Chethan Pandarinath, PhD - US grants

For decades, neuroscientists have recorded from single brain cells (neurons) to understand how the brain senses, makes decisions, and controls movements. We can now record from hundreds of neurons simultaneously but are still at an early stage in developing tools for determining how networks of neurons work together to perceive the world and to generate the control signals needed to produce coordinated movement. Focusing on movement, this project brings to bear the power of deep learning --- powerful new machine learning algorithms --- on the problem of understanding neural activity. Because deep learning thrives on big data, the investigators can leverage massive-scale brain recordings. These include month-long recordings chronicling the activity of 100 neurons as a monkey goes about its daily business, or recording from thousands of neurons for hours in the mouse, each identified with an exact location in the brain and tied to the mouse's on-going behaviors. These approaches will open new windows on how neurons act together moment-by-moment to produce movement. The investigators will develop simple descriptions of the underlying processes to be shared with the public through venues including online tutorials, a new open course that will be developed at Emory University and Georgia Tech, the Atlanta Science Festival, and Atlanta's Brain Awareness Month. They will also make their data sets publicly available, and host data tutorial and modeling competitions at key scientific meetings, to accelerate progress by engaging the broader scientific community.

In the fifty years since Ed Evarts first recorded single neurons in M1 of behaving monkeys, great effort has been devoted to understanding the relation between these individual signals and movement-related signals collected during highly constrained motor behaviors performed by over-trained monkeys. In parallel, theoreticians posited that the computations performed in the brain depend critically on network-level phenomena: dynamical laws in brain circuits that constrain the activity and dictate how it evolves over time. The goal of this project is to develop a powerful new suite of tools, based on deep learning, to analyze these dynamics at unprecedented temporal and spatial scales. The investigators will leverage recordings with month-long M1 electrophysiology, EMG, and behavioral data during natural behaviors from monkeys, and vast numbers of neurons recorded with two-photon imaging from behaving mice. Novel machine learning techniques using sequential auto-encoders will enable the investigators to learn the dynamics underlying these data. This combination will provide windows into the brain's control of motor behavior that have never before been possible. The novel analytical framework developed here will be extensible from motor behaviors to higher level problems of error processing, decision making, and learning.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

ABSTRACT People with disabling motor disorders rely on assistive devices and caregivers for many of their most basic needs. Current assistive devices are inherently limited, as they rely on (and encumber) residual motor function as a command interface. Brain-machine interfaces (BMIs) provide a pathway to more powerful assistive options by directly monitoring brain activity and using it to decipher movement intention in real-time. However, BMIs have yet to achieve performance and robustness that would warrant widespread clinical adoption. A key obstacle is that nearly all BMIs to date use direct decoding, i.e., they attempt to map the activity of brain areas like motor cortex (MC) directly onto external movement parameters such as velocity. This has resulted in BMIs that are brittle: they often fail in new contexts, and are highly sensitive to neural interface instabilities. Instead, I envision a radically different approach with the potential to impact virtually every existing BMI application. The central element is dynamical systems decoding (DSD), a framework I developed that fuses advances in motor neuroscience with cutting-edge AI methods to achieve unprecedented decoding accuracy. DSD uses neural networks to precisely reveal MC's complex internal activity patterns, known as dynamics, on a moment-by- moment basis. This enables a clean separation between activity related to internal dynamics and activity related to external movement parameters. In offline analyses, I showed that DSD enables a breakthrough in decoding, predicting movements on millisecond timescales with substantially higher accuracy than the current state-of-the- art. A key focus of this proposal is developing universal, subject-independent BMIs that harness the remarkable similarities in MC dynamics observed across subjects. Using new AI methods to model more than a decade of previously-collected monkey data, we will test whether subject-independent models can enable BMIs that work nearly `out of the box', with performance that could only be achieved through massive datasets, while still avoiding burdensome, subject-specific calibration. In parallel with offline studies, we will work directly with people who are paralyzed to develop online BMIs with unparalleled performance and robustness. Performance improvements will be achieved through hybrid decoding paradigms that capitalize on high-level movement information that is uniquely uncovered via DSD. While BMI robustness is typically limited in direct decoding ? due to gradual changes in the specific neurons being monitored ? DSD will enable robust BMIs by leveraging MC dynamics, which are stable for years and independent of whichever specific neurons are being monitored at a given time. These two innovations would enable BMIs that achieve unprecedented performance and on- demand, 24/7 reliability for years. If successful, these studies will pave the way to dramatically improving the performance, robustness, and clinical utility of nearly every BMI application.