NeurIPS 2021 论文列表

Understanding generalization in deep learning has been one of the major challenges in statistical learning theory over the last decade. While recent work has illustrated that the dataset and the training algorithm must be taken into account in order to obtain meaningful generalization bounds, it is still theoretically not clear which properties of the data and the algorithm determine the generalization performance. In this study, we approach this problem from a dynamical systems theory perspective and represent stochastic optimization algorithms as \emph{random iterated function systems} (IFS). Well studied in the dynamical systems literature, under mild assumptions, such IFSs can be shown to be ergodic with an invariant measure that is often supported on sets with a \emph{fractal structure}. As our main contribution, we prove that the generalization error of a stochastic optimization algorithm can be bounded based on the `complexity' of the fractal structure that underlies its invariant measure. Then, by leveraging results from dynamical systems theory, we show that the generalization error can be explicitly linked to the choice of the algorithm (e.g., stochastic gradient descent -- SGD), algorithm hyperparameters (e.g., step-size, batch-size), and the geometry of the problem (e.g., Hessian of the loss). We further specialize our results to specific problems (e.g., linear/logistic regression, one hidden-layered neural networks) and algorithms (e.g., SGD and preconditioned variants), and obtain analytical estimates for our bound. For modern neural networks, we develop an efficient algorithm to compute the developed bound and support our theory with various experiments on neural networks.

In this paper, we focus on auditing black-box prediction models for compliance with the GDPR’s data minimization principle. This principle restricts prediction models to use the minimal information that is necessary for performing the task at hand. Given the challenge of the black-box setting, our key idea is to check if each of the prediction model’s input features is individually necessary by assigning it some constant value (i.e., applying a simple imputation) across all prediction instances, and measuring the extent to which the model outcomes would change. We introduce a metric for data minimization that is based on model instability under simple imputations. We extend the applicability of this metric from a finite sample model to a distributional setting by introducing a probabilistic data minimization guarantee, which we derive using a Bayesian approach. Furthermore, we address the auditing problem under a constraint on the number of queries to the prediction system. We formulate the problem of allocating a budget of system queries to feasible simple imputations (for investigating model instability) as a multi-armed bandit framework with probabilistic success metrics. We define two bandit problems for providing a probabilistic data minimization guarantee at a given confidence level: a decision problem given a data minimization level, and a measurement problem given a fixed query budget. We design efficient algorithms for these auditing problems using novel exploration strategies that expand classical bandit strategies. Our experiments with real-world prediction systems show that our auditing algorithms significantly outperform simpler benchmarks in both measurement and decision problems.

A challenging aspect of the bandit problem is that a stochastic reward is observed only for the chosen arm and the rewards of other arms remain missing. The dependence of the arm choice on the past context and reward pairs compounds the complexity of regret analysis. We propose a novel multi-armed contextual bandit algorithm called Doubly Robust Thompson Sampling (DRTS) employing the doubly-robust estimator used in missing data literature to Thompson Sampling with contexts (\texttt{LinTS}). Different from previous works relying on missing data techniques (Dimakopoulou et al. [2019], Kim and Paik [2019]), the proposed algorithm is designed to allow a novel additive regret decomposition leading to an improved regret bound with the order of $\tilde{O}(\phi^{-2}\sqrt{T})$, where $\phi^2$ is the minimum eigenvalue of the covariance matrix of contexts. This is the first regret bound of \texttt{LinTS} using $\phi^2$ without $d$, where $d$ is the dimension of the context. Applying the relationship between $\phi^2$ and $d$, the regret bound of the proposed algorithm is $\tilde{O}(d\sqrt{T})$ in many practical scenarios, improving the bound of \texttt{LinTS} by a factor of $\sqrt{d}$. A benefit of the proposed method is that it uses all the context data, chosen or not chosen, thus allowing to circumvent the technical definition of unsaturated arms used in theoretical analysis of \texttt{LinTS}. Empirical studies show the advantage of the proposed algorithm over \texttt{LinTS}.

