dataset, and show that ss convolution improves the classification performance across various CNNs.

We propose spatially shuffled convolution that the regular convolution incorporates the information from outside of its receptive field.

We propose a framework to model the distribution of sequential data coming from a set of entities connected in a graph with a known topology. The method is based on a mixture of shared hidden Markov models (HMMs), which are trained in order to exploit the knowledge of the graph structure and in such a way that the obtained mixtures tend to be sparse. Experiments in different application domains demonstrate the effectiveness and versatility of the method.

A method to model the generative distribution of sequences coming from graph connected entities.

To gain high rewards in muti-agent scenes, it is sometimes necessary to understand other agents and make corresponding optimal decisions. We can solve these tasks by first building models for other agents and then finding the optimal policy with these models. To get an accurate model, many observations are needed and this can be sample-inefficient. What's more, the learned model and policy can overfit to current agents and cannot generalize if the other agents are replaced by new agents. In many practical situations, each agent we face can be considered as a sample from a population with a fixed but unknown distribution. Thus we can treat the task against some specific agents as a task sampled from a task distribution. We apply meta-learning method to build models and learn policies. Therefore when new agents come, we can adapt to them efficiently. Experiments on grid games show that our method can quickly get high rewards.

Our work applies meta-learning to multi-agent Reinforcement Learning to help our agent efficiently adapted to new coming opponents.

We characterize the singular values of the linear transformation associated with a standard 2D multi-channel convolutional layer, enabling their efficient computation. This characterization also leads to an algorithm for projecting a convolutional layer onto an operator-norm ball. We show that this is an effective regularizer; for example, it improves the test error of a deep residual network using batch normalization on CIFAR-10 from 6.2% to 5.3%.

We characterize the singular values of the linear transformation associated with a standard 2D multi-channel convolutional layer, enabling their efficient computation.

Trading off exploration and exploitation in an unknown environment is key to maximising expected return during learning. A Bayes-optimal policy, which does so optimally, conditions its actions not only on the environment state but on the agent's uncertainty about the environment. Computing a Bayes-optimal policy is however intractable for all but the smallest tasks. In this paper, we introduce variational Bayes-Adaptive Deep RL (variBAD), a way to meta-learn to perform approximate inference in an unknown environment, and incorporate task uncertainty directly during action selection. In a grid-world domain, we illustrate how variBAD performs structured online exploration as a function of task uncertainty. We also evaluate variBAD on MuJoCo domains widely used in meta-RL and show that it achieves higher return during training than existing methods.

VariBAD opens a path to tractable approximate Bayes-optimal exploration for deep RL using ideas from meta-learning, Bayesian RL, and approximate variational inference.

In a continual learning setting, new categories may be introduced over time, and an ideal learning system should perform well on both the original categories and the new categories. While deep neural nets have achieved resounding success in the classical setting, they are known to forget about knowledge acquired in prior episodes of learning if the examples encountered in the current episode of learning are drastically different from those encountered in prior episodes. This makes deep neural nets ill-suited to continual learning. In this paper, we propose a new model that can both leverage the expressive power of deep neural nets and is resilient to forgetting when new categories are introduced. We demonstrate an improvement in terms of accuracy on original classes compared to a vanilla deep neural net.

We show metric learning can help reduce catastrophic forgetting

Biomedical knowledge bases are crucial in modern data-driven biomedical sciences, but auto-mated biomedical knowledge base construction remains challenging. In this paper, we consider the problem of disease entity normalization, an essential task in constructing a biomedical knowledge base. We present NormCo, a deep coherence model which considers the semantics of an entity mention, as well as the topical coherence of the mentions within a single document. NormCo mod-els entity mentions using a simple semantic model which composes phrase representations from word embeddings, and treats coherence as a disease concept co-mention sequence using an RNN rather than modeling the joint probability of all concepts in a document, which requires NP-hard inference. To overcome the issue of data sparsity, we used distantly supervised data and synthetic data generated from priors derived from the BioASQ dataset. Our experimental results show thatNormCo outperforms state-of-the-art baseline methods on two disease normalization corpora in terms of (1) prediction quality and (2) efficiency, and is at least as performant in terms of accuracy and F1 score on tagged documents.

We present NormCo, a deep coherence model which considers the semantics of an entity mention, as well as the topical coherence of the mentions within a single document to perform disease entity normalization.

We explore the role of multiplicative interaction as a unifying framework to describe a range of classical and modern neural network architectural motifs, such as gating, attention layers, hypernetworks, and dynamic convolutions amongst others. Multiplicative interaction layers as primitive operations have a long-established presence in the literature, though this often not emphasized and thus under-appreciated. We begin by showing that such layers strictly enrich the representable function classes of neural networks. We conjecture that multiplicative interactions offer a particularly powerful inductive bias when fusing multiple streams of information or when conditional computation is required. We therefore argue that they should be considered in many situation where multiple compute or information paths need to be combined, in place of the simple and oft-used concatenation operation. Finally, we back up our claims and demonstrate the potential of multiplicative interactions by applying them in large-scale complex RL and sequence modelling tasks, where their use allows us to deliver state-of-the-art results, and thereby provides new evidence in support of multiplicative interactions playing a more prominent role when designing new neural network architectures.

We explore the role of multiplicative interaction as a unifying framework to describe a range of classical and modern neural network architectural motifs, such as gating, attention layers, hypernetworks, and dynamic convolutions amongst others.

Developing conditional generative models for text-to-video synthesis is an extremely challenging yet an important topic of research in machine learning. In this work, we address this problem by introducing Text-Filter conditioning Generative Adversarial Network (TFGAN), a GAN model with novel conditioning scheme that aids improving the text-video associations. With a combination of this conditioning scheme and a deep GAN architecture, TFGAN generates photo-realistic videos from text on very challenging real-world video datasets. In addition, we construct a benchmark synthetic dataset of moving shapes to systematically evaluate our conditioning scheme. Extensive experiments demonstrate that TFGAN significantly outperforms the existing approaches, and can also generate videos of novel categories not seen during training.

An effective text-conditioning GAN framework for generating videos from text

Over-parameterization is ubiquitous nowadays in training neural networks to benefit both optimization in seeking global optima and generalization in reducing prediction error. However, compressive networks are desired in many real world applications and direct training of small networks may be trapped in local optima. In this paper, instead of pruning or distilling over-parameterized models to compressive ones, we propose a new approach based on \emph{differential inclusions of inverse scale spaces}, that generates a family of models from simple to complex ones by coupling gradient descent and mirror descent to explore model structural sparsity. It has a simple discretization, called the Split Linearized Bregman Iteration (SplitLBI), whose global convergence analysis in deep learning is established that from any initializations, algorithmic iterations converge to a critical point of empirical risks. Experimental evidence shows that\ SplitLBI may achieve state-of-the-art performance in large scale training on ImageNet-2012 dataset etc., while with \emph{early stopping} it unveils effective subnet architecture with comparable test accuracies to dense models after retraining instead of pruning well-trained ones.

