Routing on the Stiefel Manifold: When Does Adaptive Subspace Selection Help for Cross-Domain EEG Decoding?

Researchers propose dynamic Stiefel routing, a novel machine learning approach using expert projection filters on the Stiefel manifold to improve cross-domain EEG decoding without requiring target-domain calibration data. The method addresses a fundamental degeneracy problem where naive routing collapses to ensemble averaging, introducing three structural properties that enable genuine domain-specialized routing with significant accuracy improvements across datasets.

This research tackles a persistent challenge in brain-computer interfaces and neuroscience: adapting EEG signal processing models across different subjects and experimental conditions. Cross-domain EEG decoding has practical applications in clinical diagnostics and assistive technologies, where collecting calibration data from new subjects is often impractical or costly. The paper identifies and solves a critical theoretical problem—that naive implementations of adaptive routing mechanisms mathematically reduce to static ensemble methods, providing no actual benefit over simpler baselines.

The proposed solution leverages differential geometry on the Stiefel manifold, a mathematical space representing orthogonal projection matrices. By introducing three design principles—a symmetric anchor filter, a frozen domain-discriminative encoder, and a decoupled key alignment loss—the authors create routing mechanisms that genuinely specialize for different data domains rather than merely averaging expert predictions. This represents meaningful progress in Riemannian deep learning, a subfield focused on leveraging non-Euclidean geometry for improved model performance.

The experimental results demonstrate consistent improvements across three EEG datasets, with balanced accuracy gains ranging from 5 to 3.8 percentage points. Notably, the approach requires no dataset-specific hyperparameter tuning, relying instead on a single data-driven alignment rule. This generalization capability is crucial for practical deployment in medical and research settings where transfer learning is essential.

For the broader AI community, this work contributes to understanding when and why adaptive mechanisms succeed in machine learning. The findings on routing degeneracy apply beyond EEG to any system using mixture-of-experts on manifolds, potentially influencing architecture design in specialized domains requiring geometric constraints.

• →Dynamic Stiefel routing with three structural properties enables genuine domain-specialized expert selection without calibration data
• →Naive adaptive routing provably collapses to ensemble averaging, a degeneracy solved through symmetric anchoring and decoupled training
• →Cross-domain EEG decoding accuracy improves 3.8-5 percentage points with automatic alignment rules and no hyperparameter search
• →The approach leverages Riemannian geometry on covariance matrices to respect non-Euclidean structure of EEG data distributions
• →Results generalize across three independent datasets, suggesting applicability to similar domain adaptation problems on manifolds