Structural Encoding of Human and Schematic FacesHolistic and Part-Based Processes

Noam Sagiv1 and Shlomo Bentin

Abstract

amp The range of specificity and the response properties ofthe extrastriate face area were investigated by comparing theN170 event-related potential (ERP) component elicited byphotographs of natural faces realistically painted portraitssketches of faces schematic faces and by nonface mean-ingful and meaningless visual stimuli Results showed thatthe N170 distinguished between faces and nonface stimuliwhen the concept of a face was clearly rendered by thevisual stimulus but it did not distinguish among differentface types Even a schematic face made from simple linefragments triggered the N170 However in a second experi-ment inversion seemed to have a different effect on naturalfaces in which face components were available and on the

pure gestalt-based schematic faces The N170 amplitude wasenhanced when natural faces were presented upside downbut reduced when schematic faces were inverted Inversiondelayed the N170 peak latency for both natural andschematic faces Together these results suggest that earlyface processing in the human brain is subserved by amultiple-component neural system in which both whole-faceconfigurations and face parts are processed The relativeinvolvement of the two perceptual processes is probablydetermined by whether the physiognomic value of thestimuli depends upon holistic configuration or whether theindividual components can be associated with faces evenwhen presented outside the face context amp

INTRODUCTION

Accumulated findings from neuropsychology as well asprimate electrophysiology point to some degree ofdomain specificity in the functional organization of theventral pathway of the visual system (for reviews seeFarah 1996 Logothetis amp Scheinberg 1996 see alsoFarah Wilson Drain amp Tanaka 1998 Logothesis ampPauls 1995 Logothetis Pauls amp Poggio 1995) Inparticular there is considerable evidence indicating thatthe perception of faces is mediated by extrastriatemechanisms specifically tuned to process physiognomicinformation (Allison Ginter et al 1994 Allison PuceSpencer amp McCarthy 1999 McCarthy Puce Gore ampAllison 1997 McCarthy 1999 McCarthy Puce Belger ampAllison 1999 Puce Allison amp McCarthy 1999 Kanw-isher McDermott amp Chun 1997) This hypothesis issupported by single-cell recordings in primates (for areview see Desimone 1991) by intracranial recordingsof event-related potentials (ERPs) in humans (eg Alli-son McCarthy Nobre Puce amp Belger 1994) as well asby hemodynamic imaging studies using PET (eg Haxbyet al 1993 Sergent Ohta amp MacDonald 1992) andfMRI (eg Kanwisher et al 1997 Puce Allison Gore ampMcCarthy 1995)

Selective activation of face-specific neural networks ata relative early stage of visual processing may increase

face recognition efficiency by restricting subsequentidentification stages to a particular class of internalmodelsmdashthat of faces (Ullman 1996) Indeed severalmodels of face recognition incorporate an initial rsquo rsquo struc-tural encodingrsquo rsquo mechanism with physiognomic featuresas input and an integrated but not fully identifiedrepresentation of a face as output (eg Moses Edelmanamp Ullman 1993 Bruce amp Young 1986) The resultingstructural code allows the identification of particularexemplars of faces despite changes in viewing angleexpression lighting age or paraphernalia According tosuch models face identification is achieved by thecombined operation of face recognition units and se-mantic nodes The face recognition units select theprestored face model that best fits the currently struc-tured perceptual representation and the semantic nodesprovide the entire knowledge associated in semanticmemory to the particular face model selected Thedomain specificity of the structural encoder and itscategorized output representation prevent the face rec-ognition system from attempting to match the resultingrepresentation to all possible 3-D object models Indeedin the absence of such specificity face recognition isinhibited and becomes inefficient (Bentin Deouell ampSoroker 1999)

To date the research into the rules governing theencoding of faces has focused mainly on the role ofstructural encoding in face identification (see reviews inBruce amp Young 1998 Ullman 1996 Rhodes 1995) ForHebrew University of Jerusalem

copy 2001 Massachusetts Institute of Technology Journal of Cognitive Neuroscience 137 pp 937ndash951

example these studies examined how variation in view-ing conditions influence face recognition (eg viewingposition and illumination direction Moses Adini ampUllman 1994) or stimulus configuration (eg verticalorientation Rhodes Brake amp Atkinson 1993 Tanaka ampFarah 1991 for a review see Valentine 1988 spatialrelationship between components Sergent 1984 oranalytic vs holistic encoding Tanaka amp Farah 1993)Relatively less effort has been directed toward investigat-ing the functional architecture of the neural mechanismthat mediates the structural encoding of faces in thevisual system Initial steps in this direction have beentaken however by recording ERPs elicited by faces andobjects in human volunteers (eg Bentin Allison PucePerez amp McCarthy 1996 George Evans Fiori Davidoffamp Renault 1996)

Bentin et al (1996) described a negative potential witha mean peak latency of 172 msec (N170) that was elicitedby human faces but not by animal faces cars or butter-flies This component was larger over the right than theleft hemisphere and largest in the posterior temporalareas Because the familiarity of the face was foundinconsequential for the N170 (Bentin amp Deouell 2000)Bentin and his colleagues suggested that although es-sential for efficient face recognition the mechanismassociated with the N170 acts on basic physiognomicfeatures and precedes within-category identificationHence the N170 may well constitute an electrophysio-logical manifestation of the neural analogue to thestructural encoder proposed by Bruce and Young(1986)mdasha mechanism specialized to detect physiognom-ic features and extracts the visual characteristics neededto form an internal representation of a human face

Several experimental manipulations were aimed atunderstanding how the structural face-encoding mech-anism functions Isolated eyes or combinations of innercomponents presented without the face contour elicitedan N170 significantly larger than that elicited by full faces(Bentin et al 1996 Bentin amp McCarthy unpublished)In the initial studies of N170 the response to invertedhuman faces was slightly more negative (ie larger) andsignificantly delayed relative to that elicited by uprightfaces (Bentin et al 1996) More recent studies showedthat indeed the N170 elicited by inverted faces issignificantly larger than that elicited by upright faces(Rossion et al 1999 2000) A possible interpretation ofthis pattern is that the amplitude enhancement andparticularly the delay of the N170 latency reflects a moredifficult encoding of inverted than of upright faces Thissuggestion is based on the assumption that face encod-ing is heavily based on holistic processing of the facegestalt which is distorted by inversion (Rossion et al1999) Surprisingly however distortion of the inner-facecomponent configuration did not affect the responsesignificantly in fact the amplitude of the N170 elicitedby faces with a distorted configuration of inner compo-nents was slightly smaller than for normal faces (Bentin

et al 1996) Furthermore the addition of the nose andthe lips to isolated eyes (in the absence of the facecontour) enhanced rather than reduced the amplitudeof the N170 and a face without inner components elicitsa significantly attenuated response (Bentin amp McCarthyunpublished) These data speak against the assumptionthat the enhancement of the N170 amplitude is associ-ated with a more difficult perceptual process Alterna-tively they suggest the possibility of having a complexneural system for structural encoding which involveboth a component processing system and a holistic faceprocessor According to this view the scalp-recordedN170 is influenced by the face component processormore than by the holistic face perception processorSupport for this hypothesis comes from recordings ofthe intracranial analogue of the N170 the N200 poten-tial These data show that recorded from the face areasin the fusiform the N200 is bigger in response to facesthan in response to face components Laterally from thefusiform however there are regions where the N200 islarger for isolated eyes than for faces (McCarthy et al1999) The present study was designed to explore someof the basic characteristics of the structural encodingmechanisms and provide evidence pertinent to the dualmechanisms hypothesis

EXPERIMENT 1

Schematic faces (eg Smiley) are instantly perceived asfaces Even a set of fruits vegetables or other objectsmay be perceived as a face when organized in a faceconfiguration Is the visual information conveyed bysuch stimuli sufficient to trigger the face-specific struc-tural encoder Note that alternatively it is possible thatthe perception of schematic faces relies on a general-purpose visual mechanism and the face representation isformed at a higher-level of perceptual integration In thepresent experiment we addressed this question bycomparing the ERPs elicited by four categories of facesphotographs of natural faces realistically painted por-traits sketches of faces and schematic faces Photo-graphs of flowers as well as equiluminant scrambledfaces and scrambled schematic faces were used as con-trol stimuli to determine face specificity (Figure 4)Participants were engaged in an oddball paradigmsilently counting butterflies that were sporadically pre-sented on the screen

Assuming that the N170 is indeed associated with thestructural encoding process we hypothesized that theminimal requirements for eliciting the N170 should alsoconstitute the minimal requirements for the perceptualsystem to determine that a face exists in the visual fieldand activate the face-specific perceptual module Hencethe modulation of the N170 by different face or face-likestimuli could indicate whether this domain-specificmodule in the visual system is constrained to processrigidly defined visual primitives or whether it can adapt

938 Journal of Cognitive Neuroscience Volume 13 Number 7

itself and be triggered by a variety of visual stimuli whichmay potentially represent a face

The components of a schematic face such as Smileydo not usually convey physiognomic information inisolation Their interpretation as face parts depends onthe configuration in which they are embedded There-fore apart from testing the range of the face structuralencoder sensitivity comparison of the evoked re-sponses to real faces with the responses to schematicfaces was expected to yield new insights regarding thefunctional organization of the structural encoding sys-tem with which the N170 ERP component is probablyassociated

Results

Replicating previous studies natural faces elicited arobust N170 component maximal at the posteriortemporal sites At the same time a positive peakwas measured at the anterior sites (P190 cf Jeffreysamp Tukmachi 1992) The negative (or negative going)ERPs elicited by equiluminant nonsense stimuli as

well as photographs of flowers within the same timerange were significantly smaller (Figure 1) The out-standingly large amplitude of the N170 elicited byfaces compared to flowers and meaningless stimuluscategories in the present study as well as similarlybig differences between the response to faces andthe response to other types of stimuli reported inprevious studies support our hypothesis that theN170 is a manifestation of an early face-specific visualprocessing mechanism Given our present interest inthis neural mechanism we will not discuss differencesbetween stimulus conditions that occurred later thanthe N170

The major manipulation in this experiment involvedpresenting different face types at different levels ofphysical resemblance to photographs of natural facesThe consequences of this manipulation are presented inFigure 2 which compares the N170 elicited at the rightand the left mastoid scalp locations by the different facecategories Relative to nonface stimuli (Figure 1) all fourface types elicited more conspicuous but not identicalN170 components

Faces

Flowers

Scrambled Faces

PO8PO7

LM RM

TP8TP7 CP5 CP6CP2CP1 CPz

FT8FC6

FC2FC1 FCzFC5

FT7

GND

FP1 FP2

POz

PO3 PO4P7

P5 P 3

O1 Oz O2

Iz

Pz

Cz

Fz

P4P6

P8

C4

T8

F8F6

F4F3F5

F7

C3

T7

IM1 IM2

Left Mastoid

0

10

Time (msec)0 100 200 300 400 500

Right Mastoid 10

Time (msec)0 100 200 300 400 500

10Fz

Time (msec)0 100 200 300 400 500

N170

mV 0

mV 0

Figure 1 Demonstration of the basic difference between the N170 elicited by faces and that elicited by nonface stimuli Note the polarity inversionat 170 msec between the posteriorndashlateral and the anterior sites

Sagiv and Bentin 939

The statistical reliability of the facenonface differenceand of the apparent differences between the N170elicited by each face type was assessed by a series ofwithin-subject three-factor analysis of variance (ANOVA)with repeated measures The factors were stimulus type(four face types flowers scrambled faces)2 hemisphere(left right) and site (P78 left and right mastoids PO78IM12)3 The dependent variables were the N170 ampli-tude and its latency (Table 1)

The ANOVA of the amplitude showed that there weresignificant differences among the different stimulustypes [F(5145) = 856 p lt 001] that the N170 waslarger at right (ndash 652 mV) than at left (ndash 492 mV)hemisphere sites [F(129) = 196 p lt 001] and thatthere was a significant effect of site [F(387) = 598 p lt005] Post hoc comparisons revealed that the N170elicited by faces (weighted across all four face catego-ries) was significantly more negative than that elicited bynonfaces (weighted across scrambled faces and flowers)[F(129) = 1836 p lt 001] The ERPs elicited by flowers

were significantly less negative than those elicited bysketches [F(129) = 561 p lt 001] but significantlymore negative than those elicited by scrambled faces[F(129) = 223 p lt 001] Because our present experi-ment focused on the potential difference in processingdifferent face types and because of the clear distinctionbetween the N170 elicited by faces and the ERPs elicitedby nonface stimuli in the same time range our subse-quent analyses focused only on the four face types Thedesign of these ANOVAs was similar to the one de-scribed above except that the stimulus type factorincluded only the four face categories

The ANOVA of the amplitude showed that all threemain effects were significant [F(387) = 110 p lt 001F(387) = 37 p lt 025 F(129) = 203 p lt 001 forstimulus type site and hemisphere respectively] Posthoc contrasts showed that the amplitude of the N170elicited by sketches was significantly less negative thanthat elicited by the three other face types [F(129) =468 p lt 001] There was no significant difference

- 0 +

SketchesPhotographs Paintings Schematic Faces

10

Time (msec)0 100 200 300 400 500

Left Mastoid 10

0Time (msec)

100 200 300 400 500

Right Mastoid

mV 0 mV 0

Figure 2 The N170 potentials elicited at the left and right mastoids by the four face types and the scalp distribution of the potentials at thislatency (spline interpolation)

940 Journal of Cognitive Neuroscience Volume 13 Number 7

between the amplitudes of N170 elicited by photo-graphs paintings and schematic faces The N170 ampli-tudes across face categories were largest at the mastoidsand in descending order IM12 P78 and PO78 Sig-nificant differences were found only between the mas-toids and the P78 and PO78 locations [F(129) = 50p lt 05 and F(129) = 48 p lt 05 respectively] andbetween the IM12 and the PO78 locations [F(129) =49 p lt 05]

All three first-order interactions were significant[F(9261) = 48 p lt 001 F(387) = 31 p lt 05F(387) = 63 p lt 01 for the Stimulus Type pound SiteStimulus Type pound Hemisphere and Site pound Hemisphererespectively] The second-order interaction between allthree factors was also significant [F(9261) = 34 p lt001] Post hoc examination of these interactions re-vealed the following pattern The difference betweenthe N170 elicited by sketches and that elicited by theother face types was larger at the right hemisphere sitesthan at the left hemisphere sites and slightly smaller atIM sites than at the other three sites The interactionbetween hemisphere and scalp location is probably aresult of the smaller (but clear) right hemisphere ad-vantage at scalp site IM12 [F(129) = 90 p lt 01]compared to other sites [F(129) = 153 p lt 001F(129) = 182 p lt 001 F(129) = 242 p lt 001 forPO78 mastoids and P78 respectively] The second-order interaction reflected (a) a smaller (Site pound Hemi-

sphere) interaction for the sketched faces category[F(387) = 23 vs F(387) gt 59 for each of the othercategories] (b) an insignificant (Stimulus Type pound Hemi-sphere) interaction at the IM scalp site [F(387) lt 100]versus a significant or close-to-significant effect for othersites [F(387) gt 24] and (c) a smaller (Stimulus Type poundSite) interaction in the right hemisphere [F(15435) =30] than in the left hemisphere [F(15435) = 55]

As evident in Table 1B the mean N170 latency wasshortest for schematic faces (1676 msec) followed byphotographs (1693 msec) paintings (1714 msec) andsketches (1743 msec) Scalp sites also differed on N170latency the site at which the shortest N170 latency wasrecorded was PO78 (1697 msec) followed by P78(1702 msec) IM12 (1704 msec) and RL-Mast (1723msec) ANOVA showed that both the stimulus type andthe site effects were significant [F(387) = 167 p lt001 F(387) = 50 p lt 005 respectively] Post hoccontrasts revealed that differences between the meanN170 latencies of all stimulus type pairs were highlysignificant [F(129) = 92 p lt 005] except for thedifference between face photograph and schematic facelatencies [F(129) = 30 p lt 1] In addition the N170latency at the mastoid scalp locations (1723 msec) wassignificantly delayed [F(129) = 173 p lt 001] com-pared with the three other locations (1701 msec)which did not significantly differ from each other[F(129) = 5 p = 6]

Table 1 (A) Mean Amplitudes (in mV) and (B) Mean Latencies (in msec) of the N170 Elicited by Natural Faces Painted PortraitsSketches Schematic Faces Flowers and Scrambled Faces at the Posterior Temporal Scalp Sites

Left Hemisphere Right Hemisphere

Left mastoid P7 PO7 IM1 IM2 PO8 P8 Right mastoid

A

Natural Faces ndash 72 ndash 62 ndash 60 ndash 76 ndash 89 ndash 93 ndash 91 ndash 97

Painted Portraits ndash 73 ndash 61 ndash 62 ndash 78 ndash 88 ndash 88 ndash 92 ndash 98

Sketches ndash 58 ndash 49 ndash 41 ndash 61 ndash 73 ndash 64 ndash 70 ndash 76

Schematic Faces ndash 68 ndash 63 ndash 60 ndash 70 ndash 83 ndash 88 ndash 96 ndash 94

Flowers ndash 29 ndash 32 ndash 23 ndash 28 ndash 30 ndash 29 ndash 33 ndash 34

Scrambled Faces ndash 19 ndash 17 ndash 07 ndash 13 ndash 15 ndash 07 ndash 16 ndash 22

B

Natural Faces 170 168 170 168 169 169 170 170

Painted Portraits 171 171 172 171 171 171 172 172

Sketches 177 171 171 174 175 173 176 178

Schematic Faces 170 165 165 168 167 167 169 169

Flowers 174 166 168 171 172 168 171 176

Scrambled Faces 177 167 166 173 175 166 168 178

Sagiv and Bentin 941

Hemispheric differences failed to reach significance[F(129) lt 10] A Site pound Hemisphere interaction washowever reliable [F(387) = 40 p lt 005] This inter-action is possibly the result of relatively late right hemi-sphere mean latency at P78 and PO78 scalp locationsThe interaction between stimulus type and site was alsosignificant [F(9261) = 35 p lt 001] reflecting a greatersite effect within the sketched face category The Stim-ulus Type pound Hemisphere interaction was not significant[F(387) = 14 p lt 25] The interaction between allthree factors did not reach significance [F(9261) = 17p = 1]

Discussion

Experiment 1 demonstrated that face-specific N170 com-ponents can be recorded with scalp electrodes not onlywhen photographs of natural faces are presented butalso when paintings sketches and even highly sche-matic drawings of faces are used as stimuli The ampli-tude of the N170 elicited by all face stimuli was at leasttwice as big as the amplitude of the negative potentialelicited by flowers and three times as big as the ampli-tude of the potential elicited by nonsense images duringthe same time The significant difference between flow-ers and scrambled faces could reflect either that theneural activity associated to the structural encoding ofany object is distinct from that elicited by visual patternsthat are not objects (McCarthy et al 1997) or that thereis specificity for other stimulus categories in the vicinityof the face area and that part of this specificity is pickedup by the far-field recording (Ishai Ungerleider MartinSchouten amp Haxby 1999) The present study was notdesigned to address this issue Therefore we will turnnow to the focus of this experiment that is the compar-ison between the different face types

The finding that the N170 elicited by schematic facesis not significantly smaller in amplitude is not delayedrelative to natural face photographs and has a similarscalp distribution suggests that common neural mech-anisms underlie the perception of both natural andschematic faces Assuming that these mechanisms arethe neural analogue to the hypothesized structuralencoder for faces these data also suggest that a face-specific visual mechanism is triggered whenever a stim-ulus contains sufficient information to generate theconcept of a face In other words schematically drawnfaces are categorized by an early visual mechanismrather than by a higher-level cognitive process Indeedit is possible that once triggered such a face perceptionmodule may take precedence over a more general visualprocessor A similar account for the suppression ofacoustic perception when phonetic features are de-tected (Bentin amp Mann 1990 Liberman amp Mattingly1989) led to the hypothesis that the phonetic modulemay preempt auditory information that is relevant tophonetic perception (Whalen amp Liberman 1987)

The reduced amplitude of the N170 elicited by detail-richer face sketches relative to that elicited by schematicfaces and its delayed peak latency might be interpretedprima facie as evidence against the above hypothesisHowever post hoc this result can be incorporated intoour theory Although graphically more detailed thesketches did not attempt to imitate a natural face (asthe paintings) or to emphasize those basic featuresapparently required for the formation of a face concept(as in the schematic faces) Indeed it is possible thatsome of the sketches were difficult to interpret ashuman faces without an elaborate analytic processHence the generation of a face concept on the basisof such sketches may have been somewhat more timeconsuming either because they were processed by gen-eral visual processors or because they activated thespecial purpose structural encoder later In additionunlike the schematic faces that although not identicalwere more uniform in shape4 the sketches were stylis-tically more diverse As a consequence some of thesketches may have introduced noise or a time jitter inthe average ERP which could explain both the reducedamplitude and the longer latency of the N170 associatedwith the sketched faces category5 All face categorieselicited a significantly larger N170 in the right hemi-sphere However the asymmetry was reduced forsketches and absent for nonface categories This findingis consistent with neuropsychological (eg De Renzi1997) neuroimaging (eg McCarthy et al 1997 DeRenzi Perani Carlesimo Silveri amp Fazio 1994 Haxbyet al 1993) and behavioral studies (eg Luh Rueckertamp Levy 1991) indicating a right hemisphere advantagefor face processing Furthermore it corroborates thehypothesis that some of the sketches were processed bya general rather than a face-specific visual processor orthat in general the sketches were less efficient ininitiating a face-specific process

The finding that schematic faces elicit a conspicuousN170 that does not significantly differ from the ERPselicited by natural faces is intriguing in light of theassumption that the N170 is associated primarily withthe processing of face components In previous studieswe have consistently found that eyes in isolation or eyescombined with other inner-face components elicit alarger N170 than that elicited by full faces (Bentinet al 1996 Bentin amp McCarthy unpublished) Further-more full faces elicit an N170 which is only slightlylarger than that elicited by face contours from which allinner components have been erased On the basis ofthese data Bentin et al (1996) suggested that the N170is modulated by two neural sources (both part of thestructural encoding mechanism) one possibly locatedin the posterior upper bank of the occipito-temporalsulcus (OTS) or in the inferior temporal (IT) gyrus istriggered by physiognomic information appearing in thevisual field the other probably located in the middlefusiform gyrus is responsible for holistic face processing

942 Journal of Cognitive Neuroscience Volume 13 Number 7

and integrates the face parts into a whole Assuming thatwithout the framing contour and the face-specific con-figuration inner components of schematic faces have nophysiognomic value6 the N170 elicited by schematicfaces should be modulated primarily by the fusiformsource which is responsible for the holistic processingof a face Consequently the N170 should be particularlysusceptible to manipulations affecting holistic process-ing of schematic faces An initial attempt to explore thispossibility is reported in Experiment 2 In addition thesimilar latency for the N170s elicited by natural faces andschematic faces would thus be in agreement with theprimacy of holistic processing for face stimuli the lateralcomponent primarily responsible for the N170 is trig-gered by the holistic component whether or not theface parts are recognizable as such in isolation ordefined by the global configuration

EXPERIMENT 2

The face inversion effectmdashthat is the difficulty to iden-tify familiar faces presented upside down (comparedwith other familiar objects subject to a similar manipu-lation)mdashhas been the focus of many studies for morethan two decades (reviewed by Valentine 1988) Thedifference between the effects of face inversion relativeto inversion of other objects is usually attributed to abasic difference between the processing strategies in-curred by faces versus objects According to this hypoth-esis recognition of faces relies on the formation of awhole-face representation from which relations betweenthe inner-face parts are derived that is holistic process-ing (Farah Tanaka amp Drain 1995 Tanaka amp Farah 1993versions of the holistic hypothesis are reviewed byMoscovitch Winocur amp Behrmann 1997) In contrastthe recognition of other perhaps less stereotypic ob-jects entails initial identification of basic componentswhich are then related one to the other to form anunequivocal visual percept (eg Biderman 1987 Marr1982 Marr amp Nishihara 1978) The face inversion effecthas become the hallmark of normal face recognition inhumans and has often been used to distinguish face-specific from general visual processes

Bentin et al (1996) compared the N170 elicited byupright and inverted faces and found that the N170elicited by inverted faces was slightly larger and peaked10 msec later than that elicited by upright faces only thedifference in latency was significant in that study Otherinvestigators however reported that the enhancementof the N170 amplitude by face inversion is significant aswell as the delay in latency (eg Rossion et al 19992000) A possible interpretation of the enhanced N170amplitude in response to inverted faces is that the scalp-recorded N170 is associated primarily with the percep-tion of face components a process that might not beimpeded by inversion This hypothesis conforms to thepresumed anatomical location of the putative compo-

nents detector (in lateral OTS and IT gyrus) and theorientation of these cell assemblies relative to the re-cording electrodes on the scalp In contrast to the N170the intracranially recorded N200 (on the ventral plane ofthe temporal lobe) is modulated primarily by the holis-tic face integration mechanism located in the posteriorlateral parts of the fusiform gyrus (eg Allison et al1999) Hence given the orientation of the electricaldipole fields in these two cell assemblies a manipulationthat inhibits holistic perception should enhance theN170 amplitude This hypothesis also conforms to thefinding that the N200 (recorded at the fusiform) elicitedby faces is rsquo rsquo largerrsquo rsquo than that elicited by isolated eyes(McCarthy et al 1999) whereas the N170 (recorded atthe scalp) elicited by faces is rsquo rsquosmallerrsquo rsquo than that elicitedby isolated eyes (Bentin et al 1996)

According to this model face perception is specificbecause the highly stereotypic organization of facespermits categorizing the visual stimulus before a fulldescription of its features is provided by the analyticmechanism Therefore most of the time faces activatethe holistic perception mechanism instantly Indeedonly such a model has the capacity to explain the resultsof Experiment 1 in which a visual stimulus that can beperceived as a face only if processed as a whole (ie ahighly schematic face) elicited a similar N170 to thatelicited by photographs of natural faces

Based on the above conceptualization we hypothe-size that if the detection of schematic faces is basedprimarily on the holistic face configuration compromis-ing their gestalt by presenting the stimuli upside downshould reduce the N170 In contrast turning naturalfaces upside down should enhance the amplitude of theN170 since the face component processing modulewould still be activated To test this hypothesis wepresented a new group of participants with naturaland schematic faces in upright and inverted orienta-tions

Results

Figure 3 presents the N170s elicited by photographs ofnatural faces and schematic faces presented upright orupside down Replicating the findings of Rossion et al(1999) the N170 elicited by inverted natural faces waslarger than that elicited by upright faces In contrast theN170 elicited by inverted schematic faces was reduced inamplitude relative to that elicited by upright schematicfaces Inversion delayed the peak latency of the N170 forschematic as well as for natural faces

The reliability of these observations was statisticallyanalyzed by a within-subject ANOVA using repeatedmeasures (see data in Table 2A)

The factors were stimulus type (natural faces sche-matic faces) stimulus orientation (upright inverted) site(P78 mastoids PO78 IM12) and hemisphere (leftright) As a main effect stimulus orientation was not

Sagiv and Bentin 943

significant [F(115) = 21 p = 16] However the inter-esting pattern of the face inversion effects was revealedby a significant interaction between the effects of orien-tation and stimulus type [F(115) = 248 p lt 001]followed by planed comparisons The planed compari-sons showed that whereas for natural photographs theN170 elicited by inverted faces (ndash 66 mV) was larger thatthat elicited by upright faces (ndash 55 mV) with schematicfaces inversion was associated with a reduction inthe N170 amplitude (ndash 38 and ndash 50 mV for invertedand upright presentation respectively t(15) = 2104p lt 055 and t(15) = 3000 p lt 01 for natural andschematic faces respectively) The main effect of stim-ulus type was significant [F(115) = 302 p lt 001] but itwas also qualified by its interaction with the effect oforientation Post hoc comparisons showed that whereasthe difference between the N170 elicited by invertednatural and schematic faces was significant [t(15) =4264 p lt 001] the difference between the N170elicited by natural and schematic faces presented uprightwas not [t(15) = 884 p = 39] The main effect of site

was also significant [F(345) = 68 p lt 01] revealingonly that the N170 amplitudes at P7 and P8 were smallerthan at all other sites [F(115) = 117 p lt 01] It isnoteworthy that as evidence by an insignificant interac-tion with stimulus type [F(345) = 13 p = 28] thiseffect was similar for schematic and natural faces On theother hand a significant main effect of hemisphere[F(115) = 65 p lt 05] was qualified by its interactionwith stimulus type [F(115) = 53 p lt 05] Post hocanalysis revealed that whereas for both stimulus typesthe amplitude of the N170 was larger over the rightcompared to the left hemisphere the interhemisphericdifference was significant for natural faces [t(15) = 335p lt 01] but not for schematic faces [t(15) = 1408p = 18] The effects of the Hemisphere pound Orientationinteraction was significant [F(115) = 47 p lt 5] Thisinteraction was caused by the fact that whereas acrossstimulus type the reduction and enhancement of theN170 amplitude by inversion canceled each other in theleft hemisphere in the right hemisphere the enhance-ment of N170 induced by the inversion of natural faces

Faces Inverted Faces Schematic Faces Inverted SchematicFaces

Time(msec) Time(msec)0 100 200 300 400

10

mV 0

0 100 200 300 400

10PO7 PO8

- 0 +

mV 0

Figure 3 The N170 potentials elicited at the left and right mastoids by upright and inverted natural faces and schematic faces and the scalpdistribution at this latency

944 Journal of Cognitive Neuroscience Volume 13 Number 7

weighted more than its reduction resulting from theinversion of schematic faces However the second-orderinteraction between orientation hemisphere and stim-ulus type was not significant [F(115) lt 10] Finally aseries of second-order interactions involving the effect ofsite suggested that all the effects were differently con-spicuous at different sites As revealed in Table 2 by andlarge it seems that whereas natural faces induced similarinversion effects across the analyzed sites the inversioneffect for schematic faces varied at different sites We hadno prior assumptions about the distribution of the effectswithin the posterior temporal scalp regions Further-more the spatial resolution of scalp-recorded ERPs istoo low to permit reliable interpretation of site effectswithin a relative small region unless the same distribu-tion is replicated therefore a detailed analysis of thisdistribution of effects was not attempted post hoc

ANOVA of the N170 peak latencies showed a signifi-cant main effect of stimulus orientation [F(115) =1615 p lt 001] which significantly interacted withstimulus type [F(115) = 62 p lt 025] Plannedcomparisons showed that for both natural and sche-matic faces inversion delayed the N170 significantly[t(15) = 10821 p lt 001 and t(15) = 3626 p lt 01respectively] However as evident in Table 2B thisdifference was larger for natural than for schematicfaces The latencies were similar over the left and righthemispheres [F(115) = 19 p = 185] The main effectof site was significant [F(345) = 47 p lt 01] butbecause the differences between sites were all smaller

than the sampling rate this effect was not investigatedfurther No other significant effects emerged in thisanalysis

Discussion

The results of the Experiment 2 validated our predic-tion that inversion should reduce the amplitude of theN170 elicited by schematic faces but have little effector enhance the N170 elicited by natural faces Thenatural face inversion effect observed in the presentstudy replicates previous results reported by Rossionet al (1999) These authors suggested that the en-hancement in the N170 amplitude reflects the en-hanced difficulty of the holistic processor to encodethe inverted face Indeed the delay in the N170 peaklatency elicited by inverted relative to upright faces iscongruent with such a view However if this were thecase we should have found similar effects of inversionon the N170 amplitude elicited by schematic and bynatural faces Furthermore the assumption that moredifficult face encoding induces an enhancement of theN170 amplitude is also at odds with fMRI data showingthat face inversion reduces the activity in the fusiformface areas relative to the activity elicited by uprightfaces (Kanwisher Tong amp Nakayama 1998) Indeedfinding opposite effects of inversion for schematic andnatural faces can more likely be accommodated by ourhypothesis that the larger N170 amplitude in responseto inverted natural faces reflects its stronger associa-

Table 2 (A) Amplitudes (in mV) and (B) Latencies (in msec) of the N170 Elicited by Upright and Inverted Natural and SchematicFaces at the Posterior Temporal Scalp Sites

Natural Faces Schematic Faces

Right Hemisphere Left Hemisphere Right Hemisphere Left Hemisphere

Upright Inverted Upright Inverted Upright Inverted Upright Inverted

A

Mastoid ndash 64 ndash 86 ndash 58 ndash 82 ndash 54 ndash 55 ndash 51 ndash 35

P87 ndash 48 ndash 68 ndash 45 ndash 54 ndash 41 ndash 30 ndash 32 ndash 33

PO87 ndash 68 ndash 88 ndash 48 ndash 50 ndash 56 ndash 56 ndash 50 ndash 23

IM21 ndash 53 ndash 81 ndash 49 ndash 73 ndash 50 ndash 29 ndash 65 ndash 39

Mean ndash 58 ndash 81 ndash 50 ndash 65 ndash 50 ndash 425 ndash 49 ndash 325

B

Mastoid 149 162 151 165 159 164 158 165

P87 151 163 154 166 158 164 162 166

PO87 151 160 152 164 151 161 159 166

IM21 150 162 147 161 161 164 154 162

Mean 150 162 151 164 157 163 158 165

Sagiv and Bentin 945

tion with a face component analyzer in the lateraltemporal gyri than with the holistic face processor inthe fusiform

Our model proposes that although schematic andnatural faces probably activate similar neural networksin the extrastriate visual pathway these mechanismsare not identical Natural faces probably activate twoperceptual modules one specialized to detect andprocess physiognomic features in the visual field andthe other specialized in the holistic processing of facesor forming a holistic face representation Unlike naturalfaces the components of schematic faces are notperceived as carrying physiognomic information outof the schematic face gestalt Therefore in contrastto natural faces schematic faces may trigger the holisticprocessor but they do not trigger directly the analysisof the components The latter mechanism may beactivated only subsequent to the formation of theholistic face representation or at the very least follow-ing this process rsquo rsquo in cascadersquo rsquo (McClelland 1979) Con-sequently adversely affecting the perception of the facegestalt (by inverting the stimuli) inhibited both holisticperception and the perception of individual compo-nents and this inhibition reduced the amplitude of theN170 Such a pattern is also in agreement with themyriad of studies supporting the primacy of holisticprocessing of face stimuli

GENERAL DISCUSSION

In the present study we investigated several character-istics of the extrastriate structural face-encoding mech-anisms by comparing the N170 elicited by faces differingin their naturalistic aspect and by comparing the faceinversion effects on the N170 elicited by photographs ofnatural faces and by schematically drawn faces To theextent that the N170 is indeed associated with structuralencoding we found that this process distinguishes wellbetween faces and nonfaces stimuli but not amongdifferent face types that render the concept of a faceclearly However we also found evidence that schematicfaces and natural faces are not processed identicallyWhereas inversion of natural faces enhanced the ampli-tude of the N170 inversion of schematic faces reducedits amplitude For both face types the latency of theN170 peak was significantly delayed by inversion Thispattern of results suggests that the face-specific struc-tural encoder can be triggered by a variety of stimuli ifthey include a canonical face configuration Howeverthe process of encoding the face information and form-ing a structural representation is probably differentwhen the physiognomic value of the stimuli dependsupon holistic configuration as opposed to when theindividual components can be associated with faces evenwhen presented outside the face context

Apart from the neurofunctional specificity issue anenduring question in face perception research is whether

the formation of a mental representation of the faceentails extraction of a holistic configuration directly fromthe visual input or whether it is feature-based that iswhether for faces as for other complex visual stimulistructural encoding is based on an initial analysis of parts(for a review of this debate see eg Farah 1990 Farahet al 1995) In principle because faces form a veryhomogeneous and stereotypical category they lendthemselves to holistic perception Holistic processing isadvantageous in face recognition because face compo-nents are fairly similar across individuals and thereforethe distinction among faces is usually based on theinterrelationships between the different face compo-nents and their overall configuration (eg Tanaka ampFarah 1993 Young Hellawell amp Hay 1987)7 Moreoverour memory for faces is usually based on nondecompos-able holistic images which at the interface betweenperception and memory are most efficiently addressedby integrated face representations8However if the well-learned gestalt is altered and holistic perception is notpossible it is probably replaced by a more analyticcomponent-based process of structural encoding Thisanalytic process may inhibit or at least delay the per-ception of the spatial relationships between face com-ponents and therefore may impede identificationIndeed a major source of evidence supporting theholistic view is that faces are more difficult to recognizewhen presented upside down (ie compromising theirfamiliar gestalt) than when presented in upright orien-tation (see reviews in Farah et al 1995 Rhodes et al1993 Valentine 1988)

Having to cope with different types of face input thevisual system must be able to encode the face config-uration directly from the visual display as well as byintegrating face components encoded individuallyThese two processes may concur triggered independ-ently by relevant physiognomic input and performed bydifferent neural circuits to form a complex structuralencoding mechanism in the extrastriate cortex Further-more the two systems may interact supporting orinhibiting each other Hence a regular face presentedin upright orientation may activate the holistic percep-tion process which in turn may inhibit superfluousanalysis of individual components By contrast corrupt-ing the face gestalt by presenting faces in an unusualorientation may trigger (and facilitate) face componentanalysis which should not depend on a particular con-figuration In such a case the integration of the compo-nents into a whole by the holistic processor will followtheir independent analysis assisted by the feature-detection system This hypothesis may explain for in-stance why within the context of a regular face individ-ual face parts are distinguished faster than in isolationbut if the configuration of the face is altered the partsare distinguished as fast as when presented in isolation(Tanaka amp Farah 1993 Sergent 1984)9 The ability toprocess face information both holistically and piecewise

946 Journal of Cognitive Neuroscience Volume 13 Number 7

was implemented in a neural model initially based onscalp recordings (Bentin et al 1996) and more recentlycorrected improved and elaborated on the basis intra-cranial recordings (McCarthy et al 1999 for a detaileddescription and a review of supportive functional MRIdata see McCarthy 1999) This model suggests that theneural mechanism dedicated to structural encoding offaces is a complex system composed of at least twoseparate subsystems One probably located in thelateral part of the middle fusiform gyrus is responsiblefor processing the configuration of the face as a wholeand provides the integrated face representation that isnecessary for face recognition and identification Theother possibly distributed over posterior regions in theIT and middle temporal (MT) gyri and in the OTS isresponsible for detecting physiognomic informationregardless of the spatial configuration of the items thatcarry it The fusiform gyrus is farther away from therecording sites at the scalp than where the temporalgyri are and the orientation of its cells forms anelectrical dipole source that is almost perpendicular tothe lateral aspects of the scalp Hence although bothsources contribute to the modulation of the face-spe-cific N170 the contribution of the posterior temporalface component processor should be considerably moreconspicuous

According to this hypothesis the opposite effects ofinversion on the N170 elicited by natural and schematicfaces might be explained assuming that the holisticencoding system in the fusiform gyrus indeed inhibitsthe component analysis in the lateral temporal lobe facesites Hence if the face stimulus can trigger both theholistic and the component analysis (as it is true fornatural faces) an experimental manipulation that inter-feres with holistic processing (like the face inversionprocedure in the present experiment) should reducethe amount of inhibition of component analysis andconsequently the amplitude of the N170 would beenhanced However if the rsquo rsquocomponentsrsquorsquo do not conveyphysiognomic information so that their analysis by aface-specific mechanism rsquo rsquodependsrsquo rsquo on the previousrecognition of the face gestalt (as it is true for schematicfaces) interfering with holistic perception cannot facil-itate the activity of the component analysis and there-fore the modulation of the N170 would only reflect theadditional difficulty to encode the face ie reducedamplitude and delayed latency

The observed effects of the different factors on theN170 peak latency are nicely aligned with the abovemodel Although stimulus inversion delayed the N170for both natural and schematic faces this delay was twiceas big (12 msec) for former stimuli than for the latter (6msec) This interaction suggests that the inversion had abigger effect on natural than on schematic faces Thismay be because regardless of whether they are upright orinverted the physiognomic value of schematic faces isbased on their gestalt Hence inversion does inhibit the

processing of the gestalt but does not induce a change inprocessing strategy On the other hand natural faces areanalyzed holistically if they are upright but inversionrequires a change in that strategy encouraging theanalysis of components The significantly faster responseto upright natural than to upright schematic faces inExperiment 2 suggests that the extraction of the facegestalt is easier if the face is more natural

There is ample evidence indicating that the righthemisphere is specialized in global or holistic processingof objects and particularly faces whereas the left hemi-sphere is specialized in part-based processing (egRobertson amp Delis 1986 for a review see Bradshow ampNettleton 1983 for specific reference to face perceptionsee eg Bruce 1998 Rhodes et al 1993 Corballis1988 Bradshow amp Sherlock 1982 Gilbert amp Bakan1973) Consequently interfering with holistic processingshould have had stronger effects in the right than in theleft hemisphere Our results confirmed this predictiononly partially The pronounced interhemispheric asym-metry with larger N170 amplitudes at the right than atthe left hemisphere sites for all stimulus types suggests apredominant holistic process for all face categoriesInterestingly this asymmetry was reduced (in Experi-ment 1) for sketches relative to the other stimulus typesThe relatively enhanced involvement of the left hemi-sphere in processing the sketches is in agreement withour suggestion that the face-specific structural encodingsystem was activated less efficiently by sketches than byother face types10 Hence the same factor that may haveconstrained the processing of schematic faces may alsohave constrained the processing of sketches but in theopposite direction Whereas schematic faces enhancedholistic processing by providing a clear face gestalt whileobscuring the physiognomic value of face componentsthe sketches reduced holistic processing by blurring orveiling the face gestalt Also in Experiment 2 the N170was larger over right than over left hemisphere sitesHowever because the interaction between the inversionand hemisphere effects in Experiment 2 was not modu-lated by stimulus type (ie the three-way interactionwas not significant) our interpretation of the interhemi-spheric asymmetry (particularly in Experiment 2) mustbe limited

In conclusion the present scalp-recorded N170 dataconverge with the previously published pattern of theintracranially recorded N200 data (for a comprehensivereview see Allison et al 1999 McCarthy et al 1999Puce et al 1999) supporting a multifunctional organ-ization of the structural encoding mechanism for facesApparently distinct neural circuits are responsible forthe encoding and initial analysis of face components andfor holistic perception of a full-face configuration Theformer is probably located in the lateral posterior tem-poral lobe whereas the latter is located in the middlefusiform gyrus Face perception usually entails predom-inantly holistic processing which explains the right

Sagiv and Bentin 947

hemisphere dominance in performance as well as inelectrophysiological and neuroimaging data Howeverboth mechanisms can be triggered independently byrelevant information

METHODS

Participants

Right-handed participants were recruited among volun-teer students from the Hebrew University They receivedeither experimental credit hours or payment for theirparticipation Thirty participants (13 males and 17 fe-males) were tested in the first experiment and 16participants (8 males and 8 females) were tested in thesecond experiment Participantsrsquo ages ranged from 19 to37 years (mean = 233 years) They all had normal orcorrected-to-normal vision

Stimuli

Stimuli were scanned using a desk scanner processedwith commercially available graphic software and con-

verted to 16 color grayscale 248 by 248 pixel imagesThere were eight stimulus categories four face types(photographs of natural faces paintings face sketchesand schematic faces) two nonsense patterns (formedby scrambling the natural faces and the schematicfaces) flowers and butterflies (Figure 4) Paintingsand sketches of faces were scanned from art booksPaintings were typically realistic Renaissance portraitsand face sketches did not portray details and usuallylacked shading and texture Schematic face parts weredrawn by hand scanned into the computer and latercombined into different schematic faces Face photo-graphs and paintings contained many shades of grayand the sketches and the schematic faces were basicallyblack on white All face stimuli were front views

Scrambled faces were created by scrambling thephases of the two-dimensional Fourier transform andthen smoothing the uneven edges created by this proc-ess The original and the scrambled stimuli were equallyluminous following this procedure Each stimulus cat-egory comprised 75 stimuli except for the butterflycategory which consisted of 30 stimuli In Experiment2 only natural faces schematic faces and flowers were

Face Stimuli

Photographs Painted portraits Sketches Schematic faces

Nonface Stimuli

Scrambledphotographs

Butterflies(targets) Flowers Scrambled

schematic faces

Figure 4 Example of the eight stimulus types used in the present study

948 Journal of Cognitive Neuroscience Volume 13 Number 7

used The faces were presented in both upright andupside down orientation Upright stimuli were the samein Experiments 1 and 2

All stimuli were presented for 350 msec with anaverage interstimulus interval (ISI) of 12 sec and viewedfrom a distance of 110 plusmn 10 cm subtending a viewingangle of approximately 68 Stimulus timing was con-trolled by the MEL software system (Psychology Soft-ware Tools Pittsburgh PA)

