ebook img

Deploying a quantum annealing processor to detect tree cover in aerial imagery of California PDF

22 Pages·2017·2.56 MB·English
by  
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Deploying a quantum annealing processor to detect tree cover in aerial imagery of California

RESEARCHARTICLE Deploying a quantum annealing processor to detect tree cover in aerial imagery of California EdwardBoyda1,2*,SaikatBasu3,SangramGanguly2,4,AndrewMichaelis4,5, SupratikMukhopadhyay3,RamakrishnaR.Nemani6 1DepartmentofPhysicsandAstronomy,SaintMary’sCollegeofCalifornia,Moraga,CA,UnitedStatesof America,2BayAreaEnvironmentalResearchInstitute,MoffettField,CA,UnitedStatesofAmerica, 3DepartmentofComputerScience,LouisianaStateUniversity,BatonRouge,LA,UnitedStatesofAmerica, a1111111111 4EarthScienceDivision,NASAAmesResearchCenter,MoffettField,CA,UnitedStatesofAmerica, a1111111111 5UniversityCorporationatCSUMontereyBay,Seaside,CA,UnitedStatesofAmerica,6NASAAdvanced a1111111111 SupercomputingDivision,NASAAmesResearchCenter,MoffettField,CA,UnitedStatesofAmerica a1111111111 a1111111111 *[email protected] Abstract OPENACCESS Quantumannealingisanexperimentalandpotentiallybreakthroughcomputationaltechnol- Citation:BoydaE,BasuS,GangulyS,MichaelisA, ogyforhandlinghardoptimizationproblems,includingproblemsofcomputervision.We MukhopadhyayS,NemaniRR(2017)Deployinga presentacasestudyintrainingaproduction-scaleclassifieroftreecoverinremotesensing quantumannealingprocessortodetecttreecover imagery,usingearly-generationquantumannealinghardwarebuiltbyD-waveSystems, inaerialimageryofCalifornia.PLoSONE12(2): Inc.Beginningwithinaknownboostingframework,wetraindecisionstumpsontexturefea- e0172505.doi:10.1371/journal.pone.0172505 turesandvegetationindicesextractedfromfour-band,one-meter-resolutionaerialimagery Editor:ShijoJoseph,KeralaForestResearch fromthestateofCalifornia.Wethenimposearegulatedquadratictrainingobjectiveto Institute,INDIA selectanoptimalvotingsubsetfromamongthesestumps.Thevotesofthesubsetdefine Received:June29,2016 theclassifier.Foroptimization,thelogicalvariablesintheobjectivefunctionmaptoquantum Accepted:February6,2017 bitsinthehardwaredevice,whilequadraticcouplingsencodeasthestrengthofphysical Published:February27,2017 interactionsbetweenthequantumbits.Hardwaredesignlimitsthenumberofcouplings Copyright:Thisisanopenaccessarticle,freeofall betweenthesebasicphysicalentitiestofiveorsix.Toaccountforthislimitationinmapping copyright,andmaybefreelyreproduced, largeproblemstothehardwarearchitecture,weproposeatruncationandrescalingofthe distributed,transmitted,modified,builtupon,or trainingobjectivethroughatrainablemetaparameter.Theboostingprocessonourbasic otherwiseusedbyanyoneforanylawfulpurpose. 108-and508-variableproblems,thusconstituted,returnsclassifiersthatincorporatea TheworkismadeavailableundertheCreative CommonsCC0publicdomaindedication. diverserangeofcolor-andtexture-basedmetricsanddiscriminatetreecoverwithaccura- ciesashighas92%invalidationand90%onatestsceneencompassingtheopenspace DataAvailabilityStatement:Dataareavailable fromFigshare(DOI:10.6084/m9.figshare. preservesanddensesuburbanbuildofMillValley,CA. 4644535). Funding:ThisworkwassupportedbytheNASA EarthScienceDivisionandperformedusingthe computingfacilitiesoftheNASAAdvanced Supercomputing(NAS)divisionandNASAEarth Exchange(NEX).Anyopinions,findings,and Introduction conclusionsorrecommendationsexpressedinthis Theproliferationofveryhighresolution(VHR)aerialandsatelliteimageryopensthewayto materialarethoseoftheauthorsanddonot significantimprovementsinremotesensingdataproducts.Itisnowpossibletoidentify necessarilyreflectthatofNASAortheUnited StatesGovernment.Thefundershadnorolein structuresatbetterthan1-meterresolution,downfrom30metersinexistingLandsat-based