Newsgroups: comp.speech
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!swrinde!gatech!news.mathworks.com!uunet!in1.uu.net!portal.austin.ibm.com!bocanews.bocaraton.ibm.com!io.bocaraton.ibm.com!bens
From: bens@io.bocaraton.ibm.com (Shahshahani)
Subject: Re: Baum-Welch algorithm convergence
Sender: news@bocanews.bocaraton.ibm.com (News Admin ID)
Message-ID: <D9vH6C.1B99@bocanews.bocaraton.ibm.com>
Date: Thu, 8 Jun 1995 21:03:48 GMT
X-Nntp-Posting-Host: bcrmip1.bocaraton.ibm.com
References:  <D9tC1I.8u3@ennews.eas.asu.edu>
Organization: IBM Boca Raton
Keywords: HMM training Baum-Welch EM convergence rate
Lines: 38

Some convergence properties of the EM algorithm (not the rate of
convergence though, are studied in the following.

C.F. Jeff Wu, "On the Convergence Properties of the EM algorithm",
The Annals of Statistics, 1983, Vol 11, No 1, pp 95-103

-Ben


In article <D9tC1I.8u3@ennews.eas.asu.edu>, deisher@trcsun3.eas.asu.edu (Michael
E. Deisher) writes:
|> Hi.  I'm using the Baum-Welch (EM) algorithm to train a
|> continuous-Gaussian-mixture-density HMM (isn't everybody? :-)  In
|> practice, it converges very fast (in less than 10 iterations) for a
|> variety of model sizes/topologies.
|> 
|> Are there any good references that address the rate of convergence of
|> this algorithm, perhaps showing analytically why it is (typically) so
|> fast?  If not, do any of you have intuition as to why this happens?
|> 
|> --Mike
|> 
|> 
|>
==============================================================================
|>   |  Mike Deisher                                  Arizona State University 
|> |
|>   |  deisher@dspsun.eas.asu.edu          Telecommunications Research Center 
|> |
|>   |  voice:  (602) 965-0396                    Signal Processing Laboratory 
|> |
|>   |  fax:    (602) 965-8325                           Tempe, AZ  85287-7206 
|> |
|> 
|>
==============================================================================
|>                      If Murphy's Law can go wrong it will.
|> 
