Adversarial training is an effective technique for enhancing the robustness of deep neural networks (DNNs). Prior research shows that misclassified examples influence final adversarial robustness much more than correctly classified examples. Ignoring this difference during training can hurt model performance. In crowdsourcing, varying annotator expertise causes noisy, inconsistent labels. As a result, it is hard to distinguish misclassified and correctly classified examples using only provided annotations. Thus, how to use the reliability and discrepancy between these example types to improve robustness within adversarial learning remains a critical but underexplored issue. In this work, we first explore how misclassified and correctly classified examples affect learning from crowds (LFC) in adversarial environments. Then, we formulate the problem of misclassification-aware robust learning from multiple human labelers as a bilevel min-max problem. After that, we introduce MALC, a new approach to make classifiers more robust to adversarial examples via iterative adversarial example generation and parameter estimation. We conduct an extensive evaluation of the proposed MALC, showing that MALC can outperform the state-of-the-art LFC methods in both white-box and black-box settings.

Autonomous driving has made significant progress in both academia and industry, including performance improvements in perception tasks and the development of end-to-end autonomous driving systems. However, the safety and robustness assessment of autonomous driving has not received sufficient attention. Current evaluations of autonomous driving are typically conducted in natural driving scenarios. However, accidents often occur in edge cases, also known as safety-critical scenarios. These safety-critical scenarios are difficult to collect, and there is currently no clear definition of what constitutes a safety-critical scenario. In this work, we explore the safety and robustness of autonomous driving in safety-critical scenarios. First, we provide a definition of safety-critical scenarios, including static traffic scenarios such as adversarial attack scenarios and natural distribution shifts, as well as dynamic traffic scenarios such as accident scenarios. Then, we develop an autonomous driving test framework to comprehensively evaluate autonomous driving systems, encompassing not only the assessment of perception modules but also system-level evaluations. Our work systematically constructs a safety verification process for autonomous driving, providing technical support for the industry to establish standardized test framework.

The open-set domain adaptation (DA) aims to address both covariate shift and category shift between a labeled source domain and an unlabeled target domain. Nevertheless, existing open-set DA methods always ignore the demand for discovering novel classes that are not present in the source domain and simply reject them as "unknown" sets without further exploration, which motivates us to understand the unknown sets more specifically. In this article, we present a more challenging open-world DA problem that recognizes seen classes while discovering novel classes in the target domain. To address this problem, we propose a novel framework that converts this problem into a clustering task via contrastive learning to learn pairwise relationships among the instances. More specifically, our method consists of two iterative steps. The semi-supervised clustering step clusters the unlabeled target data and separates it into seen and novel classes. In the contrastive learning step, based on the cluster assignments, we design tailored contrastive losses that learn pairwise relationships to reduce domain discrepancy and discover novel classes. Our method can be optimized as an example of expectation maximization (EM). We establish several baselines by extending related work. Our method obtains the superior performance on five public datasets, benchmarking this challenging setting for future research.

Interpreting model behavior is crucial for model evaluation and optimization. Recent research demonstrates that incorporating human intelligence into the learning process effectively improve the interpretability and performance of the machine learning models, especially for simple classification tasks. However, the image captioning task has not received much attention. Such complex sequential tasks generally contain semantic relationships between different concepts, which pose challenges for interpreting model behavior and developing optimization methods. In this paper, we present RA3(Relation-Aware Attribution Analysis), a human-in-the-loop framework, for improving the interpretability, and further boosting the performance of the image captioning model. Specifically, we first engage human participants in two types of annotation tasks to identify what the model actually focuses on (model attribution) and what it should focus on (human rationale) at the conceptual level, supported by machine learning interpretability methods. Then, we identify and filter hard instances based on relation-aware model attribution for both validating the quality of the explanation and eliminating low-quality captions (this process is also considered as a kind of data debugging). We subsequently designed an explanation loss that penalizes the difference between model attribution and human rationale to optimize the model's behavior for improving caption quality. Through extensive experiments on crowdsourced annotations and MSCOCO, the experiment results indicate that the explanations produced by RA3can accurately describe the model's behavior, effectively identify difficult instances, and significantly improve the caption quality.

Unsupervised multi-class domain adaptation (multi-class UDA) has recently been proposed to fill the gap between empirically practical methods for multi-class classification and well-founded theory with the setting of binary classification. Nevertheless, the multi-class UDA methods use model predictions to characterize the disagreement of multi-class scoring hypotheses, which is used to optimize the divergence between domain distributions. Such self-training manner may bring inaccurate model predictions, which would damage the target structure due to the absence of labels of the target domain, leading to sub-optimal performance. On the other hand, this disagreement between multi-class scoring hypotheses does not involve the relationships among all of the multiple classes. It causes that multi-class UDA cannot properly connect the advanced practical UDA methods that consider class-conditional distribution alignment. Thus, we propose to exploit the target structure information and then incorporate it into multi-class UDA to achieve class-conditional distribution alignment. We theoretically and experimentally explain the importance of accurate target structure information to reduce the expected error on the target domain. Notably, our method achieves state-of-the-art results on three commonly-used benchmarks with different scales. In addition, using the target structure information, we propose a variant to cope with noisy open-world source domains such as noisy labels and out-of-distribution samples, enhancing the robustness of our method. The source code is available at https://github.com/jingzhengli/Multi_Class_UDA .

Pre-trained Vision-Language Models (VLMs) on large-scale image-text pairs, e.g., CLIP, have shown promising performance on zero-shot knowledge transfer. Recently, fine-tuning pre-trained VLMs to downstream few-shot classification with limited image annotation data yields significant gains. However, there are two limitations. First, most of the methods for fine-tuning VLMs only update newly added parameters while keeping the whole VLM frozen. Thus, it remains unclear how to directly update the VLM itself. Second, fine-tuning VLMs to a specific set of base classes would deteriorate the well-learned representation space such that the VLMs generalize poorly on novel classes. To address these issues, we first propose Layer-wise Fine-Tuning (LiFT) which achieves average gains of 3.9%, 4.3%, 4.2% and 4.5% on base classes under 2-, 4-, 8- and 16-shot respectively compared to the baseline CoOp over 11 datasets. Alternatively, we provide a parameter-efficient LiFT-Adapter exhibiting favorable performance while updating only 1.66% of total parameters. Further, we design scalable LiFT-NCD to identify both base classes and novel classes, which boosts the accuracy by an average of 5.01% over zero-shot generalization of CLIP, exploring the potential of VLMs in discovering novel classes.

Deep Neural Networks (DNNs) have shown impressive performance on large-scale training data with high-quality annotations. However, the collected annotations inevitably contain inaccurate labels in consideration of time and money budget, which causes DNNs to generalize poorly on the test set. To combat noisy labels in deep learning, the label correction methods are dedicated to simultaneously updating model parameters and correcting noisy labels, in which the noisy labels are usually corrected based on model predictions, the topological structures of data, or the aggregation of multiple models. However, such self-training manner cannot guarantee that the direction of label correction is always reliable. In view of this, we propose a novel label correction method to supervise and guide the process of label correction. In particular, the proposed label correction is an online two-fold process at each iteration only through back-propagation. The first label correction minimizes the empirical risk on noisy training data using noise-tolerant loss function, and the second label correction adopts a meta-learning paradigm to rectify the direction of first label correction so that the model can perform optimally in the evaluation procedure. Extensive experiments demonstrate the effectiveness of the proposed method on synthetic datasets with varying noise types and noise rates. Notably, our method achieves test accuracy of 77.37% on the real-world Clothing1M dataset.