Multiclass brain-computer interface classification by Riemannian geometry
A. Barachant, S. Bonnet, M. Congedo, C. Jutten
For: developer
The problem
Traditional motor-imagery classifiers relied on spatial filtering (e.g. CSP), which extracts limited information from the spectral covariance structure of EEG.
Method
Spatial covariance matrices were used directly as EEG descriptors and classified using Riemannian geometry on the manifold of symmetric positive-definite (SPD) matrices, with no spatial filtering; two such methods were compared to a reference classifier (multiclass CSP + LDA).
Result
Both Riemannian classification methods outperformed the reference classifier on the multiclass BCI Competition IV dataset (Dataset IIa).
Why it matters
Established a mathematical framework that later became a widely adopted standard in EEG signal processing for brain-computer interfaces.
Limitations
Numerically sensitive to rank-deficient covariance estimates when channel counts exceed the recording window length.