Home Tekniikka Keinotekoiset neuroverkot

Keinotekoiset neuroverkot



Neuroni

Kuten kuvassa näkyy

a1~ana ovat tulovektorin komponentit

w1~osti neuronkosketuksen painon

bistheoffset

fisthetransferfunction,yleensäononlineaarinen funktio.Seuraava oletusarvo onhardlim()

tistheneuronoutput

Matemaattinen esityst=f(WA'+b)

Witheweightvektori

Aisthe-tulovektori,A'onvektorin transponointi

bistheoffset

fisthetransferfunction

Visible,anerveThefunctionoftheelementistoobtainascalarresultthroughanonlineartransferfunctionafterobtainingtheinnerproductoftheinputvectorandtheweightvector.

Theroleofasingleneuron:divideann-dimensionalvectorspacewithahyperplaneintotwoparts(calledthejudgmentboundary),givenaninputvector,theneuroncanjudgethatthisvectorislocatedinthehyperplane.Whichsideoftheplane.

Hypertason yhtälö: Wp+b=0

Wpainovektori

boffset

ponthehypertasovektori

Basicfeatures

Artificialneuralnetworkisanonlinear,adaptiveinformationprocessingsystemcomposedofalargenumberofinterconnectedprocessingunits.Itisproposedonthebasisoftheresultsofmodernneuroscienceresearch,andattemptstoprocessinformationbysimulatingtheprocessingandmemoryinformationofthebrain'sneuralnetwork.Artificialneuralnetworkhasfourbasiccharacteristics:

(1)Non-linearityNon-linearrelationshipisauniversalcharacteristicofnature.Thewisdomofthebrainisanon-linearphenomenon.Artificialneuronsareintwodifferentstatesofactivationorinhibition.Thisbehaviorismathematicallyexpressedasanonlinearrelationship.Anetworkcomposedofthresholdneuronshasbetterperformance,andcanimprovefaulttoleranceandstoragecapacity.

(2)Non-limitationAneuralnetworkisusuallycomposedofmultipleneuronsconnectedextensively.Theoverallbehaviorofasystemnotonlydependsonthecharacteristicsofasingleneuron,butalsomaybemainlydeterminedbytheinteractionandinterconnectionbetweentheunits.Simulatethenon-limitationofthebrainthroughalargenumberofconnectionsbetweenunits.Associativememoryisatypicalexampleofnon-limitation.

(3)VeryqualitativeArtificialneuralnetworkhastheabilityofself-adaptation,self-organization,andself-learning.Notonlycantheinformationprocessedbytheneuralnetworkhavevariouschanges,butwhileprocessingtheinformation,thenonlineardynamicsystemitselfisalsoconstantlychanging.Iterativeprocessisoftenusedtodescribetheevolutionprocessofthedynamicsystem.

(4)Non-convexityTheevolutiondirectionofasystemwilldependonaspecificstatefunctionundercertainconditions.Forexample,theenergyfunction,itsextremevaluecorrespondstotherelativelystablestateofthesystem.Non-convexitymeansthatthisfunctionhasmultipleextremevalues,sothesystemhasmultiplestableequilibriumstates,whichwillleadtothediversityofsystemevolution.

Inartificialneuralnetworks,neuronprocessingunitscanrepresentdifferentobjects,suchasfeatures,letters,concepts,orsomemeaningfulabstractpatterns.Therearethreetypesofprocessingunitsinthenetwork:inputunits,outputunitsandhiddenunits.Theinputunitacceptssignalsanddatafromtheexternalworld;theoutputunitrealizestheoutputofthesystem'sprocessingresults;thehiddenunitisaunitthatliesbetweentheinputandoutputunitsandcannotbeobservedfromtheoutsideofthesystem.Theconnectionweightbetweenneuronsreflectstheconnectionstrengthbetweenunits,andtherepresentationandprocessingofinformationisreflectedintheconnectionrelationshipofnetworkprocessingunits.Artificialneuralnetworkisakindofnon-programmed,adaptive,andbrain-styleinformationprocessing.Itsessenceistoobtainaparallelanddistributedinformationprocessingfunctionthroughnetworktransformationanddynamicbehavior,andtoimitatepeopleindifferentdegreesandlevels.Theinformationprocessingfunctionofthecranialnervoussystem.Itisaninterdisciplinarysubjectinvolvingneuroscience,thinkingscience,artificialintelligence,computerscienceandmanyotherfields.

Artificialneuralnetworkisaparalleldistributedsystem.Itadoptsacompletelydifferentmechanismfromtraditionalartificialintelligenceandinformationprocessingtechnology.Itovercomesthetraditionalartificialintelligencebasedonlogicalsymbolsinprocessingintuitiveandunstructuredinformation.Theshortcomingsofithavethecharacteristicsofself-adaptation,self-organizationandreal-timelearning.

