What is Bottom-Up and What is Top-Down in Predictive Coding?

As a first approximation, we thus define top-down processes as instances of direct information transfer from higher to lower regions that skip ... DownloadArticle DownloadPDF ReadCube EPUB XML(NLM) totalviews ViewArticleImpact SHAREON Abstract Introduction AShortHistoryofBottom-UpandTop-Down Problems Solutions Conclusion ConflictofInterestStatement Acknowledgments References Exportcitation EndNote ReferenceManager SimpleTEXTfile BibTex Peoplealsolookedat HYPOTHESISANDTHEORYarticle Front.Psychol.,17May2013Sec.PerceptionScience https://doi.org/10.3389/fpsyg.2013.00276 Whatisbottom-upandwhatistop-downinpredictivecoding? KarstenRauss1*andGillesPourtois2 1InstituteofMedicalPsychologyandBehavioralNeurobiology,FacultyofMedicine,UniversityofTübingen,Tübingen,Germany 2PsychopathologyandAffectiveNeuroscienceLab,DepartmentofExperimentalClinical&HealthPsychology,GhentUniversity,Ghent,Belgium Everyoneknowswhatbottom-upis,andhowitisdifferentfromtop-down.Atleastoneistemptedtothinkso,giventhatbothtermsareubiquitouslyused,butonlyrarelydefinedinthepsychologyandneuroscienceliterature.Inthisreview,wehighlighttheproblemsandlimitationsofourcurrentunderstandingofbottom-upandtop-downprocesses,andweproposeareformulationofthisdistinctionintermsofpredictivecoding. Introduction Acentralassumptioninpredictive-codingtheoriesisthatactivityinthenervoussystemreflectsaprocessofmatchinginternallygeneratedpredictionstoexternalstimulation(Heekerenetal.,2008;Bar,2009).Evidenceforthisassumptionhasbeencollectedatdifferentlevelsofneuralprocessing(RaoandBallard,1999;Hosoyaetal.,2005;Mucklietal.,2005;Summerfieldetal.,2006;Alinketal.,2010),whichsuggeststhatpredictive-codingoperatesacrossawiderangeofspatialandtemporalscales. Asecondimportantassumptioninpredictivecodingisthatpredictionsaretransferredfromhierarchicallyhigherlevelsofprocessingtolowerones,whereassignalstravelingintheoppositedirectionencodepredictionerrors(RaoandBallard,1999;Serences,2008;Friston,2009;Grossberg,2009).Inaccordancewithterminologycommonlyusedinresearchonperception,ithasthusbeensuggestedthatpredictivesignalingreflectstop-downprocesses,whereasprediction-errorsignalingconstitutesbottom-upprocessing(Friston,2009;Alinketal.,2010;Hesselmannetal.,2010). Inthisreview,wearguethatpredictivecodingprovidesapowerfulconceptualframeworkthatgoesbeyondthestandarddichotomyof“bottom-up”and“top-down.”Wefirstprovideanoverviewofpreviousattemptsatdefiningbottom-upandtop-downprocesses,andwehighlighttheirproblemsandlimitations.Wethenoutlinehowpredictivecodingoffersauniqueperspectiveforre-definingwhatismeantbytheseterms. Throughouttheremainderofthispaper,werefertoascendinganddescendingconnectionswhendiscussingtheanatomyofbiologicalsystemsorthearchitectureofartificialones(Friston,2005;Clark,inpress).Functionalactivityalongthesetwotypesofconnectionsisreferredtoasfeedforwardandfeedback,respectively.Weavoidadditionaltermssuchas“recurrent”or“reentrant,”becausewebelievetheycanbesubsumedunder“feedback”inmostcases. AShortHistoryofBottom-UpandTop-Down Morethan30yearsago,KinchlaandWolfe(1979)setouttotestwhethervisualprocessingisorganizedbottom-uportop-down.Followinguponearlierwork(Reicher,1969;Navon,1977),theauthorsusedcompoundstimuli,whereaglobalshapeismadeupofsmaller,localelements.Upper-caseletterswereusedasstimuli,andtop-downprocessingwasassumedtomanifestitselfinshorterresponsetimes(RTs)iftheglobalshape,ratherthanthelocalelements,correspondedtoatargetletter.Bottom-upprocessingwasdefinedas“theopposite,”i.e.,fasterresponsesifthelocalstimuluselements,ratherthantheoverallshape,correspondedtoatargetletter.Resultsindicatedthatwhetheraglobalshapeoritsconstituentelementsareprocessedfastercriticallydependsonstimulussize.KinchlaandWolfe(1979)concludedthatvisualperceptiondoesnotproceedstrictlybottom-uportop-down,but“middle-out.”Theysuggestedthatthevisualsystemusestheinformationmostreadilyavailableinthecontextof,forexample,aparticularstimulussizeandviewingdistance.Basedonthisinformation,thesystemwouldthenworkitswaytowardmoreglobalormorelocalaspects,asrequiredbythetask. Morerecently,Mellonietal.(2012)examinedthegenerationofsaliencymapsinthevisualcortexwithfunctionalmagneticresonanceimaging(fMRI).Theauthorsrefertobottom-upsalienceasthedegreeofdifferencebetweenastimulusanditsneighbors.Intheirvisualsearchtask,targetanddistractorgratingsdifferedinorientation.Bottom-upsaliencewasmanipulatedbyaddingcolorasanadditionalstimulusdimension.Thus,thetargetstimuluscouldbeeitherasingletonintermsofcolorandorientation;asingletonintermsoforientationonly;orasingletonintermsoforientationaccompaniedbyadistractorsingletonintermsofcolor.Top-downcontrolisdescribedbyMellonietal.(2012)astheinfluenceofourinnergoalsonstimulusselection.Thiswasmanipulatedbypresentingtrialsofdifferentdegreesofbottom-upsalienceeitherinblocksorrandomlyintermixed.Resultsindicatedthatprimaryvisualcortex(V1)encodesonlybottom-upsalience,V2encodesonlytop-downcontrolsettings,andV4encodestheinteractionbetweenthetwo.Theauthorsconcludedthatmultiplesaliencymapsarepresentatdifferentlevelsofprocessing.Stimulusselectioncouldthusbeflexiblyadaptedbyreferringtothesaliencymapmostrelevantinthecontextofagiventask(Weidneretal.,2009). InthethreedecadesbetweenthestudiesofKinchlaandWolfe(1979)andMellonietal.(2012),conceptualandmethodologicaladvanceshavesubstantiallyincreasedourknowledgeofsensoryprocessingingeneralandofthevisualsysteminparticular.Whathasn’tchangedduringthistimeisournotionofbottom-upandtop-downprocesses.Stayingwithourfirstexample,KinchlaandWolfe(1979)wrote: […]theelectrophysiologicalanalysisof“receptivefields”seemedtosuggestabottom-upmodeofprocessing:Cellsassociatedwithprogressivelymorecomplexfields[…]werefoundasonewentfromtheretinatothevisualcortex,asifsystemsfordetectinglow-order“features”[…]fedintosystemsfordetectingprogressivelymorecomplexpatterns.