Three studies on configural face processing by chimpanzees

Abstract

Previous studies have demonstrated the sensitivity of chimpanzees to facial configurations. Three studies further these findings by showing this sensitivity to be specific to second-order relational properties. In humans, this type of configural processing requires prolonged experience and enables subordinate-level discriminations of many individuals. Chimpanzees showed evidence of a composite-like effect for conspecific but not human faces despite extensive experience with humans. Chimpanzee face recognition was impaired only when manipulations targeted second-order properties. Finally, face processing was impaired when individual features were blurred through pixelation. Results confirm that chimpanzee face discrimination, like humans, depends on the integrity of second-order relational properties.

1. Introduction

Numerous studies have demonstrated that human faces are recognized primarily using the configurational arrangement of facial features. In their seminal article, Diamond and Carey (1986) differentiated two types of configural information present in faces. First-order relational properties refer to the relative arrangement of facial features which is similar in every face, i.e., the eyes are laterally displaced above the nose, which is above the mouth, etc. This type of configural information provides the means for recognizing faces from other categories of visual stimuli. Newborn babies, for example, are more attracted to face-like patterns than nonface like patterns, suggesting that attraction to and/or the ability to use first-order relational information might be present from birth (Goren, Sarty, & Wu, 1975; Mondloch et al., 1999; Valenza, Simion, Cassia, & Umilta, 1996). Second-order relational properties reflect the prototypical arrangement of facial features into a standard template, i.e., the spacing between the eyes and mouth, which is more or less unique to every face. This prototypical configuration of facial features establishes the formation of individual representations, enabling specific individuals to be distinguished from one another through the use of facial information alone. Thus, while first-order relational properties enable the identification of faces at a basic categorical level, second-order relational properties facilitate the subordinate-level classification of faces, i.e. the ability to distinguish Mary from Jane.

One of the most robust methods for investigating configural vs. feature-based face processing is the inversion effect. Inverting faces by rotating them 180° from their typical upright orientation results in impaired recognition response times and accuracy (Valentine, 1988; Valentine & Bruce, 1988; Yin, 1969). This is because inverting faces impairs the ability to extract various aspects of configural information, including both first-order relational features, second-order relational features and holistic information (Maurer, Le Grand, & Mondloch, 2002). Studies of the inversion effect in nonhuman primates are limited but there is strong evidence for the inversion effect in chimpanzees. Parr and colleagues, for example, demonstrated significant inversion effects for unfamiliar chimpanzee and human faces, the categories that the chimpanzee subjects were familiar with, but not for discriminations involving the faces of an unfamiliar species of monkey, the capuchin, automobiles, or abstract shapes (Parr, Dove, & Hopkins, 1998). In these studies, subjects used a joystick-controlled cursor to select one of two inverted stimuli that matched the sample, i.e., the same stimulus presented in its upright orientation. Tomonaga (1999) also found the inversion effect in chimpanzees when discriminating human faces. Moreover, a recent study found that chimpanzee's performance matching unfamiliar faces compared to houses was significantly affected by their angle of orientation: as faces were rotated away from their veridical angle (0°) towards their inverted orientation (180°), performance declined linearly (Parr & Heintz, in press). This is interpreted as evidence against a dual-processing strategy for upright and inverted face, with upright faces being processed configurally and inverted faces being processed using feature-based cues. If the dual-process account were correct, performance would not show a linear decline with angle of rotation but would be expected to return to typical upright levels with inverted images once the processing shift has occurred. This suggests that upright and inverted faces are not processed using separate perceptual mechanisms, one configural and the other feature based, but rather Hominoids must preprocess inverted faces into their typically viewed upright orientation to make use of configural features (Collishaw & Hole, 2002).

Rhesus monkeys were also tested using the same computerized matching-to-sample tasks as the chimpanzees and they showed no evidence of familiarity-dependent inversion effects for faces. These monkeys showed significant impairments discriminating inverted compared to upright faces of unfamiliar humans, the capuchin monkey faces and automobiles, but not their own species’ faces (Parr, Winslow, & Hopkins, 1999). Published studies on the inversion effect in monkeys, however, are conflicting. Some have reported evidence of the inversion effect in macaques and chimpanzees (Tomonaga, 1994; Vermeire & Hamilton, 1998), while others have failed to find evidence of orientation-specific processing for faces in monkeys (Bruce, 1982; Parr et al., 1999) or chimpanzees (Tomonaga, Itakura, & Matsuzawa, 1993). Some authors have suggested that the inversion effect occurs only with low-spatial frequency homogeneous stimuli which bias processing towards the use of configurational cues regardless of expertise (Phelps & Roberts, 1994; Wright & Roberts, 1996). These authors suggest that human faces are more homogeneous than nonhuman primate faces making the inversion effect a phenomenon specific to human faces. There is little empirical evidence, however, to support such a view in nonhuman primates. Therefore, studies of the inversion effect have consistently shown that chimpanzees are sensitive to the configurational arrangement of facial features when processing familiar categories of faces like conspecifics and humans, while data from monkeys suggest that they may use a different perceptual strategy. More comparative research is needed before any similarities and differences in the face processing strategies between apes and monkeys can be fully understood.

A second widely recognized test of configural face processing, particularly holistic, second-order relational processing, is the face composite effect (Young, Hellawell, & Hay, 1987). In this task, the top half of one individual's face is either aligned or misaligned with the bottom half of another individual's face. In the seminal study, subjects were first familiarized to the individuals (most of the individuals were already famous for subjects) and then were presented with either the aligned or misaligned composite and asked to name either the top or bottom half individuals as quickly as possible. Because faces are processed using second-order relational features, our configural face processing system fuses the identity of the individuals in the aligned condition, making it very difficult to detect that the aligned composite is actually composed of two different individuals. In this case, subjects were impaired at naming the individuals from their face parts, regardless of whether they were asked to name the top or bottom half individual. When the face parts were misaligned, however, humans had little difficulty in recognizing that the face consisted of parts from two different individuals and were especially good at naming the individual represented by their top face half (Young et al., 1987; experiment 1). The composite effect was found for faces of both famous, highly recognizable individuals and unfamiliar individuals (Young et al., 1987; experiment 3). The reaction time advantages for misaligned composites were eliminated by inverting the composite faces, as configurational cues are not readily available in inverted faces (Collishaw & Hole, 2000; Young et al., 1987).

