The Quarterly Journal of Experimental Psychology

ISSN: 1747-0218 (Print) 1747-0226 (Online) Journal homepage:

Telling faces together: Learning new faces through
exposure to multiple instances

Sally Andrews, Rob Jenkins, Heather Cursiter & A. Mike Burton

To cite this article: Sally Andrews, Rob Jenkins, Heather Cursiter & A. Mike Burton (2015) Telling
faces together: Learning new faces through exposure to multiple instances, The Quarterly
Journal of Experimental Psychology, 68:10, 2041-2050, DOI: 10.1080/17470218.2014.1003949

To link to this article:

Accepted author version posted online: 21
Jan 2015.
Published online: 17 Feb 2015.

Submit your article to this journal

Article views: 513

View related articles

View Crossmark data

Citing articles: 6 View citing articles

Full Terms & Conditions of access and use can be found at

Download by: [University of Birmingham] Date: 08 January 2017, At: 08:14

Aberdeen. York. Without constraints. Over longer periods.1003949 Telling faces together: Learning new faces through exposure to multiple instances Sally Andrews1. they are highly accurate. http://dx. and from the Economic and Social Research Council. two images of the same person can encounters. In a card-sorting task for facial identity. van Montford. Within across different instances of the same person the homogeneous category of faces. Face perception. In photographs. Rob Jenkins2. York YO10 5DD. This minimal contextual information appears to support viewers in “telling faces together”. ences in the appearance of different people allow The face of an individual may vary for a number of us to identify thousands of individuals successfully. Here we report two card-sorting experiments in which we manipulate whether participants know the number of identities present. For example. Mike Burton.323262. 10. more visually distinct than two images of different health. and A. reasons. Glasgow. that UK 2 Department of Psychology. Stable representations. 68.. 2041–2050. University of Glasgow. University of York. Face recognition has generally been approached people. the fact that two instances of the same person as does rotation of the head along horizontal or ver- might not look the same.g. nodding or turning). Within a single In Experiment 2 we show that exposure to within-person variability in the sorting task improves performance in a subsequent face- matching task. effects of ageing. E-mail: mike. No. & Burton. an additional subset of one must not only discriminate between different variability is encountered due to characteristics of Correspondence should be addressed to A. Adini. participants sort faces into many identities. first published online 17 February 2015) We are usually able to recognize novel instances of familiar faces with little difficulty. Heather Cursiter3. UK [ES/J022950/1]. UK (Received 4 August 2014.1080/17470218. different photos of the same unfamiliar face are often seen as differ- ent people. The research leading to these results has received funding from the European Research Council under the European Union’s Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreement n. White.2 1 School of Psychology. when told the number of identities present. yet recognition of unfamiliar faces can be dramatically impaired by natural within-person variability in appearance. This appears to offer a fast route to learning generalizable representations of new faces. In order to identify a face successfully. UK 3 School of Psychology. expressions Yet. However. subtle differ. Keywords: Face recognition. accepted 17 December 2014. Face learning.doi. tical axes (e. matically. taken under differ. Department of Psychology. our understanding of this ability often ignores and speech alter the relative appearance of features. Between ent conditions. and adiposity also influence appearance dra- people. University of York. Moses. but also code information that is stable from the perspective of telling faces apart. 2011). Mike Burton1. and and nature of lighting. there may be differences in the source look very different. (Jenkins. 2015 Vol. both of which affect the Ullman (1997) demonstrated that it is possible to apparent texture and general appearance of the find two instances of the same person that appear face. Identity. © 2015 The Experimental Psychology Society 2041 . University of Aberdeen.THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY.