Probabilistic Circuits (PCs) are a promising avenue for probabilistic modeling. They combine advantages of probabilistic graphical models (PGMs) with those of neural networks (NNs). Crucially, however, they are tractable probabilistic models, supporting efficient and exact computation of many probabilistic inference queries, such as marginals and MAP. Further, since PCs are structured computation graphs, they can take advantage of deep-learning-style parameter updates, which greatly improves their scalability. However, this innovation also makes PCs prone to overfitting, which has been observed in many standard benchmarks. Despite the existence of abundant regularization techniques for both PGMs and NNs, they are not effective enough when applied to PCs. Instead, we re-think regularization for PCs and propose two intuitive techniques, data softening and entropy regularization, that both take advantage of PCs' tractability and still have an efficient implementation as a computation graph. Specifically, data softening provides a principled way to add uncertainty in datasets in closed form, which implicitly regularizes PC parameters. To learn parameters from a softened dataset, PCs only need linear time by virtue of their tractability. In entropy regularization, the exact entropy of the distribution encoded by a PC can be regularized directly, which is again infeasible for most other density estimation models. We show that both methods consistently improve the generalization performance of a wide variety of PCs. Moreover, when paired with a simple PC structure, we achieved state-of-the-art results on 10 out of 20 standard discrete density estimation benchmarks. Open-source code and experiments are available at https://github.com/UCLA-StarAI/Tractable-PC-Regularization.

It is well known that vision classification models suffer from poor calibration in the face of data distribution shifts. In this paper, we take a geometric approach to this problem. We propose Geometric Sensitivity Decomposition (GSD) which decomposes the norm of a sample feature embedding and the angular similarity to a target classifier into an instance-dependent and an instance-independent com-ponent. The instance-dependent component captures the sensitive information about changes in the input while the instance-independent component represents the insensitive information serving solely to minimize the loss on the training dataset. Inspired by the decomposition, we analytically derive a simple extension to current softmax-linear models, which learns to disentangle the two components during training. On several common vision models, the disentangled model out-performs other calibration methods on standard calibration metrics in the face of out-of-distribution (OOD) data and corruption with significantly less complexity. Specifically, we surpass the current state of the art by 30.8% relative improvement on corrupted CIFAR100 in Expected Calibration Error.

Recently, a new trend of exploring sparsity for accelerating neural network training has emerged, embracing the paradigm of training on the edge. This paper proposes a novel Memory-Economic Sparse Training (MEST) framework targeting for accurate and fast execution on edge devices. The proposed MEST framework consists of enhancements by Elastic Mutation (EM) and Soft Memory Bound (&S) that ensure superior accuracy at high sparsity ratios. Different from the existing works for sparse training, this current work reveals the importance of sparsity schemes on the performance of sparse training in terms of accuracy as well as training speed on real edge devices. On top of that, the paper proposes to employ data efficiency for further acceleration of sparse training. Our results suggest that unforgettable examples can be identified in-situ even during the dynamic exploration of sparsity masks in the sparse training process, and therefore can be removed for further training speedup on edge devices. Comparing with state-of-the-art (SOTA) works on accuracy, our MEST increases Top-1 accuracy significantly on ImageNet when using the same unstructured sparsity scheme. Systematical evaluation on accuracy, training speed, and memory footprint are conducted, where the proposed MEST framework consistently outperforms representative SOTA works. A reviewer strongly against our work based on his false assumptions and misunderstandings. On top of the previous submission, we employ data efficiency for further acceleration of sparse training. And we explore the impact of model sparsity, sparsity schemes, and sparse training algorithms on the number of removable training examples. Our codes are publicly available at: https://github.com/boone891214/MEST.

Self-supervised speech representation learning (speech SSL) has demonstrated the benefit of scale in learning rich representations for Automatic Speech Recognition (ASR) with limited paired data, such as wav2vec 2.0. We investigate the existence of sparse subnetworks in pre-trained speech SSL models that achieve even better low-resource ASR results. However, directly applying widely adopted pruning methods such as the Lottery Ticket Hypothesis (LTH) is suboptimal in the computational cost needed. Moreover, we show that the discovered subnetworks yield minimal performance gain compared to the original dense network. We present Prune-Adjust-Re-Prune (PARP), which discovers and finetunes subnetworks for much better performance, while only requiring a single downstream ASR finetuning run. PARP is inspired by our surprising observation that subnetworks pruned for pre-training tasks need merely a slight adjustment to achieve a sizeable performance boost in downstream ASR tasks. Extensive experiments on low-resource ASR verify (1) sparse subnetworks exist in mono-lingual/multi-lingual pre-trained speech SSL, and (2) the computational advantage and performance gain of PARP over baseline pruning methods. In particular, on the 10min Librispeech split without LM decoding, PARP discovers subnetworks from wav2vec 2.0 with an absolute 10.9%/12.6% WER decrease compared to the full model. We further demonstrate the effectiveness of PARP via: cross-lingual pruning without any phone recognition degradation, the discovery of a multi-lingual subnetwork for 10 spoken languages in 1 finetuning run, and its applicability to pre-trained BERT/XLNet for natural language tasks1.

