f ace-lik eness and ima g e v ariability driv e responses...

16
r Human Brain Mapping 00:000–000 (2011) r Face-Likeness and Image Variability Drive Responses in Human Face-Selective Ventral Regions Nicolas Davidenko, 1 * David A. Remus, 1 and Kalanit Grill-Spector 1,2 1 Psychology Department, Stanford University, Stanford, California 2 Neuroscience Institute, Stanford University, Stanford, California r r Abstract: The human ventral visual stream contains regions that respond selectively to faces over objects. However, it is unknown whether responses in these regions correlate with how face-like stimuli appear. Here, we use parameterized face silhouettes to manipulate the perceived face-likeness of stimuli and mea- sure responses in face- and object-selective ventral regions with high-resolution fMRI. We first use ‘‘con- centric hyper-sphere’’ (CH) sampling to define face silhouettes at different distances from the prototype face. Observers rate the stimuli as progressively more face-like the closer they are to the prototype face. Paradoxically, responses in both face- and object-selective regions decrease as face-likeness ratings increase. Because CH sampling produces blocks of stimuli whose variability is negatively correlated with face-likeness, this effect may be driven by more adaptation during high face-likeness (low-variability) blocks than during low face-likeness (high-variability) blocks. We tested this hypothesis by measuring responses to matched-variability (MV) blocks of stimuli with similar face-likeness ratings as with CH sampling. Critically, under MV sampling, we find a face-specific effect: responses in face-selective regions gradually increase with perceived face-likeness, but responses in object-selective regions are unchanged. Our studies provide novel evidence that face-selective responses correlate with the perceived face-like- ness of stimuli, but this effect is revealed only when image variability is controlled across conditions. Finally, our data show that variability is a powerful factor that drives responses across the ventral stream. This indicates that controlling variability across conditions should be a critical tool in future neuroimag- ing studies of face and object representation. Hum Brain Mapp 00:000–000, 2011. V C 2011 Wiley-Liss, Inc. Key words: face-likeness; variability; face silhouettes; FFA; fMRI r r INTRODUCTION Functional magnetic resonance imaging (fMRI) studies reliably identify regions in the human ventral stream that respond selectively to faces [Kanwisher et al., 1997; Puce et al., 1995; Tong et al., 2000; Yovel and Kanwisher, 2004], including regions in the middle (mFus-faces) and posterior [pFus-faces (Weiner and Grill-Spector, 2010)] fusiform gyrus (often referred to collectively as FFA), and a region in lat- eral occipital cortex, overlapping the inferior occipital gyrus, referred to as OFA [Gauthier et al., 2000], or IOG- faces [Weiner and Grill-Spector, 2010]. These face-selective regions are thought to be critically involved in face percep- tion [Andrews et al., 2002; Hasson et al., 2001; Moutoussis and Zeki, 2002; Rotshtein et al., 2005; Tong et al., 1998] and recognition [Barton, 2002; Golarai et al., 2007; Grill- Additional Supporting Information may be found in the online version of this article. Contract grant sponsor: NIH NRSA; Contract grant number: EY18279-01A1; Contract grant sponsor: Whitehall; Contract grant number: 2005-05-111-RES; Contract grant sponsor: NSF; Contract grant numbers: BCS 0920865, 0547775, 0624345; Contract grant sponsor: NEI; Contract grant number: RO1EY019279-01-A1; Contract grant sponsor: American Philosophical Society. *Correspondence to: Nicolas Davidenko, Psychology Department, Stanford University, Stanford, California. E-mail: [email protected] Received for publication 21 December 2010; Revised 14 April 2011; Accepted 2 May 2011 DOI: 10.1002/hbm.21367 Published online in Wiley Online Library (wileyonlinelibrary. com). V C 2011 Wiley-Liss, Inc.

Upload: others

Post on 09-Aug-2020

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

r Human Brain Mapping 00:000–000 (2011) r

Face-Likeness and Image Variability DriveResponses inHumanFace-SelectiveVentral Regions

Nicolas Davidenko,1* David A. Remus,1 and Kalanit Grill-Spector1,2

1Psychology Department, Stanford University, Stanford, California2Neuroscience Institute, Stanford University, Stanford, California

r r

Abstract: The human ventral visual stream contains regions that respond selectively to faces over objects.However, it is unknown whether responses in these regions correlate with how face-like stimuli appear.Here, we use parameterized face silhouettes to manipulate the perceived face-likeness of stimuli and mea-sure responses in face- and object-selective ventral regions with high-resolution fMRI. We first use ‘‘con-centric hyper-sphere’’ (CH) sampling to define face silhouettes at different distances from the prototypeface. Observers rate the stimuli as progressively more face-like the closer they are to the prototype face.Paradoxically, responses in both face- and object-selective regions decrease as face-likeness ratingsincrease. Because CH sampling produces blocks of stimuli whose variability is negatively correlated withface-likeness, this effect may be driven by more adaptation during high face-likeness (low-variability)blocks than during low face-likeness (high-variability) blocks. We tested this hypothesis by measuringresponses to matched-variability (MV) blocks of stimuli with similar face-likeness ratings as with CHsampling. Critically, under MV sampling, we find a face-specific effect: responses in face-selective regionsgradually increase with perceived face-likeness, but responses in object-selective regions are unchanged.Our studies provide novel evidence that face-selective responses correlate with the perceived face-like-ness of stimuli, but this effect is revealed only when image variability is controlled across conditions.Finally, our data show that variability is a powerful factor that drives responses across the ventral stream.This indicates that controlling variability across conditions should be a critical tool in future neuroimag-ing studies of face and object representation. Hum Brain Mapp 00:000–000, 2011. VC 2011Wiley-Liss, Inc.

Keywords: face-likeness; variability; face silhouettes; FFA; fMRI

r r

INTRODUCTION

Functional magnetic resonance imaging (fMRI) studiesreliably identify regions in the human ventral stream thatrespond selectively to faces [Kanwisher et al., 1997; Puceet al., 1995; Tong et al., 2000; Yovel and Kanwisher, 2004],including regions in the middle (mFus-faces) and posterior[pFus-faces (Weiner and Grill-Spector, 2010)] fusiform gyrus(often referred to collectively as FFA), and a region in lat-eral occipital cortex, overlapping the inferior occipitalgyrus, referred to as OFA [Gauthier et al., 2000], or IOG-faces [Weiner and Grill-Spector, 2010]. These face-selectiveregions are thought to be critically involved in face percep-tion [Andrews et al., 2002; Hasson et al., 2001; Moutoussisand Zeki, 2002; Rotshtein et al., 2005; Tong et al., 1998]and recognition [Barton, 2002; Golarai et al., 2007; Grill-

Additional Supporting Information may be found in the onlineversion of this article.

Contract grant sponsor: NIH NRSA; Contract grant number:EY18279-01A1; Contract grant sponsor: Whitehall; Contract grantnumber: 2005-05-111-RES; Contract grant sponsor: NSF; Contractgrant numbers: BCS 0920865, 0547775, 0624345; Contract grantsponsor: NEI; Contract grant number: RO1EY019279-01-A1;Contract grant sponsor: American Philosophical Society.

*Correspondence to: Nicolas Davidenko, Psychology Department,Stanford University, Stanford, California.E-mail: [email protected]

Received for publication 21 December 2010; Revised 14 April2011; Accepted 2 May 2011

DOI: 10.1002/hbm.21367Published online in Wiley Online Library (wileyonlinelibrary.com).

VC 2011 Wiley-Liss, Inc.

Page 2: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

Spector et al., 2004], and much work in the past decadehas focused on understanding their response properties.

