ABSTRACT
Unsupervised Learning of New Network Configurations Tbuplat.e probability updating e<ut cause network structures t.o converge toward opt.i111al performance, hut how does one create new network structnrcs to clmract.eriz<! newly discover<!d phonological rnlm> or unforeseen acousticphonetic possibilities'! It is necessary to be able to detect when a correctly recognized acoustic input docs not. match the correct path through the network very well, and furthermore, establish that this acoustic data. could be generated by a. human vocal t.ra.ct obeying rules of English phonetics and phonology. Detecting a poor match may not be too difficult for a. machine, but to be able to determine whether the deviations arc worthy of inclusion as new network paths of local or global import requires expert knowledge of the rules of speech production and their acoustic consequences. It is my belief that the role of analysis by synthesis and the motor theory of speech perception arises exactly here, to serve as a constraint on the construction of alternative network paths during unsupervised lea.rning.4 Construction of a LAFS-Iike computer simulation possessing these skills must await progress in understanding the detailed relations between speech production, perception, and phonology.
