Generalization in a diluted neural network

CRISOGONO DA SILVA; FRANCISCO TAMARIT; NEY LEMKE; JEFERSON ARENZON; EVALDO CURADO

JOURNAL OF PHYSICS A-MATHEMATICAL AND THEORETICAL

IOP PUBLISHING LTD

Lugar: Londres; Año: 1995 vol. 28 p. 1593 - 1602

1751-8113

We study the generalization capability of an extreme and asymmetrically diluted version of the Hopfield model through analytical and simulation techniques. Generalization is the ability of the system for grouping a given set of correlated patterns (the examples) in distinct classes (concepts), in such a way that each concept represents the common features of a set of examples and are attractors of the network dynamics. The dynamics of the system can be solved exactly and the generalization error for the long-time regime can be evaluated. As occur for the storage capacity, here it is shown that dilution improves the performance of the network as a categorization device when confronted with the fully connected Hopfield model.