基于领域知识的神经网络泛化性能研究进展  

Research advances of neural network generalization performance based on domain knowledge

作　　者：胡铁松[1] 严铭[1] 赵萌[1] 

机构地区：[1]武汉大学水资源与水电工程科学国家重点实验室,湖北武汉430072

出　　处：《武汉大学学报（工学版）》2016年第3期321-328,共8页Engineering Journal of Wuhan University

基　　金：国家自然科学基金项目(编号:71171151;51339004;51479142)

摘　　要：从样本复杂性、结构复杂性、学习策略和建模技术4个方面对基于领域知识的神经网络泛化性能研究进展进行了评述,指出了目前基于领域知识神经网络泛化性能研究存在的主要问题是只是利用研究对象的单调性、凸性、对称性和增益等一些简单非线性特征来虚拟训练样本、形成非监督学习算法约束条件、构造节点作用函数等方面.利用关于研究复杂对象部分已知的物理机制或动力学特性来建立有一定物理基础的神经网络模型,从而有效控制网络训练存在的过学习问题是今后神经网络泛化理论与方法研究的主要发展趋势.This article reviews the research progress of neural network generalization based on domain knowledge in four aspects, i. e. sample complexity, structural complexity, learning strategy and modeling technology. It is pointed out that the research object＇s simple nonlinear features like monotonicity,convexity,symmetry and gain are using to construct training samples, to form unsupervised learning algorithms constraints,to structure functions of neural network＇s node is the major problems in the field. Building neural network models based on some physical basis by using partial known physical mechanism or dynamic characteristic of complicated research object, to control the over-fitting problem in the network training progress, can be the main development tendencies of theory and research method of neural network generalization performance in the future.

关 键 词：神经网络 泛化性能 领域知识 先验知识 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]