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:Natural Conjugate Gradient Learning for Fixed-Form
Variational Bayes - Dr Antti Honkela (Aalto Unive
rsity School of Science and Technology\, Finland)
DTSTART;TZID=Europe/London:20100616T110000
DTEND;TZID=Europe/London:20100616T120000
UID:TALK24175AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/24175
DESCRIPTION:Variational Bayesian (VB) methods are typically on
ly applied to models in the conjugate-exponential
family using the variational Bayesian expectation
maximisation (VB EM) algorithm or one of its varia
nts. Here I present an efficient algorithm for app
lying VB to more general models. The method is bas
ed on specifying the functional form of the approx
imation\, such as multivariate Gaussian. The para
meters of the approximation are optimised using a
natural conjugate gradient algorithm that utilises
the Riemannian geometry of the space of the appro
ximations. This leads to a very efficient algorith
m for suitably structured approximations. It is sh
own empirically that the proposed method is compar
able or superior in efficiency to the VB EM in a c
ase where both are applicable. The algorithm is al
so applied to learning a nonlinear state-space mod
el and a nonlinear factor analysis model for which
the VB EM is not applicable. For these models\, t
he proposed algorithm outperforms alternative grad
ient-based methods by a significant margin.\n\nThi
s is joint work with Tapani Raiko\, Mikael Kuusela
\, Matti Tornio and Juha Karhunen.
LOCATION:Engineering Department\, CBL Room 438
CONTACT:Zoubin Ghahramani
END:VEVENT
END:VCALENDAR