NeurIPS 2021 论文列表

We propose a toolkit for developing and comparing reinforcement learning (RL)-based traffic signal controllers. The toolkit includes implementation of state-of-the-art deep-RL algorithms for signal control along with benchmark control problems that are based on realistic traffic scenarios. Importantly, the toolkit allows a first-of-its-kind comparison between state-of-the-art RL-based signal controllers while providing benchmarks for future comparisons. Consequently, we compare and report the relative performance of current RL algorithms. The experimental results suggest that previous algorithms are not robust to varying sensing assumptions and non-stylized intersection layouts. When more realistic signal layouts and advanced sensing capabilities are assumed, a distributed deep-Q learning approach is shown to outperform previously reported state-of-the-art algorithms in many cases.

Vision transformer (ViT) models exhibit substandard optimizability. In particular, they are sensitive to the choice of optimizer (AdamW vs. SGD), optimizer hyperparameters, and training schedule length. In comparison, modern convolutional neural networks are easier to optimize. Why is this the case? In this work, we conjecture that the issue lies with the patchify stem of ViT models, which is implemented by a stride-p p×p convolution (p = 16 by default) applied to the input image. This large-kernel plus large-stride convolution runs counter to typical design choices of convolutional layers in neural networks. To test whether this atypical design choice causes an issue, we analyze the optimization behavior of ViT models with their original patchify stem versus a simple counterpart where we replace the ViT stem by a small number of stacked stride-two 3×3 convolutions. While the vast majority of computation in the two ViT designs is identical, we find that this small change in early visual processing results in markedly different training behavior in terms of the sensitivity to optimization settings as well as the final model accuracy. Using a convolutional stem in ViT dramatically increases optimization stability and also improves peak performance (by ∼1-2% top-1 accuracy on ImageNet-1k), while maintaining flops and runtime. The improvement can be observed across the wide spectrum of model complexities (from 1G to 36G flops) and dataset scales (from ImageNet-1k to ImageNet-21k). These findings lead us to recommend using a standard, lightweight convolutional stem for ViT models in this regime as a more robust architectural choice compared to the original ViT model design.

Vision transformer networks have shown superiority in many computer vision tasks. In this paper, we take a step further by proposing a novel generative vision transformer with latent variables following an informative energy-based prior for salient object detection. Both the vision transformer network and the energy-based prior model are jointly trained via Markov chain Monte Carlo-based maximum likelihood estimation, in which the sampling from the intractable posterior and prior distributions of the latent variables are performed by Langevin dynamics. Further, with the generative vision transformer, we can easily obtain a pixel-wise uncertainty map from an image, which indicates the model confidence in predicting saliency from the image. Different from the existing generative models which define the prior distribution of the latent variables as a simple isotropic Gaussian distribution, our model uses an energy-based informative prior which can be more expressive to capture the latent space of the data. We apply the proposed framework to both RGB and RGB-D salient object detection tasks. Extensive experimental results show that our framework can achieve not only accurate saliency predictions but also meaningful uncertainty maps that are consistent with the human perception.

Several families of continual learning techniques have been proposed to alleviate catastrophic interference in deep neural network training on non-stationary data. However, a comprehensive comparison and analysis of limitations remains largely open due to the inaccessibility to suitable datasets. Empirical examination not only varies immensely between individual works, it further currently relies on contrived composition of benchmarks through subdivision and concatenation of various prevalent static vision datasets. In this work, our goal is to bridge this gap by introducing a computer graphics simulation framework that repeatedly renders only upcoming urban scene fragments in an endless real-time procedural world generation process. At its core lies a modular parametric generative model with adaptable generative factors. The latter can be used to flexibly compose data streams, which significantly facilitates a detailed analysis and allows for effortless investigation of various continual learning schemes.