(KinchlaandWolfe,1979,p.225) […]ithasbeensuggestedthattheorderofvisualprocessingisbestdescribedasatop-downprocess,withhigher-orderformsprocessedfirst,followedbylower-orderforms[…](ibd.) ComparethistoamorerecentdefinitionisgivenbyPalmer(1999): “Bottom-up”processing[…]referstoprocessesthattakea“lower-level”representationasinputandcreateormodifya“higher-level”representationasoutput.Top-downprocessing[…]referstoprocessesthatoperateintheoppositedirection,takinga“higher-level”representationasinputandproducingormodifyinga“lower-level”representationasoutput.(Palmer,1999,pp.84–85) Thecommondenominatoroftheseandvirtuallyallotherdefinitionsofbottom-upandtop-downprocessescanbesummarizedasfollows: 1.Informationprocessingisorganizedhierarchically. 2.Lowerlevelsofthehierarchyrepresentdetailedstimulusinformation,whilehigherlevelsrepresentmoreintegratedinformation. 3.Informationexchangebetweenlevelsisbidirectional. Theapparentsimplicityoftheseassumptionsmayexplainwhythetermsbottom-upandtop-downareusedsofrequentlyinresearchonperception.However,theseassumptionscanbequestioned,andwebrieflyhighlightsomeoftheproblemssurroundingthem. ANoteonHierarchies FellemanandVanEssen(1991)advocatedadistributedprocessinghierarchyintheprimatevisualcortexbasedonanatomicalconnectivitypatterns.Morerecently,however,HegdéandFelleman(2007)notedthatahierarchicallyorganizedanatomicalstructuredoesnotimplythatvisualprocessingisitselfhierarchical,northatthefunctionalhierarchy(ifitexists)matchestheanatomicalone(Rousseletetal.,2004).Moreover,HegdéandFelleman(2007)statethattheanatomicalhierarchyitselfisclear-cutonlyuptothelevelofareasV4andMT,withthesender-receiverdistinctionbecominglessevidentathigherlevels.Thus,thenotionofhierarchicalprocessingcanbecriticizedonbothstructuralandfunctionalgrounds. Thesecriticismsmaybepartlyresolvedbyadoptingamoreflexibleviewofstructuralandfunctionalhierarchies.AnexampleforsuchaviewisprovidedbyEngeletal.(2001),whodistinguishfour“flavors”oftop-down: 1.ananatomicalone,equatingtop-downprocesseswithfunctionalactivityalongdescendingconnectionsbetweenthelevelsofthehierarchy 2.acognitivistone,wheretop-downmeanshypothesis-drivenprocessing 3.agestaltistone,viewingtop-downprocessesintermsofcontextualmodulationsofbottom-upprocessing 4.adynamicistone,describingtop-downprocessesintermsofanentrainmentoflocalneuronalpopulationsbywidespreadoscillatoryactivityindistantanddistributedbrainregions. Thelastoftheseflavorsdoesnotrequireafixedanatomicalorfunctionalhierarchy.Itaccommodatesflexiblerecruitmentofbrainregionsfordifferenttasks,withoutnegatingtheknownspecializationsoftheseregions. Indeed,thedifficultiesinestablishingclear-cutprocessinghierarchiesinthecentralnervoussystembasedoneitheranatomicalorfunctionalcriteriamaybeadirectconsequenceofthesystem’sadaptability.Thus,ratherthanrelyingonafixedhierarchyforprocessingalltypesofstimuliandperformingallsortsoftasks,itseemslikelythatthesystemusesdifferenthierarchiesfordifferenttasks.Thisdoesnotmeanthatthenotionofprocessinghierarchiesisobsolete,butthathierarchicalprocessingshouldbeseenasaconceptualsimplification. Incidentally,theproblemsassociatedwithhierarchicalprocessingmodelswerealreadyrecognizedbyKinchlaandWolfe(1979).Theystatedthatalthoughindividualimagescouldusefullybedescribedintermsofahierarchyoffeatures,theknowledgeunderlyingourperceptionoftheseimageshadneithertopnorbottom.Nevertheless,theyarguedthatconceptualizingthisknowledgeashierarchicallyorganizedremainsausefulsimplification.Intriguingly,theirdiscussionalsopointstotheroleofpredictabilityinestablishingthesehierarchies: […]amajorreasonforsuch[ahypotheticalhierarchical]organizationisthelife-longsequentialpatternofourvisualexperiencewherebyrecognitionofaformatonelevelofstructureisanalmostinvariantprecursoroftherecognitionofformsatlevelsslightlyhigherorlower.(KinchlaandWolfe,1979,p.229) Problems Evenifoneacceptsthebasicassumptionsunderlyinghierarchicalprocessingmodels,thereareanumberofproblemsassociatedwithcurrentnotionsofbottom-upandtop-downprocesses.Someexamplesfromtheliteraturewilldemonstratethatoneauthor’stop-downmaywellbeanotherone’sbottom-up. AFuzzyDichotomy Inarecentpositionpaper,Theeuwes(2010a)reviewedevidenceforthenotionthatinitialprocessingofavisualstimulusisexclusivelydrivenbybottom-upfactors.Theauthorfocusesonselectionprocessesinvisualsearchtasksanddefinestop-downselectionas: completelyundercontroloftheintentionsoftheobserver anactivevolitionalprocess basedonexpectancyandgoalset(Theeuwes,2010a,pp.77–78) Bottom-upselection,ontheotherhand,isdescribedas: determinedbythefeaturepropertiespresentintheenvironment passive[and]automatic associatedwithsalience drivenbyemotionalcontentorpreviousexperience(ibd.) IndependentofwhetherornotonesubscribestoTheeuwes’(2010a)pointofview,thenumberofconceptsinvokedtodefinetwocommonlyusedtermssuggeststhatthereisactuallymoretoconsiderthanasimpledichotomy.Inotherwords,theuseofasimplebinaryclassificationdoesnotappeartocapturetheprofusionofmentalprocessesorfunctionsdescribedinthesestatements,whichincludeintention,volition,expectation,andemotion.Accordingly,alargepartofthediscussionofTheeuwes’(2010a)targetpaperconcernsthedelimitationofbottom-upandtop-downprocesses(Egethetal.,2010;FolkandRemington,2010;Kristjánsson,2010;Theeuwes,2010b). SearchingfortheTop Importantly,theproblemswithdistinguishingbottom-upfromtop-downprocessesarenotlimitedtopsychologicalstudies.Confusionalsoseemstoreignattheapparentlymoresolidlevelofcellularneuroscience. Forexample,Rolandetal.