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 @ CUED
SUMMARY:Annealing Between Distributions by Averaging Momen
ts - Chris Maddison (U Toronto)
DTSTART;TZID=Europe/London:20130802T110000
DTEND;TZID=Europe/London:20130802T120000
UID:TALK46498AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/46498
DESCRIPTION:Many powerful Monte Carlo techniques for estimatin
g partition functions\, such as annealed importanc
e sampling (AIS)\, are based on sampling from a se
quence of intermediate distributions\, which inter
polate between a tractable initial distribution an
d an intractable target distribution. The near-uni
versal practice is to use geometric averages of th
e initial and target distributions\, but alternati
ve paths can perform substantially better. We pres
ent a novel sequence of intermediate distributions
for exponential families: averaging the moments o
f the initial and target distributions. We derive
an asymptotically optimal piecewise linear schedul
e for the moments path and show that it performs a
t least as well as geometric averages with a line
ar schedule. Moment averaging performs well empiri
cally at estimating partition functions of restric
ted Boltzmann machines (RBMs)\, which form the bui
lding blocks of many deep learning models.
LOCATION:Engineering Department\, CBL Room BE-438
CONTACT:Zoubin Ghahramani
END:VEVENT
END:VCALENDAR