A Generalist Intracortical Motor Decoder

Joel Ye, Fabio Rizzoglio, Xuan Ma, Adam Smoulder, Hongwei Mao, Gary H Blumenthal, William Hockeimer, Nicolas Guazzelli Kunigk, Dalton D. Moore, Patrick J. Marino, Raeed H. Chowdhury, J. Patrick Mayo, Aaron Batista, Steven Chase, Michael L Boninger, Charles M. Greenspon, Andrew B. Schwartz, Nicholas G. Hatsopoulos, Lee E. Miller, Kristofer Bouchard et al. (3 additional authors not shown)

Keywords: Brain-Computer Interfaces, Neuroscience, Motor Cortex

TL;DR: Scaling Transformers for brain-computer interfaces show gains restricted to low data regimes.

Abstract: Mapping the relationship between neural activity and motor behavior is a central aim of sensorimotor neuroscience and neurotechnology. While most progress to this end has relied on restricting complexity, the advent of foundation models instead proposes integrating a breadth of data as an alternate avenue for broadly advancing downstream modeling. We quantify this premise for motor decoding from intracortical microelectrode data, pretraining an autoregressive Transformer on 2000 hours of neural population spiking activity paired with diverse motor covariates from over 30 monkeys and humans. The resulting model is broadly useful, benefiting decoding on 8 downstream decoding tasks and generalizing to a variety of neural distribution shifts. However, we also highlight that scaling autoregressive Transformers seems unlikely to resolve limitations stemming from sensor variability and output stereotypy in neural datasets.

Primary Area: Neuroscience and cognitive science (e.g., neural coding, brain-computer interfaces)

Submission Number: 10528