PLOSONE|DOI:10.1371/journal.pone.0172505 February27,2017 1/22 DeployingaquantumannealingprocessortodetecttreecoverinaerialimageryofCalifornia studydesign,datacollectionandanalysis,decision solutions.Objects—individualsheds,tractors,streams,islands,rocks,trees,vines,andfur- topublish,orpreparationofthemanuscript. rows—comeintofocusfromoutofbroadswathsofforestorfield,allowingfordetailedsite- Competinginterests:Theauthorshavedeclared specificstudiesaswellasmoreaccuratedelineationsoflandcoverinthelarge.VHRdatasets thatnocompetinginterestsexist. arerichinpotentialities.Atthesametime,newlysophisticatedcomputeralgorithmsare requiredtoparsethedata. Duetohighvariabilitywithinclassesandinatmospheric,lighting,andphoto-geometric conditions,land-coverclasscognitionatveryhighresolutionremainsadifficultchallenge.In thisrealm,object-orientedtechniquesforintegratedsegmentationandclassificationhave showngreatrecentpromise.Theyofferarichersemanticsandmoreaccurateclassification whencomparedtoclusteringofspectralandtexturalprimitivesalone.(See,forexample,[1]in thecontextofcomputervisionor[2]forareviewinthecontextofremotesensing.)Object- orientedappraochesputsignificantdemandsoncomputationalinfrastructure.Themachine learningalgorithmsleadtomemory-andprocessor-intensivetraining(optimization)prob- lemsinwhichthousandsofparametersmustbedetermined,whiletherelevantVHRdatasets themselvesextendtoterabytesinsize.Giventhesepressures,itisnaturaltoaskwhatsortsof breakthroughs,algorithmicortechnological,maylieonthehorizon. Quantumcomputing(see,e.g.,[3])isonesuchpossibility.Broadlydefined,quantumcom- putingisanefforttoencodehardcomputationalproblemsinthedynamicsofquantumphysi- calsystems.Thestatespaceofquantumsystemsisexponentiallylargeinthenumberofbasic physicalvariables,andiftappedproperly,canyieldcomputationalresultsexponentiallyfaster thanthebestavailableclassicalalternatives.Thisadvantagehasbeendemonstratedformally forparticularproblems,integerfactorization[4]beingtheexamplemostoftencitedduetoits roleinthewidely-usedRSApublic-keycryptographyscheme.Thecommunityisactively workingtocharacterizethescalingadvantageswecanexpectforbroaderclassesofproblems. Withinthequantumcomputingparadigm,quantumannealing[5–7]isacomputational metaheuristicdesignedtosolveoptimizationproblems.Akintosimulatedannealing,quantum annealingseekstheminimumofacostfunctioninacomplexconfigurationspace.Physically, thecostfunctionencodesasthesystem’senergy.Thealgorithmproceedsbypreparingthesys- teminaquantumsuperpostionofallpossibleconfigurationsinthesolutionspace,allequally probable,thusinitiatingauniquelyquantumparallelprocessing.Thesystemthenisevolvedin timeuntilthesoughtminimalenergyconfigurationisoverwhelminglyprobable.Inprinciple, intheabsenceofthermalnoise,itcanbearrangedsothattheminimalenergyconfiguration willbemeasuredonread-outwithprobabilityarbitrarilyclosetoone.Ratherthansampling, physicalinteractionsbetweenquantumbitsdrivethesystemtotheenergyminimum.Aspart ofthisprocess,thesystemhasthepossibilityofquantumtunnelingthroughtall,narrowbarri- ersintheenergylandscapetoescapelocalminimainlessthanexponentialtime. AquantumannealingprocessorbuiltbyD-waveSystems,Inc.,with1152quantumbits (qubits)isnowoperatingatNASA’sAmesResearchCenter.ThedeploymentoftheD-wave 2Xfollowsearliertrialsof128-qubitand512-qubitprocessorsatLockheedMartinandat Ames.Muchworkhasgonetocharacterizetheperformanceofthesemachines.Evidenceof thepersistenceofquantumcoherenceduringcomputationhasbeenobservedinsubsystemsof eightqubits[8–10].Ontheotherhand,theprocessorhashandilybeenbeatenforspeedby desktopCPUsrunningoptimizedsimulatedannealingand/ormoretargetedsamplingalgo- rithms[11–13].Inlate2015,afirstsetofproblemswerecraftedonwhichtheD-wavequantum