a uniﬁed account of the effects of distinctiveness, inversion, and race in face recognition

8/2/2019 A unified account of the effects of distinctiveness, inversion, and race in face recognition

1/46

This article was downloaded by: [McMaster University]On: 28 February 2012, At: 11:46Publisher: Psychology PressInforma Ltd Registered in England and Wales Registered Number:1072954 Registered office: Mortimer House, 37-41 Mortimer Street,London W1T 3JH, UK

The Quarterly Journal ofExper iment al PsychologySection A: HumanExper iment al PsychologyPublication details, including inst ruct ions for

authors and subscription information:http:/ / www.tandfonline.com/ loi/ pqj a20

A unif ied account of theeffects of dist inct iveness,inversion, and race in face

recognitionTim Valent ine aa

University of Manchester, U.K.

Available online: 29 May 2007

To cit e this art icle: Tim Valentine (1991): A unified account of the effects of

distinctiveness, inversion, and race in face recognition, The Quarterly Journalof Experimental Psychology Sect ion A: Human Experimental Psychology, 43:2,161-204

To link t o this art icle: http:/ / dx.doi.org/ 10.1080/ 14640749108400966

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions

This article may be used for research, teaching, and private studypurposes. Any substantial or systematic reproduction, redistribution,reselling, loan, sub-licensing, systematic supply, or distribution in anyform to anyone is expressly forbidden.
http://www.tandfonline.com/page/terms-and-conditionshttp://www.tandfonline.com/page/terms-and-conditionshttp://www.tandfonline.com/page/terms-and-conditionshttp://dx.doi.org/10.1080/14640749108400966http://www.tandfonline.com/loi/pqja20


2/46

The publisher does not give any warranty express or implied or makeany representation that the contents will be complete or accurate orup to date. The accuracy of any instructions, formulae, and drug dosesshould be independently verified with primary sources. The publishershall not be liable for any loss, actions, claims, proceedings, demand, orcosts or damages whatsoever or howsoever caused arising directly orindirectly in connection with or arising out of the use of this material.

Downloadedby[McMasterUniversity]at11:4628February2012


3/46

THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 1991,43A (2) 161-204

A Unified Account of the Effects ofDistinctiveness, Inversion, and Race in FaceRecognitionTim Valentine

University of Manchester, U . K .

A framework is outlined in which individual faces are assumed t o be encoded asa point in a multidimensional space, defined by dimensions that serve todiscriminate faces. It is proposed that such a framework can account for theeffects of distinctiveness, inversion, and race on recognition of faces. Twospecific models within this framework are identified: a norm-based codingmodel, in which faces are encoded as vectors from a population norm orprototype; and a purely exemplar-based model. Both models make similarpredictions, albeit in different ways, concerning the interactions between theeffects of distinctiveness, inversion an d race. These predictions were suppo rtedin five experiments in which photog raphs of faces served as stimuli. The norm -based coding version and the exem plar-based version of the framew ork can no tbe distinguished o n the basis of the experiments reported, but it is argued th at amultidimensional space provides a useful heuristic framework to investigaterecognition of faces. Finally, the relationship between the specific models isconsidered an d an implementation in terms of parallel distributed processing isbriefly discussed.

T he ra ted d i s tinc tiveness of a face, th e or ientat ion in which i t is seen, and t h erace of a face a re a l l fac tors kn ow n t o in fluence the ab i l ity of an observer tosubsequent ly recognize the face . The re has been a t endency for each o f thesef ac t o r s t o be i nves ti gat ed in i sola t ion largely appeal ing t o di f ferent theoret i-cal explana t ions . In the presen t pa pe r it is p ro po sed t ha t s imi la r ity be tweenfaces can account for the e f fec t s o f all th ree fac tors . F i r s t the cur ren tl i terature o n th e effects of dis t inct iveness a n d invers ion is reviewed brief ly . In

Requests for reprints should be sent to Tim Valentine. Department of Psychology,University of Manchester, Oxfo rd Roa d, Manchester M I 3 9PL , U . K .

The experiments reported here were carried out while the author was employed by theMedical Research Council at the MRC Applied Psychology Unit , 15 Chaucer Road, Cam-bridge. I would l ike to thank Gill ian Rhodes, G lyn H ump hreys, and anony mous reviewers forhelpful comments on an earlier draft of th e manuscript.

( 1991 Th e Experimental Psychology Society



4/46

162 VALENTINEthe second section a general framework is proposed in which it is assumedthat faces can be represented by a point in a m ultidimensional space. Withinthis frame work , two models can be identified. On e model assumes a norm o rprototype is abstracted, the other model is purely exemplar-based. Predic-tions based o n these mode ls concern ing the inte raction between the effects ofdistinctiveness an d inversion in face processing tasks are tested in Experi-ments 1-4. The multidimensional space framework is then extended toaccount for the effect of race. Experiment 5 tests a prediction derived fromthe framework concerning the interaction between the effects of race andinversion. In the General Discussion the relationship between the norm-based and purely exemplar-based versions of the framework and othermodels of race recognition are discussed.Any model of visual object recognition must specify how stored knowl-edge is used t o facilitate recognition of visual stimuli (Palmer, 1975). Palm ercited faces as a n exam ple of a perceptual category t ha t w ould include in itsrepresentation information about the prototypical values (or central tend-ency) of the relevant dim ensions. Such informa tion w ould b e specific t o faceprocessing. Fodor (1983) proposes a similar view of face processing in hissuggestion that faces are favourite candidates for an eccentric stimulusdom ain-that is, a dom ain whose perceptual analysis requires inform ationthat is highly specific to the d om ain in q uestio n. ( F odo r, 1983, pp. 51-52).The effect of distinctiveness on the recognition of faces provides someindication that category-specific knowledge, presumably acquired throughexperience of the population of faces previously seen, is used to facilitaterecognition of faces encountered subsequently. Generally, previously unfa-miliar faces that are rated as distinctive or unusual are more accuratelyrecognized in a recognition memory paradigm. Conversely, typical faces aremore likely to be incorrectly identified as having been seen before (G oing &Rea d, 1974; Coh en & C arr , 1975; Light, Kayra-S tuart, & Hollander, 1979;W inograd, 1981; Bartlett, Hu rry , & Th orley , 1984). An effect of distinctive-ness has been found in three measures of recognition mem ory performance-hit rate, false positive rate, and combined measures of sensitivity based onsignal detection theory. The exact nature of the effect varies slightly fromstudy to stud y, but generally the adva ntage foun d fo r distinctive faces in falsepositive rate and sensitivity measures is robust. The effect on hit rate seems tobe dependent upon the procedure and is found most often when the initialencoding conditions are good, involving trait judgement encoding activitiesan d/ or relatively long exposure times of 8 sec (Light et al., 1979).Light et al. (1979) prop osed a two-com ponent theory of mem ory based oninter-stimulus similarity in orde r to accoun t fo r the effect of distinctiveness.They suggested that a distinctive face is more likely to access a specificmemory because i t is less similar to specific memories of other faces. Thus



5/46

FACE RECOGNITION 163access to specific mem ory gives rise to a n ad van tage in hit rate for distinctivefaces. However, in the absence of a specific memory, Light and colleaguessuggested that subjects base their recognition judgement on schematicmemory of category structure (e.g. similarity to a prototype). Use ofschematic memory gives rise to the g reater false positive rate to typical faces.In contrast, Bartlett et al. (1984) interpreted the effect of distinctiveness interms of familiarity information alone (cf. Mandler, 1980). Bartlett andcolleagues suggested that presentation of a distinctive face resulted in agreater increment in familiarity than that resulting from presentation of atypical face. Valentine a nd Bruce (1986b) examined the relationship betweenthe effects of distinctiveness and familiarity in processing highly familiarfaces. It was found that both the rated distinctiveness and familiarity of aface affected the RT to decide it was familiar in a familiarity decision task.The m ore distinctive or the more fam iliar a face was rated, the faster it couldbe recognized. However, the two effects were found to be independent,suggesting that familiarity information does not form the basis of the effectof distinctiveness. (See Valentine, 1990, for fur the r discussion of the relation-ship between distinctiveness and familiarity.)Valentine an d Bruce (1 9 86 ~ ) rgued th at if the effect of distinctiveness onthe latency to recognize familiar faces arises from the role of a facialprototype, then distinctive faces should take longer than typical faces to beclassified as a face in a task in which faces must be distinguished fromjumbled faces. This prediction was confirmed. Thus distinctive faces arerecognized faster in a familiarity decision task but classified as faces moreslowly th an ar e typical faces. Valentine an d Bruce suggest that a general faceprototyp e is extracted from faces previously encoun tered an d t ha t individualfaces are stored as a set of transform ations required to match the face to th eprototype. This proposal was termed the prototype hypothesis.

Furthe r evidence of proto type abs tractio n in face recognition comes fromthe study of caricature in recognizing faces. Perkins (1975) describes carica-ture as a process of exaggerating the distinctive features that individuate aparticular face. Rhodes, Brennan, an d C arey (1 987) studied the recognitionof computer-generated caricatures. Veridical line drawings of faces weregenerated by joining the co-ord inates o f 169 specified points on each face bylines. A norm was generated by averaging the position of these points acrossseveral faces. Caricatures were then generated by increasing the distancebetween the location at each point in an individuals face and the norm by afixed pro portion . An anti-caricature could be generated by reducing thedifferences between a face and a norm. Rhodes and colleagues found thatcaricatures were recognized faster tha n veridical line drawings, which in turnwere recognized faster than anti-caricatures; however, there were no dif-ferences in accuracy in recognizing the three classes of faces. They interpret



6/46

164 VALENTINEtheir result as evidence of norm-based coding, a holistic encod ing process inwhich the distinctive aspects of faces are encoded by com parison to a normor average face.

Both Valentine and Bruce (1986b, 1986c) an d Rh odes et al. (1987) have,quite independently, proposed very similar accounts of the role of aprototype or norm in face recognition. There are clear parallels betweenthese hypotheses an d Goldstein an d Chan ces (1980) proposa l tha t the role ofa face schema can account for the development of the effect of race andinversion in face recognition. Adults recognize other-race faces less accur-ately than own-race faces. However, young children recognize other-racefaces and own-race faces with equal accuracy (Goldstein & Chance, 1980;Chance, Tu rne r, & Goldstein, 1982). An alogous results a re also foun d fo r theeffect of inversion on face recognition. Adults show a large effect of stimulusinversion o n their ability to recognize faces. Indeed, the effect of inversion onface recognition is disproportionately large compared to the effect ofinversion on recognition of other stimulus classes (Yin, 1969). However,young children show little effect of inversion upon recognition memory forfaces (Goldstein, 1975). Goldstein an d C hance a ccoun t for the developmen tof the effects of race and inversion by arguing that with increasing agechildren become m ore efficient in their use of a face schem a, leading to betterface recognition performance. However, this increase in efficiency is accom -panied by an increase in schema rigidity, so tha t as the schem a develops, itbecomes relatively less efficient at p rocessing unusual stimuli such a s invertedor other-race faces.

Goldstein and Chances (1980) face schema theory has the basic idea thatknowledge of the pop ulation of faces is acquired and used in face processingin comm on w ith the prototype hypothesis. However the evidence, from thedevelopment of the effects of race an d inversion on face recognition, used t osu pp or t schema theory is equivocal. Goldstein an d Cha nce (1 980) foundconsiderable improvem ent in face recognition ability across the age range 6-12 years but did n ot find an interaction between the effect of race a nd age inthis range. In a later stud y C hance et a] . (1982) did find an Age ofSubject x Race of Stimulus Face interaction in this age range. However,othe r studies (Cross, Cross, & Daley , 1971; Ka gan & Klein, 1973; Feinm an &Entwisle, 1976) have failed t o find an increase in the effect of race with age.Recent work on the development of the effect of inversion on face recogni-tion has also shown th at earlier work h as overstated the effect. Althou gh theability to recognize upright faces improves faster tha n th e ability to recognizeinverted faces, young children d o show a n effect of inversion, and the abilityto recognize inverted faces doe s improve with age (Flin, 1985; see Valen tine,1988, for a review of the effect of inversion on face recognition).



7/46

FACE RECOGNITION 165Diamond and Carey (1986) have demonstrated that recognition of

another stimulus class is as adversely affected by inversion as is facerecognition. It was found that dog experts showed a similarly large effect ofinversion on their ability to recognize individual dogs. Diamond and Careyargue that a large effect of inversion will be found if: (1) the exemplars of astimulus class have a common configuration but subtle differences in thespatial relations between the features (termed second-order relational infor-mation); and (2) the observers have sufficient expertise to distinguish the ,exemplars of the stimulus class on the basis of these differences in configuralinformation. Thus the disproportionate effect of inversion is not specific tofaces but likely to be found for recognition of any highly familiar and highlyhomogeneous stimulus class.In summary, the prototype hypothesis (Valentine & Bruce, 1986b, 1986c),the norm-based coding model (Rhodes et al., 1987), schema theory (Gold-stein & Chance, 1980), and Diamond and Careys (1986) study of inversion,together, suggest that the effects of distinctiveness, inversion, and race onface recognition could all be explained by the role of knowledge of thepopulation of faces previously encountered. Valentine and Bruce, Rhodes etal., and Goldstein and Chance have all proposed that a prototype, norm, orschema is abstracted. However, alternative accounts can also be formulated,which are based on inter-stimulus similarity but do not assume the existenceof a stored face prototype. For instance, this would be in keeping with recentmodels in the concept learning literature (e.g. Medin & Schaffer, 1978;Nosofsky, 1986).In the next section a framework is proposed in which faces are assumed tobe encoded as points in a multidimensional space. Two specific models areidentified. In one model faces are encoded by reference to an abstracted normor prototype. In order to avoid possible confusion with the different ways inwhich the term prototype has been used in the concept learning and facerecognition literature, I will adopt the term used by Rhodes and colleaguesand refer to this model as the norm-based coding model. The other modelidentified assumes that a norm is not abstracted and that only specific faces(or category exemplars) are stored. This will be referred to as the exemplar-based model. The term multidimensional space framework will be usedas a generic term to refer to both the norm-based and exemplar-basedmodels.The experiments reported do not aim to distinguish between the norm-based and exemplar-based models. Indeed, the two models make similarpredictions, although based on some different assumptions. The aim is toshow that knowledge of the population of faces can provide a parsimoniousaccount of the effects of distinctiveness, inversion, and race.



8/46

166 VALENTINEA Framework for Coding Processesin Face RecognitionThe main assumption of the proposed framework is that a location in aEuclidean multidimensional space provides an ap prop riate m etapho r for themental representation of a face. The dimensions of the space represent thephysiognomic features that are used to encode faces. N o attempt will bemade to identify the aspects of a face that the dimensions represent.However, previous work using multidimensional scaling techniques suggestthat the principal dimensions needed would represent hair colour an d length,face shape, and age (at least for the Caucasian faces, which are own-racefor the subjects in the experiments reported here: see Shepherd & Dere-gowski, 1981). It is assumed that the number of dimensions could be largeenough to represent any aspect of a face that could serve to discriminatebetween faces. At this stage in the discussion it will be assumed that theoverwhelming majority of faces encountered are own-race faces. Althoughthe model to be described will be discussed in terms of dimensions rathertha n features, this does no t m ean t o exclude the possibility t ha t th e processesinvolved may be based on frequency information of discrete feature values.However, dimensions appear to cope more naturally with discriminationwithin a stimulus class of which all exemplars share a common structure,such as faces. This view is in line with Garners (1978) definition of adimension a s a n attribute th at exists for each stimulus in the relevant set atsom e positive, mu tually exclusive value (p. 104).T he origin of the m ultidimensional space is defined a s th e central tendencyof the dimensions. It is assumed that the values of the feature dimensions ofthe population of faces experienced will vary normally around the centraltendency (at least for own -race faces). The refore by definition, typical faces(close to the central tendency) will be seen mo re often than distinctive faces(distant from the central tendency). Thus the density of points (i.e. thenumber of previously seen faces) will decrease as the distance from thecentral tendency increases. The population of points will include familiarfaces in addition to many points representing faces that have been seenpreviously, but would not necessarily be familiar in the sense that theycould be identified. Thus, an implicit knowledge derived from a lifetimesexperience with faces contributes to the normal distribution of faces withinthe multidimensional space.The implicit knowledge of faces is something a subject brings to anexperiment, an d it is assumed th at th e effect of the set of faces used within a nexperiment will have an insignificant influence compared to a lifetimes

The assumption of a Euclidean metric is only made fo r simplicity in the absence of anyevidence of the most appropriate metric. As the dimensions are not identified, the metric of thespace is not an issue addressed in the current paper.