Cowan. perspective. In a demonstration of in a change detection task for facial identity how difficult unfamiliar face recognition is. the role of expectancy in face processing participants to sort a set of 40 unfamiliar faces is large enough to disrupt processing of even into piles.. introducing where many different faces may be present. & in that context.. sequent recognition. matching as actually being unfamiliar. McNeill. Bruce et al. 1982). ancy of seeing both men at the podium. Megreya & geographical location. In recog. within-person variability. Perhaps because of the expect- performed almost perfectly. the internal features from Bill Clinton’s face onto mode 9). O’Toole. 1999. such that learned faces are posed that familiar and unfamiliar face processing less likely to be recognized if the context is altered rely. different instances as the same person. face belong to the same person. Interestingly. (Dalton. 68 (10) . whereas viewers familiar with the faces the face of Al Gore. In an unconstrained situation.e. The difficulty that initially seemed familiar were correctly rejected is not restricted to memory. Yet for unfamiliar For example. photos of the same person into a single identity Different images typically contain a combination (i. (Dalton. it has been pro. 2015. & Avery. In an entertaining demon- were only two identities present. & Young. 1993. 2011) asked In fact. a few degrees of head rotation can and low-familiarity faces. It is well documented that familiar title). on separate processes (e. Hay. 1976). 1993). 2000. diary study. 2003. one pile per identity. semantic contexts (e. Approximately 16% of the occasions when nition memory tasks. Liu. 2011. This demonstration pro- containing both identities.. job Burton.g. the capture device (focal length. and exter- takes made by unfamiliar viewers were not mistakes nal features serving as a visual cue. Some context effects may be related to the expect- cation is much harder—even in seemingly ideal ancy of seeing a particular face. Palmisano. and even the presence of an additional face face recognition occurs with remarkable ease and presented at the same time as the target face accuracy (Bruce. the mis. Burton. Longmore et al. there even one dimension of variability to newly learned is no reason to suspect that two instances of a faces can dramatically reduce recognition accuracy. 1999. Changes in lighting in ordinary situations. 1996. Burton. them apart”). it is rarely a problem for rec. which is highly robust to additional cues. provides an indication of conditions have a similar disruptive effect (Hill & the importance of context (Young. Krouse. to some extent.ANDREWS ET AL. 2010). Instead. 1982. people seemed vides useful clues to the utility of context in face to have great difficulty in cohering different processing. unfamiliar stration of this. ally distinct as the same person may require ognition of familiar faces. context may provide the reduce accuracy close to baseline in simple identity mechanism to overcome variability and to perceive tasks (Favelle. simply changing expression familiar faces went unrecognized were due to from smiling to neutral between learning and test meeting in an unexpected context. most people do of confusing two people: There were very few piles not detect the alteration. (Austen & Enns..). Wilson. Sinha and Poggio (1996) duplicate viewers sorted them into many piles (mean 7. 2002. Learning context has a clear effect on sub- In the light of these findings. “telling people together” rather than “telling of variability along these dimensions. approximately 87% of simultaneously present high-quality images of the time this was because it would be implausible unfamiliar people turns out to be a very difficult or impossible for the candidate person to appear task (Bruce et al. 2005). 2008. Jenkins and colleagues (Jenkins et al. This holds for contexts such as Hancock. documenting failures of identification Edelman. Watkins. For example. Longmore. 1998). situations (e. & Ellis. 2008). & Burton. & Bültoff. White.g. whereas unfamiliar face identifi. 1999). Clutterbuck & manipulating expectancy can affect performance Johnston. 1985). 2006).. Bruce.. 2042 THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY.. Ho. While within-person variability is unavoidable The ability to reconcile instances that look visu- in normal interactions.g. When faces can introduce errors (Bruce.5.. Simons & Levin. & Tulving. An early 1981. 1998). Bruce. Although there highly familiar faces. & Bruce. However. etc.

