Predictive models are increasingly used in decision-making pipelines, and the statistical information derived from their outputs is often treated as if it preserves classical coverage guarantees. This talk examines how such guarantees can fail when estimates are constructed from synthetic or inferred data rather than directly observed samples. In these settings, conflating prediction errors with natural data variation can produce unreliable uncertainty estimates. Through an example involving inferred sensitive attributes, we show how standard confidence interval construction fails to achieve nominal coverage and how an adjusted procedure restores valid bounds. Our discussion explores when classical guarantees remain applicable and when they must be reconsidered in the presence of such inferred data.
Bio:
Blossom Metevier studies how machine learning systems can be designed to act responsibly in dynamic environments. Her research focuses on improving the reliability of sequential learning methods, including those used in large language models. She is also interested in designing general methods with provable safety and fairness guarantees for high-stakes applications. Metevier completed her Ph.D. in computer science at the University of Massachusetts Amherst. She has collaborated with industry through research internships at IBM, Microsoft Research, and Meta, where she worked on challenges of fairness in machine learning systems. Her research has been published in venues including FAccT, NeurIPS, and ICLR.
In-person attendance is open to Princeton University faculty, staff and students.
This talk will not be livestreamed or recorded.
If you need an accommodation for a disability please contact Jean Butcher at butcher@princeton.edu.
Sponsorship of an event does not constitute institutional endorsement of external speakers or views presented.