Task

Participants were instructed to keep a mental countof stimuli belonging to one category butterflies inthe first experiment and flowers in the secondHence all of the experiment-relevant categories werefrom the participantsrsquo point of view nontargets Thetargets were presented in approximately 5 (Experi-ment 1) and 6 (Experiment 2) of the trials Duringeach break and at the end of the experimentparticipants reported the number of targets countedA few participants who did not count correctly wereexcluded

EEG Recording and Data Analysis

Continuous EEG was recorded from 48 scalp locationsusing tin electrodes attached to elastic electrode caps(Electrode Caps International Ohio) A schematic illus-tration of the electrode locations and names is pre-sented in Figure 1 A ground electrode was located onthe forehead and a reference electrode was located onthe tip of the nose Two additional electrodes located atthe outer canthus and the infraorbital region of the righteye measured vertical and horizontal electrooculograms(EOGs)

The EEG was sampled at 250 Hz by a DAP2400 DataAcquisition Processor (Microstar Laboratories) ampli-fied 20000-fold using a battery-operated HU-5072 Iso-lated Bioelectric Amplifier (SA Instrumentation SanDiego) with a frequency bandpass of 01ndash30 Hz andstored on disk for off-line averaging Electrode impe-dance was below 5 k A digital filter of 08ndash16 Hz wasapplied before analysis

Codes synchronized to stimulus delivery were used toselectively average epochs associated with differentstimulus categories Epochs contaminated with saccadesor eye blinks were excluded using root-mean-squareEOG-values as criteria Typically epochs associated with55ndash70 (never less than 50) out of 75 stimuli in eachcategory were averaged

The mean and peak amplitudes and peak latencies ofspecific ERP components within bounding intervals werecomputed for each category for each participant Sig-nificance of the differences between these measures wastested using a repeated-measures ANOVA Where appro-priate the more conservative GreenhousendashGeisser ad-

justed df values were used ANOVAs were followed byunivariate F contrasts

Acknowledgments

This study was supported by grants from the USndashIsraelBinational Science Foundation and Israeli Foundations Trust-ees to S Bentin We thank Amir Katz for his help in preparingthe stimuli and running Experiment 1 as well as Leon DeouellMichael Fink Ayelet Landau Yulia Goland David Carmel GilRaviv Amir Raz and Dan Tzur for their support We are alsovery grateful to Jerry Levinson Alex List and Lynn Robertsonfor their comments on the manuscript

Reprint requests should be sent to Shlomo Bentin TheHebrew University of Jerusalem Mount Scopus JerusalemIsrael 91905 or via e-mail msbentinmscchujiacil

Notes

1 Present address University of California Berkeley2 There was no reliable difference between the ERP elicited

by scrambled natural faces and scrambled schematic facesTherefore these two categories were collapsed into one3 These were the sites at which the N170 was most

conspicuous4 So were in fact all face categories except for the sketches5 Note that this result also contradicts the view that a more

difficult encoding process elicit N170 of a higher amplitude6 This assumption was indeed supported by a current study

(Bentin Mecklinger Bosch Sagiv amp von Cramon 1999)7 Of course some faces have very distinctive components

which can be used as markers for recognition Such markersare frequently exaggerated and used as distinctive features incaricatures (Rhodes 1987)8 Distinctive features (such as unusual components birth-

marks etc) may be stored in semantic memory as part of apersonrsquo s face entry Such information may have visual qualitiesthat can be addressed directly during visual matching tofacilitate identification9 Note that this account for the rsquorsquoface superiority effectrsquo rsquo is

identical to a common explanation of the rsquorsquoword superiorityeffectrsquo rsquo in the identification of single letters Indeed ourassumption about the coexistence of rsquo rsquowhole-facersquo rsquo and rsquo rsquoface-partrsquo rsquo processors resonates to the rsquorsquodual routersquo rsquo model of visualword processing It is quite possible that identical principlesgovern all the content-specific perceptual modules in the visualsystem10 Recall that relative to other face types the reduction of theN170 amplitude elicited by sketches at right hemisphere siteswas greater then at left hemisphere sites

REFERENCES

Allison T Ginter H McCarthy G Nobre A C Puce ALuby M amp Spencer D D (1994) Face recognition in hu-man extrastriate cortex Journal of Neurophysiology 71821ndash825

Allison T McCarthy G Nobre A C Puce A amp Belger A(1994) Human extrastriate visual cortex and the perceptionof faces words numbers and colors Cerebral Cortex 5544ndash554

Allison T Puce A Spencer D D amp McCarthy G (1999)Electrophysiological studies of human face perception I

Sagiv and Bentin 949

example these studies examined how variation in view-ing conditions influence face recognition (eg viewingposition and illumination direction Moses Adini ampUllman 1994) or stimulus configuration (eg verticalorientation Rhodes Brake amp Atkinson 1993 Tanaka ampFarah 1991 for a review see Valentine 1988 spatialrelationship between components Sergent 1984 oranalytic vs holistic encoding Tanaka amp Farah 1993)Relatively less effort has been directed toward investigat-ing the functional architecture of the neural mechanismthat mediates the structural encoding of faces in thevisual system Initial steps in this direction have beentaken however by recording ERPs elicited by faces andobjects in human volunteers (eg Bentin Allison PucePerez amp McCarthy 1996 George Evans Fiori Davidoffamp Renault 1996)

Bentin et al (1996) described a negative potential witha mean peak latency of 172 msec (N170) that was elicitedby human faces but not by animal faces cars or butter-flies This component was larger over the right than theleft hemisphere and largest in the posterior temporalareas Because the familiarity of the face was foundinconsequential for the N170 (Bentin amp Deouell 2000)Bentin and his colleagues suggested that although es-sential for efficient face recognition the mechanismassociated with the N170 acts on basic physiognomicfeatures and precedes within-category identificationHence the N170 may well constitute an electrophysio-logical manifestation of the neural analogue to thestructural encoder proposed by Bruce and Young(1986)mdasha mechanism specialized to detect physiognom-ic features and extracts the visual characteristics neededto form an internal representation of a human face

Several experimental manipulations were aimed atunderstanding how the structural face-encoding mech-anism functions Isolated eyes or combinations of innercomponents presented without the face contour elicitedan N170 significantly larger than that elicited by full faces(Bentin et al 1996 Bentin amp McCarthy unpublished)In the initial studies of N170 the response to invertedhuman faces was slightly more negative (ie larger) andsignificantly delayed relative to that elicited by uprightfaces (Bentin et al 1996) More recent studies showedthat indeed the N170 elicited by inverted faces issignificantly larger than that elicited by upright faces(Rossion et al 1999 2000) A possible interpretation ofthis pattern is that the amplitude enhancement andparticularly the delay of the N170 latency reflects a moredifficult encoding of inverted than of upright faces Thissuggestion is based on the assumption that face encod-ing is heavily based on holistic processing of the facegestalt which is distorted by inversion (Rossion et al1999) Surprisingly however distortion of the inner-facecomponent configuration did not affect the responsesignificantly in fact the amplitude of the N170 elicitedby faces with a distorted configuration of inner compo-nents was slightly smaller than for normal faces (Bentin

et al 1996) Furthermore the addition of the nose andthe lips to isolated eyes (in the absence of the facecontour) enhanced rather than reduced the amplitudeof the N170 and a face without inner components elicitsa significantly attenuated response (Bentin amp McCarthyunpublished) These data speak against the assumptionthat the enhancement of the N170 amplitude is associ-ated with a more difficult perceptual process Alterna-tively they suggest the possibility of having a complexneural system for structural encoding which involveboth a component processing system and a holistic faceprocessor According to this view the scalp-recordedN170 is influenced by the face component processormore than by the holistic face perception processorSupport for this hypothesis comes from recordings ofthe intracranial analogue of the N170 the N200 poten-tial These data show that recorded from the face areasin the fusiform the N200 is bigger in response to facesthan in response to face components Laterally from thefusiform however there are regions where the N200 islarger for isolated eyes than for faces (McCarthy et al1999) The present study was designed to explore someof the basic characteristics of the structural encodingmechanisms and provide evidence pertinent to the dualmechanisms hypothesis

EXPERIMENT 1

Schematic faces (eg Smiley) are instantly perceived asfaces Even a set of fruits vegetables or other objectsmay be perceived as a face when organized in a faceconfiguration Is the visual information conveyed bysuch stimuli sufficient to trigger the face-specific struc-tural encoder Note that alternatively it is possible thatthe perception of schematic faces relies on a general-purpose visual mechanism and the face representation isformed at a higher-level of perceptual integration In thepresent experiment we addressed this question bycomparing the ERPs elicited by four categories of facesphotographs of natural faces realistically painted por-traits sketches of faces and schematic faces Photo-graphs of flowers as well as equiluminant scrambledfaces and scrambled schematic faces were used as con-trol stimuli to determine face specificity (Figure 4)Participants were engaged in an oddball paradigmsilently counting butterflies that were sporadically pre-sented on the screen

Assuming that the N170 is indeed associated with thestructural encoding process we hypothesized that theminimal requirements for eliciting the N170 should alsoconstitute the minimal requirements for the perceptualsystem to determine that a face exists in the visual fieldand activate the face-specific perceptual module Hencethe modulation of the N170 by different face or face-likestimuli could indicate whether this domain-specificmodule in the visual system is constrained to processrigidly defined visual primitives or whether it can adapt

938 Journal of Cognitive Neuroscience Volume 13 Number 7

itself and be triggered by a variety of visual stimuli whichmay potentially represent a face

The components of a schematic face such as Smileydo not usually convey physiognomic information inisolation Their interpretation as face parts depends onthe configuration in which they are embedded There-fore apart from testing the range of the face structuralencoder sensitivity comparison of the evoked re-sponses to real faces with the responses to schematicfaces was expected to yield new insights regarding thefunctional organization of the structural encoding sys-tem with which the N170 ERP component is probablyassociated

Results

Replicating previous studies natural faces elicited arobust N170 component maximal at the posteriortemporal sites At the same time a positive peakwas measured at the anterior sites (P190 cf Jeffreysamp Tukmachi 1992) The negative (or negative going)ERPs elicited by equiluminant nonsense stimuli as

well as photographs of flowers within the same timerange were significantly smaller (Figure 1) The out-standingly large amplitude of the N170 elicited byfaces compared to flowers and meaningless stimuluscategories in the present study as well as similarlybig differences between the response to faces andthe response to other types of stimuli reported inprevious studies support our hypothesis that theN170 is a manifestation of an early face-specific visualprocessing mechanism Given our present interest inthis neural mechanism we will not discuss differencesbetween stimulus conditions that occurred later thanthe N170

The major manipulation in this experiment involvedpresenting different face types at different levels ofphysical resemblance to photographs of natural facesThe consequences of this manipulation are presented inFigure 2 which compares the N170 elicited at the rightand the left mastoid scalp locations by the different facecategories Relative to nonface stimuli (Figure 1) all fourface types elicited more conspicuous but not identicalN170 components

Faces

Flowers

Scrambled Faces

PO8PO7

LM RM

TP8TP7 CP5 CP6CP2CP1 CPz

FT8FC6

FC2FC1 FCzFC5

FT7

GND

FP1 FP2

POz

PO3 PO4P7

P5 P 3

O1 Oz O2

Iz

Pz

Cz

Fz

P4P6

P8

C4

T8

F8F6

F4F3F5

F7

C3

T7

IM1 IM2

Left Mastoid

0

10

Time (msec)0 100 200 300 400 500

Right Mastoid 10

Time (msec)0 100 200 300 400 500

10Fz

Time (msec)0 100 200 300 400 500

N170

mV 0

mV 0

Figure 1 Demonstration of the basic difference between the N170 elicited by faces and that elicited by nonface stimuli Note the polarity inversionat 170 msec between the posteriorndashlateral and the anterior sites

Sagiv and Bentin 939

The statistical reliability of the facenonface differenceand of the apparent differences between the N170elicited by each face type was assessed by a series ofwithin-subject three-factor analysis of variance (ANOVA)with repeated measures The factors were stimulus type(four face types flowers scrambled faces)2 hemisphere(left right) and site (P78 left and right mastoids PO78IM12)3 The dependent variables were the N170 ampli-tude and its latency (Table 1)

The ANOVA of the amplitude showed that there weresignificant differences among the different stimulustypes [F(5145) = 856 p lt 001] that the N170 waslarger at right (ndash 652 mV) than at left (ndash 492 mV)hemisphere sites [F(129) = 196 p lt 001] and thatthere was a significant effect of site [F(387) = 598 p lt005] Post hoc comparisons revealed that the N170elicited by faces (weighted across all four face catego-ries) was significantly more negative than that elicited bynonfaces (weighted across scrambled faces and flowers)[F(129) = 1836 p lt 001] The ERPs elicited by flowers

were significantly less negative than those elicited bysketches [F(129) = 561 p lt 001] but significantlymore negative than those elicited by scrambled faces[F(129) = 223 p lt 001] Because our present experi-ment focused on the potential difference in processingdifferent face types and because of the clear distinctionbetween the N170 elicited by faces and the ERPs elicitedby nonface stimuli in the same time range our subse-quent analyses focused only on the four face types Thedesign of these ANOVAs was similar to the one de-scribed above except that the stimulus type factorincluded only the four face categories

The ANOVA of the amplitude showed that all threemain effects were significant [F(387) = 110 p lt 001F(387) = 37 p lt 025 F(129) = 203 p lt 001 forstimulus type site and hemisphere respectively] Posthoc contrasts showed that the amplitude of the N170elicited by sketches was significantly less negative thanthat elicited by the three other face types [F(129) =468 p lt 001] There was no significant difference

- 0 +

SketchesPhotographs Paintings Schematic Faces

10

Time (msec)0 100 200 300 400 500

Left Mastoid 10

0Time (msec)

100 200 300 400 500

Right Mastoid

mV 0 mV 0

Figure 2 The N170 potentials elicited at the left and right mastoids by the four face types and the scalp distribution of the potentials at thislatency (spline interpolation)

940 Journal of Cognitive Neuroscience Volume 13 Number 7

between the amplitudes of N170 elicited by photo-graphs paintings and schematic faces The N170 ampli-tudes across face categories were largest at the mastoidsand in descending order IM12 P78 and PO78 Sig-nificant differences were found only between the mas-toids and the P78 and PO78 locations [F(129) = 50p lt 05 and F(129) = 48 p lt 05 respectively] andbetween the IM12 and the PO78 locations [F(129) =49 p lt 05]

All three first-order interactions were significant[F(9261) = 48 p lt 001 F(387) = 31 p lt 05F(387) = 63 p lt 01 for the Stimulus Type pound SiteStimulus Type pound Hemisphere and Site pound Hemisphererespectively] The second-order interaction between allthree factors was also significant [F(9261) = 34 p lt001] Post hoc examination of these interactions re-vealed the following pattern The difference betweenthe N170 elicited by sketches and that elicited by theother face types was larger at the right hemisphere sitesthan at the left hemisphere sites and slightly smaller atIM sites than at the other three sites The interactionbetween hemisphere and scalp location is probably aresult of the smaller (but clear) right hemisphere ad-vantage at scalp site IM12 [F(129) = 90 p lt 01]compared to other sites [F(129) = 153 p lt 001F(129) = 182 p lt 001 F(129) = 242 p lt 001 forPO78 mastoids and P78 respectively] The second-order interaction reflected (a) a smaller (Site pound Hemi-

sphere) interaction for the sketched faces category[F(387) = 23 vs F(387) gt 59 for each of the othercategories] (b) an insignificant (Stimulus Type pound Hemi-sphere) interaction at the IM scalp site [F(387) lt 100]versus a significant or close-to-significant effect for othersites [F(387) gt 24] and (c) a smaller (Stimulus Type poundSite) interaction in the right hemisphere [F(15435) =30] than in the left hemisphere [F(15435) = 55]

As evident in Table 1B the mean N170 latency wasshortest for schematic faces (1676 msec) followed byphotographs (1693 msec) paintings (1714 msec) andsketches (1743 msec) Scalp sites also differed on N170latency the site at which the shortest N170 latency wasrecorded was PO78 (1697 msec) followed by P78(1702 msec) IM12 (1704 msec) and RL-Mast (1723msec) ANOVA showed that both the stimulus type andthe site effects were significant [F(387) = 167 p lt001 F(387) = 50 p lt 005 respectively] Post hoccontrasts revealed that differences between the meanN170 latencies of all stimulus type pairs were highlysignificant [F(129) = 92 p lt 005] except for thedifference between face photograph and schematic facelatencies [F(129) = 30 p lt 1] In addition the N170latency at the mastoid scalp locations (1723 msec) wassignificantly delayed [F(129) = 173 p lt 001] com-pared with the three other locations (1701 msec)which did not significantly differ from each other[F(129) = 5 p = 6]

Table 1 (A) Mean Amplitudes (in mV) and (B) Mean Latencies (in msec) of the N170 Elicited by Natural Faces Painted PortraitsSketches Schematic Faces Flowers and Scrambled Faces at the Posterior Temporal Scalp Sites

Left Hemisphere Right Hemisphere

Left mastoid P7 PO7 IM1 IM2 PO8 P8 Right mastoid

A

Natural Faces ndash 72 ndash 62 ndash 60 ndash 76 ndash 89 ndash 93 ndash 91 ndash 97

Painted Portraits ndash 73 ndash 61 ndash 62 ndash 78 ndash 88 ndash 88 ndash 92 ndash 98

Sketches ndash 58 ndash 49 ndash 41 ndash 61 ndash 73 ndash 64 ndash 70 ndash 76

Schematic Faces ndash 68 ndash 63 ndash 60 ndash 70 ndash 83 ndash 88 ndash 96 ndash 94

Flowers ndash 29 ndash 32 ndash 23 ndash 28 ndash 30 ndash 29 ndash 33 ndash 34

Scrambled Faces ndash 19 ndash 17 ndash 07 ndash 13 ndash 15 ndash 07 ndash 16 ndash 22

B

Natural Faces 170 168 170 168 169 169 170 170

Painted Portraits 171 171 172 171 171 171 172 172

Sketches 177 171 171 174 175 173 176 178

Schematic Faces 170 165 165 168 167 167 169 169

Flowers 174 166 168 171 172 168 171 176

Scrambled Faces 177 167 166 173 175 166 168 178

Sagiv and Bentin 941

Hemispheric differences failed to reach significance[F(129) lt 10] A Site pound Hemisphere interaction washowever reliable [F(387) = 40 p lt 005] This inter-action is possibly the result of relatively late right hemi-sphere mean latency at P78 and PO78 scalp locationsThe interaction between stimulus type and site was alsosignificant [F(9261) = 35 p lt 001] reflecting a greatersite effect within the sketched face category The Stim-ulus Type pound Hemisphere interaction was not significant[F(387) = 14 p lt 25] The interaction between allthree factors did not reach significance [F(9261) = 17p = 1]

Discussion

Experiment 1 demonstrated that face-specific N170 com-ponents can be recorded with scalp electrodes not onlywhen photographs of natural faces are presented butalso when paintings sketches and even highly sche-matic drawings of faces are used as stimuli The ampli-tude of the N170 elicited by all face stimuli was at leasttwice as big as the amplitude of the negative potentialelicited by flowers and three times as big as the ampli-tude of the potential elicited by nonsense images duringthe same time The significant difference between flow-ers and scrambled faces could reflect either that theneural activity associated to the structural encoding ofany object is distinct from that elicited by visual patternsthat are not objects (McCarthy et al 1997) or that thereis specificity for other stimulus categories in the vicinityof the face area and that part of this specificity is pickedup by the far-field recording (Ishai Ungerleider MartinSchouten amp Haxby 1999) The present study was notdesigned to address this issue Therefore we will turnnow to the focus of this experiment that is the compar-ison between the different face types

The finding that the N170 elicited by schematic facesis not significantly smaller in amplitude is not delayedrelative to natural face photographs and has a similarscalp distribution suggests that common neural mech-anisms underlie the perception of both natural andschematic faces Assuming that these mechanisms arethe neural analogue to the hypothesized structuralencoder for faces these data also suggest that a face-specific visual mechanism is triggered whenever a stim-ulus contains sufficient information to generate theconcept of a face In other words schematically drawnfaces are categorized by an early visual mechanismrather than by a higher-level cognitive process Indeedit is possible that once triggered such a face perceptionmodule may take precedence over a more general visualprocessor A similar account for the suppression ofacoustic perception when phonetic features are de-tected (Bentin amp Mann 1990 Liberman amp Mattingly1989) led to the hypothesis that the phonetic modulemay preempt auditory information that is relevant tophonetic perception (Whalen amp Liberman 1987)

The reduced amplitude of the N170 elicited by detail-richer face sketches relative to that elicited by schematicfaces and its delayed peak latency might be interpretedprima facie as evidence against the above hypothesisHowever post hoc this result can be incorporated intoour theory Although graphically more detailed thesketches did not attempt to imitate a natural face (asthe paintings) or to emphasize those basic featuresapparently required for the formation of a face concept(as in the schematic faces) Indeed it is possible thatsome of the sketches were difficult to interpret ashuman faces without an elaborate analytic processHence the generation of a face concept on the basisof such sketches may have been somewhat more timeconsuming either because they were processed by gen-eral visual processors or because they activated thespecial purpose structural encoder later In additionunlike the schematic faces that although not identicalwere more uniform in shape4 the sketches were stylis-tically more diverse As a consequence some of thesketches may have introduced noise or a time jitter inthe average ERP which could explain both the reducedamplitude and the longer latency of the N170 associatedwith the sketched faces category5 All face categorieselicited a significantly larger N170 in the right hemi-sphere However the asymmetry was reduced forsketches and absent for nonface categories This findingis consistent with neuropsychological (eg De Renzi1997) neuroimaging (eg McCarthy et al 1997 DeRenzi Perani Carlesimo Silveri amp Fazio 1994 Haxbyet al 1993) and behavioral studies (eg Luh Rueckertamp Levy 1991) indicating a right hemisphere advantagefor face processing Furthermore it corroborates thehypothesis that some of the sketches were processed bya general rather than a face-specific visual processor orthat in general the sketches were less efficient ininitiating a face-specific process

The finding that schematic faces elicit a conspicuousN170 that does not significantly differ from the ERPselicited by natural faces is intriguing in light of theassumption that the N170 is associated primarily withthe processing of face components In previous studieswe have consistently found that eyes in isolation or eyescombined with other inner-face components elicit alarger N170 than that elicited by full faces (Bentinet al 1996 Bentin amp McCarthy unpublished) Further-more full faces elicit an N170 which is only slightlylarger than that elicited by face contours from which allinner components have been erased On the basis ofthese data Bentin et al (1996) suggested that the N170is modulated by two neural sources (both part of thestructural encoding mechanism) one possibly locatedin the posterior upper bank of the occipito-temporalsulcus (OTS) or in the inferior temporal (IT) gyrus istriggered by physiognomic information appearing in thevisual field the other probably located in the middlefusiform gyrus is responsible for holistic face processing

942 Journal of Cognitive Neuroscience Volume 13 Number 7

and integrates the face parts into a whole Assuming thatwithout the framing contour and the face-specific con-figuration inner components of schematic faces have nophysiognomic value6 the N170 elicited by schematicfaces should be modulated primarily by the fusiformsource which is responsible for the holistic processingof a face Consequently the N170 should be particularlysusceptible to manipulations affecting holistic process-ing of schematic faces An initial attempt to explore thispossibility is reported in Experiment 2 In addition thesimilar latency for the N170s elicited by natural faces andschematic faces would thus be in agreement with theprimacy of holistic processing for face stimuli the lateralcomponent primarily responsible for the N170 is trig-gered by the holistic component whether or not theface parts are recognizable as such in isolation ordefined by the global configuration

EXPERIMENT 2

The face inversion effectmdashthat is the difficulty to iden-tify familiar faces presented upside down (comparedwith other familiar objects subject to a similar manipu-lation)mdashhas been the focus of many studies for morethan two decades (reviewed by Valentine 1988) Thedifference between the effects of face inversion relativeto inversion of other objects is usually attributed to abasic difference between the processing strategies in-curred by faces versus objects According to this hypoth-esis recognition of faces relies on the formation of awhole-face representation from which relations betweenthe inner-face parts are derived that is holistic process-ing (Farah Tanaka amp Drain 1995 Tanaka amp Farah 1993versions of the holistic hypothesis are reviewed byMoscovitch Winocur amp Behrmann 1997) In contrastthe recognition of other perhaps less stereotypic ob-jects entails initial identification of basic componentswhich are then related one to the other to form anunequivocal visual percept (eg Biderman 1987 Marr1982 Marr amp Nishihara 1978) The face inversion effecthas become the hallmark of normal face recognition inhumans and has often been used to distinguish face-specific from general visual processes

Bentin et al (1996) compared the N170 elicited byupright and inverted faces and found that the N170elicited by inverted faces was slightly larger and peaked10 msec later than that elicited by upright faces only thedifference in latency was significant in that study Otherinvestigators however reported that the enhancementof the N170 amplitude by face inversion is significant aswell as the delay in latency (eg Rossion et al 19992000) A possible interpretation of the enhanced N170amplitude in response to inverted faces is that the scalp-recorded N170 is associated primarily with the percep-tion of face components a process that might not beimpeded by inversion This hypothesis conforms to thepresumed anatomical location of the putative compo-

nents detector (in lateral OTS and IT gyrus) and theorientation of these cell assemblies relative to the re-cording electrodes on the scalp In contrast to the N170the intracranially recorded N200 (on the ventral plane ofthe temporal lobe) is modulated primarily by the holis-tic face integration mechanism located in the posteriorlateral parts of the fusiform gyrus (eg Allison et al1999) Hence given the orientation of the electricaldipole fields in these two cell assemblies a manipulationthat inhibits holistic perception should enhance theN170 amplitude This hypothesis also conforms to thefinding that the N200 (recorded at the fusiform) elicitedby faces is rsquo rsquo largerrsquo rsquo than that elicited by isolated eyes(McCarthy et al 1999) whereas the N170 (recorded atthe scalp) elicited by faces is rsquo rsquosmallerrsquo rsquo than that elicitedby isolated eyes (Bentin et al 1996)

According to this model face perception is specificbecause the highly stereotypic organization of facespermits categorizing the visual stimulus before a fulldescription of its features is provided by the analyticmechanism Therefore most of the time faces activatethe holistic perception mechanism instantly Indeedonly such a model has the capacity to explain the resultsof Experiment 1 in which a visual stimulus that can beperceived as a face only if processed as a whole (ie ahighly schematic face) elicited a similar N170 to thatelicited by photographs of natural faces

Based on the above conceptualization we hypothe-size that if the detection of schematic faces is basedprimarily on the holistic face configuration compromis-ing their gestalt by presenting the stimuli upside downshould reduce the N170 In contrast turning naturalfaces upside down should enhance the amplitude of theN170 since the face component processing modulewould still be activated To test this hypothesis wepresented a new group of participants with naturaland schematic faces in upright and inverted orienta-tions

Results

Figure 3 presents the N170s elicited by photographs ofnatural faces and schematic faces presented upright orupside down Replicating the findings of Rossion et al(1999) the N170 elicited by inverted natural faces waslarger than that elicited by upright faces In contrast theN170 elicited by inverted schematic faces was reduced inamplitude relative to that elicited by upright schematicfaces Inversion delayed the peak latency of the N170 forschematic as well as for natural faces

The reliability of these observations was statisticallyanalyzed by a within-subject ANOVA using repeatedmeasures (see data in Table 2A)

The factors were stimulus type (natural faces sche-matic faces) stimulus orientation (upright inverted) site(P78 mastoids PO78 IM12) and hemisphere (leftright) As a main effect stimulus orientation was not

Sagiv and Bentin 943

significant [F(115) = 21 p = 16] However the inter-esting pattern of the face inversion effects was revealedby a significant interaction between the effects of orien-tation and stimulus type [F(115) = 248 p lt 001]followed by planed comparisons The planed compari-sons showed that whereas for natural photographs theN170 elicited by inverted faces (ndash 66 mV) was larger thatthat elicited by upright faces (ndash 55 mV) with schematicfaces inversion was associated with a reduction inthe N170 amplitude (ndash 38 and ndash 50 mV for invertedand upright presentation respectively t(15) = 2104p lt 055 and t(15) = 3000 p lt 01 for natural andschematic faces respectively) The main effect of stim-ulus type was significant [F(115) = 302 p lt 001] but itwas also qualified by its interaction with the effect oforientation Post hoc comparisons showed that whereasthe difference between the N170 elicited by invertednatural and schematic faces was significant [t(15) =4264 p lt 001] the difference between the N170elicited by natural and schematic faces presented uprightwas not [t(15) = 884 p = 39] The main effect of site

was also significant [F(345) = 68 p lt 01] revealingonly that the N170 amplitudes at P7 and P8 were smallerthan at all other sites [F(115) = 117 p lt 01] It isnoteworthy that as evidence by an insignificant interac-tion with stimulus type [F(345) = 13 p = 28] thiseffect was similar for schematic and natural faces On theother hand a significant main effect of hemisphere[F(115) = 65 p lt 05] was qualified by its interactionwith stimulus type [F(115) = 53 p lt 05] Post hocanalysis revealed that whereas for both stimulus typesthe amplitude of the N170 was larger over the rightcompared to the left hemisphere the interhemisphericdifference was significant for natural faces [t(15) = 335p lt 01] but not for schematic faces [t(15) = 1408p = 18] The effects of the Hemisphere pound Orientationinteraction was significant [F(115) = 47 p lt 5] Thisinteraction was caused by the fact that whereas acrossstimulus type the reduction and enhancement of theN170 amplitude by inversion canceled each other in theleft hemisphere in the right hemisphere the enhance-ment of N170 induced by the inversion of natural faces

Faces Inverted Faces Schematic Faces Inverted SchematicFaces

Time(msec) Time(msec)0 100 200 300 400

10

mV 0

0 100 200 300 400

10PO7 PO8

- 0 +

mV 0

Figure 3 The N170 potentials elicited at the left and right mastoids by upright and inverted natural faces and schematic faces and the scalpdistribution at this latency

944 Journal of Cognitive Neuroscience Volume 13 Number 7

weighted more than its reduction resulting from theinversion of schematic faces However the second-orderinteraction between orientation hemisphere and stim-ulus type was not significant [F(115) lt 10] Finally aseries of second-order interactions involving the effect ofsite suggested that all the effects were differently con-spicuous at different sites As revealed in Table 2 by andlarge it seems that whereas natural faces induced similarinversion effects across the analyzed sites the inversioneffect for schematic faces varied at different sites We hadno prior assumptions about the distribution of the effectswithin the posterior temporal scalp regions Further-more the spatial resolution of scalp-recorded ERPs istoo low to permit reliable interpretation of site effectswithin a relative small region unless the same distribu-tion is replicated therefore a detailed analysis of thisdistribution of effects was not attempted post hoc

ANOVA of the N170 peak latencies showed a signifi-cant main effect of stimulus orientation [F(115) =1615 p lt 001] which significantly interacted withstimulus type [F(115) = 62 p lt 025] Plannedcomparisons showed that for both natural and sche-matic faces inversion delayed the N170 significantly[t(15) = 10821 p lt 001 and t(15) = 3626 p lt 01respectively] However as evident in Table 2B thisdifference was larger for natural than for schematicfaces The latencies were similar over the left and righthemispheres [F(115) = 19 p = 185] The main effectof site was significant [F(345) = 47 p lt 01] butbecause the differences between sites were all smaller

than the sampling rate this effect was not investigatedfurther No other significant effects emerged in thisanalysis

Discussion

The results of the Experiment 2 validated our predic-tion that inversion should reduce the amplitude of theN170 elicited by schematic faces but have little effector enhance the N170 elicited by natural faces Thenatural face inversion effect observed in the presentstudy replicates previous results reported by Rossionet al (1999) These authors suggested that the en-hancement in the N170 amplitude reflects the en-hanced difficulty of the holistic processor to encodethe inverted face Indeed the delay in the N170 peaklatency elicited by inverted relative to upright faces iscongruent with such a view However if this were thecase we should have found similar effects of inversionon the N170 amplitude elicited by schematic and bynatural faces Furthermore the assumption that moredifficult face encoding induces an enhancement of theN170 amplitude is also at odds with fMRI data showingthat face inversion reduces the activity in the fusiformface areas relative to the activity elicited by uprightfaces (Kanwisher Tong amp Nakayama 1998) Indeedfinding opposite effects of inversion for schematic andnatural faces can more likely be accommodated by ourhypothesis that the larger N170 amplitude in responseto inverted natural faces reflects its stronger associa-

Table 2 (A) Amplitudes (in mV) and (B) Latencies (in msec) of the N170 Elicited by Upright and Inverted Natural and SchematicFaces at the Posterior Temporal Scalp Sites

Natural Faces Schematic Faces

Right Hemisphere Left Hemisphere Right Hemisphere Left Hemisphere

Upright Inverted Upright Inverted Upright Inverted Upright Inverted

A

Mastoid ndash 64 ndash 86 ndash 58 ndash 82 ndash 54 ndash 55 ndash 51 ndash 35

P87 ndash 48 ndash 68 ndash 45 ndash 54 ndash 41 ndash 30 ndash 32 ndash 33

PO87 ndash 68 ndash 88 ndash 48 ndash 50 ndash 56 ndash 56 ndash 50 ndash 23

IM21 ndash 53 ndash 81 ndash 49 ndash 73 ndash 50 ndash 29 ndash 65 ndash 39

Mean ndash 58 ndash 81 ndash 50 ndash 65 ndash 50 ndash 425 ndash 49 ndash 325

B

Mastoid 149 162 151 165 159 164 158 165

P87 151 163 154 166 158 164 162 166

PO87 151 160 152 164 151 161 159 166

IM21 150 162 147 161 161 164 154 162

Mean 150 162 151 164 157 163 158 165

Sagiv and Bentin 945

tion with a face component analyzer in the lateraltemporal gyri than with the holistic face processor inthe fusiform

Our model proposes that although schematic andnatural faces probably activate similar neural networksin the extrastriate visual pathway these mechanismsare not identical Natural faces probably activate twoperceptual modules one specialized to detect andprocess physiognomic features in the visual field andthe other specialized in the holistic processing of facesor forming a holistic face representation Unlike naturalfaces the components of schematic faces are notperceived as carrying physiognomic information outof the schematic face gestalt Therefore in contrastto natural faces schematic faces may trigger the holisticprocessor but they do not trigger directly the analysisof the components The latter mechanism may beactivated only subsequent to the formation of theholistic face representation or at the very least follow-ing this process rsquo rsquo in cascadersquo rsquo (McClelland 1979) Con-sequently adversely affecting the perception of the facegestalt (by inverting the stimuli) inhibited both holisticperception and the perception of individual compo-nents and this inhibition reduced the amplitude of theN170 Such a pattern is also in agreement with themyriad of studies supporting the primacy of holisticprocessing of face stimuli

GENERAL DISCUSSION

In the present study we investigated several character-istics of the extrastriate structural face-encoding mech-anisms by comparing the N170 elicited by faces differingin their naturalistic aspect and by comparing the faceinversion effects on the N170 elicited by photographs ofnatural faces and by schematically drawn faces To theextent that the N170 is indeed associated with structuralencoding we found that this process distinguishes wellbetween faces and nonfaces stimuli but not amongdifferent face types that render the concept of a faceclearly However we also found evidence that schematicfaces and natural faces are not processed identicallyWhereas inversion of natural faces enhanced the ampli-tude of the N170 inversion of schematic faces reducedits amplitude For both face types the latency of theN170 peak was significantly delayed by inversion Thispattern of results suggests that the face-specific struc-tural encoder can be triggered by a variety of stimuli ifthey include a canonical face configuration Howeverthe process of encoding the face information and form-ing a structural representation is probably differentwhen the physiognomic value of the stimuli dependsupon holistic configuration as opposed to when theindividual components can be associated with faces evenwhen presented outside the face context

Apart from the neurofunctional specificity issue anenduring question in face perception research is whether

the formation of a mental representation of the faceentails extraction of a holistic configuration directly fromthe visual input or whether it is feature-based that iswhether for faces as for other complex visual stimulistructural encoding is based on an initial analysis of parts(for a review of this debate see eg Farah 1990 Farahet al 1995) In principle because faces form a veryhomogeneous and stereotypical category they lendthemselves to holistic perception Holistic processing isadvantageous in face recognition because face compo-nents are fairly similar across individuals and thereforethe distinction among faces is usually based on theinterrelationships between the different face compo-nents and their overall configuration (eg Tanaka ampFarah 1993 Young Hellawell amp Hay 1987)7 Moreoverour memory for faces is usually based on nondecompos-able holistic images which at the interface betweenperception and memory are most efficiently addressedby integrated face representations8However if the well-learned gestalt is altered and holistic perception is notpossible it is probably replaced by a more analyticcomponent-based process of structural encoding Thisanalytic process may inhibit or at least delay the per-ception of the spatial relationships between face com-ponents and therefore may impede identificationIndeed a major source of evidence supporting theholistic view is that faces are more difficult to recognizewhen presented upside down (ie compromising theirfamiliar gestalt) than when presented in upright orien-tation (see reviews in Farah et al 1995 Rhodes et al1993 Valentine 1988)

Having to cope with different types of face input thevisual system must be able to encode the face config-uration directly from the visual display as well as byintegrating face components encoded individuallyThese two processes may concur triggered independ-ently by relevant physiognomic input and performed bydifferent neural circuits to form a complex structuralencoding mechanism in the extrastriate cortex Further-more the two systems may interact supporting orinhibiting each other Hence a regular face presentedin upright orientation may activate the holistic percep-tion process which in turn may inhibit superfluousanalysis of individual components By contrast corrupt-ing the face gestalt by presenting faces in an unusualorientation may trigger (and facilitate) face componentanalysis which should not depend on a particular con-figuration In such a case the integration of the compo-nents into a whole by the holistic processor will followtheir independent analysis assisted by the feature-detection system This hypothesis may explain for in-stance why within the context of a regular face individ-ual face parts are distinguished faster than in isolationbut if the configuration of the face is altered the partsare distinguished as fast as when presented in isolation(Tanaka amp Farah 1993 Sergent 1984)9 The ability toprocess face information both holistically and piecewise

946 Journal of Cognitive Neuroscience Volume 13 Number 7

was implemented in a neural model initially based onscalp recordings (Bentin et al 1996) and more recentlycorrected improved and elaborated on the basis intra-cranial recordings (McCarthy et al 1999 for a detaileddescription and a review of supportive functional MRIdata see McCarthy 1999) This model suggests that theneural mechanism dedicated to structural encoding offaces is a complex system composed of at least twoseparate subsystems One probably located in thelateral part of the middle fusiform gyrus is responsiblefor processing the configuration of the face as a wholeand provides the integrated face representation that isnecessary for face recognition and identification Theother possibly distributed over posterior regions in theIT and middle temporal (MT) gyri and in the OTS isresponsible for detecting physiognomic informationregardless of the spatial configuration of the items thatcarry it The fusiform gyrus is farther away from therecording sites at the scalp than where the temporalgyri are and the orientation of its cells forms anelectrical dipole source that is almost perpendicular tothe lateral aspects of the scalp Hence although bothsources contribute to the modulation of the face-spe-cific N170 the contribution of the posterior temporalface component processor should be considerably moreconspicuous

According to this hypothesis the opposite effects ofinversion on the N170 elicited by natural and schematicfaces might be explained assuming that the holisticencoding system in the fusiform gyrus indeed inhibitsthe component analysis in the lateral temporal lobe facesites Hence if the face stimulus can trigger both theholistic and the component analysis (as it is true fornatural faces) an experimental manipulation that inter-feres with holistic processing (like the face inversionprocedure in the present experiment) should reducethe amount of inhibition of component analysis andconsequently the amplitude of the N170 would beenhanced However if the rsquo rsquocomponentsrsquorsquo do not conveyphysiognomic information so that their analysis by aface-specific mechanism rsquo rsquodependsrsquo rsquo on the previousrecognition of the face gestalt (as it is true for schematicfaces) interfering with holistic perception cannot facil-itate the activity of the component analysis and there-fore the modulation of the N170 would only reflect theadditional difficulty to encode the face ie reducedamplitude and delayed latency

The observed effects of the different factors on theN170 peak latency are nicely aligned with the abovemodel Although stimulus inversion delayed the N170for both natural and schematic faces this delay was twiceas big (12 msec) for former stimuli than for the latter (6msec) This interaction suggests that the inversion had abigger effect on natural than on schematic faces Thismay be because regardless of whether they are upright orinverted the physiognomic value of schematic faces isbased on their gestalt Hence inversion does inhibit the

processing of the gestalt but does not induce a change inprocessing strategy On the other hand natural faces areanalyzed holistically if they are upright but inversionrequires a change in that strategy encouraging theanalysis of components The significantly faster responseto upright natural than to upright schematic faces inExperiment 2 suggests that the extraction of the facegestalt is easier if the face is more natural

There is ample evidence indicating that the righthemisphere is specialized in global or holistic processingof objects and particularly faces whereas the left hemi-sphere is specialized in part-based processing (egRobertson amp Delis 1986 for a review see Bradshow ampNettleton 1983 for specific reference to face perceptionsee eg Bruce 1998 Rhodes et al 1993 Corballis1988 Bradshow amp Sherlock 1982 Gilbert amp Bakan1973) Consequently interfering with holistic processingshould have had stronger effects in the right than in theleft hemisphere Our results confirmed this predictiononly partially The pronounced interhemispheric asym-metry with larger N170 amplitudes at the right than atthe left hemisphere sites for all stimulus types suggests apredominant holistic process for all face categoriesInterestingly this asymmetry was reduced (in Experi-ment 1) for sketches relative to the other stimulus typesThe relatively enhanced involvement of the left hemi-sphere in processing the sketches is in agreement withour suggestion that the face-specific structural encodingsystem was activated less efficiently by sketches than byother face types10 Hence the same factor that may haveconstrained the processing of schematic faces may alsohave constrained the processing of sketches but in theopposite direction Whereas schematic faces enhancedholistic processing by providing a clear face gestalt whileobscuring the physiognomic value of face componentsthe sketches reduced holistic processing by blurring orveiling the face gestalt Also in Experiment 2 the N170was larger over right than over left hemisphere sitesHowever because the interaction between the inversionand hemisphere effects in Experiment 2 was not modu-lated by stimulus type (ie the three-way interactionwas not significant) our interpretation of the interhemi-spheric asymmetry (particularly in Experiment 2) mustbe limited

In conclusion the present scalp-recorded N170 dataconverge with the previously published pattern of theintracranially recorded N200 data (for a comprehensivereview see Allison et al 1999 McCarthy et al 1999Puce et al 1999) supporting a multifunctional organ-ization of the structural encoding mechanism for facesApparently distinct neural circuits are responsible forthe encoding and initial analysis of face components andfor holistic perception of a full-face configuration Theformer is probably located in the lateral posterior tem-poral lobe whereas the latter is located in the middlefusiform gyrus Face perception usually entails predom-inantly holistic processing which explains the right

Sagiv and Bentin 947

hemisphere dominance in performance as well as inelectrophysiological and neuroimaging data Howeverboth mechanisms can be triggered independently byrelevant information

METHODS

Participants

Right-handed participants were recruited among volun-teer students from the Hebrew University They receivedeither experimental credit hours or payment for theirparticipation Thirty participants (13 males and 17 fe-males) were tested in the first experiment and 16participants (8 males and 8 females) were tested in thesecond experiment Participantsrsquo ages ranged from 19 to37 years (mean = 233 years) They all had normal orcorrected-to-normal vision

Stimuli

Stimuli were scanned using a desk scanner processedwith commercially available graphic software and con-

verted to 16 color grayscale 248 by 248 pixel imagesThere were eight stimulus categories four face types(photographs of natural faces paintings face sketchesand schematic faces) two nonsense patterns (formedby scrambling the natural faces and the schematicfaces) flowers and butterflies (Figure 4) Paintingsand sketches of faces were scanned from art booksPaintings were typically realistic Renaissance portraitsand face sketches did not portray details and usuallylacked shading and texture Schematic face parts weredrawn by hand scanned into the computer and latercombined into different schematic faces Face photo-graphs and paintings contained many shades of grayand the sketches and the schematic faces were basicallyblack on white All face stimuli were front views

Scrambled faces were created by scrambling thephases of the two-dimensional Fourier transform andthen smoothing the uneven edges created by this proc-ess The original and the scrambled stimuli were equallyluminous following this procedure Each stimulus cat-egory comprised 75 stimuli except for the butterflycategory which consisted of 30 stimuli In Experiment2 only natural faces schematic faces and flowers were

