IJCAI 2022 Papers

Automatic and accurate retinal disease diagnosis is critical to guide proper therapy and prevent potential vision loss. Previous works simply exploit the most discriminative features while ignoring the pathological visual clues of scattered subtle lesions. Therefore, without a comprehensive understanding of features from different lesion regions, they are vulnerable to noise from complex backgrounds and suffer from misclassification failures. In this paper, we address these limitations with a novel saliency-guided abnormality-aware transformer which explicitly captures the correlation between different lesion features from a global perspective with enhanced pathological semantics. The model has several merits. First, we propose a saliency enhancement module (SEM) which adaptively integrates disease related semantics and highlights potentially salient lesion regions. Second, to the best of our knowledge, this is the first work to explore comprehensive lesion feature dependencies via a tailored efficient self-attention. Third, with the saliency enhancement module and abnormality-aware attention, we propose a new variant of Vision Transformer models, called SatFormer, which outperforms the state-of-the-art methods on two public retinal disease classification benchmarks. Ablation study shows that the proposed components can be easily embedded into any Vision Transformers via a plug-and-play manner and effectively boost the performance.

A fundamental problem in adversarial machine learning is to quantify how much training data is needed in the presence of evasion attacks. In this paper we address this issue within the framework of PAC learning, focusing on the class of decision lists. Given that distributional assumptions are essential in the adversarial setting, we work with probability distributions on the input data that satisfy a Lipschitz condition: nearby points have similar probability. Our key results illustrate that the adversary's budget (that is, the number of bits it can perturb on each input) is a fundamental quantity in determining the sample complexity of robust learning. Our first main result is a sample-complexity lower bound: the class of monotone conjunctions (essentially the simplest non-trivial hypothesis class on the Boolean hypercube) and any superclass has sample complexity at least exponential in the adversary's budget. Our second main result is a corresponding upper bound: for every fixed k the class of k-decision lists has polynomial sample complexity against a log(n)-bounded adversary. This sheds further light on the question of whether an efficient PAC learning algorithm can always be used as an efficient log(n)-robust learning algorithm under the uniform distribution.

Dominant scene text recognition models commonly contain two building blocks, a visual model for feature extraction and a sequence model for text transcription. This hybrid architecture, although accurate, is complex and less efficient. In this study, we propose a Single Visual model for Scene Text recognition within the patch-wise image tokenization framework, which dispenses with the sequential modeling entirely. The method, termed SVTR, firstly decomposes an image text into small patches named character components. Afterward, hierarchical stages are recurrently carried out by component-level mixing, merging and/or combining. Global and local mixing blocks are devised to perceive the inter-character and intra-character patterns, leading to a multi-grained character component perception. Thus, characters are recognized by a simple linear prediction. Experimental results on both English and Chinese scene text recognition tasks demonstrate the effectiveness of SVTR. SVTR-L (Large) achieves highly competitive accuracy in English and outperforms existing methods by a large margin in Chinese, while running faster. In addition, SVTR-T (Tiny) is an effective and much smaller model, which shows appealing speed at inference. The code is publicly available at https://github.com/PaddlePaddle/PaddleOCR.

As deep learning advances, there is an ever-growing demand for models capable of synthesizing information from multi-modal resources to address the complex tasks raised from real-life applications. Recently, many large multi-modal datasets have been collected, on which researchers actively explore different methods of fusing multi-modal information. However, little attention has been paid to quantifying the contribution of different modalities within the proposed models. In this paper, we propose the SHapley vAlue-based PErceptual (SHAPE) scores that measure the marginal contribution of individual modalities and the degree of cooperation across modalities. Using these scores, we systematically evaluate different fusion methods on different multi-modal datasets for different tasks. Our experiments suggest that for some tasks where different modalities are complementary, the multi-modal models still tend to use the dominant modality alone and ignore the cooperation across modalities. On the other hand, models learn to exploit cross-modal cooperation when different modalities are indispensable for the task. In this case, the scores indicate it is better to fuse different modalities at relatively early stages. We hope our scores can help improve the understanding of how the present multi-modal models operate on different modalities and encourage more sophisticated methods of integrating multiple modalities.

Heterogeneous graphs have multiple node and edge types and are semantically richer than homogeneous graphs. To learn such complex semantics, many graph neural network approaches for heterogeneous graphs use metapaths to capture multi-hop interactions between nodes. Typically, features from non-target nodes are not incorporated into the learning procedure. However, there can be nonlinear, high-order interactions involving multiple nodes or edges. In this paper, we present Simplicial Graph Attention Network (SGAT), a simplicial complex approach to represent such high-order interactions by placing features from non-target nodes on the simplices. We then use attention mechanisms and upper adjacencies to generate representations. We empirically demonstrate the efficacy of our approach with node classification tasks on heterogeneous graph datasets and further show SGAT's ability in extracting structural information by employing random node features. Numerical experiments indicate that SGAT performs better than other current state-of-the-art heterogeneous graph learning methods.

