Cracking Cross-Domain EEG: Dynamic Routing on SPD Manifolds

Cross-domain EEG decoding has long been a challenging frontier, mired by the uniqueness of covariance matrices across subjects. Even the most advanced Riemannian deep learning techniques hit a wall domain adaptation. Existing methods either cling to target-domain calibration data or rely on subject-specific components that ultimately fail to generalize. But what if there's a new way forward?

Introducing Dynamic Stiefel Routing

Enter dynamic Stiefel routing, a breakthrough in this niche. Imagine a pool of expert projection filters, each tuned for a specific region on the Stiefel manifold. This system, through cross-attention, routes each input covariance to the most suitable filter, adapting the subspace projection on a per-sample basis. Sounds like science fiction? Not quite. This approach paves the way for greater precision in EEG decoding.

Avoiding the Pitfalls

However, there's a catch. Implementing this naively can lead to a collapse into ensemble averaging. This happens when routing weights become uniform, effectively reducing the adaptive filter to an indistinguishable mix of equal contributions from all experts. So, how do we avoid this trap? Three key structural properties come into play: a symmetric anchor removing proximity bias, a frozen domain-discriminative query encoder, and a decoupled key alignment loss. Together, these elements foster genuine domain-structured routing.

Performance Gains

Let's talk numbers. Deploying this method results in solid improvements across three datasets. Balanced accuracy leaps from 0.773 to 0.823, 0.757 to 0.809, and 0.801 to 0.839. It's not just about boosting numbers. This is about achieving these gains with a single, data-driven rule, sidestepping the need for a painstaking hyperparameter search.

Why It Matters

So why should you care? Because cross-domain EEG decoding isn't just academic. It's a critical step towards more personalized and effective brain-computer interfaces. If the AI can hold a wallet, who writes the risk model? With dynamic Stiefel routing, we're one step closer to AI systems that are truly adaptable.