Four-to-six-year-old children use norm-based coding in face-space

Abstract

Children's performance on face perception tests does not reach adult levels until adolescence, a result which, a priori, could be due to qualitative change in face mechanisms with age, quantitative change in these mechanisms, or improvements in general cognitive abilities that are not face-specific (e.g., memory, attention). In adults, the major functional mechanisms of face recognition include holistic/configural processing and face-space coding. Previous research has established that holistic/configural processing is present by 4–6 years of age. Very little, however, is known about face-space coding in children. Here, we demonstrate that 4–6-year-old children show adaptation aftereffects for figural distortions (expanded/contracted, eyes up/down), providing the first evidence of aftereffects for identity-relevant information in children younger than 8 years. We also show that in 4–5 year-olds, as in adults, face aftereffects are stronger for adaptors far from the average (extreme distortions) than for adaptors closer to the average (mild distortions). This result provides the first compelling evidence that face-space coding is norm-based in children younger than 8 years of age, and rules out a qualitative shift from exemplar-based to norm-based coding as the source of developmental improvement in face identification performance beyond preschool age.

Introduction

Our social interactions are frequently guided by information that we rapidly extract from faces. This information includes the identity, gender, ethnicity and emotional state of an individual. To extract this information we must be sensitive to very subtle differences between highly similar visual patterns. This is particularly true for face identification. Our ability to discriminate thousands of faces is often argued to result from expertise acquired through extensive experience discriminating faces. For this reason there has been considerable interest in determining whether specific perceptual mechanisms of face identification develop during childhood as experience with faces accumulates. 

In the present study, we provide the first investigation of whether a change in the basic organization of face space with age—from exemplar-based coding to norm-based coding—might play an important role in explaining why preschoolers perform so poorly relative to adults in face identification tests. We contrast this hypothesis with the alternative idea that face-space coding is qualitatively adult-like in its reliance on norm-based coding at an early age. 

It is well established that children's performance in face identification tasks improves dramatically between preschool ages and late childhood and continues to improve into adolescence (e.g. Bruce et al., 2000; Carey, Diamond, & Woods, 1980; Mondloch, Geldart, Maurer, & Le Grand, 2003). A priori, there are three possible types of explanations for this finding. First, there could be qualitative changes in the way faces are processed with one or more key functional properties of adult face processing emerging during childhood, such as “special” holistic/configural processing (perceptual integration of the internal face regions; Carey et al., 1980), or norm-based coding in face-space (coding of individual faces in terms of their deviation from the average; cf. Valentine, 1991). Second, these mechanisms could be present early but one or more could quantitatively mature with age, that is, holistic/configural processing or face-space coding might be present but weaker in children than in adults (Johnston & Ellis, 1995; Mondloch, Le Grand, & Maurer, 2002). Or third, all the key face perception mechanisms could be both qualitatively and quantitatively mature in preschoolers and all further age-related improvements in laboratory tasks could reflect development in other aspects of cognition such as the ability to sustain attention or the use of memory strategies (Crookes & McKone, 2009; Pellicano, Rhodes, & Peters, 2006). 

While there is active debate about the second and third of these possibilities (e.g. Crookes & McKone, 2009; Mondloch & Thomson, 2008), most recent articles have taken it for granted that the first theoretical option—late qualitative emergence of the key functional properties of adult face processing—has been rejected by the evidence. Certainly infants are sensitive to featural information in faces from birth (Simion, Farroni, Macchi Cassia, Turati, & Barba, 2002) and young children and infants demonstrate a wide range of qualitatively adult-like holistic/configural processing phenomena. These include large inversion effects for face discrimination (infants, Rose, Jankowski, & Feldman, 2008; Turati, Sangrigoli, Ruel, & de Schonen, 2004; 3-year-olds, Picozzi, Macchi Cassia, Turati, & Vescovo, 2009), inversion effects that are larger for faces than other objects (3-year-olds, Picozzi et al., 2009; 7-year-olds, Crookes & McKone, 2009), the composite effect (3-year-olds, Macchi Cassia, Picozzi, Kuefner, Bricolo, & Turati, 2009; 4-year-olds, de Heering, Houthuys, & Rossion, 2007), the part–whole effect (4–5-year-olds, Pellicano et al., 2006; Pellicano & Rhodes, 2003), integrating external and internal facial features (infants, Cohen & Cashon, 2001) and sensitivity to spacing between facial features (infants, Bhatt, Bertin, Hayden, & Reed, 2005; Thompson, Madrid, Westbrook, & Johnston, 2001; 4–5-year-olds, McKone & Boyer, 2006; Pellicano et al., 2006; though see Mondloch, Dobson, Parsons, & Maurer, 2004; Mondloch & Thomson, 2008). 

Importantly, however, the results described above establish early presence only of holistic/configural aspects of face perception. There have been surprisingly few developmental studies of face-space coding in children. It remains unknown therefore how similar preschoolers' face-space coding is to that of adults'. 

Adult face-space is defined by dimensions that correspond to the critical information used to discriminate faces. The average face lies at the center of the face-space. The role of this average in adult face coding has historically been controversial. In norm-based coding models of face identification the average face has a special status, in that other face identities are coded in terms of how they deviate from this average (see Figure 1A). The alternative model is exemplar-based coding in which a face is coded in terms of its absolute value on each dimension in face-space ( Figure 1B). In exemplar-based coding, an average value for any dimension can be calculated, and thus an average face can still be computed, but this average has no special status. In adults, there has been a long tradition of studies attempting to distinguish norm-based and exemplar-based coding, and it has been recognized that many findings that were initially considered direct evidence of norm-based coding can also be explained by exemplar-based models (e.g., distinctiveness effects, caricature effects, for a review see Rhodes, 1996). Recently, however, compelling evidence favoring norm-based coding of identity-related information has come from the use of adaptation aftereffect techniques combined with theoretical recognition that norm-based and exemplar-based coding can be associated with, respectively, opponent two-pool and multichannel neural coding models as taken from low-level vision research (Leopold, Bondar, & Giese, 2006; Leopold, O'Toole, Vetter, & Blanz, 2001; Loffler, Yourganov, Wilkinson, & Wilson, 2005; Robbins, McKone, & Edwards, 2007; Tsao & Freiwald, 2006; Webster & MacLin, 1999; Wilson, Loffler, & Wilkinson, 2002). 

Figure 1

 

Row 1. Simplified face-space models (cf. Valentine, 1991), showing only two dimensions, illustrating A. norm-based coding, in which a face is coded as a deviation from the average, and B. exemplar-based coding, in which a face is represented by the absolute value on the dimensions. Row 2. Neural coding of a single dimension of face-space (eye height) using C. norm-based coding, implemented as a two-pool opponent neural model, and D. exemplar-based coding, implemented as a multichannel neural model, Row 3. The predicted effects of adaptation on the perception of a physically average face when the adaptors are near to, and far from, the average face for E. a norm-based coding model and F. an exemplar-based coding model.

