LAND: Latent Aligned Neural-Behavioral Dynamics via Flow Matching forGeneralizable Movement Decoding

Invasive brain-computer interfaces (BCIs) have struggled with generalizable movement decoding, as trained decoders often fail to adapt to new movement speeds, limbs, and subjects. Decoders typically rely on paired data, but integrating behavioral structure from unpaired data for robust out-of-distribution (OOD) decoding has remained a challenge.

To tackle this issue, researchers propose LAND (Latent Aligned Neural-behavioral Dynamics), a framework that aligns neural and behavioral dynamics through flow matching. LAND learns a neural-to-behavioral transport map and leverages behavioral-dynamics priors from unpaired data to promote structured neural manifolds. This regularization encourages cross-domain generalization by shaping representation geometry.

The framework was tested on synthetic neural data, epidural BCIs from a tetraplegic participant, and multi-electrode array (MEA) recordings from nonhuman primates. LAND demonstrated improved zero-shot generalization to unseen movement speeds and generated speed-modulated manifolds. Remarkably, with minimal target-domain fine-tuning, it achieved transfer across limbs and subjects.

These findings suggest that flow-based neural-behavioral alignment using unpaired kinematic priors could be a promising strategy for developing transferable neural representations and ensuring robust movement decoding across diverse behavioral and recording domains.