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:The Block Diagonal Infinite Hidden Markov Model -
Tom Stepleton (CMU)
DTSTART;TZID=Europe/London:20090121T130000
DTEND;TZID=Europe/London:20090121T140000
UID:TALK16426AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/16426
DESCRIPTION:The Infinite Hidden Markov Model (IHMM) is a nonpa
rametric Bayesian discrete time series model. By m
aintaining a posterior distribution over transitio
n and emission models for a countably infinite num
ber of hidden states\, the IHMM flexibly accommoda
tes data for which the "right" number of such stat
es is not known.\n\nThe Block Diagonal Infinite Hi
dden Markov Model (BD-IHMM) extends the IHMM to ta
rget circumstances where the data alternate over t
ime among a collection of distinct behavioral regi
mes\, or "sub-behaviors". In ordinary HMMs\, alter
nation like this arises from transition matrices w
ith nearly block-diagonal structure\, and the BD-I
HMM's prior induces similar forms in inferred dyna
mics\, here too with flexibility in the number of
blocks. By inferring a hidden state sequence for t
he data\, the model also associates all parts of t
he data with a sub-behavior\, a sequence-clusterin
g capability that may also have useful analogs in
settings besides time series.\n\nThis talk will pr
esent the BD-IHMM\, describe inference techniques
for the model\, show recent results (these involve
the Nintendo Wii and '90s USA alt-rock sensation
Collective Soul)\, and discuss extensions and new
directions under investigation now.
LOCATION:Engineering Department\, CBL Room 438
CONTACT:Zoubin Ghahramani
END:VEVENT
END:VCALENDAR