Abstract
Feature information can be distorted in visual working memory leading to characteristic biases in memory recall. Both repulsive and attractive biases have been reported, but a normative framework that reconciles these two distinct effects has been missing. Here, we introduce a framework that employs grouping and segregation as two fundamental processes necessary to create an efficient and robust memory representation of visual information. In its simplest form applied to the memorization of individual items, the framework creates memory representations that include not only sensory information but also discrete structural information about both the items' similarity (grouping) and their distinct identities (segregation). Memory retrieval consists of an active inference process that infers the original feature values from both the memorized sensory samples and structural information.
We conducted a psychophysical experiment to test the basic predictions of our framework. Subjects (N = 8) were instructed to recall the angular positions of two briefly displayed dots. The angular distances between the dots were parametrically varied in a range from 6 to 60 degrees. We found that subjects' recalls were systematically biased depending on the distance between the dots. Recalled dot distances showed a transition from being over-estimated (repulsion) for small and under-estimated (attraction) for larger actual dot distances. This bias pattern is consistent with previous results measured for color (Colomb, 2015) or orientation features (Bae and Luck, 2017). Our framework predicts this general bias pattern and also accounts for the variance and correlation structure observed in our experimental data. The proposed framework can be expanded to model hierarchical memory representations where grouping and segregation act at different scales, and thus generalizes to richer stimulus patterns.