Definition
Biologicalneuralnetworkmainlyreferstotheneuralnetworkofthehumanbrain,whichisthetechnicalprototypeoftheartificialneuralnetwork.Thehumanbrainisthematerialbasisofhumanthinking.Thefunctionofthinkingislocatedinthecerebralcortex.Thelattercontainsabout10^11neurons.Eachneuronisconnectedtoabout103otherneuronsthroughnervesynapses,formingahighlycomplexHighlyflexibledynamicnetwork.Asadiscipline,biologicalneuralnetworkmainlystudiesthestructure,functionandworkingmechanismofhumanbrainneuralnetwork,andisintendedtoexplorethelawsofhumanbrainthinkingandintelligentactivities.
Artificialneuralnetworkisthetechnicalreproductionofbiologicalneuralnetworkinasimplifiedsense.Asadiscipline,itsmaintaskistobuildpracticalartificialNeuralnetworkmodel,designthecorrespondinglearningalgorithm,simulateacertainintelligentactivityofthehumanbrain,andthenrealizeittechnicallytosolvepracticalproblems.Therefore,thebiologicalneuralnetworkmainlystudiesthemechanismofintelligence;theartificialneuralnetworkmainlystudiestherealizationoftheintelligentmechanism,andthetwocomplementeachother.
Researchcontent
Theresearchcontentofneuralnetworksisquiteextensive,reflectingthecharacteristicsofmultidisciplinaryandinterdisciplinarytechnicalfields.Themainresearchworkfocusesonthefollowingaspects:
Biologicalprototype
Fromphysiology,psychology,anatomy,brainscience,pathology,etc.Studythebiologicalprototypestructureandfunctionalmechanismofnervecells,neuralnetworks,andnervoussystems.
Establishmodel
Basedontheresearchofbiologicalprototypes,establishtheoreticalmodelsofneuronsandneuralnetworks.Includingconceptualmodels,knowledgemodels,physicalandchemicalmodels,mathematicalmodels,etc.
Algorithm
Constructaspecificneuralnetworkmodelonthebasisoftheoreticalmodelresearchtorealizecomputersimulationorpreparehardware,includingnetworklearningalgorithmResearch.Thisworkisalsocalledtechnicalmodelresearch.
Thealgorithmusedinneuralnetworksisvectormultiplication,andsymbolicfunctionsandtheirvariousapproximationsarewidelyused.Parallelism,faulttolerance,hardwareimplementation,andself-learningcharacteristicsarethebasicadvantagesofneuralnetworksandthedifferencebetweenneuralnetworkcalculationmethodsandtraditionalmethods.
Classification
Accordingtoitsmodelstructure,artificialneuralnetworkscanberoughlydividedintofeedforwardnetworks(alsocalledmultilayerperceptronnetworks)andfeedbacknetworks(alsocalledHopfieldnetworks))Twocategories,theformercanberegardedasakindoflarge-scalenonlinearmappingsysteminmathematics,andthelatterisakindoflarge-scalenonlineardynamicsystem.Accordingtothelearningmethod,theartificialneuralnetworkcanbedividedintothreetypes:supervisedlearning,unsupervisedlearningandsemi-supervisedlearning;accordingtotheworkingmethod,itcanbedividedintotwotypes:deterministicandrandom;accordingtotimecharacteristics,itcanalsobedividedintocontinuousordiscreteTypetwotypes,andsoon.
Features
Nomatterwhattypeofartificialneuralnetwork,theircommonfeaturesarelarge-scaleparallelprocessing,distributedstorage,elastictopology,highredundancyandnonlinearoperations.Therefore,ithasveryhighcomputingspeed,strongassociationability,strongadaptability,strongfaulttoleranceandself-organizationability.Thesecharacteristicsandabilitiesconstitutethetechnicalbasisforartificialneuralnetworkstosimulateintelligentactivities,andhaveobtainedimportantapplicationsinawiderangeoffields.Forexample,inthefieldofcommunications,artificialneuralnetworkscanbeusedfordatacompression,imageprocessing,vectorcoding,errorcontrol(errorcorrectionanderrordetectioncoding),adaptivesignalprocessing,adaptiveequalization,signaldetection,patternrecognition,ATMflowcontrol,Routing,communicationnetworkoptimizationandintelligentnetworkmanagement,etc.
Howitworks
"Howdoesthehumanbrainwork?"
"Canhumansmakeartificialneuronsthatmimicthehumanbrain?"
Formanyyears,peoplehavetriedtounderstandandanswertheabovequestionsfromvariousanglessuchasmedicine,biology,physiology,philosophy,informatics,computerscience,cognition,andorganizationsynergy.Intheprocessofsearchingfortheanswerstotheabovequestions,anewfieldofinterdisciplinarytechnologyhasgraduallyformed,whichiscalled"neuralnetwork".Theresearchofneuralnetworksinvolvesmanydisciplines,whicharecombined,penetratedandpromotedwitheachother.Scientistsindifferentfieldsstartfromtheinterestsandcharacteristicsoftheirrespectivedisciplines,askdifferentquestions,andconductresearchfromdifferentangles.