9/46

FACE RECOG NITION 167experience of faces. Therefore, the proposed framework is fundamentallydifferent from the theoretical basis of studies in which sets of artificial faceshave been manipulated within an experimental context (e.g. Reed, 1972;Goldman & Homa, 1977; Neumann, 1977; Das-Smaal & De Swart, 1984;Malpass & Hughes, 1986). As the distribution of faces within the multidi-mensional space is assumed to emerge as a consequence of a lifetimesexperience, it is important to use photographs of real faces in order toinvestigate the influence of such implicit knowledge. Photographs of faces arestimuli that vary along the same dimensions and with the same centraltendencies as the population of faces previously experienced. Of course thiswould not be true of an artificially constructed set of faces (e.g. schematicfaces, photo-fit faces), therefore performance in tasks based upon artificialstimulus sets might be unaffected by knowledge of the population of realfaces. It is impossible to distinguish distance-based and frequency-basedmodels because the exemplars that contribute to the implicit knowledge offaces cannot be controlled experimentally.At this point, two specific models based on the multidimensional spaceframework need to be distinguished. A norm-based coding model will bedescribed first, in which it is assumed that faces are encoded in terms of theirdeviation from a single general face norm or prototype located at the originof the space (i.e. representing the central tendency). It is assumed that there isonly a single norm (Valentine & Bruce, 1986~). his is quite different fromthe notion of a different prototype to represent each type or family of faces(cf. Ellis & Christie cited in Ellis, 1981). If the number of dimensions isdenoted by n, an n-dimensional vector from the norm (or origin) to the pointrepresenting the dimension values of a particular face can uniquely specifythat face. Figure 1 illustrates the norm-based coding model using just a two-dimensional rather than a multidimensional space for the purposes ofillustration only.The recognition process can be viewed as involving two stages. First, astimulus face is encoded as an n-dimensional vector. Second, some form ofdecision process is required to determine whether or not the stimulus matchesa vector for a known face. The encoding process is assumed to have someassociated error or noise, which will depend upon the encoding conditions.Therefore, under difficult viewing conditions, the vector derived will have arelatively large associated error estimate, which could be represented inFigure 1 as a region of uncertainty around the co-ordinates of the stimulusvector. It is assumed that the confidence signalled by the decision processdepends on: (1) the error associated with the vector derived from thestimulus face; (2) a measure of similarity between the vector derived fromthe stimulus and the vector corresponding to the nearest known face; and(3) the similarity between the stimulus vector and the vector of the nextnearest neighbour. The measure of similarity used is assumed to be some



10/46

168 VALENTINE

FIG. 1. A two-dimensional representation of a norm-based coding model of face recognition.Each point plotted represents a previously seen face located in an n-dimensional space. Theorigin represents a general face norm. Two dimensions are used in this figure for the purposes ofillustration only. The axes are unlabelled but could represent any dimension that could be usedto discriminate faces.

function of a vector similarity measure such as the dot product between twovectors. (See Micko, 1970, for a discussion of vector-based measures ofsimilarity in multidimensional scaling.)

The exemplar-based model assumes that there is no extracted norm orprototype. The model is similar to the norm-based coding model, except thati t is more appropriate to consider faces as being encoded as points ratherthan vectors. The origin of the multidimensional space plays no part inencoding stimuli, it merely indicates the point of maximum exemplar density.The exemplar-based model assumes that similarity between two faces is amonotonic function of the distance separating the representations of thefaces in the multidimensional space. The decision process is assumed to



11/46

FACE RECOGNITION 169depend on: (1) the estimate of error associated with encoding the stimulus;(2) the distance between the location o f the stimulus and th e nearest know nface; and (3) the distance between the stimulus and the next nearestneighbour.It should be noted that though both the norm-based coding and theexemplar-based models a re nearest neighbour exem plar models, they differ inthe role of a n abstracted norm in coding an d the use of a vector- or distance-based measure of similarity. (See Valentine, 1990, for fu rther discussion ofthis point.)Predictions Derived from the FrameworkBoth the norm-based coding model and the exemplar-based model accountfor the effects of distinctiveness found in recognition of faces by appealing toexem plar display. Distinctive faces ar e located in regions w here the density ofpo ints is low. The refore, when a distinctive familiar face is encountered in arecognition task, the location encoded in the multidimensional space will bemuch closer to the representation of the target face stored in memory tha nto the location of another face. (By definition, few faces will resemble adistinctive face.) In addition, the error associated with the encoded locationis likely to be small compared to the distance to the nearest neighbour,especially under good encoding conditions. Therefore, the face can beidentified accurately and rapidly. If the face presented is a typical familiarface, the location derived from this face is close to the central tendency, so itwill fall in a region in which there is a high density of points. Although thelocation of the stimulus will be close to the location of the target face, it willalso probably be close to other previously seen but unfamiliar faces.There fore the decision a s to w hether the stimulus is closer to the location o fthe familiar face than to a nearby unfamiliar face will be more difficult,rendering it slower and more error-prone. The multidimensional spaceframewo rk the refore pred icts that distinctive familiar faces will be recognizedmo re accurately or more quickly than typical familiar faces.Consider the case of a face that has not been seen before being encoun-tered. Of course there is no sto red description for the face in memory. If thestimulus face is distinctive, the density of points around i t will be low.Therefore, it is unlikely that the location of a previously seen face will beclose enough to the stim ulus with its associated erro r to give a false positive.If the stim ulus face is a typical unfamiliar face, its derived location will be in aregion of a high density of points. Therefore, it is more likely to be close tothe representation of a previously seen face. If the similarity between theencoded face and a previously seen face is sufficiently high, a false positivewill result. T hu s the framewo rk predicts that d istinctive unfam iliar faces canbe rejected more accurately or more quickly than typical unfamiliar faces.



12/46

170 VALENTINEBoth the norm -based coding model a nd the exemplar-based model clearlypredict a n effect of distinctiveness in correctly rejecting unfamiliar faces in afamiliarity decision task. H owever, Valentine an d Bruce (1 9 8 6 ~ ) oun d atrend in the opposite direction th at did n ot reach statistical significance. It ispossible that the failure to find the expected effect in rejecting unfamiliarfaces was the result of an artifact of the stimulus sets used. This prediction ofthe framework is examined in Experiment 3.The norm-based coding model and the exemplar-based model bothaccount for the effect of distinctiveness on recognition in terms of exemplardensity. However, the role of a vector similarity measure in the norm-basedcoding model prod uces a n effect th at w orks against the effect of distinctive-

ness. In the norm-based coding model the similarity between two faces thatare equi-distant in the space is dependent on the distance of the points fromthe norm. C onsider two pairs of points (A, B and C, D) located such that thedistance between A and B is equal to the distance between C and D. [Apossible arrangement is shown in Figure 2. In this figure A and B areequidistant from the origin (0), s are C and D.] If a similarity measure tha tis based on th e distance between exemplars is used (as in the exem plar-basedmodel), A and B are as equally similar to each other as are C and D.How ever, the angle between the vectors OA and OB is greater tha n the anglebetween the vectors O C a nd OD, and so the faces represented by OA and OBwill be more dissimilar tha n the faces represented by O C and OD. Th us in thenorm -based coding model, faces separated by a given distance will be mo redifficult to discriminate, the furth er they a re from the no rm. T he norm -based

FIG. 2. An illustration of the effect o f distance from th e norm on similarity in the norm-basedcoding m odel. The distance between A and B is the same as the distance between C and D, butthe angle between vectors OA and OB is greater than the angle OC and O D. See text forexplanation.



13/46

FACE RECOGNITION 171coding model can still support the prediction that distinctive faces can bemore accurately recognized than typical faces because distinctive faces areless densely clustered (i.e. further apart from each other) than typical faces.However, it is necessary to assume that the effect due to the difference inexemplar density is greater than the opposite effect due to distance from thenorm on the vector similarity measure.It was noted above that the effect of distinctiveness found in a faceclassification task (in which faces must be distinguished from jumbled faces)was the opposite to that found in a recognition task. The classification taskcan be regarded as a judgement of how closely a stimulus resembles thecentral tendency of the population of faces. In terms of the norm-basedcoding model, the RT in this task will depend upon the length of the vectorderived from the stimulus face (i.e. its distance from the norm in the n-dimensional space). The density of points in a region will be irrelevant to thetask, because there is no requirement to distinguish between faces. Thereforetypical faces should be correctly classified as a face faster than distinctivefaces, because typical faces are closer to the norm. Thus the norm-basedcoding model successfully predicts the reversal of the effect of distinctivenessbetween a recognition and classification task.According to the exemplar-based model, both recognition and classifica-tion of faces will be affected by exemplar density but in quite different ways.In a recognition task high exemplar density makes the nearest neighbourlikely to be close to a target face, thus impeding the recognition decision.The classification task has no requirement to distinguish between individualfaces. In this task exemplar-density in the region around the stimulus is usedto judge category membership, the greater the density of faces in this region,the faster the stimulus can be classified as a face. Thus in a classification taskit is assumed that RT is determined by the exemplar density in themultidimensional space around the location of the stimulus (Krumhansl,1978). As distinctive faces are located in areas of lower exemplar density,typical faces will be classified faster than distinctive faces. Both the norm-based coding and the exemplar-based models make the same prediction ofthe effect of distinctiveness on classification as a face, but the norm-basedcoding model appeals to distance from the norm and the exemplar-basedmodel appeals to exemplar density.Some researchers have suggested that inversion disrupts the normal facerecognition process to such an extent that different features are used torecognize upright and inverted faces (e.g. Carey & Diamond, 1977; Diamond& Carey, 1986). This view arose from the observation that recognition offaces is more disrupted by inversion than recognition of other stimulusclasses (Yin, 1969). Inversion makes face recognition slower and lessaccurate, but there is no compelling evidence that upright and inverted facesare processed qualitatively differently (see Valentine, 1988, for a review). One



