Opinion
Oscillations and Episodic
Memory: Addressing the
Synchronization/
Desynchronization Conundrum
Simon Hanslmayr,
1,* Bernhard P. Staresina,
1and
Howard Bowman
1,2Brain oscillations are one of the core mechanisms underlying episodic memory. However, while some studies highlight the role of synchronized oscillatory activity, others highlight the role of desynchronized activity. We here describe a framework to resolve this conundrum and integrate these two opposing
oscillatory behaviors. Specifically, we argue that the synchronization and
desynchronization reflect a division of labor between a hippocampal and a neocortical system, respectively. We describe a novel oscillatory framework that integrates synchronization and desynchronization mechanisms to explain how the two systems interact in the service of episodic memory.
The Synchronization, Desynchronization Conundrum
Brain oscillations (seeGlossary) provide temporal windows for neuralfiring and shape synaptic plasticity by synchronizing and desynchronizing neural assemblies. Oscillations therefore have a particularly high potential to shed light on the mechanisms underlying episodic memory [1,2]. However, over the past few years a conundrum has emerged regarding how brain oscillations relate to memory [3]. Whereas some studies highlight the role of synchronized activity, mainly in the theta (3–8 Hz) and gamma (40–80 Hz) frequencies[1], others highlight the role of desynchronized activity, mainly in the lower-frequency ranges (<20 Hz[4]). This conundrum exists at both an empirical and a theoretical level. Based on Hebb's seminal idea– neurons thatfire together, wire together – a strong case can be made that synchrony is required for memory formation[1,2]. However, mathematical information theory, alongside physiological studies in animals, postulates that high synchrony, especially in the lower-frequency ranges (<20 Hz), reduces information coding [4–6]. We here aim to integrate these two seemingly incompatible concepts and present a mechanistic framework to resolve this conundrum. Building on complementary learning systems [7], we assume a division of labor between a hippocampal system, which mainly binds information, and a neocortical system, which mainly represents the content of this information. Based on recentfindings, we argue that this division of labor is visible in the two opposing oscillatory behaviors, with the hippocampal system showing synchronization in the theta/gamma range[8–10], mediating binding [11], and the neocortex showing low-frequency desynchronization[4], mediating the representation of infor-mation[12,13].
Hippocampal Synchronization in the Service of Episodic Memory
The critical role of the hippocampus for intact episodic memory has beenfirmly established by neuropsychological studies, animal models, computational models, as well as human
Trends
Data from rodent as well as human studies suggest that theta/gamma syn-chronization in the hippocampus (i.e., theta phase to gamma power cross-frequency coupling) mediates the bind-ing of different elements in episodic memory.
In vivo and in vitro animal studies sug-gest that theta provides selective time windows for fast-acting synaptic mod-ifications and recent computational models have implemented these mechanisms to explain human mem-ory formation and retrieval.
Recent data from human experiments suggest that low-frequency power decreases in the neocortex, most evi-dent in the alpha/beta frequency range, mediate encoding and reinstatement of episodic memories.
The content of reinstated memories can be decoded from cortical low-frequency patterns.
1
University of Birmingham, School of Psychology, Birmingham, UK
2
University of Kent, School of Computing, Canterbury, UK
*Correspondence:
Glossary
ComplementaryLearning Systems(CLS):theCLSapproach wasconceivedasaresponsetothe plasticity–stabilitydilemma,a fundamentaltrade-offthatlearning systemshavetodealwith.CLS proposesameanstoobtaina relativelystablelearningsystem withoutforfeitingplasticityanddoes thisbypositingtwosubsystems,the neocortexandhippocampus,with complementaryrepresentational characteristics.Theneocortical systemslowlylearnshighly ‘overlapping’representations,with unitsbecomingbroadlytunedand therebycontributingtomanydifferent storedpatterns.Asaconsequence, thecorticalsystemishighly susceptibletocatastrophicforgetting. Thehippocampalsystem,by contrast,isnotlimitedinthisway. Thisisachievedbyitexhibitinga sparsecodeinwhichthe representationforeachnewpattern haslittleoverlapwithpriorlearned patterns.Thispattern-separated character,whichisparticularly evidentinthedentategyrusand CA3,ensuresthatitisnotsubjectto catastrophicforgettinginthesame wayastheneocortex.Furthermore, thislevelofoverlapofunits representingnewandolditems meansthatthissystemcanlearn quicklywithlittlecost.Thisenables theCLSsystemtoexhibit‘one-shot’ learning,aswehumansdo.See
[7,66]forarecentreview. Desynchronization:adecreasein theneuralbehaviorcausing synchronization(seeSynchronization). Information:averybroadtermwith numerousdifferentdefinitions.Rather thangivingafurthertheoretical definitionofinformation,wehererefer toinformationinanoperationalway. Thatis,informationis,forus,the degreetowhichwecanmake inferencesabouttheidentityofa particularperceivedorretrieved memoryfromaparticularobserved temporalorspatialneuralpattern. Noisecorrelations:aplethoraof single-unitrecordingsinmonkeyshas shownthatevenunderthemost controlledconditionsfiringratevaries significantlyfromtrialtotrialwhen identicalstimuliarepresented. Synchronizedlow-frequency fluctuations(<20Hz)infiringrate, whichareindependentofthe stimulus,areassumedtobethemain
neuroimaging[14].Notably,hippocampalengagementseemsparticularlyimportantforthekey characteristicofepisodicmemories;thatis,bindingdiscontiguous(i.e.,separate)eventdetails intorichassociativememorytraces[15].Howisthisfeataccomplishedmechanistically?Gamma oscillations,operatingattimescalesbelow30ms,haverecentlybeenproposedastheprime candidatetofacilitatelearning-relatedsynapticchangessuchasspike-timing-dependent plas-ticity (STDP)/long-term potentiation (LTP) [1]. Does the hippocampus recruit such gamma oscillationsintheserviceofepisodicmemoryformationandretrieval?
