Abstract
Visual working memory stores object features (e.g., locations) according to their statistical structure (Alvarez & Oliva, 2009). When recalling objects, people often use that structure information to compensate for uncertainty about the individual objects (Brady & Alvarez, 2011). Although any stimulus has its own ensemble statistics, people also have expectations from the real world about how objects are organized. Here we try to characterize Gestalt priors about the spatial arrangement of objects in an iterative visual working memory paradigm. We examined visual working memory’s priors for locations by asking participants to recall the locations of objects, and then having someone else remember and reproduce those recalled locations. A long sequence of individuals remembering the positions recalled by previous participants yields a Markov chain that will overemphasize the priors that people use to encode object locations (Sanborn & Griffiths, 2008). Across iterations, subjects recalled objects more densely packed (t(9)=8.03, p< .001) and with more similar translational errors (t(9)=9.11, p< .001), suggesting that subjects grouped objects in memory. To determine how subjects grouped objects, we designed a non-parametric clustering algorithm that infers whether objects are parts of clusters or straight lines. The clustering model revealed that subjects increasingly grouped objects as lines, going from using line groupings 2% to 22% of the time. Furthermore, consecutive subjects were more likely to group objects the same way when arranged in lines (t(9)=8.08, p< .001) or very eccentric clusters (t(9)=3.14, Bonferroni corrected p=.024). This suggests that linear arrangements are particularly stable in memory. Our results are consistent with evidence that people use priors from the real world to efficiently encode information in visual working memory (Orhan & Jacobs, 2014). Additionally, the increasing likelihood of people remembering objects as components of lines rather than clusters suggests that these priors aid the perception of higher-level constructs from ensemble statistics.
Meeting abstract presented at VSS 2015