Event details from: Business Analytics Seminar: Florian van Leeuwen (Utrecht University)

Date
9 November 2026
Time
12:00 – 13:00

Abstract:

Prediction with small samples relies heavily on prior information. Prior-Data Fitted Networks (PFNs), a recent class of tabular foundation models, offer a new route: rather than requiring an explicit prior, they learn one implicitly by pretraining on many synthetic datasets. But what does such an implicit prior actually contain, and how strong is it? In this talk, we first compare PFNs to parametric Bayesian models. We show that a PFN only sees a finite sample of tasks from its prior and can therefore fail when too few similar tasks are sampled, a failure mode that does not exist for parametric priors. Yet, PFNs can in some cases also generalize to tasks their pretraining did not cover. We then benchmark PFNs against a correctly specified regression model in small sample settings and ask how informative its priors must be to match PFN performance.

General information:

Attendance to this seminar is possible by invitation only. Please send an e-mail to secbs-abs@uva.nl if your are interested in attending this seminar.