Developmenthistory

In1943,psychologistW.S.McCullochandmathematicallogicianW.Pittsestablishedneuralnetworksandmathematicalmodels,calledMPmodels.TheyproposedaformalmathematicaldescriptionofneuronsandanetworkstructuremethodthroughtheMPmodel,andprovedthatasingleneuroncanperformlogicalfunctions,thusopeningtheeraofartificialneuralnetworkresearch.In1949,psychologistsproposedtheideaof​​variablesynapticconnectionstrength.Inthe1960s,artificialneuralnetworkswerefurtherdeveloped,andmorecompleteneuralnetworkmodelswereproposed,includingperceptronsandadaptivelinearcomponents.Aftercarefullyanalyzingthefunctionsandlimitationsoftheneuralnetworksystemrepresentedbytheperceptron,M.Minskyandotherspublishedthebook"Perceptron"in1969,pointingoutthattheperceptroncannotsolvetheproblemofhigher-orderpredicates.Theirargumentsgreatlyaffectedtheresearchofneuralnetworks,andtheachievementsofserialcomputersandartificialintelligenceatthattimeobscuredthenecessityandurgencyofdevelopingnewtypesofcomputersandnewwaysofartificialintelligence,makingtheresearchofartificialneuralnetworksatalowebb..Duringthisperiod,someartificialneuralnetworkresearchersarestillcommittedtothisresearch,andproposedadaptiveresonancetheory(ARTnetwork),self-organizingmap,cognitivemachinenetwork,andconductedresearchonneuralnetworkmathematicaltheory.Theaboveresearchhaslaidthefoundationfortheresearchanddevelopmentofneuralnetworks.In1982,J.J.Hopfield,aphysicistattheCaliforniaInstituteofTechnologyintheUnitedStates,proposedtheHopfieldneuralgridmodel,introducedtheconceptof"computationalenergy",andgaveajudgmentonnetworkstability.In1984,healsoproposedthecontinuous-timeHopfieldneuralnetworkmodel,whichmadepioneeringworkfortheresearchofneuralcomputers,andcreatedanewwayforneuralnetworkstobeusedforassociativememoryandoptimizedcalculations,whichstronglypromotedtheresearchofneuralnetworks.In1985,anotherscholarputforwardtheBoltzmannmodel,andusedstatisticalthermodynamicssimulatedannealingtechnologyinthestudytoensurethattheentiresystemtendstoagloballystablepoint.In1986,heconductedcognitivemicrostructureresearchandproposedthetheoryofparalleldistributedprocessing.In1986,Rumelhart,Hinton,WilliamsdevelopedtheBPalgorithm.RumelhartandMcClellandpublished"Paralleldistributionprocessing:explorationsinthemicrostructuresofcognition".Sofar,theBPalgorithmhasbeenusedtosolvealargenumberofpracticalproblems.In1988,Linskerproposedanewself-organizationtheoryforperceptronnetworks,andformedtheMaximumMutualInformationTheoryonthebasisofShanon'sinformationtheory,thusignitingthelightofNN-basedinformationapplicationtheory.In1988,BroomheadandLoweusedaradialbasisfunction(RBF)toproposeahierarchicalnetworkdesignmethod,thuslinkingthedesignofNNwithnumericalanalysisandlinearadaptivefiltering.Intheearly1990s,Vapniketal.proposedtheconceptofsupportvectormachines(SVM)andVC(Vapnik-Chervonenkis)dimensions.Theresearchofartificialneuralnetworkshasattractedtheattentionofvariousdevelopedcountries.TheUSCongresspassedaresolutiontodesignatethedecadebeginningonJanuary5,1990asthe"decadeofthebrain.""Year"becomesaglobalbehavior.InJapan's"RealWorldComputing(RWC)"project,theresearchofartificialintelligencehasbecomeanimportantcomponent.

Networkmodel

Theartificialneuralnetworkmodelmainlyconsidersthetopologicalstructureofthenetworkconnection,thecharacteristicsoftheneurons,andthelearningrules.Atpresent,therearenearly40neuralnetworkmodels,includingbackpropagationnetworks,perceptrons,self-organizingmaps,Hopfieldnetworks,Boltzmannmachines,adaptiveresonancetheory,etc.Accordingtothetopologicalstructureoftheconnection,theneuralnetworkmodelcanbedividedinto:

Forwardnetwork

Eachneuroninthenetworkacceptstheinputofthepreviouslevelandoutputstothenextlevel,thenetworkThereisnofeedbackinit,whichcanberepresentedbyadirectedacyclicgraph.Thiskindofnetworkrealizesthetransformationofthesignalfromtheinputspacetotheoutputspace,anditsinformationprocessingabilitycomesfromthemultiplerecombinationofsimplenon-linearfunctions.Thenetworkstructureissimpleandeasytoimplement.Thebackhaulnetworkisatypicalforwardnetwork.

Feedbacknetwork

Thereisfeedbackbetweenneuronsinthenetwork,whichcanberepresentedbyanundirectedcompletegraph.Theinformationprocessingofthiskindofneuralnetworkisthetransformationofstate,whichcanbeprocessedbydynamicsystemtheory.Thestabilityofthesystemiscloselyrelatedtotheassociativememoryfunction.HopfieldnetworkandBoltzmannmachinebelongtothistype.

Typesoflearning

Learningisanimportantcontentofneuralnetworkresearch,anditsadaptabilityisachievedthroughlearning.Accordingtochangesintheenvironment,theweightsareadjustedtoimprovethebehaviorofthesystem.TheHebblearningrulesproposedbyHebblaythefoundationfortheneuralnetworklearningalgorithm.Hebb'srulebelievesthatthelearningprocessultimatelyoccursatthesynapsesbetweenneurons,andthestrengthofsynapseschangeswiththeactivityofneuronsbeforeandaftersynapses.Onthisbasis,peoplehaveproposedvariouslearningrulesandalgorithmstomeettheneedsofdifferentnetworkmodels.Theeffectivelearningalgorithmenablestheneuralnetworktoconstructaninternalrepresentationoftheobjectiveworldthroughtheadjustmentofconnectionweights,andformauniqueinformationprocessingmethod.Informationstorageandprocessingarereflectedinthenetworkconnection.