Deep neural networks are prone to overfitting noisy labels, resulting in poor generalization performance. To overcome this problem, we present a simple and effective method self-ensemble label correction (SELC) to progressively correct noisy labels and refine the model. We look deeper into the memorization behavior in training with noisy labels and observe that the network outputs are reliable in the early stage. To retain this reliable knowledge, SELC uses ensemble predictions formed by an exponential moving average of network outputs to update the original noisy labels. We show that training with SELC refines the model by gradually reducing supervision from noisy labels and increasing supervision from ensemble predictions. Despite its simplicity, compared with many state-of-the-art methods, SELC obtains more promising and stable results in the presence of class-conditional, instance-dependent, and real-world label noise. The code is available at https://github.com/MacLLL/SELC.

Semi-Supervised Object Detection (SSOD) aims to improve performance by leveraging a large amount of unlabeled data. Existing works usually adopt the teacher-student framework to enforce student to learn consistent predictions over the pseudo-labels generated by teacher. However, the performance of the student model is limited since the noise inherently exists in pseudo-labels. In this paper, we investigate the causes and effects of noisy pseudo-labels and propose a simple yet effective approach denoted as Self-Correction Mean Teacher(SCMT) to reduce the adverse effects. Specifically, we propose to dynamically re-weight the unsupervised loss of each student's proposal with additional supervision information from the teacher model, and assign smaller loss weights to possible noisy proposals. Extensive experiments on MS-COCO benchmark have shown the superiority of our proposed SCMT, which can significantly improve the supervised baseline by more than 11% mAP under all 1%, 5% and 10% COCO-standard settings, and surpasses state-of-the-art methods by about 1.5% mAP. Even under the challenging COCO-additional setting, SCMT still improves the supervised baseline by 4.9% mAP, and significantly outperforms previous methods by 1.2% mAP, achieving a new state-of-the-art performance.

Synthetic Aperture Radar (SAR) becomes prevailing in remote sensing while SAR images are challenging to interpret by human visual perception due to the active imaging mechanism and speckle noise. Recent researches on SAR-to-optical image translation provide a promising solution and have attracted increasing attentions, though still suffering from low optical image quality with geometric distortion due to the large domain gap. In this paper, we mitigate this issue from a novel perspective, i.e., neural partial differential equations (PDE). First, based on the efficient numerical scheme for solving PDE, i.e., Taylor Central Difference (TCD), we devise a basic TCD residual block to build the backbone network, which promotes the extraction of useful information in SAR images by aggregating and enhancing features from different levels. Furthermore, inspired by the Perona-Malik Diffusion (PMD), we devise a PMD neural module to implement feature diffusion through layers, aiming at removing the noises in smooth regions while preserving the geometric structures. Assembling them together, we propose a novel SAR-to-Optical image translation network named S2O-NPDE, which delivers optical images with finer structures and less noise while enjoying an explainability advantage from explicit mathematical derivation. Experiments on the popular SEN1-2 dataset show that our model outperforms state-of-the-art methods in terms of both objective metrics and visual quality.

Superconductivity allows electrical current to flow without any energy loss, and thus making solids superconducting is a grand goal of physics, material science, and electrical engineering. More than 16 Nobel Laureates have been awarded for their contribution in superconductivity research. Superconductors are valuable for sustainable development goals (SDGs), such as climate change mitigation, affordable and clean energy, industry, innovation and infrastructure, and so on. However, a unified physics theory explaining all superconductivity mechanism is still unknown. It is believed that superconductivity is microscopically due to not only molecular compositions but also the geometric crystal structure. Hence a new dataset, S2S, containing both crystal structures and superconducting critical temperature, is built upon SuperCon and Material Project. Based on this new dataset, we propose a novel model, S2SNet, which utilizes the attention mechanism for superconductivity prediction. To overcome the shortage of data, S2SNet is pre-trained on the whole Material Project dataset with Masked-Language Modeling (MLM). S2SNet makes a new state-of-the-art, with out-of-sample accuracy of 92% and Area Under Curve (AUC) of 0.92. To the best of our knowledge, S2SNet is the first work to predict superconductivity with only information of crystal structures. This work is beneficial to superconductivity discovery and further SDGs. The code and datasets are available at https://github.com/supercond/S2SNet