To date it has not been established how face-selectiveregions respond to stimuli that vary in their degree offace-likeness. On the one hand, responses in face-selectiveregions are higher when stimuli are perceived as faces vs.when they are not. For example, when presented with am-biguous stimuli (e.g., the Rubin face-vase illusion)responses in fusiform face-selective regions are higherwhen subjects report seeing two faces rather than a vase[Andrews et al., 2002; Hasson et al., 2001]. Further, severalstudies report that face-selective regions respond similarlyto different kinds of faces, such as familiar vs. unfamiliarfaces [Ewbank and Andrews, 2008; Kanwisher et al., 1997],adult vs. child faces [Golarai et al., 2010], male vs. femalefaces [Kranz and Ishai, 2006], and front- vs. profile-viewfaces [Kanwisher et al., 1997] (though there is evidence forsome modulation of response by race [Golby et al., 2001]and distinctiveness [Loffler et al., 2005]). On the otherhand, although face-selective regions respond more tofaces than non-faces, their responses are not uniformlylow across non-face stimuli. For example, face-selectiveregions respond more strongly to images of animals andbody parts than to inanimate objects, and more strongly toobjects than to houses and scenes [Downing et al., 2006;Grill-Spector, 2003; Grill-Spector et al., 2004; Weiner andGrill-Spector, 2010]. Because these visual categories differon many dimensions, it is unclear what aspects of imagesmodulate the level of response of face-selective regions. Arecent study reported that FFA responses to images at var-ious morph levels between a face and a house decreased

monotonically as the morph level had a lower contributionof the face image, suggesting that responses may bemodulated by stimulus shape [Tootell, 2008]. However,several intermediate morphs between the face and thehouse resembled neither a face nor a house, and therewere no behavioral measurements of how face-like thestimuli appeared to subjects. Thus, it remains unknownwhat kinds of stimulus transformations modulate responselevels in face-selective regions, and how they might relateto perceived face-likeness.

To address these questions we used a novel method ofparameterized face silhouettes [Davidenko, 2007]. Thisface space is built from a database of 480 profile-view,gray-scale faces which are parameterized into silhouettesusing 18 keypoints (see Fig. 1 and Materials and Methods).The resulting face stimuli are defined entirely by theirshape. Using principal component analysis (PCA), we cap-ture the shape dimensions, or principal components (PCs),along which human face profiles vary relative to a proto-type face (the average of these 480 silhouettes, which sub-jects perceive as typical [Davidenko, 2007; Davidenko andRamscar, 2006; Davidenko et al., 2007]). Importantly, theface silhouette methodology provides a method for manip-ulating the perceived face-likeness of stimuli simply bytransforming the shape of silhouettes along the PCs thatfaces typically vary along, while controlling low-levelproperties of the stimuli, such as size, contrast, and spatialfrequency.

Here, we generated silhouettes at various distances fromthe prototype face. We first measured the perceived face-likeness of these stimuli and show that faces further away

Figure 1.Steps in parameterizing a face silhouette [adapted from Davidenko, 2007]. (a) A profile faceimage from the FERET database; (b) 18 keypoints are identified along the profile contour; (c)keypoints are normalized into a common x–y plane across all face images; (d) bi-cubic splinessmoothly interpolate between keypoints to create the parameterized face silhouette.

r Davidenko et al. r

r 2 r

Page 3: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

from the prototype are perceived as less face-like, in amonotonically decreasing way. We then measuredresponses in face- and object-selective ventral regions withhigh-resolution fMRI. We reasoned that responses in face-selective, but not object-selective, regions would be modu-lated by the degree of face-likeness of the silhouettes. Wepredicted that responses in fusiform face-selective regionsshould decrease as face-like ratings decrease, sinceresponses in these regions are known to be correlated withsubjects’ face perception.

MATERIALS AND METHODS

Subjects

Thirteen right-handed subjects (seven female) partici-pated in one or more of four fMRI experiments. In Experi-ments 1 and 2 (Face silhouette study), we verified that facesilhouettes elicit a face-selective response profile in humanventral visual cortex. Eight subjects (six female, ages 23–39) participated in Experiment 1 and 9 subjects partici-pated in Experiment 2. In Experiments 3 and 4, we manip-ulated face-likeness and measured responses in face- andobject-selective regions. Ten right-handed subjects (fivefemale, ages 21–39 years) participated in Experiment 3(CH sampling) and nine right-handed subjects (fourfemale, ages 21–39) participated in Experiment 4 (MV sam-pling); seven subjects participated in both of these. Stan-ford University’s Human Subjects IRB approved theexperimental protocol and subjects gave written informedconsent to participate.

fMRI Scans

Scans were conducted at the Richard M. Lucas Centerfor Imaging at Stanford University using an eight-channelsurface coil in a 3-Tesla GE magnet. Experiments 1–4 wereconducted on separate days between 3 and 11 monthsapart for different subjects (mean: 6 months). Each scanlasted !90 min and included an anatomical run, two func-tional localizer runs to identify face- and object-selectiveregions, and eight experimental runs. Functional scanswere conducted at high-resolution (HR-fMRI, 1.5 mm iso-tropic voxels) and covered a field of view of 19.2 cm "

19.2 cm " 3.9 cm that included occipital and temporalregions. HR-fMRI allows for better segregation betweenobject- and face-selective activations and more accurateROI definitions. We acquired 26 1.5 mm-thick oblique sli-ces using a two-shot T2*-sensitive spiral acquisitionsequence [Glover, 1999] (TE # 30 ms, TR # 2,000 ms, flipangle # 77$, and bandwidth # 125 kHz). In-plane T1-weighted anatomical images were acquired in the samesession as the functional scans using the same prescriptionas the functional scans and a standard two-dimensionalRF-spoiled GRASS (SPGR) pulse sequence (TE # 1.9 ms,flip angle # 15$, bandwidth # 15.63 kHz). The anatomical

in-planes were used to co-register each subject’s functionaland anatomical data to aid with ROI visualization and def-inition. A high-resolution anatomical volume of the wholebrain was acquired for 11 subjects with a head-coil using aT1-weighted SPGR pulse sequence (TR # 1,000 ms, FA #

45$, 2 NEX, FOV # 200 mm, resolution of 0.78 " 0.78 "

1.2 mm3). For the remaining two subjects, ROI definitionswere conducted based on the inplane images acquired ineach session.

fMRI Data Analysis

Imaging data was analyzed using MATLAB and our in-house software, mrVista (white.stanford.edu/software).The analysis included alignment of functional data towhole-brain anatomies, motion correction of the functionaldata, transformation of each voxel’s time course to percentsignal change, and application of a general linear model(GLM [Worsley et al., 1996] using the hemodynamicresponse function as implemented in SPM2 and SPM5,www.fil.ion.ucl.ac.uk/spm] to estimate b-weights for eachvoxel in each experimental condition. There was no spatialsmoothing. For ROI-based analyses, we measured the av-erage in each subject’s individually defined ROIs from thelocalizer scan from each condition of Experiments 1–4. Wethen averaged data across subjects. Data were analyzedseparately for each hemisphere.

Localizer Runs

Two 5-min localizer scans were conducted at the begin-ning of each scanning session. Subjects observed 12-sblocks of grayscale front-view faces, grayscale commonobjects (cars, instruments, and plants), gray-scalescrambled versions of these images, and fixation blocks,with 12 instances of each type of block using differentstimuli. These localizer runs were analyzed to define face-and object-selective ROIs separately in each subject.

Face-selective ROIs were defined in each subject as vox-els in the fusiform gyrus (Fus) and lateral occipital cortexoverlapping the inferior occipital gyrus (IOG) thatresponded significantly more strongly to faces than toobjects (P < 0.001, voxel level, uncorrected; see Fig. 2–red).In most subjects, right hemisphere Fus-faces consisted of aposterior (pFus-faces) and anterior (mFus-faces) regionthat we analyzed separately. The mFus-faces ROI is moremedial and tends to overlap the mid-fusiform sulcus,whereas the pFus-ROI is about 1–2 cm more posterior andsometimes overlaps the occipito-temporal sulculs (OTS),see [Weiner and Grill-Spector, 2010]. We focus our analy-ses on right- and left-hemisphere ROIs that were presentin at least five subjects in each of the experiments. Theseinclude right-hemisphere mFus-faces (minimum n # 9),pFus-faces (9), and IOG-faces (6) and left-hemispherepFus-faces (9). Left mFus-faces (minimum n # 3) and IOGfaces (4) were found less consistently in our studies.

r Face-Likeness and Image Variability r

r 3 r

Page 4: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

Object-selective regions were defined as voxels in poste-rior fusiform gyrus overlapping with the OTS (pFus/OTS)and in the lateral aspect of the occipital lobe posterior andventral to MT (LO) that responded significantly morestrongly to objects than scrambled images (P < 0.001,voxel level, uncorrected; Fig. 2–blue). We identified object-selective regions in the right hemisphere pFus/OTS (mini-mum n # 6) and LO (9) and left hemisphere pFus/OTS (6)and LO (7).

