Title: EVA-Net: Leveraging Video-Derived Motor Priors for Subject-Independent EEG Motor Decoding

Abstract:

For practical non-invasive Brain-Computer Interfaces (BCIs) to be viable, EEG decoders must exhibit robust cross-subject generalization while requiring minimal calibration. However, achieving subject-independent decoding is often hindered by inter-subject variability and signal non-stationarity, which tend to entangle motor semantics with subject-specific noise. While recent multimodal strategies have employed text as a semantic anchor, such text-based supervision is inherently sparse and static, making it ill-suited for dynamic motor processes. To overcome these limitations, we introduce EVA-Net, a novel two-stage framework that utilizes action videos as semantic priors to enhance subject-independent EEG motor decoding.

In the first stage, we align EEG and video features within a shared space, employing both cross-modal and supervised contrastive objectives to mitigate subject-specific variations. The second stage involves transferring video-derived priors to an EEG-only classifier via knowledge distillation and video category prototypes, a process designed to incur no additional inference overhead. Our experiments across two public datasets demonstrate that EVA-Net delivers superior subject-independent decoding performance, notably achieving an 8.66% increase in Leave-One-Subject-Out (LOSO) accuracy on the EEGMMI dataset. Furthermore, ablation studies indicate that video serves as a more effective semantic anchor compared to the text baseline evaluated in this study.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC