Abstract
The authors study the relation between spatial structure and information processing properties of layered Ising spin neural networks with lateral interactions. The interactions between layers are given by the Hebb rule, the interactions within layers by the so-called anti-Hebb rule. Secondly they study the development of spatial structure in such networks as the result of an unsupervized learning process (now both neurons and interactions are dynamical variables). By calculating the spectrum of the output covariance matrix as a function of the spectrum of the input covariance matrix, they show how the spatial characteristics of the input signals are reflected in the information processing properties of the equilibrated system.