BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Bayesian Inference with Kernels - Dr Arthur Gretton (UCL)
DTSTART:20100920T100000Z
DTEND:20100920T110000Z
UID:TALK26272@talks.cam.ac.uk
CONTACT:Zoubin Ghahramani
DESCRIPTION:An embedding of probability distributions into a reproducing k
 ernel Hilbert space (RKHS) has been introduced: like the characteristic fu
 nction\, this provides a unique representation of a probability distributi
 on in a high dimensional feature space. This representation forms the basi
 s of an inference procedure on graphical models\, where the likelihoods ar
 e represented as RKHS functions. The resulting algorithm is completely non
 parametric: all aspects of the model are represented implicitly\, and lear
 ned from a training sample. Both exact inference on trees and loopy BP on 
 pairwise Markov random fields are demonstrated.  \n\nKernel message passin
 g can be applied to general domains where kernels are defined\, handling c
 hallenging cases such as discrete variables with huge domains\, or very co
 mplex\, non-Gaussian continuous distributions. In experiments\, the approa
 ch outperforms state-of-the-art techniques in a cross-lingual document ret
 rieval task and a camera rotation estimation problem. Finally\, time permi
 tting\, a more general kernelized Bayes' law will be described\, in which 
 a prior distribution embedding is updated to provide a posterior distribut
 ion embedding. This last approach makes weaker assumptions on the underlyi
 ng distributions\, but is somewhat more complex to implement.\n\nJoint wor
 k with Danny Bickson\, Kenji Fukumizu\, Carlos Guestrin\, Yucheng Low\, Le
  Song\n
LOCATION:Engineering Department\, CBL Room 438
END:VEVENT
END:VCALENDAR