ICLR 2022 论文列表

Recent progress in language model pre-training has achieved a great success via leveraging large-scale unstructured textual data. However, it is still a challenge to apply pre-training on structured tabular data due to the absence of large-scale high-quality tabular data. In this paper, we propose TAPEX to show that table pre-training can be achieved by learning a neural SQL executor over a synthetic corpus, which is obtained by automatically synthesizing executable SQL queries and their execution outputs. TAPEX addresses the data scarcity challenge via guiding the language model to mimic a SQL executor on the diverse, large-scale and high-quality synthetic corpus. We evaluate TAPEX on four benchmark datasets. Experimental results demonstrate that TAPEX outperforms previous table pre-training approaches by a large margin and achieves new state-of-the-art results on all of them. This includes the improvements on the weakly-supervised WikiSQL denotation accuracy to 89.5% (+2.3%), the WikiTableQuestions denotation accuracy to 57.5% (+4.8%), the SQA denotation accuracy to 74.5% (+3.5%), and the TabFact accuracy to 84.2% (+3.2%). To our knowledge, this is the first work to exploit table pre-training via synthetic executable programs and to achieve new state-of-the-art results on various downstream tasks. Our code can be found at https://github.com/microsoft/Table-Pretraining.

Vision Transformer (ViT) has recently demonstrated promise in computer vision problems. However, unlike Convolutional Neural Networks (CNN), it is known that the performance of ViT saturates quickly with depth increasing, due to the observed attention collapse or patch uniformity. Despite a couple of empirical solutions, a rigorous framework studying on this scalability issue remains elusive. In this paper, we first establish a rigorous theory framework to analyze ViT features from the Fourier spectrum domain. We show that the self-attention mechanism inherently amounts to a low-pass filter, which indicates when ViT scales up its depth, excessive low-pass filtering will cause feature maps to only preserve their Direct-Current (DC) component. We then propose two straightforward yet effective techniques to mitigate the undesirable low-pass limitation. The first technique, termed AttnScale, decomposes a self-attention block into low-pass and high-pass components, then rescales and combines these two filters to produce an all-pass self-attention matrix. The second technique, termed FeatScale, re-weights feature maps on separate frequency bands to amplify the high-frequency signals. Both techniques are efficient and hyperparameter-free, while effectively overcoming relevant ViT training artifacts such as attention collapse and patch uniformity. By seamlessly plugging in our techniques to multiple ViT variants, we demonstrate that they consistently help ViTs benefit from deeper architectures, bringing up to 1.1% performance gains "for free" (e.g., with little parameter overhead). We publicly release our codes and pre-trained models at https://github.com/VITA-Group/ViT-Anti-Oversmoothing.

We present a novel framework to train a large deep neural network (DNN) for only $\textit{once}$, which can then be pruned to $\textit{any sparsity ratio}$ to preserve competitive accuracy $\textit{without any re-training}$. Conventional methods often require (iterative) pruning followed by re-training, which not only incurs large overhead beyond the original DNN training but also can be sensitive to retraining hyperparameters. Our core idea is to re-cast the DNN training as an explicit $\textit{pruning-aware}$ process: that is formulated with an auxiliary $K$-sparse polytope constraint, to encourage network weights to lie in a convex hull spanned by $K$-sparse vectors, potentially resulting in more sparse weight matrices. We then leverage a stochastic Frank-Wolfe (SFW) algorithm to solve this new constrained optimization, which naturally leads to sparse weight updates each time. We further note an overlooked fact that existing DNN initializations were derived to enhance SGD training (e.g., avoid gradient explosion or collapse), but was unaligned with the challenges of training with SFW. We hence also present the first learning-based initialization scheme specifically for boosting SFW-based DNN training. Experiments on CIFAR-10 and Tiny-ImageNet datasets demonstrate that our new framework named $\textbf{SFW-pruning}$ consistently achieves the state-of-the-art performance on various benchmark DNNs over a wide range of pruning ratios. Moreover, SFW-pruning only needs to train once on the same model and dataset, for obtaining arbitrary ratios, while requiring neither iterative pruning nor retraining. All codes will be released to the public.