annealerrunssignificantlyfasterthanclassicalsimulatedannealing[14]. Motivatedtoadvanceourremotesensingcapabilitiesandtobetterunderstandthepossibil- itiesofquantumannealingvisionalgorithms,wesetouttotrainaproduction-scaleclassifier ofaerialimageryontheD-waveprocessor.Webeginwithanimplementationofaboosting algorithmknownasQBoost,developedspecificallyfortheD-wavearchitecture.Itwas PLOSONE|DOI:10.1371/journal.pone.0172505 February27,2017 2/22 DeployingaquantumannealingprocessortodetecttreecoverinaerialimageryofCalifornia employedin2009toidentifycarsinphotographsofstreetscenes,havingbeentrainedona processorwith52functioningqubits[15–17].Unfortunately,QBoost,alongwithproblems fromacommongeneralclassofquadratictrainingobjectives,doesnotscalewellontheD- wavearchitectureoronanyforeseeablequantumannealingprocessor.Bytruncatingand rescalingcouplingsintheQBoosttrainingobjective,withtheintroductionofanadditional trainablemetaparameter,weareabletomapproblemsofhundredsofvariablestotheD-wave chipandtobuildthedesiredclassifieroftreecoverinaerialimagery. Thisworkisanoffshootofaprototypestudy[18]plannedeventuallytodelivertreecover estimatesforthecontinentalUnitedStatesvia1-meter-resolutionVHRdatafromtheNational AgricultureImageryProgram(NAIP)[19].Theobject-orientedplatformproducespixelwise probabilisticmapsfortreecoverastheoutputofaconditionalrandomfield,whichitselfinte- gratesoutputsfromaregion-mergingsegmentationroutineandaneuralnetworkclassifier.In theprototype,treecovermapsweregeneratedfor11,095inputNAIPtilescoveringthestateof California,withcorrectdetectionratesof85%inregionsoffragmentedforestand70%for urbanareas.Wehaveformulatedtheboostedclassifiersothatitcanworkinconcertwithor standinfortheneuralnetworkinthelargerobject-orientedplatform.Althoughthisremains workinprogress,weareaimingataviablescientificapplicationofD-waveoutputinthenear term.Ourcontributionsincludethedemonstrationoftreecoverclassification,alongwitha detailedanalysisoftrainingonourremotesensingdataandashortcutsolutiontoembedthis classofproblemsintotheD-wavearchitecture.Interalia,wediscoveredsomesimple,classi- callyfast-to-trainquadraticdecisionstumpsonderivedimagefeaturesthatthemselvespro- ducesurprisinglygoodclassificationoftreecoverinCalifornia.Forpointofreference, antecedentcasestudiesofpotentialD-waveapplicationsinclude[20–27],while[28]presentsa broadcollectionofpotentialapplicationsofinteresttoNASA. Thepaperwillproceedasfollows.Wefirstreviewthestructureofproblemsamenableto solutionsontheD-wavequantumannealingprocessor.Mathematically,theyarequadratic unconstrainedbinaryoptimization(QUBO)problems,andinphysics,theyaregeneralized Isingmodelsofaspinglass.WediscussQBoostinthiscontext,theproblemofembedding intotheD-wavearchitecture,andourproposedmodificationstoQBoost.Wepresentthe detailsofourimplementationontheNAIPdataset,layingoutthetwoproblems,oneon108 qubits,anotheron508qubits,whicharethefocusofthisstudy,alongwithourresults.We concludewithadiscussionofchallengesandpossibleimprovementstothisframework. QuantumannealingontheD-waveprocessor InquantummechanicstheenergyfunctionisknownastheHamiltonian,denotedH.It encodesalldynamicsofasystemandwillvarywithtimetalongwithambientconditions.The basicprocessofquantumannealingistointerpolatephysicallybetweenaninitialHamilitonian H ,withaneasy-to-implementminimalenergyconfiguration(orgroundstate),andaproblem 0 HamiltonianH whoseminimalconfigurationissought.Forinstance,foralinearinterpola- P tionscheduleandcomputationtimeτ, (cid:16) (cid:17) t t HðtÞ¼ 1(cid:0) H þ H : ð1Þ t 0 t P TheinterpolationiseffectedphysicallyontheD-wavechipbyadjustingcurrentsthatflow toindividualqubits,eachofwhichisatinysuperconductingcircuit.Thesystembeginsinthe groundstateofH andends,ideally,inthegroundstateofH .Forperfectlyisolatedquantum 0 P systems,thegroundstateofH canbeattainedforsufficientlylargeτwithprobabilityarbi- P trarilyclosetoone.Inpractice,duetothermalnoiseandlossofquantumcoherence,optimal PLOSONE|DOI:10.1371/journal.pone.0172505 February27,2017 3/22 DeployingaquantumannealingprocessortodetecttreecoverinaerialimageryofCalifornia computetimesintheD-wavedeviceareactuallylessthanitscurrentlyminimalallowable time,τ=20μs.