多块神经网络误差反传的新概念  

NEW CONCEPTS IN BACKPROPAGATION FOR MULTIBLOCK NETWORKS

作　　者：周苏[1] 许立俭[1] 孙学瑞 

机构地区：[1]青岛大学

出　　处：《青岛大学学报（自然科学版）》1995年第2期9-20,共12页Journal of Qingdao University(Natural Science Edition)

摘　　要：连接型网络的误差反传学习通常只是改变网络的权系数，所学的知识仅存储子所用网络内部神经元的连接之中，而神经元的作用函数在学习过程中保持不变．人脑中的神经无处理信息的方式对变化的信息环境应该具有相应的自适应性，这样的观点用于连接型网络的学习便意味着，在学习过程中，不仅网络内部神经元的连接，表示神经无处理信息方式的作用函数也应该可以变化，参与学习．本文对具有上述功能的多块神经网络以矢量一矩阵的形式给出了一般性的描述，并介绍了相应的误差反传学习算法．多块神经网络及其学习算法的矢量一矩阵描述有助于网络的稳定性分析和学习算法的收敛性分析．The connectionist learning by backpropagation makes use of weights-change-only where the learned knowledge is concentrated only in the inter-neuron connections of the applied network and the activation functions of neurons remain constant during the learning process.The new consideration, that the neurons of the human brain should adapt themselves to a changing environment, implies that,.in addition to the neuron interconnections,also the activation functions of the neurons should be tunable during the process of learning. General descriptions for multiblock networks and their learning by backpropagation in vector-matrix form, introduced here, contribute to the algorithmic improvements,and may facilitate the stability and convergence analysis of the networks.

关 键 词：误差反传 连接权系数 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]