1

Characterizing the response to face pareidolia in human

category-selective visual cortex

Susan G Wardle*1, Kiley Seymour2 & Jessica Taubert3

1Department of Cognitive Science and ARC Centre of Excellence in Cognition and its Disorders, 16 University Avenue,

Macquarie University, New South Wales, 2109, Australia.

2School of Social Sciences and Psychology, Western Sydney University, Sydney, New South Wales, Australia.

3Section on Neurocircuitry, Laboratory of Brain and Cognition, The National Institute of Mental Health. 10 Center Drive,

Bethesda, Maryland, 20892, USA.

* Corresponding Author

Dr Susan Wardle

Department of Cognitive Science

16 University Avenue

Macquarie University

New South Wales, 2109, Australia.

Ph: +61 2 9850 4072

susan.wardle@mq.edu.au

2

Abstract

Theneuralmechanismsunderlyingfaceandobjectrecognitionareunderstoodtooriginateinventral

occipital-temporalcortex.Akeyfeatureofthefunctionalarchitectureofthevisualventralpathwayis

itscategory-selectivity,yetitisunclearhowcategory-selectiveregionsprocessambiguousvisualinput

which violates category boundaries. One example is the spontaneous misperception of faces in

inanimateobjectssuchastheManintheMoon,inwhichanobjectbelongstomorethanonecategory

andfaceperceptionisdivorcedfromitsusualdiagnosticvisualfeatures.WeusedfMRItoinvestigate

therepresentationofillusoryfacesincategory-selectiveregions.Theperceptionofillusoryfaceswas

decodablefromactivationpatternsinthefusiformfacearea(FFA)andlateraloccipitalcomplex(LOC),

butnotfromothervisualareas.Further,activityinFFAwasstronglymodulatedbytheperceptionof

illusoryfaces,suchthatevenobjectswithvastlydifferentvisualfeatureswererepresentedsimilarlyif

allimagescontainedanillusoryface.TheresultsshowthattheFFAisbroadly-tunedforfacedetection,

not finely-tuned to the homogenous visual properties that typically distinguish faces from other

objects.Acompleteunderstandingofhigh-levelvisionwill requireexplanationof themechanisms

underlyingnaturalerrorsoffacedetection.

Keywords. face processing, fMRI, fusiform face area (FFA), lateral occipital complex (LOC), object

recognition,representationalsimilarityanalysis

3

INTRODUCTION

Facepareidoliaistheperceptionofillusoryfacesininanimateobjects.Recently,facepareidoliahas

been demonstrated in non-human primates (Taubert et al. 2017), evidence that it arises from a

fundamentalaspectofafacedetectionsystemsharedacrossprimatespecies.Twokeyfeaturesof

face pareidolia are that it is typically a spontaneous andpersistent phenomenon,with the object

perceivedsimultaneouslyasbothanillusoryfaceandaninanimateobject,andsecondlythatthevisual

attributes perceived as illusory facial features are highly variable. Face pareidolia thus offers the

potential toexaminehowthebrainprocessesfaces inauniquesituationwherefaceperception is

decoupledfromthevisualpropertiesthattypicallydefinefacesasacategory.Inthehumanbrain,the

lateraloccipitalcomplex(LOC)(Malachetal.1995;KourtziandKanwisher2000)andfusiformface

area(FFA)(Kanwisheretal.1997)inoccipital-temporalcortexrespondpreferentiallytoeitherobjects

orfacesrespectively.Theinvolvementofthesecategory-selectiveregionsinthevisualventralstream

inprocessingfalsepositivefacesthathaveasimultaneousdualfaceandobjectidentityisunclear.

Several lines of evidence indicate that the response of the FFA is tightly-tuned to faces

(KanwisherandYovel2006).EarlyfMRIstudiesshowedthattheFFArespondspreferentiallytofaces

comparedtootherobjectcategories,evenifthefacesareschematicorimagined(Kanwisheretal.

1997;O'CravenandKanwisher2000;Tongetal.2000).Morerecently,face-relatedfeaturesincluding

identity and perceptual differences have been successfully decoded from fMRI BOLD activation

patterns in theFFAusingmultivariatepatternanalysis (Nestoretal.2011;GoesaertanddeBeeck

2013; Anzellotti et al. 2014; Axelrod and Yovel 2015; Zhang et al. 2016; Guntupalli et al. 2017).

ConvergingevidenceindicatesthattheresponseoftheFFAisrelatedtofaceperception.Activityin

the FFA correlateswith face detection and identification, but notwith the identification of other

complexobjects(e.g.cars)thatrequireasimilar levelofvisualexpertise(Grill-Spectoretal.2004).

Similarly,theFFAdisplaysagreaterresponsetouprightthaninvertedfaces,consistentwiththerobust

behavioral observation of superior face recognition for upright faces (Yovel and Kanwisher 2005).

Although the face-selectivity of the FFA is well-established, it is not yet known whether the

spontaneousperceptionofafaceinanexemplarfromanon-preferredcategory(e.g.inaninanimate

object)wouldmodulateactivityintheFFA.

Givenitsselectivityforshapeandobjectproperties(Malachetal.1995;KourtziandKanwisher

2000;2001),itisunclearwhetheractivityintheLOCwoulddifferentiatebetweeninanimateobjects

whichcontainillusoryfacescomparedtosimilarobjectswithoutanillusoryface.Althoughtheobject-

selectiveLOCisalsoresponsivetofacestosomedegree(YovelandKanwisher2004;2005),unlikethe

FFA,activity in LOCdoesnot correlatewithbehavioralmeasures suchas the face inversioneffect

(YovelandKanwisher2005)orthedegreeofperceivedface-likenessforsimplesilhouettes(Davidenko

4

etal.2012).Thus,whilethemisperceptionofafaceinnoiseproducesanincreaseinthemagnitude

ofactivityintheFFA(Summerfieldetal.2006;Liuetal.2014)andissuggestiveofarolefortheFFA

inperceivingillusoryfaces,thepossibleroleofobject-selectivecortexinprocessingillusoryfacesis

moredifficulttopredict.

