The robustness of BP-networks

In the framework of neural network theory, a lot of research deals with designing self-organising neural networks that seem to be appropriate for a particular task domain. However, a good training accuracy does not usually guarantee a satisfactory robustness and/or generalization capability of the trained network. The aim of this paper is to contribute to better understanding the behaviour of BP-networks, their knowledge extraction and generalization capability. This is the way along which neural networks and rule-based AI-systems are generally hoped to unify. We formulate a so-called separation characteristic that can be used as a criterion for evaluating robustness of BP-networks in many “conventional” cases. Then we show that it is possible to find for every BP-network an ε-equivalent one with smaller separation characteristics