Transformer-based architectures with grid features represent the state-of-the-art in visual and language reasoning tasks, such as visual question answering and image-text matching. However, directly applying them to image captioning may result in spatial and fine-grained semantic information loss. Their applicability to image captioning is still largely under-explored. Towards this goal, we propose a simple yet effective method, Spatial- and Scale-aware Transformer (S2 Transformer) for image captioning. Specifically, we firstly propose a Spatial-aware Pseudo-supervised (SP) module, which resorts to feature clustering to help preserve spatial information for grid features. Next, to maintain the model size and produce superior results, we build a simple weighted residual connection, named Scale-wise Reinforcement (SR) module, to simultaneously explore both low- and high-level encoded features with rich semantics. Extensive experiments on the MSCOCO benchmark demonstrate that our method achieves new state-of-art performance without bringing excessive parameters compared with the vanilla transformer. The source code is available at https://github.com/zchoi/S2-Transformer

Prescription (aka Rx) drugs can be easily overprescribed and lead to drug abuse or opioid overdose. Accordingly, a state-run prescription drug monitoring program (PDMP) in the United States has been developed to reduce overprescribing. However, PDMP has limited capability in detecting patients' potential overprescribing behaviors, impairing its effectiveness in preventing drug abuse and overdose in patients. In this paper, we propose a novel model RxNet, which builds 1) a dynamic heterogeneous graph to model Rx refills that are essentially prescribing and dispensing (P&D) relationships among various patients, 2) an RxLSTM network to explore the dynamic Rx-refill behavior and medical condition variation of patients, and 3) a dosing-adaptive network to extract and recalibrate dosing patterns and obtain the refined patient representations which are finally utilized for overprescribing detection. The extensive experimental results on a one-year state-wide PDMP data demonstrate that RxNet consistently outperforms state-of-the-art methods in predicting patients at high risk of opioid overdose and drug abuse.

Chance constrained optimization problems allow to model problems where constraints involving stochastic components should only be violated with a small probability. Evolutionary algorithms have been applied to this scenario and shown to achieve high quality results. With this paper, we contribute to the theoretical understanding of evolutionary algorithms for chance constrained optimization. We study the scenario of stochastic components that are independent and Normally distributed. Considering the simple single-objective (1+1)~EA, we show that imposing an additional uniform constraint already leads to local optima for very restricted scenarios and an exponential optimization time. We therefore introduce a multi-objective formulation of the problem which trades off the expected cost and its variance. We show that multi-objective evolutionary algorithms are highly effective when using this formulation and obtain a set of solutions that contains an optimal solution for any possible confidence level imposed on the constraint. Furthermore, we prove that this approach can also be used to compute a set of optimal solutions for the chance constrained minimum spanning tree problem. Experimental investigations on instances of the NP-hard stochastic minimum weight dominating set problem confirm the benefit of the multi-objective approach in practice.

In many domains, worst-case guarantees on the performance (e.g. prediction accuracy) of a decision function subject to distributional shifts and uncertainty about the environment are crucial. In this work we develop a method to quantify the robustness of decision functions with respect to credal Bayesian networks, formal parametric models of the environment where uncertainty is expressed through credal sets on the parameters. In particular, we address the maximum marginal probability (MARmax) problem, that is, determining the greatest probability of an event (such as misclassification) obtainable for parameters in the credal set. We develop a method to faithfully transfer the problem into a constrained optimization problem on a probabilistic circuit. By performing a simple constraint relaxation, we show how to obtain a guaranteed upper bound on MARmax in linear time in the size of the circuit. We further theoretically characterize this constraint relaxation in terms of the original Bayesian network structure, which yields insight into the tightness of the bound. We implement the method and provide experimental evidence that the upper bound is often near tight and demonstrates improved scalability compared to other methods.

Vision Transformer (ViT) is becoming more popular in image processing. Specifically, we investigate the effectiveness of test-time adaptation (TTA) on ViT, a technique that has emerged to correct its prediction during test-time by itself. First, we benchmark various test-time adaptation approaches on ViT-B16 and ViT-L16. It is shown that the TTA is effective on ViT and the prior-convention (sensibly selecting modulation parameters) is not necessary when using proper loss function. Based on the observation, we propose a new test-time adaptation method called class-conditional feature alignment (CFA), which minimizes both the class-conditional distribution differences and the whole distribution differences of the hidden representation between the source and target in an online manner. Experiments of image classification tasks on common corruption (CIFAR-10-C, CIFAR-100-C, and ImageNet-C) and domain adaptation (digits datasets and ImageNet-Sketch) show that CFA stably outperforms the existing baselines on various datasets. We also verify that CFA is model agnostic by experimenting on ResNet, MLP-Mixer, and several ViT variants (ViT-AugReg, DeiT, and BeiT). Using BeiT backbone, CFA achieves 19.8% top-1 error rate on ImageNet-C, outperforming the existing test-time adaptation baseline 44.0%. This is a state-of-the-art result among TTA methods that do not need to alter training phase.

