• Non ci sono risultati.

Invariant visual object recognition and shape processing in rats

N/A
N/A
Protected

Academic year: 2021

Condividi "Invariant visual object recognition and shape processing in rats"

Copied!
24
0
0

Testo completo

(1)

ContentslistsavailableatScienceDirect

Behavioural

Brain

Research

j o ur na l h o me p a g e :w w w . e l s e v i e r . c o m / l o c a t e / b b r

Review

Invariant

visual

object

recognition

and

shape

processing

in

rats

Davide

Zoccolan

VisualNeuroscienceLab,InternationalSchoolforAdvancedStudies(SISSA),34136Trieste,Italy

h

i

g

h

l

i

g

h

t

s

•Ratsarecapableofinvariantvisualobjectrecognition.

•Ratsspontaneouslyperceivedifferentviewsofavisualobjectassimilartoeachother,thatisasinstancesofthesameobject. •Ratsarecapableofamultifeatural,shape-basedvisualprocessingstrategy.

•Ratscanlearncomplex,configuralvisualdiscriminations.

•Ratsspontaneouslyprocesscompositevisualpatternsaccordingtoperceptualgroupingcues.

a

r

t

i

c

l

e

i

n

f

o

Articlehistory:

Received21June2014 Receivedinrevisedform 19December2014 Accepted25December2014 Availableonline2January2015 Keywords: Invariantrecognition Rat Rodent Shapeprocessing Patternvision

a

b

s

t

r

a

c

t

Invariantvisualobjectrecognitionistheabilitytorecognizevisualobjectsdespitethevastlydifferent imagesthateachobjectcanprojectontotheretinaduringnaturalvision,dependingonitspositionand sizewithinthevisualfield,itsorientationrelativetotheviewer,etc.Achievinginvariantrecognition representssuchaformidablecomputationalchallengethatisoftenassumedtobeauniquehallmarkof primatevision.Historically,thishaslimitedtheinvasiveinvestigationofitsneuronalunderpinningsto monkeystudies,inspiteofthenarrowrangeofexperimentalapproachesthattheseanimalmodelsallow. Meanwhile,rodentshavebeenlargelyneglectedasmodelsofobjectvision,becauseofthewidespread beliefthattheyareincapableofadvancedvisualprocessing.However,thepowerfularrayof experimen-taltoolsthathavebeendevelopedtodissectneuronalcircuitsinrodentshasmadethesespeciesvery attractivetovisionscientiststoo,promotinganewtideofstudiesthathavestartedtosystematically explorevisualfunctionsinratsandmice.Rats,inparticular,havebeenthesubjectsofseveralbehavioral studies,aimedatassessinghowadvancedobjectrecognitionandshapeprocessingisinthisspecies. Here,Ireviewtheserecentinvestigations,aswellasearlierstudiesofratpatternvision,toprovidean historicaloverviewandacriticalsummaryofthestatusoftheknowledgeaboutratobjectvision.The pictureemergingfromthissurveyisveryencouragingwithregardtothepossibilityofusingratsas complementarymodelstomonkeysinthestudyofhigher-levelvision.

©2015TheAuthor.PublishedbyElsevierB.V.ThisisanopenaccessarticleundertheCCBYlicense (http://creativecommons.org/licenses/by/4.0/).

Contents

1. Introduction... 11

2. Visualornotvisual?... 11

2.1. Therediscoveryofrodentvision ... 12

3. Earlyinvestigationsofratpatternvision ... 12

3.1. Earliestevidenceofinvariantrecognitionandadvancedshapeprocessinginrats... 12

3.2. Gestaltvisualperceptioninrats... 13

3.3. Strengthsandlimitationsofearlyworksonratvisualperception ... 14

∗ Correspondenceto:InternationalSchoolforAdvancedStudies(SISSA),ViaBonomea,265,34136Trieste(TS),Italy.Tel.:+390403787701;fax:+390403787702. E-mailaddress:zoccolan@sissa.it

http://dx.doi.org/10.1016/j.bbr.2014.12.053

(2)

3.3.1. Limitationsofspatialtwo-alternativeforced-choicetasks... 14

3.3.2. Limitationsinthetestofgeneralization... 15

4. Patternvisionandvisualshapeprocessinginrats... 15

4.1. Insearchofthecriticalstimulusdimensionsunderlyingratpatternvision... 15

4.2. Evidenceofshape-basedvisualobjectrecognitioninrats... 16

4.3. Spatialintegrationofvisualinformationinrats... 17

4.3.1. Configuralvisualdiscriminationinrats... 18

4.3.2. Perceptualgroupinginrats... 18

4.4. Areratstrulycapableofadvancedshapeprocessing?... 19

5. Invariantvisualobjectrecognitionandadvancedshapeprocessinginrats... 20

5.1. Ratsarecapableoftransformation-tolerant(invariant)visualobjectrecognition... 21

5.1.1. Visualprimingasawaytoassesstransformation-tolerantperceptionofvisualobjectsinrats... 23

5.2. Visualprocessingstrategiesunderlyingratobjectvision... 23

5.2.1. Aflexible,mid-levelstrategyunderliesratdiscriminationofsimplegeometricalshapes... 24

5.2.2. Ashape-based,multifeaturalprocessingstrategyunderliesratinvariantvisualobjectrecognition ... 26

5.2.3. Reconcilingfindingsaboutthecomplexityofvisualobjectprocessinginrats... 26

5.3. Ratprocessingofcollinearvisualfeatures ... 28

6. Conclusionsandperspectives... 28

Acknowledgments ... 29

References... 29

1. Introduction

Thegreatestchallengethatanyvisionsystem(whether biolog-icalorartificial)hastofaceistheextractionofthebehaviorally relevantcontentofvisualscenesfromthelargelyvariableimages collected, as an input, by the retina [1–3]. Within thedomain of object recognition, this challenge manifests as the need of recognizingpreviouslyseenobjects,despitethefactthatanynew encounterwiththeseobjectswilloftenresultinradicallydifferent retinalinputimages.Theprocessofrecognizingobjectsinspiteof thissubstantialvariationintheirappearanceiscommonlyknown asinvariantrecognitionor,moreprecisely,transformation-tolerant recognition[3–6].Quiteobviously,theproblemofunderstanding howthebrainachievesinvariantrecognitionistightlylinkedtothe problemofunderstandinghowthebrainprocessesandrepresents theshapeofvisualobjects.Althoughshapeisnotawell-defined concept, intuitively, shape features are those properties of the visualobjectsthatmorereliablyallowdistinguishingeachobject fromanyotheronein theretinalinputimages. Assuch,global visualattributes that are not very diagnosticof object identity (becausetheyarelikelysharedbymanydifferentobjects),such asoverallbrightness,contrast,size,area,etc.,arecommonly con-sideredaslower-levelvisualproperties,ratherthanshapefeatures (andaprocessingstrategyrelyingonsuchfeaturesiscommonly consideredasalower-levelstrategy).Ontheotherhand,specific arrangementsof localstraight and curved boundaries, oriented edges and corners [7–10], as well as oriented and unoriented localcontrastpatterns[11–13]andimagefragments[14,15](i.e., attributesthatmoreuniquelydefinevisualobjects)aretypically viewedasproper,higher-level shape features (anda processing strategy relying onsuchfeatures is commonlyconsidered as a higher-levelstrategy).Bearinginmindthatarigorous(or agreed-upon)definitionofshapefeaturecomplexitydoesnotactuallyexist (e.g.,see[16] for adiscussion), onecould tentativelydefinean advancedvisualprocessingsystemasfollowing–asystem capa-bleofextractingfromtheretinalinputthosehigher-order(shape) features that are diagnostic of object identity acrossthe many (unpredictable)imagetransformationseachobjectundergoes dur-ingnaturalvision.Throughoutthearticle,termssuchaslower-level properties,higher-levelfeaturesandadvancedprocessingstrategywill beusedaccordingtothedefinitionsprovidedabove.

Thesedefinitionsmakeveryexplicitthechallengeatthecore ofobjectvision–therequirementofmaintainingahighdegreeof specificityforthefeaturesdefiningtheidentityofvisualobjects, while,at thesametime,disregarding thehugevariation inthe

appearanceofsuchdiagnosticfeatures[17].Computationally, bal-ancingthistrade-offbetweenspecificityandinvariancemakesthe developmentofmachinevisionsystemsextraordinarychallenging [4,5,17,18].Ontheotherhand,humansarecapableofdetecting andclassifyingobjectsoutoftensofthousandsofpossibilities[19] withinafractionofasecond[20,21],whichimpliesthattheirvisual systemimplementsanextremelyefficientandreliablemachinery toprocessobjectinformation.Studiesinnonhumanprimateshave revealedthatsuchaprocessingiscarried outalonga hierarchy of visualcorticalareasknownastheventral visualpathway (or stream)[4,6,22–25].Thispathwaystartsinprimaryvisualcortex andculminatesintheanteriorpartoftheinferotemporalcortex, whichconveysthemostexplicitrepresentationsofvisualobjects– thatis,therepresentationsthatallowtheeasiestread-outofobject identity(andgeneralizationtoidentity-preservingimage transfor-mations)usingsimplelineardecoders[16,26–28].Unfortunately, theneuronal mechanismsunderlyingthe formation of increas-inglyexplicitrepresentationsofvisualobjectsalongtheventral streamarestillpoorlyunderstood.Thisislikelyaconsequenceof theextraordinarycomplexityoftheprimatevisualsystem[29–31] andofthelimitedrangeofexperimentalmanipulationsthat pri-matestudiesallowatthemolecular,synaptic,andcircuitrylevels. Thisdisproportionbetweenthecomplexityoftheneuronal archi-tectureunderinvestigationandthelimitedarrayofexperimental approachesthat areavailabletocarryoutsuchaninvestigation hasrecentlyledseveralvisionscientiststoexplorethepotential of rodentsasmodels ofvisualfunctions.Withregard toobject recognition,whetherrodentscanactuallyserveasusefulmodelsto understanditsneuronalbasiscruciallydependsonhowadvanced theirvisualrecognitionbehavioris.Inthefollowing,Iwillcritically reviewmostoftheliteratureconcernedwithratpatternvision, shapeprocessingandvisualobjectrecognition,arguingthatthere isenough behavioralevidencetosuggestthat thisspecies does embodysomeofthecoremechanismsunderlyinginvariantobject recognition.Assuch,theratand,byhomology,themouse(given thesimilaritybetweenthevisualsystemsofthesespecies[32–35]) mayrepresentapowerfulcomplementarymodeltothenonhuman primateintheinvasiveinvestigationoftheneuronalsubstratesof objectvision.

