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:Signal Processing and Communications Lab Seminars
SUMMARY:Sequential Monte Carlo for graphical models: Graph
decompositions and Divide-and-Conquer SMC - Dr Fr
edrik Lindsten\, CUED
DTSTART;TZID=Europe/London:20141023T140000
DTEND;TZID=Europe/London:20141023T150000
UID:TALK55627AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/55627
DESCRIPTION:Probabilistic graphical models (PGMs) are widely u
sed to represent and to reason about underlying st
ructure in high-dimensional probability distributi
ons. We develop a framework for using sequential M
onte Carlo (SMC) methods for inference and learnin
g in general PGMs. Structural information from the
PGM is used to find a collection of graph decompo
sitions\, which are then used as the basis for an
SMC sampler.\n\nIn the first part of the talk we c
onsider sequential decompositions\, which results
in that standard SMC techniques can be used. In th
e second part\, we consider instead an auxiliary t
ree decomposition. Based on this we develop a new
class of SMC samplers\, Divide-and-Conquer SMC\, i
n which we maintain multiple independent populatio
ns of weighted particles. These particle populatio
ns are propagated\, merged\, and resampled as the
method progresses up the tree. We will see how thi
s method naturally extends the standard chain-base
d SMC framework to a method that naturally runs on
trees. We illustrate empirically that these appro
aches can outperform standard methods in terms of
estimation accuracy. They also open up novel paral
lel implementation options and the possibility of
concentrating the computational effort on the most
challenging parts of the problem at hand.
LOCATION:LR5\, Cambridge University Engineering Department
CONTACT:Prof. Ramji Venkataramanan
END:VEVENT
END:VCALENDAR