On the Relation Between Stability Parameters and Sizes of Domains of Attraction in Attractor Neural Networks
A C C Coolen · Europhysics Letters (EPL) · 1991
As a measure of the stability of stored patterns in attractor neural networks Gardner defined stability parameters, which are generally assumed to be related to the sizes of the domains of attraction. In this letter I investigate the validity of this assumption for the Viana model in which patterns can have variable embedding strengths and for which, if the number of patterns is small, the domain sizes can be calculated analytically in the thermodynamic limit. Surprisingly, the stability parameters turn out not to be a direct measure of the absolute sizes of the domains of attraction, but rather of the relative contributions of the individual patterns to the total fraction of phase space that is occupied by the stored patterns.