The learning and production of serial movements has recently received much attention from psychological and neuroscience experimentalists as well as modelers. A new model that focuses on how the temporal characteristics of serial behavior change in response to experimental manipulations is presented. The proposed neural network architecture specifies interactions among several distinct bases for serial movement learning and performance. The model accounts for performance of sequences both early and late in learning and provides a unified treatment of changes that occur along the learning continuum. Of particular focus behaviorally are the keypress data of Sternberg et al. , Klapp , and Verwey . Consistent with these data, amongst characteristics that the model exhibits are: 1) a list length effect on latency early in practice that disappears with extended practice; 2) a pattern of long latency followed by markedly shorter inter-response intervals (IRIs) for non-initial sequence elements under two conditions: either with foreknowledge of a novel sequence to be produced and adequate opportunity to prepare for its execution, or without preparation but after significant amounts of practice; 3) a slowing of mean production rate for longer sequences that does not disappear with practice; and 4) a serial position dependence of IRIs that disappears after extended practice. The model is also consistent with patterns of errors that are apparent during serial performance, as well as with the working memory dynamics implicated by recent word-length effects from immediate serial recall tasks (e.g., Hulme et al., 1999).

The major elements of the model are: a fronto-cortical gradient-based representation of serial-order that provides a sequence production buffer and competitive queuing; and a cerebellar-based learning module that learns both serial chunks and individual responses. Through practice, this cerebellar learning mechanism learns to anticipate and preempt slower cortical loading of the appropriate gradient into the frontal production buffer as well as to speed up the execution of individual responses within the sequence. These major components, along with others included within the model, are compatible with neuroanatomical constraints and with the major trends emerging from neurophysiological, clinical, and imaging investigations of learning and performance of serial movements.

This work was supported in part by the Office of Naval Research (ONR N00014-92-J-1309, ONR N00014-93-1-1364, and ONR N00014-95-1-0409) and The University of Ballarat.