Viewers extract mean and individual identity from sets of famous faces

Highlights

• Ensemble coding occurs across sets of different famous faces.

• Participants retain good memory for the set exemplars.

• Set average and set exemplar representations are simultaneously encoded.

Abstract

When viewers are shown sets of similar objects (for example circles), they may extract summary information (e.g., average size) while retaining almost no information about the individual items. A similar observation can be made when using sets of unfamiliar faces: Viewers tend to merge identity or expression information from the set exemplars into a single abstract representation, the set average. Here, across four experiments, sets of well-known, famous faces were presented. In response to a subsequent probe, viewers recognized the individual faces very accurately. However, they also reported having seen a merged ‘average’ of these faces. These findings suggest abstraction of set characteristics even in circumstances which favor individuation of the items. Moreover, the present data suggest that, although seemingly incompatible, exemplar and average representations co-exist for sets consisting of famous faces. This result suggests that representations are simultaneously formed at multiple levels of abstraction.

Introduction

“Set representations” have recently attracted increasing research interest. When seeing groups of perceptually similar objects, information such as size, or motion, may be coded via summary statistics in terms of a mean value across exemplars (Albrecht and Scholl, 2010, Chong and Treisman, 2003). Whenever observers can capitalize on redundancy of information – a common observation in structured sets – they can compress this information into a single representation such as the set average (Alvarez, 2011). In a seminal investigation, Ariely (2001) investigated size representations from sets containing differently sized circles. Critically, participants tended to identify a test circle as having been presented when it had a similar size to the mean of the whole set, even when such an item had not been present. Moreover, participants were near chance when they had to choose which of two circles had been presented. Taken together, these findings suggest that (i) mean size information was computed and retained for the set and (ii) size information of individual set members was unavailable. There are different potential explanations for weak exemplar representations. First, encoding of precise exemplar representations may not routinely occur, or may simply contain too much noise, perhaps due to the lack of focal attention to set exemplars. Alternatively, an individual representation may initially be computed but may then be discarded extremely fast.

Recently, statistical representations have been demonstrated for sets of perceptually complex stimuli, such as faces. When asked to compare the emotional intensity of a single image with the mean of a set (up to 16 face photographs varying in emotional intensity), participants performed highly accurately (Haberman and Whitney, 2007, Haberman and Whitney, 2009). Performance was actually comparable to a control “exemplar” condition, in which participants compared an image with a homogeneous set with constant emotional intensity. Beyond extraction of mean emotion (and gender, see Haberman & Whitney, 2007) information from sets of faces, a similar mechanism may compute the mean identity from sets of unfamiliar faces. In one recent study (de Fockert & Wolfenstein, 2009), participants initially saw sets containing photographs of 4 unfamiliar faces from different individuals. In a “match” condition, a subsequent single image could either be an exemplar image from the previous set, or an average morph created from the four set images. Strikingly, the set averages (which had never been seen) received more “present” responses than the (seen) exemplars. The authors concluded that averaging identity information might serve as the “default mode” for generating mental representations from groups of faces.

Given that facial representations should serve person recognition, this is a surprising finding, since mean identity representations should actually prevent identification of a specific person in a group. It is relatively straightforward to understand how superficial averaging of abstract shapes might take place in the visual system, but much harder to account for averaging over such high-level characteristics as someone’s identity. For this reason, it is important to note that the authors used unfamiliar faces. Crucially, unfamiliar face recognition is strongly image-dependent and sensitive to superficial picture similarity (Bruce et al., 1999), and is thus based on very different mechanisms than familiar face recognition. For example, viewers are very good at matching different images of a familiar person, but very poor at matching unfamiliar faces (Bruce et al., 2001, Burton et al., 1999, Clutterbuck and Johnston, 2004, Kemp et al., 1997). This discrepancy suggests a qualitative difference in perception of familiar and unfamiliar face identities (Hancock, Bruce, & Burton, 2000), which may also have consequences for the interpretation of the identity set averaging data. Accordingly, increased percentages of “present” responses to matching averages in the study of de Fockert and Wolfenstein (2009) could reflect image averaging across similar pictures, rather than identity averaging. If viewers are failing to differentiate between the unfamiliar people shown to them, they might plausibly construct a set average combining these images. So, while this study certainly demonstrates set averaging for a class of high-level stimuli (faces), we argue that evidence for identity set averaging would be much more compelling if it could also be shown to exist for familiar faces sets.

Another important characteristic of previous studies examining set averaging for faces was relatively small image variability within sets. For instance, set averaging for facial expressions was generally investigated by assembling sets from a single identity, using slightly different emotional intensities from a morph continuum between two veridical expressions (Haberman et al., 2009, Haberman and Whitney, 2007, Haberman and Whitney, 2010). One study on set identity averaging actually involved 4 true set photographs, but had sets deliberately arranged to comprise similar identities (de Fockert & Wolfenstein, 2009). Therefore, low recognition rates for set exemplars may have originated from participants being unable to differentiate between exemplars at encoding. It is important to see if the use of more naturally diverse sets could increase exemplar memory, and whether this would in turn affect the quality and strength of set representations.

In sum, previous studies have investigated set averaging using face sets that varied little on either identity or image properties. In the present study, we tested facial identity averaging by using diverse pictures from highly familiar identities, for which participants have rich pre-existing mental representations. We further encouraged identity processing for half of the participants by instructing them to indicate whether a specific person had been seen in a set of faces, while the other half indicated whether a specific image had occurred. We expected that set averaging would be strongly reduced or absent for highly familiar faces, and that performance would reflect accurate representation of exemplars instead; since viewers know these identities, and faces in the set were quite diverse, there appears to be no advantage in averaging across them.