Directhippocampalrecordingsfromhumanepilepsypatientshavefoundthatgammapower increasescorrelatewithbothsuccessfulmemoryencoding[16]andretrieval[17].Onecaveat relatedtotheseresultsisthat,oncloserexamination,thegammaeffectswerefoundtobepartof more broadband power increases, ranging from 30Hz to 100Hz [18,19]. The functional significanceofsuchbroadband powerincreasesremainsunderdebate,butconsensushas emergedthattheyreflectmulti-unitspikingactivityratherthantrueoscillatoryactivity([20,21], butsee[22]).However,recentrecordingsfromtheprimatehippocampushavesucceededin showingnarrow-bandgammapowerincreasesaswellasspike-fieldcoherencebetweensingle unitsandgamma-bandLFPsduringsuccessfullearning[23].
Perhaps even more compelling evidence for hippocampal gamma oscillations supporting episodicmemoryprocessescomesfromstudiesexaminingshort-rangeinter-regional synchro-nization, particularly between the hippocampus and the adjacententorhinal cortex (EC). In humanepilepsypatients,gamma-bandcoherencehasbeenfoundtoincreasebetweenthese regionsasafunctionofsuccessfulepisodicencoding[24]aswellasretrieval[25](Figure1B). Recordings from rodent models have gone further to identify a frequency-specific ‘routing function’ofgammaoscillations,suchthattheCA1subregionisfunctionallycoupledwithCA3 versustheECindifferentgammasub-bands[26].Establishingtherelevanceofthiscouplingfor learningprocesses,arecentstudyhaslinkedthegamma-bandcoherencebetweenCA1and theECtotheencodingandretrievalofodor/spaceassociations[27].
Insummary,convergingevidencehasaccumulatedacrossdifferentspeciesandacrossdifferent experimentalparadigmsforacriticalroleofsynchronizedhippocampalgammaoscillationsinthe serviceofepisodicmemory.Thepressingquestionisthenhowtheseoscillationsarethemselves regulated.Ignitedbythehallmarkobservationofthetaphaseprecessionofhippocampalplace cells[28],muchresearchhasbeendedicatedtounderstandingtheroleofthethetarhythmin orchestratingmemory-relatedneuralsignals[29].Astrikingexampleisthefinding–observedvia directrecordingsfromthehumanhippocampus–thatsuccessfulepisodicencodingmaynot necessarilyrelyonthenetincreaseinhippocampalfiringratesbutonthetemporalprecisionof single-unitfiringwithrespecttotheconcurrentthetaphase[8].
Importantly,recentresearchhasbeguntoinvestigatenotonlyhowthethetarhythmmayclock single-unitfiringbutalsohowoscillatorypatternsinthegammarangemayberelatedtothe ongoingthetaphase[9].Forinstance,theabovementionedcouplingbetweenCA1andCA3 versustheEC was notonly expressed indifferent gammasub-bands butalsooccurred at differentphases ofthethetacycle[26].Moreover(andnotwithstandingtheabovementioned questionofwhethergammapowerincreasesreflecttrueoscillatorychangesversusmulti-unit spiking),evidencehasaccumulatedinbothrodentmodelsandhumanintracranialrecordings thathippocampaltheta–gammacouplingislinkedtosuccessfulepisodicmemoryprocesses [30].Ofparticularimportanceisarecentmagnetoencephalography(MEG)studythatshowed thatitem–contextbinding,ahallmarkofepisodicmemory,criticallydependsontheta–gamma couplinginthemedialtemporallobe(MTL),andspecificallyonwhethergammaoscillationsare coupledtothepeakorthetroughofathetacycle(Figure1A)[11,30,31].Notably,thefrequency rangeshowingthestrongestmodulatoryeffectsisslightlylowerinhumans(3–4Hz)thanin
rodents (8Hz), in line with the notion that functionally homologous oscillations occur at increasinglyslowerfrequenciesasbrainsizeincreasesacrossspecies[32,33].
Takingthese findings together, the mnemonicfunctions ofthe hippocampus appear to be intimatelylinkedtoneuronalsynchronizationinthegammafrequencyband,witharegulatory influenceofthephaseofongoingthetaoscillations(Figure1A,B).
CorticalDesynchronizationintheServiceofEpisodicMemory
Oscillatory powerdecreases during the formation ofmemories (i.e., encoding) are typically observedinthelower-frequencyranges(<20Hz),especiallyduringencodingofitemsthatare laterrememberedcomparedwith laternotremembered [34,16,35–37].Forinstance,during encodingofverbalmaterial,alpha/betapowerdecreases(12–18Hz)aremostevidentinthe leftinferiorprefrontalcortex[16,37,38].Importantly,suchleftfrontalbetapowerdecreasesare notasimple(incidental)byproductofmemoryformationbutarecausallyrelevant,as demon-stratedinarecent combinedelectroencephalography (EEG)–rhythmictranscranial magnetic stimulation(rTMS)study[39](Figure1C).Giventhattheleftinferiorprefrontalcortexisstrongly involvedinsemanticprocessingingeneral,theseresultsfitwithseveralfindingsinthelanguage domainshowingthatleftprefrontalbetapowerdecreasesduringsemanticprocessing[40,41]. Memoryformationfornonverbalmaterial(i.e.,images)isalsoaccompaniedbypowerdecreases;
drivingforceofthisvariability,termed noisecorrelations[49].Onewayin whichattentionenhancestheneural signal-to-noiseratioisbyreducing (i.e.,desynchronizing)these low-frequencyfluctuations[6]. Oscillations:brainoscillationsare typicallydividedintodifferent frequencybandsandreferredtoby lettersoftheGreekalphabet, includingbutnotlimitedtodelta (1–4Hz),theta(4–8Hz),alpha (8–12Hz),beta(15–25Hz),and gamma(40–80Hz).Theexact boundariesbetweenfrequenciesare difficulttodefineandvary considerablybetweenexperiments andspecies.
Synchronization:therearemany waystodefineneuralsynchronization andevenmoremethodstomeasure it.Inthisopinionarticle,wereview studiesthatfocusedoneitherofthe followingthreetypesof
synchronization:(i)thelevelof synchronizationwithinalocalcell assembly,whichcanbemeasured viathepowerofanoscillatorysignal onEEGorMEG[67];(ii)interareal synchronization,whichisthe synchronizationbetweentwobrain regions(e.g.,thehippocampusand therhinalcortex)[24];and(iii)phase amplitudecross-frequencycoupling, whichreferstothesynchronization betweentwofrequencieswhereby thephaseofalowerfrequency modulatesthepowerofahigher frequency(e.g.,thecouplingbetween thephaseofthetaandthepowerof gamma[11,30,31]).
