BCIwiki (bciwiki.com) — A deep adversarial network built on the Symmetric Positive Definite (SPD) matrix manifold decodes motor imagery EEG across subjects without subject-specific recalibration and outperforms several state-of-the-art unsupervised domain adaptation (UDA) methods on 4 brain-computer interface (BCI) datasets, researchers at Chongqing Normal University and Chongqing University reported in Neural Networks on May 12, 2026 (PMID 42119262, DOI 10.1016/j.neunet.2026.109076).
The authors say EEG non-stationarity and inter-individual variability keep current models from generalizing across subjects without costly per-user calibration, which limits practical deployment. Earlier work treated each subject as a separate domain and learned domain-invariant features through discrepancy minimization or adversarial training, but mostly in Euclidean space, which the team argues cannot capture the non-linear structure of EEG data.
The method, called SPD-DANN, uses adversarial learning on the SPD manifold to extract subject-invariant features. The team added 2 loss terms: an SPD domain feature alignment loss to align features across subjects, and an SPD class prototype pair loss to sharpen class separation within the feature space.
The authors say the loss functions can be adapted to broader unsupervised and semi-supervised domain adaptation frameworks. The abstract does not report specific accuracy figures.