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:Probabilistic Systems\, Information\, and Inferenc
 e Group Seminars
SUMMARY:Bayesians turn to experts for advice! - Dr Steven 
 de Rooij\, Statistical Laboratory\, University of 
 Cambridge
DTSTART;TZID=Europe/London:20081112T141500
DTEND;TZID=Europe/London:20081112T151500
UID:TALK14588AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/14588
DESCRIPTION:In Bayesian model selection and model averaging\, 
 inference is normally based on a posterior distrib
 ution on the models\, usually interpreted as a mea
 sure of how likely we consider each of the models 
 to be "true"\, or at least in some sense close to 
 true\, given the observations.\n     Rather than w
 ith truth\, I will be concerned with the more prac
 tical goal of finding a "useful" model\, in the se
 nse that it predicts future outcomes of the underl
 ying process well. As it turns out\, the most usef
 ul model may well vary depending on the number of 
 available observations! For instance\, given ten s
 amples from some continuous density\, a seven-bin 
 histogram model is more useful than a 1\,000-bin m
 odel\, even though the latter is arguably closer t
 o being "true".\n     As it turns out\, methods fo
 r tracking transient performance of prediction str
 ategies have already been developed in the learnin
 g theory literature under the heading "prediction 
 with expert advice". I will illustrate how these m
 ethods can improve model selection performace usin
 g results from computer simulations on density est
 imation problems.\n
LOCATION:LR12\, Engineering\, Department of
CONTACT:Rachel Fogg
END:VEVENT
END:VCALENDAR