NeurIPS 2021 Papers

Gradient compression is a recent and increasingly popular technique for reducing the communication cost in distributed training of large-scale machine learning models. In this work we focus on developing efficient distributed methods that can work for any compressor satisfying a certain contraction property, which includes both unbiased (after appropriate scaling) and biased compressors such as RandK and TopK. Applied naively, gradient compression introduces errors that either slow down convergence or lead to divergence. A popular technique designed to tackle this issue is error compensation/error feedback. Due to the difficulties associated with analyzing biased compressors, it is not known whether gradient compression with error compensation can be combined with acceleration. In this work, we show for the first time that error compensated gradient compression methods can be accelerated. In particular, we propose and study the error compensated loopless Katyusha method, and establish an accelerated linear convergence rate under standard assumptions. We show through numerical experiments that the proposed method converges with substantially fewer communication rounds than previous error compensated algorithms.

We study whether and how can we model a joint distribution $p(x,z)$ using two conditional models $p(x|z)$ and $q(z|x)$ that form a cycle. This is motivated by the observation that deep generative models, in addition to a likelihood model $p(x|z)$, often also use an inference model $q(z|x)$ for extracting representation, but they rely on a usually uninformative prior distribution $p(z)$ to define a joint distribution, which may render problems like posterior collapse and manifold mismatch. To explore the possibility to model a joint distribution using only $p(x|z)$ and $q(z|x)$, we study their compatibility and determinacy, corresponding to the existence and uniqueness of a joint distribution whose conditional distributions coincide with them. We develop a general theory for operable equivalence criteria for compatibility, and sufficient conditions for determinacy. Based on the theory, we propose a novel generative modeling framework CyGen that only uses the two cyclic conditional models. We develop methods to achieve compatibility and determinacy, and to use the conditional models to fit and generate data. With the prior constraint removed, CyGen better fits data and captures more representative features, supported by both synthetic and real-world experiments.

A key challenge with machine learning approaches for ranking is the gap between the performance metrics of interest and the surrogate loss functions that can be optimized with gradient-based methods. This gap arises because ranking metrics typically involve a sorting operation which is not differentiable w.r.t. the model parameters. Prior works have proposed surrogates that are loosely related to ranking metrics or simple smoothed versions thereof, and often fail to scale to real-world applications. We propose PiRank, a new class of differentiable surrogates for ranking, which employ a continuous, temperature-controlled relaxation to the sorting operator based on NeuralSort [1]. We show that PiRank exactly recovers the desired metrics in the limit of zero temperature and further propose a divide-and-conquer extension that scales favorably to large list sizes, both in theory and practice. Empirically, we demonstrate the role of larger list sizes during training and show that PiRank significantly improves over comparable approaches on publicly available Internet-scale learning-to-rank benchmarks.

We study solution methods for (strongly-)convex-(strongly)-concave Saddle-Point Problems (SPPs) over networks of two type--master/workers (thus centralized) architectures and mesh (thus decentralized) networks. The local functions at each node are assumed to be \textit{similar}, due to statistical data similarity or otherwise. We establish lower complexity bounds for a fairly general class of algorithms solving the SPP. We show that a given suboptimality $\epsilon>0$ is achieved over master/workers networks in $\Omega\big(\Delta\cdot \delta/\mu\cdot \log (1/\varepsilon)\big)$ rounds of communications, where $\delta>0$ measures the degree of similarity of the local functions, $\mu$ is their strong convexity constant, and $\Delta$ is the diameter of the network. The lower communication complexity bound over mesh networks reads $\Omega\big(1/{\sqrt{\rho}} \cdot {\delta}/{\mu}\cdot\log (1/\varepsilon)\big)$, where $\rho$ is the (normalized) eigengap of the gossip matrix used for the communication between neighbouring nodes. We then propose algorithms matching the lower bounds over either types of networks (up to log-factors). We assess the effectiveness of the proposed algorithms on a robust regression problem.

