多层前馈神经网络的误差特性分析

Error feature analysis of multilayered feedforward neural networks

为分析人工神经网络（ＡＮＮ）的品质或性能对权值扰动及输入误差的容忍特性，文章基于输入及权值的随机模型，采用统计学的方法，得到了由ｓｉｇｍｏｉｄ型神经元构成的任意多层前馈神经网络在任意大小且具有任意相关性的输入误差与／或权值扰动下，输出误差特性的通用算法。仿真及对比理论计算结果表明了所提算法是正确的。

The tolerance of qualities or performance of artificial neural networks (ANNs) to weight disturbance and input errors is studied. Based on a stochastic model for inputs and weights, and in view of the disturbance of arbitrarily correlative and large input and weight errors which may exist simultaneously, a general algorithm to obtain the output error feature of the multilayered feedforward neural networks composed of sigmoidal neurons is proposed by using statistical approach. The results of computer simulation and relative theoretical computation indicate that the proposed algorithm is correct.