Simulating prosopagnosia through a lesion of lateral connections in a feed-forward neural network.

Abstract

We show that particular features of prosopagnosic impairment can be simulated by a connectionist model trained with an unsupervised learning procedure. In particular we describe a Kohonen's neural network which is able to correctly recognize and categorize a series of digitized pictures of faces when learning is characterized by certain parameter values, but which shows a prosopagnosic behavior when lateral connections are lesioned. We discuss the relationship between this result and some neurophysiological hypotheses about prosopagnosia.