Previous studies have tended to emphasize included as long as the faces were not occluded in number or duration of training encounters. Materials Results and discussion We are particularly concerned here to use images whose appearance varies naturally. reflecting a ticipants in a sorting task perform almost perfectly normal range of variability over which target when they are told that only two faces are present. such Maasland. rather than on top (SD = 9. as a result of Free sort environmental differences such as lighting. B. and C. Twenty participants were given no indication about rectly) that there are only two. and were free to move photos corrected-to-normal vision. 68 (10) 2043 . images were selected within the recognition. For this reason. we show that par. such that the whole face was visible and filled the majority of the slide. and Chantal Janzen.) and were tations. THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY.05 years were visible at the same time. find it difficult to integrate different photos of an unfamiliar face into a single identity (Jenkins Procedure et al. we examine are referred to as “ambient images” (Jenkins et al. number of identities to when they are told (cor. chosen using a random UNFAMILIAR FACES number generator. range = 18–60). FACE LEARNING WITH MULTIPLE INSTANCES In the experiments reported here. Emphasis was placed on accuracy. Using these criteria. we use the same card for whom many pictures exist.M. people are typically recognized. the acquisition of stable face representations from 2011. were then scaled to passport size (35 mm × 45 mm). but who are sorting to fast-track face learning. Here any way and had sufficient resolution (defined as we show that assimilation of inherent within. 2011). Participants Participants were 40 students (26 female) from the were encouraged to place instances of the same University of Aberdeen. 2013). These images ent identity piles (median = 6.. To test whether this difficulty can All participants were handed a pile of shuffled cards be overcome by manipulating observers’ expec. C. Images for sorting varied images.M. Sutherland et al. Photographs were con- acquisition of a new face identity. and 20 C. person variability is a key part of face learning.02. We show that unknown by our UK participants (Bridget the sorting task enhances incidental learning. taken at different Method times (two-sort). Participants with no time restriction on the task. but generate many piles when this instruction is we chose two celebrities from The Netherlands. were gathered from an internet search.M. and the remaining 20 were informed that the images were of two different people.8 differ- general health and focal distance. between piles as many times as they liked. 2015. who participated for course person alongside one another so that all images credit or a small fee.. how many identities were present (free sort).. of the two identities (B. This is a novel approach to face learn. In Experiment 1. withheld. and laminated.J.J. The mean age was 22. the first 65 images for each and that a simple technique can support efficient search target were retained. a minimum of 285 × 190 pixels for these purposes). pose range three-quarter to full face.J. Forty of these images EXPERIMENT 1: SORTING (20 B.. All had normal or of one another. Images were ing.). we compared sorting performance when asked to sort the images into separate piles so that viewers are given no information about the all the instances of the same person were together.). that participants develop robust representations of From an internet search using the names of these new identities that can be recruited in subsequent Dutch celebrities. Participants sorted photos into a mean of 6. range = 2–16). In Experiment 2. verted to greyscale and cropped to a size of 285 × 190 pixels. printed in high-quality Experiment 1 builds on the finding that observers greyscale.

We define an intrusion error as an nition might lie in failure to discriminate individ- instance of a face appearing in a pile containing uals. . Jenkins & Burton. solution. and it was counted as a single misidentifi.73. Davies-Thompson.. 2015.34. A one-sample t-test revealed that this difference of facial features is used to distinguish between was significantly greater than the two that were faces (e. they also suggest a novel Two-sort approach to face learning: the transition from unfa- Ten participants sorted the two identities perfectly. partici- median and mode intrusion errors were zero. Simons & et al. miliar to familiar. difference between familiar and unfamiliar face rec- ognition. 1998). intrusions. EXPERIMENT 2: FACE LEARNING These results confirm that free sorting is a very difficult task for unfamiliar viewers. Two Face A pictures in considerable difficulty cohering different instances a pile with majority of Face Bs are counted as two of the same face into a single representation. . there was no significant difference in the number and we briefly review this topic below. there has been faces. d = 0.56.ANDREWS ET AL.. 2000. It is well established that recognition of familiar simply instructing participants that there are only faces is considerably more robust than recognition two identities radically improves their performance. Davies-Thompson. Within this context. and across changes in image. Burton. p . and there to do so relatively quickly (i. actually present. This was the only situation the basis of visual information—but only when where there was a pile consisting of 50% of each viewers are specifically encouraged to do so.05. Maurer. Intrusions were summed across all a when given no expectation of how many individuals participant’s piles. p . generalizable representations (Carbon & example. one might expect that The two different identities were rarely confused the well-known difficulty of unfamiliar face recog- for each other. Recently. they have is counted as one intrusion. The problem of face recognition is typically renewed interest in FRUs and how they code posed as one of distinguishing between faces. the notion of “configural processing” is Leder. these findings show that. a single instance particular task. 1986. However. 2002). unfamiliar observers are largely able face recognition is much more generalizable to overcome the natural variability in faces. 2044 THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY. Le Grand.. 2005. Instead. During the two-sort task. t(38) = 1. The fact that this task suddenly across participants (median 1). In the next experiment we test the An independent-samples t-test revealed that possibility that this task can accelerate face learning. For stable. Only two participants made more different stimuli into a single representation—and than two intrusion errors per identity. However. 1990). with pants appear to learn to cohere superficially range 0–11.g. 2003.001. Hancock expectancy (Austen & Enns. imental session).e..g. & often recruited to propose that the spatial layout Andrews. with no standing interest in the nature of visual represen- errors of misidentification the most common tations underlying familiar face recognition. Newling. However. d = 2. and these ranged from 0 to 3 they will see. t(19) = 5. & Mondloch. identity. The results are important in understanding the cation error. One participant becomes very easy with the addition of this infor- made one pile containing one instance of Face A mation demonstrates that faces can be cohered on and one of Face B. participants have little difficulty in of Face A in a pile containing a majority of Face B distinguishing between faces. familiar Levin. 68 (10) . denoted face recognition units (FRUs) by some This is a particularly interesting finding.10. Gouws. over a single exper- were no piles consisting of 50% of each identity. While unfamiliar face proces- Given minimal context information in the form of sing is relatively image bound (e. 2009. Bruce it illuminates viewers’ difficulty with unfamiliar & Johnston. for this the majority of the second face. of unfamiliar faces. 2011). because researchers (Bruce & Young. So. of misidentification errors made between two- sorters and free sorters. There has been a long- discern identity-specific information.

