摘要
Ｋａｎｅｒｖａ的稀疏分布存贮（ＳＤＭ）模型由于对寻址地址采用了稀疏编码，对数据采用了分布式存贮，从而解决了大维数向量的输人问题。ＳＤＭ实际上是一个由输入层、中间层和输出层组成的三层前向网络，其中神经元间的互连权值在输入层与中间层是预置的（用矩阵Ａ表示），中间层与输出层的连接权阵Ｃ由外积法得到。文中假定在相同的学习规则下，就信噪比意义而言，Ａ的均匀预置能使ＳＤＭ获得最优性能，从而为Ａ的预置提供了理论依据。
Kanerva' s Sparse Distributed Memory (SDM)solves the large dimensional(≥ 1000)input problem by introducing sparse coding for addresses and distributed memorying for data. In fact,SDM is a three-tayer forward neural network which consists of an input layer,an immediate layer and an output layer,The connected weight matrix between the input layer and the immediate layer is represented by A which is preset,whereas the connected weight matrix C between the immediate layer and the output one is obtained by the outer-product method, In this paper, in the sense of signal-to-noise ratio (RSN), we have proved that SDM performance by presetting A uniformly is superior to that by presetting it nonuniformly,hence a theoretical basis is offered for presetting A.
出处
《南京航空航天大学学报》
CAS
CSCD
1994年第6期822-826,共5页
Journal of Nanjing University of Aeronautics & Astronautics
基金
国家863高技术资助
关键词
联想记忆
神经网络
稀疏分布存贮
连接权阵
associative memories
neural networks
sparse distributed memory
signal-tonoise ratio
connected weight matrix