Toexaminetherepresentationofillusoryfacesinhumancategory-selectivecortex,weused

functionalmagneticresonance imaging(fMRI) torecordpatternsofblood-oxygen-level-dependent

(BOLD)activationinvisualbrainareaswhilesubjectsviewedpicturesofobjectseitherwithorwithout

illusory faces (Figure 2a). We collected 56 photographs of naturally occurring examples of face

pareidoliainobjectssuchasfood,accessories,andappliances.Foreachobjectwithanillusoryface,

wefoundasimilarimageofthesamecategoryofobjectbutwithoutanillusoryface.Importantly,this

yokedimagesetof56imageswasmatchedforobjectcontentandvisualfeaturestypicaloftheobject

category but did not contain any illusory faces. In total there were 8 unique image sets, each

containing14 images (Figure1),whichwerepresented in a standard fMRIblockeddesign (Figure

2b,c).Inordertocomprehensivelycharacterizetheresponseofvisualventralareastotheperception

of illusory faces in inanimate objects, we applied both multivariate pattern analysis and

representational similarity analysis techniques to the fMRI data. The aimswere firstly to discover

whichcategory-selectiveareasofvisualventralcortexaresensitivetotheperceptionofillusoryfaces,

andsecondlytodeterminetowhatextenttherepresentationintheseregions is influencedbythe

presenceofanillusoryface.

5

Figure1.ExperimentalStimuli.All112experimentalstimulifromthe8imagesets.Fourimagesets

containimagesofillusoryfacesininanimateobjects,andfourimagesetscontainimagesofobjects

matchedforcontentand/orvisualfeatures,butwithoutanillusoryface.

6

MATERIALS AND METHODS

Participants.Nineparticipants(8naivetoexperimentalaims(6female),1author)participatedin

returnforfinancialcompensationandprovidedinformedconsent.Oneadditionalparticipantwas

excludedfromanalysisduetoexcessivemovementintheMRIscanner(>20mm).Ethicalapproval

forallexperimentalprocedureswasobtainedfromtheMacquarieUniversityHumanResearchEthics

Committee(MedicalSciences).

MRIDataacquisition.MRIdatawereacquiredwitha3TSiemensVerioMRIscanneratMacquarie

Medical Imaging,MacquarieUniversityHospital.Ahigh-resolution (1x1x1mm)T1-weighted3D

whole-brain structural MRI scan was collected for each participant at the start of the session.

Functionalscanswereacquiredwitha2DT2*-weightedEPIacquisitionsequence:TR=2s,TE=32ms,

FA=80deg,voxelsize:2.5x2.5x2.5mm,inplanematrixsize:102x102.Apartialvolumecontaining

33sliceswascollectedorientedapproximatelyparalleltothebaseofthetemporallobeoranterior

commissure-posterior commissure (AC-PC) line for each participant, ensuring coverage of both

occipitalandtemporallobes.

Visualstimuliweredisplayedusingaprojectorforonesubject(1author)withresolution1280

x800andviewingdistance150cm.Experimentalstimuli(400x400pixels)subtended10.5x10.5°,

localizer stimuli (480 x 480 pixels) 12.6 x 12.6°. For the remaining 8 naive subjects, stimuli were

displayed using a flat-panel MRI-compatible 32" Cambridge Research Systems BOLDscreen with

resolution1920x1080andviewingdistance112cm.Experimentalstimulisubtended7.4x7.4°and

localizerstimuli8.8x8.8°.BehavioralresponseswerecollectedusingaLuminaMRI-compatiblebutton

box.

ExperimentalDesignandStimuli.ExperimentalcodewaswritteninMATLABwithfunctionsfromthe

Psychtoolbox(Brainard1997;Pelli1997;Kleineretal.2007),andrunonanAppleMacBookProwith

MacOSX.Thescanningsessionforeachsubjectconsistedofahigh-resolutionstructuralMRIscan,2

functionallocalizerrunstoidentifycategory-selectiveregionsFFA,LOC,andPPA,and8experimental

runs.Thefunctionallocalizerwasrunonceatthestartofthesessionfollowingthestructuralscanand

onceattheendofthesession.Theeightexperimentalrunsforeachparticipantwerecollectedin-

betweenthelocalizerruns.

7

Figure2.fMRIMethods.(a)Examplestimuli:illusoryfaces(left)andmatchedobjects(right).(b)Trial

sequencewithineachimageblock.Whileinthescanner,participants(N=9)performeda1-backtask

tomaintainattention,pressingakeywheneveranimagewasshowntwiceinarow.(c)Sequenceof

blockdesign.Illusoryfaceblocksalternatedwithmatchedobjectblocks,separatedbyfixationperiods.

Fortheexperimentalruns,56colorimagesofnaturalexamplesofillusoryfacesininanimate

objectswerecollectedfromtheinternet.Foreachillusoryfaceimage,wesourcedaphotographofa

similarobjectwithoutanillusoryfaceandcroppedittomatchthefirstimageascloselyaspossible

(Figure2a). Imageswerecroppedtothesamesize,butnoothermanipulationwasmade.The112

photographsweregroupedinto8uniqueimagesetsof14imageseach,with4illusoryfacesetsand

4matchedobjectsets(Figure1).Stimuliwereshowncenteredonagrayscreenwithacentralblack

andwhite fixation bullseye (0.4° diameter for subject 1, 0.3° diameter for the all other subjects).

Duringfixationblocksthefixationbullseyeremainedonthegraybackground.Eachimageblockwas

16s in duration, followed by a 10s fixation period.Within each block, every imagewas shown in

random order for 800ms followed by a 200ms inter-stimulus-interval (Figure 2b). The order of

presentation of the images within each block was yoked between the face image sets and their

matchedobjectsets.Anewyokedrandomorderwasgeneratedeachtimeanimagesetblockwas

repeatedinthesession.Foreachblocktherewere16images:14uniqueimages,andanadditional2

8

imageswhichwererepeatedforthe1-backtask.The1-backtaskwasyokedbetweenillusoryfaceand

matchedobjectblocks,suchthattherepeatedimagesonagivenblockpresentationwereidentical

between the matched image sets. Two 1-back trials occurred during each 16s block to maintain

attention,onceinthefirsthalfandonceinthesecondhalf.Participantspressedakeywhentheysaw

thesameimagetwiceinarow.Feedbackontaskperformancewasgivenattheendofeachrunafter

scanningwascomplete.Overall taskperformancewashighacross subjectsandexperimental runs

(Meanaccuracy=98.8%,SD=1.5%).

Theorderoftheexperimentalblocksalternatedbetweenillusoryobjectsandmatchedobjects

inanABAB/BABAdesign(Figure2c),withthestartingblocktypecounterbalancedacrossrunsforeach

participant.Theorderofblockpresentationwithineachrunwasrandomizedwithinthisdesignwith

theadditionalrestrictionsthat(1)amatchedimagesetneverdirectlyfolloweditsyokedfacesetand

viceversa,and(2)eachimagesetwascycledthroughonceintheruninarandomorderbeforebeing

repeated. Each of the 8 unique image sets was repeated twice per 7-minute experimental run.

Experimentalrunsbeganwitha4sfixationperiod.