(2006)usedvoltage-sensitivedye(VSD)imaginginanesthetizedferrets.Theauthorsfoundthatexposuretoaluminance-definedsquareelicitedforwardandlateralactivationspreadthatwasfollowedbyafeedbackwaveofactivitytravelingfromextrastriatevisualareastoV1.Thisfeedbackwaveselectivelyhighlightedfirstthestimulusrepresentationandthentherepresentationofthebackground.Rolandetal.describethisfeedbackwaveasapossiblemechanismoftop-downinfluencesonearlyvisualcortex. Kuhnetal.(2008),ontheotherhand,useddepth-resolvedVSDimaginginmiceduringanesthesiaandwakefulness.Theauthorsobservedcharacteristicdesynchronizationsinlayer1ofthesomatosensorycortexuponawakening.Theyconsiderthisasevidenceforlong-rangecorticalandthalamicinputexertingtop-downcontroloversensoryprocessingduringwakefulness. Arguably,theterm“top-down”isconceptuallyuselessifitdescribesneuralactivityduringbothwakefulness(Kuhnetal.,2008)andanesthesia(Rolandetal.,2006),withsourcesineithercloselyneighboringregions(Rolandetal.,2006)oranywhereinthenervoussystem(Kuhnetal.,2008).However,suchaninterpretationdirectlyemergesfromwhatEngeletal.(2001)termedtheanatomicalflavoroftop-down.Thatis,ifactivityalongdescendingfibersnecessarilyreflectstop-downprocesses,thelattermayoriginatefromanysourceconnectedtothecurrentregionofinterest,providedthatthesourceislocatedatahigherlevelofthehierarchy.Andifsourceactivitypersistsduringsleeporanesthesia,thiswouldalsoqualifyastop-downactivity. Barlow(1997)succinctlynotedthisproblem,statingthatthevisualsystemhasnotop.Reviewingpotentialsourcesofknowledgeusedinvisualprocessing,hearguesthattheinteractionofthegeneticallydeterminedstructureofthevisualsystemandtheredundanciespresentinallvisualimagesmayexplainmanyoftheeffectsusuallyattributedtotop-downfactors.Heconcludes: Toavoidthetop-down/bottom-updichotomyblindingustomoreimportantquestions,theterm“top-down”shouldperhapsbechallengedwheneveritisused[…](Barlow,1997,p.1146) Barlow’s(1997)understandingoftop-downprocessesisthattheyinvolveoutsideknowledgepreviouslyacquiredthroughtrainingorteachingandbroughttoactuponcurrentsensoryinput.Here,“outside”referstoregionsofthenervoussystemnotprimarilyconcernedwithvisualprocessing. WhilewefullyagreewithBarlow’s(1997)conclusion,itseemswhathisargumentessentiallyhighlightsisthevaguenessofwhatismeantbyatop-downprocess.Hiscritiquerestsontheassumptionthatbottom-upprocessingconsistsoftheinterplaybetweenthesystem’sgeneticallypredeterminedstructureandtheorganism’scurrentenvironment.However,sincethestructureofthesystemalsodefineshowitstoresandretrievespreviouslyencodedinformation,hisunderstandingofbottom-upprocessesseemstoleavehardlyanyroomfortop-downprocesses. Bottom-Up/Top-Downvs.Feedforward/feedback Insummary,currentnotionsofbottom-upandtop-downprocesseslumptogetherlargelydisparatestructuresandfunctions.Asaconsequence,itisoftenunclearwhetherthereisadifferencebetweenbottom-upandtop-downprocessesontheonehandandfeedforwardandfeedbackactivity,ontheother.Inthiscontext,itisinterestingtonotethatanumberofauthorsspeakof“top-downfeedback”torefertoattention-inducedmodulationsofeitherstimulus-evokedactivity(Martínezetal.,2001),baselineactivity(Kastneretal.,1999),orbackgroundconnectivity(Al-Aidroosetal.,2012).Inmostofthesecases,theaimseemstobetohighlighttheparallelsbetweenthepsychologicaldescriptionofhighercognitiveprocessesandtheanatomicaldescriptionoffibersconnectinghigh-leveltolow-levelregions.However,amoreliteralinterpretationwouldsuggestthat,ifonehastospecifythatacertaintypeoffeedbackistop-down,othertypesoffeedbackarenot.Wewillreturntothisideainthefollowingsection. Plasticity Inanoriginaldemonstrationofhowcultureshapesperception,TseandCavanagh(2000)showedthatthedirectionofapparent-motionperceivedduringthestepwisepresentationofaChinesecharacterdiffersbetweenChineseandAmericanparticipants.Thecharacterinquestionwasunveiledonestrokeatatime,withoutactualmovementscontainedinthestimulus.BasedonGestaltprinciples(Wagemansetal.,2012a,b),thelaststrokewasexpectedtobeperceivedasaright-to-leftmovement,andthisiswhatAmericansubjectsreportedseeing.Ontheotherhand,amajorityofChinesesubjectsreportedseeingamovementfromlefttoright,inaccordancewiththemovementperformedwhenwritingthecharacter.Theauthorsconcludedthatbottom-upcuesformotionperceptioncanbeoverriddenbytop-downfactorslinkedtoculturallyshapedexpectations. Intheirdiscussion,TseandCavanagh(2000)notethat: The“standard”viewoftop-downprocessingisthatlatervisualareasinfluenceearlierareasviafeedbackconnections.Thereareotherpossibilities,however.Expectationandknowledgecouldinprinciplealterthecircuitryinvolvedingrouping,inwhichcaseatop-downinfluencewouldbeexertedinabottom-upmanner(TseandCavanagh,2000,p.B32) Thisinterpretationoftop-downprocessesraisesanimportantissue:beyondthedifferencesinhowbottom-upandtop-downprocessesareconceptualized,individualdifferencesmaywellleadtosituationswhereoneparticipant’sbottom-upisanotherone’stop-down. Solutions Giventheseubiquitousproblems,webelieveitistimethattheconceptsofbottom-upandtop-downberefinedbasedoncurrentevidence.Webrieflyhighlightpreviousworkaiminginthisdirection,beforeoutliningapredictive-codingmodelthatmayhelpresolvesomeoftheissuesoutlinedabove. Top-Down≠Feedback InaninsightfulcommentaryonTheeuwes(2010a),Rauschenberger(2010)arguesthatthedichotomybetweenbottom-upandtop-downprocessesshouldbeabandonedaltogetherinlightofever-increasingevidenceontheimportanceofinteractiveinformationprocessinginperception(DiLolloetal.,2000).Rauschenberger(2010)goesontoproposethatthedirectionalityofneuralpathwaysdoesnotnecessarilycorrespondtotheirprimaryorexclusiveinvolvementinbottom-uportop-downprocesses. Theideathatascendinganddescendingpathwaysmaybejointlyinvolvedinbothbottom-upandtop-downprocessesseemstousthesinglemostimportantinsightforsalvagingtheseconcepts.Accordingly,thisideaisatthecoreofourownproposalsbelow. AThirdElement Followingthelivelydiscussionsurroundinghispositionpaper(Theeuwes,2010a),Theeuwesco-authoredanarticleentitled“Top-downvs.bottom-upattentionalcontrol:afailedtheoreticaldichotomy”(Awhetal.,2012).Theauthorsmakethepointthatattentionalcontrolcannotbefullydescribedwhenonlybottom-upsalienceandtop-downgoalsaretakenintoaccount.Rather,selectionhistoryneedstobeconsideredaswell,i.e.,whetheraparticularstimuluswaspreviouslytask-relevantorrewarded.Awhetal.