Back to single-layer learning principles

Summary form only given, as follows. Simple single-layer learning principles like the perceptron rule have been proven to be of limited computational power. However, combining two such principles, the perceptron rule and a simple competitive learning rule, relaxes a great number of these limitations. Moreover, this combination preserves positive properties of simple learning rules such as fast and reliable convergence and good generalization capabilities. Computational experiments with a difficult, highly nonlinear classification task have confirmed these hypotheses: a neural network based on these two principles is superior to the classical multi layer backpropagation in misclassification rates, learning speed and reliability of convergence. In addition, its generalization capabilities are substantially better due to the smoothness enforced by the linearity of the single-layer perceptron