基于输出层权值解析修正的神经网络有效训练  

Efficient Learning Approach for Feedforward Neural Networks Based on the Analytic Modification for the Weights of Output Layer

作　　者：张德贤[1] 

机构地区：[1]郑州工程学院计算机科学系,郑州450052

出　　处：《计算机工程与应用》2005年第4期82-84,140,共4页Computer Engineering and Applications

基　　金：河南省教育厅基础研究项目资助(编号:2003520261);河南省自然科学基金项目资助(编号:994060500)

摘　　要：根据神经网络训练误差对权值的梯度特征分析,提出了网络输出层权值与网络隐含层权值轮换修正的思想,并基于网络输出层权值与网络隐含层权值之间的依赖关系,建立了网络输出层权值解析修正和隐含层权值修正的具体方法,所提出的方法通过提高网络权值修正的准确性而提高网络训练的有效性。根据网络输出节点的输出误差与其总输入误差的关系,提出了进一步提高所获得网络推广性的具体方法。实例计算结果表明,所提出的方法可以显著地提高网络的训练效率,并有效地增强网络推广性。The idea of rotation modification between the weights of the network output layer and the weights of the hide layer is proposed according to the analysis of derivative characteristics of network training errors.The techniques of the analytic modification for the weights of the output layer and the modification for the weights of the hide layer are established based on the relation between the weights of the network output layer and the weights of the hide layer.The new methods proposed effectively improve neural network's learning efficiency through the enhancement of the modification validity for the network weights.The techniques to further improving trained network's generalization are also presented according to the analysis of the relation between the output errors and total input's errors of output nodes.Actual computation cases demonstrate that the technique proposed could substantially improve network's learning efficiency and generalization.

关 键 词：神经网络 解析修正 轮换修正 推广性 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]