Sliced‑Wasserstein Importance Weighting for Robust Brain-Computer Interface Speech Decoding

Noah Cowan, Scott Linderman

Reviewer: ~Noah_Cowan2

Presenter: ~Noah_Cowan2

TL;DR: Weighting past sessions by their sliced‑Wasserstein distance to a unlabeled prefix of the current day stabilizes BCI decoding—cutting PER from 0.296 to 0.169 (−42.9%) on the first post‑train session and sustaining gains across the next three days.

Abstract: Brain–computer interfaces (BCIs) hold transformative potential, but their performance often degrades across sessions due to signal drift and calibration challenges. In this paper, we propose a method to improve cross-session robustness by reweighting training data according to their similarity to the target session, as measured with the Sliced-Wasserstein distance. We provide theoretical justification for this approach in a simplified statistical model, and we evaluate it on real BCI data. Our results show that Sliced-Wasserstein weighting improves BCI performance by reducing phoneme error rate from 0.296 to 0.169 (a 42.9\% reduction) on the first post-training session, and it maintains nearly the same level of performance over the following three sessions. Our results suggest that distributionally informed reweighting offers a principled and fully unsupervised way to mitigate session-to-session variability in BCIs, paving the way toward more reliable long-term neural decoding without the need for costly recalibration.

Length: short paper (up to 4 pages)

Domain: methods

Format Check: Yes, the presenting author will attend in person if this work is accepted to the workshop.

Author List Check: The author list is correctly ordered and I understand that additions and removals will not be allowed after the abstract submission deadline.

Anonymization Check: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and URLs that point to identifying information.

Submission Number: 62