Download - Interactions between Intrinsic and StimulusEvoked ...

Transcript
Page 1: Interactions between Intrinsic and StimulusEvoked ...

1

InteractionsbetweenIntrinsicandStimulus­Evoked

ActivityinRecurrentNeuralNetworks

L.F.AbbottandKanakaRajan

DepartmentofNeuroscienceDepartmentofPhysiologyandCellularBiophysics

ColumbiaUniversityCollegeofPhysiciansandSurgeonsNewYork,NY10032‐2695USA

and

HaimSompolinsky

RacahInstituteofPhysicsInterdisciplinaryCenterforNeuralComputation

HebrewUniversityJerusalem,Israel

Introduction

Trial‐to‐trialvariabilityisanessentialfeatureofneuralresponses,butitssourceisasubjectofactivedebate.Responsevariability(MastandVictor,1991;Arielietal.,1995&1996;Andersonetal.,2000&2001;Kenetetal.,2003;Petersenetal.,2003a&b;Fiser,ChiuandWeliky,2004;MacLeanetal.,2005;Yusteetal.,2005;Vincentetal.,2007)isoftentreatedasrandomnoise,generatedeitherbyotherbrainareas,orbystochasticprocesseswithinthecircuitrybeingstudied.Wecallsuchsourcesofvariability“external”tostresstheindependenceofthisformofnoisefromactivitydrivenbythestimulus.Variabilitycanalsobegeneratedinternallybythesamenetworkdynamicsthatgeneratesresponsestoastimulus.Howcanwedistinguishbetweenexternalandinternalsourcesofresponsevariability?Hereweshowthatinternalsourcesofvariabilityinteractnonlinearlywithstimulus‐inducedactivity,andthisinteractionyieldsasuppressionofnoiseintheevokedstate.Thisprovidesatheoreticalbasisandpotentialmechanismfortheexperimentalobservationthat,inmanybrainareas,stimulicausesignificantsuppressionofneuronalvariability(WernerandMountcastle,1963;Fortier,SmithandKalaska,1993;Andersonetal.,2000;FriedrichandLaurent,2004;Churchlandetal.,2006;Finn,PriebeandFerster,2007;Mitchell,SundbergandReynolds,2007;Churchlandetal.,2009).Thecombinedtheoreticalandexperimentalresultssuggestthatinternallygeneratedactivityisasignificantcontributortoresponsevariabilityinneuralcircuits.

Weareinterestedinuncoveringtherelationshipbetweenintrinsicandstimulus‐evokedactivityinmodelnetworksandstudyingtheselectivityofthesenetworkstofeaturesofthestimulidrivingthem.Therelationshipbetweenintrinsicand

Page 2: Interactions between Intrinsic and StimulusEvoked ...

2

extrinsicallyevokedactivityhasbeenstudiedexperimentallybycomparingactivitypatternsacrosscorticalmaps(Arielietal.,1995&1996).Wedeveloptechniquesforperformingsuchcomparisonsincaseswherethereisnoapparentsensorymap.Inadditiontorevealinghowthetemporalandspatialstructureofspontaneousactivityaffectsevokedresponses,thesemethodscanbeusedtoinferinputselectivity.Historically,selectivitywasfirstmeasuredbystudyingstimulus‐drivenresponses(HubelandWiesel,1962),andonlylaterweresimilarselectivitypatternsobservedinspontaneousactivityacrossthecorticalsurface(Arielietal.,1995&1996).Wearguethatitispossibletoworkinthereverseorder.Havinglittleinitialknowledgeofsensorymapsinournetworks,weshowhowtheirspontaneousactivitycaninformusabouttheselectivityofevokedresponsestoinputfeatures.Throughoutthisstudy,werestrictourselvestoquantitiesthatcanbemeasuredexperimentally,suchasresponsecorrelations,soouranalysismethodscanbeappliedequallytotheoreticalmodelsandexperimentaldata.

Webeginbydescribingthenetworkmodelandillustratingthetypesofactivityitproduces,usingcomputersimulations.Inparticular,weillustrateanddiscussatransitionbetweentwotypesofresponses;oneinwhichintrinsicandstimulus‐evokedactivitycoexist,andtheotherinwhichintrinsicactivityiscompletelysuppressed.Next,weexplorehowthespatialpatternsofspontaneousandevokedresponsesarerelated.Byspatialpattern,wemeanthewaythatactivityisdistributedacrossthedifferentneuronsofthenetwork.Spontaneousactivityisausefulindicatorofrecurrenteffects,becauseitiscompletelydeterminedbynetworkfeedback.Therefore,westudytheimpactofnetworkconnectivityonthespatialpatternofinput‐drivenresponsesbycomparingthespatialstructureofevokedandspontaneousactivity.Finally,weshowhowthestimulusselectivityofthenetworkcanbeinferredfromananalysisofitsspontaneousactivity.

TheModel

Neuronsinthemodelweconsideraredescribedbyfiring‐rates,theydonotfireindividualactionpotentials.Suchfiring‐ratenetworksareattractivebecausetheyareeasiertosimulatethanspikingnetworkmodelsandareamenabletomoredetailedmathematicalanalyses.Ingeneral,aslongasthereisnolarge‐scalesynchronizationofactionpotentials,firing‐ratemodelsdescribenetworkactivityadequately(Shriki,HanselandSompolinsky,2003;WongandWang,2006).WeconsideranetworkofNinterconnectedneurons,withneuronicharacterizedbyanactivationvariablexisatisfying

τdxidt

= −xi + g Jijrj + Iij=1

N

∑ .

Thetimeconstantτissetto10ms.Forallofthefigures,except5e,N=1000.TherecurrentsynapticweightmatrixJhaselementJijdescribingtheconnectionfrompresynapticneuronjtopostsynapticneuroni.Excitatoryconnectionscorrespond

Page 3: Interactions between Intrinsic and StimulusEvoked ...

