Improving Fairness and Mitigating MADness in Generative Models

TL;DR

This paper proposes a hypernetwork-based training approach for generative models that enhances fairness, stability, and reduces bias, supported by theoretical analysis and empirical validation.

Contribution

It introduces a novel hypernetwork training framework that improves fairness, stability, and bias mitigation in generative models, along with a scalable implementation for various architectures.

Findings

Models trained with hypernetworks are more fair for minority classes.

Hypernetwork training leads to more stable generative models under MAD conditions.

The approach reduces statistical bias in learned parameters.

Abstract

Generative models unfairly penalize data belonging to minority classes, suffer from model autophagy disorder (MADness), and learn biased estimates of the underlying distribution parameters. Our theoretical and empirical results show that training generative models with intentionally designed hypernetworks leads to models that 1) are more fair when generating datapoints belonging to minority classes 2) are more stable in a self-consumed (i.e., MAD) setting, and 3) learn parameters that are less statistically biased. To further mitigate unfairness, MADness, and bias, we introduce a regularization term that penalizes discrepancies between a generative model's estimated weights when trained on real data versus its own synthetic data. To facilitate training existing deep generative models within our framework, we offer a scalable implementation of hypernetworks that automatically generates a hypernetwork architecture for any given generative model.