Dynamic Stiefel Routing for Cross-Domain EEG Decoding
A recent paper on arXiv introduces dynamic Stiefel routing aimed at enhancing cross-domain EEG decoding. This approach tackles the issue of covariance matrices from various subjects residing in separate areas of the SPD manifold. It employs a set of K expert projection filters on the Stiefel manifold, each tailored to a specific SPD region, utilizing cross-attention to direct each input covariance to the most suitable filter. A significant observation is that a naive implementation results in ensemble averaging; uniform routing weights lead to an adaptive filter that merely combines experts equally, resembling a single fixed filter. The study highlights three structural characteristics that avert this collapse.
Key facts
- arXiv:2605.31043v1
- Cross-domain EEG decoding remains challenging
- Covariance matrices from different subjects occupy distinct regions of the SPD manifold
- Existing domain adaptation methods require target-domain calibration data or learn subject-specific components
- Proposes dynamic Stiefel routing with K expert projection filters on the Stiefel manifold
- Each input covariance is routed to the most appropriate filter via cross-attention
- Naive implementation provably collapses to ensemble averaging
- Uniform routing weights reduce adaptive filter to equal-contribution combination of experts
Entities
—