职业健康安全管理体系标准理解讲座知识

《职业健康安全管理体系标准理解讲座知识》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、Chapter3Maximum-LikelihoodandBayesianParameterEstimation(7,10)ProblemsofDimensionalityComputationalComplexityHiddenMarkovModelsDrDingYuxinPatterClassificationchapter3part3Problemsinvolving50or100features(binaryvalued)ClassificationaccuracydependsuponthedimensionalityandtheamountoftrainingdataCaseof

2、twoclassesmultivariatenormalwiththesamecovariance3.7ProblemsofDimensionalityDrDingYuxinPatterClassificationchapter3part3Iffeaturesareindependentthen:MostusefulfeaturesaretheonesforwhichthedifferencebetweenthemeansislargerelativetothestandarddeviationIthasfrequentlybeenobservedinpracticethat,beyondac

3、ertainpoint,theinclusionofadditionalfeaturesleadstoworseratherthanbetterperformance:wehavethewrongmodel!7DrDingYuxinPatterClassificationchapter3part3777DrDingYuxinPatterClassificationchapter3part3ComputationalComplexityOurdesignmethodologyisaffectedbythecomputationaldifficulty“bigoh”notationf(x)=O(h

4、(x))“bigohofh(x)”If:f(x)=2+3x+4x2g(x)=x2f(x)=O(x2)7DrDingYuxinPatterClassificationchapter3part3“bigoh”isnotunique!f(x)=O(x2);f(x)=O(x3);f(x)=O(x4)“bigtheta”notationf(x)=(h(x))If:f(x)=(x2)butf(x)(x3)DrDingYuxinPatterClassificationchapter3part3ComplexityoftheMLEstimationGaussianpriorsinddimensions

5、classifierwithntrainingsamplesforeachofcclassesForeachcategory,wehavetocomputethediscriminantfunctionTotal=O(d2..n)Totalforcclasses=O(cd2.n)O(d2.n)Costincreasewhendandnarelarge!DrDingYuxinPatterClassificationchapter3part3OverfittingSampleisinadequateReducedimensionality(selectasubsetoffeaturesorcom

6、binefeatures)AllcclassessharethesamecovariancematrixLookforabetterestimateforcovariancematrixPseudo-Bayesianestimation:Anextremecase:statisticalindependenceHowtogetbetterperformanceifstatisticaldependenceSufficientdataPreventoverfittingDrDingYuxinPatterClassificationchapter3part3Goal:makeasequenceof

7、decisionsAprocessthatunfoldintime,statesattimetareinfluencedbyastateattimet-1Applications:speechrecognition,gesturerecognition,partsofspeechtaggingandDNAsequencing3.10HiddenMarkovModelsDrDingYuxinPatt