Face Stimuli

Photographs Painted portraits Sketches Schematic faces

Nonface Stimuli

Scrambledphotographs

Butterflies(targets) Flowers Scrambled

schematic faces

Figure 4 Example of the eight stimulus types used in the present study

948 Journal of Cognitive Neuroscience Volume 13 Number 7

used The faces were presented in both upright andupside down orientation Upright stimuli were the samein Experiments 1 and 2

All stimuli were presented for 350 msec with anaverage interstimulus interval (ISI) of 12 sec and viewedfrom a distance of 110 plusmn 10 cm subtending a viewingangle of approximately 68 Stimulus timing was con-trolled by the MEL software system (Psychology Soft-ware Tools Pittsburgh PA)

Task

Participants were instructed to keep a mental countof stimuli belonging to one category butterflies inthe first experiment and flowers in the secondHence all of the experiment-relevant categories werefrom the participantsrsquo point of view nontargets Thetargets were presented in approximately 5 (Experi-ment 1) and 6 (Experiment 2) of the trials Duringeach break and at the end of the experimentparticipants reported the number of targets countedA few participants who did not count correctly wereexcluded

EEG Recording and Data Analysis

Continuous EEG was recorded from 48 scalp locationsusing tin electrodes attached to elastic electrode caps(Electrode Caps International Ohio) A schematic illus-tration of the electrode locations and names is pre-sented in Figure 1 A ground electrode was located onthe forehead and a reference electrode was located onthe tip of the nose Two additional electrodes located atthe outer canthus and the infraorbital region of the righteye measured vertical and horizontal electrooculograms(EOGs)

The EEG was sampled at 250 Hz by a DAP2400 DataAcquisition Processor (Microstar Laboratories) ampli-fied 20000-fold using a battery-operated HU-5072 Iso-lated Bioelectric Amplifier (SA Instrumentation SanDiego) with a frequency bandpass of 01ndash30 Hz andstored on disk for off-line averaging Electrode impe-dance was below 5 k A digital filter of 08ndash16 Hz wasapplied before analysis

Codes synchronized to stimulus delivery were used toselectively average epochs associated with differentstimulus categories Epochs contaminated with saccadesor eye blinks were excluded using root-mean-squareEOG-values as criteria Typically epochs associated with55ndash70 (never less than 50) out of 75 stimuli in eachcategory were averaged

The mean and peak amplitudes and peak latencies ofspecific ERP components within bounding intervals werecomputed for each category for each participant Sig-nificance of the differences between these measures wastested using a repeated-measures ANOVA Where appro-priate the more conservative GreenhousendashGeisser ad-

justed df values were used ANOVAs were followed byunivariate F contrasts

Acknowledgments

This study was supported by grants from the USndashIsraelBinational Science Foundation and Israeli Foundations Trust-ees to S Bentin We thank Amir Katz for his help in preparingthe stimuli and running Experiment 1 as well as Leon DeouellMichael Fink Ayelet Landau Yulia Goland David Carmel GilRaviv Amir Raz and Dan Tzur for their support We are alsovery grateful to Jerry Levinson Alex List and Lynn Robertsonfor their comments on the manuscript

Reprint requests should be sent to Shlomo Bentin TheHebrew University of Jerusalem Mount Scopus JerusalemIsrael 91905 or via e-mail msbentinmscchujiacil

Notes

1 Present address University of California Berkeley2 There was no reliable difference between the ERP elicited

by scrambled natural faces and scrambled schematic facesTherefore these two categories were collapsed into one3 These were the sites at which the N170 was most

conspicuous4 So were in fact all face categories except for the sketches5 Note that this result also contradicts the view that a more

difficult encoding process elicit N170 of a higher amplitude6 This assumption was indeed supported by a current study

(Bentin Mecklinger Bosch Sagiv amp von Cramon 1999)7 Of course some faces have very distinctive components

which can be used as markers for recognition Such markersare frequently exaggerated and used as distinctive features incaricatures (Rhodes 1987)8 Distinctive features (such as unusual components birth-

marks etc) may be stored in semantic memory as part of apersonrsquo s face entry Such information may have visual qualitiesthat can be addressed directly during visual matching tofacilitate identification9 Note that this account for the rsquorsquoface superiority effectrsquo rsquo is

identical to a common explanation of the rsquorsquoword superiorityeffectrsquo rsquo in the identification of single letters Indeed ourassumption about the coexistence of rsquo rsquowhole-facersquo rsquo and rsquo rsquoface-partrsquo rsquo processors resonates to the rsquorsquodual routersquo rsquo model of visualword processing It is quite possible that identical principlesgovern all the content-specific perceptual modules in the visualsystem10 Recall that relative to other face types the reduction of theN170 amplitude elicited by sketches at right hemisphere siteswas greater then at left hemisphere sites

REFERENCES

Allison T Ginter H McCarthy G Nobre A C Puce ALuby M amp Spencer D D (1994) Face recognition in hu-man extrastriate cortex Journal of Neurophysiology 71821ndash825

Allison T McCarthy G Nobre A C Puce A amp Belger A(1994) Human extrastriate visual cortex and the perceptionof faces words numbers and colors Cerebral Cortex 5544ndash554

Allison T Puce A Spencer D D amp McCarthy G (1999)Electrophysiological studies of human face perception I

Sagiv and Bentin 949

itself and be triggered by a variety of visual stimuli whichmay potentially represent a face

The components of a schematic face such as Smileydo not usually convey physiognomic information inisolation Their interpretation as face parts depends onthe configuration in which they are embedded There-fore apart from testing the range of the face structuralencoder sensitivity comparison of the evoked re-sponses to real faces with the responses to schematicfaces was expected to yield new insights regarding thefunctional organization of the structural encoding sys-tem with which the N170 ERP component is probablyassociated

Results

Replicating previous studies natural faces elicited arobust N170 component maximal at the posteriortemporal sites At the same time a positive peakwas measured at the anterior sites (P190 cf Jeffreysamp Tukmachi 1992) The negative (or negative going)ERPs elicited by equiluminant nonsense stimuli as

well as photographs of flowers within the same timerange were significantly smaller (Figure 1) The out-standingly large amplitude of the N170 elicited byfaces compared to flowers and meaningless stimuluscategories in the present study as well as similarlybig differences between the response to faces andthe response to other types of stimuli reported inprevious studies support our hypothesis that theN170 is a manifestation of an early face-specific visualprocessing mechanism Given our present interest inthis neural mechanism we will not discuss differencesbetween stimulus conditions that occurred later thanthe N170

The major manipulation in this experiment involvedpresenting different face types at different levels ofphysical resemblance to photographs of natural facesThe consequences of this manipulation are presented inFigure 2 which compares the N170 elicited at the rightand the left mastoid scalp locations by the different facecategories Relative to nonface stimuli (Figure 1) all fourface types elicited more conspicuous but not identicalN170 components

Faces

Flowers

Scrambled Faces

PO8PO7

LM RM

TP8TP7 CP5 CP6CP2CP1 CPz

FT8FC6

FC2FC1 FCzFC5

FT7

GND

FP1 FP2

POz

PO3 PO4P7

P5 P 3

O1 Oz O2

Iz

Pz

Cz

Fz

P4P6

P8

C4

T8

F8F6

F4F3F5

F7

C3

T7

IM1 IM2

Left Mastoid

0

10

Time (msec)0 100 200 300 400 500

Right Mastoid 10

Time (msec)0 100 200 300 400 500

10Fz

Time (msec)0 100 200 300 400 500

N170

mV 0

mV 0

Figure 1 Demonstration of the basic difference between the N170 elicited by faces and that elicited by nonface stimuli Note the polarity inversionat 170 msec between the posteriorndashlateral and the anterior sites

Sagiv and Bentin 939

The statistical reliability of the facenonface differenceand of the apparent differences between the N170elicited by each face type was assessed by a series ofwithin-subject three-factor analysis of variance (ANOVA)with repeated measures The factors were stimulus type(four face types flowers scrambled faces)2 hemisphere(left right) and site (P78 left and right mastoids PO78IM12)3 The dependent variables were the N170 ampli-tude and its latency (Table 1)

The ANOVA of the amplitude showed that there weresignificant differences among the different stimulustypes [F(5145) = 856 p lt 001] that the N170 waslarger at right (ndash 652 mV) than at left (ndash 492 mV)hemisphere sites [F(129) = 196 p lt 001] and thatthere was a significant effect of site [F(387) = 598 p lt005] Post hoc comparisons revealed that the N170elicited by faces (weighted across all four face catego-ries) was significantly more negative than that elicited bynonfaces (weighted across scrambled faces and flowers)[F(129) = 1836 p lt 001] The ERPs elicited by flowers

were significantly less negative than those elicited bysketches [F(129) = 561 p lt 001] but significantlymore negative than those elicited by scrambled faces[F(129) = 223 p lt 001] Because our present experi-ment focused on the potential difference in processingdifferent face types and because of the clear distinctionbetween the N170 elicited by faces and the ERPs elicitedby nonface stimuli in the same time range our subse-quent analyses focused only on the four face types Thedesign of these ANOVAs was similar to the one de-scribed above except that the stimulus type factorincluded only the four face categories

The ANOVA of the amplitude showed that all threemain effects were significant [F(387) = 110 p lt 001F(387) = 37 p lt 025 F(129) = 203 p lt 001 forstimulus type site and hemisphere respectively] Posthoc contrasts showed that the amplitude of the N170elicited by sketches was significantly less negative thanthat elicited by the three other face types [F(129) =468 p lt 001] There was no significant difference

- 0 +

SketchesPhotographs Paintings Schematic Faces

10

Time (msec)0 100 200 300 400 500

Left Mastoid 10

0Time (msec)

100 200 300 400 500

Right Mastoid

mV 0 mV 0

Figure 2 The N170 potentials elicited at the left and right mastoids by the four face types and the scalp distribution of the potentials at thislatency (spline interpolation)

940 Journal of Cognitive Neuroscience Volume 13 Number 7

between the amplitudes of N170 elicited by photo-graphs paintings and schematic faces The N170 ampli-tudes across face categories were largest at the mastoidsand in descending order IM12 P78 and PO78 Sig-nificant differences were found only between the mas-toids and the P78 and PO78 locations [F(129) = 50p lt 05 and F(129) = 48 p lt 05 respectively] andbetween the IM12 and the PO78 locations [F(129) =49 p lt 05]

All three first-order interactions were significant[F(9261) = 48 p lt 001 F(387) = 31 p lt 05F(387) = 63 p lt 01 for the Stimulus Type pound SiteStimulus Type pound Hemisphere and Site pound Hemisphererespectively] The second-order interaction between allthree factors was also significant [F(9261) = 34 p lt001] Post hoc examination of these interactions re-vealed the following pattern The difference betweenthe N170 elicited by sketches and that elicited by theother face types was larger at the right hemisphere sitesthan at the left hemisphere sites and slightly smaller atIM sites than at the other three sites The interactionbetween hemisphere and scalp location is probably aresult of the smaller (but clear) right hemisphere ad-vantage at scalp site IM12 [F(129) = 90 p lt 01]compared to other sites [F(129) = 153 p lt 001F(129) = 182 p lt 001 F(129) = 242 p lt 001 forPO78 mastoids and P78 respectively] The second-order interaction reflected (a) a smaller (Site pound Hemi-

sphere) interaction for the sketched faces category[F(387) = 23 vs F(387) gt 59 for each of the othercategories] (b) an insignificant (Stimulus Type pound Hemi-sphere) interaction at the IM scalp site [F(387) lt 100]versus a significant or close-to-significant effect for othersites [F(387) gt 24] and (c) a smaller (Stimulus Type poundSite) interaction in the right hemisphere [F(15435) =30] than in the left hemisphere [F(15435) = 55]

As evident in Table 1B the mean N170 latency wasshortest for schematic faces (1676 msec) followed byphotographs (1693 msec) paintings (1714 msec) andsketches (1743 msec) Scalp sites also differed on N170latency the site at which the shortest N170 latency wasrecorded was PO78 (1697 msec) followed by P78(1702 msec) IM12 (1704 msec) and RL-Mast (1723msec) ANOVA showed that both the stimulus type andthe site effects were significant [F(387) = 167 p lt001 F(387) = 50 p lt 005 respectively] Post hoccontrasts revealed that differences between the meanN170 latencies of all stimulus type pairs were highlysignificant [F(129) = 92 p lt 005] except for thedifference between face photograph and schematic facelatencies [F(129) = 30 p lt 1] In addition the N170latency at the mastoid scalp locations (1723 msec) wassignificantly delayed [F(129) = 173 p lt 001] com-pared with the three other locations (1701 msec)which did not significantly differ from each other[F(129) = 5 p = 6]

Table 1 (A) Mean Amplitudes (in mV) and (B) Mean Latencies (in msec) of the N170 Elicited by Natural Faces Painted PortraitsSketches Schematic Faces Flowers and Scrambled Faces at the Posterior Temporal Scalp Sites

Left Hemisphere Right Hemisphere

Left mastoid P7 PO7 IM1 IM2 PO8 P8 Right mastoid

A

Natural Faces ndash 72 ndash 62 ndash 60 ndash 76 ndash 89 ndash 93 ndash 91 ndash 97

Painted Portraits ndash 73 ndash 61 ndash 62 ndash 78 ndash 88 ndash 88 ndash 92 ndash 98

Sketches ndash 58 ndash 49 ndash 41 ndash 61 ndash 73 ndash 64 ndash 70 ndash 76

Schematic Faces ndash 68 ndash 63 ndash 60 ndash 70 ndash 83 ndash 88 ndash 96 ndash 94

Flowers ndash 29 ndash 32 ndash 23 ndash 28 ndash 30 ndash 29 ndash 33 ndash 34

Scrambled Faces ndash 19 ndash 17 ndash 07 ndash 13 ndash 15 ndash 07 ndash 16 ndash 22

B

Natural Faces 170 168 170 168 169 169 170 170

Painted Portraits 171 171 172 171 171 171 172 172

Sketches 177 171 171 174 175 173 176 178

Schematic Faces 170 165 165 168 167 167 169 169

Flowers 174 166 168 171 172 168 171 176

Scrambled Faces 177 167 166 173 175 166 168 178

Sagiv and Bentin 941

Hemispheric differences failed to reach significance[F(129) lt 10] A Site pound Hemisphere interaction washowever reliable [F(387) = 40 p lt 005] This inter-action is possibly the result of relatively late right hemi-sphere mean latency at P78 and PO78 scalp locationsThe interaction between stimulus type and site was alsosignificant [F(9261) = 35 p lt 001] reflecting a greatersite effect within the sketched face category The Stim-ulus Type pound Hemisphere interaction was not significant[F(387) = 14 p lt 25] The interaction between allthree factors did not reach significance [F(9261) = 17p = 1]

Discussion

Experiment 1 demonstrated that face-specific N170 com-ponents can be recorded with scalp electrodes not onlywhen photographs of natural faces are presented butalso when paintings sketches and even highly sche-matic drawings of faces are used as stimuli The ampli-tude of the N170 elicited by all face stimuli was at leasttwice as big as the amplitude of the negative potentialelicited by flowers and three times as big as the ampli-tude of the potential elicited by nonsense images duringthe same time The significant difference between flow-ers and scrambled faces could reflect either that theneural activity associated to the structural encoding ofany object is distinct from that elicited by visual patternsthat are not objects (McCarthy et al 1997) or that thereis specificity for other stimulus categories in the vicinityof the face area and that part of this specificity is pickedup by the far-field recording (Ishai Ungerleider MartinSchouten amp Haxby 1999) The present study was notdesigned to address this issue Therefore we will turnnow to the focus of this experiment that is the compar-ison between the different face types

The finding that the N170 elicited by schematic facesis not significantly smaller in amplitude is not delayedrelative to natural face photographs and has a similarscalp distribution suggests that common neural mech-anisms underlie the perception of both natural andschematic faces Assuming that these mechanisms arethe neural analogue to the hypothesized structuralencoder for faces these data also suggest that a face-specific visual mechanism is triggered whenever a stim-ulus contains sufficient information to generate theconcept of a face In other words schematically drawnfaces are categorized by an early visual mechanismrather than by a higher-level cognitive process Indeedit is possible that once triggered such a face perceptionmodule may take precedence over a more general visualprocessor A similar account for the suppression ofacoustic perception when phonetic features are de-tected (Bentin amp Mann 1990 Liberman amp Mattingly1989) led to the hypothesis that the phonetic modulemay preempt auditory information that is relevant tophonetic perception (Whalen amp Liberman 1987)

The reduced amplitude of the N170 elicited by detail-richer face sketches relative to that elicited by schematicfaces and its delayed peak latency might be interpretedprima facie as evidence against the above hypothesisHowever post hoc this result can be incorporated intoour theory Although graphically more detailed thesketches did not attempt to imitate a natural face (asthe paintings) or to emphasize those basic featuresapparently required for the formation of a face concept(as in the schematic faces) Indeed it is possible thatsome of the sketches were difficult to interpret ashuman faces without an elaborate analytic processHence the generation of a face concept on the basisof such sketches may have been somewhat more timeconsuming either because they were processed by gen-eral visual processors or because they activated thespecial purpose structural encoder later In additionunlike the schematic faces that although not identicalwere more uniform in shape4 the sketches were stylis-tically more diverse As a consequence some of thesketches may have introduced noise or a time jitter inthe average ERP which could explain both the reducedamplitude and the longer latency of the N170 associatedwith the sketched faces category5 All face categorieselicited a significantly larger N170 in the right hemi-sphere However the asymmetry was reduced forsketches and absent for nonface categories This findingis consistent with neuropsychological (eg De Renzi1997) neuroimaging (eg McCarthy et al 1997 DeRenzi Perani Carlesimo Silveri amp Fazio 1994 Haxbyet al 1993) and behavioral studies (eg Luh Rueckertamp Levy 1991) indicating a right hemisphere advantagefor face processing Furthermore it corroborates thehypothesis that some of the sketches were processed bya general rather than a face-specific visual processor orthat in general the sketches were less efficient ininitiating a face-specific process

The finding that schematic faces elicit a conspicuousN170 that does not significantly differ from the ERPselicited by natural faces is intriguing in light of theassumption that the N170 is associated primarily withthe processing of face components In previous studieswe have consistently found that eyes in isolation or eyescombined with other inner-face components elicit alarger N170 than that elicited by full faces (Bentinet al 1996 Bentin amp McCarthy unpublished) Further-more full faces elicit an N170 which is only slightlylarger than that elicited by face contours from which allinner components have been erased On the basis ofthese data Bentin et al (1996) suggested that the N170is modulated by two neural sources (both part of thestructural encoding mechanism) one possibly locatedin the posterior upper bank of the occipito-temporalsulcus (OTS) or in the inferior temporal (IT) gyrus istriggered by physiognomic information appearing in thevisual field the other probably located in the middlefusiform gyrus is responsible for holistic face processing

942 Journal of Cognitive Neuroscience Volume 13 Number 7

and integrates the face parts into a whole Assuming thatwithout the framing contour and the face-specific con-figuration inner components of schematic faces have nophysiognomic value6 the N170 elicited by schematicfaces should be modulated primarily by the fusiformsource which is responsible for the holistic processingof a face Consequently the N170 should be particularlysusceptible to manipulations affecting holistic process-ing of schematic faces An initial attempt to explore thispossibility is reported in Experiment 2 In addition thesimilar latency for the N170s elicited by natural faces andschematic faces would thus be in agreement with theprimacy of holistic processing for face stimuli the lateralcomponent primarily responsible for the N170 is trig-gered by the holistic component whether or not theface parts are recognizable as such in isolation ordefined by the global configuration

EXPERIMENT 2

The face inversion effectmdashthat is the difficulty to iden-tify familiar faces presented upside down (comparedwith other familiar objects subject to a similar manipu-lation)mdashhas been the focus of many studies for morethan two decades (reviewed by Valentine 1988) Thedifference between the effects of face inversion relativeto inversion of other objects is usually attributed to abasic difference between the processing strategies in-curred by faces versus objects According to this hypoth-esis recognition of faces relies on the formation of awhole-face representation from which relations betweenthe inner-face parts are derived that is holistic process-ing (Farah Tanaka amp Drain 1995 Tanaka amp Farah 1993versions of the holistic hypothesis are reviewed byMoscovitch Winocur amp Behrmann 1997) In contrastthe recognition of other perhaps less stereotypic ob-jects entails initial identification of basic componentswhich are then related one to the other to form anunequivocal visual percept (eg Biderman 1987 Marr1982 Marr amp Nishihara 1978) The face inversion effecthas become the hallmark of normal face recognition inhumans and has often been used to distinguish face-specific from general visual processes

Bentin et al (1996) compared the N170 elicited byupright and inverted faces and found that the N170elicited by inverted faces was slightly larger and peaked10 msec later than that elicited by upright faces only thedifference in latency was significant in that study Otherinvestigators however reported that the enhancementof the N170 amplitude by face inversion is significant aswell as the delay in latency (eg Rossion et al 19992000) A possible interpretation of the enhanced N170amplitude in response to inverted faces is that the scalp-recorded N170 is associated primarily with the percep-tion of face components a process that might not beimpeded by inversion This hypothesis conforms to thepresumed anatomical location of the putative compo-

nents detector (in lateral OTS and IT gyrus) and theorientation of these cell assemblies relative to the re-cording electrodes on the scalp In contrast to the N170the intracranially recorded N200 (on the ventral plane ofthe temporal lobe) is modulated primarily by the holis-tic face integration mechanism located in the posteriorlateral parts of the fusiform gyrus (eg Allison et al1999) Hence given the orientation of the electricaldipole fields in these two cell assemblies a manipulationthat inhibits holistic perception should enhance theN170 amplitude This hypothesis also conforms to thefinding that the N200 (recorded at the fusiform) elicitedby faces is rsquo rsquo largerrsquo rsquo than that elicited by isolated eyes(McCarthy et al 1999) whereas the N170 (recorded atthe scalp) elicited by faces is rsquo rsquosmallerrsquo rsquo than that elicitedby isolated eyes (Bentin et al 1996)

According to this model face perception is specificbecause the highly stereotypic organization of facespermits categorizing the visual stimulus before a fulldescription of its features is provided by the analyticmechanism Therefore most of the time faces activatethe holistic perception mechanism instantly Indeedonly such a model has the capacity to explain the resultsof Experiment 1 in which a visual stimulus that can beperceived as a face only if processed as a whole (ie ahighly schematic face) elicited a similar N170 to thatelicited by photographs of natural faces

Based on the above conceptualization we hypothe-size that if the detection of schematic faces is basedprimarily on the holistic face configuration compromis-ing their gestalt by presenting the stimuli upside downshould reduce the N170 In contrast turning naturalfaces upside down should enhance the amplitude of theN170 since the face component processing modulewould still be activated To test this hypothesis wepresented a new group of participants with naturaland schematic faces in upright and inverted orienta-tions

Results

Figure 3 presents the N170s elicited by photographs ofnatural faces and schematic faces presented upright orupside down Replicating the findings of Rossion et al(1999) the N170 elicited by inverted natural faces waslarger than that elicited by upright faces In contrast theN170 elicited by inverted schematic faces was reduced inamplitude relative to that elicited by upright schematicfaces Inversion delayed the peak latency of the N170 forschematic as well as for natural faces

The reliability of these observations was statisticallyanalyzed by a within-subject ANOVA using repeatedmeasures (see data in Table 2A)

The factors were stimulus type (natural faces sche-matic faces) stimulus orientation (upright inverted) site(P78 mastoids PO78 IM12) and hemisphere (leftright) As a main effect stimulus orientation was not

Sagiv and Bentin 943

significant [F(115) = 21 p = 16] However the inter-esting pattern of the face inversion effects was revealedby a significant interaction between the effects of orien-tation and stimulus type [F(115) = 248 p lt 001]followed by planed comparisons The planed compari-sons showed that whereas for natural photographs theN170 elicited by inverted faces (ndash 66 mV) was larger thatthat elicited by upright faces (ndash 55 mV) with schematicfaces inversion was associated with a reduction inthe N170 amplitude (ndash 38 and ndash 50 mV for invertedand upright presentation respectively t(15) = 2104p lt 055 and t(15) = 3000 p lt 01 for natural andschematic faces respectively) The main effect of stim-ulus type was significant [F(115) = 302 p lt 001] but itwas also qualified by its interaction with the effect oforientation Post hoc comparisons showed that whereasthe difference between the N170 elicited by invertednatural and schematic faces was significant [t(15) =4264 p lt 001] the difference between the N170elicited by natural and schematic faces presented uprightwas not [t(15) = 884 p = 39] The main effect of site

was also significant [F(345) = 68 p lt 01] revealingonly that the N170 amplitudes at P7 and P8 were smallerthan at all other sites [F(115) = 117 p lt 01] It isnoteworthy that as evidence by an insignificant interac-tion with stimulus type [F(345) = 13 p = 28] thiseffect was similar for schematic and natural faces On theother hand a significant main effect of hemisphere[F(115) = 65 p lt 05] was qualified by its interactionwith stimulus type [F(115) = 53 p lt 05] Post hocanalysis revealed that whereas for both stimulus typesthe amplitude of the N170 was larger over the rightcompared to the left hemisphere the interhemisphericdifference was significant for natural faces [t(15) = 335p lt 01] but not for schematic faces [t(15) = 1408p = 18] The effects of the Hemisphere pound Orientationinteraction was significant [F(115) = 47 p lt 5] Thisinteraction was caused by the fact that whereas acrossstimulus type the reduction and enhancement of theN170 amplitude by inversion canceled each other in theleft hemisphere in the right hemisphere the enhance-ment of N170 induced by the inversion of natural faces

Faces Inverted Faces Schematic Faces Inverted SchematicFaces

Time(msec) Time(msec)0 100 200 300 400

10

mV 0

0 100 200 300 400

10PO7 PO8

- 0 +

mV 0

Figure 3 The N170 potentials elicited at the left and right mastoids by upright and inverted natural faces and schematic faces and the scalpdistribution at this latency

944 Journal of Cognitive Neuroscience Volume 13 Number 7

weighted more than its reduction resulting from theinversion of schematic faces However the second-orderinteraction between orientation hemisphere and stim-ulus type was not significant [F(115) lt 10] Finally aseries of second-order interactions involving the effect ofsite suggested that all the effects were differently con-spicuous at different sites As revealed in Table 2 by andlarge it seems that whereas natural faces induced similarinversion effects across the analyzed sites the inversioneffect for schematic faces varied at different sites We hadno prior assumptions about the distribution of the effectswithin the posterior temporal scalp regions Further-more the spatial resolution of scalp-recorded ERPs istoo low to permit reliable interpretation of site effectswithin a relative small region unless the same distribu-tion is replicated therefore a detailed analysis of thisdistribution of effects was not attempted post hoc

ANOVA of the N170 peak latencies showed a signifi-cant main effect of stimulus orientation [F(115) =1615 p lt 001] which significantly interacted withstimulus type [F(115) = 62 p lt 025] Plannedcomparisons showed that for both natural and sche-matic faces inversion delayed the N170 significantly[t(15) = 10821 p lt 001 and t(15) = 3626 p lt 01respectively] However as evident in Table 2B thisdifference was larger for natural than for schematicfaces The latencies were similar over the left and righthemispheres [F(115) = 19 p = 185] The main effectof site was significant [F(345) = 47 p lt 01] butbecause the differences between sites were all smaller

than the sampling rate this effect was not investigatedfurther No other significant effects emerged in thisanalysis

Discussion

The results of the Experiment 2 validated our predic-tion that inversion should reduce the amplitude of theN170 elicited by schematic faces but have little effector enhance the N170 elicited by natural faces Thenatural face inversion effect observed in the presentstudy replicates previous results reported by Rossionet al (1999) These authors suggested that the en-hancement in the N170 amplitude reflects the en-hanced difficulty of the holistic processor to encodethe inverted face Indeed the delay in the N170 peaklatency elicited by inverted relative to upright faces iscongruent with such a view However if this were thecase we should have found similar effects of inversionon the N170 amplitude elicited by schematic and bynatural faces Furthermore the assumption that moredifficult face encoding induces an enhancement of theN170 amplitude is also at odds with fMRI data showingthat face inversion reduces the activity in the fusiformface areas relative to the activity elicited by uprightfaces (Kanwisher Tong amp Nakayama 1998) Indeedfinding opposite effects of inversion for schematic andnatural faces can more likely be accommodated by ourhypothesis that the larger N170 amplitude in responseto inverted natural faces reflects its stronger associa-

Table 2 (A) Amplitudes (in mV) and (B) Latencies (in msec) of the N170 Elicited by Upright and Inverted Natural and SchematicFaces at the Posterior Temporal Scalp Sites

Natural Faces Schematic Faces

Right Hemisphere Left Hemisphere Right Hemisphere Left Hemisphere

Upright Inverted Upright Inverted Upright Inverted Upright Inverted

A

Mastoid ndash 64 ndash 86 ndash 58 ndash 82 ndash 54 ndash 55 ndash 51 ndash 35

P87 ndash 48 ndash 68 ndash 45 ndash 54 ndash 41 ndash 30 ndash 32 ndash 33

PO87 ndash 68 ndash 88 ndash 48 ndash 50 ndash 56 ndash 56 ndash 50 ndash 23

IM21 ndash 53 ndash 81 ndash 49 ndash 73 ndash 50 ndash 29 ndash 65 ndash 39

Mean ndash 58 ndash 81 ndash 50 ndash 65 ndash 50 ndash 425 ndash 49 ndash 325

B

Mastoid 149 162 151 165 159 164 158 165

P87 151 163 154 166 158 164 162 166

PO87 151 160 152 164 151 161 159 166

IM21 150 162 147 161 161 164 154 162

Mean 150 162 151 164 157 163 158 165

Sagiv and Bentin 945

tion with a face component analyzer in the lateraltemporal gyri than with the holistic face processor inthe fusiform

Our model proposes that although schematic andnatural faces probably activate similar neural networksin the extrastriate visual pathway these mechanismsare not identical Natural faces probably activate twoperceptual modules one specialized to detect andprocess physiognomic features in the visual field andthe other specialized in the holistic processing of facesor forming a holistic face representation Unlike naturalfaces the components of schematic faces are notperceived as carrying physiognomic information outof the schematic face gestalt Therefore in contrastto natural faces schematic faces may trigger the holisticprocessor but they do not trigger directly the analysisof the components The latter mechanism may beactivated only subsequent to the formation of theholistic face representation or at the very least follow-ing this process rsquo rsquo in cascadersquo rsquo (McClelland 1979) Con-sequently adversely affecting the perception of the facegestalt (by inverting the stimuli) inhibited both holisticperception and the perception of individual compo-nents and this inhibition reduced the amplitude of theN170 Such a pattern is also in agreement with themyriad of studies supporting the primacy of holisticprocessing of face stimuli

GENERAL DISCUSSION

In the present study we investigated several character-istics of the extrastriate structural face-encoding mech-anisms by comparing the N170 elicited by faces differingin their naturalistic aspect and by comparing the faceinversion effects on the N170 elicited by photographs ofnatural faces and by schematically drawn faces To theextent that the N170 is indeed associated with structuralencoding we found that this process distinguishes wellbetween faces and nonfaces stimuli but not amongdifferent face types that render the concept of a faceclearly However we also found evidence that schematicfaces and natural faces are not processed identicallyWhereas inversion of natural faces enhanced the ampli-tude of the N170 inversion of schematic faces reducedits amplitude For both face types the latency of theN170 peak was significantly delayed by inversion Thispattern of results suggests that the face-specific struc-tural encoder can be triggered by a variety of stimuli ifthey include a canonical face configuration Howeverthe process of encoding the face information and form-ing a structural representation is probably differentwhen the physiognomic value of the stimuli dependsupon holistic configuration as opposed to when theindividual components can be associated with faces evenwhen presented outside the face context

Apart from the neurofunctional specificity issue anenduring question in face perception research is whether

the formation of a mental representation of the faceentails extraction of a holistic configuration directly fromthe visual input or whether it is feature-based that iswhether for faces as for other complex visual stimulistructural encoding is based on an initial analysis of parts(for a review of this debate see eg Farah 1990 Farahet al 1995) In principle because faces form a veryhomogeneous and stereotypical category they lendthemselves to holistic perception Holistic processing isadvantageous in face recognition because face compo-nents are fairly similar across individuals and thereforethe distinction among faces is usually based on theinterrelationships between the different face compo-nents and their overall configuration (eg Tanaka ampFarah 1993 Young Hellawell amp Hay 1987)7 Moreoverour memory for faces is usually based on nondecompos-able holistic images which at the interface betweenperception and memory are most efficiently addressedby integrated face representations8However if the well-learned gestalt is altered and holistic perception is notpossible it is probably replaced by a more analyticcomponent-based process of structural encoding Thisanalytic process may inhibit or at least delay the per-ception of the spatial relationships between face com-ponents and therefore may impede identificationIndeed a major source of evidence supporting theholistic view is that faces are more difficult to recognizewhen presented upside down (ie compromising theirfamiliar gestalt) than when presented in upright orien-tation (see reviews in Farah et al 1995 Rhodes et al1993 Valentine 1988)

Having to cope with different types of face input thevisual system must be able to encode the face config-uration directly from the visual display as well as byintegrating face components encoded individuallyThese two processes may concur triggered independ-ently by relevant physiognomic input and performed bydifferent neural circuits to form a complex structuralencoding mechanism in the extrastriate cortex Further-more the two systems may interact supporting orinhibiting each other Hence a regular face presentedin upright orientation may activate the holistic percep-tion process which in turn may inhibit superfluousanalysis of individual components By contrast corrupt-ing the face gestalt by presenting faces in an unusualorientation may trigger (and facilitate) face componentanalysis which should not depend on a particular con-figuration In such a case the integration of the compo-nents into a whole by the holistic processor will followtheir independent analysis assisted by the feature-detection system This hypothesis may explain for in-stance why within the context of a regular face individ-ual face parts are distinguished faster than in isolationbut if the configuration of the face is altered the partsare distinguished as fast as when presented in isolation(Tanaka amp Farah 1993 Sergent 1984)9 The ability toprocess face information both holistically and piecewise

946 Journal of Cognitive Neuroscience Volume 13 Number 7

was implemented in a neural model initially based onscalp recordings (Bentin et al 1996) and more recentlycorrected improved and elaborated on the basis intra-cranial recordings (McCarthy et al 1999 for a detaileddescription and a review of supportive functional MRIdata see McCarthy 1999) This model suggests that theneural mechanism dedicated to structural encoding offaces is a complex system composed of at least twoseparate subsystems One probably located in thelateral part of the middle fusiform gyrus is responsiblefor processing the configuration of the face as a wholeand provides the integrated face representation that isnecessary for face recognition and identification Theother possibly distributed over posterior regions in theIT and middle temporal (MT) gyri and in the OTS isresponsible for detecting physiognomic informationregardless of the spatial configuration of the items thatcarry it The fusiform gyrus is farther away from therecording sites at the scalp than where the temporalgyri are and the orientation of its cells forms anelectrical dipole source that is almost perpendicular tothe lateral aspects of the scalp Hence although bothsources contribute to the modulation of the face-spe-cific N170 the contribution of the posterior temporalface component processor should be considerably moreconspicuous

According to this hypothesis the opposite effects ofinversion on the N170 elicited by natural and schematicfaces might be explained assuming that the holisticencoding system in the fusiform gyrus indeed inhibitsthe component analysis in the lateral temporal lobe facesites Hence if the face stimulus can trigger both theholistic and the component analysis (as it is true fornatural faces) an experimental manipulation that inter-feres with holistic processing (like the face inversionprocedure in the present experiment) should reducethe amount of inhibition of component analysis andconsequently the amplitude of the N170 would beenhanced However if the rsquo rsquocomponentsrsquorsquo do not conveyphysiognomic information so that their analysis by aface-specific mechanism rsquo rsquodependsrsquo rsquo on the previousrecognition of the face gestalt (as it is true for schematicfaces) interfering with holistic perception cannot facil-itate the activity of the component analysis and there-fore the modulation of the N170 would only reflect theadditional difficulty to encode the face ie reducedamplitude and delayed latency

The observed effects of the different factors on theN170 peak latency are nicely aligned with the abovemodel Although stimulus inversion delayed the N170for both natural and schematic faces this delay was twiceas big (12 msec) for former stimuli than for the latter (6msec) This interaction suggests that the inversion had abigger effect on natural than on schematic faces Thismay be because regardless of whether they are upright orinverted the physiognomic value of schematic faces isbased on their gestalt Hence inversion does inhibit the

processing of the gestalt but does not induce a change inprocessing strategy On the other hand natural faces areanalyzed holistically if they are upright but inversionrequires a change in that strategy encouraging theanalysis of components The significantly faster responseto upright natural than to upright schematic faces inExperiment 2 suggests that the extraction of the facegestalt is easier if the face is more natural

There is ample evidence indicating that the righthemisphere is specialized in global or holistic processingof objects and particularly faces whereas the left hemi-sphere is specialized in part-based processing (egRobertson amp Delis 1986 for a review see Bradshow ampNettleton 1983 for specific reference to face perceptionsee eg Bruce 1998 Rhodes et al 1993 Corballis1988 Bradshow amp Sherlock 1982 Gilbert amp Bakan1973) Consequently interfering with holistic processingshould have had stronger effects in the right than in theleft hemisphere Our results confirmed this predictiononly partially The pronounced interhemispheric asym-metry with larger N170 amplitudes at the right than atthe left hemisphere sites for all stimulus types suggests apredominant holistic process for all face categoriesInterestingly this asymmetry was reduced (in Experi-ment 1) for sketches relative to the other stimulus typesThe relatively enhanced involvement of the left hemi-sphere in processing the sketches is in agreement withour suggestion that the face-specific structural encodingsystem was activated less efficiently by sketches than byother face types10 Hence the same factor that may haveconstrained the processing of schematic faces may alsohave constrained the processing of sketches but in theopposite direction Whereas schematic faces enhancedholistic processing by providing a clear face gestalt whileobscuring the physiognomic value of face componentsthe sketches reduced holistic processing by blurring orveiling the face gestalt Also in Experiment 2 the N170was larger over right than over left hemisphere sitesHowever because the interaction between the inversionand hemisphere effects in Experiment 2 was not modu-lated by stimulus type (ie the three-way interactionwas not significant) our interpretation of the interhemi-spheric asymmetry (particularly in Experiment 2) mustbe limited

In conclusion the present scalp-recorded N170 dataconverge with the previously published pattern of theintracranially recorded N200 data (for a comprehensivereview see Allison et al 1999 McCarthy et al 1999Puce et al 1999) supporting a multifunctional organ-ization of the structural encoding mechanism for facesApparently distinct neural circuits are responsible forthe encoding and initial analysis of face components andfor holistic perception of a full-face configuration Theformer is probably located in the lateral posterior tem-poral lobe whereas the latter is located in the middlefusiform gyrus Face perception usually entails predom-inantly holistic processing which explains the right

Sagiv and Bentin 947

hemisphere dominance in performance as well as inelectrophysiological and neuroimaging data Howeverboth mechanisms can be triggered independently byrelevant information

METHODS

Participants

Right-handed participants were recruited among volun-teer students from the Hebrew University They receivedeither experimental credit hours or payment for theirparticipation Thirty participants (13 males and 17 fe-males) were tested in the first experiment and 16participants (8 males and 8 females) were tested in thesecond experiment Participantsrsquo ages ranged from 19 to37 years (mean = 233 years) They all had normal orcorrected-to-normal vision

Stimuli

Stimuli were scanned using a desk scanner processedwith commercially available graphic software and con-

verted to 16 color grayscale 248 by 248 pixel imagesThere were eight stimulus categories four face types(photographs of natural faces paintings face sketchesand schematic faces) two nonsense patterns (formedby scrambling the natural faces and the schematicfaces) flowers and butterflies (Figure 4) Paintingsand sketches of faces were scanned from art booksPaintings were typically realistic Renaissance portraitsand face sketches did not portray details and usuallylacked shading and texture Schematic face parts weredrawn by hand scanned into the computer and latercombined into different schematic faces Face photo-graphs and paintings contained many shades of grayand the sketches and the schematic faces were basicallyblack on white All face stimuli were front views

Scrambled faces were created by scrambling thephases of the two-dimensional Fourier transform andthen smoothing the uneven edges created by this proc-ess The original and the scrambled stimuli were equallyluminous following this procedure Each stimulus cat-egory comprised 75 stimuli except for the butterflycategory which consisted of 30 stimuli In Experiment2 only natural faces schematic faces and flowers were

Face Stimuli

Photographs Painted portraits Sketches Schematic faces

Nonface Stimuli

Scrambledphotographs

Butterflies(targets) Flowers Scrambled

schematic faces

Figure 4 Example of the eight stimulus types used in the present study

948 Journal of Cognitive Neuroscience Volume 13 Number 7

used The faces were presented in both upright andupside down orientation Upright stimuli were the samein Experiments 1 and 2

All stimuli were presented for 350 msec with anaverage interstimulus interval (ISI) of 12 sec and viewedfrom a distance of 110 plusmn 10 cm subtending a viewingangle of approximately 68 Stimulus timing was con-trolled by the MEL software system (Psychology Soft-ware Tools Pittsburgh PA)

Task

Participants were instructed to keep a mental countof stimuli belonging to one category butterflies inthe first experiment and flowers in the secondHence all of the experiment-relevant categories werefrom the participantsrsquo point of view nontargets Thetargets were presented in approximately 5 (Experi-ment 1) and 6 (Experiment 2) of the trials Duringeach break and at the end of the experimentparticipants reported the number of targets countedA few participants who did not count correctly wereexcluded

EEG Recording and Data Analysis

Continuous EEG was recorded from 48 scalp locationsusing tin electrodes attached to elastic electrode caps(Electrode Caps International Ohio) A schematic illus-tration of the electrode locations and names is pre-sented in Figure 1 A ground electrode was located onthe forehead and a reference electrode was located onthe tip of the nose Two additional electrodes located atthe outer canthus and the infraorbital region of the righteye measured vertical and horizontal electrooculograms(EOGs)

The EEG was sampled at 250 Hz by a DAP2400 DataAcquisition Processor (Microstar Laboratories) ampli-fied 20000-fold using a battery-operated HU-5072 Iso-lated Bioelectric Amplifier (SA Instrumentation SanDiego) with a frequency bandpass of 01ndash30 Hz andstored on disk for off-line averaging Electrode impe-dance was below 5 k A digital filter of 08ndash16 Hz wasapplied before analysis

Codes synchronized to stimulus delivery were used toselectively average epochs associated with differentstimulus categories Epochs contaminated with saccadesor eye blinks were excluded using root-mean-squareEOG-values as criteria Typically epochs associated with55ndash70 (never less than 50) out of 75 stimuli in eachcategory were averaged

The mean and peak amplitudes and peak latencies ofspecific ERP components within bounding intervals werecomputed for each category for each participant Sig-nificance of the differences between these measures wastested using a repeated-measures ANOVA Where appro-priate the more conservative GreenhousendashGeisser ad-

justed df values were used ANOVAs were followed byunivariate F contrasts

Acknowledgments

This study was supported by grants from the USndashIsraelBinational Science Foundation and Israeli Foundations Trust-ees to S Bentin We thank Amir Katz for his help in preparingthe stimuli and running Experiment 1 as well as Leon DeouellMichael Fink Ayelet Landau Yulia Goland David Carmel GilRaviv Amir Raz and Dan Tzur for their support We are alsovery grateful to Jerry Levinson Alex List and Lynn Robertsonfor their comments on the manuscript

Reprint requests should be sent to Shlomo Bentin TheHebrew University of Jerusalem Mount Scopus JerusalemIsrael 91905 or via e-mail msbentinmscchujiacil

Notes

1 Present address University of California Berkeley2 There was no reliable difference between the ERP elicited

by scrambled natural faces and scrambled schematic facesTherefore these two categories were collapsed into one3 These were the sites at which the N170 was most

conspicuous4 So were in fact all face categories except for the sketches5 Note that this result also contradicts the view that a more

difficult encoding process elicit N170 of a higher amplitude6 This assumption was indeed supported by a current study

