BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Machine Learning Reading Group @ CUED
SUMMARY:Sampling as Optimization - Eric Nalisnick\, Univer
sity of Cambridge
DTSTART;TZID=Europe/London:20190403T134500
DTEND;TZID=Europe/London:20190403T151500
UID:TALK122527AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/122527
DESCRIPTION:Sampling and optimization are often thought of as
alternative methods for model fitting. In this me
eting of the reading group\, we summarize recent r
esults that draw connections between sampling and
optimization. The key result is the work of Jorda
n et al. (1998)\, which shows that the gradient fl
ow of the Kullback-Leibler divergence in the space
of measures follows the Fokker-Planck equation.
This Fokker-Planck equation can then be recast as
running Langevin dynamics in the space of model pa
rameters. With this result established\, we then
discuss implications for discretization schemes\,
viewing SGD as approximate Bayesian inference\, an
d models for which sampling can be faster than opt
imization.
LOCATION:Engineering Department\, CBL Room BE-438
CONTACT:Robert Pinsler
END:VEVENT
END:VCALENDAR