NeurIPS 2021 Papers

The recent explosive interest on transformers has suggested their potential to become powerful ``universal" models for computer vision tasks, such as classification, detection, and segmentation. While those attempts mainly study the discriminative models, we explore transformers on some more notoriously difficult vision tasks, e.g., generative adversarial networks (GANs). Our goal is to conduct the first pilot study in building a GAN \textit{completely free of convolutions}, using only pure transformer-based architectures. Our vanilla GAN architecture, dubbed \textbf{TransGAN}, consists of a memory-friendly transformer-based generator that progressively increases feature resolution, and correspondingly a multi-scale discriminator to capture simultaneously semantic contexts and low-level textures. On top of them, we introduce the new module of grid self-attention for alleviating the memory bottleneck further, in order to scale up TransGAN to high-resolution generation. We also develop a unique training recipe including a series of techniques that can mitigate the training instability issues of TransGAN, such as data augmentation, modified normalization, and relative position encoding. Our best architecture achieves highly competitive performance compared to current state-of-the-art GANs using convolutional backbones. Specifically, TransGAN sets \textbf{new state-of-the-art} inception score of 10.43 and FID of 18.28 on STL-10. It also reaches the inception score of 9.02 and FID of 9.26 on CIFAR-10, and 5.28 FID on CelebA $\mathbf{128} \times \mathbf{128}$, respectively: both on par with the current best results and outperforming StyleGAN-V2. When it comes to higher-resolution (e.g. $\mathbf{256} \times \mathbf{256}$) generation tasks, such as on CelebA-HQ and LSUN-Church, TransGAN continues to produce diverse visual examples with high fidelity and impressive texture details. In addition, we dive deep into the transformer-based generation models to understand how their behaviors differ from convolutional ones, by visualizing training dynamics. The code is available at: https://github.com/VITA-Group/TransGAN.

Existing semi-supervised learning (SSL) algorithms typically assume class-balanced datasets, although the class distributions of many real world datasets are imbalanced. In general, classifiers trained on a class-imbalanced dataset are biased toward the majority classes. This issue becomes more problematic for SSL algorithms because they utilize the biased prediction of unlabeled data for training. However, traditional class-imbalanced learning techniques, which are designed for labeled data, cannot be readily combined with SSL algorithms. We propose a scalable class-imbalanced SSL algorithm that can effectively use unlabeled data, while mitigating class imbalance by introducing an auxiliary balanced classifier (ABC) of a single layer, which is attached to a representation layer of an existing SSL algorithm. The ABC is trained with a class-balanced loss of a minibatch, while using high-quality representations learned from all data points in the minibatch using the backbone SSL algorithm to avoid overfitting and information loss. Moreover, we use consistency regularization, a recent SSL technique for utilizing unlabeled data in a modified way, to train the ABC to be balanced among the classes by selecting unlabeled data with the same probability for each class. The proposed algorithm achieves state-of-the-art performance in various class-imbalanced SSL experiments using four benchmark datasets.

As deep networks begin to be deployed as autonomous agents, the issue of how they can communicate with each other becomes important. Here, we train two deep nets from scratch to perform realistic referent identification through unsupervised emergent communication. We show that the largely interpretable emergent protocol allows the nets to successfully communicate even about object types they did not see at training time. The visual representations induced as a by-product of our training regime, moreover, show comparable quality, when re-used as generic visual features, to a recent self-supervised learning model. Our results provide concrete evidence of the viability of (interpretable) emergent deep net communication in a more realistic scenario than previously considered, as well as establishing an intriguing link between this field and self-supervised visual learning.

In comparison to the interpretation of classification models, the explanation of sequence generation models is also an important problem, however it has seen little attention. In this work, we study model-agnostic explanations of a representative text generation task -- dialogue response generation. Dialog response generation is challenging with its open-ended sentences and multiple acceptable responses. To gain insights into the reasoning process of a generation model, we propose a new method, local explanation of response generation (LERG) that regards the explanations as the mutual interaction of segments in input and output sentences. LERG views the sequence prediction as uncertainty estimation of a human response and then creates explanations by perturbing the input and calculating the certainty change over the human response. We show that LERG adheres to desired properties of explanations for text generation including unbiased approximation, consistency and cause identification. Empirically, our results show that our method consistently improves other widely used methods on proposed automatic- and human- evaluation metrics for this new task by $4.4$-$12.8$\%. Our analysis demonstrates that LERG can extract both explicit and implicit relations between input and output segments.

