Learning and generalization in a coulombian network

Constructs a two layer modular network in which each layer can be trained independently. The first layer connection weights are determined by the learning algorithm of the coulombian network. For the second layer the authors use the perceptron algorithm. To generate the first layer the authors use a sequential constructive process that is able to make a linearly separable transformation of the input patterns, so the perceptron can always make a successful separation. As a result the authors obtain both a tendency to generate a minimal network and an improved generalization ability. The authors apply this particular architecture to some classification problems with discrete and continuous inputs, analyzing learning and generalization behavior