Structure evolution of neural networks

We present an approach to investigate the dependence of the capabilities of neural networks on their connectivity structure based on the concept of random graphs. The evolution of structural properties of suitably defined spaces of random graphs with parameters characterizing the edge density and their impact on the performance of the representative edge weighted graphs as feed-forward neural networks will be studied. The occurrence of phase transitions in relevant structural properties will be established and the scaling behaviour of the governing parameters will be worked out. The presentation here is restricted to the case of a simple minded character recognition problem in a multi-layer feed-forward network. The approach is embedded in a larger context of an attempt to develop algorithms for structure optimization of neural networks. (orig.)Available from TIB Hannover: RO 5073(534) / FIZ - Fachinformationszzentrum Karlsruhe / TIB - Technische InformationsbibliothekSIGLEDEGerman