Abstract
Groups, gestalts, and configurations—these vaunted constructs in perceptual psychology typically connote representational contents to be distinguished from individuals, objects, and positions. We investigated a well-known configuration effect in spatial working memory to determine whether it could be derived from rational inferences applied to representations of individual objects. Specifically, we employed a spatial change detection task as in Jiang, Olson, and Chun (2000): participants remembered the positions of several identical objects and, at test, reported whether one item (' the probe', indicated with a bold outline) had changed position. Replicating earlier findings, performance was extremely poor when non-probe items occupied new, random positions in the test display, in comparison to conditions in which non-probe items maintained their previous positions or only the probe item appeared at test. These effects have been interpreted as evidence for configurations represented in SWM. In contrast, we hypothesized that they arise from an algorithm that identifies mutually exclusive correspondences among the memory and test items. Establishing correspondence is logically necessary for making comparisons between previous and current positions. Our algorithm identified correspondences via two assumptions known to operate in apparent motion: objects are more likely to make small as opposed to large position changes; and objects tend to sustain similar changes to one another. A rational model embodying these assumptions performed each of the experimental trials in multiple simulations under perceptual noise. When performance was averaged on a trial-by-trial basis, the model evidenced a strong and significant correlation with human performance (r2 = 0.90, 0.84, 0.69 for single item, preserved context, and new context conditions respectively), even reproducing severe biases observed in human behavior. These results highlight the inescapable role of correspondence algorithms in SWM, and one way that rational algorithms combined with simple inputs can produce complex dynamics.
Meeting abstract presented at VSS 2014