In fact. or reimbursed a small different instances as the same person.. It has been by viewers in daily life. If sorting provides a good way of these experiments typically involves testing the learning a new face. we tested matching performance using ning multiple sessions (Bonner. saturation. with an unexpected matching task. 68 (10) 2045 . THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY. of experimental settings vary greatly. 1982) and and Johnston (2002. We can see from sorting and matching tasks that seeing one face image is not adequate Method for successful face learning (Bruce et al. shown that seeing a face for longer (Reynolds & In Experiment 2. Jenkins. 2010. which is often “controlled people used in sorting phase. and other factors will each participants carried out a face matching task in impact learning. 2011). 2005) that such attempts may obscure matching test comprised new pictures of the important information. (two-sort) conditions. FACE LEARNING WITH MULTIPLE INSTANCES & Andrews. Hancock. & unseen photos of the faces from the sorting task McNeill. Schweinberger..7 able to learn one face from a pair (or set). Here. Two groups of participants is common between encounters and may be diag. In the second phase. The stimuli and cing variability in an individual face. while these findings (seen IDs). & Kane. challenge for any account of face learning is to accommodate this within-person variability. Burton. who all reported normal face from only one instance. did not take part in the sorting phase and pro- that changes in lighting direction. Thus. or a still in greater matching accuracy for seen than unseen from a seen video. One approach to understanding versions of the task. 2010) photos by identity. learning faces alongside semantic or 1. In the following novel identities. surprisingly little is known about how images of unfamiliar faces. 1999. Bruce (1994) proposed that by experien. range = 17–41). the fee for their time. 2015. 2013. Participants Burton et al. task followed by a matching task. (18 in each) completed the free-sort and two-sort nostic of identity. ceeded directly to Phase 2.. experiments we examine learning using ambient However. Jenkins & Burton. & showed the same or different people. Burton. However. span. then we expect this to result same image that was initially learned. Clutterbuck nition (Klatzky. observers would be or corrected-to-normal vision (mean age = 21. But faces encountered outside IDs at test. If it Participants were 54 students (20 male) from the were possible for observers to successfully learn a University of Aberdeen. 2013. of the kind recognized such representations are formed. recognition in (unseen IDs). Jenkins. under free or constrained can improve subsequent recognition. However. which years. 2005) have shown that pair- accompanying differences in neural activation wise matching of faces is a good measure of famili- (Kaufmann. we then followed up this sorting task personal information leads to more accurate recog. Jenkins et al. focal length. we become procedure for the sorting task were identical to better equipped to determine which information those in Experiment 1. age. Martin. we have argued which they judged whether pairs of photos (Burton. 2003). as discussed above. we asked participants to sort Pezdek. 2011). it is highly likely 18). & Burton. Participants who took part would then allow them to successfully identify were given course credit. information are recognized more confidently. A third group of participants (n = the stimuli used: For example.. 1992) or more often (Xue et al. arity and is a sensitive index such that accuracy Faces that are successfully associated with semantic steadily improves with increasing familiarity. all pose. respectively. as well as completely out” of experimental stimuli. Faces in the White. as in Experiment Furthermore. However. 2009). it has been suggested that exposure to Design and stimuli within-person variability may be necessary for face The experiment comprised two phases: a sorting learning. as described in face variability is to attempt a parameterization of Experiment 1. and photos of previously unseen faces provide some useful foundations.

