Enhancing Decision-Based Neural Networks Through Local Competition Gustavo Camps-Valls a,1,.pdf

Enhancing Decision-Based Neural Networks Through Local Competition Gustavo Camps-Valls a,1, Luis Go?mez-Chova a, Joan Vila-France?s a, Jose? D. Mart??n-Guerrero a, Antonio J. Serrano Lo?pez a, Emilio Soria-Olivas a aDept. Enginyeria Electro?nica, Universitat de Vale?ncia, Spain. Abstract In this paper the Decision-Based Neural Network (DBNN) learning algorithm is modified to stimulate local competition. Performance is assessed in ten UCI databases, resulting in improved results at the expense of a relatively low increase of the computational burden. Key words: Decision-based neural network; hierarchical network structure; competitive credit-assignment scheme; local competition; UCI database. Classification codes: neural networks, signal analysis. 1 Introduction Credit-assignment criteria is the fundamental guiding principle for a great variety of classification algorithms available in the literature (e.g. neural networks, decision trees, support vector machines [1]). A particularly interesting decision-driven algo- rithm for pattern recognition is the decision-based neural network (DBNN), which usually provides very fast and satisfactory learning performance, along with an easily scalable network’s structure. However, different strategies are needed when dealing with highly overlapping distributions and/or issues on false acceptance/rejection, e.g. introduction of non-linear discriminant functions, fuzzy-decision neural networks, or modular networks (see [2,3] for full details). 1 Correspondence address: Prof. Gustavo Camps-Valls. Escola Te?cnica Superior d’Enginyeria (ETSE). Dept. Enginyeria Electro?nica. Grup de Processament Digital de Senyals. C/ Dr. Moliner, 50. Burjassot (Vale?ncia). Spain. Tel.: +34 96 3160197; Fax: +34 96 3160466. E-mail address: gustavo.camps@uv.es, http://www.uv.es/～gcamps. Preprint submitted to Neurocomputing Journal (short communications)4 September 2005 In the DBNN framework, multi-classification problems are tackled by means of task div