A shared functional architecture for error-based and reinforcement-based motor learning in the human brain

Abstract

The motor system can flexibly learn from fundamentally different teaching signals, from directional errors to rewards. While computational and neurobiological models have long assigned these to distinct error-based learning (EL) and reinforcement-based learning (RL) processes, whether this separation holds at the level of whole-brain network architecture has not been directly tested. Here, we characterized whole-brain functional connectivity manifolds in the same participants as they performed separate EL and RL motor tasks -- differing in both feedback type and motor structure -- during two fMRI sessions. By jointly embedding covariance patterns from both tasks into a common low-dimensional neural space, we isolated task-general from task-specific learning-related network reconfigurations. We show that both forms of motor learning induce converging changes in the manifold structure of higher-order transmodal cortex, with comparatively limited task-specific modulation. Notably, this convergence extended to the cerebellum and basal ganglia, the canonical substrates of error-based and reward-based learning, which reconfigured comparably across both tasks. Against this shared backdrop, we found that the posterior medial cortex exhibited a unique functional geometry, selectively redistributing its connectivity across distinct brain circuits depending on feedback type and learning stage. We further demonstrate that individual differences in domain-general motor learning ability are associated with stage-dependent reconfigurations within limbic, default-mode, and attentional systems. These findings indicate that flexible motor adaptation, irrespective of the nature of the learning signal, is supported less by changes within task-specific sensorimotor circuits than by the dynamic reorganization of higher-order brain networks that coordinate them.