(A) (B)
(C) (D)
Item-context binding No item-context binding
Phase (rad.) Phase (rad.)
Theta/Gamma synchrony during EM encoding Gamma synchrony during EM retrieval
Beta desynchronisaon during EM encoding Alpha/beta desynchronisaon during EM retrieval
0 0.5 1.0 1.5 2.0 0.0 –0.4 0.4 0.8 1.2 Time (sec)
coh. change from baseline
iEEG coherence (35-40 Hz) between hippocampus and Rhinal cortex (source memory vs. item recognion)
Source Item
Power change (%) from baseline 0.0
10 20
Alpha/Beta power (10-25 Hz, 100-200 ms) Item encoded in the le VF –
Item encoded in the right VF
Le electrodes Right electrodes Le VF Right VF – pi 0 pi – pi 0 pi
Gamma power Gamma power
3.9 4.1 4.2 4.0 3.9 4.1 4.2 4.0 x 10 x 10 1 0 –1 1 0 –1 ∗ ∗ 10-25 Hz 100-200 ms Memory performance 18.7 Hz 10.7 Hz 6.8 Hz Sham 50 45 40 35 0 % Recalled Key: Key:
Figure1.StudiesShowingHippocampalTheta/GammaSynchronizationandAlpha/BetaDesynchronization
DuringEncodingandRetrievalofEpisodicMemories(EMs).(A)Gammapowercoupledtodifferentphasesoftheta
predictswhetheritem–contextbindingoccursordoesnotoccur,astestedinasubsequentmemoryparadigmwherethe contextualoverlapbetweenencodingandretrievalwasdirectlymanipulated[11].(B)Resultsfromanintracranial electro-encephalography(EEG)studyshowinggammaphasesynchronizationbetweenthehippocampusandrhinalcortexduring successfulassociativerecognition(i.e.,sourcememory)comparedwithsimpleitemrecognition.(C)Preventingbeta desynchronization(i.e.,18.7-Hzstimulation)attheleftinferiorfrontalcortexviarhythmictranscranialmagneticstimulation (rTMS)selectivelyimpairsmemoryencoding.(D)Alpha/betapowerdecreasesduringretrievalindicatethevisualfield(VF) where a stimuluswasinitially encoded (i.e., alpha/beta power decreasesindicate memory reactivation). (A,B,C,D) reproducedandmodifiedwithpermissionfrom[11,25,39,47],respectively.
however,thesedecreasesapparentlyoccurmoreinthealpharangeatparieto-occipitalregions [34]. Clearly, more studies are needed to investigate alpha/beta power decreases during memoryformationofdifferentmaterials(includingwords,images,andsounds),butthepattern ofresultssofarsuggeststhatalpha/betadesynchronizationindexesinformationprocessingin specializedcorticalmodulesduringtheperceptionofanevent[42,43]andthereforepredictits likelihood of being later remembered. Of note, this negative relationship between power decreasesandmemoryformationmightalsoextendtothethetafrequencyrange,asindexed byseveralrecentstudiesshowingthatthetapowerdecreasescorrelatewithmemoryformation [16,44].
Asimilarpicturearisesforpowerdecreasesduringmemoryretrieval.Forinstance,the topog-raphyofalpha/betapowerdecreasesvarieswiththetypeofretrievedmaterial(i.e.,wordsversus faces[45]orlocationsversusobjects[46]),suggestingthatpowerdecreasesindicate material-specificmemoryreactivation.Directsupportforthiscomesfromstudiesthatpresentedobjects duringencodingtoeithertheleftorrightvisualhemifieldandshowedthatmemoryretrievalof centrallypresentedobjectsisindeedreflectedbyalpha/betapowerdecreasescontralateralto thesiteatwhichtheitemwasencoded[13,47](Figure1D).Usingmultivariatepatternanalysisof time–frequency data, two MEG/intracranial EEG (iEEG) studies went one step further and demonstratedthatthecontentofreactivatedmaterialcanbereliablydecodedfromalpha/beta frequencies[12,48].Forinstance,applyingatemporal-pattern-analysisapproach,Staudigletal. [12]demonstratedthatthereactivationofindividualdynamiccontexts(i.e.,movieclips)canbe decoded from the temporal pattern of beta phase in material-specific cortical areas (i.e., parahippocampalarea,visualcortex;[12]).Together,theseresultsofferaninterestingpossibility forthemechanistic roleoflow-frequency powerdecreases:decreases inoscillatory activity enableaneuralassemblytoexpressastimulus-specificcodebyallowingamorecomplex(i.e., information-rich)temporalphasetrajectory.Arguably,suchacodingmechanismwouldnotwork veryefficientlyinsituationsofhighsynchrony(i.e.,astationarysignal),wherelargepopulationsof neuronsareentrainedtothesamerhythm[4].
Takingthesefindingstogether,wepostulatethatlow-frequencypowerdecreasesreflecttheactive engagementofcorticalmodulesduringencodingandretrievalofmemories.Anopenquestionis whatalow-frequencypowerdecreasemeansmechanistically(i.e.,attheneurallevel)andhowit relatestothetheta/gamma dynamicsinthehippocampalsystem.Althoughmoreresearchis neededtoanswerthiscomplexquestion,aconvergentpicturecanbederivedfromthreedifferent recentframeworks.(i)Followingthealphainhibitionframework,low-frequencypowerdecreases couldactasagatingmechanismwherebydecreasingalphapowerincreasesfiringrates[42,43]. Notably,thisnegativerelationshipbetweenoscillatoryactivityandneuralfiringmightalsoextendto higher (beta) and lower (theta) frequencies [9]. (ii) A second line of research suggests that synchronizedfiringinlowfrequenciesisamaincontributortothetrial-by-trialvarianceinneural firing(termed‘noisecorrelations’),therebyreducingthereliabilitywithwhichaneuralcodeis expressedinapopulationofneurons[6,49].Decreasingtheamplitudeoftheselow-frequency oscillationswouldreducesuchneuralnoisecorrelationsandthusincreasethereliabilityofaneural codethatis conveyedto downstreamneurons (i.e.,inthe hippocampus). (iii)Finally, power decreasesthemselvescouldbeamechanismtode-correlateneuralactivitytoenhancetheneural codingcapacityperse.Powerdecreasesmighttherebyallowflexiblephaseadjustmentsina neuralpopulationtoformatemporalcoderepresentingtheidentityofaspecificstimulus[4,12,50]. The common notion between all of these different frameworks, however, is that power decreasesenableaneuralassemblytoexpressaneuralcodeinsomeform;thatis, viaan increasein firing rate, a reductionin noise, or phase encoding. Anysuch code would be meaninglessifit werenot interpretedbya ‘reader’[2].Here weassumethatthis readeris thehippocampus(Figure2).
