Modern AI systems offer impressive predictive abilities, but often struggle to provide reliable uncertainty estimates. This challenge arises when the distribution of new inputs differs from that seen during training, for example when broadly pretrained models are adapted to specialized downstream tasks. Classical Bayesian models provide a principled framework for uncertainty quantification, but do not directly address this setting: a new covariate changes where the posterior predictive distribution is evaluated, but not the posterior over the predictive mechanism itself. At the same time, methods designed to account for distribution shifts typically respond to distributional change itself, even when predictive ability is unaffected. In this talk, Yuli Slavutsky will present covariate-dependent Bayesian models, in which the prior depends on both the training covariates and the queried covariate, but not on the outcomes. Constructing the prior from the likelihoods of the predictive models ties this adaptation directly to predictive ability, rather than to distributional change alone. The seminar will discuss two instantiations: covariate-dependent priors on prediction-layer parameters, which yield uncertainty estimates that adapt to shifts affecting predictive performance, and random selection models over fixed predictors whose posterior probabilities yield input-adaptive averaging weights.

https://upenn.zoom.us/j/97645937545

Meeting ID: 976 4593 7545