Efficient Motor Learning Through Action-Perception Cycles in Deep Kinematic Inference

Matteo Priorelli, Ivilin Peev Stoianov

Abstract: How does the brain adapt to slow changes in the body’s kinematic chain? And how can it perform complex operations that need tool use? Here, we consider both processes through the same perspective and propose that the kinematic chain is represented by an Active Inference model encoding, in a hierarchical fashion, intrinsic and extrinsic information separately. However, the several pathways through which prediction errors can be minimized introduce some optimization problems. We show that an agent can rapidly change its kinematic chain online using action-perception cycles, similar to how learning and inference processes are handled in Predictive Coding Networks.