Interference in working memory: a formal model

The authors propose a neural-network model of working memory. Contents of working memory are represented as distributed patterns of activation. Bindings between content representations and context representations are created by synchronous firing of active units. This assumption leads to the hypothesis of interference by feature overwriting as one source of the capacity limit of working memory. Initial evidence for feature overwriting comes from experiments showing impaired serial recall of words in conditions with high phonemic overlap. The authors illustrate the application of a mathematical formalization of the model to time-accuracy functions of young and old adults obtained with an arithmetic working-memory task. This application shows how theoretically motivated models can be applied through nonlinear mixed effect (NLME) regression, estimating group means and individual differences simultaneously.