Sequential classification by perceptrons and application to net pruning of multilayer perceptron

摘要

Using the important property of approximating a posteriori probability functions of the classes in the outputs of the trained multilayer perceptrons, we propose the technique for the implementation of sequential classification by a perceptron and/or multilayer perceptron, and the application to the node growing in the number of input nodes of a perceptron and the number of hidden nodes of a multilayer perceptron. A measurement for the ordering of hidden nodes of the trained multilayer perceptron is also proposed. The ordering of the hidden nodes comes from the contribution of each hidden node. Using the node growing technique, the minimum number of hidden nodes can be obtained in the training and used in the classification. The technique can also be applied to a single layer perceptron. In the experiment, a typical "XOR" problem was applied, and the balance between the reduction of hidden nodes and classification results was quite good.