Preserved Embedding of Neural Population Activity for Brain Computer Interfaces
Journal:
bioRxiv
Published Date:
Feb 9, 2026
Abstract
Inferring stable neural representations of motor cortical dynamics is essential for brain-computer interface (BCI) control. However, neural recordings during similar behaviors exhibit intrinsic variability across recording sessions, subjects, and task contexts due to neural plasticity, individual differences, and diverse kinematic patterns. This variability substantially degrades BCI performance when generalizing across these contexts. Here, we introduce Multi-Aligned Neural Data Transformer (MANDT), a transformer-based method that aligns neural dynamics into a shared embedding space to extract consistent neural representations. We validated MANDT on the motor cortical recordings, revealing a circular geometry structure during reaching movements. Moreover, the consistent representations enable generalization of BCI performance: a decoder trained on one session maintain performance on new contexts. Furthermore, we demonstrated the translational potential of MANDT by deploying our framework on a humanoid robot. We establish a novel paradigm for bio-hybrid embodied intelligence where embodied intelligence systems inherit motor capability directly from biological neural systems.
