ICML 2018 Papers

Despite the success of generative adversarial networks (GANs) for image generation, the trade-off between visual quality and image diversity remains a significant issue. This paper achieves both aims simultaneously by improving the stability of training GANs. The key idea of the proposed approach is to implicitly regularize the discriminator using representative features. Focusing on the fact that standard GAN minimizes reverse Kullback-Leibler (KL) divergence, we transfer the representative feature, which is extracted from the data distribution using a pre-trained autoencoder (AE), to the discriminator of standard GANs. Because the AE learns to minimize forward KL divergence, our GAN training with representative features is influenced by both reverse and forward KL divergence. Consequently, the proposed approach is verified to improve visual quality and diversity of state of the art GANs using extensive evaluations.

Thompson sampling (TS) is an effective approach to trade off exploration and exploration in reinforcement learning. Despite its empirical success and recent advances, its theoretical analysis is often limited to the Bayesian setting, finite state-action spaces, or finite-horizon problems. In this paper, we study an instance of TS in the challenging setting of the infinite-horizon linear quadratic (LQ) control, which models problems with continuous state-action variables, linear dynamics, and quadratic cost. In particular, we analyze the regret in the frequentist sense (i.e., for a fixed unknown environment) in one-dimensional systems. We derive the first $O(\sqrt{T})$ frequentist regret bound for this problem, thus significantly improving the $O(T^{2/3})$ bound of Abeille & Lazaric (2017) and matching the frequentist performance derived by Abbasi-Yadkori & Szepesvári (2011) for an optimistic approach and the Bayesian result Ouyang et al. (2017) We obtain this result by developing a novel bound on the regret due to policy switches, which holds for LQ systems of any dimensionality and it allows updating the parameters and the policy at each step, thus overcoming previous limitations due to lazy updates. Finally, we report numerical simulations supporting the conjecture that our result extends to multi-dimensional systems.

We consider the transfer of experience samples (i.e., tuples < s, a, s’, r >) in reinforcement learning (RL), collected from a set of source tasks to improve the learning process in a given target task. Most of the related approaches focus on selecting the most relevant source samples for solving the target task, but then all the transferred samples are used without considering anymore the discrepancies between the task models. In this paper, we propose a model-based technique that automatically estimates the relevance (importance weight) of each source sample for solving the target task. In the proposed approach, all the samples are transferred and used by a batch RL algorithm to solve the target task, but their contribution to the learning process is proportional to their importance weight. By extending the results for importance weighting provided in supervised learning literature, we develop a finite-sample analysis of the proposed batch RL algorithm. Furthermore, we empirically compare the proposed algorithm to state-of-the-art approaches, showing that it achieves better learning performance and is very robust to negative transfer, even when some source tasks are significantly different from the target task.

Recent years have seen a flurry of activities in designing provably efficient nonconvex optimization procedures for solving statistical estimation problems. For various problems like phase retrieval or low-rank matrix completion, state-of-the-art nonconvex procedures require proper regularization (e.g. trimming, regularized cost, projection) in order to guarantee fast convergence. When it comes to vanilla procedures such as gradient descent, however, prior theory either recommends highly conservative learning rates to avoid overshooting, or completely lacks performance guarantees. This paper uncovers a striking phenomenon in several nonconvex problems: even in the absence of explicit regularization, gradient descent follows a trajectory staying within a basin that enjoys nice geometry, consisting of points incoherent with the sampling mechanism. This “implicit regularization” feature allows gradient descent to proceed in a far more aggressive fashion without overshooting, which in turn results in substantial computational savings. Focusing on two statistical estimation problems, i.e. solving random quadratic systems of equations and low-rank matrix completion, we establish that gradient descent achieves near-optimal statistical and computational guarantees without explicit regularization. As a byproduct, for noisy matrix completion, we demonstrate that gradient descent enables optimal control of both entrywise and spectral-norm errors.

In this work, we build on recent advances in distributional reinforcement learning to give a generally applicable, flexible, and state-of-the-art distributional variant of DQN. We achieve this by using quantile regression to approximate the full quantile function for the state-action return distribution. By reparameterizing a distribution over the sample space, this yields an implicitly defined return distribution and gives rise to a large class of risk-sensitive policies. We demonstrate improved performance on the 57 Atari 2600 games in the ALE, and use our algorithm’s implicitly defined distributions to study the effects of risk-sensitive policies in Atari games.

Image generation has been successfully cast as an autoregressive sequence generation or transformation problem. Recent work has shown that self-attention is an effective way of modeling textual sequences. In this work, we generalize a recently proposed model architecture based on self-attention, the Transformer, to a sequence modeling formulation of image generation with a tractable likelihood. By restricting the self-attention mechanism to attend to local neighborhoods we significantly increase the size of images the model can process in practice, despite maintaining significantly larger receptive fields per layer than typical convolutional neural networks. While conceptually simple, our generative models significantly outperform the current state of the art in image generation on ImageNet, improving the best published negative log-likelihood on ImageNet from 3.83 to 3.77. We also present results on image super-resolution with a large magnification ratio, applying an encoder-decoder configuration of our architecture. In a human evaluation study, we find that images generated by our super-resolution model fool human observers three times more often than the previous state of the art.

