764 nature neuroscience • volume 3 no 8 • august 2000

commentary

How does the primate visual system processand identify objects? Central to this ques-tion is that visual recognition can occur atmany levels, from coarse categories, forexample “a car,” to highly specific individ-uals, “my dad’s 1960 TR3 convertible” (sub-ordinate-level processing). The processingdemands exerted by these two tasks seemto require separable visual recognition sys-tems, one dedicated to determining cate-gory membership and one dedicated toindividuation. The most conspicuous ver-sion of this argument is one in which thereis a functional and neuroanatomical divi-sion between the mechanisms supportingobject at the basic level and face recogni-tion at the individual level. Here we evalu-ate this claim and ask whether putativelyface-specific mechanisms in a region of thevisual cortex known as the Fusiform FaceArea (“FFA”) are selective specifically to thegeometry of faces or whether, through theoperation of other factors, they extend tonon-face stimuli.

The comparison between faces and non-face objects is critical for evaluating threecompeting models. First, Kanwisher (thisissue) suggests that face selectivity in the FFAas revealed by functional magnetic reso-nance imaging (fMRI) reflects a special-pur-pose, innate mechanism for detecting thegeometry of faces, that is, a mechanism thatdetermines only that a face is a face, but notthe individual identity of the face1–3. Evidence for this ‘detection model’ comesfrom studies in which selectivity for faces inFFA has been established over a wide vari-

similarity alone seems unlikely to accountfor category selectivity, in that task manip-ulations independent of image geometrycan lead to activation maps indistinguish-able from those produced by manipulatinggeometry.

Third, our preferred model, the flexibleprocess map19, holds that an assortment ofextrastriate areas support separable com-ponents of visual object processing andrecognition. This model is distinguishedfrom the detection and feature-map mod-els in that specialization of large extrastriateareas (that is, those detectable with fMRI)is not a priori a function of stimulus geom-etry. Rather, through experience, observersassociate particular geometries—whichdefine object categories—with task-appro-priate recognition strategies that automati-cally recruit components of the process map.In contrast to other accounts, the process-map model provides a general explanationfor category selectivity, as well as why theFFA responds to a wide range of stimuli,including faces and non-face objects.

Factors affecting FFA processingGiven that face selectivity may not beintrinsically related to the specific geome-try of faces, what else might account for thestrong activation for faces in the FFA?Behavioral work20–22 suggests two criticalfactors associated with face recognitionbecause of our particular experience withfaces. First, the level of categorization atwhich objects are recognized is different—faces generally being recognized at the indi-vidual level, but objects such as chairs orbirds being recognized at a more categoricallevel. Second, we have far greater expertisewith faces as compared to almost any othervisual category. Expertise with faces mayautomatize face processing at the individ-ual level, thereby rendering task manipula-tions less effective for faces, for example, therapid presentation rates used in some stud-

ety of conditions, including line drawings,two-tone, gray-scale or colored images—stimulus manipulations that affect face-recognition performance but have littleimpact on activation in FFA1,4,5.

For example, face inversion seems toonly minimally influence activity in theFFA1,6,7. This result appears at odds with thestrong effect of inversion on face recogni-tion performance8. However, inversion onlyimpairs, but does not abolish, the percep-tion of identity. Inverted faces are typicallyrecognized well above chance level, whichis consistent with the small but significantfMRI face inversion effects reported insome studies1,9. Moreover, this same smallinversion effect has been replicated withnon-face objects in expert subjects9. Theabsence or relatively small inversion effectin the FFA may also mean that activity asmeasured by fMRI simply is not sufficient-ly fine-grained to pick up certain effects. Asa case in point, fMRI cannot detect therobust and early effect of face inversionreported in ERPs: a ten-millisecond delayin the N170 face-sensitive potential10,11 thatis recorded at temporo-occipital elec-trodes12,13 or a similar inversion effect inintracranial recordings14.

A second model holds that categoryselectivity reflects the presence of a geo-metrically defined feature map in ventraltemporal cortex15,16. The hypothesis is thatsmall, localized brain regions are activatedby objects, generally from a single catego-ry, with similar shape and image features.Evidence for feature maps comes frommonkey neurophysiology17 suggesting atopography of features in inferior tempo-ral cortex (IT) and from human fMRI stud-ies indicating that across a single task,different stimuli selectively activate differ-ent regions of the ventral temporal cortex.For instance, distinct brain areas respondpreferentially to letter strings, chairs, build-ings or faces4,6,16,18. However, geometric

FFA: a flexible fusiform area forsubordinate-level visual processingautomatized by expertiseMichael J. Tarr and Isabel Gauthier

Much evidence suggests that the fusiform face area is involved in face processing. In contrastto the accompanying article by Kanwisher, we conclude that the apparent face selectivity ofthis area reflects a more generalized form of processing not intrinsically specific to faces.

