Investigation on Improving Generalization of Neural Network
Wei Geng · Science Technology and Engineering · 2009
Survey and comparisons are made on different methods of improving generalization of neural networks,in which especially 5 realistic approaches are analyzed and contrasted.In experiments these methods are applied to function approximation and data classification.These methods are ranked in the order of their generalization ability and computation time.It concluded that the Bayesian adaptive method is excellent,the regulation method,stepwise incremental method and pruning method are in next place.The early stopping is found as the fastest,however,it performed not well in function approximation and can only applied to data classification.