SplitLBI is applied to deep learning to explore model structural sparsity, achieving state-of-the-art performance in ImageNet-2012 and unveiling effective subnet architecture.

In this paper, we study the learned iterative shrinkage thresholding algorithm (LISTA) for solving sparse coding problems. Following assumptions made by prior works, we first discover that the code components in its estimations may be lower than expected, i.e., require gains, and to address this problem, a gated mechanism amenable to theoretical analysis is then introduced. Specific design of the gates is inspired by convergence analyses of the mechanism and hence its effectiveness can be formally guaranteed. In addition to the gain gates, we further introduce overshoot gates for compensating insufficient step size in LISTA. Extensive empirical results confirm our theoretical findings and verify the effectiveness of our method.

We propose gated mechanisms to enhance learned ISTA for sparse coding, with theoretical guarantees on the superiority of the method.

The learning of hierarchical representations for image classification has experienced an impressive series of successes due in part to the availability of large-scale labeled data for training. On the other hand, the trained classifiers have traditionally been evaluated on a handful of test images, which are deemed to be extremely sparsely distributed in the space of all natural images. It is thus questionable whether recent performance improvements on the excessively re-used test sets generalize to real-world natural images with much richer content variations. In addition, studies on adversarial learning show that it is effortless to construct adversarial examples that fool nearly all image classifiers, adding more complications to relative performance comparison of existing models. This work presents an efficient framework for comparing image classifiers, which we name the MAximum Discrepancy (MAD) competition. Rather than comparing image classifiers on fixed test sets, we adaptively sample a test set from an arbitrarily large corpus of unlabeled images so as to maximize the discrepancies between the classifiers, measured by the distance over WordNet hierarchy. Human labeling on the resulting small and model-dependent image sets reveals the relative performance of the competing classifiers and provides useful insights on potential ways to improve them. We report the MAD competition results of eleven ImageNet classifiers while noting that the framework is readily extensible and cost-effective to add future classifiers into the competition.

We present an efficient and adaptive framework for comparing image classifiers to maximize the discrepancies between the classifiers, in place of comparing on fixed test sets.

Robustness of neural networks has recently been highlighted by the adversarial examples, i.e., inputs added with well-designed perturbations which are imperceptible to humans but can cause the network to give incorrect outputs. In this paper, we design a new CNN architecture that by itself has good robustness. We introduce a simple but powerful technique, Random Mask, to modify existing CNN structures. We show that CNN with Random Mask achieves state-of-the-art performance against black-box adversarial attacks without applying any adversarial training. We next investigate the adversarial examples which “fool” a CNN with Random Mask. Surprisingly, we find that these adversarial examples often “fool” humans as well. This raises fundamental questions on how to define adversarial examples and robustness properly.

We propose a technique that modifies CNN structures to enhance robustness while keeping high test accuracy, and raise doubt on whether current definition of adversarial examples is appropriate by generating adversarial examples able to fool humans.

Supervised deep learning methods require cleanly labeled large-scale datasets, but collecting such data is difficult and sometimes impossible. There exist two popular frameworks to alleviate this problem: semi-supervised learning and robust learning to label noise. Although these frameworks relax the restriction of supervised learning, they are studied independently. Hence, the training scheme that is suitable when only small cleanly-labeled data are available remains unknown. In this study, we consider learning from bi-quality data as a generalization of these studies, in which a small portion of data is cleanly labeled, and the rest is corrupt. Under this framework, we compare recent algorithms for semi-supervised and robust learning. The results suggest that semi-supervised learning outperforms robust learning with noisy labels. We also propose a training strategy for mixing mixup techniques to learn from such bi-quality data effectively.

We propose to compare semi-supervised and robust learning to noisy label under a shared setting

Hierarchical Sparse Coding (HSC) is a powerful model to efficiently represent multi-dimensional, structured data such as images. The simplest solution to solve this computationally hard problem is to decompose it into independent layerwise subproblems. However, neuroscientific evidence would suggest inter-connecting these subproblems as in the Predictive Coding (PC) theory, which adds top-down connections between consecutive layers. In this study, a new model called Sparse Deep Predictive Coding (SDPC) is introduced to assess the impact of this inter-layer feedback connection. In particular, the SDPC is compared with a Hierarchical Lasso (Hi-La) network made out of a sequence of Lasso layers. A 2-layered SDPC and a Hi-La networks are trained on 3 different databases and with different sparsity parameters on each layer. First, we show that the overall prediction error generated by SDPC is lower thanks to the feedback mechanism as it transfers prediction error between layers. Second, we demonstrate that the inference stage of the SDPC is faster to converge than for the Hi-La model. Third, we show that the SDPC also accelerates the learning process. Finally, the qualitative analysis of both models dictionaries, supported by their activation probability, show that the SDPC features are more generic and informative.

This paper experimentally demonstrates the beneficial effect of top-down connections in Hierarchical Sparse Coding algorithm.

Explaining a deep learning model can help users understand its behavior and allow researchers to discern its shortcomings. Recent work has primarily focused on explaining models for tasks like image classification or visual question answering. In this paper, we introduce an explanation approach for image similarity models, where a model's output is a score measuring the similarity of two inputs rather than a classification. In this task, an explanation depends on both of the input images, so standard methods do not apply. We propose an explanation method that pairs a saliency map identifying important image regions with an attribute that best explains the match. We find that our explanations provide additional information not typically captured by saliency maps alone, and can also improve performance on the classic task of attribute recognition. Our approach's ability to generalize is demonstrated on two datasets from diverse domains, Polyvore Outfits and Animals with Attributes 2.

A black box approach for explaining the predictions of an image similarity model.

Adversarial examples have been shown to be an effective way of assessing the robustness of neural sequence-to-sequence (seq2seq) models, by applying perturbations to the input of a model leading to large degradation in performance. However, these perturbations are only indicative of a weakness in the model if they do not change the semantics of the input in a way that would change the expected output. Using the example of machine translation (MT), we propose a new evaluation framework for adversarial attacks on seq2seq models taking meaning preservation into account and demonstrate that existing methods may not preserve meaning in general. Based on these findings, we propose new constraints for attacks on word-based MT systems and show, via human and automatic evaluation, that they produce more semantically similar adversarial inputs. Furthermore, we show that performing adversarial training with meaning-preserving attacks is beneficial to the model in terms of adversarial robustness without hurting test performance.

