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:ML@CL Seminar Series
SUMMARY:Theoretical guarantees for sampling and inference
in generative models with latent diffusions - Beli
nda Tzen - University of Illinois at Urbana-Champ
aign
DTSTART;TZID=Europe/London:20210629T150000
DTEND;TZID=Europe/London:20210629T160000
UID:TALK161191AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/161191
DESCRIPTION:*Paper:*\n\nTalk is based on "this":http://proceed
ings.mlr.press/v99/tzen19a.html paper.\n\n*Abstrac
t:*\n\nWe introduce and study a class of probabili
stic generative models\, where the latent object i
s a finite-dimensional diffusion process on a fini
te time interval and the observed variable is draw
n conditionally on the terminal point of the diffu
sion. We make the following contributions: We prov
ide a unified viewpoint on both sampling and varia
tional inference in such generative models through
the lens of stochastic control. We quantify the e
xpressiveness of diffusion-based generative models
. Specifically\, we show that one can efficiently
sample from a wide class of terminal target distri
butions by choosing the drift of the latent diffus
ion from the class of multilayer feedforward neura
l nets\, with the accuracy of sampling measured by
the Kullback–Leibler divergence to the target dis
tribution. Finally\, we present and analyze a sche
me for unbiased simulation of generative models wi
th latent diffusions and provide bounds on the var
iance of the resulting estimators. This scheme can
be implemented as a deep generative model with a
random number of layers.\n\n*Website:* https://sch
olar.google.com/citations?user=UgdTN9UAAAAJ&hl=en\
n\nPart of ML@CL Seminar Series in topics relevant
to machine learning and statistics.\n\nMeeting ID
: 966 8293 0826\n\nPasscode: 007922
LOCATION:https://cl-cam-ac-uk.zoom.us/j/96682930826?pwd=SVV
pTFplRVRFeXhmOE1VejFVeTdzdz09
CONTACT:Francisco Vargas
END:VEVENT
END:VCALENDAR