最新6 MLP神经网络-药学医学精品资料.ppt

《最新6 MLP神经网络-药学医学精品资料.ppt》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、最新6MLP神经网络-药学医学精品资料contentsstructureuniversaltheoremMLPforclassificationmechanismofMLPforclassificationnonlinearmappingbinarycodingoftheareasMLPforregressionlearningalgorithmoftheMLPbackpropagationlearningalgorithmheuristicsinlearningprocessXORandLinearSeparabilityRevis

2、itedRememberthatitisnotpossibletofindweightsthatenableSingleLayerPerceptronstodealwithnon-linearlyseparableproblemslikeXOR:However,Multi-LayerPerceptrons(MLPs)areabletocopewithnon-linearlyseparableproblems.Historically,theproblemwasthattherewerenolearningalgorithmsfortr

3、ainingMLPs.Actually,itisnowquitestraightforward.ExpressivepowerofanMLPQuestionsHowmanyhiddenlayersareneeded?Howmanyunitsshouldbeina(the)hiddenlayer?AnswersKomogorov’smappingneuralnetworkexistencetheorem(universaltheorem)Komogorov’smappingneuralnetworkexistencetheorem(un

4、iversaltheorem)Anycontinuousfunctiong(x)definedontheunithypercubecanberepresentedinthefromForproperlychosenfunctionsandItisimpracticalthefunctionsandarenotthesimpleweightedsumspassedthroughnonlinearitiesfavoredinneuralnetworksIttellsusverylittleabouthowtofindthenonlinea

5、rfunctionsbasedondata—thecentralprobleminnetworkbasedpatternrecognitionthosefunctionscanbeextremelycomplex;theyarenotsmoothKomogorov’smappingneuralnetworkexistencetheorem(universaltheorem)Anycontinuousfunctiong(x)canbeapproximatedtoarbitraryprecisionbyforproperlychosenf

6、unctionf(.)whenNHapproachestoinfinity.MLPforclassificationMLPforregressionLearningschemeSupervisedlearningTwopropagationdirections-FunctionSignal:inforwarddirection-Errorsignal:inbackwarddirectionLearninginMLPObjectivefunctionwhereThedesiredoutputofthejthoutputneuronThe

7、realoutputofthejthoutputneuronSumsquarederrorfunctionSteepestdescentsearchmethodPartialderivativeextensionLearningrateparameterSynapticweightfromithneuronink-1thlayertothejthneuroninkthlayerofthenetworkSituationforSituationforBackpropagationlearningalgorithmofMLPUpdatin

8、gequationwherewhichisBackpropagationformulaForsigmoidalfunctionf(.)wehaveSpeedingthelearningprocessLearningrat