These two models are illustrated in Figures 1C and 1D. In opponent two-pool (norm-based) coding, all possible values along a single dimension, or trajectory, in face-space (e.g., the height of the eyes) are coded by the relative output of only two oppositely tuned pools of neurons: one that responds maximally to one end of the value range (e.g., eyes far up) and minimally to the opposite end (e.g., eyes far down), and the other that shows the reverse tuning. In this model, the norm on this trajectory (e.g., the average eye height) is perceived when the two pools produce equal output strength. ¹ In multichannel (exemplar-based) coding, values along a trajectory are coded by multiple pools of neurons each with bell shaped tuning centered over a different value. In this model, neighboring values (e.g., similar eye heights) will activate largely the same set of neurons, but very different values will activate non-overlapping sets of neurons. The average face is perceived when the pool, or pools, that respond to the average values on the trajectory are more strongly activated than pools coding values above and below the average. 

Importantly, both norm-based and exemplar-based models can explain the existence of adaptation aftereffects: in both cases, adapting to faces on one side of the average (e.g., a face with eyes higher than average, or a particular identity, say “Dan”) will alter the response ratios of the various neural pools such that the physically average face will subsequently appear distorted in the opposite direction (e.g., eyes low, or like “antiDan”, a face with all characteristics opposite to Dan's). However, there are several ways in which adaptation can be used to unambiguously dissociate the two models. First, the norm-based model predicts that adaptation on a trajectory passing through the average should produce a stronger aftereffect than adaptation between equally dissimilar faces on a trajectory not passing through the average; exemplar-based coding does not make this prediction. Second, the norm-based model predicts that adapting to the average face will not affect what is perceived as average; in contrast, the exemplar-based model predicts that adapting to the average face will distort perception of non-average faces. Third, the norm-based model predicts that a more physically extreme adaptor (e.g., a face with eyes raised much higher than average eye-height) will produce a larger change in the perception of the average face than will a less extreme adaptor (e.g., a face with eyes raised by a lesser amount). The reason for this is that in a norm-based model adaptors furthest from the norm produce the maximum change in the response ratio of the two pools after adaptation because adaptation reduces response in proportion to the initial unadapted firing rate (e.g., Maddess, McCourt, Blakeslee, & Cunningham, 1988; Movshon & Lennie, 1979). In contrast, the exemplar-based model predicts that adaptors close to the average will produce stronger changes in the perception of the average face than adaptors far from the average, because the amount of adaptation is dependent on the amount of overlap between pools coding the adapting values and the average values; indeed, sufficiently far-from-average adaptors will produce no aftereffect at all because they activate no neurons overlapping with the average face (Figures 1E and 1F). 

In adults, experimental findings have tested all three of these predictions for information related to face identity. Results clearly support an opponent norm-based coding model (opposite vs. nonopposite, Rhodes & Jeffery, 2006; adapting to average/undistorted face, Leopold et al., 2001; Webster & MacLin, 1999; distance of adaptor from average, Robbins et al., 2007). Norm-based coding has also been supported via single unit recording in adult monkeys (Freiwald, Tsao, & Livingstone, 2009; Leopold et al., 2006). 

Turning to children, what is known about their face-space coding? Some form of face-space appears to exist by preschool age. Like adults, children as young as four can judge distinctiveness in natural and manipulated faces (Johnston & Ellis, 1995; McKone & Boyer, 2006) and 6-year-olds can perceive caricatures as more distinctive than anti-caricatures (Chang, Levine, & Benson, 2002). 

These studies do not, however, address the question of whether children's face-space coding shares the fundamental adult property of coding faces as deviations from a norm because caricature and distinctiveness effects can also be explained by exemplar-coding in face-space (see Rhodes, 1996 for a review). Hence there is currently no direct evidence as to whether children younger than 8 years have even developed a face norm. The closest relevant findings are from infant studies of prototype extraction. Rubenstein, Kalakanis, and Langlois (1999) found that by 6-months-of-age, infants could average naturalistic faces to recognize a prototype. de Haan, Johnson, Maurer, and Perrett (2001) found that 3-month-old, but not 1-month-old infants were able to recognize a face that was the average of four faces with which they had previously been familiarized. Note, however, that while such studies might be taken as suggestive of norm-based coding at an early age, Nosofsky (1988) demonstrated computationally that prototype effects (recognizing a new averaged item as an old item) could be explained using a summed similarity exemplar-based approach. 

Here, we test the type of face-space coding used by children using the adaption aftereffect approach. A minimum prediction of norm-based coding of identity is that adaptation aftereffects should occur for identity-related information. This has been demonstrated in 8-year-olds (Anzures, Mondloch, & Lackner, 2009; Nishimura, Maurer, Jeffery, Pellicano, & Rhodes, 2008; Pimperton, Pellicano, Jeffery, & Rhodes, 2009) but there have been no previous tests in younger children. Further, 8 years is also the youngest age at which adaptation has been used to test for norm-based coding more directly. Consistent with such coding, Nishimura et al. (2008) showed that face identity aftereffects differed for adaptation to opposite identity faces versus non-opposite identity faces (although we note that this result is not conclusive given that the opposite and non-opposite adaptors were not matched for perceptual similarity to the target face). 

In the present study, our first aim was to seek evidence of figural aftereffects in 4-to-6-year-old children ( Experiment 1). We note that this age group is approximately the youngest for which adult-like adaptation paradigms can be used successfully. If preschoolers do show figural aftereffects, our second aim ( Experiment 2) was to provide a strong test of whether these aftereffects arise specifically from norm-based coding. We tested the norm-based coding prediction that adapting to a more extreme physical distortion (i.e., a “far” adaptor) should produce a larger figural aftereffect than adapting to a smaller physical distortion (i.e., a “near” adaptor). 

An alternative theoretical possibility is that face-space coding is not yet norm-based at 4-to-6 years, either because children have not yet developed a face norm or because they possess a norm but do not use it for coding face identity. This idea arises from the common assumption that face-space is built from ongoing exposure to faces, and that adults encode and recognize faces according to the average or norm that they have built up over their lifetime's experience (Johnston & Ellis, 1995; Valentine, 1991). It is not known how many faces need to be seen before this norm is established. Clearly, children have had less exposure than adults, and it is possible that 4–6-year-olds may not have had sufficient exposure to develop and/or use a norm. Under this view, face-space would use exemplar-based coding while still developing, and this would be followed by a developmental switch to norm-based coding when children have formed either a sufficiently good representation of the average face to function as a norm, or a sufficiently good representation of the dimensions on which to code individual faces as deviating from this norm. This developmental switch in the basic organization of face space might then play an important role in explaining why children's performance on face identification tests improves dramatically between preschool ages and late childhood. 

Experiment 1

The primary purpose of Experiment 1 was to establish whether preschool age children show figural face aftereffects. We adapted participants to either expanded or contracted faces (see Figure 2). Adapting to expanded faces causes subsequently viewed faces to appear somewhat contracted i.e. distorted in a way opposite to the adapting distortion. Adaptation to contracted faces causes subsequently viewed faces to appear expanded. The specific distortions we employed as adaptors have been used in previous studies of adaptation in adults (Jaquet & Rhodes, 2008; Jaquet, Rhodes, & Hayward, 2008; Rhodes, Jeffery, Watson, Clifford, & Nakayama, 2003; Rhodes et al., 2004) and children (Anzures et al., 2009). 