(Bentin Mecklinger Bosch Sagiv amp von Cramon 1999)7 Of course some faces have very distinctive components

which can be used as markers for recognition Such markersare frequently exaggerated and used as distinctive features incaricatures (Rhodes 1987)8 Distinctive features (such as unusual components birth-

marks etc) may be stored in semantic memory as part of apersonrsquo s face entry Such information may have visual qualitiesthat can be addressed directly during visual matching tofacilitate identification9 Note that this account for the rsquorsquoface superiority effectrsquo rsquo is

identical to a common explanation of the rsquorsquoword superiorityeffectrsquo rsquo in the identification of single letters Indeed ourassumption about the coexistence of rsquo rsquowhole-facersquo rsquo and rsquo rsquoface-partrsquo rsquo processors resonates to the rsquorsquodual routersquo rsquo model of visualword processing It is quite possible that identical principlesgovern all the content-specific perceptual modules in the visualsystem10 Recall that relative to other face types the reduction of theN170 amplitude elicited by sketches at right hemisphere siteswas greater then at left hemisphere sites

REFERENCES

Allison T Ginter H McCarthy G Nobre A C Puce ALuby M amp Spencer D D (1994) Face recognition in hu-man extrastriate cortex Journal of Neurophysiology 71821ndash825

Allison T McCarthy G Nobre A C Puce A amp Belger A(1994) Human extrastriate visual cortex and the perceptionof faces words numbers and colors Cerebral Cortex 5544ndash554

Allison T Puce A Spencer D D amp McCarthy G (1999)Electrophysiological studies of human face perception I

Sagiv and Bentin 949

The statistical reliability of the facenonface differenceand of the apparent differences between the N170elicited by each face type was assessed by a series ofwithin-subject three-factor analysis of variance (ANOVA)with repeated measures The factors were stimulus type(four face types flowers scrambled faces)2 hemisphere(left right) and site (P78 left and right mastoids PO78IM12)3 The dependent variables were the N170 ampli-tude and its latency (Table 1)

The ANOVA of the amplitude showed that there weresignificant differences among the different stimulustypes [F(5145) = 856 p lt 001] that the N170 waslarger at right (ndash 652 mV) than at left (ndash 492 mV)hemisphere sites [F(129) = 196 p lt 001] and thatthere was a significant effect of site [F(387) = 598 p lt005] Post hoc comparisons revealed that the N170elicited by faces (weighted across all four face catego-ries) was significantly more negative than that elicited bynonfaces (weighted across scrambled faces and flowers)[F(129) = 1836 p lt 001] The ERPs elicited by flowers

were significantly less negative than those elicited bysketches [F(129) = 561 p lt 001] but significantlymore negative than those elicited by scrambled faces[F(129) = 223 p lt 001] Because our present experi-ment focused on the potential difference in processingdifferent face types and because of the clear distinctionbetween the N170 elicited by faces and the ERPs elicitedby nonface stimuli in the same time range our subse-quent analyses focused only on the four face types Thedesign of these ANOVAs was similar to the one de-scribed above except that the stimulus type factorincluded only the four face categories

The ANOVA of the amplitude showed that all threemain effects were significant [F(387) = 110 p lt 001F(387) = 37 p lt 025 F(129) = 203 p lt 001 forstimulus type site and hemisphere respectively] Posthoc contrasts showed that the amplitude of the N170elicited by sketches was significantly less negative thanthat elicited by the three other face types [F(129) =468 p lt 001] There was no significant difference

- 0 +

SketchesPhotographs Paintings Schematic Faces

10

Time (msec)0 100 200 300 400 500

Left Mastoid 10

0Time (msec)

100 200 300 400 500

Right Mastoid

mV 0 mV 0

Figure 2 The N170 potentials elicited at the left and right mastoids by the four face types and the scalp distribution of the potentials at thislatency (spline interpolation)

940 Journal of Cognitive Neuroscience Volume 13 Number 7

between the amplitudes of N170 elicited by photo-graphs paintings and schematic faces The N170 ampli-tudes across face categories were largest at the mastoidsand in descending order IM12 P78 and PO78 Sig-nificant differences were found only between the mas-toids and the P78 and PO78 locations [F(129) = 50p lt 05 and F(129) = 48 p lt 05 respectively] andbetween the IM12 and the PO78 locations [F(129) =49 p lt 05]

All three first-order interactions were significant[F(9261) = 48 p lt 001 F(387) = 31 p lt 05F(387) = 63 p lt 01 for the Stimulus Type pound SiteStimulus Type pound Hemisphere and Site pound Hemisphererespectively] The second-order interaction between allthree factors was also significant [F(9261) = 34 p lt001] Post hoc examination of these interactions re-vealed the following pattern The difference betweenthe N170 elicited by sketches and that elicited by theother face types was larger at the right hemisphere sitesthan at the left hemisphere sites and slightly smaller atIM sites than at the other three sites The interactionbetween hemisphere and scalp location is probably aresult of the smaller (but clear) right hemisphere ad-vantage at scalp site IM12 [F(129) = 90 p lt 01]compared to other sites [F(129) = 153 p lt 001F(129) = 182 p lt 001 F(129) = 242 p lt 001 forPO78 mastoids and P78 respectively] The second-order interaction reflected (a) a smaller (Site pound Hemi-

sphere) interaction for the sketched faces category[F(387) = 23 vs F(387) gt 59 for each of the othercategories] (b) an insignificant (Stimulus Type pound Hemi-sphere) interaction at the IM scalp site [F(387) lt 100]versus a significant or close-to-significant effect for othersites [F(387) gt 24] and (c) a smaller (Stimulus Type poundSite) interaction in the right hemisphere [F(15435) =30] than in the left hemisphere [F(15435) = 55]

As evident in Table 1B the mean N170 latency wasshortest for schematic faces (1676 msec) followed byphotographs (1693 msec) paintings (1714 msec) andsketches (1743 msec) Scalp sites also differed on N170latency the site at which the shortest N170 latency wasrecorded was PO78 (1697 msec) followed by P78(1702 msec) IM12 (1704 msec) and RL-Mast (1723msec) ANOVA showed that both the stimulus type andthe site effects were significant [F(387) = 167 p lt001 F(387) = 50 p lt 005 respectively] Post hoccontrasts revealed that differences between the meanN170 latencies of all stimulus type pairs were highlysignificant [F(129) = 92 p lt 005] except for thedifference between face photograph and schematic facelatencies [F(129) = 30 p lt 1] In addition the N170latency at the mastoid scalp locations (1723 msec) wassignificantly delayed [F(129) = 173 p lt 001] com-pared with the three other locations (1701 msec)which did not significantly differ from each other[F(129) = 5 p = 6]

Table 1 (A) Mean Amplitudes (in mV) and (B) Mean Latencies (in msec) of the N170 Elicited by Natural Faces Painted PortraitsSketches Schematic Faces Flowers and Scrambled Faces at the Posterior Temporal Scalp Sites

Left Hemisphere Right Hemisphere

Left mastoid P7 PO7 IM1 IM2 PO8 P8 Right mastoid

A

Natural Faces ndash 72 ndash 62 ndash 60 ndash 76 ndash 89 ndash 93 ndash 91 ndash 97

Painted Portraits ndash 73 ndash 61 ndash 62 ndash 78 ndash 88 ndash 88 ndash 92 ndash 98

Sketches ndash 58 ndash 49 ndash 41 ndash 61 ndash 73 ndash 64 ndash 70 ndash 76

Schematic Faces ndash 68 ndash 63 ndash 60 ndash 70 ndash 83 ndash 88 ndash 96 ndash 94

Flowers ndash 29 ndash 32 ndash 23 ndash 28 ndash 30 ndash 29 ndash 33 ndash 34

Scrambled Faces ndash 19 ndash 17 ndash 07 ndash 13 ndash 15 ndash 07 ndash 16 ndash 22

B

Natural Faces 170 168 170 168 169 169 170 170

Painted Portraits 171 171 172 171 171 171 172 172

Sketches 177 171 171 174 175 173 176 178

Schematic Faces 170 165 165 168 167 167 169 169

Flowers 174 166 168 171 172 168 171 176

Scrambled Faces 177 167 166 173 175 166 168 178

Sagiv and Bentin 941

Hemispheric differences failed to reach significance[F(129) lt 10] A Site pound Hemisphere interaction washowever reliable [F(387) = 40 p lt 005] This inter-action is possibly the result of relatively late right hemi-sphere mean latency at P78 and PO78 scalp locationsThe interaction between stimulus type and site was alsosignificant [F(9261) = 35 p lt 001] reflecting a greatersite effect within the sketched face category The Stim-ulus Type pound Hemisphere interaction was not significant[F(387) = 14 p lt 25] The interaction between allthree factors did not reach significance [F(9261) = 17p = 1]

Discussion

Experiment 1 demonstrated that face-specific N170 com-ponents can be recorded with scalp electrodes not onlywhen photographs of natural faces are presented butalso when paintings sketches and even highly sche-matic drawings of faces are used as stimuli The ampli-tude of the N170 elicited by all face stimuli was at leasttwice as big as the amplitude of the negative potentialelicited by flowers and three times as big as the ampli-tude of the potential elicited by nonsense images duringthe same time The significant difference between flow-ers and scrambled faces could reflect either that theneural activity associated to the structural encoding ofany object is distinct from that elicited by visual patternsthat are not objects (McCarthy et al 1997) or that thereis specificity for other stimulus categories in the vicinityof the face area and that part of this specificity is pickedup by the far-field recording (Ishai Ungerleider MartinSchouten amp Haxby 1999) The present study was notdesigned to address this issue Therefore we will turnnow to the focus of this experiment that is the compar-ison between the different face types

The finding that the N170 elicited by schematic facesis not significantly smaller in amplitude is not delayedrelative to natural face photographs and has a similarscalp distribution suggests that common neural mech-anisms underlie the perception of both natural andschematic faces Assuming that these mechanisms arethe neural analogue to the hypothesized structuralencoder for faces these data also suggest that a face-specific visual mechanism is triggered whenever a stim-ulus contains sufficient information to generate theconcept of a face In other words schematically drawnfaces are categorized by an early visual mechanismrather than by a higher-level cognitive process Indeedit is possible that once triggered such a face perceptionmodule may take precedence over a more general visualprocessor A similar account for the suppression ofacoustic perception when phonetic features are de-tected (Bentin amp Mann 1990 Liberman amp Mattingly1989) led to the hypothesis that the phonetic modulemay preempt auditory information that is relevant tophonetic perception (Whalen amp Liberman 1987)

The reduced amplitude of the N170 elicited by detail-richer face sketches relative to that elicited by schematicfaces and its delayed peak latency might be interpretedprima facie as evidence against the above hypothesisHowever post hoc this result can be incorporated intoour theory Although graphically more detailed thesketches did not attempt to imitate a natural face (asthe paintings) or to emphasize those basic featuresapparently required for the formation of a face concept(as in the schematic faces) Indeed it is possible thatsome of the sketches were difficult to interpret ashuman faces without an elaborate analytic processHence the generation of a face concept on the basisof such sketches may have been somewhat more timeconsuming either because they were processed by gen-eral visual processors or because they activated thespecial purpose structural encoder later In additionunlike the schematic faces that although not identicalwere more uniform in shape4 the sketches were stylis-tically more diverse As a consequence some of thesketches may have introduced noise or a time jitter inthe average ERP which could explain both the reducedamplitude and the longer latency of the N170 associatedwith the sketched faces category5 All face categorieselicited a significantly larger N170 in the right hemi-sphere However the asymmetry was reduced forsketches and absent for nonface categories This findingis consistent with neuropsychological (eg De Renzi1997) neuroimaging (eg McCarthy et al 1997 DeRenzi Perani Carlesimo Silveri amp Fazio 1994 Haxbyet al 1993) and behavioral studies (eg Luh Rueckertamp Levy 1991) indicating a right hemisphere advantagefor face processing Furthermore it corroborates thehypothesis that some of the sketches were processed bya general rather than a face-specific visual processor orthat in general the sketches were less efficient ininitiating a face-specific process

The finding that schematic faces elicit a conspicuousN170 that does not significantly differ from the ERPselicited by natural faces is intriguing in light of theassumption that the N170 is associated primarily withthe processing of face components In previous studieswe have consistently found that eyes in isolation or eyescombined with other inner-face components elicit alarger N170 than that elicited by full faces (Bentinet al 1996 Bentin amp McCarthy unpublished) Further-more full faces elicit an N170 which is only slightlylarger than that elicited by face contours from which allinner components have been erased On the basis ofthese data Bentin et al (1996) suggested that the N170is modulated by two neural sources (both part of thestructural encoding mechanism) one possibly locatedin the posterior upper bank of the occipito-temporalsulcus (OTS) or in the inferior temporal (IT) gyrus istriggered by physiognomic information appearing in thevisual field the other probably located in the middlefusiform gyrus is responsible for holistic face processing

942 Journal of Cognitive Neuroscience Volume 13 Number 7

and integrates the face parts into a whole Assuming thatwithout the framing contour and the face-specific con-figuration inner components of schematic faces have nophysiognomic value6 the N170 elicited by schematicfaces should be modulated primarily by the fusiformsource which is responsible for the holistic processingof a face Consequently the N170 should be particularlysusceptible to manipulations affecting holistic process-ing of schematic faces An initial attempt to explore thispossibility is reported in Experiment 2 In addition thesimilar latency for the N170s elicited by natural faces andschematic faces would thus be in agreement with theprimacy of holistic processing for face stimuli the lateralcomponent primarily responsible for the N170 is trig-gered by the holistic component whether or not theface parts are recognizable as such in isolation ordefined by the global configuration

EXPERIMENT 2

The face inversion effectmdashthat is the difficulty to iden-tify familiar faces presented upside down (comparedwith other familiar objects subject to a similar manipu-lation)mdashhas been the focus of many studies for morethan two decades (reviewed by Valentine 1988) Thedifference between the effects of face inversion relativeto inversion of other objects is usually attributed to abasic difference between the processing strategies in-curred by faces versus objects According to this hypoth-esis recognition of faces relies on the formation of awhole-face representation from which relations betweenthe inner-face parts are derived that is holistic process-ing (Farah Tanaka amp Drain 1995 Tanaka amp Farah 1993versions of the holistic hypothesis are reviewed byMoscovitch Winocur amp Behrmann 1997) In contrastthe recognition of other perhaps less stereotypic ob-jects entails initial identification of basic componentswhich are then related one to the other to form anunequivocal visual percept (eg Biderman 1987 Marr1982 Marr amp Nishihara 1978) The face inversion effecthas become the hallmark of normal face recognition inhumans and has often been used to distinguish face-specific from general visual processes

Bentin et al (1996) compared the N170 elicited byupright and inverted faces and found that the N170elicited by inverted faces was slightly larger and peaked10 msec later than that elicited by upright faces only thedifference in latency was significant in that study Otherinvestigators however reported that the enhancementof the N170 amplitude by face inversion is significant aswell as the delay in latency (eg Rossion et al 19992000) A possible interpretation of the enhanced N170amplitude in response to inverted faces is that the scalp-recorded N170 is associated primarily with the percep-tion of face components a process that might not beimpeded by inversion This hypothesis conforms to thepresumed anatomical location of the putative compo-

nents detector (in lateral OTS and IT gyrus) and theorientation of these cell assemblies relative to the re-cording electrodes on the scalp In contrast to the N170the intracranially recorded N200 (on the ventral plane ofthe temporal lobe) is modulated primarily by the holis-tic face integration mechanism located in the posteriorlateral parts of the fusiform gyrus (eg Allison et al1999) Hence given the orientation of the electricaldipole fields in these two cell assemblies a manipulationthat inhibits holistic perception should enhance theN170 amplitude This hypothesis also conforms to thefinding that the N200 (recorded at the fusiform) elicitedby faces is rsquo rsquo largerrsquo rsquo than that elicited by isolated eyes(McCarthy et al 1999) whereas the N170 (recorded atthe scalp) elicited by faces is rsquo rsquosmallerrsquo rsquo than that elicitedby isolated eyes (Bentin et al 1996)

According to this model face perception is specificbecause the highly stereotypic organization of facespermits categorizing the visual stimulus before a fulldescription of its features is provided by the analyticmechanism Therefore most of the time faces activatethe holistic perception mechanism instantly Indeedonly such a model has the capacity to explain the resultsof Experiment 1 in which a visual stimulus that can beperceived as a face only if processed as a whole (ie ahighly schematic face) elicited a similar N170 to thatelicited by photographs of natural faces

Based on the above conceptualization we hypothe-size that if the detection of schematic faces is basedprimarily on the holistic face configuration compromis-ing their gestalt by presenting the stimuli upside downshould reduce the N170 In contrast turning naturalfaces upside down should enhance the amplitude of theN170 since the face component processing modulewould still be activated To test this hypothesis wepresented a new group of participants with naturaland schematic faces in upright and inverted orienta-tions

Results

Figure 3 presents the N170s elicited by photographs ofnatural faces and schematic faces presented upright orupside down Replicating the findings of Rossion et al(1999) the N170 elicited by inverted natural faces waslarger than that elicited by upright faces In contrast theN170 elicited by inverted schematic faces was reduced inamplitude relative to that elicited by upright schematicfaces Inversion delayed the peak latency of the N170 forschematic as well as for natural faces

The reliability of these observations was statisticallyanalyzed by a within-subject ANOVA using repeatedmeasures (see data in Table 2A)

The factors were stimulus type (natural faces sche-matic faces) stimulus orientation (upright inverted) site(P78 mastoids PO78 IM12) and hemisphere (leftright) As a main effect stimulus orientation was not

Sagiv and Bentin 943

significant [F(115) = 21 p = 16] However the inter-esting pattern of the face inversion effects was revealedby a significant interaction between the effects of orien-tation and stimulus type [F(115) = 248 p lt 001]followed by planed comparisons The planed compari-sons showed that whereas for natural photographs theN170 elicited by inverted faces (ndash 66 mV) was larger thatthat elicited by upright faces (ndash 55 mV) with schematicfaces inversion was associated with a reduction inthe N170 amplitude (ndash 38 and ndash 50 mV for invertedand upright presentation respectively t(15) = 2104p lt 055 and t(15) = 3000 p lt 01 for natural andschematic faces respectively) The main effect of stim-ulus type was significant [F(115) = 302 p lt 001] but itwas also qualified by its interaction with the effect oforientation Post hoc comparisons showed that whereasthe difference between the N170 elicited by invertednatural and schematic faces was significant [t(15) =4264 p lt 001] the difference between the N170elicited by natural and schematic faces presented uprightwas not [t(15) = 884 p = 39] The main effect of site

was also significant [F(345) = 68 p lt 01] revealingonly that the N170 amplitudes at P7 and P8 were smallerthan at all other sites [F(115) = 117 p lt 01] It isnoteworthy that as evidence by an insignificant interac-tion with stimulus type [F(345) = 13 p = 28] thiseffect was similar for schematic and natural faces On theother hand a significant main effect of hemisphere[F(115) = 65 p lt 05] was qualified by its interactionwith stimulus type [F(115) = 53 p lt 05] Post hocanalysis revealed that whereas for both stimulus typesthe amplitude of the N170 was larger over the rightcompared to the left hemisphere the interhemisphericdifference was significant for natural faces [t(15) = 335p lt 01] but not for schematic faces [t(15) = 1408p = 18] The effects of the Hemisphere pound Orientationinteraction was significant [F(115) = 47 p lt 5] Thisinteraction was caused by the fact that whereas acrossstimulus type the reduction and enhancement of theN170 amplitude by inversion canceled each other in theleft hemisphere in the right hemisphere the enhance-ment of N170 induced by the inversion of natural faces

Faces Inverted Faces Schematic Faces Inverted SchematicFaces

Time(msec) Time(msec)0 100 200 300 400

10

mV 0

0 100 200 300 400

10PO7 PO8

- 0 +

mV 0

Figure 3 The N170 potentials elicited at the left and right mastoids by upright and inverted natural faces and schematic faces and the scalpdistribution at this latency

944 Journal of Cognitive Neuroscience Volume 13 Number 7

weighted more than its reduction resulting from theinversion of schematic faces However the second-orderinteraction between orientation hemisphere and stim-ulus type was not significant [F(115) lt 10] Finally aseries of second-order interactions involving the effect ofsite suggested that all the effects were differently con-spicuous at different sites As revealed in Table 2 by andlarge it seems that whereas natural faces induced similarinversion effects across the analyzed sites the inversioneffect for schematic faces varied at different sites We hadno prior assumptions about the distribution of the effectswithin the posterior temporal scalp regions Further-more the spatial resolution of scalp-recorded ERPs istoo low to permit reliable interpretation of site effectswithin a relative small region unless the same distribu-tion is replicated therefore a detailed analysis of thisdistribution of effects was not attempted post hoc

ANOVA of the N170 peak latencies showed a signifi-cant main effect of stimulus orientation [F(115) =1615 p lt 001] which significantly interacted withstimulus type [F(115) = 62 p lt 025] Plannedcomparisons showed that for both natural and sche-matic faces inversion delayed the N170 significantly[t(15) = 10821 p lt 001 and t(15) = 3626 p lt 01respectively] However as evident in Table 2B thisdifference was larger for natural than for schematicfaces The latencies were similar over the left and righthemispheres [F(115) = 19 p = 185] The main effectof site was significant [F(345) = 47 p lt 01] butbecause the differences between sites were all smaller

than the sampling rate this effect was not investigatedfurther No other significant effects emerged in thisanalysis

Discussion

The results of the Experiment 2 validated our predic-tion that inversion should reduce the amplitude of theN170 elicited by schematic faces but have little effector enhance the N170 elicited by natural faces Thenatural face inversion effect observed in the presentstudy replicates previous results reported by Rossionet al (1999) These authors suggested that the en-hancement in the N170 amplitude reflects the en-hanced difficulty of the holistic processor to encodethe inverted face Indeed the delay in the N170 peaklatency elicited by inverted relative to upright faces iscongruent with such a view However if this were thecase we should have found similar effects of inversionon the N170 amplitude elicited by schematic and bynatural faces Furthermore the assumption that moredifficult face encoding induces an enhancement of theN170 amplitude is also at odds with fMRI data showingthat face inversion reduces the activity in the fusiformface areas relative to the activity elicited by uprightfaces (Kanwisher Tong amp Nakayama 1998) Indeedfinding opposite effects of inversion for schematic andnatural faces can more likely be accommodated by ourhypothesis that the larger N170 amplitude in responseto inverted natural faces reflects its stronger associa-

Table 2 (A) Amplitudes (in mV) and (B) Latencies (in msec) of the N170 Elicited by Upright and Inverted Natural and SchematicFaces at the Posterior Temporal Scalp Sites

Natural Faces Schematic Faces

Right Hemisphere Left Hemisphere Right Hemisphere Left Hemisphere

Upright Inverted Upright Inverted Upright Inverted Upright Inverted

A

Mastoid ndash 64 ndash 86 ndash 58 ndash 82 ndash 54 ndash 55 ndash 51 ndash 35

P87 ndash 48 ndash 68 ndash 45 ndash 54 ndash 41 ndash 30 ndash 32 ndash 33

PO87 ndash 68 ndash 88 ndash 48 ndash 50 ndash 56 ndash 56 ndash 50 ndash 23

IM21 ndash 53 ndash 81 ndash 49 ndash 73 ndash 50 ndash 29 ndash 65 ndash 39

Mean ndash 58 ndash 81 ndash 50 ndash 65 ndash 50 ndash 425 ndash 49 ndash 325

B

Mastoid 149 162 151 165 159 164 158 165

P87 151 163 154 166 158 164 162 166

PO87 151 160 152 164 151 161 159 166

IM21 150 162 147 161 161 164 154 162

Mean 150 162 151 164 157 163 158 165

Sagiv and Bentin 945

tion with a face component analyzer in the lateraltemporal gyri than with the holistic face processor inthe fusiform

Our model proposes that although schematic andnatural faces probably activate similar neural networksin the extrastriate visual pathway these mechanismsare not identical Natural faces probably activate twoperceptual modules one specialized to detect andprocess physiognomic features in the visual field andthe other specialized in the holistic processing of facesor forming a holistic face representation Unlike naturalfaces the components of schematic faces are notperceived as carrying physiognomic information outof the schematic face gestalt Therefore in contrastto natural faces schematic faces may trigger the holisticprocessor but they do not trigger directly the analysisof the components The latter mechanism may beactivated only subsequent to the formation of theholistic face representation or at the very least follow-ing this process rsquo rsquo in cascadersquo rsquo (McClelland 1979) Con-sequently adversely affecting the perception of the facegestalt (by inverting the stimuli) inhibited both holisticperception and the perception of individual compo-nents and this inhibition reduced the amplitude of theN170 Such a pattern is also in agreement with themyriad of studies supporting the primacy of holisticprocessing of face stimuli

GENERAL DISCUSSION

In the present study we investigated several character-istics of the extrastriate structural face-encoding mech-anisms by comparing the N170 elicited by faces differingin their naturalistic aspect and by comparing the faceinversion effects on the N170 elicited by photographs ofnatural faces and by schematically drawn faces To theextent that the N170 is indeed associated with structuralencoding we found that this process distinguishes wellbetween faces and nonfaces stimuli but not amongdifferent face types that render the concept of a faceclearly However we also found evidence that schematicfaces and natural faces are not processed identicallyWhereas inversion of natural faces enhanced the ampli-tude of the N170 inversion of schematic faces reducedits amplitude For both face types the latency of theN170 peak was significantly delayed by inversion Thispattern of results suggests that the face-specific struc-tural encoder can be triggered by a variety of stimuli ifthey include a canonical face configuration Howeverthe process of encoding the face information and form-ing a structural representation is probably differentwhen the physiognomic value of the stimuli dependsupon holistic configuration as opposed to when theindividual components can be associated with faces evenwhen presented outside the face context

Apart from the neurofunctional specificity issue anenduring question in face perception research is whether

the formation of a mental representation of the faceentails extraction of a holistic configuration directly fromthe visual input or whether it is feature-based that iswhether for faces as for other complex visual stimulistructural encoding is based on an initial analysis of parts(for a review of this debate see eg Farah 1990 Farahet al 1995) In principle because faces form a veryhomogeneous and stereotypical category they lendthemselves to holistic perception Holistic processing isadvantageous in face recognition because face compo-nents are fairly similar across individuals and thereforethe distinction among faces is usually based on theinterrelationships between the different face compo-nents and their overall configuration (eg Tanaka ampFarah 1993 Young Hellawell amp Hay 1987)7 Moreoverour memory for faces is usually based on nondecompos-able holistic images which at the interface betweenperception and memory are most efficiently addressedby integrated face representations8However if the well-learned gestalt is altered and holistic perception is notpossible it is probably replaced by a more analyticcomponent-based process of structural encoding Thisanalytic process may inhibit or at least delay the per-ception of the spatial relationships between face com-ponents and therefore may impede identificationIndeed a major source of evidence supporting theholistic view is that faces are more difficult to recognizewhen presented upside down (ie compromising theirfamiliar gestalt) than when presented in upright orien-tation (see reviews in Farah et al 1995 Rhodes et al1993 Valentine 1988)

Having to cope with different types of face input thevisual system must be able to encode the face config-uration directly from the visual display as well as byintegrating face components encoded individuallyThese two processes may concur triggered independ-ently by relevant physiognomic input and performed bydifferent neural circuits to form a complex structuralencoding mechanism in the extrastriate cortex Further-more the two systems may interact supporting orinhibiting each other Hence a regular face presentedin upright orientation may activate the holistic percep-tion process which in turn may inhibit superfluousanalysis of individual components By contrast corrupt-ing the face gestalt by presenting faces in an unusualorientation may trigger (and facilitate) face componentanalysis which should not depend on a particular con-figuration In such a case the integration of the compo-nents into a whole by the holistic processor will followtheir independent analysis assisted by the feature-detection system This hypothesis may explain for in-stance why within the context of a regular face individ-ual face parts are distinguished faster than in isolationbut if the configuration of the face is altered the partsare distinguished as fast as when presented in isolation(Tanaka amp Farah 1993 Sergent 1984)9 The ability toprocess face information both holistically and piecewise

946 Journal of Cognitive Neuroscience Volume 13 Number 7

was implemented in a neural model initially based onscalp recordings (Bentin et al 1996) and more recentlycorrected improved and elaborated on the basis intra-cranial recordings (McCarthy et al 1999 for a detaileddescription and a review of supportive functional MRIdata see McCarthy 1999) This model suggests that theneural mechanism dedicated to structural encoding offaces is a complex system composed of at least twoseparate subsystems One probably located in thelateral part of the middle fusiform gyrus is responsiblefor processing the configuration of the face as a wholeand provides the integrated face representation that isnecessary for face recognition and identification Theother possibly distributed over posterior regions in theIT and middle temporal (MT) gyri and in the OTS isresponsible for detecting physiognomic informationregardless of the spatial configuration of the items thatcarry it The fusiform gyrus is farther away from therecording sites at the scalp than where the temporalgyri are and the orientation of its cells forms anelectrical dipole source that is almost perpendicular tothe lateral aspects of the scalp Hence although bothsources contribute to the modulation of the face-spe-cific N170 the contribution of the posterior temporalface component processor should be considerably moreconspicuous

According to this hypothesis the opposite effects ofinversion on the N170 elicited by natural and schematicfaces might be explained assuming that the holisticencoding system in the fusiform gyrus indeed inhibitsthe component analysis in the lateral temporal lobe facesites Hence if the face stimulus can trigger both theholistic and the component analysis (as it is true fornatural faces) an experimental manipulation that inter-feres with holistic processing (like the face inversionprocedure in the present experiment) should reducethe amount of inhibition of component analysis andconsequently the amplitude of the N170 would beenhanced However if the rsquo rsquocomponentsrsquorsquo do not conveyphysiognomic information so that their analysis by aface-specific mechanism rsquo rsquodependsrsquo rsquo on the previousrecognition of the face gestalt (as it is true for schematicfaces) interfering with holistic perception cannot facil-itate the activity of the component analysis and there-fore the modulation of the N170 would only reflect theadditional difficulty to encode the face ie reducedamplitude and delayed latency

The observed effects of the different factors on theN170 peak latency are nicely aligned with the abovemodel Although stimulus inversion delayed the N170for both natural and schematic faces this delay was twiceas big (12 msec) for former stimuli than for the latter (6msec) This interaction suggests that the inversion had abigger effect on natural than on schematic faces Thismay be because regardless of whether they are upright orinverted the physiognomic value of schematic faces isbased on their gestalt Hence inversion does inhibit the

processing of the gestalt but does not induce a change inprocessing strategy On the other hand natural faces areanalyzed holistically if they are upright but inversionrequires a change in that strategy encouraging theanalysis of components The significantly faster responseto upright natural than to upright schematic faces inExperiment 2 suggests that the extraction of the facegestalt is easier if the face is more natural

There is ample evidence indicating that the righthemisphere is specialized in global or holistic processingof objects and particularly faces whereas the left hemi-sphere is specialized in part-based processing (egRobertson amp Delis 1986 for a review see Bradshow ampNettleton 1983 for specific reference to face perceptionsee eg Bruce 1998 Rhodes et al 1993 Corballis1988 Bradshow amp Sherlock 1982 Gilbert amp Bakan1973) Consequently interfering with holistic processingshould have had stronger effects in the right than in theleft hemisphere Our results confirmed this predictiononly partially The pronounced interhemispheric asym-metry with larger N170 amplitudes at the right than atthe left hemisphere sites for all stimulus types suggests apredominant holistic process for all face categoriesInterestingly this asymmetry was reduced (in Experi-ment 1) for sketches relative to the other stimulus typesThe relatively enhanced involvement of the left hemi-sphere in processing the sketches is in agreement withour suggestion that the face-specific structural encodingsystem was activated less efficiently by sketches than byother face types10 Hence the same factor that may haveconstrained the processing of schematic faces may alsohave constrained the processing of sketches but in theopposite direction Whereas schematic faces enhancedholistic processing by providing a clear face gestalt whileobscuring the physiognomic value of face componentsthe sketches reduced holistic processing by blurring orveiling the face gestalt Also in Experiment 2 the N170was larger over right than over left hemisphere sitesHowever because the interaction between the inversionand hemisphere effects in Experiment 2 was not modu-lated by stimulus type (ie the three-way interactionwas not significant) our interpretation of the interhemi-spheric asymmetry (particularly in Experiment 2) mustbe limited

In conclusion the present scalp-recorded N170 dataconverge with the previously published pattern of theintracranially recorded N200 data (for a comprehensivereview see Allison et al 1999 McCarthy et al 1999Puce et al 1999) supporting a multifunctional organ-ization of the structural encoding mechanism for facesApparently distinct neural circuits are responsible forthe encoding and initial analysis of face components andfor holistic perception of a full-face configuration Theformer is probably located in the lateral posterior tem-poral lobe whereas the latter is located in the middlefusiform gyrus Face perception usually entails predom-inantly holistic processing which explains the right

Sagiv and Bentin 947

hemisphere dominance in performance as well as inelectrophysiological and neuroimaging data Howeverboth mechanisms can be triggered independently byrelevant information

METHODS

Participants

Right-handed participants were recruited among volun-teer students from the Hebrew University They receivedeither experimental credit hours or payment for theirparticipation Thirty participants (13 males and 17 fe-males) were tested in the first experiment and 16participants (8 males and 8 females) were tested in thesecond experiment Participantsrsquo ages ranged from 19 to37 years (mean = 233 years) They all had normal orcorrected-to-normal vision

Stimuli

Stimuli were scanned using a desk scanner processedwith commercially available graphic software and con-

verted to 16 color grayscale 248 by 248 pixel imagesThere were eight stimulus categories four face types(photographs of natural faces paintings face sketchesand schematic faces) two nonsense patterns (formedby scrambling the natural faces and the schematicfaces) flowers and butterflies (Figure 4) Paintingsand sketches of faces were scanned from art booksPaintings were typically realistic Renaissance portraitsand face sketches did not portray details and usuallylacked shading and texture Schematic face parts weredrawn by hand scanned into the computer and latercombined into different schematic faces Face photo-graphs and paintings contained many shades of grayand the sketches and the schematic faces were basicallyblack on white All face stimuli were front views

Scrambled faces were created by scrambling thephases of the two-dimensional Fourier transform andthen smoothing the uneven edges created by this proc-ess The original and the scrambled stimuli were equallyluminous following this procedure Each stimulus cat-egory comprised 75 stimuli except for the butterflycategory which consisted of 30 stimuli In Experiment2 only natural faces schematic faces and flowers were

Face Stimuli

Photographs Painted portraits Sketches Schematic faces

Nonface Stimuli

Scrambledphotographs

Butterflies(targets) Flowers Scrambled

schematic faces

Figure 4 Example of the eight stimulus types used in the present study

948 Journal of Cognitive Neuroscience Volume 13 Number 7

used The faces were presented in both upright andupside down orientation Upright stimuli were the samein Experiments 1 and 2

All stimuli were presented for 350 msec with anaverage interstimulus interval (ISI) of 12 sec and viewedfrom a distance of 110 plusmn 10 cm subtending a viewingangle of approximately 68 Stimulus timing was con-trolled by the MEL software system (Psychology Soft-ware Tools Pittsburgh PA)

Task

Participants were instructed to keep a mental countof stimuli belonging to one category butterflies inthe first experiment and flowers in the secondHence all of the experiment-relevant categories werefrom the participantsrsquo point of view nontargets Thetargets were presented in approximately 5 (Experi-ment 1) and 6 (Experiment 2) of the trials Duringeach break and at the end of the experimentparticipants reported the number of targets countedA few participants who did not count correctly wereexcluded

EEG Recording and Data Analysis

Continuous EEG was recorded from 48 scalp locationsusing tin electrodes attached to elastic electrode caps(Electrode Caps International Ohio) A schematic illus-tration of the electrode locations and names is pre-sented in Figure 1 A ground electrode was located onthe forehead and a reference electrode was located onthe tip of the nose Two additional electrodes located atthe outer canthus and the infraorbital region of the righteye measured vertical and horizontal electrooculograms(EOGs)

The EEG was sampled at 250 Hz by a DAP2400 DataAcquisition Processor (Microstar Laboratories) ampli-fied 20000-fold using a battery-operated HU-5072 Iso-lated Bioelectric Amplifier (SA Instrumentation SanDiego) with a frequency bandpass of 01ndash30 Hz andstored on disk for off-line averaging Electrode impe-dance was below 5 k A digital filter of 08ndash16 Hz wasapplied before analysis

Codes synchronized to stimulus delivery were used toselectively average epochs associated with differentstimulus categories Epochs contaminated with saccadesor eye blinks were excluded using root-mean-squareEOG-values as criteria Typically epochs associated with55ndash70 (never less than 50) out of 75 stimuli in eachcategory were averaged

The mean and peak amplitudes and peak latencies ofspecific ERP components within bounding intervals werecomputed for each category for each participant Sig-nificance of the differences between these measures wastested using a repeated-measures ANOVA Where appro-priate the more conservative GreenhousendashGeisser ad-

justed df values were used ANOVAs were followed byunivariate F contrasts

Acknowledgments

This study was supported by grants from the USndashIsraelBinational Science Foundation and Israeli Foundations Trust-ees to S Bentin We thank Amir Katz for his help in preparingthe stimuli and running Experiment 1 as well as Leon DeouellMichael Fink Ayelet Landau Yulia Goland David Carmel GilRaviv Amir Raz and Dan Tzur for their support We are alsovery grateful to Jerry Levinson Alex List and Lynn Robertsonfor their comments on the manuscript

Reprint requests should be sent to Shlomo Bentin TheHebrew University of Jerusalem Mount Scopus JerusalemIsrael 91905 or via e-mail msbentinmscchujiacil

Notes

1 Present address University of California Berkeley2 There was no reliable difference between the ERP elicited

by scrambled natural faces and scrambled schematic facesTherefore these two categories were collapsed into one3 These were the sites at which the N170 was most

conspicuous4 So were in fact all face categories except for the sketches5 Note that this result also contradicts the view that a more

difficult encoding process elicit N170 of a higher amplitude6 This assumption was indeed supported by a current study

(Bentin Mecklinger Bosch Sagiv amp von Cramon 1999)7 Of course some faces have very distinctive components

which can be used as markers for recognition Such markersare frequently exaggerated and used as distinctive features incaricatures (Rhodes 1987)8 Distinctive features (such as unusual components birth-

marks etc) may be stored in semantic memory as part of apersonrsquo s face entry Such information may have visual qualitiesthat can be addressed directly during visual matching tofacilitate identification9 Note that this account for the rsquorsquoface superiority effectrsquo rsquo is

identical to a common explanation of the rsquorsquoword superiorityeffectrsquo rsquo in the identification of single letters Indeed ourassumption about the coexistence of rsquo rsquowhole-facersquo rsquo and rsquo rsquoface-partrsquo rsquo processors resonates to the rsquorsquodual routersquo rsquo model of visualword processing It is quite possible that identical principlesgovern all the content-specific perceptual modules in the visualsystem10 Recall that relative to other face types the reduction of theN170 amplitude elicited by sketches at right hemisphere siteswas greater then at left hemisphere sites

REFERENCES

Allison T Ginter H McCarthy G Nobre A C Puce ALuby M amp Spencer D D (1994) Face recognition in hu-man extrastriate cortex Journal of Neurophysiology 71821ndash825

Allison T McCarthy G Nobre A C Puce A amp Belger A(1994) Human extrastriate visual cortex and the perceptionof faces words numbers and colors Cerebral Cortex 5544ndash554

Allison T Puce A Spencer D D amp McCarthy G (1999)Electrophysiological studies of human face perception I

Sagiv and Bentin 949

between the amplitudes of N170 elicited by photo-graphs paintings and schematic faces The N170 ampli-tudes across face categories were largest at the mastoidsand in descending order IM12 P78 and PO78 Sig-nificant differences were found only between the mas-toids and the P78 and PO78 locations [F(129) = 50p lt 05 and F(129) = 48 p lt 05 respectively] andbetween the IM12 and the PO78 locations [F(129) =49 p lt 05]

All three first-order interactions were significant[F(9261) = 48 p lt 001 F(387) = 31 p lt 05F(387) = 63 p lt 01 for the Stimulus Type pound SiteStimulus Type pound Hemisphere and Site pound Hemisphererespectively] The second-order interaction between allthree factors was also significant [F(9261) = 34 p lt001] Post hoc examination of these interactions re-vealed the following pattern The difference betweenthe N170 elicited by sketches and that elicited by theother face types was larger at the right hemisphere sitesthan at the left hemisphere sites and slightly smaller atIM sites than at the other three sites The interactionbetween hemisphere and scalp location is probably aresult of the smaller (but clear) right hemisphere ad-vantage at scalp site IM12 [F(129) = 90 p lt 01]compared to other sites [F(129) = 153 p lt 001F(129) = 182 p lt 001 F(129) = 242 p lt 001 forPO78 mastoids and P78 respectively] The second-order interaction reflected (a) a smaller (Site pound Hemi-

sphere) interaction for the sketched faces category[F(387) = 23 vs F(387) gt 59 for each of the othercategories] (b) an insignificant (Stimulus Type pound Hemi-sphere) interaction at the IM scalp site [F(387) lt 100]versus a significant or close-to-significant effect for othersites [F(387) gt 24] and (c) a smaller (Stimulus Type poundSite) interaction in the right hemisphere [F(15435) =30] than in the left hemisphere [F(15435) = 55]

As evident in Table 1B the mean N170 latency wasshortest for schematic faces (1676 msec) followed byphotographs (1693 msec) paintings (1714 msec) andsketches (1743 msec) Scalp sites also differed on N170latency the site at which the shortest N170 latency wasrecorded was PO78 (1697 msec) followed by P78(1702 msec) IM12 (1704 msec) and RL-Mast (1723msec) ANOVA showed that both the stimulus type andthe site effects were significant [F(387) = 167 p lt001 F(387) = 50 p lt 005 respectively] Post hoccontrasts revealed that differences between the meanN170 latencies of all stimulus type pairs were highlysignificant [F(129) = 92 p lt 005] except for thedifference between face photograph and schematic facelatencies [F(129) = 30 p lt 1] In addition the N170latency at the mastoid scalp locations (1723 msec) wassignificantly delayed [F(129) = 173 p lt 001] com-pared with the three other locations (1701 msec)which did not significantly differ from each other[F(129) = 5 p = 6]

Table 1 (A) Mean Amplitudes (in mV) and (B) Mean Latencies (in msec) of the N170 Elicited by Natural Faces Painted PortraitsSketches Schematic Faces Flowers and Scrambled Faces at the Posterior Temporal Scalp Sites

Left Hemisphere Right Hemisphere

Left mastoid P7 PO7 IM1 IM2 PO8 P8 Right mastoid

A

Natural Faces ndash 72 ndash 62 ndash 60 ndash 76 ndash 89 ndash 93 ndash 91 ndash 97

Painted Portraits ndash 73 ndash 61 ndash 62 ndash 78 ndash 88 ndash 88 ndash 92 ndash 98

Sketches ndash 58 ndash 49 ndash 41 ndash 61 ndash 73 ndash 64 ndash 70 ndash 76

Schematic Faces ndash 68 ndash 63 ndash 60 ndash 70 ndash 83 ndash 88 ndash 96 ndash 94

Flowers ndash 29 ndash 32 ndash 23 ndash 28 ndash 30 ndash 29 ndash 33 ndash 34

Scrambled Faces ndash 19 ndash 17 ndash 07 ndash 13 ndash 15 ndash 07 ndash 16 ndash 22

B

Natural Faces 170 168 170 168 169 169 170 170

Painted Portraits 171 171 172 171 171 171 172 172

Sketches 177 171 171 174 175 173 176 178

Schematic Faces 170 165 165 168 167 167 169 169

Flowers 174 166 168 171 172 168 171 176

Scrambled Faces 177 167 166 173 175 166 168 178

Sagiv and Bentin 941

Hemispheric differences failed to reach significance[F(129) lt 10] A Site pound Hemisphere interaction washowever reliable [F(387) = 40 p lt 005] This inter-action is possibly the result of relatively late right hemi-sphere mean latency at P78 and PO78 scalp locationsThe interaction between stimulus type and site was alsosignificant [F(9261) = 35 p lt 001] reflecting a greatersite effect within the sketched face category The Stim-ulus Type pound Hemisphere interaction was not significant[F(387) = 14 p lt 25] The interaction between allthree factors did not reach significance [F(9261) = 17p = 1]

Discussion

Experiment 1 demonstrated that face-specific N170 com-ponents can be recorded with scalp electrodes not onlywhen photographs of natural faces are presented butalso when paintings sketches and even highly sche-matic drawings of faces are used as stimuli The ampli-tude of the N170 elicited by all face stimuli was at leasttwice as big as the amplitude of the negative potentialelicited by flowers and three times as big as the ampli-tude of the potential elicited by nonsense images duringthe same time The significant difference between flow-ers and scrambled faces could reflect either that theneural activity associated to the structural encoding ofany object is distinct from that elicited by visual patternsthat are not objects (McCarthy et al 1997) or that thereis specificity for other stimulus categories in the vicinityof the face area and that part of this specificity is pickedup by the far-field recording (Ishai Ungerleider MartinSchouten amp Haxby 1999) The present study was notdesigned to address this issue Therefore we will turnnow to the focus of this experiment that is the compar-ison between the different face types

