摘要
本文提出一种新的联想记忆模型,这种模型可根据Perceptron算法进行学习。它是一种非对称的、互连的人工神经元网络。从理论上证明了在一定条件下这种网络能记忆样本的最少个数,能够使得所要记忆的样本都能成为神经元网络动力学系统的稳定吸引子。从心理学角度看,它与人记忆采些信息的方法很相近。为了能够使网络运行时可逃离非样本吸引子,回到样本吸引子,我们又提出了一种加深联想记忆的学习及相应的运行算法。为了使任意多个样本能够存储在一个网络里,我们提出一种附加节点方法,附加节点对应于模式的一种概念,这种方法也解决了模式分类中的线性不可分问题。最后给出了计算机模拟结果。
A new associative memory network model is proposed. Its learning algorithm is based on perception, and its nodes connect with each other in asymmetrical manner. The number of samples that can be stored in a net under some conditions is proved theoretically such that the samples can become the stable attractors of the nonlinear dynamical neural system. To escape from a non-sample attractor and attain a sample attractor when system operates dynamically, a deepening impression learning algorithm is given. In order to store arbitrary number of samples in a net, an augmented node method is provided. The linear unseparable problems in pattern classification, can be also solved by the method.
出处
《计算机学报》
EI
CSCD
北大核心
1990年第5期331-339,共9页
Chinese Journal of Computers
