Minimization of Number of Neurons in Voronoi Diagram-Based Artificial Neural Networks

Artificial Neural Networks (ANNs) have been widely used to deal with various classification problems for decades. Different algorithms for synthesizing ANNs have been proposed as well. The number of neurons in an ANN usually controls the tradeoff between classification ability and computational efficiency. That is, more neurons tend to yield better results but are less efficient in either the training or recalling phase. Furthermore, if the neurons are implemented by physical devices, the implementation cost can be effectively reduced with fewer number of neurons in an ANN. In this paper, we propose a method to minimize the number of neurons used in an ANN that is built by using Voronoi diagrams without suffering any capability loss. We have conducted experiments on a set of benchmarks. The experimental results show that the resultant ANNs reduce the number of neurons by up to 94 percent.