NeurIPS 2021 Papers

The Hilbert-Schmidt Independence Criterion (HSIC) is a powerful kernel-based statistic for assessing the generalized dependence between two multivariate variables. However, independence testing based on the HSIC is not directly possible for cluster-correlated data. Such a correlation pattern among the observations arises in many practical situations, e.g., family-based and longitudinal data, and requires proper accommodation. Therefore, we propose a novel HSIC-based independence test to evaluate the dependence between two multivariate variables based on cluster-correlated data. Using the previously proposed empirical HSIC as our test statistic, we derive its asymptotic distribution under the null hypothesis of independence between the two variables but in the presence of sample correlation. Based on both simulation studies and real data analysis, we show that, with clustered data, our approach effectively controls type I error and has a higher statistical power than competing methods.

Humans and animals explore their environment and acquire useful skills even in the absence of clear goals, exhibiting intrinsic motivation. The study of intrinsic motivation in artificial agents is concerned with the following question: what is a good general-purpose objective for an agent? We study this question in dynamic partially-observed environments, and argue that a compact and general learning objective is to minimize the entropy of the agent's state visitation estimated using a latent state-space model. This objective induces an agent to both gather information about its environment, corresponding to reducing uncertainty, and to gain control over its environment, corresponding to reducing the unpredictability of future world states. We instantiate this approach as a deep reinforcement learning agent equipped with a deep variational Bayes filter. We find that our agent learns to discover, represent, and exercise control of dynamic objects in a variety of partially-observed environments sensed with visual observations without extrinsic reward.

Works on lottery ticket hypothesis (LTH) and single-shot network pruning (SNIP) have raised a lot of attention currently on post-training pruning (iterative magnitude pruning), and before-training pruning (pruning at initialization). The former method suffers from an extremely large computation cost and the latter usually struggles with insufficient performance. In comparison, during-training pruning, a class of pruning methods that simultaneously enjoys the training/inference efficiency and the comparable performance, temporarily, has been less explored. To better understand during-training pruning, we quantitatively study the effect of pruning throughout training from the perspective of pruning plasticity (the ability of the pruned networks to recover the original performance). Pruning plasticity can help explain several other empirical observations about neural network pruning in literature. We further find that pruning plasticity can be substantially improved by injecting a brain-inspired mechanism called neuroregeneration, i.e., to regenerate the same number of connections as pruned. We design a novel gradual magnitude pruning (GMP) method, named gradual pruning with zero-cost neuroregeneration (GraNet), that advances state of the art. Perhaps most impressively, its sparse-to-sparse version for the first time boosts the sparse-to-sparse training performance over various dense-to-sparse methods with ResNet-50 on ImageNet without extending the training time. We release all codes in https://github.com/Shiweiliuiiiiiii/GraNet.

Almost every single image restoration problem has a closely related parameter, such as the scale factor in super-resolution, the noise level in image denoising, and the quality factor in JPEG deblocking. Although recent studies on image restoration problems have achieved great success due to the development of deep neural networks, they handle the parameter involved in an unsophisticated way. Most previous researchers either treat problems with different parameter levels as independent tasks, and train a specific model for each parameter level; or simply ignore the parameter, and train a single model for all parameter levels. The two popular approaches have their own shortcomings. The former is inefficient in computing and the latter is ineffective in performance. In this work, we propose a novel system called functional neural network (FuncNet) to solve a parametric image restoration problem with a single model. Unlike a plain neural network, the smallest conceptual element of our FuncNet is no longer a floating-point variable, but a function of the parameter of the problem. This feature makes it both efficient and effective for a parametric problem. We apply FuncNet to super-resolution, image denoising, and JPEG deblocking. The experimental results show the superiority of our FuncNet on all three parametric image restoration tasks over the state of the arts.

Spectral normalization (SN) is a widely-used technique for improving the stability and sample quality of Generative Adversarial Networks (GANs). However, current understanding of SN's efficacy is limited. In this work, we show that SN controls two important failure modes of GAN training: exploding and vanishing gradients. Our proofs illustrate a (perhaps unintentional) connection with the successful LeCun initialization. This connection helps to explain why the most popular implementation of SN for GANs requires no hyper-parameter tuning, whereas stricter implementations of SN have poor empirical performance out-of-the-box. Unlike LeCun initialization which only controls gradient vanishing at the beginning of training, SN preserves this property throughout training. Building on this theoretical understanding, we propose a new spectral normalization technique: Bidirectional Scaled Spectral Normalization (BSSN), which incorporates insights from later improvements to LeCun initialization: Xavier initialization and Kaiming initialization. Theoretically, we show that BSSN gives better gradient control than SN. Empirically, we demonstrate that it outperforms SN in sample quality and training stability on several benchmark datasets.

