Self-organizing construction of hierarchical structure of multi-layer perceptrons

A novel algorithm for creation of a hierarchical structure of neural network classifiers for classification of large databases is suggested. Each node of the hierarchical tree is a multilayer perceptron trained by the algorithm combining self-organization with supervised learning. Thus, the problems of clustering and classification for a given node are solved in concord. Also, it allows the a priori information on similarity of grouped patterns to be naturally taken into account. The algorithm performance has been tested on model data and on real-world problems.