Learning on the Manifold of Human Brain Activity through Real-Time Neurofeedback

2051 

Abstract Submission 

Erica Busch¹, E. Chandra Fincke¹, Guillaume Lajoie², Smita Krishnaswamy¹, Nicholas Turk-Browne¹

¹Yale University, New Haven, CT, ²Université de Montréal, Montreal, quebec

Erica Busch  
Yale University
New Haven, CT

E. Chandra Fincke  
Yale University
New Haven, CT

Guillaume Lajoie  
Université de Montréal
Montreal, quebec

Smita Krishnaswamy  
Yale University
New Haven, CT

Nicholas Turk-Browne  
Yale University
New Haven, CT

Learning a new behavior is constrained by the geometry, or intrinsic manifold, of neural population activity supporting that behavior. Recent work highlights the importance of manifolds that capture low-dimensional neural dynamics for brain-computer interface learning (Sadtler et al. 2015). Through invasive BCI, non-human primates can learn to operate neural prosthetics more efficiently with a device controlled via activity on the intrinsic neural manifold (Oby et al. 2019; Sadtler et al. 2014). Recent studies have trained humans to self-modulate brain activity through real-time neurofeedback, to enhance perception (Shibata et al. 2011), attention (deBettencourt et al. 2015), or emotion regulation (Keynan et al. 2019), with variable efficacy. Prior work has not considered the neural constraints underlying their neurofeedback training. Here, we leverage manifolds in a human neurofeedback paradigm to expedite learning and unveil dimensions which facilitate learning effects.

We enrolled 20 participants (9 female; 25.8 ± 5.5 y) in a 4 session real-time fMRI experiment where they learned to use their brain activity to control an avatar's movement through a virtual world. In session one, participants practiced navigating the avatar to a goal location with a joystick while fMRI data were collected. We estimated a neural activity manifold of this task from a network of navigation-related regions using the manifold learning algorithm T-PHATE (Busch et al. 2023). In subsequent sessions, participants were trained using neurofeedback to perform the same task by controlling the avatar's movement with their brain activity.

Using a closed-loop system (Wallace et al. 2022), we acquired and transmitted fMRI volumes every 2 seconds to an HPC cluster for processing and embedding onto the T-PHATE manifold. Embedded data were mapped to the direction of the avatar's next movement in the game via one of three manifold components (i.e., intrinsic-, within-, and off-manifold mappings). These components capture the greatest, second greatest, and least variance along the manifold, respectively. Participants received feedback based on a different mapping during each neurofeedback session.

Neurofeedback training used staircasing to quantify the degree each participant's brain exerted over the avatar's movement. Higher control (referred to as "BrainControl") is a behavioral metric of learning in this task, as the parameter scales with performance. We also quantified the change in neural alignment along the manifold, as we predict learning to be driven by an increase in the variance explained by the manifold component yoked to the neurofeedback.

We find neurofeedback learning effects reflected by both behavioral and neural analyses. Behaviorally, we measured learning as the change in BrainControl across the trials of a neurofeedback session. Learning increased for the intrinsic and within-manifold conditions (Fig. 1), with a greater increase for intrinsic than within-manifold condition, but not for the off-manifold condition (Fig. 2A). We calculated neural alignment with the manifold components as the change in the variance in the neural data explained by each component at the start vs end of training. When feedback was based on the intrinsic or within-manifold components, explained variance increased over the course of training, but did not for the off-manifold component (Fig. 2B).

We introduce a framework for manifold-informed non-invasive human BCI, which affords significant learning within one session. Our variance analysis indicates that neural geometry can shift along dominant manifold components, underlying behavioral changes. With manifold-based neurofeedback, we demonstrate control and reorganization of brain activity in higher-order cognitive regions. This suggests important implications for brain-based therapeutic and behavioral interventions.

Neural Plasticity and Recovery of Function

Learning and Memory Other ²

Multivariate Approaches

Other Methods

Brain Machine Interface ¹

Computational Neuroscience

Learning

Machine Learning

Other - Neurofeedback