14/46

172 VALENTINEaim of the multidimensional space framework is to account for the effects ofinversion on face recognition and classification. It is assumed that themagnitude of the error in deriving the location of a stimulus face is dependentupon the encoding conditions. Presenting a face upside-down is an exampleof one experimental manipulation that would make the encoding conditionsdifficult, leading to a large error associated with the location in the multidi-mensional space derived from a stimulus face. The effect of inversion will berepresented in the framework in this way. Any experimental manipulationthat impairs recognition could be represented by assuming an increase in theerror associated with the location of the stimulus face. Thus the multidimen-sional space framework makes the strong prediction that other manipula-tions (e.g. adding visual noise or blurring faces) should interact withdistinctiveness in the same way as inversion. However, in this paper theconcern is to account for the effect of inversion because the inversionparadigm has had a considerable impact on the face recognition literature.Representing inversion in this way, the norm-based coding and theexemplar-based models make similar predictions concerning the interactionbetween the effects of inversion and distinctiveness in a recognition task. Atypical unfamiliar face seen inverted will be more likely to be mis-identified asfamiliar than the same face seen upright. The high density of nearby pointsand the relatively large error associated with the stimulus location derivedfrom an inverted face means that it is more likely to be sufficiently close to aneighbouring point corresponding to a previously seen face to give a falsepositive. If a typical familiar face is seen inverted, it is more likely to bemissed than when seen upright, because the increased error in deriving thelocation from the stimulus face means it is likely to be further from thelocation that forms its representation in memory. A miss is more likelybecause there is a greater chance that a previously seen but unfamiliar facewill be closer than the target. Of course the same arguments apply toinverted distinctive faces. Therefore, inversion should impair recognition ofboth typical and distinctive faces. However, because the locations derivedfrom distinctive faces will be in regions with a low density of points, a largererror can be tolerated in recognition of distinctive faces than in recognitionof typical faces. Therefore, inversion should cause greater disruption torecognition of typical faces than of distinctive faces. This prediction of aninteraction between the effects of distinctiveness and inversion was testedusing a recognition memory paradigm in Experiment 1.

There are a number of important and as yet unanswered issues raised bythe application of a multidimensional space framework to face recognition.Most notably, the dimensions of the space have not been identified, nor hasthe dimensionality of the space required. The determinants of similarity havealso not been precisely specified-for example, the metric of the space has notbeen determined. In view of our currently imprecise knowledge of the



15/46

FACE RECO GNITION 173dimensions underlying face perception, and therefore our impoverishedunderstanding of similarity between faces, it would be premature to preciselyspecify the parameters of the framework. The aim is to demonstrate that themultidimensional space framework has predictive utility and can be used toguide future research.A number of predictions can be derived from the multidimensional spaceframework. In Experiments 1-3, the prediction that recognition of typicalfaces is more impaired by inversion than recognition of distinctive faces istested. Experiment 4 demonstrates that this interaction is not found in a taskthat does not require recognition of individual faces (an intact/jumbled faceclassification task). The effect of race on face recognition is then consideredin relation to the multidimensional space framework. It is predicted that theeffect of race will interact with the effects of inversion and task demands inthe same way as distinctiveness. This analysis is consistent with data reportedby Valentine and Bruce (1986a), in which an interaction between the effectsof inversion and race on recognition memory for faces was found. InExperiment 5 it was demonstrated that the effect of race is additive with thatof inversion in a face classification task.

EXPERIMENT 1MethodThe experiment was in two parts. First, faces were rated for distinctiveness.Different subjects then carried out a recognition memory task.

Distinctiveness RatingsSubjects. Sixteen (2 male and 14 female) members of the AppliedPsychology Unit subject panel made the distinctiveness ratings; their meanage was 44.9 years.Materials . The stimuli were slides of 64 faces. All were male, photo-graphed in a full-face pose with a neutral expression. Faces with a beard,moustache and/or glasses were included. All clothing was masked.Apparatus. The slides were projected onto a wall using a KodakCarousel projector. The subject was given a remote advance button. The

faces subtended a visual angle of approximately 10" (horizontally) by 13"(vertically).Procedure. The procedure for rating the faces was that used by Valen-tine and Bruce (1986~). ubjects were asked to rate each face on a 1-7 scale.They were instructed to imagine that they had to meet each person at a



16/46

174 VALENTINErailway station and to rate each face for how easy it would be to spot in acrowd. A face that was very distinctive (or unusual) and so would berelatively easy to sp ot in a crowd shou ld be rated 7, a typical face th at wouldbe difficult to identify in a crowd should be rated 1. Subjects were allowed t oproceed through the list of slides at their own pace. Four different randomord ers of the slides were used. Based on the ratings, two sets of 16distinctivefaces (mean rating 4.86 and 4.87) were selected to serve as distinctive targe tsand distractors, respectively, in a recognition memory task. Similarly, twosets of 16 typical faces (mean rating 3.12 an d 3.13) were selected to serve astypical targets and distractors, respectively. A t-test confirmed that thedistinctive and typical faces differed significantly in their distinctiveness(t(30) = 14.39, p


17/46

FACE RECOGNITION 175items. Different pictures (involving a change of expression) of the target faceswere shown in the test lists. Half of both the targets and distractors weresmiling, half had a neutral expression. The targets and distractors in both testlists were divided into two sets of 8 faces, matched on distinctiveness. One setwas presented upright, and one set was presented inverted.

Procedure. All subjects were tested individually. They were told thatthey would be shown two lists of slides of faces, and that there would be arecognition test immediately after each list. In the study lists typical faceswere presented for 6 sec, and distinctive faces were presented for 3 sec. Therewas an interval of 2 sec between all slides. A different random order of thestudy list was constructed for every two subjects. At test all slides werepresented for 5 sec, with a 2-sec interval. A quasi-random order wasconstructed for the test lists, with the constraint that no more than threetarget or distractor faces were presented in sequence, and no more than threeconsecutive faces were presented in the same orientation. A second order wasmade by reversing the order in which the halves of the list were presented.Slide order a t test, the assignment of set of slides to orientation, and the orderin which typical and distinctive faces were presented were counter-balancedacross subjects. Subjects were warned that target faces would have a differentfacial expression at test. They were also warned that half of the faces wouldbe presented upside-down and instructed to keep their heads upright whenlooking at these items. Subjects were instructed to respond as quickly and asaccurately as possible.Resu tsFor each subject hit and false positive rates in each condition were calculatedand combined in A' scores (Rae, 1976). A' scores were calculated from hitand false positives to either upright or inverted stimuli alone. The maximumnumber of hits or false positives was 8. Mean A' scores, mean number of hitsand false positives, and the mean latency of hits and correct rejections areshown in Table 1 .An F,,, test was carried o u t on the raw data before conducting all of theanalyses reported in this paper. If F,,, indicated that the assumption ofhomogeneity of variance was violated, an appropriate transformation wasmade and the transformed data subjected to another F, test. In all analysesreported in which a transformation has been carried out , the raw dataviolated the assumption of homogeneity of variance and the transformeddata did not. When an analysis of untransformed data is reported, the rawdata did not violate the assumption of homogeneity of variance.A' scores were subjected to a sin-'Jz transformation prior to beingsubjected to an analysis of variance (McNicol, 1972, p. 117). A within-



18/46

176 VALENTINETABLE 1Mean A Scores, Hits; False Positives' and Mean RT of Hitsband Correct Rejectionsb as a

Function of Orientation and Distinctiveness (Experiment 1)Upright Inverted

Distinctive Typ ical Distinctive TypicalA' 0.905 0.919 0.802 0.699Hits 6.50 6.55 5.35 4.85F.P. I .oo 0.70 1.75 2.10RT of hits 1356 1595 1770 2050RT of C. R. 1418 I723 1767 2208

a Maximum 8.In msec.

subjects AN OV A gave a significant main effect of o rientation, F(1, 19) =43.07,p


19/46

FACE RECOGNITION 177significant m ain effect of distinctiveness, F(1, 9)= 7.86, p < 0.02. Distinctivefaces were recognized fa ster than typical faces. There was also a ma in effectof inversion, F( 1,19)= 23.74, p


20/46

178 VALENTINEE X P E R I M E N T 2

Experiment 2 was a replication of Experimen t 1, except th at in the initial liststhe same exposure time was used for all stimuli. In order to equateperformance, the number of faces included in the study list of distinctivefaces was increased. The number of stimuli included in the test list wasunchanged.M e t h o dsubject panel acted a s subjects. Their mean age was 31 years.Subjects. Twenty-eight (23 female and 5 male) members of the APU

Design and Procedure. All aspects of the design and procedure were thesame as Experiment 1, except fo r the following details. All stimuli in bo th thestudy and test lists were presented for 5 sec, with a 2-sec interstimulusinterval. The study list of distinctive faces consisted of 24 faces: 16 were thesame faces as used in Experiment 1, and 8 new faces were included to m akethe list longer. The extra 8 faces were no t included in the test list. The initiallist of typical faces was unchanged , consisting of 16 faces all of w hich wereincluded in the test list. Each of the four stimulus orders was used fo r sevensubjects.ResultsThe hit rate and false positive rate were calculated for each subject andcombined in A scores. Mean A scores, mean number of hits and falsepositives, and mean latency of hits and correct rejections are shown inTab le 2.