The finding that the N170 elicited by schematic facesis not significantly smaller in amplitude is not delayedrelative to natural face photographs and has a similarscalp distribution suggests that common neural mech-anisms underlie the perception of both natural andschematic faces Assuming that these mechanisms arethe neural analogue to the hypothesized structuralencoder for faces these data also suggest that a face-specific visual mechanism is triggered whenever a stim-ulus contains sufficient information to generate theconcept of a face In other words schematically drawnfaces are categorized by an early visual mechanismrather than by a higher-level cognitive process Indeedit is possible that once triggered such a face perceptionmodule may take precedence over a more general visualprocessor A similar account for the suppression ofacoustic perception when phonetic features are de-tected (Bentin amp Mann 1990 Liberman amp Mattingly1989) led to the hypothesis that the phonetic modulemay preempt auditory information that is relevant tophonetic perception (Whalen amp Liberman 1987)

The reduced amplitude of the N170 elicited by detail-richer face sketches relative to that elicited by schematicfaces and its delayed peak latency might be interpretedprima facie as evidence against the above hypothesisHowever post hoc this result can be incorporated intoour theory Although graphically more detailed thesketches did not attempt to imitate a natural face (asthe paintings) or to emphasize those basic featuresapparently required for the formation of a face concept(as in the schematic faces) Indeed it is possible thatsome of the sketches were difficult to interpret ashuman faces without an elaborate analytic processHence the generation of a face concept on the basisof such sketches may have been somewhat more timeconsuming either because they were processed by gen-eral visual processors or because they activated thespecial purpose structural encoder later In additionunlike the schematic faces that although not identicalwere more uniform in shape4 the sketches were stylis-tically more diverse As a consequence some of thesketches may have introduced noise or a time jitter inthe average ERP which could explain both the reducedamplitude and the longer latency of the N170 associatedwith the sketched faces category5 All face categorieselicited a significantly larger N170 in the right hemi-sphere However the asymmetry was reduced forsketches and absent for nonface categories This findingis consistent with neuropsychological (eg De Renzi1997) neuroimaging (eg McCarthy et al 1997 DeRenzi Perani Carlesimo Silveri amp Fazio 1994 Haxbyet al 1993) and behavioral studies (eg Luh Rueckertamp Levy 1991) indicating a right hemisphere advantagefor face processing Furthermore it corroborates thehypothesis that some of the sketches were processed bya general rather than a face-specific visual processor orthat in general the sketches were less efficient ininitiating a face-specific process

The finding that schematic faces elicit a conspicuousN170 that does not significantly differ from the ERPselicited by natural faces is intriguing in light of theassumption that the N170 is associated primarily withthe processing of face components In previous studieswe have consistently found that eyes in isolation or eyescombined with other inner-face components elicit alarger N170 than that elicited by full faces (Bentinet al 1996 Bentin amp McCarthy unpublished) Further-more full faces elicit an N170 which is only slightlylarger than that elicited by face contours from which allinner components have been erased On the basis ofthese data Bentin et al (1996) suggested that the N170is modulated by two neural sources (both part of thestructural encoding mechanism) one possibly locatedin the posterior upper bank of the occipito-temporalsulcus (OTS) or in the inferior temporal (IT) gyrus istriggered by physiognomic information appearing in thevisual field the other probably located in the middlefusiform gyrus is responsible for holistic face processing

942 Journal of Cognitive Neuroscience Volume 13 Number 7

and integrates the face parts into a whole Assuming thatwithout the framing contour and the face-specific con-figuration inner components of schematic faces have nophysiognomic value6 the N170 elicited by schematicfaces should be modulated primarily by the fusiformsource which is responsible for the holistic processingof a face Consequently the N170 should be particularlysusceptible to manipulations affecting holistic process-ing of schematic faces An initial attempt to explore thispossibility is reported in Experiment 2 In addition thesimilar latency for the N170s elicited by natural faces andschematic faces would thus be in agreement with theprimacy of holistic processing for face stimuli the lateralcomponent primarily responsible for the N170 is trig-gered by the holistic component whether or not theface parts are recognizable as such in isolation ordefined by the global configuration

EXPERIMENT 2

The face inversion effectmdashthat is the difficulty to iden-tify familiar faces presented upside down (comparedwith other familiar objects subject to a similar manipu-lation)mdashhas been the focus of many studies for morethan two decades (reviewed by Valentine 1988) Thedifference between the effects of face inversion relativeto inversion of other objects is usually attributed to abasic difference between the processing strategies in-curred by faces versus objects According to this hypoth-esis recognition of faces relies on the formation of awhole-face representation from which relations betweenthe inner-face parts are derived that is holistic process-ing (Farah Tanaka amp Drain 1995 Tanaka amp Farah 1993versions of the holistic hypothesis are reviewed byMoscovitch Winocur amp Behrmann 1997) In contrastthe recognition of other perhaps less stereotypic ob-jects entails initial identification of basic componentswhich are then related one to the other to form anunequivocal visual percept (eg Biderman 1987 Marr1982 Marr amp Nishihara 1978) The face inversion effecthas become the hallmark of normal face recognition inhumans and has often been used to distinguish face-specific from general visual processes

Bentin et al (1996) compared the N170 elicited byupright and inverted faces and found that the N170elicited by inverted faces was slightly larger and peaked10 msec later than that elicited by upright faces only thedifference in latency was significant in that study Otherinvestigators however reported that the enhancementof the N170 amplitude by face inversion is significant aswell as the delay in latency (eg Rossion et al 19992000) A possible interpretation of the enhanced N170amplitude in response to inverted faces is that the scalp-recorded N170 is associated primarily with the percep-tion of face components a process that might not beimpeded by inversion This hypothesis conforms to thepresumed anatomical location of the putative compo-

nents detector (in lateral OTS and IT gyrus) and theorientation of these cell assemblies relative to the re-cording electrodes on the scalp In contrast to the N170the intracranially recorded N200 (on the ventral plane ofthe temporal lobe) is modulated primarily by the holis-tic face integration mechanism located in the posteriorlateral parts of the fusiform gyrus (eg Allison et al1999) Hence given the orientation of the electricaldipole fields in these two cell assemblies a manipulationthat inhibits holistic perception should enhance theN170 amplitude This hypothesis also conforms to thefinding that the N200 (recorded at the fusiform) elicitedby faces is rsquo rsquo largerrsquo rsquo than that elicited by isolated eyes(McCarthy et al 1999) whereas the N170 (recorded atthe scalp) elicited by faces is rsquo rsquosmallerrsquo rsquo than that elicitedby isolated eyes (Bentin et al 1996)

According to this model face perception is specificbecause the highly stereotypic organization of facespermits categorizing the visual stimulus before a fulldescription of its features is provided by the analyticmechanism Therefore most of the time faces activatethe holistic perception mechanism instantly Indeedonly such a model has the capacity to explain the resultsof Experiment 1 in which a visual stimulus that can beperceived as a face only if processed as a whole (ie ahighly schematic face) elicited a similar N170 to thatelicited by photographs of natural faces

Based on the above conceptualization we hypothe-size that if the detection of schematic faces is basedprimarily on the holistic face configuration compromis-ing their gestalt by presenting the stimuli upside downshould reduce the N170 In contrast turning naturalfaces upside down should enhance the amplitude of theN170 since the face component processing modulewould still be activated To test this hypothesis wepresented a new group of participants with naturaland schematic faces in upright and inverted orienta-tions

Results

Figure 3 presents the N170s elicited by photographs ofnatural faces and schematic faces presented upright orupside down Replicating the findings of Rossion et al(1999) the N170 elicited by inverted natural faces waslarger than that elicited by upright faces In contrast theN170 elicited by inverted schematic faces was reduced inamplitude relative to that elicited by upright schematicfaces Inversion delayed the peak latency of the N170 forschematic as well as for natural faces

The reliability of these observations was statisticallyanalyzed by a within-subject ANOVA using repeatedmeasures (see data in Table 2A)

The factors were stimulus type (natural faces sche-matic faces) stimulus orientation (upright inverted) site(P78 mastoids PO78 IM12) and hemisphere (leftright) As a main effect stimulus orientation was not

Sagiv and Bentin 943

significant [F(115) = 21 p = 16] However the inter-esting pattern of the face inversion effects was revealedby a significant interaction between the effects of orien-tation and stimulus type [F(115) = 248 p lt 001]followed by planed comparisons The planed compari-sons showed that whereas for natural photographs theN170 elicited by inverted faces (ndash 66 mV) was larger thatthat elicited by upright faces (ndash 55 mV) with schematicfaces inversion was associated with a reduction inthe N170 amplitude (ndash 38 and ndash 50 mV for invertedand upright presentation respectively t(15) = 2104p lt 055 and t(15) = 3000 p lt 01 for natural andschematic faces respectively) The main effect of stim-ulus type was significant [F(115) = 302 p lt 001] but itwas also qualified by its interaction with the effect oforientation Post hoc comparisons showed that whereasthe difference between the N170 elicited by invertednatural and schematic faces was significant [t(15) =4264 p lt 001] the difference between the N170elicited by natural and schematic faces presented uprightwas not [t(15) = 884 p = 39] The main effect of site

was also significant [F(345) = 68 p lt 01] revealingonly that the N170 amplitudes at P7 and P8 were smallerthan at all other sites [F(115) = 117 p lt 01] It isnoteworthy that as evidence by an insignificant interac-tion with stimulus type [F(345) = 13 p = 28] thiseffect was similar for schematic and natural faces On theother hand a significant main effect of hemisphere[F(115) = 65 p lt 05] was qualified by its interactionwith stimulus type [F(115) = 53 p lt 05] Post hocanalysis revealed that whereas for both stimulus typesthe amplitude of the N170 was larger over the rightcompared to the left hemisphere the interhemisphericdifference was significant for natural faces [t(15) = 335p lt 01] but not for schematic faces [t(15) = 1408p = 18] The effects of the Hemisphere pound Orientationinteraction was significant [F(115) = 47 p lt 5] Thisinteraction was caused by the fact that whereas acrossstimulus type the reduction and enhancement of theN170 amplitude by inversion canceled each other in theleft hemisphere in the right hemisphere the enhance-ment of N170 induced by the inversion of natural faces

Faces Inverted Faces Schematic Faces Inverted SchematicFaces

Time(msec) Time(msec)0 100 200 300 400

10

mV 0

0 100 200 300 400

10PO7 PO8

- 0 +

mV 0

Figure 3 The N170 potentials elicited at the left and right mastoids by upright and inverted natural faces and schematic faces and the scalpdistribution at this latency

944 Journal of Cognitive Neuroscience Volume 13 Number 7

weighted more than its reduction resulting from theinversion of schematic faces However the second-orderinteraction between orientation hemisphere and stim-ulus type was not significant [F(115) lt 10] Finally aseries of second-order interactions involving the effect ofsite suggested that all the effects were differently con-spicuous at different sites As revealed in Table 2 by andlarge it seems that whereas natural faces induced similarinversion effects across the analyzed sites the inversioneffect for schematic faces varied at different sites We hadno prior assumptions about the distribution of the effectswithin the posterior temporal scalp regions Further-more the spatial resolution of scalp-recorded ERPs istoo low to permit reliable interpretation of site effectswithin a relative small region unless the same distribu-tion is replicated therefore a detailed analysis of thisdistribution of effects was not attempted post hoc

ANOVA of the N170 peak latencies showed a signifi-cant main effect of stimulus orientation [F(115) =1615 p lt 001] which significantly interacted withstimulus type [F(115) = 62 p lt 025] Plannedcomparisons showed that for both natural and sche-matic faces inversion delayed the N170 significantly[t(15) = 10821 p lt 001 and t(15) = 3626 p lt 01respectively] However as evident in Table 2B thisdifference was larger for natural than for schematicfaces The latencies were similar over the left and righthemispheres [F(115) = 19 p = 185] The main effectof site was significant [F(345) = 47 p lt 01] butbecause the differences between sites were all smaller

than the sampling rate this effect was not investigatedfurther No other significant effects emerged in thisanalysis

Discussion

The results of the Experiment 2 validated our predic-tion that inversion should reduce the amplitude of theN170 elicited by schematic faces but have little effector enhance the N170 elicited by natural faces Thenatural face inversion effect observed in the presentstudy replicates previous results reported by Rossionet al (1999) These authors suggested that the en-hancement in the N170 amplitude reflects the en-hanced difficulty of the holistic processor to encodethe inverted face Indeed the delay in the N170 peaklatency elicited by inverted relative to upright faces iscongruent with such a view However if this were thecase we should have found similar effects of inversionon the N170 amplitude elicited by schematic and bynatural faces Furthermore the assumption that moredifficult face encoding induces an enhancement of theN170 amplitude is also at odds with fMRI data showingthat face inversion reduces the activity in the fusiformface areas relative to the activity elicited by uprightfaces (Kanwisher Tong amp Nakayama 1998) Indeedfinding opposite effects of inversion for schematic andnatural faces can more likely be accommodated by ourhypothesis that the larger N170 amplitude in responseto inverted natural faces reflects its stronger associa-

Table 2 (A) Amplitudes (in mV) and (B) Latencies (in msec) of the N170 Elicited by Upright and Inverted Natural and SchematicFaces at the Posterior Temporal Scalp Sites

Natural Faces Schematic Faces

Right Hemisphere Left Hemisphere Right Hemisphere Left Hemisphere

Upright Inverted Upright Inverted Upright Inverted Upright Inverted

A

Mastoid ndash 64 ndash 86 ndash 58 ndash 82 ndash 54 ndash 55 ndash 51 ndash 35

P87 ndash 48 ndash 68 ndash 45 ndash 54 ndash 41 ndash 30 ndash 32 ndash 33

PO87 ndash 68 ndash 88 ndash 48 ndash 50 ndash 56 ndash 56 ndash 50 ndash 23

IM21 ndash 53 ndash 81 ndash 49 ndash 73 ndash 50 ndash 29 ndash 65 ndash 39

Mean ndash 58 ndash 81 ndash 50 ndash 65 ndash 50 ndash 425 ndash 49 ndash 325

B

Mastoid 149 162 151 165 159 164 158 165

P87 151 163 154 166 158 164 162 166

PO87 151 160 152 164 151 161 159 166

IM21 150 162 147 161 161 164 154 162

Mean 150 162 151 164 157 163 158 165

Sagiv and Bentin 945

tion with a face component analyzer in the lateraltemporal gyri than with the holistic face processor inthe fusiform

Our model proposes that although schematic andnatural faces probably activate similar neural networksin the extrastriate visual pathway these mechanismsare not identical Natural faces probably activate twoperceptual modules one specialized to detect andprocess physiognomic features in the visual field andthe other specialized in the holistic processing of facesor forming a holistic face representation Unlike naturalfaces the components of schematic faces are notperceived as carrying physiognomic information outof the schematic face gestalt Therefore in contrastto natural faces schematic faces may trigger the holisticprocessor but they do not trigger directly the analysisof the components The latter mechanism may beactivated only subsequent to the formation of theholistic face representation or at the very least follow-ing this process rsquo rsquo in cascadersquo rsquo (McClelland 1979) Con-sequently adversely affecting the perception of the facegestalt (by inverting the stimuli) inhibited both holisticperception and the perception of individual compo-nents and this inhibition reduced the amplitude of theN170 Such a pattern is also in agreement with themyriad of studies supporting the primacy of holisticprocessing of face stimuli

GENERAL DISCUSSION

In the present study we investigated several character-istics of the extrastriate structural face-encoding mech-anisms by comparing the N170 elicited by faces differingin their naturalistic aspect and by comparing the faceinversion effects on the N170 elicited by photographs ofnatural faces and by schematically drawn faces To theextent that the N170 is indeed associated with structuralencoding we found that this process distinguishes wellbetween faces and nonfaces stimuli but not amongdifferent face types that render the concept of a faceclearly However we also found evidence that schematicfaces and natural faces are not processed identicallyWhereas inversion of natural faces enhanced the ampli-tude of the N170 inversion of schematic faces reducedits amplitude For both face types the latency of theN170 peak was significantly delayed by inversion Thispattern of results suggests that the face-specific struc-tural encoder can be triggered by a variety of stimuli ifthey include a canonical face configuration Howeverthe process of encoding the face information and form-ing a structural representation is probably differentwhen the physiognomic value of the stimuli dependsupon holistic configuration as opposed to when theindividual components can be associated with faces evenwhen presented outside the face context

Apart from the neurofunctional specificity issue anenduring question in face perception research is whether

the formation of a mental representation of the faceentails extraction of a holistic configuration directly fromthe visual input or whether it is feature-based that iswhether for faces as for other complex visual stimulistructural encoding is based on an initial analysis of parts(for a review of this debate see eg Farah 1990 Farahet al 1995) In principle because faces form a veryhomogeneous and stereotypical category they lendthemselves to holistic perception Holistic processing isadvantageous in face recognition because face compo-nents are fairly similar across individuals and thereforethe distinction among faces is usually based on theinterrelationships between the different face compo-nents and their overall configuration (eg Tanaka ampFarah 1993 Young Hellawell amp Hay 1987)7 Moreoverour memory for faces is usually based on nondecompos-able holistic images which at the interface betweenperception and memory are most efficiently addressedby integrated face representations8However if the well-learned gestalt is altered and holistic perception is notpossible it is probably replaced by a more analyticcomponent-based process of structural encoding Thisanalytic process may inhibit or at least delay the per-ception of the spatial relationships between face com-ponents and therefore may impede identificationIndeed a major source of evidence supporting theholistic view is that faces are more difficult to recognizewhen presented upside down (ie compromising theirfamiliar gestalt) than when presented in upright orien-tation (see reviews in Farah et al 1995 Rhodes et al1993 Valentine 1988)

Having to cope with different types of face input thevisual system must be able to encode the face config-uration directly from the visual display as well as byintegrating face components encoded individuallyThese two processes may concur triggered independ-ently by relevant physiognomic input and performed bydifferent neural circuits to form a complex structuralencoding mechanism in the extrastriate cortex Further-more the two systems may interact supporting orinhibiting each other Hence a regular face presentedin upright orientation may activate the holistic percep-tion process which in turn may inhibit superfluousanalysis of individual components By contrast corrupt-ing the face gestalt by presenting faces in an unusualorientation may trigger (and facilitate) face componentanalysis which should not depend on a particular con-figuration In such a case the integration of the compo-nents into a whole by the holistic processor will followtheir independent analysis assisted by the feature-detection system This hypothesis may explain for in-stance why within the context of a regular face individ-ual face parts are distinguished faster than in isolationbut if the configuration of the face is altered the partsare distinguished as fast as when presented in isolation(Tanaka amp Farah 1993 Sergent 1984)9 The ability toprocess face information both holistically and piecewise

946 Journal of Cognitive Neuroscience Volume 13 Number 7

was implemented in a neural model initially based onscalp recordings (Bentin et al 1996) and more recentlycorrected improved and elaborated on the basis intra-cranial recordings (McCarthy et al 1999 for a detaileddescription and a review of supportive functional MRIdata see McCarthy 1999) This model suggests that theneural mechanism dedicated to structural encoding offaces is a complex system composed of at least twoseparate subsystems One probably located in thelateral part of the middle fusiform gyrus is responsiblefor processing the configuration of the face as a wholeand provides the integrated face representation that isnecessary for face recognition and identification Theother possibly distributed over posterior regions in theIT and middle temporal (MT) gyri and in the OTS isresponsible for detecting physiognomic informationregardless of the spatial configuration of the items thatcarry it The fusiform gyrus is farther away from therecording sites at the scalp than where the temporalgyri are and the orientation of its cells forms anelectrical dipole source that is almost perpendicular tothe lateral aspects of the scalp Hence although bothsources contribute to the modulation of the face-spe-cific N170 the contribution of the posterior temporalface component processor should be considerably moreconspicuous

According to this hypothesis the opposite effects ofinversion on the N170 elicited by natural and schematicfaces might be explained assuming that the holisticencoding system in the fusiform gyrus indeed inhibitsthe component analysis in the lateral temporal lobe facesites Hence if the face stimulus can trigger both theholistic and the component analysis (as it is true fornatural faces) an experimental manipulation that inter-feres with holistic processing (like the face inversionprocedure in the present experiment) should reducethe amount of inhibition of component analysis andconsequently the amplitude of the N170 would beenhanced However if the rsquo rsquocomponentsrsquorsquo do not conveyphysiognomic information so that their analysis by aface-specific mechanism rsquo rsquodependsrsquo rsquo on the previousrecognition of the face gestalt (as it is true for schematicfaces) interfering with holistic perception cannot facil-itate the activity of the component analysis and there-fore the modulation of the N170 would only reflect theadditional difficulty to encode the face ie reducedamplitude and delayed latency

The observed effects of the different factors on theN170 peak latency are nicely aligned with the abovemodel Although stimulus inversion delayed the N170for both natural and schematic faces this delay was twiceas big (12 msec) for former stimuli than for the latter (6msec) This interaction suggests that the inversion had abigger effect on natural than on schematic faces Thismay be because regardless of whether they are upright orinverted the physiognomic value of schematic faces isbased on their gestalt Hence inversion does inhibit the

processing of the gestalt but does not induce a change inprocessing strategy On the other hand natural faces areanalyzed holistically if they are upright but inversionrequires a change in that strategy encouraging theanalysis of components The significantly faster responseto upright natural than to upright schematic faces inExperiment 2 suggests that the extraction of the facegestalt is easier if the face is more natural

There is ample evidence indicating that the righthemisphere is specialized in global or holistic processingof objects and particularly faces whereas the left hemi-sphere is specialized in part-based processing (egRobertson amp Delis 1986 for a review see Bradshow ampNettleton 1983 for specific reference to face perceptionsee eg Bruce 1998 Rhodes et al 1993 Corballis1988 Bradshow amp Sherlock 1982 Gilbert amp Bakan1973) Consequently interfering with holistic processingshould have had stronger effects in the right than in theleft hemisphere Our results confirmed this predictiononly partially The pronounced interhemispheric asym-metry with larger N170 amplitudes at the right than atthe left hemisphere sites for all stimulus types suggests apredominant holistic process for all face categoriesInterestingly this asymmetry was reduced (in Experi-ment 1) for sketches relative to the other stimulus typesThe relatively enhanced involvement of the left hemi-sphere in processing the sketches is in agreement withour suggestion that the face-specific structural encodingsystem was activated less efficiently by sketches than byother face types10 Hence the same factor that may haveconstrained the processing of schematic faces may alsohave constrained the processing of sketches but in theopposite direction Whereas schematic faces enhancedholistic processing by providing a clear face gestalt whileobscuring the physiognomic value of face componentsthe sketches reduced holistic processing by blurring orveiling the face gestalt Also in Experiment 2 the N170was larger over right than over left hemisphere sitesHowever because the interaction between the inversionand hemisphere effects in Experiment 2 was not modu-lated by stimulus type (ie the three-way interactionwas not significant) our interpretation of the interhemi-spheric asymmetry (particularly in Experiment 2) mustbe limited

In conclusion the present scalp-recorded N170 dataconverge with the previously published pattern of theintracranially recorded N200 data (for a comprehensivereview see Allison et al 1999 McCarthy et al 1999Puce et al 1999) supporting a multifunctional organ-ization of the structural encoding mechanism for facesApparently distinct neural circuits are responsible forthe encoding and initial analysis of face components andfor holistic perception of a full-face configuration Theformer is probably located in the lateral posterior tem-poral lobe whereas the latter is located in the middlefusiform gyrus Face perception usually entails predom-inantly holistic processing which explains the right

Sagiv and Bentin 947

hemisphere dominance in performance as well as inelectrophysiological and neuroimaging data Howeverboth mechanisms can be triggered independently byrelevant information

METHODS

Participants

Right-handed participants were recruited among volun-teer students from the Hebrew University They receivedeither experimental credit hours or payment for theirparticipation Thirty participants (13 males and 17 fe-males) were tested in the first experiment and 16participants (8 males and 8 females) were tested in thesecond experiment Participantsrsquo ages ranged from 19 to37 years (mean = 233 years) They all had normal orcorrected-to-normal vision

Stimuli

Stimuli were scanned using a desk scanner processedwith commercially available graphic software and con-

verted to 16 color grayscale 248 by 248 pixel imagesThere were eight stimulus categories four face types(photographs of natural faces paintings face sketchesand schematic faces) two nonsense patterns (formedby scrambling the natural faces and the schematicfaces) flowers and butterflies (Figure 4) Paintingsand sketches of faces were scanned from art booksPaintings were typically realistic Renaissance portraitsand face sketches did not portray details and usuallylacked shading and texture Schematic face parts weredrawn by hand scanned into the computer and latercombined into different schematic faces Face photo-graphs and paintings contained many shades of grayand the sketches and the schematic faces were basicallyblack on white All face stimuli were front views

Scrambled faces were created by scrambling thephases of the two-dimensional Fourier transform andthen smoothing the uneven edges created by this proc-ess The original and the scrambled stimuli were equallyluminous following this procedure Each stimulus cat-egory comprised 75 stimuli except for the butterflycategory which consisted of 30 stimuli In Experiment2 only natural faces schematic faces and flowers were

Face Stimuli

Photographs Painted portraits Sketches Schematic faces

Nonface Stimuli

Scrambledphotographs

Butterflies(targets) Flowers Scrambled

schematic faces

Figure 4 Example of the eight stimulus types used in the present study

948 Journal of Cognitive Neuroscience Volume 13 Number 7

used The faces were presented in both upright andupside down orientation Upright stimuli were the samein Experiments 1 and 2

All stimuli were presented for 350 msec with anaverage interstimulus interval (ISI) of 12 sec and viewedfrom a distance of 110 plusmn 10 cm subtending a viewingangle of approximately 68 Stimulus timing was con-trolled by the MEL software system (Psychology Soft-ware Tools Pittsburgh PA)

Task

Participants were instructed to keep a mental countof stimuli belonging to one category butterflies inthe first experiment and flowers in the secondHence all of the experiment-relevant categories werefrom the participantsrsquo point of view nontargets Thetargets were presented in approximately 5 (Experi-ment 1) and 6 (Experiment 2) of the trials Duringeach break and at the end of the experimentparticipants reported the number of targets countedA few participants who did not count correctly wereexcluded

EEG Recording and Data Analysis

Continuous EEG was recorded from 48 scalp locationsusing tin electrodes attached to elastic electrode caps(Electrode Caps International Ohio) A schematic illus-tration of the electrode locations and names is pre-sented in Figure 1 A ground electrode was located onthe forehead and a reference electrode was located onthe tip of the nose Two additional electrodes located atthe outer canthus and the infraorbital region of the righteye measured vertical and horizontal electrooculograms(EOGs)

The EEG was sampled at 250 Hz by a DAP2400 DataAcquisition Processor (Microstar Laboratories) ampli-fied 20000-fold using a battery-operated HU-5072 Iso-lated Bioelectric Amplifier (SA Instrumentation SanDiego) with a frequency bandpass of 01ndash30 Hz andstored on disk for off-line averaging Electrode impe-dance was below 5 k A digital filter of 08ndash16 Hz wasapplied before analysis

Codes synchronized to stimulus delivery were used toselectively average epochs associated with differentstimulus categories Epochs contaminated with saccadesor eye blinks were excluded using root-mean-squareEOG-values as criteria Typically epochs associated with55ndash70 (never less than 50) out of 75 stimuli in eachcategory were averaged

The mean and peak amplitudes and peak latencies ofspecific ERP components within bounding intervals werecomputed for each category for each participant Sig-nificance of the differences between these measures wastested using a repeated-measures ANOVA Where appro-priate the more conservative GreenhousendashGeisser ad-

justed df values were used ANOVAs were followed byunivariate F contrasts

Acknowledgments

This study was supported by grants from the USndashIsraelBinational Science Foundation and Israeli Foundations Trust-ees to S Bentin We thank Amir Katz for his help in preparingthe stimuli and running Experiment 1 as well as Leon DeouellMichael Fink Ayelet Landau Yulia Goland David Carmel GilRaviv Amir Raz and Dan Tzur for their support We are alsovery grateful to Jerry Levinson Alex List and Lynn Robertsonfor their comments on the manuscript

Reprint requests should be sent to Shlomo Bentin TheHebrew University of Jerusalem Mount Scopus JerusalemIsrael 91905 or via e-mail msbentinmscchujiacil

Notes

1 Present address University of California Berkeley2 There was no reliable difference between the ERP elicited

by scrambled natural faces and scrambled schematic facesTherefore these two categories were collapsed into one3 These were the sites at which the N170 was most

conspicuous4 So were in fact all face categories except for the sketches5 Note that this result also contradicts the view that a more

difficult encoding process elicit N170 of a higher amplitude6 This assumption was indeed supported by a current study

(Bentin Mecklinger Bosch Sagiv amp von Cramon 1999)7 Of course some faces have very distinctive components

which can be used as markers for recognition Such markersare frequently exaggerated and used as distinctive features incaricatures (Rhodes 1987)8 Distinctive features (such as unusual components birth-

marks etc) may be stored in semantic memory as part of apersonrsquo s face entry Such information may have visual qualitiesthat can be addressed directly during visual matching tofacilitate identification9 Note that this account for the rsquorsquoface superiority effectrsquo rsquo is

identical to a common explanation of the rsquorsquoword superiorityeffectrsquo rsquo in the identification of single letters Indeed ourassumption about the coexistence of rsquo rsquowhole-facersquo rsquo and rsquo rsquoface-partrsquo rsquo processors resonates to the rsquorsquodual routersquo rsquo model of visualword processing It is quite possible that identical principlesgovern all the content-specific perceptual modules in the visualsystem10 Recall that relative to other face types the reduction of theN170 amplitude elicited by sketches at right hemisphere siteswas greater then at left hemisphere sites

REFERENCES

Allison T Ginter H McCarthy G Nobre A C Puce ALuby M amp Spencer D D (1994) Face recognition in hu-man extrastriate cortex Journal of Neurophysiology 71821ndash825

Allison T McCarthy G Nobre A C Puce A amp Belger A(1994) Human extrastriate visual cortex and the perceptionof faces words numbers and colors Cerebral Cortex 5544ndash554

Allison T Puce A Spencer D D amp McCarthy G (1999)Electrophysiological studies of human face perception I

Sagiv and Bentin 949

Hemispheric differences failed to reach significance[F(129) lt 10] A Site pound Hemisphere interaction washowever reliable [F(387) = 40 p lt 005] This inter-action is possibly the result of relatively late right hemi-sphere mean latency at P78 and PO78 scalp locationsThe interaction between stimulus type and site was alsosignificant [F(9261) = 35 p lt 001] reflecting a greatersite effect within the sketched face category The Stim-ulus Type pound Hemisphere interaction was not significant[F(387) = 14 p lt 25] The interaction between allthree factors did not reach significance [F(9261) = 17p = 1]

Discussion

Experiment 1 demonstrated that face-specific N170 com-ponents can be recorded with scalp electrodes not onlywhen photographs of natural faces are presented butalso when paintings sketches and even highly sche-matic drawings of faces are used as stimuli The ampli-tude of the N170 elicited by all face stimuli was at leasttwice as big as the amplitude of the negative potentialelicited by flowers and three times as big as the ampli-tude of the potential elicited by nonsense images duringthe same time The significant difference between flow-ers and scrambled faces could reflect either that theneural activity associated to the structural encoding ofany object is distinct from that elicited by visual patternsthat are not objects (McCarthy et al 1997) or that thereis specificity for other stimulus categories in the vicinityof the face area and that part of this specificity is pickedup by the far-field recording (Ishai Ungerleider MartinSchouten amp Haxby 1999) The present study was notdesigned to address this issue Therefore we will turnnow to the focus of this experiment that is the compar-ison between the different face types

The finding that the N170 elicited by schematic facesis not significantly smaller in amplitude is not delayedrelative to natural face photographs and has a similarscalp distribution suggests that common neural mech-anisms underlie the perception of both natural andschematic faces Assuming that these mechanisms arethe neural analogue to the hypothesized structuralencoder for faces these data also suggest that a face-specific visual mechanism is triggered whenever a stim-ulus contains sufficient information to generate theconcept of a face In other words schematically drawnfaces are categorized by an early visual mechanismrather than by a higher-level cognitive process Indeedit is possible that once triggered such a face perceptionmodule may take precedence over a more general visualprocessor A similar account for the suppression ofacoustic perception when phonetic features are de-tected (Bentin amp Mann 1990 Liberman amp Mattingly1989) led to the hypothesis that the phonetic modulemay preempt auditory information that is relevant tophonetic perception (Whalen amp Liberman 1987)

The reduced amplitude of the N170 elicited by detail-richer face sketches relative to that elicited by schematicfaces and its delayed peak latency might be interpretedprima facie as evidence against the above hypothesisHowever post hoc this result can be incorporated intoour theory Although graphically more detailed thesketches did not attempt to imitate a natural face (asthe paintings) or to emphasize those basic featuresapparently required for the formation of a face concept(as in the schematic faces) Indeed it is possible thatsome of the sketches were difficult to interpret ashuman faces without an elaborate analytic processHence the generation of a face concept on the basisof such sketches may have been somewhat more timeconsuming either because they were processed by gen-eral visual processors or because they activated thespecial purpose structural encoder later In additionunlike the schematic faces that although not identicalwere more uniform in shape4 the sketches were stylis-tically more diverse As a consequence some of thesketches may have introduced noise or a time jitter inthe average ERP which could explain both the reducedamplitude and the longer latency of the N170 associatedwith the sketched faces category5 All face categorieselicited a significantly larger N170 in the right hemi-sphere However the asymmetry was reduced forsketches and absent for nonface categories This findingis consistent with neuropsychological (eg De Renzi1997) neuroimaging (eg McCarthy et al 1997 DeRenzi Perani Carlesimo Silveri amp Fazio 1994 Haxbyet al 1993) and behavioral studies (eg Luh Rueckertamp Levy 1991) indicating a right hemisphere advantagefor face processing Furthermore it corroborates thehypothesis that some of the sketches were processed bya general rather than a face-specific visual processor orthat in general the sketches were less efficient ininitiating a face-specific process

The finding that schematic faces elicit a conspicuousN170 that does not significantly differ from the ERPselicited by natural faces is intriguing in light of theassumption that the N170 is associated primarily withthe processing of face components In previous studieswe have consistently found that eyes in isolation or eyescombined with other inner-face components elicit alarger N170 than that elicited by full faces (Bentinet al 1996 Bentin amp McCarthy unpublished) Further-more full faces elicit an N170 which is only slightlylarger than that elicited by face contours from which allinner components have been erased On the basis ofthese data Bentin et al (1996) suggested that the N170is modulated by two neural sources (both part of thestructural encoding mechanism) one possibly locatedin the posterior upper bank of the occipito-temporalsulcus (OTS) or in the inferior temporal (IT) gyrus istriggered by physiognomic information appearing in thevisual field the other probably located in the middlefusiform gyrus is responsible for holistic face processing

942 Journal of Cognitive Neuroscience Volume 13 Number 7

and integrates the face parts into a whole Assuming thatwithout the framing contour and the face-specific con-figuration inner components of schematic faces have nophysiognomic value6 the N170 elicited by schematicfaces should be modulated primarily by the fusiformsource which is responsible for the holistic processingof a face Consequently the N170 should be particularlysusceptible to manipulations affecting holistic process-ing of schematic faces An initial attempt to explore thispossibility is reported in Experiment 2 In addition thesimilar latency for the N170s elicited by natural faces andschematic faces would thus be in agreement with theprimacy of holistic processing for face stimuli the lateralcomponent primarily responsible for the N170 is trig-gered by the holistic component whether or not theface parts are recognizable as such in isolation ordefined by the global configuration

EXPERIMENT 2

The face inversion effectmdashthat is the difficulty to iden-tify familiar faces presented upside down (comparedwith other familiar objects subject to a similar manipu-lation)mdashhas been the focus of many studies for morethan two decades (reviewed by Valentine 1988) Thedifference between the effects of face inversion relativeto inversion of other objects is usually attributed to abasic difference between the processing strategies in-curred by faces versus objects According to this hypoth-esis recognition of faces relies on the formation of awhole-face representation from which relations betweenthe inner-face parts are derived that is holistic process-ing (Farah Tanaka amp Drain 1995 Tanaka amp Farah 1993versions of the holistic hypothesis are reviewed byMoscovitch Winocur amp Behrmann 1997) In contrastthe recognition of other perhaps less stereotypic ob-jects entails initial identification of basic componentswhich are then related one to the other to form anunequivocal visual percept (eg Biderman 1987 Marr1982 Marr amp Nishihara 1978) The face inversion effecthas become the hallmark of normal face recognition inhumans and has often been used to distinguish face-specific from general visual processes

Bentin et al (1996) compared the N170 elicited byupright and inverted faces and found that the N170elicited by inverted faces was slightly larger and peaked10 msec later than that elicited by upright faces only thedifference in latency was significant in that study Otherinvestigators however reported that the enhancementof the N170 amplitude by face inversion is significant aswell as the delay in latency (eg Rossion et al 19992000) A possible interpretation of the enhanced N170amplitude in response to inverted faces is that the scalp-recorded N170 is associated primarily with the percep-tion of face components a process that might not beimpeded by inversion This hypothesis conforms to thepresumed anatomical location of the putative compo-

nents detector (in lateral OTS and IT gyrus) and theorientation of these cell assemblies relative to the re-cording electrodes on the scalp In contrast to the N170the intracranially recorded N200 (on the ventral plane ofthe temporal lobe) is modulated primarily by the holis-tic face integration mechanism located in the posteriorlateral parts of the fusiform gyrus (eg Allison et al1999) Hence given the orientation of the electricaldipole fields in these two cell assemblies a manipulationthat inhibits holistic perception should enhance theN170 amplitude This hypothesis also conforms to thefinding that the N200 (recorded at the fusiform) elicitedby faces is rsquo rsquo largerrsquo rsquo than that elicited by isolated eyes(McCarthy et al 1999) whereas the N170 (recorded atthe scalp) elicited by faces is rsquo rsquosmallerrsquo rsquo than that elicitedby isolated eyes (Bentin et al 1996)

According to this model face perception is specificbecause the highly stereotypic organization of facespermits categorizing the visual stimulus before a fulldescription of its features is provided by the analyticmechanism Therefore most of the time faces activatethe holistic perception mechanism instantly Indeedonly such a model has the capacity to explain the resultsof Experiment 1 in which a visual stimulus that can beperceived as a face only if processed as a whole (ie ahighly schematic face) elicited a similar N170 to thatelicited by photographs of natural faces

Based on the above conceptualization we hypothe-size that if the detection of schematic faces is basedprimarily on the holistic face configuration compromis-ing their gestalt by presenting the stimuli upside downshould reduce the N170 In contrast turning naturalfaces upside down should enhance the amplitude of theN170 since the face component processing modulewould still be activated To test this hypothesis wepresented a new group of participants with naturaland schematic faces in upright and inverted orienta-tions

Results

Figure 3 presents the N170s elicited by photographs ofnatural faces and schematic faces presented upright orupside down Replicating the findings of Rossion et al(1999) the N170 elicited by inverted natural faces waslarger than that elicited by upright faces In contrast theN170 elicited by inverted schematic faces was reduced inamplitude relative to that elicited by upright schematicfaces Inversion delayed the peak latency of the N170 forschematic as well as for natural faces

The reliability of these observations was statisticallyanalyzed by a within-subject ANOVA using repeatedmeasures (see data in Table 2A)

The factors were stimulus type (natural faces sche-matic faces) stimulus orientation (upright inverted) site(P78 mastoids PO78 IM12) and hemisphere (leftright) As a main effect stimulus orientation was not

Sagiv and Bentin 943

significant [F(115) = 21 p = 16] However the inter-esting pattern of the face inversion effects was revealedby a significant interaction between the effects of orien-tation and stimulus type [F(115) = 248 p lt 001]followed by planed comparisons The planed compari-sons showed that whereas for natural photographs theN170 elicited by inverted faces (ndash 66 mV) was larger thatthat elicited by upright faces (ndash 55 mV) with schematicfaces inversion was associated with a reduction inthe N170 amplitude (ndash 38 and ndash 50 mV for invertedand upright presentation respectively t(15) = 2104p lt 055 and t(15) = 3000 p lt 01 for natural andschematic faces respectively) The main effect of stim-ulus type was significant [F(115) = 302 p lt 001] but itwas also qualified by its interaction with the effect oforientation Post hoc comparisons showed that whereasthe difference between the N170 elicited by invertednatural and schematic faces was significant [t(15) =4264 p lt 001] the difference between the N170elicited by natural and schematic faces presented uprightwas not [t(15) = 884 p = 39] The main effect of site

was also significant [F(345) = 68 p lt 01] revealingonly that the N170 amplitudes at P7 and P8 were smallerthan at all other sites [F(115) = 117 p lt 01] It isnoteworthy that as evidence by an insignificant interac-tion with stimulus type [F(345) = 13 p = 28] thiseffect was similar for schematic and natural faces On theother hand a significant main effect of hemisphere[F(115) = 65 p lt 05] was qualified by its interactionwith stimulus type [F(115) = 53 p lt 05] Post hocanalysis revealed that whereas for both stimulus typesthe amplitude of the N170 was larger over the rightcompared to the left hemisphere the interhemisphericdifference was significant for natural faces [t(15) = 335p lt 01] but not for schematic faces [t(15) = 1408p = 18] The effects of the Hemisphere pound Orientationinteraction was significant [F(115) = 47 p lt 5] Thisinteraction was caused by the fact that whereas acrossstimulus type the reduction and enhancement of theN170 amplitude by inversion canceled each other in theleft hemisphere in the right hemisphere the enhance-ment of N170 induced by the inversion of natural faces

Faces Inverted Faces Schematic Faces Inverted SchematicFaces

Time(msec) Time(msec)0 100 200 300 400

10

mV 0

0 100 200 300 400

10PO7 PO8

- 0 +

mV 0

Figure 3 The N170 potentials elicited at the left and right mastoids by upright and inverted natural faces and schematic faces and the scalpdistribution at this latency

944 Journal of Cognitive Neuroscience Volume 13 Number 7

weighted more than its reduction resulting from theinversion of schematic faces However the second-orderinteraction between orientation hemisphere and stim-ulus type was not significant [F(115) lt 10] Finally aseries of second-order interactions involving the effect ofsite suggested that all the effects were differently con-spicuous at different sites As revealed in Table 2 by andlarge it seems that whereas natural faces induced similarinversion effects across the analyzed sites the inversioneffect for schematic faces varied at different sites We hadno prior assumptions about the distribution of the effectswithin the posterior temporal scalp regions Further-more the spatial resolution of scalp-recorded ERPs istoo low to permit reliable interpretation of site effectswithin a relative small region unless the same distribu-tion is replicated therefore a detailed analysis of thisdistribution of effects was not attempted post hoc

ANOVA of the N170 peak latencies showed a signifi-cant main effect of stimulus orientation [F(115) =1615 p lt 001] which significantly interacted withstimulus type [F(115) = 62 p lt 025] Plannedcomparisons showed that for both natural and sche-matic faces inversion delayed the N170 significantly[t(15) = 10821 p lt 001 and t(15) = 3626 p lt 01respectively] However as evident in Table 2B thisdifference was larger for natural than for schematicfaces The latencies were similar over the left and righthemispheres [F(115) = 19 p = 185] The main effectof site was significant [F(345) = 47 p lt 01] butbecause the differences between sites were all smaller

than the sampling rate this effect was not investigatedfurther No other significant effects emerged in thisanalysis

Discussion

The results of the Experiment 2 validated our predic-tion that inversion should reduce the amplitude of theN170 elicited by schematic faces but have little effector enhance the N170 elicited by natural faces Thenatural face inversion effect observed in the presentstudy replicates previous results reported by Rossionet al (1999) These authors suggested that the en-hancement in the N170 amplitude reflects the en-hanced difficulty of the holistic processor to encodethe inverted face Indeed the delay in the N170 peaklatency elicited by inverted relative to upright faces iscongruent with such a view However if this were thecase we should have found similar effects of inversionon the N170 amplitude elicited by schematic and bynatural faces Furthermore the assumption that moredifficult face encoding induces an enhancement of theN170 amplitude is also at odds with fMRI data showingthat face inversion reduces the activity in the fusiformface areas relative to the activity elicited by uprightfaces (Kanwisher Tong amp Nakayama 1998) Indeedfinding opposite effects of inversion for schematic andnatural faces can more likely be accommodated by ourhypothesis that the larger N170 amplitude in responseto inverted natural faces reflects its stronger associa-