3

topositivematrixelements,inhibitoryconnectionstonegativeelements.Theinputterm,Iiforneuroni,takesvariousformsthatwillbedescribedasweusethem.

Thefiringrateofneuroniisgivenby

ri = R0 + φ(x i) with

φ(x) = R0 tanh(x /R0)for

x ≤ 0 and

φ(x) = (Rmax − R0)tanh(x /(Rmax − R0)) for

x > 0 .Here,R0isthebackgroundfiringrate(thefiringratewhenx=0),andRmaxisthemaximumfiringrate.Thisfunctionallowsustospecifyindependentlythemaximumfiringrate,Rmax,andthebackgroundrate,R0,andsetthemtoreasonablevalues,whileretainingthegeneralformofthecommonlyusedtanhfunction.ThisfiringratefunctionisplottedinFigure1forR0=0.1Rmax,thevalueweuse.Tofacilitatecomparisonwithexperimentaldatainavarietyofsystems,wereportallresponsesrelativetoRmax.Similarly,wereportallinputcurrentsrelativetothecurrentI1/2requiredtodriveanisolatedneurontohalfofitsmaximalfiringrate(seeFigure1).

Figure1.ThefiringratefunctionusedinthenetworkmodelversustheinputIforR0=0.1RmaxnormalizedbythemaximumfiringrateRmax.TheparameterI1/2isdefinedbythedashedlines.

Althoughaconsiderableamountisknownaboutthestatisticalpropertiesof

recurrentconnectionsincorticalcircuitry(Holmgrenetal.,2003;Songetal.,2005),wedonothaveanythinglikethespecificneuron‐to‐neuronwiringdiagramwewouldneedtobuildatrulyfaithfulmodelofacorticalcolumnorhypercolumn.Instead,weconstructtheconnectionmatrixJofthemodelnetworkonthebasisofastatisticaldescriptionoftheunderlyingcircuitry.WedothisbychoosingelementsofthesynapticweightmatrixindependentlyandrandomlyfromaGaussiandistributionwithzeromeanandvariance1/N.Wecoulddividethenetworkintoseparateexcitatoryandinhibitorysubpopulations,butthisdoesnotqualitativelychangethenetworkpropertiesthatwediscuss(vanVreeswijkandSompolinsky,1996&1998;RajanandAbbott,2006).

Theparametergcontrolsthestrengthofthesynapticconnectionsinthemodel,butbecausethesestrengthsarechosenfromadistributionwithzeromeanandnonzerovariance,gactuallycontrolsthesizeofthestandarddeviationofthesynapticstrengths(seeDiscussion).Withoutanyinput(Ii=0foralli)andforlargenetworks(largeN),twospontaneouspatternsofactivityareseen.Ifg<1,thenetworkisinatrivialstateinwhichxi=0andri=R0forallneurons(alli).Thecaseg>1ismoreinterestinginthatthespontaneousactivityofthenetworkischaotic,meaningthatitisirregular,non‐repeatingandhighlysensitivetoinitialconditions

Page 4: Interactions between Intrinsic and StimulusEvoked ...

4

(Sompolinsky,Crisanti,andSommers,1988;vanVreeswijkandSompolinsky,1996&1998).Wetypicallyuseavalueofg=1.5,meaningthatournetworksareinthischaoticstatepriortoactivatinganyinputs.

Figure2.Firingratesandresponsevariabilitynormalizedbythemaximumfiringrate,Rmax,forastimulusinputsteppingfromzerotoaconstant,non‐zerovalueatt=1000ms.Leftcolumnshowsthefiringrateofatypicalneuroninanetworkwith1000neurons.Rightcolumnshowstheaveragefiringrate(redtraces)andthesquarerootoftheaveragefiring‐ratevariance(blacktraces)acrossthenetworkneurons.a‐b)Anetworkwithchaoticspontaneousactivityreceivingnonoiseinput.Theresponsevariability(b,blacktrace)dropstozerowhenthestimulusinputispresent.c‐d)Anetworkwithoutspontaneousactivitybutreceivingnoiseinput.Theresponsevariability(d,blacktrace)risesslightlywhenthestepinputisturnedon.Thestochasticinputinthisexamplewasindependentwhite‐noisetoeachneuron,low‐passfilteredwitha500mstimeconstant.ResponsestoStepInputTobeginourexaminationoftheeffectsofinputonchaoticspontaneousnetworkactivity,weconsidertheeffectofastepofinput(from0toapositivevalue),applieduniformlytoeveryneuron(Ii=Iforalli).Beforetheinputisturnedon(Figure2a&b,t<1000ms),atypicalneuronofthenetworkshowsthehighlyirregularactivitycharacteristicofthechaoticspontaneousstate.However,whenasufficientlystrongstimulusisapplied,theinternallygeneratedfluctuationsarecompletelysuppressed(Figure2a&b,t>1000ms).Wecontrastthisbehaviortothatofexternalnoise,byturningofftherecurrentdynamicsandgeneratingfluctuationswithexternal

Page 5: Interactions between Intrinsic and StimulusEvoked ...

5

stochasticinputs(Figure1c&d,t<1000ms).Inthiscase,thereisnoreductionintheamplitudeoftheneuronalfluctuationswhenastimulusisapplied,infactthereisasmallincrease.Notethattheincreaseinthemeanactivityissimilarinbothcases.Theseresultsrevealacriticaldistinctionbetweeninternallyandexternallygeneratedfluctuations–theformercanbesuppressedbyastimulusandthereforedonotnecessarilyinterferewithsensoryprocessing.

