NeurIPS 2018 Papers

We describe a neural network-based system for text-to-speech (TTS) synthesis that is able to generate speech audio in the voice of many different speakers, including those unseen during training. Our system consists of three independently trained components: (1) a speaker encoder network, trained on a speaker verification task using an independent dataset of noisy speech from thousands of speakers without transcripts, to generate a fixed-dimensional embedding vector from seconds of reference speech from a target speaker; (2) a sequence-to-sequence synthesis network based on Tacotron 2, which generates a mel spectrogram from text, conditioned on the speaker embedding; (3) an auto-regressive WaveNet-based vocoder that converts the mel spectrogram into a sequence of time domain waveform samples. We demonstrate that the proposed model is able to transfer the knowledge of speaker variability learned by the discriminatively-trained speaker encoder to the new task, and is able to synthesize natural speech from speakers that were not seen during training. We quantify the importance of training the speaker encoder on a large and diverse speaker set in order to obtain the best generalization performance. Finally, we show that randomly sampled speaker embeddings can be used to synthesize speech in the voice of novel speakers dissimilar from those used in training, indicating that the model has learned a high quality speaker representation.

How to leverage the temporal dimension is a key question in video analysis. Recent works suggest an efficient approach to video feature learning, i.e., factorizing 3D convolutions into separate components respectively for spatial and temporal convolutions. The temporal convolution, however, comes with an implicit assumption – the feature maps across time steps are well aligned so that the features at the same locations can be aggregated. This assumption may be overly strong in practical applications, especially in action recognition where the motion serves as a crucial cue. In this work, we propose a new CNN architecture TrajectoryNet, which incorporates trajectory convolution, a new operation for integrating features along the temporal dimension, to replace the existing temporal convolution. This operation explicitly takes into account the changes in contents caused by deformation or motion, allowing the visual features to be aggregated along the the motion paths, trajectories. On two large-scale action recognition datasets, namely, Something-Something and Kinetics, the proposed network architecture achieves notable improvement over strong baselines.

Recently developed deep-learning-based denoisers often outperform state-of-the-art conventional denoisers, such as the BM3D. They are typically trained to minimizethe mean squared error (MSE) between the output image of a deep neural networkand a ground truth image. In deep learning based denoisers, it is important to use high quality noiseless ground truth data for high performance, but it is often challenging or even infeasible to obtain noiseless images in application areas such as hyperspectral remote sensing and medical imaging. In this article, we propose a method based on Stein’s unbiased risk estimator (SURE) for training deep neural network denoisers only based on the use of noisy images. We demonstrate that our SURE-based method, without the use of ground truth data, is able to train deep neural network denoisers to yield performances close to those networks trained with ground truth, and to outperform the state-of-the-art denoiser BM3D. Further improvements were achieved when noisy test images were used for training of denoiser networks using our proposed SURE-based method.

Training a neural network using backpropagation algorithm requires passing error gradients sequentially through the network. The backward locking prevents us from updating network layers in parallel and fully leveraging the computing resources. Recently, there are several works trying to decouple and parallelize the backpropagation algorithm. However, all of them suffer from severe accuracy loss or memory explosion when the neural network is deep. To address these challenging issues, we propose a novel parallel-objective formulation for the objective function of the neural network. After that, we introduce features replay algorithm and prove that it is guaranteed to converge to critical points for the non-convex problem under certain conditions. Finally, we apply our method to training deep convolutional neural networks, and the experimental results show that the proposed method achieves {faster} convergence, {lower} memory consumption, and {better} generalization error than compared methods.

