Robust data augmentation and contrast learning for debiased visual question answering

The language prior problem in VQA causes the model to make predictions directly based on spurious correlations between questions and answers, causing the model's performance to drop sharply outside the distribution. Current debiasing methods often achieve good out-of-distribution generalization capabilities at the expense of significant in-distribution performance degradation, while non-debiasing methods sacrifice a large amount of out-of-distribution performance to achieve high in-distribution performance. We propose a novel method combining multi-teacher knowledge distillation and contrastive learning (MKDCL) to solve the language prior problem in VQA. We propose a Question Answer Selection (QAS) module to select reasonable questions for images, which also determines the pseudo answers with multi-teacher's weighted predictions. The Contrastive Learning Samples Generation (CLSG) module we propose synthesizes four types of positive and negative samples in visual and language modalities for contrastive learning, effectively increases the semantic dependency of the images while avoiding performance degradation due to spurious correlations between questions and answers. Our method is model-agnostic and achieves state-of-the-art performance (62.93%) on the language prior-sensitive VQA-CP v2 dataset while maintaining performance (65.43%) on the VQA v2 dataset.