Abstract

During passive listening, the brain maintains a hierarchy of predictive models to monitor the statistics of its surroundings. The automatic discovery of regular patterns has been associated with a gradual increase in sustained tonic magnetoencephalography (MEG)/electroencephalography activity, sourced in auditory, hippocampal, and frontal areas—reflecting evidence accumulation and establishment of a regularity model. Conversely, when a regular pattern is interrupted, the sustained activity drops—indicating disengagement from the model. However, how such models are established in and retrieved from memory and the conditions under which they are activated and interrupted remain underexplored. In this MEG experiment (N = 26 human participants; both sexes), we examined how neural responses related to model “establishment” and “interruption” are influenced by (1) the rate of stimulus presentation (tone presentation rate 20 vs 40 Hz) and (2) the novelty of the experienced acoustic structure (novel vs resumed regular pattern). The results show that (1) the dynamics of model interruption and establishment are independent of stimulus presentation rate, and that (2) model establishment occurred much faster when an experienced versus novel pattern was presented after pattern interruption, suggesting reactivation of the stored original model facilitated by the hippocampus. (3) Finally, sustained-response rises in response to pattern establishment and interruption were localized in auditory, hippocampal, and frontal sources, supporting top–down model information flow. These results unveil the temporal dynamics and neural network underlying the brain's construction and selection of predictive models to monitor changes in sensory statistics.