How you should evaluate adversarial attacks on seq2seq

We introduce a new normalization technique that exhibits the fast convergence properties of batch normalization using a transformation of layer weights instead of layer outputs. The proposed technique keeps the contribution of positive and negative weights to the layer output in equilibrium. We validate our method on a set of standard benchmarks including CIFAR-10/100, SVHN and ILSVRC 2012 ImageNet.

An alternative normalization technique to batch normalization

We present a framework for building unsupervised representations of entities and their compositions, where each entity is viewed as a probability distribution rather than a fixed length vector. In particular, this distribution is supported over the contexts which co-occur with the entity and are embedded in a suitable low-dimensional space. This enables us to consider the problem of representation learning with a perspective from Optimal Transport and take advantage of its numerous tools such as Wasserstein distance and Wasserstein barycenters. We elaborate how the method can be applied for obtaining unsupervised representations of text and illustrate the performance quantitatively as well as qualitatively on tasks such as measuring sentence similarity and word entailment, where we empirically observe significant gains (e.g., 4.1% relative improvement over Sent2vec and GenSen). The key benefits of the proposed approach include: (a) capturing uncertainty and polysemy via modeling the entities as distributions, (b) utilizing the underlying geometry of the particular task (with the ground cost), (c) simultaneously providing interpretability with the notion of optimal transport between contexts and (d) easy applicability on top of existing point embedding methods. In essence, the framework can be useful for any unsupervised or supervised problem (on text or other modalities); and only requires a co-occurrence structure inherent to many problems. The code, as well as pre-built histograms, are available under https://github.com/context-mover.

Represent each entity as a probability distribution over contexts embedded in a ground space.

Over the last few years, the phenomenon of adversarial examples --- maliciously constructed inputs that fool trained machine learning models --- has captured the attention of the research community, especially when the adversary is restricted to making small modifications of a correctly handled input. At the same time, less surprisingly, image classifiers lack human-level performance on randomly corrupted images, such as images with additive Gaussian noise. In this work, we show that these are two manifestations of the same underlying phenomenon. We establish this connection in several ways. First, we find that adversarial examples exist at the same distance scales we would expect from a linear model with the same performance on corrupted images. Next, we show that Gaussian data augmentation during training improves robustness to small adversarial perturbations and that adversarial training improves robustness to several types of image corruptions. Finally, we present a model-independent upper bound on the distance from a corrupted image to its nearest error given test performance and show that in practice we already come close to achieving the bound, so that improving robustness further for the corrupted image distribution requires significantly reducing test error. All of this suggests that improving adversarial robustness should go hand in hand with improving performance in the presence of more general and realistic image corruptions. This yields a computationally tractable evaluation metric for defenses to consider: test error in noisy image distributions.

Small adversarial perturbations should be expected given observed error rates of models outside the natural data distribution.

Recent developments in natural language representations have been accompanied by large and expensive models that leverage vast amounts of general-domain text through self-supervised pre-training. Due to the cost of applying such models to down-stream tasks, several model compression techniques on pre-trained language representations have been proposed (Sun et al., 2019; Sanh, 2019). However, surprisingly, the simple baseline of just pre-training and fine-tuning compact models has been overlooked. In this paper, we first show that pre-training remains important in the context of smaller architectures, and fine-tuning pre-trained compact models can be competitive to more elaborate methods proposed in concurrent work. Starting with pre-trained compact models, we then explore transferring task knowledge from large fine-tuned models through standard knowledge distillation. The resulting simple, yet effective and general algorithm, Pre-trained Distillation, brings further improvements. Through extensive experiments, we more generally explore the interaction between pre-training and distillation under two variables that have been under-studied: model size and properties of unlabeled task data. One surprising observation is that they have a compound effect even when sequentially applied on the same data. To accelerate future research, we will make our 24 pre-trained miniature BERT models publicly available.

Studies how self-supervised learning and knowledge distillation interact in the context of building compact models.

In this paper, we investigate lossy compression of deep neural networks (DNNs) by weight quantization and lossless source coding for memory-efficient deployment. Whereas the previous work addressed non-universal scalar quantization and entropy coding of DNN weights, we for the first time introduce universal DNN compression by universal vector quantization and universal source coding. In particular, we examine universal randomized lattice quantization of DNNs, which randomizes DNN weights by uniform random dithering before lattice quantization and can perform near-optimally on any source without relying on knowledge of its probability distribution. Moreover, we present a method of fine-tuning vector quantized DNNs to recover the performance loss after quantization. Our experimental results show that the proposed universal DNN compression scheme compresses the 32-layer ResNet (trained on CIFAR-10) and the AlexNet (trained on ImageNet) with compression ratios of $47.1$ and $42.5$, respectively.

We introduce the universal deep neural network compression scheme, which is applicable universally for compression of any models and can perform near-optimally regardless of their weight distribution.

What would be learned by variational autoencoder(VAE) and what influence the disentanglement of VAE? This paper tries to preliminarily address VAE's intrinsic dimension, real factor, disentanglement and indicator issues theoretically in the idealistic situation and implementation issue practically through noise modeling perspective in the realistic case. On intrinsic dimension issue, due to information conservation, the idealistic VAE learns and only learns intrinsic factor dimension. Besides, suggested by mutual information separation property, the constraint induced by Gaussian prior to the VAE objective encourages the information sparsity in dimension. On disentanglement issue, subsequently, inspired by information conservation theorem the clarification on disentanglement in this paper is made. On real factor issue, due to factor equivalence, the idealistic VAE possibly learns any factor set in the equivalence class. On indicator issue, the behavior of current disentanglement metric is discussed, and several performance indicators regarding the disentanglement and generating influence are subsequently raised to evaluate the performance of VAE model and to supervise the used factors. On implementation issue, the experiments under noise modeling and constraints empirically testify the theoretical analysis and also show their own characteristic in pursuing disentanglement.

This paper tries to preliminarily address the disentanglement theoretically in the idealistic situation and practically through noise modelling perspective in the realistic case.

