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:Microsoft Research Machine Learning and Perception
Seminars
SUMMARY:Bayesian Inference with Kernels - Arthur Gretton\,
UCL
DTSTART;TZID=Europe/London:20101110T140000
DTEND;TZID=Europe/London:20101110T150000
UID:TALK26857AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/26857
DESCRIPTION:An embedding of probability distributions into a r
eproducing kernel Hilbert space (RKHS) has been in
troduced: like the characteristic function\, this
provides a unique representation of a probability
distribution in a high dimensional feature space.
This representation forms the basis of an inferenc
e procedure on graphical models\, where the likeli
hoods are represented as RKHS functions. The resul
ting algorithm is completely nonparametric: all as
pects of the model are represented implicitly\, an
d learned from a training sample. Both exact infer
ence on trees and loopy BP on pairwise Markov rand
om fields are demonstrated. \n\nKernel message pa
ssing can be applied to general domains where kern
els are defined\, handling challenging cases such
as discrete variables with huge domains\, or very
complex\, non-Gaussian continuous distributions. I
n experiments\, the approach outperforms state-of-
the-art techniques in a cross-lingual document ret
rieval task and in the prediction of depth from 2-
D images. Finally\, time permitting\, a more gener
al kernelized Bayes` law will be described\, in wh
ich a prior distribution embedding is updated to p
rovide a posterior distribution embedding. This la
st approach makes weaker assumptions on the underl
ying distributions\, but is somewhat more complex
to implement.\n\nJoint work with Danny Bickson\, K
enji Fukumizu\, Carlos Guestrin\, Yucheng Low\, Le
Song\n
LOCATION:Small public lecture room\, Microsoft Research Ltd
\, 7 J J Thomson Avenue (Off Madingley Road)\, Cam
bridge
CONTACT:Dr. Pushmeet Kohli
END:VEVENT
END:VCALENDAR