版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
SergiosTheodoridis
KonstantinosKoutroumbasVersion2ACourseon1PATTERNRECOGNITIONTypicalapplicationareasMachinevisionCharacterrecognition(OCR)ComputeraideddiagnosisSpeechrecognitionFacerecognitionBiometricsImageDataBaseretrievalDataminingBionformaticsThetask:Assignunknownobjects–patterns–intothecorrectclass.Thisisknownasclassification.2Features:
Thesearemeasurablequantitiesobtainedfromthepatterns,andtheclassificationtaskisbasedontheirrespectivevalues.
Featurevectors:
Anumberoffeatures
constitutethefeaturevector
Featurevectorsaretreatedasrandomvectors.3Anexample:4Theclassifierconsistsofasetoffunctions,whosevalues,computedat,determinetheclasstowhichthecorrespondingpatternbelongsClassificationsystemoverviewsensorfeaturegenerationfeatureselectionclassifierdesignsystemevaluationPatterns5Supervised–unsupervisedpatternrecognition:
ThetwomajordirectionsSupervised:Patternswhoseclassisknowna-prioriareusedfortraining.Unsupervised:Thenumberofclassesis(ingeneral)unknownandnotrainingpatternsareavailable.6CLASSIFIERSBASEDONBAYESDECISIONTHEORYStatisticalnatureoffeaturevectorsAssignthepatternrepresentedbyfeaturevector
tothemostprobableoftheavailableclasses
Thatis
maximum7Computationofa-posterioriprobabilitiesAssumeknowna-prioriprobabilities
Thisisalsoknownasthelikelihoodof8
TheBayesrule(Μ=2)where9TheBayesclassificationrule(fortwoclassesM=2)Given classifyitaccordingtotheruleEquivalently:classify accordingtotheruleForequiprobableclassesthetestbecomes1011Equivalentlyinwords:Dividespaceintworegions
ProbabilityoferrorTotalshadedarea
BayesianclassifierisOPTIMALwithrespecttominimisingtheclassificationerrorprobability!!!!12Indeed:MovingthethresholdthetotalshadedareaINCREASESbytheextra“grey〞area.13TheBayesclassificationruleformany(M>2)classes:Givenclassifyittoif:SuchachoicealsominimizestheclassificationerrorprobabilityMinimizingtheaverageriskForeachwrongdecision,apenaltytermisassignedsincesomedecisionsaremoresensitivethanothers14ForM=2Definethelossmatrix
penaltytermfordecidingclass,
althoughthepatternbelongsto,etc.Riskwithrespectto15Riskwithrespectto
AverageriskProbabilitiesofwrongdecisions,weightedbythepenaltyterms16Chooseandsothatr
isminimizedThenassigntoifEquivalently:
assignxin if :likelihoodratio17If18Anexample:19Thenthethresholdvalueis:Threshold forminimumr20Thusmovestotheleftof(WHY?)21DISCRIMINANTFUNCTIONS
DECISIONSURFACESIf arecontiguous: isthesurfaceseparatingtheregions.Ononesideispositive(+),ontheotherisnegative(-).ItisknownasDecisionSurface+
-22Iff(.)monotonic,theruleremainsthesameifweuse: isadiscriminantfunctionIngeneral,discriminantfunctionscanbedefinedindependentof
theBayesianrule.Theyleadtosuboptimalsolutions,yetifchosenappropriately,canbecomputationallymoretractable.23BAYESIANCLASSIFIERFORNORMALDISTRIBUTIONSMultivariateGaussianpdfcalledcovariancematrix24ismonotonic.Define:
Example:
25
Thatis, isquadraticandthesurfacesquadrics,ellipsoids,parabolas,hyperbolas,
pairsoflines.Forexample:26DecisionHyperplanesQuadraticterms: IfALL (thesame)thequadratictermsarenotofinterest.Theyarenotinvolvedincomparisons.Then,equivalently,wecanwrite: DiscriminantfunctionsareLINEAR
27Letinaddition:
28Nondiagonal:
Decisionhyperplane29MinimumDistanceClassifiersequiprobable
EuclideanDistance:
smaller
MahalanobisDistance:
smaller3031Example:32MaximumLikelihood
ESTIMATIONOFUNKNOWNPROBABILITY
DENSITYFUNCTIONS33
3435
Asymptoticallyunbiasedandconsistent36Example:37MaximumAposterioriProbabilityEstimationInMLmethod,θwasconsideredasaparameterHereweshalllookatθasarandomvectordescribedbyapdfp(θ),assumedtobeknownGiven ComputethemaximumofFromBayestheorem38Themethod:3940Example:41BayesianInference
4243TheaboveisasequenceofGaussiansas
MaximumEntropyEntropy
44Example:
x
isnonzerointheinterval
andzerootherwise.ComputetheMEpdfTheconstraint:LagrangeMultipliers
45
MixtureModels
Assumeparametricmodeling,i.e.,Thegoalistoestimate givenasetWhynotML?Asbefore?46Thisisanonlinearproblemduetothemissinglabelinformation.Thisisatypicalproblemwithanincompletedataset.TheExpectation-Maximisation(EM)algorithm.Generalformulation
whichare
notobserveddirectly.
