A discriminative SPD feature learning approach on Riemannian manifolds for EEG classification

•We propose a new Riemannian-based deep learning network to generate more discriminative features for electroencephalogram (EEG) classification.•The proposed Riemannian Barycenter layer block pulls SPD features of the same class toward their barycenters and pushes other deep features of different classes farther away.•Our experimental results demonstrate the superiority of the proposed framework for learning the non-stationary nature of EEG signals.•We demonstrate the increased power of deep features over the state-of-the-art methods in different EEG-based tasks.