Backward propagation of errors (backpropagation) is a method to minimize objective functions (e.g., loss functions) of deep neural networks by identifying optimal sets of weights and biases. Imposing constraints on weight precision is often required to alleviate prohibitive workloads on hardware. Despite the remarkable success of backpropagation, the algorithm itself is not capable of considering such constraints unless additional algorithms are applied simultaneously. To address this issue, we propose the constrained backpropagation (CBP) algorithm based on the pseudo-Lagrange multiplier method to obtain the optimal set of weights that satisfy a given set of constraints. The defining characteristic of the proposed CBP algorithm is the utilization of a Lagrangian function (loss function plus constraint function) as its objective function. We considered various types of constraints — binary, ternary, one-bit shift, and two-bit shift weight constraints. As a post-training method, CBP applied to AlexNet, ResNet-18, ResNet-50, and GoogLeNet on ImageNet, which were pre-trained using the conventional backpropagation. For most cases, the proposed algorithm outperforms the state-of-the-art methods on ImageNet, e.g., 66.6\%, 74.4\%, and 64.0\% top-1 accuracy for ResNet-18, ResNet-50, and GoogLeNet with binary weights, respectively. This highlights CBP as a learning algorithm to address diverse constraints with the minimal performance loss by employing appropriate constraint functions. The code for CBP is publicly available at \url{https://github.com/dooseokjeong/CBP}.

Domain adaptation (DA) benefits from the rigorous theoretical works that study its insightful characteristics and various aspects, e.g., learning domain-invariant representations and its trade-off. However, it seems not the case for the multiple source DA and domain generalization (DG) settings which are remarkably more complicated and sophisticated due to the involvement of multiple source domains and potential unavailability of target domain during training. In this paper, we develop novel upper-bounds for the target general loss which appeal us to define two kinds of domain-invariant representations. We further study the pros and cons as well as the trade-offs of enforcing learning each domain-invariant representation. Finally, we conduct experiments to inspect the trade-off of these representations for offering practical hints regarding how to use them in practice and explore other interesting properties of our developed theory.

The time dimension of datasets and the long-term performance of machine learning models have received little attention. With extended deployments in the wild, models are bound to encounter novel scenarios and concept drift that cannot be accounted for during development and training. In order for long-term patterns and cycles to appear in datasets, the datasets must cover long periods of time. Since this is rarely the case, it is difficult to explore how computer vision algorithms cope with changes in data distribution occurring across long-term cycles such as seasons. Video surveillance is an application area clearly affected by concept drift. For this reason, we publish the Long-term Thermal Drift (LTD) dataset. LTD consists of thermal surveillance imaging from a single location across 8 months. Along with thermal images we provide relevant metadata such as weather, the day/night cycle, and scene activity. In this paper, we use the metadata for in-depth analysis of the causal and correlational relationships between environmental variables and the performance of selected computer vision algorithms used for anomaly and object detection. Long-term performance is shown to be most correlated with temperature, humidity, the day/night cycle, and scene activity level. This suggests that the coverage of these variables should be prioritised when building datasets for similar applications. As a baseline, we propose to mitigate the impact of concept drift by first detecting points in time where drift occurs. At this point, we collect additional data that is used to retraining the models. This improves later performance by an average of 25% across all tested algorithms.

Recent advances in computationally efficient non-myopic Bayesian optimization offer improved query efficiency over traditional myopic methods like expected improvement, with only a modest increase in computational cost. These advances have been largely limited to unconstrained BO methods with only a few exceptions which require heavy computation. For instance, one existing multi-step lookahead constrained BO method (Lam & Willcox, 2017) relies on computationally expensive unreliable brute force derivative-free optimization of a Monte Carlo rollout acquisition function. Methods that use the reparameterization trick for more efficient derivative-based optimization of non-myopic acquisition functions in the unconstrained setting, like sample average approximation and infinitesimal perturbation analysis, do not extend: constraints introduce discontinuities in the sampled acquisition function surface. Moreover, we argue here that being non-myopic is even more important in constrained problems because fear of violating constraints pushes myopic methods away from sampling the boundary between feasible and infeasible regions, slowing the discovery of optimal solutions with tight constraints. In this paper, we propose a computationally efficient two-step lookahead constrained Bayesian optimization acquisition function (2-OPT-C) supporting both sequential and batch settings. To enable fast acquisition function optimization, we develop a novel likelihood ratio-based unbiased estimator of the gradient of the two-step optimal acquisition function that does not use the reparameterization trick. In numerical experiments, 2-OPT-C typically improves query efficiency by 2x or more over previous methods, and in some cases by 10x or more.

