提高前馈神经网络泛化能力的新算法  

A New Approach for Improving Generalization Ability of MultiayerFeedfoward Networks

作　　者：鲁子奕[1] 杨绿溪[1] 吴球 何振亚[1] 

机构地区：[1]东南大学无线电系

出　　处：《电路与系统学报》1997年第4期7-12,共6页Journal of Circuits and Systems

基　　金：国家自然科学基金

摘　　要：利用神经网络提取含噪数据的特征时，如何提高泛化能力是亟需解决的重要问题。通常的解决方法是通过结构优化和正则化来控制网络的复杂性。本文从熵函数的概念出发，提出了一项新的加性惩罚因子以消除过训练现象，从而有效的提高网络的泛化能力。Generalization ability is a major problem encountered when using neural network to find the structures in a noisy data sets.Controlling the network complexity is a commom method to solve the problem.In this paper,however,a novel additive penalty term which represents the features extracted by hidden units is introduced to eliminate the overtraining of multiayer feedfoward networks.Computer simulations demonstrate that the generalization ability is greatly improved.

关 键 词：泛化能力 过训练 熵 神经网络 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]