Neural networks and deep learningNeural networks and deep learning PDFDeep learning neural networkneural networks and deep learning中文neural network教學neural network應用神經網絡Neural network classifierneural network中文neural network介紹neural network python教學Artificial neural network

What are Neural Networks? | IBM

文章推薦指數： 80 %
投票人數：10人

Neural networks, also known as artificial neural networks (ANNs) or simulated neural networks (SNNs), are a subset of machine learning and are ... Skiptocontent IBMCloudLearnHub WhatareNeuralNetworks? NeuralNetworks By: IBMCloudEducation 17August2020 Artificialintelligence Whatareneuralnetworks? Howdoneuralnetworkswork? Typesofneuralnetworks Neuralnetworksvs.deeplearning Historyofneuralnetworks NeuralnetworksandIBMCloud Jumpto... Whatareneuralnetworks? Howdoneuralnetworkswork? Typesofneuralnetworks Neuralnetworksvs.deeplearning Historyofneuralnetworks NeuralnetworksandIBMCloud menuicon NeuralNetworks Neuralnetworksreflectthebehaviorofthehumanbrain,allowingcomputerprogramstorecognizepatternsandsolvecommonproblemsinthefieldsofAI,machinelearning,anddeeplearning. Whatareneuralnetworks? Neuralnetworks,alsoknownasartificialneuralnetworks(ANNs)orsimulatedneuralnetworks(SNNs),areasubsetofmachinelearningandareattheheartofdeeplearningalgorithms.Theirnameandstructureareinspiredbythehumanbrain,mimickingthewaythatbiologicalneuronssignaltooneanother. Artificialneuralnetworks(ANNs)arecomprisedofanodelayers,containinganinputlayer,oneormorehiddenlayers,andanoutputlayer.Eachnode,orartificialneuron,connectstoanotherandhasanassociatedweightandthreshold.Iftheoutputofanyindividualnodeisabovethespecifiedthresholdvalue,thatnodeisactivated,sendingdatatothenextlayerofthenetwork.Otherwise,nodataispassedalongtothenextlayerofthenetwork. Neuralnetworksrelyontrainingdatatolearnandimprovetheiraccuracyovertime.However,oncetheselearningalgorithmsarefine-tunedforaccuracy,theyarepowerfultoolsincomputerscienceandartificialintelligence,allowingustoclassifyandclusterdataatahighvelocity.Tasksinspeechrecognitionorimagerecognitioncantakeminutesversushourswhencomparedtothemanualidentificationbyhumanexperts.Oneofthemostwell-knownneuralnetworksisGoogle’ssearchalgorithm. Howdoneuralnetworkswork? Thinkofeachindividualnodeasitsownlinearregressionmodel,composedofinputdata,weights,abias(orthreshold),andanoutput.Theformulawouldlooksomethinglikethis: ∑wixi+bias=w1x1+w2x2+w3x3+bias output=f(x)=1if∑w1x1+b>=0;0if∑w1x1+b<0 Onceaninputlayerisdetermined,weightsareassigned.Theseweightshelpdeterminetheimportanceofanygivenvariable,withlargeronescontributingmoresignificantlytotheoutputcomparedtootherinputs.Allinputsarethenmultipliedbytheirrespectiveweightsandthensummed.Afterward,theoutputispassedthroughanactivationfunction,whichdeterminestheoutput.Ifthatoutputexceedsagiventhreshold,it“fires”(oractivates)thenode,passingdatatothenextlayerinthenetwork.Thisresultsintheoutputofonenodebecomingintheinputofthenextnode.Thisprocessofpassingdatafromonelayertothenextlayerdefinesthisneuralnetworkasafeedforwardnetwork. Let’sbreakdownwhatonesinglenodemightlooklikeusingbinaryvalues.Wecanapplythisconcepttoamoretangibleexample,likewhetheryoushouldgosurfing(Yes:1,No:0).Thedecisiontogoornottogoisourpredictedoutcome,ory-hat.Let’sassumethattherearethreefactorsinfluencingyourdecision-making: Arethewavesgood?(Yes:1,No:0) Istheline-upempty?(Yes:1,No:0) Hastherebeenarecentsharkattack?