基于子空间分析的前馈神经网络隐层评测方法  

A hidden layer evaluation method for feedforward neural networks based on subspace analysis

作　　者：朴相范[1] 崔荣一 洪炳熔[2] 李白雅[3] 

机构地区：[1]延边大学理工学院,吉林延吉133002 [2]哈尔滨工业大学计算机科学与技术学院,黑龙江哈尔滨150001 [3]湖南科技大学信息学院,湖南湘潭411201

出　　处：《湖南科技大学学报（自然科学版）》2004年第4期59-63,共5页Journal of Hunan University of Science And Technology：Natural Science Edition

基　　金：国家自然科学基金资助项目(编号:69362001);湖南自然科学基金资助项目(编号:01JJY2141)

摘　　要：针对最佳平方逼近3层前馈神经网络模型,采用子空间分析方法,讨论了隐单元的误差补偿性能,提出了隐层评测方法.研究结果表明隐单元选取策略应遵循其输出向量有效分量位于误差空间、回避耗损空间和尽可能靠近某一能量空间的原则,这一结果与隐单元采用什么激发函数无关,也允许各隐单元采用不同激发函数.网络的隐层性能评价可以通过隐层品质因子、隐层有效系数、隐单元剩余度来进行,而总体结果可采用隐层评价因子进行评测.评测实验表明,所提出的隐层评测方法是合理有效的.图1,表1,参11.The error compensation performance of the hidden unit was discussed according to the model of least-squares approximation feedforward neural networks using the method of subspace analysis,and further more,a hidden layer evaluation method was proposed.It was revealed that the efficient component of output vector of a hidden unit should be lying in error space,avoiding expend space and closing an energy space,which were independent of the nonlinear functions used by the hidden units,and it was permitted that different hidden units have different activation functions.The quality factor of hidden layer,efficient coefficient of hidden layer,the redundancy of hidden units and the evaluation factor of hidden layer,for total evaluation,were proposed to evaluate the performance of the hidden layer.The rationality and validity of proposed method is validated by the results of evaluation experiment.1fig.,1tab.,11refs.

关 键 词：三层前馈神经网络 隐层生长 评测方法 误差补偿 子空间分析 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]