A critical review of the development of face recognition: Experience is less important than previously believed

Abstract

Historically, it has been argued that face individuation develops very slowly, not reaching adult levels until adolescence, with experience being the driving force behind this protracted improvement. Here, we challenge this view based on extensive review of behavioural and neural findings. Results demonstrate qualitative presence of all key phenomena related to face individuation (encoding of novel faces, holistic processing effects, face-space effects, face-selective responses in neuroimaging) at the earliest ages tested, typically 3–5 years of age and in many cases even infancy. Results further argue for quantitative maturity by early childhood, based on an increasing number of behavioural studies that have avoided the common methodological problem of restriction of range, as well as event-related potential (ERP), but not functional magnetic resonance imaging (fMRI) studies. We raise a new possibility that could account for the discrepant fMRI findings—namely, the use of adult-sized head coils on child-sized heads. We review genetic and innate contributions to face individuation (twin studies, neonates, visually deprived monkeys, critical periods, perceptual narrowing). We conclude that the role of experience in the development of the mechanisms of face identification has been overestimated. The emerging picture is that the mechanisms supporting face individuation are mature early, consistent with the social needs of children for reliable person identification in everyday life, and are also driven to an important extent by our evolutionary history.

Keywords:

• Face recognition
• Development
• Age of maturity
• Behaviour
• Functional magnetic resonance imaging
• Event-related potential

Acknowledgments

Preparation of this article was supported by Australian Research Council Grants DP0984558 to E.M. and DP0770923 to L.J., plus salary funding for K.C. from Hong Kong Research Grants Council (HKU7440/08H, William Hayward), salary funding for L.J. from ARC Centre of Excellence Grant (CE110001021), and salary funding for D.D. from Ellison Medical Foundation (Nancy Kanwisher). We thank Laura Germine for her unpublished data referred to in Footnote 9.

Notes

¹ Many studies exclude hair because it is a simple cue that can be used to recognize photographs without requiring face processing (e.g., prosopagnosics use hair cues in experiments even when they cannot recognize people in real life; Duchaine & Nakayama, 2006).

² Where a study tested separate age groups and found a clear effect present in each, we list all ages separately (e.g., “3, 6, 9” m.o.); where a study showed an effect in a combined age group but did not split by exact age, we indicate the age range of the group (e.g., “3–6” y.o.). Note, m.o. = month old; y.o. = year old.

³ This result holds when the opposite and nonopposite adaptors are equally dissimilar from the target and when the test trajectory for nonopposite adaptors is between nonopposite-and-target rather than opposite-and-target (8–9 y.o., Jeffery et al., 2011).

⁴ Although note that contingency in 5-year-olds was driven by adaptation only to one of the races.

⁵ Thus, even showing that conditions are significantly greater than chance in the youngest group does not ensure that restriction of range has been avoided.

⁶ Matched accuracy requires that two conditions have equivalent means and that these means are not approaching ceiling or floor. It is not sufficient to show that two conditions have near-perfect accuracy in adults (e.g., 97% versus 99%): The “matching” (unless also present on reaction times) could reflect simply a ceiling effect.

⁷ We review studies where the object class is theoretically similar to faces (animate objects with exemplars sharing a first-order configuration). Interestingly, however, face memory might develop faster than memory for houses or scenes (which activate different brain areas from discrete objects, i.e., the parahippocampal place area). In two studies, accuracy was similar for faces and houses/scenes in the youngest age group (and above floor), making comparison of developmental trends valid. Both these studies found stronger development for faces than houses (across 6 to 10 years, Carey & Diamond, 1977) and scenes (across 7–11 to 12–16 years, Golarai et al., 2007). Our conclusion that there is no development of face-specific mechanisms is based on the presumption that discrete objects form a better control class for faces than do houses/scenes. The results reviewed do not rule out the possibility that mechanisms common to both faces and objects might develop at a different rate than those used to code houses and scenes.

⁸ Children do show a larger after-effect than adults for an asymmetric manipulation of eye height that breaks the first-order face configuration (Hills et al., 2010). This could suggest greater flexibility in the range of faces that can be coded (see “Effects of age and early experience on the flexibility of tuning to different face subtypes”), but it could arise from mid-level-vision responses to the vernier offset of the eyes: vernier offset is less precisely coded in children than in adults (Skoczenski & Norcia, 2002) and could plausibly produce after-effects.

⁹ In conflict, a recent adult study appears to suggest very late quantitative maturity: Germine et al. (2011) reported face memory peaks at 32 y.o. Because general cognitive function should not change over 20–32 years, this appears to imply that the amount of lifetime experience with faces is a key driver of face recognition ability. However, the “age” effect could reflect a confound with the ORE. Data were internet-collected via the http://www.testmybrain.org site. Face stimuli were Caucasian. Participants' race or country of origin were not recorded, and they could have participated from anywhere in the world. More recent data from testmybrain (2011 sample, four years later than the original; L. T. Germine, personal communication, August 26, 2011) reported the country in which participants grew up. Grouping participants into majority-Caucasian countries (US, Europe, etc.) or other countries (Asia, Africa, etc.), results showed noticeable variation across age in proportion from majority-Caucasian countries. In some age ranges, this could not explain face recognition performance (e.g., recognition declined across 34–65 years despite proportion-Caucasian increasing). However, across the crucial 20–32-year-old range, proportion-Caucasian increased significantly with age (from approximately .71 at 20 years to .81 at 32 years, r = .61, p = .027). If this confound was present in the original sample, then the improvement in face recognition between 20 and 32 could have reflected reduction in the ORE (noting that this would also predict the observed changes for upright and not inverted) rather than an age/experience effect.