Last month we welcomed Charles Jones to present his latest work at the Statistical Methods for Health Equity webinar series.
Charles Jones is a fourth-year PhD student at Imperial College London, advised by Professor Ben Glocker. Charles' research is at the intersection of fairness and causality in machine learning for medical imaging. He is interested in how causality may be used as a unifying language to understand (and begin to solve) core problems in machine learning, such as fairness, robustness, and distribution shift.
In this talk, Charles covered his most recent preprint, "Rethinking Fair Representation Learning for Performance-Sensitive Tasks", which investigated the validity of prominent fairness methods when applied in settings such as medical imaging. The talk provided an organizing perspective on the fairness literature and demonstrated how to use causal reasoning to define and formalize different sources of dataset bias. Using this understanding, he presented fundamental limitations on fair representation learning when evaluation data was drawn from the same distribution as training data. He further provided two hypotheses for the potential validity of these methods under distribution shift and presented experimental evidence in their favor. The results explained apparent contradictions in the existing literature and revealed how rarely considered causal and statistical aspects of the underlying data affected the validity of fair representation learning.
Don't want to miss out on any future Statistical Methods theme activities? Stay up to date by joining the #theme-statistical-methods channel on our Slack workspace.
Comentarios