BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Machine Learning @ CUED
SUMMARY:Frank-Wolfe optimization insights in machine learn
 ing - Simon Lacoste-Julien (INRIA\, ENS\, Paris)
DTSTART;TZID=Europe/London:20120824T110000
DTEND;TZID=Europe/London:20120824T120000
UID:TALK39391AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/39391
DESCRIPTION:The Frank-Wolfe optimization algorithm (also calle
 d\nconditional gradient) is a very simple and intu
 itive optimization\nalgorithm proposed in the 1950
 s by Marguerite Frank and Phil Wolfe. It\nwas part
 ly forgotten as it became superseded by faster alg
 orithms\, but\nit is making a recent revival in ma
 chine learning\, thanks to its\nability to exploit
  well the structure of the machine learning\noptim
 ization problems. In this talk\, I will mention tw
 o recent\nadvances making use of Frank-Wolfe. In t
 he first part\, I will describe\nhow it can be eff
 iciently applied to large margin learning for\nstr
 uctured prediction. I will show how several previo
 us algorithms\nwere special cases of Frank-Wolfe\,
  and I will present a new\nblock-coordinate versio
 n of Frank-Wolfe which yields a simple\nalgorithm 
 which outperforms the state-of-the-art. In the sec
 ond part\,\nI will describe how the herding algori
 thm recently proposed by Max\nWelling is actually 
 equivalent to the Frank-Wolfe optimization of a\nq
 uadratic moment discrepancy. This link enables us 
 to obtain a\nweighted version of herding which con
 verge faster for the task of\napproximating integr
 als (obtaining adaptive quadrature rules). On the\
 nother hand\, our experiments indicate that herdin
 g could still be\nbetter for the task learning\, s
 hedding more light on the properties of\nthe herdi
 ng algorithm.\n\nThis is joint work with Francis B
 ach\, Martin Jaggi\, Guillaume\nObozinski\, Mark S
 chmidt and Patrick Pletscher.
LOCATION:Engineering Department\, CBL Room BE-438
CONTACT:Zoubin Ghahramani
END:VEVENT
END:VCALENDAR