Neural networks have a remarkable capacity for contextual processing—using recent or nearby inputs to modify processing of current input. For example, in natural language, contextual processing is necessary to correctly interpret negation (e.g. phrases such as "not bad"). However, our ability to understand how networks process context is limited. Here, we propose general methods for reverse engineering recurrent neural networks (RNNs) to identify and elucidate contextual processing. We apply these methods to understand RNNs trained on sentiment classification. This analysis reveals inputs that induce contextual effects, quantifies the strength and timescale of these effects, and identifies sets of these inputs with similar properties. Additionally, we analyze contextual effects related to differential processing of the beginning and end of documents. Using the insights learned from the RNNs we improve baseline Bag-of-Words models with simple extensions that incorporate contextual modification, recovering greater than 90% of the RNN's performance increase over the baseline. This work yields a new understanding of how RNNs process contextual information, and provides tools that should provide similar insight more broadly.

Learning from unordered sets is a fundamental learning setup, which is attracting increasing attention. Research in this area has focused on the case where elements of the set are represented by feature vectors, and far less emphasis has been given to the common case where set elements themselves adhere to certain symmetries. That case is relevant to numerous applications, from deblurring image bursts to multi-view 3D shape recognition and reconstruction. In this paper, we present a principled approach to learning sets of general symmetric elements. We first characterize the space of linear layers that are equivariant both to element reordering and to the inherent symmetries of elements, like translation in the case of images. We further show that networks that are composed of these layers, called Deep Sets for Symmetric elements layers (DSS), are universal approximators of both invariant and equivariant functions. DSS layers are also straightforward to implement. Finally, we show that they improve over existing set-learning architectures in a series of experiments with images, graphs, and point-clouds.

Many AI problems, in robotics and other domains, are goal-directed, essentially seeking a trajectory leading to some goal state. Reinforcement learning (RL), building on Bellman's optimality equation, naturally optimizes for a single goal, yet can be made goal-directed by augmenting the state with the goal. Instead, we propose a new RL framework, derived from a dynamic programming equation for the all pairs shortest path (APSP) problem, which naturally solves goal-directed queries. We show that this approach has computational benefits for both standard and approximate dynamic programming. Interestingly, our formulation prescribes a novel protocol for computing a trajectory: instead of predicting the next state given its predecessor, as in standard RL, a goal-conditioned trajectory is constructed by first predicting an intermediate state between start and goal, partitioning the trajectory into two. Then, recursively, predicting intermediate points on each sub-segment, until a complete trajectory is obtained. We call this trajectory structure a sub-goal tree. Building on it, we additionally extend the policy gradient methodology to recursively predict sub-goals, resulting in novel goal-based algorithms. Finally, we apply our method to neural motion planning, where we demonstrate significant improvements compared to standard RL on navigating a 7-DoF robot arm between obstacles.

Recently there have been several attempts to extend Nesterov's accelerated algorithm to smooth stochastic and variance-reduced optimization. In this paper, we show that there is a simpler approach to acceleration: applying optimistic online learning algorithms and querying the gradient oracle at the online average of the intermediate optimization iterates. In particular, we tighten a recent result of Cutkosky (2019) to demonstrate theoretically that online iterate averaging results in a reduced optimization gap, independently of the algorithm involved. We show that carefully combining this technique with existing generic optimistic online learning algorithms yields the optimal accelerated rates for optimizing strongly-convex and non-strongly-convex, possibly composite objectives, with deterministic as well as stochastic first-order oracles. We further extend this idea to variance-reduced optimization. Finally, we also provide ``universal'' algorithms that achieve the optimal rate for smooth and non-smooth composite objectives simultaneously without further tuning, generalizing the results of Kavis et al. (2019) and solving a number of their open problems.