Reconciliation:ASynchronized Hippocampusanda DesynchronizedCortex
TheComplementaryLearningSystems(CLS)theoryhaschampionedthepositionthatthe neocortexand hippocampus provide differentbut complementary representational formats, withtheneocortexsupportingarichintegratedrepresentationlearnedovermanyexperiences andthehippocampusprovidingasparserepresentationlearnedin‘oneshot’(oratleastfew shots).Thetheoryalsoproposesthatthehippocampus,sittingattheendoftheprocessing pathway,providesonlinelearning ofconjunctive representations,whichbindtheconstituent elementsofanepisodicmemory.Thisisimplementedbystrengtheningthesynapsesbetween contributingneuronsthroughLTP.
WearguethattheCLStheoryprovidesaframeworkwithinwhichthedisparityinsynchronyof oscillationsbetweencortexandhippocampuscanbereconciled.Statedexplicitly,wepropose thatthecomplementarityofcomponentsubsystemsisreflectedinthedisparityofoscillatory dynamicsthatthesesubsystemsexhibit.Ourproposalismostclearlyunderstoodby consider-ingthesequenceofstepsthatweenvisagerealizeencodingofepisodicmemories(illustratedin Figure2).(i)Sensorystimulation(e.g.,seeingapersoninaparticularplace)inducesareductionin alpha/betaoscillationstoallowneocorticalunitstoencodethecontentofthatstimulation[42,43] bypushing their firing rates upsignificantly.Whether alpha/betapower reductions drivean increaseinneuralfiringorwhetheranincreaseinneuralfiringdrivesareductioninalpha/beta powerisanopenquestion;however,thenetresultisinbothcasesanobservabledecreaseof alpha/betapowertogetherwithincreasedanddesynchronizedfiringofbroadlytunedneurons. (ii)This stimulation-specificfiringrateincreaseintheneocortexinturn drivescorresponding stimulation-specifichippocampal unitsto move their firing forwardinthe phaseoftheta. In thecasewherehippocampal unitsarenot receivingstimulus-specificdrive fromneocortical units,theywouldfireintheexcitatoryphaseoftheta;thatis,inthetroughofthetaasrecordedin
No Input Hipp. NC Alpha (10 Hz) Theta (4 Hz) LTP LTD LTD LTP Input Hipp. NC Alpha (10 Hz) Theta (4 Hz) LTP LTD LTD LTP No memory formaon Memory formaon (A) (B) + - + - LFP LFP + LFP +
Figure2.ASchematicoftheCurrentFramework.(A)Intheabsenceofexternalinput,highalphapowerisevidentin
theneocortex(NC),withNCneuronsfiringatalowrate.Inthisstate,theNCneuronshavelittleimpactontheirdownstream neuronsinthehippocampus.long-termpotentiation(LTP)/long-termdepression(LTD)isregulatedbythethetaphase, whereby,intheabsenceofinput,thesynapticconnectionsbecomeweakened(LTD).(B)Inputreducesthe(effective)firing thresholdforstimulation-specificpopulationsintheNC,which,inthiscase,respondtotwostimuli.Onthepopulationlevel, neuronalfiringincreases,whichleadstoadecreaseinsynchronyandalphapower.TheincreasedfiringinthetwoNC populationsof neuronswouldinturnreducethe(effective)firingthreshold oftheirhippocampalstimulation-specific downstreamneurons,causingphaseprecession(indicatedbyarrows)andhenceLTP,whichthenformsanassociation betweenthetwostimuli.Notethatthephaseofthetachangesdependingonwhichhippocampalsubregionitisrecorded from[29],being180o
phaseshiftedbetweenthehippocampalfissureandCA1/CA3.Wehereplotthetaaswouldbe recordedinCA1(andCA3),inkeepingwithstudiesthatshowedLTPtooccuratthepeak(+)andLTDtooccuratthetrough ( )oftheta[53,55].Moreover,aspresentedhere,thepeakofthetaisinfactthefunctionallyinhibitoryphase,whileitstrough isthefunctionallyexcitatoryphase.
CA1/CA3[51](Figure2A).Inthecaseofstimulation(i.e.,encoding),thoseunitsthatrespondto thepresentedstimuliwouldadvancetheirspikingwithrespecttotheongoingthetaoscillation, showingapatternthatissimilartothewell-documentedphenomenonofphaseprecession[28]. Thatis,theincreasedexternaldrivingexcitationenablesthestimulation-specifichippocampal unitstohittheirfiringthresholdearlierinthethetaphase;thatis,closertotheinhibitoryphaseof theta(i.e.,thepeakinCA1/CA3)–atimepointwheremosthippocampalunitsaresilent[51,52]. (iii)Thisadvancementinitiatesatemporalsegregationinwhichprecessedunitscometofireinan earlierandseparategammacycleepisode.Thisagaintakesinspirationfromphaseprecession, wherebyprecessionspecificallyadvancesthegammacycleofavolleyofspikesinthephaseof theta.Importantly,weproposethatthesegregationofstimulation-sensitivefromnon-sensitive unitsenablesLTPtobeselectivelyapplied.Thereby,theearlierphaseofthetathat stimulation-specifichippocampalunitsaredriventofireinisspecificallyoneinwhichLTPoccurs.SeeBox1 forpossiblemechanismsthatcouldunderliethisincreaseinLTP.(iv)Thisincreasedsegregation intogammacycleswouldexplainhowthecouplingofthetaphaseandgammapowerwould correlatewithmemoryformationanditem–contextbindinginthehippocampus[11,30,31].(v)As aresultofthisselectiveLTP,stimulation-specificunitswouldbecomestronglyinterconnectedin thehippocampus,layingdownamemoryoftheepisode.FollowingCLStheory,these asso-ciativeconnectionswouldfacilitateretrievalthroughcompletionofapartialstimulationpattern. Importantly,thesparse,pattern-separatednatureofthehippocampalrepresentationsensures thatonlyasmallproportionoftheentirepopulationofunitswouldprecessforwardinthetheta oscillation,whichinturnensuresthatlearningisselective.
