Interpretable neural networks with BP-SOM

Artificial Neural Networks (ANNS) are used successfully in industry and commerce. This is not surprising since neural networks are especially competitive for complex tasks for which insufficient domain-specific knowledge is available. However, interpretation of models induced by ANNS is often extremely difficult. BP-SOM is an relatively novel neural network architecture and learning algorithm which offers possibilities to overcome this limitation. BP-SOM is a combination of a multi-layered feed-forward network (MFN) trained with the back-propagation learning rule (BP), and Kohonen's self-organizing maps (sorts). In earlier reports, it has been shown that BP-SOM improved the generalization performance as compared to that of BP, while at the same time it decreased the number of necessary hidden units without loss of generalization performance. In this paper we demonstrate that BP-SOM trained networks results in uniform and clustered hidden layer representations appropriate for interpretation of the networks functionality.