Learning and encoding motor primitives for limb actions in a brain-like computation approach

Abstract

Recent neurophysiological studies discovered the sparse rotational patterns in the dynamics of neural population during motor control. In this work, we show that a computational model guided by the dynamical system theory of motor coding can successfully generate the similar network behaviors as found in the electrophysiological studies. The RNN-based model learns the arm reaching control policy from self-generated movements. Essential biomechanical and neural properties including multiphasic neural response and the sparse rotation naturally emerge after training for the movement control tasks. The temporal dynamics in the trained network is analyzed to illustrate how the sparse rotational patterns correlate to the generalization capability of the control policy. We find that the trial-and-error motor learning, which naturally brings in the generalization capability, lead to the existence of low-dimensional manifold in the population dynamics of the motor network.