Stochastic Comple ity of ariational Bayesian
Tikara Hosino, TokyoInstitute ofTechnology · 2005
ariational Bayesian Learning wasproposed asthe approximation methodofBayesian learning. Inspite ofeffi- ciency andexperimental goodperformance, their mathematical property hasnotyetbeenclarified. Inthis paperweanalyze variational Bayesian hidden Markovmodels whichinclude the trueonethusthemodels arenon-identifiable. We derive their asymptotic stochastic complexity. Itisshownthat, insomeprior condition, thestochastic complexity ismuchsmaller thanthose ofidentifiable models.