TABLE 2Mean A Scores, Hits, False Positives and Me an RT of Hitsband Correct Rejectionsb as aFunction of Orientation and Distinctiveness (Experiment 2)Upright In verted

Distinctive Tvpical Distinctive TypicalA 0.915 0.894 0.824 0.693Hits 6.75 6.33 5.75 4.82F.P. 1 O 1.14 1.93 2.75R T of hits I242 1471 I626 1851RT of C . R . I382 I647 1830 1910

Maximum 8 .In msec.



21/46

FACE RECOGNITION 179Transformed A' scores (sin- ',/A') ere subjected to a 2 X 2 within-subjectsANOVA. There was a significant main effect of orientation, F(1, 27)=

127.49, p


22/46

180 VALENTINEinversion was significant for upright and inverted faces ( p < O . O 5 ) , and thatthe effect of distinctiveness was significant for up right faces (p


23/46

FACE RECOGNITION 181ha s been used extensively in recent face recognition research. T he subject isshown a series of faces, some of which a re of fam ous (or personally familiar)faces, othe rs ar e unfamiliar faces. The dependent variable is the subjects RTto decide whether each face is familiar or unfamiliar. A problem found inpilot work using a fam iliarity decision task to exp lore the effect of inversionwas the range of performance foun d for upright an d inverted faces. Accuracyfor inverted faces was too low to allow RT to be reliably used as thedependent variable, but accuracy was effectively 100% for upright faces, soaccuracy was also an unreliable measure. In o rde r to overcom e this problem ,accuracy was enhanced by giving subjects a list of the names of the famouspeople who appeared in the experiment and using RT as the dependentvariable. R eading a fam iliar persons name does not produce a priming effecton the subsequent R T in a familiarity decision task (Bruce & Valentine, 1985;Ellis, Young, Flude, & Hay, 1987).An additional aim of Experiment 3 was to investigate the effect ofdistinctiveness upon the R T t o reject unfam iliar faces. T he m ultidimensionalspace fram ewo rk predicts tha t subjects should be faster to reject distinctiveunfamiliar faces than to reject typical unfamiliar faces. In a previous studythis effect was n ot found in a fam iliarity decision task , bu t the distinctivenessof familiar an d unfamiliar faces had not been m atched (Valentine & Bruce,1986c). Therefo re, in Experiment 3 distinctiveness of familiar and unfamiliarfaces was matched.MethodThe experiment was in two parts. First faces were rated for distinctivenessan d fam iliarity. Different subjects then ca rried o u t a familiarity decision task.

Distinctiveness and Familiarity RatingsSubjects. Eighteen (14 female an d 4 male) members of the A PU subjectpanel r ated the stimulus faces for distinctiveness an d familiarity. The ir meanage was 42.4 years.Materials. Pictures of 54 famous faces and 39 unfamiliar faces wererated . All of the pictures were of males. Pictures of the fam ous faces had beencollected from a variety of sources (picture libraries, political parties,

magazines, etc.). The poses varied from full-face to three -qua rters profile. Allpictures were copied onto monochrome slides through a circular mask,which excluded the majority of the backgroun d. T he pho tog ra ph s ofunfam iliar faces were either copied f rom a directory of actors or were studioportraits of academics. They were prepared in an identical format to thefamous faces and included a similar range of poses. It was known from



24/46

182 VALENTINEprevious w ork tha t all the actors were likely to be un familiar to the subjects.Th e academics lived in an oth er pa rt of the country a nd so were very unlikelyto be familiar. Som e pictures of faces with beards an d glasses were includedin the s timulus sets.

Apparatus. As fo r the distinctiveness ratings collected in Exp erimen t I .Procedure. Subjects rated the unfamiliar faces first. The procedure forthese ratings was identical to that described for Experiment 1. The subjectsthen rated the famous faces. When making distinctiveness ratings of famousfaces, subjects were instructed to treat the faces as if they were unfam iliar-

that is, to base their judgement entirely on the information available in thepicture shown and to ignore any distinctive feature which they knew thecelebrity had, but which could not be seen in the photograph (e.g. adistinctive hair colour). The subjects also rated the famous faces forfamiliarity on a 1-7 scale. In all other aspects the procedure was the same asth at described above.The stimulus sets to be used in the second phase of the experiment wereselected on the basis of these ratings. These consisted of 16 typical unfamiliarfaces (mean distinctiveness rating = 3.21) an d 16 distinctive unfamiliar faces(mean distinctiveness rating = 5.05). A f-test for independent samples con-firmed th at there was a significant difference in distinctiveness between thesetwo sets, (30) = 9.41, p


25/46

FACE RECOGNITION 183Design. A list of 64 slides of faces was presen ted. H alf of the slides wereof fam ous people, half were of unfam iliar faces. The subject was instructed t orespond YES if a face was familiar an d NO if it was unfam iliar. There were 16faces in each of the following four categories of stimuli: distinctive fa m ousfaces, typical fam ous faces, distinctive unfam iliar faces, an d typical unfami-liar faces. Eight faces in each category were presented upright and 8 werepresented upside-down. Each face was presented upright to half of thesubjects and inverted to the remainder.Procedure. All subjects were tested individually. Subjects were given alist of the nam es of the fam ous people who app eare d in the experiment. They

were asked whether they thought they would be able to recognize all of thepeople on the list. Tw o subjects who were not fam iliar with m any of the facesdid not take part in the experiment. All of the remaining subjects werefamiliar with all but two o r three faces a t most. If subjects said they wouldnot recognize a particular face, the experimenter provided some othersemantic information ab ou t the individual (e.g. a film they had appea red in, adescription of a pa rt played in a TV series, etc.). Often subjects would then bemore confident of recognizing the face.Subjec ts were told that they wou ld see a series of slides of faces. They wereinformed tha t a bo ut half of the faces of celebrities, an d th at the names of allof the celebrities in the series were included in the list they had just read.Subjects were instructed to press the YES bu tton if a face was familiar and torespond NO if it was unfamiliar. They were informed that RT was beingmeasured and were instructed to respond as quickly and as accurately aspossible. Subjects were warned that half of the faces would be presentedupside-down an d w ere instructed to keep their head upright while looking a tupside-down faces. T he slides were presented for 5 sec, with a 2-sec intervalbetween slides. A quasi-random slide order was constructed with theconstraint that there were no more than three fam ous o r unfamiliar faces insequence and no more than three consecutive faces presented in the sameorientation. A second slide order was generated by reversing the order inwhich the two halves of the first series were presented. T he slide orde r a nd theassignment of slides to the o rienta tion were counter-balanced across subjects.ResultsF o r each subject, hit and false positive rates were calculated an d comb ined inA scores. Mean A scores, mean n um ber of hits and false positives, and meanR T of hits and correct rejections a re shown in Tab le 3.Assessment of accuracy in a familiarity decision task is always slightlyproblematic as i t is possible th at a subject does not know a famous face orgenuinely recognizes an unfamiliar face. The former disagreement is



26/46

184 VALENTINETABLE 3Mean Latency and Accuracy of Responses in a Familiarity Decision Task as a Functionof Orientation and Distinctiveness (Experiment 3)

Upright InvertedDistinctive Typical Distinctive Typical

Mean RT of correct 983 1076 1552 2312familiar responsesMean RT o f correct 1526 1698 2197 2372unfamiliar responsesMean n o. o f correct 7.5 7.3 5.0 3.0familiar responsesMean no. of familiar 0.3 0.9 0.7 1.3responses to unfamiliarfacesMean A of familiarity 0.975 0.944 0.865 0.7 15decisions