A fundamental limitation of applying semi-supervised learning in real-world settings is the assumption that unlabeled test data contains only classes previously encountered in the labeled training data. However, this assumption rarely holds for data in-the-wild, where instances belonging to novel classes may appear at testing time. Here, we introduce a novel open-world semi-supervised learning setting that formalizes the notion that novel classes may appear in the unlabeled test data. In this novel setting, the goal is to solve the class distribution mismatch problem between labeled and unlabeled data, where at the test time every input instance either needs to be classified into one of the existing classes or a new unseen class needs to be initialized and the instance assigned to it. To tackle this challenging problem, we propose ORCA, an end-to-end approach that assigns instances to previously seen classes or forms novel classes by grouping similar instances without assuming any prior knowledge. The key idea in ORCA is to utilize uncertainty adaptive margin to circumvent the bias towards seen classes caused by learning seen classes faster than the novel classes. In this way, ORCA gradually increases the discriminability of the model during the training and reduces the gap between intra-class variance of seen with respect to novel classes. Extensive experiments on image classification datasets and a single-cell dataset demonstrate that ORCA consistently outperforms alternative baselines, achieving 25% improvement on seen and 96% improvement on novel classes of the ImageNet dataset.

Aiming to find a program satisfying the user intent given input-output examples, program synthesis has attracted increasing interest in the area of machine learning. Despite the promising performance of existing methods, most of their success comes from the privileged information of well-designed input-output examples. However, providing such input-output examples is unrealistic because it requires the users to have the ability to describe the underlying program with a few input-output examples under the training distribution. In this work, we propose a query-based framework that trains a query neural network to generate informative input-output examples automatically and interactively from a large query space. The quality of the query depends on the amount of the mutual information between the query and the corresponding program, which can guide the optimization of the query framework. To estimate the mutual information more accurately, we introduce the functional space (F-space) which models the relevance between the input-output examples and the programs in a differentiable way. We evaluate the effectiveness and generalization of the proposed query-based framework on the Karel task and the list processing task. Experimental results show that the query-based framework can generate informative input-output examples which achieve and even outperform well-designed input-output examples.

Some neurons in deep networks specialize in recognizing highly specific perceptual, structural, or semantic features of inputs. In computer vision, techniques exist for identifying neurons that respond to individual concept categories like colors, textures, and object classes. But these techniques are limited in scope, labeling only a small subset of neurons and behaviors in any network. Is a richer characterization of neuron-level computation possible? We introduce a procedure (called MILAN, for mutual information-guided linguistic annotation of neurons) that automatically labels neurons with open-ended, compositional, natural language descriptions. Given a neuron, MILAN generates a description by searching for a natural language string that maximizes pointwise mutual information with the image regions in which the neuron is active. MILAN produces fine-grained descriptions that capture categorical, relational, and logical structure in learned features. These descriptions obtain high agreement with human-generated feature descriptions across a diverse set of model architectures and tasks, and can aid in understanding and controlling learned models. We highlight three applications of natural language neuron descriptions. First, we use MILAN for analysis, characterizing the distribution and importance of neurons selective for attribute, category, and relational information in vision models. Second, we use MILAN for auditing, surfacing neurons sensitive to human faces in datasets designed to obscure them. Finally, we use MILAN for editing, improving robustness in an image classifier by deleting neurons sensitive to text features spuriously correlated with class labels.

