神经网络模型的透明化及输入变量约简  

Visualize Black-box of NN Model and its Application in Dimensionality Reduction

作　　者：姚立忠[1] 李太福[2] 易军[2] 苏盈盈[2] 胡文金[2] 肖大志[2] 

机构地区：[1]西安石油大学电子工程学院,西安710065 [2]重庆科技学院电气与信息工程学院,重庆401331

出　　处：《计算机科学》2012年第9期247-251,278,共6页Computer Science

基　　金：国家自然科学基金(61174015;51075418);重庆市自然科学基金(CSTC2010BB2285)资助

摘　　要：由于神经网络很容易实现从输入空间到输出空间的非线性映射,因此,神经网络应用者往往未考虑输入变量和输出变量之间的相关性,直接用神经网络来实现输入变量与输出变量之间的黑箱建模,致使模型中常存在冗余变量,并造成模型可靠性和鲁棒性差。提出一种透明化神经网络黑箱特性的方法,并用它剔除模型中的冗余变量。该方法首先利用神经网络释义图可视化网络;再利用连接权法计算神经网络输入变量的相对贡献率,判断其对输出变量的重要性;最后利用改进的随机化测验对连接权和输入变量贡献率进行显著性检验,修剪模型,并以综合贡献度和相对贡献率均不显著的输入变量的交集为依据,剔除冗余变量,实现NN模型透明化及变量选择。实验结果表明,该方法增加了模型的透明度,选择出了最佳输入变量,剔除了冗余输入变量,提高了模型的可靠性和鲁棒性。因此,该研究为神经网络模型的透明化及变量约简提供了一种新的方法。Since the neural network can easily fit the nonlinear mapping from the input space to the output space,the users of artificial neural network directly use it to get the black-box model with data pairs including input variables and output variables,often without taking dependencies between the input variables and output variables into account.So,there are often redundant variables in the model which would result in poor reliability and robustness.An approach to increase visual capability for black-box properties of neural network was proposed.Firstly,the network interpretation diagram is employed to make the network transparency.Then,connection weights method is used to compute the relative contribution of each input variable for estimating the importance to the output variable.Lastly,the significance tests of the connection weights and contribution ratios of input variables are implemented using the improved randomization tests for trimming the model,and the redundant variables can be eliminated by the intersection of the variables which are not significant for the overall contribution and the relative contribution rate to realize dimensionality reduction of neural network model.The experimental results indicate that the method can increase the transparency of model,select the best input variable set,eliminate redundant input variables,and improve the reliability and robustness of the model.Therefore,the study provides a new approach to visualize neural network model and eliminates redundant input variables.

关 键 词：神经网络模型 透明化 网络释义图 变量选择 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]