Structured representations in visual working memory: Using results from individual displays to constrain cognitive theory

摘要

Influential models of visual working memory assume we encode items independently and examine how many individual items can be remembered. However, observers remember not only individual items but also ensemble statistics, like mean size or color.Here we show that these levels of representation are not independent, even when observers remember displays of random colors. Exp. 1:Observers remembered several colored dots, and after a delay, reported the color of a dot selected at random. We estimated the rate of random guessing, and whether it depends on how items are clustered in color space. At set size 3, even subtle differences in clustering resulted in different guess rates: with distractors 90 and 180?° in color space from the tested item, the guess rate was 24%; at 60/120?°, the guess rate was only 13% (p0.05). In fact, with distractors somewhat clustered, observers rarely guess even at set size 5 (distractors: 30/60/90/120?°; guesses: 7%). These clustering effects suggest that observers make use of the ensemble statistics to know something about all of the items. Importantly, clustering in color space is inherently more likely at low set sizes, and thus guess rates at low set sizes are systematically underestimated in all experiments using these methods (up to 30%). Exp. 2:In a second experiment, we tested hundreds of observers on the same displays (N=465 on each), and found that models that do not take into account the relationship between items fail to fit the data on any particular display, even though they fit on average across all displays. Models that take into account how clustered in hue the items are (e.g., ensembles) are required. We suggest that observersa?? use of ensembles, which vary from display-to-display, make studying capacity by averaging across displays impossible, and requires a new framework based on large amounts of data on individual displays.