Theobservationcentraltotheconundrumweseektoresolve,thatgammasynchronizestothetain the hippocampus during episodic memoryencoding, is critical to the phase advancement argumentjustmade.Thatis,thespecificpresenceofstronggamma-to-thetasynchronyprovides themechanismbywhichfiringofdrivenunitsstepsoutandforwardinthethetacycle,enabling relevantsynapsestobeselectivelysubjecttoLTP.Inthissense,onecouldarguethat gamma-to-thetasynchronizationiskeytothedistinctivelearningcapacityofthehippocampuswhilealpha/ betadesychronizationiskeytotheneocortex'scapacitytorepresentinformation.
Box1.MechanismsforTheta-Phase-DependentSynapticModifications
WeenvisagethreepossiblewaysinwhichadvancementoffiringinthephaseofthetacouldinduceincreasedLTP. ThereisconsiderableevidencethatthestrengthofLTPinthehippocampus(orindeedwhetherthereisLTPorLTD) varieswiththephaseoftheta[53–55].Invivo[55]aswellasinvitro[53]studiesshowedthatstimulationduringthe positivepeakofthetafavorsLTP.Notably,thispositivethetapeakinCA1iseffectivelythehyperpolarizingphase(i.e., inhibitoryformosthippocampalpyramidalcells[29,51]).Together,theseresultsfitwithourmodel,wherecorticallydriven neuronswouldbepushedtofireclosertotheinhibitorythetaphase(Figure2).Accordingly,advancingthefiringofa subsetofunitscouldsimplyplacethoseunitsinaphaseofthetawithstrongLTP.
TheoriesofSTDP[68]holdthattheorderofspikingdeterminesthedirectionofsynapticchange.Inparticular,asynapse increasesinstrengthifthepresynapticunitspikesbeforethepostsynapticunit.Thisprovidescausalitytolearning;thatis, earlierspikefiringsuggeststhatthepresynapticunitiscontributingtodrivingthepostsynapticunittofire,indicatingthat thesynapseshouldincrease.However,forsuchlearningtocausesustainedLTPofasynapse,aconsistentpre–post orderoffiringneedstoobtain.Thus,itcouldbethatadvancingthefiringofstimulation-specificunitsinthehippocampus introducesasystematicordertothatfiringonwhichSTDPcanact.
ThereareformsofHebbian learningthatincorporateanormalizationfortheamountofweightchangeforeach postsynapticunit(e.g.,[56],seepp.289,290,and291).Asaresult,aboundedformofHebbianlearningisobtained, accordingtowhichweightscannotbecomearbitrarilylarge.Aconsequenceofthenormalization,however,isthat competitionisgeneratedbetweenweightchangesonlinksenteringapostsynapticunit,withchangesbeingbiggerif feweroftheseweightsaresubjecttochange.Thiswouldoccurwhenfewerpresynapticunitsareactiveandwould generategreaterLTPforsynapsesthatchangewhenfiringisselective(i.e.,duringaninhibitoryphase).Thisisexactly whattheadvancementofspikingforstimulation-sensitiveunitstowardtheinhibitorythetaphase,asweareproposing, woulddo:itwouldrestrictthesetofpresynapticunitsthatarefiring,ensuringgreaterLTPontheirspecificsynapses. Similarly,theamountofLTDforstimulation-unspecificunitswouldberathersmall,evennegligible.
CaveatsandLinks toOtherMemoryModels/Theories
Oneseeminglycounterintuitiveconsequenceofourframework,andoftheanimalworkitbuilds on[53–55],isthatbothLTPandlong-termdepression(LTD)(i.e.,learningandforgetting)occur sequentiallyinanalternatingmannerdependingonthetaphase.Thiswouldbeaproblemforour frameworkiftheamountofthepositiveweightchangeatanysinglesynapseduringLTPwere equaltotheamountofthenegativeweightchangeatasynapseduringLTD.However,thisis unlikelytobethecase.Inourframework,LTDoccursforsynapsesof‘irrelevant’units;thatis, unitsthatdonotcurrentlyreceiveinputfromtheircorticalcounterpartsandthereforefireduring thethetatrough(asrecordedinCA1/CA3).Notably, mosthippocampal pyramidalcellsare activeduringthattime[51];consequently,theactualweightchangethatisappliedtotheseunits canbeassumedtobefairlysmall(Box1)[56].Thus,ourframeworkpredictsaweakamountof forgettingovertime,especiallyforunitsthatarenotreactivated,whichresonateswellwiththe factthatmemoriesfadeovertime.Moreover,thisassumptionisalsoinlinewithaplethoraof findingsshowingthatreactivationofmemoriesduringwakefulness[57]andsleep[58] counter-acts this forgetting. This aspect of our framework also resonates well with recent ideas highlightingtheroleofforgettingasahighlyorganizedprocessthatkeepsourmemorysystem flexible,goaloriented,andorganized[59,60].
Inthepresentedframework,wefocusedontheinteractionbetweentwoofthemostubiquitous oscillatorydynamicsinthebrain,thetaoscillationsinthehippocampusandalpha/beta oscil-lationsin the neocortex. Therefore we did not include a detaileddescription of how theta oscillationsregulateencodingandretrievaldynamicswithinthedifferenthippocampalsubfields.
Box2.PredictionsoftheSynchronization/DesynchronizationFramework(SDF)
Thehere-presentedSDFmakesseveralpredictions,whicharelistedbelow.
Thelevelof corticaldesynchronization should correlatepositively withcorticalinformation coding; thatis,more desynchronizationshouldleadtostrongerfidelityofneuralrepresentationsasdetectedwithmultivariatepattern-analysis methods.
