brain emulation roadmap report

Upload: -

Post on 11-Oct-2015

24 views

Category:

Documents


0 download

DESCRIPTION

dvsvgdfvgzs

TRANSCRIPT

  • Whole Brain Emulation A Roadmap

    (2008)TechnicalReport#20083

    AndersSandberg*NickBostrom

    FutureofHumanityInstitute

    FacultyofPhilosophy&JamesMartin21stCenturySchoolOxfordUniversity

    CITE:Sandberg,A.&Bostrom,N.(2008):WholeBrainEmulation:ARoadmap,TechnicalReport#20083,Futureof

    HumanityInstitute,OxfordUniversityURL:www.fhi.ox.ac.uk/reports/20083.pdf

    (*)Correspondingauthor:[email protected]

  • 2

    In memoriam: Bruce H. McCormick (1930 2007)

  • 3

    ContentsWholeBrainEmulation............................................................................................................................1 ARoadmap ................................................................................................................................................1

    Inmemoriam:BruceH.McCormick(19302007)...........................................................................2 Contents..................................................................................................................................................3 Introduction ...........................................................................................................................................5 Thanksto ............................................................................................................................................6

    Theconceptofbrainemulation..........................................................................................................7 Emulationandsimulation...............................................................................................................7 Littleneedforwholesystemunderstanding...............................................................................8 Levelsofemulationandsuccesscriteria.....................................................................................10 Scaleseparation...............................................................................................................................12 Simulationscales.............................................................................................................................13 WBEassumptions ...........................................................................................................................15

    Roadmap ..............................................................................................................................................16 Requirements...................................................................................................................................16 Linkages............................................................................................................................................19 Roadmap ..........................................................................................................................................20 Technologydrivers.........................................................................................................................23 Uncertaintiesandalternatives......................................................................................................24 Alternativepathways.....................................................................................................................27 Relatedtechnologiesandspinoffs ..............................................................................................28

    Issues.....................................................................................................................................................30 Emulationsystems..........................................................................................................................30 Complicationsandexotica ............................................................................................................31 Summary ..........................................................................................................................................39

    Scanning ...............................................................................................................................................40 Embedding,fixationandstainingtechniques ...........................................................................52 Conclusion .......................................................................................................................................53

    Imageprocessingandscaninterpretation......................................................................................55 Geometricadjustment....................................................................................................................55 Noiseremoval .................................................................................................................................56 Datainterpolation...........................................................................................................................56 Celltracing.......................................................................................................................................57 Synapseidentification....................................................................................................................59 Identificationofcelltypes .............................................................................................................60 Estimationofparametersforemulation.....................................................................................61 Connectivityidentification............................................................................................................62 Conclusion .......................................................................................................................................63

    Neuralsimulation ...............................................................................................................................64 Howmuchneurondetailisneeded?...........................................................................................64 Neuralmodels .................................................................................................................................66 Simulators ........................................................................................................................................70 Parallelsimulation..........................................................................................................................70 Currentlargescalesimulations....................................................................................................71 Conclusion .......................................................................................................................................72

    Bodysimulation ..................................................................................................................................74 Conclusion .......................................................................................................................................75

    Environmentsimulation....................................................................................................................76

  • 4

    Vision ................................................................................................................................................76 Hearing .............................................................................................................................................77 SmellandTaste ...............................................................................................................................77 Haptics..............................................................................................................................................77 Conclusion .......................................................................................................................................78

    Computerrequirements ....................................................................................................................79 Conclusions......................................................................................................................................81

    Validation.............................................................................................................................................82 Discussion ............................................................................................................................................83 AppendixA:Estimatesofthecomputationalcapacity/demandsofthehumanbrain ..........84 AppendixB:ComputerPerformanceDevelopment ....................................................................86 ProcessingPower............................................................................................................................86 Memory ............................................................................................................................................95 Discdrives........................................................................................................................................97 Future................................................................................................................................................98

    AppendixC:Largescaleneuralnetworksimulations...............................................................101 AppendixD:Historyandpreviouswork.....................................................................................105 AppendixE:Nondestructiveandgradualreplacement...........................................................107 NonDestructiveScanning ..........................................................................................................107 Gradualreplacement....................................................................................................................108

    AppendixF:Glossary.......................................................................................................................110 References ..........................................................................................................................................113

  • 5

    IntroductionWholebrainemulation(WBE),thepossiblefutureonetoonemodellingofthefunctionofthehumanbrain,isacademicallyinterestingandimportantforseveralreasons:

    Researcho Brainemulationisthelogicalendpointofcomputationalneurosciences

    attemptstoaccuratelymodelneuronsandbrainsystems.o Brainemulationwouldhelpustounderstandthebrain,bothintheleadup

    tosuccessfulemulationandafterwardsbyprovidinganidealtestbedforneuroscientificexperimentationandstudy.

    o Neuromorphicengineeringbasedonpartialresultswouldbeusefulinanumberofapplicationssuchaspatternrecognition,AIandbraincomputerinterfaces.

    o Asalongtermresearchgoalitmightbeastrongvisiontostimulatecomputationalneuroscience.

    o Asacaseoffuturestudiesitrepresentsacasewherearadicalfuturepossibilitycanbeexaminedinthelightofcurrentknowledge.

    Economicso Theeconomicimpactofcopyablebrainscouldbeimmense,andcouldhave

    profoundsocietalconsequences(Hanson,1994,2008b).Evenlowprobabilityeventsofsuchmagnitudemeritinvestigation.

    Individuallyo Ifemulationofparticularbrainsispossibleandaffordable,andifconcerns

    aboutindividualidentitycanbemet,suchemulationwouldenablebackupcopiesanddigitalimmortality.

    Philosophyo Brainemulationwoulditselfbeatestofmanyideasinthephilosophyof

    mindandphilosophyofidentity,orprovideanovelcontextforthinkingaboutsuchideas.

    o Itmayrepresentaradicalnewformofhumanenhancement.WBErepresentsaformidableengineeringandresearchproblem,yetonewhichappearstohaveawelldefinedgoalandcould,itwouldseem,beachievedbyextrapolationsofcurrenttechnology.Thisisunlikemanyothersuggestedradicallytransformativetechnologieslikeartificialintelligencewherewedonothaveanyclearmetricofhowfarwearefromsuccess.InordertodevelopideasaboutthefeasibilityofWBE,groundtechnologyforesightandstimulateinterdisciplinaryexchange,theFutureofHumanityInstitutehostedaworkshoponMay26and27,2007,inOxford.Invitedexpertsfromareassuchascomputationalneuroscience,brainscanningtechnology,computing,nanotechnology,andneurobiologypresentedtheirfindingsanddiscussedthepossibilities,problemsandmilestonesthatwouldhavetobereachedbeforeWBEbecomesfeasible.Theworkshopavoideddealingwithsocioeconomicramificationsandwithphilosophicalissuessuchastheoryofmind,identityorethics.Whileimportant,suchdiscussionswouldundoubtedlybenefitfromamorecomprehensiveunderstandingofthebrainanditwasthisunderstandingthatwewishedtofocusonfurtheringduringtheworkshop.Suchissueswilllikelybedealtwithatfutureworkshops.

  • 6

    Thisdocumentcombinesanearlierwhitepaperthatwascirculatedamongworkshopparticipants,andadditionssuggestedbythoseparticipantsbefore,duringandaftertheworkshop.ItaimsatprovidingapreliminaryroadmapforWBE,sketchingoutkeytechnologiesthatwouldneedtobedevelopedorrefined,andidentifyingkeyproblemsoruncertainties.Brainemulationiscurrentlyonlyatheoreticaltechnology.Thismakesitvulnerabletospeculation,handwavinganduntestableclaims.AsproposedbyNickSzabo,falsifiabledesignisawayofcurbingtheproblemswiththeoreticaltechnology:

    thedesignersofatheoreticaltechnologyinanybutthemostpredictableofareasshouldidentifyitsassumptionsandclaimsthathavenotalreadybeentestedinalaboratory.Theyshoulddesignnotonlythetechnologybutalsoamapoftheuncertaintiesandedgecasesinthedesignandaseriesofsuchexperimentsandteststhatwouldprogressivelyreducetheseuncertainties.Aproposalthatlacksthisadmissionofuncertaintiescoupledwithdesignsofexperimentsthatwillreducesuchuncertaintiesshouldnotbedeemedcredibleforthepurposesofanyimportantdecision.Wemightcallthisrequirementarequirementforafalsifiabledesign.(Szabo,2007)

    Inthecaseofbrainemulationthiswouldmeannotonlysketchinghowabrainemulatorwouldworkifitcouldbebuiltandaroadmapoftechnologiesneededtoimplementit,butalsoalistofthemainuncertaintiesinhowitwouldfunctionandproposedexperimentstoreducetheseuncertainties.Itisimportanttoemphasizethelongtermandspeculativenatureofmanyaspectsofthisroadmap,whichinanycaseistoberegardedonlyasafirstdrafttobeupdated,refined,andcorrectedasbetterinformationbecomesavailable.Giventhedifficultiesanduncertaintiesinherentinthistypeofwork,onemayaskwhetherourstudyisnotpremature.Ourviewisthatwhenthestakesarepotentiallyextremelyhigh,itisimportanttoapplythebestavailablemethodstotrytounderstandtheissue.Evenifthesemethodsarerelativelyweak,itisthebestwecando.Thealternativewouldbetoturnablindeyetowhatcouldturnouttobeapivotaldevelopment.Withoutfirststudyingthequestion,howisonetoformanywellgroundedviewonewayortheotherastothefeasibilityandproximityofaprospectlikeWBE?

    Thanks to Wewouldliketowarmlythankthemanypeoplewhohavecommentedonthepaperandhelpedextendandimproveit:Workshopparticipants:JohnFiala,RobinHanson,KennethJeffreyHayworth,ToddHuffman,EugeneLeitl,BruceMcCormick,RalphMerkle,TobyOrd,PeterPassaro,NickShackel,RandallA.Koene,RobertA.FreitasJrandRebeccaRoache.Otherusefulcomments:StuartArmstrong.

  • 7

    TheconceptofbrainemulationWholebrainemulation,ofteninformallycalleduploadingordownloading,hasbeenthesubjectofmuchsciencefictionandalsosomepreliminarystudies(seeAppendixDforhistoryandpreviouswork).Thebasicideaistotakeaparticularbrain,scanitsstructureindetail,andconstructasoftwaremodelofitthatissofaithfultotheoriginalthat,whenrunonappropriatehardware,itwillbehaveinessentiallythesamewayastheoriginalbrain.

    Emulation and simulation Thetermemulationoriginatesincomputerscience,whereitdenotesmimickingthefunctionofaprogramorcomputerhardwarebyhavingitslowlevelfunctionssimulatedbyanotherprogram.Whileasimulationmimicstheoutwardresults,anemulationmimicstheinternalcausaldynamics(atsomesuitablelevelofdescription).Theemulationisregardedassuccessfuliftheemulatedsystemproducesthesameoutwardbehaviourandresultsastheoriginal(possiblywithaspeeddifference).Thisissomewhatsofterthanastrictmathematicaldefinition1.AccordingtotheChurchTuringthesis,aTuringmachinecanemulateanyotherTuringmachine.ThephysicalChurchTuringthesisclaimsthateveryphysicallycomputablefunctioncanbecomputedbyaTuringmachine.Thisisthebasisforbrainemulation:ifbrainactivityisregardedasafunctionthatisphysicallycomputedbybrains,thenitshouldbepossibletocomputeitonaTuringmachine.Eveniftrue,however,itdoesnotdemonstratethatitisacomputationallyfeasibleprocess.Inthefollowing,emulationwillrefertoa1to1modelwhereallrelevantpropertiesofasystemexist,whileasimulationwilldenoteamodelwhereonlysomepropertiesexist.Emulationsmaybehavedifferentlyfromeachotherortheoriginalduetonoiseorintrinsicchaos,butbehavewithintherangeofwhatonewouldexpectfromtheoriginalifithadexperiencedthesamenoiseorchaos.Byanalogywithasoftwareemulator,wecansaythatabrainemulatorissoftware(andpossiblydedicatednonbrainhardware)thatmodelsthestatesandfunctionaldynamicsofabrainatarelativelyfinegrainedlevelofdetail.Inparticular,amindemulationisabrainemulatorthatisdetailedandcorrectenoughtoproducethephenomenologicaleffectsofamind.