The state-of-the-art hardware platforms for training deep neural networks are moving from traditional single precision (32-bit) computations towards 16 bits of precision - in large part due to the high energy efficiency and smaller bit storage associated with using reduced-precision representations. However, unlike inference, training with numbers represented with less than 16 bits has been challenging due to the need to maintain fidelity of the gradient computations during back-propagation. Here we demonstrate, for the first time, the successful training of deep neural networks using 8-bit floating point numbers while fully maintaining the accuracy on a spectrum of deep learning models and datasets. In addition to reducing the data and computation precision to 8 bits, we also successfully reduce the arithmetic precision for additions (used in partial product accumulation and weight updates) from 32 bits to 16 bits through the introduction of a number of key ideas including chunk-based accumulation and floating point stochastic rounding. The use of these novel techniques lays the foundation for a new generation of hardware training platforms with the potential for 2-4 times improved throughput over today's systems.

In theory, importance sampling speeds up stochastic gradient algorithms for supervised learning by prioritizing training examples. In practice, the cost of computing importances greatly limits the impact of importance sampling. We propose a robust, approximate importance sampling procedure (RAIS) for stochastic gradient de- scent. By approximating the ideal sampling distribution using robust optimization, RAIS provides much of the benefit of exact importance sampling with drastically reduced overhead. Empirically, we find RAIS-SGD and standard SGD follow similar learning curves, but RAIS moves faster through these paths, achieving speed-ups of at least 20% and sometimes much more.

The wide adoption of DNNs has given birth to unrelenting computing requirements, forcing datacenter operators to adopt domain-specific accelerators to train them. These accelerators typically employ densely packed full-precision floating-point arithmetic to maximize performance per area. Ongoing research efforts seek to further increase that performance density by replacing floating-point with fixed-point arithmetic. However, a significant roadblock for these attempts has been fixed point's narrow dynamic range, which is insufficient for DNN training convergence. We identify block floating point (BFP) as a promising alternative representation since it exhibits wide dynamic range and enables the majority of DNN operations to be performed with fixed-point logic. Unfortunately, BFP alone introduces several limitations that preclude its direct applicability. In this work, we introduce HBFP, a hybrid BFP-FP approach, which performs all dot products in BFP and other operations in floating point. HBFP delivers the best of both worlds: the high accuracy of floating point at the superior hardware density of fixed point. For a wide variety of models, we show that HBFP matches floating point's accuracy while enabling hardware implementations that deliver up to 8.5x higher throughput.

Learning in small sample regimes is among the most remarkable features of the human perceptual system. This ability is related to robustness to transformations, which is acquired through visual experience in the form of weak- or self-supervision during development. We explore the idea of allowing artificial systems to learn representations of visual stimuli through weak supervision prior to downstream supervised tasks. We introduce a novel loss function for representation learning using unlabeled image sets and video sequences, and experimentally demonstrate that these representations support one-shot learning and reduce the sample complexity of multiple recognition tasks. We establish the existence of a trade-off between the sizes of weakly supervised, automatically obtained from video sequences, and fully supervised data sets. Our results suggest that equivalence sets other than class labels, which are abundant in unlabeled visual experience, can be used for self-supervised learning of semantically relevant image embeddings.

It is widely believed that learning good representations is one of the main reasons for the success of deep neural networks. Although highly intuitive, there is a lack of theory and systematic approach quantitatively characterizing what representations do deep neural networks learn. In this work, we move a tiny step towards a theory and better understanding of the representations. Specifically, we study a simpler problem: How similar are the representations learned by two networks with identical architecture but trained from different initializations. We develop a rigorous theory based on the neuron activation subspace match model. The theory gives a complete characterization of the structure of neuron activation subspace matches, where the core concepts are maximum match and simple match which describe the overall and the finest similarity between sets of neurons in two networks respectively. We also propose efficient algorithms to find the maximum match and simple matches. Finally, we conduct extensive experiments using our algorithms. Experimental results suggest that, surprisingly, representations learned by the same convolutional layers of networks trained from different initializations are not as similar as prevalently expected, at least in terms of subspace match.