Torevealthenatureofinternallygeneratedvariability,wehaveconsideredanidealizedscenarioinFigure2inwhichtherewasnoexternalsourceofnoise.Inreality,weexpectbothexternalandinternalsourcesofnoisetocoexistinlocalcorticalcircuits.Aslongastheinternalnoiseprovidesasubstantialcomponentoftheoverallvariability,ourqualitativeresultsremainvalid.Wecansimulatethissituationbyaddingexternalnoise(asinFigure2c&d)toamodelthatexhibitschaoticspontaneousactivity(asinFigure2a&b).Theresultshowsasharpdropinvarianceatstimulusonset,butwithonlypartial,ratherthancomplete,suppressionofresponsevariability(Figure3).Thisresultisingoodagreementwithexperimentaldata(Churchlandetal.,2009).

Figure3.Firingratesandresponsevariabilitynormalizedbythemaximumfiringrate,Rmax,forthesamenetwork,stimulusandnoiseasinFigure2,butforanetworkwithbothspontaneousactivityandinjectednoise.a)Theresponseofatypicalneuron.b)Theaveragefiringrate(redtrace)increasesandtheresponsevariability(blacktrace)decreaseswhenthestimulusinputispresent,butitdoesnotgotozeroasinFigure2b.

ResponsetoPeriodicInputForFigures2&3,thestimulusconsistedofastepinputappliedidenticallytoallneurons.Toinvestigatetheeffectofmoreinterestingandrealisticstimulionthechaoticactivityofarecurrentnetwork,weconsiderinputswithnon‐homogeneousspatio‐temporalstructure.Specifically,weintroduceinputsthatoscillateinasinusoidalmannerwithamplitudeIandfrequencyfandexaminehowthesuppressionoffluctuationsdependsontheiramplitudeandfrequency.Inmanycases,neuronsinalocalpopulationhavediversestimulusselectivities,soa

Page 6: Interactions between Intrinsic and StimulusEvoked ...

6

particularstimulusmayinducelittlechangeinthetotalactivityacrossthenetwork.Tomimicthissituation,wegivetheseoscillatinginputsadifferentphaseforeachneuron(intermsofavisualstimulus,thisisequivalenttopresentingastationary,counterphasegratingtoapopulationofsimplecellswithdifferentspatial‐phaseselectivities).Specifically,Ii=Icos(2πft+θi),whereθiischosenrandomlyfromauniformdistributionbetween0and2π.Therandomlyassignedphasesensurethatthespatialpatternofinputinourmodelnetworkisnotcorrelatedwiththepatternofrecurrentconnectivity.

Figure4.Achaoticnetworkof1000neuronsreceivingsinusoidal5Hzinput.a)Firingratesoftypicalnetworkneurons(normalizedbyRmax).b)Thelogarithmofthepowerspectrumoftheactivityacrossthenetwork.i)Withnoinput(I=0),networkactivityischaotic.ii)Inthepresenceofaweakinput(I/I1/2=0.1),anoscillatoryresponseissuperposedonchaoticfluctuations.iii)Forastrongerinput(I/I1/2=0.5),thenetworkresponseisperiodic.

Intheabsenceofastimulusinput,thefiringratesofindividualneuronsfluctuateirregularly,asseeninFigure2(Figure4a‐i),andthepowerspectrumacrossnetworkneuronsiscontinuousanddecaysexponentiallyasafunctionoffrequency(Figure4b‐i),acharacteristicfeaturesofthechaoticstateofthisnetwork(Sompolinsky,Crisanti,andSommers,1988).Whenthenetworkisdrivenbyaweakoscillatoryinput,thesingle‐neuronresponseisasuperpositionofaperiodicpatterninducedbytheinputandachaoticbackground(Figure4a‐ii).Thepowerspectrumshowsacontinuouscomponentduetotheresidualchaosandpeaksatthefrequencyoftheinputanditsharmonics,reflectingtheperiodicbutnon‐sinusoidalcomponentoftheresponse(Figure4b‐ii).Foraninputwithalargeramplitude,thefiringratesofnetworkneuronsareperiodic(Figure4a‐iii),andthepowerspectrumshowsonlypeaksattheinputfrequencyanditsharmonics,withnocontinuous

Page 7: Interactions between Intrinsic and StimulusEvoked ...

7

spectrum(Figure4b‐iii).ThisindicatesacompletesuppressionoftheinternallygeneratedfluctuationsasinFigure2a&b.

Wehaveusedamean‐fieldapproachsimilartothatdevelopedbySompolinsky,Crisanti,andSommers(1988)toanalyzepropertiesofthetransitionbetweenchaoticaperiodicresponsestoaperiodicstimulus(Rajan,AbbottandSompolinsky,2009).Thisextendspreviousworkontheeffectofinputonchaoticnetworkactivity(Molgedey,SchuchhardtandSchuster,1992;BertchingerandNatschläger,2004)tocontinuoustimemodelsandperiodicinputs.Wefindthatthereisacriticalinputintensity(acriticalvalueofI)thatdependsonfandg,belowwhichnetworkactivityischaoticthoughdrivenbytheinput(asinFigures4a‐ii&4b‐ii)andabovewhichitisperiodic(asinFigures4a‐iii&4b‐iii).Asurprisingfeatureofthiscriticalamplitudeisthatitisanon‐monotonicfunctionofthefrequencyfoftheinput.Asaresult,thereisa“best”frequencyatwhichitiseasiesttoentrainthenetworkandsuppresschaos.Fortheparametersweuse,the“best”frequencyisaround5Hz,afrequencyweremanysensorysystemstendtooperate,andtherearesomeinitialexperimentalindicationsthatthisisindeedtheoptimalfrequencyforsuppressingbackgroundactivitybyvisualstimulation(WhiteandFiser,2008).Itisinterestingthatapreferredinputfrequencyforentrainmentariseseventhoughthepowerspectrumofthespontaneousactivitydoesnotshowanyresonantfeatures(Figure4b‐i).