    1AstrictdefinitionofsimulationmightbethatasystemSconsistsofastatex(t)evolvingbyaparticulardynamicsf,influencedbyinputsandproducingoutputs:x(t+1)=f(I,x(t)),O(t)=g(x(t)).AnothersystemTsimulatesSifitproducesthesameoutput(withinatolerance)forthesameinputtimeseriesstartingwithagivenstate(withinatolerance):X(t+1)=F(I,X(t)),O(t)=G(X(t))where|x(t)X(t)|

  • 8

    Apersonemulationisamindemulationthatemulatesaparticularmind.Whattherelevantpropertiesareisacrucialissue.Intermsofsoftwareemulationthisisoftenthebitsstoredinmemoryandhowtheyareprocessed.Acomputeremulatormayemulatetheprocessor,memory,I/Oandsoonoftheoriginalcomputer,butdoesnotsimulatetheactualelectronicworkingsofthecomponents,onlytheirqualitativefunctiononthestoredinformation(anditsinteractionwiththeoutsideworld).Whilelowerlevelemulationofcomputersmaybepossibleitwouldbeinefficientandnotcontributemuchtothefunctionsthatinterestus.Dependingonthedesiredsuccesscriterionemulationmayrequiredifferentlevelsofdetail.Itmightalsousedifferentlevelsofdetailindifferentpartsofthesystem.Inthecomputerexample,emulatingtheresultofamathematicalcalculationmaynotrequiresimulatingtheexecutionofalloperatingsystemcallsformathfunctions(sincethesecanbedonemoreefficientlybytheemulatingcomputersprocessor)whileemulatingthebehaviourofananaloguevideoeffectmayrequireadetailedelectronicssimulation.

    Little need for whole-system understanding AnimportanthypothesisforWBEisthatinordertoemulatethebrainwedonotneedtounderstandthewholesystem,butratherwejustneedadatabasecontainingallnecessarylowlevelinformationaboutthebrainandknowledgeofthelocalupdaterulesthatchangebrainstatesfrommomenttomoment.Afunctionalunderstanding(whyisaparticularpieceofcortexorganizedinacertainway)islogicallyseparatefromdetailknowledge(howisitorganised,andhowdoesthisstructurerespondtosignals).FunctionalunderstandingmaybeapossibleresultfromdetailknowledeanditmayhelpgatheronlytherelevantinformationforWBE,butitisentirelypossiblethatwecouldacquirefullknowledgeofthecomponentpartsandinteractionsofthebrainwithoutgaininganinsightintohowtheseproduce(say)consciousnessorintelligence.Evenadatabasemerelycontainingthecompletepartslistofthebrain,includingthemorphologyofitsneurons,thelocations,sizesandtypesofsynapticconnections,wouldbeimmenselyusefulforresearch.Itwouldenabledatadrivenresearchinthesamewayasgenomicshasdoneinthefieldofcellbiology(Fiala,2002).Computationalneuroscienceattemptstounderstandthebrainbymakingmathematicalorsoftwaremodelsofneuralsystems.Currently,themodelsareusuallyfarsimplerthanthestudiedsystems,withtheexceptionofsomesmallneuralnetworkssuchasthelobsterstomatogastricganglion(NusbaumandBeenhakker,2002)andthelocomotornetworkofthelampreyspinalcord(Kozlov,Lansneretal.,2007).Oftenmodelsinvolveacombinationofsimplifiedparts(simulatedneuronsandsynapticlearningrules)andnetwork

    structures(subsamplingofbiologicalneurons,simpletopologies).Suchnetworkscanthemselvesconstitute

    Figure1:Understandingoffunctionvs.understandingofdetails.

  • 9

    learningorpatternrecognizingsystemsontheirown,artificialneuralnetworks(ANNs).ANNmodelscanbeusedtoqualitativelymodel,explainandanalyzethefunctionsofbrainsystems(Rumelhart,McClellandetal.,1986).Connectionistmodelsbuildmorecomplexmodelsofcognitionorbrainfunctiononthesesimplerparts.Theendpointofthispursuitwouldbemodelsthatencompassafullunderstandingofthefunctionofallbrainsystems.Suchqualitativemodelsmightnotexhibitintelligenceorthecomplexityofhumanbehaviourbutwouldenableaformalizedunderstandingofhowtheycomeaboutfromsimpleparts.Anotherapproachincomputationalneuroscienceinvolvescreatingmorebiologicallyrealisticmodels,whereinformationaboutthebiologicaldetailsofneuronssuchastheirelectrochemistry,biochemistry,detailedmorphologyandconnectivityareincluded.Atitssimplestwefindcompartmentmodelsofindividualneuronsandsynapses,whilemorecomplexmodelsincludemultiplerealisticneuronsconnectedintonetworks,possiblytakinginteractionssuchaschemicalvolumetransmissionintoaccount.Thisapproachcanbeseenasaquantitativeunderstandingofthebrain,aimingforacompletelistofthebiologicalparts(chemicalspecies,neuronmorphologies,receptortypesanddistributionetc.)andmodellingasaccuratelyaspossiblethewayinwhichthesepartsinteract.Giventhisinformationincreasinglylargeandcomplexsimulationsofneuralsystemscanbecreated.WBErepresentsthelogicalconclusionofthiskindofquantitativemodel:a1to1modelofbrainfunction.Notethattheamountoffunctionalunderstandingneededtoachievea1to1modelissmall.Itsbehaviourisemergentfromthelowlevelproperties,andmayormaynotbeunderstoodbytheexperimenters.Forexample,ifcoherentoscillationsareimportantforconceptualbindingandtheseemergefromthelowlevelpropertiesofneuronsandtheirnetworks,acorrectandcompletesimulationofthesepropertieswillproducethecoherence.Inpracticecomputationalneuroscienceworksinbetweenquantitativeandqualitativemodels.Qualitativemodelsareusedtoabstractcomplex,uncertainandpotentiallyirrelevantbiologicaldata,andoftenprovidesignificantimprovementsinsimulationprocessingdemands(inturnenablinglargersimulations,whichmayenableexplorationofdomainsofmoreinterest).Quantitativemodelsaremoreconstrainedbyknownbiology,chemistryandphysicsbutoftensufferfromanabundanceoffreeparametersthathavetobeset(Herz,Gollischetal.,2006).Hybridmodelsmayincludepartsusingdifferentlevelsofabstraction,orexistasafamilyofmodelsrepresentingthesamesystematdifferentlevelsofabstraction.Theinterplaybetweenbiologicalrealism(attemptingtobefaithfultobiology),completeness(usingallavailableempiricaldataaboutthesystem),tractability(thepossibilityofquantitativeorqualitativesimulation)andunderstanding(producingacompressedrepresentationofthesalientaspectsofthesysteminthemindoftheexperimenter)willoftendeterminewhatkindofmodelisused.Theappropriatelevelofabstractionandmethodofimplementationdependsontheparticulargoalofthemodel.InthecaseofWBE,thesuccesscriteriadiscussedbelowplacelittleemphasisonunderstanding,butmuchemphasisonqualitativelycorrectdynamics,requiringmuchbiologicalrealism(uptoapoint,setbyscaleseparation)andtheneedfordatadrivenmodels.Whethersuchmodelsforwholebrainsystemsaretractablefromamodellingandsimulationstandpointisthecrucialissue.Brainemulationcannotbeachievedwithoutsomefunctionalunderstanding.Itneedsmodelsandtheoriesforrecognizingwhatdataisrelevant,andwouldprovidedatafordevelopingandtestingthesefurther.WhileintheorybrainemulationmighthugthelowerlineinFigure1,inpracticeitwilllikelyoccursomewherealongtherightedgestillfarbelowafullunderstandingofthetoplevelphenomena,butincludingabroadunderstandingofmany

  • 10

    kindsoflowlevelphenomena.Wealsoneedsomeunderstandingofhigherlevelphenomenatotestoursimulationsandknowwhatkindofdataweneedtopursue.Fosteringtherightresearchcyclefordevelopingtherightunderstanding,collectingdata,improvinginstrumentation,andexperimentingwithlimitedemulationsinadditiontoprovidingusefulservicestorelatedfieldsandbeneficialspinoffswouldbeindispensableforthedevelopmentofWBE.

    Levels of emulation and success criteria Forthebrain,severallevelsofsuccesscriteriaforemulationcanbeused.Theyformahierarchyextendingfromlowleveltargetstocompleteemulation.SeeTable1onpage11.Notshowninthishierarchyareemulationsofsubsystemsorsmallvolumesofthebrain,partialbrainemulations.Properlyspeaking,acompletescan,partslistandbraindatabase(1a,1band2)donotconstitutesuccessfulbrainemulation,butsuchachievements(andpartialbrainemulations)wouldinanycasebeimportantmilestonesandusefulinthemselves.Similarly,thehighlevelachievementsrelatedtosocialroles,mentalstates,andpersonalidentify(6a,6band6c)arebothpoorlyunderstoodandhardtooperationalize,butgiventhe

    philosophicalinterestinWBEwehaveincludedthemhereforcompleteness.Itisnotobvioushowthesecriteriarelatetooneanother,ortowhatextenttheymightbeentailedbythecriteriafor4and5.Achievingthethirdsuccesscriterionbeyondacertainresolutionwould,assumingsomesuperveniencethesis,implysuccessofsomeoralloftheothercriteria.AfullquantummechanicalNbodyorfieldsimulationencompassingeveryparticlewithinabrainwouldplausiblysufficeevenifquantummindtheoriesarecorrect.Attheveryleasta1to1materialcopyofthebrain(asomewhatinflexibleandveryparticularkindofemulatingcomputer)appearstoachieveallcriteria,possiblyexceptingthosefor6c.However,thisis

    1a Parts List

    1b Complete scan

    2 Brain database

    3 Functional

    brain emulation

    4 Species

    generic brain emulation

    5 Individual brain

    emulation

    6a Social role-fit

    emulation

    6b Mind emulation

    6c Personal identity

    emulation

    Figure2:SuccesslevelsforWBE.

  • 11

    likelyanexcessivelydetailedlevelsincetheparticularphenomenaweareinterestedin(brainfunction,psychology,mind)appeartobelinkedtomoremacroscopicphenomenathandetailedatomicactivity.Giventhecomplexitiesandconceptualissuesofconsciousnesswewillnotexaminecriteria6abc,butmainlyexamineachievingcriteria15.Table1:SuccessCriteria

    Level Successcriterion Relevantproperties1a Partslist Aninventoryofallobjectsonaparticular

    sizescale,theirpropertiesandinteractions.Lowlevelneuralstructure,chemistry,dynamicsaccuratetoresolutionlevel.

    1b Completescan Acomplete3Dscanofabrainathighresolution.

    Resolution,informationenablingstructuretofunctionmapping.

    2 Braindatabase Combiningthescanandpartslistintoadatabasemappingthelowlevelobjectsofabrain.

    1to1mappingofscantosimulation/emulationobjects.

    3 Functionalbrainemulation

    Theemulationsimulatestheobjectsinabraindatabasewithenoughaccuracytoproduce(atleast)asubstantialrangeofspeciestypicalbasicemergentactivityofthesamekindasabrain(e.g.aslowwavesleepstateoranawakestate).

    Genericallycorrectcausalmicrodynamics.

    4 Speciesgenericbrainemulation

    Theemulationproducesthefullrangeof(human)speciestypicalemergentbehaviorandlearningcapacity.

    Longtermdynamicsandadaptation.Appropriatebehaviourresponses.Fullrangelearningcapacity.

    5 Individualbrainemulation

    Theemulationproducesemergentactivitycharacteristicofthatofoneparticular(fullyfunctioning)brain.Itismoresimilartotheactivityoftheoriginalbrainthananyotherbrain.

