BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Multi-view Learning of Speech Feature Spaces - Karen Livescu (TTI-
 Chicago)
DTSTART:20090915T150000Z
DTEND:20090915T160000Z
UID:TALK19955@talks.cam.ac.uk
CONTACT:Dr Marcus Tomalin
DESCRIPTION:Many learning tasks (classification\, regression\,\nclustering
 ) can be improved when multiple views of the data are\navailable.  The mea
 ning of "views" may be a natural one like audio vs.\nimages vs. text\, or 
 more abstract like arbitrary subsets of the\nobservation vector.  Multi-vi
 ew learning algorithms\, such as\nco-training\, take advantage of the rela
 tionships between the views.\nIn this work\, we explore two-view learning 
 of feature spaces:  Given\ntwo views of the training data\, we learn a tra
 nsformation of each view\nthat\, in some sense\, best predicts the other v
 iew.  Importantly\, we\ncan then apply the learned transformations even wh
 en only one view\n(e.g. audio) is available at test time.  For this talk\,
  I will focus\non work using canonical correlation analysis (CCA)\, in whi
 ch a linear\nprojection of each view is learned\, such that the two views'
 \nprojections are maximally correlated.  I will describe recent\nexperimen
 ts showing improvements on clustering tasks (speaker\nclustering of audio 
 and/or video and topic clustering of Wikipedia\npages) and on a speaker id
 entification task.  Time permitting\, I will\ndescribe additional ongoing 
 work in speech and language at TTI-C.\n\nJoint work with Kamalika Chaudhur
 i (UCSD)\, Sham Kakade (TTI-C)\,\nKarthik Sridharan (TTI-C)\, and Mark Sto
 ehr (U. Chicago)
LOCATION:LR4\, Engineering Department\, Baker Building
END:VEVENT
END:VCALENDAR