Table 2 (A) Amplitudes (in mV) and (B) Latencies (in msec) of the N170 Elicited by Upright and Inverted Natural and SchematicFaces at the Posterior Temporal Scalp Sites

Natural Faces Schematic Faces

Right Hemisphere Left Hemisphere Right Hemisphere Left Hemisphere

Upright Inverted Upright Inverted Upright Inverted Upright Inverted

A

Mastoid ndash 64 ndash 86 ndash 58 ndash 82 ndash 54 ndash 55 ndash 51 ndash 35

P87 ndash 48 ndash 68 ndash 45 ndash 54 ndash 41 ndash 30 ndash 32 ndash 33

PO87 ndash 68 ndash 88 ndash 48 ndash 50 ndash 56 ndash 56 ndash 50 ndash 23

IM21 ndash 53 ndash 81 ndash 49 ndash 73 ndash 50 ndash 29 ndash 65 ndash 39

Mean ndash 58 ndash 81 ndash 50 ndash 65 ndash 50 ndash 425 ndash 49 ndash 325

B

Mastoid 149 162 151 165 159 164 158 165

P87 151 163 154 166 158 164 162 166

PO87 151 160 152 164 151 161 159 166

IM21 150 162 147 161 161 164 154 162

Mean 150 162 151 164 157 163 158 165

Sagiv and Bentin 945

tion with a face component analyzer in the lateraltemporal gyri than with the holistic face processor inthe fusiform

Our model proposes that although schematic andnatural faces probably activate similar neural networksin the extrastriate visual pathway these mechanismsare not identical Natural faces probably activate twoperceptual modules one specialized to detect andprocess physiognomic features in the visual field andthe other specialized in the holistic processing of facesor forming a holistic face representation Unlike naturalfaces the components of schematic faces are notperceived as carrying physiognomic information outof the schematic face gestalt Therefore in contrastto natural faces schematic faces may trigger the holisticprocessor but they do not trigger directly the analysisof the components The latter mechanism may beactivated only subsequent to the formation of theholistic face representation or at the very least follow-ing this process rsquo rsquo in cascadersquo rsquo (McClelland 1979) Con-sequently adversely affecting the perception of the facegestalt (by inverting the stimuli) inhibited both holisticperception and the perception of individual compo-nents and this inhibition reduced the amplitude of theN170 Such a pattern is also in agreement with themyriad of studies supporting the primacy of holisticprocessing of face stimuli

GENERAL DISCUSSION

In the present study we investigated several character-istics of the extrastriate structural face-encoding mech-anisms by comparing the N170 elicited by faces differingin their naturalistic aspect and by comparing the faceinversion effects on the N170 elicited by photographs ofnatural faces and by schematically drawn faces To theextent that the N170 is indeed associated with structuralencoding we found that this process distinguishes wellbetween faces and nonfaces stimuli but not amongdifferent face types that render the concept of a faceclearly However we also found evidence that schematicfaces and natural faces are not processed identicallyWhereas inversion of natural faces enhanced the ampli-tude of the N170 inversion of schematic faces reducedits amplitude For both face types the latency of theN170 peak was significantly delayed by inversion Thispattern of results suggests that the face-specific struc-tural encoder can be triggered by a variety of stimuli ifthey include a canonical face configuration Howeverthe process of encoding the face information and form-ing a structural representation is probably differentwhen the physiognomic value of the stimuli dependsupon holistic configuration as opposed to when theindividual components can be associated with faces evenwhen presented outside the face context

Apart from the neurofunctional specificity issue anenduring question in face perception research is whether

the formation of a mental representation of the faceentails extraction of a holistic configuration directly fromthe visual input or whether it is feature-based that iswhether for faces as for other complex visual stimulistructural encoding is based on an initial analysis of parts(for a review of this debate see eg Farah 1990 Farahet al 1995) In principle because faces form a veryhomogeneous and stereotypical category they lendthemselves to holistic perception Holistic processing isadvantageous in face recognition because face compo-nents are fairly similar across individuals and thereforethe distinction among faces is usually based on theinterrelationships between the different face compo-nents and their overall configuration (eg Tanaka ampFarah 1993 Young Hellawell amp Hay 1987)7 Moreoverour memory for faces is usually based on nondecompos-able holistic images which at the interface betweenperception and memory are most efficiently addressedby integrated face representations8However if the well-learned gestalt is altered and holistic perception is notpossible it is probably replaced by a more analyticcomponent-based process of structural encoding Thisanalytic process may inhibit or at least delay the per-ception of the spatial relationships between face com-ponents and therefore may impede identificationIndeed a major source of evidence supporting theholistic view is that faces are more difficult to recognizewhen presented upside down (ie compromising theirfamiliar gestalt) than when presented in upright orien-tation (see reviews in Farah et al 1995 Rhodes et al1993 Valentine 1988)

Having to cope with different types of face input thevisual system must be able to encode the face config-uration directly from the visual display as well as byintegrating face components encoded individuallyThese two processes may concur triggered independ-ently by relevant physiognomic input and performed bydifferent neural circuits to form a complex structuralencoding mechanism in the extrastriate cortex Further-more the two systems may interact supporting orinhibiting each other Hence a regular face presentedin upright orientation may activate the holistic percep-tion process which in turn may inhibit superfluousanalysis of individual components By contrast corrupt-ing the face gestalt by presenting faces in an unusualorientation may trigger (and facilitate) face componentanalysis which should not depend on a particular con-figuration In such a case the integration of the compo-nents into a whole by the holistic processor will followtheir independent analysis assisted by the feature-detection system This hypothesis may explain for in-stance why within the context of a regular face individ-ual face parts are distinguished faster than in isolationbut if the configuration of the face is altered the partsare distinguished as fast as when presented in isolation(Tanaka amp Farah 1993 Sergent 1984)9 The ability toprocess face information both holistically and piecewise

946 Journal of Cognitive Neuroscience Volume 13 Number 7

was implemented in a neural model initially based onscalp recordings (Bentin et al 1996) and more recentlycorrected improved and elaborated on the basis intra-cranial recordings (McCarthy et al 1999 for a detaileddescription and a review of supportive functional MRIdata see McCarthy 1999) This model suggests that theneural mechanism dedicated to structural encoding offaces is a complex system composed of at least twoseparate subsystems One probably located in thelateral part of the middle fusiform gyrus is responsiblefor processing the configuration of the face as a wholeand provides the integrated face representation that isnecessary for face recognition and identification Theother possibly distributed over posterior regions in theIT and middle temporal (MT) gyri and in the OTS isresponsible for detecting physiognomic informationregardless of the spatial configuration of the items thatcarry it The fusiform gyrus is farther away from therecording sites at the scalp than where the temporalgyri are and the orientation of its cells forms anelectrical dipole source that is almost perpendicular tothe lateral aspects of the scalp Hence although bothsources contribute to the modulation of the face-spe-cific N170 the contribution of the posterior temporalface component processor should be considerably moreconspicuous

According to this hypothesis the opposite effects ofinversion on the N170 elicited by natural and schematicfaces might be explained assuming that the holisticencoding system in the fusiform gyrus indeed inhibitsthe component analysis in the lateral temporal lobe facesites Hence if the face stimulus can trigger both theholistic and the component analysis (as it is true fornatural faces) an experimental manipulation that inter-feres with holistic processing (like the face inversionprocedure in the present experiment) should reducethe amount of inhibition of component analysis andconsequently the amplitude of the N170 would beenhanced However if the rsquo rsquocomponentsrsquorsquo do not conveyphysiognomic information so that their analysis by aface-specific mechanism rsquo rsquodependsrsquo rsquo on the previousrecognition of the face gestalt (as it is true for schematicfaces) interfering with holistic perception cannot facil-itate the activity of the component analysis and there-fore the modulation of the N170 would only reflect theadditional difficulty to encode the face ie reducedamplitude and delayed latency

The observed effects of the different factors on theN170 peak latency are nicely aligned with the abovemodel Although stimulus inversion delayed the N170for both natural and schematic faces this delay was twiceas big (12 msec) for former stimuli than for the latter (6msec) This interaction suggests that the inversion had abigger effect on natural than on schematic faces Thismay be because regardless of whether they are upright orinverted the physiognomic value of schematic faces isbased on their gestalt Hence inversion does inhibit the

processing of the gestalt but does not induce a change inprocessing strategy On the other hand natural faces areanalyzed holistically if they are upright but inversionrequires a change in that strategy encouraging theanalysis of components The significantly faster responseto upright natural than to upright schematic faces inExperiment 2 suggests that the extraction of the facegestalt is easier if the face is more natural

There is ample evidence indicating that the righthemisphere is specialized in global or holistic processingof objects and particularly faces whereas the left hemi-sphere is specialized in part-based processing (egRobertson amp Delis 1986 for a review see Bradshow ampNettleton 1983 for specific reference to face perceptionsee eg Bruce 1998 Rhodes et al 1993 Corballis1988 Bradshow amp Sherlock 1982 Gilbert amp Bakan1973) Consequently interfering with holistic processingshould have had stronger effects in the right than in theleft hemisphere Our results confirmed this predictiononly partially The pronounced interhemispheric asym-metry with larger N170 amplitudes at the right than atthe left hemisphere sites for all stimulus types suggests apredominant holistic process for all face categoriesInterestingly this asymmetry was reduced (in Experi-ment 1) for sketches relative to the other stimulus typesThe relatively enhanced involvement of the left hemi-sphere in processing the sketches is in agreement withour suggestion that the face-specific structural encodingsystem was activated less efficiently by sketches than byother face types10 Hence the same factor that may haveconstrained the processing of schematic faces may alsohave constrained the processing of sketches but in theopposite direction Whereas schematic faces enhancedholistic processing by providing a clear face gestalt whileobscuring the physiognomic value of face componentsthe sketches reduced holistic processing by blurring orveiling the face gestalt Also in Experiment 2 the N170was larger over right than over left hemisphere sitesHowever because the interaction between the inversionand hemisphere effects in Experiment 2 was not modu-lated by stimulus type (ie the three-way interactionwas not significant) our interpretation of the interhemi-spheric asymmetry (particularly in Experiment 2) mustbe limited

In conclusion the present scalp-recorded N170 dataconverge with the previously published pattern of theintracranially recorded N200 data (for a comprehensivereview see Allison et al 1999 McCarthy et al 1999Puce et al 1999) supporting a multifunctional organ-ization of the structural encoding mechanism for facesApparently distinct neural circuits are responsible forthe encoding and initial analysis of face components andfor holistic perception of a full-face configuration Theformer is probably located in the lateral posterior tem-poral lobe whereas the latter is located in the middlefusiform gyrus Face perception usually entails predom-inantly holistic processing which explains the right

Sagiv and Bentin 947

hemisphere dominance in performance as well as inelectrophysiological and neuroimaging data Howeverboth mechanisms can be triggered independently byrelevant information

METHODS

Participants

Right-handed participants were recruited among volun-teer students from the Hebrew University They receivedeither experimental credit hours or payment for theirparticipation Thirty participants (13 males and 17 fe-males) were tested in the first experiment and 16participants (8 males and 8 females) were tested in thesecond experiment Participantsrsquo ages ranged from 19 to37 years (mean = 233 years) They all had normal orcorrected-to-normal vision

Stimuli

Stimuli were scanned using a desk scanner processedwith commercially available graphic software and con-

verted to 16 color grayscale 248 by 248 pixel imagesThere were eight stimulus categories four face types(photographs of natural faces paintings face sketchesand schematic faces) two nonsense patterns (formedby scrambling the natural faces and the schematicfaces) flowers and butterflies (Figure 4) Paintingsand sketches of faces were scanned from art booksPaintings were typically realistic Renaissance portraitsand face sketches did not portray details and usuallylacked shading and texture Schematic face parts weredrawn by hand scanned into the computer and latercombined into different schematic faces Face photo-graphs and paintings contained many shades of grayand the sketches and the schematic faces were basicallyblack on white All face stimuli were front views

Scrambled faces were created by scrambling thephases of the two-dimensional Fourier transform andthen smoothing the uneven edges created by this proc-ess The original and the scrambled stimuli were equallyluminous following this procedure Each stimulus cat-egory comprised 75 stimuli except for the butterflycategory which consisted of 30 stimuli In Experiment2 only natural faces schematic faces and flowers were

Face Stimuli

Photographs Painted portraits Sketches Schematic faces

Nonface Stimuli

Scrambledphotographs

Butterflies(targets) Flowers Scrambled

schematic faces

Figure 4 Example of the eight stimulus types used in the present study

948 Journal of Cognitive Neuroscience Volume 13 Number 7

used The faces were presented in both upright andupside down orientation Upright stimuli were the samein Experiments 1 and 2

All stimuli were presented for 350 msec with anaverage interstimulus interval (ISI) of 12 sec and viewedfrom a distance of 110 plusmn 10 cm subtending a viewingangle of approximately 68 Stimulus timing was con-trolled by the MEL software system (Psychology Soft-ware Tools Pittsburgh PA)

Task

Participants were instructed to keep a mental countof stimuli belonging to one category butterflies inthe first experiment and flowers in the secondHence all of the experiment-relevant categories werefrom the participantsrsquo point of view nontargets Thetargets were presented in approximately 5 (Experi-ment 1) and 6 (Experiment 2) of the trials Duringeach break and at the end of the experimentparticipants reported the number of targets countedA few participants who did not count correctly wereexcluded

EEG Recording and Data Analysis

Continuous EEG was recorded from 48 scalp locationsusing tin electrodes attached to elastic electrode caps(Electrode Caps International Ohio) A schematic illus-tration of the electrode locations and names is pre-sented in Figure 1 A ground electrode was located onthe forehead and a reference electrode was located onthe tip of the nose Two additional electrodes located atthe outer canthus and the infraorbital region of the righteye measured vertical and horizontal electrooculograms(EOGs)

The EEG was sampled at 250 Hz by a DAP2400 DataAcquisition Processor (Microstar Laboratories) ampli-fied 20000-fold using a battery-operated HU-5072 Iso-lated Bioelectric Amplifier (SA Instrumentation SanDiego) with a frequency bandpass of 01ndash30 Hz andstored on disk for off-line averaging Electrode impe-dance was below 5 k A digital filter of 08ndash16 Hz wasapplied before analysis

Codes synchronized to stimulus delivery were used toselectively average epochs associated with differentstimulus categories Epochs contaminated with saccadesor eye blinks were excluded using root-mean-squareEOG-values as criteria Typically epochs associated with55ndash70 (never less than 50) out of 75 stimuli in eachcategory were averaged

The mean and peak amplitudes and peak latencies ofspecific ERP components within bounding intervals werecomputed for each category for each participant Sig-nificance of the differences between these measures wastested using a repeated-measures ANOVA Where appro-priate the more conservative GreenhousendashGeisser ad-

justed df values were used ANOVAs were followed byunivariate F contrasts

Acknowledgments

This study was supported by grants from the USndashIsraelBinational Science Foundation and Israeli Foundations Trust-ees to S Bentin We thank Amir Katz for his help in preparingthe stimuli and running Experiment 1 as well as Leon DeouellMichael Fink Ayelet Landau Yulia Goland David Carmel GilRaviv Amir Raz and Dan Tzur for their support We are alsovery grateful to Jerry Levinson Alex List and Lynn Robertsonfor their comments on the manuscript

Reprint requests should be sent to Shlomo Bentin TheHebrew University of Jerusalem Mount Scopus JerusalemIsrael 91905 or via e-mail msbentinmscchujiacil

Notes

1 Present address University of California Berkeley2 There was no reliable difference between the ERP elicited

by scrambled natural faces and scrambled schematic facesTherefore these two categories were collapsed into one3 These were the sites at which the N170 was most

conspicuous4 So were in fact all face categories except for the sketches5 Note that this result also contradicts the view that a more

difficult encoding process elicit N170 of a higher amplitude6 This assumption was indeed supported by a current study

(Bentin Mecklinger Bosch Sagiv amp von Cramon 1999)7 Of course some faces have very distinctive components

which can be used as markers for recognition Such markersare frequently exaggerated and used as distinctive features incaricatures (Rhodes 1987)8 Distinctive features (such as unusual components birth-

marks etc) may be stored in semantic memory as part of apersonrsquo s face entry Such information may have visual qualitiesthat can be addressed directly during visual matching tofacilitate identification9 Note that this account for the rsquorsquoface superiority effectrsquo rsquo is

identical to a common explanation of the rsquorsquoword superiorityeffectrsquo rsquo in the identification of single letters Indeed ourassumption about the coexistence of rsquo rsquowhole-facersquo rsquo and rsquo rsquoface-partrsquo rsquo processors resonates to the rsquorsquodual routersquo rsquo model of visualword processing It is quite possible that identical principlesgovern all the content-specific perceptual modules in the visualsystem10 Recall that relative to other face types the reduction of theN170 amplitude elicited by sketches at right hemisphere siteswas greater then at left hemisphere sites

REFERENCES

Allison T Ginter H McCarthy G Nobre A C Puce ALuby M amp Spencer D D (1994) Face recognition in hu-man extrastriate cortex Journal of Neurophysiology 71821ndash825

Allison T McCarthy G Nobre A C Puce A amp Belger A(1994) Human extrastriate visual cortex and the perceptionof faces words numbers and colors Cerebral Cortex 5544ndash554

Allison T Puce A Spencer D D amp McCarthy G (1999)Electrophysiological studies of human face perception I

Sagiv and Bentin 949

and integrates the face parts into a whole Assuming thatwithout the framing contour and the face-specific con-figuration inner components of schematic faces have nophysiognomic value6 the N170 elicited by schematicfaces should be modulated primarily by the fusiformsource which is responsible for the holistic processingof a face Consequently the N170 should be particularlysusceptible to manipulations affecting holistic process-ing of schematic faces An initial attempt to explore thispossibility is reported in Experiment 2 In addition thesimilar latency for the N170s elicited by natural faces andschematic faces would thus be in agreement with theprimacy of holistic processing for face stimuli the lateralcomponent primarily responsible for the N170 is trig-gered by the holistic component whether or not theface parts are recognizable as such in isolation ordefined by the global configuration

EXPERIMENT 2

The face inversion effectmdashthat is the difficulty to iden-tify familiar faces presented upside down (comparedwith other familiar objects subject to a similar manipu-lation)mdashhas been the focus of many studies for morethan two decades (reviewed by Valentine 1988) Thedifference between the effects of face inversion relativeto inversion of other objects is usually attributed to abasic difference between the processing strategies in-curred by faces versus objects According to this hypoth-esis recognition of faces relies on the formation of awhole-face representation from which relations betweenthe inner-face parts are derived that is holistic process-ing (Farah Tanaka amp Drain 1995 Tanaka amp Farah 1993versions of the holistic hypothesis are reviewed byMoscovitch Winocur amp Behrmann 1997) In contrastthe recognition of other perhaps less stereotypic ob-jects entails initial identification of basic componentswhich are then related one to the other to form anunequivocal visual percept (eg Biderman 1987 Marr1982 Marr amp Nishihara 1978) The face inversion effecthas become the hallmark of normal face recognition inhumans and has often been used to distinguish face-specific from general visual processes

Bentin et al (1996) compared the N170 elicited byupright and inverted faces and found that the N170elicited by inverted faces was slightly larger and peaked10 msec later than that elicited by upright faces only thedifference in latency was significant in that study Otherinvestigators however reported that the enhancementof the N170 amplitude by face inversion is significant aswell as the delay in latency (eg Rossion et al 19992000) A possible interpretation of the enhanced N170amplitude in response to inverted faces is that the scalp-recorded N170 is associated primarily with the percep-tion of face components a process that might not beimpeded by inversion This hypothesis conforms to thepresumed anatomical location of the putative compo-

nents detector (in lateral OTS and IT gyrus) and theorientation of these cell assemblies relative to the re-cording electrodes on the scalp In contrast to the N170the intracranially recorded N200 (on the ventral plane ofthe temporal lobe) is modulated primarily by the holis-tic face integration mechanism located in the posteriorlateral parts of the fusiform gyrus (eg Allison et al1999) Hence given the orientation of the electricaldipole fields in these two cell assemblies a manipulationthat inhibits holistic perception should enhance theN170 amplitude This hypothesis also conforms to thefinding that the N200 (recorded at the fusiform) elicitedby faces is rsquo rsquo largerrsquo rsquo than that elicited by isolated eyes(McCarthy et al 1999) whereas the N170 (recorded atthe scalp) elicited by faces is rsquo rsquosmallerrsquo rsquo than that elicitedby isolated eyes (Bentin et al 1996)

According to this model face perception is specificbecause the highly stereotypic organization of facespermits categorizing the visual stimulus before a fulldescription of its features is provided by the analyticmechanism Therefore most of the time faces activatethe holistic perception mechanism instantly Indeedonly such a model has the capacity to explain the resultsof Experiment 1 in which a visual stimulus that can beperceived as a face only if processed as a whole (ie ahighly schematic face) elicited a similar N170 to thatelicited by photographs of natural faces

Based on the above conceptualization we hypothe-size that if the detection of schematic faces is basedprimarily on the holistic face configuration compromis-ing their gestalt by presenting the stimuli upside downshould reduce the N170 In contrast turning naturalfaces upside down should enhance the amplitude of theN170 since the face component processing modulewould still be activated To test this hypothesis wepresented a new group of participants with naturaland schematic faces in upright and inverted orienta-tions

Results

Figure 3 presents the N170s elicited by photographs ofnatural faces and schematic faces presented upright orupside down Replicating the findings of Rossion et al(1999) the N170 elicited by inverted natural faces waslarger than that elicited by upright faces In contrast theN170 elicited by inverted schematic faces was reduced inamplitude relative to that elicited by upright schematicfaces Inversion delayed the peak latency of the N170 forschematic as well as for natural faces

The reliability of these observations was statisticallyanalyzed by a within-subject ANOVA using repeatedmeasures (see data in Table 2A)

The factors were stimulus type (natural faces sche-matic faces) stimulus orientation (upright inverted) site(P78 mastoids PO78 IM12) and hemisphere (leftright) As a main effect stimulus orientation was not

Sagiv and Bentin 943

significant [F(115) = 21 p = 16] However the inter-esting pattern of the face inversion effects was revealedby a significant interaction between the effects of orien-tation and stimulus type [F(115) = 248 p lt 001]followed by planed comparisons The planed compari-sons showed that whereas for natural photographs theN170 elicited by inverted faces (ndash 66 mV) was larger thatthat elicited by upright faces (ndash 55 mV) with schematicfaces inversion was associated with a reduction inthe N170 amplitude (ndash 38 and ndash 50 mV for invertedand upright presentation respectively t(15) = 2104p lt 055 and t(15) = 3000 p lt 01 for natural andschematic faces respectively) The main effect of stim-ulus type was significant [F(115) = 302 p lt 001] but itwas also qualified by its interaction with the effect oforientation Post hoc comparisons showed that whereasthe difference between the N170 elicited by invertednatural and schematic faces was significant [t(15) =4264 p lt 001] the difference between the N170elicited by natural and schematic faces presented uprightwas not [t(15) = 884 p = 39] The main effect of site

was also significant [F(345) = 68 p lt 01] revealingonly that the N170 amplitudes at P7 and P8 were smallerthan at all other sites [F(115) = 117 p lt 01] It isnoteworthy that as evidence by an insignificant interac-tion with stimulus type [F(345) = 13 p = 28] thiseffect was similar for schematic and natural faces On theother hand a significant main effect of hemisphere[F(115) = 65 p lt 05] was qualified by its interactionwith stimulus type [F(115) = 53 p lt 05] Post hocanalysis revealed that whereas for both stimulus typesthe amplitude of the N170 was larger over the rightcompared to the left hemisphere the interhemisphericdifference was significant for natural faces [t(15) = 335p lt 01] but not for schematic faces [t(15) = 1408p = 18] The effects of the Hemisphere pound Orientationinteraction was significant [F(115) = 47 p lt 5] Thisinteraction was caused by the fact that whereas acrossstimulus type the reduction and enhancement of theN170 amplitude by inversion canceled each other in theleft hemisphere in the right hemisphere the enhance-ment of N170 induced by the inversion of natural faces

Faces Inverted Faces Schematic Faces Inverted SchematicFaces

Time(msec) Time(msec)0 100 200 300 400

10

mV 0

0 100 200 300 400

10PO7 PO8

- 0 +

mV 0

Figure 3 The N170 potentials elicited at the left and right mastoids by upright and inverted natural faces and schematic faces and the scalpdistribution at this latency

944 Journal of Cognitive Neuroscience Volume 13 Number 7

weighted more than its reduction resulting from theinversion of schematic faces However the second-orderinteraction between orientation hemisphere and stim-ulus type was not significant [F(115) lt 10] Finally aseries of second-order interactions involving the effect ofsite suggested that all the effects were differently con-spicuous at different sites As revealed in Table 2 by andlarge it seems that whereas natural faces induced similarinversion effects across the analyzed sites the inversioneffect for schematic faces varied at different sites We hadno prior assumptions about the distribution of the effectswithin the posterior temporal scalp regions Further-more the spatial resolution of scalp-recorded ERPs istoo low to permit reliable interpretation of site effectswithin a relative small region unless the same distribu-tion is replicated therefore a detailed analysis of thisdistribution of effects was not attempted post hoc

ANOVA of the N170 peak latencies showed a signifi-cant main effect of stimulus orientation [F(115) =1615 p lt 001] which significantly interacted withstimulus type [F(115) = 62 p lt 025] Plannedcomparisons showed that for both natural and sche-matic faces inversion delayed the N170 significantly[t(15) = 10821 p lt 001 and t(15) = 3626 p lt 01respectively] However as evident in Table 2B thisdifference was larger for natural than for schematicfaces The latencies were similar over the left and righthemispheres [F(115) = 19 p = 185] The main effectof site was significant [F(345) = 47 p lt 01] butbecause the differences between sites were all smaller

than the sampling rate this effect was not investigatedfurther No other significant effects emerged in thisanalysis

Discussion

The results of the Experiment 2 validated our predic-tion that inversion should reduce the amplitude of theN170 elicited by schematic faces but have little effector enhance the N170 elicited by natural faces Thenatural face inversion effect observed in the presentstudy replicates previous results reported by Rossionet al (1999) These authors suggested that the en-hancement in the N170 amplitude reflects the en-hanced difficulty of the holistic processor to encodethe inverted face Indeed the delay in the N170 peaklatency elicited by inverted relative to upright faces iscongruent with such a view However if this were thecase we should have found similar effects of inversionon the N170 amplitude elicited by schematic and bynatural faces Furthermore the assumption that moredifficult face encoding induces an enhancement of theN170 amplitude is also at odds with fMRI data showingthat face inversion reduces the activity in the fusiformface areas relative to the activity elicited by uprightfaces (Kanwisher Tong amp Nakayama 1998) Indeedfinding opposite effects of inversion for schematic andnatural faces can more likely be accommodated by ourhypothesis that the larger N170 amplitude in responseto inverted natural faces reflects its stronger associa-

Table 2 (A) Amplitudes (in mV) and (B) Latencies (in msec) of the N170 Elicited by Upright and Inverted Natural and SchematicFaces at the Posterior Temporal Scalp Sites

Natural Faces Schematic Faces

Right Hemisphere Left Hemisphere Right Hemisphere Left Hemisphere

Upright Inverted Upright Inverted Upright Inverted Upright Inverted

A

Mastoid ndash 64 ndash 86 ndash 58 ndash 82 ndash 54 ndash 55 ndash 51 ndash 35

P87 ndash 48 ndash 68 ndash 45 ndash 54 ndash 41 ndash 30 ndash 32 ndash 33

PO87 ndash 68 ndash 88 ndash 48 ndash 50 ndash 56 ndash 56 ndash 50 ndash 23

IM21 ndash 53 ndash 81 ndash 49 ndash 73 ndash 50 ndash 29 ndash 65 ndash 39

Mean ndash 58 ndash 81 ndash 50 ndash 65 ndash 50 ndash 425 ndash 49 ndash 325

B

Mastoid 149 162 151 165 159 164 158 165

P87 151 163 154 166 158 164 162 166

PO87 151 160 152 164 151 161 159 166

IM21 150 162 147 161 161 164 154 162

Mean 150 162 151 164 157 163 158 165

Sagiv and Bentin 945

tion with a face component analyzer in the lateraltemporal gyri than with the holistic face processor inthe fusiform

Our model proposes that although schematic andnatural faces probably activate similar neural networksin the extrastriate visual pathway these mechanismsare not identical Natural faces probably activate twoperceptual modules one specialized to detect andprocess physiognomic features in the visual field andthe other specialized in the holistic processing of facesor forming a holistic face representation Unlike naturalfaces the components of schematic faces are notperceived as carrying physiognomic information outof the schematic face gestalt Therefore in contrastto natural faces schematic faces may trigger the holisticprocessor but they do not trigger directly the analysisof the components The latter mechanism may beactivated only subsequent to the formation of theholistic face representation or at the very least follow-ing this process rsquo rsquo in cascadersquo rsquo (McClelland 1979) Con-sequently adversely affecting the perception of the facegestalt (by inverting the stimuli) inhibited both holisticperception and the perception of individual compo-nents and this inhibition reduced the amplitude of theN170 Such a pattern is also in agreement with themyriad of studies supporting the primacy of holisticprocessing of face stimuli

GENERAL DISCUSSION

In the present study we investigated several character-istics of the extrastriate structural face-encoding mech-anisms by comparing the N170 elicited by faces differingin their naturalistic aspect and by comparing the faceinversion effects on the N170 elicited by photographs ofnatural faces and by schematically drawn faces To theextent that the N170 is indeed associated with structuralencoding we found that this process distinguishes wellbetween faces and nonfaces stimuli but not amongdifferent face types that render the concept of a faceclearly However we also found evidence that schematicfaces and natural faces are not processed identicallyWhereas inversion of natural faces enhanced the ampli-tude of the N170 inversion of schematic faces reducedits amplitude For both face types the latency of theN170 peak was significantly delayed by inversion Thispattern of results suggests that the face-specific struc-tural encoder can be triggered by a variety of stimuli ifthey include a canonical face configuration Howeverthe process of encoding the face information and form-ing a structural representation is probably differentwhen the physiognomic value of the stimuli dependsupon holistic configuration as opposed to when theindividual components can be associated with faces evenwhen presented outside the face context

Apart from the neurofunctional specificity issue anenduring question in face perception research is whether

the formation of a mental representation of the faceentails extraction of a holistic configuration directly fromthe visual input or whether it is feature-based that iswhether for faces as for other complex visual stimulistructural encoding is based on an initial analysis of parts(for a review of this debate see eg Farah 1990 Farahet al 1995) In principle because faces form a veryhomogeneous and stereotypical category they lendthemselves to holistic perception Holistic processing isadvantageous in face recognition because face compo-nents are fairly similar across individuals and thereforethe distinction among faces is usually based on theinterrelationships between the different face compo-nents and their overall configuration (eg Tanaka ampFarah 1993 Young Hellawell amp Hay 1987)7 Moreoverour memory for faces is usually based on nondecompos-able holistic images which at the interface betweenperception and memory are most efficiently addressedby integrated face representations8However if the well-learned gestalt is altered and holistic perception is notpossible it is probably replaced by a more analyticcomponent-based process of structural encoding Thisanalytic process may inhibit or at least delay the per-ception of the spatial relationships between face com-ponents and therefore may impede identificationIndeed a major source of evidence supporting theholistic view is that faces are more difficult to recognizewhen presented upside down (ie compromising theirfamiliar gestalt) than when presented in upright orien-tation (see reviews in Farah et al 1995 Rhodes et al1993 Valentine 1988)

Having to cope with different types of face input thevisual system must be able to encode the face config-uration directly from the visual display as well as byintegrating face components encoded individuallyThese two processes may concur triggered independ-ently by relevant physiognomic input and performed bydifferent neural circuits to form a complex structuralencoding mechanism in the extrastriate cortex Further-more the two systems may interact supporting orinhibiting each other Hence a regular face presentedin upright orientation may activate the holistic percep-tion process which in turn may inhibit superfluousanalysis of individual components By contrast corrupt-ing the face gestalt by presenting faces in an unusualorientation may trigger (and facilitate) face componentanalysis which should not depend on a particular con-figuration In such a case the integration of the compo-nents into a whole by the holistic processor will followtheir independent analysis assisted by the feature-detection system This hypothesis may explain for in-stance why within the context of a regular face individ-ual face parts are distinguished faster than in isolationbut if the configuration of the face is altered the partsare distinguished as fast as when presented in isolation(Tanaka amp Farah 1993 Sergent 1984)9 The ability toprocess face information both holistically and piecewise

946 Journal of Cognitive Neuroscience Volume 13 Number 7

was implemented in a neural model initially based onscalp recordings (Bentin et al 1996) and more recentlycorrected improved and elaborated on the basis intra-cranial recordings (McCarthy et al 1999 for a detaileddescription and a review of supportive functional MRIdata see McCarthy 1999) This model suggests that theneural mechanism dedicated to structural encoding offaces is a complex system composed of at least twoseparate subsystems One probably located in thelateral part of the middle fusiform gyrus is responsiblefor processing the configuration of the face as a wholeand provides the integrated face representation that isnecessary for face recognition and identification Theother possibly distributed over posterior regions in theIT and middle temporal (MT) gyri and in the OTS isresponsible for detecting physiognomic informationregardless of the spatial configuration of the items thatcarry it The fusiform gyrus is farther away from therecording sites at the scalp than where the temporalgyri are and the orientation of its cells forms anelectrical dipole source that is almost perpendicular tothe lateral aspects of the scalp Hence although bothsources contribute to the modulation of the face-spe-cific N170 the contribution of the posterior temporalface component processor should be considerably moreconspicuous

According to this hypothesis the opposite effects ofinversion on the N170 elicited by natural and schematicfaces might be explained assuming that the holisticencoding system in the fusiform gyrus indeed inhibitsthe component analysis in the lateral temporal lobe facesites Hence if the face stimulus can trigger both theholistic and the component analysis (as it is true fornatural faces) an experimental manipulation that inter-feres with holistic processing (like the face inversionprocedure in the present experiment) should reducethe amount of inhibition of component analysis andconsequently the amplitude of the N170 would beenhanced However if the rsquo rsquocomponentsrsquorsquo do not conveyphysiognomic information so that their analysis by aface-specific mechanism rsquo rsquodependsrsquo rsquo on the previousrecognition of the face gestalt (as it is true for schematicfaces) interfering with holistic perception cannot facil-itate the activity of the component analysis and there-fore the modulation of the N170 would only reflect theadditional difficulty to encode the face ie reducedamplitude and delayed latency

The observed effects of the different factors on theN170 peak latency are nicely aligned with the abovemodel Although stimulus inversion delayed the N170for both natural and schematic faces this delay was twiceas big (12 msec) for former stimuli than for the latter (6msec) This interaction suggests that the inversion had abigger effect on natural than on schematic faces Thismay be because regardless of whether they are upright orinverted the physiognomic value of schematic faces isbased on their gestalt Hence inversion does inhibit the

processing of the gestalt but does not induce a change inprocessing strategy On the other hand natural faces areanalyzed holistically if they are upright but inversionrequires a change in that strategy encouraging theanalysis of components The significantly faster responseto upright natural than to upright schematic faces inExperiment 2 suggests that the extraction of the facegestalt is easier if the face is more natural

There is ample evidence indicating that the righthemisphere is specialized in global or holistic processingof objects and particularly faces whereas the left hemi-sphere is specialized in part-based processing (egRobertson amp Delis 1986 for a review see Bradshow ampNettleton 1983 for specific reference to face perceptionsee eg Bruce 1998 Rhodes et al 1993 Corballis1988 Bradshow amp Sherlock 1982 Gilbert amp Bakan1973) Consequently interfering with holistic processingshould have had stronger effects in the right than in theleft hemisphere Our results confirmed this predictiononly partially The pronounced interhemispheric asym-metry with larger N170 amplitudes at the right than atthe left hemisphere sites for all stimulus types suggests apredominant holistic process for all face categoriesInterestingly this asymmetry was reduced (in Experi-ment 1) for sketches relative to the other stimulus typesThe relatively enhanced involvement of the left hemi-sphere in processing the sketches is in agreement withour suggestion that the face-specific structural encodingsystem was activated less efficiently by sketches than byother face types10 Hence the same factor that may haveconstrained the processing of schematic faces may alsohave constrained the processing of sketches but in theopposite direction Whereas schematic faces enhancedholistic processing by providing a clear face gestalt whileobscuring the physiognomic value of face componentsthe sketches reduced holistic processing by blurring orveiling the face gestalt Also in Experiment 2 the N170was larger over right than over left hemisphere sitesHowever because the interaction between the inversionand hemisphere effects in Experiment 2 was not modu-lated by stimulus type (ie the three-way interactionwas not significant) our interpretation of the interhemi-spheric asymmetry (particularly in Experiment 2) mustbe limited

In conclusion the present scalp-recorded N170 dataconverge with the previously published pattern of theintracranially recorded N200 data (for a comprehensivereview see Allison et al 1999 McCarthy et al 1999Puce et al 1999) supporting a multifunctional organ-ization of the structural encoding mechanism for facesApparently distinct neural circuits are responsible forthe encoding and initial analysis of face components andfor holistic perception of a full-face configuration Theformer is probably located in the lateral posterior tem-poral lobe whereas the latter is located in the middlefusiform gyrus Face perception usually entails predom-inantly holistic processing which explains the right

Sagiv and Bentin 947

hemisphere dominance in performance as well as inelectrophysiological and neuroimaging data Howeverboth mechanisms can be triggered independently byrelevant information

METHODS

Participants

Right-handed participants were recruited among volun-teer students from the Hebrew University They receivedeither experimental credit hours or payment for theirparticipation Thirty participants (13 males and 17 fe-males) were tested in the first experiment and 16participants (8 males and 8 females) were tested in thesecond experiment Participantsrsquo ages ranged from 19 to37 years (mean = 233 years) They all had normal orcorrected-to-normal vision

Stimuli

Stimuli were scanned using a desk scanner processedwith commercially available graphic software and con-

verted to 16 color grayscale 248 by 248 pixel imagesThere were eight stimulus categories four face types(photographs of natural faces paintings face sketchesand schematic faces) two nonsense patterns (formedby scrambling the natural faces and the schematicfaces) flowers and butterflies (Figure 4) Paintingsand sketches of faces were scanned from art booksPaintings were typically realistic Renaissance portraitsand face sketches did not portray details and usuallylacked shading and texture Schematic face parts weredrawn by hand scanned into the computer and latercombined into different schematic faces Face photo-graphs and paintings contained many shades of grayand the sketches and the schematic faces were basicallyblack on white All face stimuli were front views

Scrambled faces were created by scrambling thephases of the two-dimensional Fourier transform andthen smoothing the uneven edges created by this proc-ess The original and the scrambled stimuli were equallyluminous following this procedure Each stimulus cat-egory comprised 75 stimuli except for the butterflycategory which consisted of 30 stimuli In Experiment2 only natural faces schematic faces and flowers were

Face Stimuli

Photographs Painted portraits Sketches Schematic faces

Nonface Stimuli

Scrambledphotographs

Butterflies(targets) Flowers Scrambled

schematic faces

Figure 4 Example of the eight stimulus types used in the present study

948 Journal of Cognitive Neuroscience Volume 13 Number 7

used The faces were presented in both upright andupside down orientation Upright stimuli were the samein Experiments 1 and 2

All stimuli were presented for 350 msec with anaverage interstimulus interval (ISI) of 12 sec and viewedfrom a distance of 110 plusmn 10 cm subtending a viewingangle of approximately 68 Stimulus timing was con-trolled by the MEL software system (Psychology Soft-ware Tools Pittsburgh PA)

Task

Participants were instructed to keep a mental countof stimuli belonging to one category butterflies inthe first experiment and flowers in the secondHence all of the experiment-relevant categories werefrom the participantsrsquo point of view nontargets Thetargets were presented in approximately 5 (Experi-ment 1) and 6 (Experiment 2) of the trials Duringeach break and at the end of the experimentparticipants reported the number of targets countedA few participants who did not count correctly wereexcluded

EEG Recording and Data Analysis

Continuous EEG was recorded from 48 scalp locationsusing tin electrodes attached to elastic electrode caps(Electrode Caps International Ohio) A schematic illus-tration of the electrode locations and names is pre-sented in Figure 1 A ground electrode was located onthe forehead and a reference electrode was located onthe tip of the nose Two additional electrodes located atthe outer canthus and the infraorbital region of the righteye measured vertical and horizontal electrooculograms(EOGs)

The EEG was sampled at 250 Hz by a DAP2400 DataAcquisition Processor (Microstar Laboratories) ampli-fied 20000-fold using a battery-operated HU-5072 Iso-lated Bioelectric Amplifier (SA Instrumentation SanDiego) with a frequency bandpass of 01ndash30 Hz andstored on disk for off-line averaging Electrode impe-dance was below 5 k A digital filter of 08ndash16 Hz wasapplied before analysis

Codes synchronized to stimulus delivery were used toselectively average epochs associated with differentstimulus categories Epochs contaminated with saccadesor eye blinks were excluded using root-mean-squareEOG-values as criteria Typically epochs associated with55ndash70 (never less than 50) out of 75 stimuli in eachcategory were averaged

The mean and peak amplitudes and peak latencies ofspecific ERP components within bounding intervals werecomputed for each category for each participant Sig-nificance of the differences between these measures wastested using a repeated-measures ANOVA Where appro-priate the more conservative GreenhousendashGeisser ad-

justed df values were used ANOVAs were followed byunivariate F contrasts

Acknowledgments

This study was supported by grants from the USndashIsraelBinational Science Foundation and Israeli Foundations Trust-ees to S Bentin We thank Amir Katz for his help in preparingthe stimuli and running Experiment 1 as well as Leon DeouellMichael Fink Ayelet Landau Yulia Goland David Carmel GilRaviv Amir Raz and Dan Tzur for their support We are alsovery grateful to Jerry Levinson Alex List and Lynn Robertsonfor their comments on the manuscript

Reprint requests should be sent to Shlomo Bentin TheHebrew University of Jerusalem Mount Scopus JerusalemIsrael 91905 or via e-mail msbentinmscchujiacil

Notes

1 Present address University of California Berkeley2 There was no reliable difference between the ERP elicited

by scrambled natural faces and scrambled schematic facesTherefore these two categories were collapsed into one3 These were the sites at which the N170 was most

conspicuous4 So were in fact all face categories except for the sketches5 Note that this result also contradicts the view that a more

difficult encoding process elicit N170 of a higher amplitude6 This assumption was indeed supported by a current study

(Bentin Mecklinger Bosch Sagiv amp von Cramon 1999)7 Of course some faces have very distinctive components

which can be used as markers for recognition Such markersare frequently exaggerated and used as distinctive features incaricatures (Rhodes 1987)8 Distinctive features (such as unusual components birth-

marks etc) may be stored in semantic memory as part of apersonrsquo s face entry Such information may have visual qualitiesthat can be addressed directly during visual matching tofacilitate identification9 Note that this account for the rsquorsquoface superiority effectrsquo rsquo is

identical to a common explanation of the rsquorsquoword superiorityeffectrsquo rsquo in the identification of single letters Indeed ourassumption about the coexistence of rsquo rsquowhole-facersquo rsquo and rsquo rsquoface-partrsquo rsquo processors resonates to the rsquorsquodual routersquo rsquo model of visualword processing It is quite possible that identical principlesgovern all the content-specific perceptual modules in the visualsystem10 Recall that relative to other face types the reduction of theN170 amplitude elicited by sketches at right hemisphere siteswas greater then at left hemisphere sites

REFERENCES

Allison T Ginter H McCarthy G Nobre A C Puce ALuby M amp Spencer D D (1994) Face recognition in hu-man extrastriate cortex Journal of Neurophysiology 71821ndash825

Allison T McCarthy G Nobre A C Puce A amp Belger A(1994) Human extrastriate visual cortex and the perceptionof faces words numbers and colors Cerebral Cortex 5544ndash554

Allison T Puce A Spencer D D amp McCarthy G (1999)Electrophysiological studies of human face perception I