Figure 2

 

Experiment 1. A. Sample adapting stimuli showing an extremely contracted (−50, left) and extremely expanded (+50, right) face. B. Sample test stimuli showing the five distortion levels for one individual ranging from very contracted (−40) to very expanded (+40).

Method

Participants

Twenty-six adults (M = 21:11 years, range = 18–27, 17 female) and thirty-two children participated (M = 5:3, range = 4:1–6:5, 9 female). The adults were undergraduates at the University of Western Australia and completed the study for course-credit or a $5 honorarium. The children were recruited from kindergartens at the Child Study Centre, University of Western Australia. Participant's ethnicity was not formally recorded, but the majority of participants were Caucasian. 

Design

The adaptation task was presented as a game in which participants first learned to discriminate between faces that had been distorted in opposite ways, specifically, faces that had their inner features either expanded or contracted (see Figure 2). Pre-adaptation classification performance was measured for test stimuli presented at 5 distortion values (manipulated within-subjects) ranging from highly contracted to highly expanded (−40, −20, 0, +20 and +40; negative numbers indicate contraction and positive indicate expansion, see Figure 2). The dependent measure was the proportion of trials on which the participant judged the face to be expanded (relative to the participant's idea of a normal face). An adaptation phase followed in which participants were exposed to a series of faces all extremely distorted in one direction (e.g. extremely contracted). Adaptor condition (expanded, contracted) was varied between subjects, with half of each age group participating in each condition. In the post-adaptation classification phase participants classified the same stimuli as in the pre-adaptation phase. Adaptation to contracted faces, for example, should result in more faces being classified as “expanded” after adaptation than before. To avoid aftereffects arising from low-level retinotopic vision, the adaptor faces differed in identity from the test face stimuli, and eye movements around the faces were allowed. 

We also included an adult control group who completed an identical task to children. If children experience figural aftereffects then this should allow us to compare the size of children's and adult's aftereffects to seek further evidence of differences in children's and adults' face-space coding. The only previous study of figural face aftereffects in children (8-year-olds, Anzures et al., 2009) was unable to directly compare children's aftereffects with those of adults because different distortion levels were used for both adapt and test for the two age groups. 

Stimuli

The adapting and test stimuli were created from grayscale photographs of 24 faces (12 female). Each face was pasted onto a black rectangle measuring 11.3 × 14.2 cm, and an oval mask was applied that hid the outer hairline but showed face outline and the inner hairline. The oval mask had an average visual angle of 6.0° (w) × 7.5° (h) when viewed from an average distance of 58 cm. 

Eighteen of these faces (9 male) were used to make the adapting stimuli. The inner facial features were either expanded, or contracted using the Photoshop CS1 Spherize function, set to either the 50% expanded (+50) or 50% contracted (−50) setting (see Figure 2A). The remaining six faces (3 male) were used to create the test stimuli. Five versions of each face were created by varying distortion type (contracted/expanded) and amount (0, 20%, 40%) to produce a continuum from highly contracted (−40) to highly expanded (+40), see Figure 2B, resulting in 30 test stimuli (6 faces × 5 distortion levels). Forty additional faces were used to create practice stimuli (16 +50 distortions, 16 −50 distortions, 4 +30 distortions and 4 −30 distortions) which were printed on cards. 

Procedure

Participants were tested individually in a quiet room at their preschool or at the University. The same procedure was used for children and adults. Participants were randomly assigned to either the contracted or expanded adaptation condition. Stimuli were presented using Superlab 1.75 on a MacBook G2 laptop computer running OS9. 

Practice trials. The session began with a series of practice trials to ensure that children could discriminate the two distortions (contracted and expanded). Participants were told a brief story about two imaginary planets (contracted: red; expanded: blue). The people who lived on the red planet had ‘squashed’ faces (contracted inner facial features) and the people who lived on the blue planet had ‘stretched’ faces (expanded inner facial features). A red ball was placed left of the computer screen and a blue ball was placed to the right to represent the planets. Two cards showing highly contracted stimuli (−50), were placed in front of the red ball, one at a time, and participants were told that these were people from the red planet. They were then shown two extremely expanded stimuli (+50) depicting people from the blue planet. Participants sorted eight cards featuring extremely expanded and contracted stimuli (half of each) into two piles corresponding to the two planets. Errors were corrected, the cards shuffled and participants sorted the cards a second time. After correctly sorting all the cards participants were told that, “it would now get a bit trickier”, and they were given eight additional cards which depicted stimuli with weaker distortions (4 −30 and 4 +30). Once participants had correctly sorted these cards they proceeded to the pre-adaptation classification task. 

Pre-adaptation phase. In this phase we measured participants' classification performance in the absence of any adaptation. Participants were told that the faces of people from the blue and red planets would appear on the computer screen and their task was to decide which planet each person was from. They were told that some would be easy and some would be hard. The 30 test stimuli were then presented one at a time in random order. Each test stimulus was displayed for two seconds, followed by a response screen that was displayed until a response was made. Adult participants responded using colored keys on the computer keyboard. Children either responded verbally, by saying “red” or “blue”, and/or pointed at the red or blue ball, and the experimenter recorded their response by making the appropriate key-press. A response initiated the next trial. 

Adaptation phase. After completing the pre-adaptation trials participants were told they were now going to play a different game that was a lot like “snap”. This game was designed to encourage participants to attend to the adapting stimuli. Participants were told that faces would be shown, one at a time on the screen, and their job was to call “snap” whenever they saw the same face (i.e., identity) presented twice in a row. The experimenter stated that the faces would all be from one planet (either red/contracted or blue/expanded, depending on condition). The task was demonstrated using 20 pre-sorted cards showing extremely distorted stimuli (+50 or −50), depending on the condition. Four “snaps” occurred. The practice snap game was repeated until the experimenter was confident that the participant understood the game. The adaptation phase then began. 

The eighteen adapting stimuli were each shown four times for 1500 ms, with an inter-stimulus-interval of 200 ms. Total adaptation time was approximately two minutes. The adapting stimuli were presented in one of two pre-determined pseudo-random orders to ensure that all participants received the same number of snaps and that snaps were well spaced over the adaptation phase. Ten snaps took place in each sequence (averaging one snap every 12 seconds). Half the participants in each adaptation condition received each order. The experimenter monitored children's gaze and encouraged them to attend to the faces. 

Post-adaptation phase. In this phase, participants' classification performance on the 30 test stimuli was re-assessed. The procedure was similar to the pre-adaptation trials but to maintain adaptation the test trials were divided into three equal blocks, which were interspersed with one minute periods of top-up adaptation (snap game). Prior to starting each test block participants were required to classify two face cards, one highly expanded (+50) and one highly contracted (−50) to ensure that they remembered the planet associated with each distortion. 

The entire task took approximately 10–15 minutes to complete. Children were offered stickers upon completion and adults were debriefed. 