    Correctinternalandbehaviourresponses.Retainsmostmemoriesandskillsoftheparticularbrainthatwasemulated.(Inanemulationofananimalbrain,itshouldbepossibletorecognizetheparticular(familiar)animal.)

    6a Socialrolefitemulation/Personemulation

    Theemulationisabletofillandbeacceptedintosomeparticularsocialrole,forexampletoperformallthetasksrequiredforsomenormallyhumanjob.(Socioeconomiccriteriainvolved)

    Propertiesdependonwhich(rangeof)socialrolestheemulationwouldbeabletofit.Inalimitingcase,theemulationwouldbeabletopassapersonalizedTuringtest:outsidersfamiliarwiththeemulatedpersonwouldbeunabletodetectwhetherresponsescamefromoriginalpersonoremulation.

    6b Mindemulation Theemulationproducessubjectivementalstates(qualia,phenomenalexperience)ofthesamekindthatwouldhavebeenproducedbytheparticularbrainbeingemulated.(Philosophicalcriteriainvolved)

    Theemulationistrulyconsciousinthesamewayasanormalhumanbeing.

    6c

    Personalidentityemulation

    Theemulationiscorrectlydescribedasacontinuationoftheoriginalmind;eitherasnumericallythesameperson,orasasurvivingcontinuerthereof.(Philosophicalcriteriainvolved)

    Theemulationisanobjectofprudentiallyrationalselfconcernforthebraintobeemulated.

  • 12

    Scale separation Atfirstitmayappearunlikelythatacomplexsystemwithmanydegreesoffreedomlikethebraincouldbemodelledwiththerightcausaldynamics,butwithouttakingintoaccountthesmallestparts.Microstimulationofindividualneuronscaninfluencesensorydecisions(HouwelingandBrecht,2008),showingthatverysmalldisturbancescanundertherightcircumstancesscaleuptobehaviouraldivergences.However,statevariablesofcomplexsystemscanbequantitativelypredictedwhenthereisscaleseparation:whendifferentaspectsofthesystemexistonsufficiently(ordersofmagnitude)differentscales(ofsize,energy,timeetc),theycanbecomeuncoupled(Hillerbrand,2008).Atypicalexampleishowthemicroscopicdynamicsofalaser(atomsinteractingwithanoscillatingelectromagneticfield)givesrisetoamacroscopicdynamics(thegrowthanddecayofdifferentlasermodes)insuchawaythatanaccuratesimulationofthesystemusingonlyelementsonthemacroscaleispossible.Anotherexampleisthescaleseparationbetweenelectriccurrentsandlogicoperationsinacomputer,whichenablesbitbasedemulation.Whenthereisnoscaleseparation(suchasinfluidturbulence)macroscalepredictionsbecomeimpossiblewithoutsimulatingtheentiremicroscale.Animportantissuetobedeterminediswhethersuchacutoffexistsinthecaseofthehumanbrainand,ifitdoesexist,atwhatlevel.Whilethispaperphrasesitintermsofsimulation/emulation,itisencounteredinarangeoffields(AI,cognitiveneuroscience,philosophyofmind)inotherforms:whatleveloforganisationisnecessaryforintelligent,personal,orconsciousbehaviour?AkeyassumptionofWBEisthat,atsomeintermediarylevelofsimulationresolutionbetweentheatomicandthemacroscopic,thereexistsatleastonecutoffsuchthatmeetingcriteria1aand1batthislevelofresolutionalsoenablesthehighercriteriatobemet.Atsuchaspatial,temporal,ororganisationalscale,thedynamicsonthelarger/slowerscaleisnotfunctionallysensitivetothedynamicsofthesmaller/fasterscale.Suchscaleseparationmightoccuratthesynapticscale,wherethedetailedchemicaldynamicsunderlyingsynapticfunctioncouldbereplacedbyasimplifiedqualitativemodelofitseffectsonsignalsandsynapticstrengths.Anotherpossiblescaleseparationlevelmightoccurbetweenindividualmoleculesandmolecularconcentrationscales:moleculardynamicscouldbereplacedwithmassactioninteractionsofconcentrations.Aperhapslesslikelyseparationcouldalsooccuronhigherlevelsifwhatmattersistheactivityofcorticalminicolumnsratherthanindividualneurons.Afinallikelybutcomputationallydemandingscaleorseparationwouldbetheatomicscale,treatingthebrainemulationasaNbodysystemofatoms.Conversely,ifitcouldbedemonstratedthatthereisnosuchscale,itwoulddemonstratetheinfeasibilityofwholebrainemulation.Duetocausallyimportantinfluencefromsmallerscalesinthiscase,asimulationataparticularscalecannotbecomeanemulation.Thecausaldynamicsofthesimulationisnotinternallyconstrained,soitisnota1to1modeloftherelevantdynamics.Biologicallyinterestingsimulationsmightstillbepossible,buttheywouldbe

    localtoparticularscalesandphenomena,andtheywouldnotfullyreproducetheinternalcausalstructureofthewholebrain.

    Figure3:Sizescalesinthenervoussystem.

  • 13

    Simulation scales Thewidelyreproduceddiagramfrom(ChurchlandandSejnowski,1992)inFigure3depictsthevariouslevelsoforganisationinthenervoussystemorderedbysizescale,runningfromthemolecularleveltotheentiresystem.Simulations(andpossiblyemulations)canoccuronalllevels:

    Molecularsimulation(individualmolecules) Chemistrysimulation(concentrations,lawofmassaction) Geneticexpression Compartmentmodels(subcellularvolumes) Wholecellmodels(individualneurons) Localnetworkmodels(replacesneuronswithnetworkmodulessuchas

    minicolumns) Systemmodels

    AnotherhierarchywasintroducedbyJohnFialaduringtheworkshop,andwillbeusedwithsomeadaptationsinthisdocument.Table2:LevelsofemulationLevel 1 Computational

    moduleClassicAI,highlevelrepresentationsofinformationandinformationprocessing.

    2 Brainregionconnectivity

    Eacharearepresentsafunctionalmodule,connectedtoothersaccordingtoa(speciesuniversal)connectome(Sporns,Tononietal.,2005).

    3 Analognetworkpopulationmodel

    Neuronspopulationsandtheirconnectivity.Activityandstatesofneuronsorgroupsofneuronsarerepresentedastheirtimeaverages.ThisissimilartoconnectionistmodelsusingANNs,ratemodelneuralsimulationsandcascademodels.

    4 Spikingneuralnetwork

    Asabove,plusfiringproperties,firingstateanddynamicalsynapticstates.Integrateandfiremodels,reducedsinglecompartmentmodels(butalsosomeminicolumnmodels,e.g.(JohanssonandLansner,2007)).

    5 Electrophysiology Asabove,plusmembranestates(ionchanneltypes,properties,state),ionconcentrations,currents,voltagesandmodulationstates.Compartmentmodelsimulations.

    6 Metabolome Asabove,plusconcentrationsofmetabolitesandneurotransmittersincompartments.

    7 Proteome Asabove,plusconcentrationsofproteinsandgeneexpressionlevels.8 Statesofprotein

    complexesAsabove,plusquaternaryproteinstructure.

    9 Distributionofcomplexes

    Asabove,pluslocomeinformationandinternalcellulargeometry.

    10 Stochasticbehaviourofsinglemolecules

    Asaboveplusmoleculepositions,oramolecularmechanicsmodeloftheentirebrain.

    11 Quantum Quantuminteractionsinandbetweenmolecules.

    Theamountofunderstandingneededtoaccuratelysimulatetherelevantobjectstendstoincreaseradicallyforhigher(here,lownumbered)levels:whilebasicmechanicsiswellunderstood,membranebiophysicsiscomplex,andthecomputationalfunctionsofbrainareasarelikelyexceedinglymultifaceted.Conversely,theamountofcomputingpowerneededincreasesrapidlyaswedescendtowardslowerlevelsofsimulation,andmaybecomefundamentallyinfeasibleonlevel112.Theamountandtypedataneededtofullyspecifya

    2Butseealsothefinalchapterof(Hameroff,1987).Themainstumblingblockoflevel11simulationmaynotbecomputinghardwareorunderstandingbutfundamentalquantumlimitationsonscanning.

  • 14

    modelalsochangescharacterbetweenthedifferentlevels.Lowlevelsimulationsrequiremassivequantitiesofsimpleinformation(molecularpositionsandtypes)whereashigherlevelsrequireasmalleramountofverycomplexinformation(contentofmentalprocesses).Eachlevelhasitsowncharacteristicsizeandtimescale,restrictingtherequiredimagingresolutionandsimulationtimestep.Forexample,synapticspinenecksandthethinnestaxonscanbeontheorderof50nmorsmaller,requiringimagingontheorderofthe5nanometerscaletoresolvethem.AninformalpollamongworkshopattendeesproducedarangeofestimatesoftherequiredresolutionforWBEis.Theconsensusappearedtobelevel46.Twoparticipantsweremoreoptimisticabouthighlevelmodels,whiletwosuggestedthatelementsonlevel89maybenecessaryatleastinitially(butthatthebulkofmatureemulation,oncethebasicswereunderstood,couldoccuronlevel45).Toachieveemulationonthislevel,theconsensuswasthat5550nmscanningresolutionwouldbeneeded.Thisroadmapwillhencefocusonlevel46models,whilebeingopenforthatdeeperlevelsmayturnouttobeneeded.AsnotedbyFiala,WBElikelyrequiresatleastlevel4tocapturethespecificityofindividualbrains,butprobablyrequirescomplexityatlevel6orlowertofullycapturethecomputationalcontributionsofionchannels,secondmessengers,proteinleveladaptation,andstochasticsynaptictransmission.Otherparticipantsthoughtthatatleastlevel5wouldbeneededforindividualbrainproperties.

    ForecastingAnalysingtherequirementsforemulation(intermsofscanningmethod,numberofentitiestosimulate,resultingstoragerequirementsandcomputationaldemands)ateachofthelevelsprovidesawayofboundingprogresstowardsWBE.GiventheseestimatesandscenariosoffutureprogressinscanningandcomputingitispossibletocalculatetheearliestpointintimewherethereisenoughresourcestoproduceaWBEonagivenlevelatacertainprice.Asbetterinformationbecomesavailablesuchestimatescanberefined.Althoughanytimeestimateswillbesubjecttostronguncertainties,theycanbehelpfulinestimatinghowfarawayWBEisfrompolicyrelevanttimescales,aswellaslikelytimeframesforearlysmallscaleemulations.Theyalsoallowcomparisonstoothertechnologyforecasts,enablingestimationofthechancesforsynergies(e.g.thedevelopmentofmolecularnanotechnology,whichwouldaccelerateWBEprogress),reliability(e.g.thefurtherintothefuture,themoreunlikelyMooreslawistohold),andtheriskofothertechnologiesovertakingWBE(e.g.artificialintelligence).EarlyWBEmayrequirelowerlevelsimulationthanlaterforms,astheremightnotyetbeenoughexperience(andtestsystems)todeterminewhichsimulationelementsarestrictlynecessaryforsuccess.ThemainconcerninthisdocumentisestimatingwhenandhowWBEwillfirstbeachievedratherthanitseventualmatureorbestform.

  • 15

    WBE assumptions

    PhilosophicalassumptionsPhysicalism(everythingsupervenesonthephysical)isaconvenientbutnotnecessaryassumption,sincesomenonphysicalisttheoriesofmentalpropertiescouldallowthemtoappearinthecaseofWBE.Successcriterion6bemulationassumesmultiplerealizability(thatthesamementalproperty,state,oreventcanbeimplementedbydifferentphysicalproperties,states,andevents).SufficientapparentsuccesswithWBEwouldprovidepersuasiveevidenceformultiplerealizability.Generally,emulationuptoandincludinglevel6adoesnotappeartodependonanystrongmetaphysicalassumptions.

