A computational model of canonical cortical microcircuits for dynamic Bayesian inference and control as inference

Abstract

Canonical cortical microcircuits (CCMs), a six-layer architecture conserved across the mammalian neocortex, play a crucial role in cognitive functions such as sensory inference, decision-making, and motor planning. However, the computational mechanisms underlying these functions remain unclear, and existing models often lack detailed representations of cell types and interlayer connectivity. In this study, we applied the structure-constrained interface decomposition (SCID) method (Yamakawa, 2021) to construct a biologically plausible computational model of CCMs for dynamic Bayesian inference (DBI) and control as inference (CAI). Our model explicitly assigns computational roles to all major excitatory neuronal populations and incorporates the contributions of inhibitory neurons via circuit motifs. Based on this framework, we constructed CCM models of the somatosensory cortex for DBI and motor cortex for CAI. Through simulations of a mouse lever-push/pull perceptual decision-making and control task, our model reproduced key behavioral features, including psychometric curves and “change of mind” behaviors. Furthermore, we conducted layer-specific perturbation simulations that produced experimentally testable predictions about the functional roles of different cortical layers. Based on the duality of inference and control, this study addresses the long-standing challenge of giving a unified account of CCMs’ functions in sensory and motor cortices.