Not all summary statistics are made equal: Evidence from extracting summaries across time

Over the past 15 years, a number of behavioral studies have shown that the human visual system can extract the average value of a set of items along a variety of feature dimensions, often with great facility and accuracy. These efficient representations of sets of items are commonly referred to as summary representations, but very little is known about whether their computation constitutes a single unitary process or if it involves different mechanisms in different domains. Here, we asked participants to report the average value of a set of items presented serially over time in four different feature dimensions. We then measured the contribution of different parts of the information stream to the reported summaries. We found that this temporal weighting profile differs greatly across domains. Specifically, summaries of mean object location (Experiment 1) were influenced approximately 2.5 times more by earlier items than by later items. Summaries of mean object size (Experiment 1), mean facial expression (Experiment 2), and mean motion direction (Experiment 3), however, were more influenced by later items. These primacy and recency effects show that summary representations computed across time do not incorporate all items equally. Furthermore, our results support the hypothesis that summary representations operate differently in different feature domains, and may be subserved by distinct mechanisms.