Max n o. o f responses in each cel l=8 .

mu ch m ore likely, bu t th e occurrence of such disagreements in this experi-ment should have been very much reduced by checking at the start thatsubjects were familiar with the names of the celebrities. Despite thisprecaution, misses occurred on 28.75% of trials in which a famous facewas presented, and false positives occurred on 10% of trials in which anunfamiliar face was presented. In view of the high error rate, which reflectsthe extreme difficulty of recognizing upside-down faces, an analysis of theaccuracy da ta is reported below.Analysis of RT of correct responses does not raise such problems. T heRT data of correct responses to famous faces were subjected to a logtransform before an ANOVA was carried out. There was a main effect ofdistinctiveness, F( 1, 23) = 50.19, p


27/46

FACE RECOGNITION 185order to test this interaction using a non-parametric test, the differencebetween RT to upright and inverted faces was calculated separately fortypical and distinctive faces. A Wilcoxon test carried out on these differencesin RT for typical and distinctive faces was significant, T(24)= 17,p


28/46

186 VALENTINEDiscussionThe RT data supported the predictions derived from the multidimensionalspace framework. Familiar distinctive faces were accepted faster thanfamiliar typical faces. Thus the effect of distinctiveness in a familiaritydecision task, found in previous work, was replicated. It was also predictedthat an effect of distinctiveness should be found on the RT to rejectunfam iliar faces. Th is prediction w as supported: unfam iliar distinctive faceswere rejected faster th an unfam iliar typical faces.The predicted Inversion x Distinctiveness interaction was found in theRT s to familiar faces. R T to accept fam iliar typical faces showed a greaterincrease due to inversion tha n did the R T to accept fam iliar distinctive faces.There was no Distinctiveness x Inversion interaction in the RT to rejectunfamiliar faces. It is not clear why the interaction between distinctivenessan d inversion sho uld no t be found in rejection latencies, but it is possible thatwhen, in the context of looking for highly familiar faces, a match to aninverted face is no t found , some form of a checking process may be initiatedwhich obscures the Distinctiveness x Inversion interaction. The RT datawere supported by the accuracy data. The predicted main effect of distinc-tiveness and the Distinctiveness x Inversion interaction were found in theanalysis of the A scores an d correct fam iliar responses. This contrasts withthe results of Experiments 1 and 2 in which the interaction was found in theA an d false positive rate d at a, but no t in the hit rate d ata . The difference inthe task demands between the two experiments may be a cause for thisdifference in the results. Experiments 1 and 2 required recognition ofpreviously unfamiliar faces seen once a few minutes earlier, whereas asExperiment 3 required recognition of highly familiar faces.

EXPERIMENT 4Experiments 1-3 have demonstrated an interaction between distinctivenessand inversion in two tasks that require recognition of individual faces. InExperiment 4 the interaction between these facto rs in a face class$cation taskwas investigated. As described above, th e effect of distinctiveness reverses ina classification task. Typical faces can be classified as a face more rapidlyth an distinctive faces. In terms of the norm -based cod ing model, it is assumedthat the length of the derived vector will determine the RT in a faceclassification task. Typical faces are closer to the norm and so can beclassified faster than distinctive faces. In the exemplar-based model theexemplar density will determine the RT of classification. High exemplardensity will lead to faster classification decisions than low exemplar density.As typical faces will be located in regions of higher exemplar density thandistinctive faces, the exemplar-based model also predicts that typical faceswill be classified faster than distinctive faces. If the stimuli in this task are

Downloadedby[McMasterUniversit

y]at11:4628February2012


29/46

FACE RECOGNITION 187presented upside-down, the effect would be represented in the multidimen-sional space by increased random error associated with each encoded face. Interms of the norm-based coding model, an increased random error is as likelyto move the vector encoded from the stimulus closer to the norm, as it is tomove it further away. In terms of the exemplar-based model, an increasedrandom error is as likely to move the encoded location of the stimulus to anarea of greater exemplar density, as it is to move it to an area of lowerexemplar density. Therefore, in both models the effects of the increasedvariability of encoding should tend to cancel each other out. Thus, it ispredicted that the effect of distinctiveness should be found fo r both uprightand inverted faces, with no interaction between distinctiveness and inversion.

The prediction of a Distinctiveness X Inversion interaction in a recognitiontask arises because inversion causes a greater mismatch between the encodingof a face and its representation in memory. The effect of this greatermismatch will depend on the exemplar density in the region. No interactionbetween distinctiveness and inversion is predicted in a face cIassiJication askbecause there is no requirement to discriminate between the representation ofindividual faces in memory.

MethodThe experiment was in two parts. First faces were rated for distinctiveness.Different subjects then carried out a face classification task.

Distinctive RatingsSubjects. Fifteen (2 male and 13 female) of the sixteen subjects whorated the faces used in Experiment 1 also rated the faces used in this

experiment. Their mean age was 44.9 years.Materials. Thirty-two faces were prepared as monochrome prints ap-proximately 130 mm x I 10 mm. The faces were all male, in a full-face posewith a neutral expression. None had a beard or a moustache and none woreglasses. They were copied from a directory of actors, and on the basis ofprevious work were not likely to be familiar to subjects.

Procedure. The procedure was the same as that described above, exceptthat the prints rather than slides were used. Based on the ratings, the faceswere split into a set of 16 distinctive faces (mean rating, 4.9) and 16 typicalfaces (mean rating, 3.3). A t-test confirmed that there was a significantdifference in perceived distinctiveness between the two sets of faces,(30) = 7.88 , p


30/46

188 VALENTINEFace Classification TaskSubjects. Twenty-one members of the APU subject panel acted as

subjects. Data from one subject were discarded because of an error rate of10%. Thus usable data were obtained from 20 (4 male and 16 female). Theirmean age was 41.4 years.

Materials . Stimuli were prepared from the 32 faces described above,together with 4 aces for use as practice items. An intact and a jumbledversion of each face was made. A grid of lines was drawn on the intact facesbefore being copied onto slides, so that the intact faces had the same lines onthem as the jumbled faces. To make the jumbled faces, a rectangle was drawnaround the eyes-nose-mouth region of each face and was divided byhorizontal lines into three regions corresponding to each of these features.The features were cut out and reassembled either in the order mouth-eyes-nose or nose-mouth-eyes (from top to bottom). Figure 3 shows an exampleof an intact and a jumble face.

Apparatus. The apparatus was the same as that used in Experiment 1.Design. The intact and jumbled versions of the extra four faces com-prised eight practice items at the beginning of the list. The experimental trials

FIG. 3. An example of an intact and a jumbled face used in the face classification task inExperiment 4.



31/46

FACE RECOGNITION 189consisted of 16 distinctive intact faces, 16 distinctive jumbled faces, 16 typicalintact faces, and 16 typical jumbled faces. The faces were split into two sets,each consisting of half of each of these categories of stimuli. The meandistinctiveness ratings of faces in the two sets of distinctive and typical faceswere matched. Set A was presented upright, and Set B was presented invertedto half of the subjects. The remaining half saw the stimuli presented in theopposite orientation. The subject was asked to classify each stimulus as it waspresented as either a face or a jumbled face. The design had two within-subjects factors, the distinctiveness and orientation of the face. The depen-dent variable was the RT of correct responses.

Procedure. Subjects were tested individually. They were shown anexample print of an intact and a jumbled face, and the nature of the task wasexplained. They were informed that half of the faces would be presentedupside-down and were instructed simply to classify each stimulus as a face ora jumbled face regardless of its orientation. The response box had one buttonlabelled face and one labelled jumbled face. Subjects were informed thatresponse latency was being measured and instructed to respond as quicklyand as accurately as possible. The stimuli were presented in a quasi-randomorder, with the following constraints: (a) no more than three consecutiveintact or jumbled faces, (b) no more than three typical or three distinctivefaces occurred in sequence, and (c) no more than three consecutive stimuliwere presented in the same orientation. Two orders were used, one beinggenerated by reversing the order of presentation of the two halves of the firstlist. The slide order and the assignment of slide set to orientation werecounter-balanced across subjects. Slides were shown for 2.5 sec, with a 2.5-sec interval between slides. Response latencies greater than 2 sec were scoredas errors.ResultsErrors were made on 2.3% of trials in which an intact face was presented.There were too few errors to analyse. Mean RTs of correct responses to theintact faces are shown in Table 4. These data were subjected to a 2 x 2A N O V A with distinctiveness and orientation as within-subjects factors.There was a main effect of orientation, F( 1, 19)= 47.47, p


32/46

190 VALENTINETABLE 4

Mean RTs to Classify Correctly Intact and Jum bled Faces, as a Function ofDistinctiveness and O rientation (Experiment 4)Intact fac es Jumbled Faces

