NetworksforSupervisedLearning.PDF

Neural Comput Applic (1997)6:19-41 9 1997 Springer-Verlag London Limited Neural Computing Applications Combining Linear Discriminant Functions with Neural Networks for Supervised Learning Ke Chen 12, Xiang Yu 1 and Huisheng Chi 1 ~National Laboratory of Machine Perception and Center for Information Science, Peking University, Beijing, China; 2Department of Computer and Information Science and The Center for Cognitive Science, The Ohio State University, Columbus, OH, USA A novel supervised learning method is proposed by 1. Introduction combining linear discriminant functions with neural networks. The proposed method results in a tree- Neural networks, particularly Multi-Layered Per- structured hybrid architecture. Due to constructive ceptrons (MLPs), have already been found to be learning, the binary tree hierarchical architecture successful for various supervised learning tasks [1- is automatically generated by a controlled growing 11]. Both theoretical and empirical studies have process for a specific supervised learning task. shown that the neural network is of powerful capa- Unlike the classic decision tree, the linear discrimin- bilities for pattern classification and universal ant functions are merely employed in the intermedi- approximation which are typical supervised learning ate level of the tree for heu