Statistics Seminar - 02/12/26

Abstract:

Bayesian machine learning provides flexible modeling and principled uncertainty quantification, yet ensuring trustworthy inference remains challenging in high-dimensional and complex settings. In this talk, I propose a framework for trustworthy Bayesian learning in which model and prior design are guided by frequentist principles to achieve calibrated uncertainty and reliable decision-making. I illustrate this framework through three examples: uniform false discovery rate control for high-dimensional data, a multiscale theoretical framework for validating Bayesian uncertainty, and ongoing work on scalable and interpretable Bayesian semi-supervised learning for large-scale electronic health record data. Together, these examples demonstrate how frequentist validation can serve as a design constraint for trustworthy Bayesian learning.

Note:

Contact Prof. JZ at jzhang@math.msstate.edu for additional information.