2. Visualornotvisual?

Rats and mice are the more widespread laboratory animal species,accountingforover 80%ofallresearchanimalsusedin theEuropeanUnion[36].Vision,ontheotherhand,isarguably

(3)

oneofthemainresearchtopicsinneuroscience,targetedby disci-plinesasdiverseassystem,cognitive,computational,cellularand developmentalneuroscience [37,38]. Giventhese premises,one couldexpectrodent-basedstudiestobeattheforefrontofvision research,but,historically,thishasnotbeenthecase.Infact,until veryrecently,rodentshavebeenlargelyoverlookedbythevision sciencecommunity,becausetheirbrainshavebeenassumedtolack advancedvisualprocessingmachinery.Suchanassumptionrests ontheobservationthatratsandmicearenocturnal/crepuscular species[36],havemuchlowervisualacuitythanprimates(e.g.,∼1 cycle/deginpigmentedrats[39–46],comparedto30–60cycles/deg inhumanandmacaquefovea[47–53]),andmakeextensiveuseof othersensorymodalities,suchastouch[54–56]andsmell[57–61], whenexploringandinteractingwiththeirenvironment.In addi-tion,neurons inrodent primary visualcortex, while displaying manyofthetuningpropertiesfoundinhighermammalianspecies (e.g.,orientationtuning[62–74]),arenotarrangedintothe func-tionalcorticalmodules,suchastheorientationcolumns[68,73,75], whicharetypicalofnon-humanprimatesandsmallcarnivores. Historically,this hasconfined theresearch onrodent vision to developmentalstudiesofcorticalplasticitymechanisms[76–80] (e.g.,oculardominanceplasticityduringthecriticalperiod)and behavioral/lesionstudiesoflearning,familiarityandmemory(e.g., theanatomical substrates of object recognition memory; for a reviewsee[81–85]).Inotherwords,thevisualabilitiesofratsand micehavebeenthesubjectofinvestigationnotbecauseofan inter-estintorodentvisualprocessingperse,butonlyinsofarasthese speciesprovidedaneasieraccesstotheneuronal/anatomical sub-stratesofplasticityandmemory,ascomparedtomonkeyandcat cortex.

Atthesametime,someofthemostcompelling,although indi-rect,evidenceabouttheabilityofrodentstoextractbehaviorally relevantinformation from thevisualenvironment and perform non-trivialcomputationsonthevisualinput,canbefoundin non-visionstudies.Infact,rodentshavebeenextensivelystudiedfor theirimpressive spatial navigationabilities, typically displayed whenforagingoverlargerangesaroundtheirnests[86],alsoat dusk,whenvisualcuesare availableandcanplaya keyrole in theirorientingbehavior.Laboratoryrats,inparticular,havebeen found to make not only excellent use of visual cues [87–90], buttopreferentiallyrelyonsuchcues(overolfactoryand audi-torycuesand pathintegration[91–93]), when testedin spatial navigation tasks. At the neuronal level, the most remarkable signature of such a strong dependence of rat spatial behavior onvision is the locking of hippocampal place fields to spatial visualcues, and theirremapping when suchcues are changed [94–99].Aspointedoutinarecentreview[1],althoughthevisual processing implied in navigation cannot simply be equated to theone underlyingobject recognition,still itmustengage sim-ilarlyadvancedmechanisms interms of extractionof ascene’s content (e.g., visual landmarks and their relative spatial loca-tion)from a largely viewpoint- and lighting-dependent retinal input.

2.1. Therediscoveryofrodentvision

Thenotionthatrodentscanactuallyserveasusefulmodelsto studyvision hasonly recentlytaken holdin thevision science community,witha newtide of studiesthat, over thepast few years,havestartedtosystematicallyexplorevisualfunctionsinrats andmice.Theturningpointintheappreciationofrodent mod-elsbyvisionscientistscanbeascribedtotwomainfactors.The firstistheevergrowingarrayofpowerfulexperimentaltools(e.g., optogenetics[100–102]andinvivotwo-photonimaging[75,103]) thathavebeendevelopedtodissectneuronalcircuitsinrodents [104]and,whilepossibletoapplyinotherspecies,arecurrently

far less advanced in (e.g.) non-human primates. The second is the increasingappreciation for rodent cognitiveabilities [105], whichismakingthemattractivemodelstostudycognitive func-tionsascomplex asperceptual decision-making [106,107],rule learning[108]and workingmemory[109].While theaccessto geneticmanipulationhasinspiredaconsiderablenumberof neu-roanatomy[32,33], imaging [73,110,111] and electrophysiology [72,112–115]studiesofmousevisualcortex(culminatingwiththe endowment, bythe AllenInstitutefor Brain Science,of a $300 million budget toa 10-year projectto study themouse visual system[116,117]), thepossibility of testing highly manageable animalmodelsincomplexvisualtasks,suchasshapeprocessing andinvariantobjectrecognition,hasfosteredthebehavioral inves-tigation of rat visual abilities [118–128]. Several reviews have summarized themost recentadvances in theunderstandingof themouse visualsystem [129–131].In this report,I willfocus insteadonthemountingbehavioralevidenceaboutratcapability foradvancedprocessingofvisualshape/objectinformation (Sec-tion5).Thisdiscussionwillbeintroducedbyanhistoricaloverview ofearlyinvestigationsofratvision(Section3),uptomorerecent studiesofshapeprocessingandobjectrecognitioninthisspecies (Section4).

3. Earlyinvestigationsofratpatternvision

Thefirstbehavioralstudiesonratvisiondatebacktoone cen-turyago.Someoftheseearlyinvestigatorsfoundeitherveryscarce [132]orno[133,134]evidenceforpatternvision inalbinorats, whilesomeothersfoundthatwhiteratswerecapableofform dis-crimination(e.g.,discriminationofupwardvs.invertedtriangles [135,136]).Readingthisearlydebateaboutratvisualcapabilities isveryinstructive,becausemanyoftheissuestheseearlyauthors discussedarethesamelaterstudiesalsofacedwhentestingrat visualbehavior:(1) thechoiceof theratstrain (albino vs. pig-mented);(2)thedesign ofthetestingapparatus(e.g.,imposing, ornot,afixedviewingdistance);(3)thepossiblecontributionof othermodalities(e.g.,tactileandolfactory)inexperimentswhere ratsareallowedtoapproachphysicalshapes/objects;and(4)the confoundsduetolow-levelvisualproperties,suchasbrightness and sizeofthe stimuli.Thisdebatewaseventually settledin a seriesofseminalstudiespublishedbyKarlLashley[137,138]and otherauthors[139–145]inthe30s.Lashley,inparticular,using aningeniousjumpingstandapparatus,notonlyobtainedthefirst behavioralmeasureofvisualacuityinpigmentedandalbinorats [43] (laterconfirmedby many studies[39–42,44–46]), butalso providedthemostcompellingevidenceofhistimeaboutthe abil-ityof pigmentedratstosuccessfully discriminatevisualshapes [137,138].

3.1. Earliestevidenceofinvariantrecognitionandadvanced shapeprocessinginrats

Thejumpingstandapparatus(showninFig.1A)implemented a spatial two-alternative forced-choice procedure, with a rat requiredtojumpagainstatarget(positive)stimuluscardfroma distanceof20cm,whileignoringaflankingnegativestimuluscard. Thecardbearingthenegativestimuluswasfixedrigidly,causing therat tofallinto anet foran incorrectjumptoward it (pun-ishment),whilethecardbearingthepositivestimuluswasheld byalightspring,sothatwhena ratjumpedagainstit,thecard fellbackandallowedtheanimaltoreachalandingplatformwith food(reward).Usingthisapproach,Lashleyfoundthatpigmented ratswere able to discriminatedifferent pairs of 2-dimensional geometricalshapes,suchasanupwardvs. aninvertedtriangle, atrianglevs.a circle,a trianglevs.across,etc.[137,138].More

(4)

Fig.1.ExtractsfromLashley’sseminalstudyofratpatternvision.(A)AsketchofthejumpingstandapparatusintroducedbyLashleytotestratsinvisualdiscrimination tasks(seeSection3.1foradescription).(B)Anexampleofvisualpatterndiscrimination,inwhichratsweretrainedinoneofLashley’sexperiments.(C)Alterationsofthe trainedvisualpatterns(showninB)toprobegeneralizationofratrecognitionintransfertrials.(Forinterpretationofthereferencestocolorinthisfigurelegend,thereader isreferredtothewebversionofthisarticle.)

ModifiedfromLashley[138].

importantly,hediscoveredthatonceananimalhadlearnedto dis-criminatetwoshapes,hewasabletogeneralizethisdiscrimination “inspiteofalterationinsize,continuityofsurface,orofoutline[of theshapesandwas]abletodiscoverthefigureinvarious combina-tionswithirrelevantlines”[138](seeFig.1BandC).Inaddition,he alsofoundthat“markedchangesintheluminousintensityofeither figureorground[did]notdisturbreactionsolongasthebrightness relationsoffigureandground[were]notreversed”[138](e.g.,see Fig.1C,toprow).ThesefindingsledLashleytoconcludethatin rats“theremustbesomeprimitivegeneralizationofformwhich goesbeyondtherecognitionofidenticalelements”[138],which iswhat moderninvestigatorswouldrefer toas transformation-invariantor,moreproperly,transformation-tolerantvisualobject recognition(namely,size-tolerant,clutter-tolerantand luminance-tolerantrecognition).Ontheotherhand,Lashleyfoundonlylimited tolerancetoin-planerotation(e.g.,seeFig.1C, centralpanelin thebottomrow)and,byobscuringpartsofthevisualtargets,he found“thatthemajorityofanimals[did]notreacttotheentire figurepresentedinthestimuluspattern.Thelowerorinner mar-ginofthefiguresmostfrequentlydetermine[d]thereaction.[...] Animalsreact[ed]tosuchvarious characters asthe distanceof onefigurefromtheframe,therelative surfaceareasof the fig-ures,a conspicuousprojectingpoint,andthelike”[138].Atthe sametime,Lashleyfoundthatratscouldlearntodiscriminate dis-continuousorientedpatternsmadeofdistinctelements(e.g.,two circlesdefiningeitheradiagonalorahorizontal“line”)andthat, moreimportantly,thisdiscriminationgeneralizedtocontinuous orientedlineswiththesameorientationofthepreviouslylearned discontinuouspatterns. The reversewas alsotrue: rats trained todiscriminatebetweenaverticalandahorizontalgrating(i.e., patternsmadeofmultiple,alternatingblackandwhitestripes, ori-entedeitherverticallyorhorizontally)generalizedwellnotonlyto singlecontinuouslineswiththesameorientation,butalsoto pat-ternsmadeofsingleormultiplediscontinuouslineswiththesame orientation(i.e.,patternsmadeofeitherverticallyorhorizontally alignedarraysofsmallsquares/rectangles).Thesefindingsinduced Lashleytoconcludethat“theconstantstowhich[arat]reactsin otherwisevariablesituationarepropertieswhichcanbederived onlyfromatotalfigure,hencethathisreactionisdependentupon somesortofunificationoftheelementswithinapartofthefield” [138].