    ComputationalassumptionsComputability:brainactivityisTuringcomputable,orifitisuncomputable,theuncomputableaspectshavenofunctionallyrelevanteffectsonactualbehaviour.Nonorganicism:totalunderstandingofthebrainisnotrequired,justcomponentpartsandtheirfunctionalinteractions.Scaleseparation:atsomeintermediarylevelofsimulationresolutionbetweentheatomicandthemacroscopicthereexistsone(ormore)cutoffssuchthatmeetingcriterion2atthislevelissufficientformeetingoneormoreofthehighercriteria.Componenttractability:theactualbraincomponentsatthelowestemulatedlevelcanbeunderstoodwellenoughtoenableaccuratesimulation.Simulationtractability:simulationofthelowestemulatedleveliscomputationallytractablewithapracticallyrealizablecomputer.

    NeuroscienceassumptionsBraincenteredness:inordertoproduceaccuratebehaviouronlythebrainandsomepartsofthebodyneedtobesimulated,nottheentirebody.WBEappearstobeawayoftestingmanyoftheseassumptionsexperimentally.Inacquiringaccuratedataaboutthestructureandfunctionofthebrainandrepresentingitasemulationsitshouldbepossibletofindmajordiscrepanciesif,forexample,Computabilityisnottrue.

  • 16

    Roadmap

    Requirements WBErequiresthreemaincapabilities:theabilitytophysicallyscanbrainsinordertoacquirethenecessaryinformation,theabilitytointerpretthescanneddatatobuildasoftwaremodel,andtheabilitytosimulatethisverylargemodel.Theseinturnrequireanumberofsubcapabilities(Table3:CapabilitiesneededforWBE).Plausiblescanningmethodsrequirewaysofpreparingthebrains,inparticularseparationfromothertissue,fixationandpossiblydyeing.Thereisalsoaneedformethodsofphysicallyhandlingandstoringpiecesoftissue:sincemostscanningmethodscannotimagelargevolumesthebrainswillhavetobesectionedintomanageablepieces.Thismustallowcorrespondingcellsanddendritestobeidentifiedonbothsides.Whilefixationandsectioningmethodsarecommonlyusedinneuroscience,thedemandsforwholebrainemulationarestricter:muchlargervolumesmustbehandledwithfarlesstolerancefordamage.Imagingmethodsarediscussedinmoredetailinthechapteronscanning.Thethreekeyissuesareachievingthenecessaryresolutiontoimagethesmallestsystemsneededforanemulation,theabilitytoimage(notnecessarilysimultaneously)theentirebrain,andtheabilitytoacquirethefunctionallyrelevantinformation.Translatingthedatafromtheimagingprocessintosoftwarerequiressophisticatedimageprocessing,theabilitytointerprettheimageryintosimulationrelevantparameters,andhavingacomputationalneurosciencemodelofsufficientprecision.Theimageprocessingwillhavetodealwiththeunavoidableartefactsfromscanningsuchasdistortionsandnoise,aswellasoccasionallostdata.Itwilllikelyincludemethodsofconvertingdirectscandataintomorecompressedforms,suchastracedstructures,inordertoavoidexcessivestorageneeds.Thescaninterpretationprocessmakesuseofthisdatatoestimatetheconnectivity,andtoidentifysynapticconnections,celltypesandsimulationparameters.Itthenplacesthisinformationinaninventorydatabasefortheemulation.Thesestepsarediscussedintheimageprocessingchapter.Thesoftwaremodelrequiresbothamathematicalmodelofneuralactivityandwaysofefficientlyimplementingsuchmodelsoncomputers(discussedinthechapteronneuralsimulation).Computationalneuroscienceaimsatmodellingthebehaviourofneuralentitiessuchasnetworks,neurons,synapsesandlearningprocesses.ForWBE,itneedstohavesufficientlygoodmodelsofallrelevantkindsofsubsystems,alongwiththerelevantparameterssetfromscandatainordertoconstructacomputationalmodeloftheactualbrainthatwasscanned.Toemulateabrain,weneedenoughcomputingpowertorunthebasicemulationsoftware,asufficientlyrealisticbodysimulation,andpossiblyasimulatedenvironment.Thekeydemandsareformemorystoragetoholdtheinformationandprocessorpowertorunitatasuitablespeed.Themassiveparallelismoftheproblemwillputsomesignificantdemandsontheinternalbandwidthofthecomputingsystem.Inaddition,WBElikelyrequiresthedevelopmentofthreesupportingtechnologyareas,withwhichithasasymbioticrelationship.First,validationmethodstocheckthatothersteps

  • 17

    produceaccuratedataandmodels.Thisincludesvalidationofscanning,validationofscaninterpretation,validationofneurosciencemodels,validationsofimplementation,andwaysoftestingthesuccessofWBE.Whileordinaryneuroscienceresearchcertainlyaimsatvalidation,itdoesnotsystematizeit.ForacomplexmultistepresearcheffortlikeWBE,integratedvalidationislikelynecessarytoensurethatbaddataormethodsdonotconfuselaterstepsintheprocess.Second,WBErequiressignificantlowlevelunderstandingofneuroscienceinordertoconstructthenecessarycomputationalmodelsandscaninterpretationmethods.Thisisessentiallyacontinuationandstrengtheningofsystemsbiologyandcomputationalneuroscienceaimingataverycompletedescriptionofthebrainonsomesizeorfunctionalscale.Third,WBEislargescaleneuroscience,requiringmethodsofautomatingneuroscientificinformationgatheringandexperimentation.Thiswillreducecostsandincreasethroughput,andisnecessaryinordertohandlethehugevolumesofdataneeded.Largescale/industrialneuroscienceisclearlyrelevantforotherneuroscienceprojectstoo.

  • 18

    Figure4:TechnologicalcapabilitiesneededforWBE.

    Whole brain

    emulation

    Simulation

    Scanning

    Translation

    Preparation

    Physical handling

    Imaging Volume

    Resolution

    Functional information

    Storage

    Software model of neural system

    Scan interpretation

    Image processing

    Bandwidth

    Efficient implementation

    Mathematical model

    Environment simulation

    CPU

    Noise removal

    Tracing

    Synapse identification

    Cell type identification

    Geometric adjustment

    Data interpolation

    Databasing

    Connectivity identification

    Parameter estimation

    Body Simulation

  • 19

    Table3:CapabilitiesneededforWBE

    Preprocessing/fixationPreparingbrainsappropriately,retainingrelevantmicrostructureandstate

    PhysicalhandlingMethodsofmanipulatingfixedbrainsandtissuepiecesbefore,during,andafterscanning

    VolumeCapabilitytoscanentirebrainvolumesinreasonabletimeandexpense.

    ResolutionScanningatenoughresolutiontoenablereconstruction

    Scanning

    Imaging

    FunctionalinformationScanningisabletodetectthefunctionallyrelevantpropertiesoftissue

    GeometricadjustmentHandlingdistortionsduetoscanningimperfection

    Datainterpolation HandlingmissingdataNoiseremoval ImprovingscanqualityImageprocessing

    TracingDetectingstructureandprocessingitintoaconsistent3Dmodelofthetissue

    Celltypeidentification Identifyingcelltypes

    SynapseidentificationIdentifyingsynapsesandtheirconnectivity

    ParameterestimationEstimatingfunctionallyrelevantparametersofcells,synapses,andotherentities

    Scaninterpretation

    DatabasingStoringtheresultinginventoryinanefficientway

    MathematicalmodelModelofentitiesandtheirbehaviour

    Translation

    Softwaremodelofneuralsystem

    Efficientimplementation Implementationofmodel

    StorageStorageoforiginalmodelandcurrentstate

    BandwidthEfficientinterprocessorcommunication

    CPUProcessorpowertorunsimulation

    BodysimulationSimulationofbodyenablinginteractionwithvirtualenvironmentorthroughrobot

    Simulation

    EnvironmentsimulationVirtualenvironmentforvirtualbody

    Linkages MostofthecapabilitiesneededforWBEareindependentofeachother,orformsynergisticclusters.Clustersoftechnologiesdeveloptogether,supportingandmotivatingeachotherwiththeiroutput.Atypicalexampleisbettermathematicalmodelsstimulatinganeedforbetterimplementationsandcomputingcapacity,whileimprovementsinthelattertwostimulateinterestinmodelling.Anotherkeyclusteris3Dmicroscopyandimageprocessing,whereimprovementsinonemakestheothermoreuseful.Therearefewclearcaseswhereacapabilityneedsacompletedearliercapabilityinordertobegindevelopment.Currentfixationandhandlingmethodsarelikelyunabletomeetthe

  • 20

    demandsofWBElevel3Dmicroscopy,butaregoodenoughtoenableearlydevelopmentforcertainsmallsystems.Scaninterpretationneedsenoughscandatatodevelopmethods,butgivencurrentresearchthebottleneckappearstobemoreontheimageprocessingandinterpretationsidethandataavailability.Achievinglargevolumescanningrequiresparallelizationandscalinguppreviousscanningmethods,forexamplebyusingroboticworkandparallelmicroscopy.Thisrequiresresearchersthinkingintermsof,andhavingexperiencewith,anindustrialapproachtodatacollection.Thisinterlinkednatureofthefieldavoidsanyobvioustechnologythresholdsandbottlenecks.Thereisnoonetechnologythatmustbedevelopedbeforeothertechnologiescanadvance.Developmentcanoccuronabroadfrontsimultaneously,andrapidprogressinafieldcanpromotefeedbackprogressinrelatedfields.Unfortunately,italsomeansthatslowprogressinoneareamayholdbackotherareas,notjustduetolackofresultsbutbyreduceddemandfortheirfindings,reducedfunding,andfocusonresearchthatdoesnotleadinthedirectionofWBE.

    Roadmap Basedontheseconsiderationswecansketchoutaroadmapwithmilestones,requiredtechnologies,keyuncertaintiesandexternaltechnologyinteractions.ApproachtoWBEhastwophases.Thefirstphaseconsistsofdevelopingthebasiccapabilitiesandsettlingkeyresearchquestionsthatdeterminethefeasibility,requiredlevelofdetailandoptimaltechniques.Thisphasemainlyinvolvespartialscans,simulationsandintegrationoftheresearchmodalities.Thesecondphasebeginsoncethecoremethodshavebeendevelopedandanautomatedscaninterpretationsimulatepipelinehasbeenachieved.Atthispointthefirstemulationsbecomepossible.Ifthedevelopedmethodsprovetobescalabletheycanthenbeappliedtoincreasinglycomplexbrains.Herethemainissueisscalinguptechniquesthathavealreadybeenprovenonthesmallscale.

  • 21

    Figure5:WBEroadmap.Thekeymilestonesare:Groundtruthmodels:asetofcaseswherethebiologicalgroundtruthisknownandcanbecomparedtoscans,interpretationsandsimulationsinordertodeterminetheiraccuracy.Determiningappropriatelevelofsimulation:thisincludesdeterminingwhetherthereexistsanysuitablescaleseparationinbrains(ifnot,theWBEeffortmaybeseverelylimited),andifso,onwhatlevel.Thiswouldthenbetherelevantscaleforscanningandsimulation.Fullcellsimulation:acompletesimulationofacellorsimilarlycomplexbiologicalsystem.WhilestrictlynotnecessaryforWBEitwouldbeatestcaseforlargescalesimulations.

