BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:VIKING: Deep variational inference with stochastic projections - S
 amuel Fadel (Technical University of Denmark)
DTSTART:20251125T130000Z
DTEND:20251125T140000Z
UID:TALK241198@talks.cam.ac.uk
CONTACT:Mateja Jamnik
DESCRIPTION:Variational mean field approximations tend to struggle with\nc
 ontemporary overparameterised deep neural networks. Where a Bayesian\ntrea
 tment is usually associated with high-quality predictions and\nuncertainti
 es\, the practical reality has been the opposite\, with\nunstable training
 \, poor predictive power\, and subpar\ncalibration. Building upon recent w
 ork on reparameterisations of\nneural networks\, we propose a simple varia
 tional family that considers\ntwo independent linear subspaces of the para
 meter space. These\nrepresent functional changes inside and outside the su
 pport of\ntraining data. This allows us to build a fully-correlated approx
 imate\nposterior reflecting the overparameterisation that tunes\neasy-to-i
 nterpret hyperparameters. We develop scalable numerical\nroutines that max
 imize the associated evidence lower bound (ELBO) and\nsample from the appr
 oximate posterior. Our results show that\napproximate Bayesian inference a
 pplied to deep neural networks is far\nfrom a lost cause when constructing
  inference mechanisms that reflect\nthe geometry of reparametrisations.\n\
 n*Bio:* Samuel is a postdoc at Søren Hauberg's group at the Technical\nUn
 iversity of Denmark and an affiliated researcher at the Pioneer\nCentre fo
 r AI. Samuel received his PhD in Computer Science from the\nUniversity of 
 Campinas\, Brazil and an MSc in Applied Mathematics and\nComputer Science 
 from the University of São Paulo\, Brazil. His\nresearch interests includ
 e representation learning and uncertainty\nquantification\, both with a ge
 ometric perspective.\n\n*This talk is co-hosted with the Machine Learning 
 Theory Group in the Maths Department and the Informed-AI Hub.*
LOCATION:Computer Laboratory\, William Gates Building\, Room FW11 (note ch
 ange of room)
END:VEVENT
END:VCALENDAR