An Ultra-Low Power Wearable BMI System With Continual Learning Capabilities

Area of Science:

• Neuroscience
• Machine Learning
• Embedded Systems

Background:

• Wearable Brain-Machine Interfaces (BMIs) are increasingly utilizing embedded processing for enhanced portability, privacy, and battery life.
• Challenges in wearable BMIs include managing electroencephalographic (EEG) signal variability and limited onboard resources, impacting latency and performance.

Purpose of the Study:

• To develop and deploy a Brain-Machine Interface (BMI) workflow incorporating a Convolutional Neural Network (CNN)-based Continual Learning (CL) framework.
• To enable adaptive performance in wearable BMIs by addressing inter-session signal variability.

Main Methods:

• A CNN-based Continual Learning (CL) framework was developed for adaptive BMI operation.
• The CL workflow was implemented on a wearable, parallel ultra-low power BMI platform (BioGAP).
• Performance was evaluated using two in-house datasets (Dataset A and Dataset B).

Main Results:

• The CL workflow demonstrated significant improvements in average accuracy, reaching up to 30.36% (Dataset A) and 10.17% (Dataset B).
• On a Parallel Ultra-Low Power (PULP) microcontroller (GAP9), the system achieved low energy consumption (0.45 mJ/inference) and rapid adaptation (21.5 ms).
• The BioGAP platform provides approximately 25 hours of battery life with a 100 mAh battery.

Conclusions:

• The proposed on-device CL framework enhances inter-session performance for wearable BMIs.
• The system successfully balances low latency, high accuracy, privacy, and extended battery life for practical applications.
• This approach shows significant promise for advanced, real-world embedded BMI systems.