Weight decay is one of the standard tricks in the neural network toolbox, but the reasons for its regularization effect are poorly understood, and recent results have cast doubt on the traditional interpretation in terms of $L_2$ regularization. Literal weight decay has been shown to outperform $L_2$ regularization for optimizers for which they differ. We empirically investigate weight decay for three optimization algorithms (SGD, Adam, and K-FAC) and a variety of network architectures. We identify three distinct mechanisms by which weight decay exerts a regularization effect, depending on the particular optimization algorithm and architecture: (1) increasing the effective learning rate, (2) approximately regularizing the input-output Jacobian norm, and (3) reducing the effective damping coefficient for second-order optimization. Our results provide insight into how to improve the regularization of neural networks.

We investigate weight decay regularization for different optimizers and identify three distinct mechanisms by which weight decay improves generalization.

In this paper we present the first freely available dataset for the development and evaluation of domain adaptation methods, for the sound event detection task. The dataset contains 40 log mel-band energies extracted from $100$ different synthetic sound event tracks, with additive noise from nine different acoustic scenes (from indoor, outdoor, and vehicle environments), mixed at six different sound-to-noise ratios, SNRs, (from -12 to -27 dB with a step of -3 dB), and totaling to 5400 (9 * 100 * 6) sound files and a total length of 30 564 minutes. We provide the dataset as is, the code to re-create the dataset and remix the sound event tracks and the acoustic scenes with different SNRs, and a baseline method that tests the adaptation performance with the proposed dataset and establishes some first results.

The very first freely available domain adaptation dataset for sound event detection.

This paper aims to address the limitations of mutual information estimators based on variational optimization. By redefining the cost using generalized functions from nonextensive statistical mechanics we raise the upper bound of previous estimators and enable the control of the bias variance trade off. Variational based estimators outperform previous methods especially in high dependence high dimensional scenarios found in machine learning setups. Despite their performance, these estimators either exhibit a high variance or are upper bounded by log(batch size). Our approach inspired by nonextensive statistical mechanics uses different generalizations for the logarithm and the exponential in the partition function. This enables the estimator to capture changes in mutual information over a wider range of dimensions and correlations of the input variables whereas previous estimators saturate them.

Mutual information estimator based nonextensive statistical mechanics

Generative adversarial networks (GANs) are a widely used framework for learning generative models. Wasserstein GANs (WGANs), one of the most successful variants of GANs, require solving a minmax problem to global optimality, but in practice, are successfully trained with stochastic gradient descent-ascent. In this paper, we show that, when the generator is a one-layer network, stochastic gradient descent-ascent converges to a global solution in polynomial time and sample complexity.

We show that stochastic gradient descent ascent converges to a global optimum for WGAN with one-layer generator network.

Classifiers such as deep neural networks have been shown to be vulnerable against adversarial perturbations on problems with high-dimensional input space. While adversarial training improves the robustness of classifiers against such adversarial perturbations, it leaves classifiers sensitive to them on a non-negligible fraction of the inputs. We argue that there are two different kinds of adversarial perturbations: shared perturbations which fool a classifier on many inputs and singular perturbations which only fool the classifier on a small fraction of the data. We find that adversarial training increases the robustness of classifiers against shared perturbations. Moreover, it is particularly effective in removing universal perturbations, which can be seen as an extreme form of shared perturbations. Unfortunately, adversarial training does not consistently increase the robustness against singular perturbations on unseen inputs. However, we find that adversarial training decreases robustness of the remaining perturbations against image transformations such as changes to contrast and brightness or Gaussian blurring. It thus makes successful attacks on the classifier in the physical world less likely. Finally, we show that even singular perturbations can be easily detected and must thus exhibit generalizable patterns even though the perturbations are specific for certain inputs.

We empirically show that adversarial training is effective for removing universal perturbations, makes adversarial examples less robust to image transformations, and leaves them detectable for a detection approach.

We address the challenging problem of efficient deep learning model deployment, where the goal is to design neural network architectures that can fit different hardware platform constraints. Most of the traditional approaches either manually design or use Neural Architecture Search (NAS) to find a specialized neural network and train it from scratch for each case, which is computationally expensive and unscalable. Our key idea is to decouple model training from architecture search to save the cost. To this end, we propose to train a once-for-all network (OFA) that supports diverse architectural settings (depth, width, kernel size, and resolution). Given a deployment scenario, we can then quickly get a specialized sub-network by selecting from the OFA network without additional training. To prevent interference between many sub-networks during training, we also propose a novel progressive shrinking algorithm, which can train a surprisingly large number of sub-networks ($> 10^{19}$) simultaneously. Extensive experiments on various hardware platforms (CPU, GPU, mCPU, mGPU, FPGA accelerator) show that OFA consistently outperforms SOTA NAS methods (up to 4.0% ImageNet top1 accuracy improvement over MobileNetV3) while reducing orders of magnitude GPU hours and $CO_2$ emission. In particular, OFA achieves a new SOTA 80.0% ImageNet top1 accuracy under the mobile setting ($<$600M FLOPs). Code and pre-trained models are released at https://github.com/mit-han-lab/once-for-all.

We introduce techniques to train a single once-for-all network that fits many hardware platforms.

A deep generative model is a powerful method of learning a data distribution, which has achieved tremendous success in numerous scenarios. However, it is nontrivial for a single generative model to faithfully capture the distributions of the complex data such as images with complicate structures. In this paper, we propose a novel approach of cascaded boosting for boosting generative models, where meta-models (i.e., weak learners) are cascaded together to produce a stronger model. Any hidden variable meta-model can be leveraged as long as it can support the likelihood evaluation. We derive a decomposable variational lower bound of the boosted model, which allows each meta-model to be trained separately and greedily. We can further improve the learning power of the generative models by combing our cascaded boosting framework with the multiplicative boosting framework.

Propose an approach for boosting generative models by cascading hidden variable models

Contextualized representation models such as ELMo (Peters et al., 2018a) and BERT (Devlin et al., 2018) have recently achieved state-of-the-art results on a diverse array of downstream NLP tasks. Building on recent token-level probing work, we introduce a novel edge probing task design and construct a broad suite of sub-sentence tasks derived from the traditional structured NLP pipeline. We probe word-level contextual representations from four recent models and investigate how they encode sentence structure across a range of syntactic, semantic, local, and long-range phenomena. We find that existing models trained on language modeling and translation produce strong representations for syntactic phenomena, but only offer comparably small improvements on semantic tasks over a non-contextual baseline.

We probe for sentence structure in ELMo and related contextual embedding models. We find existing models efficiently encode syntax and show evidence of long-range dependencies, but only offer small improvements on semantic tasks.

