Abstract. It is well known, that online neural gas (NG) possesses a magnification exponent different from the information theoretically optimum one in adaptive map formation. The exponent can explicitely be controlled by a small change of the learning algo
Magni cationcontrolforbatchneuralgas
BarbaraHammer1,AlexanderHasenfuß1,andThomasVillmann2
1-ClausthalUniversityofTechnology,InstituteofComputerScience,
Clausthal-Zellerfeld,Germany,{hammer|hasenfuss}@in.tu-clausthal.de
2-UniversityofLeipzig-ClinicforPsychotherapyandPsychosomatic
Medicine,Leipzig,Germany,villmann@informatik.uni-leipzig.de
Abstract.Itiswellknown,thatonlineneuralgas(NG)possessesamag-ni cationexponentdi erentfromtheinformationtheoreticallyoptimum
oneinadaptivemapformation.Theexponentcanexplicitelybecontrolled
byasmallchangeofthelearningalgorithm.BatchNGconstitutesafast
alternativeoptimizationschemeforNGvectorquantizerswhichpossesses
thesamemagni cationfactorasstandardonlineNG.Inthispaper,we
proposeamethodtointegratemagni cationcontrolbylocallearninginto
batchNGbylinkingmagni cationcontroltoanunderlyingcostfunction.
Wevalidatethelearningruleinanexperimentalsetting.
1Introduction
Vectorquantizationconstitutesanimportanttechnicalproblemindi erentar-easofapplicationsuchasdatamining,control,imagecompression,orinfor-mationrepresentation.Thereby,thetasksarediversesuchasminimizingthequantizationerror,optimuminformationtransfer,classi cation,visualization,ortopographicmapformation.Self-organizingquantizationprocessesareacom-monpropertyofmanyregionsofthebrain,includingthevisual,auditory,andsomatosensoriccortex.Neuralgas(NG)asproposedin[9]constitutesapar-ticularlyrobustvectorquantizationmethodwhichdynamicscanbeinterpretedasanoverlayofstandardvectorquantizationanddi usion.Itcanbeextendedtowardstopographicmapswithdataoptimumtopology[10].
Acharacteristicpropertyofvectorquantizersconsistsinaselectivemagni- cationofregionsofinterest.Thiscorrespondstoaspeci cconnectionofthedensityofprototypesandstimuli.Aninformationtheoreticallyoptimummagni- cationfactoronecorrespondstoanexactadjustmentoftheprototypesaccord-ingtotheunderlyingdatadistribution.Thismagni cationlawisachievedbyapproacheswhichexplicitelyoptimizetheinformationtransferorrelatedquanti-ties[8].Foravarietyofpopularalternatives,however,themagni cationfollowsapowerlawwithexponentdi erentfromone[3,9,16].Startingwiththework
[1],schemestocontrolthemagni cationfactorhavebeenproposedintheliter-ature,forrecentresultsseee.g.[13].Magni cationcontrolchangesthelearningschemeandallowstoachieveamagni cationfactoroneorbeyond.Anexplicitcontrolisparticularlyinterestingforapplicationareaswhererareeventsshouldbesuppressedor,contrarily,emphasized,whichhasprovenbene cialinseveraltasksinroboticsandimageinspection[11,15,14].Inaddition,thise ectcor-respondstobiologicalphenomenasuchastheperceptualmagnete ectwhichleadstoanemphasisofrarelyoccurringstimuli[5,7].Themagni cationfactorofonlineNGisD/(D+2),Dbeingtheintrinsic(Hausdor -)dimensionofthedatamanifoldofstimuli.Thiscanbecontrolledusinge.g.locallearning,whichchangesthelearningratebyafactordependingonthelocaldatadensity[12].
7
Magnification control in winner relaxing neural gas Abstract. We transfer the idea of winner relaxing learning from the self-organizing map to the neural ...
Neural Gas for Surface R... 暂无评价 10页 免费 Magnification control fo... 6页 免费 A Recruiting Neural-Gas ... 暂无评价 6页 免费 Double Growing...
presents systematic SOM simulations with the explicit magnification control ... ESANN'2004 proceedings - European Symposium on Artificial Neural Networks ...
Neural Gas for Sequences 暂无评价 6页 免费 Neural Gas for Surface R... 暂无评价 10页 免费 Magnification control fo... 6页 免费 A Recruiting Neural...
Magnification control fo... 6页 免费 Relating Sampling Period... 暂无评价...Hierarchical supervisory control for batch processes. Technical report, Control ...
Recent extensions -Insensitivity of initialization: batch neural gas -Optimum information transfer: magnification control -Incorporating label information: ...
Recent extensions -Insensitivity of initialization: batch neural gas -Optimum information transfer: magnification control -Incorporating label information: ...
热门文档