2、htscoefficientsandinputsignals.Thesecondunitrealisenonlinearfunction,calledneuronactivationfunction.Signal e isadderoutputsignal,and y=f(e) isoutputsignalofnonlinearelement.Signal y isalsooutputsignalofneuron. Toteachtheneuralnetworkweneedtrainingdataset.Thetrainingd
3、atasetconsistsofinputsignals(x1 and x2 )assignedwithcorrespondingtarget(desiredoutput) z.Thenetworktrainingisaniterativeprocess.Ineachiterationweightscoefficientsofnodesaremodifiedusingnewdatafromtrainingdataset.Modificationiscalculatedusingalgorithmdescribedbelow:Ea
5、t xm andneuron n ininputlayer.Symbols yn representsoutputsignalofneuron n. Propagationofsignalsthroughthehiddenlayer.Symbols wmn representweightsofconnectionsbetweenoutputofneuron m andinputofneuron n inthenextlayer. Propagationofsignalsthroughtheoutputlayer. Inthene
6、xtalgorithmsteptheoutputsignalofthenetwork y iscomparedwiththedesiredoutputvalue(thetarget),whichisfoundintrainingdataset.Thedifferenceiscallederrorsignal d ofoutputlayerneuron. Itisimpossibletocomputeerrorsignalforinternalneuronsdirectly,becauseoutputvaluesofthesene
7、uronsareunknown.Formanyyearstheeffectivemethodfortrainingmultiplayernetworkshasbeenunknown.Onlyinthemiddleeightiesthebackpropagationalgorithmhasbeenworkedout.Theideaistopropagateerrorsignal d (computedinsingleteachingstep)backtoallneurons,whichoutputsignalswereinputf