[11]Inthiscontext,itshouldbenotedthattheparameterτcapturesonlythe actualannealingtimeanddoesnotincludetimesforcooling,initialization,andreadoutofthe device. Becauseofthefacilityofphysicalcontrolattainablewithbinaryqubitsandpairwiseinterac- tionsbetweenthem,theproblemHamiltoniantakestheform: X X H ¼(cid:0) hs (cid:0) J ss: P i i ij i j ð2Þ i2V fi;jg2E InphysicsthisHamiltonianwasfirststudiedastheIsingmodelofamagnet.Thebinary variabless 2{−1,+1}arethuscalledspins,fixedinalatticegraphGwithverticesandedges i ðV;EÞ.Theprogrammableelementsarethelocalmagneticfields,h,andthecouplings i betweenspins,J .Bothareinprinciplecontinuumrealvariablesbutareinpracticelimitedto ij adiscretumbynoiseinthedevice.TheoptimizationseekstheminimumofH overallconfig- P urationsofthespins{s}. i Theintuitionfortheoptimizationisasfollows:Thenegativesigninthefirsttermindicates thattheenergyislowerwhenaspins alignswith(hasthesamesignas)themagneticfieldh at i i latticesitei;thisimperativecompeteswiththedemandthats alignoranti-alignwithneighbor- i ingspinss,accordingtothesignofthecouplingJ .IfJ >0,thecouplingbetweenspinsisfer- j ij ij romagnetic,drivingthemtoalign.IfJ <0,thecouplingisantiferromagnetic,drivingthemto ij anti-align.TheproblemofminimizingtheIsingenergyfunctionwithantiferromagneticcou- plingsisknowntobeNP-hard,meaningthatthecomputationaleffortrequiredforthehardest instancesscalesexponentiallywithproblemsizeforallknownclassicalalgorithms[29,30]. ComputationontheD-waveisfirstaprocessofmappingtheproblemtotheIsingstruc- ture,binaryandquadratic,thenembeddingitintotheavailablequbitlattice.OntheD-wave thequbitsarearrangedaccordingtoachimeragraph,asillustratedinFig1.Eachqubitcouples tofiveorsixothers,exceptwheretherearedefectsduetofaultyqubits.Iftheproblemdoesn’t Fig1.ChimerastructureofqubitconnectivityontheD-wave2Xprocessor.Thefull1152-qubitgraph extendstoa12x12latticeofgroupsofeightqubits.Withintheillustratedsubset,currentlyinoperablequbits aremarkedinred. doi:10.1371/journal.pone.0172505.g001 PLOSONE|DOI:10.1371/journal.pone.0172505 February27,2017 4/22 DeployingaquantumannealingprocessortodetecttreecoverinaerialimageryofCalifornia embeddirectly,auxiliaryqubitscanbeintroducedtoaugmenttheavailablecouplings,atasig- nificantcostinqubits.Bothmappingandembeddingimplyrestrictionsonthetypesofprob- lemsthatcanprofitablybetackledwiththeD-waveprocessor.Wewillinvestigatetheseissues inthecontextoftheQBoostalgorithm.Forathoroughrecentstudy,andformoredetailson quantumannealingintheD-waveprocessor,see[22,28]. Boosting Boostingisthetacticofbuildingastrongclassifierasanoptimallyweightedcombinationof weakclassifiers,eachofwhichmayclassifyonlymoderatelybetterthanrandomguessingon itsown.Iftheweakclassifiersarelinearintheinputfeatures,theboostedclassifiercarvesouta piecewise-planardecisionsurfacethatis,ifnottothesamedegreeasthatexpressedbyaneural network,effectivelynonlinear.In2008Neven,Denchev,Rose,andMacreadyproposeda boostedclassifier,christenedQBoost,thatcouldbetrainedonaD-waveprocessor[15].Given Nbinaryweakclassifiersc,i=1...N,eachofwhichclassifiesadatasampletaccordingto i c(t)2{−1,+1},theysoughtastrongclassifierofform i ! XN CðtÞ¼sign wcðtÞ : ð3Þ i i i¼1 Theauthorsachievedtheirbesttestresultswithbinaryweights,w 2{0,1},inwhichcase i thestrongclassifierissimplyanoptimalvotingsubsetofweakclassifiers.Thenaturalcost