Michael Tarr is at the Department of Cognitiveand Linguistic Sciences, Brown University, Box1978, Providence, Rhode Island 02912, USA.Isabel Gauthier is at the Department ofPsychology, Vanderbilt University, 301 WilsonHall, Nashville, Tennessee 37240, USA.e-mail: isabel.gauthier@vanderbilt.edu orMichael_Tarr@brown.edu

© 2000 Nature America Inc. • http://neurosci.nature.com©

200

0 N

atur

e A

mer

ica

Inc.

• h

ttp:

//neu

rosc

i.nat

ure.

com

nature neuroscience • volume 3 no 8 • august 2000 765

ies (such as 300 ms with a 500 ms ISI5) maynot affect face processing nearly as much asobject processing.

Because extreme values for these twodimensions are typically associated withface recognition, any comparison betweenfaces and non-face objects confounds thesefactors unless they are independentlymanipulated. When we do such manipula-tions, it is the level of categorization and thelevel of expertise, not the stimulus geome-try, that are critical in determining selectiv-ity in the FFA. Given such domain-generalprocesses, we argue that the FFA may bebetter thought of as a ‘flexible fusiformarea’—a mechanism for visual recognitionthat exhibits a great deal of plasticity inresponse to both task demands and expe-rience.

Evidence from fMRIFunctional MRI provides an ideal tool forevaluating these models because brainareas can be identified in a functional man-ner. In comparison to cytoarchitectonic cri-teria (like Brodmann areas) or grossanatomical landmarks (such as theTalairach system), a region performing thesame function in different individuals canbe defined, allowing for variability in themapping of function to anatomy. Buildingon a long history of neuropsychologicalfindings indicating specialization in the pri-mate visual system for face processing20,both PET and fMRI studies have localizedface-selective regions in the ventral tem-poral lobe4,18,23,24. Functional MRI has beenused to identify a small number of face-selective voxels—the FFA—through a com-parison of faces and non-face objects in acommon task4. Importantly, this neuralpattern replicates within the same studyand generalizes across experimental con-ditions. That is, candidate face-selectivevoxels are recruited more for faces than fornon-face objects even when the task andstimuli are varied. These results from imag-ing seem to support earlier accounts of faceselectivity in the brain (and provide amethod of localizing it). Different inter-pretations are possible, however, when cat-egorization level and experience are takeninto account.

To specifically address the roles of thesetwo factors, several studies have manipulat-ed the level of categorization and the levelof expertise for non-face objects. These stud-ies find that the response of FFA to non-faceobjects is affected by these manipulationsindependent of particular stimulus geome-tries25–27. Logically then, these others factorscan account for most of the strong faceselectivity in FFA activity.

Greebles never shown during training. Suchexpertise effects indicate a high degree offlexibility with regard to acceptable imagegeometries in the neural network mediat-ing object recognition. Moreover, this flexi-bility does not have a critical period and canbe harnessed for relatively novel learningexperiences (such as learning to read brainscans or individuate nonsense objects suchas Greebles).

Automatization by expertiseObject processing in the FFA can be autom-atized by expertise. Consider that Greebleexperts show FFA activation in response toGreebles even under passive viewinginstructions9. Moreover, whereas novicescan easily focus on the top half of a com-posite Greeble (made out of the top of oneand the bottom of another), Greebleexperts cannot ignore the bottom distrac-tor part. The behavioral changes measuredby this ‘composite effect’ show a significantcorrelation across subjects and training ses-sions with activity in the right FFA of Gree-ble experts (I.G. and M.J.T., unpublisheddata). Thus, neural changes with expertise,at least in the right FFA, seem to be tiedstrongly to holistic processing strategies.This is consistent with a PET study32 report-ing that the right FFA is more active whenobservers attend to whole faces than faceparts, whereas the opposite pattern is foundin the left FFA.

An fMRI study of bird and car expertsfurther supports the idea that expertiseautomatizes processing in the right FFA25.In novices, the homogeneous classes of birdsand cars elicit more activity in the FFA thana set of various familiar objects, and this

Manipulating categorization-levelThe FFA bilaterally shows an increasedresponse when familiar artifactual and nat-ural kind objects are seen in the context ofvisually similar stimuli25. The FFA shows asimilar response for animals, even whentheir faces are obscured28 (but see ref. 5).Moreover, when the level of categorization ismanipulated by different labels that subjectsmatch to the same image (for example,‘TR3’ versus ‘car’ for the same picture of aTR3), the FFA is more active for judgmentsrequiring classification at a subordinate levelcompared to more categorical judgmentsfor a large variety of objects, living or arti-factual26,27. Thus, living things, includinganimals and animals with hidden faces15,may recruit the FFA because of their visualhomogeneity29 and the level of visual dis-crimination that is most often applied tothem (such as distinguishing a cow from ahorse by default).