Results

We measured the effects of adaptation in two ways. The first was the pre- versus post-adaptation shift in the Point of Subjective Equality (PSE, namely the distortion level that was classified as expanded and contracted equally often.) This method has the advantage that it directly represents the change in the physical distortion level perceived as undistorted (“normal”). It is the method typically used to measure aftereffects in adults (e.g., Leopold et al., 2001; Rhodes & Jeffery, 2006; Robbins et al., 2007). In children, however, it has the potential disadvantage that it requires fitting psychometric functions (cumulative Gaussians, with minimum and maximum values fixed at 0 and 1, respectively) to each individual's data. Good fits to an individual child's data cannot always be obtained because children can perform only a small number of trials, and are also more likely to make random responses. These factors can result in erratic individual plots from which the PSE cannot be reliably estimated, so the participant's data must be excluded from analysis. 

To address this issue, our second measure was the difference in the overall proportion of “expanded” responses before and after adaptation. This method has previously been used in children (see Nishimura et al., 2008; Vida & Mondloch, 2009). It does not give a value to the face perceived as most “normal”, but it has the advantage that data from all participants are used. Both measures produced similar results and both are presented below. 

Shift in the psychometric function (PSE). For each individual participant, the mean of each cumulative Gaussian curve was taken as the Point of Subjective Equality (PSE). Figure 3 shows sample individual data and curves. Where poor fits were obtained, inspection of the curves indicated that this was typically due to erratic patterns of responding and/or poor accuracy for the most extreme distortions. Participants were excluded from the PSE analyses if R ² was less than 0.5 for either the pre- or post-adaptation function. Given the requirement to have a good fit both before adaptation curve and after adaptation, this resulted in exclusion of 14 children and 2 adults. This left 24 adults (11 in contracted and 13 in expanded conditions) and 18 children (9 in each condition) with good fits (Adults: Mean R ² = 0.94, SD = 0.08, range = 0.65–1.00; Children: Mean R ² = 0.81, SD = 0.14, range = 0.57–1.00). 

Figure 3

 

Experiment 1. Sample data and curve fits (Cumulative Gaussian) for four individual participants including a) an adult in the contracted adaptation condition (pre-adaptation PSE = −1.9, R ² = 0.95; post-adaptation PSE = −14.3, R ² = 0.95) with the PSEs indicated by the arrows, b) an adult in the expanded adaptation condition (pre-adaptation PSE = 8.1, R ² = 0.98; post-adaptation PSE = 25.0, R ² = 0.97), c) a child in the contracted adaptation condition (pre-adaptation PSE = 22.1, R ² = 0.90; post-adaptation PSE = 1.2, R ² = 0.97), and d) a child in the expanded adaptation condition (pre-adaptation PSE = 1.6, R ² = 0.93; post-adaptation PSE = 17.8, R ² = 0.79).

For each participant we calculated the size of the aftereffect by taking the difference between the PSE before and after adaptation, calculated as post-adaptation PSE − pre-adaptation PSE. Note that the direction of shift corresponding to an aftereffect is opposite for contracted and expanded adaptors: adaptation to contracted faces should result in the PSE shifting toward the contracted (−ve) distortions (i.e., contracted faces now look more normal than they did prior to adaptation); and the reverse is true for expanded. Thus, if aftereffects are present, then the aftereffect scores in Figure 4 should be negative for contracted and positive for expanded. 

Figure 4

 

Experiment 1. The size of the adaptation aftereffect as a function of age group and adapting condition measured via: A. the shift in the face perceived as most normal (i.e., PSE measure, N = 18 children, 24 adults) and B. the change in mean difference in the overall proportion “expanded” responses (all participants, N = 32 children, 26 adults). In both A. and B. an aftereffect in the contracted condition should result in a negative difference whereas an aftereffect in the expanded condition should result in a positive difference. To keep the direction of the aftereffects the same for both measures we reversed the direction of subtraction in the Proportion measure (B.). Error bars show one standard error either side of the mean.

The aftereffects on the PSE measure are shown in Figure 4A. Both adults and children showed shifts in the directions corresponding to adaptation. A two-way ANOVA was conducted with age group (children, adults) and adapting condition (expanded, contracted) as factors. We found a significant main effect of adapting condition, F(1, 38) = 35.92, p < .001, η _p ² = 0.49. This effect suggests that adaptation was present (i.e., as predicted, contracted shifted in one direction, M = −13.2, SD = 11.7, while expanded shifted in the other, M = 8.8, SD = 11.9. Note that this does not necessarily show that the magnitude of contracted and expanded adaptation differed.) The main effect of age group was not significant, F(1, 38) = 2.28, p = .14, η _p ² = 0.06. Nor was the interaction between age group and condition, F(1, 38) = 0.20, p = .66, η _p ² < 0.01. Planned single-sample t-tests, comparing the size of each aftereffect with zero, confirmed that significant aftereffects were obtained for both age groups, in both expanded and contracted conditions (all four ts > 2.23, all ps < .05, all ds > 0.60). 

Difference in the overall proportion of “expanded” responses. Data from all participants were included in analysis of the overall proportion measure. Inspection of the group means for proportion “expanded” responses (see Figure 5) shows that adaptation affected responses at all but the most extreme levels of distortion for both children and adults. The aftereffect was calculated as the change in overall proportion “expanded” ( pre-adaptation − post-adaptation). Note that the direction of subtraction is the reverse of that used for the PSE measure. This was done so that the direction of the shift corresponding to an aftereffect would be consistent across both the PSE and Proportion measures. Contracted adaptation should increase the overall proportion of expanded responses, relative to pre-adaptation, resulting in a negative difference. Adaptation to expanded faces should result in a positive difference. 

Figure 5

 

Experiment 1. Mean proportion of “expanded” responses at each distortion level pre-adaptation (gray circles) and post-adaptation (black open circles) for all participants (N = 13 adults per condition, 16 children per condition) as a function of expanded versus contracted adaptation condition. In the contracted condition an aftereffect increases the number of “expanded” responses, so the post-adaptation means (open circles) should be above the pre-adaptation (gray circles) means. The reverse is true for the expanded condition. Error bars show one standard error either side of the mean.

The mean aftereffects are shown in Figure 4B and replicate the pattern of results obtained for the PSE measure (see Figure 4A). Both adults and children showed shifts in the directions corresponding to aftereffects. A two-way ANOVA with age group (children, adults) and adapting condition (contracted, expanded) as factors showed a significant main effect of adapting condition, F(1, 54) = 41.95, p < .001, η _p ² = 0.44, and a significant interaction between age group and condition, F(1, 54) = 4.59, p < .05, η _p ² = 0.08. The main effect of age was not significant, F(1, 54) = 2.59, p = .11, η _p ² = 0.05. Planned t-tests showed that in the contracted condition adults' aftereffects were larger than children's, t(27) = 3.43, p < .01, d = 1.25. The age groups did not differ significantly in the expanded condition, t(27) = 0.32, p = .75, d = 0.12. Planned single-sample t-tests, comparing the size of each difference to zero, showed that significant aftereffects were obtained for both age groups in the contracted condition (both t's > 3.40, both p's < .01, both ds > 0.85) but only for adults in the expanded condition, t(12) = 2.40, p < .05, d = 0.67. The effect of expanded adaptation was marginal for children, t(15) = 1.91, p = .08, d = 0.48. 