We study the following learning problem with dependent data: Given a trajectory of length $n$ from a stationary Markov chain with $k$ states, the goal is to predict the distribution of the next state. For $3 \leq k \leq O(\sqrt{n})$, the optimal prediction risk in the Kullback-Leibler divergence is shown to be $\Theta(\frac{k^2}{n}\log \frac{n}{k^2})$, in contrast to the optimal rate of $\Theta(\frac{\log \log n}{n})$ for $k=2$ previously shown in Falahatgar et al in 2016. These nonparametric rates can be attributed to the memory in the data, as the spectral gap of the Markov chain can be arbitrarily small. To quantify the memory effect, we study irreducible reversible chains with a prescribed spectral gap. In addition to characterizing the optimal prediction risk for two states, we show that, as long as the spectral gap is not excessively small, the prediction risk in the Markov model is $O(\frac{k^2}{n})$, which coincides with that of an iid model with the same number of parameters.

It is well known that modern deep neural networks are powerful enough to memorize datasets even when the labels have been randomized. Recently, Vershynin(2020) settled a long standing question by Baum(1988), proving that deep threshold networks can memorize $n$ points in $d$ dimensions using $\widetilde{\mathcal{O}}(e^{1/\delta^2}+\sqrt{n})$ neurons and $\widetilde{\mathcal{O}}(e^{1/\delta^2}(d+\sqrt{n})+n)$ weights, where $\delta$ is the minimum distance between the points. In this work, we improve the dependence on $\delta$ from exponential to almost linear, proving that $\widetilde{\mathcal{O}}(\frac{1}{\delta}+\sqrt{n})$ neurons and $\widetilde{\mathcal{O}}(\frac{d}{\delta}+n)$ weights are sufficient. Our construction uses Gaussian random weights only in the first layer, while all the subsequent layers use binary or integer weights. We also prove new lower bounds by connecting memorization in neural networks to the purely geometric problem of separating $n$ points on a sphere using hyperplanes.

The effectiveness of machine learning algorithms arises from being able to extract useful features from large amounts of data. As model and dataset sizes increase, dataset distillation methods that compress large datasets into significantly smaller yet highly performant ones will become valuable in terms of training efficiency and useful feature extraction. To that end, we apply a novel distributed kernel-based meta-learning framework to achieve state-of-the-art results for dataset distillation using infinitely wide convolutional neural networks. For instance, using only 10 datapoints (0.02% of original dataset), we obtain over 65% test accuracy on CIFAR-10 image classification task, a dramatic improvement over the previous best test accuracy of 40%. Our state-of-the-art results extend across many other settings for MNIST, Fashion-MNIST, CIFAR-10, CIFAR-100, and SVHN. Furthermore, we perform some preliminary analyses of our distilled datasets to shed light on how they differ from naturally occurring data.

Task-oriented dialogue (ToD) benchmarks provide an important avenue to measure progress and develop better conversational agents. However, existing datasets for end-to-end ToD modeling are limited to a single language, hindering the development of robust end-to-end ToD systems for multilingual countries and regions. Here we introduce BiToD, the first bilingual multi-domain dataset for end-to-end task-oriented dialogue modeling. BiToD contains over 7k multi-domain dialogues (144k utterances) with a large and realistic bilingual knowledge base. It serves as an effective benchmark for evaluating bilingual ToD systems and cross-lingual transfer learning approaches. We provide state-of-the-art baselines under three evaluation settings (monolingual, bilingual, and cross-lingual). The analysis of our baselines in different settings highlights 1) the effectiveness of training a bilingual ToD system comparing to two independent monolingual ToD systems, and 2) the potential of leveraging a bilingual knowledge base and cross-lingual transfer learning to improve the system performance in the low resource condition.

To deal with ambiguities in partial multilabel learning (PML), state-of-the-art methods perform disambiguation by identifying ground-truth labels directly. However, there is an essential question:“Can the ground-truth labels be identified precisely?". If yes, “How can the ground-truth labels be found?". This paper provides affirmative answers to these questions. Instead of adopting hand-made heuristic strategy, we propose a novel Mutual Information Label Identification for Partial Multilabel Learning (MILI-PML), which is derived from a clear probabilistic formulation and could be easily interpreted theoretically from the mutual information perspective, as well as naturally incorporates the feature/label relevancy considerations. Extensive experiments on synthetic and real-world datasets clearly demonstrate the superiorities of the proposed MILI-PML.

