A hierarchical neural network architecture for classification

Abstract

In this paper, a hierarchical neural network with cascading architecture is proposed and its application to classification is analyzed. This cascading architecture consists of multiple levels of neural network structure, in which the outputs of the hidden neurons in the higher hierarchical level are treated as an equivalent input data to the input neurons at the lower hierarchical level. The final predictive result is obtained through a modified weighted majority vote scheme. In this way, it is hoped that new patterns could be learned from hidden layers at each level and thus the combination result could significantly improve the learning performance of the whole system. In simulation, a comparison experiment is carried out among our approach and two popular ensemble learning approaches, bagging and AdaBoost. Various simulation results based on synthetic data and real data demonstrate this approach can improve the classification performance. © 2012 Springer-Verlag.