BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Contrastive Learning Using Spectral Methods - James Zou (Harvard)
DTSTART:20131106T133000Z
DTEND:20131106T143000Z
UID:TALK48749@talks.cam.ac.uk
CONTACT:Zoubin Ghahramani
DESCRIPTION:In many natural settings\, the analysis goal is not to charact
 erize a single data set in isolation\, but rather to understand the differ
 ence between one set of observations and another. For example\, given a ba
 ckground corpus of news articles together with writings of a particular au
 thor\, one may want a topic model that explains word patterns and themes s
 pecific to the author. Another example comes from genomics\, in which biol
 ogical signals may be collected from different regions of a genome\, and o
 ne wants a model that captures the differential statistics observed in the
 se regions. This paper formalizes this notion of contrastive learning for 
 mixture models\, and develops spectral algorithms for inferring mixture co
 mponents specific to a foreground data set when contrasted with a backgrou
 nd data set. The method builds on recent moment-based estimators and tenso
 r decompositions for latent variable models\, and has the intuitive featur
 e of using background data statistics to appropriately modify moments esti
 mated from foreground data. A key advantage of the method is that the back
 ground data need only be coarsely modeled\, which is important when the ba
 ckground is too complex\, noisy\, or not of interest. The method is demons
 trated on applications in contrastive topic modeling and genomic sequence 
 analysis.\n\nThis is joint work with Daniel Hsu\, David Parkes and Ryan Ad
 ams. If there's time\, I will also briefly describe a statistical approach
  to identify epigenomic variations in human populations that are associate
 d with diseases.  
LOCATION:Engineering Department\, CBL Room BE-438
END:VEVENT
END:VCALENDAR