Structured priors in visual working memory revealed through iterated learning.

What hierarchical structures do people use to encode visual displays? We examined visual working memory’s priors for locations by asking participants to recall the locations of objects in an iterated learning task. We designed a nonparametric clustering algorithm that infers the clustering structure of objects and encodes individual items within this structure. Over many iterations, participants recalled objects with more similar displacement errors, especially for objects our clustering algorithm grouped together, suggesting that subjects grouped objects in memory. Additionally, participants increasingly remembered objects as lines with similar orientations and lengths, consistent with the Gestalt grouping principles of continuity and similarity. Furthermore, the increasing tendency of participants to remember objects as components of hierarchically organized lines rather than individual objects or clusters suggests that these priors aid the perception of higher-level structures from ensemble statistics.