Human learning of noninvasive brain-computer interfaces via manifold geometry.

Area of Science:

• Neuroscience
• Human-Computer Interaction
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) offer potential for restoring and enhancing human capabilities.
• BCI adoption is hindered by slow and inconsistent user learning.
• Existing BCI paradigms do not fully utilize the inherent structure of neural activity.

Purpose of the Study:

• To investigate if leveraging the intrinsic geometry of brain activity can accelerate BCI learning.
• To determine how neural manifolds constrain learning in real-time functional magnetic resonance imaging (rtfMRI)-BCI tasks.
• To identify principles for improving neurotechnologies by understanding brain activity geometry.

Main Methods:

• Participants trained using rtfMRI to control a video game avatar via self-modulation of spatial navigation brain regions.
• Data diffusion techniques were employed to extract the intrinsic manifold of brain activity.
• The mapping between brain activity and avatar movement was perturbed to assess the influence of neural manifolds on learning.

Main Results:

• BCI learning was significantly accelerated when new control mappings aligned with directions of high variance on the intrinsic neural manifold.
• Participants could successfully learn to control the avatar by realigning brain activity along manifold-guided directions.
• Learning failed when new mappings deviated from the intrinsic manifold structure, indicating its constraining role.

Conclusions:

• The intrinsic geometry of brain activity, or neural manifold, plays a crucial role in guiding human learning of complex cognitive tasks within BCIs.
• Leveraging neural manifold geometry offers a promising principle for enhancing the speed and consistency of BCI learning.
• This research identifies a fundamental mechanism for improving the design and efficacy of future neurotechnologies.