Results of Experiments 1–4 are shown as the averagepercentage signal for each condition across participants,shown separately in each face- and object-selective ROI.For robustness and to complement our ROI analyses, wemeasured the extent of activation within each region tothe different experimental conditions. We defined a voxelcount measure in Experiments 3 and 4 as the number ofvoxels within each functional ROI responding significantly(P < 0.01) more strongly to each active condition ofExperiments 3 and 4 versus fixation. This analysis wasdone separately for each subject and ROI. Average voxelcounts across subjects are shown in Supporting Informa-tion Figures 5 and 6.

Stimuli

Our stimuli were constructed using the parameterizedface silhouettes methodology [Davidenko, 2007]. Silhouetteface space is generated by placing 18 keypoints along thecontour of 480 left-facing profile photographs of male and

female faces from the FERET face database [Phillips, 1998,2000] (Fig. 1a). We fixed the top and bottom points to con-trol for stimulus size and orientation. The XY coordinatesof the remaining 16 keypoints form the basis of a principalcomponents analysis (PCA) that produces 32 orthogonaldimensions, or principal components (PCs) that fullydescribe the space of face silhouettes [Davidenko, 2007].The prototype face silhouette—the average of all 480 faceprofiles—is obtained by setting 0-coefficients along all PCsand novel face silhouettes can be generated by samplingthe PC coefficients from a multi-normal distribution [seeDavidenko, 2007].

In previous work [Davidenko, 2007], we have shownface silhouettes are perceived much like regular face stim-uli, eliciting accurate judgments of gender and age, reliableratings of attractiveness and distinctiveness, successfulidentification with their front-view, gray-scale counter-parts, and better recognition when upright than upside-down. In Experiments 1 and 2, we further show that para-meterized face silhouettes elicit a face-selective responseprofile in the ventral stream.

Face Silhouette Study (Experiments 1 and 2)

To determine whether face silhouettes elicit a face-selec-tive response profile, we measured responses with high-re-solution fMRI as subjects observed blocks of facesilhouettes, two-tone shapes, and scrambled two-toneimages (Experiment 1), and blocks of upright and upside-

Figure 2.Face- and object-selective activations on a ventral surface of arepresentative subject’s inflated cortical surface. Face-selectiveregions (responding more to front-view faces than objects) areshown in red and object-selective regions (responding more toobjects than scrambled images) are shown in blue. Color bar

indicates the statistical significance (P-value) of the contrast ofinterest. Abbreviations: mFus: mid-fusiform; pFus: posterior fusi-form: OTS: occipito-temporal sulcus; IOG: inferior occipitalgyrus. LO: lateral occipital object-selective region.

r Davidenko et al. r

r 4 r

Page 5: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

down face silhouettes (Experiment 2). In Experiment 1, wegenerated multiple face silhouettes by sampling the coeffi-cients of the first 6 PCs from a multi-normal distributionaround the prototype face. Each PC coefficient wasweighted by its z-score, thus producing normally distrib-uted values along each dimension, resulting in realisticlooking face silhouettes. We used six dimensions in orderto generate a wide variety of silhouette stimuli that wouldnot repeat across blocks and runs. Using a similar key-point construction method, we generated two-tone shapesthat did not resemble faces but were matched with facesilhouettes in their size, contrast, and number of key-points. We generated scrambled two-tone images by split-ting the face and shape silhouette images into 8 by 8 sub-images and randomizing the positions of the sub-images.In Experiment 2, we generated upright face silhouettes byvarying the coefficients of PCs 2 and 6 by up to %1.5 and%2 standard deviations (PC units), respectively, awayfrom the prototype face silhouette and upside-down sil-houettes by vertically inverting the upright silhouettes.

For each subject in Experiments 1 and 2, we acquired 8runs where subjects observed six 12-s image blocks alter-nating with 12-s fixation blocks. Each run included oneinstance of each type of image block. Stimuli were dis-played in MATLAB (Mathworks.com) using Psychtoolbox2[Brainard, 1997] and projected onto a screen inside thescanner. Images subtended !5.7$ of subjects’ visual fieldand were presented at 1 Hz, for 750 ms with 250 ms fixa-tion interstimulus intervals. During all image blocks, sub-jects fixated on a central cross and performed a one-back

task, pressing a button whenever two consecutive imagesrepeated (!8% of trials). Subjects’ performance was moni-tored to ensure they were paying attention to the stimuli.

Measuring Responses in the Ventral Stream as aFunction of Face-Likeness

Critically, we contrast two methods of defining blocks ofsilhouettes at different levels of face-likeness. Concentrichyper-sphere (CH) sampling (Fig. 3a and Supporting In-formation Fig. 1) is a method that has been used previ-ously to manipulate face distinctiveness [Loffler et al.,2005; Valentine, 1991]. CH sampling captures a wide rangeof face stimuli but results in a correlation between distan-ces from the prototype face and image variability, wherebyblocks of stimuli far from the prototype are more variablethan blocks of stimuli near the prototype. To account forthis potential confound, we introduce a novel method of‘‘matched variability’’ (MV) sampling, where silhouettesare defined primarily along one PC of face space at thesame distances from the prototype, but with each block ofsilhouettes spanning an equal-size region of face spaceand thus matching variability across blocks (Fig. 3b).

Measures of Variability

We defined a metric of image variability to describe theheterogeneity of images in each block. Image variabilitywas defined as the mean distance in face space between

Figure 3.

(a) Schematics of CH and MV sampling. Each line represents a different PC and each shaded diskrepresents a block of stimuli at a given distance from the prototype to be shown in the fMRIexperiment. (b) Schematic diagram of MV sampling for Experiment 4. Silhouettes are definedalong PC3 and each equal-size shaded disk represents a block of matched-variability stimuli at agiven distance from the prototype to be shown in the fMRI experiment.

r Face-Likeness and Image Variability r

r 5 r

Page 6: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

pairs of 32-dimensional vectors corresponding to pairs offace silhouettes within each block. The units of image vari-ability are in standard deviations of the PC coefficients (orPC units). This measure was nearly identical to a measurebased on the sum of XY distances between pairs of corre-sponding keypoints that define each silhouette stimulus (r# 0.9993, P < 10&10). We considered two additional met-rics of variability in our analyses: pixel-based variabilityand perceptual variability. Pixel-based variability wasdefined as the pixel-by-pixel difference between the two-tone images in each block, expressed as a proportion ofdifferent pixels in the image area. Perceptual variabilitywas measured in a post-scan study outside the scanner.Nine of our subjects provided ratings of dissimilarityamong pairs of silhouettes sampled from each of theblocks, on a seven-point scale, with ‘‘1’’ representing‘‘identical’’ and ‘‘7’’ representing ‘‘maximally dissimilar.’’

Concentric Hyper-Sphere (CH) Sampling(Experiment 3)

We generated silhouettes by sampling from 12 orthogo-nal face space dimensions (PCs 1, 3–5, and 7–14) at 9 dif-ferent distances (0, 1.5, 3, 4.5, 6, 7.5, 9, 10.5, and 12 PCunits) from the prototype (see Fig. 3a), and blocking thesilhouettes by their distance from the prototype. As withimage variability, distances from the prototype are meas-ured along each PC as a z-score (number of standard devi-ations) of that PCs coefficients in the silhouette face spaceparameterization [Davidenko, 2007] and we refer to theseas PC units. The coefficients of PCs 2 and 6 also varied byup to %1.5 PC and %2 PC units, respectively, to generatedifferent exemplar silhouettes across blocks and runs. Weselected PCs 2 and 6 for this because they were found tobe least correlated with social attributes of faces such asgender and race [Davidenko and Ramscar, 2006]. Eachblock thus contained stimuli sampled from 12 differentdimensions of face space, covering a broad range of facesilhouettes. Importantly, image variability was notmatched across CH blocks. Instead, mean image variabilityhad values of 3.2, 4.0, 5.5, 7.3, 9.2, 11.2, 13.2, 15.3, and 17.3PC units: as blocks of stimuli deviated more from the pro-totype face silhouette, within-block image variabilityincreased (see Fig. 3a and Supporting Information Fig. 2).