Thefunctionallocalizerstimuliwerecolorpicturesoffaces,places,andobjects.54imagesfor

eachcategorywere selected fromTheCenter forVital Longevity FaceDatabase (MinearandPark

2004),theSUN397database(Xiaoetal.2010),andtheBOSSdatabase(Brodeuretal.2010;2014)

respectively. Scrambled objects for localizing object-selective region LOC (Kourtzi and Kanwisher

2000)werepre-generatedinMATLABbyrandomlyscramblingeachobjectimageinan8x8gridand

savingtheresultingimage.Foreachstimulusclass,therewere3uniqueblocksof18images.Every

timeablockwasrun,theimageswerepresentedinarandomorderandtworandomimageswere

repeatedtwiceforthe1-backtask.Participantsperformeda1-backtaskasfortheexperimentalruns,

pressingakeyeachtimeanimagewasrepeatedtwiceinarow(Meanaccuracy=96.8%,SD=3.8%).

Eachof the20 imageswithinablock (18unique+2 repeats)wasshownfor600ms followedbya

200msinter-stimulusinterval.The16sstimulusblocksalternatedwith10sfixationblocks.Eachofthe

4stimuluscategorieswasrepeatedthreetimesper5-minutelocalizerrun,onceperuniqueimageset.

The order of block presentation was in a pseudorandom order, with two different orders

counterbalancedacrossrunsforeachparticipant.Localizerrunsbeganwitha4sfixationperiodbefore

thefirststimulusblock.

fMRIdatapreprocessing.MinimalpreprocessingoftheMRIdatawasconductedusingSPM8.Foreach

observer,fMRIdataforallexperimentalandlocalizerrunswasmotioncorrectedandco-registeredto

their structural scan. No normalization or spatial smoothing was applied, and all analyses were

conductedinthenativebrainspaceofeachsubject.

9

Region-of-interest(ROI)definition.CorticalreconstructionwasperformedusingFreesurfer5.3from

thestructuralscanforeachsubject.V1(range:886-1244voxels,M=1094)wasanatomicallydefined

ineachobserverbasedontheirindividualsurfacetopologyusingcorticalsurfacetemplatesapplied

totheirinflatedcorticalsurface(Bensonetal.2012).Thecategory-selectiveregions-of-interest(ROIs)

were functionally defined using standard procedures from the independent localizer runs

(Kriegeskorteetal.2009).FFA(range:29-78voxels,M=58)wasdefinedasthecontiguousclusterof

voxels inthefusiformgyrusproducedfromthecontrastbetweenfacesversusobjectsandscenes.

LOCscr(range:7-64voxels,M=39)wasdefinedasactivationonthelateraloccipitalsurfacefromthe

contrast between objects and scrambled objects. For comparison, LOCofs (range: 43-245 voxels,

M=113)wasdefinedasactivationonthelateraloccipitalsurfaceforthecontrastofobjectsversus

faces and places. Finally, PPA (range: 12-170 voxels,M=75) was defined as the peak cluster of

activationintheparahippocampalgyrusproducedbythecontrastbetweenscenesversusfacesand

objects(Kauffmannetal.2015).OverlappingvoxelswerepermittedbetweenLOCscrandLOCofs(range:

2-31voxels,M=12),butnotbetweentheotherregions.

Generallinearmodel.Toestimatethebetaweightsforeachexperimentalcondition,thefunctional

MRIdatafromtheexperimentalrunswasenteredintoaGLMinSPM8withaseparateregressorfor

eachofthe8stimulusblocks,producingoneparameterestimate(βweight)perconditionperrun(64

intotalperparticipant).Fixationwasnotexplicitlymodelledbutprovidedanimplicitbaseline.Percent

BOLDsignalchange(Figure3a)wascalculatedforeachparticipantandregion-of-interestacrossall

facevs.matchedobjectblocksusingSPM8andthemarsbar(Brettetal.2002)package.

Decodinganalysis.Decodinganalyseswereperformedseparatelyforeachparticipantintheirnative

brainspaceinMATLABusingfunctionsfromTheDecodingToolbox(Hebartetal.2014).Classification

usingalinearSVMwasconductedusingthebetaweightsforeachofthe8uniquestimulusblocks.For

leave-one-run-outclassification(Figure3b),theclassifierwastrainedontheactivationpatternsfor

each pair of the 8 conditions for 7 runs, and tested on the one run left out of the training set.

Classification accuracywas averaged across all permutations of leave-one-run-out cross-validation

folds(n=28).Forfacecross-decoding(Figure3c),theclassifierwastrainedonalldataforayoked

stimulusset (e.g.FacesSet1vs.MatchedObjectSet1)andtestedonalldata forasecondyoked

stimulusset(e.g.FacesSet2vs.MatchedObjectsSet2),andthenclassificationwasrepeatedwiththe

test and training sets swapped. This process was repeated for all possible yoked pairings, with

classificationaccuracyaveragedacrossallpermutations(n=6).Forobjectcross-decoding(Figure3d),

10

two-way cross classificationwas conducted for each relevant pair as for face cross-decoding. For

example,theclassifierwastrainedontwodifferentimagesetseitherbothwithorwithoutfaces(e.g.

Faces Set 1 vs. Faces Set 2) and tested on their yokedmatched sets (e.g.MatchedObjects 1 vs.

MatchedObjects2),andthenviceversa.Classifieraccuracywasaveragedacrossallpermutationsof

possiblepairs(n=6).

Representationalsimilarityanalysis.Representationalsimilarityanalyses(Kriegeskorteetal.2008;

Kriegeskorte and Kievit 2013) was performed in MATLAB using functions from the toolbox for

representationalsimilarityanalysis(rsatoolbox)(Nilietal.2014).Categoricalmodelswereconstructed

basedonthestimulusdesign,todistinguishbetweenthetwopotentialoutcomesofinterest:whether

thesimilarityofbrainactivationpatternswasdeterminedprimarilybythepresenceofanillusoryface

orobjectcontent(Figure4c).Blueregionsindicatepredictedsimilarpatternsofactivationbetween

image set pairs, yellow regions indicate predicted dissimilar activation patterns, and grey regions

indicatemoderatedissimilarity.Theempiricaldissimilaritymatricesforeachbrainregion(Figure4a,b)

werecalculatedby1-correlationoftheBOLDactivationpatternsforeachimagesetpair.Dissimilarity

was scaled separately for each brain region to fall between 0 and 1 (min andmax dissimilarity).

CorrespondencebetweenthecategoricalRSAmodelsandthedissimilaritymatriciesforeachbrain

region(Figure4d)wereassessedbycomputingthecorrelation(Kendalltau-a)andapplyingcondition-

labelrandomizationwithFDRcorrectionappliedformultiplecomparisons(p<.05)(Nilietal.2014).