    Organism simulation

    Complete inventory

    Partial emulations

    Eutelic organism emulation

    Small mammal emulation

    Large mammal emulation

    Human emulation

    Automated pipeline

    Scanning development

    Interpretation development

    Low-level neuroscience

    Validation methods

    Full cell simulation

    Ground truth models

    Body simulation

    Simulation hardware

    Appropriate level

    Deducing function

    Invertebrate emulation

  • 22

    Bodysimulation:anadequatesimulationforthemodelanimalsbody(andenvironment).Ideallydemonstratedbyfoolingarealanimalconnectedtoit.Simulationhardware:specialpurposesimulation/emulationcomputerhardwaremaybefoundtobenecessaryoreffective.Organismsimulation:asimulationofanentireorganismintermsofneuralcontrol,bodystateandenvironmentalinteraction.Thiswouldnotbeatrueemulationsinceitisnotbasedonanyindividualbutratherknownphysiologicaldataforthespecies.Thiswouldenablemorerealisticandindividualmodelsasscans,modelsandcomputerpowerimproves.Demonstrationoffunctiondeduction:demonstratingthatallrelevantfunctionalpropertiesonalevelcanbededucedfromscandata.Completeinventory:acompletedatabaseofentitiesatsomelevelofresolutionforaneuralsystem,e.g.notjusttheconnectivityoftheC.elegansnervoussystembutalsotheelectrophysiologyofthecellsandsynapses.Thiswouldenablefullemulationifalltheupdaterulesareknown.Itdemonstratesthatthescanningandtranslationmethodshavematured.Automatedpipeline:asystemabletoproduceasimulationbasedonaninputtissuesample,goingthroughthescan,interpretationandsimulationstepswithoutmajorhumanintervention.Theresultingsimulationwouldbebasedontheparticulartissueratherthanbeingagenericmodel.Partialemulation:Acompleteemulationofneuralsystemsuchastheretina,invertebrategangliaoraV1circuitbasedonscannedandinterpreteddatafromabrainratherthanspeciesdata.Thiswoulddemonstratethefeasibilityofdatadrivenbrainemulation.Eutelicorganismemulation:acompleteemulationofasimpleorganism,suchasC.elegansoranothereutelic(fixednervoussystem)organismusingdatafrompipelinescanning.Itmayturnoutthatitisunnecessarytostartwithaeutelicorganismandthefirstorganismemulationwouldbeamorecomplexinvertebrate.InvertebrateWBE:Emulationofaninvertebratesuchasasnailoraninsect,withlearning.ThiswouldtestwhethertheWBEapproachcanproduceappropriatebehaviours.Ifthescannedindividualwastrainedbeforescanning,retentionoftrainedresponsescanbechecked.SmallmammalWBE:DemonstrationofWBEinmiceorrats,provingthattheapproachcanhandlemammalianneuroanatomy.LargemammalWBE:Demonstrationinhighermammals,givingfurtherinformationabouthowwellindividuality,memoryandskillsarepreservedaswellasinvestigationofsafetyconcerns.HumanWBE:Demonstrationofaninteractivehumanemulation.

  • 23

    Technology drivers

    Figure6:TechnologydriversforWBEnecessarytechnologies.Differentrequiredtechnologieshavedifferentsupportanddriversfordevelopment.

    Whole brain

    emulation

    Simulation

    Scanning

    Translation

    Preparation

    Physical handling

    Imaging Volume

    Resolution

    Functional information

    Storage

    Software model of neural system

    Scan interpretation

    Image processing

    Bandwidth

    Efficient implementation

    Mathematical model

    Environment simulation

    CPU

    Noise removal

    Tracing

    Synapse identification

    Cell type identification

    Geometric adjustment

    Data interpolation

    Databasing

    Connectivity identification

    Parameter estimation

    Body Simulation

    Moores law driven

    WBE specific? Research drivers

    Commercial drivers

  • 24

    Computersaredevelopedindependentlyofanyemulationgoal,drivenbymassmarketforcesandtheneedforspecialhighperformancehardware.Mooreslawandrelatedexponentialtrendsappearlikelytocontinuesomedistanceintothefuture,andthefeedbackloopspoweringthemareunlikelytorapidlydisappear(seefurtherdiscussioninAppendixB:ComputerPerformanceDevelopment).Thereisindependent(andoftensizeable)investmentintocomputergames,virtualreality,physicssimulationandmedicalsimulations.Likecomputers,thesefieldsproducetheirownrevenuestreamsanddonotrequireWBEspecificorscientificencouragement.Alargenumberoftheothertechnologies,suchasmicroscopy,imageprocessing,andcomputationalneurosciencearedrivenbyresearchandnicheapplications.Thismeanslessfunding,morevariabilityofthefunding,anddependenceonsmallergroupsdevelopingthem.Scanningtechnologiesaretiedtohowmuchmoneythereisinresearch(includingbrainemulationresearch)unlessmedicalorotherapplicationscanbefound.Validationtechniquesarenotwidelyusedinneuroscienceyet,butcould(andshould)becomestandardassystemsbiologybecomesmorecommonandwidelyapplied.FinallythereareafewareasrelativelyspecifictoWBE:largescaleneuroscience,physicalhandlingoflargeamountsoftissueblocks,achievinghighscanningvolumes,measuringfunctionalinformationfromtheimages,automatedidentificationofcelltypes,synapses,connectivityandparameters.TheseareasaretheonesthatneedmostsupportinordertoenableWBE.Thelattergroupisalsothehardesttoforecast,sinceithasweakdriversandasmallnumberofresearchers.Thefirstgroupiseasiertoextrapolatebyusingcurrenttrends,withtheassumptionthattheyremainunbrokensufficientlyfarintothefuture.

    Uncertainties and alternatives Themainuncertaintiesintheroadmapare:DoesscaleseparationenablingWBEoccur?Thisisabasicsciencequestion,andifscaleseparationdoesnotoccuratasufficientlyhighscalelevelthenWBEwouldbeseverelylimitedorinfeasible.Itispossiblethatprogressinunderstandingcomplexsystemsingeneralwillhelpclarifythesituation.Thequestionofwhichcomplexsystemsaresimulableunderwhichconditionsisofgeneralinterestformanyfields.However,inrelationtoWBEtheanswerseemsmostlikelytocomefromadvancesincomputationalneuroscienceandfromtrialanderror.Byexperimentingwithvariousspecializedneuralemulations,atdifferentlevelsofresolution,andcomparingthefunctionallyrelevantcomputationalpropertiesoftheemulationwiththoseoftheemulatedsubsystem,wecantestwhetheragivenemulationissuccessful.Ifso,wecaninferthatasufficientscaleseparationforthatsubsystemexistsat(orabove)thescalelevel(granularity)usedintheemulation.Weemphasizethatasuccessfulemulationneednotpredictalldetailsoftheoriginalbehavioroftheemulatedsystem;itneedonlyreplicatecomputationallyrelevantfunctionalityatthedesiredlevelofemulationWhatlevelsarepossible/mostappropriateforemulation?Thiswilldetermineboththerequirementsforscanningandemulationsoftware.Inordertodiscoverit,smallscalescanningandsimulationprojectsneedtobeundertaken,developingskillsandmethods.Later,fullscaleemulationsofsmallsystemswilltestwhethertheestimateshold.Ifanearlyanswercanbefound,effortscanfocusonthislevel;otherwisetheWBEresearchfrontwouldhavetoworkonmultiplelevels.

  • 25

    Howmuchofthefunctionallyrelevantinformationcanbededucedfromscanninginaparticularmodality(e.g.electronmicroscopy)?Atpresent,electronmicroscopyappearstobetheonlyscanningmethodthathastherightresolutiontoreachsynapticconnectivity,butitislimitedinwhatchemicalstateinformationitcanreveal.Ifitispossibletodeducethefunctionofaneuron,synapseorotherstructurethroughimageinterpretationmethods,thenscanningwouldbefarsimplerthanifthisisnot(inwhichcasesomeformofhybridmethodorentirelynewscanningmodalitywouldhavetobedeveloped).Thisissueappearstoformapotentiallywelldefinedresearchquestionthatcouldbepursued.Answeringitwouldrequirefindingasuitablemodelsystemforwhichgroundtruth(thecomputationalfunctionalityoftargetsystem)wasknown,usingthescanningmodalitytoproduceimageryandthentestingoutvariousformsofinterpretationonthedata.

    Figure7:WBEcomputationalbiologyresearchcycle(basedon(Takahashi,Yugietal.2002)).WBEintroduceslargescalescanningandsimulationintothecycle.Developingtherightresearchcycle.Thecomputationalbiologyresearchcycletodayinvolveswetexperimentsprovidingcellulardataandhypotheses,whichdrivequalitativemodelling.Thismodellinginturnisusedinquantitativemodelling,whichusingsimulationsgeneratedatathatcanbeanalysedandemployedtorefinethemodels,comparewiththeexperiments,andsuggestnewexperiments(Takahashi,Yugietal.,2002).TheWBEparadigmincorporatesthisresearchcycle(especiallyinthesoftwaremodellingpart),butincludestwonewfactors.Oneislargescalescanningandprocessingofbraintissue,providingmassiveamountsofdataasinputtothecycle,butalsorequiringthemodelstoguidethedevelopmentofscanningmethods.Thesecondisthelargescaledatadrivensimulationsthatdonotaimatproducingjusthypothesistestingandmodelrefinement,butalsoataccuratelymimickingthewetsystem.Bothfactorswillincreaseandchangethedemandsfordatamanagement,hypothesissearchingandsimulationanalysis/interpretation.Theywillalsointroducesociologicalandinterdisciplinaryfactors,asdifferentacademicdisciplineswithverydifferentmethodologieswillhavetolearnhowtocommunicateandcooperate.Inordertobeviable,fieldresearchmethodsoftesting,datasharing,validation,andstandardsforwhatconstitutesaresultmustbedevelopedsothattheextendedcycleprovidesincentivesforallparticipantstocooperateandpushthetechnologyforward.Thisiscloselylinkedtothelikelymove

    Interpretation

    Wet experiments

    Cellular data and

    hypotheses

    Qualitative modelling

    Quantitative modelling

    Cell programming

    Simulation

    Analysis

    Large-scale scanning

    Large-scale simulation

  • 26

    towardslargescaleneuroscience,whereautomatedmethodswillplayanincreasinglyprominentroleastheyhavedoneingenomics.

    Figure8:Caenorhabditiselegans,apopularmodelorganismwithafullymapped302neuronnervoussystem.Selectionofsuitablemodelsystems.Selectingtherighttargetsforscanningandmodellingwillhavetotakeintoaccountexistingknowledge,existingresearchcommunities,likelihoodoffundingandacademicimpactaswellaspracticalfactors.WhiletheC.elegansnervoussystemhasbeencompletelymapped(White,Southgateetal.,1986),westilllackdetailedelectrophysiology,likelybecauseofthedifficultyofinvestigatingthesmallneurons.Animalswithlargerneuronsmayprovelessrestrictiveforfunctionalandscanninginvestigationbutmaylacksizeableresearchcommunities.Molluscssuchasfreshwatersnailsandinsectssuchasfruitfliesmaybepromisingtargets.Theyhavewellcharacterisedbrains,existingresearchcommunitiesandneuralnetworkswellwithincurrentcomputationalcapabilities.Similarly,theselectionofsubsystemsofthebraintostudyrequirescarefulconsideration.Someneuralsystemsareheavilystudied(corticalcolumns,thevisualsystem,thehippocampus)andbetterdataaboutthemwouldbewarmlyreceivedbytheresearchcommunity,yetthelackofcharacterizationoftheirinputs,outputsandactualfunctionmaymakedevelopmentofemulationmethodsveryhard.Onesystemthatmaybeverypromisingistheretina,whichhasanaccessiblegeometry,iswellstudiedandsomewhatwellunderstood,isnotexcessivelycomplex,andbetterinsightsintowhichwouldbeusefultoawideresearchcommunity.Buildingonretinalmodels,modelsofthelateralgeniculatenucleusandvisualcortexmaybeparticularlyuseful,sincetheywouldbothhaverelativelywelldefinedinputsfromthepreviousstages.AtwhatpointwillthepotentialbeclearenoughtobringmajoreconomicactorsintoWBE?GiventhepotentiallyhugeeconomicimpactofhumanWBE(Hanson,1994,2008a,2004,2008b),ifthefieldshowssufficientpromise,majoreconomicactorswillbecomeinterestedinfundinganddrivingtheresearchasaninvestment.Itisunclearhowfaradvancedthefieldwouldneedtobeinordertogarnerthisattention.Soliddemonstrationsofkeytechnologiesarelikelyrequired,aswellasaplausiblepathtowardsprofitableWBE.Theimpactoffundingonprogresswilldependonthemainremainingbottlenecksandtheirfundingelasticity.Ifscanningthroughputorcomputerpoweristhelimitingfactor,extrafundingcanrelativelyeasilyscaleupfacilities.Bycontrast,limitationsinneuroscienceunderstandingarelessresponsivetoinvestment.Iffundingarriveslate,whenthefundamentalproblemshavealreadybeensolved,theamplifiedresourceswouldbeusedtoscaleuppreexistingsmallscaledemonstrationprojects.

