智能神经元网络前向传播过程的硬件实现  

Hardware implementation of multilayer feed forward network with intelligent neuron

作　　者：张海燕[1] 李欣[1] 

机构地区：[1]中国海洋大学电子工程系,山东青岛266071

出　　处：《哈尔滨工程大学学报》2006年第B07期40-45,共6页Journal of Harbin Engineering University

基　　金：国家863课题基金资助项目（2001AA612030）.

摘　　要：在离线训练多层前馈网络的基础上，在FPGA中实现了智能神经元网络的前向传播过程．神经网络的学习采用改进的遗传算法，其中对遗传算法交叉、变异后所得结果的处理方面做了改进，并将整个进化过程分为渐进和骤变2个阶段．所构造的多层前馈网络具有智能神经元模型，从而提高了神经网络的收敛速度和泛化能力．在硬件实现中采用对称表求和算法逼近网络激励函数，用较少的存储空间在FPGA中构造智能神经元，使神经网络具有占用硬件资源少和工作速度高的优点．该方法应用于10000m同轴电缆仿真线的研究中，收到了良好的效果．Multi-layer feed forward network with intelligent neuron is designed and then trained using the improved genetic algorithm （GA）. Applying STAM （Symmetric Table and Addition Method） algorithm to approximate the activation function of neural network, the forward propagation process is realized by field programmable gate array （FPGA）. In the proposed improved GA. the results of crossover and permutation operation in GA must be selected and the whole evolution process is divided into two stages. So the improved GA has the ability of global optimization and higher convergence speed. In the hardware implementation of the activation function, STAM algorithm is used to reduce the amount of memory and make the operation simpler. Finally, experiment results are given based on our proposed approach. For an underwater coaxial cable, a neural network with one hidden layer is trained to correspond with the line＇s transfer function.

关 键 词：智能神经元 遗传算法 FPGA STAM算法 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]