RESULTS

We focused our analysis on specific predefined regions-of-interest in ventral occipital-temporal

cortex.Ineachparticipant,welocalizedthreecategory-selectiveregions:thefusiformfacearea(FFA),

thelateraloccipitalcomplex(twoversions:LOCscrandLOCofs),andtheparahippocampalplacearea

(PPA),usingstandardfunctionallocalizers(Kanwisheretal.1997;EpsteinandKanwisher1998;Kourtzi

andKanwisher2000;Kravitzetal.2011;Parketal.2011;Kauffmannetal.2015).Asanadditional

comparisonregion,wedefinedV1anatomicallyineachsubject(Bensonetal.2012).

Univariateanalysis

Tocharacterizetheamountofactivationineachregion,wecalculatedBOLDpercentsignalchangefor

allillusoryfaceblocksversusallnon-faceblocks(Figure2a).Statisticalsignificancewasassessedfor

percentsignalchangeforillusoryfaceversusnon-faceblocksonthemeanactivationaveragedacross

participantsforeachregionofinterest(Figure2a)usingtwo-tailedpairedt-testswiththeBonferroni

11

correctionappliedformultiplecomparisons(k=5)tomaintainα<0.05(α/k=.01).Illusoryfaceblocks

produced significantlyhigheractivation thannon-faceblocksasmeasuredbypercentBOLDsignal

changeforLOCscr(t(8)=4.25,p=.002814)andFFA(t(8)=8.15,p=.000038).Therewasnosignificant

differenceinactivationforillusoryfacesversusnon-facesforLOCofs(t(8)=2.89,p=.020278),PPA(t(8)=

-2.68,p=.027728),orV1(t(8)=-2.41,p=.042692).

Figure3.fMRIresultsaveragedacrosssubjects(N=9)foreachregion-of-interest.(a)BOLD%signal

change for illusory face blocks andmatched object blocks. Error bars are between-subjects SEM.

AsterisksdenoteasignificantdifferenceassessedwithBonferronipairedt-tests(*α<.05,**α<.01).

(b) Leave-one-run-out pairwise decoding of the 8 image sets. For panels b-d, asterisks denote a

significantdifferencefromchancedecodingperformance(50%)assessedwithBonferronione-sample

t-tests(*α<.05,**α<.01)anderrorbarsarebetween-subjectsSD.(c)Illusoryfacecross-decoding

results.Cross-classificationofillusoryfacesversusyokedobjectswithoutillusoryfacesacrossimage

sets(e.g.trainclassifieronFacesSet1vs.MatchedSet1;testclassifieronFacesSet2vs.MatchedSet

2).(d)Objectcross-decodingresults.Cross-classificationofobject identityacrossyokedimagesets

(e.g.trainclassifieronFacesSet1vs.FacesSet2;testclassifieronMatchedSet1vs.MatchedSet2).

12

Multivariatepatternclassification

Leave-one-runoutclassification

Foraninitialsanitycheck,weperformedastandardleave-one-run-outclassificationanalysisonthe

patternsofBOLDactivationevokedby the8unique image sets.Usingpairwise classificationwith

cross-validation,alinearsupportvectormachinewastrainedonthedataforallrunsminusone,and

testedontheremainingset(Hebartetal.2014).Figure2bshowsthemeanclassificationaccuracyfor

each region-of-interest, averaged across subjects.Decoding accuracywas assessed statistically for

eachROIusingtwo-tailedone-samplet-testsonthemeanclassificationaccuraciesacrossparticipants

against chance decoding performance of 50%with the Bonferroni correction applied formultiple

comparisons(k=5)tomaintainα<0.05(α/k=.01).Decodingperformancewassignificantlyabove

chanceforallfiveROIs(Figure2b).V1:(t(8)=7.27,p=.000086),LOCscr(t(8)=4.35,p=.002444),LOCofs

(t(8)=7.40,p=.000076),PPA(t(8)=6.33,p=.000226),FFA(t(8)=11.16,p=.000004).Notethatasthe

classifieristrainedonexamplesofBOLDactivationpatternsevokedbyeachoftheuniquestimulus

blocksusingthisclassificationmethod,successfulclassificationcouldbebasedonsensitivitytolow-

levelvisualfeaturesratherthantheobjectorfacecontentoftheimagesets.Itislikelythatsuccessful

classification was based on different visual features in each ROI, however the consistently high

classifierperformancevalidatesthehighpowerofthefMRIdesignandverifiestherobustnessofthe

dataforthesubsequentanalysesofinterest.

Illusoryfacecross-classification

Weusedcross-decodingasthecriticaltestfordefiningwhethereachbrainregionwassensitivetothe

presenceorabsenceofillusoryfacesininanimateobjects(Figure2c).Alinearclassifierwastrained

onthepatternsofBOLDactivationevokedbyoneyokedsetofimages(e.g.FaceSet1vs.Matched

Set1),andtestedonthebrainactivationpatternsforanewyokedset(e.g.FaceSet2vs.MatchedSet

2). This analysis requires that the classifier be able to generalize across new images in order to

correctly classify the test set, excluding the possibility of illusory face decoding based upon brain

responsestothelow-levelvisualfeaturesofparticularimages.Cross-decodingaccuracywasassessed

usingtwo-tailedone-samplet-testswiththeBonferronicorrectionappliedformultiplecomparisons

(k=5)tomaintainα<0.05(α/k= .01).Cross-classificationof illusoryfaceswassignificantlyabove

chance in LOCscr (t(8)=4.97, p = .001096), LOCofs (t(8)= 7.21, p = .000092), and FFA (t(8)= 13.29, p =

.000001),butnotforV1(t(8)=3.10,p=.014737)orPPA(t(8)=2.44,p=.040749).

13

Objectcross-classification

As our stimulus set included matched objects, we also tested whether object identity could be

decodedfromeachbrainregionacrossdifferentphotographs(Figure2d).Inthisanalysis,weignored

thepresence/absenceofanillusoryface,andinsteadtrainedtheclassifiertodiscriminatebetween

twodifferentobjectsetsthateitherbothdidordidnotcontainillusoryfaces(e.g.FaceSet1vs.Face

Set2).Theclassifierwasthentestedonthematchedsetscorrespondingtothetwotrainingsets(e.g.

