Statistical methods for construction of neural networks.

Wlodzislaw Duch, Rafal Adamczak,
Department of Informatics, Nicolaus Copernicus University,
Grudziadzka 5, 87-100 Torun, Poland.

International Conference on Neural Information Processing, ICONIP'98, Kitakyushu, Japan, Oct. 1998, Vol. 2, pp. 629-642

Despite all the progress in neural networks the technology is still brittle and sometimes difficult to apply. Automatic construction of networks and proper initialization of their adaptive parameters are the key factors to create robust neural networks. Methods of initialization of MLPs are reviewed and new methods based on statistical discriminants and logical rules are suggested. These methods in many cases achieve higher accuracy before learning starts than random initialization achieves after the learning process finishes.

Paper in PDF format, 99KB