FAST PRUNING STRATEGY FOR NEURAL NETWORK SIZE OPTIMIZATION AND ITS APPLICATIONS

A new pruning strategy used to evaluate the optimal size of a multi-layered feed-forward neural network is proposed in this paper.The strategy is designed for arbitrary neural network converging after normal training.The output matrix of the hidden layer is calculated and used to extract characteristics.The number of the significant characteristics corresponds to the optimal number of hidden nodes.The redundant hidden nodes are deleted and their previous contributions are loaded on to the remaining ones.Therefore,the initial weights of the optimal structure can be mathematically calculated and assigned.After retrain the optimal structure, all the weights will be determined.Finally,this strategy is satisfactorily applied to solving a real pattern classification problem.The results indicate that this new method requires less computing time,and the accuracy of predictions is improved.Moreover,each stage of the proposed strategy possesses definite mathematical explanations so that the strategy could be generalized.