How lateral interaction develops in a self-organizing feature map

Abstract

A biologically motivated mechanism for self-organizing a neural network with modifiable lateral connections is presented. The weight modification rules are purely activity-dependent, unsupervised, and local. The lateral interaction weights are initially random, but develop into a 'Mexican hat' shape around each neuron. At the same time, the external input weights self-organize to form a topological map of the input space. The algorithm demonstrates how self-organization can bootstrap itself using input information. Predictions of the algorithm agree very well with experimental observations on the development of lateral connections in cortical feature maps. >