It was estimated that the world produced $59 ZB$ ($5.9 \times 10^{13} GB$) of data in 2020, resulting in the enormous costs of both data storage and transmission. Fortunately, recent advances in deep generative models have spearheaded a new class of so-called "neural compression" algorithms, which significantly outperform traditional codecs in terms of compression ratio. Unfortunately, the application of neural compression garners little commercial interest due to its limited bandwidth; therefore, developing highly efficient frameworks is of critical practical importance. In this paper, we discuss lossless compression using normalizing flows which have demonstrated a great capacity for achieving high compression ratios. As such, we introduce iFlow, a new method for achieving efficient lossless compression. We first propose Modular Scale Transform (MST) and a novel family of numerically invertible flow transformations based on MST. Then we introduce the Uniform Base Conversion System (UBCS), a fast uniform-distribution codec incorporated into iFlow, enabling efficient compression. iFlow achieves state-of-the-art compression ratios and is $5 \times$ quicker than other high-performance schemes. Furthermore, the techniques presented in this paper can be used to accelerate coding time for a broad class of flow-based algorithms.

What advantage do sequential procedures provide over batch algorithms for testing properties of unknown distributions? Focusing on the problem of testing whether two distributions $\mathcal{D}_1$ and $\mathcal{D}_2$ on $\{1,\dots, n\}$ are equal or $\epsilon$-far, we give several answers to this question. We show that for a small alphabet size $n$, there is a sequential algorithm that outperforms any batch algorithm by a factor of at least $4$ in terms sample complexity. For a general alphabet size $n$, we give a sequential algorithm that uses no more samples than its batch counterpart, and possibly fewer if the actual distance between $\mathcal{D}_1$ and $\mathcal{D}_2$ is larger than $\epsilon$. As a corollary, letting $\epsilon$ go to $0$, we obtain a sequential algorithm for testing closeness (with no a priori bound on the distance between $\mathcal{D}_1$ and $\mathcal{D}_2$) with a sample complexity $\tilde{\mathcal{O}}(\frac{n^{2/3}}{TV(\mathcal{D}_1, \mathcal{D}_2)^{4/3}})$: this improves over the $\tilde{\mathcal{O}}(\frac{n/\log n}{TV(\mathcal{D}_1, \mathcal{D}_2)^{2} })$ tester of [Daskalakis and Kawase 2017] and is optimal up to multiplicative constants. We also establish limitations of sequential algorithms for the problem of testing closeness: they can improve the worst case number of samples by at most a constant factor.

Most conventional Neural Architecture Search (NAS) approaches are limited in that they only generate architectures without searching for the optimal parameters. While some NAS methods handle this issue by utilizing a supernet trained on a large-scale dataset such as ImageNet, they may be suboptimal if the target tasks are highly dissimilar from the dataset the supernet is trained on. To address such limitations, we introduce a novel problem of Neural Network Search (NNS), whose goal is to search for the optimal pretrained network for a novel dataset and constraints (e.g. number of parameters), from a model zoo. Then, we propose a novel framework to tackle the problem, namely Task-Adaptive Neural Network Search (TANS). Given a model-zoo that consists of network pretrained on diverse datasets, we use a novel amortized meta-learning framework to learn a cross-modal latent space with contrastive loss, to maximize the similarity between a dataset and a high-performing network on it, and minimize the similarity between irrelevant dataset-network pairs. We validate the effectiveness and efficiency of our method on ten real-world datasets, against existing NAS/AutoML baselines. The results show that our method instantly retrieves networks that outperform models obtained with the baselines with significantly fewer training steps to reach the target performance, thus minimizing the total cost of obtaining a task-optimal network. Our code and the model-zoo are available at https://anonymous.4open.science/r/TANS-33D6

We propose new width-based planning and learning algorithms inspired from a careful analysis of the design decisions made by previous width-based planners. The algorithms are applied over the Atari-2600 games and our best performing algorithm, Novelty guided Critical Path Learning (N-CPL), outperforms the previously introduced width-based planning and learning algorithms $\pi$-IW(1), $\pi$-IW(1)+ and $\pi$-HIW(n, 1). Furthermore, we present a taxonomy of the Atari-2600 games according to some of their defining characteristics. This analysis of the games provides further insight into the behaviour and performance of the algorithms introduced. Namely, for games with large branching factors, and games with sparse meaningful rewards, N-CPL outperforms $\pi$-IW, $\pi$-IW(1)+ and $\pi$-HIW(n, 1).

