Online Neural-to-Movement Mapping Transfer for Task Switching and Retention in Brain-Machine Interfaces

Area of Science:

• Neuroscience
• Machine Learning
• Robotics

Background:

• Current brain-machine interfaces (BMIs) often require task-specific decoders, limiting real-world adaptability.
• Task switching in BMIs presents challenges for maintaining decoding accuracy and retaining learned information.

Purpose of the Study:

• To develop an online learning framework for BMIs that facilitates knowledge transfer across tasks.
• To enhance flexibility and performance of BMIs in dynamic, multi-task environments.

Main Methods:

• Projecting neural activity into a dynamical jPCA space to separate variant and invariant components.
• Utilizing Gradient-based Kullback-Leibler Divergence Minimization (GKLD) for online neural signal alignment.
• Employing a kernel reinforcement learning (KRL) model for decoding with prior knowledge reuse.

Main Results:

• The framework demonstrated over four times faster convergence compared to baseline methods.
• Achieved approximately 35% improvement in decoding accuracy during task switching.
• Successfully retained knowledge from previously learned tasks after switching back.

Conclusions:

• The proposed online learning framework enables efficient neural-to-movement knowledge transfer across tasks.
• Demonstrates robust performance and general applicability for long-term, real-world BMI applications.
• Highlights potential for adaptive and flexible control in complex environments.