3.2. Gestaltvisualperceptioninrats

Thesecontrastingobservations(i.e.,thelackof“react[ion]to theentire[stimulus]figure”insometestsandtheprocessingof “properties...derived...fromatotalfigure”insomeothertests [138])anticipatedcontemporaryattemptsatdeterminingthe crit-icalvisualfeaturesunderlyingratdiscriminationofvisualpatterns, andtheensuingdebateaboutwhetherratsareactuallycapableof processingshapeinformationacrossalargespanofanobject’s sur-faceor,instead,simplyrelyonlocal,lower-levelimageproperties [120,121,146,147](seeSections4.2,4.4and5.2).Inthelanguage ofLashley’stime,thesefindingscouldbeinterpretedasevidence against or in favor of rat ability to process the total organiza-tionof visualpatterns,by relyingonthespontaneousgrouping ofsensoryelementsintowholeformpercepts,aspostulatedby Gestaltpsychology.Thistopicwasthesubjectofaseriesof stud-ies by two of Lashley’s contemporaries:Paul Fieldand Isadore Krechevsky,concerned,respectively,withtheroleofconcept for-mation [140,141,145]and perceptual grouping[143,144] in rat visualperception.

Inparticular,Krechevsky[143,144]carriedoutaseriesof exper-imentsinwhich,usingLashley’sjumpingstand,ratsweretrained todiscriminatebetweentwoarraysmadeofthesamenumberof identicalconstituentelements(smallsquares),whoseproximity waslargeralongtheverticaldimensioninonearray(thus form-ingapatternofcolumns)andalongthehorizontaldimensionin theotherarray(thusformingapatternofrows).Notonlydidrats learntoperformthisdiscriminationveryeffectively(thusshowing acapabilitytoprocessthespatialdistributionoftheconstituent elementsof thetrained patterns), but,when tested in general-ization trials, theyshowed a preference for a grating made of continuouslineswiththesameorientationoftheoriginallylearned training-positivearrayoverthepositivearrayitself.Forinstance, ratstrainedtorespondto(i.e.,jump toward)apattern ofrows (the positivestimulus; seeFig.2A,right)and avoid amatching patternofcolumns(thenegativestimulus;seeFig.2A,left),when presentedwiththesamepatternofrowsandanever-seen-before horizontalgrating(seeFig.2B),spontaneouslychosethesecond overthefirstinthelargemajorityofthetransfer/generalization trials.Interestingly,suchapreferenceforthecontinuousgratings wasnotobservedincasesinwhichratshadbeenoriginallytrained

(5)

Fig.2.ThevisualpatternsusedbyKrechevskyinhisseminalstudyofperceptual groupinginrats.(A)Ratswereinitiallytrainedtodiscriminateanegative(−) col-umnpatternfromapositive(+)rowpattern.(B)Whenrequiredtochoosebetween thetraining-positiverowpatternandacontinuoushorizontalgrating,mostrats consistentlychosethelatter.

ModifiedfromKrechevsky[143,144].

todiscriminateapositivecolumn(orrow)patternfroma nega-tivepatternmadeofasmallernumberofelements,arrangedso astoformanX[143].Basedonthesefindings,Krechevsky con-cludedthatrat“perceptualprocess[...]involvestheoperationof ‘forcesofattraction’betweenthemembersofavisualgroupofsuch anatureastomaketherattopreferthecontinuousgestaltover thatofthediscontinuousgropingeventhoughtheoriginal train-ing[was]onthediscontinuousstimulus-complex”[143].However, theemergenceofsuchapreferenceforthecontinuousgestaltwas notfullyspontaneous,i.e.,purelyrootedintothe“autochthonous forcesoforganization”[143]ofthestimuluspatterns,butdepended ontheneed oftheanimal to extractsucha gestalt to success-fullyaccomplish the task. Hence, thefailure of the continuous gestalttoemergewhenthecolumn/rowpatternshadtobe dis-criminatedfromtheXpatterns,wherealower-levelrecognition strategy(e.g.,onebasedoncomparingtheoverallbrightnessofthe patterns)wouldsufficetosucceedinthetask.Translatedinto con-temporaryterms,Krechevsky’sfindings(laterconfirmedbyseveral studies[148–153];seeSection4.3.2)canbeconsideredasthefirst solidevidenceofratabilitytointegratetheconstituentfeatures ofvisualpatternsintoconfiguralrepresentations,uptothepoint ofextractingfromthetrainingpatternsabstractshapedimensions, suchasverticalnessandhorizontalness.Perhapsmoreimportantly, Krechevsky’sinterpretationoftheimpactoftaskcomplexityonthe emergenceofgestaltperceptsanticipatedtheconclusionsreached bythemostrecentstudiesofratobjectrecognitionaboutthecrucial roleoftaskdemands/constraintsindeterminingthecomplexityof ratrecognitionstrategy[118,120].Ontheotherhand,Krechevsky’s conclusionsareobviouslylimitedbyhislackofknowledgeabout theexistenceoforientation-selectivefiltersinratprimaryvisual cortex(V1)[62–66,74].InSection4.3.2,Iwilldiscusstowhatextent

thetuningpropertiesofV1neuronscanaccountforKrechevsky’s findings,aswellasformorerecentstudiesofperceptualgrouping inrats.

3.3. Strengthsandlimitationsofearlyworksonratvisual perception

Overall,theearlystudiesbyLashleyand hiscontemporaries indicateacapabilityofratstousetheirvisiontodiscriminatevisual patternsandgeneralizetovariationsofpreviouslylearned discrim-inations(e.g.,duetosizechangesofthetargetstimulioraddition ofbackgroundclutter).Thesefindingssuggestthatratsareable to,atleasttosomedegree,abstracttheirrecognitionfromthe spe-cificretinalinputpatternsproducedbytheirencounterswithvisual objects.Assuch,Lashley’sandrelatedworkcanbeconsideredasthe firstdemonstrationthatratsarecapableofanon-trivialprocessing strategyofvisualpatterns,i.e.,astrategythatisnotsimplybased onmatchingidentical(ornearlyidentical) retinalinputs across consecutiveencounters.However,oneshouldbecarefulto inter-pretthesefindingsasconclusiveevidenceforratinvariantobject recognition.

Infact,Lashley’sstudieswereaffectedbyalackofsystematicity –althoughalargenumberofexperimentswereperformed,with manydifferentvisualstimuliandmanipulations,onlyafewratsper conditionweretestedandasmallnumberoftrialswascollectedper animal.Inaddition,onlyasmallnumberoftransformationsofthe trainedvisualpatternswastested,andacrossalimitedvariation range.Krechevsky’sexperimentsweremorefocused,more thor-oughintermsofthenumberoftestedratsandcollectedtrialsper animal,and,assuch,statisticallymorerigorous.Still,Krechevsky’s conclusionswerealsolimitedbythelittlevariationofthepatterns testedinhisexperiments(e.g.,thespatialfrequencyandphaseof thegratingswerenotvaried).Morecritically,alltheseearlystudies basedonLashley’sjumpingstandwereaffectedbytwo method-ologicalissuesthatlimitedthescopeoftheirconclusions,asfar asinvariantrecognitionisconcerned.Itisworthdiscussingthese methodologicallimitations,sincetheycanbefoundinmanystudies ofratvision,includingsomeveryrecentones[121,146].

3.3.1. Limitationsofspatialtwo-alternativeforced-choicetasks ThemajorissueinLashley’sandrelatedworkistheuseofa spatialtwo-alternativeforced-choiceprocedure.Byconstruction, thiskindoftaskrequiresasubjecttodiscriminatebetweentwo simultaneouslypresentedvisualpatterns.Asaconsequence, dis-criminationcanbebasedonthedirectcomparisonoflower-level visual attributes (e.g., local luminance or contrast), whose rel-ativemagnitudeis, generally,preservedacrosstransformations, especiallywhenbothpatternsareequallytransformed,asinthe caseofLashley’sstudy.Asanexample,ratdiscriminationofa tri-angle froma square could bebased oncomparingthe relative luminanceofthebottomhalvesofthesepatterns,without nec-essarilyprocessinganyactualshapeattribute(e.g.,orientationof edges,presenceandsharpnessofcorners,etc.).Ifthesepatternsare equallyscaledortranslated,ortheiroverallluminosityisequally reduced/increased,theratwillstillcorrectlydiscriminatethem, sincetherelative brightnessoftheirbottomhalveswillremain unchanged(thesquare’sbottomhalfwillalwaysbebrighterthan thetriangle’sbottomhalf), defactoacting asa transformation-invariant,low-level,diagnosticimagefeature.Sucharecognition strategy,whiledefinitelymoresophisticatedthanasimple match-ing of retinalinputs to fixedtemplates, shouldanywaynot be consideredasanadvancedprocessingstrategy,becauseitwould notengagehigher-order,transformation-tolerantrepresentations ofshapefeatures.Lashley’swork,aswellasotherratvision stud-iesbasedontwo-alternativeforced-choicetaskswithidentically transformedstimuli[121,146],cannotdifferentiatebetweensuch

(6)

alow-levelstrategyandatrulyinvariant,shape-basedprocessingof visualobjects.Thisdoesnotmeanthat,ingeneral,two-alternative forced-choiceprocedurescannotbeusedtostudyshapeprocessing (see,asanexample,Simpson’sandGaffan’swork[147],discussed inSection4.2).Alsoinvariantrecognitioncan,inprinciple,betested usingtheseprocedures,provided thatthestimulishown tothe animals are independently transformed in each trial.However, this isvery impractical,becauseofthecombinatorial explosion thatwouldresultfrombuildingeverypossiblepairoftransformed stimuli(especiallywhenthestimuliaretransformedalongmany differentdimensions–size,position,in-planerotation, etc.).As showninSection5,amorepracticalandeffectivesolutionisto useataskwhereasingle,isolatedstimulusispresentedineach trial[118–120].