Matched Variability (MV) Sampling(Experiment 4)

Here, we defined nine different blocks of face silhouettesby sampling along both directions of PC 3, a dimensionassociated with face gender [Davidenko, 2007], at 0, %3,%6, %9, and %12 PC units from the prototype (see Fig. 3b).We selected PC 3 because it produced a representativerange of face-likeness ratings in a pilot study (see Support-ing Information Fig. 3), but we would predict similarresults if other dimensions were used. Different exemplars

in each block were generated by varying the coefficient ofPCs 2 and 6 by up to %1.5 and %2 PC units, respectively,as with CH sampling. Because the size of each samplingregion was kept constant, image variability was matchedat 3.2 PC units across all blocks, equivalent to the lowestvariability blocks in CH sampling (Experiment 3).

For each subject in Experiments 3 and 4, we acquiredeight runs where subjects observed 9 12-s image blocksalternating with 12-s fixation blocks. Each run includedone block at each distance from the prototype face. In eachrun, the blocks were presented in an order that minimizedany cross-block adaptation. Specifically, in Experiment 3(CH sampling), the general order of blocks (labeled by dis-tance from the prototype face, in PC units) was 0, 7.5, 1.5,9, 3, 10.5, 4.5, 12, and 6. Across runs, this order was alter-nately reversed, and the starting block was randomized.Similarly, in Experiment 4 (MV sampling), the generalorder of blocks was &12, 3, &9, 6, &6, 9, &3, 12, 0, keepingthe distance in face-space between consecutive blocksnearly constant to avoid any systematic biases in block-to-block similarity that might differentially affect adaptationin some blocks. The stimulus display method, timing, andone-back task were identical to those in Experiments 1 and2.

Perceived Face-Likeness

Following the fMRI scans, nine of the subjects weregiven a questionnaire that included 54 face silhouettes(three from each of the MV and CH blocks) as well asthree additional two-tone shape stimuli that matched theface silhouettes in size, contrast, and number of keypointsbut did not resemble faces (Fig. 4c). Subjects provided a 1–5 face-likeness rating, where 1 indicated ‘‘Not at all face-like’’ and 5 indicated ‘‘Completely face-like.’’ As shown inFigure 4a,b, stimuli near the prototype were rated ashighly face-like, and ratings decreased gradually with dis-tance from the prototype. Although this was true for allPCs, equal distances along different PCs (or along differ-ent directions of the same PC) can yield slightly differentsubjective ratings of face-likeness. For example, face-like-ness ratings decrease faster along the positive direction ofPC3 than along the negative direction (Fig. 4b). This sug-gests that face-likeness ratings may be better than distancefrom the prototype in characterizing brain responses inrelation to face perception. Thus, in subsequent fMRI anal-yses, responses are plotted as a function of the mean face-likeness rating associated with each silhouette block.Importantly, even though face silhouettes in CH and MVsampling were defined along different PCs, face-like rat-ings covered a similar range across studies (Fig. 4a,b),from a mean % SEM rating of 4.9 % 0.1 for stimuli closestto the prototype to 2.5 % 0.2 for stimuli farthest from theprototype. We note that the lowest ratings on face silhou-ettes did not reach the floor of the scale. Only the nonface

r Davidenko et al. r

r 6 r

Page 7: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

shapes received a rating of 1 (‘‘not at all face-like’’), andthis was observed for every subject (Fig. 4c).

RESULTS

Face Silhouettes Elicit a Face-Selective ResponseProfile in Face-Selective Regions

To test the utility of face silhouettes as a tool for fMRIinvestigations of face perception, we first examinedwhether face silhouettes elicit a face-selective responseprofile. In Experiment 1 (8 subjects), we comparedresponses in face- and object-selective regions to face sil-houettes, two-tone shapes that were controlled for low-level properties, and scrambled two-tone images (seeMaterials and Methods). In Experiment 2 (9 subjects), wecompared responses to blocks of upright face silhouettesto the same silhouettes presented upside-down.

We extracted the mean response in each image condi-tion from each subject’s ROIs defined from the independ-ent localizer (see Materials and Methods) and averagedthese responses across subjects. In Experiment 1, all face-selective ROIs responded significantly more strongly dur-ing blocks of face silhouettes than during blocks of two-tone shapes (ts > 2.7, Ps < 0.03; paired t-test, two-tailed)and scrambled two-tone images (ts > 4.3, P < 0.02, Fig.5a). The voxels responding significantly more to face ver-sus shape silhouettes (P < 0.01) overlapped with the inde-pendently localized face-selective ROIs (see SupportingInformation Fig. 4).

In Experiment 2, all face-selective ROIs responded sig-nificantly more strongly to upright than to upside-downface silhouettes (ts > 3.3, Ps < 0.01, Fig. 5a), demonstratingan fMRI face-inversion effect with face silhouettes. In con-trast, responses in object-selective regions (pFus/OTS andLO) were higher for intact vs. scrambled two-tone images(ts > 2.1, Ps < 0.05) but similar across face silhouettes andtwo-tone shapes in Experiment 1 (|ts| < 1.1, Ps > 0.3)and between upright and upside-down face silhouettes inExperiment 2 (|ts| < 0.6, Ps > 0.5; Fig. 5b). These datacomplement our work showing face silhouettes’ effective-ness as stimuli for behavioral tasks [Davidenko, 2007] andprevious fMRI studies that have shown higher responsesin face-selective regions when subjects report a face per-cept in Rubin’s face/vase stimuli [Andrews et al., 2002;Hasson et al., 2001]. Taken together, these results indicatethat face silhouettes elicit a face-selective response profileand are therefore useful stimuli for fMRI studies of facerepresentation.

Concentric Hypersphere Sampling: Responses inBoth Face- and Object-Selective are Negatively

Correlated With Face-Likeness

We measured responses in face-selective regions,defined from the independent localizer, as subjectsobserved silhouettes at nine different levels of face-likenesswithin a five-point scale, ranging from 2.5 (low face-like-ness) to 4.9 (high face-likeness; see Fig. 4a and Materialsand Methods). To our surprise, we found that average

Figure 4.Face-like ratings decrease as stimuli deviate from the prototypeface for both CH and MV sampling. (a) Face-like ratings on stim-uli from the CH blocks (Experiment 3) provided outside thescanner by nine subjects who participated in the fMRI experi-ments, plotted as a function of distance from the prototypeface, measured in PC units; (b) Face-like ratings on stimuli fromthe MV blocks (in Experiment 4) provided by the same nine sub-

jects. X-axis indicates distance from the prototype face alongPC3 in the negative (gray) and positive (black) directions. (c)Ratings on similarly constructed nonface shapes were uniformly1 across subjects. Face-likeness scale: 1 # Not at all face-like to5 # Completely face-like. Error bars denote between-subjectsSEM.

r Face-Likeness and Image Variability r

r 7 r

Page 8: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

responses across subjects in fusiform and IOG face-selec-tive ROIs were significantly negatively correlated with per-ceived face-likeness (rRight mFus-faces # &0.91, P < 0.001;rRight pFus-faces # &0.88, P < 0.01; rRight IOG-faces # &0.9, P <

0.001; rLeft pFus-faces # &0.85, P < 0.01; see Fig. 6a). That is,responses in face-selective regions decreased as face-like-ness increased. In fact, the strongest responses in rightmFus-faces and right pFus-faces were observed duringblocks of stimuli rated by the same observers as being theleast face-like. Moreover, this puzzling pattern ofresponses was also present in object-selective regions. Av-erage responses in object-selective regions were also nega-tively correlated with perceived face-likeness (rRight pFus/

OTS # &0.94, P < 10&4; rRight LO # &0.97, P < 10&4; rLeftpFus/OTS # &0.85, P < 0.01; rLeft LO # &0.91, P < 0.001; seeFig. 6b) even though these regions did not show face-selec-tive responses in Experiments 1 and 2. When we calcu-lated the number of voxels within these ROIs activated byeach stimulus condition versus fixation (P < 0.01), wefound similar results: the number of active voxelsdecreased as perceived face-likeness increased (Ps < 0.01in every ROI; see Materials and Methods and SupportingInformation Fig. 5).

