Simple contrastive learning in a self-supervised manner for robust visual question answering

Shuwen Yang, Luwei Xiao, Xingjiao Wu, Junjie Xu, Linlin Wang, Liang He

Abstract: Highlights•SCLSM mitigates the issue of language bias in VQA through self-supervised learning.•DCL enhances the focus of model on hard negatives by modifying the contrastive loss.•Our model can achieve comparable results for VQA debiasing tasks.