Manipulating level of expertiseThe second factor influencing FFA activityfor a wide range of objects is subject exper-tise. The FFA and a similarly selective occip-ital area (‘OFA’) in the righthemisphere4,7,30,31 are both recruited whenobservers become experts in discriminatingobjects from a visually homogeneous cate-gory (objects that share a common geo-metric structure). As detailed below, thisoccurs in subjects trained in the laboratoryto be experts with novel objects called ‘Gree-bles,’ as well as bird and car experts with 20years of experience9,25. Expertise here meansmore than practice or priming with specif-ic instances of a class because the FFAexhibits an equivalent response for new

commentary

0.60

0.70

0.80

0.90

1.00

1 2 3Test Number

Pro

port

ion

Cor

rect

0.60

0.70

0.80

0.90

1.00

1100

1400

1700

2000

2300

Mea

n R

espo

nse

Tim

e (m

s)

1100

1400

1700

2000

2300

Mea

n R

espo

nse

Tim

e (m

s)

Pro

port

ion

Cor

rect

1 2 3Test Number

a

b

Fig. 1. Examples of Greebles. (a) Greebles from a set that Greeble experts could learn to recog-nize faster than novices. (b) Another set of Greebles, which the same experts could not learnfaster than novices, presumably because they are more visually homogeneous than Greebles in thetraining set. Filled squares denote data from novices, and open circles denote data from experts43.

a

b

© 2000 Nature America Inc. • http://neurosci.nature.com©

200

0 N

atur

e A

mer

ica

Inc.

• h

ttp:

//neu

rosc

i.nat

ure.

com

766 nature neuroscience • volume 3 no 8 • august 2000

effect is more pronounced when subjectsattend to the identity of the objects thanwhen they attend to their location. In con-trast, experts show more activity in the rightFFA for their category of expertise, birds orcars, regardless of whether they attend to theidentity or location of the objects. In addi-tion, a behavioral test of relative perfor-mance in matching birds or cars is highlycorrelated (r ~ 0.8) with activity in the rightFFA for another set of birds or cars duringthe location judgments. Similarly, an ERPstudy of bird and dog experts found anenhanced N170—the negative potential typ-ically associated with face-specific process-ing10,11—for categorizing objects withintheir respective domains of expertise33.Thus, both fMRI25 and ERP33 studies ofexperts provide evidence that the neuralbasis of face recognition extends to otherdomains of expertise.

Although these findings would alreadyseem sufficient to reject the hypothesis ofthe FFA as a face detector1,2, additional evi-dence is provided by fMRI habituation pro-cedures assessing the nature of automatizedprocessing for faces in the FFA. A task in

with brain injury is that of prosopagnosia.Patients with this disorder usually have suf-fered injury in the lingual and fusiformgyri, which typically encompass the FFA35.The pattern of sparing and loss is strik-ing—although they can recognize a face asa face, these patients are remarkablyimpaired at face recognition at the indi-vidual level, but sometimes seem to be ableto recognize non-face objects. Althoughthis deficit is often cited as evidence for aface-specific neural substrate, such as theFFA, it actually says little about the originsof face selectivity. Thus, prosopagnosia asa face-specific deficit does not distinguishamong the detection, feature-map andprocess-map models—all three of whichacknowledge the existence of and, indeed,attempt to explain category-selective neur-al substrates. To the extent that prosopag-nosia speaks to these models at all, it isbecause of an intact ability. Even thoughthe FFA is typically lesioned in prosopag-nosia, face detection is spared when test-ed35,36; thus, the detection model cannotaccount for face-selective activity in theFFA. Interestingly, autistic individuals donot show normal FFA activity when view-ing faces and also have few problems withface detection37.

At the same time, a patient (CK) withthe complementary pattern of sparing andloss—preserved face recognition withimpaired object recognition—is widelyheld as the other side of a double dissoci-ation demonstrating distinct face andobject recognition mechanisms38. As withprosopagnosia, at issue is not whether thereexist category-selective brain regions, butrather, whether these regions are domain-general or domain-specific. That is, allthree models predict that a patient such asCK could occur; the question is simplywhy. Unfortunately, testing the domain-specificity of CK’s spared recognition abil-ities would be difficult in that it is hisholistic processing mechanisms thatremain intact (given that it is generallyagreed that the FFA is involved in holisticprocessing for faces and possibly otherobjects9,25,27,31). Within the framework ofthe process-map model, the recruitment ofthese holistic mechanisms results from theinteraction between subordinate-levelrecognition and expertise training. CK,however, seems to have lost the ability torecognize objects in a non-holistic mannerand, in particular, the ability to move froma novice to an expert for any new objectclass. Thus, unlike normal subjects9,39, CKcannot be bootstrapped into applying theintact components of his visual recogni-tion system to new, non-face objects.

which subjects attended to the location of aface revealed that FFA and OFA activityhabituates more to the repeated presenta-tion of a single face than to different faces31.Thus, this activity seems to reflect more thangeneric face detection. Also supporting thehypothesis that the FFA is involved in indi-viduation of faces, attention to face identi-ty (rather than to eye gaze) increases activityin the vicinity of the FFA34. It does not fol-low, however, that this area is not involvedin individual face processing when subjectsattend to dimensions other than identity.Indeed, several studies find that the FFAactivity level in response to faces, as well asto other categories of expertise, is relativelyimmune to whether the task is recognitionor not4,16,25. Thus, it seems that the FFAautomatically processes objects for whichobservers are experts at the subordinatelevel, but, as with most visual processes, theefficacy of the system may be further mod-ulated by attention.

