A spiking neural network with continuous local learning for robust online brain machine interface.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Spiking Neural Networks (SNNs) offer computational efficiency and biological plausibility for Brain-Machine Interfaces (BMIs).
• Existing SNN training methods demand substantial memory and cannot handle continuous input streams without periodic interruptions for backpropagation.
• Continuous, uninterrupted training is crucial for ideal BMIs to minimize user disruption and adapt to dynamic neural environments.

Purpose of the Study:

• To develop a continuous SNN weight update algorithm for regression learning.
• To eliminate the need for storing past spiking events, thereby reducing memory requirements.
• To evaluate the algorithm's performance in real-world neural data and simulated closed-loop BMI settings.

Main Methods:

• Proposed a novel continuous SNN weight update algorithm enabling regression learning.
• Implemented the algorithm without requiring memory for past spiking events, achieving constant memory usage.
• Evaluated the SNN on primate premotor cortex neural recordings during reaching tasks and in a simulated closed-loop environment.

Main Results:

• Achieved peak correlation (ρ=0.7) comparable to offline training methods while reducing memory usage by 92%.
• Demonstrated state-of-the-art accuracy in a closed-loop simulation.
• Showcased adaptability to neural input disruptions and successful online training without prior offline training.

Conclusions:

• The developed continuous learning SNN presents a rapid neural decoding algorithm suitable for closed-loop BMI applications.
• This algorithm enhances user acclimation speed and adapts to neural behavior changes with minimal disruption.
• The findings pave the way for more responsive and user-friendly brain-machine interfaces.