BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Deep Gaussian Processes - Prof. Neil Lawrence (Sheffield)
DTSTART:20130501T140000Z
DTEND:20130501T150000Z
UID:TALK44984@talks.cam.ac.uk
CONTACT:Zoubin Ghahramani
DESCRIPTION:[NOTE: rescheduled\, now 15:00-16:00] \n\nIn this talk we will
  introduce deep Gaussian process (GP) models. Deep\nGPs are a deep belief 
 network based on Gaussian process mappings. The\ndata is modeled as the ou
 tput of a multivariate GP. The inputs to that\nGaussian process are then g
 overned by another GP. A single layer model\nis equivalent to a standard G
 P or the GP latent variable model\n(GPLVM). We perform inference in the mo
 del by approximate variational\nmarginalization. This results in a strict 
 lower bound on the marginal\nlikelihood of the model which we use for mode
 l selection (number of\nlayers and nodes per layer). Deep belief networks 
 are typically\napplied to relatively large data sets using stochastic grad
 ient\ndescent for optimization. Our fully Bayesian treatment allows for th
 e\napplication of deep models even when data is scarce. Model selection\nb
 y our variational bound shows that a five layer hierarchy is\njustified ev
 en when modelling a digit data set containing only 150\nexamples. In the s
 eminar we will briefly review dimensionality reduction\nvia Gaussian proce
 sses\, before showing how this framework can be\nextended to build deep mo
 dels.
LOCATION:Engineering Department\, CBL Room BE-438
END:VEVENT
END:VCALENDAR