Asynchronous momentum stochastic gradient descent algorithms (Async-MSGD) have been widely used in distributed machine learning, e.g., training large collaborative filtering systems and deep neural networks. Due to current technical limit, however, establishing convergence properties of Async-MSGD for these highly complicated nonoconvex problems is generally infeasible. Therefore, we propose to analyze the algorithm through a simpler but nontrivial nonconvex problems --- streaming PCA. This allows us to make progress toward understanding Aync-MSGD and gaining new insights for more general problems. Specifically, by exploiting the diffusion approximation of stochastic optimization, we establish the asymptotic rate of convergence of Async-MSGD for streaming PCA. Our results indicate a fundamental tradeoff between asynchrony and momentum: To ensure convergence and acceleration through asynchrony, we have to reduce the momentum (compared with Sync-MSGD). To the best of our knowledge, this is the first theoretical attempt on understanding Async-MSGD for distributed nonconvex stochastic optimization. Numerical experiments on both streaming PCA and training deep neural networks are provided to support our findings for Async-MSGD.

Recent progress in deep generative models has been fueled by two paradigms -- autoregressive and adversarial models. We propose a combination of both approaches with the goal of learning generative models of text. Our method first produces a high-level sentence outline and then generates words sequentially, conditioning on both the outline and the previous outputs. We generate outlines with an adversarial model trained to approximate the distribution of sentences in a latent space induced by general-purpose sentence encoders. This provides strong, informative conditioning for the autoregressive stage. Our quantitative evaluations suggests that conditioning information from generated outlines is able to guide the autoregressive model to produce realistic samples, comparable to maximum-likelihood trained language models, even at high temperatures with multinomial sampling. Qualitative results also demonstrate that this generative procedure yields natural-looking sentences and interpolations.

Most recent work on interpretability of complex machine learning models has focused on estimating a-posteriori explanations for previously trained models around specific predictions. Self-explaining models where interpretability plays a key role already during learning have received much less attention. We propose three desiderata for explanations in general -- explicitness, faithfulness, and stability -- and show that existing methods do not satisfy them. In response, we design self-explaining models in stages, progressively generalizing linear classifiers to complex yet architecturally explicit models. Faithfulness and stability are enforced via regularization specifically tailored to such models. Experimental results across various benchmark datasets show that our framework offers a promising direction for reconciling model complexity and interpretability.

Although the recent progress is substantial, deep learning methods can be vulnerable to the maliciously generated adversarial examples. In this paper, we present a novel training procedure and a thresholding test strategy, towards robust detection of adversarial examples. In training, we propose to minimize the reverse cross-entropy (RCE), which encourages a deep network to learn latent representations that better distinguish adversarial examples from normal ones. In testing, we propose to use a thresholding strategy as the detector to filter out adversarial examples for reliable predictions. Our method is simple to implement using standard algorithms, with little extra training cost compared to the common cross-entropy minimization. We apply our method to defend various attacking methods on the widely used MNIST and CIFAR-10 datasets, and achieve significant improvements on robust predictions under all the threat models in the adversarial setting.

There has been growing interest in using neural networks and deep learning techniques to create dialogue systems. Conversational recommendation is an interesting setting for the scientific exploration of dialogue with natural language as the associated discourse involves goal-driven dialogue that often transforms naturally into more free-form chat. This paper provides two contributions. First, until now there has been no publicly available large-scale data set consisting of real-world dialogues centered around recommendations. To address this issue and to facilitate our exploration here, we have collected ReDial, a data set consisting of over 10,000 conversations centered around the theme of providing movie recommendations. We make this data available to the community for further research. Second, we use this dataset to explore multiple facets of conversational recommendations. In particular we explore new neural architectures, mechanisms and methods suitable for composing conversational recommendation systems. Our dataset allows us to systematically probe model sub-components addressing different parts of the overall problem domain ranging from: sentiment analysis and cold-start recommendation generation to detailed aspects of how natural language is used in this setting in the real world. We combine such sub-components into a full-blown dialogue system and examine its behavior.

