Abstract

Modern neural networks often produce overconfident predictions, particularly when facing out-of-distribution data or under computational constraints. In this talk, I will explore recent advances in uncertainty-aware modelling that tackle this issue from two complementary angles. First, I present how periodic activation functions can induce stationary priors in Bayesian neural networks, drawing a direct connection to Gaussian process models and enabling models that better “know what they don’t know”. Second, I discuss a lightweight, post-hoc method to correct overconfidence in dynamic neural networks—models that adapt their computational effort based on input complexity—by probabilistically modelling their final layers to account for uncertainty. Together, these contributions provide principled tools to improve calibration, reliability, and resource-aware decision-making in modern deep learning systems.