Set prediction tasks require the matching between predicted set and ground truth set in order to propagate the gradient signal. Recent works have performed this matching in the original feature space thus requiring predefined distance functions. We propose a method for learning the distance function by performing the matching in the latent space learned from encoding networks. This method enables the use of teacher forcing which was not possible previously since matching in the feature space must be computed after the entire output sequence is generated. Nonetheless, a naive implementation of latent set prediction might not converge due to permutation instability. To address this problem, we provide sufficient conditions for permutation stability which begets an algorithm to improve the overall model convergence. Experiments on several set prediction tasks, including image captioning and object detection, demonstrate the effectiveness of our method.

Tensor network methods have been a key ingredient of advances in condensed matter physics and have recently sparked interest in the machine learning community for their ability to compactly represent very high-dimensional objects. Tensor network methods can for example be used to efficiently learn linear models in exponentially large feature spaces [Stoudenmire and Schwab, 2016]. In this work, we derive upper and lower bounds on the VC dimension and pseudo-dimension of a large class of tensor network models for classification, regression and completion. Our upper bounds hold for linear models parameterized by arbitrary tensor network structures, and we derive lower bounds for common tensor decomposition models~(CP, Tensor Train, Tensor Ring and Tucker) showing the tightness of our general upper bound. These results are used to derive a generalization bound which can be applied to classification with low rank matrices as well as linear classifiers based on any of the commonly used tensor decomposition models. As a corollary of our results, we obtain a bound on the VC dimension of the matrix product state classifier introduced in [Stoudenmire and Schwab, 2016] as a function of the so-called bond dimension~(i.e. tensor train rank), which answers an open problem listed by Cirac, Garre-Rubio and Pérez-García in [Cirac et al., 2019].

Mitigating the dependence on spurious correlations present in the training dataset is a quickly emerging and important topic of deep learning. Recent approaches include priors on the feature attribution of a deep neural network (DNN) into the training process to reduce the dependence on unwanted features. However, until now one needed to trade off high-quality attributions, satisfying desirable axioms, against the time required to compute them. This in turn either led to long training times or ineffective attribution priors. In this work, we break this trade-off by considering a special class of efficiently axiomatically attributable DNNs for which an axiomatic feature attribution can be computed with only a single forward/backward pass. We formally prove that nonnegatively homogeneous DNNs, here termed $\mathcal{X}$-DNNs, are efficiently axiomatically attributable and show that they can be effortlessly constructed from a wide range of regular DNNs by simply removing the bias term of each layer. Various experiments demonstrate the advantages of $\mathcal{X}$-DNNs, beating state-of-the-art generic attribution methods on regular DNNs for training with attribution priors.

When faced with distribution shift at test time, deep neural networks often make inaccurate predictions with unreliable uncertainty estimates. While improving the robustness of neural networks is one promising approach to mitigate this issue, an appealing alternate to robustifying networks against all possible test-time shifts is to instead directly adapt them to unlabeled inputs from the particular distribution shift we encounter at test time. However, this poses a challenging question: in the standard Bayesian model for supervised learning, unlabeled inputs are conditionally independent of model parameters when the labels are unobserved, so what can unlabeled data tell us about the model parameters at test-time? In this paper, we derive a Bayesian model that provides for a well-defined relationship between unlabeled inputs under distributional shift and model parameters, and show how approximate inference in this model can be instantiated with a simple regularized entropy minimization procedure at test-time. We evaluate our method on a variety of distribution shifts for image classification, including image corruptions, natural distribution shifts, and domain adaptation settings, and show that our method improves both accuracy and uncertainty estimation.

Bilevel optimization has been widely applied in many important machine learning applications such as hyperparameter optimization and meta-learning. Recently, several momentum-based algorithms have been proposed to solve bilevel optimization problems faster. However, those momentum-based algorithms do not achieve provably better computational complexity than $\mathcal{\widetilde O}(\epsilon^{-2})$ of the SGD-based algorithm. In this paper, we propose two new algorithms for bilevel optimization, where the first algorithm adopts momentum-based recursive iterations, and the second algorithm adopts recursive gradient estimations in nested loops to decrease the variance. We show that both algorithms achieve the complexity of $\mathcal{\widetilde O}(\epsilon^{-1.5})$, which outperforms all existing algorithms by the order of magnitude. Our experiments validate our theoretical results and demonstrate the superior empirical performance of our algorithms in hyperparameter applications.

