Reinforcement Learning Based Fast Self-Recalibrating Decoder for Intracortical Brain-Machine Interface.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Neural recordings in brain-machine interfaces (BMIs) are nonstationary, necessitating daily retraining.
• Reinforcement learning (RL) offers a self-recalibrating decoder approach but struggles with rapid exploration and performance maintenance.

Purpose of the Study:

• To develop an attention-gated RL-based algorithm to address challenges in RL-based decoders for BMIs.
• To accelerate weight updating and prevent overfitting in self-recalibrating decoders.

Main Methods:

• Proposed an attention-gated RL algorithm integrating transfer learning, mini-batch, and weight updating schemes.
• Tested the algorithm on intracortical neural data from monkeys for decoding reaching and grasping movements.

Main Results:

• Achieved a ~20% increase in classification accuracy compared to non-retrained classifiers.
• Outperformed daily retraining classifiers and improved accuracy by ~10% over conventional RL methods.
• Increased online weight updating speed by approximately 70 times compared to conventional RL.

Conclusions:

• Developed a self-recalibrating decoder with robust and efficient decoding performance.
• The algorithm demonstrates fast weight updating, crucial for practical BMI applications.
• Potential for facilitating applications in wearable devices and clinical settings.