神经网络课件.ppt

《神经网络课件.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、PerceptronsandBPnetworksLectureNoteZhaoJieYuComputerScienceDepartmentNingboUniversityPerceptronThesimplestformofaneuralnetworkconsistsofasingleneuronwithadjustablesynapticweightsandbias.performspatternclassificationwithonlytwoclassesperceptronconvergencetheorem:Patterns(vectors)aredrawnf

2、romtwolinearlyseparableclassesDuringtraining,theperceptronalgorithmconvergesandpositionsthedecisionsurfaceintheformofhyperplanebetweentwoclassesbyadjustingsynapticweightsWithmorethanoneneuronatoutputlayer,wemaycorrespondinglyformclassificationwithmorethantwoclasseswhicharelinearlyseparab

3、leInputsignalSynapticweightsSummingfunctionBiasbActivationfunctionLocalFieldvOutputyx1x2xmw2wmw1McCulloch—Pitts（M-P）模型TypesofActivationFunctioniniOj+1iniOj+1tThresholdFunction阈值函数Piecewise-linearFunction非线性斜面函数SigmoidFunctionS形函数(Logistic)iniOj+1t对称性激活函数vi+1-1SignumFunctionvi+1Hyperbolic

4、tangentFunctionNetworkArchitecture网络结构Single-layerFeedforwardNetworks（单层前向）inputlayerandoutputlayersingle(computation)layerfeedforward,acyclicMultilayerFeedforwardNetworks（多层前向）hiddenlayers-hiddenneuronsandhiddenunitsenablestoextracthighorderstatistics10-4-2network,100-30-10-3networkfull

5、yconnectedlayerednetworkRecurrentNetworks（反馈网络）atleastonefeedbackloopwithorwithouthiddenneuronSinglelayerMultiplelayerfullyconnectedRecurrentnetworkwithouthiddenunitsinputsoutputsRecurrentnetworkwithhiddenunitsUnitdelayoperator不同的网络结构感知机210.50.8b=-1w1x1+w2x2+b>0?1,02*(0.5)+1*(0.8)+(-1)=0

6、.8>0,O=10.5x1+0.8x2-1=0决策线OXORProblem:InputsOutputx1x2y000011101110W?B?XORnetwork:output0/1typeLearningProcedure:Randomlyassignweights(between0-1)PresentinputsfromtrainingdataGetoutputO,nudgeweightstogivesresultstowardourdesiredoutputTRepeat;stopwhennoerrors,orenoughepochscompletedPercep

7、tron(I)GoalclassifyingappliedInputintooneoftwoclassesProcedureifoutputofhardlimiteris+1,toclassifitis-1,toclass(±1type)inputofhardlimiter:weightedsumofinputeffectofbiasbismerelytoshiftdecisionboundaryawayfromoriginsynapticweightsadaptedoniterationbyiterationbasisPerceptro