MatchedSet1vs.MatchedSet2).Successfulclassificationrequiresextrapolationbasedonresponses

toeitherobject identityor sharedvisual featuresacrossdifferent images,but it isnotpossible to

distinguishbetweenthetwopossibilities.Objectcross-decodingwassignificantlyabovechancefor

LOCscr(t(8)=4.14,p=.003250),LOCofs(t(8)=5.63,p=.000491),andPPA(t(8)=4.37,p=.002379),butnot

forV1(t(8)=1.12,p=.296317)orFFA(t(8)=1.34,p=.216760).Thusbothfaceandobjectcross-decoding

weresuccessfulinLOC,butonlyfacesandnotobjectscouldbecross-decodedacrossimagesetsfrom

activityinFFA,andonlyobjectsbutnotfacesfromactivityinPPA.Neitherformofcross-decodingwas

successfulinV1,yettheidentityofthe8uniqueimagesetscouldbedecodedfromV1activityusing

leave-one-run-outclassification,consistentwithitsknownsensitivitytolow-levelvisualproperties.

14

Figure4.Representationalsimilarityanalysis.(a)Dissimilaritymatrixshowingthenormalizedpairwise

dissimilarity(1-correlation)inBOLDactivationpatternsevokedbyeachofthe8uniqueimagesetsin

theFFA.Blueindicatesmoresimilaractivationpatterns,yellowmoredissimilar.Notethatthefour

uniquefaceblocksproducesimilarpatternsofactivity(upperleftquadrantisblue).(b)Dissimilarity

matrices for LOCscr, LOCofs,PPAandV1 (legendas forpanela). (c)Categoricalmodelsofexpected

patternsof similarity in theBOLDactivationpatternsevokedbyeach imageset if facesorobjects

dominatetheunderlyingrepresentation.(d)Correlation(Kendalltau-a)betweenthecategoricalRSA

models and the FFA RDM. Mean correlation across subjects (N=9) is shown, asterisks denote a

significantdifferencefromchancecalculatedusingcondition-labelrandomizationwithFDRcorrection

appliedformultiplecomparisons(p<.05).ErrorbarsareSEM.Thegraybandrepresentstheupper

andlowerboundariesofthenoiseceiling,indicatingthemaximumpossiblecorrelationpossiblefor

any model given the noise in the data (Nili et al. 2014). Note that no correlations between the

categoricalmodelsandtheotherregionsofinterest(LOCscr,LOCofs,PPA,V1)weresignificant.

15

RepresentationalSimilarityAnalysis

To further characterize the response to illusory faces in visual ventral cortex, we applied

representationalsimilarityanalysis(KriegeskorteandKievit2013)toexaminesimilaritiesinactivation

patternsforthedifferentimagesets.Therepresentationaldissimilaritymatricesforeachregionof

interestquantifythedegreeofsimilarityintheBOLDactivationpatternsacrossvoxelsbetweeneach

pair of image sets (Figure 4a,b). Different models of the underlying representational space are

predictediffacesorobjectsarethedominantorganizingprincipleinaparticularbrainregion(Figure

4c). Notably, the dissimilarity matrix for the FFA (Figure 4a) reveals a strong dominance of the

presence of an illusory face on the brain activation patterns (upper left quadrant). Image sets

containingillusoryfacesproducemoresimilarresponsesinFFAthantheirmatchedimagesets,even

thoughthematchedimagesetssharemanymorevisualfeatureswiththefacesetsthanispresent

betweendifferentfacesets.Thisrelationshipiswell-describedbythefacemodel(Figure4c),which

significantlycorrelateswiththeFFAdissimilaritymatrix(Figure4d).

Beyond a similar representation of illusory faces irrespective of their visual features, the

responseoftheFFAisstronglymodulatedbythepresenceorabsenceofaface.Objectswhichdonot

haveanillusoryfaceelicitdissimilaractivationpatternstoeachotherinFFA(bottomrightquadrant

inFigure4a),aswouldbeexpectedfromdifferencesintheirobjectidentityand/orvisualfeatures.

Thisisinsharpcontrasttotheactivationpatternselicitedbytheiryokedobjectswithillusoryfaces,

whichelicithighlysimilaractivationpatternstoeachotherintheFFAdespitetheirsubstantialvisual

differences (compare upper left and bottom right quadrants in Figure 4a). In contrast, objects

containingillusoryfacesaremoderatelydissimilartothosewithoutillusoryfaces(toprightquadrant

inFigure 4a), despite their visual similarity resulting from the yoked stimulus design.Overall, the

results indicatethattheresponseofFFAisdominatedbythepresenceorabsenceofafacetothe

extent that itoverridesany influenceof thesimilarityofvisualpropertiesorobject identity in the

representation.

These relationships are well-captured by the face-nonface model (Figure 4c), which

significantlycorrelateswiththeFFAdissimilaritymatrix(Figure4d)andapproachesthetheoretical

limitofthemaximumpossiblecorrelationwiththefMRIdataasdefinedbythenoiseceiling(Nilietal.

2014).Notablytheobjectmodelfailstocapturetherepresentationalstructureofthevisualstimuliin

FFA,indicatedbyalackofasignificantcorrelationwiththeFFAdissimilaritymatrix(Figure4d).Insum,

theperceptionofanillusoryfacemediatesactivityinFFA,regardlessoftheparticularvisualproperties

that compose the facial features. Although the presence of an illusory face could be successfully

decodedfromactivationpatternsinLOCscrandLOCofs(Figure4b),noneofthedissimilaritymatrices

16

fortheotherregionsofinterest(Figure4b)significantlycorrelatedwiththecategoricalmodels(Figure

4c),suggestingthisorganizationoftherepresentationalspaceisuniquetotheFFA.

DISCUSSION

We investigated the representation of illusory faces in inanimate objects throughout category-

selectivehumanventralvisualcortex.Importantly,weusedayokedfMRIdesigntocomparepatterns

ofBOLDactivationevokedbyimagesofobjectscontainingillusoryfaceswiththatevokedbysimilar

matched objects without illusory faces. We selected only naturally-occurring examples of face

pareidoliainordertoinvestigatespontaneousfalsepositivesoffacedetectionwithahighdegreeof

varianceintheirspecificvisualproperties.Thecross-decodingresultsrevealedthatactivity inboth

face (FFA) andobject-selective (LOC) regionswasmodulatedby images that elicit facepareidolia,

regardlessof theparticularvisualelements thatcomposethe facial featuresand itsconfiguration.

GiventhatV1isparticularlysensitivetolow-levelvisualfeaturesandisimplicatedintheearlystages

ofretinotopicvisualprocessing,itisunsurprisingthattheclassifierfailedtodecodethepresenceof

an illusory face fromactivationpatterns in this region.Wealso foundnoevidence thatactivity in

scene-selectivePPAismediatedbytheperceptionofillusoryfaces.ThisisnotableasPPAisfrequently

functionally defined as voxelswhich exhibit a greater response to scenes than faces (Epstein and

Kanwisher 1998; Kravitz et al. 2011; Park et al. 2011; Kauffmann et al. 2015), thus a reduction in

activity for illusory facesmighthavebeenexpected.However, therewasalsonodifference in the

magnitudeofBOLDactivationinPPAforillusoryfacesversusmatchedobjects.