Luokittelu

Accordingtodifferentlearningenvironments,neuralnetworklearningmethodscanbedividedintosupervisedlearningandunsupervisedlearning.Insupervisedlearning,thedataoftrainingsamplesisaddedtotheinputofthenetwork,andthecorrespondingexpectedoutputiscomparedwiththeoutputofthenetworktoobtainanerrorsignaltocontroltheadjustmentoftheweightconnectionstrength.Aftermultipletrainings,itconvergestooneDeterminedweight.Whenthesamplesituationchanges,theweightcanbemodifiedafterlearningtoadapttothenewenvironment.Neuralnetworkmodelsthatusesupervisedlearningincludebackpropagationnetworksandperceptrons.Inunsupervisedlearning,nostandardsamplesaregiveninadvance,andthenetworkisdirectlyplacedintheenvironment,andthelearningphaseandtheworkingphaseareintegrated.Atthistime,thechangeofthelearninglawobeystheevolutionequationoftheconnectionweight.ThesimplestexampleofunsupervisedlearningistheHebblearningrules.Competitivelearningrulesareamorecomplexexampleofunsupervisedlearning,whichadjustsweightsbasedonestablishedclusters.Self-organizingmapping,adaptiveresonancetheorynetwork,etc.arealltypicalmodelsrelatedtocompetitivelearning.

Analysismethod

Tostudythenonlineardynamicpropertiesofneuralnetworks,mainlyusingdynamicsystemtheory,nonlinearprogrammingtheoryandstatisticaltheorytoanalyzetheevolutionandattractionofneuralnetworksThenatureofthechild,explorethecooperativebehaviorandcollectivecomputingfunctionofneuralnetworks,andunderstandthemechanismofneuralinformationprocessing.Inordertoexplorethepossibilityofneuralnetworkprocessinginformationintermsofintegrityandambiguity,theconceptsandmethodsofchaostheorywillplayarole.Chaosisamathematicalconceptthatisquitedifficulttodefineprecisely.Generallyspeaking,"chaos"referstothenon-deterministicbehaviorinthedynamicsystemdescribedbythedeterministicequation,ordeterministicrandomness."Determinism"isbecauseitisproducedbyinternalreasonsratherthanexternalnoiseorinterference,while"randomness"referstoitsirregularandunpredictablebehavior,whichcanonlybedescribedbystatisticalmethods.Themaincharacteristicofachaoticdynamicsystemisitsstate'ssensitivedependenceoninitialconditions,andchaosreflectsitsinherentrandomness.Chaostheoryreferstothebasictheories,concepts,andmethodsthatdescribethenonlineardynamicsystemwithchaoticbehavior.Itunderstandsthecomplexbehaviorofthedynamicsystemasitsinherentexistenceintheexchangeofmatter,energy,andinformationwiththeoutsideworld.Structuralbehavior,notforeignandaccidentalbehavior,thechaoticstateisasteadystate.Thestationarystateofachaoticdynamicsystemincludes:static,stationaryquantity,periodicity,quasi-simultaneityandchaoticsolution.Thechaotictrajectoryistheresultofthecombinationofoverallstabilityandlocalinstability,andiscalledasingularattractor.Asingularattractorhasthefollowingcharacteristics:(1)asingularattractorisanattractor,butitisneitherafixedpointnoraperiodicsolution;(2)asingularattractorisindivisible,thatis,itcannotbedividedintotwoAndtwoormoreattractors;(3)Itisverysensitivetotheinitialvalue,anddifferentinitialvalues​​willleadtoverydifferentbehaviors.

Ominaisuudet ja edut

Tehollisten hermoverkkojen ominaisuudet ja edut ilmenevät pääasiassa kolmesta näkökulmasta:

First,ithasaself-learningfunction.Forexample,whenimplementingimagerecognition,onlyinputmanydifferentimagetemplatesandcorrespondingrecognitionresultsintotheartificialneuralnetwork,andthenetworkwillslowlylearntorecognizesimilarimagesthroughitsself-learningfunction.Theself-learningfunctionisparticularlyimportantforprediction.Itisexpectedthattheartificialneuralnetworkcomputerinthefuturewillprovidehumanbeingswitheconomicforecasts,marketforecasts,andprofitforecasts,anditsapplicationprospectsareverypromising.

Second,ithasLenovostoragefunction.Thiskindofassociationcanberealizedwiththefeedbacknetworkofartificialneuralnetwork.

Thirdly,ithastheabilitytofindoptimalsolutionsatahighspeed.Findinganoptimizedsolutiontoacomplexproblemoftenrequiresalotofcalculations.Usingafeedbackartificialneuralnetworkdesignedforacertainproblemandusingthecomputer'shigh-speedcomputingcapabilities,itmaybepossibletoquicklyfindanoptimizedsolution.

Tutkimuksen suunta

Neuroverkon tutkimus voidaan jakaa kahteen osa-alueeseen: teoreettiseen tutkimukseen ja soveltavaan tutkimukseen.

Theoreticalresearchcanbedividedintothefollowingtwocategories:

1.Usingneurophysiologicalandcognitivescientificresearchonhumanthinkingandintelligencemechanisms.