Deep reinforcement learning has succeeded in sophisticated games such as Atari, Go, etc. Real-world decision making, however, often requires reasoning with partial information extracted from complex visual observations. This paper presents Discriminative Particle Filter Reinforcement Learning (DPFRL), a new reinforcement learning framework for partial and complex observations. DPFRL encodes a differentiable particle filter with learned transition and observation models in a neural network, which allows for reasoning with partial observations over multiple time steps. While a standard particle filter relies on a generative observation model, DPFRL learns a discriminatively parameterized model that is training directly for decision making. We show that the discriminative parameterization results in significantly improved performance, especially for tasks with complex visual observations, because it circumvents the difficulty of modelling observations explicitly. In most cases, DPFRL outperforms state-of-the-art POMDP RL models in Flickering Atari Games, an existing POMDP RL benchmark, and in Natural Flickering Atari Games, a new, more challenging POMDP RL benchmark that we introduce. We further show that DPFRL performs well for visual navigation with real-world data.

We introduce DPFRL, a framework for reinforcement learning under partial and complex observations with a fully differentiable discriminative particle filter

Extending models with auxiliary latent variables is a well-known technique to in-crease model expressivity. Bachman & Precup (2015); Naesseth et al. (2018); Cremer et al. (2017); Domke & Sheldon (2018) show that Importance Weighted Autoencoders (IWAE) (Burda et al., 2015) can be viewed as extending the variational family with auxiliary latent variables. Similarly, we show that this view encompasses many of the recent developments in variational bounds (Maddisonet al., 2017; Naesseth et al., 2018; Le et al., 2017; Yin & Zhou, 2018; Molchanovet al., 2018; Sobolev & Vetrov, 2018). The success of enriching the variational family with auxiliary latent variables motivates applying the same techniques to the generative model. We develop a generative model analogous to the IWAE bound and empirically show that it outperforms the recently proposed Learned Accept/Reject Sampling algorithm (Bauer & Mnih, 2018), while being substantially easier to implement. Furthermore, we show that this generative process provides new insights on ranking Noise Contrastive Estimation (Jozefowicz et al.,2016; Ma & Collins, 2018) and Contrastive Predictive Coding (Oord et al., 2018).

Monte Carlo Objectives are analyzed using auxiliary variable variational inference, yielding a new analysis of CPC and NCE as well as a new generative model.

Stochastic Gradient Descent or SGD is the most popular optimization algorithm for large-scale problems. SGD estimates the gradient by uniform sampling with sample size one. There have been several other works that suggest faster epoch wise convergence by using weighted non-uniform sampling for better gradient estimates. Unfortunately, the per-iteration cost of maintaining this adaptive distribution for gradient estimation is more than calculating the full gradient. As a result, the false impression of faster convergence in iterations leads to slower convergence in time, which we call a chicken-and-egg loop. In this paper, we break this barrier by providing the first demonstration of a sampling scheme, which leads to superior gradient estimation, while keeping the sampling cost per iteration similar to that of the uniform sampling. Such an algorithm is possible due to the sampling view of Locality Sensitive Hashing (LSH), which came to light recently. As a consequence of superior and fast estimation, we reduce the running time of all existing gradient descent algorithms. We demonstrate the benefits of our proposal on both SGD and AdaGrad.

We improve the running of all existing gradient descent algorithms.

In recent years we have made significant progress identifying computational principles that underlie neural function. While not yet complete, we have sufficient evidence that a synthesis of these ideas could result in an understanding of how neural computation emerges from a combination of innate dynamics and plasticity, and which could potentially be used to construct new AI technologies with unique capabilities. I discuss the relevant principles, the advantages they have for computation, and how they can benefit AI. Limitations of current AI are generally recognized, but fewer people are aware that we understand enough about the brain to immediately offer novel AI formulations.

Limitations of current AI are generally recognized, but fewer people are aware that we understand enough about the brain to immediately offer novel AI formulations.

Recent work has demonstrated how predictive modeling can endow agents with rich knowledge of their surroundings, improving their ability to act in complex environments. We propose question-answering as a general paradigm to decode and understand the representations that such agents develop, applying our method to two recent approaches to predictive modeling – action-conditional CPC (Guo et al., 2018) and SimCore (Gregor et al., 2019). After training agents with these predictive objectives in a visually-rich, 3D environment with an assortment of objects, colors, shapes, and spatial configurations, we probe their internal state representations with a host of synthetic (English) questions, without backpropagating gradients from the question-answering decoder into the agent. The performance of different agents when probed in this way reveals that they learn to encode detailed, and seemingly compositional, information about objects, properties and spatial relations from their physical environment. Our approach is intuitive, i.e. humans can easily interpret the responses of the model as opposed to inspecting continuous vectors, and model-agnostic, i.e. applicable to any modeling approach. By revealing the implicit knowledge of objects, quantities, properties and relations acquired by agents as they learn, question-conditional agent probing can stimulate the design and development of stronger predictive learning objectives.

We use question-answering to evaluate how much knowledge about the environment can agents learn by self-supervised prediction.

In most real-world scenarios, training datasets are highly class-imbalanced, where deep neural networks suffer from generalizing to a balanced testing criterion. In this paper, we explore a novel yet simple way to alleviate this issue via synthesizing less-frequent classes with adversarial examples of other classes. Surprisingly, we found this counter-intuitive method can effectively learn generalizable features of minority classes by transferring and leveraging the diversity of the majority information. Our experimental results on various types of class-imbalanced datasets in image classification and natural language processing show that the proposed method not only improves the generalization of minority classes significantly compared to other re-sampling or re-weighting methods, but also surpasses other methods of state-of-art level for the class-imbalanced classification.

We develop a new method for imbalanced classification using adversarial examples

Active matter consists of active agents which transform energy extracted from surroundings into momentum, producing a variety of collective phenomena. A model, synthetic active system composed of microtubule polymers driven by protein motors spontaneously forms a liquid-crystalline nematic phase. Extensile stress created by the protein motors precipitates continuous buckling and folding of the microtubules creating motile topological defects and turbulent fluid flows. Defect motion is determined by the rheological properties of the material; however, these remain largely unquantified. Measuring defects dynamics can yield fundamental insights into active nematics, a class of materials that include bacterial films and animal cells. Current methods for defect detection lack robustness and precision, and require fine-tuning for datasets with different visual quality. In this study, we applied Deep Learning to train a defect detector to automatically analyze microscopy videos of the microtubule active nematic. Experimental results indicate that our method is robust and accurate. It is expected to significantly increase the amount of video data that can be processed.

An interesting application of CNN in soft condensed matter physics experiments.

