资源描述
单击此处编辑母版标题样式,单击此处编辑母版文本样式,第二级,第三级,第四级,第五级,*,基于局部泛化误差界的RBF网络训练方法研究,报告人:刘晓艳,基于局部泛化误差界的RBF网络训练方法研究报告人:刘晓艳,1,主要内容,课题来源及背景和意义,研究现状及分析,所做的工作,遇到的问题及进一步的工作,参考文献,主要内容课题来源及背景和意义,2,课题来源及背景和意义,RBF神经网络的结构选择中,即隐含层神经元个数的确定问题,一直是难点。合理的选择其结构会提高RBF神经网络的泛化能力。,局部泛化误差模型,考虑分类器在输入空间局部区域上的泛化能力。对于量化的考察对于网络的容错能力(error-tolerance)和泛化能力(generalization ability)有一定启发意义。,神经网络的敏感性标示着这种分类器的variance特性,而经验误差的大小则是标示着分类器的bias特性,将两者有机的结合起来作为一种评价分类器泛化能力的标准可能会有很好的效果。(,criteria,),课题来源及背景和意义RBF神经网络的结构选择中,即隐含层,3,研究现状及分析,介绍局部泛化误差模型现状,敏感性SM(sensitivity measure),敏感性定义及其计算,敏感性用途,Constructive for neural network,Center selection,Feature/sample/weight accuracy selection,研究现状及分析介绍局部泛化误差模型现状敏感性定义及其计算敏感,4,敏感性定义及其计算,定义:衡量网络输出对于输入或权重(或其他的参数)的扰动而改变程度的定量度量。,对象上:,Sensitivity to input perturbation,Sensitivity to weight perturbation,Sensitivity to neuron perturbation,计算方式上:,Partial derivative sensitivity analysis,stochastic sensitivity analysis,要求激活函数对于输入是可微的并且输入扰动必须很小,考察输出变化的期望或方差概率特性,敏感性定义及其计算定义:衡量网络输出对于输入或权重(或其他的,5,敏感性的应用,正是由于敏感性考察网络各参数的变化对于网络输出的影响程度,因而,基于敏感性分析来优化或调整各参数的选择即成为它的主要应用方向。,敏感性引用于RBF神经网络的中心选择(结构选择),1.“,Sensitivity analysis applied to the construction of radial basis function networks”D.Shia,D.S.Yeung,J.Gao,2.“,LOCALIZED GENERALIZATION ERROR AND ITS APPLICATION TO RBFNN TRAINING”,WING W.Y.NG,DANIEL S.YEUNG,DE-FENG WANG,ERIC C.C.TSANG,XI-ZHAO WANG,3.“,Hidden neuron pruning multilayer perceptrons using a sensitivity measure”Daniel s.yeung,xiaoqin zeng,敏感性用于sample selection(Active learning),“Active Learning Using Localized Generalization Error of Candidate Sample as Criterion”Patrick P.K.Chan,Wing W.Y.Ng,Daniel S.Yeung,敏感性用于feature selection,wing,敏感性的应用正是由于敏感性考察网络各参数的变化对于网络输出的,6,研究现状及分析,现存的局部泛化误差模型理论,differences between the maximum and minimum values of the target output,分类问题中,目标输出的最大最小值之差至少为1,那么将该模型用于结构选择时就会出现问题。,研究现状及分析现存的局部泛化误差模型理论difference,7,研究现状及分析,现存的局部泛化误差模型用于RBFNN的结构选择,思想:两个,分类器f1,f2,如果存在Q1,使得,f2 has a,better,Generalization,capability,RSM(Q1)a,for f1,RSM(Q2)a,for f2,Q1 Q2,在相同误差界标准下,设计分类器使得它覆盖的Q邻域比较大,认为覆盖的邻域面积越大,得到的分类器的泛化能力越好。,分析:界的阈值a的取值标准难以确定,现存的方法建议a取0.25,这样在解上述二次方程时就会出现问题。,研究现状及分析现存的局部泛化误差模型用于RBFNN的结构选择,8,研究现状及分析,分类问题中取值大于1,0.25,1.由于在解方程时存在矛盾之处,造成该模型用于RBFNN结构设计时存在问题。,2.有关界的表达式,存在常数A其值是否相对过大的问题,相对于前两项如果取值过大的话,其失去意义。,3.单纯的将经验误差作为训练RBF分类器的标准的话,存在过拟和 以及得到的分类器的泛化能力不高的缺点。,研究现状及分析分类问题中取值大于10.251.由于在解方程,9,所做的工作,将经验误差项和敏感性项的加和做为一种新的评价分类器泛化能力的标准(QNBQ neighborhood balance)。考察其合理性。,将QNB用于RBFNN的结构选择,设计网络结构。,用范数形式简化现有的局部泛化误差模型的分析表达式。得到一种基于范数的局部泛化误差界的分析式。