Riemannian potato

An online artifact-detection method that operates on trial covariance matrices rather than raw channels. Each short data segment is summarized by its spatial covariance matrix, a point on the manifold of symmetric positive-definite matrices; a reference mean and a dispersion are estimated from clean data, and a new segment is rejected when its Riemannian (geodesic) distance from the mean exceeds a z-score threshold — the accepted region forms a blob (the potato) on the manifold. Because a covariance captures both channel power and inter-channel structure, a single criterion catches diverse artifacts (blinks, muscle, movement, a bad channel) that change the spatial covariance.

The method is unsupervised, fast and naturally online, which fits it to real-time BCI quality control and complements the Riemannian classifiers used for decoding. Its assumptions are those of the manifold framework — reasonably full-rank covariances and a meaningful clean reference — and, like ASR, it flags whole segments rather than surgically removing an artifact, so it is a gatekeeper (accept or reject) rather than a corrector.