BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Safe Learning: How to Modify Bayesian Inference when All Models ar
 e Wrong - Peter Grunwald\, Centrum voor Wiskunde en Informatica\, Amsterda
 m
DTSTART:20111020T150000Z
DTEND:20111020T160000Z
UID:TALK32922@talks.cam.ac.uk
CONTACT:Richard Samworth
DESCRIPTION:Standard Bayesian inference can behave suboptimally if the mod
 el under\nconsideration is wrong:\nin some simple settings\, the posterior
  may fail to concentrate even in the\nlimit of infinite sample size.\nWe i
 ntroduce a test that can tell from the data whether we are in such a\nsitu
 ation. If we are\, we can adjust\nthe learning rate (equivalently: make th
 e prior lighter-tailed) in a\ndata-dependent way. The\nresulting "safe" es
 timator continues to achieve good rates with wrong\nmodels. When applied t
 o classification\, the approach achieves optimal\nrates under Tsybakov's c
 onditions\, thereby creating a bridge between\nBayes/MDL and statistical l
 earning-style inference.\n
LOCATION:MR12\, CMS\, Wilberforce Road\, Cambridge\, CB3 0WB
END:VEVENT
END:VCALENDAR