Categorical Perception - Computational and Neural Models

Computational and Neural Models

Computational modeling (Tijsseling & Harnad 1997; Damper & Harnad 2000) has shown that many types of category-learning mechanisms (e.g. both back-propagation and competitive networks) display CP-like effects. In back-propagation nets, the hidden-unit activation patterns that "represent" an input build up within-category compression and between-category separation as they learn; other kinds of nets display similar effects. CP seems to be a means to an end: Inputs that differ among themselves are "compressed" onto similar internal representations if they must all generate the same output; and they become more separate if they must generate different outputs. The network's "bias" is what filters inputs onto their correct output category. The nets accomplish this by selectively detecting (after much trial and error, guided by error-correcting feedback) the invariant features that are shared by the members of the same category and that reliably distinguish them from members of different categories; the nets learn to ignore all other variation as irrelevant to the categorization.