McClelland, J. & Kaitan, P.

Stanford University

The interactive activation model of context effects in letter perception, and the related TRACE model of speech perception, have been criticized for failing to conform to principles of optimal Bayesian information integration. Critics have argued that this shortcoming is a consequence of the bi-directional propagation of information -- the claim is that interactivity inherently distorts perceptual inference. Here we address this conclusion by demonstrating that an version of the IA model called the multinomial IA model actually samples exactly from the posterior of the probability distribution of a generative model that is captured by the units, connections, and connection weights in the model.