Training of neural network classifier by combining hyperplane with exemplar approach

A neural network classifier which combines hyperplane with exemplar approach is presented. The network structure does not have to be specified before training. An appropriate network structure is built during training. The perceptron-based algorithm is applied to train a linear threshold unit (LTU). The LTU builds a hyperplane that classifies as many training instances as possible. HB nodes that represent hyperboxes are generate to classify the training instances that cannot be classified by the hyperplane. The proposed model works well on both clustered and strip-distributed instances. The number of HB nodes generated depends on the overlapping degree of training instances. This classifier can classify continuous-valued and nonlinearly separable instances. Online learning is supplied, and the learning speed is very fast. The parameters used are few and insensitive