Sensorimotor learning under switching dynamics

Humans have a remarkable capacity to learn new motor behaviours without forgetting old ones. This capacity relies on the ability to acquire and express multiple motor memories without interference. Here we combine behavioural experiments and computational modelling to investigate how the sensorimotor system uses contextual information to create, update and recall motor memories. We first examine the role of muscle co-contraction in the learning of novel movement dynamics. We show that muscle co-contraction, as measured by surface electromyography, accelerates motor learning. We then explore the role of control points on objects in the formation of motor memories during object manipulation. We show that opposing dynamic perturbations, which interfere when controlling a single location on an object, can be learned when each is associated with a separate control point. To account for these results, we develop a parametric switching state-space model, in which the association between cues (control points) and contexts (dynamics) is learned from experience rather than fixed. We then extend this model to a Bayesian nonparametric switching state-space model, in which the number of contexts and cues are learned online rather than specified in advance. This model can instantiate new memories when novel perturbations are experienced and exhibits spontaneous recovery of a memory that has been ostensibly overwritten. To test the model, we perform an experiment in which we briefly present a previously experienced perturbation after behaviour has returned to baseline. As predicted, we observe a qualitatively distinct and more pronounced form of recovery, which we refer to as evoked recovery. Finally, we investigate Bayesian context estimation using single-trial learning. We show that people are able to learn novel associations between cues and contexts and that they use both contextual cues and state feedback to infer the current context and partition learning between memories. Taken together, these findings further the understanding of the behaviour and computational principles of sensorimotor learning under switching dynamics.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2020

Author dc:creator

• Heald, James Barry

Advisor dc:contributor.advisor

• Wolpert, Daniel Mark

• motor learning
• motor control
• context estimation
• Bayesian inference
• nonparametric switching state-space model

Licence

• https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/67bf89e8-e7a7-4ac3-866a-1a243d7a5993/download
• https://creativecommons.org/licenses/by-nc-sa/4.0/

Language dc:language

en

Author Identifier

0000-0002-7293-7914
0000-0003-2011-2790

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/301826