The human developmental literature has historically reported a strong relationship between age and face processing skills, with the critical period falling between 6 and 10 years of age (Diamond & Carey, 1977; Goldstein & Chance, 1964; Mondloch, Geldart, Maurer, & Le Grand, 2003). Young children, for example, were more distracted when required to match faces that contained paraphernalia, such as hats, glasses and other external features, compared to children over 10 years of age. This was interpreted as evidence of a shift from piecemeal, or feature based processing, to configural face processing that requires attention to inner facial details (Diamond & Carey, 1977). Some studies have continued to support a general shift in face processing skills in children over 10 years of age. When presented with the task of recognizing classmates from inner compared to outer facial features, children under 7 years of age showed the best performance when the whole face was presented but were significantly better on the external compared to internal feature manipulations. After 9 years of age (9–11 years age group), performance was better for discriminations involving inner compared to external features (Campbell, Walker, & Baron-Cohen, 1995). Sensitivity to second-order relational features was supported in a recent study by Mondloch, Le Grand, and Maurer (2002) who found children between 6 and 8 years of age to be more impaired when discriminating faces that were altered slightly with regard to the spacing between the eyes and the mouth, compared to children 10 years of age. In contrast, when the manipulations involved only the external contour of the faces, the youngest children (6 years) performed comparably to the adults (Mondloch et al., 2002). This latter study was based on an earlier set of manipulations by Freire and Lee (2001) that showed children between 4 and 7 years of age to perform better on faces that differed in terms of feature identity compared to the spacing between features.

Finally, the importance of early exposure to faces for the development of adult-like face processing skills has been demonstrated in an intriguing study of 12 individuals born with dense bilateral cataracts, determined to be clinically blind from birth (Le Grand, Mondloch, Maurer, & Brent, 2001). The cataracts were surgically repaired in this population after a period of 3–6 months. A minimum of 8 years after the corrective surgery, these individuals were tested on a variety of face recognition tasks including feature processing, the inversion effect, and the face composite effect. Results showed that these visually deprived individuals performed normally when discriminating faces that only differed in terms of a few facial features, i.e. altered eyes and mouths, but were significantly impaired on upright compared to inverted configuration trials in which the spacing between the eyes was altered, i.e. an alteration of second-order relational features (Le Grand et al., 2001). In a follow-up study, the face composite effect was virtually nonexistent in this deprived population. Subjects performed equally well on aligned and misaligned composites compared to control subjects, those not experiencing visual deprivation, who showed the composite effect by performing significantly better on misaligned than aligned trials (Le Grand, Mondloch, Maurer, & Brent, 2004). Collectively, these studies suggest that children differ from adults in their ability to use configural information in face recognition and that a period of exposure to faces during early development is critical.

More recent studies, however, have raised doubts as to the length of exposure that is necessary for the development of adult-like face perception, i.e. reliance on second-order relational features, or the extent to which infants and neonates show a reliance on the spatial configuration of facial features. Bonner and Burton (2004) reported recognition advantages for discriminating familiar faces from inner vs. outer features in children as young as 7 years that was comparable to that of adult subjects (Nachson & Shechory, 2002). Similarly, Flin (1985) reported an advantage for face recognition in upright compared to inverted faces in children as young as 7 years of age. Even studies of young infants suggest that the shift from feature to configural processing may not be as well-defined as the studies cited above would suggest. Neonates, for example, show spontaneous preference for looking at upright compared to inverted faces (Cassia, Turati, & Simion, 2004), suggesting that configural cues might be important earlier than previously thought. Similarly, Cohen and Cashon (2001) habituated 7-month-old infants to two faces in either their upright or inverted orientations and then presented them with one of the original habituated faces and a face that combined the inner and outer features of the two. Their results showed that the infants dishabituated, i.e. looked longer, to the novel face that combined features suggesting that they saw it as different configuration, since its parts had already been familiarized. Infants habituated to the inverted faces did not show this effect, suggesting that these infants were not familiarized to the configuration typically biased in upright faces. Thus, while the majority of the data suggest that children and adults differ in their face processing strategies, researchers acknowledge that data are lacking on infant's abilities during critical developmental periods (Want, Pascalis, Coleman, & Blades, 2002).

Data on the evolution of face recognition skills in species that are closely related to humans are extremely lacking. In this regard, nonhuman studies can be advanced considerably by attending to the debates posed in the developmental literature. To this end, this paper presents the results of three experiments that go beyond the inversion effect in examining the sensitivity of chimpanzees to configural face processing. These studies represent the first to systematically explore the importance of both first- and second-order relational features in the face processing of a nonhuman species. The first experiment tested chimpanzees using a modified version of the well-known face composite effect by presenting subjects with normal faces, aligned and misaligned composite faces, and upper and lower face parts of unfamiliar humans and chimpanzees. No studies to date have examined configural face processing in nonhuman primates using the composite effect. We hypothesized that if chimpanzees are sensitive to second-order relational features when discriminating faces, they should show clear evidence of the composite effect: greater impairment in detecting the presence of two individuals in the aligned vs. misaligned conditions. Additionally, we expected that subjects would show similar results for both conspecific and human faces, as they have considerable expertise with both species faces having been born in captivity. Thus, there would be no reason to presume that human faces and chimpanzee faces would be processed using different perceptual mechanisms in this population of subjects.

The second and third experiments further analyzed subjects’ configural processing biases for conspecific's faces by manipulating both first- and second-order relational properties in faces (Experiment 2), and by manipulating the spatial frequency of these faces (Experiment 3). Experiment 2 presented three types of facial manipulations. The inner features were extracted, preserving both the first- and second-order relational properties, the features were split apart, disrupting only second-order relational properties, and finally the features were split apart and rearranged, disrupting both first- and second-order relational properties. We hypothesized that if subjects are sensitive only to the first-order relational properties of faces, like inexperienced human children, they should perform well on the inner and split feature faces and show the greatest impairments on the split and rearranged feature faces. However, if subjects are sensitive to both first- and second-order relational properties, they should show the best performance on the inner feature trials, show some impairment for the split feature faces, but show the greatest impairments on the split and rearranged faces. Experiment 3 manipulated spatial frequency by pixelating faces using a small and large filter. Additionally, faces were presented with the eyes masked. We hypothesized that subjects would show greater difficulty processing faces pixilated over a large diameter compared to small pixel diameters (Harmon & Julesz, 1973). Additionally, we anticipated that subjects would be impaired in processing faces in which the eyes had been masked.

2. General methods

2.1. Subjects

Data were collected from 6 chimpanzees (Pan troglodytes), 4 males and 2 females in Experiment 1. Three of the males and one female were approximately 18 years of age and one male and one female were approximately 10 years of age. Experiments 2 and 3 used five subjects (the 10 year old female was not available for testing at that time). These ages represent reproductively mature chimpanzees. All subjects were socially housed in adjacent indoor/outdoor enclosures at the Yerkes National Primate Research Center, Atlanta, Georgia. The chimpanzees were raised by humans in peer groups at the Yerkes Primate Center nursery and then moved into permanent social groups with older animals at 4 years of age. This provided them with appropriate peer contact and social interaction early in their development. Later, they were housed with different combinations of adult animals and had considerable exposure to a range of neighbors with whom they shared auditory contact and some physical contact through mesh. All subjects had prior experience with a range of computerized tasks using matching-to-sample procedures, including face recognition tasks, and represent a subset of the Yerkes Primate Center's Chimpanzee Research Core (Parr, Hopkins, & de Waal, 1998; Parr, Winslow, Hopkins, & de Waal, 2000).

2.2. Procedure

Chimpanzees were tested voluntarily in their home cage. In brief, the computer was housed in an audio-visual cart (90.0 × 67.5 × 100.0 cm) enclosed in clear Plexiglas with external speakers located on a shelf approximately 10 cm from the ground. This was wheeled to the front of subjects’ home cage and positioned approximately 30 cm from the cage mesh. A joystick (approximately 5.0 × 2.5 × 5.0 cm) was then attached vertically to the front of their home cage so that the stick protruded approximately 4 cm into the mesh. The task was then initiated.

According to the matching-to-sample (MTS) procedure, subjects were first required to make an orienting response to a single photograph, hereafter referred to as the sample, by moving the cursor to contact this image presented on a black background. After subjects oriented to the sample image, two comparison stimuli appeared. These three images are collectively referred to as a stimulus set and represent one trial. In all experiments, the sample stimuli consisted of the manipulated face image (see specific descriptions below), while comparison stimuli consisted of unaltered face images. The comparison stimuli were presented in the corners of the monitor such that the images were across from one another and equidistant from the sample, i.e. bottom left and right, upper and lower left, upper left and right, or upper and lower right corners. All images were formatted so that their presentation size was 250 pixels high. The delay between the presentation of the sample image and the two comparisons was approximately 500 ms. Subjects were required to move the joystick-controlled cursor to contact the comparison image that matched the sample. If this occurred, subjects were given a small food reward (with the exception of the composite trials described in Experiment 1), such as a piece of fruit or squirt of juice. If the trial was incorrect, no food reward was given.

3. Experiment 1: The composite effect

3.1. Methods

3.1.1. Stimuli

Stimuli consisted of digitized photographs depicting chimpanzee and human faces. Chimpanzee photographs were taken from the chimpanzee colonies at the MD Anderson Cancer Center, Bastrop TX, none of whom were familiar to the subjects of this study. The human photos were taken from the NimStim Face Set.¹ Only the neutral photographs in this stimulus set were used. These individuals were also unfamiliar to the chimpanzees and photographs of these particular individuals had never been shown to the subjects in any previous task, so both the chimpanzee and human faces were novel in terms of the individual portrayed and the specific photograph. All stimuli were converted to 256 grayscale, placed on a black background, and cropped to a height of 250 pixels using Adobe Photoshop 7.0.

Photographs of 20 individual chimpanzees and humans were used. An equal number of males and females were represented for the human stimuli. Chimpanzee stimuli depicted 7 males and 13 females. All photographs showed frontal face views. Chimpanzee and human faces were subjected to five different manipulations for this study. Nineteen images were control trials (N = 9 chimpanzee and N = 10 human trials). These showed the faces with the standard cropping described above. Ten chimpanzee and 10 human images were additionally edited to form a group of composites. The fourth and fifth image categories showed only the top or bottom halves of the faces. The individual who prepared these stimulus manipulations (U.A.) was, at the time, naïve to the hypotheses of the study. Specifics about these manipulations and their correct choices in the MTS procedure are described below.

3.2. Control trials

Sample control trials showed the faces of 9 chimpanzees and 10 humans cropped to form an outline around the head. Each of these was combined with another individual from the same species category to represent the nonmatching comparison, or foil image. Thus, the correct pair of photographs showed identical images and the nonmatching stimulus was a different individual. Correct responses were to select the comparison photograph that matched the individual represented in the sample. These responses were rewarded with food, while incorrect responses were not food rewarded.

3.3. Composites

The composite stimuli were created by splitting the top and bottom halves of each face horizontally below the eyes. The 20 split human and chimpanzee faces were then recombined to form 10 new faces that contained the top half of one individual's face and the bottom half of another individual's face. Two types of composites were made using the same face part re-combinations. Ten trials were aligned where there the top and bottom face halves were matched as closely as possible, and 10 trials were misaligned so that approximately 1/4 of the width of the top face part was moved to the left above the bottom portion. Fig. 1 shows an example of an unaltered face and the two composite face manipulation for each species. In the matching task, the composites were presented as the sample stimuli while the two comparison stimuli showed the unaltered faces of each individual represented in the composite. Due to this arrangement, there were no incorrect responses on composite trials because subjects could choose the individual who matched the bottom face part, or the individual who matched the top face half in the composite. Because of this, composite trials were nondifferentially reinforced: subjects received a food reward for any response they made.

Fig. 1.

An example of the stimuli used in the face composite task, Experiment 1. The two faces on the far left are unaltered. The two faces on the right are the aligned and misaligned composites created by taking the bottom half of the far left individual's face and combining it with the top half of the other individual's face.

