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 Cambridge\, public talks
SUMMARY:Some recent developments in approximate inference:
learning and control - David Barber\, University
College London
DTSTART;TZID=Europe/London:20110706T140000
DTEND;TZID=Europe/London:20110706T150000
UID:TALK31717AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/31717
DESCRIPTION:I'll discuss two pieces of work on inference in pr
obabilistic models:\n\nThe first concerns a very g
eneral class of Bayesian Linear Models that are wi
dely used in statistics and machine learning. A gr
eat deal of research has been carried out on devel
oping approximate inference techniques for this im
portant class. In particular I'll discuss methods
that bound the model likelihood\, which is of inte
rest in parameter learning. The well-known `local
' variational methods lower bound the marginal li
kelihood. Despite their popularity over the last d
ecade\, I'll discuss our recent result that shows
that local methods result in weaker bounds than al
ternative `mean-field' variational methods. In add
ition\, I'll discuss the perhaps surprising result
that the mean-field bound is concave and discuss
how one may make computationally efficient approxi
mations in large-scale models with many thousands
of variables.\n\nLagrange Duality is being increas
ingly exploited across machine learning but to dat
e has received comparatively little attention in p
lanning and control. For the second part of the ta
lk I'll discuss an application of Lagrange Duality
in learning Markov Decision Process policies. In
particular\, I'll discuss the computationally diff
icult finite-horizon time-independent policy case\
, and demonstrate how our method exhibits substant
ially improved performance compared to policy grad
ients and more recent `EM' style procedures.
LOCATION:Small lecture theatre\, Microsoft Research Ltd\, 7
J J Thomson Avenue (Off Madingley Road)\, Cambrid
ge
CONTACT:Microsoft Research Cambridge Talks Admins
END:VEVENT
END:VCALENDAR