,所做的工作将经验误差项和敏感性项的加和做为一种新的评价分类器,10,QNB作为一个衡量分类器评价标准的合理性,measure for classifier complexity,图示(1):“,simple”classifier,Low SM,but bad training error,QNB作为一个衡量分类器评价标准的合理性measure fo,11,QNB作为一个衡量分类器评价标准的合理性,图示(2),:“,complex”classifier,high SM,but bad generalization capability and maybe overfitting,VC维较大,QNB作为一个衡量分类器评价标准的合理性图示(2):“com,12,QNB作为一个衡量分类器评价标准的合理性,图示(3),:“,good fit”classifier,what we expected,Good balance,between,Training error,SM,QNB作为一个衡量分类器评价标准的合理性图示(3):“goo,13,QNB作为一个衡量分类器评价标准的合理性(实验),Sensitivity measure 衡量RBFNN复杂程度,Iris dataset,Ionosphere dataset,QNB作为一个衡量分类器评价标准的合理性(实验)Sensit,14,QNB作为一个衡量分类器评价标准的合理性(实验),Hidden number(K),QNB作为一个衡量分类器评价标准的合理性(实验)Hidden,15,QNB用于RBFNN的结构选择(architecture selection),Algorithm:,Step 1:,Start with the number of the hidden neurons by 1.,Step 2:,Perform k-means clustering to find the location of centers for the hidden numbers.,Step 3:,Select the width of each neuron to be half of the maximum distance between the,center itself and other neurons.,Step 4:,Using pseudo-inverse method to obtain the weight.,Step 5:,For a selected Q value,compute the current neural networks error bound by the following equation:,Step 6:,Find the minimum error bound,and output the corresponding hidden neurons number.,QNB用于RBFNN的结构选择(architecture s,16,初步实验情况,Hidden number,9,13,8,8,7,10,9,7,7,9,8.7,(average),Train accuracy,0.9619,0.9810,0.9238,0.9714,0.9524,0.9619,0.9714,0.9619,0.9810,0.9714,0.9638,(average),Test accuracy,0.9333,0.9556,0.8667,1,0.9333,0.9778,0.9333,0.9111,0.9556,0.9333,0.9400,(average),(Iris,Q=0.1)information,:,4 150,3 classes,(Pima,Q=0.1)information,:,8768,2 classes,Hidden number,23,22,15,18,17,22,22,26,18,21,20.4,(average),Train accuracy,0.7989,0.7877,0.7914,0.7803,0.7877,0.7952,0.8082,0.8007,0.8007,0.7952,0.7946,(average),Test accuracy,0.7749,0.7706,0.7662,0.7489,0.7662,0.7792,0.7359,0.7489,0.7749,0.7619,0.7628,(average),初步实验情况Hidden number91388710977,17,初步实验情况,Hidden number,7,8,8,7,9,7,9,7,8,8,7.8000,(average),Train accuracy,0.9597,0.9919,0.9758,0.9839,0.9839,0.9597,0.9758,0.9597,0.9758,0.9758,0.9742,(average),Test accuracy,0.9259,0.9444,1,0.9074,0.9815,0.9815,0.9444,0.9630,0.9815,0.9630,0.9593,(average),(Wine,Q=1.5)information,:,13178,3 classes,Hidden number,17,18,16,15,19,18,14,18,16,16,16.7(average),Train accuracy,0.9184,0.9388,0.9347,0.9143,0.9469,0.9347,0.9429,0.9510,0.9347,0.9224,0.9339(average),Test accuracy,0.9245,0.9245,0.9528,0.9340,0.9151,0.9245,0.9245,0.9057,0.9434,0.9340,0.9283(average),(Ionosphere,Q=1.0)information,:,34 351,2 classes,初步实验情况Hidden number78879797887,18,初步
展开阅读全文