  • 27

    Intellectualpropertyconstitutesanotherimportantconsiderationforcommercialfunding:whatcouldearlydevelopersown,howsecureandlongtermwouldtheirinvestmentbe?Withoutsolidprospectsofhavingpreferentialownershipofwhatitdevelops,afirmisunlikelytopursuetheproject.

    Alternative pathways SpecialhardwareforWBE.ItispossiblethatWBEcanbeachievedmoreefficientlyusingdedicatedhardwareratherthangenerichardware.Suchperformancegainsarepossibleif,forexample,thereisaclosemappingbetweenthehardwareandthebrain,orifthefunctionsoftheemulationsoftwarecouldbeimplementedefficientlyphysically.Developingsuchdedicatedhardwarewouldbecostlyunlessotherapplicationsexisted(whichiswhyinterestindedicatedneuralnetworkchipspeakedintheearly1990s).Dedicatedneuralnetworkchipshavereachedupto1.7billionsynapticupdates(and337millionsynapticadjustments)persecondforANNmodels(Kondo,Koshibaetal.,1996),whichisapproachingcurrentsupercomputingspeedsformorecomplexmodels.Recently,therehasbeensomedevelopmentofFPGAs(FieldProgrammableGateArrays)forrunningcomplexneuronsimulations,producinganorderofmagnitudefastersimulationforamotorneuronthanasoftwareimplementation(fourtimesrealtime,8Mcompartments/s)(WeinsteinandLee,2005).AFPGAimplementationhastheadvantageofbeingprogrammable,notrequiringWBEspecialpurposehardware.Anotheradvantageincludethataslongasthereischipspace,morecomplexmodelsdonotrequiremoreprocessingtimeandthatprecisioncanbeadjustedtosuitthemodelandreducespacerequirements.However,scalinguptolargeanddenselyinterconnectednetworkswillrequiredevelopingnewtechniques(WeinsteinandLee,2006).Abetterunderstandingoftheneocorticalarchitecturemayservetoproducehardwarearchitecturesthatfititwell(Daisyproject,2008).IthasbeensuggestedthatusingFPGAscouldincreasecomputationalspeedsinnetworksimulationsbyuptotwoordersofmagnitude,andinturnenabletestinggroundsfordevelopingspecialpurposeWBEchips(Markram,2006).Itmayalsobepossibletouseembeddedprocessortechnologytomanufacturelargeamountsofdedicatedhardwarerelativelycheaply.Astudyofhighresolutionclimatemodellinginthepetafloprangefounda24to34foldreductionofcostandabouttwoordersofmagnitudesmallerpowerrequirementsusingacustomvariantofembeddedprocessorchips(Wehner,Olikeretal.,2008).Thisroadmapisroughlycentredontheassumptionthatscanningtechnologywillbesimilartocurrentmicroscopy,developedforlargescaleneuroscience,automatedsectioningoffixatedtissueandlocalimagetomodelconversion.Forreasonsdiscussedinthescanningsection,nondestructivescanningoflivingbrainsappearstobehardcomparedtothesliceanddiceapproachwherewehavevarioussmallscaleexistenceproofs.However,aspointedoutbyRobertFreitasJr.,nanomedicaltechniquescouldpossiblyenablenondestructivescanningbyuseofinvasivemeasurementdevices.Evenifsuchdevicesproveinfeasible,molecularnanotechnologycouldlikelyprovidemanynewscanningmethodologiesaswellasradicalimprovementofneuroscientificresearchmethodsandtheefficiencyofmanyroadmaptechnologies.Evenfarmoremodestdevelopmentssuchassinglemoleculeanalysis,nanosensors,artificialantibodiesandnanoparticlesforimaging(whichareexpectedtobeinuseby2015(NanoroadmapProject,2006))wouldhaveanimportantimpact.Henceearlyorverysuccessfulnanotechnologywouldofferfasterandalternativeroutestoachievethe

  • 28

    roadmap.Analysingthelikelihood,timeframe,andabilitiesofsuchnanomedicineisoutsidethescopeofthisdocument.OnepossiblescanningalternativenotexaminedmuchhereishighresolutionscanningoffrozenbrainsusingMRI.Thismightbeacomplementtodestructivescanning,butcouldpossiblygainenoughinformationtoenableWBE.However,wecurrentlyhavelittleinformationonthelimitsandpossibilitiesofthetechnique(seediscussioninAppendixE:Nondestructiveandgradualreplacement).Asdiscussedintheoverview,WBEdoesnotassumeanyneedforhighlevelunderstandingofthebrainormind.Infact,shouldsuchunderstandingbereacheditislikelythatitcouldbeusedtoproduceartificialintelligence(AI).HumanlevelAI(orsuperintelligentAI)wouldnotnecessarilyprecludeWBE,butsomeofthescientificandeconomicreasonswouldvanish,possiblymakingthefieldlessrelevant.Ontheotherhand,powerfulAIcouldgreatlyaccelerateneuroscienceadvancesandperhapshelpdevelopWBEforotherpurposes.Conversely,successinsomepartsoftheWBEendeavourcouldhelpAI,forexampleifcorticalmicrocircuitryandlearningrulescouldbesimulatedefficientlyasagenerallearning/behavioursystem.TheimpactanddevelopmentofWBEwilldependonwhichofthemaincapabilities(scanning,interpretation,simulation)developlast.Iftheydeveloprelativelyindependentlyitwouldbeunlikelyforallthreetomatureenoughtoenablehumanlevelemulationsatthesametime.Ifcomputingpoweristhelimitingfactor,increasinglycomplexanimalemulationsarelikelytoappear.SocietyhastimetoadapttotheprospectofhumanlevelWBEinthenearfuture.Ifscanningresolution,imageinterpretation,orneuralsimulationisthelimitingfactor,arelativelysuddenbreakthroughispossible:thereisenoughcomputingpower,scanningtechnology,andsoftwaretogorapidlyfromsimpletocomplexorganisms,usingrelativelysmallcomputersandprojects.ThiscouldleadtoasurprisescenariowhereinsocietyhaslittletimetoconsiderthepossibilityofhumanlevelWBE.Ifcomputingpoweristhelimitingfactor,orifscanningisthebottleneckduetolackofthroughput,thenthepaceofdevelopmentwouldlikelybeeconomicallydetermined:ifenoughinvestmentweremade,WBEcouldbeachievedrapidly.ThiswouldplaceWBEenablementunderpoliticaloreconomicalcontroltoagreaterdegreethaninthealternativescenarios.

    Related technologies and spin-offs ThetechnologiesneededtoachieveWBEincludetheabilitytoscanorganictissueonalowlevel,interpretthefindingsintofunctionalmodels,andrunextremelylargescalesimulations.WBEalsorequiressufficientknowledgeoflowlevelneuralfunction.Thedesireforrunningextremelylargesimulationshasbeenastrongmotivatorforsupercomputing.Applicationsinnanotechnology,virtualreality,cryptography,signalprocessing,mathematics,genomics,andsimulationoftransportation,societies,business,physics,biology,andclimatesciencewillrequirepetaflopsperformanceinthenextdecade.Itisunlikelythatthiswillbetheend,andexaflopperformanceisalreadybeingdiscussedinthesupercomputingcommunity.However,scalingupcurrentarchitecturestotheexascalemaybeproblematic,andmayrequirenewwaysofthinkingabouthowtomanagecomplexconcurrentsystems.Thesewilltoalargedegreebeshapedbytheproblemsthecomputersareintendedtosolve(andtherelativerankingoftheseproblemsbysociety,affecting

  • 29

    funding)aswellasbytradeoffsbetweenperformancewithprice,energyrequirements3,andotherconstraints.TherelatedareaofverylargescaleinformationstorageisalsoakeyissueforWBE.Anobviousrelatedtechnologyisthecreationofvirtualbodymodelsforuseinmedicine.Theycanbeusedastrainingandstudyobjects,or,atamoreadvancedstage,aspersonalmedicalmodels.Suchmodelsmightenabledoctorstoinvestigatethecurrentstateofthepatient,compareittopreviousdata,testoroptimisevariousformsofsimulatedtreatment,trainparticularsurgicalapproaches,etc.Thisisatechnologyofsignificantpracticalimportancethatisdrivenbycurrentadvancesinmedicalimaging,thepersonalisationofhealthcare,andimprovingphysiologicalmodels.Whilemostcurrentmodelsareeitherstaticdatasetsorhighlevelphysiologicalmodels,personalisationrequiresdevelopingmethodsofphysiologicalparameterestimationfromthedata.SimulationwillpromotethedevelopmentofmorerealisticbodymodelsusableforWBE.Conversely,theWBEfocusondatadrivenbottomupsimulationappearstofitinwithdevelopingpersonalisedbiochemicalandtissuemodels.Virtualtestanimals,iftheycouldbedevelopedtoasufficientdegreeofrealism,wouldbehighindemandasfasterandlessexpensivetestinggroundsforbiomedicalhypotheses(MichelsonandCole,2007;Zheng,Kreuweletal.,2007).Theymayperhapsalsobeawayofavoidingtheethicalcontroversiessurroundinganimaltesting(althoughitisnotinconceivablethatconcernsaboutanimalwelfarewouldintimebeextendedtoemulatedanimals).Thiscouldprovideanimpetusforthedevelopmentoforganismemulationtechniquesandespeciallyvalidationmethodsthathelpguaranteethatthevirtualanimalshavethesamepropertiesasrealanimalswould.Overall,thereisincreasinginterestandcapabilityinquantitativemodellingofbiology(DiVentura,Lemerleetal.,2006).Whilemuchefforthasgoneintometabolicorcontrolnetworksofparticularinterest,thereisapushtowardscellmodelsencompassingmetabolism,thegenomeandproteomics(Ortoleva,Berryetal.,2003;Tomita,2001;Schaff,Slepchenkoetal.,2001;Tyson,2001).Thereisalsointerestinbuildingsimulatorsformodellingselfassemblyinsubcellularsystems(Ortoleva,Berryetal.,2003).Besidespredictingbiologicalresponsesandhelpingtounderstandbiology,modellingwilllikelybecomeimportantforsyntheticbiology(Serrano,2007).Inordertoproducerealisticvirtualstimuliforabrainemulation,accuratesimulationsofthesensoryinputtothebrainareneeded.Thesametechniquesusedtoinvestigateneuralfunctioncanbeappliedtothesenses.Ithasalsobeenproposedthattechnologytorecordnormalneuralactivityalongthebrainnervesandspinalcordcouldbehelpful.Suchdata,forexamplerecordedusingmassivelyparallelelectrodearraysornanoimplants,wouldprovidegoodtestdatatovalidateabodymodel.Recordedsensorydatawouldalsoberepeatable,enablingcomparisonsbetweendifferentvariantsofthesameemulation.Neuralinterfacingisalsousefulfordevelopingbetterroboticbodymodels.Currently,mostinterestinneuralinterfacesisfocusedonhelpingpeoplewithdisabilitiesusesensoryormotoricprosthetics.Whileneuralinterfacingforenhancementpurposesispossible,itisunlikelytobecomeasignificantdriveruntilprostheticsystemshavebecomecheap,safe,andveryeffective.

    3Anexascalecomputerusing2008technologywouldrequiretensofmegawatts(SandiaNationalLaboratories,2008).

