Transfer learning and SpecAugment applied to SSVEP based BCI classification

Objective: We used deep convolutional neural networks (DCNNs) to classify electroencephalography (EEG) signals in a steady-state visually evoked potentials (SSVEP) based single-channel brain-computer interface (BCI), which does not require calibration on the user. Methods: EEG were converted spectrograms and served as input train DCNNs using transfer learning technique. also modified applied data augmentation method, SpecAugment, generally employed for speech recognition. Furthermore, comparison purposes, we classified SSVEP dataset Support-vector machines (SVMs) Filter Bank canonical correlation analysis (FBCCA). Results: Excluding evaluated user's from fine-tuning process, reached 82.2% mean test accuracy 0.825 F1-Score 35 subjects an open dataset, small length (0.5 s), only one electrode (Oz) DCNN with learning, window slicing (WS) SpecAugment's time masks. Conclusion: The results surpassed SVM FBCCA performances, single length. Transfer provided minimal change, but made training faster. SpecAugment created performance improvement was successfully combined WS, yielding higher accuracies. Significance: present new methodology solve problem of classification DCNNs. recognition technique it context BCIs. presented performances obtained SVMs (more traditional methods) BCIs lengths electrode. This type BCI can be develop fast systems.