To understand the relationship between behavior and neural activity, experiments in neuroscience often include an animal performing a repeated behavior such as a motor task. Recent progress in computer vision and deep learning has shown great potential in the automated analysis of behavior by leveraging large and high-quality video datasets. In this paper, we design Disentangled Behavior Embedding (DBE) to learn robust behavioral embeddings from unlabeled, multi-view, high-resolution behavioral videos across different animals and multiple sessions. We further combine DBE with a stochastic temporal model to propose Variational Disentangled Behavior Embedding (VDBE), an end-to-end approach that learns meaningful discrete behavior representations and generates interpretable behavioral videos. Our models learn consistent behavior representations by explicitly disentangling the dynamic behavioral factors (pose) from time-invariant, non-behavioral nuisance factors (context) in a deep autoencoder, and exploit the temporal structures of pose dynamics. Compared to competing approaches, DBE and VDBE enjoy superior performance on downstream tasks such as fine-grained behavioral motif generation and behavior decoding.

Large-scale, two-sided matching platforms must find market outcomes that align with user preferences while simultaneously learning these preferences from data. But since preferences are inherently uncertain during learning, the classical notion of stability (Gale and Shapley, 1962; Shapley and Shubik, 1971) is unattainable in these settings. To bridge this gap, we develop a framework and algorithms for learning stable market outcomes under uncertainty. Our primary setting is matching with transferable utilities, where the platform both matches agents and sets monetary transfers between them. We design an incentive-aware learning objective that captures the distance of a market outcome from equilibrium. Using this objective, we analyze the complexity of learning as a function of preference structure, casting learning as a stochastic multi-armed bandit problem. Algorithmically, we show that "optimism in the face of uncertainty," the principle underlying many bandit algorithms, applies to a primal-dual formulation of matching with transfers and leads to near-optimal regret bounds. Our work takes a first step toward elucidating when and how stable matchings arise in large, data-driven marketplaces.

We study the power of learning via mini-batch stochastic gradient descent (SGD) on the loss of a differentiable model or neural network, and ask what learning problems can be learnt using this paradigm. We show that SGD can always simulate learning with statistical queries (SQ), but its ability to go beyond that depends on the precision $\rho$ of the gradients and the minibatch size $b$. With fine enough precision relative to minibatch size, namely when $b \rho$ is small enough, SGD can go beyond SQ learning and simulate any sample-based learning algorithm and thus its learning power is equivalent to that of PAC learning; this extends prior work that achieved this result for $b=1$. Moreover, with polynomially many bits of precision (i.e. when $\rho$ is exponentially small), SGD can simulate PAC learning regardless of the batch size. On the other hand, when $b \rho^2$ is large enough, the power of SGD is equivalent to that of SQ learning.

We tackle sequential learning under label noise in applications where a human supervisor can be queried to relabel suspicious examples. Existing approaches are flawed, in that they only relabel incoming examples that look "suspicious" to the model. As a consequence, those mislabeled examples that elude (or don't undergo) this cleaning step end up tainting the training data and the model with no further chance of being cleaned. We propose CINCER, a novel approach that cleans both new and past data by identifying \emph{pairs of mutually incompatible examples}. Whenever it detects a suspicious example, CINCER identifies a counter-example in the training set that - according to the model - is maximally incompatible with the suspicious example, and asks the annotator to relabel either or both examples, resolving this possible inconsistency. The counter-examples are chosen to be maximally incompatible, so to serve as \emph{explanations} of the model's suspicion, and highly influential, so to convey as much information as possible if relabeled. CINCER achieves this by leveraging an efficient and robust approximation of influence functions based on the Fisher information matrix (FIM). Our extensive empirical evaluation shows that clarifying the reasons behind the model's suspicions by cleaning the counter-examples helps in acquiring substantially better data and models, especially when paired with our FIM approximation.

