Internal Representations for the Generalization of Motor Memories

Abstract

Movement and memory are two of the most fundamental components of our existence. From the moment of birth, we rely on a variety of movements to interact with people and objects around us, and as we grow, we continuously form new motor memories to improve the fidelity of these interactions by exploring and learning more about our environment, especially in unfamiliar situations, ultimately becoming better equipped to handle novel and unknown environments. In this dissertation, we explore four facets of motor memory associated with voluntary movement and postural control in the upper limbs: (1) Optimal motor memory formation via sensorimotor integration. We ask whether the motor system combines prior memories with new sensory information to produce statistically-optimal weight estimates. We find that the weight estimate that the motor system makes in order to re-stabilize one’s arm posture when an object is rapidly removed from the hand that supports it, reflected information integration in a Bayesian, statistically-optimal fashion. Remarkably, we demonstrate that when experiencing the well-known size-weight illusion, the motor and perceptual system’s weight estimates are biased in opposite directions, suggesting two divergent modes for information integration within the central nervous system. (2) Movement features important for the learning and generalization of motor memories. We show that, velocity-dependent adaptation generalizes across different movements, even from discrete straight point-to-point to continuous circular movements, however the amount of generalization is limited and context-dependent. In a series of experiments, we quantified the contributions of different movement features to the elicited adaptation transfer. In particular, we show that other movement states (i.e. position and acceleration) make only minor contributions whereas, the contexts provided by movement geometry and movement continuity are critical. (3) Internal representation of motor memories in intrinsic-extrinsic coordinates. We show that motor memories are based not on fully intrinsic or extrinsic representations but on a gain-field (multiplicative) combination the two. This gain-field representation generalizes between actions by effectively computing movement similarity based on the Mahalanobis distance across both intrinsic and extrinsic coordinates, in line with neural recordings showing mixed intrinsic-extrinsic representations in motor and parietal cortices. (4) Motor memories with local and global generalization. We demonstrate the existence of two distinct components of motor memory displaying different generalization footprints: One generalizes only locally, around the trained movement direction and with the trained end-effector, whereas the other generalizes broadly across both., We proceed to show that broad generalization results from a rapidly-learning adaptive process, dominates on easier-to-learn tasks, and performs high-level processing, producing adaptation vectors that integrate multiple sources of information, in line with a recent theory for perceptual learning.