3.3.2. Limitationsinthetestofgeneralization

Anotherissue, when probinginvariant object recognitionin animalsubjects,istotest generalizationof recognitiontonovel viewsofpreviouslylearnedobjects,withoutgivinganyfeedback abouttheidentityoftheseviews(bymeansofrewardand/or pun-ishment).Thisisobviously verychallenging,because,ideally, it wouldrequireananimaltoperformlargenumbersofunrewarded trials.Howeverimpractical,failuretowithholdfeedbackimplies aninabilitytotestpuregeneralizationofrecognition,sincethe testedanimalscould,inprinciple, learnand memorizethe cor-rectassociationbetweennewobjectviewsandreward,without necessarilyperceivingsuchnewviewsassimilartotheonesthey originallylearned.Inotherwords,asetofperceptuallyunrelated (fortheanimals)objectviewscouldbearbitrarilyassociated,in memory, tothesamecategory, onlybecauseof thecontinuous trainingreceivedduringthetask. Lashley’sapproach toaddress thisissue,soastotestgeneralizationofpattern discrimination, wasactuallywelldesigned,giventhetechnologicallimitsofhis time.Aftertrainingarattodiscriminateapairofvisualpatterns (usingrewardand punishment asdescribed above), theanimal wastested ina series of“critical”(i.e.,generalization)trials, in whichhewas“allowedafreechoicebetween[transformed ver-sionsofthepatterns],receivingfoodateverytrial,regardlessof thepattern chosen”[138].Usingthesepunishment-freetrialsis obviouslyagoodapproximationtoapuretestofgeneralization, yetis notcompletelyrobust, asacknowledgedbyLashely him-self[138]andKrechevsky[143].Sinceananimalreceivesreward nomatterwhatstimulushechooses,hecould,inprinciple,learn toconsistently respondto(i.e.,jump toward) a noveltest pat-tern,basedonthepositivefeedbackhereceivesonthefirst(or thefirstfew) trials.Thatis,a rat thathappens torespond cor-rectlyon thefirst trial, purely by chance,would be reinforced inhisinitial(random)choicebytherewardhereceives.Onthe otherhand, as reportedby Lashley,an animal responding ran-domly,wheninitiallyconfrontedwithtwonovelpatterns,could learn“thatnochoiceisnecessarywhennewfiguresarepresented andpromptly[adopt]apositionhabit[...]evenwithfigureswhich inearliertestsgave perfecttransfer” [138].Theseshortcomings donotfullyunderminethevalidityofLashley’stransfertests,but couldhave ledto eitheran overestimationor an underestima-tionofthegeneralizationability ofratsthatwere,respectively, eitherinitiallysuccessfulorunsuccessfulwithnewtestpatterns. Moreingeneral,theapproachoftestinggeneralizationby reward-ingwhateverchoiceananimalmakesonnovelstimuli,whilenot perfect,isacceptable,aslongasdatafrommultipleanimalsare pooledtoobtaingroupaverageperformances(inwhichpossible overestimations and underestimations are averaged out). Nev-ertheless, as shown in Section 5, novel approaches have been recentlyproposedthatallowacleanerassessmentofgeneralization [118,119].

4. Patternvisionandvisualshapeprocessinginrats

Theencouragingfindingsaboutratvisualprocessingabilities providedbyLashleyandhiscontemporariesinthe30sdidnot suc-ceedintriggeringasustainedinterestinthisspecies(oranyother rodentspecies)amongvision scientists.Intheensuingdecades, thevisionsciencecommunitybecameprogressivelymore inter-estedinmonkeys andsmallcarnivores(e.g.,cats) asmodelsto studythevisualsystem,andvirtuallyabandonedanysystematic attemptatstudyingvisualprocessinginrodents.Still,evenbefore theveryrecentsurgeofinterestinrodentvision(seeSection5),a fewbehavioralstudiesfurtherexploredtheperceptualstrategies underlyingratpatternvisionandobjectrecognition.Amongthese studies,themostnoteworthyarethosepublishedbySutherland andcolleaguesinthe60s[154–157]andbyGaffanandcolleagues inthe90s[147,158–160].

4.1. Insearchofthecriticalstimulusdimensionsunderlyingrat patternvision

Sutherlanddevelopedanewapparatustotrainratsinvisual dis-criminationtasks[154–156].Itconsistedinattachingtotherear wallofanoperantboxtwoshapescutoutofwhitePerspex(the discriminanda),withadrinkingnozzleforrewarddelivery protrud-ingthroughthemiddleofeachshape.Ratsweretrainedtoleavea startingbox(locatedattheoppositeendofthewallbearingthe dis-criminanda)andapproachafixedpositivepattern(whileignoring theothernegativepattern)tocollectreward.Usingthis appara-tus,Sutherlandsuccessfullytrainedpigmentedratstodiscriminate betweenaverticalandahorizontalrectangle[154].More impor-tantly,healsorantransfer(i.e.,generalization)testsinwhichrats werepresentedwithalteredversionsofthelearnedpatternsand wererewardedindependentlyoftheirchoice(i.e.,sameapproach totestgeneralizationasLashley’s;seeSection3.3.2).Thesetests confirmedand extended Lashley’searlyreports aboutrat capa-bilitytodiscriminatevisualshapesinspiteofsizevariation(i.e., bothlargerandsmallerversionsoftherectangles)andchangesto theirsurface/outline(i.e.,breakingtherectanglesinto discontinu-oussegments).

Inanotherseriesofstudies[155,156],Sutherlandand cowork-ers tried to infer the critical shape dimensions underlying rat discrimination of 2-dimensional visual patterns. They initially trained rats to discriminate between an “open” and a “close” shape(respectively,a diamond-likeandahourglass-like silhou-ette,eitherhorizontallyorverticallyoriented),andthenmeasured ratperceivedsimilaritybetweeneachof20newtransfershapes andtheoriginallytrainedshapes.Thesetransfertestswererunnot onlyusingthespatialtwo-alternativeforced-choicetaskdescribed above,butalsobypresentingasingleshapeatthetimeinthewall bearingthediscriminanda(leavingtheotherstimulusslotinthe wallempty,withjustthedrinkingnozzleprotrudingfromthebare wall).Again,beingthesetransfertests,ratswererewardedno mat-terwhethertheychosetoapproachthenozzleprotrudingfromthe testshapeorfromtheemptyslot.Noticeably,thisisoneofthe earliestattemptstoovercomethelimitationsoftwo-alternative forced-choiceprocedures(seediscussioninSection3.3.1), soas toallowamorereliableassessmentoftheperceptualsimilarity betweenthetransfershapesandtheoriginallylearned diamond-likeandhourglass-likepatterns.Theshapesthatratsperceivedas beingmoresimilartothediamond-likepatternwerethosethat, qualitatively,lookedmoreas“closed”(i.e.,morecompactand with-outprotruding elements),suchas circles,ellipses,trianglesand squares.Bycontrast,theshapesthatratsperceivedasbeingmore similartothehourglass-lookingpatternwerethose thatlooked more“open”(i.e.,morespread-out,withalotofcontourand pro-jections), suchascrosses, Hs,Us,etc.Asa waytoquantifythis

(7)

tendencyofratstorankandclassifythetransfershapesalong a “closeness–openness”stimulusdimension,theauthorscomputed theratiobetweencontourlengthandsquarerootoftheareafor eachshapeandfoundastrong,positivecorrelationbetweenthe rankingoftheshapesaccordingtotheirperceivedsimilaritytothe diamond-like(“closed”)shapeandtheirrankingaccordingtothis ratio.Sincethisratioprovidesawaytoquantifythecomplexity(in termsofopenness/closeness,i.e.,presenceorabsenceofprotruding features)of2-dimensionalpatterns(withmorecomplex, spread-outshapeshavingalongercontourthansimpler,morecompact shapes,given thesamearea), Sutherland’sand coauthors’ find-ingssuggestacapabilityofratstoabstracttheirrecognitionfrom thespecificretinalinputpatternstheywere originallyexposed to,soastobuildquiteexplicitrepresentationsofabstractshape dimensions.This conclusionis in agreementwith Krechevsky’s [143,144]andrelatedstudies[148–153]aboutratabilitytolearn thediagnosticdimensionsofverticalnessandhorizontalnessfrom discontinuousorientedpatterns bymeansofperceptual group-ing(see Sections3.2 and 4.3.2), and witha more recent study [123]aboutratcapabilityofcategorizing naturalimagesonthe baseofglobalstimulusdimensions,suchasconvexityandaspect ratio.

Overall,Sutherland’s studies suggest that therat visual sys-temiscapableofthekindofadvancedshape processingthatis thehallmark of high-level vision:the extrapolation of abstract shape dimensions. However, once again, care should be taken beforeconsideringthesefindingsasconclusiveevidencefor invari-ant object recognition in rats, because of several limitations in Sutherland’s and coworkers’ experimental approach (mostly pointed out bythe authorsthemselves). First, only few shapes weretestedandveryfewtrialswerecollectedpershape(10trials, inthetransferexperimentwithsingle-shapepresentation),thus greatlylimitingtheaccuracyofthemeasurementofratrecognition performanceand,therefore, ofratperceivedsimilaritybetween transfer and trained shapes. In addition, the single-shape pre-sentationdevisedbySutherland,whileovercomingsomeofthe issuesconcerningtwo-alternativeforced-choicetasks(seeSection 3.3.1),wasitselfaffectedbyonemainlimitation:theratsshowed aconsistentbiastoward thesideonwhichtheshape was pre-sented(nomatterwhethertheyweresupposedtoapproachitor avoidit),thus furtherreducingthepowertoestimaterat pref-erence/perceivedsimilarity[156].Morecritically,itisdifficultto understandtowhatextentotherlower-levelstimulusproperties mayhaveplayedaroleindeterminingratsimilarityjudgments. Many different feature dimensions could,potentially, correlate withtheshape–complexityratiocomputedin[156]:shapearea, averageluminosity,overallluminosity,luminosity/contrastinthe bottom/tophalvesof the shapes,etc. As shown in Section 5.2, classification image methods are arguably the bestapproaches touncoverwhatvisualfeaturesaratreliesuponwhen perform-inganobjectdiscriminationtaskand,therefore,revealthenature andcomplexityofratrecognitionstrategy[120,121].Ontheother hand,ensuringthatratsdonottriviallydiscriminatevisualobjects becauseofdifferencesinlow-levelpropertiescanonlybeachieved intwoways:either bytestingratswithavery largebattery of objecttransformationsorbycarefullymatchingasmanydifferent low-levelvisualattributesaspossible.Whiletheformerapproach wasappliedinthemostrecentstudiesofratinvariantrecognition [118–120],thelatterwasfollowedbySimpsonandGaffanintheir verythoroughinvestigationofvisualsceneandobjectprocessing inrats[147].

4.2. Evidenceofshape-basedvisualobjectrecognitioninrats Simpson’s and Gaffan’s study [147] was based on a proce-dure/apparatustotestratsinvisualtasksthatGaffanandcolleagues

haddescribedinpreviousreports[158,159],andthattheylater applied in a series of lesion studies to identify the neuronal substrates of visual object discrimination and memory in rats [160–164].TheapparatusconsistedofaYmaze,withthethreearms havingequallengthandthethreeanglesformedbyeacharmpair havingequalamplitude.Apairofcomputermonitors(jointedbya 117◦angle,soastosubtend∼94◦horizontally,whenviewedfrom

themazecenter)waslocatedattheendofeacharm,wherereward wasalsodelivered.Visualscenesofvariouscomplexity(containing anumberof2-dimensionalobjects,suchasellipses,ASCIItext char-acters,polygons,rectangles,etc.)weredisplayedonthemonitors, withthescenesshownintheadjacentmonitorsattheendofanarm beingthemirrorversionofeachotheralongtheverticalaxis(each pairofmirrorscenesonadjacentmonitorsconstitutedone stimu-lusscene;seeexamplesinFig.3A).Eachratwastrainedinaspatial two-alternativeforced-choicetask,inwhichtwodifferent stimu-lusscenesweredisplayedattheendofthetwoarmsnotcurrently occupiedbytheanimal,andtherathadtolearntoeitherapproacha targetpositive-constantoravoidatargetnegative-constantscene. Thisconstantscenewaspaired,ineverytrial,toatrial-unique vari-ablescene,containingvisualobjectsthatcoulddifferfromthoseof theconstantsceneinavarietyofproperties(examplesofaconstant sceneandsomevariablescenesthatwerepairedtoitinSimpson’s andGaffan’sstudy[147]areshowninFig.3A).