Evidence from neuropsychologyThe most salient neuropsychological exam-ple of impaired visual object recognition

commentary

Fig. 2. Complex selectivity of IT cells. Adapted from ref. 49. (a) This cell responded most to the pic-ture of a teacup, and also responded more to the picture of a balloon than all other images tested.Although the selectivity of this cell may be based on color, this set of responses illustrates that anyhypothesis about the selectivity of a neuron is highly dependent upon the stimulus set used in anexperiment. (b) A cell that responds to a subset of animals, with no obvious measure of similarityseparating the effective from ineffective stimuli. The authors49 argue that selectivity emerges fromexperience and that images that activate IT neurons need not be especially similar to one another.

a

b

© 2000 Nature America Inc. • http://neurosci.nature.com©

200

0 N

atur

e A

mer

ica

Inc.

• h

ttp:

//neu

rosc

i.nat

ure.

com

nature neuroscience • volume 3 no 8 • august 2000 767

Beyond its specific inconsistency withthe detection model, the syndrome ofprosopagnosia reinforces the point that pro-cessing in the FFA is domain-general. Inparticular, patients with prosopagnosia areimpaired with more than faces. Althoughthis was raised as a possibility some timeago20, it was rejected by some researchers40

based on a study in which a prosopagnosicpatient was differentially impaired at recog-nizing previously studied faces as comparedto recognizing previously studied homoge-neous non-face objects (such as eyeglasses).This experiment, however, did not considerpotential response biases, for instance, thefrequency with which the patient responded“old” to individual items actually shownduring study relative to those items notshown during study, and the amount oftime and effort expended by the patient forfaces as compared to objects. Indeed, if thepatient believed (perhaps incorrectly) thathe was more impaired with faces relative toother objects, he may spend less time andallocate fewer attentional resources inencoding and recalling faces36. These con-cerns were addressed in a study that mea-sured recognition response times inconjunction with a bias-free (d’) measureof recognition accuracy36. Data from twopatients with prosopagnosia were collectedfor their recognition of faces, visually simi-lar non-face objects and Greebles. Althoughthe patients did about as well as control sub-jects in recognizing objects and Greebles,they showed a speed–accuracy tradeoff toachieve this level of performance. That is,the patients spent far more time than thecontrols to achieve the same level of recog-nition accuracy. Moreover, when patientswere forced to spend the same amount oftime for each stimulus category, accuracy asmeasured by d’—a statistic capturing a sub-ject’s ability to discriminate or rememberitems independent of response bias—revealed that patients with prosopagnosiaare most impaired whenever visually similarstimuli must be discriminated, regardless ofcategory. Thus, prosopagnosia may actuallybe a domain-general deficit when assessedusing a complete picture of performance.

Critiques of the process mapIt is our contention that behavioral, neu-roimaging, and neuropsychological resultsprovide strong evidence in support of adomain-general model—the process-mapmodel—for face processing and recognition.In particular, manipulations of the factorsof level of categorization and level of exper-tise are able to account for nearly all of theapparent differences between face and objectrecognition. There are, however, skeptics.

mechanism efficient for face processingmight gain some of its efficiency by makingassumptions regarding stimulus geometry.However, it is crucial that we be clear aboutwhat such claims imply, given that we havealready established selectivity of the FFA inresponse to non-face stimuli.

As a specific and oft cited example, con-sider FFA selectivity for Greebles. Severalfindings indicate that Greebles are not treat-ed as face-like in terms of their geometry.First, Greeble novices do not show Greeble-specific activation in FFA, but Greebleexperts do show such activation9. Second,Greeble novices do not show face-likebehavioral effects, but Greeble experts doshow such effects39,43. Third, CK, the agnosicpatient who shows intact face recognition,but is dramatically impaired in non-faceobject recognition38 cannot recognize Gree-bles any better than other objects—evenwhen he is told to think of Greebles as facesor little people (M. Behrmann, I. G. & M. J. T.,unpublished observation). At most, subjectscan only learn to treat Greebles ‘like faces’following an intensive training process thattakes about 10 hours.

A second point is that subjects can seeobjects as face-like without recruiting thoseprocesses thought to be face-specific. Forexample, inverted faces are not recognizedusing the same configural processes asupright faces, yet anybody, includingprosopagnosics, can identify inverted facesas faces. Likewise, schematic faces (madeout of small geometric shapes) clearly looklike faces but elicit about half as muchactivity as shaded faces do (equal to theactivation obtained with the back of headsin a different experiment44). Finally, per-haps the best example comes from patientswith prosopagnosia, who can identify facesas such, but have severe problems perceiv-ing and recognizing them as individuals35.This is consistent with our experiencetraining Greeble experts. Neither behav-ioral nor neural effects of expertise withGreebles are found even when subjects arewell advanced in the training but beforethey meet our criterion for expertise (torecognize Greebles as quickly at the indi-vidual as at the family level).