Since the same set of 40 faces (20 chimpanzees and 20 humans) were used in both the control trials and in the composite trial format, subjects might display a bias for certain individuals they had been previously reinforced for selecting, i.e. they were presented as samples in the control trials. For example, subjects get reinforced for selecting Jane in a control trial and may then remember Jane's face when she is presented in a composite trial (whether it's Jane's top or bottom face half). They would not then be responding to the configural nature of the composite, but simply recognize Jane. To control for this, half of the face parts used in the composite trials were taken from individuals who were presented as correct responses and half from those presented as incorrect comparisons in the control task. So, the composite trials presented face parts of individuals who were both correct and incorrect in the control trials. If subjects were simply selecting individuals who they had received reinforcement for in the control trials, their performance would only be 50%.

3.4. Upper and lower face

The upper and lower face trials showed the face part as the sample image, the correct response was a photograph of the individual whose face part was shown, while the nonmatching comparison was another individual. The individuals whose face parts were presented in these trials were balanced so that each individual had been the correct response on an equal number of previous trials. For example, if an upper face trial showed the top half of Individual A's face, the correct response would be Individual A while the incorrect choice would be Individual B. Then, in the lower face trial, Individual B's face part would be shown and the incorrect comparison would be Individual A. Thus, subjects saw each individual the same number of times as the correct and incorrect comparisons. These trials provided an important control for whether the top or bottom half of specific individual's faces were more or less salient than other individuals.

The order of trials was pseudo random so that each trial was seen once before any individual trial was repeated. Subjects were given trials in blocks of 99 per day, representing one exposure to each of the 99 stimulus sets. Six blocks were given, representing a total of 594 trials, or six repetitions of each unique stimulus manipulation. No food deprivation or restraint of any kind was used. Testing was voluntary and initiated by the subject. Sessions lasted approximately 20 min per subject.

3.5. Data analysis

No a priori percentage correct criterion was set for this task. Subjects were given six different exposures to the 99 stimulus sets and analyses then compared performance across the 10 stimulus categories. Data were analyzed using repeated measures ANOVAs where species (chimpanzee and human), stimulus type (either control, upper and lower face halves or aligned vs. misaligned composites,) and trial (N = 6) were the within-subject factors. Follow-up comparisons were performed where appropriate. Statistical significance was set at p < .05 for two-tailed tests and the analyses were adjusted for multiple comparisons using Bonferroni's correction procedure where alpha was divided by the number of comparisons to produce the adjusted alpha level, α′ = α (.05)/# comparisons.

4. Experiment 1

4.1. Results

An initial repeated measures ANOVA comparing stimulus type, trial, and species revealed no effects of trial order, therefore all subsequent analyses were performed on the mean response to each of the 10 stimulus categories. Fig. 2 shows the mean performance matching control faces, upper faces and lower faces. As there was no incorrect response for the composite trials, performance on these trials is graphed as the percentage of trials in which subjects matched the individual whose upper face was shown in the sample.

Fig. 2.

The performance matching chimpanzee and human faces, face parts and composites. The graph on the left is plotted as the percentage correct matching the control and upper and lower face parts. The right side shows the performance on the aligned and misaligned composite trials. These are plotted as the percentage matching the individual represented by the upper face part.

To examine the effect of species on matching faces and face parts, a second ANOVA was performed with species (humans and chimpanzee) and face trials (control, upper and lower) as the within-subject factors. This revealed significant main effects for stimulus type, F(2,10) = 5.80, p < .03, and species, F(1,5) = 7.86, p < .04. No significant interaction was observed between species and stimulus type in this analysis. The logical follow-up comparisons for these main effects are to compare human vs. chimpanzee faces regardless of feature manipulation, and to compare each feature manipulation regardless of species. Bonferroni's correction procedure for this last comparison was p < .017. Overall, there was no significant difference discriminating chimpanzee vs. human stimuli (p = .073), despite the main effect. Subjects were, however, significantly better in the control condition compared to the upper face condition, t(5) = 3.64, p < .015. An additional series of comparisons were performed to further identify differences observed in Fig. 2. Three t tests for each species compared the control faces with upper and lower face trials and the upper vs. lower face trials to each other. The corrected alpha value using the Bonferroni correction procedures was p < .017 (.05/3). Significant differences were found only for the chimpanzee faces. Subjects were significantly better matching chimpanzee control faces (76.85%) compared to the upper portion of the chimpanzee faces (65.83%), t(5) = 8.22, p < .001, and the lower portion (57.22%), t(5) = 3.57, p < .016. No significant differences were observed between the upper and lower face parts. Thus, subjects showed a significant advantage for discriminations involving whole chimpanzee faces compared to each face part, while no such advantage was observed for matching human faces.

A final ANOVA compared subjects’ performance on the nondifferentially reinforced composite trials. This analysis revealed a significant interaction between stimulus type and species, F(1,5) = 6.60, p < .05. Three sets of follow-up comparisons were performed for each species, control vs. aligned composite, control vs. misaligned composite, and each composite type. The corrected alpha value was the same as in the previous analysis. These analyses revealed significantly better performance on control chimpanzee faces (76.85%) compared to misaligned chimpanzee faces (56.94%), t(5) = 5.42, p < .003. Differences between the aligned vs. misaligned composites neared significance, t(5) = 2.60, p < .05. No significant difference was observed for control faces compared to the aligned composite faces. No differences were observed for the human faces for any of these comparisons.

5. Experiment 1

5.1. Discussion

Although the composite effect in this study was performed using a slightly different methodology than previous studies in humans, utilizing the chimpanzees’ expertise with matching-to-sample tasks compared to the same/different and naming methodologies used in human studies, the results suggest the presence of a composite effect that may be interpreted as comparable to that shown in humans. In humans, the composite effect suggests that the aligned face parts are perceived as a separate individual, a new gestalt, compared to the misaligned composite, for which subjects accurately identify the presence of two individuals. Chimpanzees also processed aligned composites much differently than they did the misaligned composites, and this effect was found only for chimpanzee faces, not for human faces. Chimpanzees spontaneously matched aligned composite faces most often by selecting the individual represented in the upper compared to lower face part. This suggests that information pertaining to the identity of the lower face individual was not detected as well in the aligned trials. When presented with misaligned composites, the chimpanzees were as likely to select the individual represented by either the top or bottom face parts indicating that they saw each distinct individual.

Replicating the human composite effect precisely, a fused, gestalt perception of the different individuals represented in the top and bottom face halves, was not possible given the limitations of adapting the composite effect for testing in chimpanzees. However, our data are suggestive that the presence of a different individual in the bottom half of the aligned composite decreased recognition of the individual represented in the top half. This was because subjects’ performance matching the aligned composites to the upper face individual was lower than control trial performance, indicating that the default pattern was not simply a bias towards matching all faces using the top half individual. This difference, however, did not reach significance and is therefore only suggestive of chimpanzees interpreting the aligned composite as a single individual, or fusing the face parts into a single gestalt. Similar attentional biases towards upper face features and “top-heavy” face patterns have recently been reported in human newborns and are interpreted as supporting a general perceptual process that becomes more specialized for meaningful face-like patterns through repeated exposure to those face categories (Cassia et al., 2004). This is because faces have naturally more energy in the upper visual field.

It is particularly important to stress that these patterns were not learned across repeated trials, nor were they due to differential reinforcement history. First, performance across the six testing sessions showed no significant change, thus no overall learning occurred. These tasks were, after all, not designed to train subjects to perform a certain way, but instead relied on the combination of different trial types to reveal inherent perceptual biases in face processing. Many more trials could have been given in this task but this simply would have been unnecessary as clear preferences emerged from the onset of testing. Second, chimpanzees were reinforced for whichever comparison face they selected, the top face or bottom face individual. Thus, subjects were not receiving differential reinforcement for attending to only top half or bottom half individuals, nor were they being differentially reinforced for selecting particular individuals. Finally, the preferences for the top half individual in the aligned composites was not due to an overall bias or preference for matching faces using the upper face part, i.e. the eye region. Subjects performed significantly better on the control trials compared to either the top half faces or bottom half faces. However, no differences were observed between the top and bottom face part trials. The chimpanzees were not any better matching unfamiliar chimpanzees from their upper face parts compared to their lower face parts. Therefore, the results presented here appear to reflect inherent perceptual mechanisms involved in face processing in this species.

What remains to be explained is the nature of expertise required to elicit configural face processing as it was hypothesized that chimpanzees would show the composite effect for both chimpanzee and human faces as they have had a lifetime of expertise with both species. The chimpanzees in this study were born in captivity, raised by humans in the Yerkes Primate Center nursery for 4 years, and then cared for by numerous human providers. They had daily exposure to caretakers, veterinary staff, researchers, maintenance staff and landscaping crews, among others. Although it is difficult to estimate the number of different individuals known to these subjects, their experience with humans is as close to natural expertise as is possible for a captive species. Yet, despite their prolonged exposure with humans, chimpanzees’ performance on all trials involving human faces showed no evidence of configural processing. Subjects performed as well on the unaltered human control trials as they did on any of the face manipulation trials. Several explanations for these findings may be possible. First, chimpanzees’ experience matching unfamiliar chimpanzee faces using computerized tasks far outweighs their experience matching human faces. Only twice previously had chimpanzees been presented with human faces for discrimination using computer tasks. Once was the inversion effect studies published by Parr et al. (1998) and the other experience came twice in unpublished studies of facial expression processing conducted in 1997 and again in 2004. Second, in the majority of the interactions that the chimpanzees have had with humans, the human wears personal protective equipment (PPE), i.e. face masks, face shields, and hair nets, as required by Yerkes and Federal guidelines for working with nonhuman primates. Before all of the subjects moved from the nursery to their permanent housing facility around 4 years of age, they also had periodic exposure to humans without their PPE. So although it is difficult to characterize exactly their experience seeing human faces, it seems reasonable to conclude that for the first 4 years of life, these subjects had ample opportunity to view humans without face masks. After this time, however, their exposure was primarily with humans wearing PPE. To this day, the chimpanzees occasionally have visual access to different individuals at a distance who do not wear this equipment. What is interesting is that the chimpanzees’ performance matching human faces from the lower face part was exceptionally low, the lowest of any stimulus category (51.4%). Additionally, their performance matching control human faces compared to upper face parts was almost identical, near 60% for both. Therefore, it may be that chimpanzee's expertise with human faces is limited to parts, particularly the upper face, and not the entire facial configuration. Learning second-order relational properties under these conditions would be doubtful. Human studies have shown that the reliance on second-order configural cues requires a period of exposure to the stimuli, and subjects can even learn to use these cues when nonface images become highly familiar (Gauthier, Skudlarski, Gore, & Anderson, 2000).

Because of the lack of significant results using human stimuli from this experiment, the next two experiments performed manipulations of first- and second-order relational features on chimpanzee faces only.

6. Experiment 2: First- vs. second-order feature manipulations

6.1. Methods

6.1.1. Stimuli

Stimuli consisted of digitized photographs of chimpanzee faces. These were taken mostly from colonies living outside of the United States and were supplemented from individuals living at the Yerkes National Primate Center field station, Lawrenceville, GA. All individuals were unfamiliar to subjects. All faces were presented as 256 grayscale images and cropped to a height of 250 pixels using Adobe Photoshop 7.0.

Photographs of 25 different individuals were presented with three different manipulations, totaling 75 unique trials. The inner features of the faces were extracted (inner), the facial features were split apart maintaining first-order relational properties but distorting second-order relational properties, and finally the location of the eyes and nose were reversed, so that the nose was relocated above the eyes. This latter manipulation disrupted both first- and second-order relational properties. Fig. 3 provides an example of each of these manipulations, along with the unaltered face from which these manipulations were derived (far left). The individual performing these manipulations was, at the time, naïve as to the hypotheses of the experiments (U.A.).

Fig. 3.

An example of the stimuli used in Experiment 2. The face on the far left shows the original face from which the three manipulations were created. The first manipulation shows the extracted inner features (inner), the second shows the features split apart (split) and the face on the far right shows the features split apart and rearranged (rearranged).