In self-supervised visual representation learning, a feature extractor is trained on a "pretext task" for which labels can be generated cheaply, without human annotation. A central challenge in this approach is that the feature extractor quickly learns to exploit low-level visual features such as color aberrations or watermarks and then fails to learn useful semantic representations. Much work has gone into identifying such "shortcut" features and hand-designing schemes to reduce their effect. Here, we propose a general framework for mitigating the effect shortcut features. Our key assumption is that those features which are the first to be exploited for solving the pretext task may also be the most vulnerable to an adversary trained to make the task harder. We show that this assumption holds across common pretext tasks and datasets by training a "lens" network to make small image changes that maximally reduce performance in the pretext task. Representations learned with the modified images outperform those learned without in all tested cases. Additionally, the modifications made by the lens reveal how the choice of pretext task and dataset affects the features learned by self-supervision.

Learning a good representation is an essential component for deep reinforcement learning (RL). Representation learning is especially important in multitask and partially observable settings where building a representation of the unknown environment is crucial to solve the tasks. Here we introduce Predictions of Bootstrapped Latents (PBL), a simple and flexible self-supervised representation learning algorithm for multitask deep RL. PBL builds on multistep predictive representations of future observations, and focuses on capturing structured information about environment dynamics. Specifically, PBL trains its representation by predicting latent embeddings of future observations. These latent embeddings are themselves trained to be predictive of the aforementioned representations. These predictions form a bootstrapping effect, allowing the agent to learn more about the key aspects of the environment dynamics. In addition, by defining prediction tasks completely in latent space, PBL provides the flexibility of using multimodal observations involving pixel images, language instructions, rewards and more. We show in our experiments that PBL delivers across-the-board improved performance over state of the art deep RL agents in the DMLab-30 multitask setting.

Initialization, normalization, and skip connections are believed to be three indispensable techniques for training very deep convolutional neural networks and obtaining state-of-the-art performance. This paper shows that deep vanilla ConvNets without normalization nor skip connections can also be trained to achieve surprisingly good performance on standard image recognition benchmarks. This is achieved by enforcing the convolution kernels to be near isometric during initialization and training, as well as by using a variant of ReLU that is shifted towards being isometric. Further experiments show that if combined with skip connections, such near isometric networks can achieve performances on par with (for ImageNet) and better than (for COCO) the standard ResNet, even without normalization at all. Our code is available at https://github.com/HaozhiQi/ISONet.

Likelihood-free (a.k.a. simulation-based) inference problems are inverse problems with expensive, or intractable, forward models. ODE inverse problems are commonly treated as likelihood-free, as their forward map has to be numerically approximated by an ODE solver. This, however, is not a fundamental constraint but just a lack of functionality in classic ODE solvers, which do not return a likelihood but a point estimate. To address this shortcoming, we employ Gaussian ODE filtering (a probabilistic numerical method for ODEs) to construct a local Gaussian approximation to the likelihood. This approximation yields tractable estimators for the gradient and Hessian of the (log-)likelihood. Insertion of these estimators into existing gradient-based optimization and sampling methods engenders new solvers for ODE inverse problems. We demonstrate that these methods outperform standard likelihood-free approaches on three benchmark-systems.

In order to compute fast approximations to the singular value decompositions (SVD) of very large matrices, randomized sketching algorithms have become a leading approach. However, a key practical difficulty of sketching an SVD is that the user does not know how far the sketched singular vectors/values are from the exact ones. Indeed, the user may be forced to rely on analytical worst-case error bounds, which may not account for the unique structure of a given problem. As a result, the lack of tools for error estimation often leads to much more computation than is really necessary. To overcome these challenges, this paper develops a fully data-driven bootstrap method that numerically estimates the actual error of sketched singular vectors/values. Furthermore, the method is computationally inexpensive, because it operates only on sketched objects, and hence it requires no extra passes over the full matrix being factored.