Overfitting widely exists in adversarial robust training of deep networks. An effective remedy is adversarial weight perturbation, which injects the worst-case weight perturbation during network training by maximizing the classification loss on adversarial examples. Adversarial weight perturbation helps reduce the robust generalization gap; however, it also undermines the robustness improvement. A criterion that regulates the weight perturbation is therefore crucial for adversarial training. In this paper, we propose such a criterion, namely Loss Stationary Condition (LSC) for constrained perturbation. With LSC, we find that it is essential to conduct weight perturbation on adversarial data with small classification loss to eliminate robust overfitting. Weight perturbation on adversarial data with large classification loss is not necessary and may even lead to poor robustness. Based on these observations, we propose a robust perturbation strategy to constrain the extent of weight perturbation. The perturbation strategy prevents deep networks from overfitting while avoiding the side effect of excessive weight perturbation, significantly improving the robustness of adversarial training. Extensive experiments demonstrate the superiority of the proposed method over the state-of-the-art adversarial training methods.

Subset selection, which aims to select a subset from a ground set to maximize some objective function, arises in various applications such as influence maximization and sensor placement. In real-world scenarios, however, one often needs to find a subset which is robust against (i.e., is good over) a number of possible objective functions due to uncertainty, resulting in the problem of robust subset selection. This paper considers robust subset selection with monotone objective functions, relaxing the submodular property required by previous studies. We first show that the greedy algorithm can obtain an approximation ratio with respect to the correlation and submodularity ratios of the objective functions; and then propose EPORSS, an evolutionary Pareto optimization algorithm that can utilize more time to find better subsets. We prove that EPORSS can also be theoretically grounded, achieving a similar approximation guarantee to the greedy algorithm. In addition, we derive the lower bound of the correlation ratio for the application of robust influence maximization, and further conduct experiments to validate the performance of the greedy algorithm and EPORSS.

Multi-defender Stackelberg Security Games (MSSG) have recently gained increasing attention in the literature. However, the solutions offered to date are highly sensitive, wherein even small perturbations in the attacker's utility or slight uncertainties thereof can dramatically change the defenders' resulting payoffs and alter the equilibrium. In this paper, we introduce a robust model for MSSGs, which admits solutions that are resistant to small perturbations or uncertainties in the game's parameters. First, we formally define the notion of robustness, as well as the robust MSSG model. Then, for the non-cooperative setting, we prove the existence of a robust approximate equilibrium in any such game, and provide an efficient construction thereof. For the cooperative setting, we show that any such game admits a robust approximate (alpha) core, and provide an efficient construction thereof. Lastly, we show that stronger types of the core may be empty. Interestingly, the robust solutions can substantially increase the defenders' utilities over those of the non-robust ones.

Single image dehazing as a fundamental low-level vision task, is essential for the development of robust intelligent surveillance system. In this paper, we make an early effort to consider dehazing robustness under variational haze density, which is a realistic while under-studied problem in the research filed of singe image dehazing. To properly address this problem, we propose a novel density-variational learning framework to improve the robustness of the image dehzing model assisted by a variety of negative hazy images, to better deal with various complex hazy scenarios. Specifically, the dehazing network is optimized under the consistency-regularized framework with the proposed Contrast-Assisted Reconstruction Loss (CARL). The CARL can fully exploit the negative information to facilitate the traditional positive-orient dehazing objective function, by squeezing the dehazed image to its clean target from different directions. Meanwhile, the consistency regularization keeps consistent outputs given multi-level hazy images, thus improving the model robustness. Extensive experimental results on two synthetic and three real-world datasets demonstrate that our method significantly surpasses the state-of-the-art approaches.

Robust Reinforcement Learning (RL) focuses on improving performances under model errors or adversarial attacks, which facilitates the real-life deployment of RL agents. Robust Adversarial Reinforcement Learning (RARL) is one of the most popular frameworks for robust RL. However, most of the existing literature models RARL as a zero-sum simultaneous game with Nash equilibrium as the solution concept, which could overlook the sequential nature of RL deployments, produce overly conservative agents, and induce training instability. In this paper, we introduce a novel hierarchical formulation of robust RL -- a general-sum Stackelberg game model called RRL-Stack -- to formalize the sequential nature and provide extra flexibility for robust training. We develop the Stackelberg Policy Gradient algorithm to solve RRL-Stack, leveraging the Stackelberg learning dynamics by considering the adversary's response. Our method generates challenging yet solvable adversarial environments which benefit RL agents' robust learning. Our algorithm demonstrates better training stability and robustness against different testing conditions in the single-agent robotics control and multi-agent highway merging tasks.