BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Beam Sampling for Infinite Hidden Markov Models - Jurgen Van Gael
DTSTART:20080402T130000Z
DTEND:20080402T140000Z
UID:TALK11310@talks.cam.ac.uk
CONTACT:Carl Edward Rasmussen
DESCRIPTION:The Infinite Hidden Markov Model (iHMM) [1\,2] is an extension
  of the\nclassical Hidden Markov Model widely used in machine learning and
 \nbioinformatics. As a tool to model sequential data\, Hidden Markov\nMode
 ls suffer from the need to specify the number of hidden states.\nAlthough 
 model selection and model averaging are widely used in this\ncontext\, the
  Infinite Hidden Markov Model offers a nonparametric\nalternative. The cor
 e idea of the iHMM is to use Dirichlet Processes\nto define the distributi
 on of the rows of a Markov Model transition\nmatrix. As such\, the number 
 of used states can automatically be\nadapted during learning\; or can be i
 ntegrated over for prediction.\nUntil now\, the Gibbs sampler was the only
  known inference algorithm\nfor the iHMM. This is unfortunate as the Gibbs
  sampler is known to be\nweak for strongly correlated data\; which is ofte
 n the case in\nsequential or time series data. Moreover\, it is suprising 
 that we have\npowerful inference algorithms for finite HMM's (the forward-
 backward\nor Baum-Welch dynamic programming algorithms) but cannot apply t
 hese\nmethods for the iHMM. In this work\, we propose a method called the\
 n"Beam Sampler" which combines ideas from slice sampling and dynamic\nprog
 ramming for inference in the iHMM. We show that the beam sampler\nhas some
  interesting properties such as: (1) it is less susceptible to\nstrong cor
 relations in the data than the Gibbs sampler\, (2) it can\nhandle non-conj
 ugacy in the model more easily than the Gibbs sampler.\nWe also show that 
 the scope of the beam sampler idea goes beyond\ntraining the Infinite Hidd
 en Markov Model\, but can also be used to\nefficiently train finite HMM's.
LOCATION:Engineering Department\, CBL Room 438
END:VEVENT
END:VCALENDAR