Enhancing FR-Train in Data Imbalance

Abstract

This report presents a comprehensive analysis of the FR-Train model, an architecture designed to concurrently enhance fairness and robustness in artificial intelligence (AI) systems. Our study offers a detailed investigation into the discriminator modules of FR-Train, complementing the original authors’ ablation study. We re-examined the key theorems using information theory principles, reproduced the model to verify its claims, and conducted additional experiments. Our main contributions are threefold: i) we provided additional information theory insights on the function of the fair discriminator; ii) we introduced an alternative approach by replacing the model’s fair discriminator with a naive mutual information estimate, yielding comparable performance with a faster and more stable training process; and iii) we evaluated FR-Train’s performance in extreme data imbalance scenarios and proposed a “normalized mutual information” loss function that can significantly alleviate the performance deterioration.