Generative Bias for Robust Visual Question Answering

TL;DR

This paper introduces GenB, a generative approach to learn and mitigate biases in Visual Question Answering models by directly modeling the bias from the target model, leading to improved robustness and state-of-the-art results.

Contribution

The paper proposes a novel generative bias model, GenB, that learns biases directly from the target VQA model using adversarial training and knowledge distillation.

Findings

GenB improves robustness on multiple VQA bias datasets

Achieves state-of-the-art results on VQA-CP2 with LXMERT

Effectively reduces bias exploitation in VQA models

Abstract

The task of Visual Question Answering (VQA) is known to be plagued by the issue of VQA models exploiting biases within the dataset to make its final prediction. Various previous ensemble based debiasing methods have been proposed where an additional model is purposefully trained to be biased in order to train a robust target model. However, these methods compute the bias for a model simply from the label statistics of the training data or from single modal branches. In this work, in order to better learn the bias a target VQA model suffers from, we propose a generative method to train the bias model directly from the target model, called GenB. In particular, GenB employs a generative network to learn the bias in the target model through a combination of the adversarial objective and knowledge distillation. We then debias our target model with GenB as a bias model, and show through extensive experiments the effects of our method on various VQA bias datasets including VQA-CP2, VQA-CP1, GQA-OOD, and VQA-CE, and show state-of-the-art results with the LXMERT architecture on VQA-CP2.