Duringencoding,phaseprecessioninthehippocampusshouldcorrelatewithmemoryformation;thatis,itemsthatare laterrememberedshouldshowstrongerphaseprecessionthanitemsthatarelaternotremembered.Thethetaphase shouldalsodiscriminatebetween stimulus-selectiveandstimulus-unselective neuronsduringboth encoding and retrieval.
Desynchronizationinthecortexhasadownstreameffectonthestimulus-selectiveneuronsinthehippocampusthat shouldexpressitselfin:(i)phaseprecessionofneuronsinthehippocampus;(ii)highertheta/gammacross-frequency coupling;and(iii)highermetabolicdemands[blood-oxygen-leveldependent(BOLD)]signalinthehippocampus. Arelativelysmallproportionofhippocampalneuronsshouldprecessforwardduringencoding,reflectingtherelatively sparsenatureofhippocampalrepresentations.
Thesepredictionscanbetestedinmultiplewaysusingvariousapproachesandrecordingtechniques.Onatheoretical level,thepredictionscanbetestedincomputationalmodelsthatcombinebasicideasofCLStheorywithoscillationsand neuralfiringintheneocortexandhippocampus(see,forexample,[52]).Implementingtheproposedsynchronization/ desynchronizationmechanismsinsuchmodelswillshowwhattypeofbehaviorcanbemodeledand,moreimportantly, whattypeofbehaviorcannotbemodeled.Thereby,computationalmodelingwouldrevealthelimitsofourframework.On anempiricalbasis,thesepredictionscanbetestedusinginvasiverecordingmethodscombininglocalfieldpotential(LFP) withsingle-unitrecordingsinanimalmodels(i.e.,rodentsandmacaques)andhumans(i.e.,epilepsypatientsundergoing presurgicalmonitoring).Asthecoreideaofourframeworkreliesonaninteractionbetweenthehippocampusand neocortex,itiskeytonotonlyrecordinthehippocampusortheneocortexbuttorecordhippocampalandneocortical activitysimultaneously.Finally,someofthepredictionscanalsobetestedusingnoninvasiverecordingmethodssuchas EEG/MEGorEEG–fMRI.Recentdevelopmentsinsourcelocalizationhavemadeitpossibletorecord electrophysio-logicalsignalsfromtheMTL[11].AlthoughitmightbedifficultforEEG/MEGstudiestodisentanglehippocampalfrom otheractivitywithintheMTL(e.g.,parahippocampalgyrus,EC),oscillatoryinteractionsbetweentheMTLandthe neocortexinprinciplecanbestudied.Additionally,multivariateanalysisapproachesusingtemporalandspatialpattern analysis can be utilized to test for the representation of information, which can then be linked to cortical desynchronization.
Thesemechanismsaredescribedingreatdetailinothertheoreticalpapers(e.g.,[29])andhave beenimplementedinrecentcomputationalmodels[52].Notably,althoughtheexact mecha-nismsofhowLTP/LTDandthetaphaseprecessionareimplementeddifferslightly,thebasic characteristicsofourframeworkandthesepreviousstudiesareverywellinlinewitheachother. Ingeneral, itwouldbehelpfulifmorecomputationalmemorymodelstake intoaccountthe intrinsicoscillatorybehaviorofthebraintorelatetheelectrophysiologicalpatternstomemory processes.Similarly,ourunderstandingofthemechanismsthatneurallydrivedesynchronization inthecortexduringmemoryformationisratherlimitedtodate.Previousanimalstudiesinthe attention[61]andmotor[62]domainsuggestpossiblecandidatemechanisms(i.e., thalamo-corticalinteractions),butwhetherthesecanbeextendedtomemoryremainstobeinvestigated.
Asforanytheoreticalframework,therearepotentiallimitationsandchallengesthatneedtobe considered.Forinstance,nohumandatacurrentlyexistthatsupportsthathippocampaltheta oscillationsregulateLTP/LTDinasimilarwayastheydoinrodents[53–55].Similarly,although placecells[63]andgridcells[64]havebeendiscoveredrecentlyinhumans,nostudyhasyet shownthephenomenonofphaseprecession inthehumanhippocampus. Anymechanistic frameworkthataimstoexplainepisodicmemoryneedstoaddresshowsynapticmodifications occur ata rapidtimescale. The theta-mediated LTP/LTDmechanisms observed inanimals provide amechanism thatfulfilsthese criteria. However, STDP (Box1),which, it has been argued,operatesonlongertimescales,mightbeamorequestionablemechanismforepisodic memoryformation. Interestingly, animal studiessuggest thatfor STDP protocolsto reliably induce LTP within a few seconds the stimulation has to occur in the theta rhythm [65], suggestinganinteractionbetween STDPandthetaoscillations,which wouldthenbeinline withtheframeworkproposedhere.
ConcludingRemarks
Theaimofthisopinionarticlewastoprovideamechanisticframeworkwithinwhichwecan understandthedifferentrolesofneuralsynchronizationanddesynchronizationintheserviceof episodicmemories.Linkingfindingsfromtherecentelectrophysiological literatureand inves-tigationsofoscillatorycorrelatesofepisodicmemoryformationandretrievalwiththewell-known CLSframework, weargue thathippocampal theta/gamma synchronizationisnecessaryfor bindingepisodeswhereascorticallow-frequencydesynchronizationisnecessarytorepresent thecontentoftheseepisodes. Thisframeworkmakesseveralclear predictions(Box2) and raisesimportantopenquestions(seeOutstandingQuestions)thatshouldbetestedinthefuture.
Acknowledgments
S.H.thankshiscollaboratorsfromtheCognitionandOscillationsLab,inparticular(inalphabeticalorder)Marie-Christin Fellner,TobiasStaudigl,andGerdWaldhauser,whoprovidedmostoftheempiricalgroundworkthatledtothisopinion article.ThisworkwassupportedbygrantsfromtheDeutscheForschungsgemeinschaft(HA5622/1-1),theEuropean ResearchCouncil(GrantAgreementno647954),andaSirHenryDaleFellowshipjointlyfundedbytheWellcomeTrustand theRoyalSociety(107672/Z/15/Z).