Weobserve
amanytoonetransformation47LetWhatweneedistocomputeBut arenotobserved.HerecomestheEM.Maximizetheexpectationoftheloglikelihood
conditionedontheobservedsamplesandthecurrentiterationestimateof48Thealgorithm:E-step:M-step:ApplicationtothemixturemodelingproblemCompletedataObserveddata
Assumingmutualindependence49UnknownparametersE-stepM-step
50NonparametricEstimation
51ParzenWindowsDividethemultidimensionalspaceinhypercubes52DefineThatis,itis1insideaunitsidehypercubecenteredat0
Theproblem:Parzenwindows-kernels-potentialfunctions53Meanvalue
Henceunbiasedinthelimit54VarianceThesmallerthehthehigherthevarianceh=0.1,N=1000h=0.8,N=100055h=0.1,N=10000ThehighertheNthebettertheaccuracy56If
asymptoticallyunbiasedThemethodRemember:
57CURSEOFDIMENSIONALITYInallthemethods,sofar,wesawthatthehighestthenumberofpoints,
N,thebettertheresultingestimate.Ifintheone-dimensionalspaceaninterval,filledwith
N
points,isadequately(forgoodestimation),inthetwo-dimensionalspacethecorrespondingsquarewillrequireN2
andintheℓ-dimensionalspacetheℓ-dimensionalcubewillrequireNℓpoints.Theexponentialincreaseinthenumberofnecessarypointsinknownasthecurseofdimensionality.Thisisamajorproblemoneisconfrontedwithinhighdimensionalspaces.58NAIVE–BAYESCLASSIFIERLetandthegoalistoestimate i=1,2,…,M.Fora“good〞estimateofthepdfonewouldneed,say,Nℓpoints.Assumex1,x2,…,xℓmutuallyindependent.Then:Inthiscase,onewouldrequire,roughly,Npointsforeachpdf.Thus,anumberofpointsoftheorderN·ℓwouldsuffice.ItturnsoutthattheNaïve–Bayesclassifierworksreasonablywellevenincasesthatviolatetheindependenceassumption.59KNearestNeighborDensityEstimationInParzen:ThevolumeisconstantThenumberofpointsinthevolumeisvaryingNow:Keepthenumberofpoints
constantLeavethevolumetobevarying
60
61TheNearestNeighborRuleChoosekoutoftheNtrainingvectors,identifytheknearestonestoxOutofthesekidentifykithatbelongtoclassωi
Thesimplestversion
k=1!!!ForlargeNthisisnotbad.Itcanbeshownthat:
ifPBistheoptimalBayesianerrorprobability,then:62
ForsmallPB:63Voronoitesselation64BayesProbabilityChainRuleAssumenowthattheconditionaldependenceforeachxi
islimitedtoasubsetofthefeaturesappearingineachoftheproductterms.Thatis: where
BAYESIANNETWORKS65Forexample,ifℓ=6,thenwecouldassume: Then:TheaboveisageneralizationoftheNaïve–Bayes.FortheNaïve–Bayestheassumptionis:Ai=Ø,fori=1,2,…,ℓ66AgraphicalwaytoportrayconditionaldependenciesisgivenbelowAccordingtothisfigurewehavethat:x6isconditionallydependentonx4,x5.x5
on
x4
x4
on
x1,x2x3
on
x2x1,x2
areconditionally
independentonothervariables.Forthiscase:67BayesianNetworksDefinition:ABayesianNetworkisadirectedacyclic
graph(DAG)wherethenodescorrespondtorandomvariables.Eachnodeisassociatedwithasetofconditionalprobabilities(densities),p(xi|Ai),wherexi
isthevariableassociatedwiththenodeandAi
isthesetofitsparentsinthegraph.ABayesianNetworkisspecifiedby:Themarginalprobabilitiesofitsrootnodes.Theconditionalprobabilitiesofthenon-rootnodes,giventheirparents,forALLpossiblecombinations.68ThefigurebelowisanexampleofaBayesianNetworkcorrespondingtoaparadigmfromthemedicalapplicationsfield.ThisBayesiannetworkmodelsconditionaldependenciesforanexampleconcerningsmokers(S),tendenciestodevelopcancer(C)andheartdisease(H),togetherwithvariablescorrespondingtoheart(H1,H2)andcancer(C1,C2)medicaltests.69OnceaDAGhasbeenconstructed,thejointprobabilitycanbeobtainedbymultiplyingthemarginal(rootnodes)andtheconditional(non-rootnodes)probabilities.Training:Onceatopologyisgiven,probabilitiesareestimatedvia
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 洗胃知识考核题目及答案
- 等效内阻专项试题与答案分享
- 酒店委托管理合同(范本)
- 六年级下册数学北师大含答案 整数2
- 四年级下册数学北师大含答案 探索与发现:三角形内角和1
- 湖理工机械设计基础课件08回转件的平衡
- 合格不合格产品分区存放细则
- 2026中国基因治疗设备产业化路径与商业化前景展望报告
- 小升初教育考试题目与答案
- 9.八上数学1.3.2-1.3.5课时练习
- 人力公司劳动知识竞赛
- 非ST段抬高型心肌梗死诊疗指南(2025年版)
- 煤矿事故应急预案与处置培训课件
- 养殖建房合同
- 配网调控培训知识课件
- DB65T 4633-2022 棉花消防安全管理规范
- 2026届福建省宁德市八年级物理第一学期期末联考试题含解析
- 在建工程转固课件
- 2020典型精密零件机械加工工艺分析实例
- 教育机构经营情况说明范文
- 小学英语教师进城考试试题及答案
评论
0/150
提交评论