Title: Reframing Algorithmic Fairness: A paradigm for fair, accurate, and flexible model development

Live stream: https://uwmadison.zoom.us/j/95853702992?pwd=aX58aqk3yOoAsmq9bpeDPu4b3xSpQp.1 

Abstract: The mitigation of unfairness in real-world machine learning systems poses a variety of challenges. Unfair outcomes are often only discovered post-deployment by diverse users, the tensions between standard notions of fairness and model accuracy present challenges for model developers, and industry-scale models are deployed in different downstream contexts with competing fairness concerns. This talk, grounded in the algorithmic foundations of responsible computing, will present a collection of techniques to resolve these challenges. Drawing on a number of results, it will consider fairness interventions throughout the development pipeline, building a paradigm where fairness and accuracy are aligned goals and where models can be efficiently updated or post-processed to fit different downstream contexts. Along the way, we will resolve an open problem from the learning theory literature and explore mechanisms for auditing and ameliorating models. 

Bio: Ira Globus-Harris is a PhD student in computer and information sciences at the University of Pennsylvania, advised by Michael Kearns and Aaron Roth. Their work broadly looks at mechanisms to resolve harms incurred by AI-driven decision-making, focusing on using algorithmic techniques which are scalable for real-world use, holistically consider an algorithm in its broader context, and which flexibly incorporate human input. They were a 2024 AWS AI ASSET Fellow and a 2023 EECS Rising Star.