形态双向联想记忆网络的摄动鲁棒性研究  

Perturbation Robustness of Morphological Bidirectional Associative Memories Networks

作　　者：何春梅[1] 叶征春[2] 韩牟[3] 叶有培[3] 

机构地区：[1]华东交通大学信息工程学院,江苏南京210094 [2]华东交通大学机电工程学院,江西南昌330013 [3]南京理工大学计算机科学与技术学院,江苏南京210094

出　　处：《南京理工大学学报》2011年第5期664-669,共6页Journal of Nanjing University of Science and Technology

基　　金：华东交通大学科研基础项目(09102018);南京理工大学科研计划项目(ZYTS067)

摘　　要：针对训练模式对的小幅摄动可能会对神经网络的输出产生副作用的问题,该文研究了形态双向联想记忆(MBAM)网络和模糊形态双向联想记忆(FMBAM)网络训练模式对的摄动鲁棒性。比较了训练模式对摄动前后2种网络输出范围的变化,理论证明了MBAM和FMBAM 2种网络的摄动鲁棒性。理论和仿真实例表明:MBAM网络对训练模式对的摄动全局拥有好的鲁棒性,其训练模式对的选取可以适度粗糙;而FMBAM网络对训练模式对的摄动不具有好的鲁棒性,则其训练模式对的选取精度要求较高。In view of the problem that the small perturbations of training pattern pairs may cause side effects on the outputs of neural networks,the influences of perturbations of training pattern pairs on the morphological bidirectional associative memories（MBAM）networks and fuzzy morphological bidirectional associative memories（FMBAM）networks are researched here.By comparing the output variations of the two training sample pairs before and after perturbations,the perturbation robustness of the MBAM networks and FMBAM networks is theoretically proved.The simulations show that the MBAM network has good robustness on the perturbation of training pattern pairs,and its rough training pattern pairs are acceptable.The FMBAM network robustness is not so good,and its training pattern pairs should have higher accuracy.

关 键 词：神经网络 形态双向联想记忆 模糊 摄动 训练模式对 鲁棒性 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]