br Author contributions br Financial disclosures This resear
Financial disclosures This research was supported by a grant from the Fuller Theological Seminary/Thrive Center in concert with the John Templeton Foundation awarded to J.H.D. and J.S.M.. H.S.S. was supported by a National Science Foundation Graduate Research Fellowship (NSF Award No DGE-0802267) during manuscript preparation. J.S.M. was funded by National Institutes of Health K12 grant (HD065879). Any opinions, findings, conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of these funding agencies. The authors declare there are no competing financial interests.
Introduction This ability to assess automatically the summary or ‘gist’ of large amounts of information presented in visual scenes, often referred to as ensemble perception or ensemble encoding, is crucial for navigating an inherently complex world (Chong and Treisman, 2003, 2005; Haberman and Whitney, 2009; Sweeny et al., 2013). Given the processing limitations of the brain, it is often efficient to sacrifice representations of individual elements in the interest of concise, summary representations, which become available as the mi2 rapidly encodes statistical regularities in notions of a ‘mean’ or a ‘texture’ (Haberman and Whitney, 2012; Whitney et al., 2013). Ensemble perception has been demonstrated consistently for low-level visual attributes, including size, orientation, motion, speed, position and texture (Ariely, 2001; Chong and Treisman, 2003; Parkes et al., 2001). More recently, studies have also demonstrated ensemble perception in high-level vision. In Haberman and Whitney (2007)’s initial work on ensemble perception − and on which the current study was based, three adult observers viewed sets of morphs (computer-generated continuous variations of expressions of the same face) ranging from sad to happy. Observers were then asked to indicate whether a subsequent test face was happier or sadder than the average expression of the set, a task that required creating an internal representation of an average of facial expressions in the first set. The precision with which the three observers completed this task was remarkably good. In fact, two of the three observers were as precise in discriminating ensemble emotions as Transformation were in identifying the emotions of single faces (in a control task). In another task, the same observers viewed sets of emotional morphs and were subsequently asked to indicate which of two new morphs was a member of the preceding set. All three observers were unable to perform above chance in this condition, suggesting that observers were unable to encode information about individual face emotions, despite being able to encode seemingly effortlessly information about average emotions. Subsequent work has shown these effects for a range of facial attributes (gender, ethnicity, identity, emotion, attractiveness; Haberman et al., 2009; Haberman and Whitney, 2007, 2009, 2010, 2011; Neumann et al., 2013). Sweeny et al. (2014) have also shown that ensemble perception of size is also present, though not yet fully developed early in development, in 4–6 year-old children. In the primary condition of their child-friendly task, participants saw two trees, each containing eight differently sized oranges, and were asked to determine which tree had the largest oranges overall. A secondary condition (see Sweeny et al., 2014) included experimental manipulations that allowed for the empirical simulation of performance in the primary condition with no ensemble coding strategies available−that is, as if participants gave their response after comparing the sizes of a single, randomly-chosen orange from each tree. The difference in accuracy between the primary and secondary conditions provided an estimate of the extent to which participants benefited from the use of ensemble perception strategies, the ‘ensemble coding advantage’ (Sweeny et al., 2014). They found significant ensemble coding advantages in both young children and adults, although children presented smaller such advantages than adults. An ideal observer model, which was also used to predict the minimum number of items integrated in the primary condition, suggested that both children and adults did not necessarily derive ensemble codes from the entire set of items (N=16), while children integrated fewer items than adults (4.24 vs. 7.18 items, correspondingly, across both trees), consistent with the smaller ensemble coding advantage they exhibited.