Many clustering algorithms lead to cluster assignments that are hard to explain, partially because they depend on all the features of the data in a complicated way. To improve interpretability, we consider using a small decision tree to partition a data set into clusters, so that clusters can be characterized in a straightforward manner. We study this problem from a theoretical viewpoint, measuring cluster quality by the k-means and k-medians objectives. In terms of negative results, we show that popular top-down decision tree algorithms may lead to clusterings with arbitrarily large cost, and any clustering based on a tree with k leaves must incur an Omega(log k) approximation factor compared to the optimal clustering. On the positive side, for two means/medians, we show that a single threshold cut can achieve a constant factor approximation, and we give nearly-matching lower bounds; for general k > 2, we design an efficient algorithm that leads to an O(k) approximation to the optimal k-medians and an O(k^2) approximation to the optimal k-means. Prior to our work, no algorithms were known with provable guarantees independent of dimension and input size.

We consider learning a multi-class classification model in the federated setting, where each user has access to the positive data associated with only a single class. As a result, during each federated learning round, the users need to locally update the classifier without having access to the features and the model parameters for the negative classes. Thus, naively employing conventional decentralized learning such as distributed SGD or Federated Averaging may lead to trivial or extremely poor classifiers. In particular, for embedding based classifiers, all the class embeddings might collapse to a single point. To address this problem, we propose a generic framework for training with only positive labels, namely Federated Averaging with Spreadout (FedAwS), where the server imposes a geometric regularizer after each round to encourage classes to be spreadout in the embedding space. We show, both theoretically and empirically, that FedAwS can almost match the performance of conventional learning where users have access to negative labels. We further extend the proposed method to settings with large output spaces.

Space-filling designs such as Low Discrepancy Sequence (LDS), Latin Hypercube Sampling (LHS) and Jittered Sampling (JS) were proposed for fully parallel hyperparameter search, and were shown to be more effective than random and grid search. We prove that LHS and JS outperform random search only by a constant factor. Consequently, we introduce a new sampling approach based on the reshaping of the search distribution, and we show both theoretically and numerically that it leads to significant gains over random search. Two methods are proposed for the reshaping: Recentering (when the distribution of the optimum is known), and Cauchy transformation (when the distribution of the optimum is unknown). The proposed methods are first validated on artificial experiments and simple real-world tests on clustering and Salmon mappings. Then we demonstrate that they drive performance improvement in a wide range of expensive artificial intelligence tasks, namely attend/infer/repeat, video next frame segmentation forecasting and progressive generative adversarial networks.

We present a description of the function space and the smoothness class associated with a convolutional network using the machinery of reproducing kernel Hilbert spaces. We show that the mapping associated with a convolutional network expands into a sum involving elementary functions akin to spherical harmonics. The functional decomposition can be related to functional ANOVA decompositions in nonparametric statistics. Building off this functional characterization, we obtain statistical bounds which highlight an interesting trade-off between the approximation error and the estimation error.

Optimally solving decentralized partially observable Markov decision processes under either full or no information sharing received significant attention in recent years. However, little is known about how partial information sharing affects existing theory and algorithms. This paper addresses this question for a team of two agents, with one-sided information sharing---\ie both agents have imperfect information about the state of the world, but only one has access to what the other sees and does. From the perspective of a central planner, we show that the original problem can be reformulated into an equivalent information-state Markov decision process and solved as such. Besides, we prove that the optimal value function exhibits a specific form of uniform continuity. We also present a heuristic search algorithm utilizing this property and providing the first results for this family of problems.

In model-based reinforcement learning, planning with an imperfect model of the environment has the potential to harm learning progress. But even when a model is imperfect, it may still contain information that is useful for planning. In this paper, we investigate the idea of using an imperfect model selectively. The agent should plan in parts of the state space where the model would be helpful but refrain from using the model where it would be harmful. An effective selective planning mechanism requires estimating predictive uncertainty, which arises out of aleatoric uncertainty, parameter uncertainty, and model inadequacy, among other sources. Prior work has focused on parameter uncertainty for selective planning. In this work, we emphasize the importance of model inadequacy. We show that heteroscedastic regression can signal predictive uncertainty arising from model inadequacy that is complementary to that which is detected by methods designed for parameter uncertainty, indicating that considering both parameter uncertainty and model inadequacy may be a more promising direction for effective selective planning than either in isolation.

