A new fast algorithm for effective training of neural classifiers

Abstract A neural classifier whose training can be executed very effectively is proposed to overcome the disadvantages of the method of potential functions. The power of the method of potential functions is limited by the severe requirements for computation time and storage. After mapping the computational structure of the method of potential functions to a three-layered feedforward network, a new fast learning algorithm is applied to train the net. By the characteristics of the new training algorithm, the network's complexity can be largely reduced without degrading its performance too much. That is to say, the algorithm will enable the network to learn in a very fast and, moreover, very effective manner.