On Modality Bias Recognition and Reduction

TL;DR

This paper systematically studies modality bias in multi-modal classification, highlighting how spurious correlations cause dominance of certain modalities, and proposes a loss function to mitigate this bias, improving model performance.

Contribution

It introduces a comprehensive analysis of modality bias, creates new OoD datasets for evaluation, and proposes a plug-and-play loss to reduce bias and enhance multi-modal learning.

Findings

The proposed method improves performance across multiple datasets.

Existing methods suffer from modality bias in OoD settings.

The loss function effectively reduces modality dominance in models.

Abstract

Making each modality in multi-modal data contribute is of vital importance to learning a versatile multi-modal model. Existing methods, however, are often dominated by one or few of modalities during model training, resulting in sub-optimal performance. In this paper, we refer to this problem as modality bias and attempt to study it in the context of multi-modal classification systematically and comprehensively. After stepping into several empirical analysis, we recognize that one modality affects the model prediction more just because this modality has a spurious correlation with instance labels. In order to primarily facilitate the evaluation on the modality bias problem, we construct two datasets respectively for the colored digit recognition and video action recognition tasks in line with the Out-of-Distribution (OoD) protocol. Collaborating with the benchmarks in the visual question answering task, we empirically justify the performance degradation of the existing methods on these OoD datasets, which serves as evidence to justify the modality bias learning. In addition, to overcome this problem, we propose a plug-and-play loss function method, whereby the feature space for each label is adaptively learned according to the training set statistics. Thereafter, we apply this method on eight baselines in total to test its effectiveness. From the results on four datasets regarding the above three tasks, our method yields remarkable performance improvements compared with the baselines, demonstrating its superiority on reducing the modality bias problem.