functiontomatewiththeD-wavearchitectureisaregulatedquadraticloss.ForasetToftrain- ingsamples,witheachelementthavingbeenassignedatraininglabely(t)2{±1},atraining problemcanbeposedasfollows: ( ! ) X XN 2 XN Find : min wcðtÞ(cid:0) yðtÞ þl w : ð4Þ fwi;lg i i i t2T i¼1 i¼1 Theregularizationtermgovernedbytheparameterλisintendedtoimprovegeneralization andspeedinexecutionbykeepingthefinalclassifiercompact.Thenormalizationoftheweak classifiersisthenadjustedsoasnottoundulypenalizelargepositivemarginsfromthedecision hypersurface, XN cðtÞ2f(cid:0) 1=N;þ1=Ng $ (cid:0) 1(cid:20) wcðtÞ(cid:20)1: ð5Þ i i i i¼1 Thetrainingproblemthusformulatedisoneofquadraticunconstrainedbinaryoptimiza- tion(QUBO).Intheirinitialtestsofthealgorithm,Nevenetal.optimizedtheQUBOproblem directlyusingclassicalheuristicsolvers.ComparingwithAdaboost,theyfoundmodest improvementsinclassificationaccuracyandsignificantimprovment(oforder50%)incom- pactnessoftheboostedclassifiers. ToconverttheQUBOtoIsingform,onemakesthetransformations =2w −1.Thenew i i variabless takevaluess =±1.Expandingthequadratic,thecostfunctionbecomes i i ! ! X X X X l(cid:0) 2 cðtÞyðtÞ w þ cðtÞcðtÞ ww þconst i i i j i j i t2T i;j t2T ! ! ð6Þ X l X 1X 1X X ! (cid:0) cðtÞyðtÞþ cðtÞcðtÞ s þ cðtÞcðtÞ ss þconst0: 2 i 2 i j i 2 i j i j i t2T j;t2T i>j t2T PLOSONE|DOI:10.1371/journal.pone.0172505 February27,2017 5/22 DeployingaquantumannealingprocessortodetecttreecoverinaerialimageryofCalifornia Inthelatterequation,anextrafactoroftwointhequadratictermcompensatesforrewrit- ingthesumtopassoverallindexpairs(i,j)onceonly.Wecanthenidentifythemagneticfields andcouplingsoftheIsingframeHamiltonian(Eq(2)), l X 1X h ¼(cid:0) þ cðtÞyðtÞ(cid:0) cðtÞcðtÞ ð7Þ i 2 i 2 i j t2T j;t2T 1X J ¼(cid:0) cðtÞcðtÞ ð8Þ ij 2 i j t2T TheconstantsdroppedfromEq(6)donotaffecttheoptimization.Onecanreadilyinter- prethowvarioustermsinfluencetheconstructionofthestrongclassifier.Thecontribution ∑ c(t)y(t)toh describeshowwelltheoutputc(t)ofaweakclassifiercorrelatestothe t2T i i i traininglabelsy(t)overthetrainingsetT.Iftheycorrelatewell,theygiveastrongpositive contributiontothemagneticfield,drivingthespintobepositive.Apositivespinindicates thatthecorrespondingweightisequaltoone:Theweakclassifier’svoteistabulatedinthe P finalstrongclassifier.ThecouplingJ ¼(cid:0) 1 cðtÞcðtÞlikewisedescribesthecorrelation ij 2 t2T i j ofweakclassifiersc andc overthetrainingset.Ifthetwoweakclassifierscorrelatewell, i j J <0.Thespinss ands tendtooppositevalues,meaningoneandnottheotherwouldbe ij i j includedinthefinalstrongclassifier.Thisisasitshouldbe.Towhateverextenttheycorre- late,theysupplyredundantinformationonthedata. Embeddingintothechimeragraph TheQBoostprocedureresultsinafully-connectedIsingproblem,witheachs coupledtoevery i others bya(generically)non-zeroJ .TorunontheD-waveprocessortheproblemneedsto j ij beembeddedintothechimeragraph.Themaximaldegreeofthechimeragraphissix.The fullyconnectedIsingproblemonNspinsconstitutesagraphofdegreeN−1.Nonethelessthe lattercanbeembeddedintotheformerbymappingeachspinnottoanindividualqubitbutto aconnectedsubgraphofqubits,suchthateverysubgraph(correspondingtoans)isconnected i byatleastonechimeragraphedgetoeveryothersubgraph(correspondingtoans)[31].The j graphedgesbewteensubgraphscanbeassignedtheproblemcouplingsJ .Withinasubgraph, ij internalgraphedgescanbeassignedlarge,ferromagneticcouplingsJ toimposethecondition F thatallqubitsassociatedtoagivenspinalign,encodingoneandthesamespinstate. Thisembeddingcomesatahighcostinqubits.Sinceeachauxiliaryqubitinachimerasub- graphcouplestoatmostdotherqubits,thesubgraphsizemustscalewithNtoprovidesuffi- cientcouplingstoothersubgraphs.AstherearenecessarilyNsubgraphs,theembedding overheadinqubitsscalesquadraticallywiththenumberofspinsN.Fortheexplicitexamples studiedrecentlyin[32],N=30wasthelargestfully-connectedproblemembedableina 