Results p .85. Different-person pairs Table 1. two-sort. Free sort . h2r = . significant difference in the time taken to viously unseen identities—Wendy van Dyke (45 sort stimuli in two-sort and free sort conditions W. range nificant difference between no sort and two-sort = 3–16).33. their similar age and hair colour to the four target identities (B.11.M. (mean = 1. . (p = . p . these were randomly mixed-design analysis of variance (ANOVA) selected independently for each participant. mode = 0. None of these 240 photos had samples t-test was conducted to establish whether been presented in the card-sorting task. misidentification errors were low (p = .01) . p .001. F(2.01) . . 51) = 5.D.22. one of the target ID.03. Tukey’s HSD revealed a sig- median number of 6. pendently for each participant.001.D. Simple main effects analyses showed significant (novel identities). A one-sample t-test revealed that partici. These were 15 same-ID and p .17. h2r = .5 identities (mode = 3. sorting-task identities) one of a foil ID.02. For the matching task. which was not evident pants in this condition generated significantly more between no sort and free sort conditions identities than the two that were actually present.44 vs.S. and ID type. F(2. F(1. showed main effects of sort type.46. completed in total. t(34) = 1. and C.D. 102) = 9. The remaining (7.. respectively). F(2.01.36. (sorting-task identities).73 (.198).ANDREWS ET AL. .). 2015. observers generated a For sorting-task IDs. One hundred and twenty matching trials were p .03. h2r = . 60 were photos of foil identities. which were randomly selected indepen. which was not present for Sorting phase novel IDs. h2r = .. There was also no significant difference t(17) = 23. d = 0. A 3 (sort type: free sort.) and Tatjana Simic (45 T. An Table 1. there was a significant effect of sort type for sorting-task IDs. Note: Standard errors in parentheses. in one condition than the other. median = 0.5 (mode = 0.05. and sort) × 2 (ID type: novel vs.22. Order of presentation was effects of ID-type for free sort.S. . h2r = .15. Accuracy for the face matching test is shown in dently for each participant.01.47. and T. In the free sort condition. 102) = 1.001. d = 11.01. range = 0–7).15. and 15 same-ID and 15 p .17. F(1.).J. and T. 240 additional images While response time was not a primary depen- were collected using the method described in dent variable in this study.009). 51) = 45.73 (.76 (. t(34) = 0.001. .40. inde. but not for the no sort condition. As in between two-sort and free sort conditions Experiment 1.05. In neither the free sort nor the two-sort No sort . 7. W.98. There was also a significant 15 different-ID trials for each of B. p .01) d = 0. h2r = .S..01) condition did any participant make any piles con- sisting of 50% of each identity.86 (.).J. p . Same- person pairs comprised two images of the same Face-matching phase person. Once again.75 (. F(1.373). p = . C.J. mixed (unblocked) and randomly ordered.02) . 51) = 20. Condition Novel IDs Sorting-task IDs dentification errors made between the two-sort and the free sort conditions. F(1. 2046 THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY.81 (.52. F(2.5 min. 45 C. This revealed no M. 90 were participants spent longer looking at the images novel photos of the sorting-task identities (45 B. 68 (10) . h2r = . no comprised two images. .29.001.83. . interaction between these factors. .01) Two-sort . an independent- Experiment 1. h2r = . Mean face matching accuracy following different types of sort independent-samples t-test revealed that there was no reliable difference in the number of misi. different-ID trials for each of W. range = 0–11). 51) = 33. a further 90 were photos of two pre.. p . and two-sort conditions.63. selected based on p = .M. conditions (p = . . Further. 51) = 5.32. The median number of errors in the two-sort condition was 0. 51) = 1.07.

