Vision-language foundation models have exhibited remarkable success across a multitude of downstream tasks due to their scalability on extensive image-text paired datasets. However, these models display significant limitations when applied to long-tail tasks, such as fine-grained image classification, as a result of "decision shortcuts" that hinders their generalization capabilities. In this work, we find that the CLIP model possesses a rich set of features, encompassing both \textit{desired invariant causal features} and \textit{undesired decision shortcuts}. Moreover, the underperformance of CLIP on downstream tasks originates from its inability to effectively utilize pre-trained features in accordance with specific task requirements. To address this challenge, this paper introduces a test-time prompt tuning paradigm that optimizes a learnable prompt, thereby compelling the model to exploit genuine causal invariant features while disregarding decision shortcuts during the inference phase. The proposed method effectively alleviates excessive dependence on potentially misleading, task-irrelevant contextual information, while concurrently emphasizing critical, task-related visual cues. We conduct comparative analysis of the proposed method against various approaches which validates its effectiveness.
Miscalibration in deep learning refers to there is a discrepancy between the predicted confidence and performance. This problem usually arises due to the overfitting problem, which is characterized by learning everything presented in the training set, resulting in overconfident predictions during testing. Existing methods typically address overfitting and mitigate the miscalibration by adding a maximum-entropy regularizer to the objective function. The objective can be understood as seeking a model that fits the ground-truth labels by increasing the confidence while also maximizing the entropy of predicted probabilities by decreasing the confidence. However, previous methods lack clear guidance on confidence adjustment, leading to conflicting objectives (increasing but also decreasing confidence). Therefore, we introduce a method called Dynamic Regularization (DReg), which aims to learn what should be learned during training thereby circumventing the confidence adjusting trade-off. At a high level, DReg aims to obtain a more reliable model capable of acknowledging what it knows and does not know. Specifically, DReg effectively fits the labels for in-distribution samples (samples that should be learned) while applying regularization dynamically to samples beyond model capabilities (e.g., outliers), thereby obtaining a robust calibrated model especially on the samples beyond model capabilities. Both theoretical and empirical analyses sufficiently demonstrate the superiority of DReg compared with previous methods.
Recent advancements in large vision-language models (LVLMs) have demonstrated impressive capability in visual information understanding with human language. Despite these advances, LVLMs still face challenges with multimodal hallucination, such as generating text descriptions of objects that are not present in the visual information. However, the underlying fundamental reasons of multimodal hallucinations remain poorly explored. In this paper, we propose a new perspective, suggesting that the inherent biases in LVLMs might be a key factor in hallucinations. Specifically, we systematically identify a semantic shift bias related to paragraph breaks (\n\n), where the content before and after '\n\n' in the training data frequently exhibit significant semantic changes. This pattern leads the model to infer that the contents following '\n\n' should be obviously different from the preceding contents with less hallucinatory descriptions, thereby increasing the probability of hallucinatory descriptions subsequent to the '\n\n'. We have validated this hypothesis on multiple publicly available LVLMs. Besides, we find that deliberately inserting '\n\n' at the generated description can induce more hallucinations. A simple method is proposed to effectively mitigate the hallucination of LVLMs by skipping the output of '\n'.
Recent advancements in large vision-language models (LVLMs) have demonstrated impressive capability in visual information understanding with human language. Despite these advances, LVLMs still face challenges with multimodal hallucination, such as generating text descriptions of objects that are not present in the visual information. However, the underlying fundamental reasons of multimodal hallucinations remain poorly explored. In this paper, we propose a new perspective, suggesting that the inherent biases in LVLMs might be a key factor in hallucinations. Specifically, we systematically identify a semantic shift bias related to paragraph breaks ('$\textbackslash n\textbackslash n$'), where the content before and after '$\textbackslash n\textbackslash n$' in the training data frequently exhibit significant semantic changes. This pattern leads the model to infer that the contents following '$\textbackslash n\textbackslash n$' should be obviously different from the preceding contents with less hallucinatory descriptions, thereby increasing the probability of hallucinatory descriptions subsequent to the '$\textbackslash n\textbackslash n$'. We have validated this hypothesis on multiple publicly available LVLMs. Besides, we find that deliberately inserting '$\textbackslash n\textbackslash n$' at the generated description can induce more hallucinations. A simple method is proposed to effectively mitigate the hallucination of LVLMs by skipping the output of `\textbackslash n'.