PrincipalComponentAnalysisofSpontaneousandEvokedActivityTheresultsoftheprevioustwosectionsrevealedaregimeinwhichaninputgeneratesanon‐chaoticnetworkresponse,eventhoughthenetworkischaoticintheabsenceofinput.Althoughthechaoticintrinsicactivityhasbeencompletelysuppressedinthisnetworkstate,itsimprintcanstillbedetectedinthespatialpatternofthenon‐chaoticactivity.

ThenetworkstateatanyinstantcanbedescribedbyapointinanN‐dimensionalspacewithcoordinatesequaltothefiringratesoftheNneurons.Overtime,activitytraversesatrajectoryinthisN‐dimensionalspace.Principalcomponentanalysiscanbeusedtodelineatethesubspaceinwhichthistrajectorypredominantlylies.Theanalysisisdonebydiagonalizingtheequal‐timecross‐correlationmatrixofnetworkfiringrates,<ri(t)rj(t)>,wheretheanglebracketsdenoteanaverageovertimeTheeigenvaluesofthismatrixexpressedasafractionoftheirsum(denotedby

˜ λ a ),indicatethedistributionofvariancesacrossdifferentorthogonaldirectionsintheactivitytrajectory.Inthespontaneousstate,thereareanumberofsignificantcontributorstothetotalvariance,asindicatedinFigure5a.Forthisvalueofg,theleading10%ofthecomponentsaccountfor90%ofthetotalvariance.Thevarianceassociatedwithhighercomponentsfallsoffexponentially.Itisinterestingtonotethattheprojectionsofthenetworkactivityontotheprincipalcomponentdirectionsfluctuatemorerapidlyforhighercomponents(Figure5c),revealingtheinteractionbetweenthespatialandtemporalstructureofthechaoticfluctuations.

Page 8: Interactions between Intrinsic and StimulusEvoked ...

8

Figure5:Principalcomponentanalysisofthechaoticspontaneousstateandnon‐chaoticdrivenstate.a)Percentvarianceaccountedforbydifferentprincipalcomponentsforchaoticspontaneousactivity.b)Sameasa,butfornon‐chaoticdrivenactivity.c)Projectionsofthechaoticspontaneousactivityontoprincipalcomponentvectors1,10and50(indecreasingorderofvariance).d)Projectionsofperiodicdrivenactivityontoprincipalcomponents1,3,and5.Projectionsontocomponents2,4,and6aresimilarexceptforbeingphaseshiftedbyπ/2.e)Theeffectivedimension,Neff,ofthetrajectoryofchaoticspontaneousactivity(definedinthetext)asafunctionofgfornetworkswith1000(solidcircles)or2000(opencircles)neurons.Parameters:g=1.5fora‐d,andf=5andI/I1/2=0.7forbandd.

Thenon‐chaoticdrivenstateisapproximatelytwodimensional(Figure5b),with

thetwodimensionsdescribingacircularoscillatoryorbit.Projectionsofthisorbitcorrespondtotheoscillationsπ/2apartinphase.Theresidualvarianceinthehigherdimensionsreflectshigherharmonicsarisingfromnetworknonlinearity,asillustratedbytheprojectionsinFigure5d.

Toquantifythedimensionofthesubspacecontainingthechaotictrajectoryinmoredetail,weintroducethequantity

Neff = ˜ λ a2

a=1

N

−1

.

Thisprovidesameasureoftheeffectivenumberofprincipalcomponentsdescribingatrajectory.Forexample,ifnprincipalcomponentssharethetotalvarianceequally,andtheremainingN­nprincipalcomponentshavezerovariance,Neff=n.Forthechaoticspontaneousstateinthenetworkswestudy,Neffincreaseswithg(Figure5e),duetothehigheramplitudeandfrequencycontentofthechaoticactivityforlargeg.NotethatNeffscalesapproximatelywithN,whichmeansthatlargenetworkshaveproportionallyhigher‐dimensionalchaoticactivity(comparethetwotracesinFigures5e).Thefactthatthenumberofactivatedmodesisonly2%ofthesystemdimensionality,evenforgashighas2.5,isanothermanifestationofthedeterministicnatureofthefluctuations.Forcomparison,wecalculatedNeffforasimilarnetworkdrivenbyexternalwhitenoise,withgsetbelowthechaotictransitionatg=1.Inthiscase,Neffonlyassumesuchlowvalueswhengiswithinafewpercentofthecriticalvalue1.TheresultsinFigure5illustrateanotherfeature

Page 9: Interactions between Intrinsic and StimulusEvoked ...

9

ofthesuppressionofspontaneousactivitybyinput,whichisthatthePCAdimensionNeffisreduceddramaticallybythepresenceoftheinput.

NetworkEffectsontheSpatialPatternofEvokedActivityInthenon‐chaoticregime,thetemporalstructureofnetworkresponsesislargelydeterminedbytheinput;theybothoscillateatthesamefrequency,althoughthenetworkactivityincludesharmonicsnotpresentintheinput.Theinputdoesnot,however,exertnearlyasstrongcontrolonthespatialstructureofthenetworkresponse.Thephasesofthefiring‐rateoscillationsofnetworkneuronsareonlypartiallycorrelatedwiththephasesoftheinputsthatdrivethem,andtheyarestronglyinfluencedbytherecurrentfeedback.

