训练多层网络的样本数问题

THE SAMPLE SIZE FOR TRAINING MULTI-LAYERED NEURAL NETWORK

本文分析多层网络的映射增长函数,以经验风险最小和期望风险最小之间的偏差来定义网络的泛化能力。基于Vapnik-Chervonenkis的事件出现频率一致收敛于其概率的理论,讨论网络的结构、训练样本数和网络泛化能力间的关系。分析在最不利的情况下为保证一定泛化能力所需要的训练样本数。

Based on the theory of uniform convergence of frequencies of events to their probabilities given by Vapnik and Chervonenkis, the relation-ship among the number of mapping functions, the size of training samples, and the ability of generalization of the multilayered neural network is discussed. The minimum training sample size, which guarantees valid generalization in the worst case, is analysed.