Out-of-distribution (OOD) detection methods often exploit auxiliary outliers to train model identifying OOD samples, especially discovering challenging outliers from auxiliary outliers dataset to improve OOD detection. However, they may still face limitations in effectively distinguishing between the most challenging OOD samples that are much like in-distribution (ID) data, i.e., ID-like samples. To this end, we propose a novel OOD detection framework that discovers ID-like outliers using CLIP from the vicinity space of the ID samples, thus helping to identify these most challenging OOD samples. Then a prompt learning framework is proposed that utilizes the identified ID-like outliers to further leverage the capabilities of CLIP for OOD detection. Benefiting from the powerful CLIP, we only need a small number of ID samples to learn the prompts of the model without exposing other auxiliary outlier datasets. By focusing on the most challenging ID-like OOD samples and elegantly exploiting the capabilities of CLIP, our method achieves superior few-shot learning performance on various real-world image datasets (e.g., in 4-shot OOD detection on the ImageNet-1k dataset, our method reduces the average FPR95 by 12.16% and improves the average AUROC by 2.76%, compared to state-of-the-art methods).
Nowadays, billions of people engage in communication and express their opinions on the internet daily. Unfortunately, not all of these expressions are friendly or compliant, making content moderation an indispensable task. With the successful development of Large Language Models (LLMs) in recent years, LLM-based methods have become a feasible solution for handling tasks in various domains. However, in the field of content moderation, there is still a lack of detailed work that systematically introduces implementation details. In this paper, we introduce how to fine-tune an LLM model that can be privately deployed for content moderation. Specifically, we discuss whether incorporating reasons during the fine-tuning process would be better or if it should be treated as a classification task directly. We also explore the benefits of utilizing reasons generated by more powerful LLMs for fine-tuning privately deployed models and the impact of different processing approaches when the answers generated by the more powerful LLMs are incorrect. We report the entire research process and the key findings in this paper, hoping to provide valuable experience for researchers who are fine-tuning privately deployed models in their domain-specific research.
Mixup is a well-established data augmentation technique, which can extend the training distribution and regularize the neural networks by creating ''mixed'' samples based on the label-equivariance assumption, i.e., a proportional mixup of the input data results in the corresponding labels being mixed in the same proportion. However, previous mixup variants may fail to exploit the label-independent information in mixed samples during training, which usually contains richer semantic information. To further release the power of mixup, we first improve the previous label-equivariance assumption by the semantic-equivariance assumption, which states that the proportional mixup of the input data should lead to the corresponding representation being mixed in the same proportion. Then a generic mixup regularization at the representation level is proposed, which can further regularize the model with the semantic information in mixed samples. At a high level, the proposed semantic equivariant mixup (sem) encourages the structure of the input data to be preserved in the representation space, i.e., the change of input will result in the obtained representation information changing in the same way. Different from previous mixup variants, which tend to over-focus on the label-related information, the proposed method aims to preserve richer semantic information in the input with semantic-equivariance assumption, thereby improving the robustness of the model against distribution shifts. We conduct extensive empirical studies and qualitative analyzes to demonstrate the effectiveness of our proposed method. The code of the manuscript is in the supplement.
The inherent challenge of multimodal fusion is to precisely capture the cross-modal correlation and flexibly conduct cross-modal interaction. To fully release the value of each modality and mitigate the influence of low-quality multimodal data, dynamic multimodal fusion emerges as a promising learning paradigm. Despite its widespread use, theoretical justifications in this field are still notably lacking. Can we design a provably robust multimodal fusion method? This paper provides theoretical understandings to answer this question under a most popular multimodal fusion framework from the generalization perspective. We proceed to reveal that several uncertainty estimation solutions are naturally available to achieve robust multimodal fusion. Then a novel multimodal fusion framework termed Quality-aware Multimodal Fusion (QMF) is proposed, which can improve the performance in terms of classification accuracy and model robustness. Extensive experimental results on multiple benchmarks can support our findings.
Cutting out an object and estimating its opacity mask, known as image matting, is a key task in image and video editing. Due to the highly ill-posed issue, additional inputs, typically user-defined trimaps or scribbles, are usually needed to reduce the uncertainty. Although effective, it is either time consuming or only suitable for experienced users who know where to place the strokes. In this work, we propose a decomposed-uncertainty-guided matting (dugMatting) algorithm, which explores the explicitly decomposed uncertainties to efficiently and effectively improve the results. Basing on the characteristic of these uncertainties, the epistemic uncertainty is reduced in the process of guiding interaction (which introduces prior knowledge), while the aleatoric uncertainty is reduced in modeling data distribution (which introduces statistics for both data and possible noise). The proposed matting framework relieves the requirement for users to determine the interaction areas by using simple and efficient labeling. Extensively quantitative and qualitative results validate that the proposed method significantly improves the original matting algorithms in terms of both efficiency and efficacy.