Abstract
It is broadly agreed that the capacity of working memory (WM) is 4±1 items. In the vision literature these items are objects (i.e., bound collections of object features), whereas in higher cognition they are role bindings. The distinction between these accounts becomes clearer in the context of the spatial relations among an object's parts: If parts are items then the WM capacity required to store the spatial relations among an object's parts should scale with the number, n, of parts (i.e., load = n); but if part-relation bindings are items, then WM load should scale as r*n2, where r is the number of relations to be remembered. An intermediate account, according to which relational roles can be ""stacked"" on object parts, predicts that load should scale simply as n2. We ran an experiment investigating WM for spatial relations among object parts, orthogonally varying both the number of parts composing and object and the number of relations the subject was required to remember. The results clearly support the intermediate model, which accounts for 85% of the variance in Ss accuracy. The higher cognitive model accounts for only 71% and the ""parts as objects"" model accounts for 0%. Visual WM load thus appears to scale with n2 with no additional cost imposed for additional relations.
AFOSR Grant # FA9550-07-1-0147.