Thus, seeing Greebles as ‘face-like’(which any novice can do easily) does notpredict expertise effects, including recruit-ment of the FFA. In contrast, a specificmeasure of holistic processing (the com-posite effect) with Greebles correlates withthe recruitment of the FFA for Greebles ina different task (I.G. & M.J.T., unpublisheddata). This method, combining precisepsychophysics and fMRI measurements ofexpertise acquisition, indicates that the

Next we review and address some critiquesof a domain-general account.

Several groups have attempted to refutethe hypothesis that the reason that FFAseems face selective is because faces are rec-ognized at the individual level, whereasother objects are recognized at a coarserlevel. One study compared passive viewingof faces and visually similar flowers appear-ing on a continuously changing backgroundof nonsense patterns or non-face objects.When compared to a baseline of non-faceobjects, faces, but not flowers, producedactivation in right fusiform gyrus41. Simi-larly, another study compared identity judg-ments for faces and hands and found higheractivation for faces than hands as comparedto a fixation baseline4.

How can these results be reconciled withother studies indicating a role for subordi-nate-level processing in the FFA? First, con-sider that the first study used passive viewingand did not require subjects to process flow-ers at the subordinate level. The secondstudy showed faces and hands for 500 mseach. As reviewed above, subordinate-leveljudgments not automatized by expertisetypically require additional processingtime42. Second, these experiments only indi-cate that categorization level is not sufficientto account for all the specificity in the FFA.Note that this is already assumed by ourmodel, in which the interaction of multiplefactors (such as homogeneity, categoriza-tion level and expertise) constrain face-recognition performance.

In sum, experiments using non-faceobjects can provide some evidence that agiven factor influences FFA response. How-ever, the exact contribution of a factor to thecombination that leads to the maximumFFA response cannot be inferred based onthe same experiment (unless it can be pre-cisely equated for faces and objects and weknow how it interacts with other factors).This also means that we cannot refute thepossibility that image geometry is one ofseveral factors that contribute to maximumselectivity—at least not until maximumselectivity is obtained with non-face objects.

Even accepting that the FFA is involvedin subordinate-level recognition, it remainspossible that this area is preferentially selec-tive for a specific geometry—that definingwhat sort of an image ‘counts’ as a face.Refuting this argument is complex—although we know that the FFA can becomeselective for other domains of expertise, forexample, Greebles, birds and cars9,25, it isoften stated that “Greebles look like faces”and that experts can learn to apply face-selective mechanisms to other categories.These two claims seem reasonable—a

commentary© 2000 Nature America Inc. • http://neurosci.nature.com

© 2

000

Nat

ure

Am

eric

a In

c. •

htt

p://n

euro

sci.n

atur

e.co

m

768 nature neuroscience • volume 3 no 8 • august 2000

FFA is recruited by Greebles becauseexperts acquire a more efficient strategyfor discriminating Greebles, not becausesubjects learn to see Greebles as face-like.Reinforcing this point, even when Greebleexperts successfully transfer their expertabilities to the recognition of new Greebleexemplars, their expertise may not applyto new Greebles that are more homoge-neous than the original training set, eventhough the new Greebles clearly lookequivalently face-like (Fig. 1).

The evidence for selectivity on a purelygeometric level in the FFA is not strong,although it is possible that the FFA tends torespond to objects that are face-like becauseour perceptual expertise with faces makesthis geometry a cue that the computationsexecuted in the FFA are required. However,our results indicate experience recognizingGreebles, cars or birds at the individual levelcan make the geometries of these objectsequally valid cues. Selectivity for particularimage geometry is learned—a point drivenhome because objects of acceptable geom-etry do not activate the FFA by default.Thus, selectivity at a purely geometric leveldoes not provide an adequate account ofFFA specialization. In contrast, as alreadyreviewed, we have established that level ofcategorization and expertise in a givenobject domain do affect the response of FFAindependent of stimulus geometry. Weargue that it is the manipulation of these fac-tors, not stimulus geometry, that leads torecruitment of the FFA.

Could different interpretations for thefunction of the FFA depend on how wedefine this area? One method consists ofdefining a face-selective area in a group ofsubjects and measuring its behavior inother conditions (in the same or anothergroup). This averaging method is not opti-mal, but it may be useful when individuallocalizers are not available. A study com-paring group-averaged with individualfunctional definitions revealed that eventhough a FFA can be found in the largemajority of subjects, it is small enough andvariable enough in its location that the for-mer approach may be inadequate27. Specif-ically, if the defined region is large enoughto encompass the FFA for all subjects, it alsoincludes many non-face-selective voxels,and if it is small enough to contain onlyface-selective voxels, it may not include theFFA of many subjects.

