Structured patterns retrieval using a metric attractor network: Application to fingerprint recognition

The ability of a metric attractor neural networks (MANN) to learn structured patterns is analyzed. In particular we consider collections of fingerprints, which present some local features, rather than being modeled by random patterns. The network retrieval proved to be robust to varying the pattern activity, the threshold strategy, the topological arrangement of the connections, and for several types of noisy configuration. We found that the lower the fingerprint patterns activity is, the higher the load ratio and retrieval quality are. A simplified theoretical framework, for the unbiased case, is developed as a function of five parameters: the load ratio, the finiteness connectivity, the density degree of the network, randomness ratio, and the spatial pattern correlation. Linked to the latter appears a new neural dynamics variable: the spatial neural correlation. The theory agrees quite well with the experimental results.