References
1. Fell,J.andAxmacher,N.(2011)Theroleofphasesynchronization inmemoryprocesses.Nat.Rev.Neurosci.12,105–118 2. Buzsaki,G.(2010)Neuralsyntax:cellassemblies,synapsembles,
andreaders.Neuron68,362–385
3. Hanslmayr,S.andStaudigl,T.(2014)Howbrainoscillationsform memories–aprocessingbasedperspectiveonoscillatory sub-sequentmemoryeffects.Neuroimage85,648–655 4. Hanslmayr,S.etal.(2012)Oscillatorypowerdecreasesand
long-termmemory:theinformationviadesynchronizationhypothesis. Front.Hum.Neurosci.6,74
5. Schneidman,E.etal.(2011)Synergyfromsilenceina combina-torialneuralcode.J.Neurosci.31,15732–15741
6. Mitchell,J.F.etal.(2009)Spatialattentiondecorrelatesintrinsic activityfluctuationsinmacaqueareaV4.Neuron63,879–888 7. O’Reilly, R.C.etal.(2014)Complementarylearningsystems.
Cogn.Sci.38,1229–1248
8. Rutishauser,U.etal.(2010)Humanmemorystrengthispredicted bytheta-frequencyphase-lockingofsingleneurons.Nature464, 903–907
9. Lisman,J.E.andJensen,O.(2013)Thetheta–gammaneural code.Neuron77,1002–1016
10.Rey,H.G.etal.(2014)Timingofsingle-neuronandlocalfield potentialresponsesinthehumanmedialtemporallobe.Curr.Biol. 24,299–304
OutstandingQuestions
What are the neural mechanisms underlying memory-related low-fre-quencypowerdecreases?
Ishippocampaltheta/gamma synchro-nization directly related to cortical desynchronization as reflected in alpha/betapowerdecreases? Doescorticaldesynchronization sup-porttherepresentation(i.e., reinstate-ment)ofmemories?
Isthetaphaseprecessionin hippocam-palareasCA1andC3relatedtocortical low-frequencypowerdecreases? Doesphaseprecession occurin the humanhippocampusanddoesit sep-arate stimulation-specific from non-stimulation-specificunits?
Can a computational model that implementsthehere-proposed hippo-campalsynchronizationandneocortical desynchronizationmechanismsaccount forexperimentaldata?
11.Staudigl,T.andHanslmayr,S.(2013)Thetaoscillationsat encod-ingmediatethecontext-dependentnatureofhumanepisodic memory.Curr.Biol.23,1101–1106
12.Staudigl, T.et al.(2015) Temporal-patternsimilarity analysis revealsthebeneficialanddetrimentaleffectsofcontext reinstate-mentonhumanmemory.J.Neurosci.35,5373–5384 13.Waldhauser,G.T.etal.(2012)Alpha/betaoscillationsindicate
inhibitionofinterferingvisualmemories.J.Neurosci.32,1953– 1961
14.Eichenbaum,H.etal.(2007)Themedialtemporallobeand rec-ognitionmemory.Annu.Rev.Neurosci.30,123–152 15.Staresina,B.P.andDavachi,L.(2009)Mindthegap:binding
experiencesacrossspaceandtimeinthehumanhippocampus. Neuron63,267–276
16.Long,N.M.etal.(2014)Subsequentmemoryeffectinintracranial andscalpEEG.Neuroimage84,488–494
17.Burke,J.F.etal.(2014)Thetaandhigh-frequencyactivitymark spontaneousrecallofepisodicmemories.J.Neurosci.34,11355– 11365
18.Burke,J.F.etal.(2015)Humanintracranialhigh-frequencyactivity duringmemoryprocessing:neuraloscillationsorstochastic vola-tility?Curr.Opin.Neurobiol.31,104–110
19.Lega,B.C.etal.(2012)Humanhippocampalthetaoscillations andtheformationofepisodicmemories.Hippocampus22,748– 761
20.Ray,S.andMaunsell,J.H.(2015)Dogammaoscillationsplaya roleincerebralcortex?TrendsCogn.Sci.19,78–85 21.Buzsaki,G.etal.(2012)Theoriginofextracellularfieldsand
currents–EEG,ECoGLFPandspikes.Nat.Rev.Neurosci.13, 407–420
22.Scheffer-Teixeira,R.etal.(2013)Onhigh-frequencyfield oscil-lations(>100Hz)andthespectralleakageofspikingactivity.J. Neurosci.33,1535–1539
23.Jutras,M.J.etal.(2009)Gamma-bandsynchronizationinthe macaquehippocampusand memory formation.J.Neurosci. 29,12521–12531
24.Fell,J.etal.(2001)Humanmemoryformationisaccompaniedby rhinal–hippocampalcouplinganddecoupling.Nat.Neurosci.4, 1259–1264
25.Staresina,B.P.etal.(2012)Memorysignalsaretemporally dis-sociatedinandacrosshumanhippocampusandperirhinalcortex. Nat.Neurosci.15,1167–1173
26.Colgin,L.L.etal.(2009)Frequencyofgammaoscillationsroutes flowofinformationinthehippocampus.Nature462,353–357 27.Igarashi,K.M.etal.(2014)Coordinationof
entorhinal–hippocam-palensembleactivityduringassociativelearning.Nature510, 143–147
28.O’Keefe,J.andRecce,M.L.(1993)Phaserelationshipbetween hippocampalplaceunitsandtheEEGthetarhythm.Hippocampus 3,317–330
29.Hasselmo,M.E.(2005)Whatisthefunctionofhippocampaltheta rhythm?Linkingbehavioraldatatophasicpropertiesoffield potentialandunitrecordingdata.Hippocampus15,936–949 30.Lega,B.etal.(2014)Slow-theta-to-gammaphase–amplitude
couplinginhumanhippocampussupportstheformationofnew episodicmemories.Cereb.Cortex.PublishedonlineOctober14, 2014.http://dx.doi.org/10.1093/cercor/bhu232
31.Tort,A.B.etal.(2009)Theta–gammacouplingincreasesduring thelearningofitem–contextassociations.Proc.Natl.Acad.Sci. U.S.A.106,20942–20947
32.Buzsáki,G.etal.(2013)Scalingbrainsize,keepingtiming: evolu-tionarypreservationofbrainrhythms.Neuron80,751–764 33.Jacobs,J.(2014)Hippocampalthetaoscillationsareslowerin