The partially observable Markov decision process (POMDP) provides a general framework for modeling an agent's decision process with state uncertainty, and online planning plays a pivotal role in solving it. A belief is a distribution of states representing state uncertainty. Methods for large-scale POMDP problems rely on the same idea of sampling both states and observations. That is, instead of exact belief updating, a collection of sampled states is used to approximate the belief; instead of considering all possible observations, only a set of sampled observations are considered. Inspired by this, we take one step further and propose an online planning algorithm, Adaptive Online Packing-guided Search (AdaOPS), to better approximate beliefs with adaptive particle filter technique and balance estimation bias and variance by fusing similar observation branches. Theoretically, our algorithm is guaranteed to find an $\epsilon$-optimal policy with a high probability given enough planning time under some mild assumptions. We evaluate our algorithm on several tricky POMDP domains, and it outperforms the state-of-the-art in all of them.

Unsupervised domain adaptation aims to align a labeled source domain and an unlabeled target domain, but it requires to access the source data which often raises concerns in data privacy, data portability and data transmission efficiency. We study unsupervised model adaptation (UMA), or called Unsupervised Domain Adaptation without Source Data, an alternative setting that aims to adapt source-trained models towards target distributions without accessing source data. To this end, we design an innovative historical contrastive learning (HCL) technique that exploits historical source hypothesis to make up for the absence of source data in UMA. HCL addresses the UMA challenge from two perspectives. First, it introduces historical contrastive instance discrimination (HCID) that learns from target samples by contrasting their embeddings which are generated by the currently adapted model and the historical models. With the historical models, HCID encourages UMA to learn instance-discriminative target representations while preserving the source hypothesis. Second, it introduces historical contrastive category discrimination (HCCD) that pseudo-labels target samples to learn category-discriminative target representations. Specifically, HCCD re-weights pseudo labels according to their prediction consistency across the current and historical models. Extensive experiments show that HCL outperforms and state-of-the-art methods consistently across a variety of visual tasks and setups.

Simulations of complex physical systems are typically realized by discretizing partial differential equations (PDEs) on unstructured meshes. While neural networks have recently been explored for the surrogate and reduced order modeling of PDE solutions, they often ignore interactions or hierarchical relations between input features, and process them as concatenated mixtures. We generalize the idea of conditional parameterization -- using trainable functions of input parameters to generate the weights of a neural network, and extend them in a flexible way to encode critical information. Inspired by discretized numerical methods, choices of the parameters include physical quantities and mesh topology features. The functional relation between the modeled features and the parameters is built into the network architecture. The method is implemented on different networks and applied to frontier scientific machine learning tasks including the discovery of unmodeled physics, super-resolution of coarse fields, and the simulation of unsteady flows with chemical reactions. The results show that the conditionally-parameterized networks provide superior performance compared to their traditional counterparts. The CP-GNet - an architecture that can be trained on very few data snapshots - is proposed as the first deep learning model capable of standalone prediction of reacting flows on irregular meshes.

Sequence-to-sequence learning with neural networks has become the de facto standard for sequence modeling. This approach typically models the local distribution over the next element with a powerful neural network that can condition on arbitrary context. While flexible and performant, these models often require large datasets for training and can fail spectacularly on benchmarks designed to test for compositional generalization. This work explores an alternative, hierarchical approach to sequence-to-sequence learning with synchronous grammars, where each node in the target tree is transduced by a subset of nodes in the source tree. The source and target trees are treated as fully latent and marginalized out during training. We develop a neural parameterization of the grammar which enables parameter sharing over combinatorial structures without the need for manual feature engineering. We apply this latent neural grammar to various domains---a diagnostic language navigation task designed to test for compositional generalization (SCAN), style transfer, and small-scale machine translation---and find that it performs respectably compared to standard baselines.