Machine learning models that are developed to be invariant under certain types of data transformations have shown improved generalization in practice. However, a principled understanding of why invariance benefits generalization is limited. Given a dataset, there is often no principled way to select "suitable" data transformations under which model invariance guarantees better generalization. This paper studies the generalization benefit of model invariance by introducing the sample cover induced by transformations, i.e., a representative subset of a dataset that can approximately recover the whole dataset using transformations. For any data transformations, we provide refined generalization bounds for invariant models based on the sample cover. We also characterize the "suitability" of a set of data transformations by the sample covering number induced by transformations, i.e., the smallest size of its induced sample covers. We show that we may tighten the generalization bounds for "suitable" transformations that have a small sample covering number. In addition, our proposed sample covering number can be empirically evaluated and thus provides a guidance for selecting transformations to develop model invariance for better generalization. In experiments on multiple datasets, we evaluate sample covering numbers for some commonly used transformations and show that the smaller sample covering number for a set of transformations (e.g., the 3D-view transformation) indicates a smaller gap between the test and training error for invariant models, which verifies our propositions.

Adder neural networks (ANNs) are designed for low energy cost which replace expensive multiplications in convolutional neural networks (CNNs) with cheaper additions to yield energy-efficient neural networks and hardware accelerations. Although ANNs achieve satisfactory efficiency, there exist gaps between ANNs and CNNs where the accuracy of ANNs can hardly be compared to CNNs without the assistance of other training tricks, such as knowledge distillation. The inherent discrepancy lies in the similarity measurement between filters and features, however how to alleviate this difference remains unexplored. To locate the potential problem of ANNs, we focus on the property difference due to similarity measurement. We demonstrate that unordered heavy tails in ANNs could be the key component which prevents ANNs from achieving superior classification performance since fatter tails tend to overlap in feature space. Through pre-defining Multivariate Skew Laplace distributions and embedding feature distributions into the loss function, ANN features can be fully controlled and designed for various properties. We further present a novel method for tackling existing heavy tails in ANNs with only a modification of classifier where ANN features are clustered with their tails well-formulated through proposed angle-based constraint on the distribution parameters to encourage high diversity of tails. Experiments conducted on several benchmarks and comparison with other distributions demonstrate the effectiveness of proposed approach for boosting the performance of ANNs.

Tabular datasets are the last "unconquered castle" for deep learning, with traditional ML methods like Gradient-Boosted Decision Trees still performing strongly even against recent specialized neural architectures. In this paper, we hypothesize that the key to boosting the performance of neural networks lies in rethinking the joint and simultaneous application of a large set of modern regularization techniques. As a result, we propose regularizing plain Multilayer Perceptron (MLP) networks by searching for the optimal combination/cocktail of 13 regularization techniques for each dataset using a joint optimization over the decision on which regularizers to apply and their subsidiary hyperparameters. We empirically assess the impact of these regularization cocktails for MLPs in a large-scale empirical study comprising 40 tabular datasets and demonstrate that (i) well-regularized plain MLPs significantly outperform recent state-of-the-art specialized neural network architectures, and (ii) they even outperform strong traditional ML methods, such as XGBoost.

Methods for inference and simulation of linearly constrained Gaussian Markov Random Fields (GMRF) are computationally prohibitive when the number of constraints is large. In some cases, such as for intrinsic GMRFs, they may even be unfeasible. We propose a new class of methods to overcome these challenges in the common case of sparse constraints, where one has a large number of constraints and each only involves a few elements. Our methods rely on a basis transformation into blocks of constrained versus non-constrained subspaces, and we show that the methods greatly outperform existing alternatives in terms of computational cost. By combining the proposed methods with the stochastic partial differential equation approach for Gaussian random fields, we also show how to formulate Gaussian process regression with linear constraints in a GMRF setting to reduce computational cost. This is illustrated in two applications with simulated data.

Label noise generally degenerates the performance of deep learning algorithms because deep neural networks easily overfit label errors. Let $X$ and $Y$ denote the instance and clean label, respectively. When $Y$ is a cause of $X$, according to which many datasets have been constructed, e.g., \textit{SVHN} and \textit{CIFAR}, the distributions of $P(X)$ and $P(Y|X)$ are generally entangled. This means that the unsupervised instances are helpful to learn the classifier and thus reduce the side effect of label noise. However, it remains elusive on how to exploit the causal information to handle the label-noise problem. We propose to model and make use of the causal process in order to correct the label-noise effect. Empirically, the proposed method outperforms all state-of-the-art methods on both synthetic and real-world label-noise datasets.