unfa- these people helps to form a representation of miliar observers are able to categorize different face them. 2011) can be have encountered previously. 2015. (b) the matching test was (the unseen identities).. These data give quite a clear indication that the Using sorting and matching tasks. instances as the same person. 2005) we observe that stable face resentations are sufficiently flexible to be useful in representations can form through experience of recognition of novel instances. By using a sensitive measure of familiarity evidence to suggest that stable representations have (simultaneous matching task. differences in per- but were focused on the sorting task. There are three important points to the experiment. note that the “no sort” unexpected. In Experiment 2 we used the Furthermore. and face matching between those identities we used so this advantage is not brought about by memory for the sorting task. and However. Instead. 2002. reconcil- ticipants to show better performance than the free ing different instances as the same person requires sort participants—a trend that is present only for that transient within-person variability that occurs the (already seen) sorting-task identities. that rather simply overcome. though less conclusive. The comparison of the different types of sort is If stable representations form through experi- also interesting. we note that the design used here means that participants learn only a small number of iden- The clearest finding from this experiment is that tities. the present sorting task leads to incidental learning of the iden. we find that the matching task Experiment 1. merely to say whether pairs of simul- taneously presented faces were the same or GENERAL DISCUSSION different. and that famili. effects of within-person variability when identify- arity is enhanced for the identities that participants ing unfamiliar faces (Jenkins et al. and that these rep. this occurs without any deliberate sorting task as an incidental learning procedure. This provides tations. In Johnston (2005). within-person variability. the were matched more accurately during the later same identities were used for learning throughout test phase. with quite a high unseen images when they are encountered later. Johnston. FACE LEARNING WITH MULTIPLE INSTANCES Discussion Finally. one might expect some evidence that the two-sort might lead to that reconciling more instances as the same better learning of the identities than the free sort. support for this idea is weakened by the that the stable information is retained into a rep- nonsignificant pairwise comparisons and can there. formance at test seems to rely entirely on previous ing test was ostensibly independent of the sorting exposure to these identities. degree of accuracy. but as a side-effect to explore the development of stable represen- of encountering them in this context. attempt to learn the new faces. THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY. phase—participants were not asked to remember anything. 2011). (c) the match. To address the possibility that note in order to understand these data: (a) The there is an inherent difference in the difficulty of matching test employed entirely new images. resentation. There is ence with natural variability. person results in stronger individual face represen- This is suggested by the trend for the two-sort par. 68 (10) 2047 . Clutterbuck & begun to form for these faces. tations (Jenkins & Burton. With additional infor- exposure to a range of very different images of mation about the number of targets to expect. we observe that the debilitating is sensitive to levels of familiarity. Consistent with Clutterbuck and person variability on face identification. Because the experimental constraints require the faces of people seen during the sorting phase a large number of instances for four people. That is. Experience of more instances should fore only be regarded as tentative from these data. between different instances is disregarded. participants were not trying to group showed no difference in performance commit faces to memory during the initial phase. It seems then. which can be used to recognize new. therefore promote the extraction of stable. experiments explore the influence of within- tities involved. between these groups. and those used for comparison for particular photos.

that even context in the form of expectancy is ade. & Jeffery. Müller. Context information is 2007.. 68 (10) . context information. 2013. that are highly tolerant to variability even at an 2008). Jenkins & Burton. identities and how we are able explicitly to remem- Familiar faces. we observe accuracy is at ceiling (Burton et al. & Bruce. 2011). and within-person variability make different.. contributions to the problem determine which information is consistent—and of face recognition. 1999). 2011. past research has focused person—and in doing so rarely mistake one largely on how we are able to differentiate between person for another. as Reynolds & Pezdek. If few instances are required in order to experimental conditions variability becomes an determine which information is identity-specific obvious problem (Bruce et al. Burton (Bruce. 1999. This trend for free sorters to perform less well at match. more stable representation. yet under stability. as encounters of a face actually form a reasonable rep- opposed to simple exposure to instances. This is consistent with Bruce’s faces. We observe a free sorters and no sorters was not found. while familiarity was measured person’s face. We find some evidence to suggest using entirely new instances of the faces. Similarly to until faces are highly familiar. Free sorters generate far more we can be confident that these effects are not identities than two-sorters. the corresponding difference between number of images of each identity). Here that experience of variability is necessary in we show that such representations can form forming stable face representations that are useful through incidental learning. Caharel. Watkins et al. Notably. It therefore appears evident that between- (1994) earlier proposal that experience of within. rarely require ber faces over different encounters (Bonner. 2005). context information for recognition (Burton et al. 1976).. When new instances are used to suggest that this is because of the addition of update existing representations (Carbon et al. this would result in representations et al. 2009. Yet it is notable that However. 2008.ANDREWS ET AL. Leopold. More flexible face is transient. while additional instances that are sub- We rarely notice the influence of within-person sequently included then produce small increases in variability in normal social situations. but person variability is necessary. & Rossion. from experience of for recognizing previously unseen instances of natural variability.. This suggests that perhaps the first few encouraged with top-down constraints. condition (even though all observers see the same However. 1994).. stable representations were representations would then be more likely to be formed in the absence of explicit instructions to activated by a completely novel instance of that remember faces. (Dalton. on the other hand. these data are somewhat nuanced. and identification Young et al.’s (1985) early diary study. as it allows us to equally important. In order comprehensively to therefore identity-specific—and which information understand face learning processes. In Experiment 1 we show evidence to early stage. 1992. Xue et al. Ramon. known to influence the likelihood that the same this seemingly results in finer tuned representations. means that each “identity” generated comprises Matching accuracy for previously sorted faces fewer instances in the free sort than the two-sort was greater for two-sorters than for no sorters. Rhodes. observation is consistent with the idea that experi- ing learnt faces than did two-sorters. stable representations enable recognition of pre. Longmore et al.. 2010. Longmore et al. Here we address the problem of how we identify viously unseen instances (Davies-Thompson difference instances as the same person and show et al. encing a wide range of different images results in a forming no differently in accuracy to novel faces.. identity-specific information. and thus this from our data. These data offer a novel perspective on under- quate to identify different instances as the same standing face learning. and accuracy is high for learnt faces. resentation. person will be recognized on a subsequent occasion which show incremental tolerance to variability. 2003. 2011). 1993... Jenkins et al. 2010). we need to 2048 THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY. 1999. which necessarily merely the result of conscious image learning.. Burton. regardless of it will be an important topic of future research to whether they are learnt from a few or many establish the degree to which learning can be instances. while per. 2015..