Overall, results of the proportion-expanded measure closely replicated those of the PSE measure, with the exception that the tendency for contracted adaptation to produce a larger aftereffect in adults than in children reached significance only on the proportion measure (which had larger sample sizes). We also checked that there was no evidence of any systematic differences between the children who were excluded from the PSE analysis due to poor curve fits and those who were not. To do so, we compared the size of these two subgroups' aftereffects on the proportion measure. To boost power we collapsed across adapting condition (expanded, contracted). No significant difference was found (Excluded: M = 0.05, SD = 0.10, n = 14, Included: M = 0.06, SD = 0.09, n = 18, t(30) = 0.46, p = .65, d = 0.16). 

Pre-adaptation classification performance. We examined pre-adaptation performance for children and adults to determine whether age affected classification performance in the absence of adaptation. Inspection of Figures 3 and 5 suggests that adults more accurately classified distorted faces prior to adaptation, as indicated by the steeper slopes. We assessed the difference between the age groups using a statistic derived from the curve-fitting analysis. The standard deviation of the cumulative Gaussians provides a measure that is analogous to the slope, with a smaller standard deviation indicating more sensitive discrimination of the distortions. We calculated the standard deviations for the Gaussians fitted to each participant's baseline data for the 24 adults and 18 children included in the PSE analysis described earlier. ANOVA with age group (adults, children) and condition (contracted, expanded) as factors showed a main effect of age group, F(1, 38) = 14.71, p < .001, η _p ² = 0.28, with adults showing more sensitive discrimination (M = 19.3, SD = 11.6) than children (M = 35.4, SD = 15.7). There was no main effect of subsequent adaptation condition, F(1, 38) = 1.36, p = .25, η _p ² = 0.03 and no significant interaction, F(1, 38) = 1.30, p = .26, η _p ² = 0.03. 

Discussion

We have shown, for the first time, that like adults, 4–6 year-old children experience figural face aftereffects. A previous study has demonstrated that 5-year-old children show aftereffects for facial expressions (happy–sad; Vida & Mondloch, 2009), but the present finding is the first demonstration of aftereffects for identity-related facial information in children this young. Our finding confirms that preschoolers use face-space coding and is consistent with other evidence that children of this age have a face-space that has at least some adult-like properties (e.g., distinctiveness effects, Johnston & Ellis, 1995; McKone & Boyer, 2006). 

We also note that the adapting distortions we used were quite extreme, that is, far from the average. The fact that children experienced figural aftereffects on the PSE from distortions this far from the average face is suggestive of norm-based coding. This type of coding predicts strong aftereffects from extreme distortions; in contrast, within an exemplar-based model using fairly narrow bell-shaped tuning, it would be unlikely that neurons responsive to the extreme distortion would also be responsive to the average face, as would be necessary to produce an aftereffect on the PSE. The results of Experiment 1 thus provide the necessary basis for more direct tests that seek to distinguish norm-based from exemplar based coding by also testing the effects of a less extreme adaptor (see Experiment 2). 

Experiment 1 is also the first to compare the magnitude of figural aftereffects between preschoolers and adults. We found no significant age effects on the PSE measure. However, on the proportion measure adults showed significantly larger contracted aftereffects than did children, though there was no developmental difference in the expanded adaptor condition. In Experiment 2 we examine whether an age difference emerges for adaptation to a different figural distortion to determine whether children show smaller figural face aftereffects than adults for other kinds of distortions. 

In the absence of any adaptation, children less accurately discriminated the distortions than adults. This is unsurprising, and is consistent with both Anzures et al. (2009) and Crookes and McKone (2009) who found that children required more extreme expanded/contracted distortions than adults to achieve comparable levels of performance. Theoretically, the interpretation of the age-related improvement is ambiguous. Adults' superior discrimination performance could reflect improvements in the sensitivity of face-space tuning with age, that is, a perceptual change within the face system. Alternatively, it could equally reflect general cognitive improvement with age, with no underlying change in perceptual ability: for example, if poorer attentional control in children than adults results in children experiencing concentration lapses more during testing, then this would make children's performance less accurate than adults' due to guessing, even when discriminating extreme distortions. 

Experiment 2

In this experiment we sought more direct evidence that preschool children's figural face aftereffects arise specifically from norm-based coding and not from exemplar-based coding. The norm-based (two-pool) coding model predicts that adaptors that are far from the average will affect the perception of stimuli at the physical average more than will adaptors that are near the average. In contrast, the exemplar-based (multichannel) coding model predicts that an adaptor near the average (i.e., with only a mild distortion level) will affect the perception of stimuli at the physical average more than will a far adaptor (i.e., with an extreme distortion level; see Figures 1E and 1F). In Experiment 2, we therefore compared the size of children's aftereffects when adapting to near-the-average and far-from-the-average adaptors to determine which of these two kinds of coding preschool children use. 

We used similar methods to those used in Experiment 1 but with the following two changes. First we added a size change between adapt and test images to further rule out low-level retinotopic adaptation as the source of children's aftereffects (Pimperton et al., 2009; Zhao & Chubb, 2001). Second, we used a different figural distortion, raising or lowering the position of the eyes in the face relative to the naturally occurring eye-position (see Figure 6). We varied the degree to which the eyes were raised in the adapting faces to create “far” and “near” adapting conditions. Children in the “far” condition adapted to faces with eyes raised by an extreme amount (50 pixels above their original location; henceforth “50-up”) and children in the “near” condition adapted to faces with eyes raised by a less extreme amount (10 pixels above their original location; henceforth “10-up”). Manipulations of eye height have previously been tested in adults (Robbins et al., 2007; Susilo, McKone, & Edwards, 2010), where it has been shown that adults produced significantly larger aftereffects on perception of the average face from 50-up adaptors than from adaptors closer to the average (e.g. 5-up, 12-up). We predicted that, if children show qualitative similarity to adults in using norm-based coding, they should similarly experience larger aftereffects for the “far” adaptors than “near” adaptors. As in Experiment 1 an adult group were tested under the same conditions as children to allow comparison of children's and adults' aftereffects. 

Figure 6

 

Experiment 2. A: Sample adapting stimuli showing a 50-pixels-up (left) and 10-pixels-up (right) face. B: Sample test stimuli showing the five different pixel displacement levels for one target face.

Method

Participants

Sixty-two children were recruited from preschools in Canberra, Australia. Nine did not complete all sections of the experiment and we excluded data from one further participant due to failure to score above chance in the pre-adaptation phase. The final sample of 52 children (29 female) had a mean age of 5:0 years (range = 4:5–5:8 years). Forty-one adults were recruited at ANU, Canberra and UWA, Perth, Australia. Two participants (both female) showed a strong response bias to say “eyes down” (≥80%) on pre-adaptation trials, making it difficult to detect an aftereffect in the predicted direction (i.e. an increase in eyes down responses after adaptation) so their data were removed prior to any analysis. This left 39 adults (21 female) with a mean age of 26 years (range = 17–64 years). All children and adults were Caucasian (the same race as the face stimuli). 