Inadditiontosuccessfulclassificationof illusory facesacross imagesets,activity inobject-

selectiveregionswasalsoinformativeaboutobjectcontent.SuccessfulobjectcrossdecodinginLOC

is consistent with previous results suggesting patterns of activation in these regions are strongly

associatedwithobjectidentity(MacEvoyandEpstein2009)andobjectshape/features(Kourtziand

Kanwisher2000).Notably,objectcrossdecodingwasalsosuccessful inPPA,butnot inFFAorV1.

Thereisconsiderableinfluenceoflow-levelvisualfeaturesontheresponseofscene-selectiveregions

(Groenetal.2017),thuscross-classificationofobjectsinPPAmaybebaseduponresponsesevoked

bysharedvisualfeaturesbetweendifferentpicturesofthesameobjectclass.

AlthoughbothFFAandLOCcontaineddecodableinformationabouttheperceptionofillusory

facesininanimateobjects,itislikelythatthestructureofthecorrespondingrepresentationsofthese

categoricallyambiguous stimuli is substantiallydifferentbetween these two regions.Beyond their

difference in preferred category (faces vs. objects) as measured by the magnitude of the BOLD

response (Kanwisher et al. 1997; Kourtzi and Kanwisher 2000), there are other known functional

differences.Previousstudieshaveobservedthatactivity inLOCdoesnotcorrelatewithbehavioral

17

measuresoffaceprocessing(YovelandKanwisher2005;Davidenkoetal.2012),whileactivityinFFA

does(Grill-Spectoretal.2004;YovelandKanwisher2005).Thisisconsistentwithourfindingthatthe

facemodelssignificantlycorrelatedwiththeFFAdissimilaritymatrix,andnotwiththat forobject-

selectiveLOC.Thissuggeststhatthepresenceofillusoryfacesisastrongerpredictoroftheunderlying

representationoftheseobjectsintheFFA,whiletherepresentationinLOCislikelytoinvolveother

factors.Further,theobjectcross-decodinganalysisrevealedthatactivityinLOCwasalsoinformative

aboutobjectidentity,whileobjectcontentcouldnotbedecodedfromtheFFA.Togethertheseresults

point to genuine functional differences in the representationof illusory faces in face- andobject-

selectivecorticalregions.

Thefunctionalarchitectureofhumanventraltemporalcortexisunderstoodtoinvolveboth

clustered category-selective regions such as the FFA, and more distributed patterns of category-

selectivityacrosslargerregionsofcortex(Grill-SpectorandWeiner2014).Animportantoutstanding

questionistounderstandthefunctionalsignificanceoftheinformationtheFFAcontainsaboutnon-

face objects (Haxby et al. 2001; O'Toole et al. 2005), which is a challenge to its face-specificity

(Kanwisher2017).ArelevantpointraisedbyourresultsisthattheresponseoftheFFAtoinanimate

objects is stronglymodulated by the spontaneous perception of an illusory face. Image setswith

diversevisualpropertiesandobjectidentitieselicitedsimilarrepresentationsiftheycontainedillusory

faces.Conversely,imagesetswiththesamedegreeofvarianceinvisualpropertiesandobjectidentity

elicitedhighlydissimilarrepresentationsifnoillusoryfaceswerepresent.Notablyevenyokedimage

setsmatchedforvisualfeaturesandobjectidentityproducedmarkedlydissimilarrepresentationsif

onlyonesetcontainedillusoryfaces,despitetheirconsiderablevisualandsemanticsimilarityonother

dimensions.Thissuggeststhatanynon-faceobjectinformationcontainedintheFFAissecondaryto

thedominanceoffaceinformationinitsrepresentationalorganization.

A significant current issue in studying the neural mechanisms of object recognition is in

disentangling the degree towhich category-selective responses in inferior temporal cortex reflect

categoryrepresentationsperse,versussensitivitytothelowlevelvisualfeatures(e.g.shape,color,

symmetry)thatco-varywithcategorymembership(CoxandSavoy2003;Baldassietal.2013;Riceet

al.2014;WardleandRitchie2014;Andrewsetal.2015;BracciandOpdeBeeck2016;Cogganetal.

2016;Kaiseretal.2016;Proklovaetal.2016;Braccietal.2017).Itisintuitivethatbydefinition,any

high-levelvisualrepresentationsmustberelatedtothevisualfeaturesthatdefineobjectcategories

to some degree. However, the impossibility of comprehensively accounting for all possible

combinations of visual features at different levels of description entails that it is empirically

challenging to provide evidence of category-representations that are not in principle reducible to

somesetofco-varyingvisualelements.Thenovelapproachwetakehereistouseastimulusclassin

18

whichfaceperceptionisdecoupledfromthetypicalvisualfeaturesthatdefinerealfaces.Secondly,

theuseofayokedexperimentaldesignmeansthatthesetsofillusoryfacessharemorevisualfeatures

withtheircorrespondingmatchedobjectsthanwitheachother,ensuringthatthevisualvarianceis

higherwithin-categorythanacross-categoryinourdesign.Thisisanextremeexampleofintroducing

highvarianceintothecategoryexemplars(BracciandOpdeBeeck2016;Braccietal.2017),asthe

visualfeaturesinourstimulussetareinsteaddiagnosticofnon-preferredobjectcategories(forthe

FFA).Afurtheradvantageofthisapproachisthatitdoesnotrequireartificiallycontrollingthevisual

contentoftheexperimentalstimuliandthereforereducingtheirecologicalvalidity(TalebiandBaker

2012).Instead,inourdesignweexploitthehighvariabilityinnaturallyoccurringfalsepositivesofface

detectionmadebytheprimatevisualsystem(Taubertetal.2017).

