Improvement of the Generalization Ability of the Single-Layered Perceptron with Many Outputs

Abstract

In order to improve the ability of generalization of the single-layered perceptron with many outputs, two new recognition models are proposed: (1) by activations, (2) by distances. In those models, a whole input space is divided into K (number of categories) convex polyhedrons each of which includes the same category patterns.
Experimental results on the 2-dimensional space and on the on-line hand-written Chinese character recognition are reported, which shows that the rates of recognition for unknown data by the proposed models are about 6_??_18% higher than by the conventional model.