humansthaninrodents:implicationsformodelsofspatial naviga-tionandmemory.Philos.Trans.R.Soc.Lond.BBiol.Sci.369, 20130304
34.Noh,E.etal.(2014)Usingsingle-trialEEGtopredictandanalyze subsequentmemory.Neuroimage84,712–723
35.Burke,J.F.etal.(2014)Humanintracranialhigh-frequencyactivity mapsepisodicmemoryformationinspaceandtime.Neuroimage 85,834–843
36.Fellner,M.C.etal.(2013)Brainoscillatorysubsequentmemory effects differ in power and long-range synchronization betweensemanticandsurvivalprocessing.Neuroimage79, 361–370
37.Hanslmayr,S.etal.(2011)Therelationshipbetweenbrain oscil-lationsandBOLDsignalduringmemoryformation:acombined EEG–fMRIStudy.J.Neurosci.31,15674–15680
38.Meeuwissen,E.B.etal.(2011)Evidenceforhumanfronto-central gammaactivityduring long-termmemory encodingof word sequences.PLoSONE6,e21356
39.Hanslmayr,S.etal.(2014)Entrainmentofprefrontalbeta oscil-lationsinducesanendogenousechoandimpairsmemory forma-tion.Curr.Biol.24,904–909
40.Kielar, A. et al. (2015) Localization of electrophysiological responsestosemanticandsyntacticanomaliesinlanguage com-prehensionwithMEG.Neuroimage105,507–524
41.Singh,K.D.(2012)Which“neuralactivity”doyoumean?fMRI MEG,oscillationsandneurotransmitters.Neuroimage62,1121– 1130
42.Jensen,O.andMazaheri,A.(2010)Shapingfunctional architec-turebyoscillatoryalphaactivity:gatingbyinhibition.Front.Hum. Neurosci.4,186
43.Klimesch,W.(2012)Alpha-bandoscillations,attention,and con-trolledaccesstostoredinformation.TrendsCogn.Sci.16,606– 617
44.Greenberg,J.A.etal.(2015)Decreasesinthetaandincreasesin highfrequencyactivityunderlieassociativememoryencoding. Neuroimage114,257–263
45.Burgess,A.P.andGruzelier,J.H.(2000)Shortdurationpower changesintheEEGduringrecognitionmemoryforwordsand faces.Psychophysiology37,596–606
46.Khader,P.H.andRosler,F.(2011)EEGpowerchangesreflect distinctmechanismsduringlong-termmemoryretrieval. Psycho-physiology48,362–369
47.Waldhauser,G.T.etal.(2016)Episodicmemoryretrieval function-allyreliesonveryrapidreactivationofsensoryinformation.J. Neurosci.PublishedonlineNovember23,2015.http://dx.doi. org/10.1523/JNEUROSCI.2101-15.2016
48.Jafarpour,A.etal.(2014)Replayofveryearlyencoding repre-sentationsduringrecollection.J.Neurosci.34,242–248 49.Cohen,M.R.andKohn,A.(2011)Measuringandinterpreting
neuronalcorrelations.Nat.Neurosci.14,811–819
50.Ng,B.S.etal.(2013)EEGphasepatternsreflecttheselectivityof neuralfiring.Cereb.Cortex23,389–398
51.Csicsvari,J.etal.(1999)Oscillatorycouplingofhippocampal pyramidalcellsandinterneuronsinthebehavingrat.J.Neurosci. 19,274–287
52.Ketz,N.etal.(2013)Thetacoordinatederror-drivenlearninginthe hippocampus.PLoSComput.Biol.9,e1003067
53.Huerta,P.T.andLisman,J.E.(1995)Bidirectionalsynaptic plas-ticityinducedbyasingleburstduringcholinergicthetaoscillationin CA1invitro.Neuron15,1053–1063
54.Pavlides,C.etal.(1988)Long-termpotentiationinthedentate gyrusisinducedpreferentiallyonthepositivephaseof theta-rhythm.BrainRes.439,383–387
55.Hyman,J.M.etal.(2003)StimulationinhippocampalregionCA1 inbehavingratsyieldslong-termpotentiationwhendeliveredto thepeakofthetaandlong-termdepressionwhendeliveredtothe trough.J.Neurosci.23,11725–11731
56.Dayan,P.and Abbott,L.F.(2001)TheoreticalNeuroscience: ComputationalandMathematicalModelingofNeuralSystems, MITPress
57.Karpicke,J.D.andRoediger,H.L.,III (2008)Thecritical impor-tanceofretrievalforlearning.Science319,966–968 58.Westermann,J.etal.(2015)Systemconsolidationduringsleep-a
commonprincipleunderlyingpsychologicalandimmunological memoryformation.TrendsNeurosci.38,585–597
59.Hardt,O.etal.(2013)Decayhappens:theroleofactiveforgetting inmemory.TrendsCogn.Sci.17,111–120
60.Anderson,M.C.andHanslmayr,S.(2014)Neuralmechanismsof motivatedforgetting.TrendsCogn.Sci.18,279–292
61.Saalmann,Y.B.etal.(2012)Thepulvinarregulatesinformation transmissionbetweencorticalareasbasedonattentiondemands. Science337,753–756
62.Murthy,V.N.andFetz,E.E.(1996)Synchronizationofneurons duringlocalfieldpotentialoscillationsinsensorimotorcortexof awakemonkeys.J.Neurophysiol.76,3968–3982
63.Miller,J.F.etal.(2013)Neuralactivityinhumanhippocampal formationrevealsthespatialcontextofretrievedmemories. Sci-ence342,1111–1114
64.Doeller,C.F.etal.(2010)Evidenceforgridcellsinahuman memorynetwork.Nature463,657–661
65.Wittenberg,G.M.andWang,S.S.(2006)Malleabilityof spike-timing-dependentplasticityattheCA3-CA1synapse.J.Neurosci. 26,6610–6617
66.Norman,K.A.(2010)Howhippocampusandcortexcontributeto recognitionmemory:revisitingthecomplementarylearning sys-temsmodel.Hippocampus20,1217–1227
67.LopesdaSilva,F.(2013)EEGandMEG:relevanceto neurosci-ence.Neuron80,1112–1128
68.Bi,G.andPoo,M.(2001)Synapticmodificationbycorrelated activity:Hebb'spostulaterevisited.Annu.Rev.Neurosci.24,139– 166