Given these problems, many researchershave chosen to rely on individual localiz-ers, leading to very small and highly selec-tive FFAs4,9. This approach sustains acertain illusion: that of a sharply definedarea that is stable across conditions. Typi-

these neurons serve similar functional roles,but with somewhat different tuning (forinstance, adjacent cortical columns aretuned to specific orientations, but the func-tional role of the entire area is a coding oforientation space—arguing that faces arespecial in this context is equivalent to sug-gesting that vertical lines are special, too).

Moreover, the finding that a given neu-ron responds most strongly to a specificstimulus does not imply that this same neu-ron is not functionally implicated in the pro-cessing of other stimuli. For example,neurophysiologists often rely on a reductionmethod17 in which, beginning with animage from a restricted basis set, they incre-mentally simplify the test stimulus whileattempting to maintain the same level of cellresponse. Thus, the reduction method isbound to produce a single preferred featureset per cell. In contrast, IT cells can haveresponses that defy simple explanations interms of singular preferred features (Fig. 2).

CONCLUSIONSFrom grandmothers to scientists (and thetwo are not mutually exclusive), there is acertain fondness for the idea that the dif-ficulty of face recognition combined withits social importance has exerted uniqueadaptive pressures. Thus, it is often argued,it would not be particularly surprising ifthere were a distinct neural and function-al module for face recognition. Althoughthe adaptive pressures may be real, thisimplies only that the primate visual sys-tem should be capable of efficiently per-forming this task, not that whateverperceptual mechanisms represent the pre-sent end-state of evolution must be exclu-sive to faces. Put another way, speculatingon the evolutionary reasons for why a par-ticular neural system might have arisendoes not define the limits of this system.In an effort to explore these limits, we havecollected evidence that FFA processing isdomain-general. Specifically, it is involvedin processing subordinate-level informa-tion for all objects, including faces. Thisprocessing is not restricted to objects of acertain geometry (that is, faces), as this fac-tor cannot account for the majority ofresponses in the FFA to non-face objects.On the other hand, manipulations of thelevel of categorization and the level ofexpertise are able to predict when and ifFFA activity will be found in a given task.Thus, even if stimulus geometry is some-how involved in FFA selectivity, it is likelyto remain only one of several factors thatcontribute to the specialization of thisbrain area. That faces generally elicit themaximum response in the FFA is interest-

cally, a localizer is used to define a FFA atthe outset of the study, and all furtheranalyses reflect the response of only thesevoxels. Little consideration is given to thepotential variability of the FFA. Forinstance, both the peaks of face selectivityand the extent of the FFA as defined usinga faces-minus-objects comparison can dif-fer substantially within subjects during pas-sive viewing as compared to an activeidentification task25. This issue is highly rel-evant to the nature of the selectivity in theFFA, as differences between faces and non-face objects must be weighed against thevariability of face selectivity itself. To theextent the microstructure of this variabil-ity can be assessed with current techniques,there is no indication that the face-selec-tive part of cortex is in any sense mutuallyexclusive with the rest of the object-recog-nition system. In a fMRI study of bird andcar experts, several criteria for the defini-tion of the FFA were compared (seehttp://www.nature.com/neuro/web_specials).The criterion most stringent in terms ofsize and face selectivity (face to object activ-ity ratio, 12.8) led to a right FFA region ofinterest with an average size of 3 voxels(each voxel was 3.125 mm2 in plane) butcould be found in all 19 subjects. In com-parison, the criterion used in many priorstudies4 led to a larger area (average size of6 voxels), which displayed more moderateface selectivity (face to object activity ratio,5.2) but was less inclusive subject-wise (aright FFA only found in 7 of 19 subjects).Crucially, regardless of criterion, the FFAshowed sensitivity to the level of catego-rization and expertise manipulations. Thisand several other analyses, such as centerof mass comparisons and inspection of thepeaks in unsmoothed activation maps,indicate that face expertise cannot be dis-sociated from object expertise at the spa-tial resolution of fMRI.

The existence of ‘face cells’ in IT45 is oftencited as evidence for a distinct recognitionmechanism for faces. On the other hand,studies in which monkeys learn to recognizenovel objects have revealed a remarkablesimilarity between properties of cells tunedto these trained objects and face cells46,47.One interpretation is that “…it would besurprising if neurons exist that respondstrongly to both faces and cars (but notother complex objects) in car experts (forexample)”48. The implication is that facesand objects would selectively activate inde-pendent populations of neurons, inter-spersed within the same fMRI voxel. Evenif this were the case, the intermingling of atleast two populations of neurons in a rela-tively small region of IT would suggest that

commentary© 2000 Nature America Inc. • http://neurosci.nature.com

© 2

000

Nat

ure

Am

eric

a In

c. •

htt

p://n

euro

sci.n

atur

e.co

m

nature neuroscience • volume 3 no 8 • august 2000 769

commentary

ing, but only as an effect that is inherentlyconfounded with the processing biasesthat, through experience, are automatical-ly engaged when we see a face.

