Object-Based Benefits without Object-Based Representations

Abstract

The organization of visual information into objects strongly influences visual memory: Displays with objects defined by two features (e.g. color, orientation) are easier to remember than displays with twice as many objects defined by one feature (Olson & Jiang, 2002). Existing theories suggest that this ‘object-benefit’ is based on object-based limitations in working memory: because a limited number of objects can be stored, packaging features together so that fewer objects have to be remembered improves memory performance. This view predicts that memory for "packaged features" should be correlated (if you remember one feature of an object you should remember the object’s other features). Counter to this prediction, we show that some object features are stored largely independently. Participants were instructed to remember the colors and orientations of 5 colorful isosceles triangles (five-object condition) or the color of 5 colorful circles and the orientation of 5 black isosceles triangles (ten-object condition). After encoding (1200ms) and retention (900ms), memory was assessed with a continuous report for both features. Critically, participants reported both features of the same item in the five-object condition, allowing us to determine whether features were stored in an integrated fashion. Here we replicate the object-benefit: participants remembered twice as many features when arranged in 5 versus 10 objects. However, in the five-object condition memory for the color and orientation of an object was largely independent—when participants failed to accurately report the color or orientation of an object they were often quite accurate at judging the object's other feature. These results challenge the claim that the object-benefit is driven by the storage of integrated object representations, and require a revision of the concept of object-based memory representations. We propose that working memory is object-based in regard to the factors that enhance performance, but is feature-based at the level of representational failure.