A Generalist Intracortical Motor Decoder.

Area of Science:

• Neuroscience
• Machine Learning
• Bioengineering

Background:

• Understanding the neural basis of motor control is crucial for advancing neurotechnology.
• Traditional methods often simplify neural data, limiting broad applicability.
• Foundation models offer a new approach by integrating vast datasets.

Purpose of the Study:

• To evaluate the efficacy of foundation models, specifically autoregressive Transformers, for motor decoding using large-scale neural population spiking activity.
• To assess the generalizability of pretrained models across diverse motor decoding tasks and neural data variations.

Main Methods:

• Pretrained an autoregressive Transformer model on 2000 hours of intracortical microelectrode data from monkeys and humans.
• Paired neural spiking activity with extensive motor covariates.
• Evaluated model performance on 8 downstream motor decoding tasks and across various neural distribution shifts.

Main Results:

• The pretrained foundation model demonstrated broad utility, improving performance on multiple downstream decoding tasks.
• The model exhibited generalization capabilities across different neural data distributions.
• Identified limitations related to sensor variability and output stereotypy inherent in neural datasets.

Conclusions:

• Foundation models show promise for motor decoding by leveraging large, diverse neural datasets.
• While powerful, current autoregressive Transformer scaling may not fully overcome inherent data limitations.
• Further research is needed to address sensor variability and output stereotypy for enhanced neural decoding.