(2012)highlighthowdifferencesinselectionhistoryareoftenconfoundedwithtop-downeffectssuchasselectiveattention,andhowthismayexplainsomeofthecontradictoryfindingsinthevisualsearchliterature(e.g.,MaljkovicandNakayama,1996;Wolfeetal.,2003). Thecentralandsometimesparadoxicalroleofselectionhistoryonattentionalcontrolisalsoinaccordancewiththeideathatactivepredictionsofincominginformationmayaffectsensoryprocessingatmultiplelevels.AsAwhetal.putit: […]webelievethat[selectionhistoryeffects]shareonecorefeature:ineachcase,pastselectionepisodesarerecapitulatedinsubsequenttrialswhentherelevantcontextisencounteredagain.(Awhetal.,2012,p.440) Notwithstandingtheimportanceofselectionhistoryeffects,weareunsurewhetheraddingathirdconcepttothediscussionishelpful.Indeed,thedefinitiongivenbyAwhandco-workersissobroadthatitspansvirtuallyalltypesofneuralactivity,asselectionhistorymaybeencodedatverydifferenthierarchicallevels(Treue,2003)andonverydifferenttime-scales(Barlow,1997). APredictive-CodingAccount Inthefollowing,weoutlineapredictive-codingmodelfromwhichwederivesimpleandunequivocaldefinitionsofbottom-upandtop-downprocesses.Ourgoalisnottoprovidedetailedguidelinesfordistinguishingtheseprocessesincomplexbiologicalsystems.Rather,weaimtore-establishtop-downandbottom-upasusefulheuristiccategories.Inordertodoso,westartfromtwosimplepremises: 1.Bottom-upandtop-downarenotopposites. 2.Withinhierarchicalsystems,bothascendinganddescendingconnectionsareinvolvedinbottom-upandtop-downprocesses. Thesepremisesareanextensionofthemainideaunderlyingpredictive-codingtheoriesandmoregeneralaccountsofbrainfunction(SummerfieldandEgner,2009;Friston,2010):namely,thatbidirectionalinformationexchangebetweenthelevelsofhierarchicalsystemsservestoreconcileincominginformationwithinternallygeneratedpredictions.Ifthisisthecase,therearetwoclassesofquestionsonemayaskaboutthesystem: 1.Howisstimulusinformationcomparedtopredictions? 2.Wheredopredictionscomefrom,andwherearepredictionerrorsroutedto? Forthesecondclassofquestions,webelievethemostinterestingfunctionalunitistheloopformedbyacentralregionthatgeneratespredictionsandalower-levelregionthatreceivesstimulusinformationand/orprediction-errorsignalsfromtheperiphery. Nowconsiderabiologicallyinspired,yet-to-beprogramedcomputermodelsimilartopreviouslypublishedcomputationalaccountsofthevisualcortex(RaoandBallard,1999).Themodelsystemishierarchicallyorganizedoverseverallevels,withincreasinglycomplexstimulusattributesrepresentedathigherlevels.Forthetimebeing,wefurtherassumethatascendinganddescendingconnectionsonlylinkneighboringareaswithinthehierarchy. CombiningtheseassumptionsleavesuswithaseriesofloopsasshowninFigure1.Weproposethattheseloopsconstitutethebasisofbottom-upprocessing.Inotherwords,bottom-upprocessingconsistsofoneormorecyclesoffeedforward-feedbackactivityalongascendinganddescendingconnections,instantiatedacrosspairsofneighboringareasinthehierarchy.Thisiterative,multi-layeredprocessreflectsthecurrent,semi-hardwiredarchitectureofthesystem,whichinturnreflectsacombinationofphylogeneticandontogeneticinformationavailabletotheorganism. FIGURE1 Figure1.Bottom-upprocessingasaseriesofloopsformedbyascendinganddescendingconnectionsbetweenneighboringlevelsinahierarchicallyorganizedsystem.Forsimplicity,levelsL1–L4areshownasconsistingofthreelayersonly(ascendinginput,descendinginput,andoutput). Thecrucialpointhereisthatweexplicitlyassignaroleinbottom-upprocessingtodescendingconnectionsbetweenneighboringareasinthehierarchy.Bottom-upprocessingalongthisseriesofloopsisreliableandrelativelyfast(Bullier,2001;Bacon-Maceetal.,2005).However,itisnotveryflexible. Itisnowgenerallyappreciatedthateventheearliestlevelsofsensoryprocessingremainmalleablethroughoutlife(NevilleandBavelier,2002;Bavelieretal.,2010).Animalstudieshavedemonstratedthatintensetrainingfollowedbyappropriateconsolidationcanleadtolong-lastingmodificationsofneuralresponsesinprimarysensorycorticesvirtuallyfromthefirstspikeonward(Cristetal.,2001;Lietal.,2004).Recentevidencefromhumanstudiespointsinthesamedirection(Schwartzetal.,2002;Pourtoisetal.,2008;Baoetal.,2010). Thisiswhatwerefertoasasemi-hardwiredarchitecture:onethatissufficientlystabletoallowforrapidprocessingofstimulithathavebeenfrequentlyencountered,eitheroverthecourseofaspecies’evolutionorduringanindividual’slifespan;andthatisatthesametimesufficientlyflexibletoadapttolastingandpervasivechangesintheorganism’senvironment.However,neuralplasticityatthetime-scalesexaminedintheabove-mentionedstudiesdoesnotofferenoughflexibilitytoanorganisminacompetitiveandrapidlychangingenvironment. Thus,ataconceptuallevel,top-downprocessescanbeconceivedofasinfluencesthatconfermoment-to-momentflexibilityontothesemi-hardwirednetworkthatatanygiventimeensuresefficientandreliablesensoryprocessinginknownenvironments.Thisrelatestotheassumptionthatthehierarchicalorganizationofsensorysystemsseenacrossmodalitiesandspeciesreflectstwobasicimperativesintermsofpredictivecoding(Barlow,1985):first,totakeinamaximumofnewinformationinordertodetectcontingenciesintheenvironment;andsecond,toexploitthesecontingencies,onceextracted,toconstructpredictionsabouttheenvironmentwhichcanberapidlyappliedtoguideadaptivebehavior. Contingenciesandthepredictionsbaseduponthemvaryincomplexity,andthereisrichevidenceshowingthatlow-levelcontingenciesmaybereflectedinthebasicorganizationoflow-levelsensorycortices(RaoandBallard,1999;Barlow,2001),andevenattheleveloftheretina(Srinivasanetal.,1982;Barlow,1997;Hosoyaetal.,2005).AparticularlystrikingexampleisgivenbytheseminalworkofRaoandBallard(1999):theauthorstrainedanartificialneuralnetworkwithnaturalimagesandexplainedend-stopping,anextra-classicalreceptivefieldpropertyofV1neurons,intermsofthefrequentoccurrenceintheirstimulusmaterialoforientedlinesextendingbeyondthesmallreceptivefieldsofthemodel’sV1neurons. High-levelcontingencies,ontheotherhand,relatetothehigher-ordercausesofwhatiscurrentlyperceived(Friston,2005)andcanoftenonlybedetectedbyintegratinginformationacrosslargeportionsofspace,longintervalsoftime,andmultiplemodalities(Clark,inpress).Asanexample,considernaturallanguages:understandingaspokenorwrittenphraseinGermancanbecomplicatedbythefactthattheverbmaybepositionedmoreflexiblythanforexampleinEnglish.Thus,crucialaspectsoftheinformationtransmittedremainunspecifiedfordifferentperiodsoftime,requiringdifferentprocessesofverification,interpretation,andprediction(Dambacheretal.,2009).Importantly,suchdifferencesarecompletelyindependentoftheactualcontentofthemessageandalso,tosomeextent,independentofsensorymodality.Thesehigher-ordercontingenciesmaybesubjecttomorerapidchanges(e.g.,switchingbetweenlanguages)thantheregularitiestowhichthelowerlevelsofsensorysystemsaretuned(Kerstenetal.,2004),thusrenderingadaptationsbasedonlong-lastingstructuralreconfigurationsorstimulus-specificfunctionalchangesinappropriate. Weassumethatthehigherlevelsinourmodelsystemextractsuchhigher-ordercontingenciesanddynamicallyusethemtocreateasetofrelativelyabstractpredictionsthatcanberapidlyadaptedorexchanged.However,fortheseabstractpredictionstobeuseful,theremustbeawayforthemtoeffectivelymodulateiterativebottom-upprocessingatmanydifferentlevelsofthehierarchy.Therefore,weassumethattop-downprocessesincludesomeformofbypassingorshort-circuitingofbottom-upprocesses.Thefastestandmostdirectroutetoachievethiswouldbeviadirectconnectionsbetweenhigh-levelregionL(z)andlow-levelregionL(x)whicharenothierarchicalneighborsandthereforedonotformabottom-upprocessingloop. Asafirstapproximation,wethusdefinetop-downprocessesasinstancesofdirectinformationtransferfromhighertolowerregionsthatskipatleastonelevelinthehierarchy.Thisanatomicalcriterionfordistinguishingtop-downfrombottom-upprocessesspecifiestheminimalrequirementsfortheconceptualframeworkoutlinedabove.Thatis,inordertorenderbottom-upprocessingmoreflexible,high-levelregionsareassumedtomodulateactivityinlow-levelregionsusinginformationthatisnotrepresentedatintermediatelevels. ThisbasicideaisdepictedinFigure2:ahigh-levelregionrelaysinformationdirectlytoalowertierofthesystem,bypassinganareasituatedmoreimmediatelyupstreamofthistargetregion.Giventhereciprocityofmostcorticalconnections(FellemanandVanEssen,1991),weassumethattop-downprocessesalsoinvolvebidirectionalinformationexchange,i.e.,theyrelyonloopsofascendinganddescendingconnectionsdistinctfromthosemediatingbottom-upprocessing. FIGURE2 Figure2.Top-downeffectsasdirectinfluencesofasourceregionlocatedatleasttwolevelsabovethetargetregioninahierarchicalpredictive-codingsystem. Themainpointhereisthatinourmodel,backwardinformationtransferbetweenimmediateneighborsinthehierarchydoesnotqualifyastop-downprocessing.Thisisimportantbecauseitseparatesourviewoftop-downprocessesfromanapparentlysimpleronebasedonlyondirectionofinformationflow. Insummary,weproposeananatomicalcriterionfortheseparationoffunctionalactivitypatternsintocategoriesofbottom-upandtop-down.Givenastrictlyhierarchicalsystem,thiscriterionspecifiestheminimalrequirementsforclassifyingaparticularpatternofeffectiveconnectivityastop-down:namely,theremustbeadirectanatomicalconnectionbetweenasourceandatargetregion,andthesourcemustbelocatedatleasttwolevelsabovethetarget. TestingthePredictive-CodingAccount Itisimportanttonotethatourproposalsarenotmeanttobeacomprehensivetheoryofhierarchicalprocessing.Rather,weprovideaxiomaticdefinitionsoftwoequivocaltermscommonlyinvokedtobuildtheoriesandtointerpretempiricalresults.Ourideasshouldbecriticallydiscussedandempiricallytested.Ideally,thesetestswoulddemonstratetwothings: 1.Ourproposalsprovideabetterheuristicforinterpretingpreviousempiricalresultsthancurrentloosenotionsofbottom-upandtop-downprocesses. 2.Ourproposalsprovidenovelhypothesesandcorrectlypredicttheoutcomeoffutureexperiments. Explainingpreviousresults Animportantlimitationofcurrentdescriptionsofbottom-upandtop-downprocessesisthattheylargelyfailtoexplainfundamentaleffectsonperceptualprocessing,suchaspriming(Grill-Spectoretal.,2006).ThislimitationclearlyemergesinthediscussionofthepositionpaperbyTheeuwes(2010a)alreadymentioned.Alargepartofthisdiscussionfocusesonwhetherprimingshouldbeconceivedofasabottom-up(Theeuwes,2010a,b)oratop-downprocess(deFockert,2010;Egethetal.,2010;EimerandKiss,2010;Kristjánsson,2010;Mülleretal.,2010). Ontheonehand,primingcanbeseenasaverybasicprocessthatoneistemptedtoclassifyaspurelybottom-up:itisrapid(Dehaeneetal.,2001),oftenautomaticinthesenseofbeingunrelatedtotheobserver’sgoals(MoorsandDeHouwer,2006),anditcanapplytoverybasicstimuluscharacteristicscodedatthelowestlevelsofsensoryprocessing(MaljkovicandNakayama,1994,1996).Ontheotherhand,anincreasingbodyofevidenceshowsthatprimingeffectscanbeobject-related(ChunandJiang,1998;Kristjanssonetal.,2008)aswellaslinkedtotheobserver’sgoals(KieferandMartens,2010).Moreimportantly,itisdifficulttodescribeprimingaspurelystimulus-driven,asitreflectsaneffectofpreviousstimulation,andhencesomekindofmemorytracehastobeinvolved(Kristjánsson,2010). Webelieveourproposalscouldprovevaluableinresolvingthisdebate.Theysuggestthatprimingcanbeinstantiatedbothalongtheloopsofascendinganddescendingconnectionsbetweenhierarchicalneighborsthatwehavedescribedastheanatomicalbasisofbottom-upprocessing,andviatop-downprocessesthatoperatealonglong-rangecortico-corticalconnections(Summerfieldetal.,2006).Inthisview,primingeffectsarenotseenasamonolithicphenomenonreflectingaunitaryprocess(forasimilararguments,seeHenson,2003;KristjánssonandCampana,2010).Rather,differentformsofprimingmayarisedependingontheprocessesandpathwaysrecruitedtoyieldthissensoryfacilitation.Morespecifically,discrepanciesintheliteratureconcerningtheinfluenceofgoals,taskset,etc.,onprimingcouldbelinkedtotheinvolvement(ornot)ofhigh-levelcontrolregionsthatrepresenttask-relevantcontingencies. Derivingnovelhypotheses