BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Frank-Wolfe optimization insights in machine learning - Simon Laco
 ste-Julien (INRIA\, ENS\, Paris)
DTSTART:20120824T100000Z
DTEND:20120824T110000Z
UID:TALK39391@talks.cam.ac.uk
CONTACT:Zoubin Ghahramani
DESCRIPTION:The Frank-Wolfe optimization algorithm (also called\ncondition
 al gradient) is a very simple and intuitive optimization\nalgorithm propos
 ed in the 1950s by Marguerite Frank and Phil Wolfe. It\nwas partly forgott
 en as it became superseded by faster algorithms\, but\nit is making a rece
 nt revival in machine learning\, thanks to its\nability to exploit well th
 e structure of the machine learning\noptimization problems. In this talk\,
  I will mention two recent\nadvances making use of Frank-Wolfe. In the fir
 st part\, I will describe\nhow it can be efficiently applied to large marg
 in learning for\nstructured prediction. I will show how several previous a
 lgorithms\nwere special cases of Frank-Wolfe\, and I will present a new\nb
 lock-coordinate version of Frank-Wolfe which yields a simple\nalgorithm wh
 ich outperforms the state-of-the-art. In the second part\,\nI will describ
 e how the herding algorithm recently proposed by Max\nWelling is actually 
 equivalent to the Frank-Wolfe optimization of a\nquadratic moment discrepa
 ncy. This link enables us to obtain a\nweighted version of herding which c
 onverge faster for the task of\napproximating integrals (obtaining adaptiv
 e quadrature rules). On the\nother hand\, our experiments indicate that he
 rding could still be\nbetter for the task learning\, shedding more light o
 n the properties of\nthe herding algorithm.\n\nThis is joint work with Fra
 ncis Bach\, Martin Jaggi\, Guillaume\nObozinski\, Mark Schmidt and Patrick
  Pletscher.
LOCATION:Engineering Department\, CBL Room BE-438
END:VEVENT
END:VCALENDAR