Wehaveseenthattheorbitdescribingtheactivityinthenon‐chaoticdrivenstateconsistsprimarilyofacircleinatwo‐dimensionalsubspaceofthefullN‐dimensionsdescribingneuronalactivities.Wenowaskhowthiscirclealignsrelativetosubspacesdefinedbydifferentnumbersofprincipalcomponentsthatcharacterizethespontaneousactivity.Thisrelationshipisdifficulttovisualizebecauseboththechaoticsubspaceandthefullspaceofnetworkactivitiesarehighdimensional.Toovercomethisdifficulty,wemakeuseofthenotionof“principalangles”betweensubspaces(IpsenandMeyer,1995).

Thefirstprincipalangleistheanglebetweentwounitvectors(calledprincipalvectors),oneineachsubspace,thathavethemaximumoverlap(dotproduct).Higherprincipalanglesaredefinedrecursivelyastheanglesbetweenpairsofunitvectorswiththehighestoverlapthatareorthogonaltothepreviouslydefinedprincipalvectors.Specifically,fortwosubspacesofdimensiond1andd2definedbytheorthogonalunitvectorsV1a,fora=1,2,...,d1andV2b,forb=1,2,...,d2,thecosinesoftheprincipalanglesareequaltothesingularvaluesofthed1byd2matrixformedfromallthepossibledotproductsofthesetwovectors.Theresultingprincipalanglesvarybetween0andπ/2withzeroanglesappearingwhenpartsofthetwosubspacesoverlapandπ/2correspondingtodirectionsinwhichthetwosubspacesarecompletelynon‐overlapping.Theanglebetweentwosubspacesisthelargestoftheirprincipalangles.ThisdefinitionisillustratedinFigure6awhereweshowtheirregulartrajectoryofthechaoticspontaneousactivity,describedbyitstwoleadingprincipalcomponents(blackcurveinFigure6a).Thecircularorbitoftheperiodicactivity(redcurveinFigure6a)hasbeenrotatedbythesmallerofitstwoprincipalangles.Theanglebetweenthesetwosubspaces(theangledepictedinFigure6a)isthentheremaininganglethroughwhichtheperiodicorbitwouldhavetoberotatedtobringitintoalignmentwiththehorizontalplanecontainingthetwo‐dimensionalprojectionofthechaotictrajectory.

Figure6ashowstheanglebetweenthesubspacesdefinedbythefirsttwoprincipalcomponentsoftheorbitofperiodicdrivenactivityandthefirsttwoprincipalcomponentsofthechaoticspontaneousactivity.Wenowextendthisideatoacomparisonofthetwo‐dimensionalsubspaceoftheperiodicorbitandsubspacesdefinedbythefirstmprincipalcomponentsofthechaoticspontaneousactivity.ThisallowsustoseehowtheorbitliesinthefullN‐dimensionalspaceof

Page 10: Interactions between Intrinsic and StimulusEvoked ...

10

neuronalactivitiesrelativetothetrajectoryofthechaoticspontaneousactivity.Theresults(Figure6b,reddots)showthatthisangleisclosetoπ/2forsmallm,equivalenttotheanglebetweentworandomlychosensubspaces.However,thevaluedropsquicklyforsubspacesdefinedbyprogressivelymoreoftheleadingprincipalcomponentsofthechaoticactivity.Ultimately,thisangleapproacheszerowhenallNofthechaoticprincipalcomponentvectorsareconsidered,asitmust,becausethesespantheentirespaceofnetworkactivities.

Figure6:Spatialpatternofnetworkresponses.a)Definitionoftheanglebetweenthesubspacedefinedbythefirsttwocomponentsofthechaoticactivity(blackcurve)andatwo‐dimensionaldescriptionoftheperiodicorbit(redcurve).b)Relationshipbetweentheorientationofperiodicandchaotictrajectories.Anglesbetweenthesubspacedefinedbythetwoprincipalcomponentsofthenon‐chaoticdrivenstateandsubspacesformedbyprincipalcomponents1throughmofthechaoticspontaneousactivity,wheremappearsonthehorizontalaxis(reddots).Blackdotsshowtheanalogousanglesbutwiththetwo‐dimensionalsubspacedefinedbyrandominputphasesreplacingthesubspaceofthenon‐chaoticdrivenactivity.c)Effectofinputfrequencyontheorientationoftheperiodicorbit.Theangle(verticalaxis)betweenthesubspacesdefinedbythetwoleadingprincipalcomponentsofnon‐chaoticdrivenactivityatdifferentfrequencies(horizontalaxis)andthesetwovectorsfora5Hzinputfrequency.d)Networkselectivitytodifferentspatialpatternsofinput.Signal(dashedcurvesandopencircles)andnoise(solidcurvesandfilledcircles)amplitudesinresponsetoinputsalignedtotheleadingprincipalcomponentsofthespontaneousactivityofthenetwork.Theinsetshowsalargerrangeonacoarserscale.Parameters:I/I1/2=0.7andf=5Hzforb,I/I1/2=1.0forc,andI/I1/2=0.2andf=2Hzford.

Page 11: Interactions between Intrinsic and StimulusEvoked ...

11

Intheperiodicstate,thetemporalphasesofthedifferentneuronsdeterminetheorientationoftheorbitinthespaceofneuronalactivities.Therapidlyfallinganglebetweenthisorbitandthesubspacesdefinedbyspatialpatternsdominatingthechaoticstate(Figure6b,reddots)indicatesthatthesephasesarestronglyinfluencedbytherecurrentconnectivitythatinturndeterminesthespatialpatternofthespontaneousactivity.Asanindicationofthemagnitudeofthiseffect,wenotethattheanglesbetweentherandomphasesinusoidaltrajectoryoftheinputtothenetworkandthesamechaoticsubspacesaremuchlargerthanthoseassociatedwiththeperiodicnetworkactivity(Figure6b,blackdots).