The integration and transfer of information from multiple sources to multiple targets is a core motive of neural systems. The emerging field of partial information decomposition (PID) provides a novel information-theoretic lens into these mechanisms by identifying synergistic, redundant, and unique contributions to the mutual information between one and several variables. While many works have studied aspects of PID for Gaussian and discrete distributions, the case of general continuous distributions is still uncharted territory. In this work we present a method for estimating the unique information in continuous distributions, for the case of one versus two variables. Our method solves the associated optimization problem over the space of distributions with fixed bivariate marginals by combining copula decompositions and techniques developed to optimize variational autoencoders. We obtain excellent agreement with known analytic results for Gaussians, and illustrate the power of our new approach in several brain-inspired neural models. Our method is capable of recovering the effective connectivity of a chaotic network of rate neurons, and uncovers a complex trade-off between redundancy, synergy and unique information in recurrent networks trained to solve a generalized XOR~task.

Variational autoencoders (VAEs) are one of the powerful likelihood-based generative models with applications in many domains. However, they struggle to generate high-quality images, especially when samples are obtained from the prior without any tempering. One explanation for VAEs' poor generative quality is the prior hole problem: the prior distribution fails to match the aggregate approximate posterior. Due to this mismatch, there exist areas in the latent space with high density under the prior that do not correspond to any encoded image. Samples from those areas are decoded to corrupted images. To tackle this issue, we propose an energy-based prior defined by the product of a base prior distribution and a reweighting factor, designed to bring the base closer to the aggregate posterior. We train the reweighting factor by noise contrastive estimation, and we generalize it to hierarchical VAEs with many latent variable groups. Our experiments confirm that the proposed noise contrastive priors improve the generative performance of state-of-the-art VAEs by a large margin on the MNIST, CIFAR-10, CelebA 64, and CelebA HQ 256 datasets. Our method is simple and can be applied to a wide variety of VAEs to improve the expressivity of their prior distribution.

Knowledge distillation (KD) has recently been identified as a method that can unintentionally leak private information regarding the details of a teacher model to an unauthorized student. Recent research in developing undistillable nasty teachers that can protect model confidentiality has gained significant attention. However, the level of protection these nasty models offer has been largely untested. In this paper, we show that transferring knowledge to a shallow sub-section of a student can largely reduce a teacher’s influence. By exploring the depth of the shallow subsection, we then present a distillation technique that enables a skeptical student model to learn even from a nasty teacher. To evaluate the efficacy of our skeptical students, we conducted experiments with several models with KD on both training data-available and data-free scenarios for various datasets. While distilling from nasty teachers, compared to the normal student models, skeptical students consistently provide superior classification performance of up to ∼59.5%. Moreover, similar to normal students, skeptical students maintain high classification accuracy when distilled from a normal teacher, showing their efficacy irrespective of the teacher being nasty or not. We believe the ability of skeptical students to largely diminish the KD-immunity of potentially nasty teachers will motivate the research community to create more robust mechanisms for model confidentiality. We have open-sourced the code at https://github.com/ksouvik52/Skeptical2021

Many complex time series can be effectively subdivided into distinct regimes that exhibit persistent dynamics. Discovering the switching behavior and the statistical patterns in these regimes is important for understanding the underlying dynamical system. We propose the Recurrent Explicit Duration Switching Dynamical System (RED-SDS), a flexible model that is capable of identifying both state- and time-dependent switching dynamics. State-dependent switching is enabled by a recurrent state-to-switch connection and an explicit duration count variable is used to improve the time-dependent switching behavior. We demonstrate how to perform efficient inference using a hybrid algorithm that approximates the posterior of the continuous states via an inference network and performs exact inference for the discrete switches and counts. The model is trained by maximizing a Monte Carlo lower bound of the marginal log-likelihood that can be computed efficiently as a byproduct of the inference routine. Empirical results on multiple datasets demonstrate that RED-SDS achieves considerable improvement in time series segmentation and competitive forecasting performance against the state of the art.

