Description
The process of skill learning can span a wide timeframe, but this thesis focuses on the "early” part of learning, in which an animal initially acquires and begins to refine a skill. Here, we define a skill as any set of intended behaviors — whether covert or overt — that improve in consistency over time as a result of training and reward. Within this thesis, we’ll first explore which skills can be learned in the human peripheral nervous system by investigating the level of flexibility latent in our muscles, with an eye towards designing a high-performance, non-invasive neural-machine interface. In particular, classical experiments suggest motor unit activity within a single muscle lies along a low-dimensional manifold due to a common descending input, theoretically severely constraining the throughput of an assistive device utilizing motor unit activity. However, by utilizing a neural-machine interface paradigm across six days of training in eight adults, we found that motor unit dimensionality is higher than previously theorized, showing promise that a motor-unit-based neural machine interface could provide clinical benefit.
Second, we’ll ask the “how” of skill learning: how do we learn skills, whether motor or cognitive, in the brain? Here, we focus on studying a core mechanism of learning, known as “credit assignment”, which enables activity in the cortex to become more stereotyped over training. Utilizing a neural-machine interface paradigm in rats, we investigate the role of the striatum, an area well-known for housing associations between behavior and reward, in credit assignment. We found that activity of neurons in the striatum form an internal model of cortical activity, continuously estimating both the proximity of current cortical activity to reward and the change in this proximity, with this model of cortical activity differing in complementary yet distinct ways between dorsal and ventral striatum. Such a model could be critical to implementing credit assignment in the brain.
Finally, we’ll conclude this thesis on a practical note by discussing the closed-loop, performant software built for the aforementioned studies, open-sourced to benefit the broader neuroscience community. Neuroscience is an increasingly complex, performance-sensitive discipline, with neural data volume exploding over the past few decades. Software developed throughout this thesis work empowers researchers to build more flexible, more scalable systems with relative ease.
Therefore, by utilizing neural-machine interfaces to interrogate the functionality of the nervous system, the results described in this thesis contribute to our understanding of skill learning, demonstrating novel levels of flexibility in the peripheral nervous system with applications in clinical translation and suggesting a critical role of the striatum in skill learning as a part of credit assignment, a core neural mechanism of learning.