Analysis of a learning algorithm for neural network classifiers

The authors provide a convergence analysis of a learning rule derived for the adaptation of the neurons´ synaptic weight vectors representing the prototype vectors of the class distribution in a classifier. The analysis also provides a theoretical foundation for the Kohonen learning vector quantization (LVQ1 and LVQ2) algorithms. The convergence of the learning rules is proved under certain conditions. More specifically, it is shown that the algorithm will converge to an error-free solution when the input patterns are linearly separable