Human learning of noninvasive brain-computer interfaces via manifold geometry

Brain-computer interfaces (BCIs) promise to restore and enhance human capabilities. Yet, their adoption has been limited by slow and inconsistent learning across users. We show that BCI learning is accelerated by leveraging the naturally occurring geometry, or intrinsic manifold, of brain activity, extracted using data diffusion. Participants were trained with real-time functional magnetic resonance imaging to control an avatar in a video game by self-modulating activity in brain regions supporting spatial navigation. We perturbed the mapping between brain activity and avatar movement to test how neural manifolds constrain human BCI learning. When new mappings relied on directions of significant variance on the intrinsic manifold, participants successfully gained control by realigning brain activity along these directions. When new mappings did not follow the intrinsic manifold, participants could not learn to control the avatar. These findings show how manifold geometry in higher-order brain regions guides human learning of complex cognitive tasks, identifying a principle for improving future neurotechnologies.