Rijksuniversiteit Te Leiden

This paper presents a method for designing artificial neural network architectures. The method implies a reverse engineering of the processes resulting in the mammalian brain. The method extends the brain metaphor in neural network design with genetic algorithms and L-systems, modelling natural evolution and growth. It will be argued that a principle of modularity, which is inherent to the design method as well as the resulting network architectures, improves network performance. 1. Introduction Several neural network simulation studies show that many problems can not be solved by a learning algorithm in conventional fully connected layered neural networks [e.g. 2, 5, 7, 10, 14, 18, 24]. Two problems that occur frequently are: a lack of generalization and a problem called interference. A trained network is said to generalize if it gives correct responses to input patterns not seen during training. Many networks fail to generalize because they have too many degrees of freedom. A large..