Numéro |
J. Phys. I France
Volume 3, Numéro 8, August 1993
|
|
---|---|---|
Page(s) | 1693 - 1715 | |
DOI | https://doi.org/10.1051/jp1:1993210 |
J. Phys. I France 3 (1993) 1693-1715
Effect of synapse dilution on the memory retrieval in structured attractor neural networks
N. BrunelLaboratoire de Physique Statistique, Ecole Normale Supérieure, 24 rue Lhomond, F-75231 Paris Cedex 05, France
(Received 22 February 1993, accepted 27 April 1993)
Abstract
We investigate a simple model of structured attractor neural network (ANN). In this network a module codes for the category
of the stored information, while another group of neurons codes for the remaining information. The probability distribution
of stabilities of the patterns and the prototypes of the categories are calculated, for two different synaptic structures.
The stability of the prototypes is shown to increase when the fraction of neurons coding for the category goes down. Then
the effect of synapse destruction on the retrieval is studied in two opposite situations : first analytically in sparsely
connected networks, then numerically in completely connected ones. In both cases the behaviour of the structured network and
that of the usual homogeneous networks are compared. When lesions increase, two transitions are shown to appear in the behaviour
of the structured network when one of the patterns is presented to the network. After the first transition the network recognizes
the category of the pattern but not the individual pattern. After the second transition the network recognizes nothing. These
effects are similar to syndromes caused by lesions in the central visual system, namely prosopagnosia and agnosia. In both types of networks (structured or homogeneous) the stability of the prototype is greater than the stability of individual
patterns, however the first transition, for completely connected networks, occurs only when the network is structured.
© Les Editions de Physique 1993