We derive information-theoretic generalization bounds for supervised learning algorithms based on the information contained in predictions rather than in the output of the training algorithm. These bounds improve over the existing information-theoretic bounds, are applicable to a wider range of algorithms, and solve two key challenges: (a) they give meaningful results for deterministic algorithms and (b) they are significantly easier to estimate. We show experimentally that the proposed bounds closely follow the generalization gap in practical scenarios for deep learning.

Real world applications such as economics and policy making often involve solving multi-agent games with two unique features: (1) The agents are inherently *asymmetric* and partitioned into leaders and followers; (2) The agents have different reward functions, thus the game is *general-sum*. The majority of existing results in this field focuses on either symmetric solution concepts (e.g. Nash equilibrium) or zero-sum games. It remains open how to learn the *Stackelberg equilibrium*---an asymmetric analog of the Nash equilibrium---in general-sum games efficiently from noisy samples. This paper initiates the theoretical study of sample-efficient learning of the Stackelberg equilibrium, in the bandit feedback setting where we only observe noisy samples of the reward. We consider three representative two-player general-sum games: bandit games, bandit-reinforcement learning (bandit-RL) games, and linear bandit games. In all these games, we identify a fundamental gap between the exact value of the Stackelberg equilibrium and its estimated version using finitely many noisy samples, which can not be closed information-theoretically regardless of the algorithm. We then establish sharp positive results on sample-efficient learning of Stackelberg equilibrium with value optimal up to the gap identified above, with matching lower bounds in the dependency on the gap, error tolerance, and the size of the action spaces. Overall, our results unveil unique challenges in learning Stackelberg equilibria under noisy bandit feedback, which we hope could shed light on future research on this topic.

The solution of a partial differential equation can be obtained by computing the inverse operator map between the input and the solution space. Towards this end, we introduce a $\textit{multiwavelet-based neural operator learning scheme}$ that compresses the associated operator's kernel using fine-grained wavelets. By explicitly embedding the inverse multiwavelet filters, we learn the projection of the kernel onto fixed multiwavelet polynomial bases. The projected kernel is trained at multiple scales derived from using repeated computation of multiwavelet transform. This allows learning the complex dependencies at various scales and results in a resolution-independent scheme. Compare to the prior works, we exploit the fundamental properties of the operator's kernel which enable numerically efficient representation. We perform experiments on the Korteweg-de Vries (KdV) equation, Burgers' equation, Darcy Flow, and Navier-Stokes equation. Compared with the existing neural operator approaches, our model shows significantly higher accuracy and achieves state-of-the-art in a range of datasets. For the time-varying equations, the proposed method exhibits a ($2X-10X$) improvement ($0.0018$ ($0.0033$) relative $L2$ error for Burgers' (KdV) equation). By learning the mappings between function spaces, the proposed method has the ability to find the solution of a high-resolution input after learning from lower-resolution data.