Design

Adapting condition was varied between subjects, with data from 28 children and 20 adults in the far adaptor condition (50-up), and 24 children and 19 adults in the near (10-up) condition. In the pre- and post-adaptation phases, test stimuli were presented at 5 different pixel displacement values (manipulated within-subjects) ranging from eyes very low to eyes very high (−25, −8, 0, +8 and +25 pixels; negative numbers indicate eyes lowered and positive indicate eyes raised, see Figure 6). The dependent variable was the proportion of ‘eyes up’ responses. 

Stimuli

The adapt and test stimuli were made from grayscale photographs of male faces following the methods of Robbins et al. (2007). The eyes, nose and mouth region was cut from these photographs and pasted onto a common background face from which the features had been removed but the face outline (hairline, chin, ears) and skin texture remained. For the ‘zero’ (undistorted) face, the natural spacing of internal features was maintained. The distorted versions were created by raising or lowering the eye region and blending this region into the background face. The adapting stimuli were made from nine faces. Two versions were made of each face. The eyes were raised by 50 pixels for the 50-up (+50) condition and raised by 10 pixels for the 10-up (+10) versions. The noses were lengthened to preserve the face configuration in the 50-up versions. Six different male faces were used to create the test stimuli. Four different eye height versions were made of each face (−25, −8, +8 and +25 pixels) which, along with the original undistorted face (0), resulted in 30 test faces (6 identities × 5 eye heights). A negative number indicates that the eyes were moved down. Figure 6 shows examples of the adapting and test stimuli. Adapting stimuli subtended a visual angle of 7.2° (w) × 8.7° (h) and test stimuli subtended a visual angle 9.0° (w) × 10.8° (h). 

Twenty additional practice stimuli were made from 20 grayscale photographs of females. Six had their eyes raised by 50 pixels, four had their eyes raised by 30 pixels, six had their eyes lowered by 50 pixels and the remaining four had their eyes lowered by 30 pixels. These were presented at the same size as the test stimuli. 

Procedure

Participants were individually tested in a quiet room at their preschool or at the university. The trials were presented using PsyScope software (Cohen, MacWhinney, Flatt, & Provost, 1993), on an i-mac computer running OS9 with a CRT screen. Children responded verbally and the experimenter recorded these responses on the keyboard. Adults used the keyboard directly. 

The procedure was similar to Experiment 1 with three main exceptions. First, all practice stimuli were presented on the computer rather than on cards. Second, the test stimuli remained on the screen until a response was made. Third, the snap game during adaptation was replaced by a simpler task in which children called out whenever they saw a “red” face. The red faces were created by adding a red tint to the grayscale photos in Photoshop. 

Participants were told that they would see faces that had been in stretching-up and stretching-down machines and they had to identify which machine each face had been in. They were then shown four face stimuli simultaneously on the computer screen. Two had their eyes moved up by 50 pixels and two had their eyes moved down by 50 pixels. The experimenter pointed to these faces and explained which machines they had each been in. Participants were then asked to identify which machine each had been in. If the participant was incorrect or did not know, the machines were explained again. 

Participants then completed practice trials in which they were shown four +50 and four −50 practice stimuli sequentially and in random order. The stimuli remained on screen until a response was recorded by a key press, after which the next trial began. Eight practice trials with less extreme distortions (four +30 and four −30 pixels stimuli) followed. Participants were given verbal feedback. 

Pre-adaptation phase. The thirty test faces were presented, one at a time, in random order and each remained on the screen until a response was recorded. For each face the experimenter asked children “Was this face in the “stretching-up” or “stretching- down” machine?”. The experimenter made the appropriate key-press on the child's behalf. Adults were told to classify each face according to which machine it had been in and respond using the keyboard. No feedback was provided. 

Adaptation phase. In this phase, participants were told that they would play a new game. They were told that all the faces had been in the stretching-up machine, but this time some of the faces would be red, and it was their job to call out “red” (children) or press the spacebar (adults) when they saw a red face. To confirm that each child could recognize the color red each child was asked to identify something in the room that was red. 

Adapting stimuli were presented in a pseudo-random, pre-determined order. Seventy-two adapting stimuli were shown for 1500 ms each with an interstimulus interval of 500 ms. The adaptation phase ran for approximately two minutes. Red stimuli appeared on average every seven faces (10 in total). The experimenter monitored children's gaze and encouraged them to attend to the faces. 

Post-adaptation phase. This was similar to the pre-adaptation phase except that three blocks of test trials (10 trials in each) were interspersed with 1 minute periods of top-up adaptation. Each time participants switched games they were reminded of the procedure. To confirm that they remembered the differences between the two machines, each set of test trials began with a brief classification test in which participants were asked to classify two practice faces (−50 and +50). 

The entire task took approximately 15 minutes. Upon completion children were offered a sticker and adult participants were thanked for their participation. 

Results

The same measures and analyses were used as in Experiment 1. 

Shift in psychometric function (PSE). Cumulative Gaussians were fit to each participant's pre- and post-adaptation data. Figure 7 shows sample individual data and curves for representative children and adults in each condition. Participants' data were removed from analysis if the fits were poor (R ² < 0.5) for either curve. Data from eight children (of 52) were excluded on this basis leaving 23 children in the 50-up condition, and 21 children in the 10-up condition. The remaining fits were good (mean R ² = 0.91, SD = 0.07, range 0.75–1.00). No adult data were removed (mean R ² = 0.99, SD = 0.02, range 0.91–1.00). 

Figure 7

 

Experiment 2. Sample data and curve fits (Cumulative Gaussian) for four participants, a) Adult in 10-up adaptation condition (pre-adaptation PSE = 0.5, R ² = 1.00; post-adaptation PSE = 0.8, R ² = 0.99), b) Adult in 50-up adaptation condition, (pre-adaptation PSE = 2.5, R ² = 1.00; post-adaptation PSE = 7.5, R ² = 1.00), c) Child in 10-up adaptation condition (pre-adaptation PSE = −1.6, R ² = 0.97; post-adaptation PSE = −0.8, R ² = 0.99), d) Child in 50-up adaptation condition, (pre-adaptation PSE = 5.1, R ² = 0.92; post-adaptation PSE = 16.6, R ² = 1.00).

For each participant we calculated the size of the aftereffect as PSE post- minus pre-adaptation. Adapting to “eyes up” faces should result in a positive difference if an aftereffect is present. The predictions of norm-based coding are that a clear aftereffect should be present from the extreme adaptor (50-up), and that this should be larger than the aftereffect present from a mildly distorted adaptor (10-up). 

As illustrated in Figure 8A, both children and adults showed this pattern. ANOVA with age group (adults, children) and adapting condition (10-up, 50-up) as factors confirmed this, showing a main effect of adapting condition, F(1, 79) = 5.30, p < .05, η _p ² = 0.06. Neither the main effect of age group, F(1,79) = 0.16, p = .69, η _p ² < 0.01, nor the interaction between age group and adapting condition were significant, F(1, 79) = 2.57, p = .11, η _p ² = .03. To confirm that children, considered alone, show norm-based coding we also performed an a priori comparison of near and far adaptors for children. Results demonstrated norm-based coding, showing that children's aftereffects were significantly larger for far adaptors than near adaptors, t(42) = 2.26, p < .05, d = 0.68. 

