The BP-SOM architecture and learning rule

For some problems, the back-propagation learning rule often used for training multilayer feedforward networks appears to have serious limitations. In this paper we describe BP-SOM, an alternative training procedure. In BP-SOM the traditional back-propagation learning rule is combined with unsupervised learning in self-organizing maps. While the multilayer feedforward network is trained, the hidden-unit activations of the feedforward network are used as training material for the accompanying self-organizing maps. After a few training cycles, the maps develop, to a certain extent, self-organization. The information in the maps is used in updating the connection weights of the feedforward network. The effect is that during BP-SOM learning, hidden-unit activations of patterns, associated with the same class, becomemore similar to each other. Results on two hard to learn classification tasks show that the BP-SOM architecture and learning rule offer a strong alternative for training multilayer feedforward networks with back-propagation.