VQA-PDF: Purifying Debiased Features for Robust Visual Question Answering Task

With the widespread adoption of deep learning, the performance of Visual Question Answering (VQA) tasks has seen significant improvements. Nonetheless, this progress has unveiled significant challenges concerning their credibility, primarily due to the susceptibility of linguistic biases. Such biases can result in considerable declines in performance when faced with out-of-distribution scenarios. Therefore, various debiasing methods have been developed to reduce the impact of linguistic biases, where causal theory-based methods have attracted great attention due to their theoretical underpinnings and superior performance. However, traditional debiased causal strategies typically remove biases through simple subtraction, which neglects the fine-grained bias information, resulting in incomplete debiasing. To tackle this issue, we propose a fine-grained debiasing method named as VQA-PDF, which utilizes the features of the base model to guide the identification of biased features, purifying the debiased features and aiding the base learning process. This method has shown significant improvements on VQA-CP v2, VQA v2 and VQA-CE datasets.