2.Usetheresearchresultsofneuralbasictheories,usemathematicalmethodstoexploreneuralnetworkmodelswithmorecompletefunctionsandsuperiorperformance,andin-depthstudyofnetworkalgorithmsandperformance,suchas:stability,convergence,andfaulttolerance,Robustness,etc.;developnewnetworkmathematicaltheories,suchasneuralnetworkdynamics,nonlinearneuralfields,etc.

Applicationresearchcanbedividedintothefollowingtwocategories:

1.Researchonsoftwaresimulationandhardwarerealizationofneuralnetworks.

2.Researchontheapplicationofneuralnetworksinvariousfields.Thesefieldsmainlyinclude

:kuvioiden tunnistus,signaalinkäsittely,tietämyssuunnittelu,asiantuntijajärjestelmät,optimoitu yhdistelmä,robottiohjaus jne.Neuraaliverkkoteorian jatkuvan kehittämisen, liittyvien teorioiden ja tekniikoiden ansiosta hermoverkkojen sovellus tulee varmasti syvemmälle.

Developmenttrend

Theuniquenonlinearadaptiveinformationprocessingcapabilitiesofartificialneuralnetworksovercometheintuitionoftraditionalartificialintelligencemethods,suchaspattern,speechrecognition,andunstructuredinformationprocessing.Theshortcomingsof,makeitsuccessfullyappliedinthefieldsofneuralexpertsystem,patternrecognition,intelligentcontrol,combinationoptimization,predictionandsoon.Thecombinationofartificialneuralnetworkandothertraditionalmethodswillpromotethecontinuousdevelopmentofartificialintelligenceandinformationprocessingtechnology.Inrecentyears,artificialneuralnetworksaredevelopingmoredeeplyontheroadofsimulatinghumancognition,combiningwithfuzzysystems,geneticalgorithms,evolutionarymechanisms,etc.,toformcomputationalintelligence,becominganimportantdirectionofartificialintelligence,andwillbedevelopedinpracticalapplications..Theapplicationofinformationgeometrytotheresearchofartificialneuralnetworkshasopenedupanewwayforthetheoreticalresearchofartificialneuralnetworks.Theresearchofneuralcomputershasdevelopedrapidly,andproductshaveenteredthemarket.Thephotoelectriccombinedneuralcomputerprovidesgoodconditionsforthedevelopmentofartificialneuralnetworks.

Neuralnetworkshavebeenwellappliedinmanyfields,buttherearestillmanyaspectsthatneedtobestudied.Amongthem,thecombinationofneuralnetworkswiththeadvantagesofdistributedstorage,parallelprocessing,self-learning,self-organization,andnonlinearmapping,andothertechnologies,andtheresultinghybridmethodsandhybridsystems,havebecomeamajorresearchhotspot.Sinceothermethodsalsohavetheirownadvantages,combiningneuralnetworkswithothermethodscanlearnfromeachother'sstrengths,andthenobtainbetterapplicationeffects.Thecurrentworkinthisareaincludesthefusionofneuralnetworksandfuzzylogic,expertsystems,geneticalgorithms,waveletanalysis,chaos,roughsettheory,fractaltheory,evidencetheoryandgreysystems.

Seuraavassa analysoidaan pääasiassa hermoverkon ja aaltoanalyysin, kaaoksen, karkeasarjateorian ja fraktaaliteorian fuusiota.

jaaaltoanalyysin yhdistelmä

In1981,theFrenchgeologistMorletwasseekingForgeologicaldata,throughcreativeresearchonthesimilaritiesanddifferences,characteristicsandfunctionstructureofFouriertransformandwindowedFouriertransform,theconceptof"waveletanalysis"wasfirstproposedandtheMorletwaveletnamedafterhimwasestablished.Since1986,duetothefoundationworkofYMeyer,S.MallatandIDaubechies,waveletanalysishasrapidlydevelopedintoanemergingdiscipline."WaveletsandOperators"byMeyerand"TenLecturesonWavelets"byDaubechiesarethemostauthoritativeworksinthefieldofwaveletresearch.

WavelettransformisabreakthroughinFourieranalysismethod.Itnotonlyhasgoodlocalizationpropertiesinboththetimedomainandthefrequencydomain,butalsohasgoodresolutionforlow-frequencysignalsinthefrequencydomainandforhigh-frequencysignalsinthetimedomain,sothatitcangatheranydetailsoftheobject.Waveletanalysisisequivalenttoamathematicalmicroscope,withfunctionsofzoomingin,zoomingout,andpanning.Itcanstudythedynamiccharacteristicsofthesignalbycheckingthechangesunderdifferentmagnifications.Therefore,waveletanalysishasbecomeapowerfultoolinmanyfieldssuchasgeophysics,signalprocessing,imageprocessing,andtheoreticalphysics.

Waveletneuralnetworkcombinesthegoodtime-frequencylocalizationcharacteristicsofwavelettransformwiththeself-learningfunctionofneuralnetwork,soithasstrongapproximationabilityandfaulttolerance.Inthecombinationmethod,thewaveletfunctioncanbeusedasthebasisfunctiontoconstructtheneuralnetworktoformthewaveletnetwork,orthewavelettransformcanbeusedastheinputpre-processingtoolofthefeedforwardneuralnetwork,thatis,theprocessstatesignalisprocessedwiththemulti-resolutioncharacteristicsofthewavelettransform.Realizethesignal-to-noiseseparation,andextractthestatecharacteristicsthathavethegreatestimpactontheprocessingerror,astheinputoftheneuralnetwork.