(Yes:0,No:1) Then,let’sassumethefollowing,givingusthefollowinginputs: X1=1,sincethewavesarepumping X2=0,sincethecrowdsareout X3=1,sincetherehasn’tbeenarecentsharkattack Now,weneedtoassignsomeweightstodetermineimportance.Largerweightssignifythatparticularvariablesareofgreaterimportancetothedecisionoroutcome. W1=5,sincelargeswellsdon’tcomearoundoften W2=2,sinceyou’reusedtothecrowds W3=4,sinceyouhaveafearofsharks Finally,we’llalsoassumeathresholdvalueof3,whichwouldtranslatetoabiasvalueof–3.Withallthevariousinputs,wecanstarttopluginvaluesintotheformulatogetthedesiredoutput. Y-hat=(1*5)+(0*2)+(1*4)–3=6 Ifweusetheactivationfunctionfromthebeginningofthissection,wecandeterminethattheoutputofthisnodewouldbe1,since6isgreaterthan0.Inthisinstance,youwouldgosurfing;butifweadjusttheweightsorthethreshold,wecanachievedifferentoutcomesfromthemodel.Whenweobserveonedecision,likeintheaboveexample,wecanseehowaneuralnetworkcouldmakeincreasinglycomplexdecisionsdependingontheoutputofpreviousdecisionsorlayers. Intheexampleabove,weusedperceptronstoillustratesomeofthemathematicsatplayhere,butneuralnetworksleveragesigmoidneurons,whicharedistinguishedbyhavingvaluesbetween0and1.Sinceneuralnetworksbehavesimilarlytodecisiontrees,cascadingdatafromonenodetoanother,havingxvaluesbetween0and1willreducetheimpactofanygivenchangeofasinglevariableontheoutputofanygivennode,andsubsequently,theoutputoftheneuralnetwork. Aswestarttothinkaboutmorepracticalusecasesforneuralnetworks,likeimagerecognitionorclassification,we’llleveragesupervisedlearning,orlabeleddatasets,totrainthealgorithm.Aswetrainthemodel,we’llwanttoevaluateitsaccuracyusingacost(orloss)function.Thisisalsocommonlyreferredtoasthemeansquarederror(MSE).Intheequationbelow, irepresentstheindexofthesample, y-hatisthepredictedoutcome, yistheactualvalue,and misthenumberofsamples. 𝐶𝑜𝑠𝑡𝐹𝑢𝑛𝑐𝑡𝑖𝑜𝑛=𝑀𝑆𝐸=1/2𝑚∑129_(𝑖=1)^𝑚▒(𝑦 ̂^((𝑖))−𝑦^((𝑖)))^2 Ultimately,thegoalistominimizeourcostfunctiontoensurecorrectnessoffitforanygivenobservation.Asthemodeladjustsitsweightsandbias,itusesthecostfunctionandreinforcementlearningtoreachthepointofconvergence,orthelocalminimum.Theprocessinwhichthealgorithmadjustsitsweightsisthroughgradientdescent,allowingthemodeltodeterminethedirectiontotaketoreduceerrors(orminimizethecostfunction).Witheachtrainingexample,theparametersofthemodeladjusttograduallyconvergeattheminimum.  SeethisIBMDeveloperarticleforadeeperexplanationofthequantitativeconceptsinvolvedinneuralnetworks. Mostdeepneuralnetworksarefeedforward,meaningtheyflowinonedirectiononly,frominputtooutput.However,youcanalsotrainyourmodelthroughbackpropagation;thatis,moveintheoppositedirectionfromoutputtoinput.Backpropagationallowsustocalculateandattributetheerrorassociatedwitheachneuron,allowingustoadjustandfittheparametersofthemodel(s)appropriately. Typesofneuralnetworks Neuralnetworkscanbeclassifiedintodifferenttypes,whichareusedfordifferentpurposes.Whilethisisn’tacomprehensivelistoftypes,thebelowwouldberepresentativeofthemostcommontypesofneuralnetworksthatyou’llcomeacrossforitscommonusecases: Theperceptronistheoldestneuralnetwork,createdbyFrankRosenblattin1958.Ithasasingleneuronandisthesimplestformofaneuralnetwork: Feedforwardneuralnetworks,ormulti-layerperceptrons(MLPs),arewhatwe’veprimarilybeenfocusingonwithinthisarticle.Theyarecomprisedofaninputlayer,ahiddenlayerorlayers,andanoutputlayer.WhiletheseneuralnetworksarealsocommonlyreferredtoasMLPs,it’simportanttonotethattheyareactuallycomprisedofsigmoidneurons,notperceptrons,asmostreal-worldproblemsarenonlinear.Datausuallyisfedintothesemodelstotrainthem,andtheyarethefoundationforcomputervision,naturallanguageprocessing,andotherneuralnetworks. Convolutionalneuralnetworks(CNNs)aresimilartofeedforwardnetworks,butthey’reusuallyutilizedforimagerecognition,patternrecognition,and/orcomputervision.Thesenetworksharnessprinciplesfromlinearalgebra,particularlymatrixmultiplication,toidentifypatternswithinanimage. Recurrentneuralnetworks(RNNs)areidentifiedbytheirfeedbackloops.Theselearningalgorithmsareprimarilyleveragedwhenusingtime-seriesdatatomakepredictionsaboutfutureoutcomes,suchasstockmarketpredictionsorsalesforecasting. Neuralnetworksvs.deeplearning