Preserved Embedding of Neural Population Activity for Brain Computer Interfaces

Abstract Inferring neural population dynamics from motor cortex recordings is essential for brain-computer interface (BCI) control. However, neural activity patterns during similar behaviors exhibit systematic variability across recording sessions, subjects, and task contexts due to neural plasticity, anatomical differences, and varying movement distances. This variability degrades BCI decoder performanc when applied to different sessions, subjects, and task contexts. Here, we introduce Multi-Aligned Neural Data Transformer (MANDT) that extracts consistent embeddings of the motor cortical dynamics. By learning these consistent embeddings, MANDT enables generalization of BCI decoding: a decoder trained on data from one session maintains robust performance when applied to new sessions without recalibration. We validated our approach on neural recordings from rhesus monkeys performing the delayed-reach task. Lastly, we show that CEBRA can be used for the robotic control, enabling direct neural control of a humanoid robot, allowing monkeys to guide complex robotic manipulation tasks through brain signal.