BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Can Bayesian neural networks make confident predictions? - Kathari
 ne Fisher (Massachusetts Institute of Technology)
DTSTART:20250606T091000Z
DTEND:20250606T093000Z
UID:TALK230833@talks.cam.ac.uk
DESCRIPTION:Bayesian neural networks (BNN) promise a principled approach t
 o quantifying uncertainty in overparameterized models. Evaluating this pro
 mise is a challenge because in most practical settings\, BNN predictive di
 stributions can only be accessed through approximate inference. To systema
 tically investigate the calibration of BNN predictive distributions\, we a
 rgue for the use of discrete priors on interior parameters. We demonstrate
  that networks can be reverse engineered to determine which parameter &lsq
 uo\;candidates&rsquo\; should be given prior weight. This approach reveals
  that multimodal distributions in parameter space map to multimodal distri
 butions in prediction space which are often only partially captured by app
 roximate methods. We also find that uncertainty metrics have non-intuitive
  dependence on network dimensions\, including cases where network capacity
  increases but uncertainty decreases. These results raise questions of whe
 ther some approximate methods may perform &lsquo\;better&rsquo\; than the 
 true BNN predictive distribution.\nCo-author: Youssef Marzouk
LOCATION:Seminar Room 1\, Newton Institute
END:VEVENT
END:VCALENDAR