Usingthisapproach,SimpsonandGaffanshowedthatratswere abletodiscriminatepairsofstimulusscenescontainingtwovisual objectspermonitor,evenwhentheseobjectsoccupiedthesame locationwithinthevisualspace(i.e.,withineachscene),thus con-cludingthat rats“mustencodemorethanjust thegrossspatial distributionofpatches(objects)thatcontrastwiththebackground” [147].Moreimportantly,theyfoundthatratsmaintainedahigh dis-criminationperformance(above80%correct)nomatterwhether position-matchedobjects in thestimulusscenes were approxi-matelyequatedintermsofarea,luminance(i.e.,graylevel,since eachobjectwasrenderedinauniformgrayintensityvalue)and luminousflux(i.e.,area×luminance).Thesethreelow-level fea-tureswerematchedindependentlyinseparatetests,thusshowing thatnoneofthemwascritical,byitself,forthesuccessful discrimi-nationofthevisualscenes(comparebarsinFig.3B).Whenallthese propertiesweresimultaneouslymatched(thusproducingstimulus scenescontainingobjectswithsimilararea,luminanceand lumi-nousflux)ratsstillsuccessfullydiscriminatedthescenes(compare barsinFig.3C),thussuggestingthattheybasedtheirrecognitionon shapeinformation.Thisconclusionwasconfirmedbyrequiringrats todiscriminatescenescontainingobjectsofthesameshape cate-goryinmatchingvisualfieldlocations(e.g.,ellipseswithdifferent aspectratioand/orslightsizedifferences;comparetheconstant scenetothelastvariablesceneinFig.3A).Thisproducedadrop inratperformance(seemiddlebarinFig.3D)thatledGaffanand Simpsontoconclude“first,thatratsperceivedtheobjectsas dis-tinctshapes,notasluminousblobs,andsecond,thattheyencoded themin termsof visualfeaturesthat arecommontoa classof shape”[147].Inaddition,whenratsweregiventhechancetobase theirdiscriminationoneithershapedifferencesorsizedifferences amongthevisualobjectscontainedinapairofstimulusscenes, theyappearedtorelyonlyontheformer(higher-level)property, whileignoringthelatter(lower-level)attribute.Thatis,rat per-formancewasequallygoodwhentheobjectsdifferedinshape(no matterwhethertheyalsodifferedinsize;seepurplebarsinFig.3E), andequallyworsewhentheobjectshadsimilarshape(nomatter whethersizedifferenceswerealsopresentthatcouldbeusedas low-levelcues;seebluebarsinFig.3E).

Overall,Gaffan’sandSimpson’sstudyprovidescompelling evi-denceaboutthecapabilityof ratstodiscriminatevisualobjects onthebaseofshapedifferences.Althoughnoattemptwasdone toinfer what shape features theanimals relied upon (as done

(8)

Fig.3.SummaryofSimpson’sandGaffan’sexperimentaldesignandresults.(A)ExamplesofstimulusscenesusedtoproberatpatternvisioninSimpson’sandGaffan’s study.Ineachtrial,ratswererewardedforavoidinganegative-constantscene(rightpanel)andapproachingatrial-specificvariablescene(examplesshownintheleft panels).Theobjectsinthevariablescenescouldbematchedtothoseintheconstantscenewithrespecttodifferentvisualproperties:luminance,area,luminousflux(i.e., area×luminance),andshapeclass.(B–E)Ratrecognitionperformancefordifferentkindsoftestedvariablescenes.Thecolorofthebarsmatchesthecolorofthecorresponding examplevariablescenesin(A).SeeSection4.2foradetaileddescription.(Forinterpretationofthereferencestocolorinthisfigurelegend,thereaderisreferredtotheweb versionofthisarticle.)

ModifiedfromSimpsonandGaffan[147].

bySutherlandandcolleagues[156]andby morerecentstudies [120,121,123]),thesystematic(althoughpartial)matchingof vari-ouslow-levelpropertiesconvincinglyshowsthatratscanrecognize visualpatternswithoutresortingtogrossdifferencesinarea, lumi-nosity and luminous flux. In addition, theindependence of rat performanceonthearea/sizeoftheobjectscanbetakenasa fur-therdemonstrationofthesizetoleranceofratobjectrecognition. However,tolerancetoparametricobjectvariationsalong differ-enttransformationaxeswasnotsystematicallyinvestigated,thus leavingopenthequestionofhowmuchvariabilityinobject appear-ance,andalongwhatdimensions,ratsareabletotolerate.Finally, theexperimentsofGaffanandSimpsonwerestillaffectedbythe limitationsinherentinspatialtwo-alternativeforced-choicetasks (seeSection3.3.1),althoughtoalesserextentthanotherstudies, becauseofthelargevarietyofshapesagainstwhicheachlearned targetobjecthadtobecompared(e.g.,comparetheobjectsinthe constantscenetothoseinthevariablescenesofFig.3A),which madeitmoreunlikelyfortheratstorelyonlocallow-levelcues (e.g.,luminance differences in the top orbottom halves ofthe objects,aslaterreportedbyMininiandJeffery[146];seeSection 4.4).

Gaffan and colleagues further investigated rat visual object recognitioninaseriesoflesionstudiesaimedatassessingtheroleof perirhinalcortexinobjectrecognitionmemoryandobject identifi-cation[160–164].Giventhatthisreviewismainlyconcernedwith behavioralinvestigationsofratvisualobjectrecognition,Iwillnot discusstheneuroanatomicalimplicationsof thesestudies(fora review,see[81–85]).Still,leavingasidetheneuroanatomical find-ings,someofthesestudiesprovideadditionalbehavioralevidence aboutratvisualprocessingabilities.Forinstance,in[160],ratswere testedinahighlydemandingtask,whichrequiredthemto discrim-inatetheconstantscenefromvariablescenesthatdifferedfromthe constantinfourpossibleways:(1)bothidentityandpositionofthe objectswithinthescenes;(2)onlyobjectidentity(sameposition);

(3)onlyobjectposition(sameidentity);and(4)arrangementof objectswithfixedidentityacrossfixedpositions.Sincethesefour typesofscenecomparisonswererandomlyinterleaved,thefact thatcontrol(i.e.,non-lesioned)animalsachievedanaverage per-formanceof∼70%correct(orgreater)ineachcomparisonsuggests thatratsarecapableofahighlyflexible recognitionstrategy,in whichpositionandshapecuescanbeinterchangeablyextracted, dependingontaskdemands[160].

4.3. Spatialintegrationofvisualinformationinrats

Many other lesion studies of rat recognition memory have reporteddataaboutobjectdiscriminationbehaviorinthisspecies. Unfortunately,mostofthesestudieshaveusedsolidobjectsthat ratswerefreetoapproachand touch,ratherthan purelyvisual stimuli(e.g.,see[165–167]).Inaddition,recognitionperformance hasbeentypicallymeasuredintermsofthetimespentbyan ani-malexploringagivenobject,ratherthanasthefractionofexplicit correct identifications he made. This makes extremely hard to infer anythingaboutrat visualprocessing,because,under such experimentalsettings,notonlythevisualpropertiesoftheobjects (e.g.,size,brightness,etc.)arenotcontrolled,butratscaninspect thediscriminandausingamixtureofvisual,tactileandolfactory cues[168,169].Morerecently,severalinvestigatorshavedeveloped approachesthat,similarlytotheonedevisedbyGaffanand col-leagues,allowtestingratsinpurelyvisualtasks.Theseapproaches arebasedeitheronV-shapedwatermazesequippedwithmonitors forstimuluspresentation[170–172]oronoperantboxesequipped withtouchscreens[146,173–177].Still,mostofthesestudiesare notmuchconcernedwithratvisualprocessingperse,but,rather, ratvisualbehaviorisexploitedasameantotestandunderstand learningandmemoryfunctions.Assuch,thesestudies,while pro-vidingadditionalevidenceofratcapabilitytousevisiontoextract task-relevantinformation fromvisualpatterns, do nottypically

(9)

Fig.4. Visualpatternsusedthetestconfiguraldiscriminationinrats.(A)Thestimuli usedbyEacottetal.[161],whichdidnotenforceastrictlyconfiguralprocessing strategy,becauseoftheuniquefeaturescreatedbytheintersectionoftheir con-stituentelements.(B)ThestimuliusedbyDaviesetal.[172],whichmoreproperly enforcedaconfiguralprocessingstrategy.

PanelAismodifiedfromEacottetal.[161].PanelBismodifiedfromDaviesetal. [172].

includemanipulations ofthevisualstimuli that allowinferring howadvancedratobjectvisionisintermsofshapeprocessingand invariantrecognition.

4.3.1. Configuralvisualdiscriminationinrats

Exceptionstothisruleareafewreportsabouttheabilityofrats tolearncomplexconfiguralvisualdiscriminations[161,171,172], thatisdiscriminationsthatrequiretheconcurrentprocessingof multiplefeatures(orparts)ofthevisualstimuli.Themostwidely appliedofsuchtasksisthebiconditionaldiscrimination,inwhich fourdifferentvisualfeatures(i.e.,featuresA,B,CandD)are com-binedtoobtainfourdifferentcompoundstimuli(i.e.,compound patternsAB,AC,BC,andCD)thatcanonlybedistinguishedfrom eachotherbyconcurrentlyconsideringbothconstituentelements ofeachpattern. Whileinmonkey studiesthebiconditional dis-criminationistypicallyimplementedbypresentingonecompound patternsatthetimeandthenaskingthemonkeysubjecttoreport itsidentitybymeansofsaccadesorleverresponses[178,179],in thecontextofratstudiesthisparadigmhasbeenimplementedin theformofmultipletwo-alternativeforced-choicetasks(i.e.,AB+ vs.AC−,DC+vs.DB−,AB+vs.DB−andDC+vs.AC−,withthe+ denotingtherewarded pattern),administeredtotheanimalsin interleavedtrialsorsmallblocksoftrials[161,172].