Typ ical Distinctive Mean Typical Distinctive MeanUpright 683 737 710 74 1 738 739

(0.6) (1.9) (1.3) (0) ( 0 . 6 ) (0.3)Inverted 817 860 839 866 84 1 855

(1.3) (5 .6 ) (3.5) (7.5) (5.0) (6.3)Mean 750 798

(0.9) (3.7)803 79 1

(3.7) (2.8)Note: a YO rrors are shown in parentheses

within-subjects ANOVA. There was a significant effect of orientation,F( I , 19)= 14.23,p


33/46

FACE RECOGNITION 191discriminating another race of faces. For example, there is evidence thatdifferent facial features are used to describe black and white faces (Ellis,Deregowski, & Shep herd, 1975), an d to judge similarity of a set of simu ltan-eously presented black or white faces (Shepherd & Deregowski, 1981).Although subjects are c apable of using ap pro pr iate cues to distinguishsimultaneously presented other-race faces, the evidence available suggeststha t they resort to using the fam iliar but inap propriate own-race facial cuesin a recognition m emory paradigm (Shepherd, 1981). Therefore, other-racefaces will be encoded o n dimensions tha t d o not serve to discriminate wellam ongst faces of that race. Th e dimension values tha t ar e salient will tend tobe those that are characteristic of the other race, rather than those that arecharacteristic of the individual. Therefore, other-race faces will be distantfrom the central tendency but have many feature values in common and sowill be densely clustered. Thus other-race faces form an exception to th eassumption that the exemplar density decreases with increasing distancefrom the central tendency.An illustration of a two-dimensional version of the norm-based codingmodel including own an d other-race faces is shown in F igure 4.A number of predictions can be derived from this interpretation of therace effect. The norm -based coding m odel will be considered first. Increaseddistance fro m the norm will result in the vectors of faces sepa rated by a givendistance being more similar to each other and therefore more difficult todiscriminate (see Figure 2). Assuming that other-race faces will be moredistant from the norm than own-race faces, the norm-based coding modelpredicts that other-race faces will be more difficult to recognize. Thisprediction holds even if othe r-race faces are equally densely clustered as ow n-race faces. The other-race effect will be enhanced if other-race faces areindeed m ore densely clustered because the dimensions defining the space areinappropriate. The difficulty of discrimination between similar vectors, dueto the distance from the norm and/or higher density of other-race faces,means that recognition of other-race faces should be severely disrupted by anincrease in the erro r of deriving a stimu lus vector. By the same argum ent usedabove for typical and distinctive faces, this leads to the prediction thatrecognition of other-race faces will be more disrupted by inversion thanrecognition of own-race faces (i.e. in a recognition task own-race faces areanalogo us to distinctive faces an d o ther-race faces are a nalogou s to typicalfaces.)

The exemplar-based model assumes that similarity is determined by thedistance separating faces and so is independent of the location of the norm.Therefore, the exemplar-based model can only accoun t for better recognitionof own-race faces than of other-race faces if it is assumed that exemplardensity is greater .for other-race faces than for own-race faces. The workreviewed above suggests tha t there is some evidence to s up po rt this view. As



34/46

192 VALENTINE

FIG. 4. A two-dimensional representation of a norm-based coding model of face recognition,showing populations of own a nd other-race faces. As in Figure I , faces are represented by pointsin an n-dimensiona! space. Two dimensions have been plotted for the purpose of illustrationonly. The axes are unlabelled but could represent any dimensions that serve to discriminatebetween faces and differ in the cen tral tendency of ow n an d other-race faces.

the exemplar-based mo del uses differences in exemplar density to acc oun t forthe other-rac e effect, it also leads to the prediction tha t recognition of other-race faces will be more disrupted by inversion th an recognition of own-racefaces. This result has been reported by Valentine and Bruce (1986a). Theresults of this study are given in Table 5.The design of this experiment was identical to Experiment 1, except thatlists of white faces an d black African faces were used instead of lists of typicaland distinctive white faces. T he only differences in the procedure were that anidentical picture of the target faces was used at test and exposure times of2 sec (white faces) an d 5 sec (black faces) in the initial list were used to equateperformance. Twenty white subjects took part in the experiment. Separate



35/46

FACE RECOG NITION 193TABLE 5Mean A' Scores, Hits," False Positives" and M ea n RT of Hitsband Correct Rejectionsb as aFunction of Orientation and Race in a Recognition M emory Task

Upright InvertedWhite Black Wh ite Black

A' 0.900 0.918 0.723 0.609Hits 6.5 6 .7 4 .9 4.1F.P. 1.2 0 .9 2.3 3.6RT of Hits 1222 1553 1685 1933RT of C . R . 1502 1792 2002 2355'Maximum 8.Note:In msec.Accuracy data from: Valentine & Bruce (1986a).

analysis of the hit and false positive data, and analysis of the latency datawere not reported by Valentine and Bruce ( 1 986a), so they will be reportedhere briefly. The results of both the accuracy and latency measures weredirectly analogous to the results of Experiment 1 in all aspects. There w as asignificant interaction between the effects of race and orientation in theanalyses of transform ed A' scores (sin - 'JA ), F(1,16)= 16.52,p


36/46

194 VALENTINEIt is interesting to note that the multidimensional space framework andthe distinction between com pon ent a nd configural information (see introduc -tion) lead to different predictions of the interaction between inversion andrace. R hodes, T an , Brake, a nd Tay lor (1989) base their analysis of th e effectof race on Diamond and Careys (1986) work on the role of expertise inencoding second-order relational inform ation. Rh odes a nd colleagues arguethat lack of expertise in encoding other-race faces would mean that second-order relational information canno t be encoded from other-race faces. As i tis assumed that disruption to encoding of second-order relational informa-tion in own-race faces causes the disp rop ortionate effect of inversion o n own-race face recognition, R hod es an d colleagues argue th at recognition of other-

race faces will be fess disrupted by inversion than recognition of own-racefaces. They found support for their hypothesis using a somewhat differentprocedure to that used by Valentine a nd Bruce (1 986a). These ap pare ntlyconflicting results will be discussed further below. In Experiment 5, theexploration of the interaction between the effects of distinctiveness andinversion is extended to a face classification task.EXPERIMENT 5

Th e norm-based coding m odel makes a clear prediction tha t other-race facesshould b e classified as a face mo re slowly th an own-race faces in an intact/jumbled face classification task. T he length of the vector is assumed to affectRT in a face classification task. As other-race faces will, as a group, belocated further from the norm than own-race faces, other-race faces will beclassified more slowly than own-race faces. The exemplar-based model doesnot make a clear prediction. Classification RT is assumed to be determinedby the exemplar density in a region of a given radius around the stimulus(Krumhansl, 1978). Higher exemplar density will lead to faster classificationRT s. In ord er to account fo r the effect of race on recognition memory, theexemplar-based model must assume that other-race faces are more denselyclustered than own-race faces. However, there will be fewer other-race facesin memory than own-race faces. Therefore, there will be two opposingfactors tha t determine R T in the exemplar-based model. T he effect of race o nclassification R T will depend o n the value of unspecified param eters such a sthe radius of the region in question, the density of own and other-race faces,and the number of faces in memory. The exemplar-based model couldaccount for faster classification of own-race faces if the exemplar densityaround a stimulus is greater for own-race faces, despite being less denselyclustered, because there are many more own-race faces falling within theregion concerned. The norm-based coding model and the exemplar-basedmodel b oth predict t ha t the effect of race on classification of faces should notbe affected by inversion of the stimulus. An increase in the random encod ing



37/46

FACE RECOGNITION 195error caused by stimulus inversion would not have a systematic effect oneither the distance from the norm or the exemplar density around thelocation of the stimulus. The effect of race and inversion in a classificationtask were investigated in Experiment 5 .Method

Sixteen (10 female and 6 male) members of the APU subjectpanel acted as subjects. Their mean age was 40.5 years. All were Caucasian.

Subjects.

Muteriuls. Sets of 18 white faces and 18 black African faces wereselected. None had a beard or wore glasses. The 18 white faces were a subsetof those used in Experiment 4. For each face a slide of an intact face and ajumbled face were prepared as described above.

Apparatus. This was the same as for Experiment I .Design. Two of the 18 faces of each race were assigned to be practice

items. The intact and jumbled versions of these faces comprised eight practiceitems at the beginning of the list. The experimental trials consisted of 16black intact faces, 16 black jumbled faces, 16 white intact faces, and 16 whitejumbled faces. As in Experiment 4, the stimuli were split into sets, and theassignment of stimulus set to orientation was counter-balanced acrosssubjects. The design had two within-subjectsfactors, the race and orientationof the face. The dependent variable was the RT to classify intact facescorrectly.