We propose a novel prediction interval (PI) method for uncertainty quantification, which addresses three major issues with the state-of-the-art PI methods. First, existing PI methods require retraining of neural networks (NNs) for every given confidence level and suffer from the crossing issue in calculating multiple PIs. Second, they usually rely on customized loss functions with extra sensitive hyperparameters for which fine tuning is required to achieve a well-calibrated PI. Third, they usually underestimate uncertainties of out-of-distribution (OOD) samples leading to over-confident PIs. Our PI3NN method calculates PIs from linear combinations of three NNs, each of which is independently trained using the standard mean squared error loss. The coefficients of the linear combinations are computed using root-finding algorithms to ensure tight PIs for a given confidence level. We theoretically prove that PI3NN can calculate PIs for a series of confidence levels without retraining NNs and it completely avoids the crossing issue. Additionally, PI3NN does not introduce any unusual hyperparameters resulting in a stable performance. Furthermore, we address OOD identification challenge by introducing an initialization scheme which provides reasonably larger PIs of the OOD samples than those of the in-distribution samples. Benchmark and real-world experiments show that our method outperforms several state-of-the-art approaches with respect to predictive uncertainty quality, robustness, and OOD samples identification.

Neural networks have enabled learning over examples that contain thousands of dimensions. However, most of these models are limited to training and evaluating on a finite collection of \textit{points} and do not consider the hypervolume in which the data resides. Any analysis of the model's local or global behavior is therefore limited to very expensive or imprecise estimators. We propose to formulate neural networks as a composition of a bijective (flow) network followed by a learnable, separable network. This construction allows for learning (or assessing) over full hypervolumes with precise estimators at tractable computational cost via integration over the \textit{input space}. We develop the necessary machinery, propose several practical integrals to use during training, and demonstrate their utility.

Audio-visual navigation task requires an agent to find a sound source in a realistic, unmapped 3D environment by utilizing egocentric audio-visual observations. Existing audio-visual navigation works assume a clean environment that solely contains the target sound, which, however, would not be suitable in most real-world applications due to the unexpected sound noise or intentional interference. In this work, we design an acoustically complex environment in which, besides the target sound, there exists a sound attacker playing a zero-sum game with the agent. More specifically, the attacker can move and change the volume and category of the sound to make the agent suffer from finding the sounding object while the agent tries to dodge the attack and navigate to the goal under the intervention. Under certain constraints to the attacker, we can improve the robustness of the agent towards unexpected sound attacks in audio-visual navigation. For better convergence, we develop a joint training mechanism by employing the property of a centralized critic with decentralized actors. Experiments on two real-world 3D scan datasets, Replica, and Matterport3D, verify the effectiveness and the robustness of the agent trained under our designed environment when transferred to the clean environment or the one containing sound attackers with random policy. Project: https://yyf17.github.io/SAAVN .

Learning a good transfer function to map the word vectors from two languages into a shared cross-lingual word vector space plays a crucial role in cross-lingual NLP. It is useful in translation tasks and important in allowing complex models built on a high-resource language like English to be directly applied on an aligned low resource language. While Procrustes and other techniques can align language models with some success, it has recently been identified that structural differences (for instance, due to differing word frequency) create different profiles for various monolingual embedding. When these profiles differ across languages, it correlates with how well languages can align and their performance on cross-lingual downstream tasks. In this work, we develop a very general language embedding normalization procedure, building and subsuming various previous approaches, which removes these structural profiles across languages without destroying their intrinsic meaning. We demonstrate that meaning is retained and alignment is improved on similarity, translation, and cross-language classification tasks. Our proposed normalization clearly outperforms all prior approaches like centering and vector normalization on each task and with each alignment approach.