Artificialneuralnetworksmustfirstlearnwithcertainlearningcriteriabeforetheycanwork.Nowtaketheartificialneuralnetwork'srecognitionofthetwoletters"A"and"B"asanexample.Itisstipulatedthatwhen"A"isinputtothenetwork,itshouldoutput"1",andwhentheinputis"B",theoutputis"0".
Sothecriterionofnetworklearningshouldbe:ifthenetworkmakesawrongdecision,throughthenetworklearning,thenetworkshouldreducethepossibilityofmakingthesamemistakenexttime.First,assignarandomvalueintheintervalof(0,1)toeachconnectionweightofthenetwork,andinputtheimagemodecorrespondingto"A"tothenetwork.Thenetworkaddstheweightoftheinputmode,comparesitwiththethreshold,andperformsnon-Linearoperation,gettheoutputofthenetwork.Inthiscase,theprobabilitythatthenetworkoutputis"1"and"0"is50%,whichmeansitiscompletelyrandom.Atthistime,iftheoutputis"1"(theresultiscorrect),theconnectionweightisincreasedsothatwhenthenetworkencountersthe"A"modeinputagain,itcanstillmakeacorrectjudgment.
Thefunctionofanordinarycomputerdependsontheknowledgeandabilitygivenintheprogram.Obviously,itwillbeverydifficulttosummarizeandcompileprogramsforsmartactivities.
Artificialneuralnetworksalsohavepreliminaryadaptiveandself-organizingcapabilities.Changethesynapticweightvalueinthelearningortrainingprocesstoadapttotherequirementsofthesurroundingenvironment.Thesamenetworkcanhavedifferentfunctionsduetodifferentlearningmethodsandcontent.Artificialneuralnetworkisasystemwithlearningability,whichcandevelopknowledgesoastoexceedtheoriginalknowledgelevelofthedesigner.Generally,itslearningandtrainingmethodscanbedividedintotwotypes,oneissupervisedormentoredlearning,wherethegivensamplestandardisusedtoclassifyorimitate;theotherisunsupervisedlearningornon-supervisedlearning.Atthistime,onlythelearningmethodorcertainrulesarespecified,andthespecificlearningcontentvarieswiththeenvironmentofthesystem(ie,theinputsignalsituation).Thesystemcanautomaticallydiscovertheenvironmentalcharacteristicsandregularity,andhasafunctionmoresimilartothehumanbrain.
Neuralnetworkislikeachildwholovestolearn.Shewillnotforgettheknowledgeyouteachherandwillapplywhatshehaslearned.WeaddeachinputintheLearningSettotheneuralnetworkandtelltheneuralnetworkwhatclassificationshouldbeoutput.Afterallthelearningsetsarerun,theneuralnetworksumsupherownthoughtsbasedontheseexamples.Howshesummedthemupisablackbox.Afterthat,wecanusetheneuralnetworktotestthetestexamplesinthetestingset.Ifthetestpasses(forexample,80%or90%accuracy),thentheneuralnetworkisconstructedsuccessfully.Wecanthenusethisneuralnetworktodeterminetheclassificationofthetransaction.
Neuralnetworkistoexplorethemodelthatsimulatesthefunctionofthehumanbrainnervoussystemthroughthemodelingandconnectionofthebasicunitofthehumanbrain-neurons,anddevelopamodelwithlearning,association,memoryandpatternrecognitionArtificialsystemswithintelligentinformationprocessingfunctions.Animportantfeatureofaneuralnetworkisthatitcanlearnfromtheenvironmentandstoretheresultsofthelearninginthesynapticconnectionsofthenetwork.Thelearningofaneuralnetworkisaprocess.Undertheexcitationofitsenvironment,somesamplepatternsaresuccessivelyinputtothenetwork,andtheweightmatrixofeachlayerofthenetworkisadjustedaccordingtocertainrules(learningalgorithm).Convergestoacertainvalue,andthelearningprocessends.Thenwecanusethegeneratedneuralnetworktoclassifytherealdata.
Historyofdevelopment
In1943,psychologistW·MccullochandmathematicallogicianW·PittsfirstproposedneuronmathematicsbasedonanalyzingandsummarizingthebasiccharacteristicsofneuronsModel.Thismodelisstillinusetoday,anddirectlyaffectstheprogressofresearchinthisfield.Therefore,thetwoofthemcanbecalledthepioneersofartificialneuralnetworkresearch.
In1945,thedesignteamledbyvonNeumannsuccessfullytrial-producedstored-programelectroniccomputers,markingthebeginningoftheelectroniccomputerera.In1948,inhisresearchwork,hecomparedthefundamentaldifferencebetweenthestructureofthehumanbrainandthestored-programcomputer,andproposedanetworkstructureofregenerativeautomatacomposedofsimpleneurons.However,duetotherapiddevelopmentofinstructionstoragecomputertechnology,hewasforcedtoabandonthenewapproachofneuralnetworkresearch,continuetodevotehimselftotheresearchofinstructionstoragecomputertechnology,andmadegreatcontributionsinthisfield.AlthoughvonNeumann'snameisassociatedwithordinarycomputers,heisalsooneofthepioneersofartificialneuralnetworkresearch.
