BCIwiki (bciwiki.com) — According to arXiv, the preprint "Teaching Neural Decoders with Language," uploaded to the platform on 2026-10-09, introduces Language as Supervision (LaS) to assist neural decoding for brain-computer interfaces (BCIs) using natural-language class descriptions. The preprint has not been peer reviewed. It addresses the fact that conventional class-label supervision (CLS) leaves relationships among classes implicit, whereas descriptions may provide richer references for learning.
LaS aligns neural features with natural-language class descriptions through learned projections, then predicts classes using candidate text prototypes. In practical terms, classification depends on how neural representations match the text representations of candidate classes. The study evaluated seven encoder architectures with independently trained models for visual, motor, and speech decoding from scalp and intracranial recordings. The authors report that LaS improved speech-decoding accuracy by up to 17.15 percentage points over matched CLS baselines. This is the maximum reported improvement, not a gain shared by every task or model.
The authors also report that the learned representations support candidate-set reuse and label-supervised subject and session adaptation. For steady-state visual evoked potential (SSVEP) responses, LaS enables unseen-frequency decoding under specific conditions, with correct class-description correspondence supporting this generalization. Dual-sided attribution visualizations show neural and textual contributions to the same class-matching score. The authors suggest that text descriptions could serve as general-purpose references for learning and interpretation in neural decoding; these findings do not establish clinical effectiveness.