Overparameterized neural networks generalize well but are expensive to train. Ideally one would like to reduce their computational cost while retaining their generalization benefits. Sparse model training is a simple and promising approach to achieve this, but there remain challenges as existing methods struggle with accuracy loss, slow training runtime, or difficulty in sparsifying all model components. The core problem is that searching for a sparsity mask over a discrete set of sparse matrices is difficult and expensive. To address this, our main insight is to optimize over a continuous superset of sparse matrices with a fixed structure known as products of butterfly matrices. As butterfly matrices are not hardware efficient, we propose simple variants of butterfly (block and flat) to take advantage of modern hardware. Our method (Pixelated Butterfly) uses a simple fixed sparsity pattern based on flat block butterfly and low-rank matrices to sparsify most network layers (e.g., attention, MLP). We empirically validate that Pixelated Butterfly is $3\times$ faster than Butterfly and speeds up training to achieve favorable accuracy--efficiency tradeoffs. On the ImageNet classification and WikiText-103 language modeling tasks, our sparse models train up to 2.3$\times$ faster than the dense MLP-Mixer, Vision Transformer, and GPT-2 small with no drop in accuracy.

Online learning via Bayes' theorem allows new data to be continuously integrated into an agent's current beliefs. However, a naive application of Bayesian methods in non-stationary environments leads to slow adaptation and results in state estimates that may converge confidently to the wrong parameter value. A common solution when learning in changing environments is to discard/downweight past data; however, this simple mechanism of "forgetting" fails to account for the fact that many real-world environments involve revisiting similar states. We propose a new framework, Bayes with Adaptive Memory (BAM), that takes advantage of past experience by allowing the agent to choose which past observations to remember and which to forget. We demonstrate that BAM generalizes many popular Bayesian update rules for non-stationary environments. Through a variety of experiments, we demonstrate the ability of BAM to continuously adapt in an ever-changing world.

We study the problem of learning verifiably safe parameters for programs that use neural networks as well as symbolic, human-written code. Such neurosymbolic programs arise in many safety-critical domains. However, because they need not be differentiable, it is hard to learn their parameters using existing gradient-based approaches to safe learning. Our method, Differentiable Symbolic Execution (DSE), samples control flow paths in a program, symbolically constructs worst-case "safety loss" along these paths, and backpropagates the gradients of these losses through program operations using a generalization of the REINFORCE estimator. We evaluate the method on a mix of synthetic tasks and real-world benchmarks. Our experiments show that DSE significantly outperforms the state-of-the-art DiffAI method on these tasks.

Popular approaches for minimizing loss in data-driven learning often involve an abstraction or an explicit retention of the history of gradients for efficient parameter updates. The aggregated history of gradients nudges the parameter updates in the right direction even when the gradients at any given step are not informative. Although the history of gradients summarized in meta-parameters or explicitly stored in memory has been shown effective in theory and practice, the question of whether $all$ or only a subset of the gradients in the history are sufficient in deciding the parameter updates remains unanswered. In this paper, we propose a framework of memory-augmented gradient descent optimizers that retain a limited view of their gradient history in their internal memory. Such optimizers scale well to large real-life datasets, and our experiments show that the memory augmented extensions of standard optimizers enjoy accelerated convergence and improved performance on a majority of computer vision and language tasks that we considered. Additionally, we prove that the proposed class of optimizers with fixed-size memory converge under assumptions of strong convexity, regardless of which gradients are selected or how they are linearly combined to form the update step.

This paper presents a simple MLP-like architecture, CycleMLP, which is a versatile backbone for visual recognition and dense predictions. As compared to modern MLP architectures, e.g. , MLP-Mixer, ResMLP, and gMLP, whose architectures are correlated to image size and thus are infeasible in object detection and segmentation, CycleMLP has two advantages compared to modern approaches. (1) It can cope with various image sizes. (2) It achieves linear computational complexity to image size by using local windows. In contrast, previous MLPs have $O(N^2)$ computations due to fully spatial connections. We build a family of models which surpass existing MLPs and even state-of-the-art Transformer-based models, e.g. Swin Transformer, while using fewer parameters and FLOPs. We expand the MLP-like models’ applicability, making them a versatile backbone for dense prediction tasks. CycleMLP achieves competitive results on object detection, instance segmentation, and semantic segmentation. In particular, CycleMLP-Tiny outperforms Swin-Tiny by 1.3% mIoU on ADE20K dataset with fewer FLOPs. Moreover, CycleMLP also shows excellent zero-shot robustness on ImageNet-C dataset.

