Neural systems mediating recognition of changes in statistical regularities

摘要

Neuroimaging research has identified several brain systems sensitive to statistical regularities within environmental input. However, the continuous input impinging on sensory organs is rarely stationary and its degree of regularity may itself change over time. The goals of the current fMRI study were to identify systems sensitive to changes in statistical regularities within an ongoing stimulus, and determine to what extent sensitivity to such changes depends on intentional monitoring of order. We predicted that changes in regularity would be coded for in systems previously associated with statistical coding (hippocampus and middle frontal regions) or event segmentation (posterior medial regions). Participants listened to a rapid train of four different tones whose order levels fluctuated over time. In an active task, participants monitored the tones and indicated when they perceived a change in regularity; in a passive task, they performed a concurrent visuo-motor task and could ignore the auditory input. Behavioral responses in the active task were used to define points of consensus between participants regarding changes in regularity. Activity in 7.5. s epochs that preceded these order-change points was contrasted with activity during matched-length epochs where no participant indicated a change in order. We found that brain regions differentiating these two types of epochs matched those identified in prior research as mediating event segmentation in narratives and movies. These consisted mainly of medial posterior parietal and occipital regions, with limited involvement of temporal and lateral frontal cortices and no hippocampal involvement. In both tasks, order-change epochs were associated with a higher BOLD response than stable-order epochs, but the specific regions showing this pattern varied across tasks. We suggest that partitioning an input stream on the basis of statistical shifts constitutes a basic neural function underlying the ability to segment both semantic and non-semantic inputs. We further discuss the implications of these findings for neurobiological theories of statistical coding and event segmentation.