Characterizing the structure of motor cortex activity across multiple behaviors for improved brain-machine interfaces

Project Abstract. Candidate and career goals: I am an engineer by training, with a strong background in neural engineering and the development of motor brain-machine interfaces (BMIs). My career goal is to establish an independent nonhuman primate (NHP) laboratory with two primary aims. First, I will advance our fundamental understanding of the motor system via the combination of electrophysiology with novel statistical and computational methods. Second, I will leverage this knowledge to develop frameworks for superior BMI systems. Throughout my academic and research career I have developed expertise in engineering, computation, and neuroscience with the goal of pursuing these aims. Advances in machine learning, large-scale neural recordings, and deep learning in neural networks are happening quickly (in part via the BRAIN Initiative), and are very promising for the field. Yet very few researchers have the correct combination of skills to make use of them in my areas of interest. In completing the proposed training, I will be uniquely positioned to perform the innovative work necessary to advance our understanding of the planning and execution of cortically controlled movements. I will train the next generation of scientists and engineers in the experimental and computational methods necessary to understand fundamental principles of cortical computations. Research plan: In this project, I will employ multiple computational approaches to understand the structure of population activity in motor cortex (M1) across multiple kinds of behaviors. I will then use that knowledge to create high performance BMI decoders that will be applicable to a wide range of movements. Recent empirical observations are changing our view of the structure of M1 activity. During one particular task (e.g., reaching), neural activity may seem to exist within a small space that it explores completely. Yet as more tasks are observed, it becomes clear that activity comprises a highly structured geometry within a much larger space. This means that activity patters for different movements do not come ?near? one another or overlap. While counterintuitive, this geometry yields new opportunities. By exploiting the separation of activity patterns, movements can be readily distinguished, even when unfolding simultaneously. I will further explore this geometry across multiple behaviors, both in primates and neural network models, to develop new BMI methods. The specific aims of the plan are to (1) create a high-performance decoder for a novel wheelchair-relevant navigation task, (2) build network models to understand M1 activity structure and identify decoding principles that will generalize across tasks (reaching, navigation), and (3) implement a multitask BMI using a unified decoder that allows animals to both navigate and interact with objects. Career development plan: I will be trained by Dr. Mark Churchland and Dr. Larry Abbott at Columbia University.