TemporalFrequencyModulationofSpatialPatternsAlthoughrecurrentfeedbackinthenetworkplaysanimportantroleinthespatialstructureofdrivennetworkresponses,thespatialpatternoftheactivityisnotfixedbutinsteadisshapedbyacomplexinteractionbetweenthedrivinginputandtheintrinsicnetworkdynamics.Itisthereforesensitivetoboththeamplitudeandthefrequencyofthisdrive.Toseethis,weexaminehowtheorientationoftheapproximatelytwo‐dimensionalperiodicorbitofdrivennetworkactivityinthenon‐chaoticregimedependsoninputfrequency.Weusethetechniqueofprincipalanglesdescribedintheprevioussection,toexaminehowtheorientationoftheoscillatoryorbitchangeswhentheinputfrequencyisvaried.Forcomparisonpurposes,wechoosethedominanttwo‐dimensionalsubspaceofthenetworkoscillatoryresponsestoadrivinginputat5Hzasareference.Wethencalculatetheprincipalanglesbetweenthissubspaceandthecorrespondingsubspacesevokedbyinputswithdifferentfrequencies.TheresultshowninFigure6cindicatesthattheorientationoftheorbitforthesedrivenstatesrotatesastheinputfrequencychanges.

ThefrequencydependenceoftheorientationoftheevokedresponseislikelyrelatedtotheeffectseeninFigure6cinwhichhigherfrequencyactivityisprojectedontohigherprincipalcomponentsofthespontaneousactivity.Thiscausestheorbitofdrivenactivitytorotateinthedirectionofhigher‐orderprincipalcomponentsofthespontaneousactivityasthestimulusfrequencyincreases.Inaddition,thelargerthestimulusamplitude,theclosertheresponsephasesoftheneuronswillbetotherandomphasesoftheirexternalinputs(resultsnotshown).

NetworkSelectivityWehaveshownthattheresponseofanetworktorandom‐phaseinputisstronglyaffectedbythespatialstructureofspontaneousactivity(Figure6b).Wenowaskifthespatialpatternsthatdominatethespontaneousactivityinanetworkcorrespondtothespatialinputpatternstowhichthenetworkrespondsmostvigorously.Ratherthanusingrandom‐phaseinputs,wenowalignedtheinputstoournetworkalongthedirectionsdefinedbydifferentprincipalcomponentsofitsspontaneousactivity.Specifically,theinputtoneuroniissettoIViacos(2πft),whereIistheamplitudefactorandViaistheithcomponentofprincipalcomponentvectoraofthe

Page 12: Interactions between Intrinsic and StimulusEvoked ...

12

spontaneousactivity.Theindexaisorderedsothata=1correspondstotheprincipalcomponentwithlargestvarianceanda=Ntheleast.Toanalyzetheresultsofusingthisinput,wedividetheresponseintoasignalcomponentcorrespondingtothetrial‐averagedresponse,andanoisecomponentconsistingofthefluctuationsaroundthisaverageresponse.Wecalltheamplitudeofthesignalcomponentoftheresponsethe“signalamplitude”andthestandarddeviationofthefluctuationsthe“noiseamplitude”.

AsseeninFigure6dtheamplitudeofthesignalcomponentoftheresponsedecreasesslowlyasafunctionofwhichprincipalcomponentisusedtodefinetheinput.Amoredramaticeffectisseenonthenoisecomponentoftheresponse.FortheinputamplitudeusedinFigure6d,inputsalignedtothefirst5principalcomponentsofthespontaneousactivitycompletelysuppressthechaoticnoise,resultinginperiodicdrivenactivity.Forhigher‐orderprincipalcomponents,thenetworkactivityischaotic.Thus,the“noise”showsmoresensitivitytothespatialstructureoftheinputthanthesignal.

DiscussionOurresultssuggestthatexperimentsthatstudythestimulus‐dependenceofthetypicallyignorednoisecomponentofresponsesshouldbeinterestingandcouldprovideinsightintothenatureandoriginofactivityfluctuations.Responsevariabilityandongoingactivityissometimesmodeledasarisingfromastochasticprocessexternaltothenetworkgeneratingtheresponses.Thisstochasticnoiseisthenaddedlinearlytothesignaltocreatethetotalneuronalactivityintheevokedstate.Ourresultsindicatethatrecurrentdynamicsofthecorticalcircuitislikelytocontributesignificantlytotheemergenceofirregularneuronalactivity,andthattheinteractionbetweensuchdeterministic“noise”andexternaldriveishighlynonlinear.Inourwork(Rajan,AbbottandSompolinsky,2009),wehaveshownthatthestimuluscausesastrongsuppressionofactivityfluctuationsandfurthermorethatthenonlinearinteractionbetweentherelativelyslowchaoticfluctuationsandthestimulusresultsinanon‐monotonicfrequencydependenceofthenoisesuppression.

Animportantfeatureofthenetworkswestudyisthatthevarianceofthesynapticstrengthsacrossthenetworkcontrolstheemergenceofinterestingcomplexdynamics.Thishasimportantimplicationsforexperimentsbecauseitsuggeststhatthemostinterestingandrelevantmodulatorsofnetworksmaybesubstancesoractivity‐dependentmodulationsthatdonotnecessarilychangepropertiesofsynapsesonaverage,butratherchangesynapticvariance.Synapticvariancecanbechangedeitherbymodifyingtherangeoverwhichsynapticstrengthsvaryacrossapopulationofsynapses,aswehavedonehere,orbymodifyingthereleaseprobabilityandvariabilityofquantalsizeatsinglesynapses.Suchmodulatorsmightbeviewedaslesssignificantbecausetheydonotchangethenetbalancebetweenexcitationandinhibition.However,networkmodelingsuggeststhatsuchmodulationsareofgreatimportanceincontrollingthestateoftheneuronalcircuit.

Page 13: Interactions between Intrinsic and StimulusEvoked ...

