Computer Engineering and Applications ›› 2010, Vol. 46 ›› Issue (5): 36-39.DOI: 10.3778/j.issn.1002-8331.2010.05.012

• 研究、探讨 • Previous Articles     Next Articles

Sensitivity analysis of multi-layer feed-forward neural networks based on approximating functions

WU Yue-bo,YANG Jing-shu   

  1. Department of 702,PLA Electronic Engineering Institute,Hefei 230037,China
  • Received:2008-09-10 Revised:2008-12-08 Online:2010-02-11 Published:2010-02-11
  • Contact: WU Yue-bo

基于函数逼近的多层前馈神经网络灵敏度分析

吴跃波,杨景曙   

  1. 解放军电子工程学院 702室,合肥 230037
  • 通讯作者: 吴跃波

Abstract: Sensitivity analysis is vital in the design of neural networks.The exiting approaches to the sensitivity analysis impose some limitations on network input and weight perturbations or can’t apply exact enough result.This paper gives a better sensitivity formula of multi-layer feed-forward neural networks by using two approximating functions based on piché’s stochastic model.It has the form convenient for the computation of sensitivity without introducing addition limitations,and brings less error than other methods.The result of computer simulation proves this formula is correct and exact.

Key words: neural networks, sensitivity, approximating function

摘要: 神经网络灵敏度分析对网络结构设计、硬件实现等具有重要的指导意义,已有的灵敏度计算公式对权值和输入扰动有一定限制或者计算误差较大。基于Piché的随机模型,通过使用两个逼近函数对神经网络一类Sigmoid激活函数进行高精度逼近,获得了新的神经网络灵敏度计算公式,公式取消了对权值扰动和输入扰动的限制,与其他方法相比提高了计算精度,实验证明了公式的正确性和精确性。

关键词: 神经网络, 灵敏度, 逼近函数

CLC Number: