Causal inference, prediction and state estimation in sensorimotor learning.

Abstract

The sensorimotor system must constantly decide which errors to learn from and which to ignore. Recent work has shown that humans are remarkably precise in parsing movement errors into internally and externally generated components for this purpose: participants automatically ignore internally generated reaching errors caused by motor noise, yet implicitly adapt to size-matched externally generated errors caused by visual perturbations. Following replication of these results with 16 neurotypical adults, we formalized our understanding of this behaviour with a novel Bayesian decision-making model. The Parsing of Internal and External Causes of Error (PIECE) model frames adaptation as a process of causal inference regarding the source of error, with the magnitude of motor corrections reflecting a combination of state estimation and the observer's degree-of-belief that their movement was externally perturbed. Thus, PIECE challenges current computational theories that posit adaptation as a process of re-aligning the perceived hand position with the movement goal. When formally compared with three representative models of this hand-to-target alignment view, we show that only PIECE can capture the precise parsing of internal versus external errors observed. Combined, this work provides a normative explanation of how the nervous system discounts intrinsic motor noise and adapts to perturbations, keeping movements finely calibrated.