We develop differentially private methods for estimating various distributional properties. Given a sample from a discrete distribution p, some functional f, and accuracy and privacy parameters alpha and epsilon, the goal is to estimate f(p) up to accuracy alpha, while maintaining epsilon-differential privacy of the sample. We prove almost-tight bounds on the sample size required for this problem for several functionals of interest, including support size, support coverage, and entropy. We show that the cost of privacy is negligible in a variety of settings, both theoretically and experimentally. Our methods are based on a sensitivity analysis of several state-of-the-art methods for estimating these properties with sublinear sample complexities

In this work we aim to solve a large collection of tasks using a single reinforcement learning agent with a single set of parameters. A key challenge is to handle the increased amount of data and extended training time. We have developed a new distributed agent IMPALA (Importance Weighted Actor-Learner Architecture) that not only uses resources more efficiently in single-machine training but also scales to thousands of machines without sacrificing data efficiency or resource utilisation. We achieve stable learning at high throughput by combining decoupled acting and learning with a novel off-policy correction method called V-trace. We demonstrate the effectiveness of IMPALA for multi-task reinforcement learning on DMLab-30 (a set of 30 tasks from the DeepMind Lab environment (Beattie et al., 2016)) and Atari57 (all available Atari games in Arcade Learning Environment (Bellemare et al., 2013a)). Our results show that IMPALA is able to achieve better performance than previous agents with less data, and crucially exhibits positive transfer between tasks as a result of its multi-task approach.

Learning graph representations via low-dimensional embeddings that preserve relevant network properties is an important class of problems in machine learning. We here present a novel method to embed directed acyclic graphs. Following prior work, we first advocate for using hyperbolic spaces which provably model tree-like structures better than Euclidean geometry. Second, we view hierarchical relations as partial orders defined using a family of nested geodesically convex cones. We prove that these entailment cones admit an optimal shape with a closed form expression both in the Euclidean and hyperbolic spaces, and they canonically define the embedding learning process. Experiments show significant improvements of our method over strong recent baselines both in terms of representational capacity and generalization.

This paper considers the generation of prediction intervals (PIs) by neural networks for quantifying uncertainty in regression tasks. It is axiomatic that high-quality PIs should be as narrow as possible, whilst capturing a specified portion of data. We derive a loss function directly from this axiom that requires no distributional assumption. We show how its form derives from a likelihood principle, that it can be used with gradient descent, and that model uncertainty is accounted for in ensembled form. Benchmark experiments show the method outperforms current state-of-the-art uncertainty quantification methods, reducing average PI width by over 10%.

The computation of convolution layers in deep neural networks typically rely on high performance routines that trade space for time by using additional memory (either for packing purposes or required as part of the algorithm) to improve performance. The problems with such an approach are two-fold. First, these routines incur additional memory overhead which reduces the overall size of the network that can fit on embedded devices with limited memory capacity. Second, these high performance routines were not optimized for performing convolution, which means that the performance obtained is usually less than conventionally expected. In this paper, we demonstrate that direct convolution, when implemented correctly, eliminates all memory overhead, and yields performance that is between 10% to 400% times better than existing high performance implementations of convolution layers on conventional and embedded CPU architectures. We also show that a high performance direct convolution exhibits better scaling performance, i.e. suffers less performance drop, when increasing the number of threads.

End-to-end models for goal-orientated dialogue are challenging to train, because linguistic and strategic aspects are entangled in latent state vectors. We introduce an approach to learning representations of messages in dialogues by maximizing the likelihood of subsequent sentences and actions, which decouples the semantics of the dialogue utterance from its linguistic realization. We then use these latent sentence representations for hierarchical language generation, planning and reinforcement learning. Experiments show that our approach increases the end-task reward achieved by the model, improves the effectiveness of long-term planning using rollouts, and allows self-play reinforcement learning to improve decision making without diverging from human language. Our hierarchical latent-variable model outperforms previous work both linguistically and strategically.

One of the most challenging machine learning problems is a particular case of data classification in which classes are hierarchically structured and objects can be assigned to multiple paths of the class hierarchy at the same time. This task is known as hierarchical multi-label classification (HMC), with applications in text classification, image annotation, and in bioinformatics problems such as protein function prediction. In this paper, we propose novel neural network architectures for HMC called HMCN, capable of simultaneously optimizing local and global loss functions for discovering local hierarchical class-relationships and global information from the entire class hierarchy while penalizing hierarchical violations. We evaluate its performance in 21 datasets from four distinct domains, and we compare it against the current HMC state-of-the-art approaches. Results show that HMCN substantially outperforms all baselines with statistical significance, arising as the novel state-of-the-art for HMC.

