Exploring the trade-off between deep-learning and explainable models for brain-machine interfaces.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-machine interfaces (BMIs) offer a potential solution for individuals with paralysis by decoding brain activity into movement commands.
• Deep learning decoders achieve high performance in BMIs but raise safety concerns due to their 'black-box' nature.
• Explainable decoders like the Kalman filter (KF) are often used despite lower performance.

Purpose of the Study:

• To develop and evaluate KalmanNet, a novel BMI decoder that integrates KF explainability with recurrent neural networks.
• To compare KalmanNet's performance against traditional KF and deep learning models (tcFNN, LSTM) in predicting finger movements.
• To assess the trade-offs between performance, explainability, and generalization in BMI decoders.

Main Methods:

• Developed KalmanNet, an extension of the KF using recurrent neural networks to compute the Kalman gain, allowing dynamic trust shifts between inputs and dynamics.
• Predicted finger movements from monkey brain activity using KalmanNet.
• Compared KalmanNet's offline and online performance against KF, tcFNN, and LSTM.
• Validated the mechanism with a heteroscedastic KF.

Main Results:

• KalmanNet achieved comparable or superior performance to deep learning models in both offline and online settings.
• KalmanNet demonstrated a flexible reliance on dynamical models for movement initiation and neural inputs for stopping.
• A heteroscedastic KF using the same strategy also approached state-of-the-art performance while maintaining explainability.
• KalmanNet exhibited limitations in generalization and performance with unseen noise distributions, similar to other deep learning decoders.

Conclusions:

• KalmanNet successfully integrates traditional and deep learning approaches for high-performing, explainable BMI decoders.
• The study highlights the potential for hybrid models to enhance BMI safety and reliability.
• Further research is needed to address generalization limitations for broader clinical application.