Figure 8

 

Experiment 2. The size of the adaptation aftereffect as a function of age group and adapting condition measured via: A. the shift in the face perceived as most normal (i.e., PSE measure, N = 44 children, 39 adults) and B. the change in mean difference in the overall proportion “eyes up” responses (all participants, N = 52 children, 39 adults). In both A. and B. an aftereffect should result in a positive difference for both the 10-up and 50-up conditions. To keep the direction of the aftereffects the same for both measures we reversed the direction of subtraction in the Proportion measure (B.). Error bars show one standard error either side of the mean.

Finally, planned t-tests showed that aftereffects were significant for both adults and children in the 50-up condition (adults; t(19) = 2.34, p < .05 d = 0.52, children; t(22) = 2.78, p < .05, d = 0.51), but not in the 10-up condition (adults; t(18) = 1.68, p = .11, d = 0.22, children; t(20) = 0.31, p = .76, d = 0.07). 

Difference in overall proportion “eyes up” responses. For this analysis, all participant's data were included (Children N = 52, Adults N = 39). The mean proportion of “eyes up” responses at each level of pixel displacement both before and after adaptation are shown in Figure 9. Adapting to 50-up faces decreased the number of “eyes up” responses, as predicted for both adults and children. Adapting to 10-up faces appeared to have little effect on “eyes up” responses. 

Figure 9

 

Experiment 2. Mean proportion of “eyes-up” responses for each different pixel displacement level for pre-adaptation (gray circles) and post-adaptation (open circles) phases for each adaptation condition for adults and children. For both 10-up and 50-up conditions an aftereffect will decrease the number of “eyes-up” responses, so the post-adaptation means (black) should be below the pre-adaptation (gray circles) means. Error bars show one standard error either side of the mean.

The size of the aftereffect was calculated as overall proportions of “eyes up” responses for pre-adaptation minus post-adaptation. Adaptation to eyes-up faces should result in a positive value (i.e. fewer “eyes up” responses after eyes up adaptation). The distribution of aftereffect scores in the adult 10-up group was significantly non-normal (Kolmogorov–Smirnov test, p < .05), showing a negative skew. The scores for children in the 10-up condition were similarly skewed, though the deviation from normality was not significant ( p = .069). To correct for this skew all scores were square-root transformed, resulting in distributions with no significant deviations from normality. Untransformed means and standard deviations are reported. Results ( Figure 8B) showed the pattern predicted by norm-based coding, with larger aftereffects for the 50-up than 10-up adaptors, for both adults and children. ANOVA confirmed this pattern showing a significant main effect of adapting condition, F(1, 87) = 6.36, p < .05, η _p ² = .07. Neither the main effect of age group, F(1, 87) = 0.40, p = .53, η _p ² < 0.01, nor the interaction were significant, F(1, 87) = 1.01, p = .32, η _p ² = 0.01. Again, a priori testing for norm-based coding in children showed that children's aftereffects were significantly larger for far adaptors than near adaptors, t(50) = 2.25, p < .05, d = 0.56. 

Planned t-tests showed the aftereffects were significantly different from zero for both adults and children in the 50-up condition (adults; t(19) = 2.71, p < .05, d = 0.61, children; t(27) = 2.95, p < .01, d = 0.57) but not in the 10-up condition (adults; t(18) = 1.19, p = .25, d = 0.27, children; t(23) = 0.37, p = .72, d = 0.07). 

Pre-adaptation classification performance. In the absence of adaptation, age again affected classification sensitivity (standard deviation of pre-adaptation curves; see Figure 9). ANOVA with age group (adults, children) and condition (10-up, 50-up) as factors showed a main effect of age group, F(1, 79) = 27.33, p < .001, η _p ² = 0.26, with adults showing more sensitive discrimination (M = 3.96, SD = 3.89) than children (M = 10.34, SD = 6.61). There was no main effect of condition, F(1, 79) = 1.11, p = .29, η _p ² = 0.01 and no significant interaction, F(1, 79) < 0.01, p = .94, η _p ² < 0.01. 

Given that a significant aftereffect was not found for the 10-up adaptor groups it was also of interest to determine whether participants could perceptually distinguish the +10 distorted adaptor faces from undistorted (0) faces. Reliable discrimination of +10 from 0 has previously been demonstrated using these face stimuli for adults (Robbins et al., 2007; Susilo et al., 2010). Here, although we did not measure participants' perceptions specifically of the 10-up faces, inspection of pre-adaptation responses in Figure 9 shows that both children and adults reliably discriminated +8 from 0 distortions. We confirmed that adults in the 10-up adaptor condition classified +8 distortions as having “eyes up” (M = 0.85, SD = 0.20) significantly more often than they did for 0 faces (M = 0.38, SD = 0.27), t(18) = 9.26, p < .001. Likewise children in the 10-up adaptor condition classified +8 distortions as having “eyes up” (M = 0.62, SD = 0.30) significantly more often than they did for 0 faces (M = 0.43, SD = 0.31), t(23) = 3.04, p < .01. Therefore, both children and adults perceived the 10-up adapting faces as distorted, relative to undistorted faces. 

Discussion

Experiment 2 has shown that five-year old children's figural face aftereffects are larger when the adaptor is far from the average than when it is near the average, as predicted by norm-based (two-pool) coding and contrary to the predictions of exemplar-based (multichannel) coding. Further, there were no interactions between adapting condition and age, on either measure. Therefore, this experiment provides strong evidence that five-year old children use norm-based coding in face-space and hence do not differ qualitatively from adults in this fundamental respect. 

Adaptation to the 10-up faces failed to produce significant aftereffects in either children or adults. This pattern of results is not surprising under a norm-based coding model. Norm-based coding predicts that aftereffects will become consistently weaker as the adaptor distortion is made milder, reaching zero for an average-face adaptor (e.g. Webster & MacLin, 1999), and this consistent reduction is found for adults (Susilo et al., 2010). In the present experiment, it may be that a 10-up adaptor is close enough to the average to have produced no measurable aftereffect. Importantly, the lack of significant effect cannot be attributed to inability of the participants to perceptually distinguish 10-up faces from the average, given that both adults and children showed significant discrimination of even 8-up stimuli as different from the average face. We also note that, in adults, Robbins et al. (2007) and Susilo et al. (2010) found significant aftereffects using even less distorted (5 up) adaptors. However, the procedures in those studies used far more trials and/or top-up adaptation before every test trial, enhancing the chance of detecting a small aftereffect. These procedures are suitable for adults but could not be implemented in 5-year-olds. 

Turning to comparison across age groups of the magnitude of the aftereffects, we found no evidence to suggest that children and adults differ systematically in the size of their aftereffects. Neither the main effect of age nor the age interaction with adaptor value approached significance. Finally, we again found the standard result that adults showed more sensitive discrimination of the distortions than did children. This result could reflect increasing sensitivity in face-space tuning with age or general cognitive improvements (see Experiment 1 discussion). 

