BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Two Approximate Sampling Methods for Bayesian Deep Learning - Wesl
 ey Maddox (New York University)
DTSTART:20190823T100000Z
DTEND:20190823T110000Z
UID:TALK129037@talks.cam.ac.uk
CONTACT:Adrià Garriga Alonso
DESCRIPTION:Deep neural networks have become popular for many tasks\, espe
 cially object classification\, in computer vision and machine learning. Ho
 wever\, these classes of models are known to be have poor uncertainty repr
 esentations – e.g. they do not know what they do not know. To address th
 is challenge\, we propose two Bayesian approaches to approximate the poste
 rior distribution of the models' parameters. The first\, termed stochastic
  weight averaging Gaussian (SWAG)\, fits a Gaussian approximation around t
 he iterates of the stochastic gradient descent trajectory from standard tr
 aining of DNNs. The second\, subspace inference\, instead reduces the high
  dimensionality of DNNs to very low dimensions\, before performing Bayesia
 n model averaging in that low dimensional subspace. Both methods draw on e
 xisting theory and are demonstrated to have strong empirical results on bo
 th regression and classification\, scaling to even ImageNet-sized datasets
 .
LOCATION:Engineering Department\, CBL Room BE-438.
END:VEVENT
END:VCALENDAR