Training a kind of hybrid universal learning networks with classification problems
D. Li, Koutaro Hirasawa, Jinglu Hu, Junichi Murata · 2003
In the search for even better parsimonious neural network modeling, this paper describes a novel approach which attempts to exploit redundancy found in the conventional sigmoidal networks. A hybrid universal learning network constructed by the combination of proposed multiplication units with summation units is trained for several classification problems. It is clarified that the multiplication units in different layers in the network improve the performance of the network.