Transfer Learning in EEG-based Reinforcement Learning Brain Machine Interfaces via Q-learning Kernel Temporal...

Area of Science:

• Neuroscience and Machine Learning
• Brain-Computer Interfaces (BCIs)
• Signal Processing

Background:

• Reinforcement learning based brain machine interfaces (RLBMIs) offer potential for real-time applications.
• Transfer learning (TL) enhances machine learning by reusing knowledge from similar tasks.
• Existing BMI applications of TL primarily use supervised learning.

Purpose of the Study:

• Investigate the impact of TL on RLBMIs for decoding free will movement intentions using electroencephalogram (EEG).
• Apply TL strategies to the Q-learning Kernel Temporal Difference (Q-KTD) algorithm for enhanced decoding.
• Utilize a public EEG dataset from a key pressing task with freewill choices.

Main Methods:

• Applied TL strategies to Q-KTD for decoding movement intentions from multichannel scalp EEG.
• Utilized a dataset of participants performing a key pressing task with freewill key selection.
• Implemented TL both between and within subjects for intention decoding.

Main Results:

• Achieved significant increases in decoding success rates (p < 0.01) in 96% of cases.
• Observed success rate improvements ranging from 1.39% to 10.69%.
• Demonstrated TL as an effective method for improving RL-based neural decoder learning efficiency.

Conclusions:

• Transfer learning effectively enhances the performance of RLBMIs for decoding movement intentions.
• Improved neural decoder performance offers an efficient modeling strategy for RLBMIs.
• This approach holds clinical relevance for assisting patients with neurological disorders.