Convergence in a learning network with pattern feedback

Abstract

An architecture based on networks of logic functions is proposed with the output of the network forming a model or archetype image of the class of pattern stimulating the net. Frequency of occurrence information which relates the output of each function to the number of times the stimulus sampled in the test pattern has occurred during the training phase is used. This is shown to have significant advantages over simple binary systems and networks with progressively decreasing connectivity during training-the so called 'ageing 'effect. The frequency of occurrence networks are no longer sensitive to the sequence or order of the training categories and have a greatly improved noise stability. The network is assessed with real world data made up of some hundreds of examples of machine printed numerals.