Abstract
In this paper, we propose an architecture and learning algorithm for a growing neural network. Drawing inspiration from the idea of electrical potentials, we develop a classifier based on a set of synthesised potential fields over the domain of input space using symmetrical functions (kernels). We propose a multilayer, multiclass potential function generators classifier (POFGEC) utilising growing architecture and a training algorithm to sequentially add potential functions created by the training patterns, if the addition improves the NN classification performance. We also present a pruning algorithm to achieve compact architecture. POFGEC incorporates the electrical potentials concept in the two main neural net building blocks: potential function generators (PFGs) and potential function entities (PFEs), which perform a non-linear transformation of the input data and create the decision rules by constructing the cumulative potential functions and adjusting the weights. The implementation of the presented method with several datasets demonstrates its capabilities in generating classification solutions for datasets of various shapes independent from the number of predefined classes. We also offer substantial comparative analysis with other known approaches in order to fully illustrate the capabilities of the proposed method and its relation with other existing techniques.