Positive and Unlabeled (PU) learning targets inducing a binary classifier from weak training datasets of positive and unlabeled instances, which arise in many real-world applications. In this paper, we propose a novel PU learning method, namely Positive and unlabeled learning with Partially Positive Mixup (P3Mix), which simultaneously benefits from data augmentation and supervision correction with a heuristic mixup technique. To be specific, we take inspiration from the directional boundary deviation phenomenon observed in our preliminary experiments, where the learned PU boundary tends to deviate from the fully supervised boundary towards the positive side. For the unlabeled instances with ambiguous predictive results, we select their mixup partners from the positive instances around the learned PU boundary, so as to transform them into augmented instances near to the boundary yet with more precise supervision. Accordingly, those augmented instances may push the learned PU boundary towards the fully supervised boundary, thereby improving the classification performance. Comprehensive experimental results demonstrate the effectiveness of the heuristic mixup technique in PU learning and show that P3Mix can consistently outperform the state-of-the-art PU learning methods.

Recent work in equivariant deep learning bears strong similarities to physics. Fields over a base space are fundamental entities in both subjects, as are equivariant maps between these fields. In deep learning, however, these maps are usually defined by convolutions with a kernel, whereas they are partial differential operators (PDOs) in physics. Developing the theory of equivariant PDOs in the context of deep learning could bring these subjects even closer together and lead to a stronger flow of ideas. In this work, we derive a $G$-steerability constraint that completely characterizes when a PDO between feature vector fields is equivariant, for arbitrary symmetry groups $G$. We then fully solve this constraint for several important groups. We use our solutions as equivariant drop-in replacements for convolutional layers and benchmark them in that role. Finally, we develop a framework for equivariant maps based on Schwartz distributions that unifies classical convolutions and differential operators and gives insight about the relation between the two.

In the online continual learning paradigm, agents must learn from a changing distribution while respecting memory and compute constraints. Experience Replay (ER), where a small subset of past data is stored and replayed alongside new data, has emerged as a simple and effective learning strategy. In this work, we focus on the change in representations of observed data that arises when previously unobserved classes appear in the incoming data stream, and new classes must be distinguished from previous ones. We shed new light on this question by showing that applying ER causes the newly added classes’ representations to overlap significantly with the previous classes, leading to highly disruptive parameter updates. Based on this empirical analysis, we propose a new method which mitigates this issue by shielding the learned representations from drastic adaptation to accommodate new classes. We show that using an asymmetric update rule pushes new classes to adapt to the older ones (rather than the reverse), which is more effective especially at task boundaries, where much of the forgetting typically occurs. Empirical results show significant gains over strong baselines on standard continual learning benchmarks.

Parallelizing Gated Recurrent Unit (GRU) is a challenging task, as the training procedure of GRU is inherently sequential. Prior efforts to parallelize GRU have largely focused on conventional parallelization strategies such as data-parallel and model-parallel training algorithms. However, when the given sequences are very long, existing approaches are still inevitably performance limited in terms of both training time and model accuracy. In this paper, we present a novel parallel training scheme (called parallel-in-time) for GRU based on a multigrid reduction in time (MGRIT) solver. MGRIT partitions a sequence into multiple shorter sub-sequences and trains the sub-sequences on different processors in parallel. The key to achieving speedup is a hierarchical correction of the hidden state to accelerate end-to-end communication in both the forward and backward propagation phases of gradient descent. Experimental results on the HMDB51 dataset, where each video is an image sequence, demonstrate that a new parallel training scheme of GRU achieves up to $6.5 \times$ speedup over a serial approach. As efficiency of our new parallelization strategy is associated with the sequence length, our parallel GRU algorithm achieves significant performance improvement as the length of sequence increases. Further, the proposed approach can be applied simultaneously with batch and other forms of model parallelism.