Online Neural-to-Movement Mapping Transfer for Task Switching and Retention in Brain–Machine Interfaces

Abstract

Current brain-machine interfaces (BMIs) often rely on decoders trained for single tasks, limiting their flexibility in real-world applications. We propose an online learning framework that enables the transfer of neural-to-movement (knowledge) across tasks, supporting both task switching and memory retention. In our framework, neural activity is projected into a dynamical jPCA space to effectively dissociate into variant and invariant components. The variant components of the neural patterns are then aligned by deriving Gradient-based Kullback–Leibler Divergence Minimization (GKLD) for efficient online adaptation. A kernel reinforcement learning (KRL) model then decodes aligned neural signals while reusing prior neural-to-movement mapping. Evaluated on rats switching between a one-lever pressing and a two-lever discrimination task, the framework shows rapid convergence, over four times faster than the baseline method, and improves decoding accuracy by around 35% during task switching. Furthermore, when switching back to the original task, the framework successfully retains knowledge from the old task. Our method demonstrates general applicability to multiple task switching scenarios and maintains stable decoding across three representative days over a 21-day period, highlighting its potential for long-term, real-world use.