QUALITY AND EFFICIENCY OF RETRIEVAL FOR WILLSHAW-LIKE AUTOASSOCIATIVENETWORKS .1. CORRECTION

The informational properties of a neural network model of an autoassoc iative memory based on binary Hebbian synapses are investigated. The m odel is a modification of the Willshaw network with a floating thresho ld which keeps approximately constant the number of active neurons (wi nners) at each time step. In the asymptotic case of large number of ne urons, informational characteristics have been calculated analytically for single-step correction. Comparison with simulations shows that th e maximal correction efficiency attains its asymptotic values for netw orks with surprisingly small number of neurons. Simulation results for multistep correction show considerable improvement over the single-st ep case.