Waveletneuralnetworkhasapplicationsinmotorfaultdiagnosis,high-voltagepowergridfaultsignalprocessingandprotectionresearch,bearingandothermechanicalfaultdiagnosis,andmanyotheraspects.Waveletneuralnetworkisusedforintelligentcontrolofinductionservomotors.Thesystemhasgoodtrackingcontrolperformanceandgoodrobustness.Ituseswaveletpacketneuralnetworkforintelligentdiagnosisofcardiovasculardiseases,waveletlayerperformsadaptivefeatureextractionintimeandfrequencydomain,andforwardneuralnetworkisusedforclassification.Thecorrectclassificationratereached94%.

Whilewaveletneuralnetworkisusedinmanyaspects,itstillhassomeshortcomings.Startingfromtherequirementsofextractionaccuracyandreal-timewavelettransform,itisnecessarytoconstructsomespecialwaveletbasesadaptedtotheapplicationrequirementsaccordingtotheactualsituationinordertoachievebetterresultsintheapplication.Inaddition,thereal-timerequirementsintheapplicationalsoneedtocombinethedevelopmentofDSPanddevelopspecialprocessingchipstomeetthisrequirement.

Chaosneuralnetwork

ThefirstdefinitionofchaoswasfirstproposedbyLi-Yorkeinthe1970s.Becauseofitswideapplicationvalue,ithasreceivedwidespreadattentionfromallaspectssinceitsappearance.Chaosisanirregularmovementthatappearsinacertainsystem.Chaosisarelativelycommonphenomenonthatexistsinnonlinearsystems.Chaoticmovementhasthecharacteristicsofergodicityandrandomness,whichcanbeadjustedwithinacertainrange.Ittraversesallstateswithoutrepeatingitsownlaws.Thechaostheorydeterminesthechaosofnonlineardynamics.Thepurposeistorevealthesimplelawsthatmaybehiddenbehindseeminglyrandomphenomena,inordertodiscoverthecommonlawsthatalargeclassofcomplexproblemsgenerallyfollow.

In1990,Kaihara,T.Takabe,andM.Toyodaetal.firstproposedachaoticneuralnetworkmodelbasedonthechaoticcharacteristicsofbiologicalneurons,andintroducedchaosintotheneuralnetwork,makingtheartificialneuralnetworkhavechaoticbehavior.Itisclosertotheactualhumanbrainneuralnetwork,sothechaoticneuralnetworkisconsideredtobeoneoftheintelligentinformationprocessingsystemsthatcanrealizeitsreal-worldcalculations,andithasbecomeoneofthemainresearchdirectionsofneuralnetworks.

ComparedwiththeconventionaldiscreteHopfieldneuralnetwork,thechaoticneuralnetworkhasrichernonlineardynamiccharacteristics,mainlyasfollows:theintroductionofchaoticdynamicbehaviorintotheneuralnetwork;thesynchronizationofthechaoticneuralnetworkCharacteristics;attractorsofchaoticneuralnetworks.

Intheactualapplicationofneuralnetwork,whentheinputofthenetworkchangesgreatly,theinherentfaulttoleranceoftheapplicationnetworkisofteninsufficient,andamnesiaoftenoccurs.Thedynamicmemoryofchaoticneuralnetworkbelongstothedeterministicdynamicmotion.Thememoryoccursonthetrajectoryofthechaoticattractor.Throughcontinuousmotion(recallprocess),thememorypatternisassociatedonebyone,especiallyforthosestatespacedistributionsthatarecloseorpartiallyoverlapped.Thememorymodelofthechaoticneuralnetworkcanalwaysbereproducedandidentifiedthroughdynamicassociativememorywithoutconfusion.Thisistheuniqueperformanceofthechaoticneuralnetwork,whichwillgreatlyimprovethememoryabilityoftheHopfieldneuralnetwork.Theattractiondomainofthechaoticattractorexists,whichformstheinherentfault-tolerantfunctionofthechaoticneuralnetwork.Thiswillplayanimportantroleincomplexpatternrecognition,imageprocessingandotherengineeringapplications.

Anotherreasonfortheattentionofchaoticneuralnetworksisthatchaosexistsintherealneuronsandneuralnetworksoforganismsandplaysacertainrole.Theelectrophysiologicalexperimentsofzoologyhaveconfirmedthis.

Becauseofitscomplexdynamiccharacteristics,chaoticneuralnetworkshavereceivedgreatattentioninthefieldsofdynamicassociativememory,systemoptimization,informationprocessing,andartificialintelligence.Aimingatthechaoticneuralnetworkwithassociativememoryfunction,butitssearchprocessisunstable,acontrolmethodisproposedtocontrolthechaoticphenomenoninthechaoticneuralnetwork.Theapplicationofchaoticneuralnetworkincombinatorialoptimizationproblemisstudied.

Inordertobetterapplythedynamiccharacteristicsofthechaoticneuralnetworkandeffectivelycontrolthechaoticphenomenonthatexists,itisstillnecessarytofurtherimproveandadjustthestructureofthechaoticneuralnetwork,aswellasthechaoticneuralnetwork.Furtherresearchonnetworkalgorithms.

BasedonRoughSetTheory

TheRoughSetstheorywasfirstproposedbyProfessorZ.PawlakoftheWarsawUniversityofTechnologyin1982.ItisaAnalyzethemathematicaltheoryofdata,andstudythemethodsofexpression,learning,andinductionofincompletedataandinaccurateknowledge.Roughsettheoryisanewmathematicaltoolfordealingwithfuzzyanduncertainknowledge.Itsmainideaistoderivethedecision-makingorclassificationrulesoftheproblemthroughknowledgereductionunderthepremiseofkeepingtheclassificationabilityunchanged.Atpresent,roughsettheoryhasbeensuccessfullyappliedinthefieldsofmachinelearning,decisionanalysis,processcontrol,patternrecognitionanddatamining.

