The Relevant Signals for Sensorimotor Adaptation in Human Reaching
- Author(s): Parvin, Darius Euge
- Advisor(s): Ivry, Richard B
- et al.
Motor control is an essential part of what makes us human. It’s important for learning how to walk and talk, for professional athletes and performers, and for recovery after injury, such as being able to function independently after a stroke. Indeed, one can even argue that the ultimate function of the brain is to dynamically select and execute movements. Without being able to move, we wouldn’t be able to express ourselves or affect the world around us. By improving our understanding the motor system, we may not only be able to optimize rehabilitation schedules or brain machine interfaces, but also discover general principles about how the brain is structured.
Our lack of understanding of the motor system is made apparent by the current state of robotics. Humans have a number of mechanical disadvantages compared to robots: The forces generated by our muscles are variable and subject to fatigue, and there is considerable noise and delays in the motor signals sent to our muscles as well as in the processing of sensory feedback from our movements. Despite these challenges, we can perform complex and dexterous tasks, controlling our legs and body to walk along highly variable surfaces or our hands to crack an egg, tasks we do far more gracefully than the most sophisticated robots. Computers can easily outperform humans in games like chess where there are well defined rules and objectives, however, when it comes to learning new movements, it’s unclear what to optimize for. What makes humans great at learning and performing movements is likely not to do with our muscles or processing power, but the learning rules instantiated in our brains.
Human motor learning is thought to be comprised of multiple learning processes. Consider a child learning to tie her shoes: The initial step involves declaratively learning a specific sequence of steps. Through trial and error, reinforcement learning would help favor certain actions over other possible actions. And sensorimotor adaptation can fine-tune the movements, allowing the child to accomplish the task for laces of different thickness or when wearing gloves on a wintry day. These different processes are thought to rely on different brain regions, and the extent to which they are independent processes or interact with each other is an open question.
In order to move accurately throughout life, the motor system must compensate for changes in the body and environment. Sensorimotor adaptation refers to the automatic and implicit process, one this is dependent on the integrity of the cerebellum, essential for keeping the sensorimotor system calibrated. For this form of supervised learning, the actual sensory feedback is compared to the predicted sensory consequences of a movement, with the difference constituting a sensory prediction error. The sensory prediction error is used to make rapid adjustments in an on-going movement and as a learning signal to alter the next movement in order to reduce the movement error.
A key insight concerning sensorimotor adaptation is that this process appears to be impervious to task goals. In a seminal study, Mazzoni and Krakauer (2006) perturbed the visual feedback while participants performed a reaching task. By providing instructions about the perturbation, the participants were able to immediately adjust their behavior such that the cursor hit the target. However, the adaptation system continued to respond to the mismatch between the predicted and observed limb position, even though in this context, the consequences of adaptation were actually detrimental to task performance. This conclusion has been reinforced with a number of different tasks over the past decade, underscoring the automatic nature of adaptation and, computationally, that this learning process is concerned with ensuring that a selected movement is executed properly, rather than that the selected movement is appropriate.
This body of work raises the question of the constraints on adaptation: If not task performance, what sources of feedback influence adaptation? The central purpose of this thesis is to characterize the features of sensory feedback that influence sensorimotor adaptation. By understanding the relevant features or inputs that constrain adaptation, we may be able to better understand the information that is communicated to this learning process. From these we could hypothesize how brain regions involved in adaptation interact with other brain regions that contribute to motor control. This should also lead to new insights and testable predictions about how the motor system will respond in various situations, meaning we could design more effective learning environments where these features are pronounced. In this thesis, I study the inputs to sensorimotor adaptation by testing cases with multiple or ambiguous sources of feedback, exploring how these signals affect performance.
In Chapter 1, I investigate the potential interaction between sensory prediction errors and reward prediction errors. To study this, I use a task in which participants reach towards targets in order to earn points. I manipulate both the task feedback (reward or no reward) and movement feedback, asking how these different feedback signals interact with each other. Contrary to previous hypotheses, we find that sensorimotor adaptation and reinforcement learning operate in parallel. Furthermore, I show that a key feature determining choice behavior is a sense of agency: Does the participant believe they have control over the outcome. If so, they are more likely to seek higher payoffs (riskier behavior) compared to when they believe that they do not control the outcome.
In Chapter 2, I present multiple feedback signals linked to a single movement. By varying which sources of feedback are task relevant and which are task-irrelevant, I ask if adaptation is sensitive to task relevance (even if insensitive to task outcome). The alternative hypothesis is that all feedback information is treated in a similar manner by the adaptation system. The results show that the adaptation system is sensitive to task relevance. However, overall adaptation is attenuated in the presence of irrelevant feedback signals. These results highlight a novel role of task relevance for sensorimotor adaptation, particularly in situations with multiple or redundant sources of feedback.
In Chapter 3, I explore the interaction of vision and proprioception in sensorimotor adaptation. Previous experiments have shown that adaptation in response to visual errors has a limited capacity; the system can be recalibrated up a certain point, beyond which accurate performance requires some sort of change in the movement plan. The basis for this limited capacity is unclear. One hypothesis is that the upper bound on adaptation reflects an equilibrium point between signals concerning visual and proprioceptive sensory prediction errors. This hypothesis predicts that, as proprioceptive information becomes less reliable, the sensitivity to visual errors should be relatively greater, and thus produce a larger upper bound on adaptation. I test hypothesis by asking if variation across individuals in terms of their sensitivity to proprioception is predictive of their response to a visual perturbation. The results show a negative correlation between proprioceptive acuity and the magnitude of adaptation to a visual error, consistent with the idea that adaptation entails the integration of different sensory prediction error signals.
In summary, by studying the response of adaptation of multiple sources of feedback, we have furthered our understanding of the constraints on sensorimotor adaptation. We find that sensorimotor adaptation not only integrates feedback from vision and proprioception, but that under situations it is also sensitive to the task relevance. Although we did not find evidence for a direct interaction between adaptation and decision making, our finding of adaptation’s sensitivity to task relevance suggests an exciting possibility for the interaction with other goal directed systems.