Backpropagation and the chain rule of derivatives have been prominent; however, the total derivative rule has not enjoyed the same amount of attention. In this work we show how the total derivative rule leads to an intuitive visual framework for creating gradient estimators on graphical models. In particular, previous ”policy gradient theorems” are easily derived. We derive new gradient estimators based on density estimation, as well as a likelihood ratio gradient, which ”jumps” to an intermediate node, not directly to the objective function. We evaluate our methods on model-based policy gradient algorithms, achieve good performance, and present evidence towards demystifying the success of the popular PILCO algorithm.

Identifying the top-K frequent items is one of the most common and important operations in large data processing systems. As a result, several solutions have been proposed to solve this problem approximately. In this paper, we identify that in modern distributed settings with both multi-node as well as multi-core parallelism, existing algorithms, although theoretically sound, are suboptimal from the performance perspective. In particular, for identifying top-K frequent items, Count-Min Sketch (CMS) has fantastic update time but lack the important property of reducibility which is needed for exploiting available massive data parallelism. On the other end, popular Frequent algorithm (FA) leads to reducible summaries but the update costs are significant. In this paper, we present Topkapi, a fast and parallel algorithm for finding top-K frequent items, which gives the best of both worlds, i.e., it is reducible as well as efficient update time similar to CMS. Topkapi possesses strong theoretical guarantees and leads to significant performance gains due to increased parallelism, relative to past work.

Online learning to rank is a sequential decision-making problem where in each round the learning agent chooses a list of items and receives feedback in the form of clicks from the user. Many sample-efficient algorithms have been proposed for this problem that assume a specific click model connecting rankings and user behavior. We propose a generalized click model that encompasses many existing models, including the position-based and cascade models. Our generalization motivates a novel online learning algorithm based on topological sort, which we call TopRank. TopRank is (a) more natural than existing algorithms, (b) has stronger regret guarantees than existing algorithms with comparable generality, (c) has a more insightful proof that leaves the door open to many generalizations, (d) outperforms existing algorithms empirically.

Real-world learning systems have practical limitations on the quality and quantity of the training datasets that they can collect and consider. How should a system go about choosing a subset of the possible training examples that still allows for learning accurate, generalizable models? To help address this question, we draw inspiration from a highly efficient practical learning system: the human child. Using head-mounted cameras, eye gaze trackers, and a model of foveated vision, we collected first-person (egocentric) images that represents a highly accurate approximation of the "training data" that toddlers' visual systems collect in everyday, naturalistic learning contexts. We used state-of-the-art computer vision learning models (convolutional neural networks) to help characterize the structure of these data, and found that child data produce significantly better object models than egocentric data experienced by adults in exactly the same environment. By using the CNNs as a modeling tool to investigate the properties of the child data that may enable this rapid learning, we found that child data exhibit a unique combination of quality and diversity, with not only many similar large, high-quality object views but also a greater number and diversity of rare views. This novel methodology of analyzing the visual "training data" used by children may not only reveal insights to improve machine learning, but also may suggest new experimental tools to better understand infant learning in developmental psychology.

Knowing when a classifier's prediction can be trusted is useful in many applications and critical for safely using AI. While the bulk of the effort in machine learning research has been towards improving classifier performance, understanding when a classifier's predictions should and should not be trusted has received far less attention. The standard approach is to use the classifier's discriminant or confidence score; however, we show there exists an alternative that is more effective in many situations. We propose a new score, called the {\it trust score}, which measures the agreement between the classifier and a modified nearest-neighbor classifier on the testing example. We show empirically that high (low) trust scores produce surprisingly high precision at identifying correctly (incorrectly) classified examples, consistently outperforming the classifier's confidence score as well as many other baselines. Further, under some mild distributional assumptions, we show that if the trust score for an example is high (low), the classifier will likely agree (disagree) with the Bayes-optimal classifier. Our guarantees consist of non-asymptotic rates of statistical consistency under various nonparametric settings and build on recent developments in topological data analysis.