Geometric recentering rivals deep sequence models for cross-session EEG motor-imagery decoding

• A bioRxiv Neuroscience preprint argues that a simple geometric recentering approach can rival deep sequence models for cross-session EEG motor-imagery (MI) decoding.¹
• The authors run a controlled benchmark across eight public MI datasets spanning 3–128 channels, 2–3 classes, and both single- and multi-session settings.¹
• The benchmark holds the feature representation fixed and varies only the decoder, isolating whether added model complexity improves decoding.¹
• The central method is a compact tangent-space geometric pipeline on Riemannian features, used as a strong simple baseline against deep architectures.¹
• Under matched features, the geometric pipeline matches deep sequence models across the eight datasets.¹
• The result challenges claims that increasingly complex deep architectures are necessary for EEG-MI decoding under identical conditions.¹
• The study positions the geometric recentering pipeline as a strong cross-session baseline for BCI decoder design.^{1 1}

Footnotes

1. https://www.biorxiv.org/content/10.64898/2026.07.07.736991v1?rss=1 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