BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Overconfidence in Neural Networks: Fixing Uncertainty with Structu
 red Priors and Post-hoc Calibration - Arno Solin (Aalto University)
DTSTART:20250605T130000Z
DTEND:20250605T140000Z
UID:TALK230845@talks.cam.ac.uk
DESCRIPTION:Modern neural networks often produce overconfident predictions
 \, particularly when facing out-of-distribution data or under computationa
 l constraints. In this talk\, I will explore recent advances in uncertaint
 y-aware modelling that tackle this issue from two complementary angles. Fi
 rst\, I present how periodic activation functions can induce stationary pr
 iors in Bayesian neural networks\, drawing a direct connection to Gaussian
  process models and enabling models that better "know what they don&rsquo\
 ;t know". Second\, I discuss a lightweight\, post-hoc method to correct ov
 erconfidence in dynamic neural networks&mdash\;models that adapt their com
 putational effort based on input complexity&mdash\;by probabilistically mo
 delling their final layers to account for uncertainty. Together\, these co
 ntributions provide principled tools to improve calibration\, reliability\
 , and resource-aware decision-making in modern deep learning systems.
LOCATION:Seminar Room 1\, Newton Institute
END:VEVENT
END:VCALENDAR