UC Berkeley

The brain has an incredible capacity to learn how to control various effectors, ranging from those endogenous the body to those that are artificially implanted. Moreover, changes to these effectors can be readily adapted by the brain, albeit on varying timescales depending on the amount of perturbation. For example, many studies have shown adaptation to force perturbations or visual-motor rotations to occur within minutes, attributing the adaptation to inputs from the cerebellum. This flexible adaptation of the brain makes the use of brain-machine interfaces (BMIs) an attractive rehabilitative option for a variety of motor pathologies such as stroke and amyotrophic lateral sclerosis. BMIs allow users to interact with their environments by using signals recorded directly from the brain and transforming them into actions taken on by an external effector such as a computer cursor or robotic arm.

Past work has shown neurons to adapt to decoders over time, suggesting BMIs as an exciting paradigm for neural rehabilitation. Modern decoding methods involve using closed-loop decoder adaptation (CLDA) allowing for decoder weights to converge quickly, and users to gain high levels of control with very little training. Furthermore, these weights can be updated over learning in the event there are changes in the neural population. While great for assistive devices, it is unclear how these decoders impact neural adaptation over time. To answer this question, we conducted four studies examining the interplay between decoder adaptation and changes in neural activity over learning.

Previous studies in the field have shown neural activity to fire in increasingly correlated patterns, converging onto low-dimensional spaces as performance with a BMI improves, indicating consolidation of coordinated firing patterns. However, these studies were conducted with static decoders. To examine how neurons adapt alongside CLDA, we used nonhuman primate models to observe how neurons in motor cortex adapt when decoder weights change in a BMI task. We found that there is a concomitant exploration-exploitation strategy that occurs where changes in decoder weights yield increases in exploratory neural activity, which in turn consolidates over many days. These results are remarkably paralleled by those in the psychophysics literature, suggesting similar mechanisms of adaptation between the two regimes. Furthermore, we found that initial levels of decoder adaptation bear no impact on the rate of neural adaptation. That is, decoders resulting in better initial performance of a BMI did not affect how quickly neurons adapted over multiple days.

However, we found that there are differences in adaptation between neurons used as decoder inputs compared to other neurons in the local supporting network. While neurons that were delegated as decoder inputs explored and eventually collapse onto a low-dimensional space, neurons in the supporting local network had comparatively less exploration and consolidation of activity. Strikingly though, these support neurons show task-relevant behavior throughout learning. In combination, these results suggest that neural adaptation to BMIs is driven by network solutions that aim to improve performance through exploring new patterns rather than suppressing preexisting ones. Rather than improving control of the BMI by decreasing noise, motor cortex seems to boost the signal quality instead. Together, the results in this body of work highlight mechanisms of neural adaptation to neuroprosthetic devices that mirror those of natural motor adaptation. Exploiting these mechanisms may lead to better methods of utilizing BMIs for neural rehabilitation.