BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Fast yet Simple Natural-Gradient Variational Inference in Complex 
 Models - Emtiyaz Khan\, team leader (equivalent to Full Professor) at the 
 RIKEN center for Advanced Intelligence Project (AIP) in Tokyo
DTSTART:20180716T100000Z
DTEND:20180716T110000Z
UID:TALK108382@talks.cam.ac.uk
CONTACT:Dr R.E. Turner
DESCRIPTION:Approximate Bayesian inference is promising in improving gener
 alization and reliability of deep learning\, but is computationally challe
 nging. Modern variational-inference (VI) methods circumvent the challenge 
 by formulating Bayesian inference as an optimization problem and then solv
 ing it using gradient-based methods. In this talk\, I will argue in favor 
 of natural-gradient approaches which can improve convergence of VI by expl
 oiting the information geometry of the solutions. I will\ndiscuss a fast y
 et simple natural-gradient method obtained by using a duality associated w
 ith exponential-family distributions. I will summarize some of our recent 
 results on Bayesian deep learning\, where natural-gradient methods lead to
  an approach which gives simpler updates than existing VI methods while pe
 rforming comparably to them.\n\nJoint work with Wu Lin (UBC)\, Didrik Niel
 sen (RIKEN)\, Voot Tangkaratt (RIKEN)\, Yarin Gal (UOxford)\, Akash Srivas
 tva (UEdinburgh)\, Zuozhu Liu (SUTD).\n\nBased on:\n\nhttps://emtiyaz.gith
 ub.io/papers/isita2018_preprint.pdf\n\nhttps://arxiv.org/abs/1806.04854\n\
 nhttps://arxiv.org/abs/1703.04265
LOCATION:Engineering Department\, CBL Room BE-438.
END:VEVENT
END:VCALENDAR