Large scale distributed optimization has become the default tool for the training of supervised machine learning models with a large number of parameters and training data. Recent advancements in the field provide several mechanisms for speeding up the training, including {\em compressed communication}, {\em variance reduction} and {\em acceleration}. However, none of these methods is capable of exploiting the inherently rich data-dependent smoothness structure of the local losses beyond standard smoothness constants. In this paper, we argue that when training supervised models, {\em smoothness matrices}---information-rich generalizations of the ubiquitous smoothness constants---can and should be exploited for further dramatic gains, both in theory and practice. In order to further alleviate the communication burden inherent in distributed optimization, we propose a novel communication sparsification strategy that can take full advantage of the smoothness matrices associated with local losses. To showcase the power of this tool, we describe how our sparsification technique can be adapted to three distributed optimization algorithms---DCGD, DIANA and ADIANA---yielding significant savings in terms of communication complexity. The new methods always outperform the baselines, often dramatically so.

Determinantal Point Process (DPPs) are statistical models for repulsive point patterns. Both sampling and inference are tractable for DPPs, a rare feature among models with negative dependence that explains their popularity in machine learning and spatial statistics. Parametric and nonparametric inference methods have been proposed in the finite case, i.e. when the point patterns live in a finite ground set. In the continuous case, only parametric methods have been investigated, while nonparametric maximum likelihood for DPPs -- an optimization problem over trace-class operators -- has remained an open question. In this paper, we show that a restricted version of this maximum likelihood (MLE) problem falls within the scope of a recent representer theorem for nonnegative functions in an RKHS. This leads to a finite-dimensional problem, with strong statistical ties to the original MLE. Moreover, we propose, analyze, and demonstrate a fixed point algorithm to solve this finite-dimensional problem. Finally, we also provide a controlled estimate of the correlation kernel of the DPP, thus providing more interpretability.

Pretrained language models have achieved state-of-the-art performance when adapted to a downstream NLP task. However, theoretical analysis of these models is scarce and challenging since the pretraining and downstream tasks can be very different. We propose an analysis framework that links the pretraining and downstream tasks with an underlying latent variable generative model of text -- the downstream classifier must recover a function of the posterior distribution over the latent variables. We analyze head tuning (learning a classifier on top of the frozen pretrained model) and prompt tuning in this setting. The generative model in our analysis is either a Hidden Markov Model (HMM) or an HMM augmented with a latent memory component, motivated by long-term dependencies in natural language. We show that 1) under certain non-degeneracy conditions on the HMM, simple classification heads can solve the downstream task, 2) prompt tuning obtains downstream guarantees with weaker non-degeneracy conditions, and 3) our recovery guarantees for the memory-augmented HMM are stronger than for the vanilla HMM because task-relevant information is easier to recover from the long-term memory. Experiments on synthetically generated data from HMMs back our theoretical findings.

We show a generic reduction from multiclass differentially private PAC learning to binary private PAC learning. We apply this transformation to a recently proposed binary private PAC learner to obtain a private multiclass learner with sample complexity that has a polynomial dependence on the multiclass Littlestone dimension and a poly-logarithmic dependence on the number of classes. This yields a doubly exponential improvement in the dependence on both parameters over learners from previous work. Our proof extends the notion of $\Psi$-dimension defined in work of Ben-David et al. [JCSS, 1995] to the online setting and explores its general properties.

Understanding how the brain constructs statistical models of the sensory world remains a longstanding challenge for computational neuroscience. Here, we derive an unsupervised local synaptic plasticity rule that trains neural circuits to infer latent structure from sensory stimuli via a novel loss function for approximate online Bayesian inference. The learning algorithm is driven by a local error signal computed between two factors that jointly contribute to neural activity: stimulus drive and internal predictions --- the network's 'impression' of the stimulus. Physiologically, we associate these two components with the basal and apical dendrites of pyramidal neurons, respectively. We show that learning can be implemented online, is capable of capturing temporal dependencies in continuous input streams, and generalizes to hierarchical architectures. Furthermore, we demonstrate both analytically and empirically that the algorithm is more data-efficient than a three-factor plasticity alternative, enabling it to learn statistics of high-dimensional, naturalistic inputs. Overall, the model provides a bridge from mechanistic accounts of synaptic plasticity to algorithmic descriptions of unsupervised probabilistic learning and inference.

