Optimal, matching-score network for pattern classification
Anson M. Y. Luk, W.F. Leung · 2002
This paper presents a design method for an optimal matching-score (MS) network with exponential and sigmoid activation functions. By using a new orthogonal learning process, the proposed net is capable of learning new patterns by growing its hidden layer and without recomputing the entire interconnection weight matrices. Simulation results on signal classification and character recognition show that the MS net is highly robust to noise. Besides, the generalization capability of the net is shown superior than that of the backpropagation net.>