BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:On Casting Importance Weighted Autoencoder to an EM Algorithm to L
 earn Deep Generative Models
DTSTART;VALUE=DATE-TIME:20191107T081500Z
DTEND;VALUE=DATE-TIME:20191107T083000Z
DTSTAMP;VALUE=DATE-TIME:20260724T053721Z
UID:indico-contribution-19@sshep.snu.ac.kr
DESCRIPTION:Speakers: Dongha Kim (Seoul National University)\nWe propose a
  new and general approach to learn deep generative models. \n  Our approac
 h is based on a new observation that the importance weighted autoencoders 
 (IWAE) can be understood as a procedure of estimating the MLE with an EM a
 lgorithm. \n  Utilizing this interpretation\, we develop a new learning al
 gorithm called importance weighted EM algorithm (IWEM). \nIWEM is an EM al
 gorithm with importance sampling (IS)\nwhere the proposal distribution is 
 carefully selected to reduce the variance\ndue to IS. In addition\, we dev
 ise an annealing strategy to stabilize the learning algorithm. For missing
  data problems\, we propose a modified\nIWEM algorithm called miss-IWEM. \
 n  Using multiple benchmark datasets\, we demonstrate empirically that our
  proposed methods outperform IWAE with significant margins for both fully-
 observed and missing data cases.\n\nhttps://sshep.snu.ac.kr/event/107/cont
 ributions/19/
LOCATION:Seoul National University\, College of Natural Sciences\, Buildin
 g 25-1 국제회의실 (International Conference Hall)
URL:https://sshep.snu.ac.kr/event/107/contributions/19/
END:VEVENT
END:VCALENDAR