Thecommonpointofroughsetandneuralnetworkisthattheycanworkwellinthenaturalenvironment.However,theroughsettheorymethodsimulatestheabstractlogicalthinkingofhumanbeings,whiletheneuralnetworkmethodsimulatesvisualintuitivethinking.Thetwohavedifferentcharacteristics.Theroughsettheorymethodtakesasinputvariousqualitative,quantitativeormixedinformationthatisclosertohowpeopledescribethings.Themappingrelationshipbetweeninputspaceandoutputspaceissimplifiedthroughasimpledecisiontable,whichtakesintoaccountthedifferencesinknowledgeexpressionTheimportanceofattributesdetermineswhichknowledgeisredundantandwhichknowledgeisuseful.Neuralnetworkusestheideaof​​non-linearmappingandparallelprocessingmethodstoexpresstheimplicitfunctioncodingofinputandoutputrelatedknowledgewiththestructureofneuralnetworkitself.

Therearetwobigdifferencesbetweentheroughsettheorymethodandtheneuralnetworkmethodinprocessinginformation:oneisthatneuralnetworkprocessinginformationgenerallycannotsimplifythedimensionoftheinputinformationspace.Whenthespatialdimensionislarge,thenetworkisnotonlycomplexinstructure,butalsolongintrainingtime;whiletheroughsetmethodcannotonlyremoveredundantinputinformationbydiscoveringtherelationshipbetweendata,butalsosimplifytheexpressionspacedimensionofinputinformation.Thesecondisthattheroughsetmethodismoresensitivetonoiseintheprocessingofactualproblems,sotheresultsoflearninginferencewithnoise-freetrainingsamplesarenoteffectiveinnoisyenvironments.Theneuralnetworkmethodhasabetterabilitytosuppressnoiseinterference.

Sothetwoarecombined,andtheroughsetmethodisusedtopreprocesstheinformationfirst,thatis,theroughsetnetworkisusedasthefrontsystem,andthentheinformationstructureaftertheroughsetmethodispreprocessedtoformaneuralnetworkInformationprocessingsystem.Throughthecombinationofthetwo,notonlythenumberofattributesexpressedbytheinformationcanbereduced,andthecomplexityoftheneuralnetworkconstitutingthesystem,butalsohasstrongfaulttoleranceandanti-interferencecapabilities,providingapowerfultoolforhandlinguncertainandincompleteinformationway.

Atpresent,thecombinationofroughsetandneuralnetworkhasbeenappliedinspeechrecognition,expertsystem,datamining,faultdiagnosisandotherfields.Neuralnetworkandroughsetareusedforautomaticrecognitionofsoundsourcelocation,andneuralnetworkSumandroughsetsareusedintheknowledgeacquisitionofexpertsystemstoachievebetterresultsthantraditionalexpertsystems.Roughsetsareusedforprocessinguncertainandinaccuratedata,andneuralnetworksareusedforclassification.

Althoughthecombinationofroughsetandneuralnetworkhasbeenappliedinmanyfieldsofresearch,inordertomakethismethodplayagreaterrole,thefollowingissuesneedtobeconsidered:theroughsettheorymethodthatsimulateshumanabstractlogicalthinkingandTheneuralnetworkmethodofsimulatingvisualandintuitivethinkingismoreeffectivecombination;thedevelopmentoftheintegratedsoftwareandhardwareplatformofthetwoimprovesitspracticability.

Yhdistelmä fraktaaliteorian kanssa

SinceBenoitB.Mandelbrot,aprofessorintheDepartmentofMathematicsatHarvardUniversity,introducedtheconceptoffractalinthemid-1970s,fractalFractalgeometryhasdevelopedintoascientificmethodology—fractaltheory,andisknownasthecreationofanimportantstageofmathematicsinthe20thcentury.Ithasbeenwidelyusedinalmostallfieldsofnaturalsciencesandsocialsciences,andhasbecomeoneofthefrontierresearchtopicsinmanydisciplinesintheworldtoday.

Duetotherapiddevelopmentinmanydisciplines,fractalhasbecomeadisciplinethatdescribestheregularityofmanyirregularthingsinnature.Ithasbeenwidelyusedinvariousfieldssuchasbiology,geography,astronomy,andcomputergraphics.

Usingfractaltheorytoexplaintheirregular,unstableandhighlycomplexstructurephenomenainnaturecanreceivesignificantresults,andthecombinationofneuralnetworkandfractaltheorymakesfulluseofneuralnetworknon-Theadvantagesoflinearmapping,computingpower,andself-adaptationcanachievebetterresults.

Theapplicationareasoffractalneuralnetworksincludeimagerecognition,imagecoding,imagecompression,andfaultdiagnosisofmechanicalequipmentsystems.Thefractalimagecompression/decompressionmethodhastheadvantagesofhighcompressionrateandlowlossrate,butitscomputingpowerisnotstrong.Becausetheneuralnetworkhasthecharacteristicsofparalleloperation,theneuralnetworkisusedinthefractalimagecompression/decompression,whichimprovestheoriginalThecomputingpowerofthemethod.Combiningneuralnetworkandfractalfortherecognitionoffruitshape.Firstly,theirregularityofseveralfruitcontourdataisobtainedbyusingfractal,andthenthethree-layerneuralnetworkisusedtoidentifythedata,andthentheirregularityisevaluated.

