A neural network algorithm for fast pruning based on remarkable analysis
Fujin Li, Meijie Huo, Hongge Ren, Wenbin Zhao · 2014
Neural network architecture designed for large-scale and the generalization is poor, presents a neural network algorithm for fast pruning based on significance analysis. The essence of the method is based on large-scale neural network perceptron as the research object, the constructor error curved surface model to analyze the network connection weights of disturbance on the network output error caused by the impact of hidden layer neurons carry remarkable analysis, direct remove redundant hidden layer neurons, reach pruning the neural network structure while improving its generalization ability pruning purposes. Experimental results show that the conventional algorithm, the optimal pruning neurosurgery in quick pruning network structure has a simpler and faster learning speed.