Sagiv and Bentin 949

significant [F(115) = 21 p = 16] However the inter-esting pattern of the face inversion effects was revealedby a significant interaction between the effects of orien-tation and stimulus type [F(115) = 248 p lt 001]followed by planed comparisons The planed compari-sons showed that whereas for natural photographs theN170 elicited by inverted faces (ndash 66 mV) was larger thatthat elicited by upright faces (ndash 55 mV) with schematicfaces inversion was associated with a reduction inthe N170 amplitude (ndash 38 and ndash 50 mV for invertedand upright presentation respectively t(15) = 2104p lt 055 and t(15) = 3000 p lt 01 for natural andschematic faces respectively) The main effect of stim-ulus type was significant [F(115) = 302 p lt 001] but itwas also qualified by its interaction with the effect oforientation Post hoc comparisons showed that whereasthe difference between the N170 elicited by invertednatural and schematic faces was significant [t(15) =4264 p lt 001] the difference between the N170elicited by natural and schematic faces presented uprightwas not [t(15) = 884 p = 39] The main effect of site

was also significant [F(345) = 68 p lt 01] revealingonly that the N170 amplitudes at P7 and P8 were smallerthan at all other sites [F(115) = 117 p lt 01] It isnoteworthy that as evidence by an insignificant interac-tion with stimulus type [F(345) = 13 p = 28] thiseffect was similar for schematic and natural faces On theother hand a significant main effect of hemisphere[F(115) = 65 p lt 05] was qualified by its interactionwith stimulus type [F(115) = 53 p lt 05] Post hocanalysis revealed that whereas for both stimulus typesthe amplitude of the N170 was larger over the rightcompared to the left hemisphere the interhemisphericdifference was significant for natural faces [t(15) = 335p lt 01] but not for schematic faces [t(15) = 1408p = 18] The effects of the Hemisphere pound Orientationinteraction was significant [F(115) = 47 p lt 5] Thisinteraction was caused by the fact that whereas acrossstimulus type the reduction and enhancement of theN170 amplitude by inversion canceled each other in theleft hemisphere in the right hemisphere the enhance-ment of N170 induced by the inversion of natural faces

Faces Inverted Faces Schematic Faces Inverted SchematicFaces

Time(msec) Time(msec)0 100 200 300 400

10

mV 0

0 100 200 300 400

10PO7 PO8

- 0 +

mV 0

Figure 3 The N170 potentials elicited at the left and right mastoids by upright and inverted natural faces and schematic faces and the scalpdistribution at this latency

944 Journal of Cognitive Neuroscience Volume 13 Number 7

weighted more than its reduction resulting from theinversion of schematic faces However the second-orderinteraction between orientation hemisphere and stim-ulus type was not significant [F(115) lt 10] Finally aseries of second-order interactions involving the effect ofsite suggested that all the effects were differently con-spicuous at different sites As revealed in Table 2 by andlarge it seems that whereas natural faces induced similarinversion effects across the analyzed sites the inversioneffect for schematic faces varied at different sites We hadno prior assumptions about the distribution of the effectswithin the posterior temporal scalp regions Further-more the spatial resolution of scalp-recorded ERPs istoo low to permit reliable interpretation of site effectswithin a relative small region unless the same distribu-tion is replicated therefore a detailed analysis of thisdistribution of effects was not attempted post hoc

ANOVA of the N170 peak latencies showed a signifi-cant main effect of stimulus orientation [F(115) =1615 p lt 001] which significantly interacted withstimulus type [F(115) = 62 p lt 025] Plannedcomparisons showed that for both natural and sche-matic faces inversion delayed the N170 significantly[t(15) = 10821 p lt 001 and t(15) = 3626 p lt 01respectively] However as evident in Table 2B thisdifference was larger for natural than for schematicfaces The latencies were similar over the left and righthemispheres [F(115) = 19 p = 185] The main effectof site was significant [F(345) = 47 p lt 01] butbecause the differences between sites were all smaller

than the sampling rate this effect was not investigatedfurther No other significant effects emerged in thisanalysis

Discussion

The results of the Experiment 2 validated our predic-tion that inversion should reduce the amplitude of theN170 elicited by schematic faces but have little effector enhance the N170 elicited by natural faces Thenatural face inversion effect observed in the presentstudy replicates previous results reported by Rossionet al (1999) These authors suggested that the en-hancement in the N170 amplitude reflects the en-hanced difficulty of the holistic processor to encodethe inverted face Indeed the delay in the N170 peaklatency elicited by inverted relative to upright faces iscongruent with such a view However if this were thecase we should have found similar effects of inversionon the N170 amplitude elicited by schematic and bynatural faces Furthermore the assumption that moredifficult face encoding induces an enhancement of theN170 amplitude is also at odds with fMRI data showingthat face inversion reduces the activity in the fusiformface areas relative to the activity elicited by uprightfaces (Kanwisher Tong amp Nakayama 1998) Indeedfinding opposite effects of inversion for schematic andnatural faces can more likely be accommodated by ourhypothesis that the larger N170 amplitude in responseto inverted natural faces reflects its stronger associa-

Table 2 (A) Amplitudes (in mV) and (B) Latencies (in msec) of the N170 Elicited by Upright and Inverted Natural and SchematicFaces at the Posterior Temporal Scalp Sites

Natural Faces Schematic Faces

Right Hemisphere Left Hemisphere Right Hemisphere Left Hemisphere

Upright Inverted Upright Inverted Upright Inverted Upright Inverted

A

Mastoid ndash 64 ndash 86 ndash 58 ndash 82 ndash 54 ndash 55 ndash 51 ndash 35

P87 ndash 48 ndash 68 ndash 45 ndash 54 ndash 41 ndash 30 ndash 32 ndash 33

PO87 ndash 68 ndash 88 ndash 48 ndash 50 ndash 56 ndash 56 ndash 50 ndash 23

IM21 ndash 53 ndash 81 ndash 49 ndash 73 ndash 50 ndash 29 ndash 65 ndash 39

Mean ndash 58 ndash 81 ndash 50 ndash 65 ndash 50 ndash 425 ndash 49 ndash 325

B

Mastoid 149 162 151 165 159 164 158 165

P87 151 163 154 166 158 164 162 166

PO87 151 160 152 164 151 161 159 166

IM21 150 162 147 161 161 164 154 162

Mean 150 162 151 164 157 163 158 165

Sagiv and Bentin 945

tion with a face component analyzer in the lateraltemporal gyri than with the holistic face processor inthe fusiform

Our model proposes that although schematic andnatural faces probably activate similar neural networksin the extrastriate visual pathway these mechanismsare not identical Natural faces probably activate twoperceptual modules one specialized to detect andprocess physiognomic features in the visual field andthe other specialized in the holistic processing of facesor forming a holistic face representation Unlike naturalfaces the components of schematic faces are notperceived as carrying physiognomic information outof the schematic face gestalt Therefore in contrastto natural faces schematic faces may trigger the holisticprocessor but they do not trigger directly the analysisof the components The latter mechanism may beactivated only subsequent to the formation of theholistic face representation or at the very least follow-ing this process rsquo rsquo in cascadersquo rsquo (McClelland 1979) Con-sequently adversely affecting the perception of the facegestalt (by inverting the stimuli) inhibited both holisticperception and the perception of individual compo-nents and this inhibition reduced the amplitude of theN170 Such a pattern is also in agreement with themyriad of studies supporting the primacy of holisticprocessing of face stimuli

GENERAL DISCUSSION

In the present study we investigated several character-istics of the extrastriate structural face-encoding mech-anisms by comparing the N170 elicited by faces differingin their naturalistic aspect and by comparing the faceinversion effects on the N170 elicited by photographs ofnatural faces and by schematically drawn faces To theextent that the N170 is indeed associated with structuralencoding we found that this process distinguishes wellbetween faces and nonfaces stimuli but not amongdifferent face types that render the concept of a faceclearly However we also found evidence that schematicfaces and natural faces are not processed identicallyWhereas inversion of natural faces enhanced the ampli-tude of the N170 inversion of schematic faces reducedits amplitude For both face types the latency of theN170 peak was significantly delayed by inversion Thispattern of results suggests that the face-specific struc-tural encoder can be triggered by a variety of stimuli ifthey include a canonical face configuration Howeverthe process of encoding the face information and form-ing a structural representation is probably differentwhen the physiognomic value of the stimuli dependsupon holistic configuration as opposed to when theindividual components can be associated with faces evenwhen presented outside the face context

Apart from the neurofunctional specificity issue anenduring question in face perception research is whether

the formation of a mental representation of the faceentails extraction of a holistic configuration directly fromthe visual input or whether it is feature-based that iswhether for faces as for other complex visual stimulistructural encoding is based on an initial analysis of parts(for a review of this debate see eg Farah 1990 Farahet al 1995) In principle because faces form a veryhomogeneous and stereotypical category they lendthemselves to holistic perception Holistic processing isadvantageous in face recognition because face compo-nents are fairly similar across individuals and thereforethe distinction among faces is usually based on theinterrelationships between the different face compo-nents and their overall configuration (eg Tanaka ampFarah 1993 Young Hellawell amp Hay 1987)7 Moreoverour memory for faces is usually based on nondecompos-able holistic images which at the interface betweenperception and memory are most efficiently addressedby integrated face representations8However if the well-learned gestalt is altered and holistic perception is notpossible it is probably replaced by a more analyticcomponent-based process of structural encoding Thisanalytic process may inhibit or at least delay the per-ception of the spatial relationships between face com-ponents and therefore may impede identificationIndeed a major source of evidence supporting theholistic view is that faces are more difficult to recognizewhen presented upside down (ie compromising theirfamiliar gestalt) than when presented in upright orien-tation (see reviews in Farah et al 1995 Rhodes et al1993 Valentine 1988)

Having to cope with different types of face input thevisual system must be able to encode the face config-uration directly from the visual display as well as byintegrating face components encoded individuallyThese two processes may concur triggered independ-ently by relevant physiognomic input and performed bydifferent neural circuits to form a complex structuralencoding mechanism in the extrastriate cortex Further-more the two systems may interact supporting orinhibiting each other Hence a regular face presentedin upright orientation may activate the holistic percep-tion process which in turn may inhibit superfluousanalysis of individual components By contrast corrupt-ing the face gestalt by presenting faces in an unusualorientation may trigger (and facilitate) face componentanalysis which should not depend on a particular con-figuration In such a case the integration of the compo-nents into a whole by the holistic processor will followtheir independent analysis assisted by the feature-detection system This hypothesis may explain for in-stance why within the context of a regular face individ-ual face parts are distinguished faster than in isolationbut if the configuration of the face is altered the partsare distinguished as fast as when presented in isolation(Tanaka amp Farah 1993 Sergent 1984)9 The ability toprocess face information both holistically and piecewise

946 Journal of Cognitive Neuroscience Volume 13 Number 7

was implemented in a neural model initially based onscalp recordings (Bentin et al 1996) and more recentlycorrected improved and elaborated on the basis intra-cranial recordings (McCarthy et al 1999 for a detaileddescription and a review of supportive functional MRIdata see McCarthy 1999) This model suggests that theneural mechanism dedicated to structural encoding offaces is a complex system composed of at least twoseparate subsystems One probably located in thelateral part of the middle fusiform gyrus is responsiblefor processing the configuration of the face as a wholeand provides the integrated face representation that isnecessary for face recognition and identification Theother possibly distributed over posterior regions in theIT and middle temporal (MT) gyri and in the OTS isresponsible for detecting physiognomic informationregardless of the spatial configuration of the items thatcarry it The fusiform gyrus is farther away from therecording sites at the scalp than where the temporalgyri are and the orientation of its cells forms anelectrical dipole source that is almost perpendicular tothe lateral aspects of the scalp Hence although bothsources contribute to the modulation of the face-spe-cific N170 the contribution of the posterior temporalface component processor should be considerably moreconspicuous

According to this hypothesis the opposite effects ofinversion on the N170 elicited by natural and schematicfaces might be explained assuming that the holisticencoding system in the fusiform gyrus indeed inhibitsthe component analysis in the lateral temporal lobe facesites Hence if the face stimulus can trigger both theholistic and the component analysis (as it is true fornatural faces) an experimental manipulation that inter-feres with holistic processing (like the face inversionprocedure in the present experiment) should reducethe amount of inhibition of component analysis andconsequently the amplitude of the N170 would beenhanced However if the rsquo rsquocomponentsrsquorsquo do not conveyphysiognomic information so that their analysis by aface-specific mechanism rsquo rsquodependsrsquo rsquo on the previousrecognition of the face gestalt (as it is true for schematicfaces) interfering with holistic perception cannot facil-itate the activity of the component analysis and there-fore the modulation of the N170 would only reflect theadditional difficulty to encode the face ie reducedamplitude and delayed latency

The observed effects of the different factors on theN170 peak latency are nicely aligned with the abovemodel Although stimulus inversion delayed the N170for both natural and schematic faces this delay was twiceas big (12 msec) for former stimuli than for the latter (6msec) This interaction suggests that the inversion had abigger effect on natural than on schematic faces Thismay be because regardless of whether they are upright orinverted the physiognomic value of schematic faces isbased on their gestalt Hence inversion does inhibit the

processing of the gestalt but does not induce a change inprocessing strategy On the other hand natural faces areanalyzed holistically if they are upright but inversionrequires a change in that strategy encouraging theanalysis of components The significantly faster responseto upright natural than to upright schematic faces inExperiment 2 suggests that the extraction of the facegestalt is easier if the face is more natural

There is ample evidence indicating that the righthemisphere is specialized in global or holistic processingof objects and particularly faces whereas the left hemi-sphere is specialized in part-based processing (egRobertson amp Delis 1986 for a review see Bradshow ampNettleton 1983 for specific reference to face perceptionsee eg Bruce 1998 Rhodes et al 1993 Corballis1988 Bradshow amp Sherlock 1982 Gilbert amp Bakan1973) Consequently interfering with holistic processingshould have had stronger effects in the right than in theleft hemisphere Our results confirmed this predictiononly partially The pronounced interhemispheric asym-metry with larger N170 amplitudes at the right than atthe left hemisphere sites for all stimulus types suggests apredominant holistic process for all face categoriesInterestingly this asymmetry was reduced (in Experi-ment 1) for sketches relative to the other stimulus typesThe relatively enhanced involvement of the left hemi-sphere in processing the sketches is in agreement withour suggestion that the face-specific structural encodingsystem was activated less efficiently by sketches than byother face types10 Hence the same factor that may haveconstrained the processing of schematic faces may alsohave constrained the processing of sketches but in theopposite direction Whereas schematic faces enhancedholistic processing by providing a clear face gestalt whileobscuring the physiognomic value of face componentsthe sketches reduced holistic processing by blurring orveiling the face gestalt Also in Experiment 2 the N170was larger over right than over left hemisphere sitesHowever because the interaction between the inversionand hemisphere effects in Experiment 2 was not modu-lated by stimulus type (ie the three-way interactionwas not significant) our interpretation of the interhemi-spheric asymmetry (particularly in Experiment 2) mustbe limited

In conclusion the present scalp-recorded N170 dataconverge with the previously published pattern of theintracranially recorded N200 data (for a comprehensivereview see Allison et al 1999 McCarthy et al 1999Puce et al 1999) supporting a multifunctional organ-ization of the structural encoding mechanism for facesApparently distinct neural circuits are responsible forthe encoding and initial analysis of face components andfor holistic perception of a full-face configuration Theformer is probably located in the lateral posterior tem-poral lobe whereas the latter is located in the middlefusiform gyrus Face perception usually entails predom-inantly holistic processing which explains the right

Sagiv and Bentin 947

hemisphere dominance in performance as well as inelectrophysiological and neuroimaging data Howeverboth mechanisms can be triggered independently byrelevant information

METHODS

Participants

Right-handed participants were recruited among volun-teer students from the Hebrew University They receivedeither experimental credit hours or payment for theirparticipation Thirty participants (13 males and 17 fe-males) were tested in the first experiment and 16participants (8 males and 8 females) were tested in thesecond experiment Participantsrsquo ages ranged from 19 to37 years (mean = 233 years) They all had normal orcorrected-to-normal vision

Stimuli

Stimuli were scanned using a desk scanner processedwith commercially available graphic software and con-

verted to 16 color grayscale 248 by 248 pixel imagesThere were eight stimulus categories four face types(photographs of natural faces paintings face sketchesand schematic faces) two nonsense patterns (formedby scrambling the natural faces and the schematicfaces) flowers and butterflies (Figure 4) Paintingsand sketches of faces were scanned from art booksPaintings were typically realistic Renaissance portraitsand face sketches did not portray details and usuallylacked shading and texture Schematic face parts weredrawn by hand scanned into the computer and latercombined into different schematic faces Face photo-graphs and paintings contained many shades of grayand the sketches and the schematic faces were basicallyblack on white All face stimuli were front views

Scrambled faces were created by scrambling thephases of the two-dimensional Fourier transform andthen smoothing the uneven edges created by this proc-ess The original and the scrambled stimuli were equallyluminous following this procedure Each stimulus cat-egory comprised 75 stimuli except for the butterflycategory which consisted of 30 stimuli In Experiment2 only natural faces schematic faces and flowers were

Face Stimuli

Photographs Painted portraits Sketches Schematic faces

Nonface Stimuli

Scrambledphotographs

Butterflies(targets) Flowers Scrambled

schematic faces

Figure 4 Example of the eight stimulus types used in the present study

948 Journal of Cognitive Neuroscience Volume 13 Number 7

used The faces were presented in both upright andupside down orientation Upright stimuli were the samein Experiments 1 and 2

All stimuli were presented for 350 msec with anaverage interstimulus interval (ISI) of 12 sec and viewedfrom a distance of 110 plusmn 10 cm subtending a viewingangle of approximately 68 Stimulus timing was con-trolled by the MEL software system (Psychology Soft-ware Tools Pittsburgh PA)

Task

Participants were instructed to keep a mental countof stimuli belonging to one category butterflies inthe first experiment and flowers in the secondHence all of the experiment-relevant categories werefrom the participantsrsquo point of view nontargets Thetargets were presented in approximately 5 (Experi-ment 1) and 6 (Experiment 2) of the trials Duringeach break and at the end of the experimentparticipants reported the number of targets countedA few participants who did not count correctly wereexcluded

EEG Recording and Data Analysis

Continuous EEG was recorded from 48 scalp locationsusing tin electrodes attached to elastic electrode caps(Electrode Caps International Ohio) A schematic illus-tration of the electrode locations and names is pre-sented in Figure 1 A ground electrode was located onthe forehead and a reference electrode was located onthe tip of the nose Two additional electrodes located atthe outer canthus and the infraorbital region of the righteye measured vertical and horizontal electrooculograms(EOGs)

The EEG was sampled at 250 Hz by a DAP2400 DataAcquisition Processor (Microstar Laboratories) ampli-fied 20000-fold using a battery-operated HU-5072 Iso-lated Bioelectric Amplifier (SA Instrumentation SanDiego) with a frequency bandpass of 01ndash30 Hz andstored on disk for off-line averaging Electrode impe-dance was below 5 k A digital filter of 08ndash16 Hz wasapplied before analysis

Codes synchronized to stimulus delivery were used toselectively average epochs associated with differentstimulus categories Epochs contaminated with saccadesor eye blinks were excluded using root-mean-squareEOG-values as criteria Typically epochs associated with55ndash70 (never less than 50) out of 75 stimuli in eachcategory were averaged

The mean and peak amplitudes and peak latencies ofspecific ERP components within bounding intervals werecomputed for each category for each participant Sig-nificance of the differences between these measures wastested using a repeated-measures ANOVA Where appro-priate the more conservative GreenhousendashGeisser ad-

justed df values were used ANOVAs were followed byunivariate F contrasts

Acknowledgments

This study was supported by grants from the USndashIsraelBinational Science Foundation and Israeli Foundations Trust-ees to S Bentin We thank Amir Katz for his help in preparingthe stimuli and running Experiment 1 as well as Leon DeouellMichael Fink Ayelet Landau Yulia Goland David Carmel GilRaviv Amir Raz and Dan Tzur for their support We are alsovery grateful to Jerry Levinson Alex List and Lynn Robertsonfor their comments on the manuscript

Reprint requests should be sent to Shlomo Bentin TheHebrew University of Jerusalem Mount Scopus JerusalemIsrael 91905 or via e-mail msbentinmscchujiacil

Notes

1 Present address University of California Berkeley2 There was no reliable difference between the ERP elicited

by scrambled natural faces and scrambled schematic facesTherefore these two categories were collapsed into one3 These were the sites at which the N170 was most

conspicuous4 So were in fact all face categories except for the sketches5 Note that this result also contradicts the view that a more

difficult encoding process elicit N170 of a higher amplitude6 This assumption was indeed supported by a current study

(Bentin Mecklinger Bosch Sagiv amp von Cramon 1999)7 Of course some faces have very distinctive components

which can be used as markers for recognition Such markersare frequently exaggerated and used as distinctive features incaricatures (Rhodes 1987)8 Distinctive features (such as unusual components birth-

marks etc) may be stored in semantic memory as part of apersonrsquo s face entry Such information may have visual qualitiesthat can be addressed directly during visual matching tofacilitate identification9 Note that this account for the rsquorsquoface superiority effectrsquo rsquo is

identical to a common explanation of the rsquorsquoword superiorityeffectrsquo rsquo in the identification of single letters Indeed ourassumption about the coexistence of rsquo rsquowhole-facersquo rsquo and rsquo rsquoface-partrsquo rsquo processors resonates to the rsquorsquodual routersquo rsquo model of visualword processing It is quite possible that identical principlesgovern all the content-specific perceptual modules in the visualsystem10 Recall that relative to other face types the reduction of theN170 amplitude elicited by sketches at right hemisphere siteswas greater then at left hemisphere sites

REFERENCES

Allison T Ginter H McCarthy G Nobre A C Puce ALuby M amp Spencer D D (1994) Face recognition in hu-man extrastriate cortex Journal of Neurophysiology 71821ndash825

Allison T McCarthy G Nobre A C Puce A amp Belger A(1994) Human extrastriate visual cortex and the perceptionof faces words numbers and colors Cerebral Cortex 5544ndash554

Allison T Puce A Spencer D D amp McCarthy G (1999)Electrophysiological studies of human face perception I

Sagiv and Bentin 949

weighted more than its reduction resulting from theinversion of schematic faces However the second-orderinteraction between orientation hemisphere and stim-ulus type was not significant [F(115) lt 10] Finally aseries of second-order interactions involving the effect ofsite suggested that all the effects were differently con-spicuous at different sites As revealed in Table 2 by andlarge it seems that whereas natural faces induced similarinversion effects across the analyzed sites the inversioneffect for schematic faces varied at different sites We hadno prior assumptions about the distribution of the effectswithin the posterior temporal scalp regions Further-more the spatial resolution of scalp-recorded ERPs istoo low to permit reliable interpretation of site effectswithin a relative small region unless the same distribu-tion is replicated therefore a detailed analysis of thisdistribution of effects was not attempted post hoc

ANOVA of the N170 peak latencies showed a signifi-cant main effect of stimulus orientation [F(115) =1615 p lt 001] which significantly interacted withstimulus type [F(115) = 62 p lt 025] Plannedcomparisons showed that for both natural and sche-matic faces inversion delayed the N170 significantly[t(15) = 10821 p lt 001 and t(15) = 3626 p lt 01respectively] However as evident in Table 2B thisdifference was larger for natural than for schematicfaces The latencies were similar over the left and righthemispheres [F(115) = 19 p = 185] The main effectof site was significant [F(345) = 47 p lt 01] butbecause the differences between sites were all smaller

than the sampling rate this effect was not investigatedfurther No other significant effects emerged in thisanalysis

Discussion

The results of the Experiment 2 validated our predic-tion that inversion should reduce the amplitude of theN170 elicited by schematic faces but have little effector enhance the N170 elicited by natural faces Thenatural face inversion effect observed in the presentstudy replicates previous results reported by Rossionet al (1999) These authors suggested that the en-hancement in the N170 amplitude reflects the en-hanced difficulty of the holistic processor to encodethe inverted face Indeed the delay in the N170 peaklatency elicited by inverted relative to upright faces iscongruent with such a view However if this were thecase we should have found similar effects of inversionon the N170 amplitude elicited by schematic and bynatural faces Furthermore the assumption that moredifficult face encoding induces an enhancement of theN170 amplitude is also at odds with fMRI data showingthat face inversion reduces the activity in the fusiformface areas relative to the activity elicited by uprightfaces (Kanwisher Tong amp Nakayama 1998) Indeedfinding opposite effects of inversion for schematic andnatural faces can more likely be accommodated by ourhypothesis that the larger N170 amplitude in responseto inverted natural faces reflects its stronger associa-

Table 2 (A) Amplitudes (in mV) and (B) Latencies (in msec) of the N170 Elicited by Upright and Inverted Natural and SchematicFaces at the Posterior Temporal Scalp Sites

Natural Faces Schematic Faces

Right Hemisphere Left Hemisphere Right Hemisphere Left Hemisphere

Upright Inverted Upright Inverted Upright Inverted Upright Inverted

A

Mastoid ndash 64 ndash 86 ndash 58 ndash 82 ndash 54 ndash 55 ndash 51 ndash 35

P87 ndash 48 ndash 68 ndash 45 ndash 54 ndash 41 ndash 30 ndash 32 ndash 33

PO87 ndash 68 ndash 88 ndash 48 ndash 50 ndash 56 ndash 56 ndash 50 ndash 23

IM21 ndash 53 ndash 81 ndash 49 ndash 73 ndash 50 ndash 29 ndash 65 ndash 39

Mean ndash 58 ndash 81 ndash 50 ndash 65 ndash 50 ndash 425 ndash 49 ndash 325

B

Mastoid 149 162 151 165 159 164 158 165

P87 151 163 154 166 158 164 162 166

PO87 151 160 152 164 151 161 159 166

IM21 150 162 147 161 161 164 154 162

Mean 150 162 151 164 157 163 158 165

Sagiv and Bentin 945

tion with a face component analyzer in the lateraltemporal gyri than with the holistic face processor inthe fusiform

Our model proposes that although schematic andnatural faces probably activate similar neural networksin the extrastriate visual pathway these mechanismsare not identical Natural faces probably activate twoperceptual modules one specialized to detect andprocess physiognomic features in the visual field andthe other specialized in the holistic processing of facesor forming a holistic face representation Unlike naturalfaces the components of schematic faces are notperceived as carrying physiognomic information outof the schematic face gestalt Therefore in contrastto natural faces schematic faces may trigger the holisticprocessor but they do not trigger directly the analysisof the components The latter mechanism may beactivated only subsequent to the formation of theholistic face representation or at the very least follow-ing this process rsquo rsquo in cascadersquo rsquo (McClelland 1979) Con-sequently adversely affecting the perception of the facegestalt (by inverting the stimuli) inhibited both holisticperception and the perception of individual compo-nents and this inhibition reduced the amplitude of theN170 Such a pattern is also in agreement with themyriad of studies supporting the primacy of holisticprocessing of face stimuli

GENERAL DISCUSSION

In the present study we investigated several character-istics of the extrastriate structural face-encoding mech-anisms by comparing the N170 elicited by faces differingin their naturalistic aspect and by comparing the faceinversion effects on the N170 elicited by photographs ofnatural faces and by schematically drawn faces To theextent that the N170 is indeed associated with structuralencoding we found that this process distinguishes wellbetween faces and nonfaces stimuli but not amongdifferent face types that render the concept of a faceclearly However we also found evidence that schematicfaces and natural faces are not processed identicallyWhereas inversion of natural faces enhanced the ampli-tude of the N170 inversion of schematic faces reducedits amplitude For both face types the latency of theN170 peak was significantly delayed by inversion Thispattern of results suggests that the face-specific struc-tural encoder can be triggered by a variety of stimuli ifthey include a canonical face configuration Howeverthe process of encoding the face information and form-ing a structural representation is probably differentwhen the physiognomic value of the stimuli dependsupon holistic configuration as opposed to when theindividual components can be associated with faces evenwhen presented outside the face context

Apart from the neurofunctional specificity issue anenduring question in face perception research is whether

the formation of a mental representation of the faceentails extraction of a holistic configuration directly fromthe visual input or whether it is feature-based that iswhether for faces as for other complex visual stimulistructural encoding is based on an initial analysis of parts(for a review of this debate see eg Farah 1990 Farahet al 1995) In principle because faces form a veryhomogeneous and stereotypical category they lendthemselves to holistic perception Holistic processing isadvantageous in face recognition because face compo-nents are fairly similar across individuals and thereforethe distinction among faces is usually based on theinterrelationships between the different face compo-nents and their overall configuration (eg Tanaka ampFarah 1993 Young Hellawell amp Hay 1987)7 Moreoverour memory for faces is usually based on nondecompos-able holistic images which at the interface betweenperception and memory are most efficiently addressedby integrated face representations8However if the well-learned gestalt is altered and holistic perception is notpossible it is probably replaced by a more analyticcomponent-based process of structural encoding Thisanalytic process may inhibit or at least delay the per-ception of the spatial relationships between face com-ponents and therefore may impede identificationIndeed a major source of evidence supporting theholistic view is that faces are more difficult to recognizewhen presented upside down (ie compromising theirfamiliar gestalt) than when presented in upright orien-tation (see reviews in Farah et al 1995 Rhodes et al1993 Valentine 1988)

Having to cope with different types of face input thevisual system must be able to encode the face config-uration directly from the visual display as well as byintegrating face components encoded individuallyThese two processes may concur triggered independ-ently by relevant physiognomic input and performed bydifferent neural circuits to form a complex structuralencoding mechanism in the extrastriate cortex Further-more the two systems may interact supporting orinhibiting each other Hence a regular face presentedin upright orientation may activate the holistic percep-tion process which in turn may inhibit superfluousanalysis of individual components By contrast corrupt-ing the face gestalt by presenting faces in an unusualorientation may trigger (and facilitate) face componentanalysis which should not depend on a particular con-figuration In such a case the integration of the compo-nents into a whole by the holistic processor will followtheir independent analysis assisted by the feature-detection system This hypothesis may explain for in-stance why within the context of a regular face individ-ual face parts are distinguished faster than in isolationbut if the configuration of the face is altered the partsare distinguished as fast as when presented in isolation(Tanaka amp Farah 1993 Sergent 1984)9 The ability toprocess face information both holistically and piecewise

946 Journal of Cognitive Neuroscience Volume 13 Number 7

was implemented in a neural model initially based onscalp recordings (Bentin et al 1996) and more recentlycorrected improved and elaborated on the basis intra-cranial recordings (McCarthy et al 1999 for a detaileddescription and a review of supportive functional MRIdata see McCarthy 1999) This model suggests that theneural mechanism dedicated to structural encoding offaces is a complex system composed of at least twoseparate subsystems One probably located in thelateral part of the middle fusiform gyrus is responsiblefor processing the configuration of the face as a wholeand provides the integrated face representation that isnecessary for face recognition and identification Theother possibly distributed over posterior regions in theIT and middle temporal (MT) gyri and in the OTS isresponsible for detecting physiognomic informationregardless of the spatial configuration of the items thatcarry it The fusiform gyrus is farther away from therecording sites at the scalp than where the temporalgyri are and the orientation of its cells forms anelectrical dipole source that is almost perpendicular tothe lateral aspects of the scalp Hence although bothsources contribute to the modulation of the face-spe-cific N170 the contribution of the posterior temporalface component processor should be considerably moreconspicuous

According to this hypothesis the opposite effects ofinversion on the N170 elicited by natural and schematicfaces might be explained assuming that the holisticencoding system in the fusiform gyrus indeed inhibitsthe component analysis in the lateral temporal lobe facesites Hence if the face stimulus can trigger both theholistic and the component analysis (as it is true fornatural faces) an experimental manipulation that inter-feres with holistic processing (like the face inversionprocedure in the present experiment) should reducethe amount of inhibition of component analysis andconsequently the amplitude of the N170 would beenhanced However if the rsquo rsquocomponentsrsquorsquo do not conveyphysiognomic information so that their analysis by aface-specific mechanism rsquo rsquodependsrsquo rsquo on the previousrecognition of the face gestalt (as it is true for schematicfaces) interfering with holistic perception cannot facil-itate the activity of the component analysis and there-fore the modulation of the N170 would only reflect theadditional difficulty to encode the face ie reducedamplitude and delayed latency

The observed effects of the different factors on theN170 peak latency are nicely aligned with the abovemodel Although stimulus inversion delayed the N170for both natural and schematic faces this delay was twiceas big (12 msec) for former stimuli than for the latter (6msec) This interaction suggests that the inversion had abigger effect on natural than on schematic faces Thismay be because regardless of whether they are upright orinverted the physiognomic value of schematic faces isbased on their gestalt Hence inversion does inhibit the

processing of the gestalt but does not induce a change inprocessing strategy On the other hand natural faces areanalyzed holistically if they are upright but inversionrequires a change in that strategy encouraging theanalysis of components The significantly faster responseto upright natural than to upright schematic faces inExperiment 2 suggests that the extraction of the facegestalt is easier if the face is more natural

There is ample evidence indicating that the righthemisphere is specialized in global or holistic processingof objects and particularly faces whereas the left hemi-sphere is specialized in part-based processing (egRobertson amp Delis 1986 for a review see Bradshow ampNettleton 1983 for specific reference to face perceptionsee eg Bruce 1998 Rhodes et al 1993 Corballis1988 Bradshow amp Sherlock 1982 Gilbert amp Bakan1973) Consequently interfering with holistic processingshould have had stronger effects in the right than in theleft hemisphere Our results confirmed this predictiononly partially The pronounced interhemispheric asym-metry with larger N170 amplitudes at the right than atthe left hemisphere sites for all stimulus types suggests apredominant holistic process for all face categoriesInterestingly this asymmetry was reduced (in Experi-ment 1) for sketches relative to the other stimulus typesThe relatively enhanced involvement of the left hemi-sphere in processing the sketches is in agreement withour suggestion that the face-specific structural encodingsystem was activated less efficiently by sketches than byother face types10 Hence the same factor that may haveconstrained the processing of schematic faces may alsohave constrained the processing of sketches but in theopposite direction Whereas schematic faces enhancedholistic processing by providing a clear face gestalt whileobscuring the physiognomic value of face componentsthe sketches reduced holistic processing by blurring orveiling the face gestalt Also in Experiment 2 the N170was larger over right than over left hemisphere sitesHowever because the interaction between the inversionand hemisphere effects in Experiment 2 was not modu-lated by stimulus type (ie the three-way interactionwas not significant) our interpretation of the interhemi-spheric asymmetry (particularly in Experiment 2) mustbe limited

In conclusion the present scalp-recorded N170 dataconverge with the previously published pattern of theintracranially recorded N200 data (for a comprehensivereview see Allison et al 1999 McCarthy et al 1999Puce et al 1999) supporting a multifunctional organ-ization of the structural encoding mechanism for facesApparently distinct neural circuits are responsible forthe encoding and initial analysis of face components andfor holistic perception of a full-face configuration Theformer is probably located in the lateral posterior tem-poral lobe whereas the latter is located in the middlefusiform gyrus Face perception usually entails predom-inantly holistic processing which explains the right

Sagiv and Bentin 947

hemisphere dominance in performance as well as inelectrophysiological and neuroimaging data Howeverboth mechanisms can be triggered independently byrelevant information

METHODS

Participants

Right-handed participants were recruited among volun-teer students from the Hebrew University They receivedeither experimental credit hours or payment for theirparticipation Thirty participants (13 males and 17 fe-males) were tested in the first experiment and 16participants (8 males and 8 females) were tested in thesecond experiment Participantsrsquo ages ranged from 19 to37 years (mean = 233 years) They all had normal orcorrected-to-normal vision

Stimuli

Stimuli were scanned using a desk scanner processedwith commercially available graphic software and con-

verted to 16 color grayscale 248 by 248 pixel imagesThere were eight stimulus categories four face types(photographs of natural faces paintings face sketchesand schematic faces) two nonsense patterns (formedby scrambling the natural faces and the schematicfaces) flowers and butterflies (Figure 4) Paintingsand sketches of faces were scanned from art booksPaintings were typically realistic Renaissance portraitsand face sketches did not portray details and usuallylacked shading and texture Schematic face parts weredrawn by hand scanned into the computer and latercombined into different schematic faces Face photo-graphs and paintings contained many shades of grayand the sketches and the schematic faces were basicallyblack on white All face stimuli were front views

Scrambled faces were created by scrambling thephases of the two-dimensional Fourier transform andthen smoothing the uneven edges created by this proc-ess The original and the scrambled stimuli were equallyluminous following this procedure Each stimulus cat-egory comprised 75 stimuli except for the butterflycategory which consisted of 30 stimuli In Experiment2 only natural faces schematic faces and flowers were

Face Stimuli

Photographs Painted portraits Sketches Schematic faces

Nonface Stimuli

Scrambledphotographs

Butterflies(targets) Flowers Scrambled

schematic faces

Figure 4 Example of the eight stimulus types used in the present study

948 Journal of Cognitive Neuroscience Volume 13 Number 7

used The faces were presented in both upright andupside down orientation Upright stimuli were the samein Experiments 1 and 2

All stimuli were presented for 350 msec with anaverage interstimulus interval (ISI) of 12 sec and viewedfrom a distance of 110 plusmn 10 cm subtending a viewingangle of approximately 68 Stimulus timing was con-trolled by the MEL software system (Psychology Soft-ware Tools Pittsburgh PA)

Task

Participants were instructed to keep a mental countof stimuli belonging to one category butterflies inthe first experiment and flowers in the secondHence all of the experiment-relevant categories werefrom the participantsrsquo point of view nontargets Thetargets were presented in approximately 5 (Experi-ment 1) and 6 (Experiment 2) of the trials Duringeach break and at the end of the experimentparticipants reported the number of targets countedA few participants who did not count correctly wereexcluded

EEG Recording and Data Analysis

Continuous EEG was recorded from 48 scalp locationsusing tin electrodes attached to elastic electrode caps(Electrode Caps International Ohio) A schematic illus-tration of the electrode locations and names is pre-sented in Figure 1 A ground electrode was located onthe forehead and a reference electrode was located onthe tip of the nose Two additional electrodes located atthe outer canthus and the infraorbital region of the righteye measured vertical and horizontal electrooculograms(EOGs)

The EEG was sampled at 250 Hz by a DAP2400 DataAcquisition Processor (Microstar Laboratories) ampli-fied 20000-fold using a battery-operated HU-5072 Iso-lated Bioelectric Amplifier (SA Instrumentation SanDiego) with a frequency bandpass of 01ndash30 Hz andstored on disk for off-line averaging Electrode impe-dance was below 5 k A digital filter of 08ndash16 Hz wasapplied before analysis

Codes synchronized to stimulus delivery were used toselectively average epochs associated with differentstimulus categories Epochs contaminated with saccadesor eye blinks were excluded using root-mean-squareEOG-values as criteria Typically epochs associated with55ndash70 (never less than 50) out of 75 stimuli in eachcategory were averaged

The mean and peak amplitudes and peak latencies ofspecific ERP components within bounding intervals werecomputed for each category for each participant Sig-nificance of the differences between these measures wastested using a repeated-measures ANOVA Where appro-priate the more conservative GreenhousendashGeisser ad-

justed df values were used ANOVAs were followed byunivariate F contrasts

Acknowledgments

This study was supported by grants from the USndashIsraelBinational Science Foundation and Israeli Foundations Trust-ees to S Bentin We thank Amir Katz for his help in preparingthe stimuli and running Experiment 1 as well as Leon DeouellMichael Fink Ayelet Landau Yulia Goland David Carmel GilRaviv Amir Raz and Dan Tzur for their support We are alsovery grateful to Jerry Levinson Alex List and Lynn Robertsonfor their comments on the manuscript

Reprint requests should be sent to Shlomo Bentin TheHebrew University of Jerusalem Mount Scopus JerusalemIsrael 91905 or via e-mail msbentinmscchujiacil

Notes

1 Present address University of California Berkeley2 There was no reliable difference between the ERP elicited

by scrambled natural faces and scrambled schematic facesTherefore these two categories were collapsed into one3 These were the sites at which the N170 was most

conspicuous4 So were in fact all face categories except for the sketches5 Note that this result also contradicts the view that a more

difficult encoding process elicit N170 of a higher amplitude6 This assumption was indeed supported by a current study

(Bentin Mecklinger Bosch Sagiv amp von Cramon 1999)7 Of course some faces have very distinctive components

which can be used as markers for recognition Such markersare frequently exaggerated and used as distinctive features incaricatures (Rhodes 1987)8 Distinctive features (such as unusual components birth-

marks etc) may be stored in semantic memory as part of apersonrsquo s face entry Such information may have visual qualitiesthat can be addressed directly during visual matching tofacilitate identification9 Note that this account for the rsquorsquoface superiority effectrsquo rsquo is

identical to a common explanation of the rsquorsquoword superiorityeffectrsquo rsquo in the identification of single letters Indeed ourassumption about the coexistence of rsquo rsquowhole-facersquo rsquo and rsquo rsquoface-partrsquo rsquo processors resonates to the rsquorsquodual routersquo rsquo model of visualword processing It is quite possible that identical principlesgovern all the content-specific perceptual modules in the visualsystem10 Recall that relative to other face types the reduction of theN170 amplitude elicited by sketches at right hemisphere siteswas greater then at left hemisphere sites

REFERENCES

Allison T Ginter H McCarthy G Nobre A C Puce ALuby M amp Spencer D D (1994) Face recognition in hu-man extrastriate cortex Journal of Neurophysiology 71821ndash825

Allison T McCarthy G Nobre A C Puce A amp Belger A(1994) Human extrastriate visual cortex and the perceptionof faces words numbers and colors Cerebral Cortex 5544ndash554

Allison T Puce A Spencer D D amp McCarthy G (1999)Electrophysiological studies of human face perception I

Sagiv and Bentin 949

tion with a face component analyzer in the lateraltemporal gyri than with the holistic face processor inthe fusiform

Our model proposes that although schematic andnatural faces probably activate similar neural networksin the extrastriate visual pathway these mechanismsare not identical Natural faces probably activate twoperceptual modules one specialized to detect andprocess physiognomic features in the visual field andthe other specialized in the holistic processing of facesor forming a holistic face representation Unlike naturalfaces the components of schematic faces are notperceived as carrying physiognomic information outof the schematic face gestalt Therefore in contrastto natural faces schematic faces may trigger the holisticprocessor but they do not trigger directly the analysisof the components The latter mechanism may beactivated only subsequent to the formation of theholistic face representation or at the very least follow-ing this process rsquo rsquo in cascadersquo rsquo (McClelland 1979) Con-sequently adversely affecting the perception of the facegestalt (by inverting the stimuli) inhibited both holisticperception and the perception of individual compo-nents and this inhibition reduced the amplitude of theN170 Such a pattern is also in agreement with themyriad of studies supporting the primacy of holisticprocessing of face stimuli

GENERAL DISCUSSION

In the present study we investigated several character-istics of the extrastriate structural face-encoding mech-anisms by comparing the N170 elicited by faces differingin their naturalistic aspect and by comparing the faceinversion effects on the N170 elicited by photographs ofnatural faces and by schematically drawn faces To theextent that the N170 is indeed associated with structuralencoding we found that this process distinguishes wellbetween faces and nonfaces stimuli but not amongdifferent face types that render the concept of a faceclearly However we also found evidence that schematicfaces and natural faces are not processed identicallyWhereas inversion of natural faces enhanced the ampli-tude of the N170 inversion of schematic faces reducedits amplitude For both face types the latency of theN170 peak was significantly delayed by inversion Thispattern of results suggests that the face-specific struc-tural encoder can be triggered by a variety of stimuli ifthey include a canonical face configuration Howeverthe process of encoding the face information and form-ing a structural representation is probably differentwhen the physiognomic value of the stimuli dependsupon holistic configuration as opposed to when theindividual components can be associated with faces evenwhen presented outside the face context

Apart from the neurofunctional specificity issue anenduring question in face perception research is whether

the formation of a mental representation of the faceentails extraction of a holistic configuration directly fromthe visual input or whether it is feature-based that iswhether for faces as for other complex visual stimulistructural encoding is based on an initial analysis of parts(for a review of this debate see eg Farah 1990 Farahet al 1995) In principle because faces form a veryhomogeneous and stereotypical category they lendthemselves to holistic perception Holistic processing isadvantageous in face recognition because face compo-nents are fairly similar across individuals and thereforethe distinction among faces is usually based on theinterrelationships between the different face compo-nents and their overall configuration (eg Tanaka ampFarah 1993 Young Hellawell amp Hay 1987)7 Moreoverour memory for faces is usually based on nondecompos-able holistic images which at the interface betweenperception and memory are most efficiently addressedby integrated face representations8However if the well-learned gestalt is altered and holistic perception is notpossible it is probably replaced by a more analyticcomponent-based process of structural encoding Thisanalytic process may inhibit or at least delay the per-ception of the spatial relationships between face com-ponents and therefore may impede identificationIndeed a major source of evidence supporting theholistic view is that faces are more difficult to recognizewhen presented upside down (ie compromising theirfamiliar gestalt) than when presented in upright orien-tation (see reviews in Farah et al 1995 Rhodes et al1993 Valentine 1988)