(2005). 10(3). White. activation model. D. (2009). Memory & Cognition. Why has research in face recog. T. (2003). Hancock. Carbon. (1993). lecture. 2015. & Burton.. A. Jenkins. (2010). D. A. British Journal of Psychology. 10(5). Cerebral Cortex. V.. P. & Andrews. Perception. 40(7). R. A. Philosophical Transactions of the Royal Burton. (1994). V. Journal of Vision. B.. 1966–1976... & Bruce. 5–28. Henderson. Cowan. & Johnston. A. (1997).. D. & McNeill. S. A. bility. 51. J. R. (1999). & Andrews. 761– Bruce. Y. Bruce.. Journal of Burton. 47. Perception. M. G. Burton.D. 22(4). Image-invariant responses in face-selective Bruce. M. 1671–1683. Langton.. 81(3). V. (1996). B. Jenkins.-C. & Avery. British Journal of Psychology. M. H. 73. 313–323. images. M. J. Behavior Research A. A. (1986). Neuropsychologia. A. 21(4). R... J. Stable face rep- 361–380. A. Wilson. Psychology. (2005).. Psychological other factors. coding processes in face recognition. S. A. to know you: How we learn new faces. V. 17(1). M.. 68 (10) 2049 . Verification of face identities from images captured Favelle. 3. Hancock. T. S. Perception. J. 1–12. S. R.. J. Burton.. L. L. 366. M. T. Face on video. (2002). V. J. & Leder.. Leder.. (2011). J.. H. nition: The problem of compensating for changes in Adaptation effects of highly familiar faces: illumination direction. Psychology Section A. V. Cognitive Psychology... European Journal of Bonner. 47(1). Palmisano. M. 97–116... R. S. A. Face recognition in poor-quality video: 625–641. Burton. Science... P. Understanding face rec. R. 21(2). 42(1).. 77. & Burton. (2003). Greenwood. B.. & Burton. Journal of Cognitive Neuroscience. H. R. (1999). 105–116. (2013). Stability from variation: The case of iliar and unfamiliar faces in the human ventral face recognition . 330–337. A. Moses... Hancock. IEEE Transactions on Pattern Immediate and long lasting. The role of stimulus familiarity in effects in face learning. Change detection in Clutterbuck. resentations.The M. matching measure.. Analysis and Machine Intelligence. Burton.. G. Harsányi. Jenkins. R. A. V. (2007).. van Montford. A. A. Trends in Cognitive nition progressed so slowly? The importance of varia. 23. C.. Gouws. The Quarterly Journal of Experimental Davies-Thompson. (2005). S. V. 223–234. (2000). (1990). L. A. 256–284. & Enns. M. (2011). Variability in photos of the same face. Face recog. glasgow face matching test. & Cognitive Psychology.. R. P. & Young. & Kovács. regions do not explain the perceptual advantage for fam- P. T. J. sition of face representations across different Burton. 1467–1485. Adini. & Bruce. & Burton. Experimental Psychology: Human Perception and Understanding face recognition with an interactive Performance. 305–327. A. Exploring an attended face depends on the expectation of the levels of face familiarity by using an indirect face- observer. Austen... 4(9). N250 ERP correlates of the acqui- Methods. 1(1). (2013). V. Sciences.. Face adaptation: Changing stable representations of familiar faces within minutes? Advances in Experimental REFERENCES Psychology. K. & Ullman. 1159–1168.. Bruce.. M. 1627–1635. Bonner. J. ognition. Carbon. An image-dependent representation of fam- Bruce. M. & Johnston. viewpoint effects about three axes: The role of config- 5(4). the perceptual of facial surfaces. 35(8).-C. E. K. Schweinberger. 121(3). P. The effects of lighting on 66(8). M. 19(7). M. Vernon memorial stream. 31.. Z. ural and featural processing.. Getting Clutterbuck. Hill. 339–360. using a face matching task. Changing faces: Visual and non-visual Cognition. K. 1–7. McNeill. A. iliar face recognition.. THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY. White. Burton. X. R. nition: The power of averages. Bruce.. 784. Visual Demonstrating how unfamiliar faces become familiar Cognition. Quarterly Journal of Experimental Psychology. 985–994. Recognition of unfamiliar faces. 243–248. (1982). M.. Society B: Biological Sciences. FACE LEARNING WITH MULTIPLE INSTANCES consider experience of natural variability alongside Evidence from security surveillance. British Journal of Davies-Thompson. Newling. & Johnston. Y. A.. 527–536. 370–377. Jenkins. Robust representations for face recog. Memory and Bruce. & White.. G.. & Bruce.. R. Meet the simpsons: Top-down Dalton. context-dependent recognition. (2011). 32. (2003).. H. C. (2009). 986–1004. V. & Miller. Journal of Experimental Psychology: Applied. Cognition.. 286–291. M. The Kaufmann. 64–74. Strobach.