Fractalneuralnetworkshaveachievedmanyapplications,buttherearestillsomeissuesworthyoffurtherstudy:thephysicalmeaningofthefractaldimension;thecomputersimulationandpracticalapplicationresearchoffractals.Withthecontinuousdeepeningofresearch,thefractalneuralnetworkwillsurelybecontinuouslyimprovedandachievebetterapplicationeffects.?

Applicationanalysis

Afterdecadesofdevelopment,neuralnetworktheoryhasachievedextensivesuccess.Thefollowingdescribestheapplicationstatusofneuralnetworksinsomefields.

ApplicationofArtificialNeuralNetworksintheInformationField

Indealingwithmanyproblems,thesourceofinformationisneithercomplete,butalsocontainsillusions.Decision-makingrulessometimescontradicteachotherandsometimeshavenorulestofollow.Thisbringsgreatdifficultiestotraditionalinformationprocessingmethods,butneuralnetworkscanhandletheseproblemswellandgivereasonablerecognitionandjudgment.

1.Tiedonkäsittely

Theproblemstobesolvedbymoderninformationprocessingareverycomplex.Artificialneuralnetworkshavethefunctionofimitatingorreplacinghumanthinking,andcanrealizeautomaticdiagnosis,Problemsolving,tosolveproblemsthattraditionalmethodscannotoraredifficulttosolve.Theartificialneuralnetworksystemhashighfaulttolerance,robustnessandself-organization.Eveniftheconnectionlineisdamagedtoahighdegree,itcanstillbeinanoptimizedworkingstate.Thisiswidelyusedinmilitarysystemelectronicequipment.application.Theexistingintelligentinformationsystemsincludeintelligentinstruments,automatictrackingandmonitoringinstrumentsystems,automaticcontrolandguidancesystems,automaticfaultdiagnosisandalarmsystems,etc.

2.Kuvion tunnistus

Patternrecognitionistheprocessingandanalysisofvariousformsofinformationthatcharacterizethingsorphenomenatodescribe,identify,classifyandexplainthingsorphenomenatheprocessof.ThetechnologyisbasedonBayesianprobabilitytheoryandShennong'sinformationtheory,andtheinformationprocessingprocessisclosertothelogicalthinkingprocessofthehumanbrain.Therearetwobasicpatternrecognitionmethods,namelystatisticalpatternrecognitionmethodsandstructuralpatternrecognitionmethods.Artificialneuralnetworkisacommonmethodinpatternrecognition.Theartificialneuralnetworkpatternrecognitionmethoddevelopedinrecentyearshasgraduallyreplacedthetraditionalpatternrecognitionmethod.Afteryearsofresearchanddevelopment,patternrecognitionhasbecomethecurrentrelativelyadvancedtechnology,whichiswidelyusedintextrecognition,voicerecognition,fingerprintrecognition,remotesensingimagerecognition,facerecognition,handwrittencharacterrecognition,industrialfaultdetection,precisionguidance,etc.aspect.

ApplicationofArtificialNeuralNetworkinMedicine

Duetothecomplexityandunpredictabilityofthehumanbodyanddisease,themanifestationofbiologicalsignalsandinformation,thelawofchange(self-changeandChangesaftermedicalintervention),itsdetectionandsignalexpression,dataandinformationanalysis,decision-makingandmanyotheraspectsareverycomplexnon-linearconnections,suitablefortheapplicationofartificialneuralnetworks.Thecurrentresearchinvolvesalmosteveryaspectfrombasicmedicinetoclinicalmedicine,mainlyusedinthedetectionandautomaticanalysisofbiologicalsignals,medicalexpertsystems,etc.

1.Biologisten signaalien havaitseminen ja analysointi

Mostmedicaltestingequipmentoutputsdataintheformofcontinuouswaveforms,whicharethebasisfordiagnosis.Artificialneuralnetworkisanadaptivedynamicsystemconnectedbyalargenumberofsimpleprocessingunits.Ithasthefunctionsofmassiveparallelism,distributedstorage,andself-organizationofadaptivelearning.Itcanbeusedtosolvebiomedicalsignalanalysisandprocessing.Problemsthataredifficultorimpossibletosolvebyconventionallaw.Theapplicationofneuralnetworkinthedetectionandprocessingofbiomedicalsignalsmainlyfocusesontheanalysisofbrainelectricalsignals,theextractionofauditoryevokedpotentialsignals,therecognitionofmyoelectricandgastrointestinalelectricalsignals,thecompressionofelectrocardiographicsignals,andtherecognitionofmedicalimages.Andprocessingetc.

2.Lääketieteellinen asiantuntijajärjestelmä

Thetraditionalexpertsystemistostoretheexpert’sexperienceandknowledgeinthecomputerintheformofrules,buildaknowledgebase,anduselogicalreasoningtoproceed.Medicaldiagnosis.However,inpracticalapplications,asthescaleofthedatabaseincreases,itwillleadtoan"explosion"ofknowledge,andtherewillalsobea"bottleneck"probleminthewayofacquiringknowledge,resultinginlowworkefficiency.Theneuralnetworkbasedonnonlinearparallelprocessinghaspointedoutanewdevelopmentdirectionfortheresearchofexpertsystem,solvedtheaboveproblemsofexpertsystem,andimprovedtheabilityofreasoning,self-organization,andself-learningofknowledge,sothatneuralnetworkisusedbymedicalexperts.Thesystemhasbeenwidelyusedanddeveloped.Intheresearchofanesthesiaandcriticalmedicineandotherrelatedfields,itinvolvestheanalysisandpredictionofmultiplephysiologicalvariables.Therearesomeundiscoveredornoevidenceofrelationshipsandphenomenainclinicaldata,signalprocessing,andautomaticdiscriminationanddetectionofinterferencesignals.,Thepredictionofvariousclinicalconditions,etc.,canbeappliedtoartificialneuralnetworktechnology.

