A Multinomial Model-Based Analysis of Bindings in Working Memory

We examined how elements are integrated into larger units in working memory (WM). Four contrasting conceptual models exist with regard to this question: (1) a Unitization model, in which there is a single integrated representation which is retrieved in an all-or-none fashion; (2) a Unitization-With-Element-Failure model, in which a single integrated representation is retrieved as a whole, but access to its elements can still fail individually; (3) a Pairwise-Binding model, in which elements of a unit are represented separately and are bound together in pairs; (4) a Hybrid model that includes an integrated representation as well as pairwise bindings between element representations. We developed four multinomial process tree (MPT) models to test these theories. In three experiments, participants memorized multiple units which were random combinations of three elements. They were given one element as a cue and prompted to report the other two elements. The model-comparison analysis revealed that the Hybrid model provides the best quantitative fit to the data. We conclude that multi-element units are represented on two levels, as an integrated unit retrieved in an all-or-none manner, and in addition through pairwise bindings between their elements. Moreover, the assumption that bindings of non-spatial elements are mediated through their shared spatial location – a special case of the Pairwise-Binding model – was not supported by the data. Keywords: working memory, binding, multinomial modeling