Abstract：The learning issue of feedforward neural networks whose activation function of hidden neurons satisfies Mercer condition is discussed in theory. The approach to improving learning speed is investigated. Then a fast learning algorithm of feedforward neural networks based on screening samples dynamically is proposed, which improves learning speed, solves the abuses of those learning algorithm based on gradient decent method and has the self-configuring advantage by determining the number of hidden neuron dynamically. The reliability and advantage of the proposed algorithm are illustrated concretely through test.
2005, Vol. 27 