Why are responses in these ventral regions negativelycorrelated with perceived face-likeness?

We suggest that these counterintuitive results are notdriven by the perceived face-likeness of stimuli itself, butby the image variability that differs across stimulusblocks. Research on fMRI-adaptation [Andrews andEwbank, 2004; Avidan et al., 2002; Grill-Spector and Mal-ach, 2001; Grill-Spector et al., 1999, 2006; Li et al., 1993;Yovel and Kanwisher, 2004] has shown that responses inhigh-level visual areas, including these face- and object-selective regions, are reduced (adapted) when presentedrepeatedly with identical stimuli. Further, fMRI-A effectsgeneralize across similar stimuli. In other words,researchers report more fMRI-A to similar stimuli than todissimilar stimuli, although precisely how the degree ofstimulus similarity determines the level of fMRI-A isdebated and differs across brain regions [Gilaie-Dotanand Malach, 2007; Gilaie-Dotan et al., 2010; Weiner et al.,2010]. In CH sampling, highly face-like silhouettes aresampled from relatively small regions of face space nearthe prototype (Fig. 3a), resulting in low variability blocksthat may generate considerable fMRI-A. In contrast, lowface-likeness blocks further from the prototype aredefined in large regions of face space and are more vari-able, which may lead to less fMRI-A and consequently astronger signal. Indeed, there is a high negative correla-tion between mean image variability and mean rated

Figure 5.Responses in face-selective (but not object-selective) regions arehigher for face silhouettes than both shape silhouettes andupside-down face silhouettes. Responses in face-selective (a)and object-selective ROIs (b) to upright face silhouettes (black),two-tone shapes (white), scrambled two-tones (striped), and

upside-down silhouettes (gray) in the face silhouette study(Experiments 1 and 2). Ns indicate number of subjects in whicheach ROI was localized in the independent localizer scan. Aster-isks indicate responses significantly different than upright silhou-ettes (*P < 0.05; **P < 0.001; paired t-test, two-tailed).

r Davidenko et al. r

r 8 r

Page 9: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

face-likeness ratings (r # &0.99, P < 10&6). Therefore, itis possible that image variability contributed to the differ-ential signal amplitudes across blocks. To test this hy-pothesis and isolate the effect of face-likeness fromeffects of variability on responses, we conducted anotherexperiment where we manipulated face-likeness as in CHsampling but kept image variability constant acrossblocks.

Matched Variability Sampling: Responses inFace- (But Not Object-) Selective Regions

are Positively Correlated With Face-Likeness

We measured activity in the same face- and object-selec-tive ROIs in nine subjects (seven overlapping with Experi-ment 3) while they observed blocks of face silhouettesdefined at the same distances from the prototype and with

Figure 6.CH sampling; responses in both face- and object-selective ROIsdecrease as stimuli become more face-like. Each point reflectsthe average response across subjects to a block of stimuli. r- andP-values denote correlations and their significance, respectively,between average face-likeness ratings and average percentagesignal across subjects; error bars denote between-subjects SEM.R2 indicates proportion variance explained. (a) Average percent-age signal change across subjects in face-selective ROIs in CH

sampling across subjects as a function of perceived face-likenessof stimuli; Number of subjects in which each ROI was localized:Right mFus-faces (n # 9), Right pFus-faces (n # 10), Right IOG-faces (n # 8); Left pFus-faces (n # 10); (b) Average percentagesignal change across subjects in object-selective ROIs in CHsampling. Right pFus/OTS (n # 8), Right LO (n # 10), Left pFus/OTS (n # 6), Left LO (n # 8).

r Face-Likeness and Image Variability r

r 9 r

Page 10: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

comparable ratings of face-likeness as during CH sam-pling. The key difference here was that silhouettes in eachblock were defined in equal-size regions in face spacesuch that image variability was matched across face-like-ness levels (MV sampling, see Fig. 3b and Materials andMethods). As anticipated, and in stark contrast to the CHsampling results, responses in face-selective ROIsincreased with the perceived face-likeness of the stimuli(Fig. 7a). That is, blocks with higher face-like ratings of sil-

houettes produced higher responses in face-selectiveregions. Average responses across nine subjects were sig-nificantly positively correlated with the average perceivedface-likeness of stimuli in each of these face-selective ROIs(rRight mFus-faces # 0.83, P < 0.01; rRight pFus-faces # 0.97, P <

10&5; rRight IOG-faces # 0.87, P < 0.01; rLeft pFus-faces # 0.91, P< 0.001). In contrast, mean responses in object-selectiveROIs were not significantly modulated by perceived face-likeness of silhouettes (all Ps > 0.1; Fig. 7b) and were

Figure 7.MV sampling; responses in face-selective (but not object-selec-tive) ROIs increase as stimuli become more face-like. Each pointreflects the average response across subjects to a block of stim-uli. r- and P-values denote correlations and their significance,respectively, between face-likeness and average percentage signalacross subjects; error bars denote between-subjects SEM. (a)Average percentage signal change in face-selective ROIs in MV

sampling across subjects as a function of rated face-likeness ofstimuli; Number of subjects in which each ROI was localized:Right mFus-faces (n # 9), Right pFus-faces (n # 9), Right IOG-faces (n # 6); Left pFus-faces (n # 9); (b) Average percentagesignal change in object-selective ROIs; Right pFus/OTS (n # 6),Right LO (n # 9), Left pFus/OTS (n # 7), Left LO (n # 7).

r Davidenko et al. r

r 10 r

Page 11: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

consistently high for all blocks. Thus, in MV-sampling wefind a correlation between average face-likeness ratingsand average responses of face-selective regions specifically.Our voxel-count measure within each ROI produced simi-lar results: more face-selective voxels were active duringhigh face-likeness blocks than during low face-likenessblocks, but the number of active voxels in object-selectiveROIs was unaffected by perceived face-likeness (see Sup-porting Information Fig. 6).

A Powerful Effect of Image Variability on Face-and Object-Selective Responses

At first glance, the response patterns of face-selectiveROIs during CH vs. MV sampling of face-space appear con-tradictory: with CH sampling, average responses acrosssubjects in face-selective regions were negatively correlatedwith perceived face-likeness, whereas with MV sampling,the correlation was highly positive. We suggest that theapparent contradictory responses are due to the differencesin image variability across experiments. To test this possibil-ity and quantify the role of image variability, we examinedresponses across the seven subjects who participated in bothstudies as a function of image variability. We found thatblocks with more variable silhouettes images elicited higherresponses than blocks with less variable (more similar) sil-houettes, across most face- and object-selective ROIs (seeFig. 8). In right-hemisphere object-selective ROIs, averageresponses across subjects were significantly correlated withaverage image variability of stimuli in a block across the 18conditions in the CH and MV sampling experiments (rRightpFus/OTS # 0.82, P < 10&7; rRight LO # 0.93, P < P < 10&8),whereas these correlations in left-hemisphere object-selec-tive ROIs were smaller (rLeft pFus/OTS # 0.33, P # 0.2; rLO #

0.54, P # 0.02; see Fig. 8b). Average responses in right-hemi-sphere face-selective regions were also positively correlatedwith image variability, but to a lesser degree than in rightobject-selective ROIs (rRight mFus-faces # 0.76, P < 0.001; rRightpFus-faces # 0.52, P < 0.01; rRight IOG-faces # 0.66, P < 10&4; rLeftpFus-faces # 0.29, P # 0.3; see Fig. 8a). These data suggest thatimage variability accounts for some of the response modula-tions across the CH and MV experiments.

