Statistical Challenges in Uncertainty Quantification for Supervised Learning
Rafael Izbicki (UFSCar)
Abstract: Uncertainty quantification is becoming increasingly important in supervised learning. In this talk, I will discuss several statistical challenges that arise when we aim not only to predict accurately, but also to evaluate predictive uncertainty.
I will focus on three problems: conditional density estimation, goodness-of-fit for predictive distributions, and prediction sets with finite-sample guarantees. I will also discuss recent developments, including tabular foundation models and conformal prediction, with examples from scientific applications.