Men Also Like Shopping: Reducing Gender Bias Amplification using Corpus-level Constraints

TL;DR

This paper addresses gender bias in web-sourced visual recognition datasets and models, proposing a corpus-level constraint method that significantly reduces bias amplification with minimal impact on accuracy.

Contribution

It introduces a novel corpus-level constraint approach using Lagrangian relaxation to reduce gender bias amplification in structured prediction models for visual recognition.

Findings

Bias in datasets and models is significant, with over 33% disparity in activity gender distribution.

The proposed method reduces bias amplification by approximately 40-50%.

Performance on recognition tasks remains nearly unchanged after applying the constraints.

Abstract

Language is increasingly being used to define rich visual recognition problems with supporting image collections sourced from the web. Structured prediction models are used in these tasks to take advantage of correlations between co-occurring labels and visual input but risk inadvertently encoding social biases found in web corpora. In this work, we study data and models associated with multilabel object classification and visual semantic role labeling. We find that (a) datasets for these tasks contain significant gender bias and (b) models trained on these datasets further amplify existing bias. For example, the activity cooking is over 33% more likely to involve females than males in a training set, and a trained model further amplifies the disparity to 68% at test time. We propose to inject corpus-level constraints for calibrating existing structured prediction models and design an algorithm based on Lagrangian relaxation for collective inference. Our method results in almost no performance loss for the underlying recognition task but decreases the magnitude of bias amplification by 47.5% and 40.5% for multilabel classification and visual semantic role labeling, respectively.