ATwo Factor Model Explains Responsesin Face-Selective Regions

We note that image variability alone cannot account forresponse modulations in face-selective regions, since imagevariability was constant across the nine MV samplingblocks and yet responses in face-selective regions weremodulated. We therefore asked whether responses in face-selective ROIs across the two studies could be explainedas a linear combination of two factors: (1) the mean per-ceived face-likeness (FL) of the stimuli and (2) the imagevariability (IV) in each block. We coded each of the 18blocks across the CH and MV sampling conditions accord-

ing to their average face-likeness rating and their imagevariability and analyzed brain responses from the 7 partic-ipants who completed both studies. In face-selective ROIs,we found that both factors contribute a significant positiveweight on responses. Furthermore, image variability con-tributed a numerically larger weight than face-likeness(right mFus-faces: WIV # 0.24, WFL # 0.12; right pFus-faces: WIV # 0.20, WFL # 0.18; right LO-faces: WIV # 0.40,WFL # 0.28; left pFus-faces: WIV # 0.25, WFL # 0.24). To-gether, face-likeness and image variability explained 78%of the variance of the responses across the 18 blocks inright mFus-faces, 87% in right pFus-faces, 79% in rightIOG-faces, and 51% in left pFus-faces (Fig. 9). In contrast,responses in object-selective ROIs were accounted for bythe single factor of image variability, as face-likeness didnot contribute additional explanation of the variance ofobject-selective responses (Ps > 0.1). Image variability pro-vided a better account of responses in right than leftobject-selective regions. It accounted for 68 and 87% of thevariance in responses in right pFus/OTS and LO, respec-tively and 11 and 29% for the left pFus/OTS and LO.

These analyses illustrate (1) a positive effect of face-like-ness on responses of face-selective regions only and (2) apositive effect of image variability on responses of bothface- and object-selective ROIs. With CH sampling, thesetwo factors were negatively correlated and produced com-peting effects on the responses of face-selective regions,with the effect of image variability overriding that of face-likeness. Thus, during CH sampling responses ultimatelyincreased with image variability despite decreasing face-likeness because the weight of image variability wasnumerically larger. This analysis also explains the unex-pected negative relationship between face-likeness and av-erage responses in object-selective cortex during CHsampling, suggesting that this correlation was in factdriven by differences in image variability across blocks(which were lowest for the most face-like blocks).

DISCUSSION

By manipulating stimuli in a parameterized silhouetteface space, we showed that two independent factors—face-likeness and image variability—drive up responses inface-selective ventral regions. Four face-selective regions(right mFus-faces, pFus-faces, IOG-faces, and left pFus-faces) showed a monotonically graded response to the per-ceived face-likeness of stimuli, responding more stronglyto stimuli that were more face-like. However, this responseprofile was only evident when image variability was keptconstant across blocks via MV sampling. When blockswere defined with CH sampling, and variability was nega-tively correlated with face-likeness, the effect of image var-iability dominated and responses were highest during thehigh-variability blocks despite their low face-likeness.Notably, this driving effect of image variability was pres-ent across multiple face- and object-selective ventral

r Face-Likeness and Image Variability r

r 11 r

Page 12: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

regions, whereas the effect of face-likeness was onlyobserved in face-selective ROIs.

Responses in Face-Selective Regions TrackPerceived Face-Likeness

When image variability was matched across blocks, wefound a striking correlation between face-likeness and av-

erage responses in face-selective regions. Moreover, theseregions tracked the face-likeness of stimuli monotonically.These results have important implications for the neuralmechanisms underlying face representation in the ventralstream. One possibility is that face-selective neurons inthese regions show graded responses, responding morestrongly to stimuli that more closely resemble a prototypi-cal face. A second possibility is that this graded response

Figure 8.Responses in face- and object-selective regions across subjectswho participated in both the CH sampling (open circles) andMV sampling (black dots) experiments, plotted as a function ofimage variability. Each point reflects the average response acrosssubjects to a block of stimuli. r- and P-values denote correlationsand their significance, respectively, between image variability andaverage percentage signal across subjects; error bars denote

between-subjects SEM. (a) Average percentage signal changeacross subjects in four face-selective ROIs; Number of subjectsin which each ROI was localized across both studies: RightmFus-faces (n # 6), Right pFus-faces (n # 7), Right IOG-faces (n# 4); Left pFus-faces (n # 7); (b) Average percentage signal inobject-selective ROIs; Right pFus/OTS (n # 4), Right LO (n #

7), Left pFus/OTS (n # 3), Left LO (n # 5).

r Davidenko et al. r

r 12 r

Page 13: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

profile is the product of different neural populations tunedto different types of stimuli [Jiang et al., 2006], with moreneurons tuned to highly face-like stimuli and fewer neu-rons tuned to stimuli that are less face-like. These possibil-ities are not mutually exclusive: it may be that moreneurons are tuned close to the prototypical face, and allneurons show a preference for face-like stimuli. Althoughour studies cannot distinguish between these possibilities,any viable neural model of face representation mustaccount for a BOLD response that correlates with per-ceived face-likeness of stimuli.

Face-Specific and General Factors

Examining responses across multiple brain regionsallowed us to determine which manipulations of our stimulihave face-specific consequences and which manipulationsaffect responses more generally across ventral occipito-tem-poral regions. We show that face likeness is a face-specificfactor: When variability was controlled (MV sampling) face-selective responses increased with face-likeness, but object-selective responses were not modulated by face-likeness. Byequating image variability across different conditions, wewere able to study the effect of face-likeness in isolation andshow a modulation in face-selective responses only.

In contrast, image variability is a general factor affectingresponses in object- as well as face-selective regions.Responses in LO and pFus/OTS, which show no selectivityto faces, increased as a function of image variability (Fig. 8).This effect of variability was large and prevalent across ven-tral occipito-temporal regions. For example, in mFus-facesand pFus-faces, stimuli that were rated as least face-like eli-cited twice as large of a response in CH sampling than inMV sampling, because the low face-like stimuli in CH sam-

pling were much more variable (Figs. 6 and 7, lowest face-likeness point). In fact, image variability had such a largeeffect on responses of face-selective regions that it reversedtheir preference for face-like stimuli in CH sampling, wherevariability was anti-correlated with face-likeness (Fig. 6).

Our data indicate that responses in face-selective regionsare driven by at least two factors: face-likeness and imagevariability, and studying their effects requires dissociatingthe two factors experimentally. The face silhouette method-ology is well suited for this purpose because the sameshape-based parameterization can be used to manipulate aface-specific factor (in this case, face-likeness), while control-ling the general factor of image variability. Face silhouettescapture a subset of the face dimensions that are relevant forface perception tasks; specifically, dimensions related toface profile shape [Davidenko, 2007]. Nevertheless, the facesilhouette methodology cannot address every aspect of facerepresentation. Previous research indicates that facial char-acteristics other than shape, including local features, tex-tures, and surface reflectance information [O’Toole et al.,1999; Russell and Sinha, 2007; Schyns et al., 2002] also influ-ence face perception. Thus, future work is needed to under-stand the effects of other face characteristics on face-selective responses. Critically, however, our data indicatethat future examinations of the functional characteristics offace-selective regions, as well as ventral stream regionsmore generally, must include a method of measuring andcontrolling the variability of stimuli across conditions.