512-qubitchimeragraph.Muchrecentwork[22,24,33–35]hasgoneintothisandrelated embeddingschemes,examiningmappingsoflogicalqubitstophysicalqubitsubgraphs,opti- malsettingsfortheinternalcouplingsJ ,thedistributionofproblemcouplingsJ among F ij graphedges,andmoregenerallyseekingproblemsthatarelessthanfullyconnectedandthere- foremoreamenabletoembeddinginthechimeragraph.Improvingtheconnectivityofhard- waregraphswillbecriticaltobroadeningthescopeofproblemssolvableonfuturequantum annealers. Intheir2009demonstrationofaQBoostclassifiertrainedtodetectcarsinstreetscenes [17],Nevenetal.embeddedviaadifferentapproach.TheymappedeachIsingspintoasingle qubitanddiscardedvaluesJ thatdidn’tembedintothechimeragraph.Tothispurposethey ij PLOSONE|DOI:10.1371/journal.pone.0172505 February27,2017 6/22 DeployingaquantumannealingprocessortodetecttreecoverinaerialimageryofCalifornia designedagreedyheuristicthatassignsspinstoqubitsinsuccession,eachspintothequbit whichwillmaximizetheedgeweightretained(thesumofthemagnitudesoftheembeddedJ ) ij withrespecttothepreviouslyembeddedspins.Underthisschemetheyretained11%oftotal edgeweightona52-qubitembedding.(Only52qubitswerefunctioningontheavailableD- waveprocessor,andtheyiteratedtrainingstepstogrowalargerclassifier.) Thisstrategydoesnotscale.Droppingtoohighaproportionofcouplingsleadstoasce- narioinwhicheachspinvariablecanbeoptimizedindependentoftheothers.If,foragiven spins ,themagneticfieldh isbiggerthanthesumofcouplingstootherspinsjretainedinthe a a embeddedlatticegraph,i.e., X if jh j> jJ j; a aj ð9Þ fa;jg2E thevalueofs intheoptimalsolutionisdeterminedsimplybythesignofh .Thiscanbeseen a a byconsideringthetotalcontributiontotheenergyduetospins ,namely, a X E ¼(cid:0) h s (cid:0) J s s: a a a aj a j ð10Þ fa;jg2E Asinthepreceedingequation,thesumhererunsoverthecoupledspinsj.Ifthespins is a anti-alignedwithitsmagneticfield,thefirsttermcontributes−h s =+|h |totheenergy.Flip- a a a pingthesignofs willdecreasethecontributionfromthattermby−2|h |.Atthesametime, a a thesecondtermisbounded, X X X (cid:0) jJ j(cid:20)(cid:0) J s s (cid:20) jJ j; aj aj a j aj ð11Þ fa;jg2E fa;jg2E fa;jg2E andsoflippingthesignofs ,regardlessoftheconfigurationoftheotherspins{s},imposesan Pa j energycostofatmostþ2 jJ j.Whenthecondition(9)holds,flippingthespinleadsto fa;jg2E aj anetdecereaseofenergy,andsothespinnecessarilyalignswithitsmagneticfield. Theconsequencesaretwo-fold.First,onecandeterminetheoptimalconfigurationofsuch spinssimplybycheckingthesignsoftheirmagneticfields.Thisisnotataskthatcallsfora quantumcomputer.Theimplicationfortheclassifieristhelossoffinebalancethatwastobe achievedamongallpossibleweakclassifiers.Weseektoretainonlytheminimalsetofweak classifiersthatcapturestheimportantfeaturesofthedata,butweakclassifierswhosespinssat- isfycondition(9)willbeincludedorexcludedirrespectiveoftheinclusionofothers. Unfortunately,thisscenarioistobeexpectedasthetotalnumberNofinputweakclassifiers growslarge.Thebasemotivationforquantumcomputingisthehopethatruntimeswillscale betterthanforclassicalalternativeswiththenumberofinputvariables.Theeffortonly becomesjustifiedonproblemswiththousandsortensofthousandsofbinaryvariables.Atthe sametime,thenumberofconnectionsbetweenqubits(fiveorsixinthecaseoftheD-wave chimeragraph)islikelytoremainsmall,duetothechallengeofbuildingandcontrollinginter- actionsbetweenmorethanafewbasicphysicalentities.Foraproblemwithaninitiallyfully connectedgraph,asimpleone-variable-to-one-qubitembeddingwilldiscardthousandsor tensofthousandsofcouplingsagainstsomesomesmallfinitenumberretained.Anycomputa- tionalproblemthatbeginsbyimposingaquadraticlossfunctiononalinearcombinationof binaryvariables,asinEq(4),resultsinafullyconnectedgraph.Whilesomecouplingsmay