A Kernel Reinforcement Learning Decoding Framework Integrating Neural and Feedback Signals for Brain Control.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-Machine Interfaces (BMIs) enable brain control (BC) of external devices using reinforcement learning (RL) decoders.
• Co-adaptation occurs between subjects and decoders, but current decoders don't fully utilize feedback cues for enhanced learning.

Purpose of the Study:

• To develop a novel kernel RL decoding method that integrates both neural signals and feedback cues.
• To address the challenge of combining signals with different temporal scales for improved BC training efficiency.

Main Methods:

• Proposed a kernel RL decoding method projecting neural signals and feedback cues into separate Reproducing Kernel Hilbert Spaces (RKHSs).
• Created a joint feature space for linear decoding of neuro-prosthesis actions.
• Evaluated the method on a simulated brain control cursor-reaching task.

Main Results:

• The proposed method demonstrated faster learning speed compared to traditional kernel RL using only neural signals.
• Achieved better decoding accuracy by effectively integrating feedback cue information.
• Successfully facilitated the training procedure for the BC task.

Conclusions:

• The novel kernel RL decoding framework effectively integrates neural signals and feedback cues.
• This integration significantly increases learning speed and accuracy in BC tasks.
• Enables subjects to learn BC tasks more easily, improving clinical relevance.