Having to cope with different types of face input thevisual system must be able to encode the face config-uration directly from the visual display as well as byintegrating face components encoded individuallyThese two processes may concur triggered independ-ently by relevant physiognomic input and performed bydifferent neural circuits to form a complex structuralencoding mechanism in the extrastriate cortex Further-more the two systems may interact supporting orinhibiting each other Hence a regular face presentedin upright orientation may activate the holistic percep-tion process which in turn may inhibit superfluousanalysis of individual components By contrast corrupt-ing the face gestalt by presenting faces in an unusualorientation may trigger (and facilitate) face componentanalysis which should not depend on a particular con-figuration In such a case the integration of the compo-nents into a whole by the holistic processor will followtheir independent analysis assisted by the feature-detection system This hypothesis may explain for in-stance why within the context of a regular face individ-ual face parts are distinguished faster than in isolationbut if the configuration of the face is altered the partsare distinguished as fast as when presented in isolation(Tanaka amp Farah 1993 Sergent 1984)9 The ability toprocess face information both holistically and piecewise

946 Journal of Cognitive Neuroscience Volume 13 Number 7

was implemented in a neural model initially based onscalp recordings (Bentin et al 1996) and more recentlycorrected improved and elaborated on the basis intra-cranial recordings (McCarthy et al 1999 for a detaileddescription and a review of supportive functional MRIdata see McCarthy 1999) This model suggests that theneural mechanism dedicated to structural encoding offaces is a complex system composed of at least twoseparate subsystems One probably located in thelateral part of the middle fusiform gyrus is responsiblefor processing the configuration of the face as a wholeand provides the integrated face representation that isnecessary for face recognition and identification Theother possibly distributed over posterior regions in theIT and middle temporal (MT) gyri and in the OTS isresponsible for detecting physiognomic informationregardless of the spatial configuration of the items thatcarry it The fusiform gyrus is farther away from therecording sites at the scalp than where the temporalgyri are and the orientation of its cells forms anelectrical dipole source that is almost perpendicular tothe lateral aspects of the scalp Hence although bothsources contribute to the modulation of the face-spe-cific N170 the contribution of the posterior temporalface component processor should be considerably moreconspicuous

According to this hypothesis the opposite effects ofinversion on the N170 elicited by natural and schematicfaces might be explained assuming that the holisticencoding system in the fusiform gyrus indeed inhibitsthe component analysis in the lateral temporal lobe facesites Hence if the face stimulus can trigger both theholistic and the component analysis (as it is true fornatural faces) an experimental manipulation that inter-feres with holistic processing (like the face inversionprocedure in the present experiment) should reducethe amount of inhibition of component analysis andconsequently the amplitude of the N170 would beenhanced However if the rsquo rsquocomponentsrsquorsquo do not conveyphysiognomic information so that their analysis by aface-specific mechanism rsquo rsquodependsrsquo rsquo on the previousrecognition of the face gestalt (as it is true for schematicfaces) interfering with holistic perception cannot facil-itate the activity of the component analysis and there-fore the modulation of the N170 would only reflect theadditional difficulty to encode the face ie reducedamplitude and delayed latency

The observed effects of the different factors on theN170 peak latency are nicely aligned with the abovemodel Although stimulus inversion delayed the N170for both natural and schematic faces this delay was twiceas big (12 msec) for former stimuli than for the latter (6msec) This interaction suggests that the inversion had abigger effect on natural than on schematic faces Thismay be because regardless of whether they are upright orinverted the physiognomic value of schematic faces isbased on their gestalt Hence inversion does inhibit the

processing of the gestalt but does not induce a change inprocessing strategy On the other hand natural faces areanalyzed holistically if they are upright but inversionrequires a change in that strategy encouraging theanalysis of components The significantly faster responseto upright natural than to upright schematic faces inExperiment 2 suggests that the extraction of the facegestalt is easier if the face is more natural

There is ample evidence indicating that the righthemisphere is specialized in global or holistic processingof objects and particularly faces whereas the left hemi-sphere is specialized in part-based processing (egRobertson amp Delis 1986 for a review see Bradshow ampNettleton 1983 for specific reference to face perceptionsee eg Bruce 1998 Rhodes et al 1993 Corballis1988 Bradshow amp Sherlock 1982 Gilbert amp Bakan1973) Consequently interfering with holistic processingshould have had stronger effects in the right than in theleft hemisphere Our results confirmed this predictiononly partially The pronounced interhemispheric asym-metry with larger N170 amplitudes at the right than atthe left hemisphere sites for all stimulus types suggests apredominant holistic process for all face categoriesInterestingly this asymmetry was reduced (in Experi-ment 1) for sketches relative to the other stimulus typesThe relatively enhanced involvement of the left hemi-sphere in processing the sketches is in agreement withour suggestion that the face-specific structural encodingsystem was activated less efficiently by sketches than byother face types10 Hence the same factor that may haveconstrained the processing of schematic faces may alsohave constrained the processing of sketches but in theopposite direction Whereas schematic faces enhancedholistic processing by providing a clear face gestalt whileobscuring the physiognomic value of face componentsthe sketches reduced holistic processing by blurring orveiling the face gestalt Also in Experiment 2 the N170was larger over right than over left hemisphere sitesHowever because the interaction between the inversionand hemisphere effects in Experiment 2 was not modu-lated by stimulus type (ie the three-way interactionwas not significant) our interpretation of the interhemi-spheric asymmetry (particularly in Experiment 2) mustbe limited

In conclusion the present scalp-recorded N170 dataconverge with the previously published pattern of theintracranially recorded N200 data (for a comprehensivereview see Allison et al 1999 McCarthy et al 1999Puce et al 1999) supporting a multifunctional organ-ization of the structural encoding mechanism for facesApparently distinct neural circuits are responsible forthe encoding and initial analysis of face components andfor holistic perception of a full-face configuration Theformer is probably located in the lateral posterior tem-poral lobe whereas the latter is located in the middlefusiform gyrus Face perception usually entails predom-inantly holistic processing which explains the right

Sagiv and Bentin 947

hemisphere dominance in performance as well as inelectrophysiological and neuroimaging data Howeverboth mechanisms can be triggered independently byrelevant information

METHODS

Participants

Right-handed participants were recruited among volun-teer students from the Hebrew University They receivedeither experimental credit hours or payment for theirparticipation Thirty participants (13 males and 17 fe-males) were tested in the first experiment and 16participants (8 males and 8 females) were tested in thesecond experiment Participantsrsquo ages ranged from 19 to37 years (mean = 233 years) They all had normal orcorrected-to-normal vision

Stimuli

Stimuli were scanned using a desk scanner processedwith commercially available graphic software and con-

verted to 16 color grayscale 248 by 248 pixel imagesThere were eight stimulus categories four face types(photographs of natural faces paintings face sketchesand schematic faces) two nonsense patterns (formedby scrambling the natural faces and the schematicfaces) flowers and butterflies (Figure 4) Paintingsand sketches of faces were scanned from art booksPaintings were typically realistic Renaissance portraitsand face sketches did not portray details and usuallylacked shading and texture Schematic face parts weredrawn by hand scanned into the computer and latercombined into different schematic faces Face photo-graphs and paintings contained many shades of grayand the sketches and the schematic faces were basicallyblack on white All face stimuli were front views

Scrambled faces were created by scrambling thephases of the two-dimensional Fourier transform andthen smoothing the uneven edges created by this proc-ess The original and the scrambled stimuli were equallyluminous following this procedure Each stimulus cat-egory comprised 75 stimuli except for the butterflycategory which consisted of 30 stimuli In Experiment2 only natural faces schematic faces and flowers were

Face Stimuli

Photographs Painted portraits Sketches Schematic faces

Nonface Stimuli

Scrambledphotographs

Butterflies(targets) Flowers Scrambled

schematic faces

Figure 4 Example of the eight stimulus types used in the present study

948 Journal of Cognitive Neuroscience Volume 13 Number 7

used The faces were presented in both upright andupside down orientation Upright stimuli were the samein Experiments 1 and 2

All stimuli were presented for 350 msec with anaverage interstimulus interval (ISI) of 12 sec and viewedfrom a distance of 110 plusmn 10 cm subtending a viewingangle of approximately 68 Stimulus timing was con-trolled by the MEL software system (Psychology Soft-ware Tools Pittsburgh PA)

Task

Participants were instructed to keep a mental countof stimuli belonging to one category butterflies inthe first experiment and flowers in the secondHence all of the experiment-relevant categories werefrom the participantsrsquo point of view nontargets Thetargets were presented in approximately 5 (Experi-ment 1) and 6 (Experiment 2) of the trials Duringeach break and at the end of the experimentparticipants reported the number of targets countedA few participants who did not count correctly wereexcluded

EEG Recording and Data Analysis

Continuous EEG was recorded from 48 scalp locationsusing tin electrodes attached to elastic electrode caps(Electrode Caps International Ohio) A schematic illus-tration of the electrode locations and names is pre-sented in Figure 1 A ground electrode was located onthe forehead and a reference electrode was located onthe tip of the nose Two additional electrodes located atthe outer canthus and the infraorbital region of the righteye measured vertical and horizontal electrooculograms(EOGs)

The EEG was sampled at 250 Hz by a DAP2400 DataAcquisition Processor (Microstar Laboratories) ampli-fied 20000-fold using a battery-operated HU-5072 Iso-lated Bioelectric Amplifier (SA Instrumentation SanDiego) with a frequency bandpass of 01ndash30 Hz andstored on disk for off-line averaging Electrode impe-dance was below 5 k A digital filter of 08ndash16 Hz wasapplied before analysis

Codes synchronized to stimulus delivery were used toselectively average epochs associated with differentstimulus categories Epochs contaminated with saccadesor eye blinks were excluded using root-mean-squareEOG-values as criteria Typically epochs associated with55ndash70 (never less than 50) out of 75 stimuli in eachcategory were averaged

The mean and peak amplitudes and peak latencies ofspecific ERP components within bounding intervals werecomputed for each category for each participant Sig-nificance of the differences between these measures wastested using a repeated-measures ANOVA Where appro-priate the more conservative GreenhousendashGeisser ad-

justed df values were used ANOVAs were followed byunivariate F contrasts

Acknowledgments

This study was supported by grants from the USndashIsraelBinational Science Foundation and Israeli Foundations Trust-ees to S Bentin We thank Amir Katz for his help in preparingthe stimuli and running Experiment 1 as well as Leon DeouellMichael Fink Ayelet Landau Yulia Goland David Carmel GilRaviv Amir Raz and Dan Tzur for their support We are alsovery grateful to Jerry Levinson Alex List and Lynn Robertsonfor their comments on the manuscript

Reprint requests should be sent to Shlomo Bentin TheHebrew University of Jerusalem Mount Scopus JerusalemIsrael 91905 or via e-mail msbentinmscchujiacil

Notes

1 Present address University of California Berkeley2 There was no reliable difference between the ERP elicited

by scrambled natural faces and scrambled schematic facesTherefore these two categories were collapsed into one3 These were the sites at which the N170 was most

conspicuous4 So were in fact all face categories except for the sketches5 Note that this result also contradicts the view that a more

difficult encoding process elicit N170 of a higher amplitude6 This assumption was indeed supported by a current study

(Bentin Mecklinger Bosch Sagiv amp von Cramon 1999)7 Of course some faces have very distinctive components

which can be used as markers for recognition Such markersare frequently exaggerated and used as distinctive features incaricatures (Rhodes 1987)8 Distinctive features (such as unusual components birth-

marks etc) may be stored in semantic memory as part of apersonrsquo s face entry Such information may have visual qualitiesthat can be addressed directly during visual matching tofacilitate identification9 Note that this account for the rsquorsquoface superiority effectrsquo rsquo is

identical to a common explanation of the rsquorsquoword superiorityeffectrsquo rsquo in the identification of single letters Indeed ourassumption about the coexistence of rsquo rsquowhole-facersquo rsquo and rsquo rsquoface-partrsquo rsquo processors resonates to the rsquorsquodual routersquo rsquo model of visualword processing It is quite possible that identical principlesgovern all the content-specific perceptual modules in the visualsystem10 Recall that relative to other face types the reduction of theN170 amplitude elicited by sketches at right hemisphere siteswas greater then at left hemisphere sites

REFERENCES

Allison T Ginter H McCarthy G Nobre A C Puce ALuby M amp Spencer D D (1994) Face recognition in hu-man extrastriate cortex Journal of Neurophysiology 71821ndash825

Allison T McCarthy G Nobre A C Puce A amp Belger A(1994) Human extrastriate visual cortex and the perceptionof faces words numbers and colors Cerebral Cortex 5544ndash554

Allison T Puce A Spencer D D amp McCarthy G (1999)Electrophysiological studies of human face perception I

Sagiv and Bentin 949

was implemented in a neural model initially based onscalp recordings (Bentin et al 1996) and more recentlycorrected improved and elaborated on the basis intra-cranial recordings (McCarthy et al 1999 for a detaileddescription and a review of supportive functional MRIdata see McCarthy 1999) This model suggests that theneural mechanism dedicated to structural encoding offaces is a complex system composed of at least twoseparate subsystems One probably located in thelateral part of the middle fusiform gyrus is responsiblefor processing the configuration of the face as a wholeand provides the integrated face representation that isnecessary for face recognition and identification Theother possibly distributed over posterior regions in theIT and middle temporal (MT) gyri and in the OTS isresponsible for detecting physiognomic informationregardless of the spatial configuration of the items thatcarry it The fusiform gyrus is farther away from therecording sites at the scalp than where the temporalgyri are and the orientation of its cells forms anelectrical dipole source that is almost perpendicular tothe lateral aspects of the scalp Hence although bothsources contribute to the modulation of the face-spe-cific N170 the contribution of the posterior temporalface component processor should be considerably moreconspicuous

According to this hypothesis the opposite effects ofinversion on the N170 elicited by natural and schematicfaces might be explained assuming that the holisticencoding system in the fusiform gyrus indeed inhibitsthe component analysis in the lateral temporal lobe facesites Hence if the face stimulus can trigger both theholistic and the component analysis (as it is true fornatural faces) an experimental manipulation that inter-feres with holistic processing (like the face inversionprocedure in the present experiment) should reducethe amount of inhibition of component analysis andconsequently the amplitude of the N170 would beenhanced However if the rsquo rsquocomponentsrsquorsquo do not conveyphysiognomic information so that their analysis by aface-specific mechanism rsquo rsquodependsrsquo rsquo on the previousrecognition of the face gestalt (as it is true for schematicfaces) interfering with holistic perception cannot facil-itate the activity of the component analysis and there-fore the modulation of the N170 would only reflect theadditional difficulty to encode the face ie reducedamplitude and delayed latency

The observed effects of the different factors on theN170 peak latency are nicely aligned with the abovemodel Although stimulus inversion delayed the N170for both natural and schematic faces this delay was twiceas big (12 msec) for former stimuli than for the latter (6msec) This interaction suggests that the inversion had abigger effect on natural than on schematic faces Thismay be because regardless of whether they are upright orinverted the physiognomic value of schematic faces isbased on their gestalt Hence inversion does inhibit the

processing of the gestalt but does not induce a change inprocessing strategy On the other hand natural faces areanalyzed holistically if they are upright but inversionrequires a change in that strategy encouraging theanalysis of components The significantly faster responseto upright natural than to upright schematic faces inExperiment 2 suggests that the extraction of the facegestalt is easier if the face is more natural

There is ample evidence indicating that the righthemisphere is specialized in global or holistic processingof objects and particularly faces whereas the left hemi-sphere is specialized in part-based processing (egRobertson amp Delis 1986 for a review see Bradshow ampNettleton 1983 for specific reference to face perceptionsee eg Bruce 1998 Rhodes et al 1993 Corballis1988 Bradshow amp Sherlock 1982 Gilbert amp Bakan1973) Consequently interfering with holistic processingshould have had stronger effects in the right than in theleft hemisphere Our results confirmed this predictiononly partially The pronounced interhemispheric asym-metry with larger N170 amplitudes at the right than atthe left hemisphere sites for all stimulus types suggests apredominant holistic process for all face categoriesInterestingly this asymmetry was reduced (in Experi-ment 1) for sketches relative to the other stimulus typesThe relatively enhanced involvement of the left hemi-sphere in processing the sketches is in agreement withour suggestion that the face-specific structural encodingsystem was activated less efficiently by sketches than byother face types10 Hence the same factor that may haveconstrained the processing of schematic faces may alsohave constrained the processing of sketches but in theopposite direction Whereas schematic faces enhancedholistic processing by providing a clear face gestalt whileobscuring the physiognomic value of face componentsthe sketches reduced holistic processing by blurring orveiling the face gestalt Also in Experiment 2 the N170was larger over right than over left hemisphere sitesHowever because the interaction between the inversionand hemisphere effects in Experiment 2 was not modu-lated by stimulus type (ie the three-way interactionwas not significant) our interpretation of the interhemi-spheric asymmetry (particularly in Experiment 2) mustbe limited

In conclusion the present scalp-recorded N170 dataconverge with the previously published pattern of theintracranially recorded N200 data (for a comprehensivereview see Allison et al 1999 McCarthy et al 1999Puce et al 1999) supporting a multifunctional organ-ization of the structural encoding mechanism for facesApparently distinct neural circuits are responsible forthe encoding and initial analysis of face components andfor holistic perception of a full-face configuration Theformer is probably located in the lateral posterior tem-poral lobe whereas the latter is located in the middlefusiform gyrus Face perception usually entails predom-inantly holistic processing which explains the right

Sagiv and Bentin 947

hemisphere dominance in performance as well as inelectrophysiological and neuroimaging data Howeverboth mechanisms can be triggered independently byrelevant information

METHODS

Participants

Right-handed participants were recruited among volun-teer students from the Hebrew University They receivedeither experimental credit hours or payment for theirparticipation Thirty participants (13 males and 17 fe-males) were tested in the first experiment and 16participants (8 males and 8 females) were tested in thesecond experiment Participantsrsquo ages ranged from 19 to37 years (mean = 233 years) They all had normal orcorrected-to-normal vision

Stimuli

Stimuli were scanned using a desk scanner processedwith commercially available graphic software and con-

verted to 16 color grayscale 248 by 248 pixel imagesThere were eight stimulus categories four face types(photographs of natural faces paintings face sketchesand schematic faces) two nonsense patterns (formedby scrambling the natural faces and the schematicfaces) flowers and butterflies (Figure 4) Paintingsand sketches of faces were scanned from art booksPaintings were typically realistic Renaissance portraitsand face sketches did not portray details and usuallylacked shading and texture Schematic face parts weredrawn by hand scanned into the computer and latercombined into different schematic faces Face photo-graphs and paintings contained many shades of grayand the sketches and the schematic faces were basicallyblack on white All face stimuli were front views

Scrambled faces were created by scrambling thephases of the two-dimensional Fourier transform andthen smoothing the uneven edges created by this proc-ess The original and the scrambled stimuli were equallyluminous following this procedure Each stimulus cat-egory comprised 75 stimuli except for the butterflycategory which consisted of 30 stimuli In Experiment2 only natural faces schematic faces and flowers were

Face Stimuli

Photographs Painted portraits Sketches Schematic faces

Nonface Stimuli

Scrambledphotographs

Butterflies(targets) Flowers Scrambled

schematic faces

Figure 4 Example of the eight stimulus types used in the present study

948 Journal of Cognitive Neuroscience Volume 13 Number 7

used The faces were presented in both upright andupside down orientation Upright stimuli were the samein Experiments 1 and 2

All stimuli were presented for 350 msec with anaverage interstimulus interval (ISI) of 12 sec and viewedfrom a distance of 110 plusmn 10 cm subtending a viewingangle of approximately 68 Stimulus timing was con-trolled by the MEL software system (Psychology Soft-ware Tools Pittsburgh PA)

Task

Participants were instructed to keep a mental countof stimuli belonging to one category butterflies inthe first experiment and flowers in the secondHence all of the experiment-relevant categories werefrom the participantsrsquo point of view nontargets Thetargets were presented in approximately 5 (Experi-ment 1) and 6 (Experiment 2) of the trials Duringeach break and at the end of the experimentparticipants reported the number of targets countedA few participants who did not count correctly wereexcluded

EEG Recording and Data Analysis

Continuous EEG was recorded from 48 scalp locationsusing tin electrodes attached to elastic electrode caps(Electrode Caps International Ohio) A schematic illus-tration of the electrode locations and names is pre-sented in Figure 1 A ground electrode was located onthe forehead and a reference electrode was located onthe tip of the nose Two additional electrodes located atthe outer canthus and the infraorbital region of the righteye measured vertical and horizontal electrooculograms(EOGs)

The EEG was sampled at 250 Hz by a DAP2400 DataAcquisition Processor (Microstar Laboratories) ampli-fied 20000-fold using a battery-operated HU-5072 Iso-lated Bioelectric Amplifier (SA Instrumentation SanDiego) with a frequency bandpass of 01ndash30 Hz andstored on disk for off-line averaging Electrode impe-dance was below 5 k A digital filter of 08ndash16 Hz wasapplied before analysis

Codes synchronized to stimulus delivery were used toselectively average epochs associated with differentstimulus categories Epochs contaminated with saccadesor eye blinks were excluded using root-mean-squareEOG-values as criteria Typically epochs associated with55ndash70 (never less than 50) out of 75 stimuli in eachcategory were averaged

The mean and peak amplitudes and peak latencies ofspecific ERP components within bounding intervals werecomputed for each category for each participant Sig-nificance of the differences between these measures wastested using a repeated-measures ANOVA Where appro-priate the more conservative GreenhousendashGeisser ad-

justed df values were used ANOVAs were followed byunivariate F contrasts

Acknowledgments

This study was supported by grants from the USndashIsraelBinational Science Foundation and Israeli Foundations Trust-ees to S Bentin We thank Amir Katz for his help in preparingthe stimuli and running Experiment 1 as well as Leon DeouellMichael Fink Ayelet Landau Yulia Goland David Carmel GilRaviv Amir Raz and Dan Tzur for their support We are alsovery grateful to Jerry Levinson Alex List and Lynn Robertsonfor their comments on the manuscript

Reprint requests should be sent to Shlomo Bentin TheHebrew University of Jerusalem Mount Scopus JerusalemIsrael 91905 or via e-mail msbentinmscchujiacil

Notes

1 Present address University of California Berkeley2 There was no reliable difference between the ERP elicited

by scrambled natural faces and scrambled schematic facesTherefore these two categories were collapsed into one3 These were the sites at which the N170 was most

conspicuous4 So were in fact all face categories except for the sketches5 Note that this result also contradicts the view that a more

difficult encoding process elicit N170 of a higher amplitude6 This assumption was indeed supported by a current study

(Bentin Mecklinger Bosch Sagiv amp von Cramon 1999)7 Of course some faces have very distinctive components

which can be used as markers for recognition Such markersare frequently exaggerated and used as distinctive features incaricatures (Rhodes 1987)8 Distinctive features (such as unusual components birth-

marks etc) may be stored in semantic memory as part of apersonrsquo s face entry Such information may have visual qualitiesthat can be addressed directly during visual matching tofacilitate identification9 Note that this account for the rsquorsquoface superiority effectrsquo rsquo is

identical to a common explanation of the rsquorsquoword superiorityeffectrsquo rsquo in the identification of single letters Indeed ourassumption about the coexistence of rsquo rsquowhole-facersquo rsquo and rsquo rsquoface-partrsquo rsquo processors resonates to the rsquorsquodual routersquo rsquo model of visualword processing It is quite possible that identical principlesgovern all the content-specific perceptual modules in the visualsystem10 Recall that relative to other face types the reduction of theN170 amplitude elicited by sketches at right hemisphere siteswas greater then at left hemisphere sites

REFERENCES

Allison T Ginter H McCarthy G Nobre A C Puce ALuby M amp Spencer D D (1994) Face recognition in hu-man extrastriate cortex Journal of Neurophysiology 71821ndash825

Allison T McCarthy G Nobre A C Puce A amp Belger A(1994) Human extrastriate visual cortex and the perceptionof faces words numbers and colors Cerebral Cortex 5544ndash554

Allison T Puce A Spencer D D amp McCarthy G (1999)Electrophysiological studies of human face perception I

Sagiv and Bentin 949

hemisphere dominance in performance as well as inelectrophysiological and neuroimaging data Howeverboth mechanisms can be triggered independently byrelevant information

METHODS

Participants

Right-handed participants were recruited among volun-teer students from the Hebrew University They receivedeither experimental credit hours or payment for theirparticipation Thirty participants (13 males and 17 fe-males) were tested in the first experiment and 16participants (8 males and 8 females) were tested in thesecond experiment Participantsrsquo ages ranged from 19 to37 years (mean = 233 years) They all had normal orcorrected-to-normal vision

Stimuli

Stimuli were scanned using a desk scanner processedwith commercially available graphic software and con-

verted to 16 color grayscale 248 by 248 pixel imagesThere were eight stimulus categories four face types(photographs of natural faces paintings face sketchesand schematic faces) two nonsense patterns (formedby scrambling the natural faces and the schematicfaces) flowers and butterflies (Figure 4) Paintingsand sketches of faces were scanned from art booksPaintings were typically realistic Renaissance portraitsand face sketches did not portray details and usuallylacked shading and texture Schematic face parts weredrawn by hand scanned into the computer and latercombined into different schematic faces Face photo-graphs and paintings contained many shades of grayand the sketches and the schematic faces were basicallyblack on white All face stimuli were front views

Scrambled faces were created by scrambling thephases of the two-dimensional Fourier transform andthen smoothing the uneven edges created by this proc-ess The original and the scrambled stimuli were equallyluminous following this procedure Each stimulus cat-egory comprised 75 stimuli except for the butterflycategory which consisted of 30 stimuli In Experiment2 only natural faces schematic faces and flowers were

Face Stimuli

Photographs Painted portraits Sketches Schematic faces

Nonface Stimuli

Scrambledphotographs

Butterflies(targets) Flowers Scrambled

schematic faces

Figure 4 Example of the eight stimulus types used in the present study

948 Journal of Cognitive Neuroscience Volume 13 Number 7

used The faces were presented in both upright andupside down orientation Upright stimuli were the samein Experiments 1 and 2

All stimuli were presented for 350 msec with anaverage interstimulus interval (ISI) of 12 sec and viewedfrom a distance of 110 plusmn 10 cm subtending a viewingangle of approximately 68 Stimulus timing was con-trolled by the MEL software system (Psychology Soft-ware Tools Pittsburgh PA)

Task

Participants were instructed to keep a mental countof stimuli belonging to one category butterflies inthe first experiment and flowers in the secondHence all of the experiment-relevant categories werefrom the participantsrsquo point of view nontargets Thetargets were presented in approximately 5 (Experi-ment 1) and 6 (Experiment 2) of the trials Duringeach break and at the end of the experimentparticipants reported the number of targets countedA few participants who did not count correctly wereexcluded

EEG Recording and Data Analysis

Continuous EEG was recorded from 48 scalp locationsusing tin electrodes attached to elastic electrode caps(Electrode Caps International Ohio) A schematic illus-tration of the electrode locations and names is pre-sented in Figure 1 A ground electrode was located onthe forehead and a reference electrode was located onthe tip of the nose Two additional electrodes located atthe outer canthus and the infraorbital region of the righteye measured vertical and horizontal electrooculograms(EOGs)

The EEG was sampled at 250 Hz by a DAP2400 DataAcquisition Processor (Microstar Laboratories) ampli-fied 20000-fold using a battery-operated HU-5072 Iso-lated Bioelectric Amplifier (SA Instrumentation SanDiego) with a frequency bandpass of 01ndash30 Hz andstored on disk for off-line averaging Electrode impe-dance was below 5 k A digital filter of 08ndash16 Hz wasapplied before analysis

Codes synchronized to stimulus delivery were used toselectively average epochs associated with differentstimulus categories Epochs contaminated with saccadesor eye blinks were excluded using root-mean-squareEOG-values as criteria Typically epochs associated with55ndash70 (never less than 50) out of 75 stimuli in eachcategory were averaged

The mean and peak amplitudes and peak latencies ofspecific ERP components within bounding intervals werecomputed for each category for each participant Sig-nificance of the differences between these measures wastested using a repeated-measures ANOVA Where appro-priate the more conservative GreenhousendashGeisser ad-

justed df values were used ANOVAs were followed byunivariate F contrasts

Acknowledgments

This study was supported by grants from the USndashIsraelBinational Science Foundation and Israeli Foundations Trust-ees to S Bentin We thank Amir Katz for his help in preparingthe stimuli and running Experiment 1 as well as Leon DeouellMichael Fink Ayelet Landau Yulia Goland David Carmel GilRaviv Amir Raz and Dan Tzur for their support We are alsovery grateful to Jerry Levinson Alex List and Lynn Robertsonfor their comments on the manuscript

Reprint requests should be sent to Shlomo Bentin TheHebrew University of Jerusalem Mount Scopus JerusalemIsrael 91905 or via e-mail msbentinmscchujiacil

Notes

1 Present address University of California Berkeley2 There was no reliable difference between the ERP elicited

by scrambled natural faces and scrambled schematic facesTherefore these two categories were collapsed into one3 These were the sites at which the N170 was most

conspicuous4 So were in fact all face categories except for the sketches5 Note that this result also contradicts the view that a more

difficult encoding process elicit N170 of a higher amplitude6 This assumption was indeed supported by a current study

(Bentin Mecklinger Bosch Sagiv amp von Cramon 1999)7 Of course some faces have very distinctive components

which can be used as markers for recognition Such markersare frequently exaggerated and used as distinctive features incaricatures (Rhodes 1987)8 Distinctive features (such as unusual components birth-

marks etc) may be stored in semantic memory as part of apersonrsquo s face entry Such information may have visual qualitiesthat can be addressed directly during visual matching tofacilitate identification9 Note that this account for the rsquorsquoface superiority effectrsquo rsquo is

identical to a common explanation of the rsquorsquoword superiorityeffectrsquo rsquo in the identification of single letters Indeed ourassumption about the coexistence of rsquo rsquowhole-facersquo rsquo and rsquo rsquoface-partrsquo rsquo processors resonates to the rsquorsquodual routersquo rsquo model of visualword processing It is quite possible that identical principlesgovern all the content-specific perceptual modules in the visualsystem10 Recall that relative to other face types the reduction of theN170 amplitude elicited by sketches at right hemisphere siteswas greater then at left hemisphere sites

REFERENCES

Allison T Ginter H McCarthy G Nobre A C Puce ALuby M amp Spencer D D (1994) Face recognition in hu-man extrastriate cortex Journal of Neurophysiology 71821ndash825

Allison T McCarthy G Nobre A C Puce A amp Belger A(1994) Human extrastriate visual cortex and the perceptionof faces words numbers and colors Cerebral Cortex 5544ndash554

Allison T Puce A Spencer D D amp McCarthy G (1999)Electrophysiological studies of human face perception I

Sagiv and Bentin 949

used The faces were presented in both upright andupside down orientation Upright stimuli were the samein Experiments 1 and 2

All stimuli were presented for 350 msec with anaverage interstimulus interval (ISI) of 12 sec and viewedfrom a distance of 110 plusmn 10 cm subtending a viewingangle of approximately 68 Stimulus timing was con-trolled by the MEL software system (Psychology Soft-ware Tools Pittsburgh PA)

Task

Participants were instructed to keep a mental countof stimuli belonging to one category butterflies inthe first experiment and flowers in the secondHence all of the experiment-relevant categories werefrom the participantsrsquo point of view nontargets Thetargets were presented in approximately 5 (Experi-ment 1) and 6 (Experiment 2) of the trials Duringeach break and at the end of the experimentparticipants reported the number of targets countedA few participants who did not count correctly wereexcluded

EEG Recording and Data Analysis

Continuous EEG was recorded from 48 scalp locationsusing tin electrodes attached to elastic electrode caps(Electrode Caps International Ohio) A schematic illus-tration of the electrode locations and names is pre-sented in Figure 1 A ground electrode was located onthe forehead and a reference electrode was located onthe tip of the nose Two additional electrodes located atthe outer canthus and the infraorbital region of the righteye measured vertical and horizontal electrooculograms(EOGs)

The EEG was sampled at 250 Hz by a DAP2400 DataAcquisition Processor (Microstar Laboratories) ampli-fied 20000-fold using a battery-operated HU-5072 Iso-lated Bioelectric Amplifier (SA Instrumentation SanDiego) with a frequency bandpass of 01ndash30 Hz andstored on disk for off-line averaging Electrode impe-dance was below 5 k A digital filter of 08ndash16 Hz wasapplied before analysis

Codes synchronized to stimulus delivery were used toselectively average epochs associated with differentstimulus categories Epochs contaminated with saccadesor eye blinks were excluded using root-mean-squareEOG-values as criteria Typically epochs associated with55ndash70 (never less than 50) out of 75 stimuli in eachcategory were averaged

The mean and peak amplitudes and peak latencies ofspecific ERP components within bounding intervals werecomputed for each category for each participant Sig-nificance of the differences between these measures wastested using a repeated-measures ANOVA Where appro-priate the more conservative GreenhousendashGeisser ad-

justed df values were used ANOVAs were followed byunivariate F contrasts

Acknowledgments

This study was supported by grants from the USndashIsraelBinational Science Foundation and Israeli Foundations Trust-ees to S Bentin We thank Amir Katz for his help in preparingthe stimuli and running Experiment 1 as well as Leon DeouellMichael Fink Ayelet Landau Yulia Goland David Carmel GilRaviv Amir Raz and Dan Tzur for their support We are alsovery grateful to Jerry Levinson Alex List and Lynn Robertsonfor their comments on the manuscript

Reprint requests should be sent to Shlomo Bentin TheHebrew University of Jerusalem Mount Scopus JerusalemIsrael 91905 or via e-mail msbentinmscchujiacil

Notes

1 Present address University of California Berkeley2 There was no reliable difference between the ERP elicited

by scrambled natural faces and scrambled schematic facesTherefore these two categories were collapsed into one3 These were the sites at which the N170 was most

conspicuous4 So were in fact all face categories except for the sketches5 Note that this result also contradicts the view that a more

difficult encoding process elicit N170 of a higher amplitude6 This assumption was indeed supported by a current study

(Bentin Mecklinger Bosch Sagiv amp von Cramon 1999)7 Of course some faces have very distinctive components

which can be used as markers for recognition Such markersare frequently exaggerated and used as distinctive features incaricatures (Rhodes 1987)8 Distinctive features (such as unusual components birth-

marks etc) may be stored in semantic memory as part of apersonrsquo s face entry Such information may have visual qualitiesthat can be addressed directly during visual matching tofacilitate identification9 Note that this account for the rsquorsquoface superiority effectrsquo rsquo is

identical to a common explanation of the rsquorsquoword superiorityeffectrsquo rsquo in the identification of single letters Indeed ourassumption about the coexistence of rsquo rsquowhole-facersquo rsquo and rsquo rsquoface-partrsquo rsquo processors resonates to the rsquorsquodual routersquo rsquo model of visualword processing It is quite possible that identical principlesgovern all the content-specific perceptual modules in the visualsystem10 Recall that relative to other face types the reduction of theN170 amplitude elicited by sketches at right hemisphere siteswas greater then at left hemisphere sites

REFERENCES

Allison T Ginter H McCarthy G Nobre A C Puce ALuby M amp Spencer D D (1994) Face recognition in hu-man extrastriate cortex Journal of Neurophysiology 71821ndash825

Allison T McCarthy G Nobre A C Puce A amp Belger A(1994) Human extrastriate visual cortex and the perceptionof faces words numbers and colors Cerebral Cortex 5544ndash554

Allison T Puce A Spencer D D amp McCarthy G (1999)Electrophysiological studies of human face perception I

Sagiv and Bentin 949

Potentials generated in occipitotemporal cortex by face andnon-face stimuli Cerebral Cortex 9 416ndash430

Bentin S Allison T Puce A Perez A amp McCarthy G(1996) Electrophysiological studies of face perception inhumans Journal of Cognitive Neuroscience 8 551ndash565

Bentin S amp Deouell L Y (2000) Structural encoding andidentification in face processing ERP evidence for separatemechanisms Cognitive Neuropsychology 17 35ndash54

Bentin S Deouell L Y amp Soroker N (1999) Selective visualstreaming in face recognition Evidence from developmentalprosopagnosia NeuroReport 10(4) 823ndash827

Bentin S amp Mann V A (1990) Masking and stimulus inten-sity effects on duplex perception A confirmation of thedissociation between speech and non-speech modes Jour-nal of Acoustical Society of America 88 64ndash74

Bentin S Mecklinger A Bosch V Sagiv N amp von CramonD Y (1999) In the eyes of the beholder Context-inducedactivation of face-specific perceptual mechanisms Sixth An-nual Meeting of the Cognitive Neuroscience Society April11ndash13 Washington DC

Biderman I (1987) Recognition-by-components A theory ofhuman image interpretation Psychological Review 94 115ndash147

Bradshow J L amp Nettleton N C (1983) Human cerebralasymmetry Englewood Cliffs NJ Prentice-Hall

Bradshow J L amp Sherlock D (1982) Bugs and faces in thetwo visual fields Task order difficulty practice and analyticholistic dichotomy Cortex 18 211ndash226

Bruce V amp Young A (1986) Understanding face recognitionBritish Journal of Psychology 77 305ndash327

Bruce V amp Young A (1998) In the eyes of the beholder Thescience of face perception Oxford Oxford University Press

Corballis M C (1988) Recognition of disoriented shapesPsychological Review 95 115ndash123

De Renzi E (1997) Prosopagnosia In T E Feinberg amp M JFarah (Eds) Behavioral neurology and neuropsychology(pp 245ndash256) New York McGraw-Hill

De Renzi E Perani D Carlesimo G A Silveri M C amp FazioF (1994) Prosopagnosia can be associated with damageconfined to the right hemisphere An MRI and PET study anda review of the literature Neuropsychologia 32 893ndash902

Desimone R (1991) Face-selective cells in the temporal cor-tex of monkeys Journal of Cognitive Neuroscience 3 1ndash8

Farah M J (1990) Visual agnosia Cambridge MIT PressFarah M J (1996) Is face recognition rsquospecialrsquo Evidence

from neuropsychology Behavioral Brain Research 76181ndash189

Farah M J Tanaka J W amp Drain H M (1995) What causesthe face inversion effect Journal of Experimental Psychol-ogy Human Perception and Performance 21 628ndash634

Farah M J Wilson K D Drain M amp Tanaka J N (1998)What is rsquo rsquo specialrsquo rsquo about face perception Psychological Re-view 105 482ndash498

George N Evans J Fiori N Davidoff J amp Renault B (1996)Brain events related to normal and moderately scrambledfaces Cognitive Brain Research 4 65ndash76

Gilbert C amp Bakan P (1973) Visual asymmetry in perceptionof faces Neuropsychologia 11 355ndash362

Haxby J V Grady C L Horwitz B Salerno J UngerleiderL G Mishkin M amp Schapiro M B (1993) Dissociation ofobject and spatial visual processing pathways in human ex-trastriate cortex In B Gulyas D Ottoson amp P E Roland(Eds) Functional organization of the human visual cortex(pp 329ndash340) Oxford Pergamon

Ishai A Ungerlieder L G Martin A Schouten J L amp HaxbyJ V (1999) Distributed representation of objects in thehuman ventral visual pathway Proceedings of the NationalAcademy of Sciences USA 96 9379ndash9384

Jeffreys D A amp Tukmachi E S A (1992) The vertex-positivescalp potential evoked by faces and by objects Experimen-tal Brain Research 91 340ndash350

Kanwisher N McDermott J amp Chun M M (1997) The fu-siform face area A module in human extrastriate cortexspecialized for face perception Journal of Neuroscience 174302ndash4311

Kanwisher N Tong F amp Nakayama K (1998) The effect offace inversion on the human fusiform face area Cognition68 1ndash11

Liberman A M amp Mattingly I G (1989) A specialization forspeech perception Science 243 489ndash494

Logothesis N K amp Pauls J (1995) Psychophysical and phy-siological evidence for viewer-centered object representa-tion in the primate Cerebral Cortex 5 270ndash288

Logothesis N K Pauls J amp Poggio T (1995) Shape repre-sentation in the inferior temporal cortex of monkeys Cur-rent Biology 5 552ndash563

Logothetis N K amp Scheinberg D L (1996) Visual objectrecognition Annual Reviews of Neuroscience 19 577ndash621

Luh K E Rueckert L M amp Levy J (1991) Perceptualasymmetries for free viewing of several types of chimericstimuli Brain and Cognition 16 83ndash103

Marr D (1982) Vision San Francisco FreemanMarr D amp Nishihara H K (1978) Representation and re-

cognition of 3-dimensional shapes Proceedings of the RoyalSociety of London Series B 200 269ndash294

McCarthy G (1999) Physiological studies of face processing inhumans In M S Gazzaniga (Ed) The new cognitive neu-rosciences (chap 28 pp 393ndash409) Cambridge MIT Press

McCarthy G Puce A Belger A amp Allison T (1999) Elec-trophysiological studies of human face perception II Re-sponse properties of face-specific potentials generated inoccipitotemporal cortex Cerebral Cortex 9 431ndash444

McCarthy G Puce A Gore J C amp Allison T (1997) Face-specific processing in the human fusiform gyrus Journal ofCognitive Neuroscience 9 605ndash610

McClelland J L (1979) On the time-relations of mental pro-cesses An examination of systems of processes in cascadePsychological Review 86 287ndash330

Moscovitch M Winocur G amp Behrmann M (1997) What isspecial about face recognition Nineteen experiments on aperson with visual object agnosia and dyslexia but normal facerecognition Journal of Cognitive Neuroscience 9 555ndash604

Moses Y Adini Y amp Ullman S (1994) Face recognition Theproblem of compensating for illumination changes Pro-ceedings of the European Conference on Compute Vision286ndash296

Moses Y Edelman S amp Ullman S (1993) Generalization tonovel images in upright and inverted faces The WeizmannInstitute of Science Technical Report CS93-14

Puce A Allison T Gore J C amp McCarthy G (1995) Face-sensitive regions in human extrastriate cortex studied byfunctional MRI Journal of Neurophysiology 74 1192ndash1199

Puce A Allison T amp McCarthy G (1999) Electrophysiolo-gical studies of human face perception III Effects of top-down processing on face-specific potentials CerebralCortex 9 445ndash458

Rhodes G (1995) Face recognition and configural coding InT Valentine (Ed) Cognitive and computational aspects offace recognition London Rutledge

Rhodes G Brake S amp Atkinson A (1993) Whatrsquos lost ininverted faces Cognition 47 25ndash57

Robertson L C amp Delis D C (1986) rsquoPartndashwholersquo processingin unilateral brain-damaged patients Dysfunction of hier-archical organization Neuropsychologia 24 363ndash370

Rossion B Delvenne J -F Debatisse D Goffaux VBruyer R Crommelinck M amp Guerit J -M (1999) Spa-

950 Journal of Cognitive Neuroscience Volume 13 Number 7

tio-temporal localization of the face inversion effect Anevent-related potentials study Biological Psychology 50173ndash189

Rossion B Gauthier I Tarr M J Despland P Bruyer R ampCrommenlink M (2000) The N170 occipito-temporal com-ponent is delayed and enhanced to inverted faces but not toinverted objects An electrophysiological account of specificprocesses in the human brain NeuroReport 11 69ndash74

Sergent J (1984) An investigation into component and con-figural processes underlying face perception British Journalof Psychology 75 221ndash242

Sergent J Ohta S amp MacDonald B (1992) Functionalneuroanatomy of face and object processing A positronemission tomography study Brain 115 15ndash36

Tanaka J W amp Farah M J (1991) Second-order relational

properties and the inversion-effect Testing a theory of faceperception Perception and Psychophysics 50 367ndash372

Tanaka J W amp Farah M J (1993) Parts and wholes in facerecognition Quarterly Journal of Experimental Psychol-ogy 46A 225ndash245

Ullman S (1996) High-level vision Object recognition andvisual cognition Cambridge MIT Press

Valentine T (1988) Upside down faces A review of the effectof inversion upon face recognition British Journal of Ex-perimental Psychology 87 116ndash124

Whalen D H amp Liberman A M (1987) Speech perceptiontakes precedence over nonspeech perception Science 237169ndash171

Young A W Hellawell D J amp Hay D C (1987) Configuralinformation in face perception Perception 16 747ndash759

Sagiv and Bentin 951