论文部分内容阅读
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 error feedback learning of connected networks usually only changes the weight coefficient of the network. The learned knowledge only stores the connections of the neurons within the network used by the sub-networks. The function of the neurons remains the same during the learning process. Nerves in the human brain have no way to deal with the information on the changing information environment should have the appropriate adaptability, such a view for the learning of connected networks means that in the learning process, not only the network neurons within the connection, The neural function that represents the way nerves do not process information should also change and participate in learning. In this paper, a general description of multi-block neural networks with the above functions is given in the form of vector matrix, and the corresponding error back propagation learning algorithm is introduced. The vector-matrix description of multiple neural networks and their learning algorithms contributes to the stability analysis of the network and the convergence analysis of learning algorithms.