Ourproposalsrestontheassumptionthattop-downprocessesconferflexibilityontothesemi-hardwirednetworkunderlyingbottom-upprocessesbyshort-circuitingthelatter.Onewaytotestthiscentralassumptionistoshowthatintheabsenceoftop-downinput,bottom-upprocessingiseithersubstantiallydelayedorfailsentirely(Superetal.,2001).Asanexamplefortestingthisidea,considerastudybyMucklietal.(2005),whoshowedthattheperceptionofapparent-motionleadstoBOLDactivityinV1alongtheapparent-motiontrajectory.TheauthorsconcludedthatthemostlikelysourceforthiseffectwasfeedbackfromareaMTtoV1. GiventhatMTislocatedseverallevelsaboveV1inthevisualhierarchy(FellemanandVanEssen,1991),anunequivocaldemonstrationthattheefficientdetectionofthe(illusory)motionstimulusdependsontop-downprocessesshort-circuitingbottom-upprocesseswouldrequireselectivelydeactivatingtheconnectionsbetweenV1andMTinbothdirections.Thisshouldpreserveprocessingalongthehierarchyofbottom-uploopsuptoandbeyondMT.Ourmodelsuggeststhatinthiscase,theperceptionofillusorymotionshouldbeabsent,oratleastsubstantiallydelayedandreduced. Whilethisexperimentmaybedifficulttocarryoutwithcurrentmethods,morerealisticapproachesexist.Forexample,inanextensionofpreviouswork(Pascual-LeoneandWalsh,2001)usingtranscranialmagneticstimulation(TMS),onecouldtemporarilydeactivateMT,whichshouldnotonlyreducetheperceptionofapparentmotion,butalsotheconcomitantactivityalongtheapparent-motiontrajectoryinV1.SuchanapproachcannotisolatethespecificcontributionsofdirectconnectionsbetweenMTandV1.However,ifsimilareffectscanbeobtainedforothertypesofillusions,involvinghigh-levelvisualareasotherthanMT,thiswouldofferconvergingevidenceforourview. Conclusion Basedonapredictive-codingmodel,wehaveoutlinedaconceptuallyunequivocaldistinctionbetweenbottom-upandtop-downprocessesthataddressessomeofthelimitationsofourcurrentunderstandingoftheseterms.Ourproposalshighlightthemutualinterdependenceandconstantinteractionbetweenbottom-upandtop-downprocesses.Thus,ratherthansearchingforcasesofpurebottom-uportop-downprocessing,futureeffortsshouldaddresstheirrelativecontributionsaswellasthemechanismsoftheirinteractioninthecontextofagiventask. ConflictofInterestStatement Theauthorsdeclarethattheresearchwasconductedintheabsenceofanycommercialorfinancialrelationshipsthatcouldbeconstruedasapotentialconflictofinterest. Acknowledgments WewouldliketothankSophieSchwartz,PatrikVuilleumier,ChristophMichel,andMichaelHerzogforstimulatingdiscussions,aswellasLauraHerdeandthereviewersforthoughtfulcommentsonearlierversionsofthemanuscript.KarstenRausswassupportedbyafortunejuniorgrantfromtheMedicalFacultyattheUniversityofTuebingenandapostdoctoralfellowshipforleadingearlycareerresearchersfromtheBaden-WürttembergStiftung.GillesPourtoisisfundedbytheEuropeanResearchCouncil(StartingGrant#200758)andGhentUniversity(BOFGrant#05Z01708). References Al-Aidroos,N.,Said,C.P.,andTurk-Browne,N.B.(2012).Top-downattentionswitchescouplingbetweenlow-levelandhigh-levelareasofhumanvisualcortex.Proc.Natl.Acad.Sci.U.S.A.109,14675–14680. PubmedAbstract|PubmedFullText|CrossRefFullText Alink,A.,Schwiedrzik,C.M.,Kohler,A.,Singer,W.,andMuckli,L.(2010).Stimuluspredictabilityreducesresponsesinprimaryvisualcortex.J.Neurosci.30,2960–2966. PubmedAbstract|PubmedFullText|CrossRefFullText Awh,E.,Belopolsky,A.V.,andTheeuwes,J.(2012).Top-downversusbottom-upattentionalcontrol:afailedtheoreticaldichotomy.TrendsCogn.Sci.(Regul.Ed.)16,437–443. PubmedAbstract|PubmedFullText|CrossRefFullText Bacon-Mace,N.,Mace,M.J.,Fabre-Thorpe,M.,andThorpe,S.J.(2005).Thetimecourseofvisualprocessing:backwardmaskingandnaturalscenecategorisation.VisionRes.45,1459–1469. PubmedAbstract|PubmedFullText|CrossRefFullText Bao,M.,Yang,L.,Rios,C.,He,B.,andEngel,S.A.(2010).Perceptuallearningincreasesthestrengthoftheearliestsignalsinvisualcortex.J.Neurosci.30,15080–15084. PubmedAbstract|PubmedFullText|CrossRefFullText Bar,M.(2009).Theproactivebrain:memoryforpredictions.Philos.Trans.R.Soc.Lond.BBiol.Sci.364,1235–1243. PubmedAbstract|PubmedFullText|CrossRefFullText Barlow,H.B.(1997).Theknowledgeusedinvisionandwhereitcomesfrom.Philos.Trans.R.Soc.Lond.BBiol.Sci.352,1141–1147. PubmedAbstract|PubmedFullText|CrossRefFullText Barlow,H.B.(2001).Redundancyreductionrevisited.Network12,241–253. PubmedAbstract|PubmedFullText|CrossRefFullText Barlow,H.B.(1985).“Cerebralcortexasmodelbuilder,”inModelsoftheVisualCortex,edsD.RoseandV.G.Dobson(Chichester:JohnWiley&Sons),37–46. Bavelier,D.,Levi,D.M.,Li,R.W.,Dan,Y.,andHensch,T.K.(2010).Removingbrakesonadultbrainplasticity:frommoleculartobehavioralinterventions.J.Neurosci.30,14964–14971. PubmedAbstract|PubmedFullText|CrossRefFullText Bullier,J.(2001).Integratedmodelofvisualprocessing.BrainRes.Rev.36,96–107. CrossRefFullText Chun,M.M.,andJiang,Y.(1998).Contextualcueing:implicitlearningandmemoryofvisualcontextguidesspatialattention.Cogn.Psychol.36,28–71. PubmedAbstract|PubmedFullText|CrossRefFullText Clark,A.(inpress).Whatevernext?Predictivebrains,situatedagents,andthefutureofcognitivescience.Behav.BrainSci. Crist,R.E.,Li,W.,andGilbert,C.D.(2001).Learningtosee:experienceandattentioninprimaryvisualcortex.Nat.Neurosci.4,519–525. PubmedAbstract|PubmedFullText Dambacher,M.,Rolfs,M.,Göllner,K.,Kliegl,R.,andJacobs,A.M.