神经网络结构设计的准则和方法
A RECENT SURVEY OF NEURAL NETWORK ARCHITECTURE DESIGN
-
摘要: 神经网络结构设计一直是一个很有意义但又难以解决的问题,文中回顾并总结了近年来神经网络结构设计的研究状况.首先分析了在神经网络结构设计中应考虑的4个准则,即神经网络的函数逼近误差、网络结构的复杂性、网络的泛化能力和网络的容错性,随后从构造法、删除法和进化方法3个方面介绍了各种神经网络的设计,并提出了今后研究的展望.Abstract: Designing neural network architecture is an interesting and tough task,this paper reviews recent researches in this area.First we analysis four criteria while designing neural network topology:minimizing the function approximation error,complexity,generalization and fault tolerance.After that,we give a detailed discussion of three designing methods,namely constructive algorithms,destructive algorithms and evolutionary algorithms.Finally suggestions for further research are presented.