Thefirstattemptat trainingratsinsuchataskwasaffected byapossibleconfound:theconstituentelementswere superim-posedontopofeach othertoproducethecompoundpatterns, thuscreatinguniquefeaturesattheirintersectionsthataffordeda nonconfiguralsolution,inspiteofthesubstantialfeatures’overlap amongthediscriminanda(seeFig.4A)[161].Therefore,thefactthat ratssuccessfullylearnedthistasktellsmoreabouttheirremarkable abilitytoextractbehaviorallyrelevantvisualfeaturesamongvery

similarpatterns,thanabouttheircapabilitytolearnandadopta strictlyconfiguralprocessingstrategy.Alaterstudy[172]overcame thislimitationbydesigningcompoundconfiguralstimulimadeof flankingrectangularelements(eachbearingaspecificvisual pat-tern;seeFig.4B).Ratssuccessfullylearnedthistask,achievinga performancehigherthan70%correctinabout12trainingsessions, thusshowingthecapabilitytoconcurrentlyprocessmultiple con-stituentpartsof visualobjects.Thisconclusionissupportedby anotherstudy,inwhichratsweretestedinadifferentkindof con-figuraldiscrimination,knownastransversepatterning[171].The taskconsistedofthreesimultaneoustwo-alternativeforced-choice discriminations(i.e.,A+versusB−,B+versusC−,andC+versusA−), witheachofthethreestimuli(i.e.,A,BandC)appearingintwo dis-criminationsandbeingassociatedwithrewardinonecase(+)but notintheother(−).Similarlytothebiconditionaldiscrimination, transversepatterningalsorequiresthatsubjectsbuilda configu-ralrepresentationofstimuluselements.Thefactthatratslearned toperformthistaskwithhighaccuracy(>90%correct)confirms thattheycanreliablyacquireconjunctiverepresentationsofvisual patterns.

4.3.2. Perceptualgroupinginrats

Thatratscouldlearntointegratevisualinformationover mul-tiplelocationsofthevisualfieldwasalreadyamajorconclusionof Kreschevsky’sexperiments[143,144]–ratsdiscriminated discon-tinuousorientedpatternsmadeofidenticalconstituentelements, bygroupingsuchelementsaccordingtotheirproximity(see Sec-tion3.2andFig.2).Theseresultshavebeenconfirmedandextended bymorerecentstudies[148–153],thusfurtherdemonstratingthe abilityofthisspeciestointegratedistinctvisualfeaturesintoglobal percepts.Dodwell[148]wasthefirsttoreplicateKreschevsky’s findings,showingthattheyarenottheresultofanoveltyeffect, i.e.,theycannotbeexplainedbyratpreferenceforthelessfamiliar betweentwovisualpatterns.Infact,ratstrainedtochoosea posi-tivecolumnpatternoveranegativerowpattern,notonlypreferred alessfamiliarverticalgratingoverathetraining-positivecolumn pattern(asoriginallyshownbyKreschevsky),butalsopreferred themorefamiliartraining-negativerowpatternoveracontinuous horizontalgrating.Inotherwords,ratsperceivedaslessnegative theoriginallylearnednegative,discontinuouspattern,when com-paredtoanovelcontinuouspattern(thegrating)withthesame orientation.Stillmorestrikingwasthefindingofanother experi-ment,inwhichratstrainedtochoosebetweenapositivediagonal grating(D+)andanegativeverticalgrating(V−)werelatertested inadiagonal(D)vs.horizontal(H)gratingcomparison.Noticeably, someoftherats,ratherthanchoosingthetraining-positive stimu-lus(i.e.,D),chosethehorizontalgrating.Thatis,theyappearedto alwayschoosethelessverticalstimulus,whichwasD+inthecase ofthetrainedV−vs.D+comparison,butwasH,inthecaseofthe Dvs.Hgeneralizationtest.Similarresultswerefoundforrats orig-inallytrainedinaH+vs.D−comparisonandlatertestedinaDvs. Vcomparison,withsomeanimalschoosingthetraining-negative stimulusDoverV,asiftheirchoicewasbasedoncomparingthe horizontalnessofthestimuli,ratherthanavoidingarigidlylearned, negative,retinalinputpattern.

Dodwell and colleagues further explored the integration of discreteelementsintoglobalperceptsofverticalnessand horizon-talness[150],showingthatratscandiscriminateacolumnarray fromarowarray,evenwhentheseparationbetweenadjacent ele-mentsalongthehorizontaldimensionismadeveryclosetotheir separation alongtheverticaldimension(thusdrastically reduc-ingthestrengthoftheproximitycue)–columnandrowarrays witharatiobetweenthelargestandthesmallestelement separa-tionassmallas1.3:1couldstillbediscriminatedbyratswithan averageperformanceof∼74%correct.Thisisaninteresting find-ing,becauseitsuggeststhatratsmayintegratevisualinformation

(10)

overlargespansofthestimulusarrays,soastoaccumulate evi-denceaboutwhatdimension(i.e.,horizontalorvertical)hasthe shorterseparationbetweenadjacentpatternelements.However, thisstudyonlyprovidesalimitedassessmentofthedependence ofperceptualgroupingfromthepropertiesofthegroping cues, mainlybecauseitlacksincontroloverimportantstimulus proper-ties,suchasthesizeandseparation,indegreesofvisualangle,ofthe constituentelementsofthestimulusarrays.Amorerecentand rig-orousstudy[151]hasfoundthatthelargestelementseparationhad tobeatleasttwicethesmallerseparationforratstoperceiverow orcolumnpatternswithanaverageaccuracyofatleast75% cor-rect(althoughlaterstudiesbysomeofthesameauthorssuggesta somewhatsmallerseparationratiotoattainthesamecriterion per-formance[152,153]).Moreover,thestrengthoftheproximitycue thatwasnecessarytoproduceareliableperceptualgroupingwas foundtobestronglydependentontheoveralldensityofthevisual patterns–givenconstituentelementsof1.2◦ofvisualangle,the ratiobetweenthelargerandthesmallerelementseparationthat wasnecessarytoachievethecriterionperformanceincreasedfrom 2:1to4:1,whenthelargerelementseparationwasincreasedfrom 5.7◦to16.8◦ofvisualangle(thusmakingthevisualpatternsparser). Interestingly,inthisstudy,notonlyelementproximitybutalso elementalignmentwasusedasagroupingcue,withtheelements beingperfectlyalignedonly alongonedimension.Althoughthe alignmentcue,byitself,failedtoproduceanyperceptualgrouping, whenitwaspairedtotheproximitycue,itenhancedthegrouping producedbythelatter–forthesparsestpattern(seeabove),the ratiobetweenthelargerandthesmallerelementseparationthat wasnecessarytoachievecriteriondecreased to2.5:1,whenthe alignmentcuewasalsopresent.Asspeculatedbytheauthors,this findingsuggeststhat,inrats,notonlyproximitybutalso“pattern regularityinfluencestheperceptualorganizationofthestimulus” [151],althoughtoamuchlesserextentthaninhumans.

Thefindingspresentedinthissection,aswellasKrechevsky’s earlyinvestigationsonperceptualgrouping(seeSection3.2),could potentiallybeaccountedbythetuningpropertiesofratV1neurons. Asfoundinothermammalianspecies,ratprimaryvisualcortex containsneuronsactingasorientation-selectivefilters,spanning thefullspectrumofpossibleorientations[62–66,74].Manyofthese orientedfiltersperformanapproximatelylinearweightedsumof thelightintensitypatternsfallingintheirreceptivefields.Assuch, aratV1neuronwithagivenorientationpreference(e.g.,vertical) wouldbeactivatedbyadiscontinuousorientedpatternmatching itspreferredorientation.Suchanactivationwouldbemoreorless strong,dependingonthedensityofthepattern,withthe maxi-malresponseachievedforcontinuousorientedgratingsmatching thepreferredorientationandspatialfrequencyoftheneuron.Itis easytoseehowcertaindecisionstrategiesrelyingontheresponse strengthoftheseorientedfilterscouldlead(e.g.)toapreferencefor anovelcontinuousverticalgratingoverapreviouslylearned dis-continuouscolumnpattern(asoriginallyreportedbyKrechevsky [143]).Forinstance,thiswouldbethecase,ifratdecision strat-egyconsistedinchoosingthestimulusthatmaximallyactivated thebankofverticallyorientedfilters(i.e.,thegrating),ratherthan choosingthestimulusthatevokedapreviouslylearned,template responsepatternacrossthefilters’bank(i.e.,thecolumnpattern). Inotherwords,thespatiallyorientedreceptivefieldsofV1neurons, whencombinedwithapreferenceforthepatternsthatmaximize theactivityofagivensetofneurons,couldeasilyaccountforrat perceptualgrouping.Thisdoesnotimplythat Krechevsky’sand Dodwell’sfindingsaretrivial,becauseotherdecisionstrategiesare possiblethatwouldleadtoadifferentbehavioraloutcome(i.e.,a differentstimuluspreference),giventhesamepatternofactivation inV1(e.g.,atemplate-matchingstrategy,aspreviouslymentioned). ThebehavioralfindingsdescribedhereandinSection3.2strongly suggestthat,amongasetofencodinganddecodingstrategies,rats

prefertorelyonanabstractrepresentationoforientationthana precisematchingofstoredtemplatepatterns.Asdiscussedin Sec-tion5.2.3,thiscanhaveprofoundimplicationsintheassessment ofhowadvancedratpatternvision is.Finally,itisimportantto mentionthatnotallKrechevsky’sandDodwell’sfindingscanbe easilyaccountedbythemechanisticargumentsdiscussedabove. Forinstance,thelackofpreferenceforacontinuousgratingover adiscontinuouspatternwiththesameorientation,inthecasea ratwastrainedtodiscriminatethediscontinuouspatternfroman Xpattern[143],confirmsthatthewaytheV1representationofa visualpatternisreadout(toleadtodecision)isnotaunique, auto-maticprocess.Whetherorientationwillemergeasthediscriminant attributewillstronglydependontaskdemands.Similarly, Dod-well’sfindingaboutthepreferenceforahorizontalgratingovera diagonalgrating,forsomeoftheratsthatwereoriginallytrainedto preferadiagonalvs.averticalgrating[148],arguesforan idiosyn-cratic,rat-dependentdecodingschemeofV1responsepatterns–a schemethatcaneitherweightmoretheactivationofagiven sub-populationofneurons(e.g.,withdiagonalpreference),orweight moretheinactivationofanotherneuronalsubpopulation(e.g.,the cellspreferringverticalstimuli).

4.4. Areratstrulycapableofadvancedshapeprocessing?

Overall,thepictureemergingfromthestudiesreviewedinthe previoussectionsindicatesthatratsarecapableofaquiteadvanced processingofvisualobjects.Tosummarize,ratshavebeenfound ableto:(1)toleratesizevariationsandothertransformationsinthe appearanceofpreviouslylearnedpatterns;(2)discriminatevisual patterns with matched low-level visual properties; (3) flexibly extractpositionandshapecues,dependingontaskdemands;(4) learncomplexconfiguralvisualdiscriminations;(5)spontaneously processcompositevisualpatternsaccordingtoperceptual group-ingcues;and (6)extract abstractstimulusdimensions(suchas verticalnessandhorizontalness)fromtrainedpatterns.However, manyofthesefindingshavebeenrecentlychallengedbyastudyof MininiandJeffery[146],whoconcludedthatrats,whentestedin asimultaneous(i.e.,atwo-alternativeforced-choice) discrimina-tiontask,donotspontaneouslyrelyonashape-basedrecognition strategy,atleastwhenotherlower-levelvisualcuesareavailable tosuccessfullyaccomplishthetask.