(2009).Event-relatedpotentialsrevealrapidverificationofpredictedvisualinput.PLoSONE4:e5047.doi:10.1371/journal.pone.0005047 CrossRefFullText deFockert,J.W.(2010).Earlytop-downattentionalmodulationinvisualprocessing.ActaPsychol.(Amst.)135,112–113. PubmedAbstract|PubmedFullText|CrossRefFullText Dehaene,S.,Naccache,L.,Cohen,L.,Bihan,D.L.,Mangin,J.-F.,Poline,J.-B.,etal.(2001).Cerebralmechanismsofwordmaskingandunconsciousrepetitionpriming.Nat.Neurosci.4,752–758. PubmedAbstract|PubmedFullText|CrossRefFullText DiLollo,V.,Enns,J.T.,andRensink,R.A.(2000).Competitionforconsciousnessamongvisualevents:thepsychophysicsofreentrantvisualprocesses.J.Exp.Psychol.Gen.129,481–507. PubmedAbstract|PubmedFullText|CrossRefFullText Egeth,H.E.,Leonard,C.J.,andLeber,A.B.(2010).Whysalienceisnotenough:reflectionsontop-downselectioninvision.ActaPsychol.(Amst.)135,130–132. PubmedAbstract|PubmedFullText|CrossRefFullText Eimer,M.,andKiss,M.(2010).Thetop-downcontrolofvisualselectionandhowitislinkedtotheN2pccomponent.ActaPsychol.(Amst.)135,100–102. PubmedAbstract|PubmedFullText|CrossRefFullText Engel,A.K.,Fries,P.,andSinger,W.(2001).Dynamicpredictions:oscillationsandsynchronyintop-downprocessing.Nat.Rev.Neurosci.2,704–716. PubmedAbstract|PubmedFullText|CrossRefFullText Felleman,D.J.,andVanEssen,D.C.(1991).Distributedhierarchicalprocessingintheprimatecerebralcortex.Cereb.Cortex1,1–47. PubmedAbstract|PubmedFullText|CrossRefFullText Folk,C.L.,andRemington,R.(2010).Acriticalevaluationofthedisengagementhypothesis.ActaPsychol.(Amst.)135,103–105. PubmedAbstract|PubmedFullText|CrossRefFullText Friston,K.(2005).Atheoryofcorticalresponses.Philos.Trans.R.Soc.Lond.BBiol.Sci.360,815–836. PubmedAbstract|PubmedFullText|CrossRefFullText Friston,K.(2009).Thefree-energy,principle:aroughguidetothebrain?TrendsCogn.Sci.(Regul.Ed.)13,293–301. PubmedAbstract|PubmedFullText|CrossRefFullText Friston,K.(2010).Thefree-energyprinciple:aunifiedbraintheory?Nat.Rev.Neurosci.11,127–138. PubmedAbstract|PubmedFullText|CrossRefFullText Grill-Spector,K.,Henson,R.,andMartin,A.(2006).Repetitionandthebrain:neuralmodelsofstimulus-specificeffects.TrendsCogn.Sci.(Regul.Ed.)10,14–23. PubmedAbstract|PubmedFullText|CrossRefFullText Grossberg,S.(2009).Corticalandsubcorticalpredictivedynamicsandlearningduringperception,cognition,emotionandaction.Philos.Trans.R.Soc.Lond.BBiol.Sci.364,1223–1234. PubmedAbstract|PubmedFullText|CrossRefFullText Heekeren,H.R.,Marrett,S.,andUngerleider,L.G.(2008).Theneuralsystemsthatmediatehumanperceptualdecisionmaking.Nat.Rev.Neurosci.9,467–479. PubmedAbstract|PubmedFullText|CrossRefFullText Hegdé,J.,andFelleman,D.J.(2007).Reappraisingthefunctionalimplicationsoftheprimatevisualanatomicalhierarchy.Neuroscientist13,416–421. PubmedAbstract|PubmedFullText|CrossRefFullText Henson,R.N.(2003).Neuroimagingstudiesofpriming.Prog.Neurobiol.70,53–81. PubmedAbstract|PubmedFullText|CrossRefFullText Hesselmann,G.,Sadaghiani,S.,Friston,K.J.,andKleinschmidt,A.(2010).Predictivecodingorevidenceaccumulation?Falseinferenceandneuronalfluctuations.PLoSONE5:e9926.doi:10.1371/journal.pone.0009926 CrossRefFullText Hosoya,T.,Baccus,S.A.,andMeister,M.(2005).Dynamicpredictivecodingbytheretina.Nature436,71–77. PubmedAbstract|PubmedFullText|CrossRefFullText Kastner,S.,Pinsk,M.A.,DeWeerd,P.,Desimone,R.,andUngerleider,L.G.(1999).Increasedactivityinhumanvisualcortexduringdirectedattentionintheabsenceofvisualstimulation.Neuron22,751–761. PubmedAbstract|PubmedFullText|CrossRefFullText Kersten,D.,Mamassian,P.,andYuille,A.(2004).ObjectperceptionasBayesianinference.Annu.Rev.Psychol.55,271–304. PubmedAbstract|PubmedFullText|CrossRefFullText Kiefer,M.,andMartens,U.(2010).Attentionalsensitizationofunconsciouscognition:tasksetsmodulatesubsequentmaskedsemanticpriming.J.Exp.Psychol.Gen.139,464–489. PubmedAbstract|PubmedFullText|CrossRefFullText Kinchla,R.A.,andWolfe,J.M.(1979).Theorderofvisualprocessing:“top-down,”“bottom-up,”or“middle-out”.Percept.Psychophys.25,225–231. PubmedAbstract|PubmedFullText|CrossRefFullText Kristjánsson,Á.(2010).Priminginvisualsearch:aspannerintheworksforTheeuwes’sbottom-upattentionsweeps?ActaPsychol.(Amst.)135,114–116. PubmedAbstract|PubmedFullText|CrossRefFullText Kristjánsson,Á.,andCampana,G.(2010).Whereperceptionmeetsmemory:areviewofrepetitionpriminginvisualsearchtasks.Atten.Percept.Psychophys.72,5–18. PubmedAbstract|PubmedFullText|CrossRefFullText Kristjansson,Á.,Ingvarsdottir,A.,andTeitsdottir,U.D.(2008).Object-andfeature-basedpriminginvisualsearch.Psychon.Bull.Rev.15,378–384. PubmedAbstract|PubmedFullText|CrossRefFullText Kuhn,B.,Denk,W.,andBruno,R.M.(2008).Invivotwo-photonvoltage-sensitivedyeimagingrevealstop-downcontrolofcorticallayers1and2duringwakefulness.Proc.Natl.Acad.Sci.U.S.A.105,7588–7593. PubmedAbstract|PubmedFullText|CrossRefFullText Li,W.,Piech,V.,andGilbert,C.D.(2004).Perceptuallearningandtop-downinfluencesinprimaryvisualcortex.Nat.Neurosci.7,651–657. PubmedAbstract|PubmedFullText|CrossRefFullText Maljkovic,V.,andNakayama,K.(1994).Primingofpop-out:I.Roleoffeatures.Mem.Cognit.22,657–672. PubmedAbstract|PubmedFullText|CrossRefFullText Maljkovic,V.,andNakayama,K.(1996).Primingofpop-out:II.Theroleofposition.Percept.Psychophys.58,977–991. PubmedAbstract|PubmedFullText|CrossRefFullText Martínez,A.,Dirusso,F.,Anllo-Vento,L.,Sereno,M.I.,Buxton,R.B.,andHillyard,S.A.(2001).Puttingspatialattentiononthemap:timingandlocalizationofstimulusselectionprocessesinstriateandextrastriatevisualareas.VisionRes.41,1437–1457. PubmedAbstract|PubmedFullText|CrossRefFullText Melloni,L.,VanLeeuwen,S.,Alink,A.,andMuller,N.G.(2012).Interactionbetweenbottom-upsaliencyandtop-downcontrol:howsaliencymapsarecreatedinthehumanbrain.Cereb.Cortex22,2943–2952. PubmedAbstract|PubmedFullText|CrossRefFullText Moors,A.,andDeHouwer,J.