Much of recent research has been devoted to video prediction and generation, yet most of the previous works have demonstrated only limited success in generating videos on short-term horizons. The hierarchical video prediction method by Villegas et al. (2017) is an example of a state-of-the-art method for long-term video prediction, but their method is limited because it requires ground truth annotation of high-level structures (e.g., human joint landmarks) at training time. Our network encodes the input frame, predicts a high-level encoding into the future, and then a decoder with access to the first frame produces the predicted image from the predicted encoding. The decoder also produces a mask that outlines the predicted foreground object (e.g., person) as a by-product. Unlike Villegas et al. (2017), we develop a novel training method that jointly trains the encoder, the predictor, and the decoder together without highlevel supervision; we further improve upon this by using an adversarial loss in the feature space to train the predictor. Our method can predict about 20 seconds into the future and provides better results compared to Denton and Fergus (2018) and Finn et al. (2016) on the Human 3.6M dataset.

We study how to effectively leverage expert feedback to learn sequential decision-making policies. We focus on problems with sparse rewards and long time horizons, which typically pose significant challenges in reinforcement learning. We propose an algorithmic framework, called hierarchical guidance, that leverages the hierarchical structure of the underlying problem to integrate different modes of expert interaction. Our framework can incorporate different combinations of imitation learning (IL) and reinforcement learning (RL) at different levels, leading to dramatic reductions in both expert effort and cost of exploration. Using long-horizon benchmarks, including Montezuma’s Revenge, we demonstrate that our approach can learn significantly faster than hierarchical RL, and be significantly more label-efficient than standard IL. We also theoretically analyze labeling cost for certain instantiations of our framework.

Multi-Rate Multivariate Time Series (MR-MTS) are the multivariate time series observations which come with various sampling rates and encode multiple temporal dependencies. State-space models such as Kalman filters and deep learning models such as deep Markov models are mainly designed for time series data with the same sampling rate and cannot capture all the dependencies present in the MR-MTS data. To address this challenge, we propose the Multi-Rate Hierarchical Deep Markov Model (MR-HDMM), a novel deep generative model which uses the latent hierarchical structure with a learnable switch mechanism to capture the temporal dependencies of MR-MTS. Experimental results on two real-world datasets demonstrate that our MR-HDMM model outperforms the existing state-of-the-art deep learning and state-space models on forecasting and interpolation tasks. In addition, the latent hierarchies in our model provide a way to show and interpret the multiple temporal dependencies.

Hierarchical clustering is a popular unsupervised data analysis method. For many real-world applications, we would like to exploit prior information about the data that imposes constraints on the clustering hierarchy, and is not captured by the set of features available to the algorithm. This gives rise to the problem of hierarchical clustering with structural constraints. Structural constraints pose major challenges for bottom-up approaches like average/single linkage and even though they can be naturally incorporated into top-down divisive algorithms, no formal guarantees exist on the quality of their output. In this paper, we provide provable approximation guarantees for two simple top-down algorithms, using a recently introduced optimization viewpoint of hierarchical clustering with pairwise similarity information (Dasgupta, 2016). We show how to find good solutions even in the presence of conflicting prior information, by formulating a constraint-based regularization of the objective. Furthemore, we explore a variation of this objective for dissimilarity information (Cohen-Addad et al., 2018) and improve upon current techniques. Finally, we demonstrate our approach on a real dataset for the taxonomy application.

We analyze the performance of the greedy algorithm, and also a discrete semi-gradient based algorithm, for maximizing the sum of a suBmodular and suPermodular (BP) function (both of which are non-negative monotone non-decreasing) under two types of constraints, either a cardinality constraint or $p\geq 1$ matroid independence constraints. These problems occur naturally in several real-world applications in data science, machine learning, and artificial intelligence. The problems are ordinarily inapproximable to any factor. Using the curvature $\curv_f$ of the submodular term, and introducing $\curv^g$ for the supermodular term (a natural dual curvature for supermodular functions), however, both of which are computable in linear time, we show that BP maximization can be efficiently approximated by both the greedy and the semi-gradient based algorithm. The algorithms yield multiplicative guarantees of $\frac{1}{\curv_f}\left[1-e^{-(1-\curv^g)\curv_f}\right]$ and $\frac{1-\curv^g}{(1-\curv^g)\curv_f + p}$ for the two types of constraints respectively. For pure monotone supermodular constrained maximization, these yield $1-\curvg$ and $(1-\curvg)/p$ for the two types of constraints respectively. We also analyze the hardness of BP maximization and show that our guarantees match hardness by a constant factor and by $O(\ln(p))$ respectively. Computational experiments are also provided supporting our analysis.

We consider the problem of learning high-dimensional Gaussian graphical models. The graphical lasso is one of the most popular methods for estimating Gaussian graphical models. However, it does not achieve the oracle rate of convergence. In this paper, we propose the graphical nonconvex optimization for optimal estimation in Gaussian graphical models, which is then approximated by a sequence of convex programs. Our proposal is computationally tractable and produces an estimator that achieves the oracle rate of convergence. The statistical error introduced by the sequential approximation using a sequence of convex programs is clearly demonstrated via a contraction property. The proposed methodology is then extended to modeling semiparametric graphical models. We show via numerical studies that the proposed estimator outperforms other popular methods for estimating Gaussian graphical models.