TogetherourresultsshowthattheFFAstronglyrespondstofalsepositivefacesininanimate

objects and is tolerant to a high degree of visual variance in what constitutes a "face". Previous

arguments for the face-selectivity of the FFA have been predominantly based on its preferred

responsetofaces,typicallyquantifiedbyahighermagnitudeoffMRIBOLDpercentsignalchangeto

faces compared to other categories (Kanwisher and Yovel 2006). In addition to human faces

(Kanwisheretal.1997),theFFAisknowntorespondtoanimalfaces(Tongetal.2000),schematic

faces(Tongetal.2000),andimaginedfaces(O'CravenandKanwisher2000).Morerecently,further

corroboratingevidencehasarisenfromsuccessfuldecodingofface-relatedpropertiesfromactivation

patternsintheFFAsuchasidentity(Nestoretal.2011;Zhangetal.2016),famousfaces(Axelrodand

Yovel2015),differencesinfacialfeaturesandtheirconfiguration(GoesaertanddeBeeck2013),and

tolerancetoviewpointrotation(Anzellottietal.2014).Importantly,ourresultsshowthattheface-

selective response in the FFA is not reducible to the specific low-level visual features that often

distinguish real faces from inanimateobjects. TheFFA isnot finely-tuned toahomogenous setof

visualfeatures.Instead,wefindthattheFFAisbroadly-tunedtothedetectionofaface,independently

oftheparticularvisualfeaturesthatformthisperception.

Object recognition and face processing are two of the most impressive computational

achievements of the human visual system, and the underlying neural mechanisms are still being

revealed.Todate thedevelopmentof the theoretical framework forunderstanding the functional

architecture of ventral temporal cortex and its role in high-level visual processing has focused on

exemplarswithdistinctcategorymembership(Kravitzetal.2013;Grill-SpectorandWeiner2014).Our

results with categorically ambiguous exemplars emphasize the complexity with which visual

informationismodulatedbyhigherlevelperceptionincategory-selectiveregionsoftheventralvisual

stream.Furtherexaminationofnatural'errors'orfalsepositivessuchasinthecaseoffacepareidolia

mayhelpconstrainmodelsofhigh-levelvisualprocessinginthefuture.

19

Acknowledgements

WethankJeffMcIntoshandthestaffofMacquarieMedicalImagingatMacquarieUniversityHospital

for assistance with the operation of the MRI scanner, and Robert Keys for assistance with data

collection.ThisworkwassupportedbyanAustralianNHMRCEarlyCareerFellowship(APP1072245)

andMacquarieUniversityResearchDevelopmentGrantawardedtoS.G.W.

References

AndrewsTJ,WatsonDM,RiceGE,HartleyT.2015.Low-levelpropertiesofnaturalimagespredict

topographicpatternsofneuralresponseintheventralvisualpathway.JVis.15:3.

AnzellottiS,FairhallSL,CaramazzaA.2014.Decodingrepresentationsoffaceidentitythatare

toleranttorotation.CerebralCortex.24:1988–1995.

AxelrodV,YovelG.2015.Successfuldecodingoffamousfacesinthefusiformfacearea.PLoSONE.

10:e0117126.

BaldassiC,Alemi-NeissiA,PaganM,DiCarloJJ,ZecchinaR,ZoccolanD.2013.Shapesimilarity,better

thansemanticmembership,accountsforthestructureofvisualobjectrepresentationsina

populationofmonkeyinferotemporalneurons.PLoSComputBiol.9:e1003167.

BensonNC,ButtOH,DattaR,RadoevaPD,BrainardDH,AguirreGK.2012.Theretinotopic

organizationofstriatecortexiswellpredictedbysurfacetopology.CurrBiol.22:2081–2085.

BracciS,OpdeBeeckH.2016.DissociationsandAssociationsbetweenShapeandCategory

RepresentationsintheTwoVisualPathways.JNeurosci.36:432–444.

BracciS,RitchieJB,deBeeckHO.2017.Onthepartnershipbetweenneuralrepresentationsof

objectcategoriesandvisualfeaturesintheventralvisualpathway.Neuropsychologia.

BrainardD.1997.ThePsychophysicsToolbox.SpatVis.10:433–436.

BrettM,AntonJL,ValabregueR,PolineJB.2002.RegionofinterestanalysisusinganSPMtoolbox

[abstract]Presentedatthe8thInternationalConferenceonFunctionalMappingoftheHuman

Brain,June2-6,2002,Sendai,Japan.NeuroImage.16.

BrodeurMB,Dionne-DostieE,MontreuilT,LepageM.2010.TheBankofStandardizedStimuli

(BOSS),anewsetof480normativephotosofobjectstobeusedasvisualstimuliincognitive

research.PLoSONE.5:e10773.

BrodeurMB,GuérardK,BourasM.2014.BankofStandardizedStimuli(BOSS)phaseII:930new

normativephotos.PLoSONE.9:e106953.

CogganDD,LiuW,BakerDH,AndrewsTJ.2016.Category-selectivepatternsofneuralresponsein

theventralvisualpathwayintheabsenceofcategoricalinformation.NeuroImage.135:107–

20

114.

CoxDD,SavoyRL.2003.Functionalmagneticresonanceimaging(fMRI)“brainreading”:detecting

andclassifyingdistributedpatternsoffMRIactivityinhumanvisualcortex.NeuroImage.

19:261–270.

DavidenkoN,RemusDA,Grill-SpectorK.2012.Face-likenessandimagevariabilitydriveresponsesin

humanface-selectiveventralregions.HumBrainMapp.33:2334–2349.

EpsteinR,KanwisherN.1998.Acorticalrepresentationofthelocalvisualenvironment.Nature.

392:598–601.

GoesaertE,deBeeckHPO.2013.RepresentationsofFacialIdentityInformationintheVentralVisual

StreamInvestigatedwithMultivoxelPatternAnalyses.JNeurosci.33:8549–8558.

Grill-SpectorK,KnoufN,KanwisherN.2004.Thefusiformfaceareasubservesfaceperception,not

genericwithin-categoryidentification.NatNeurosci.7:555–562.

Grill-SpectorK,WeinerKS.2014.Thefunctionalarchitectureoftheventraltemporalcortexandits

roleincategorization.NatRevNeurosci.15:536–548.

GroenIIA,SilsonEH,BakerCI.2017.Contributionsoflow-andhigh-levelpropertiestoneural

processingofvisualscenesinthehumanbrain.PhilosTransRSocLond,B,BiolSci.

372:20160102.

GuntupalliJS,WheelerKG,GobbiniMI.2017.DisentanglingtheRepresentationofIdentityfrom

HeadViewAlongtheHumanFaceProcessingPathway.CerebralCortex.27:46–53.

HaxbyJV,GobbiniMI,FureyML,IshaiA,SchoutenJL,PietriniP.2001.Distributedandoverlapping

representationsoffacesandobjectsinventraltemporalcortex.Science.293:2425–2430.

HebartMN,GörgenK,HaynesJ-D.2014.TheDecodingToolbox(TDT):aversatilesoftwarepackage

formultivariateanalysesoffunctionalimagingdata.FrontNeuroinform.8:88.

