An emerging view of neural geometry in motor cortex supports high-performance decoding

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Current brain-computer interface (BCI) decoders rely on assumptions about neural activity constraints.
• Recent findings suggest these assumptions may not accurately reflect the true geometry and statistics of neural activity.
• This mismatch can limit the performance of existing BCI decoders.

Purpose of the Study:

• To develop a novel decoder, MINT (Model-based Inference of Neural Trajectories), that incorporates more appropriate statistical constraints for neural activity.
• To evaluate MINT's performance against standard decoders and advanced machine learning methods.
• To assess the simplicity, scalability, and interpretability of MINT.

Main Methods:

• Designed the MINT decoder based on potentially more accurate statistical constraints of neural activity.
• Compared MINT's performance against traditional interpretable BCI decoders.
• Benchmarked MINT against expressive, data-driven machine learning methods across various tasks.

Main Results:

• MINT demonstrated strong performance across multiple BCI tasks, indicating its assumptions align well with neural data.
• MINT consistently outperformed other interpretable decoding methods in all comparisons.
• MINT achieved superior performance compared to expressive machine learning methods in 37 out of 42 comparisons.
• MINT's computational approach is simple and scales efficiently with increasing numbers of neurons.

Conclusions:

• The MINT decoder's assumptions appear well-suited for current neural data in BCI.
• MINT offers a significant performance improvement over existing interpretable BCI methods.
• MINT presents a competitive and potentially superior alternative to complex machine learning models for BCI.
• MINT's efficiency, interpretability, and strong performance make it a promising candidate for widespread BCI applications.