MininiandJefferycarriedouttwodifferentexperiments,each requiringratstoapproachapositivetargetstimuluspresentedona touchscreen,whileavoidingasimultaneouslypresentednegative stimulus.Inthefirstexperiment,ratsweretrainedtodiscriminate betweenawhitesquareandawhiteequilateraltriangle(with ver-texdown)ofequalareaandequalluminance,shownsidebyside onthestimulusdisplay(i.e.,thetouchscreen)againstablack back-ground(seeFig.5A).Followingacquisitionofthediscrimination, ratsweretestedinaseriesoftransfertrials,where thetraining shapesweretransformedinavarietyofways.Ratsfailedto gen-eralizetoillusoryKanizsafiguresandtocontrast-reversedstimuli (thusconfirmingtheearlyobservationsofLashley[138]andField [140]),but succeeded atdiscriminatingthetarget shapeswhen theseweresimultaneouslyrotatedupto50–60◦,thussuggesting somedegreeofrotationinvariance.However,wheneach shape wasrotatedorverticallyshiftedofadifferentamountcompared totheothershape,soastoreversetheirbrightnessrelationshipin thelowerhemifieldofthestimulusdisplay,ratperformancefell tochanceorbelowit.Thiswasthecasewhenthetriangleandthe squarewererotated,respectively,of180◦and45◦,thusresultingin acomparisonbetweenatrianglewithvertexupandadiamond(see Fig.5B,firstbar),orwhenthesquareonlywasshiftedupward(see Fig.5B,secondbar).Asimilartrendwasfoundwhenthedifference betweenthebrightensofthetwoshapesinthelowerhemifield wasreduced,e.g.,whenonlythetrianglewasrotatedof180◦,so

(11)

Fig.5. SummaryofMinini’sandJeffery’sexperimentaldesignandresults.(A)The trianglevs.squarediscriminationinwhichratsweretrainedinoneofMinini’s andJeffery’sexperiments.(B)Ratdiscriminationperformancefordifferentkinds ofmanipulationsoftheoriginallylearnedpatterns(shownbelowtheabscissaaxis). Manipulationsalteringtherelativebrightnessofthestimuliinthelowerhemifield ofthestimulusdisplaymaderatperformancefallingtochance(dashedline)or belowit(seeSection4.4fordetails).

ModifiedfromMininiandJeffery[146].

astobringbothshapestohaveanhorizontaledgeinthelower hemifield(see Fig.5B,thirdbar). Ontheotherhand, manipula-tionsthatdidnotalterthebrightnessrelationshipofthestimuli inthelowerhemifield(e.g.,whenthetriangle onlywasshifted upward;seeFig.5B,fourthbar)producednodropin discrimina-tionperformance.Theseledtheauthorstoconcludethatrats“used, asthediscriminativecue,nottheconfigurationofedges,but dif-ferencesintheluminanceofthetwoshapesinthelowerregionof thevisualfield”[146].Whensuchadifferencereversed(compared towhat ratsexperiencedduringtraining withthedefault stim-ulusconfiguration,wherethebottomofthesquarewasbrighter thanthebottomofthetriangle;seeFig.5A),ratsfailedto discrim-inatetheshapesorevensystematicallymistookoneshapeforthe other(seeFig.5B,firstbar).However,theauthorsalsonoticedthat, whentheluminanceofthesquarewasreduced,soastomatchthe luminanceofthetriangleinthelowerhemifield,ratrecognition performancewasnotaffected(seeFig.5B,lastbar),thus suggest-ingthat“whattheratscomputedwasnotabsoluteluminanceper sesomuchastheareaofbrightpixels,irrespectiveoftheactual levelofbrightness”[146].

Obviously,thesefindingschallengetheviewthatratsare capa-ble of processing visual patters by relying primarily on shape features.Assuch,theyareatoddwithmanyearlierandlater find-ings, presented in thepreviousand next sections, aboutvisual shapeprocessinginrats,especiallythoseofGaffanandcolleagues [147,160](seeSection4.2)andZoccolanandcolleagues[118–120] (seeSections5.1 and 5.2).A waytoreconcile theseapparently conflictingconclusionsis toconsiderthecrucialrole thattask’s structure,constraintsanddemandslikelyhaveindeterminingthe complexityof thevisualrecognitionstrategyin rats,aswellas inother species(including nonhumanprimates; e.g.,see [180], discussedinSection5.2.3).Thattask’sstructurecouldaffectthe outcomeoftheirexperimentswasacknowledgedbyMininiand Jeffery,who pointedout how ratfailure to“discover shape”as thediagnosticstimulusdimension“mayreflecttheconstraintsof discriminationtasksratherthanalimitationinratvision”[146].

Specifically,inthecaseofMinini’sandJeffery’sexperiment,rats weretrainedinatwo-alternativeforced-choicediscriminationof

twogeometricalshapeswithfixedsize,positionandbrightness (thoseshowninFig.5A).Althoughduringthepre-training pro-cedurethesizeofthesquareandtriangleundergosomechanges insize[146],becauseofthenatureofthetwo-alternative forced-choicetask(seediscussioninSection3.3.1),suchsizevariations weresimultaneouslyappliedtoboththediscriminanda,thus pre-servingtheirrelativebrightnessinthelowerhemifield.Aspointed outbyZoccolanandcolleagues,underthesecircumstances, “com-putation of lower hemifield luminance was a perfectly ‘valid’ solutiontothetaskathand(i.e.,itwasoneofmanypossible strate-giesformaximizingrewardwithinthecontextoftheexperiment)” [118].Bycontrast,inGaffanandSimpsonexperiments[147],rats werepresentedineachtrialwithadifferentscenecomparison(i.e., aconstantnegativescenepairedtoatrial-uniquevariablescene; seeFig.3A),thusforcingtheanimalstorelyonthespecificshape oftheobjectscontainedinthescenes,ratherthanontheir low-levelproperties,suchasbrightnessorarea(seeFig.3C–E).Asafinal remark,itisworthnoticingthat,inspiteofthelackoftaskpressure touseshapeinformation,ratsinMinini’sandJeffery’sfirst exper-imentdidnotexclusivelyrelyonlower-levelvisualpropertiesto carryouttheirdiscrimination.Ifthiswasthecase,thena manipu-lationproducingadramaticreversalofthediscriminandarelative brightness,suchasshiftingupwardthesquare,shouldhave pro-ducedafullreversalofratstimuluspreference,while,infact,rat averageperformanceonlydroppedslightly(andnotsignificantly) belowchance(seeFig.5B,secondbar).Instead,suchastrong rever-salofthestimuluspreferencewasobservedonlyforthevertex-up trianglevs.diamondcomparison,i.e.,foramanipulationthat,in additiontochangingtherelativebrightensofthestimuli,alsomade thebottompartofthesquare(nowa diamond)very similarto theoriginallylearnedvertex-downtriangle,andthebottompartof thetriangle(nowahorizontaledge)verysimilartotheoriginally learnedsquare.Inotherwords,itlookslikeratsdidnotfullyignore shape–rather,theyreliedonastrategythatwaslargelybasedon abrightness/areacomparison,butwhereshapeinformation(for instance,edgeorientation)stillplayedarole.

Aspreviouslymentioned,MininiandJefferyalsocarriedouta secondexperiment,inwhichratswererequiredtodiscriminate asquarefromarectangle,basedonthedifferenceintheiraspect ratio.Theanimalseventuallylearnedthetask,butonlyaftermany sessions(∼70)andonlyreachedalow,albeitsignificantlyhigher thanchance,performance(∼65%correct).Thisledtheauthorsto concludethatratsarequitepooratdiscriminatingshapesbasedon suchanabstractmetricpropertyasaspectratio.Atthesametime, thefactthattheratseventuallylearnedthetasksuggestedtothe authors“thatratsmightbesomewhatcapableofusingaspectratio, andbyimplicationshape,tosolvevisualdiscriminations”[146].

5. Invariantvisualobjectrecognitionandadvancedshape processinginrats

Takentogether,thestudiesdiscussedinSection4,strongly sug-gest that ratsare capable of processingshape information and recognizingvisualobjectsinspiteofsomevariationintheir appear-ance (e.g., becauseof size and position changes). However,till veryrecently,conclusiveevidenceaboutinvariantrecognitionand advancedshapeprocessinginratswaslacking,becauseoffourmain limitations.First,toleranceofratrecognitiontotransformationsin objectappearancehadnotbeensystematicallytested.Thatis,none ofthestudiespresentedinSection4hadmeasuredthe depend-enceofratrecognitiononthetype oftransformation anobject mayundergo,aswellasonthemagnitudeofthetransformation alongdifferentvariationaxes(e.g.,size,position,in-planeand in-depthrotation,lighting,etc.).Thesecondlimitationisthatpure generalizationofratrecognitiontonovelobjectviewshadnever

(12)

Fig.6.BehavioralrigandexperimentaldesigndevelopedbyZoccolanetal.[118].(A)Apictureofthebehavioralrig(left),showingthreeoperantboxes,eachequippedwith monitorsforstimuluspresentation,computer-controlledpumpsforliquidrewarddelivery,andtouchsensorsforacquisitionofbehavioralresponses.Ratslearnedtoinsert theirheadthroughaviewinghole,locatedinthewallfacingthemonitor(bottom-rightpicture),andinteractwiththearrayofsensors(top-rightpicture)totriggerstimulus presentation,reportstimulusidentityandcollectthereward.ThisrigiscurrentlyusedinD.D.Cox’lab(Harvard)andD.Zoccolan’slab(SISSA).Asimilarhigh-throughputrig hasalsobeendevelopedbyP.Reinagel’slab(UCSD).(B)Schematicoftheobjectdiscriminationtask.Aftertriggeringstimuluspresentationbylickingthecentralsensor,arat hadtolickeithertherightorleftsensor,dependingonobjectidentity.