13

Therandomcharacteroftheconnectivityinournetworkprecludesasimpledescriptionofthespatialactivitypatternsintermsoftopographicallyorganizedmaps.Ouranalysisshowsthatevenincorticalareaswheretheunderlyingconnectivitydoesnotexhibitsystematictopography,dissectingthespatialpatternsoffluctuationsinneuronalactivitycanrevealimportantinsightaboutbothintrinsicnetworkdynamicsandstimulusselectivity.Principalcomponentanalysisrevealedthatdespitethefactthatthenetworkconnectivitymatrixisfullrank,theeffectivedimensionalityofthechaoticfluctuationsismuchsmallerthanthenumberofneuronsinthenetwork.Thissuppressionofspatialmodesismuchstrongerthanexpectedfromalinearnetworklow‐passfilteringaspatio‐temporalwhitenoiseinput.Furthermore,asinthetemporaldomain,activespatialpatternsexhibitstrongnonlinearinteractionbetweenexternaldrivinginputsandintrinsicdynamics.Surprisingly,evenwhenthestimulusamplitudeisstrongenoughtofullyentrainthetemporalpatternofnetworkactivity,spatialorganizationoftheactivityisstillstronglyinfluencedbyrecurrentdynamics,asshowninFigures6cand6d.

Wehavepresentedtoolsforanalyzingthespatialstructureofchaoticandnon‐chaoticpopulationresponsesbasedonprincipalcomponentanalysisandanglesbetweentheresultingsubspaces.Principalcomponentanalysishas,beenappliedprofitablytoneuronalrecordings(see,forexample,Broome,JayaramanandLaurent,2006).Theseanalysesoftenplotactivitytrajectoriescorrespondingtodifferentnetworkstatesusingthefixedprincipalcomponentcoordinatesderivedfromcombinedactivitiesunderallconditions.Ouranalysisoffersacomplementaryapproachwherebyprincipalcomponentsarederivedforeachstimulusconditionseparately,andprincipalanglesareusedtorevealnotonlythedifferencebetweentheshapesoftrajectoriescorrespondingtodifferentnetworkstates,butalsothedifferenceintheorientationofthelowdimensionalsubspacesofthesetrajectorieswithinthefullspaceofneuronalactivity.

Manymodelsofselectivityincorticalcircuitsrelyonknowledgeofthespatialorganizationofafferentinputsaswellascorticalconnectivity.However,inmanycorticalareas,suchinformationisnotavailable.Ourresultsshowthatexperimentallyaccessiblespatialpatternsofspontaneousactivity(e.g.fromvoltage‐orcalcium‐sensitiveopticalimagingexperiments)canbeusedtoinferthestimulusselectivityinducedbythenetworkdynamicsandtodesignspatiallyextendedstimulithatevokestrongresponses.Thisisparticularlytruewhenselectivityismeasuredintermsoftheabilityofastimulustoentraintheneuraldynamics,asinFigure6d.Ingeneral,ourresultsindicatethattheanalysisofspontaneousactivitycanprovidevaluableinformationaboutthecomputationalimplicationsofneuronalcircuitry.

AcknowledgmentsResearchofKRandLAsupportedbyNationalScienceFoundationgrantIBN‐0235463andanNIHDirector'sPioneerAward,partoftheNIHRoadmapforMedicalResearch,throughgrantnumber5‐DP1‐OD114‐02.HSispartiallysupportedbygrantsfromtheIsraelScienceFoundationandtheMcDonnell

Page 14: Interactions between Intrinsic and StimulusEvoked ...

14

Foundation.ThisresearchwasalsosupportedbytheSwartzFoundationthroughtheSwartzCentersatColumbiaandHarvardUniversities.

ReferencesAnderson,J.S.,Lampl,I.,Gillespie,D.C.andFerster,D.Thecontributionofnoiseto

contrastinvarianceoforientationtuningincatvisualcortex.Science290,1968‐1972(2000).

Anderson,J.S.,Lampl,I.,Gillespie,D.C.andFerster,D.Membranepotentialandconductancechangesunderlyinglengthtuningofcellsincatprimaryvisualcortex.J.Neurosci.21,2104‐2112(2001).

Arieli,A.,Shoham,D.,Hildesheim,R.andGrinvald,A.Coherentspatiotemporalpatternsofongoingactivityrevealedbyreal‐timeopticalimagingcoupledwithsingle‐unitrecordinginthecatvisualcortex.J.Neurophysiol.73,2072‐2093(1995).

Arieli,A.,Sterkin,A.,Grinvald,A.andAertsen,A.Dynamicsofongoingactivity:explanationofthelargevariabilityinevokedcorticalresponses.Science273,1868‐1871(1996).

Bertchinger,N.andNatschläger,T.Real‐timecomputationattheedgeofchaosinrecurrentneuralnetworks.NeuralComput.16,1413‐1436(2004).

Broome,B.M.,Jayaraman,V.andLaurent,G.Encodinganddecodingofoverlappingodorsequences.Neuron51,467‐482(2006).

Churchland,M.M.,Yu,B.M.,Cunningham,J.P.,Sugrue,L.P.,Cohen,M.R.,Corrado,G.S.,Newsome,W.T.,Clark,A.M.,Hosseini,P.Scott,B.B.Bradley,D.C.,Smith,M.A.,Kohn,A.,Movshon,J.A.,Armstrong,K.M.,Moore,T.,Chang,S.W.,Snyder,L.H.,Ryu,S.I.,Santhanam,G.,Sahani,M.andShenoy,K.V.Stimulusonsetquenchesneuralvariability:awidespreadcorticalphenomenon.(2009,submitted).

Churchland,M.M.,Yu,B.M.,Ryu,S.I.,Santhanam,G.andShenoy,K.V.Neuralvariabilityinpremotorcortexprovidesasignatureofmotorpreparation.J.Neurosci26,3697‐3712(2006).

