BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:A New Geometric Approach to Topic Modeling and Discovery
DTSTART:20131127T161500
DTEND:20131127T171500
DTSTAMP:20260417T143133Z
UID:155d330d6cf1f68c9d21e6b998f84bcfe05785acce69f978e6aadc58
CATEGORIES:Conferences - Seminars
DESCRIPTION:Prof. Prakash Ishwar\, Boston University\nIn this talk I will 
 present a new algorithm for topic discovery based\non the geometry of cros
 s-document word-frequency patterns. The\ngeometric perspective gains signi
 ficance under the so called\nseparability condition that posits the existe
 nce of novel-words that\nare unique to each topic. The algorithm utilizes 
 random projections to\nidentify novel words and associated topics. The key
  insight here is\nthat the maximum and minimum values of cross-document fr
 equency\npatterns projected along any direction are associated with novel\
 nwords.  In contrast to ML and Bayesian approaches that require solving\n
 non-convex optimization problems using approximations or heuristics\,\nthe
  new algorithm is convex\, asymptotically consistent\, and has\nprovable p
 erformance guarantees. While our sample complexity bounds\nfor topic recov
 ery are similar to the state-of-art\, the computational\ncomplexity of our
  scheme scales linearly with the number of documents\nand the number of wo
 rds per document. We present several experiments\non synthetic and realwor
 ld datasets to demonstrate qualitative and\nquantitative merits of our sch
 eme. This talk is based on joint work\nwith Ding\, Rohban\, and Saligrama 
 at Boston University.
LOCATION:INR219 http://plan.epfl.ch/?lang=en&zoom=20&recenter_y=5863811.95
 815&recenter_x=730517.55838&layerNodes=fonds\,batiments\,labels\,informati
 on\,parkings_publics\,arrets_metro&floor=2&q=INR219
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR