Quantized Attention-Gated Kernel Reinforcement Learning for Brain-Machine Interface Decoding

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Reinforcement learning (RL) decoders in brain-machine interfaces (BMIs) interpret neural activity for movement control.
• Conventional RL is time-consuming and ill-suited for rapid learning in BMI tasks.
• Existing attention-gated RL methods can get stuck in local minima.

Purpose of the Study:

• To develop an efficient RL-based decoder for BMIs that can infer goals spatially.
• To overcome the limitations of conventional RL in fast-learning BMI scenarios.
• To improve the decoding accuracy and stability of BMI systems.

Main Methods:

• Proposed Quantized Attention-Gated Kernel Reinforcement Learning (QAGKRL).
• Utilized spatial credit assignment with instantaneous rewards.
• Employed a quantized attention mechanism to avoid local minima and sparsify network topology.

Main Results:

• QAGKRL demonstrated higher successful rates compared to previous methods.
• The proposed method achieved more stable performance in decoding neural activity.
• QAGKRL shows powerful decoding ability for complex BMI tasks.

Conclusions:

• QAGKRL offers a more effective approach for spatial credit assignment in RL-based BMIs.
• This method enhances the potential for sophisticated clinical applications of BMIs.
• The QAGKRL approach addresses key challenges in rapid policy learning for neural decoding.