An Adaptive Counter Propagation Network

Abstract

In traditional CPN network and its learning method, the number of neurons in competitive layer is difficult to decide. Too many neurons in the competitive layer will generate "dead neurons", while too few neurons will make the competitive layer unstable. In this paper, an adaptive counter propagation network named ACPN and its approach are proposed, where the number of competitive neurons can be decided adaptively. In ACPN, the neurons in competitive layer are generated dynamically, so each neuron in the competitive layer can do its best in training. Because of the efficiency of neurons in competitive is improved sufficiently, ACPN can work well with the least amount of neurons and realize the required capability of network. The experiment shows that the improved model ACPN runs faster and is more efficient than other CPN networks.