J. Phys. France
Volume 51, Numéro 15, août 1990
Page(s) 1585 - 1594
J. Phys. France 51, 1585-1594 (1990)
DOI: 10.1051/jphys:0199000510150158500

Quasi-optimized memorization and retrieval dynamics in sparsely connected neural network models

Karl E. Kürten1, 2

1  Institut für Theoretische Physik, Universität zu Köln, D-5000 Köln, F.R.G. and
2  HLRZ c/o KFA Jülich, D-5170 Jülich 1, F.R.G.

Several network topologies with sparse connectivity suitable for information processing in neural network models are studied. A trial and error scheme for quasi-optimal neighbour search is shown to successfully overcome instabilities giving rise to chaotic behaviour. Moreover, computer simulations reveal that for random unbiased patterns the connectivity of the network can be adapted to the specific structure of the information the network is asked to capture such that it can be quasi-optimally stabilized. On the other hand, networks with random or purely nearest neighbour interactions are not competitive candidates for the realization of associative memories. It is shown further that sparsely connected network models with quasi-optimal neighbour search techniques substantially outperform their fully connected counterparts.

8710 - General theory and mathematical aspects.
0705M - Neural networks, fuzzy logic, artificial intelligence.

Key words
neural nets