The lack of accessibility to data on policing has severely limited researchers’ ability to conduct thorough quantitative analyses on police activity and behavior, particularly with regard to predicting and explaining police violence. In the present work, we provide a new dataset that contains information on the personnel, activities, use of force, and complaints in the Chicago Police Department (CPD). The raw data, obtained from the CPD via a series of requests under the Freedom of Information Act (FOIA), consists of 35 unlinked, inconsistent, and undocumented spreadsheets. Our paper provides a cleaned, linked, and documented version of this data that can be reproducibly generated via open source code. We provide a detailed description of the dataset contents, the procedures for cleaning the data, and summary statistics. The data have a rich variety of uses, such as prediction (e.g., predicting misconduct from officer traits, experience, and assigned units), network analysis (e.g., detecting communities within the social network of officers co-listed on complaints), spatiotemporal data analysis (e.g., investigating patterns of officer shooting events), causal inference (e.g., tracking the effects of new disciplinary practices, new training techniques, and new oversight on complaints and use of force), and much more. Access to this dataset will enable the machine learning community to meaningfully engage with the problem of police violence.

In this paper, we devise identity tests for ranking data that is generated from Mallows model both in the \emph{asymptotic} and \emph{non-asymptotic} settings. First we consider the case when the central ranking is known, and devise two algorithms for testing the spread parameter of the Mallows model. The first one is obtained by constructing a Uniformly Most Powerful Unbiased (UMPU) test in the asymptotic setting and then converting it into a sample-optimal non-asymptotic identity test. The resulting test is, however, impractical even for medium sized data, because it requires computing the distribution of the sufficient statistic. The second non-asymptotic test is derived from an optimal learning algorithm for the Mallows model. This test is both easy to compute and is sample-optimal for a wide range of parameters. Next, we consider testing Mallows models for the unknown central ranking case. This case can be tackled in the asymptotic setting by introducing a bias that exponentially decays with the sample size. We support all our findings with extensive numerical experiments and show that the proposed tests scale gracefully with the number of items to be ranked.

We show that bringing intermediate layers' representations of two augmented versions of an image closer together in self-supervised learning helps to improve the momentum contrastive (MoCo) method. To this end, in addition to the contrastive loss, we minimize the mean squared error between the intermediate layer representations or make their cross-correlation matrix closer to an identity matrix. Both loss objectives either outperform standard MoCo, or achieve similar performances on three diverse medical imaging datasets: NIH-Chest Xrays, Breast Cancer Histopathology, and Diabetic Retinopathy. The gains of the improved MoCo are especially large in a low-labeled data regime (e.g. 1% labeled data) with an average gain of 5% across three datasets. We analyze the models trained using our novel approach via feature similarity analysis and layer-wise probing. Our analysis reveals that models trained via our approach have higher feature reuse compared to a standard MoCo and learn informative features earlier in the network. Finally, by comparing the output probability distribution of models fine-tuned on small versus large labeled data, we conclude that our proposed method of pre-training leads to lower Kolmogorov–Smirnov distance, as compared to a standard MoCo. This provides additional evidence that our proposed method learns more informative features in the pre-training phase which could be leveraged in a low-labeled data regime.

Methods to find counterfactual explanations have predominantly focused on one-step decision making processes. In this work, we initiate the development of methods to find counterfactual explanations for decision making processes in which multiple, dependent actions are taken sequentially over time. We start by formally characterizing a sequence of actions and states using finite horizon Markov decision processes and the Gumbel-Max structural causal model. Building upon this characterization, we formally state the problem of finding counterfactual explanations for sequential decision making processes. In our problem formulation, the counterfactual explanation specifies an alternative sequence of actions differing in at most k actions from the observed sequence that could have led the observed process realization to a better outcome. Then, we introduce a polynomial time algorithm based on dynamic programming to build a counterfactual policy that is guaranteed to always provide the optimal counterfactual explanation on every possible realization of the counterfactual environment dynamics. We validate our algorithm using both synthetic and real data from cognitive behavioral therapy and show that the counterfactual explanations our algorithm finds can provide valuable insights to enhance sequential decision making under uncertainty.