6.2. Data analysis

Subjects were tested until they received five presentations of each trial over consecutive days, or until their overall performance exceeded 80% correct. This was to ensure that subject's performance did not reach ceiling levels or that subjects were learning the correct comparison choices after repeated presentations. Performance was first assessed for whether subjects were performing above chance levels. To some extent, poor performance was expected for trials involving some of the manipulations. However, performance was assessed relative to control trials (inner trials in Experiment 2), on which performance was expected to exceed chance levels after minimal repetition. Data were then analyzed using repeated measures ANOVAs where stimulus type was the within-subject factor.

7. Experiment 2

7.1. Results

Only two of the five subjects met the 80% criterion and this was on the fifth testing session, so all subjects received the same number of trials, but achieved different levels of final performance. The other three subjects were given all five sessions (125 trials). The overall performance on this task, however, was not very high. For performance on a stimulus type to be above chance over a block of 25 trials, assessed using a binomial z score test [1.96 = x – 1/2T/√(T)*(.5)(.5)] was 69.7%, just over 17 correct responses out of 25. Performance across the five testing sessions can be seen in Table 1. Of the 25 possible sessions (5 sessions for 5 subjects), 14 showed performance significantly above chance for inner trials (56% of the sessions), and only 3 of the split feature and rearranged feature trials (12% of the sessions). Furthermore, on only 6 of the 75 total sessions (8%) did subjects perform better on a second-order feature manipulation compared to the inner trials.

Table 1.

Number of trials correct out of 25 for each stimulus manipulation in Experiment 2, inner features (inner), split features (split) and rearranged eyes and mouth (rearranged) for the 5 testing sessions

Subject/session	Feature manipulation
	Inner	Split	Rearranged
Jarred1	18*	12	14
Jarred2	18*	16	15
Jarred3	15	15	17
Jarred4	18*	16	15
Jarred5	21*	17	22*
Katrina1	15	16	10
Katrina2	18*	12	16
Katrina3	18*	14	17
Katrina4	20*	13	10
Katrina5	17	18*	18*
Lamar1	16	14	12
Lamar2	16	10	14
Lamar3	15	8	11
Lamar4	10	12	15
Lamar5	18*	15	12
Patrick1	14	13	13
Patrick2	16	13	11
Patrick3	21*	19*	15
Patrick4	16	15	16
Patrick5	20*	11	14
Scott1	9	15	15
Scott2	21*	10	17
Scott3	18*	16	16
Scott4	20*	12	16
Scott5	23*	18*	19*
Means (+SEM)	68.96 (2.31)	59.20 (3.08)	56.00 (2.32)

The means (+SEM) are presented at the end and the asterisk (*) indicates sessions on which subject's performance was significantly above chance levels.

A repeated measures ANOVA assessed whether subject's performance varied as a function of the type of stimulus manipulation. This showed a significant main effect of stimulus type, F(2,8) = 31.70, p < .001. Follow-up comparisons using paired t tests revealed that this was due to significant differences between performance on the inner feature extractions compared with both the split features, t(4) = 5.38, p < .006, and the split plus rearranged feature manipulations, t(4) = 15.89, p < .001. These were significant at the Bonferroni's adjusted alpha level, p < .017. There was no significant difference between the rearranged and split feature manipulations.

8. Experiment 2

8.1. Discussion

Results from Experiment 2 support the importance of second-order relational features in chimpanzee face processing. Subjects performed best on trials that presented the inner face features, a manipulation that preserved both first- and second-order relational properties. Significant impairments were found on all trials in which the faces were manipulated so as to disrupt second-order relational features, both the split feature trials and the split plus rearranged feature trials. Performance did not differ between the split feature and rearranged feature trials, suggesting that chimpanzees are unable to rely solely on first-order relational properties to discriminate among unfamiliar conspecific's faces.

Overall, the performance on inner feature trials was not extremely high, especially when compared to those trials in which faces were presented without any manipulation. Thus, the simplicity of the simultaneous matching task is not likely to have contributed to the lack of first-order feature manipulations affecting performance. Subject's performance on chimpanzee face control trials in Experiment 1, for example, was 77% compared to 69% on the inner feature trials in Experiment 2. Despite this, subjects performed consistently better on inner face trials than any of the other stimulus manipulations, and were above chance on the vast majority of sessions, suggesting at least a consistent advantage when faces contain both veridical first- and second-order relational properties. No differences were observed for manipulations that affected only second-order relational properties and those that affected both first- and second-order relational properties. It is unclear how a manipulation could affect only first- and not second-order features at the same time. Regardless, subjects in this experiment were equally as affected for manipulations that influenced both first- and second-order relational properties compared to only second-order manipulations. Therefore, the general conclusions is that the presence of second-order relational features is necessary for conspecific face recognition in chimpanzees.

Future studies will compare directly the role of both internal and external features in chimpanzee face processing. It would also be particularly interesting to examine these differences in species other than chimpanzees. Discriminations of human faces, for example, may show a greater reliance on the external face components, as it was shown in Experiment 1 that the information from some internal features, such as the mouth, appeared to provide little if any benefit to chimpanzees’ performance on human face processing trials. Additionally, faces that chimpanzees have no expertise with whatsoever, such as sheep faces, would provide another interesting level of comparison.

9. Experiment 3: Spatial frequency manipulations

9.1. Methods

9.1.1. Stimuli

Stimuli consisted of digitized photographs of chimpanzee faces. As stated in Experiment 2, these were taken from individuals who were unfamiliar to the subjects of the study. All faces were presented as 256 grayscale images and cropped to be 250 pixels wide using Adobe Photoshop 7.0.

Ten examples of four face categories were presented, totaling 40 unique trials. Ten of these were unaltered control trials in which the sample and correct comparison pair showed identical photographs of different individuals. In 10 trials, these same photographs were edited so that the eyes of the individuals were masked by covering them with a black rectangle. Ten photographs were subjected to two levels of pixilation, large and small pixels. These photographs represented a novel set of 10 individuals who were not included in the eye masked trials or control trials. The small pixel manipulation subjected the photographs to an averaging of 5 pixels in diameter while the large pixel manipulation subjected the photographs to an averaging of 10 pixels in diameter using Photoshop 7.0. This resulted in images that contained approximately 25 or 50 vertical pixels per face for the large and small pixel manipulations, respectively. Fig. 4 provides an example of each of these manipulations. The individual performing these manipulations using Photoshop (U.A.) was, as in the other experiments, naïve as to the specific experimental hypotheses.

