Construction of Artificial Neural Networks for Pattern Recognition Using a Successive Geometric Segmentation Method

Abstract

This work presents a new method for artificial neural networks construction and training based on successive geometric segmentation (SGSM). This methodology is capable of generating both the network topology and neurons weights without parameters’ specification. The SGSM groups the data of each class into hyper-boxes (HBs) aligned in accordance with the largest axis of its points’ distribution. If the HBs are linearly separable, then a separating hyperplane may be identified resulting in a neuron. If it is not, then a segmentation technique is applied to divide the data into smaller classes for new HBs. For each subdivision, new neurons are added to the network. The tests show a rapid method with a high success rate.