  • 30

    Issues

    Emulation systems Afunctioningbrainemulationwillinclude,inadditiontothebrainmodel(themainpart),somewayforthebrainmodeltoexperiencebodilyinteractionswithanenvironment.Therearetwodifferentwaysinwhichthiscouldbeaccomplished:viaasimulatedvirtualbodyinhabitingavirtualreality(whichcanbelinkedtotheoutsideworld);orviaahardwarebodyconnectedtothebrainmodelviaabodyinterfacemodule.EntrylevelWBEdoesnotrequirethecapacitytoaccommodatealloftheoriginalsensorymodalitiesortoprovideafullynaturalisticbodyexperience.Simulatedbodiesandworlds,or

    hardwarebodyinterfaces,aswellascommunicationswiththeoutsideworld,arenotnecessaryperseforbrainemulationexceptinsofartheyareneededtomaintainshorttermfunctionofthebrain.Forlongtermfunction,especiallyofhumanmindemulations,embodimentandcommunicationareimportant.Sensoryormotordeprivationappearstoproduceintellectualandperceptualdeficitswithinafewdaystime(ZubekandMacneill,1966).Thebrainemulatorperformstheactualemulationofthebrainandcloselylinkedsubsystemssuchasbrainchemistry.Theresultofitsfunctionisaseriesofstatesofemulatedbrainactivity.Theemulationproducesandreceivesneuralsignalscorrespondingtomotoractionsandsensoryinformation(inaddition,somebodystateinformationsuchasglucoselevelsmaybeincluded).Thebodysimulatorcontainsamodelofthebodyanditsinternalstate.Itproducessensorysignalsbasedonthestateofthebodymodelandtheenvironment,sendingthemtothebrainemulation.Itconvertsmotorsignalstomusclecontractionsordirectmovementsinthe

    bodymodel.Thedegreetowhichdifferentpartsofthebodyrequireaccuratesimulationislikelyvariable.Theenvironmentsimulatormaintainsamodelofthesurroundingenvironment,respondingtoactionsfromthebodymodelandsendingbacksimulatedsensoryinformation.Thisisalso

    Figure9:Emulationsystemforcompletelyvirtualemulation.

    Figure10:Emulationsystemforembodiedemulations.

  • 31

    themostconvenientpointofinteractionwiththeoutsideworld.Externalinformationcanbeprojectedintotheenvironmentmodel,virtualobjectswithrealworldaffordancescanbeusedtotriggersuitableinteractionetc.Theoverallemulationsoftwaresystem(theexoselftoborrowGregEgansterm)wouldregulatethefunctionofthesimulatorsandemulator,allocatecomputationalresources,collectdiagnosticinformation,providesecurity(e.g.backups,firewalls,errordetection,encryption)andsoon.Itcouldprovidesoftwareservicestoemulatedminds(accessedthroughthevirtualenvironment)and/oroutsideexperimenters.Avariantoftheabovesystemwouldbeanembodiedbrainemulation,inwhichcasethebodysimulatorwouldmerelycontainthetranslationfunctionsbetweenneuralactivityandphysicalsignals,andthesewouldthenbeactuatedusingahardwarebody.Thebodymightbecompletelyartificial(inwhichcasemotorsignalshavetobemappedontoappropriatebodybehaviours)orbiologicalbutequippedwithnervecomputerinterfacesenablingsensingandcontrol.Thecomputersystemrunningtheemulationdoesnothavetobephysicallypresentinthebody.Itiscertainlypossibletointroducesignalsfromtheoutsideonhigherlevelsthaninasimulatedorrealbody.Itwouldberelativelytrivialtoaddvisualorauditoryinformationdirectlytothebodymodelandhavethemappearasvirtualoraugmentedreality.Introducingsignalsdirectlyintothebrainemulationwouldrequirethemtomakesenseasneuralsignals(e.g.brainstimulationorsimulateddrugs).Virtualbraincomputerinterfaceswithperfectclarityandnoriskofsideeffectscouldbeimplementedasextensionsofthebodysimulation/interface.Ifcomputingpowerturnsouttobeabottleneckresource,thenearlyemulationsarelikelytorunmoreslowlythanthebiologicalsystemtheyaimatemulating.Thiswouldlimittheabilityoftheemulationstointeractinrealtimewiththephysicalworld.Indistributedemulationsdelaysbetweencomputingnodesputastronglimitonhowfasttheycanbecome.Theshortestdelayusing100m/saxonsacrossthebrainisabout1ms.Assuminglightspeedcommunication,processingnodescannotbefurtherawaythan300kmiflongerdelaysaretobeavoided.

    Complications and exotica Besidestraightneuraltransmissionthroughsynapsestheremaybenumerousotherformsofinformationprocessinginthebrainthatmayhavetobeemulated.Howimportanttheyareforsuccessinemulationremainsuncertain.Animportantapplicationofearlybrainemulationsandtheirprecursorswillbetoenabletestingoftheirinfluence.

    SpinalcordWhiletraditionallythevertebratespinalcordisoftenregardedaslittlemorethanabundleofmotorandsensoraxonstogetherwithacentralcolumnofstereotypicalreflexcircuitsandpatterngenerators,thereisevidencethattheprocessingmaybemorecomplex(Berg,Alaburdaetal.,2007)andthatlearningprocessesoccuramongspinalneurons(Crown,Fergusonetal.,2002).Thenetworksresponsibleforstandingandsteppingareextremelyflexibleandunlikelytobehardwired(Cai,Courtineetal.,2006).

  • 32

    Thismeansthatemulatingjustthebrainpartofthecentralnervoussystemwilllosemuchbodycontrolthathasbeenlearnedandresidesinthenonscannedcord.Ontheotherhand,itispossiblethatagenericspinalcordnetworkwould,whenattachedtotheemulatedbrain,adapt(requiringonlyscanningandemulatingonespinalcord,aswellasfindingawayofattachingthespinalemulationtothebrainemulation).Butevenifthisistrue,thetimetakenmaycorrespondtorehabilitationtimescalesof(subjective)months,duringwhichtimethesimulatedbodywouldbeessentiallyparalysed.Thismightnotbeamajorproblemforpersonalidentityinmindemulations(sincepeoplesufferingspinalinjuriesdonotlosepersonalidentity),butitwouldbeamajorlimitationtotheirusefulnessandmightlimitdevelopmentofanimalmodelsforbrainemulation.Asimilarconcerncouldexistforotherperipheralsystemssuchastheretinaandautonomicnervoussystemganglia.Thehumanspinalcordweighs2.5%ofthebrainandcontainsaround104ofthenumberofneuronsinthebrain(13.5millionneurons).Henceaddingthespinalcordtoanemulationwouldaddanegligibleextrascanandsimulationload.

    SynapticadaptationSynapsesareusuallycharacterizedbytheirstrength,thesizeofthepostsynapticpotentialtheyproduceinresponsetoagivenmagnitudeofincomingexcitation.Many(most?)synapsesintheCNSalsoexhibitdepressionand/orfacilitation:atemporarychangeinreleaseprobabilitycausedbyrepeatedactivity(Thomson,2000).Thisrapiddynamicslikelyplaysaroleinavarietyofbrainfunctions,suchastemporalfiltering(FortuneandRose,2001),auditoryprocessing(Macleod,Horiuchietal.,2007)andmotorcontrol(NadimandManor,2000).Thesechangesoccurontimescaleslongerthanneuralactivity(tensofmilliseconds)butshorterthanlongtermsynapticplasticity(minutestohours).Adaptationhasalreadybeenincludedinnumerouscomputationalmodels.Thecomputationalloadisusually13extrastatevariablesineachsynapse.

    UnknownneurotransmittersandneuromodulatorsNotallneuromodulatorsareknown.Atpresentabout10majorneurotransmittersand200+neuromodulatorsareknown,andthenumberisincreasing.(Thomas,2006)lists272endogenousextracellularneuroactivesignaltransducerswithknownreceptors,2gases,19substanceswithputativeorunknownbindingsitesand48endogenoussubstancesthatmayormaynotbeneuroactivetransducers(manyofthesemaybemoreinvolvedingeneralbiochemicalsignallingthanbrainspecificsignals).Plottingtheyearofdiscoveryfordifferentsubstances(orfamiliesofsubstances)suggestsalinearorpossiblysigmoidalgrowthovertime(Figure11).

  • 33

    GhrelinKisspeptin

    OrexinNPFF

    NociceptinAnandamide

    ApelinPACAPNO

    FGFANF

    MCHGalaninBombesin

    GHRHNPY

    DynorphinPDML

    VIPEndorphinCCKHistamine

    NeurotensinEnkephalin

    SomatostatinSubstance PGnRH

    TRHGlycine

    DopamineAngiotensin

    CalcitoninSerotonin

    GlutamateGABA

    CRFACTH

    Noradrenaline Acetylcholine

    1920 1930 1940 1950 1960 1970 1980 1990 2000 2010

    Figure11:Timeofdiscoveryoftheneurotransmitterorneuromodulatoractivityofanumberofsubstancesorfamiliesofsubstances.Datatakenfrom(vonBohlenundHalbachandDermietzel2006),likelyunderrepresentingthedevelopmentafter2000.Anupperboundonthenumberofneuromodulatorscanbefoundusinggenomics.About800Gproteincoupledreceptorscanbefoundinthehumangenome,ofwhichabouthalfweresensoryreceptors.Manyareorphansthatlackknownligands,andmethodsofdeorphanizingreceptorsbyexpressingthemanddeterminingwhattheybindtohavebeendeveloped.Inthemiddle1990sabout150receptorshadbeenpairedto75transmitters,leavingaround150200orphansin2003(Wise,Jupeetal.,2004).Atpresent,78receptorsaredeorphanizedeachyear(vonBohlenundHalbachandDermietzel,2006);atthisrateallorphansshouldbeadoptedwithin20years,leadingtothediscoveryofaround50moretransmitters(Civelli,2005).Similarlyguanylylcyclasecoupledreceptors(fourorphans,(WedelandGarbers,1998)),tyrosinekinasecoupledreceptors(

  • 34

    ForWBEmodellingallmodulatoryinteractionsisprobablycrucial,sinceweknowthatneuromodulationdoeshaveimportanteffectsonmood,consciousness,learningandperception.Thismeansnotjustdetectingtheirexistencebuttocreatequantitativemodelsoftheseinteractions,asizeablechallengeforexperimentalandcomputationalneuroscience.

    UnknownionchannelsSimilartoreceptors,therearelikelyunknownionchannelsthataffectneurondynamics.TheLigandGatedIonChannelDatabasecurrentlycontains554entrieswith71designatedaschannelsubunitsfromHomosapiens(EMBLEBI,2008;Donizelli,Djiteetal.,2006).Voltagegatedionchannelsformasuperfamilywithatleast143genes(Yu,YarovYarovoyetal.,2005).Thisdiversityisincreasedbymultimerization(combinationsofdifferentsubunits),modifiersubunitsthatdonotformchannelsontheirownbutaffectthefunctionofchannelstheyareincorporatedinto,accessoryproteinsaswellasalternatemRNAsplicingandposttranslationalmodification(Gutman,Chandyetal.,2005).Thiswouldenableatleastanorderofmagnitudemorevariants.Ionchanneldiversityincreasesthediversityofpossibleneuronelectrophysiology,butnotnecessarilyinalinearmanner.Seethediscussionofinferringelectrophysiologyfromgenetranscriptsintheinterpretationchapter.

    VolumetransmissionSurroundingthecellsofthebrainistheextracellularspace,onaverage200acrossandcorrespondingto20%ofbrainvolume(Nicholson,2001).Ittransportsnutrientsandbuffersions,butmayalsoenablevolumetransmissionofsignallingmolecules.Volumetransmissionofsmallmoleculesappearsfairlywellestablished.Nitrousoxideishydrophobicandhaslowmolecularweightandcanhencediffuserelativelyfreelythroughmembranes:itcanreachupto0.10.2mmawayfromareleasepointunderphysiologicalconditions(Malinski,Tahaetal.,1993;SchumanandMadison,1994;WoodandGarthwaite,1994).Whilemainlybelievedtobeimportantforautoregulationofbloodsupply,itmayalsohavearoleinmemory(Ledo,Fradeetal.,2004).ThismightexplainhowLTP(LongTermPotentiation)caninducecrosstalkthatreducesLTPinductionthresholdsoveraspanof10mandtenminutes(HarveyandSvoboda,2007).Signalsubstancessuchasdopamineexhibitvolumetransmission(Rice,2000)andthismayhaveeffectforpotentiationofnearbysynapsesduringlearning:simulationsshowthatasinglesynapticreleasecanbedetectedupto20mawayandwitha100mshalflife(Cragg,Nicholsonetal.,2001).Largermoleculeshavetheirrelativediffusionspeedreducedbythelimitedgeometryoftheextracellularspace,bothintermsofitstortuosityanditsanisotropy(Nicholson,2001).AssuggestedbyRobertFreitas,theremayalsoexistactiveextracellulartransportmodes.DiffusionratesarealsoaffectedbylocalflowoftheCSFandcandifferfromregiontoregion(FenstermacherandKaye,1988);ifthisisrelevantthenlocaldiffusionandflowmeasurementsmaybeneededtodevelopatleastageneralbraindiffusionmodel.Thegeometricpartofsuchdatacouldberelativelyeasilygainedfromthehighresolution3DscansneededforotherWBEsubproblems.

