摘要
提出了一种神经网络参数的初始化方法——向量单位化方法.从理论上分析了该方法的优点:(1)具有明确的数学意义和良好的数学性质;(2)能有效避免网络进入饱和区,提高网络的收敛性能.最后通过BP网络的实例对该方法的优点进行了实证研究,结果表明方法同样适用于其他类型的神经网络,具有积极的推广意义.
The paper puts out a worthy popularized method of Parameter Initialization for neural network-vector unitization method. It is theoretically demonstrated that the method has two advantages at least: (1)It has obviously mathematical significance and excellent mathematical property; (2)By the method,ANN can effectively avoid getting into the saturated region, and convergences better. A practical example is also manifested to justify the method in the end.
出处
《陕西科技大学学报(自然科学版)》
2008年第6期124-127,共4页
Journal of Shaanxi University of Science & Technology
基金
南京铁道职业技术学院重点课题项目资助(08Q001-Z)
关键词
神经网络
参数初始化
向量单位化
收敛性能
neural network
parameter initialization
vector unitization
convergence performance