PanelBismodifiedfromAlemi-Neissietal.[120].

beenproperlytested,becauseofthemethodologicalchallengeof withholdinganyfeedbacktoratsaboutthecorrectnessof their responseingeneralization/transfertrials(seediscussioninSection 3.3.2).Crucially,failuretotestpuregeneralizationpreventsarobust assessmentofwhatcomponentofratinvariantrecognitioncanbe ascribedtoaspontaneous,purelyperceptualevaluationofvisual similarity(i.e.,toatruegeneralizationprocess),andwhat compo-nent,instead,dependsonlearningarbitraryassociativerelations amongtrainedobjectviews.Thethirdlimitationisthat,inspiteof themany(ofteningenious)stimulusmanipulationsappliedinthe studiesdescribedinSection4,nodirectassessmentofthevisual featuresuponwhichratsbasedtheirrecognitionwasobtainedin theseinvestigations.Thishaspreventeda fullunderstandingof thecomplexityofratvisualprocessingstrategy,leavingopenthe questionofwhetherthisspeciesistrulycapableofadvancedshape processing(seeSection4.4).Finally,withtheexceptionof percep-tualgrouping(seeSections3.2and4.3.2),nostudyhadtestedrats inthosebenchmarkpsychophysicsexperimentsthat,inhumans, havelinkedvisualperceptiontofundamentalpropertiesofvisual corticalprocessing,suchassurroundsuppressionanddivisive nor-malization[181,182].

InSections5.1–5.3,Iwillpresentrecentexperimentalworkthat hasstartedtoaddressmany of theseopenissues. Asdiscussed inSection6,thesenewfindingsareveryencouragingaboutthe possibilityofusingtheratasamodeltoinvestigatetheneuronal substratesofhigher-levelvision.

5.1. Ratsarecapableoftransformation-tolerant(invariant) visualobjectrecognition

Thefirstsystematicinvestigationofratinvariantrecognition wascarriedoutbyZoccolanandcolleagues[118]inastudywhere ratswerefirsttrained todiscriminatethedefault viewsof two visualobjects(seeFig.6BandFig.7A,greenframes),thenrequired totoleratesomevariationintheirappearance(seeFig.7A,light blueframes),andfinallytestedwithmanynoveltransformations ofthelearnedobjects(seeFig.7A,allremainingobjectconditions).

Inthisstudy,severalratsweretrainedinparallel,usinga high-throughputbehavioralrigthatconsistedofmultipleoperantboxes, eachequippedwithaviewinghole,threetouchsensors(alsoacting asjuicetubesforliquidrewarddelivery)andacomputermonitor (seeFig.6A).Theratslearnedtoinserttheirheadintotheviewing hole,soastofacethecomputermonitor(usedasthestimulus display),andinteractwiththesensorstotriggerstimulus presen-tation(centralsensor)andreporttheidentityofthevisualobjects (leftand rightsensors). Thisrig, and the stimuluspresentation design, alloweda few importantmethodological advancesover earlierbehavioralinvestigationsofrat patternvision.First,they allowedaverytightcontroloversomecrucialstimulusproperties, suchassize,becausethestimulusviewingdistancewasconstant andreproducibleacrosstrialsandsessions.Second,itwaspossible tocarry outa veryintensive, fast-pacetraining, withcollection ofupto500behavioraltrialspersessionpersubject.Third,and more importantly,theuseofthetouchsensorsallowedtesting ratswithoutresortingtothespatialtwo-alternativeforced-choice procedureusedinthelargemajorityofrodentvisionstudies–in anygiventrial,aratwaspresentedwitha singlestimulusonly, andhadtoreportitsidentitybyrememberingitsassociationto eithertheleftorrightsensor/rewardport (seeFig.6B).Thus,it waspossibletoovercomethelimitationofprobingratinvariant recognitionwithsimultaneouslypresentedshapes,whichmakesit extremelychallengingtopreventratsfromsolvingdiscrimination tasksusinglower-levelvisualfeatures(seediscussioninSection 3.3.1andbelow).Anothermajordifferencewithpreviousstudies wasthatthevisualstimuli,ratherthanbeingsimplegeometrical shapes/patterns,wererenderingsof3-dimensionalobjectmodels madeofmultiplestructuralparts(seeFigs.6Band7A).Thisallowed transformingtheobjectsalongnaturalvariationaxes,suchas in-depthrotationsandlightingchanges,whichareimpossibletotest with2-dimensionalshapes.Therefore,thevarietyandcomplexity ofthetransformationsthestimuliunderwentinthisstudywere muchlargerthaninpreviouswork.

In [118], ratseasily succeeded at discriminating the default viewsofthevisualobjectswitha>70%correctaccuracy,andquickly

(13)

Fig.7.ObjectconditionsandratgroupaverageperformanceinZoccolanetal.[118].(A)Thefullmatrixofsizeandazimuth-rotationcombinationsusedtotestratinvariant recognitioninZoccolanetal.[118].Thegreenframesshowthedefaultobjectviewsthatratsoriginallylearned.Thelightblueframesshowthetransformationaxesthatrats weretrainedtolerate,beforebeingbestedwiththefullsetoftransformations.(B)Ratgroupaverageperformanceforeachoftheobjecttransformationsshownin(A).The percentageofcorrecttrialsisbothcolor-codedandreportedineachcell,alongwithitssignificance.(Forinterpretationofthereferencestocolorinthisfigurelegend,the readerisreferredtothewebversionofthisarticle.)

ModifiedfromZoccolanetal.[118].

learnedtotoleratesizechangesbetween15◦and40◦ofvisualangle (seeFig.7A,verticallightblueframes)andin-depthazimuth rota-tionsovera±60◦span(seeFig.7A,horizontallightblueframes). Followingthisexposuretosizeandazimuthchanges,ratswere testedwithalargebatteryofobjectconditions,obtainedby com-biningallpossiblepreviouslytrainedsizeandrotationvalues(i.e., allobjectconditionsshown inFig.7A).Crucially, mostofthese objectconditionswerenewfortherats(i.e.,thoseoutsidethelight blueframesinFig.7A).Ratrecognitionperformancewashighand significantlylargerthanchanceforvirtuallyalltested transforma-tions(seeFig.7B),theonlysubstantialdropbeingobservedforthe smallesttestedsizeof15◦ofvisualangle(likelybecauseofratpoor visualacuity).Thisresultshowshowratscantoleratesubstantial variationinobjectappearanceacrossverydifferentvariationaxes andtheircombination.It alsosuggeststhatthis abilityis likely rootedinageneralizationprocess,basedontheperceived simi-laritybetweenthetransformedandtheoriginallylearnedobject views.Infact,giventhelargenumberoftestedconditions(108),it wasunlikelythatratssucceededinthetaskbylearningand memo-rizingthecorrectassociationbetweeneachnewlypresentedview andthecorrespondingrewardport.

Thatratinvariantrecognitionwastheresultofpure general-izationwasalsoexplicitlyandrigorouslytested,bywithholding anyfeedbacktotheratsaboutthecorrectnessoftheirresponsefor afractionofthenewtransformations(approximately11%, cover-ingacontiguousquadrantofthesize-viewmatrix,witheachrat beingtestedwithadifferentno-feedbackquadrant).It isworth noticingthat,differentlyfrompreviousstudies,these generaliza-tiontrialswerenotjust“alwaysrewarded”(atrainingstrategythat, asdiscussedinSection3.3.2,canleadtoeitheranoverestimation orunderestimationofratrecognitionaccuracy,dependingonthe choicesofananimalonthefirstfewgeneralizationtrials).Rather, afteraratmadearesponseinageneralizationtrial,thestimulus wasimmediatelyremovedfromthedisplay,thetrialended,and theanimalwasallowedtotriggerthenexttrialrightaway.Thefact thatratrecognitionperformancefortheseno-feedbackconditions wasashighasforthe(size-matched)feedbackconditions(inboth cases,itwas∼75%correct),andbothwereashighasthe perfor-manceovertheoriginallytrainedsizeandazimuthaxes,indicates

thatratswerecapableofpuregeneralizationtonovelobjectviews. Thisconclusionwasfurthercorroboratedbyobservingthat “per-formancewashighandsignificantlyabovechance[...],evenfor theveryfirstpresentationofeachnovelstimulus,andremained stableoverthecourseoftheexperiment”[118].Asafinal assess-ment ofrat generalizationability, Zoccolanand colleaguesalso introducedtransformationsalongtwonewaxes–in-depth ele-vationrotationandlighting(allthesenewconditions,15peraxis, werepresentedinno-feedback,generalizationtrials).Performance wasabovechanceforallbutoneofthenovellightingconditions (withanoverallrecognitionaccuracyof∼75%correct),andallthe novelelevationconditions.

Overall,theseresultsledZoccolanandcolleaguestoconclude “thatratswereabletogeneralizeacrossa widerangeof previ-ouslyunseenobjecttransformations,includingnovelcombinations of sizeand in-depthrotationsof thelearned objects[...],new lighting/shadingpatternsovertheobjects’surface[...],andalso substantialvariationinobjectsilhouette[...]”[118].Althoughthis studydidnotexplicitlyaddresswhatobjectprocessingstrategy underlayratinvariantrecognition,severalfactorsargueinfavor ofhigher-level, shape-basedprocessing,rather thanrelianceon lower-levelvisualproperties.First,becauseeachviewofanobject waspresentedinisolation,aratcouldnotdirectlycompareittoa similarlytransformedviewoftheotherobject,asittypically hap-pensinspatialtwo-alternativeforced-choicetasks(seediscussion inSection3.3.1).Rather,theanimalhadtoimplicitlycomparethis viewtoallotherpossibletransformedviewsoftheotherobject tosucceedin thetask. Giventhesubstantialsizevariationsthe objectunderwent,thismademanylower-levelcues(suchasthe relativebrightnessorcontrastofthediscriminanda)completely unreliableaspredictorsofobjectidentity.Inotherwords,theuse ofalower-levelstrategy,suchastheonereportedbyMininiand Jeffery[146],wasruledoutbytheexperimentaldesignitself.In addition,Zoccolanandcolleaguesperformedastimulusanalysis showingthat,atthelevelofpixelorGaborfilterrepresentations (simulating,respectively,neuronalrepresentationsinretinaand primaryvisualcortex),theaverageimagevariationproducedby changingtheappearanceofanobjectwaslargerthanthe aver-agevariationbetweenmatchingviewsofthetwoobjects.Inother

Riferimenti

Documenti correlati

Dopo aver esplicitato sommariamente le fonti da cui Tucker attinse per la costruzione del suo pensiero, si andrà, in questo capitolo, a delinearne le eredità. Gli echi del

While the event analysis looks for the subjective identifi- cation of a confirmed event of change in a visual field series compared to a baseline exam, the trend analysis uses

to m identical parallel processors by minimizing the makespan) arising when the number of jobs that can be assigned to each processor cannot exceed a given integer k. The problem

Consequently, the research activities will be more and more focused on the following topics: smart grids and polygeneration microgrids, renewable power plants and storage

At shorter wavelengths (370–470 nm), light absorp- tion by brown carbon from primary organic aerosol (POA) and secondary organic aerosol (SOA) formed during ag- ing was around 10 %

This approach has been applied to two-dimensional and three-dimensional cases in regular waves: the evaluation of ship mo- tions and added resistance in time domain

Nel 1990, in seguito alla progettazione di un nuovo Museo Comunale, il quale avrebbe ospitato sia reperti provenienti da scavi effettuati nella zona negli ultimi anni sia