Procedure. This was as described for Experiment 4, except that subjectswere informed that they would see black and white faces.ResultsMean RTs to classify the intact faces correctly were calculated for eachsubject. These data and the error rates are shown in Table 6 . As inExperiment 4, responses over 2 sec were scored as errors. Errors were madeon 2.9% of trials.

The RT data were subjected to an ANOVA with race and orientation ofthe face as within-subject factors. This analysis gave a main effect of race,F(1, 15)=26.13,p


38/46

196 VALENTINETABLE 6

Mean RTs to Classify Intact and Jumbled Faces, as a Function of Race and Orientation(Experiment 5)Intact Faces Jumbled Faces

White Black Mean White Black Mean~-

Upright 61 1 765 688 739 707 723Inverted 719 843 78 1 796 724 759

(0.8) (3.1) (1.9) (4.7) (0) (2.3)(1.6) (6.3) (3.9 ) (7.0) (0) (3.5)

Mean 665 804(1.2) (4.7)

Note: YO rrors are shown in parentheses

carried out. Errors were made on 2.9% of trials. The mean RT of correctresponses are shown in T able 6.A 2 x 2 ANO VA revealed a main effect of race, F( 1, 15) = 1 1.17, p


39/46

FACE RECOGNITION 197number of other-race faces in mem ory m akes the exemplar density of oth er-race faces relatively low.One limitation of Experiment 5 and Valentine and Bruces (1986a)experiment is that only subjects from one race took part. The interactionfound by Valentine a nd Bruce was interpreted a s evidence tha t recognition ofother-race faces is more impaired by inversion than recognition of own-racefaces. However, the possibility ca nn ot be excluded t ha t the black faces usedin the experiment would show a greater effect of inversion even to blacksubjects. A cross-cultural stud y would be needed t o eliminate this possibility.How ever, it should be noted tha t Ellis an d Deregowski (1981) found tha t achange in pose affected recognition of othe r-race faces more tha n recognitionof own-race faces using a cross-cultural design. Similarly a cross-culturalstudy would be needed to confirm that the effect of race on classificationfound in Experiment 5 was genuinely d ue to a n o ther-race effect.The results of Experiment 5 an d V alentine and Bruce (1986a) provide a tleast tentative evidence that the multidimensional framework can be appliedto the effect of race in face processing. Two studies of recognition tasksprovide a note of caution, however Buckhout and Regan (1988) examinedthe interaction between the effect of race and inversion in recognitionmemory for unfamiliar faces. Both black and white subjects took part(presumably residents of New York). Buckhout and Regan found an own-race advantage but no Race x Orientation interaction. They do not reportseparate analysis of false positives and hits but use a percent correct m easure.It is possible that the use of this measure may obscure an effect on falsepositives alone a s found by V alentine and Bruce ( 1 986a).Rhodes et al. (1989) studied the effect of race and inversion in a cross-cultural study using Chinese and Europ ean faces. They used a two -alternat-ive forced choice test proce dure and found a greater inversion effect for own-race faces than other-race faces (i.e. the opposite interaction to th at found byValentine & Bruce, 1986a). In their first experiment the significant Race x Or-ientation interaction was found in forced choice RT but not accuracy. In asecond ex periment the interac tion was fou nd only in forced choice accuracyan d only when the test face pairs were not highly similar. Th e use of a forcedchoice test proce dure in R hodes and colleagues study may have obscured theeffect in false positives found by Valentine and Bruce, but this does notexplain why the opposite interaction should be found. A number ofparam eters m ay a ccoun t for the different results of the three studies that haveinvestigated the interaction of race an d inversion, e.g. the difficulty of tasks,the test procedures an d measures used, the subjects experience of other-ra cefaces, hom ogene ity of stimu lus sets, etc. Fur th er research is needed t o resolvethe issues raised by these studies.



40/46

198 VALENTINEGENERAL DISCUSSION

An account of structural encoding of faces has been outlined in which themental representation of faces is considered in terms of locations in amultidimensional space. W ithin this general fram ewo rk, tw o specific modelswere identified, a norm-based coding model and a purely exemplar-basedmodel. Although the detailed assum ptions of th e models differ, bo th m odelspredict a grea ter effect of inversion for typical faces than for distinctive facesin recognition tasks but not in a classification task. These predictions weresupp orted. It w as found t ha t inversion w as more disruptive to recognition oftypical faces th an to recognition of distinctive faces in a recognition mem orytask using previously unfamiliar faces (Experiments I and 2), and in afamiliarity decision task using fam ous faces (Experiment 3). In E xperiment 4,it was fou nd th at th e effects of inversion an d distinctiveness were additive in aface classification task . These results are consistent with inversion causing a nincreased error of encoding that only has a differential effect on distinctivean d typical faces when th e tasks req uire individual faces to be discriminated.Th e multidimensional space framewo rk was also extended to account forth e effects of race. It was propo sed tha t other-race faces form a class of facesthat violate the statistics of the own-race face population. Other-race faceshave been found t o behave like typical faces in a recognition task (Valentine& Bruce, 1986a) and like distinctive faces in a face classification task(Experiment 5 ) . The extension of the multidimensional space framework toaccount for the effects of race is, at this stage, somewhat tentative as*datafrom studies that include full cross-cultural controls are conflicting. How-ever, the multidimensional space framework potentially provides a unifiedaccount of the effects of distinctiveness, inversion, and race. A furtheradv antage is tha t it is applicable to recognition of both familiar (e.g. famous)faces an d previously unfamiliar faces.The multidimensional framework (in either its norm-based coding orpurely exemplar-based forms) provides a principled framework that canaccount for the effects found in the experiments reported here. Can someother theory account for these results rendering the multidimensional spaceframewo rk unnecessary?It is possible that the interaction between distinctiveness and orientationfound in Experiments 1-3 could be explained in terms of disruption toprocessing of configural information in inverted faces (Carey & Diamond,1977; Diam ond & Carey , 1986). Recognition of distinctive faces may be lessaffected b y inversion because distinctive faces can be m ore easily encoded interms of component information (i.e. information based on isolated featuresrather than global properties). T his interpretation of the da ta is no t mutuallyexclusive to the multidimensional space framework, but it begs the questionof the basis of distinctive information in faces. Th e multidimensional space



41/46

FACE RECOG NITION 199framework is neutral in respect to the issue of whether distinctive informa-tion is conveyed predominately by com pone nt o r configural information. Itshould be possible to separate an account based solely on disruption toprocessing of configural information from the more general distinctivenessaccount offered by the multidimensional space framework by exploring theinteraction between distinctiveness an d transform ations o ther than inversion(e.g. masking with visual noise). The multidimensional framework predictsthat any experimental manipulation that impairs encoding of faces shouldinteract with distinctiveness in the same way as inversion. The account interms of disruption to processing of configural inform ation is specific to theeffect of inversion; different accounts wou ld be required for each tran sform a-tion found to interact with distinctiveness. Further research is required toexplore the effects of other trans form ation s on recognition of distinctive andtypical faces.The predictions of the multidimensional space framework a nd the compo-nent/configural information differ in regard to the interaction between theeffects of race and orientation on recognition memory for faces. Based onDiamond and Carey's (1986) distinction between isolated features andsecond-order relational information, Rhodes et al. (1 989) predict thatrecognition of other-race faces will be less disrupted by inversion thanrecognition of own -race faces. Although the compo nent/configural inform a-tion distinction and the multidimensional space framework make differentpredictions concerning the Race x Orientation interaction in recognitionmem ory, the available d a ta ar e conflicting (see discussion of Experiment 5) .In order to provide a full account for the present data in terms of acom ponent/configural inform ation distinction, it is also necessary to explainwhy distinctiveness affects a face classification task and to account for theadditivity of the effects of distinctiveness an d inversion in such a task. It isplausible to argue that second-order relational information is irrelevant tothe task because only discrimination of first-order relational properties isrequired (i.e. Is the configuration face-like?). This argument could accountfor the lack of an interaction between distinctiveness and inversion. How-ever, it is then not clear why the presence of the more distinctive isolatedfeatures of a distinctive face should affect the classification decision at all iftheir arrangement conforms to the first order relational properties of a face.A th ird possible accoun t would be to argue tha t th e da ta reported reflectresponse bias. The dat a in recognition tasks could arise if subjects are morelikely to respond positively: ( 1 ) to typical faces than to distinctive faces;(2) to inverted faces than to upright faces; and (3 ) to other-race than toown -race faces. However, a simple response bias model can no t account forthe data in recognition tasks (Experiments 1-3) because in all experimentsthe critical orientation by distinctiveness interaction was found in A ', a


8/2/2019 A unified account of the

a uniﬁed account of the effects of distinctiveness, inversion, and race in face recognition

Documents