Otimização do desempenho de Redes Neurais do tipo Funções de Base Radial utilizando Vetores Bipolares Ortogonais

Abstract

This work proposes the use of orthogonal bipolar vectors (VBO) as new targets for Artificial Neural Networks (ANN) of the Radial Base Functions (RBF) type. Such vectors provide the expansion of the distance between the points of the output space, also understood as targets of RNA. The expansion of this distance reduces the chances of incorrect classification of patterns. The network was trained and tested with three sets of biometric data (human iris, handwritten digits and signs of the Australian sign language). The objective was to verify the network performance with the use of OBVs and compare the results obtained with those presented for the Multilayer Perceptron (MLP) networks. In addition, it is desired to compare two training techniques for RBF-type networks. Datasets used in the experiments were obtained from the CASIA Iris Image Database developed by the Chinese Academy of Sciences - Institute of Automation, Semeion Handwritten Digit of Machine Learning Repository and UCI - Machine Learning Repository. The networks were modeled using OBVs and conventional bipolar vectors for the purpose of comparing the results and the classification of the patterns in the output layer was based on the Euclidean distance. The results show that the use of OBVs in the network training process improved the hit rate and reduced the amount of cycles required for convergence.