In imperfect-information games, subgame solving is significantly more challenging than in perfect-information games, but in the last few years, such techniques have been developed. They were the key ingredient to the milestone of superhuman play in no-limit Texas hold'em poker. Current subgame-solving techniques analyze the entire common-knowledge closure of the player's current information set, that is, the smallest set of nodes within which it is common knowledge that the current node lies. While this is acceptable in games like poker where the common-knowledge closure is relatively small, many practical games have more complex information structure, which renders the common-knowledge closure impractically large to enumerate or even reasonably approximate. We introduce an approach that overcomes this obstacle, by instead working with only low-order knowledge. Our approach allows an agent, upon arriving at an infoset, to basically prune any node that is no longer reachable, thereby massively reducing the game tree size relative to the common-knowledge subgame. We prove that, as is, our approach can increase exploitability compared to the blueprint strategy. However, we develop three avenues by which safety can be guaranteed. First, safety is guaranteed if the results of subgame solves are incorporated back into the blueprint. Second, we provide a method where safety is achieved by limiting the infosets at which subgame solving is performed. Third, we prove that our approach, when applied at every infoset reached during play, achieves a weaker notion of equilibrium, which we coin affine equilibrium, and which may be of independent interest. We show that affine equilibria cannot be exploited by any Nash strategy of the opponent, so an opponent who wishes to exploit must open herself to counter-exploitation. Even without the safety-guaranteeing additions, experiments on medium-sized games show that our approach always reduced exploitability in practical games even when applied at every infoset, and a depth-limited version of it led to---to our knowledge---the first strong AI for the challenge problem dark chess.

Numerical solutions to high-dimensional partial differential equations (PDEs) based on neural networks have seen exciting developments. This paper derives complexity estimates of the solutions of $d$-dimensional second-order elliptic PDEs in the Barron space, that is a set of functions admitting the integral of certain parametric ridge function against a probability measure on the parameters. We prove under some appropriate assumptions that if the coefficients and the source term of the elliptic PDE lie in Barron spaces, then the solution of the PDE is $\epsilon$-close with respect to the $H^1$ norm to a Barron function. Moreover, we prove dimension-explicit bounds for the Barron norm of this approximate solution, depending at most polynomially on the dimension $d$ of the PDE. As a direct consequence of the complexity estimates, the solution of the PDE can be approximated on any bounded domain by a two-layer neural network with respect to the $H^1$ norm with a dimension-explicit convergence rate.

Graphs have been widely used in data mining and machine learning due to their unique representation of real-world objects and their interactions. As graphs are getting bigger and bigger nowadays, it is common to see their subgraphs separately collected and stored in multiple local systems. Therefore, it is natural to consider the subgraph federated learning setting, where each local system holds a small subgraph that may be biased from the distribution of the whole graph. Hence, the subgraph federated learning aims to collaboratively train a powerful and generalizable graph mining model without directly sharing their graph data. In this work, towards the novel yet realistic setting of subgraph federated learning, we propose two major techniques: (1) FedSage, which trains a GraphSage model based on FedAvg to integrate node features, link structures, and task labels on multiple local subgraphs; (2) FedSage+, which trains a missing neighbor generator along FedSage to deal with missing links across local subgraphs. Empirical results on four real-world graph datasets with synthesized subgraph federated learning settings demonstrate the effectiveness and efficiency of our proposed techniques. At the same time, consistent theoretical implications are made towards their generalization ability on the global graphs.

Many of the challenges facing today's reinforcement learning (RL) algorithms, such as robustness, generalization, transfer, and computational efficiency are closely related to compression. Prior work has convincingly argued why minimizing information is useful in the supervised learning setting, but standard RL algorithms lack an explicit mechanism for compression. The RL setting is unique because (1) its sequential nature allows an agent to use past information to avoid looking at future observations and (2) the agent can optimize its behavior to prefer states where decision making requires few bits. We take advantage of these properties to propose a method (RPC) for learning simple policies. This method brings together ideas from information bottlenecks, model-based RL, and bits-back coding into a simple and theoretically-justified algorithm. Our method jointly optimizes a latent-space model and policy to be self-consistent, such that the policy avoids states where the model is inaccurate. We demonstrate that our method achieves much tighter compression than prior methods, achieving up to 5$\times$ higher reward than a standard information bottleneck when constrained to use just 0.3 bits per observation. We also demonstrate that our method learns policies that are more robust and generalize better to new tasks.