ACKNOWLEDGEMENTSThis work was a collaborative effort, and theorder of authorship is arbitrary. We thank D. Sheinberg for his careful reading of this review.M.J.T. thanks S. Pinker for a challengingdiscussion on some of the issues raised here; I.G.thanks N. Kanwisher for many stimulatingdiscussions on our diverging opinions. This workwas supported by NSF Award SBR-9615819.

RECEIVED 4 APRIL; ACCEPTED 20 JUNE, 2000

1. Kanwisher, N., Tong, F. & Nakayama, K. Theeffect of face inversion on the human fusiformface area. Cognition 68, B1–11 (1998).

2. Bentin, S., Deouell, L. Y. & Soroker, N.Selective visual streaming in face recognition:evidence from developmental prosopagnosia.Neuroreport 10, 823–827 (1999).

3. Puce, A., Allison, T. & McCarthy, G.Electrophysiological studies of human faceperception. III: Effects of top-down processingon face-specific potentials. Cereb. Cortex 9,445–458 (1999).

4. Kanwisher, N., McDermott, J. & Chun, M. M.The fusiform face area: A module in humanextrastriate cortex specialized for faceperception. J. Neurosci. 17, 4302–4311 (1997).

5. Kanwisher, N., Stanley, D. & Harris, A. Thefusiform face area is selective for faces notanimals. Neuroreport 10, 183–187 (1999).

6. Aguirre, G. K., Zarahn, E. & D’Esposito, M. Anarea within human ventral cortex sensitive to“building” stimuli: Evidence and implications.Neuron 21, 373–383 (1998).

7. Haxby, J. V. et al. The effect of face inversion onactivity in human neural systems for face andobject perception. Neuron 22, 189–199 (1999).

8. Yin, R. K. Looking at upside-down faces. J. Exp.Psychol. Gen. 81, 141–145 (1969).

9. Gauthier, I., Tarr, M. J., Anderson, A. W.,Skudlarski, P. & Gore, J. C. Activation of themiddle fusiform ‘face area’ increases withexpertise in recognizing novel objects. Nat.Neurosci. 2, 568–573 (1999).

10. Bentin, S., Allison, T., Puce, A., Perez, E. &McCarthy, G. Electrophysiological studies offace perception in humans. J. Cognit. Neurosci. 8,551–565 (1996).

11. Jeffreys, D. A. A face-responsive potentialrecorded from the human scalp. Exp. Brain Res.78, 193–202 (1989).

12. Liu, J., Higuchi, M., Marantz, A. & Kanwisher, N.The selectivity of the occipitotemporal M170 forfaces. Neuroreport 7, 337–341 (2000).

13. Rossion, B. et al. The N170 occipito-temporalcomponent is delayed and enhanced to invertedfaces but not to inverted objects: Anelectrophysiological account of face-specificprocesses in the human brain. Neuroreport 11,69–74 (2000).

32. Rossion, B. et al. Hemispheric asymmetriesfor whole-based and part-based faceprocessing in the human brain. J. Cogn.Neurosci. (in press).

33. Tanaka, J. W. & Curran, T. A neural basis forexpert object recognition. Psychol. Sci. (in press).

34. Hoffman, E. A. & Haxby, J. V. Distinctrepresentations of eye gaze and identity in thedistributed human neural system for faceperception. Nat. Neurosci. 3, 80–84 (2000).

35. Tranel, D., Damasio, A. R. & Damasio, H. Intactrecognition of facial expression, gender and agein patients with impaired recognition of faceidentity. Neurology 38, 690–696 (1998).

36. Gauthier, I., Behrmann, M. & Tarr, M. J. Canface recognition really be dissociated from objectrecognition? J. Cogn. Neurosci. 11, 349–370(1999).

37. Schultz, R. T. et al. Abnormal ventral temporalcortical activity during face discriminationamong individuals with autism and aspergersyndrome. Arch. Gen. Psychiatry 57, 331–340(2000).

38. Moscovitch, M., Winocur, G. & Behrmann, M.What is special about face recognition? Nineteenexperiments on a person with visual objectagnosia and dyslexia but normal facerecognition. J. Cogn. Neurosci. 9, 555–604(1997).

39. Gauthier, I. & Tarr, M. J. Becoming a “Greeble”expert: Exploring the face recognitionmechanism. Vision Res. 37, 1673–1682 (1997).

40. Farah, M. J., Levinson, K. L. & Klein, K. L. Faceperception and within-category discriminationin prosopagnosia. Neuropsychologia 33, 661–674(1995).

41. McCarthy, G., Puce, A., Gore, J. C. & Allison, T.Face-specific processing in the human fusiformgyrus. J. Cogn. Neurosci. 9, 605–610 (1997).

42. Jolicoeur, P., Gluck, M. & Kosslyn, S. M. Picturesand names: Making the connection. Cognit.Psychol. 16, 243–275 (1984).

43. Gauthier, I., Williams, P., Tarr, M. J. & Tanaka, J.Training “Greeble” experts: A framework forstudying expert object recognition processes.Vision Res. 38, 2401–2428 (1998).

