Toward a network model of the articulatory loop*1

Abstract

The basic features of verbal short-term memory for serially ordered lists are reviewed. A feed-forward network model based on Baddeley's concept of an “articulatory loop” is presented. One of its aims was to explore mechanisms for the storage of serial order information in the articulatory loop. Information is represented locally, learning is by “one-shot” Hebbian adjustment of weighted connections, corresponding to item-item and item-context associations, which decay with time. Items are modeled at the level of phonemes and phonemic output is fed back to the next phonemic input. At recall, items are selected serially by “competitive queuing.” Noisy activation values are used, resulting in errors during recall. Simulations of recall showed good agreement with human performance with respect to memory span, phonemic similarity, word length, and patterns of error. There was good but incomplete agreement on the shape of the serial position curve and on the effects of articulatory suppression. A simple modification is shown to produce the correct serial position curve. However, the model was unable to simulate human memory for sequences containing mixtures of phonemically similar and dissimilar items. A suggested modification which retains the central idea of using competitive queuing to select among noisy activation values is described.