In this work we study locality and compositionality in the context of learning representations for Zero Shot Learning (ZSL). In order to well-isolate the importance of these properties in learned representations, we impose the additional constraint that, differently from most recent work in ZSL, no pre-training on different datasets (e.g. ImageNet) is performed. The results of our experiment show how locality, in terms of small parts of the input, and compositionality, i.e. how well can the learned representations be expressed as a function of a smaller vocabulary, are both deeply related to generalization and motivate the focus on more local-aware models in future research directions for representation learning.

An analysis of the effects of compositionality and locality on representation learning for zero-shot learning.

It is becoming increasingly clear that many machine learning classifiers are vulnerable to adversarial examples. In attempting to explain the origin of adversarial examples, previous studies have typically focused on the fact that neural networks operate on high dimensional data, they overfit, or they are too linear. Here we show that distributions of logit differences have a universal functional form. This functional form is independent of architecture, dataset, and training protocol; nor does it change during training. This leads to adversarial error having a universal scaling, as a power-law, with respect to the size of the adversarial perturbation. We show that this universality holds for a broad range of datasets (MNIST, CIFAR10, ImageNet, and random data), models (including state-of-the-art deep networks, linear models, adversarially trained networks, and networks trained on randomly shuffled labels), and attacks (FGSM, step l.l., PGD). Motivated by these results, we study the effects of reducing prediction entropy on adversarial robustness. Finally, we study the effect of network architectures on adversarial sensitivity. To do this, we use neural architecture search with reinforcement learning to find adversarially robust architectures on CIFAR10. Our resulting architecture is more robust to white \emph{and} black box attacks compared to previous attempts.

Adversarial error has similar power-law form for all datasets and models studied, and architecture matters.

Reinforcement learning (RL) has led to increasingly complex looking behavior in recent years. However, such complexity can be misleading and hides over-fitting. We find that visual representations may be a useful metric of complexity, and both correlates well objective optimization and causally effects reward optimization. We then propose curious representation learning (CRL) which allows us to use better visual representation learning algorithms to correspondingly increase visual representation in policy through an intrinsic objective on both simulated environments and transfer to real images. Finally, we show better visual representations induced by CRL allows us to obtain better performance on Atari without any reward than other curiosity objectives.

We present a formulation of curiosity as a visual representation learning problem and show that it allows good visual representations in agents.

This paper introduces the task of semantic instance completion: from an incomplete RGB-D scan of a scene, we aim to detect the individual object instances comprising the scene and infer their complete object geometry. This enables a semantically meaningful decomposition of a scanned scene into individual, complete 3D objects, including hidden and unobserved object parts. This will open up new possibilities for interactions with object in a scene, for instance for virtual or robotic agents. To address this task, we propose 3D-SIC, a new data-driven approach that jointly detects object instances and predicts their completed geometry. The core idea of 3D-SIC is a novel end-to-end 3D neural network architecture that leverages joint color and geometry feature learning. The fully-convolutional nature of our 3D network enables efficient inference of semantic instance completion for 3D scans at scale of large indoor environments in a single forward pass. In a series evaluation, we evaluate on both real and synthetic scan benchmark data, where we outperform state-of-the-art approaches by over 15 in mAP@0.5 on ScanNet, and over 18 in mAP@0.5 on SUNCG.

From an incomplete RGB-D scan of a scene, we aim to detect the individual object instances comprising the scene and infer their complete object geometry.

Style transfer usually refers to the task of applying color and texture information from a specific style image to a given content image while preserving the structure of the latter. Here we tackle the more generic problem of semantic style transfer: given two unpaired collections of images, we aim to learn a mapping between the corpus-level style of each collection, while preserving semantic content shared across the two domains. We introduce XGAN ("Cross-GAN"), a dual adversarial autoencoder, which captures a shared representation of the common domain semantic content in an unsupervised way, while jointly learning the domain-to-domain image translations in both directions. We exploit ideas from the domain adaptation literature and define a semantic consistency loss which encourages the model to preserve semantics in the learned embedding space. We report promising qualitative results for the task of face-to-cartoon translation. The cartoon dataset we collected for this purpose will also be released as a new benchmark for semantic style transfer.

XGAN is an unsupervised model for feature-level image-to-image translation applied to semantic style transfer problems such as the face-to-cartoon task, for which we introduce a new dataset.

Training neural networks on large datasets can be accelerated by distributing the workload over a network of machines. As datasets grow ever larger, networks of hundreds or thousands of machines become economically viable. The time cost of communicating gradients limits the effectiveness of using such large machine counts, as may the increased chance of network faults. We explore a particularly simple algorithm for robust, communication-efficient learning---signSGD. Workers transmit only the sign of their gradient vector to a server, and the overall update is decided by a majority vote. This algorithm uses 32x less communication per iteration than full-precision, distributed SGD. Under natural conditions verified by experiment, we prove that signSGD converges in the large and mini-batch settings, establishing convergence for a parameter regime of Adam as a byproduct. Aggregating sign gradients by majority vote means that no individual worker has too much power. We prove that unlike SGD, majority vote is robust when up to 50% of workers behave adversarially. The class of adversaries we consider includes as special cases those that invert or randomise their gradient estimate. On the practical side, we built our distributed training system in Pytorch. Benchmarking against the state of the art collective communications library (NCCL), our framework---with the parameter server housed entirely on one machine---led to a 25% reduction in time for training resnet50 on Imagenet when using 15 AWS p3.2xlarge machines.

Workers send gradient signs to the server, and the update is decided by majority vote. We show that this algorithm is convergent, communication efficient and fault tolerant, both in theory and in practice.

Profiling cellular phenotypes from microscopic imaging can provide meaningful biological information resulting from various factors affecting the cells. One motivating application is drug development: morphological cell features can be captured from images, from which similarities between different drugs applied at different dosages can be quantified. The general approach is to find a function mapping the images to an embedding space of manageable dimensionality whose geometry captures relevant features of the input images. An important known issue for such methods is separating relevant biological signal from nuisance variation. For example, the embedding vectors tend to be more correlated for cells that were cultured and imaged during the same week than for cells from a different week, despite having identical drug compounds applied in both cases. In this case, the particular batch a set of experiments were conducted in constitutes the domain of the data; an ideal set of image embeddings should contain only the relevant biological information (e.g. drug effects). We develop a general framework for adjusting the image embeddings in order to `forget' domain-specific information while preserving relevant biological information. To do this, we minimize a loss function based on distances between marginal distributions (such as the Wasserstein distance) of embeddings across domains for each replicated treatment. For the dataset presented, the replicated treatment is the negative control. We find that for our transformed embeddings (1) the underlying geometric structure is not only preserved but the embeddings also carry improved biological signal (2) less domain-specific information is present.