(2006).Automaticity:atheoreticalandconceptualanalysis.Psychol.Bull.132,297–326. PubmedAbstract|PubmedFullText|CrossRefFullText Muckli,L.,Kohler,A.,Kriegeskorte,N.,andSinger,W.(2005).Primaryvisualcortexactivityalongtheapparent-motiontracereflectsillusoryperception.PLoSBiol.3:e265.doi:10.1371/journal.pbio.0030265 CrossRefFullText Müller,H.J.,Töllner,T.,Zehetleitner,M.,Geyer,T.,Rangelov,D.,andKrummenacher,J.(2010).Dimension-basedattentionmodulatesfeed-forwardvisualprocessing.ActaPsychol.(Amst.)135,117–122. PubmedAbstract|PubmedFullText|CrossRefFullText Navon,D.(1977).Forestbeforetrees:theprecedenceofglobalfeaturesinvisualperception.Cogn.Psychol.9,353–383. CrossRefFullText Neville,H.,andBavelier,D.(2002).Humanbrainplasticity:evidencefromsensorydeprivationandalteredlanguageexperience.Prog.BrainRes.138,177–188. PubmedAbstract|PubmedFullText|CrossRefFullText Palmer,S.E.(1999).VisionScience.Cambridge,MA:MITPress. Pascual-Leone,A.,andWalsh,V.(2001).Fastbackprojectionsfromthemotiontotheprimaryvisualareanecessaryforvisualawareness.Science292,510–512. PubmedAbstract|PubmedFullText|CrossRefFullText Pourtois,G.,Rauss,K.,Vuilleumier,P.,andSchwartz,S.(2008).Effectsofperceptuallearningonprimaryvisualcortexactivityinhumans.VisionRes.48,55–62. PubmedAbstract|PubmedFullText|CrossRefFullText Rao,R.P.N.,andBallard,D.H.(1999).Predictivecodinginthevisualcortex:afunctionalinterpretationofsomeextra-classicalreceptive-fieldeffects.Nat.Neurosci.2,79–87. PubmedAbstract|PubmedFullText|CrossRefFullText Rauschenberger,R.(2010).Reentrantprocessinginattentionalguidance–timetoabandonolddichotomies.ActaPsychol.(Amst.)135,109–111. PubmedAbstract|PubmedFullText|CrossRefFullText Reicher,G.M.(1969).Perceptualrecognitionasafunctionofmeaninfulnessofstimulusmaterial.J.Exp.Psychol.81,275–280. PubmedAbstract|PubmedFullText|CrossRefFullText Roland,P.E.,Hanazawa,A.,Undeman,C.,Eriksson,D.,Tompa,T.,Nakamura,H.,etal.(2006).Corticalfeedbackdepolarizationwaves:amechanismoftop-downinfluenceonearlyvisualareas.Proc.Natl.Acad.Sci.U.S.A.103,12586–12591. PubmedAbstract|PubmedFullText|CrossRefFullText Rousselet,G.A.,Thorpe,S.J.,andFabre-Thorpe,M.(2004).Howparallelisvisualprocessingintheventralpathway?TrendsCogn.Sci.(Regul.Ed.)8,363–370. PubmedAbstract|PubmedFullText|CrossRefFullText Schwartz,S.,Maquet,P.,andFrith,C.(2002).Neuralcorrelatesofperceptuallearning:afunctionalMRIstudyofvisualtexturediscrimination.Proc.Natl.Acad.Sci.U.S.A.99,17137–17142. PubmedAbstract|PubmedFullText|CrossRefFullText Serences,J.T.(2008).Value-basedmodulationsinhumanvisualcortex.Neuron60,1169–1181. PubmedAbstract|PubmedFullText|CrossRefFullText Srinivasan,M.V.,Laughlin,S.B.,andDubs,A.(1982).Predictivecoding:afreshviewofinhibitionintheretina.Proc.R.Soc.Lond.BBiol.Sci.216,427–459. PubmedAbstract|PubmedFullText|CrossRefFullText Summerfield,C.,andEgner,T.(2009).Expectation(andattention)invisualcognition.TrendsCogn.Sci.(Regul.Ed.)13,403–409. PubmedAbstract|PubmedFullText|CrossRefFullText Summerfield,C.,Egner,T.,Greene,M.,Koechlin,E.,Mangels,J.,andHirsch,J.(2006).Predictivecodesforforthcomingperceptioninthefrontalcortex.Science314,1311–1314. PubmedAbstract|PubmedFullText|CrossRefFullText Super,H.,Spekreijse,H.,andLamme,V.A.(2001).Twodistinctmodesofsensoryprocessingobservedinmonkeyprimaryvisualcortex(V1).Nat.Neurosci.4,304–310. PubmedAbstract|PubmedFullText|CrossRefFullText Theeuwes,J.(2010a).Top-downandbottom-upcontrolofvisualselection.ActaPsychol.(Amst.)135,77–99. CrossRefFullText Theeuwes,J.(2010b).Top-downandbottom-upcontrolofvisualselection:replytocommentaries.ActaPsychol.(Amst.)135,133–139. CrossRefFullText Treue,S.(2003).Visualattention:thewhere,what,howandwhyofsaliency.Curr.Opin.Neurobiol.13,428–432. PubmedAbstract|PubmedFullText|CrossRefFullText Tse,P.U.,andCavanagh,P.(2000).ChineseandAmericansseeoppositeapparentmotionsinaChinesecharacter.Cognition74,B27–B32. PubmedAbstract|PubmedFullText|CrossRefFullText Wagemans,J.,Elder,J.H.,Kubovy,M.,Palmer,S.E.,Peterson,M.A.,Singh,M.,etal.(2012a).AcenturyofGestaltpsychologyinvisualperception:I.Perceptualgroupingandfigure-groundorganization.Psychol.Bull.138,1172–1217. CrossRefFullText Wagemans,J.,Feldman,J.,Gepshtein,S.,Kimchi,R.,Pomerantz,J.R.,VanDerHelm,P.A.,etal.(2012b).AcenturyofGestaltpsychologyinvisualperception:II.Conceptualandtheoreticalfoundations.Psychol.Bull.138,1218–1252. CrossRefFullText Weidner,R.,Krummenacher,J.,Reimann,B.,Muller,H.J.,andFink,G.R.(2009).Sourcesoftop-downcontrolinvisualsearch.J.Cogn.Neurosci.21,2100–2113. PubmedAbstract|PubmedFullText|CrossRefFullText Wolfe,J.M.,Butcher,S.J.,Lee,C.,andHyle,M.(2003).Changingyourmind:onthecontributionsoftop-downandbottom-upguidanceinvisualsearchforfeaturesingletons.J.Exp.Psychol.Hum.Percept.Perform.29,483–502. PubmedAbstract|PubmedFullText|CrossRefFullText Keywords:bottom-up,predictivecoding,top-down,V1,vision Citation:RaussKandPourtoisG(2013)Whatisbottom-upandwhatistop-downinpredictivecoding?Front.Psychol.4:276.doi:10.3389/fpsyg.2013.00276 Received:26April2012;Accepted:28April2013;Publishedonline:17May2013. Editedby: LarsMuckli,UniversityofGlasgow,UK Reviewedby: SzonyaDurant,RoyalHolloway,UniversityofLondon,UKPeterNeri,UniversityofAberdeen,UK Copyright:©2013RaussandPourtois.Thisisanopen-accessarticledistributedunderthetermsoftheCreativeCommonsAttributionLicense,whichpermitsuse,distributionandreproductioninotherforums,providedtheoriginalauthorsandsourcearecreditedandsubjecttoanycopyrightnoticesconcerninganythird-partygraphicsetc. *Correspondence:KarstenRauss,InstituteofMedicalPsychologyandBehavioralNeurobiology,Otfried-Müller-Str.25,72076Tübingen,Germany.e-mail:[email protected] ThisarticleispartoftheResearchTopic Predictivecoding Viewall 11Articles Peoplealsolookedat Download