Finn,I.M.,Priebe,N.J.andFerster,D.Theemergenceofcontrast‐invariantorientationtuninginsimplecellsofcatvisualcortex.Neuron54,137‐152(2007).

Fiser,J.,Chiu,C.andWeliky,M.Smallmodulationofongoingcorticaldynamicsbysensoryinputduringnaturalvision.Nature431,573‐578(2004).

Fortier,P.A.Smith,A.M.andKalaska,J.F.Comparisonofcerebellarandmotorcortexactivityduringreaching:directionaltuningandresponsevariability.J.Neurophysiol.69,1136‐1149(1993).

Friedrich,R.W.andLaurent,G.Dynamicsofolfactorybulbinputandoutputactivityduringodorstimulationinzebrafish.J.Neurophysiol.91,2658‐2669(2004).

Holmgren,C.,Harkany,T.,Svennenfors,B.andZilberter,Y.Pyramidalcellcommunicationwithinlocalnetworksinlayer2/3orratneocortex.J.Physiol.551,139‐153(2003).

Hubel,D.H.andWiesel,T.N.Receptivefields,binocularinteractionandfunctionalarchitectureinthecat’svisualcortex.J.Physiol.160,106‐154(1962).

Page 15: Interactions between Intrinsic and StimulusEvoked ...

15

Ipsen,I.C.F.andMeyer,C.D.Theanglebetweencomplementarysubspaces.Amer.Math.Monthly102,904‐911(1995).

Kenet,T.,Bibitchkov,D.,Tsodyks,M.,Grinvald,A.andArieli,A.Spontaneouslyemergingcorticalrepresentationsofvisualattributes.Nature425,954‐956(2003).

MacLean,J.N.,Watson,B.O.,Aaron,G.B.andYuste,R.Internaldynamicsdeterminethecorticalresponsetothalamicstimulation.Neuron48,811‐823(2005).

Mast,J.andVictor,J.D.Fluctuationsofsteady‐stateVEPs:interactionofdrivenevokedpotentialsandtheEEG.Electroenceph.andClinic.Neurophysiol.78,389‐401(1991).

Mitchell,J.F.,Sundberg,K.A.andReynolds,J.J.Differentialattention‐dependentresponsemodulationacrosscellclassesinmacaquevisualareaV4.Neuron55,131‐41(2007).

Molgedey,L.,Schuchhardt,J.andSchuster,H.G.Suppressingchaosinneuralnetworksbynoise.Phys.Rev.Lett.69,3717‐3719(1992).

Petersen,C.C.,Grinvald,A.andSakmann,B.Spatiotemporaldynamicsofsensoryresponsesinlayer2/3ofratbarrelcortexmeasuredinvivobyvoltage‐sensitivedyeimagingcombinedwithwhole‐cellvoltagerecordingsandneuronreconstructions.J.Neurosci.23,1298‐1309(2003).

Petersen,C.C.,Hahn,T.T.,Mehta,M.,Grinvald,A.andSakmann,B.Interactionofsensoryresponseswithspontaneousdepolarizationinlayer2/3barrelcortex.Proc.Natl.Acad.Sci.USA100,13638‐13643(2003).

Rajan,K.andAbbott,L.F.EigenvalueSpectraofRandomMatricesforNeuralNetworks.Phys.Rev.Lett.97,188104(2006).

Rajan,K.,Abbott,L.F.andSompolinsky,H.Frequency‐DependentSuppressionofChaosinRecurrentNeuralNetworks.(2009,submitted).

Shriki,O.,Hansel,D.andSompolinsky,H.Ratemodelsforconductance‐basedcorticalneuronalnetworks.NeuralComput.15,1809‐1841(2003).

Sompolinsky,H.,Crisanti,A.andSommers,H.J.ChaosinRandomNeuralNetworks.Phys.Rev.Lett.61,259‐262(1988).

Song,S.,Sjöström,P.J.,Reigl,M.,Nelson,S.B.andChklovskii,D.B.Highlynon‐randomfeaturesofsynapticconnectivityinlocalcorticalcircuits.PLoSBiol.3,e68(2005).

vanVreeswijk,C.andSompolinsky,H.Chaosinneuronalnetworkswithbalancedexcitatoryandinhibitoryactivity.Science24,1724‐1726(1996).

vanVreeswijk,C.andSompolinsky,H.Chaoticbalancedstateinamodelofcorticalcircuits.NeuralComput.10,1321‐1371(1998).

Vincent,J.L.,Patel,G.H.,Fox,M.D.,Snyder,A.Z.,Baker,J.T.,VanEssen,D.C.,Zempel,J.M.,Snyder,L.H.,Corbetta,M.andRaichle,M.E.Intrinsicfunctionalarchitectureintheanaesthetizedmonkeybrain.Nature447,83‐86(2007).

Werner,G.andMountcastle,V.B.TheVariabilityOfCentralNeuralActivityInASensorySystem,AndItsImplicationsForTheCentralReflectionOfSensoryEvents.J.Neurophysiol.26,958‐977(1963).

Wong,K.‐F.andWang,X.‐J.ARecurrentNetworkMechanismofTimeIntegrationinPerceptualDecisions.J.Neurosci.26,1314‐1328(2006).

Page 16: Interactions between Intrinsic and StimulusEvoked ...

16

White,B.L.andFiser,J.Therelationshipbetweenawakeandanesthetizedneuralresponsesintheprimaryvisualcortexoftherat.Soc.Neurosci.Abstracts38:459.12(2008).

Yuste,R.,MacLean,J.N.,Smith,J.andLansner,A.Thecortexasacentralpatterngenerator.Nat.Rev.Neurosci.6,477‐483(2005).