In many sequential decision making applications, the change of decision would bring an additional cost, such as the wear-and-tear cost associated with changing server status. To control the switching cost, we introduce the problem of online convex optimization with continuous switching constraint, where the goal is to achieve a small regret given a budget on the \emph{overall} switching cost. We first investigate the hardness of the problem, and provide a lower bound of order $\Omega(\sqrt{T})$ when the switching cost budget $S=\Omega(\sqrt{T})$, and $\Omega(\min\{\frac{T}{S},T\})$ when $S=O(\sqrt{T})$, where $T$ is the time horizon. The essential idea is to carefully design an adaptive adversary, who can adjust the loss function according to the cumulative switching cost of the player incurred so far based on the orthogonal technique. We then develop a simple gradient-based algorithm which enjoys the minimax optimal regret bound. Finally, we show that, for strongly convex functions, the regret bound can be improved to $O(\log T)$ for $S=\Omega(\log T)$, and $O(\min\{T/\exp(S)+S,T\})$ for $S=O(\log T)$.

Vision Transformers (ViT) have achieved remarkable success in large-scale image recognition. They split every 2D image into a fixed number of patches, each of which is treated as a token. Generally, representing an image with more tokens would lead to higher prediction accuracy, while it also results in drastically increased computational cost. To achieve a decent trade-off between accuracy and speed, the number of tokens is empirically set to 16x16 or 14x14. In this paper, we argue that every image has its own characteristics, and ideally the token number should be conditioned on each individual input. In fact, we have observed that there exist a considerable number of “easy” images which can be accurately predicted with a mere number of 4x4 tokens, while only a small fraction of “hard” ones need a finer representation. Inspired by this phenomenon, we propose a Dynamic Transformer to automatically configure a proper number of tokens for each input image. This is achieved by cascading multiple Transformers with increasing numbers of tokens, which are sequentially activated in an adaptive fashion at test time, i.e., the inference is terminated once a sufficiently confident prediction is produced. We further design efficient feature reuse and relationship reuse mechanisms across different components of the Dynamic Transformer to reduce redundant computations. Extensive empirical results on ImageNet, CIFAR-10, and CIFAR-100 demonstrate that our method significantly outperforms the competitive baselines in terms of both theoretical computational efficiency and practical inference speed. Code and pre-trained models (based on PyTorch and MindSpore) are available at https://github.com/blackfeather-wang/Dynamic-Vision-Transformer and https://github.com/blackfeather-wang/Dynamic-Vision-Transformer-MindSpore.

The convergence of stochastic gradient descent is highly dependent on the step-size, especially on non-convex problems such as neural network training. Step decay step-size schedules (constant and then cut) are widely used in practice because of their excellent convergence and generalization qualities, but their theoretical properties are not yet well understood. We provide convergence results for step decay in the non-convex regime, ensuring that the gradient norm vanishes at an $\mathcal{O}(\ln T/\sqrt{T})$ rate. We also provide near-optimal (and sometimes provably tight) convergence guarantees for general, possibly non-smooth, convex and strongly convex problems. The practical efficiency of the step decay step-size is demonstrated in several large-scale deep neural network training tasks.

A Forster transform is an operation that turns a multivariate distribution into one with good anti-concentration properties. While a Forster transform does not always exist, we show that any distribution can be efficiently decomposed as a disjoint mixture of few distributions for which a Forster transform exists and can be computed efficiently. As the main application of this result, we obtain the first polynomial-time algorithm for distribution-independent PAC learning of halfspaces in the Massart noise model with strongly polynomial sample complexity, i.e., independent of the bit complexity of the examples. Previous algorithms for this learning problem incurred sample complexity scaling polynomially with the bit complexity, even though such a dependence is not information-theoretically necessary.

Deep neural networks have achieved impressive performance in many areas. Designing a fast and provable method for training neural networks is a fundamental question in machine learning. The classical training method requires paying $\Omega(mnd)$ cost for both forward computation and backward computation, where $m$ is the width of the neural network, and we are given $n$ training points in $d$-dimensional space. In this paper, we propose two novel preprocessing ideas to bypass this $\Omega(mnd)$ barrier: * First, by preprocessing the initial weights of the neural networks, we can train the neural network in $\widetilde{O}(m^{1-\Theta(1/d)} n d)$ cost per iteration. * Second, by preprocessing the input data points, we can train neural network in $\widetilde{O} (m^{4/5} nd )$ cost per iteration. From the technical perspective, our result is a sophisticated combination of tools in different fields, greedy-type convergence analysis in optimization, sparsity observation in practical work, high-dimensional geometric search in data structure, concentration and anti-concentration in probability. Our results also provide theoretical insights for a large number of previously established fast training methods. In addition, our classical algorithm can be generalized to the Quantum computation model. Interestingly, we can get a similar sublinear cost per iteration but avoid preprocessing initial weights or input data points.