44. Tong, F., Nakayama, K., Moscovitch, M.,Weinrib, O. & Kanwisher, N. Responseproperties of the human fusiform face area.Cognit. Neuropsychol. 17, 257–279 (2000).

45. Perrett, D. I., Rolls, E. T. & Caan, W. Visualneurones responsive to faces in the monkeytemporal cortex. Exp. Brain Res. 47, 329–342(1982).

46. Kobatake, E., Wang, G. & Tanaka, K. Effects ofshape-discrimination training on the selectivityof inferotemporal cells in adult monkeys. J. Neurophysiol. 80, 324–330 (1998).

47. Logothetis, N. K. & Pauls, J. Psychophysical andphysiological evidence for viewer-centeredobject representation in the primate. Cereb.Cortex 3, 270–288 (1995).

48. Kanwisher, N., Downing, P., Epstein, R. &Kourtzi, Z. in Handbook of FunctionalNeuroimaging of Cognition (eds. Cabeza, R. &Kingstone, A.) (MIT Press, Cambridge, MA, inpress).

49. Sheinberg, D. L. & Logothetis, N. K. inPerceptual Learning (ed. Fahle, M.) (MIT Press,Cambridge, Massachusetts, 2000).

14. McCarthy, G., Puce, A., Belger, A. & Allison, T.Electrophysiological studies of human faceperception. II: Response properties of face-specific potentials generated inoccipitotemporal cortex. Cereb. Cortex 5,431–444 (1999).

15. Chao, L. L., Haxby, J. V. & Martin, A. Attribute-based neural substrates in temporal cortex forperceiving and knowing about objects. Nat.Neurosci. 2, 913–919 (1999).

16. Ishai, A., Ungerleider, L. G., Martin, A.,Schouten, J. L. & Haxby, J. Distributedrepresentation of objects in the human ventralvisual pathway. Proc. Natl. Acad. Sci. USA 96,9379–9384 (1999).

17. Tanaka, K. Inferotemporal cortex and objectvision. Annu. Rev. Neurosci. 19, 109–139 (1996).

18. Puce, A., Allison, T., Asgari, M., Gore, J. C. &McCarthy, G. Differential sensitivity of humanvisual cortex to faces, letterstrings, and textures:A functional magnetic resonance imaging study.J. Neurosci. 16, 5205–5215 (1996).

19. Gauthier, I. What constrains the organization ofthe ventral temporal cortex? Trends Cogn. Sci. 4,1–2 (2000).

20. Damasio, A. R., Damasio, H. & Van Hoesen, G. W.Prosopagnosia: Anatomical basis and behavioralmechanisms. Neurology 32, 331–341 (1982).

21. Diamond, R. & Carey, S. Why faces are and arenot special: An effect of expertise. J. Exp. Psychol.Gen. 115, 107–117 (1986).

22. Tanaka, J. W. & Gauthier, I. in Mechanisms ofPerceptual Learning (eds. Goldstone, R. L.,Medin, D. L. & Schyns, P. G.) 83–125 (Academic,San Diego, California, 1997).

23. Haxby, J. V. et al. Dissociation of object andspatial visual processing pathways in humanextrastriate cortex. Proc. Natl. Acad. Sci. USA 88,1621–1625 (1991).

24. Sergent, J., Ohta, S. & MacDonald, B. Functionalneuroanatomy of face and object processing: Apositron emission tomography study. Brain 115,15–36 (1992).

25. Gauthier, I., Skudlarski, P., Gore, J. C. &Anderson, A. W. Expertise for cars and birdsrecruits brain areas involved in face recognition.Nat. Neurosci. 3, 191–197 (2000).

26. Gauthier, I., Anderson, A. W., Tarr, M. J.,Skudlarski, P. & Gore, J. C. Levels ofcategorization in visual object studied withfunctional MRI. Curr. Biol. 7, 645–651(1997).

27. Gauthier, I. et al. Does visual subordinate-levelcategorisation engage the functionally definedFusiform Face Area? Cognit. Neuropsychol. 17,143–163 (2000).

28. Chao, L. L., Martin, A. & Haxby, J. V. Are face-responsive regions selective only for faces?Neuroreport 10, 2945–2950 (1999).

29. Gaffan, D. & Heywood, C. A. A spuriouscategory-specific agnosia for living things innormal human and nonhuman primates. J. Cogn. Neurosci. 5, 118–128 (1994).

30. Halgren, E. et al. Location of human face-selective cortex with respect to retinotopic areas.Hum. Brain Mapp. 7, 29–37 (1999).

31. Gauthier, I. et al. The fusiform “face area” is partof a network that processes faces at theindividual level. J. Cogn. Neurosci. 12, 495–504(2000).

© 2000 Nature America Inc. • http://neurosci.nature.com©

200

0 N

atur

e A

mer

ica

Inc.

• h

ttp:

//neu

rosc

i.nat

ure.

com