CAN^2:构件组合式神经网络  

CAN^2:component-assembled neural network

作　　者：吴和生[1,2] 王崇骏[1,2] 谢俊元[1,2] 

机构地区：[1]南京大学计算机软件新技术国家重点实验室 [2]南京大学软件学院,江苏南京210093

出　　处：《山东大学学报（工学版）》2010年第5期171-178,共8页Journal of Shandong University（Engineering Science）

基　　金：国家自然科学基金资助项目(60807538)

摘　　要：探索有广阔应用前景的工程化神经计算技术是促进智能计算技术进展的一种有效途径。为有效解决工程化神经计算面临的难题:神经网络的"黑箱性",提出了构件组合式神经网络(component-assembled neural network,CAN2)。基于CAN2技术,构建了易于理解和复用的数字逻辑神经构件库(digital logic neuro-component library,DLNL)。应用DLNL实现复杂数字逻辑功能、解决任意分类问题,展示CAN2技术的优越性。实验表明,CAN2能有效减少神经网络的"黑箱性",有较强的可复用性,为神经计算工程化作出了一种有效的尝试。Engineering neuro-computing,as an effective approach to boost intelligent computing technology,focus on a puzzle：the ＇black box＇ property of a neural network.It means that knowledge learning from a neural network implicates the vast connected weights.A user cannot understand what the neural network learns and what task the neural network can deal with.What is more,the user cannot know how the neural network predicts and why the neural network reasons these or those conclusions.In order to effectively solve this puzzle,a component-assembled neural network（CAN2）was proposed.Based on CAN2 technology,comprehensible and reused digital logic neuro-component library（DLNL）was constructed.The complex digital logic function was implemented and random classification problems was solved by applying DLNL.Experiment indicates that CAN2 could effectively reduce the ＇black box＇ property of a neural network and had powerful reusability.It is an effective attempt in engineering neuro-computing,and which can improve user confidence for constructing an intelligent system by applying a neural network.

关 键 词：神经计算 构件 神经元 神经网络 可复用性 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]