Annotating datasets is one of the main costs in nowadays supervised learning. The goal of weak supervision is to enable models to learn using only forms of labelling which are cheaper to collect, as partial labelling. This is a type of incomplete annotation where, for each datapoint, supervision is cast as a set of labels containing the real one. The problem of supervised learning with partial labelling has been studied for specific instances such as classification, multi-label, ranking or segmentation, but a general framework is still missing. This paper provides a unified framework based on structured prediction and on the concept of {\em infimum loss} to deal with partial labelling over a wide family of learning problems and loss functions. The framework leads naturally to explicit algorithms that can be easily implemented and for which proved statistical consistency and learning rates. Experiments confirm the superiority of the proposed approach over commonly used baselines.

Machine learning and deep learning in particular has been recently used to successfully address many tasks in the domain of code including -- finding and fixing bugs, code completion, decompilation, malware detection, type inference and many others. However, the issue of adversarial robustness of models for code has gone largely unnoticed. In this work, we explore this issue by: (i) instantiating adversarial attacks for code (a domain with discrete and highly structured inputs), (ii) showing that, similar to other domains, neural models for code are vulnerable to adversarial attacks, and (iii) developing a set of novel techniques that enable training robust and accurate models of code.

Attribution methods aim to explain a neural network's prediction by highlighting the most relevant image areas. A popular approach is to backpropagate (BP) a custom relevance score using modified rules, rather than the gradient. We analyze an extensive set of modified BP methods: Deep Taylor Decomposition, Layer-wise Relevance Propagation (LRP), Excitation BP, PatternAttribution, DeepLIFT, Deconv, RectGrad, and Guided BP. We find empirically that the explanations of all mentioned methods, except for DeepLIFT, are independent of the parameters of later layers. We provide theoretical insights for this surprising behavior and also analyze why DeepLIFT does not suffer from this limitation. Empirically, we measure how information of later layers is ignored by using our new metric, cosine similarity convergence (CSC). The paper provides a framework to assess the faithfulness of new and existing modified BP methods theoretically and empirically.

Acquiring abilities in the absence of a task-oriented reward function is at the frontier of reinforcement learning research. This problem has been studied through the lens of empowerment, which draws a connection between option discovery and information theory. Information-theoretic skill discovery methods have garnered much interest from the community, but little research has been conducted in understanding their limitations. Through theoretical analysis and empirical evidence, we show that existing algorithms suffer from a common limitation -- they discover options that provide a poor coverage of the state space. In light of this, we propose Explore, Discover and Learn (EDL), an alternative approach to information-theoretic skill discovery. Crucially, EDL optimizes the same information-theoretic objective derived from the empowerment literature, but addresses the optimization problem using different machinery. We perform an extensive evaluation of skill discovery methods on controlled environments and show that EDL offers significant advantages, such as overcoming the coverage problem, reducing the dependence of learned skills on the initial state, and allowing the user to define a prior over which behaviors should be learned.

We propose a class of kernel-based two-sample tests, which aim to determine whether two sets of samples are drawn from the same distribution. Our tests are constructed from kernels parameterized by deep neural nets, trained to maximize test power. These tests adapt to variations in distribution smoothness and shape over space, and are especially suited to high dimensions and complex data. By contrast, the simpler kernels used in prior kernel testing work are spatially homogeneous, and adaptive only in lengthscale. We explain how this scheme includes popular classifier-based two-sample tests as a special case, but improves on them in general. We provide the first proof of consistency for the proposed adaptation method, which applies both to kernels on deep features and to simpler radial basis kernels or multiple kernel learning. In experiments, we establish the superior performance of our deep kernels in hypothesis testing on benchmark and real-world data. The code of our deep-kernel-based two-sample tests is available at github.com/fengliu90/DK-for-TST.

Machine-learned predictors, although achieving very good results for inputs resembling training data, cannot possibly provide perfect predictions in all situations. Still, decision-making systems that are based on such predictors need not only to benefit from good predictions but also to achieve a decent performance when the predictions are inadequate. In this paper, we propose a prediction setup for arbitrary metrical task systems (MTS) (e.g., caching, k-server and convex body chasing) and online matching on the line. We utilize results from the theory of online algorithms to show how to make the setup robust. Specifically for caching, we present an algorithm whose performance, as a function of the prediction error, is exponentially better than what is achievable for general MTS. Finally, we present an empirical evaluation of our methods on real world datasets, which suggests practicality.

