Events

Date: 27 March 2026

Location: Room 516 G.O. Jones Building, 14:30-15:30

Title: Bayesian PINNs and overconfidence

Abstract: Bayesian physics informed neural networks (B-PINNs) merged data with the governing equations of a physical system, to solve differential equations under uncertainty. However, interpretation of uncertainty and overconfidence in B-PINNs can be subtle. Overconfidence can reflect warranted precision, enforced by physical constraints, rather than miscalibration. In this talk we will explore overconfidence in B-PINNs through several physical systems and introduce new information theoretic approaches to characterise overconfidence.