基于混合算法训练的神经网络同步二进制计数器(IJIGSP-V9-N10-5)

I.J.Image,GraphicsandSignalProcessing,2017,10,38-49PublishedOnlineOctober2017inMECS()DOI:10.5815/ijigsp.2017.10.05Copyright©2017MECSI.J.Image,GraphicsandSignalProcessing,2017,10,38-49NeuralNetworkSynchronousBinaryCounterUsingHybridAlgorithmTrainingRaviTejaYakkaliNXPSemiconductorsNoida,IndiaEmail:[e-mail]calUniversity,Delhi,IndiaEmail:[e-mail]:21July2017;Accepted:07August2017;Published:08October2017Abstract—InformationprocessingusingNeuralNetworkCountercanresultinfasterandaccuratecomputationofdataduetotheirparallelprocessing,learningandadaptabilitytovariousenvironments.Inthispaper,anovel4-BitNegativeEdgeTriggeredBinarySynchronousUp/DownCounterusingArtificialNeuralNetworkstrainedwithhybridalgorithmsisproposed.TheCounterwasbuiltsolelyusinglogicgatesandflipflops,andthentheyaretrainedusingdifferentevolutionaryalgorithms,withamultiobjectivefitnessfunctionusingthebackpropagationlearning.Thus,thedeviceislesspronetoerrorwithaveryfastconvergencerate.Thesimulationresultsofproposedhybridalgorithmsarecomparedintermsofnetworkweights,bit-value,percentageerrorandvariancewithrespecttotheoreticaloutputswhichshowthattheproposedcounterhasvaluesclosetothetheoreticaloutputs.IndexTerms—ArtificialNeuralNetworks,HybridAlgorithms,SynchronousBinaryCounter,BackPropagationAlgorithm,EvolutionaryAlgorithms.I.INTRODUCTIONArtificialNeuralNetworksandevolutionaryalgorithmtechniqueshavebeenusedinfindingoptimumsolutionsformanyengineeringapplicationsrangingfromwirelesscommunication[1],patternrecognitiontomedicalapplications[2,31-32].EffortshavebeenmadetoimproveperformanceofLogicCircuitsinapplicationsofSatelliteCommunication,medicalapplicationsandradiationproneenvironment[3-5],wherethedevicesareexposedtoahugeamountofnoiseanddistortionwithaverylowsignalpowerbutwestillrequirepreciseandaccuratesignalswithfastoutput.Soinordertoaddresstheseissues,wehaveusedNeuralNetworkasthebasicbuildingblockforthedesignofSynchronousCounter.NeuralNetworksfindapplicationsincriticalandsensitiveenvironments[6].TheadvantageofusingNeuralNetworkisthattheylearnbyexampleandalsocanadapttovariousenvironments[7-9],becauseoftheavailabilityofrobusttrainingalgorithms,theirprecision,accuracycanbecontrolledanddataprocessingbytheNeuralNetworkcanresultinfasteroutputduetotheirparallelprocessingnaturewhichcanbeusedinthedesignofhighlyreliableandfastlogiccircuits.ManyapplicationsofNeuralNetworkshavebeenproposedinDigitalCircuits.ANeuralNetworkhasbeenusedforDigitalSignalProcessingin[10]andtheyarealsousedintheirfaultdiagnosis[11-12].CantoneseSpeechRecognitionwasdoneusingNeuralNetworklogicgateCircuit(AND,OR,NOT)whichwastrainedusingGeneticAlgorithm[13].HopfieldneuralnetworkswereusedforthedesignofANDCircuit,thefulladder,D-Latchwith4-BitShiftRegisterand4-BitAsynchronouscounter[14].ABinaryLogicalNeuralNetworkwasconstructed[15]inwhichactivationfunctionofeachneuronwasacombinationallogicgate.Faultandradiationanalysis,whichinvolvescreatingareplicaoforiginaldigitalcircuitusingNeuralNetwork,isdonein[16].ABi-stableMemoryDeviceandBinaryCounterswerecreatedusingSpikingNeuralNetworksin[17].Geneticalgorithmismetaheuristicalgorithmbasedonnaturalselectionofhumanbeings.ParticleSwarmOptimizationisinspiredbythemovementoforganismsinabirdflock[18].ArtificialBeeColonyalgorithmisaswarmintelligencealgorithm[19]basedontheintelligentforagingbehaviorofhoneybees.BackPropagationAlgorithm[20]isappliedtomultilayerfeed-forwardNeuralNetwithcontinuousActivationFunction[21].Since,BinarySigmoidFunction[22]isusedforbinaryinputforcomputationofoutput,soback-propagationalgorithmiswellsuitedforourapplication.Eventhoughevolutionaryalgorithmscanbeusedtofindaglobaloptimizedsolutionbutthesealgorithmsareunabletofindalocaloptimizedsolutionbecauseofthecreationofalargesolutionspaceintrainingneuralnetworkweights[23]andtheirsubsequentevaluationprocesswhilebackpropagationalgorithmisoftenstuckatalocalbestsolution.Hybridalgorithmshavegainedinsightingivingbettersolutionsformanyengineeringapplications.AHybridGeneticAlgorithmwithhypermutationandelitistNeuralNetworkSynchronousBinaryCounterUsingHybridAlgorithmTraining39Copyright©2017MECSI.J.Image,GraphicsandSignalProcessing,2017,10,38-49strategyhasbeenappliedtodesignofAutomatedAnalogCircuit[24],ahybridPSOforMedicalImageRegistration[25]andhybridABCalgorithmwithaLMalgorithmfortrainingNeuralNetworksin[26].Also,Geneticalgorithmhasbeenusedtodesignreconfigurablehardwarein[6,27]usingANNwhichadaptsitshardwareaccordingtotheenvironment.AhybridvariantofparticleSwarmOptimizationandBackPropagationAlgorithmwhichinvolvesPSOsearchingagloballyoptimalsolutionandafterthat,BPissearchingforalocallyoptimalsolutionhasbeenappliedin[23]fortrainingfeedforwardANNinapplicationofIrisrecognition,functionapproximationandthreebitsparityproblem.Butsomeofthehybridalgorithmsalsohavesomelimitations.AhybridalgorithmwhichinvolvesGAfindingoptimizedweightsafterlearningfromBackPropagationAlgorithmispresentedin[28],butsuffersfromtheproblemthatBPfirstfinds