Another interesting finding concerned the ability of 4–5 year-olds in Experiment 2 to make above-chance discriminations of subtle differences in spacing between features in unadapted faces. Ability to discriminate spacing changes has traditionally been associated with holistic/configural processing (for a review see McKone & Yovel, 2009). In the literature on development of holistic/configural processing, there is general agreement that preschoolers can discriminate large changes of spacing information that fall outside the range that is normal in front-view faces. However, there has been controversy about whether they can also discriminate spacing changes within the normal range (McKone & Boyer, 2006; Mondloch & Thomson, 2008). Our present results showed that 4–5-year-olds, like adults, could discriminate the subtle difference in spacing between our +8 pixels versus 0 pixel stimuli. Importantly, the +8 pixel deviation falls within the normal range both physically (less than one SD higher than average according to the head measurements of Farkas, Hreczko, & Katic, 1994) and perceptually (although adults discriminate +8 from 0, they rate both as “normal”; McKone, Aitken, & Edwards, 2005). 

General discussion

Our primary findings were that 4–6-year-old children experience figural face aftereffects, for two different types of figural distortions, and that figural aftereffects were larger when the adaptor was far from the average face than when it was near. These results demonstrate that young children use norm-based (two-pool opponent) coding for faces. We can therefore rule out a switch from exemplar-based (multichannel) to norm-based coding as an explanation of the subsequent improvements in performance on face discrimination tasks with age. We also confirmed that the source of the aftereffects is not low-level retinotopic visual processing, given that aftereffects generalized across identities and eye movements ( Experiment 1) and identities, eye movements and changes in size ( Experiment 2). These results argue that children's and adults' aftereffects likely result from adaptation of the same visual mechanisms, presumably mid- and high-level mechanisms crucial to coding structural face information. In summary, we conclude that children have a face-space and that this space is qualitatively similar to that of adults in that it uses norm-based rather than exemplar-based coding. 

Our findings therefore set the upper age range for the emergence of norm-based coding in face-space at 4–5 years of age. Minimally then, norm-based coding must emerge some time during the first five years of life, though it is possible that it could be innate. There is suggestive evidence that face-space and norm-based coding is present early in infancy. Five-to-eight month old infants discriminate faces differing in averageness/distinctiveness (Rhodes, Geddes, Jeffery, Dziurawiec, & Clark, 2002). Further, both human and monkey infants respond to an average face as if it is familiar within a few months of birth (de Haan et al., 2001; Myowa-Yamakoshi, Yamaguchi, Tomonaga, Tanaka, & Matsuzawa, 2005). These findings argue for the early presence of some form of face-space. Given the difficulties of interpretation of prototype effects (Nosofsky, 1988) it remains an open question whether this face-space is norm-based, or whether there is, at some point in a child's very early development, a switch from exemplar-based to norm-based coding as the child experiences an increasing number of faces. 

Regarding the origin of the improvement in face identification beyond the preschool years, our results rule out a major late (post 5-years) reorganization in face-space coding, from exemplar-based to norm-based coding. However, the results leave open the possibility that more subtle changes in face-space coding could contribute to the ongoing improvement. While we did not find consistent differences between the magnitude of aftereffects in children and adults, results that agree with previous findings of equal identity-aftereffects in 8 year olds and adults (Nishimura et al., 2008; Pimperton et al., 2009), we did find that, in Experiment 1, children showed significantly less contracted adaptation than did adults on one measure. This suggests that children perceive contracted faces as less distorted than do adults. This could occur, for example, if adults' coding of all or some dimensions in face-space was finer-grained than children's, resulting in greater overall sensitivity to the contracted distortion. Alternatively, this difference in the size of the contracted aftereffect could result from qualitative differences between children's and adults' face-space, if children and adults differ in the number/or nature of the dimensions used and these dimensions code face attributes that are reflected in the contraction distortion (see Johnston & Ellis, 1995; Nishimura, Maurer, & Gao, 2009). The distortions used in the present study (contracted, expanded, eyes-up) likely affect some but not all possible dimensions in face-space hence it remains possible that future studies, manipulating different figural information, may be informative in determining if there are differences in children's and adults' face-space. 

Children were consistently poorer than adults at discriminating the distortions in the unadapted conditions in both Experiments 1 and 2. This developmental improvement in discrimination accuracy is consistent with many previous findings (e.g. Anzures et al., 2009; Crookes & McKone, 2009; Mondloch & Thomson, 2008; Mondloch et al., 2002, 2003; Nishimura et al., 2008; Pimperton et al., 2009). The interpretation of this developmental improvement, however, is intrinsically ambiguous. The age differences in discrimination performance could reflect subtle differences in sensitivity of face-space coding but could equally reflect differences in general cognitive mechanisms that influence accuracy more broadly (e.g., concentration on the task). 

Finally, we note that, to date, studies of face aftereffects in children, including the present one, have primarily used adult faces as stimuli. In other face perception tasks interactions between the age of the observer (child versus adult) and the age of the faces used as stimuli (child versus adult) have been found (Anastasi & Rhodes, 2005; Kuefner, Macchi Cassia, Picozzi, & Bricolo, 2008; Susilo, Crookes, McKone, & Turner, 2009). For example, children's memory for children's faces is better than their memory for adult faces and vice versa for adults (Anastasi & Rhodes, 2005). It may be that similar interactions occur in the magnitude of aftereffects. The one study that has used children's faces as stimuli was not able to address this issue because they were unable to compare the size of children's and adult's aftereffects due to the different distortion levels used for children and adults (Anzures et al., 2009). Using child face stimuli together with exactly the same procedure in children and adults may cast further light on whether there are systematic differences in the size of face aftereffects for child and adult observers. 

Conclusions

We have shown that, like adults, 4-to-6 year-old children use norm-based coding for faces, as indicated by figural face aftereffects. We have therefore ruled out one major qualitative change in the type of face-space coding used as the source of improvement in face identification performance beyond ages 4–6 and set the upper age limit by which norm-based coding must emerge during development. Overall, the literature now demonstrates that two key functional properties of adult face perception, holistic/configural processing and norm-based face-space coding, are present by ages 4–6. Consequently, all further improvement in face identification skills must reflect (a) more subtle changes in the use, or refinement of, these or other face perception mechanisms and/or (b) more general cognitive development. 

Acknowledgments

We thank the children, parents and staff at the Child Study Centre at The University of Western Australia, the Aranda, Turner, Campbell and North Ainslie Preschools for participating in this study and the WA and ACT Departments of Education. We also thank Kate Crookes and Libby Taylor for assistance in data collection and two anonymous reviewers for their helpful comments on the manuscript. This research was supported by an Australian Research Council Discovery Grant (DP0770923) to the first, second and fifth and sixth authors and by ARC grants to EM (DP0984558) and GR (DP0877379). 

Commercial relationships: none. 

Corresponding author: Linda Jeffery. 

Email: [e-mail]. 

Address: School of Psychology, M304, University of Western Australia, 35 Stirling Highway, Crawley, WA 6009, Australia. 

Footnote

Footnotes

1  Note that in Figure 1 we have drawn the tuning functions as linear. Both single unit recording from face-selective cells (Freiwald et al., 2009) and psychophysical evidence in humans (Susilo et al., 2010) argue that tuning is often, but not always, linear.