BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:MLE-Struct: Bethe Learning of Graphical Models - Kui Tang\, Columb
 ia University
DTSTART:20160513T100000Z
DTEND:20160513T110000Z
UID:TALK65214@talks.cam.ac.uk
CONTACT:Adrian Weller
DESCRIPTION:Many machine learning tasks require fitting probabilistic mode
 ls over\nstructured objects\, such as pixel grids\, matchings\, and graph 
 edges.\nMaximum likelihood estimation (MLE) for such domains is challengin
 g\ndue to the intractability of computing partition functions. One can\nre
 sort to approximate marginal inference in conjunction with gradient\ndesce
 nt\, but such algorithms require careful tuning. Alternatively\, in\nframe
 works such as the structured support vector machine (SVM-Struct)\,\ndiscri
 minative functions are learned by iteratively applying efficient\nmaximum 
 a posteriori (MAP) decoders. We introduce MLE-Struct\, a method\nfor learn
 ing discrete exponential family models using the Bethe\napproximation to t
 he partition function. Remarkably\, this problem can\nalso be reduced to i
 terative (MAP) decoding. This connection emerges\nby combining the Bethe a
 pproximation with the Frank-Wolfe (FW)\nalgorithm on a convex dual objecti
 ve\, which circumvents the\nintractable partition function. Our method can
  learn both generative\nand conditional models and is substantially faster
  and easier to\nimplement than existing MLE approaches while relying only 
 on the same\nblack-box interface to MAP decoding as SVM-Struct. We perform
 \ncompetitively on problems in denoising\, segmentation\, matching\, and\n
 new datasets of roommate assignments and news and financial time\nseries.
LOCATION:Engineering Department\, CBL Room BE-438
END:VEVENT
END:VCALENDAR