Overcoming language priors via shuffling language bias for Robust Visual Question Answering

Recent research has revealed the notorious language prior problem in visual question answering (VQA) tasks based on visual-textual interaction, which indicates that well-developed VQA models rely learning shortcuts from questions without fully considering evidence.To tackle this problem, most existing methods focus decreasing incentive to learn knowledge by adding a question-only branch and becoming complacent mechanically improving accuracy. However, these over-correct positive biases useful for generalization, leading degradation of performance v2 dataset when cumulating their into other architecture. In paper, we propose robust shuffling bias (SLB) approach explicitly balance prediction distribution, hopefully alleviating increasing training opportunities models.Experiment results demonstrate our method is cumulative with data augmentation large-scale pre-training architectures achieves competitive both in-domain benchmark out-of-distribution VQA-CP v2.