模糊汉明神经网络及其实现的研究

Fuzzy Hamming Neural Networks and Its Implementation

由于传统汉明神经网络未解决模式重叠和识别算法是否一定收敛的问题 ,也未充分利用输入模式与其他神经元之间的靠近程度信息 ,本文提出一种模糊汉明神经网络 .模糊汉明神经网络可接受二值和非二值输入 ;使用模糊类隶属度子网解决模式重叠问题和充分利用靠近程度信息 ;采用比较子网保证算法的收敛和减少互连 .

Since typical Hamming networks has not solved the problem of pattern overlap and convergence, and not made full use of the near degree information of the input pattern with other neurons in the network,a fuzzy Hamming neural networks (FHNN) is proposed in this paper.FHNN replaces matching subnet with a fuzzy class membership subnet to solve problems of pattern overlap.It is also a three layer feed forward network.The n elements of the input pattern are presented in parallel to the n nodes of the first layer of the subnet.The number of the nodes in the hidden layer is equal to the amount of the exemplar patterns,and the weights of hidden neurons are the components of exemplar.And the number of the nodes in the output layer is equal to that of the classes to classify.The weights are the fuzzy class membership function of exemplar pattern to the classes.Only the training of the threshold T of the hidden layer and the fuzzy class membership weights of the output layer are needed.FHNN replaces competitive subnet with a comparing subnet to solve the problem of not converging and having too many interconnections.The fuzzy membership degrees of an unknown pattern to the classes are compared in parallel with a gradually decreasing reference voltage as a dynamic threshold in the comparing subnet.When the reference voltage decreases to the level of certain fuzzy class membership degree,the corresponding binary outputs skips to 1.Using modular circuit design,this network is easily extended and implemented in VLSI technology.FHNN is composed of three separate chips.The first matching chip gets the matching scores; the second calculates the fuzzy class membership degree.These two chips construct the fuzzy class membership subnet.And the third chip is a comparing subnet.As for the implementation of input mapping subnet,conventional feed forward network circuits can be used.