BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Oracle Variational Inference - James McInerney (Columbia Universit
 y)
DTSTART:20141126T110000Z
DTEND:20141126T120000Z
UID:TALK53695@talks.cam.ac.uk
CONTACT:Dr Jes Frellsen
DESCRIPTION:Variational inference is a powerful approach to performing Bay
 esian inference on large datasets with complex probabilistic models. But s
 tandard methods assume the data are fixed\, which is a limitation when app
 lying inference to modern data sources. In this talk\, I will abandon the 
 assumption of fixed data and develop oracle variational inference algorith
 ms that work with a possibly unbounded number of data points. I then expla
 in how to adaptively adjust the data sampling distribution to improve infe
 rence. Recent findings\, using latent Dirichlet allocation on text corpora
  where the oracle distribution uses keyword searches to query documents\, 
 indicate that both types of oracle variational inference converge faster a
 nd to better local optima than stochastic variational inference.
LOCATION:Engineering Department\, CBL Room BE-438.
END:VEVENT
END:VCALENDAR