BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Learning and Representing: a Jointly Optimal Approach - Xinhua Zha
 ng\, University of Alberta
DTSTART:20120410T090000Z
DTEND:20120410T100000Z
UID:TALK37275@talks.cam.ac.uk
CONTACT:Microsoft Research Cambridge Talks Admins
DESCRIPTION:Data representations\, and transformations of data representat
 ions\, are fundamental to the design of effective machine learning methods
 . Previous research has established that expressing complex data objects s
 uch as documents or images as feature vectors reveals important structure\
 , both in collections of data and in individual data items. For any partic
 ular application\, however\, one does not often know which features to use
 . Automatically discovering useful features as part of training has theref
 ore been a long standing goal of machine learning research. Unfortunately 
 the resulting training problem -- simultaneously learning a feature repres
 entation and a data reconstruction model -- has been deemed intractable in
  general.\n\nI will demonstrate in this talk a fundamental reformulation o
 f representation learning that enables the training problem to be solved b
 oth globally and efficiently\, even when a feature representation and data
  reconstruction model are learned simultaneously. This is a major advance 
 over the current state-of-the-art\, where globally optimal representations
  cannot be guaranteed in general. I will show that this work has led to si
 gnificant improvements in both generalization accuracy and training time o
 ver state-of-the-art methods. \n\nIn addition\, we have taken a step towar
 ds scaling up the method to large datasets. Arguably\, optimization underl
 ies almost all branches in machine learning and a major difficulty is the 
 nonsmooth objectives. I will show that a novel framework based on smoothin
 g\, pioneered by Nesterov\, provably improves the convergence rates. In pa
 rticular\, I will show applications of our idea to optimizing multivariate
  performance measures and structure prediction. Empirical evaluation on so
 me of the largest publicly available datasets from a variety of domains sh
 ows that our method learns the optimal model significantly faster than the
  state-of-the-art solvers. A broader application of the smoothing techniqu
 e includes graphical model inference and compressive sensing.
LOCATION:Large lecture theatre\, Microsoft Research Ltd\, 7 J J Thomson Av
 enue (Off Madingley Road)\, Cambridge
END:VEVENT
END:VCALENDAR