BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Hybrids of generative and discriminative models - Julia Lasserre (
 University of Cambridge/Max Planck Institute)
DTSTART:20090630T140000Z
DTEND:20090630T142000Z
UID:TALK18945@talks.cam.ac.uk
CONTACT:Dr Fabien Petitcolas
DESCRIPTION:*Abstract*: When labelled training data is plentiful\, discrim
 inative techniques are widely used since they give excellent classificatio
 n results. However\, hand-labelling of data can get expensive\, and there 
 is considerable interest in semi-supervised techniques based on generative
  models. Although the generalisation performance of generative models can 
 often be improved by `training them discriminatively'\, they can then no l
 onger make use of unlabelled data. In an attempt to exploit the benefits o
 f both generative and discriminative approaches\, methods have been propos
 ed which interpolate between these two extremes by taking a convex combina
 tion of the generative and discriminative objective functions. In this art
 icle\, we consider that there is only one correct way to train a given mod
 el\, and that a `discriminatively trained' generative model is fundamental
 ly a new model. From this viewpoint\, generative and discriminative models
  correspond to specific choices for the prior over parameters\, which open
 s the door to principled ways of interpolating between generative and disc
 riminative extremes through alternative choices of prior. We illustrate th
 is framework on semi- supervised learning.
LOCATION:Large public lecture room\, Microsoft Research\, Roger Needham Bu
 ilding\, 7 J J Thomson Avenue\, Cambridge CB3 0FB
END:VEVENT
END:VCALENDAR