Fig. 4.

An example of the stimuli used in Experiment 3. The image on the far left is the unaltered control face. The next three images show the masked eyes (eyes), the small pixel manipulation (smpix) and the large pixel manipulation (lgpix).

Stimuli were presented twice per session, totaling 80 trials. Subjects received five consecutive days of testing, 400 trials, or until their overall performance exceeded 80% correct.

9.2. Data analysis

Subjects were tested until they received five sessions over consecutive days (10 presentations of each trial) or until their overall performance exceeded 80% correct. This was to ensure that subject's performance did not reach ceiling levels or that subjects were not learning the correct comparison choices from repeated exposures to each trial. Performance was first assessed for whether subjects were performing above chance levels on each stimulus category. To some extent, poor performance was expected for trials involving difficult manipulations, such as the large pixelations. However, performance was assessed relative to control trials (unaltered faces in Experiment 3), on which performance was expected to exceed chance levels after minimal repetition. Data were then analyzed using repeated measures ANOVAs where stimulus type (mask, small pixels, large pixels) was the within-subjects factor. Follow-up comparisons were performed where appropriate and all alpha levels were adjusted using Bonferroni's correction procedure.

10. Experiment 3

10.1. Results

Four of five subjects met the 80% criterion. One subject did this on the first testing session, one on session 3, and two on session 5. One subject had still not reached the 80% criterion by the fifth session. For performance on a stimulus type to be above chance over a block of 20 trials, assessed using a binomial z score test [1.96 = x – 1/2T/√(T)*(.5)(.5)] was 72%, just over 14 correct responses out of 20. Mean performance on each trial type can be seen in Table 2. All subjects performed significantly above chance on the unaltered control faces on the first test session. Of the 18 individual sessions given to subjects, performance was significantly above chance on 17/18 (94%) control trial sessions, 14/18 (78%) masked eye sessions, 6/18 (33%) small pixel manipulation sessions and 7/18 (39%) large pixel manipulation sessions.

Table 2.

Number of trials correct out of 20 for each stimulus manipulation in Experiment 3, control faces, masked eyes (eye), large pixel manipulations (lgpix) and small pixel manipulations (smpix) across the 5 testing sessions (where necessary)

Subject	Feature manipulations
	Control	Eye	Lgpix	Smpix
Jarred1	16*	14	15*	9
Jarred2	20*	19*	17*	14
Katrina1	16*	12	14	16*
Katrina2	19*	14	14	9
Katrina3	17*	18*	14	10
Katrina4	13	17*	14	10
Katrina5	19*	16*	15*	12
Lamar1	16*	15*	11	13
Lamar2	15*	13	13	15*
Lamar3	16*	17*	14	11
Lamar4	19*	15*	15*	13
Lamar5	18*	18*	14	16*
Patrick1	16*	17*	19*	14
Scott1	16*	16*	17*	15*
Scott2	17*	15*	11	12
Scott3	19*	19*	18*	18*
Scott4	17*	15*	14	12
Scott5	17*	15*	14	16*
Means (+SEM)	84.14 (1.17)	80.00 (1.38)	66.26 (2.80)	77.40 (4.89)

The means (+SEM) are presented at the end and the asterisk (*) indicates sessions on which subject's performance was significantly above chance levels.

A repeated measures ANOVA assessed whether subject's performance varied as a function of the type of stimulus manipulation. This showed a significant main effect of stimulus type, F(3,12) = 7.89, p < .004. Follow-up comparisons using paired t tests revealed that this was due to a significant difference in performance between the control, or unaltered faces, compared with the large pixel manipulations, t(4) = 5.61, p < .005. No other post-hoc comparisons were warranted based on the a priori hypotheses. This was significant at the Bonferroni's adjusted alpha level, p < .017. No significant differences were observed for comparisons between control trials and small pixel or eye-masked manipulations.

11. Experiment 3

11.1. Discussion

Pixelating faces using a large radius filter introduces high-frequency noise where the boundaries of the pixels overlap. The resulting image is spatially coarse and humans find it difficult to extract relevant visual information to identify these images due to the low signal-to-noise ratio (Costen, Parker, & Craw, 1994; Harmon & Julesz, 1973). Therefore, this type of manipulation was hypothesized to affect both first- and second-order relational properties, but particularly second-order relational properties when the pixelations are so large as to distort the detection of any individual features. Results supported this hypothesis: subjects were significantly worse matching faces that were grossly pixilated compared to the small radius pixelated faces. While this study was not designed to determine the range of spatial information required for face processing in the chimpanzee, future studies will provide a range of filtering to determine the minimum spatial frequency necessary for face recognition in this species.

The second proposed hypothesis was that performance would be impaired on face trials in which the eyes had been masked. This was based on previous findings in this species that masking the eyes distorts individual recognition (Parr et al., 2000). This hypothesis, however, was not supported in the present experiment: subjects performed as well on the masked eyes trials as they did on the control trials. The original report, however, presented face images in a slightly different manner than the present study, and these differences might account for the improved performance of chimpanzees in the current study. The previous study required subjects to match the face of one individual presented with some combination of features masked from another photograph of the same individual in which no features were masked (Parr et al., 2000). Thus, subjects had to recognize similarities in the facial features of a single individual, presented in two different photographs, only one of which contained occluded features. The present study, however, presented two identical photographs of unfamiliar chimpanzees where the sample photograph had the eyes masked. Thus, subjects were able to use external features, such as hairline, mouth and chin shape, etc. to discriminate this individual from the comparison individual. This suggests, as was stated from Experiment 2, that external features are likely to have an important role in chimpanzee face processing.

