{A decomposition technique for modular neural networks

R. Moddemeijer and L. Spaanenburg

University of Groningen, Department of Computing Science,
P.O. Box 800, NL-9700 AV Groningen, The Netherlands,
phone: +31.50.363 3940 - fax: +31.50.363 38005 - e-mail: rudy@cs.rug.nl


Modularity is required to resolve the data interference problems that usually plague monolithic designs. The presence of data interference shows from the learning curve of the net. Re-learning does not always remove the problem; then a modular structure can be used. However, there is no guarantee that data interference can be removed in the simple decomposition. Ensemble networks and mixed experts can be used for specific decompositions. On the other hand, there is not a generalized constructive method to achieve an efficient modularity. In this paper an information-theoretical approach to the decomposition of modular neural classifiers is presented, that allows for arbitrary hierarchies.


Order estimation, mutual information, data interference, classification, modular neural networks

Full paper , pdf-format

Poster presentation


Proceedings of the 11th workshop on Circuits, Systems and Signal Processing, (IEEE/ ProRISC ), 30 November - 1 December 2000, Mierlo (NL), pp. 427-433, eds. Jean Pierre Veen, STW Technology Foundation, Utrecht (NL), ISBN: 90-73461-24-3, BibTeX , Note: proceeding are published on CD-rom only