Learning disentangled representations for user intentions from multi-feedback (i.e., positive and negative feedback) can enhance the accuracy and explainability of recommendation algorithms. However, learning such disentangled representations from multi-feedback data is challenging because i) multi-feedback is complex: there exist complex relations among different types of feedback (e.g., click, unclick, and dislike, etc) as well as various user intentions, and ii) multi-feedback is noisy: there exists noisy (useless) information both in features and labels, which may deteriorate the recommendation performance. Existing works on disentangled representation learning only focus on positive feedback, failing to handle the complex relations and noise hidden in multi-feedback data. To solve this problem, in this work we propose a Curriculum Disentangled Recommendation (CDR) model that is capable of efficiently learning disentangled representations from complex and noisy multi-feedback for better recommendation. Concretely, we design a co-filtering dynamic routing mechanism that simultaneously captures the complex relations among different behavioral feedback and user intentions as well as denoise the representations in the feature level. We then present an adjustable self-evaluating curriculum that is able to evaluate sample difficulties for better model training and conduct denoising in the label level via disregarding useless information. Our extensive experiments on several real-world datasets demonstrate that the proposed CDR model can significantly outperform several state-of-the-art methods in terms of recommendation accuracy.

Training deep neural networks on large datasets can often be accelerated by using multiple compute nodes. This approach, known as distributed training, can utilize hundreds of computers via specialized message-passing protocols such as Ring All-Reduce. However, running these protocols at scale requires reliable high-speed networking that is only available in dedicated clusters. In contrast, many real-world applications, such as federated learning and cloud-based distributed training, operate on unreliable devices with unstable network bandwidth. As a result, these applications are restricted to using parameter servers or gossip-based averaging protocols. In this work, we lift that restriction by proposing Moshpit All-Reduce — an iterative averaging protocol that exponentially converges to the global average. We demonstrate the efficiency of our protocol for distributed optimization with strong theoretical guarantees. The experiments show 1.3x speedup for ResNet-50 training on ImageNet compared to competitive gossip-based strategies and 1.5x speedup when training ALBERT-large on preemptible compute nodes.

We propose to personalize a 2D human pose estimator given a set of test images of a person without using any manual annotations. While there is a significant advancement in human pose estimation, it is still very challenging for a model to generalize to different unknown environments and unseen persons. Instead of using a fixed model for every test case, we adapt our pose estimator during test time to exploit person-specific information. We first train our model on diverse data with both a supervised and a self-supervised pose estimation objectives jointly. We use a Transformer model to build a transformation between the self-supervised keypoints and the supervised keypoints. During test time, we personalize and adapt our model by fine-tuning with the self-supervised objective. The pose is then improved by transforming the updated self-supervised keypoints. We experiment with multiple datasets and show significant improvements on pose estimations with our self-supervised personalization. Project page with code is available at https://liyz15.github.io/TTP/.

The accuracy of simulation-based forecasting in chaotic systems is heavily dependent on high-quality estimates of the system state at the beginning of the forecast. Data assimilation methods are used to infer these initial conditions by systematically combining noisy, incomplete observations and numerical models of system dynamics to produce highly effective estimation schemes. We introduce a self-supervised framework, which we call \textit{amortized assimilation}, for learning to assimilate in dynamical systems. Amortized assimilation combines deep learning-based denoising with differentiable simulation, using independent neural networks to assimilate specific observation types while connecting the gradient flow between these sub-tasks with differentiable simulation and shared recurrent memory. This hybrid architecture admits a self-supervised training objective which is minimized by an unbiased estimator of the true system state even in the presence of only noisy training data. Numerical experiments across several chaotic benchmark systems highlight the improved effectiveness of our approach compared to widely-used data assimilation methods.

Three Operator Splitting (TOS) (Davis & Yin, 2017) can minimize the sum of multiple convex functions effectively when an efficient gradient oracle or proximal operator is available for each term. This requirement often fails in machine learning applications: (i) instead of full gradients only stochastic gradients may be available; and (ii) instead of proximal operators, using subgradients to handle complex penalty functions may be more efficient and realistic. Motivated by these concerns, we analyze three potentially valuable extensions of TOS. The first two permit using subgradients and stochastic gradients, and are shown to ensure a $\mathcal{O}(1/\sqrt{t})$ convergence rate. The third extension AdapTOS endows TOS with adaptive step-sizes. For the important setting of optimizing a convex loss over the intersection of convex sets AdapTOS attains universal convergence rates, i.e., the rate adapts to the unknown smoothness degree of the objective. We compare our proposed methods with competing methods on various applications.