With continual miniaturization ever more applications of deep learning can be found in embedded systems, where it is common to encounter data with natural representation in the complex domain. To this end we extend Sparse Variational Dropout to complex-valued neural networks and verify the proposed Bayesian technique by conducting a large numerical study of the performance-compression trade-off of C-valued networks on two tasks: image recognition on MNIST-like and CIFAR10 datasets and music transcription on MusicNet. We replicate the state-of-the-art result by Trabelsi et al. (2018) on MusicNet with a complex-valued network compressed by 50-100x at a small performance penalty.

Typical architectures of Generative Adversarial Networks make use of a unimodal latent/input distribution transformed by a continuous generator. Consequently, the modeled distribution always has connected support which is cumbersome when learning a disconnected set of manifolds. We formalize this problem by establishing a "no free lunch" theorem for the disconnected manifold learning stating an upper-bound on the precision of the targeted distribution. This is done by building on the necessary existence of a low-quality region where the generator continuously samples data between two disconnected modes. Finally, we derive a rejection sampling method based on the norm of generator’s Jacobian and show its efficiency on several generators including BigGAN.

We present a framework for autonomously learning a portable representation that describes a collection of low-level continuous environments. We show that these abstract representations can be learned in a task-independent egocentric space specific to the agent that, when grounded with problem-specific information, are provably sufficient for planning. We demonstrate transfer in two different domains, where an agent learns a portable, task-independent symbolic vocabulary, as well as operators expressed in that vocabulary, and then learns to instantiate those operators on a per-task basis. This reduces the number of samples required to learn a representation of a new task.

Generating training examples for supervised tasks is a long sought after goal in AI. We study the problem of heart signal electrocardiogram (ECG) synthesis for improved heartbeat classification. ECG synthesis is challenging: the generation of training examples for such biological-physiological systems is not straightforward, due to their dynamic nature in which the various parts of the system interact in complex ways. However, an understanding of these dynamics has been developed for years in the form of mathematical process simulators. We study how to incorporate this knowledge into the generative process by leveraging a biological simulator for the task of ECG classification. Specifically, we use a system of ordinary differential equations representing heart dynamics, and incorporate this ODE system into the optimization process of a generative adversarial network to create biologically plausible ECG training examples. We perform empirical evaluation and show that heart simulation knowledge during the generation process improves ECG classification.

Off-policy policy estimators that use importance sampling (IS) can suffer from high variance in long-horizon domains, and there has been particular excitement over new IS methods that leverage the structure of Markov decision processes. We analyze the variance of the most popular approaches through the viewpoint of conditional Monte Carlo. Surprisingly, we find that in finite horizon MDPs there is no strict variance reduction of per-decision importance sampling or stationary importance sampling, comparing with vanilla importance sampling. We then provide sufficient conditions under which the per-decision or stationary estimators will provably reduce the variance over importance sampling with finite horizons. For the asymptotic (in terms of horizon $T$) case, we develop upper and lower bounds on the variance of those estimators which yields sufficient conditions under which there exists an exponential v.s. polynomial gap between the variance of importance sampling and that of the per-decision or stationary estimators. These results help advance our understanding of if and when new types of IS estimators will improve the accuracy of off-policy estimation.

We propose a direct-to-word sequence model which uses a word network to learn word embeddings from letters. The word network can be integrated seamlessly with arbitrary sequence models including Connectionist Temporal Classification and encoder-decoder models with attention. We show our direct-to-word model can achieve word error rate gains over sub-word level models for speech recognition. We also show that our direct-to-word approach retains the ability to predict words not seen at training time without any retraining. Finally, we demonstrate that a word-level model can use a larger stride than a sub-word level model while maintaining accuracy. This makes the model more efficient both for training and inference.