We correct nuisance variation for image embeddings across different domains, preserving only relevant information.

This paper presents a Mutual Information Neural Estimator (MINE) that is linearly scalable in dimensionality as well as in sample size. MINE is back-propable and we prove that it is strongly consistent. We illustrate a handful of applications in which MINE is succesfully applied to enhance the property of generative models in both unsupervised and supervised settings. We apply our framework to estimate the information bottleneck, and apply it in tasks related to supervised classification problems. Our results demonstrate substantial added flexibility and improvement in these settings.

A scalable in sample size and dimensions mutual information estimator.

Reinforcement learning methods have recently achieved impressive results on a wide range of control problems. However, especially with complex inputs, they still require an extensive amount of training data in order to converge to a meaningful solution. This limitation largely prohibits their usage for complex input spaces such as video signals, and it is still impossible to use it for a number of complex problems in a real world environments, including many of those for video based control. Supervised learning, on the contrary, is capable of learning on a relatively small number of samples, however it does not take into account reward-based control policies and is not capable to provide independent control policies. In this article we propose a model-free control method, which uses a combination of reinforcement and supervised learning for autonomous control and paves the way towards policy based control in real world environments. We use SpeedDreams/TORCS video game to demonstrate that our approach requires much less samples (hundreds of thousands against millions or tens of millions) comparing to the state-of-the-art reinforcement learning techniques on similar data, and at the same time overcomes both supervised and reinforcement learning approaches in terms of quality. Additionally, we demonstrate the applicability of the method to MuJoCo control problems.

The new combination of reinforcement and supervised learning, dramatically decreasing the number of required samples for training on video

A typical experiment to study cognitive function is to train animals to perform tasks, while the researcher records the electrical activity of the animals neurons. The main obstacle faced, when using this type of electrophysiological experiment to uncover the circuit mechanisms underlying complex behaviors, is our incomplete access to relevant circuits in the brain. One promising approach is to model neural circuits using an artificial neural network (ANN), which can provide complete access to the “neural circuits” responsible for a behavior. More recently, reinforcement learning models have been adopted to understand the functions of cortico-basal ganglia circuits as reward-based learning has been found in mammalian brain. In this paper, we propose a Biologically-plausible Actor-Critic with Episodic Memory (B-ACEM) framework to model a prefrontal cortex-basal ganglia-hippocampus (PFC-BG) circuit, which is verified to capture the behavioral findings from a well-known perceptual decision-making task, i.e., random dots motion discrimination. This B-ACEM framework links neural computation to behaviors, on which we can explore how episodic memory should be considered to govern future decision. Experiments are conducted using different settings of the episodic memory and results show that all patterns of episodic memories can speed up learning. In particular, salient events are prioritized to propagate reward information and guide decisions. Our B-ACEM framework and the built-on experiments give inspirations to both designs for more standard decision-making models in biological system and a more biologically-plausible ANN.

Fast learning via episodic memory verified by a biologically plausible framework for prefrontal cortex-basal ganglia-hippocampus (PFC-BG) circuit

Understanding the representational power of Deep Neural Networks (DNNs) and how their structural properties (e.g., depth, width, type of activation unit) affect the functions they can compute, has been an important yet challenging question in deep learning and approximation theory. In a seminal paper, Telgarsky high- lighted the benefits of depth by presenting a family of functions (based on sim- ple triangular waves) for which DNNs achieve zero classification error, whereas shallow networks with fewer than exponentially many nodes incur constant error. Even though Telgarsky’s work reveals the limitations of shallow neural networks, it doesn’t inform us on why these functions are difficult to represent and in fact he states it as a tantalizing open question to characterize those functions that cannot be well-approximated by smaller depths. In this work, we point to a new connection between DNNs expressivity and Sharkovsky’s Theorem from dynamical systems, that enables us to characterize the depth-width trade-offs of ReLU networks for representing functions based on the presence of a generalized notion of fixed points, called periodic points (a fixed point is a point of period 1). Motivated by our observation that the triangle waves used in Telgarsky’s work contain points of period 3 – a period that is special in that it implies chaotic behaviour based on the celebrated result by Li-Yorke – we proceed to give general lower bounds for the width needed to represent periodic functions as a function of the depth. Technically, the crux of our approach is based on an eigenvalue analysis of the dynamical systems associated with such functions.

In this work, we point to a new connection between DNNs expressivity and Sharkovsky’s Theorem from dynamical systems, that enables us to characterize the depth-width trade-offs of ReLU networks

We investigate low-bit quantization to reduce computational cost of deep neural network (DNN) based keyword spotting (KWS). We propose approaches to further reduce quantization bits via integrating quantization into keyword spotting model training, which we refer to as quantization-aware training. Our experimental results on large dataset indicate that quantization-aware training can recover performance models quantized to lower bits representations. By combining quantization-aware training and weight matrix factorization, we are able to significantly reduce model size and computation for small-footprint keyword spotting, while maintaining performance.

We investigate quantization-aware training in very low-bit quantized keyword spotters to reduce the cost of on-device keyword spotting.

Single-cell RNA-sequencing (scRNA-seq) is a powerful tool for analyzing biological systems. However, due to biological and technical noise, quantifying the effects of multiple experimental conditions presents an analytical challenge. To overcome this challenge, we developed MELD: Manifold Enhancement of Latent Dimensions. MELD leverages tools from graph signal processing to learn a latent dimension within the data scoring the prototypicality of each datapoint with respect to experimental or control conditions. We call this dimension the Enhanced Experimental Signal (EES). MELD learns the EES by filtering the noisy categorical experimental label in the graph frequency domain to recover a smooth signal with continuous values. This method can be used to identify signature genes that vary between conditions and identify which cell types are most affected by a given perturbation. We demonstrate the advantages of MELD analysis in two biological datasets, including T-cell activation in response to antibody-coated beads and treatment of human pancreatic islet cells with interferon gamma.

A novel graph signal processing framework for quantifying the effects of experimental perturbations in single cell biomedical data.

Models of user behavior are critical inputs in many prescriptive settings and can be viewed as decision rules that transform state information available to the user into actions. Gaussian processes (GPs), as well as nonlinear extensions thereof, provide a flexible framework to learn user models in conjunction with approximate Bayesian inference. However, the resulting models may not be interpretable in general. We propose decision-rule GPs (DRGPs) that apply GPs in a transformed space defined by decision rules that have immediate interpretability to practitioners. We illustrate this modeling tool on a real application and show that structural variational inference techniques can be used with DRGPs. We find that DRGPs outperform the direct use of GPs in terms of out-of-sample performance.