DeepLearningandneuralnetworkstendtobeusedinterchangeablyinconversation,whichcanbeconfusing.Asaresult,it’sworthnotingthatthe“deep”indeeplearningisjustreferringtothedepthoflayersinaneuralnetwork.Aneuralnetworkthatconsistsofmorethanthreelayers—whichwouldbeinclusiveoftheinputsandtheoutput—canbeconsideredadeeplearningalgorithm.Aneuralnetworkthatonlyhastwoorthreelayersisjustabasicneuralnetwork. Tolearnmoreaboutthedifferencesbetweenneuralnetworksandotherformsofartificialintelligence, likemachinelearning,pleasereadtheblogpost“AIvs.MachineLearningvs.DeepLearningvs.NeuralNetworks:What’stheDifference?” Historyofneuralnetworks Thehistoryofneuralnetworksislongerthanmostpeoplethink.Whiletheideaof“amachinethatthinks”canbetracedtotheAncientGreeks,we’llfocusonthekeyeventsthatledtotheevolutionofthinkingaroundneuralnetworks,whichhasebbedandflowedinpopularityovertheyears: 1943:WarrenS.McCullochandWalterPittspublished“Alogicalcalculusoftheideasimmanentinnervousactivity(PDF,1MB)(linkresidesoutsideIBM)”Thisresearchsoughttounderstandhowthehumanbraincouldproducecomplexpatternsthroughconnectedbraincells,orneurons.OneofthemainideasthatcameoutofthisworkwasthecomparisonofneuronswithabinarythresholdtoBooleanlogic(i.e.,0/1ortrue/falsestatements).   1958:FrankRosenblattiscreditedwiththedevelopmentoftheperceptron,documentedinhisresearch,“ThePerceptron:AProbabilisticModelforInformationStorageandOrganizationintheBrain”(PDF,1.6MB)(linkresidesoutsideIBM).HetakesMcCullochandPitt’sworkastepfurtherbyintroducingweightstotheequation.LeveraginganIBM704,Rosenblattwasabletogetacomputertolearnhowtodistinguishcardsmarkedontheleftvs.cardsmarkedontheright. 1974:Whilenumerousresearcherscontributedtotheideaofbackpropagation,PaulWerboswasthefirstpersonintheUStonoteitsapplicationwithinneuralnetworkswithinhisPhDthesis(PDF,8.1MB)(linkresidesoutsideIBM). 1989:YannLeCunpublishedapaper(PDF,5.7MB)(linkresidesoutsideIBM)illustratinghowtheuseofconstraintsinbackpropagationanditsintegrationintotheneuralnetworkarchitecturecanbeusedtotrainalgorithms.Thisresearchsuccessfullyleveragedaneuralnetworktorecognizehand-writtenzipcodedigitsprovidedbytheU.S.PostalService. NeuralnetworksandIBMCloud Fordecadesnow,IBMhasbeenapioneerinthedevelopmentofAItechnologiesandneuralnetworks,highlightedbythedevelopmentandevolutionofIBMWatson.WatsonisnowatrustedsolutionforenterpriseslookingtoapplyadvancednaturallanguageprocessinganddeeplearningtechniquestotheirsystemsusingaproventieredapproachtoAIadoptionandimplementation. WatsonusestheApacheUnstructuredInformationManagementArchitecture(UIMA)frameworkandIBM’sDeepQAsoftwaretomakepowerfuldeeplearningcapabilitiesavailabletoapplications.UtilizingtoolslikeIBMWatsonStudio,yourenterprisecanseamlesslybringopensourceAIprojectsintoproductionwhiledeployingandrunningmodelsonanycloud. Formoreinformationonhowtogetstartedwithdeeplearningtechnology,exploreIBMWatsonStudio andtheDeepLearningservice. SignupforanIBMidandcreateyourIBMCloudaccount. ArtificialIntelligenceebook Downloadtheebook Featuredproducts WatsonStudio WatsonMachineLearningAccelerator Relatedlinks Linearregression Predictiveanalytics

請為這篇文章評分？

延伸文章資訊