. Xue. & Jeffery.. (1981). 255–260. (2002). & Bültoff.. T. M. I think I know that face faces. Psychonomic Bulletin and Review. 105–118. R.. Learning faces from photographs. The Ramon. M. C. delay on face recognition performance. A. The faces that launched a thousand slips: Everyday diffi- speed of recognition of personally familiar faces.. Sutherland. M. 495–523... (1998). A. (2008). Sciences. R. P. (2011). C. M. H. Cognitive Sciences. Z. Q.. A. R.. A.. 2015. A. J. B.. D. Nature.. Dong. larity across repetitions is associated with better Stimulus-specific effects in face recognition over memory. M. Megreya. L. J. Oldmeadow. 644–649.. D. 865–876. J. Le Grand. H. & Tulving. H. Poldrack. 279– Krouse. Journal of Psychology. J. D. W. T. 68 (10) . Science. J. 66(5). (1985). pose change. K. Hay. F. L... K. Müller.. (1976). & Memory and Cognition. A. Martin. G. W.. A. S.. S. 38.. Rhodes. 127(1). 34(4). & Burton. W. (1996). 384(6608). Journal of Verbal Learning and Verbal Behavior.. 77–100. memory: The effects of exposure duration and encod- Memory and Cognition. Santos.. (1982). 897–904.-M. Leopold. changes to people during a real-world interaction... Chen. I. faces are not faces: Evidence from a matching task. E. 2050 THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY. Failure to detect Applied Psychology. A. 272(1566).. 505–517. 404. Watkins.ANDREWS ET AL. Trends in Context effects in recognition memory for faces.. Towler. Vision Research. J. Cognition. Klatzky. A. British Perception. changes in viewpoint. and 292. Liu. M. D. D. 437–449. Mumford. Applied Cognitive Psychology. J. A.. D.. Edelman. Effects of pose. Ho. & Pezdek. G. 6(6).. 330. Journal of Simons. Reynolds. Proceedings of the Royal Society B: Biological … . (1992). C. (2006). culties and errors in recognizing people. & Kane. C. C. A. (2005). Performance. E. G. Unfamiliar 15. & Levin. Greater neural pattern simi- O’Toole. 76. Burt. A. C. 97–101. (2010). & Mondloch. & Ellis. 651–654. A. W. 5(4). 2351–2363. & Young. L. L. Caharel. Journal of Social inferences from faces: Ambient images gener- Experimental Psychology: Human Perception and ate a three-dimensional model.. The dynamics of visual adaptation to Sinha. Longmore. Lu. 6. & Rossion. ing instruction. (2013). A.. Young. 40(4). R. & Poggio. (1998). The many faces of configural processing. 195–206. 10(3). M. 34(1). Maurer. J. Face recognition Semantic interpretation effects on memory for faces.. K. & Young.