We propose a class of user models based on using Gaussian processes applied to a transformed space defined by decision rules

While Bayesian optimization (BO) has achieved great success in optimizing expensive-to-evaluate black-box functions, especially tuning hyperparameters of neural networks, methods such as random search (Li et al., 2016) and multi-fidelity BO (e.g. Klein et al. (2017)) that exploit cheap approximations, e.g. training on a smaller training data or with fewer iterations, can outperform standard BO approaches that use only full-fidelity observations. In this paper, we propose a novel Bayesian optimization algorithm, the continuous-fidelity knowledge gradient (cfKG) method, that can be used when fidelity is controlled by one or more continuous settings such as training data size and the number of training iterations. cfKG characterizes the value of the information gained by sampling a point at a given fidelity, choosing to sample at the point and fidelity with the largest value per unit cost. Furthermore, cfKG can be generalized, following Wu et al. (2017), to settings where derivatives are available in the optimization process, e.g. large-scale kernel learning, and where more than one point can be evaluated simultaneously. Numerical experiments show that cfKG outperforms state-of-art algorithms when optimizing synthetic functions, tuning convolutional neural networks (CNNs) on CIFAR-10 and SVHN, and in large-scale kernel learning.

We propose a Bayes-optimal Bayesian optimization algorithm for hyperparameter tuning by exploiting cheap approximations.

Neural networks trained only to optimize for training accuracy can often be fooled by adversarial examples --- slightly perturbed inputs misclassified with high confidence. Verification of networks enables us to gauge their vulnerability to such adversarial examples. We formulate verification of piecewise-linear neural networks as a mixed integer program. On a representative task of finding minimum adversarial distortions, our verifier is two to three orders of magnitude quicker than the state-of-the-art. We achieve this computational speedup via tight formulations for non-linearities, as well as a novel presolve algorithm that makes full use of all information available. The computational speedup allows us to verify properties on convolutional and residual networks with over 100,000 ReLUs --- several orders of magnitude more than networks previously verified by any complete verifier. In particular, we determine for the first time the exact adversarial accuracy of an MNIST classifier to perturbations with bounded l-∞ norm ε=0.1: for this classifier, we find an adversarial example for 4.38% of samples, and a certificate of robustness to norm-bounded perturbations for the remainder. Across all robust training procedures and network architectures considered, and for both the MNIST and CIFAR-10 datasets, we are able to certify more samples than the state-of-the-art and find more adversarial examples than a strong first-order attack.

We efficiently verify the robustness of deep neural models with over 100,000 ReLUs, certifying more samples than the state-of-the-art and finding more adversarial examples than a strong first-order attack.

Uncertainty estimation is an essential step in the evaluation of the robustness for deep learning models in computer vision, especially when applied in risk-sensitive areas. However, most state-of-the-art deep learning models either fail to obtain uncertainty estimation or need significant modification (e.g., formulating a proper Bayesian treatment) to obtain it. None of the previous methods are able to take an arbitrary model off the shelf and generate uncertainty estimation without retraining or redesigning it. To address this gap, we perform the first systematic exploration into training-free uncertainty estimation. We propose three simple and scalable methods to analyze the variance of output from a trained network under tolerable perturbations: infer-transformation, infer-noise, and infer-dropout. They operate solely during inference, without the need to re-train, re-design, or fine-tune the model, as typically required by other state-of-the-art uncertainty estimation methods. Surprisingly, even without involving such perturbations in training, our methods produce comparable or even better uncertainty estimation when compared to other training-required state-of-the-art methods. Last but not least, we demonstrate that the uncertainty from our proposed methods can be used to improve the neural network training.

A set of methods to obtain uncertainty estimation of any given model without re-designing, re-training, or to fine-tuning it.

Capturing spatiotemporal dynamics is an essential topic in video recognition. In this paper, we present learnable higher-order operation as a generic family of building blocks for capturing higher-order correlations from high dimensional input video space. We prove that several successful architectures for visual classification tasks are in the family of higher-order neural networks, theoretical and experimental analysis demonstrates their underlying mechanism is higher-order. On the task of video recognition, even using RGB only without fine-tuning with other video datasets, our higher-order models can achieve results on par with or better than the existing state-of-the-art methods on both Something-Something (V1 and V2) and Charades datasets.

Proposed higher order operation for context learning

Presently the most successful approaches to semi-supervised learning are based on consistency regularization, whereby a model is trained to be robust to small perturbations of its inputs and parameters. To understand consistency regularization, we conceptually explore how loss geometry interacts with training procedures. The consistency loss dramatically improves generalization performance over supervised-only training; however, we show that SGD struggles to converge on the consistency loss and continues to make large steps that lead to changes in predictions on the test data. Motivated by these observations, we propose to train consistency-based methods with Stochastic Weight Averaging (SWA), a recent approach which averages weights along the trajectory of SGD with a modified learning rate schedule. We also propose fast-SWA, which further accelerates convergence by averaging multiple points within each cycle of a cyclical learning rate schedule. With weight averaging, we achieve the best known semi-supervised results on CIFAR-10 and CIFAR-100, over many different quantities of labeled training data. For example, we achieve 5.0% error on CIFAR-10 with only 4000 labels, compared to the previous best result in the literature of 6.3%.

Consistency-based models for semi-supervised learning do not converge to a single point but continue to explore a diverse set of plausible solutions on the perimeter of a flat region. Weight averaging helps improve generalization performance.

In this paper, we find that by designing a novel loss function entitled, ''tracking loss'', Convolutional Neural Network (CNN) based object detectors can be successfully converted to well-performed visual trackers without any extra computational cost. This property is preferable to visual tracking where annotated video sequences for training are always absent, because rich features learned by detectors from still images could be utilized by dynamic trackers. It also avoids extra machinery such as feature engineering and feature aggregation proposed in previous studies. Tracking loss achieves this property by exploiting the internal structure of feature maps within the detection network and treating different feature points discriminatively. Such structure allows us to simultaneously consider discrimination quality and bounding box accuracy which is found to be crucial to the success. We also propose a network compression method to accelerate tracking speed without performance reduction. That also verifies tracking loss will remain highly effective even if the network is drastically compressed. Furthermore, if we employ a carefully designed tracking loss ensemble, the tracker would be much more robust and accurate. Evaluation results show that our trackers (including the ensemble tracker and two baseline trackers), outperform all state-of-the-art methods on VOT 2016 Challenge in terms of Expected Average Overlap (EAO) and robustness. We will make the code publicly available.

We successfully convert a popular detector RPN to a well-performed tracker from the viewpoint of loss function.