Project Summary Nearly 80% of persons with quadriplegia would consent to brain surgery to regain some use of their hands. Yet, despite advances made in the past 15 years, Intracortical Brain-Computer Interfaces (iBCIs) remain largely limited to controlling only the kinematics of simple robotic hand movements. Furthermore, decoder training data consists of a trajectory the user observes and attempts to imitate, together with the accompanying neural activity. But this “observation-based” decoding is not feasible for iBCIs that decode force or muscle activity (EMG), signals which will be needed for more nearly biomimetic iBCIs controlling contact forces and joint stiffness. We now propose a radical new approach, based on machine learning techniques used in the artificial vison and image manipulation fields, to “transfer” a decoder computed from data collected from a monkey to a human with a paralyzed hand. The key to the approach is the observation that the neural activity accompanying movement lies largely on a low-dimensional manifold within the neural state space of recorded neurons. The “latent signals” within the manifold bear remarkably stable information about behavior that is potentially useful for iBCIs. However, any given sample of neurons embeds the manifold in a different coordinate system. We have previously used Canonical Correlation Analysis (CCA) to realign these manifolds for stereotypic, trial- based tasks. We used a fixed BCI decoder with CCA-aligned latent-signal inputs to predict arm movement for as long as two years (“cross-time alignment”). Similarly, we made EMG predictions from a monkey using a decoder computed for a different monkey (“cross-subject alignment”), and even from M1 signals recorded from a person with a spinal cord injury (“cross-species alignment”). But CCA can work only on behaviors that can be trial-aligned, which is not possible for most movements typical of a person's daily life. Here, we propose to develop a new class of tools based on cycle-GAN, a Generative Adversarial Network. Because cycle-GAN works by minimizing the distance between point clouds within the manifold, it can be applied to unconstrained movements. In our initial tests, Cycle-GAN outperformed CCA when used for cross-time alignment of simple behaviors, but it failed for unconstrained behaviors and cross-subject alignment. We propose to optimize cycle- GAN by incorporating information about dynamics (Aim 1), and by initially constraining the particular regions of two partially overlapping manifolds that we subject to the alignment process (Aim 2). We will develop and validate these methods on data recorded during stereotypical motor behaviors in the lab, on unconstrained data collected wirelessly from monkeys housed in a large, plastic cage, and from humans with spinal cord injury through an on- going collaboration with groups at the Universities of Pittsburgh and Chicago. The final result will be a set of novel analytical tools that can be...