KaiserD,AzzaliniDC,PeelenMV.2016.Shape-independentobjectcategoryresponsesrevealedby

MEGandfMRIdecoding.JNeurophysiol.jn.01074.2015.

KanwisherN.2017.TheQuestfortheFFAandWhereItLed.JNeurosci.37:1056–1061.

KanwisherN,McDermottJ,ChunMM.1997.Thefusiformfacearea:amoduleinhumanextrastriate

cortexspecializedforfaceperception.JNeurosci.17:4302–4311.

KanwisherN,YovelG.2006.Thefusiformfacearea:acorticalregionspecializedfortheperception

offaces.PhilosTransRSocLond,B,BiolSci.361:2109–2128.

KauffmannL,RamanoëlS,GuyaderN,ChauvinA,PeyrinC.2015.Spatialfrequencyprocessingin

scene-selectivecorticalregions.NeuroImage.112:86–95.

KleinerM,BrainardD,PelliD.2007.What’snewinPsychtoolbox-3.Perception.36(ECVPAbstract

Supplement}.

21

KourtziZ,KanwisherN.2000.Corticalregionsinvolvedinperceivingobjectshape.JNeurosci.

20:3310–3318.

KourtziZ,KanwisherN.2001.Representationofperceivedobjectshapebythehumanlateral

occipitalcomplex.Science.293:1506–1509.

KravitzDJ,PengCS,BakerCI.2011.Real-worldscenerepresentationsinhigh-levelvisualcortex:it's

thespacesmorethantheplaces.JNeurosci.31:7322–7333.

KravitzDJ,SaleemKS,BakerCI,UngerleiderLG,MishkinM.2013.Theventralvisualpathway:an

expandedneuralframeworkfortheprocessingofobjectquality.TrendsCognSci.17:26–49.

KriegeskorteN,KievitRA.2013.Representationalgeometry:integratingcognition,computation,and

thebrain.TrendsCognSci.17:401–412.

KriegeskorteN,MurM,BandettiniP.2008.Representationalsimilarityanalysis-connectingthe

branchesofsystemsneuroscience.FrontSystNeurosci.2:4.

KriegeskorteN,SimmonsWK,BellgowanPSF,BakerCI.2009.Circularanalysisinsystems

neuroscience:thedangersofdoubledipping.NatNeurosci.12:535–540.

LiuJ,LiJ,FengL,LiL,TianJ,LeeK.2014.SeeingJesusintoast:neuralandbehavioralcorrelatesof

facepareidolia.Cortex.53:60–77.

MacEvoySP,EpsteinRA.2009.Decodingtherepresentationofmultiplesimultaneousobjectsin

humanoccipitotemporalcortex.CurrBiol.19:943–947.

MalachR,ReppasJB,BensonRR,KwongKK,JiangH,KennedyWA,LeddenPJ,BradyTJ,RosenBR,

TootellRB.1995.Object-relatedactivityrevealedbyfunctionalmagneticresonanceimagingin

humanoccipitalcortex.ProcNatlAcadSciUSA.92:8135–8139.

MinearM,ParkDC.2004.Alifespandatabaseofadultfacialstimuli.BehavResMethodsInstrum

Comput.36:630–633.

NestorA,PlautDC,BehrmannM.2011.Unravelingthedistributedneuralcodeoffacialidentity

throughspatiotemporalpatternanalysis.ProcNatlAcadSciUSA.108:9998–10003.

NiliH,WingfieldC,WaltherA,SuL,Marslen-WilsonW,KriegeskorteN.2014.Atoolboxfor

representationalsimilarityanalysis.PLoSComputBiol.10:e1003553.

O'CravenKM,KanwisherN.2000.Mentalimageryoffacesandplacesactivatescorresponding

stimulus-specificbrainregions.JCognNeurosci.12:1013–1023.

O'TooleAJ,JiangF,AbdiH,HaxbyJV.2005.Partiallydistributedrepresentationsofobjectsandfaces

inventraltemporalcortex.JCognNeurosci.17:580–590.

ParkS,BradyTF,GreeneMR,OlivaA.2011.Disentanglingscenecontentfromitsspatialboundary:

ComplementaryrolesforthePPAandLOCinrepresentingreal-worldscenes.JNeurosci.

31:1333–1340.

22

PelliDG.1997.TheVideoToolboxsoftwareforvisualpsychophysics:transformingnumbersinto

movies.SpatVis.10:437–442.

ProklovaD,KaiserD,PeelenMV.2016.DisentanglingRepresentationsofObjectShapeandObject

CategoryinHumanVisualCortex:TheAnimate-InanimateDistinction.JCognNeurosci.28:680–

692.

RiceGE,WatsonDM,HartleyT,AndrewsTJ.2014.Low-LevelImagePropertiesofVisualObjects

PredictPatternsofNeuralResponseacrossCategory-SelectiveRegionsoftheVentralVisual

Pathway.JNeurosci.34:8837–8844.

SummerfieldC,EgnerT,MangelsJ,HirschJ.2006.Mistakingahouseforaface:neuralcorrelatesof

misperceptioninhealthyhumans.CerebCortex.16:500–508.

TalebiV,BakerCL.2012.Naturalversussyntheticstimuliforestimatingreceptivefieldmodels:a

comparisonofpredictiverobustness.JNeurosci.32:1560–1576.

TaubertJ,WardleSG,FlessertM,LeopoldDA,UngerleiderLG.2017.Facepareidoliaintherhesus

monkey.CurrBiol,27:2505–2509.e2.

TongF,NakayamaK,MoscovitchM,WeinribO,KanwisherN.2000.Responsepropertiesofthe

humanfusiformfacearea.CognNeuropsychol.17:257–280.

WardleSG,RitchieJB.2014.Canobjectcategory-selectivityintheventralvisualpathwaybe

explainedbysensitivitytolow-levelimageproperties?JNeurosci.34:14817–14819.

XiaoJ,HaysJ,EhingerKA,OlivaA,TorralbaA.2010.SUNdatabase:Large-scalescenerecognition

fromabbeytozoo.In:.Presentedatthe2010IEEEConferenceonComputerVisionandPattern

Recognition(CVPR).IEEE.p.3485–3492.

YovelG,KanwisherN.2004.Faceperception:domainspecific,notprocessspecific.Neuron.44:889–

898.

YovelG,KanwisherN.2005.Theneuralbasisofthebehavioralface-inversioneffect.CurrentBiology.

15:2256–2262.

ZhangH,JapeeS,NolanR,ChuC,LiuN,UngerleiderLG.2016.Face-selectiveregionsdifferintheir

abilitytoclassifyfacialexpressions.NeuroImage.130:77–90.