Ensemble of diluted attractor networks with optimized topology for fingerprint retrieval

Abstract

The present study analyzes the retrieval capacity of an Ensemble of diluted Attractor Neural Networks for real patterns (i.e., non-random ones), as it is the case of human fingerprints. We explore the optimal number of Attractor Neural Networks in the ensemble to achieve a maximum fingerprint storage capacity. The retrieval performance of the ensemble is measured in terms of the network connectivity structure, by comparing 1D ring to 2D cross grid topologies for the random shortcuts ratio. Given the nature of the network ensemble and the different characteristics of patterns, an optimization can be carried out considering how the pattern subsets are assigned to the ensemble modules. The ensemble specialization splitting into several modules of attractor networks is explored with respect to the activities of patterns and also in terms of correlations of the subsets of patterns assigned to each module in the ensemble network.