BEGIN:VCALENDAR
VERSION:2.0
PRODID:ILLC Website
X-WR-TIMEZONE:Europe/Amsterdam
BEGIN:VTIMEZONE
TZID:Europe/Amsterdam
X-LIC-LOCATION:Europe/Amsterdam
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700329T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701025T030000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
UID:/NewsandEvents/Archives/2003/newsitem/440/26-M
 ay-2003-MDL-and-classification-revisited-Troy-Lee
DTSTAMP:20030521T000000
SUMMARY:MDL and classification, revisited, Troy Le
 e
ATTENDEE;ROLE=Speaker:Troy Lee
DTSTART;TZID=Europe/Amsterdam:20030526T160000
DTEND;TZID=Europe/Amsterdam:20030526T000000
LOCATION:CWI portacabins (Kruislaan 413c), downsta
 irs seminar room (C001), Amsterdam
DESCRIPTION:Abstract:   The Minimum Description Le
 ngth (MDL) Principle is a powerful method for mode
 l selection. The theoretical development of MDL ha
 s mostly centered around probabilistic modeling. Y
 et practical applications of MDL often involve mod
 els which are best viewed as predictors rather tha
 n probability distributions. The standard example 
 is classification, one of the most popular applica
 tions of MDL ever since its inception. MDL has bee
 n applied to such non-probabilistic models in vari
 ous ways. We review these approaches and show that
 , contrary to what is often thought, they can exhi
 bit some rather problematic behaviour: on the theo
 retical side, it is not known whether the resultin
 g procedures are consistent (none of the existing 
 proof techniques can be applied). On the practical
  side, the methods can behave quite unreasonably f
 or small data samples. We analyze the reasons for 
 this undesirable behaviour and propose a radical, 
 general and surprising solution to the problem.
X-ALT-DESC;FMTTYPE=text/html:\n      <p>\n        
 Abstract: <br />\nThe Minimum Description Length (
 MDL) Principle is a powerful\nmethod for model sel
 ection. The theoretical development of MDL\nhas mo
 stly centered around probabilistic modeling. Yet p
 ractical\napplications of MDL often involve models
  which are best viewed\nas predictors rather than 
 probability distributions. The standard\nexample i
 s classification, one of the most popular applicat
 ions of\nMDL ever since its inception. MDL has bee
 n applied to such\nnon-probabilistic models in var
 ious ways. We review these approaches\nand show th
 at, contrary to what is often thought, they can ex
 hibit\nsome rather problematic behaviour: on the t
 heoretical side, it is not\nknown whether the resu
 lting procedures are consistent (none of\nthe exis
 ting proof techniques can be applied). On the prac
 tical side,\nthe methods can behave quite unreason
 ably for small data samples.\nWe analyze the reaso
 ns for this undesirable behaviour and propose\na r
 adical, general and surprising solution to the pro
 blem.\n      </p>\n    
URL:/NewsandEvents/Archives/2003/newsitem/440/26-M
 ay-2003-MDL-and-classification-revisited-Troy-Lee
END:VEVENT
END:VCALENDAR