  • 35

    Rapidandbroadvolumetransmissionsuchasfromnitrousoxidecanbesimulatedusingarelativelycoarsespatiotemporalgridsize,whilelocaltransmissionrequiresagridwithaspatialscaleclosetotheneuralscaleifdiffusionisseverelyhindered.Forconstrainingbrainemulationitmightbeusefultoanalysetheexpecteddiffusionanddetectiondistancesofthe200knownchemicalsignallingmoleculesbasedontheirmolecularweight,diffusionconstantanduptake(fordifferentlocalneuralgeometriesandsource/sinkdistributions).Thiswouldprovideinformationondiffusiontimesthatconstrainthediffusionpartoftheemulationandpossiblyshowwhichchemicalspeciesneedtobespatiallymodelled.

    BodychemicalenvironmentThebodyactsasaninput/outputunitthatinteractswithourperceptionandmotoractivity.Italsoactsasachemicalenvironmentthataffectsthebrainthroughnutrients,hormones,salinity,dissolvedgases,andpossiblyimmunesignals.Mostofthesechemicalsignalsoccuronasubconsciouslevelandonlybecomeapparentwhentheyinfluencee.g.hypothalamustoproducehungerorthirstsensations.Forbrainemulation,someaspectsofthischemicalenvironmenthastobesimulated.Thiswouldrequiremappingthehumanmetabolome,atleastinregardstosubstancesthatcrossthebloodbrainbarrier.Themetabolomeislikelyontheorderof2,0002,500compounds(Beecher,2003;Wishart,Tzuretal.,2007)andlargelydoesnotchangemorerapidlythanonthesecondtimescale.ThissuggeststhatcomparedtothedemandsoftheWBE,thebodychemistrymodel,whileinvolved,wouldberelativelysimple.Ifaproteininteractionmodelisneededratherthanmetabolism,thencomplexityincreases.Accordingtooneestimatethehumaninteractomeisaround650,000proteinproteininteractions(Stumpf,Thorneetal.,2008).

    NeurogenesisandremodellingRecentresultsshowthatneurogenesispersistsinsomebrainregionsinadulthood,andmighthavenontrivialfunctionalconsequences(Saxe,Malleretetal.,2007).Duringneuriteoutgrowth,andpossiblyafterwards,celladhesionproteinscanaffectgeneexpressionandpossibleneuronfunctionbyaffectingsecondmessengersystemsandcalciumlevels(CrossinandKrushel,2000).However,neurogenesisismainlyconfinedtodiscreteregionsofthebrainanddoesnotoccurtoagreatextentinadultneocortex(Bhardwaj,Curtisetal.,2006).Sinceneurogenesisoccursonfairlyslowtimescales(>1week)comparedtobrainactivityandnormalplasticity,itcouldprobablybeignoredinbrainemulationifthegoalisanemulationthatisintendedtofunctionfaithfullyforonlyafewdaysandnottoexhibittrulylongtermmemoryconsolidationoradaptation.Arelatedissueisremodellingofdendritesandsynapses.Overthespanofmonthsdendritescangrow,retractandaddnewbranchtipsinacelltypespecificmanner(Lee,Huangetal.,2006).Similarlysynapticspinesintheadultbraincanchangewithinhourstodays,althoughthemajorityremainstableovermultimonthtimespans(Grutzendler,Kasthurietal.,2002;Holtmaat,Trachtenbergetal.,2005;Zuo,Linetal.,2005).Evenifneurogenesisisignoredandtheemulationisofanadultbrain,itislikelythatsuchremodellingisimportanttolearningandadaptation.

  • 36

    Simulatingstemcellproliferationwouldrequiredatastructuresrepresentingdifferentcellsandtheirdifferentiationstatus,dataonwhattriggersneurogenesis,andmodelsallowingforthegradualintegrationofthecellsintothenetwork.Suchasimulationwouldinvolvemodellingthegeometryandmechanicsofcells,possiblyeventissuedifferentiation.Dendriticandsynapticremodellingwouldalsorequireageometryandmechanicsmodel.Whiletechnicallyinvolvedandrequiringatleastageometrymodelforeachdendriticcompartmentthecomputationaldemandsappearsmallcomparedtoneuralactivity.

    GliacellsGliacellshavetraditionallybeenregardedasmerelysupportingactorstotheneurons,butrecentresultssuggestthattheymayplayafairlyactiveroleinneuralactivity.Besidetheimportantroleofmyelinizationforincreasingneuraltransmissionspeed,attheveryleasttheyhavestrongeffectsonthelocalchemicalenvironmentoftheextracellularspacesurroundingneuronsandsynapses.Glialcellsexhibitcalciumwavesthatspreadalongglialnetworksandaffectnearbyneurons(NewmanandZahs,1998).Theycanbothexciteandinhibitnearbyneuronsthroughneurotransmitters(Kozlov,Anguloetal.,2006).Conversely,thecalciumconcentrationofglialcellsisaffectedbythepresenceofspecificneuromodulators(PereaandAraque,2005).Thissuggeststhattheglialcellsactsasaninformationprocessingnetworkintegratedwiththeneurons(FellinandCarmignoto,2004).Onerolecouldbeinregulatinglocalenergyandoxygensupply.Ifglialprocessingturnsouttobesignificantandfinegrained,brainemulationwouldhavetoemulatethegliacellsinthesamewayasneurons,increasingthestoragedemandsbyatleastoneorderofmagnitude.However,thetimeconstantsforglialcalciumdynamicsisgenerallyfarslowerthanthedynamicsofactionpotentials(ontheorderofsecondsormore),suggestingthatthetimeresolutionwouldnothavetobeasfine,makingthecomputationaldemandsincreasefarlesssteeply.

    EphapticeffectsElectricaleffectsmayalsoplayaroleviasocalledephaptictransmission.Inahighresistanceenvironment,currentsfromactionpotentialsareforcedtoflowthroughneighbouringneurons,changingtheirexcitability.Ithasbeenclaimedthatthisprocessconstitutesaformofcommunicationinthebrain,inparticularthehippocampus(Krnjevic,1986).However,inmostpartsofthebrainthereisalargeextracellularspaceandblockingmyelin,soevenifephapticinteractionsplayarole,theydosoonlylocally,e.g.intheolfactorysystem(Bokil,Laarisetal.,2001),densedemyelinatednervebundles(Reutskiy,Rossonietal.,2003),ortrigeminalpainsyndromes(LoveandCoakham,2001).Itshouldbenotedthatthenervoussystemappearsrelativelyinsensitivetoeverydayexternalelectricfields(Valberg,Kavetetal.,1997;SwansonandKheifets,2006).Ifephapticeffectswereimportant,theemulationwouldneedtotakethelocallyinducedelectromagneticfieldsintoaccount.Thiswouldplausiblyinvolvedividingtheextracellularspace(possiblyalsotheintracellularspace)intofiniteelementswherethefieldcanbeassumedtobeconstant,linearorotherwiseeasilyapproximable.Thecorticalextracellularlengthconstantisonorderof100m(GardnerMedwin,1983),whichwouldnecessitateontheorderof1.41012suchcompartmentsifeachcompartmentis1/10ofthelengthconstant.

  • 37

    Eachcompartmentwouldneedatleasttwovectorstatevariablesand6componentsofaconductivitytensor;assumingonebyteforeach,thetotalmemoryrequirementswouldbeontheorderof10terabytes.Comparedtoestimatesofneuralsimulationcomplexity,thisisrelativelymanageable.Theprocessingneededtoupdatethesecompartmentswouldbeonthesameorderasadetailedcompartmentmodelofeveryneuronandgliacell.

    DynamicalstateThemethodsforcreatingthenecessarydataforbrainemulationdiscussedinthispaperdealwithjustthephysicalstructureofthebraintissue,notitsstateofactivity.Someinformationsuchasworkingmemorymaybestoredjustasongoingpatternsofneuralexcitationandwouldbelost.Similarly,informationincalciumconcentrations,synapticvesicledepletion,anddiffusingneuromodulatorsmaybelostduringscanning.Alikelyconsequencewouldbeamnesiaofthetimeclosesttothescanning.However,lossofbrainactivitydoesnotseemtopreventthereturnoffunctionandpersonalidentity.Thisisdemonstratedbythereawakeningofcomapatients,andbycoldwaterneardrowningcasesinwhichbrainactivitytemporarilyceasedduetohypothermia(Elixson,1991).

    QuantumcomputationWhilepracticallyallneuroscientistssubscribetothedogmathatneuralactivityisaphenomenonthatoccursonaclassicalscale,therehavebeenproposals(mainlyfromphysicists)thatquantumeffectsplayanimportantroleinthefunctionofthebrain(Penrose,1989;Hameroff,1987).Sofarthereisnoevidenceforquantumeffectsinthebrainbeyondquantumchemistry,andnoevidencethatsucheffectsplayanimportantroleforintelligenceorconsciousness(Litt,Eliasmithetal.,2006).Thereisnolackofpossiblecomputationalprimitivesinneurobiologynoranyphenomenathatappearunexplainableintermsofclassicalcomputations(KochandHepp,2006).Quantitativeestimatesfordecoherencetimesforionsduringactionpotentialsandmicrotubulessuggestthattheydecohereonatimescaleof10201013s,abouttenordersofmagnitudefasterthanthenormalneuralactivitytimescales.Hencequantumeffectsareunlikelytopersistlongenoughtoaffectprocessing(Tegmark,2000).This,however,hasnotdeterredsupportersofquantumconsciousness,whoarguethattheremaybemechanismsprotectingquantumsuperpositionsoversignificantperiods(RosaandFaber,2004;Hagan,Hameroffetal.,2002).Ifthesequantummindhypothesesweretrue,brainemulationwouldbesignificantlymorecomplex,butnotimpossiblegiventheright(quantum)computer.In(Hameroff,1987)mindemulationisconsideredbasedonquantumcellularautomata,whichinturnarebasedonthemicrotubulenetworkthattheauthorsuggestsunderliesconsciousness.Assuming7.1microtubulespersquaremand768.9minaveragelength(Cash,Alievetal.,2003)andthat1/30ofbrainvolumeisneurons(althoughgiventhatmicotubulinetworksoccursinallcells,gliaandanyothercelltype!maycounttoo)gives1016microtubules.Ifeachstoresjustasinglequantumbitthiswouldcorrespondtoa1016qubitsystem,requiringaphysicallyintractable210^16bitclassicalcomputertoemulate.Ifonlythemicrotubulesinsideacellactasaquantumcomputingnetwork,theemulationwouldhavetoinclude1011connected130,000qubitquantumcomputers.Anothercalculation,assumingmerelyclassicalcomputationinmicrotubules,suggests1019bytesperbrainoperatingat1028FLOPS(Tuszynski,2006).Oneproblemwiththesecalculationsisthattheyimputesuchaprofoundly

  • 38

    largecomputationalcapacityatasubneurallevelthatamacroscopicbrainseemsunnecessary(especiallysinceneuronsaremetabolicallycostly).

    AnalogcomputationAsurprisinglycommondoubtexpressedaboutthepossibilityofsimulatingevensimpleneuralsystemsisthattheyareanalogratherthandigital.Thedoubtisbasedontheassumptionthatthereisanimportantqualitativedifferencebetweencontinuousanddiscretevariables.Ifcomputationsinthebrainmakeuseofthefullpowerofcontinuousvariablesthebrainmayessentiallybeabletoachievehypercomputation,enablingittocalculatethingsanordinaryTuringmachinecannot(Ord,2006;SiegelmannandSontag,1995).See(ZenilandHernandezQuiroz,2007)forare