Abstract
Visual short-term memory (VSTM) is a central component of many human activities, but remains a poorly understood process. While previous theories have posited mechanisms intended to account for observed phenomena, in the present research we develop an ideal observer framework to uncover the expected behavior of an optimally performing memory system. We use information-theoretic constructs to characterize the optimal memory precision for visual features of a capacity-limited VSTM in a range of conditions. An ideal observer uses a flexible coding strategy determined by the task context (e.g. the distribution of feature attributes in a stimulus set), the precision of sensory signals and the number of items to be remembered. A fixed capacity limit makes specific predictions about how these factors trade-off in determining memory task performance. To evaluate the predictions of our analysis, we conducted two experiments to measure the precision with which subjects remember stimulus features (visual location or orientation) in arrays of varying set size and in conditions with different variances of the feature properties to be remembered. Subjects' precision decreased with increasing set size. For each set size, it also decreased with increasing feature variance. Both of these results are predicted by an ideal observer that flexibly allocates a continuous memory resource. Subjects' guess rates, however also increased with increasing set size, even at the smallest set sizes, indicating that subjects did not evenly distribute memory resources across all items in a display. We demonstrate that both findings fit naturally within our ideal observer framework, with the additional assumption that although VSTM has a fixed capacity, this capacity can be flexibly allocated among items in the scene. This generalized framework spans an entire range of models, with the continuous resource (uniform allocation among items) and discrete-slot (focused allocation on a subset of items) models at opposing extremes.
This research was supported by grants NIH R01-EY13319 to David Knill and NSF DRL-0817250 to Robert Jacobs.