Quantifying Representation Fairness in Diffusion Models for Solving Inverse Problems

Abstract

Generative models have been used to solve inverse problems in various applications. However, quantifying representation bias in these models to analyze their capabilities in reconstruction tasks is challenging, and limited research has been performed in this direction. In this work, a metric to quantify representation bias in a generative model for inverse tasks is proposed. The proposed metric focuses on known distinct subgroups present in the dataset. A diffusion model is used to perform conditional sampling for inverse tasks and the metric is used to evaluate how biased the diffusion model is in achieving Conditional Proportional Representation. The metric can be extended to the unconditional case and other generative models.

Date of Award	14 Sept 2023
Original language	English
Awarding Institution	Eindhoven University of Technology
Supervisor	Ruud J.G. van Sloun (Supervisor 1) & Hans van Gorp (Supervisor 2)