12. General discussion

Results from these experiments both complement and extend previous studies on configural face processing in chimpanzees and provide important comparative data on the cognitive and perceptual mechanisms underlying Hominoid face processing. First, numerous studies have now confirmed that faces are highly salient stimuli for both humans and chimpanzees. Both human and chimpanzee infants show heightened interest and attention towards face-like stimuli (Bard, Platzman, Lester, & Suomi, 1992; Goren et al., 1975; Myowa-Yamakoshi, Tomonaga, Tanaka, & Matsuzawa, 2003; Valenza et al., 1996). Additionally, chimpanzees show a high-level of skill discriminating black and white faces of unfamiliar chimpanzees using computerized tasks (Parr & de Waal, 1999; Parr et al., 2000).

Second, these studies confirm the importance of configurational cues in chimpanzee face processing. Moreover, these tasks go beyond the inversion effect in demonstrating qualitatively similar processing strategies in chimpanzees and humans. Experiment 1 showed that these similarities are based on the sensitivity for detecting second-order relational features when processing categories of faces that are highly familiar, although caution should be used in interpreting these results because of the different methodology used for testing the composite effect in chimpanzees compared to humans. The data clearly showed that chimpanzees discriminated aligned and misaligned composite of unfamiliar chimpanzee faces differently. There was also a strong suggestion that the presence of a different individual in bottom half of the aligned composite altered discrimination performance by biasing subjects to attend only to the individual represented in the top face part. This was suggestive of, but not evidence for, a gestalt perception of the aligned composites. Therefore, similar to previous findings in humans, the chimpanzees appeared sensitive to the alignment of face parts and the presence of different individuals (Young et al., 1987). This effect, however, was not found for human faces despite the fact that chimpanzee have a lifetime of experience with humans. This was also somewhat inconsistent with previous findings in this species that showed impaired processing of inverted compared to upright human and chimpanzee faces, the categories for which subjects had expertise (Parr et al., 1998). However, as described in Experiment 1, chimpanzees’ expertise with human faces is quite restricted as humans are typically encountered with face masks and other protective equipment that occludes important facial information. Results from Experiment 1 showed that the chimpanzees were extremely poor at being able to extract relevant information from the lower part of human faces: they were just as accurate discriminating whole human faces as they were the upper face alone. There was no advantage in seeing the whole face, and performance matching individuals from their lower face part was at chance levels. Therefore, the composite effect appears to be a more sensitive task than inversion for demonstrating the necessity of expertise in recruiting second-order relational features in chimpanzee face processing.

Experiments 2 and 3 provide further support for the use of configurational strategies in chimpanzees face processing using manipulations that, in human developmental studies, have been shown to be particularly sensitive to second-order relational properties. These studies have demonstrated that sensitivity to second-order relational properties is not simply a maturational process but requires prolonged exposure to many different category exemplars.² In a series of studies, for example, Mondloch and colleagues demonstrated adult-like performance in young children when asked to detect whether two faces were different when the face manipulations were feature based, such as occluding specific facial features, but were much less sensitive to alterations that involved manipulating the spacing of individual features, i.e. second-order manipulations (Mondloch, Dobson, Parsons, & Maurer, 2004). The children were only able to detect the feature-spacing manipulations when the two faces were presented simultaneously compared to sequentially (<400 ms delay). Thus, while the neural systems for detecting changes in second-order relational properties appear to be present by 8 years of age, these children were much less efficient in using these skills in a real-world context compared to adults (Mondloch et al., 2004). It appears that these children do not yet have sufficient expertise with faces to enable rapid identification of many different individuals. Some studies claim that such expertise requires at least 10 years (Goldstein & Chance, 1964; Mondloch et al., 2003), although the literature is not in agreement on the specific development of adult-like face processing skills (Cohen & Cashon, 2001; Flin, 1985; Want et al., 2002). Therefore, face processing skills may not be achieved solely through general maturational processes, but rather emerge in conjunction with a proficiency for second-order configural processing and prolonged exposure to different individuals. Together, this allows for the rapid identification and representation of a myriad of different individuals, such as recognizing Bob from Ted. Chimpanzees in these studies performed consistently above chance discriminating faces from inner features, as has been shown in human children with minimal face expertise (Campbell et al., 1995), but showed significant impairments when the facial features were split apart and rearranged, distorting second-order relational properties.

Finally, results from Experiment 3 showed that masking the eyes in unfamiliar chimpanzee faces did little to impair performance when subjects were simply required to match identical photographs. This is counter to the results from a pervious study in which subjects were required to match unfamiliar individuals from two different photographs, when only one of these photographs showed the individual with features masked (Parr et al., 2000). This latter study, however, was dependent on individual recognition while the task in Experiment 3 could be solved by matching other aspects of the particular photograph. This experiment also confirmed the sensitivity of chimpanzees to second-order relational properties in that gross pixilation impaired face discriminations, similar to what has been shown in human studies (Costen et al., 1994; Harmon & Julesz, 1973).

Demonstrating the necessity of second-order configural properties in chimpanzee face recognition has important consequences for understanding the evolution of primate social cognition. First, chimpanzees are highly skilled at discriminating among the faces of numerous unfamiliar individuals (Parr & de Waal, 1999; Parr et al., 2000). They demonstrate the inversion effect, which is more robust for highly familiar faces. Thus, the inversion effect in this species does not generalize to faces per se, but is restricted to categories of faces for which subjects have expertise, as is the case for humans. Second, the social organization of the chimpanzee, while not unique to nonhuman primates, is rather specialized. Their social groupings are organized around what is referred to as a fission–fusion society. Individuals travel together in small groups that are part of a larger society that comes together at different times for different reasons, such as at a seasonal feeding site. Thus, daily groupings are not absolute as they are in many primate societies, rather individual group composition is highly fluid. Because of this, remembering social relationships and kinships among numerous individuals in this type of flexible social environment is mentally very challenging. The ability to rapidly and accurately recognize and remember many different individuals over varied time intervals and in different social combinations would be a very useful skill and one that, in humans, requires prolonged experience (Mondloch et al., 2004). As was described previously, these skills take time to develop, particularly when individuals are not always observed together in groups. Chimpanzees do not always have this luxury and the ability to extract second-order relational properties from faces may be an evolutionary specialization that chimpanzees share with humans to deal with their complex and highly fluid social organization.