BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Bayesian canonical correlation analysis - Seppo Virtanen (Aalto Un
 iversity)
DTSTART:20140228T110000Z
DTEND:20140228T120000Z
UID:TALK51054@talks.cam.ac.uk
CONTACT:Zoubin Ghahramani
DESCRIPTION:Canonical correlation analysis (CCA) is a classical method for
 \nseeking correlations between two multivariate data sets. During the\nlas
 t ten years\, it has received more and more attention in the\nmachine lear
 ning community in the form of novel computational\nformulations and a plet
 hora of applications. \nBayesian treatments of CCA-type\nlatent variable m
 odels have been recently proposed for coping with\noverfitting in small sa
 mple sizes\, as well as for producing\nfactorizations of the data sources 
 into correlated and non-shared\neffects. However\, all of the current impl
 ementations of Bayesian CCA\nand its extensions are computationally ineffi
 cient for high-dimensional \ndata. Furthermore\, they cannot reliably sepa
 rate the correlated effects from non-shared\nones. We propose a new Bayesi
 an CCA variant that is computationally\nefficient and works for high-dimen
 sional data\, while also learning the\nfactorization more accurately. The 
 improvements are gained by\nintroducing a group sparsity assumption and an
  improved variational\napproximation. The method is demonstrated to work w
 ell on\nmulti-label prediction tasks and in analyzing brain correlates of\
 nnaturalistic audio stimulation.
LOCATION:Engineering Department\, CBL Room BE-438
END:VEVENT
END:VCALENDAR