A Multidimensional Analysis of Social Biases in Vision Transformers

TL;DR

This paper investigates the sources of social biases in Vision Transformers, analyzing how data, architecture, and training objectives influence bias, and explores mitigation strategies like diffusion-based augmentation.

Contribution

It provides a detailed analysis of factors affecting social biases in ViTs and demonstrates how model design choices can reduce biases, offering new insights into bias mitigation.

Findings

Larger models are less biased than smaller ones.

Counterfactual augmentation can mitigate but not eliminate biases.

Different training objectives can lead to opposite biases.

Abstract

The embedding spaces of image models have been shown to encode a range of social biases such as racism and sexism. Here, we investigate specific factors that contribute to the emergence of these biases in Vision Transformers (ViT). Therefore, we measure the impact of training data, model architecture, and training objectives on social biases in the learned representations of ViTs. Our findings indicate that counterfactual augmentation training using diffusion-based image editing can mitigate biases, but does not eliminate them. Moreover, we find that larger models are less biased than smaller models, and that models trained using discriminative objectives are less biased than those trained using generative objectives. In addition, we observe inconsistencies in the learned social biases. To our surprise, ViTs can exhibit opposite biases when trained on the same data set using different self-supervised objectives. Our findings give insights into the factors that contribute to the emergence of social biases and suggests that we could achieve substantial fairness improvements based on model design choices.