Voting schemes for cooperative neural network classifiers

Multiple neural network modules cooperating in taking a classification decision are modeled as multiple voters electing one candidate in a single ballot election assuming the availability of votes' preferences and intensities. All modules are considered as candidates as well as voters. Voting bids are the output-activations of the modules forming the cooperative modular structure. Different voting schemes are compared according to the accuracy of defining classification decision boundaries. A higher classification accuracy implies a better representation of the information available at different preferences (output values). The effect of the modules' voting power on the accuracy of the decision is studied and integrated in the network's design strategy.