Multiwinner feedforward self-organizing neural network with a distributed representation of information

Abstract

In order to expand the range of application of various unique functions of the neural network, it is necessary to attempt hardware implementation of the model. The hardware implementations of models hitherto proposed have various problems due to the complexity of the algorithm. This paper aims at hardware implementation for easily implementing the neural network on a hardware model and proposes a layered structure model based on a simple multiwinner algorithm. The multiwinner algorithm is a competition algorithm based on the model of von der Malsburg, in which multiple neurons are able to become winners in each layer. In each layer, the neurons located in the neighborhood of the winner neuron perform learning by Hebbian rules, affected by interaction due to lateral coupling with a fixed coupling coefficient. The following experiment is presented. Learning is performed for bar patterns with six different directions. After that, when character patterns of 0 to 9 are presented, the winner neurons appear locally and distributed on the two‐dimensional plane, indicating that the feature information of the presented patterns is mapped onto the distributed representation. Then the directional response of the neuron is examined. As regards the spatial distribution of the response direction of the neuron, it is verified that a neuron having an orientation selectivity with gradually changing response direction is formed. This indicates that the distributed representation is achieved and that an orientation‐selective structure is formed in the proposed multiwinner algorithm, which can easily be implemented in hardware. © 2006 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 89(12): 31–41, 2006; Published online in Wiley InterScience (http://www.interscience.wiley.com). DOI 10.1002/ecjc.20289