Effects of Different Measures of Variability inFace Silhouettes

In our stimulus space, image variability was defined asthe average distance in face-space between stimuli in each

Figure 9.Correlation between predicted responses based on the two-fac-tor model and observed percentage signal change in face-selec-tive ROIs across subjects who participated in both the CHsampling (open circles) and MV sampling (black dots) experi-ments. Each point reflects the average response across subjectsto a block of stimuli. X-axis: predicted % signal based on twofactors (face-likeness and image variability); Y-axis: observed aver-

age % signal across subjects; R2 indicates proportion of varianceexplained across the 18 blocks from the two studies, and p-val-ues indicate significance of the linear fit; error bars indicatebetween-subjects SEM; Number of subjects in which each ROIwas localized across both studies: Right mFus-faces (n # 6),Right pFus-faces (n # 7), Right IOG-faces (n # 4); Left pFus-faces (n # 7).

r Face-Likeness and Image Variability r

r 13 r

Page 14: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

block, but other physical and perceptual measures of vari-ability can be defined. For example, a ‘‘pixel-based’’ metric(see Materials and Methods) is generalizable to other stim-ulus spaces, as it can be defined on gray-scale or colorimages. For our stimuli, pixel-based variability was highlycorrelated with the face-space measure of image variability(r # 0.96, P < 10&9 across the 18 stimulus blocks; see Sup-porting Information Fig. 7a) and consequently similarlyexplained our fMRI data. Since recent studies suggest thatresponses in some high-level visual regions are coupledwith perceptual rather than physical variability amongstimuli [Gilaie-Dotan et al., 2010; Jiang et al., 2006; Rotsh-tein et al., 2005], we also examined whether perceptualmetrics of variability better explain our data. We consid-ered a perceptual metric of variability by behaviorallymeasuring dissimilarity ratings among stimuli in eachblock (see Materials and Methods). We found that meandissimilarity ratings were highly correlated with imagevariability (r # 0.97, P < 10&10; Supporting InformationFig. 7b) as well as with pixel-based variability (r # 0.88, P< 10&5; Supporting Information Fig. 7c). Thus, in our stim-ulus space, these three metrics of variability are tightlycoupled. However, this coupling may not generalize toother stimuli. For instance, manipulating image size withina block of stimuli will result in a large pixel-based vari-ability but relatively small perceptual variability. Further,different metrics of variability may affect various high-level visual areas differentially. Thus, in future studies,characterizing which metric of variability is relevant tocontrol will depend on both the stimulus domain and thebrain regions being investigated.

Implications of Our Data on UnderstandingFunctional Properties of High Level Visual

Cortex

Our results show that stimulus variability, whethermeasured in physical or perceptual units, is a critical fac-tor to consider in fMRI investigations of high-level visualregions. Because high-level visual areas adapt to similarstimuli, responses will generally be stronger to blocks ofhighly variable images compared to blocks of similarimages. Despite its prevalent effect in the ventral stream,many studies of face and object representation fail to con-sider the role of variability. When researchers examine theresponse properties of face-selective regions, they oftenfocus on one factor: e.g., gender or distinctiveness. Ourresults indicate that focusing on one factor is insufficientunless stimuli are matched for variability. When variabilityis not controlled, results attributed to the manipulated fac-tor may be misinterpreted if the factor is confounded withstimulus variability. For example, a recent study [Freemanet al., 2010] reported that FFA responds more strongly togendered (as opposed to androgynous) blocks of faces.However, since blocks of gendered faces included bothmale and female faces, they were likely more variable than

the blocks of androgynous faces. This raises the possibilitythat the apparent stronger FFA responses to genderedfaces were in fact driven by image variability. Another im-portant example is a study by Loffler et al. [2005] thatused synthetic faces [Wilson et al., 2002] to manipulateface distinctiveness and found that FFA responses werestronger to blocks of distinctive faces (far from the proto-type) as compared to blocks of typical faces (close to theprototype). This stronger FFA response to distinctive faceshas been interpreted as evidence for a norm-based neuralcoding of faces in which individual neurons respond morestrongly as faces deviate more from the prototype, ornorm face [Panis et al., 2011; Rhodes, 2006]. However, theblocks in Loffler et al.’s study were generated using thesame CH sampling method we employed in Experiment 3,rendering distinctiveness highly correlated with variability.In other words, blocks of distinctive faces were more vari-able than blocks of typical faces. As our data suggest, Lof-fler et al.’s results may have been driven by differences invariability across blocks rather than differences in face dis-tinctiveness itself. Further, because they did not reportresponses in other ventral regions such as LO and pFus/OTS, it is unknown whether their effects reflect a face-spe-cific mechanism or instead reflect a more general adapta-tion effect across the ventral stream. In an effort to ruleout an adaptation-based explanation, Loffler and col-leagues showed V1 responses were not modulated by theirmanipulation of distinctiveness. However, we note thatearly visual areas, including hV4, are not effective controlregions because they do not show fMRI-A to shapes orfaces [Sayres and Grill-Spector, 2006; Weiner et al., 2010].Indeed, in the present experiment, we identified hV4based on visual field mapping [Sayres and Grill-Spector,2008] in five subjects in CH sampling and six subjects inMV sampling and extracted responses in each stimuluscondition. First, we found no significant modulation ofhV4 responses in either of the CH or MV experiments (seeSupporting Information Fig. 8a,b). Second, in contrast toLO and pFus responses which show fMRI-A thatdepended on image variability, hV4 responses were notmodulated by image variability (Supporting InformationFig. 8c). This is because image variability modulatesresponses only if there is fMRI-A across similar stimuli.

As researchers continue to debate the mechanismsunderlying the neural coding of faces and compare the ef-ficacy of norm-based [Leopold et al., 2006; Loffler et al.,2005; Rhodes, 2006] versus exemplar-based [Gilaie-Dotanand Malach, 2007; Gilaie-Dotan et al., 2010; Jiang et al.,2006] accounts, it is critical to decouple distinctivenessfrom image variability when examining its effect on face-selective responses. As our studies show, one way thismay be accomplished is by generating matched-variabilityblocks of stimuli.

Our experiments show that stimulus variability affectsresponses across many ventral regions, not just face-selec-tive ROIs. This suggests that measuring and controllingvariability should become a critical tool in studies of face

r Davidenko et al. r

r 14 r

Page 15: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

perception as well as object representation more generally.Future studies may ask whether an object-selective regionsuch as LO responds more strongly to animate versus in-animate objects, or whether face-selective regions respondpreferentially to familiar versus unfamiliar faces. Thesequestions cannot be well addressed without properlyaccounting for different degrees of variability across stimu-lus conditions. This challenge pertains not only to block-design experiments. While adaptation effects are smalleracross events separated temporally relative to blocks, theyare nevertheless present and persist across many interven-ing stimuli [Andresen et al., 2009; Sayres and Grill-Spector,2006; Weiner et al., 2010]. Although a recent method ofcarry-over design [Aguirre, 2007] partially ameliorates im-mediate adaptation effects across stimuli, it does not con-trol for adaptation effects across many intervening stimuli.Therefore, the effect of image variability is important toconsider in any experiment in which responses are meas-ured across a group of stimuli, including event-relatedexperiments and multi-voxel pattern analyses. Determin-ing appropriate measures of image variability and describ-ing their effects on the responses of different brain regionswill remain a key research goal for future neuroimagingstudies of high-level visual regions.

ACKNOWLEDGMENTS

The authors thank Nathan Witthoft, Jon Winawer, andGolijeh Golarai for useful discussions during the design ofthe experiments and preparation of the manuscript.

REFERENCES

Aguirre J (2007): Continuous carry-over designs for fMRI. Neuro-image 35:1480–1494.

Andresen DR, Vinberg J, Grill-Spector K (2009): The representa-tion of object viewpoint in human visual cortex. Neuroimage45:522–536.

Andrews TJ, Ewbank MP (2004): Distinct representations for facialidentity and changeable aspects of faces in the human tempo-ral lobe. Neuroimage 23:905–913.

Andrews TJ, Schluppeck D, Homfray D, Matthews P, BlakemoreC (2002): Activity in the fusiform gyrus predicts conscious per-ception of Rubin’s vase-face illusion. Neuroimage 17:890–901.

Avidan G, Harel M, Hendler T, Ben-Bashat D, Zohary E, MalachR (2002): Contrast sensitivity in human visual areas and itsrelationship to object recognition. J Neurophysiol 87:3102–3116.

Barton JJS, Press DZ, Keenan JP, O’Connor M (2002): Lesions ofthe fusiform face area impair perception of facial configurationin prosopagnosia. Neurology 58:71–78.

Brainard DH (1997): The psychophysics toolbox. Spatial Vis10:433–436.

Davidenko N (2007): Silhouetted face profiles: a new methodologyfor face perception research. J Vis 7:6.

Davidenko N ( 2007): The use of parameterized stimulus spacesfor the study of face representation. Dissertation AbstractsInternational: Section B: The Sciences and Engineering.67:5434.

DavidenkoN, RamscarMJA (2006): The distinctiveness effect reverseswhen usingwell-controlled distractors. Vis Cogn 14:89–92.

