Details of the talk:
- Date: December 7
- Time: 11:00 a.m. - 12:00 p.m.
- Location: Ground-floor lecture hall, Tübingen AI Center (Maria-von-Linden-Str. 6, 72076 Tübingen)
Alexandru Tifreu is a PhD candidate at the Department of Computer Science of ETH Zürich.
Talk title: Best of both worlds: A post-processing framework for in-processing fairness algorithms
Abstract: Post-processing mitigation techniques for group fairness adjust the outputs of a pre-trained model in order to induce fairness. This class of methods exhibits several advantages that make it appealing in practice: post-processing requires no access to the model training pipeline, is agnostic to the pre-trained model family, and has a significantly reduced computation cost compared to in-processing. Despite these benefits, current methods face key challenges that limit their applicability. Unlike in-processing, existing post-processing techniques require that all inference-time samples have known sensitive attributes, they cannot work with continuous group labels and are oftentimes outperformed by in-processing. In this paper, we propose a general framework that transforms an in-processing method with a penalized objective into a post-processing procedure. The resulting method is specifically designed to overcome the shortcomings of prior post-processing approaches. Furthermore, we show theoretically and experimentally on real-world data that the resulting post-processing method can match the fairness-error trade-off of the in-processing counterpart. Finally, we uncover a novel shortcoming of in-processing that leads to poor performance in the low-sample regime and show how our corresponding post-processing technique can mitigate it.