Generality and Generalisation in Human Sensorimotor Skill Learning

Generalisation is an essential feature of human sensorimotor skill learning. Without the ability to generalise, each variation of a sensorimotor control task would have to be learned anew. Using a new conceptual model of learning in humans and machines, we show how learning that benefits one skill can simultaneously harm other skills dependent on the same underlying parameters. This observation motivates the central question of the thesis: how do humans achieve generalisation in sensorimotor skill learning while avoiding its potentially detrimental effects?

To investigate this question, we developed and tested a range of myoelectric human-computer interface devices to support empirical studies of novel sensorimotor skill learning. Using these devices, we conducted two studies to probe the mechanisms underlying interference-resistant skill learning in humans. In a single-session myoelectric control study, we found evidence that participants learned to adjust their mapping from task goals to motor actions, effectively re-using existing control mechanisms to rapidly achieve good performance on the novel task. In a five-session myoelectric control study, we showed that participants can generate a rich variety of behaviours simply by controlling the relative timing of two constant motor outputs.

Based on these results, we argue that re-use of existing learning may be an important means through which humans learn novel sensorimotor skills while preserving previously learned skills. By preferentially modifying behaviour-defining parameters responsible for the selection and timing of output-generating processes, skills which depend upon the latter processes can be preserved during new learning. We suggest that the hierarchical nature of human motor control may naturally induce this bias towards re-use, mitigating the potential harms of shared-parameter generalisation while supporting its beneficial effects.