Davidenko N, Witthoft N, Winawer J (2007): Gender aftereffectsin face silhouettes reveal face-specific mechanisms. Vis Cogn16:99–103.

Downing PE, Chan AW-Y, Peelen MV, Dodds CM, Kanwisher N(2006): Domain specificity in visual cortex. Cerebral Cortex16:1453–1461.

Ewbank MP, Andrews TJ (2008): Differential sensitivity for view-point between familiar and unfamiliar faces in human visualcortex. Neuroimage 40:1857–1870.

Freeman JB, Rule NO, Adams RB, Ambady N (2010): The neuralbasis of categorical face perception: Graded representations offace gender in fusiform and orbitofrontal cortices. Cereb Cor-tex 20:1314–1322.

Gauthier I, Tarr MJ, Moylan J, Skudlarski P, Gore JC, AndersonAW (2000): The fusiform ‘‘face area’’ is part of a network thatprocesses faces at the individual level. J Cogn Neurosci12:495–504.

Gilaie-Dotan S, Malach R (2007): Sub-exemplar shape tuning inhuman face-related areas. Cereb Cortex 17:325–338.

Gilaie-Dotan S, Gelbard-Sagiv H, Malach R (2010): Perceptualshape sensitivity to upright and inverted faces is reflected inneuronal adaptation. NeuroImage 50:383–395.

Glover GH (1999): Simple analytic spiral K-space algorithm. MagnReson Med 42:412–415.

Golarai G, Ghahremani DG, Whitfield-Gabrieli S, Reiss A,Eberhardt JL, Gabrieli JD, Grill-Spector K (2007): Differentialdevelopment of high-level visual cortex correlates withcategory-specific recognition memory. Nat Neurosci 10:512–522.

Golarai G, Liberman A, Yoon JM, Grill-Spector K (2010): Differen-tial development of the ventral visual cortex extends throughadolescence. Front Hum Neurosci 3:80.

Golby AJ, Gabrieli JD, Chiao JY, Eberhardt JL (2001): Differentialresponses in the fusiform region to same-race and other-racefaces. Nat Neurosci 4:845–850.

Grill-Spector K (2003): The neural basis of object perception. CurrOpin Neurobiol 13:159–166.

Grill-Spector K, Malach R (2001): fMR-adaptation: A tool forstudying the functional properties of human cortical neurons.Acta Psychol (Amst) 107:293–321.

Grill-Spector K, Kushnir T, Edelman S, Avidan G, Itzchak Y, Mal-ach R (1999): Differential processing of objects under variousviewing conditions in the human lateral occipital complex.Neuron 24:187–203.

Grill-Spector K, Knouf N, Kanwisher N (2004): The fusiform facearea subserves face perception, not generic within-categoryidentification. Nat Neurosci 7:555–562.

Grill-Spector K, Henson R, Martin A (2006): Repetition and thebrain: Neural models of stimulus-specific effects. Trends CognSci 10:14–23.

Hasson U, Hendler T, Ben Bashat D, Malach R (2001): Vase orface? A neural correlate of shape-selective grouping processesin the human brain. J Cogn Neurosci 13:744–753.

Jiang X, Rosen E, Zeffiro T, Vanmeter J, Blanz V, Riesenhuber M(2006): Evaluation of a shape-based model of human face dis-crimination using FMRI and behavioral techniques. Neuron50:159–172.

Kanwisher N, McDermott J, Chun MM (1997): The fusiform facearea: A module in human extrastriate cortex specialized forface perception. J Neurosci 17:4302–4311.

Kranz F, Ishai A (2006): Face perception is modulated by sexualpreference. Curr Biol 16:63–68.

r Face-Likeness and Image Variability r

r 15 r

Page 16: F ace-Lik eness and Ima g e V ariability Driv e Responses ...vpnl.stanford.edu/papers/Davidenko_Remus_GrillSp... · r H uman B rain M apping 00:000Ð000 (2011) r F ace-Lik eness and

Leopold DA, Bondar IV, Giese MA (2006): Norm-based faceencoding by single neurons in the monkey inferotemporal cor-tex. Nature 442:572–575.

Li L, Miller EK, Desimone R (1993): The representation of stimu-lus familiarity in anterior inferior temporal cortex. J Neurophy-siol 69:1918–1929.

Loffler G, Yourganov G, Wilkinson F, Wilson HR (2005): fMRI evi-dence for the neural representation of faces. Nat Neurosci8:1386–1390.

Moutoussis K, Zeki S (2002): The relationship between corticalactivation and perception investigated with invisible stimuli.Proc Natl Acad Sci USA 99:9527–9532.

O’Toole AJ, Vetter T, Blanz V (1999): Three-dimensional shapeand two-dimensional surface reflectance contributions to facerecognition: An application of three-dimensional morphing.Vis Res 39:3145–3155.

Panis S, Wagemans J, Op de Beeck HP (2011): Dynamic norm-based encoding for unfamiliar shapes in human visual cortex.J Cogn Neurosci 23:1829–1843.

Phillips PJ, Wechsler H, Huang J, Rauss P (1998): The FERETdatabase and evaluation procedure for face recognition algo-rithms. Image Vis Comput J 16:295–306.

Phillips PJ, Moon H, Rizvi SA, Rauss PJ (2000): The FERET evalu-ation methodology for face recognition algorithms. IEEE TransPattern Anal Machine Intell 22:090–1104.

Puce A, Allison T, Gore JC, McCarthy G (1995): Face-sensitiveregions in human extrastriate cortex studied by functionalMRI. J Neurophysiol 74:1192–1199.

Rhodes G, Jeffery L (2006): Adaptive norm-based coding of facialidentity. Vis Res 46:2977–2987.

Rotshtein P, Henson RN, Treves A, Driver J, Dolan RJ (2005):Morphing Marilyn into Maggie dissociates physical andidentity face representations in the brain. Nat Neurosci 8:107–113.

Russell R, Sinha P (2007): Real world face recognition: The impor-tance of surface reflectance properties. Perception 36:1368–1374.

Sayres R, Grill-Spector K (2006): Object-selective cortex exhibitsperformance-independent repetition suppression. J Neurophy-siol 95:995–1007.

Sayres R, Grill-Spector K (2008): Relating retinotopic and object-selective responses in human lateral occipital cortex. J Neuro-physiol 100:249–267.

Schyns PG, Bonnar L, Gosselin F (2002): Show me the features!Understanding recognition from the use of visual information.Psychol Sci 13:402–409.

Tong F, Nakayama K, Vaughan J, Kanwisher N (1998): Binocularrivlary and visual awareness in human extrastriate cortex.Neuron 21:753–759.

Tong F, Nakayama K, Moscovitch M, Weinrib O, Kanwisher N (2000): Response properties of the human fusiform face area.Cogn Neuropsychol (Special Issue: The cognitive neuroscienceof face processing) 17:257–279.

Tootell RBH, Devaney KJ, Young JC, Postelnicu G, Rajimehr R,Ungerleider LG (2008): fMRI mapping of a morphed contin-uum of 3D shapes within inferior temporal cortex. Proc NatlAcad Sci USA 105:3605–3609.

Valentine T (1991): A unified account of the effects of distinctive-ness, inversion, and race in face recognition. Q J Exp PsycholA 43:161–204.

Weiner KS, Grill-Spector K (2010): Sparsely-distributed organiza-tion of face and limb activations in human ventral temporalcortex. Neuroimage 52:1559–1573.

Weiner K, Sayres R, Vinberg J, Grill-Spector K (2010): fMRI-adaptationand category selectivity in human ventral temporal cortex: Regionaldifferences across time scales. J Neurophysiol 103:3349–3365.

Wilson HR, Loffler G, Wilkinson F (2002): Synthetic faces, facecubes, and the geometry of face space. Vis Res 42:2909–2923.

Worsley KJ, Marrett S, Neelin P, Vandal AC, Friston KJ, Evans AC(1996): A unified statistical approach for determining significantsignals in images of cerebral activation. Hum Brain Mapp 4:58–73.

Yovel G, Kanwisher N (2004): Face perception: Domain specific,not process specific. Neuron 44:889–898.

r Davidenko et al. r

r 16 r