Few-shot Algorithms for Consistent Neural Decoding (FALCON) Benchmark

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Intracortical brain-computer interfaces (iBCIs) restore function for paralyzed individuals by decoding neural activity.
• Neural data non-stationarity causes decoder failure, necessitating frequent recalibration.
• Current few-shot and zero-shot approaches for recalibration lack standardized evaluation.

Purpose of the Study:

• Introduce the FALCON benchmark suite to standardize the evaluation of iBCI decoder robustness.
• Provide standardized datasets and evaluation metrics for comparing iBCI decoding methods.
• Facilitate the development of robust iBCI decoders for real-world applications.

Main Methods:

• Curated five datasets of neural and behavioral data for movement and communication tasks.
• Developed a flexible evaluation platform for user-submitted code.
• Implemented baseline methods across various decoding approaches for seeding the benchmark.

Main Results:

• The FALCON suite enables standardized comparison of few-shot and zero-shot iBCI decoding algorithms.
• The benchmark focuses on behaviors relevant to current iBCI applications.
• A flexible platform allows for efficient evaluation of submitted algorithms.

Conclusions:

• FALCON provides a standardized framework for evaluating and selecting robust iBCI decoders.
• This benchmark aims to reduce the burden of decoder recalibration for users.
• Standardization through FALCON will accelerate the translation of iBCI technology.