A New Hybrid Neural Network Classifier Based on Adaptive Neuron and Multiplicative Neuron

Abstract

Neural network (NN) classifiers are very popular tools for solving classification tasks. Mostly known NN classifier is a multilayer perceptron (MLP). Although MLP has a good correct classification ratio, its structure could be very complex and network training may work for a long time. Pi-sigma NN (PSNN) is higher-order NN (HONN), which used higher-order correlations among the input components to establish a HONN, and the PSNN utilizes the product of neurons as the output units. By contrast with MLP and PSNN, single multiplicative neuron (SMN) is simple concerning its structure and mathematical model. The absence of the hidden layer(s) could be an advantage for easy implementation, and the mathematical model can be easily interpreted. In this paper, we propose a new hybrid NN classifier based on simple adaptive neurons and SMN which form the SMN as a whole, where input units are constituted by adaptive neurons. In contrast with conventional NN, our proposed classifier can use fewer learning parameters. To train this network, we use a modified particle swarm optimization (MPSO) algorithm. For the investigation of the generalization capability of the proposed classifier, we compare this method to other NN classifiers: MLP and PSNN together with other classification procedure classifiers.