turnouttobezero,genericallyeveryspincouplestoeveryotherspin. WecanmaketheseconsiderationsmoreexplicitbyconsideringthescalingwithNofthe varioustermsinIsingHamiltonian.Exceptinthecasethattheaccuracyofweakclassifiersis tunedcloseto50%,thecorrelations∑ c(t)y(t)willbeO(|T|/N),with|T|thesizeofthe t2T i PLOSONE|DOI:10.1371/journal.pone.0172505 February27,2017 7/22 DeployingaquantumannealingprocessortodetecttreecoverinaerialimageryofCalifornia trainingset.Forinstance,inourimplementationfortreecoverclassification,theaveragetrain- ingerrorofthelinearweakclassifiersis25%.Aweakclassifierwith25%trainingerrorwould have X ciðtÞyðtÞ¼:25jTjð(cid:0) 1=NÞþ:75jTjðþ1=NÞ¼:5jTj=N: ð12Þ t2T TheNappearsherethroughthenormalizationgiveninEq(5).Thisleveloftrainingaccu- racyimpliesalsothattheweakclassifiersarewellcorrelatedamongthemselves,withcorrela- tionsthatscaleas (cid:18) (cid:19) X jTj c ðtÞcðtÞ(cid:24)O : ð13Þ a j N2 t2T LettingkbethemaximumnumberofcouplingsbetweenqubitsinthegraphG ¼ðV;EÞ, forlargeNwehavetheoverallscalingrules: (cid:12) (cid:12) (cid:12)(cid:12) l X 1X (cid:12)(cid:12) (cid:18)jTj(cid:19) l jh j¼(cid:12)(cid:0) þ cðtÞyðtÞ(cid:0) cðtÞcðtÞ(cid:12)(cid:24)O (cid:6) ð14Þ a (cid:12) 2 i 2 i j (cid:12) N 2 t2T j;t2T X X(cid:12)(cid:12)(cid:12) 1X (cid:12)(cid:12)(cid:12) (cid:18)kjTj(cid:19) jJ j¼ (cid:12)(cid:0) c ðtÞcðtÞ(cid:12)(cid:24)O : ð15Þ aj (cid:12) 2 a j (cid:12) N2 fa;jg2E fa;jg2E t2T Sincetheregulatorisfixedonceforallspinsandkisfinite,agenericspinwillsatisfythe decouplingcondition(9), X jh j> jJ j; a aj fa;jg2E asNgrowslarge. Wecircumventedthesedifficulties,intheheuristicembeddingschemeofNevenetal.,by rescalingtheretainedcouplingsJ tocompensateforthoselost.ThedynamicsofIsingferro- ij magnets,inwhichlong-rangeorderappearsinsystemswithonlylimited,localinteractions, gaveusreasontohopethatasubsetoffiveorsixofN−1couplings,ifappropriatelyrescaled, wouldbesufficienttomaintainthecharacteristicbalancesoughtbetweentheweakclassifiers. Absentaprincipledwaytocomputearescalingonaspin-by-spinbasis,werescaledallcou- plingsbyaconstantfactorαwhichwetreatedasanewvariationalmetaparameter.Intuitively, αshouldworkouttobetheratiooflosttoretainedcouplings,α*N/5.(Thecurrentproces- sorisconstructedonan1152-vertexchimeragraph,with55currentlyinoperablequbits,mak- ingtheaveragenumberofviableedges5.6.Becausetheembeddingheuristicmaximizesthe sumofmagnitudesofretainedcouplingsinpreferencetotheirnumber,theresultingembed- dingsarenotmaximallydense.Ourembeddingstypicallyretainanaverageofbetweenfour andfivecouplingsperqubit.)Aplotofvalidationerroragainstthemetaparametersofour 108-qubitproblem,definedbelowinthesection“TreeCoverClassification,”isshowninFig2. pffiffiffi Steppingαbyfactorsof 2fromN/64toN,wefindthesolutionofoveralllowestvalidation pffiffiffi pffiffiffi errorfora2fN 2=8;N=4;N 2=4g.Thismatcheswellwithourexpectationsforαandsitu- atestheoptimalclassifierintheregimewherethecouplingsandmagneticfieldsshouldhave comparable,competinginfluenceontheoptimization.Moreover,wecanseeinthereturned classifierstheincreasinginfluenceofthecouplingswithincreasingα.Whenαisverysmall, theoptimizationisgovernedbythemagneticfieldsandtheresultingclassifiersconsist PLOSONE|DOI:10.1371/journal.pone.0172505 February27,2017 8/22 DeployingaquantumannealingprocessortodetecttreecoverinaerialimageryofCalifornia Fig2.Validationerrorasafunctionofthecouplingrescalingfactorandregulatorforthe108-qubit problem.Theregulatorisexpressedintermsofanewparameterf:λ=2f|T|/N.Foreachpair(α,f),the problemwasoptimizedwith1000callstotheD-waveprocessorandtheclassifierofminimalvalidationerror pffiffi pffiffi recorded.Theoverallminimalerrorof9%,indeepestblue,isattainedfora2fN 2;N;N 2g. 8 4 4 doi:10.1371/journal.pone.0172505.g002 