Large-Scale Neural Consolidation in BMI Learning*

Abstract

Brain-machine interfaces (BMIs) use signals acquired from the brain to control actuators such as computer cursors or robotic arms, with potential to restore motor function to individuals with disabilities. While the process of learning and controlling a BMI is complex, involving cortico-striatal networks, it has been well-established that the brain is able to learn to control BMI actuators using relatively few neurons as direct inputs into the decoder. In particular, neurons that are used as inputs to a BMI decoder (direct neurons) experience changes in direction tuning and modulation depth, eventually forming a stable neuroprosthetic map. Furthermore, previous work has shown that indirect neurons (those that are not inputs to the decoder) also form a stable neuroprosthetic map that differs from manual reaching. However, it is still unclear how these changes in indirect units are formed over the course of learning. We found that indirect neurons adapted similarly to that of direct neurons over learning. Indirect neurons formed a stabilized tuning map, decreased neural dimensionality, and consolidated firing activity into more correlated patterns. Furthermore, direct and indirect neurons adapted together, not only coordinating activity within each population, but across populations as well. Together, our results show that indirect neurons change alongside direct neurons, suggesting a large-scale neural search and adaptation for direct neurons.