Increasing variance in emotional expression in a crowd of faces reduces sensitivity to the average face

We often encounter groups of objects that are featurally similar and difficult to distinguish. For example a tree contains thousands of leaves that produce an average ‘leafiness’ texture. Instead of coding every leaf individually, it is more efficient for the visual system to quickly extract the mean of the group, in essence driving texture perception. We use this process, known as ensemble coding, to perceive the mean size of a group of circles (Ariely, 2001; Chong & Treisman, 2003), average orientation (Parkes, et al., 2001), the average speed (Watamaniuk & Duchon, 1992) and location (Alvarez & Oliva, 2008) of a set of moving dots, and even the mean emotion of a group of faces (Haberman & Whitney, 2007; 2008). A question remains, however, as to how increased variance within the set affects our ability to extract summary statistics; that is, what happens when some of the leaves' colors start to turn? We explored this by manipulating the variance in sets of faces. In each trial, observers saw a set of 16 faces that varied in emotional expression, nominally separated from one another by emotional units. We manipulated set variance by increasing or decreasing the emotional units separating set members. Using method-of-adjustment, observers adjusted a subsequent test face to the mean emotion of the set of 16. We measured how far, on average, observers' responses were from the mean; the smaller the average difference, the more precise the set mean representation. As set variance increased, mean representation precision decreased. The process of extracting a mean representation breaks down at high levels of set variance because the mean no longer functions as a reliable summary representation.