Research has consistently shown that observers are able to extract a summary of visual input. For instance, observers are able to extract an average of low-level visual features, such as the mean size and colour of a set of objects. More recently, research has investigated whether the same could occur during face perception, such as the processing of facial identity. The majority of evidence suggests that facial averages are not recognised more frequently than the individual face members of a learning set, raising doubt as to whether an internal average of these identities is actually encoded. Instead, it is possible that facial averages simply resemble the encoded set members sufficiently to be recognised as a familiar face. To investigate this possibility, the experiments in this thesis explored the ability of observers to detect the similarity of a facial average to its constituent identities. This was done by comparing the recognition of an average against its constituent exemplars under optimised conditions where set size was reduced to a logical minimum, and by using both same and different images at encoding and recognition. The exemplars and averages were also put into direct competition by asking participants to decide which face most resembled the learning set in a two-alternative forced-choice (Chapter 2). These experiments demonstrate that the average and exemplar are consistently identified at a comparable rate. The visual similarity of these stimuli was then assessed further by investigating whether participants are able to identify which faces were used to create an average (Chapter 3). Observers found this task challenging, and so a more consistent measure of similarity - facial recognition algorithms - were used to determine how much similarity an average has to its constituent identities (Chapter 4). This showed that an average created from four different faces contains minimal similarity to its constituent identities, and below a level that human observers can easily detect. These results are discussed in the context of ensemble coding and point to a cognitive mechanism that supports the formation of an internal face average.
Laura J. Hunt (Mon,) studied this question.