摘要
目标问题的复杂程度和网络处理能力的适合程度是影响人工神经网络推广性的本质因素。为了衡量前向神经网络（ Ｆ Ｎ Ｎ）的处理能力，该文对 Ｆ Ｎ Ｎ 的插值误差进行了研究，得到了统计意义下 Ｆ Ｎ Ｎ 处理能力的估计值，进而定义了能间接反映神经网络推广性的推广性量度。该方法能够估计出适合目标问题的网络规模，应用于函数逼近和样本分类问题，仿真结果证实了该方法的有效性。
The complexity of a given problem and the capacity and the ability in the sense of statistics of artificial neural networks (ANN) are the key factors that influence the generalization of ANN. By investigating the capacity and the ability in the sense of statistics of feedforward neural networks (FNN), A measurement of capacity and generalization of FNN, which can be used to estimate the proper number of FNN hidden nodes is defined. This method is applicable to functional approximation and classification problems.
出处
《清华大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
1999年第7期54-58,共5页
Journal of Tsinghua University(Science and Technology)
基金
国家"八六三"高技术项目
关键词
神经网络
前向神经网络
处理能力
推广性量度
artificial neural networks
feedforward neural networks
generalization
interpolation error