Attheendofthe1950s,F·Rosenblattdesignedandproducedthe"perceptron",whichisamulti-layerneuralnetwork.Thisworkputstheresearchofartificialneuralnetworkfromtheoreticaldiscussionintoengineeringpracticeforthefirsttime.Atthattime,manylaboratoriesintheworldimitatedtheproductionofperceptrons,whichwereappliedtothestudyoftextrecognition,voicerecognition,sonarsignalrecognition,andlearningandmemoryproblems.However,theresearchclimaxofartificialneuralnetworksdidnotlastlong.Manypeoplegaveupresearchworkinthisareaoneafteranother.Thiswasbecausethedevelopmentofdigitalcomputerswasinitsheyday,andmanypeoplemistakenlybelievedthatdigitalcomputerscouldsolveartificialintelligenceandpatterns.Alltheproblemsofrecognitionandexpertsystemsmadetheworkoftheperceptronnotbetakenseriously;secondly,thelevelofelectronictechnologyatthattimewasrelativelybackward,andthemaincomponentswereelectrontubesortransistors.Theneuralnetworksmadebythemwerebulkyandexpensive.Itiscompletelyimpossibletomakeaneuralnetworksimilarinscaletoarealneuralnetwork;inaddition,ina1968bookcalled"Perceptron",itwaspointedoutthatthefunctionoflinearperceptronwaslimited,anditcouldnotsolvesuchXOR.Thebasicproblemsofmulti-layernetworksandtheinabilitytofindeffectivecalculationmethodsformulti-layernetworkshavepromptedalargenumberofresearcherstoloseconfidenceintheprospectsofartificialneuralnetworks.Inthelate1960s,theresearchonartificialneuralnetworksenteredalowebb.
Inaddition,intheearly1960s,Widrowproposedanadaptivelinearelementnetwork,whichisalinearweightedsumthresholdnetworkwithcontinuousvalues.Later,anonlinearmultilayeradaptivenetworkwasdevelopedonthisbasis.Atthattime,althoughtheseworksdidnotmarkthenameoftheneuralnetwork,itwasactuallyanartificialneuralnetworkmodel.
Aspeople’sinterestinperceptronsdeclines,theresearchonneuralnetworkshasbeensilentforalongtime.Intheearly1980s,themanufacturingtechnologyofVLSIwithamixtureofanaloganddigitalwasraisedtoanewlevelandwascompletelyputintopracticaluse.Inaddition,thedevelopmentofdigitalcomputersencountereddifficultiesinanumberofapplicationareas.Thisbackgroundindicatesthatthetimetoseekawayoutofartificialneuralnetworksisripe.AmericanphysicistHopfieldpublishedtwopapersontheresearchofartificialneuralnetworksintheProceedingsoftheNationalAcademyofSciencesin1982and1984,whicharousedahugeresponse.Peoplehavere-recognizedthepowerofneuralnetworksandtherealityoftheirapplications.Immediately,alargenumberofscholarsandresearcherscarriedoutfurtherworkaroundthemethodproposedbyHopfield,whichformedtheresearchboomofartificialneuralnetworkssincethemid-1980s.
Commontools
Amongmanyneuralnetworktools,NeuroSolutionsisalwaysintheleadingpositionintheindustry.ItisahighlygraphicalneuralnetworkdevelopmenttoolthatcanbeusedinwindowsXP/7.Itcombinesmodularity,icon-basedwebdesigninterface,advancedlearningprogramandgeneticoptimization.Thisneuralnetworkdesigntool,whichcanbeusedtostudyandsolvecomplexproblemsintherealworld,isalmostunlimitedinuse.
Researchdirection
Theresearchofneuralnetworkcanbedividedintotwoaspects:theoreticalresearchandappliedresearch.
Theoreticalresearchcanbedividedintothefollowingtwocategories:
1.Theuseofneurophysiologicalandcognitivescientificresearchonhumanthinkingandintelligencemechanisms.
2.Makeuseoftheresearchresultsofbasicneuraltheory,usemathematicalmethodstoexploreneuralnetworkmodelswithmorecompletefunctionsandsuperiorperformance,andin-depthstudyofnetworkalgorithmsandperformance,suchas:stability,convergence,andfaulttolerance,Robustness,etc.;developnewnetworkmathematicaltheories,suchasneuralnetworkdynamics,nonlinearneuralfields,etc.
Applicationresearchcanbedividedintothefollowingtwocategories:
1.Researchonsoftwaresimulationandhardwarerealizationofneuralnetworks.
2.Researchontheapplicationofneuralnetworksinvariousfields.Thesefieldsmainlyinclude:
Patternrecognition,signalprocessing,knowledgeengineering,expertsystem,optimizationcombination,robotcontrol,etc.Withthecontinuousdevelopmentofneuralnetworktheoryitself,relatedtheories,andrelatedtechnologies,theapplicationofneuralnetworkswillsurelybecomemorein-depth.