We address the estimation of conditional average treatment effects (CATEs) for structured treatments (e.g., graphs, images, texts). Given a weak condition on the effect, we propose the generalized Robinson decomposition, which (i) isolates the causal estimand (reducing regularization bias), (ii) allows one to plug in arbitrary models for learning, and (iii) possesses a quasi-oracle convergence guarantee under mild assumptions. In experiments with small-world and molecular graphs we demonstrate that our approach outperforms prior work in CATE estimation.

The task of assigning semantic classes and track identities to every pixel in a video is called video panoptic segmentation. Our work is the first that targets this task in a real-world setting requiring dense interpretation in both spatial and temporal domains. As the ground-truth for this task is difficult and expensive to obtain, existing datasets are either constructed synthetically or only sparsely annotated within short video clips. To overcome this, we introduce a new benchmark encompassing two datasets, KITTI-STEP, and MOTChallenge-STEP. The datasets contain long video sequences, providing challenging examples and a test-bed for studying long-term pixel-precise segmentation and tracking under real-world conditions. We further propose a novel evaluation metric Segmentation and Tracking Quality (STQ) that fairly balances semantic and tracking aspects of this task and is more appropriate for evaluating sequences of arbitrary length. Finally, we provide several baselines to evaluate the status of existing methods on this new challenging dataset. We have made our datasets, metric, benchmark servers, and baselines publicly available, and hope this will inspire future research.

Decision-time planning is the process of constructing a transient, local policy with the intent of using it to make the immediate decision. Monte Carlo tree search (MCTS), which has been leveraged to great success in Go, chess, shogi, Hex, Atari, and other settings, is perhaps the most celebrated decision-time planning algorithm. Unfortunately, in its original form, MCTS can degenerate to one-step search in domains with stochasticity. Progressive widening is one way to ameliorate this issue, but we argue that it possesses undesirable properties for some settings. In this work, we present a method, called abstraction refining, for extending MCTS to stochastic environments which, unlike progressive widening, leverages the geometry of the state space. We argue that leveraging the geometry of the space can offer advantages. To support this claim, we present a series of experimental examples in which abstraction refining outperforms progressive widening, given equal simulation budgets.

Integrating physics models within machine learning models holds considerable promise toward learning robust models with improved interpretability and abilities to extrapolate. In this work, we focus on the integration of incomplete physics models into deep generative models. In particular, we introduce an architecture of variational autoencoders (VAEs) in which a part of the latent space is grounded by physics. A key technical challenge is to strike a balance between the incomplete physics and trainable components such as neural networks for ensuring that the physics part is used in a meaningful manner. To this end, we propose a regularized learning method that controls the effect of the trainable components and preserves the semantics of the physics-based latent variables as intended. We not only demonstrate generative performance improvements over a set of synthetic and real-world datasets, but we also show that we learn robust models that can consistently extrapolate beyond the training distribution in a meaningful manner. Moreover, we show that we can control the generative process in an interpretable manner.

Although recent advances in voice conversion have shown significant improvement, there still remains a gap between the converted voice and target voice. A key factor that maintains this gap is the insufficient decomposition of content and voice style from the source speech. This insufficiency leads to the converted speech containing source speech style or losing source speech content. In this paper, we present VoiceMixer which can effectively decompose and transfer voice style through a novel information bottleneck and adversarial feedback. With self-supervised representation learning, the proposed information bottleneck can decompose the content and style with only a small loss of content information. Also, for adversarial feedback of each information, the discriminator is decomposed into content and style discriminator with self-supervision, which enable our model to achieve better generalization to the voice style of the converted speech. The experimental results show the superiority of our model in disentanglement and transfer performance, and improve audio quality by preserving content information.

Modern approaches typically formulate semantic segmentation as a per-pixel classification task, while instance-level segmentation is handled with an alternative mask classification. Our key insight: mask classification is sufficiently general to solve both semantic- and instance-level segmentation tasks in a unified manner using the exact same model, loss, and training procedure. Following this observation, we propose MaskFormer, a simple mask classification model which predicts a set of binary masks, each associated with a single global class label prediction. Overall, the proposed mask classification-based method simplifies the landscape of effective approaches to semantic and panoptic segmentation tasks and shows excellent empirical results. In particular, we observe that MaskFormer outperforms per-pixel classification baselines when the number of classes is large. Our mask classification-based method outperforms both current state-of-the-art semantic (55.6 mIoU on ADE20K) and panoptic segmentation (52.7 PQ on COCO) models.