Unsupervised adaptation using structural Bayes approach

It is well-known that t hc p~~rforrnancc: Of rm-ognit.ion syst.crns is often largely degraded when t.here is a mismatch Mween the training and testing environment. It, is desirahlr to cornpensat,e for the mismatch when the system is iu operation without. any supervised learning. Recently. a sl.ruct ural maximum a posteriori (SM.\P) adaptation appreach. in which a hierarchical structure in the parameter space is assurnetl. was proposed. In this paper. this SNAP method is applied to unsuprrvisrd atlapt.abiOn. .A novel normalization trchniquc is also int.rotlucctl as a front end for the adapt ation procxss. ‘I‘hc rc~c:ognit.ion rc~sults showed t.hat r he proposctl Inct hod was cff(~crivc CWII when Only One ut,reranc‘c from it new sprilk(,r was used for atlap~,atiOn. FiirI hrrmorc. ali c:lfcctivc~ way to cx’IIit)inc* t hc sup(:rvih(d adapt iLt iou i~lld t hc: Iln?;up<‘rvisc:d il.tlil[)l.itt ion was irlwsl igiltcd I0 rducx: the tieed for a large amount of suprrvisd learriiug dat.a.