Talousalan muodollisten hermoverkkotöiden sovellus

1. Markkinahintaennuste

Theanalysisofcommoditypricechangescanbeattributedtotheinfluenceofmarketsupplyanddemand.Comprehensiveanalysisofmanyfactors.Traditionalstatisticaleconomicsmethodsaredifficulttomakescientificpredictionsofpricechangesduetotheirinherentlimitations,whileartificialneuralnetworksareeasytodealwithincomplete,fuzzy,uncertainorunobviousdata,soartificialneuralnetworksareusedforPricepredictionhasanadvantagethattraditionalmethodscannotcompare.Startingfromthemechanismfordeterminingmarketprices,establishingamoreaccurateandreliablemodelbasedoncomplexandchangeablefactorssuchasthenumberofhouseholds,percapitadisposableincome,loaninterestrates,andurbanizationlevelsthataffectcommodityprices.Themodelcanscientificallypredictthetrendofcommoditypricesandobtainaccurateandobjectiveevaluationresults.

2.Riskassessment

Riskreferstotheeconomicorfinancialloss,naturaldamageornaturaldamagecausedbytheuncertaintyintheprocessofengaginginaparticularactivity.Possibilityofinjury.Thebestwaytopreventrisksistomakescientificpredictionsandassessmentsofrisksinadvance.Thepredictionideaof​​applyingartificialneuralnetworkistoconstructthestructureandalgorithmofthecreditriskmodelsuitablefortheactualsituationaccordingtothespecificrealisticrisksources,obtaintheriskevaluationcoefficient,andthendeterminethesolutiontotheactualproblem.Usingthismodelforempiricalanalysiscanmakeupfortheinsufficiencyofsubjectiveevaluationandachievesatisfactoryresults.

Theapplicationofartificialneuralnetworkinthecontrolfield

Theartificialneuralnetworkisoutstandingduetoitsuniquemodelstructureandinherentnonlinearsimulationcapabilities,aswellashighdegreeofself-adaptationandfaulttolerance.Characteristic,ithasbeenwidelyusedinthecontrolsystem.Basedontheframestructureofvariouscontrollers,anonlinearadaptivelearningmechanismisaddedtomakethecontrollerhavebetterperformance.Thebasiccontrolstructureincludessupervisorycontrol,directinversemodelcontrol,modelreferencecontrol,internalmodelcontrol,predictivecontrol,optimaldecisioncontrolandsoon.

ApplicationofArtificialNeuralNetworkintheTransportationField

Thisyear,peoplehavebegunin-depthresearchontheapplicationofneuralnetworkinthetransportationsystem.Transportationproblemsarehighlynon-linear,andtheavailabledataareusuallylargeandcomplex.Usingneuralnetworkstodealwithrelatedproblemshasitshugeadvantages.Thescopeofapplicationinvolvesthesimulationofcardriverbehavior,parameterestimation,roadmaintenance,vehicledetectionandclassification,trafficpatternanalysis,cargooperationmanagement,trafficflowprediction,transportationstrategyandeconomy,transportationenvironmentalprotection,airtransportation,automaticnavigationofshipsandGoodresultshavebeenachievedinareassuchasshipidentification,subwayoperationandtrafficcontrol.

Applicationofartificialneuralnetworkinthefieldofpsychology

Fromtheformationofneuralnetworkmodel,ithasaninseparableconnectionwithpsychology.Theneuralnetworkisabstractedfromtheinformationprocessingfunctionoftheneuron,andthetrainingoftheneuralnetworkreflectsthecognitiveprocessessuchasperception,memory,andlearning.Throughcontinuousresearch,peoplearechangingthestructuralmodelandlearningrulesofartificialneuralnetworks,discussingthecognitivefunctionsofneuralnetworksfromdifferentangles,andlayingasolidfoundationfortheirresearchinpsychology.Inrecentyears,theartificialneuralnetworkmodelhasbecomeanindispensabletoolforexploringthemechanismsofadvancedmentalprocessessuchassocialcognition,memory,andlearning.Theartificialneuralnetworkmodelcanalsostudythecognitivedeficitsofpatientswithbraininjuryandchallengethetraditionalcognitivepositioningmechanism.

Althoughartificialneuralnetworkshavemadesomeprogress,therearestillmanyshortcomings,suchas:theapplicationareaisnotwideenough,theresultsarenotaccurateenough;thetrainingspeedofexistingmodelalgorithmsisnothighenough;theintegrationofalgorithmsNothighenough;atthesametime,wehopetofindnewbreakthroughsintheoryandestablishnewgeneralmodelsandalgorithms.Furtherresearchonthebiologicalneuronsystemisneededtocontinuouslyenrichpeople'sunderstandingofhumanbrainnerves.

This article is from the network, does not represent the position of this station. Please indicate the origin of reprint
TOP