论文部分内容阅读
讨论了输入层和输出层为恒等变换神经元.隐层为非线性变换神经元的三层神经网络的统计灵敏度,给出了权值和输入样本的加性扰动灵敏度计算的理论公式.当输入扰动较小时,通过选择多种不同的隐层神经元函数,按照本文理论结果计算出来的灵敏度数值与仿真实验的结果十分吻合,说明了本文理论分析结果的正确性.当输入扰动较小时,隐层取不同函数时灵敏度的值是不同的,但数值基本上稳定不变.对于某些隐层函数,当输入扰动值较大时,网络的灵敏度数值迅速下降,但是网络输出的平均相对误差却迅速增加.在求灵敏度时,网络的权值是用代价函数为0的精确算法求得的.
The input layer and the output layer are invariably transformed neurons, and the hidden layer is the statistical sensitivity of the three-layer neural network with non-linear transformation neurons, and the theoretical formulas of weight and additive perturbation sensitivity of the input samples are given. When the input perturbation is small, the sensitivity values calculated according to the theoretical results are in good agreement with the simulation results by selecting a variety of different hidden layer neuron functions, which shows the correctness of the theoretical analysis results in this paper.When the input disturbance is small, When the hidden layer takes different functions, the sensitivity values are different, but the values are basically stable. For some hidden layer functions, when the input disturbance value is large, the sensitivity value of the network drops rapidly, but the average relative error of network output But rapidly increasing.When seeking the sensitivity, the weight of the network is obtained by the exact algorithm whose cost function is zero.