Self-organizing map

Definition

Itsimulatesthedifferentdivisionoflaborofnervecellsindifferentareasofthehumanbrain,thatis,differentareashavedifferentresponsecharacteristics,andthisprocessisdoneautomatically.Theself-organizingmappingnetworkclassifiesthesetofinputpatternsbyfindingtheoptimalsetofreferencevectors.Eachreferencevectorisaconnectionweightvectorcorrespondingtoanoutputunit.Comparedwiththetraditionalpatternclusteringmethod,theclustercenterformedbyitcanbemappedtoacurvedsurfaceorplanewhilekeepingthetopologicalstructureunchanged.Theproblemofidentifyingunknownclustercenterscanberealizedbyself-organizingmapping.

Features

Self-organizingneuralnetworkisoneofthemostattractiveresearchfieldsofneuralnetwork.Itcanlearntodetectitsregularityandtherelationshipbetweeninputsamplesthroughitsinputsamples,Andadjustthenetworkadaptivelyaccordingtotheinformationoftheseinputsamples,sothatthefutureresponseofthenetworkadaptstotheinputsamples.Theneuronsofthecompetitiveneuralnetworkcanrecognizegroupsofsimilarinputvectorsthroughinputinformation;theself-organizingmapneuralnetworkcanalsorecognizegroupsofsimilarinputvectorsthroughlearning,sothatthepairsofneuronsinthenetworklayerthatareveryclosetoeachotherSimilarinputvectorsproducearesponse.Differentfromcompetitiveneuralnetworks,self-organizingmapneuralnetworkscannotonlylearnthedistributionofinputvectors,butalsolearnthetopologicalstructureofinputvectors.Itssingleneurondoesnotplayadecisiveroleinpatternclassification,anditdependsonmultipleneurons.Thesynergyofmeta-unitscancompletethepatternclassification.

Role

Learningvectorquantization(LVQ)isasupervisedlearningmethodfortrainingthecompetitivelayer.Thecompetitionlayerneuralnetworkcanautomaticallylearntheclassificationofinputvectorpatterns,buttheclassificationperformedbythecompetitionlayeronlydependsonthedistancebetweentheinputvectors.Whentwoinputvectorsareveryclose,thecompetitionlayermayclassifythemintoonecategory.Inthedesignofthecompetitionlayer,thereisnosuchmechanismtojudgewhetheranytwoinputvectorsbelongtothesameclassordifferentclassesstrictlyaccordingtotheregion.FortheLVQnetworkuser-specifiedtargetclassificationresult,thenetworkcancompletetheaccurateclassificationoftheinputvectorpatternthroughsupervisedlearning.