Mean-Field Representation for EEG Classifications
Suryatapa Roy, Yaoping Hu, Robert John Martinuzzi · 2024
Transformer-based classifiers (e.g., Conformer) are state-of-the-art for classifying electroencephalographic (EEG) signals. One main drawback of these classifiers is their lack of cross-individual generalization. Hence, we proposed a novel mean-field (MF) representation to remedy this drawback. Being quasi-stationary within certain brain regions over a time period, this representation enabled constrained learning for classifying EEG signals. We implemented a transformer – MFT – by cascading Conformer to the MF representation. Using 6 EEG datasets, we conducted a comparison between MFT and Conformer. This comparison revealed that MFT yielded similar outcomes as Conformer in individual-specific classifications but better performance than Conformer in cross-individual classifications. Importantly, the representation enabled MFT classifications to be robust with reduced overfitting and to endorse cross-individual generalization. Moreover, the classifications were interpretable in terms of the MF representation. Such interpretability may be beneficial for EEG-based brain-machine interfaces to equip telepresence.