前馈神经网络结构新型剪枝算法研究

A New Pruning Algorithm for Feedforward Neural Network

研究神经网络的结构优化,提出采用基于贡献值与输出连接的权重来修剪节点,节点是直接剪枝而不是消除存有内在联系的节点;该方法认为神经元贡献值低于阈值,那么此神经元就是毫无意义的,同时将该算法应用于非线性函数逼近,实验结果表明,在不牺牲网络性能的情况下,采用新型剪枝算法来修剪神经网络节点是非常有意义的,所提出的算法也是非常有效的。

In this paper ,based on the value and contribution of weights connected to the output node pruning ,pruning nodes are not directly linked to eliminate the inherent node ;neurons contribute to the method that is below the threshold value ,then the neuron is meaningless .Meanwhile ,the algo-rithm is applied to nonlinear function approximation ,the results show that the network performance without sacrificing the case ,using new pruning algorithm neural network node pruning is very mean-ingful that the proposed algorithm is also very effective .