Balancing Memorization and Generalization in RNNs for High Performance Brain-Machine Interfaces.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Brain-machine interfaces (BMIs) offer potential for restoring motor function in paralyzed individuals.
• Current BMIs are limited by the accuracy of real-time decoding algorithms.
• Recurrent neural networks (RNNs) show promise but require rigorous evaluation in closed-loop systems.

Approach:

• Compared RNNs against other neural network architectures for real-time, continuous decoding of finger movements.
• Utilized intracortical signals from nonhuman primates in one and two-finger tasks.
• Evaluated performance across varying movement set complexities and degraded input signals.

Key Points:

• LSTMs, a type of RNN, outperformed convolutional and transformer networks, achieving 18% higher throughput.
• RNN decoders demonstrated the ability to memorize movement patterns, matching able-bodied control on simplified tasks.
• Functional control was recovered even with poor input signals by training RNNs as both classifiers and continuous decoders.

Conclusions:

• RNNs show significant potential for enabling functional, real-time BMI control.
• These algorithms can learn and generate accurate movement patterns, overcoming current decoding limitations.
• The findings suggest a pathway to more effective neuroprosthetic devices for paralysis.