Contrastive self-supervised learning (CSL) has attracted increasing attention for model pre-training via unlabeled data. The resulted CSL models provide instance-discriminative visual features that are uniformly scattered in the feature space. During deployment, the common practice is to directly fine-tune CSL models with cross-entropy, which however may not be the best strategy in practice. Although cross-entropy tends to separate inter-class features, the resulting models still have limited capability for reducing intra-class feature scattering that exists in CSL models. In this paper, we investigate whether applying contrastive learning to fine-tuning would bring further benefits, and analytically find that optimizing the contrastive loss benefits both discriminative representation learning and model optimization during fine-tuning. Inspired by these findings, we propose Contrast-regularized tuning (Core-tuning), a new approach for fine-tuning CSL models. Instead of simply adding the contrastive loss to the objective of fine-tuning, Core-tuning further applies a novel hard pair mining strategy for more effective contrastive fine-tuning, as well as smoothing the decision boundary to better exploit the learned discriminative feature space. Extensive experiments on image classification and semantic segmentation verify the effectiveness of Core-tuning.

We propose ACProp (Asynchronous-centering-Prop), an adaptive optimizer which combines centering of second momentum and asynchronous update (e.g. for $t$-th update, denominator uses information up to step $t-1$, while numerator uses gradient at $t$-th step). ACProp has both strong theoretical properties and empirical performance. With the example by Reddi et al. (2018), we show that asynchronous optimizers (e.g. AdaShift, ACProp) have weaker convergence condition than synchronous optimizers (e.g. Adam, RMSProp, AdaBelief); within asynchronous optimizers, we show that centering of second momentum further weakens the convergence condition. We demonstrate that ACProp has a convergence rate of $O(\frac{1}{\sqrt{T}})$ for the stochastic non-convex case, which matches the oracle rate and outperforms the $O(\frac{logT}{\sqrt{T}})$ rate of RMSProp and Adam. We validate ACProp in extensive empirical studies: ACProp outperforms both SGD and other adaptive optimizers in image classification with CNN, and outperforms well-tuned adaptive optimizers in the training of various GAN models, reinforcement learning and transformers. To sum up, ACProp has good theoretical properties including weak convergence condition and optimal convergence rate, and strong empirical performance including good generalization like SGD and training stability like Adam. We provide the implementation at \url{ https://github.com/juntang-zhuang/ACProp-Optimizer}.

Few-shot segmentation aims to train a segmentation model that can fast adapt to novel classes with few exemplars. The conventional training paradigm is to learn to make predictions on query images conditioned on the features from support images. Previous methods only utilized the semantic-level prototypes of support images as the conditional information. These methods cannot utilize all pixel-wise support information for the query predictions, which is however critical for the segmentation task. In this paper, we focus on utilizing pixel-wise relationships between support and target images to facilitate the few-shot semantic segmentation task. We design a novel Cycle-Consistent Transformer (CyCTR) module to aggregate pixel-wise support features into query ones. CyCTR performs cross-attention between features from different images, i.e. support and query images. We observe that there may exist unexpected irrelevant pixel-level support features. Directly performing cross-attention may aggregate these features from support to query and bias the query features. Thus, we propose using a novel cycle-consistent attention mechanism to filter out possible harmful support features and encourage query features to attend to the most informative pixels from support images. Experiments on all few-shot segmentation benchmarks demonstrate that our proposed CyCTR leads to remarkable improvement compared to previous state-of-the-art methods. Specifically, on Pascal-5^i and COCO-20^i datasets, we achieve 66.6% and 45.6% mIoU for 5-shot segmentation, outperforming previous state-of-the-art by 4.6% and 7.1% respectively.

Most prior approaches to offline reinforcement learning (RL) have taken an iterative actor-critic approach involving off-policy evaluation. In this paper we show that simply doing one step of constrained/regularized policy improvement using an on-policy Q estimate of the behavior policy performs surprisingly well. This one-step algorithm beats the previously reported results of iterative algorithms on a large portion of the D4RL benchmark. The one-step baseline achieves this strong performance while being notably simpler and more robust to hyperparameters than previously proposed iterative algorithms. We argue that the relatively poor performance of iterative approaches is a result of the high variance inherent in doing off-policy evaluation and magnified by the repeated optimization of policies against those estimates. In addition, we hypothesize that the strong performance of the one-step algorithm is due to a combination of favorable structure in the environment and behavior policy.