predominantlyofthoseweakclassifierswhichindividuallyhavelowesttrainingerror.Towit, theclassifiersreturnedatthefoursmallestvaluesofαshareincommonthetwelveweak classifierswiththetwelvelowesttrainingerrors;whereas,theoptimalclassifierrealizedfor pffiffiffi pffiffiffi a2fN 2=8;N=4;N 2=4gincludesonlytwoofthosetwelve;andtheclassifieratα=N includesoneofthetwelve.Whenwecometoourresults,wewillexploretheseeffectsandthe propertiesoftheoptimalclassifierinmoredetail. Incorporatingthenewrescalingfactor,theenergyfunctiontobeminimizedacrossvari- ables{s,α,λ},becomes,finally, i X X H ¼(cid:0) hs (cid:0) a J ss; P i i ij i j i2V fi;jg2E l X 1X h ¼(cid:0) þ cðtÞyðtÞ(cid:0) cðtÞcðtÞ ð16Þ i 2 i 2 i j t2T j;t2T 1X J ¼(cid:0) cðtÞcðtÞ: ij 2 i j t2T WewillrefertotheprocessoftruncationandrescalingoftheproblemHamiltonianasa renormalization,anabuseofasuggestivetermfromstatisticalphysics.Inthinkingthrough thisapproach,itisworthrememberingthatwehadalreadydeviatedfromthemostnatural definitionofthetrainingproblematthepointofimposingaquadraticobjectivefunctionin placeofthetotalnumberofmisclassifiedtrainingsamples(L vs.L norm).Wedeviatedagain 2 0 whenweregularizedthequadraticfunction.ThechoiceofL overL normismadehabitually 2 0 ongroundsofcomputationaltractabilityandjustifiedexpostfactobytheutilityofthesolu- tionsthatresult.Likewisehere,welooktotheaccuracyoftheresultingclassifierstojustifythis reformulationoftheoriginaloptimizationproblem.Themostaccurateclassifierfoundforour 108-qubitproblemusingtherenormalizedHamiltonianEq(16)hasavalidationerrorrateof PLOSONE|DOI:10.1371/journal.pone.0172505 February27,2017 9/22 DeployingaquantumannealingprocessortodetecttreecoverinaerialimageryofCalifornia 9.00%.Thiscomparestoanerrorrateof10.13%forthebestsolutionfoundviasimulated annealingontheoriginalQBoostcostfunction.Wehavefoundthatthefinalclassifiercanbe improvedifselectedbyvalidationinpost-processingfromamongtheoutputsreturnedbythe annealingprocess,andwedosoasmatterofcourse,althoughourresultsindicatethatthe effectdiminishesforclassifiersoflargercardinality. Twofinaldetailsoftheimplementationbearmentioninthecontextoftheembedding,for bothofwhichwetakecuesfromtheoriginalreportonQBoost[15].Alongwiththerescaling factorα,theregulatorλmustbedeterminedintraining.Beforesubmittingaproblemforopti- mization,wespecifytheregulatorintermsofanewparameterf, l fjTj ¼ : ð17Þ 2 N Here,again,|T|isthenumberoftrainingsamplesandNthenumberofinputweakclassifi- ers.Themetaparameters(α,f)arechosenbyactingtheoutputstrongclassifiersona 3000-samplevalidationset.(Thisstepiscoincidentwiththepost-validationstepmentionedin thepreviousparagraph.)Ourpracticehasbeentodeterminethefractionfinitiallybyacoarse parameterscanandthentoretestwithfinerstepsizesaroundtheminimuminf.Thecardinal- ityofweakclassifiersinthestrongclassifieranditserrorratedependstronglyonf,asshown inFig3withαfixedatN/4.TheeffectoftheregulatorforgeneralαcanbeseeninFig2. Beyondenforcingcompactness,theregulatorevidentlyplaysanimportantroleinminimizing classifiertrainingorvalidationerror.Withweakclassifiersnormalizedsothatc(t)2{−1/N, i +1/N},thequadraticloss, ! XN 2 L¼ wcðtÞ(cid:0) yðtÞ ; ð18Þ i i i¼1 Fig3.Minimumvalidationerrorandweakclassifiersretainedasafunctionoftheregulatorλ=2f|T|/N. Foreachftheproblemwasoptimizedwith500callstotheD-waveprocessor,andsubsequently(inset),with 1000callsinresolvingtheminimum,withαfixedatN/4. doi:10.1371/journal.pone.0172505.g003 PLOSONE|DOI:10.1371/journal.pone.0172505 February27,2017 10/22

Description:
quantum annealing processor to detect tree cover in aerial imagery of California. PLoS ONE 12(2): e0172505. doi:10.1371/journal.pone.0172505. Editor: Shijo Joseph, Kerala Forest Research. Institute, INDIA. Received: June 29, 2016. Accepted: February 6, 2017. Published: February 27, 2017.
See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.