FACTS: First Amplify Correlations and Then Slice to Discover Bias

TL;DR

FACTS is a method that amplifies dataset correlations and then slices data to identify bias-related subsets, improving bias detection in computer vision datasets.

Contribution

It introduces a novel two-step approach combining correlation amplification and mixture modeling to better discover bias-conflicting data slices.

Findings

Significantly outperforms prior bias identification methods by up to 35% precision@10.

Effectively identifies diverse bias-conflicting data slices in computer vision datasets.

Provides a simple yet powerful framework for bias detection that can inform mitigation strategies.

Abstract

Computer vision datasets frequently contain spurious correlations between task-relevant labels and (easy to learn) latent task-irrelevant attributes (e.g. context). Models trained on such datasets learn "shortcuts" and underperform on bias-conflicting slices of data where the correlation does not hold. In this work, we study the problem of identifying such slices to inform downstream bias mitigation strategies. We propose First Amplify Correlations and Then Slice to Discover Bias (FACTS), wherein we first amplify correlations to fit a simple bias-aligned hypothesis via strongly regularized empirical risk minimization. Next, we perform correlation-aware slicing via mixture modeling in bias-aligned feature space to discover underperforming data slices that capture distinct correlations. Despite its simplicity, our method considerably improves over prior work (by as much as 35% precision@10) in correlation bias identification across a range of diverse evaluation settings. Our code is available at: https://github.com/yvsriram/FACTS.