Batch Normalization modifications for Improving Steady-State Visual Evoked Potentials Classification in Brain-Computer Interfacing

Accurate and robust classification of Steady-state visual evoked potentials (SSVEPs) signals for brain-computer interface (BCI) applications remains a challenge, particularly with short signal durations. The non-stationarity of Electroencephalography (EEG) data requires taking into consideration the individual samples differences. This study investigates the improvement of Batch normalization layers in deep neural networks (DNNs). We employed two recent modifications on state-of-the art models. The proposed modifications are evaluated on a challenging SSVEP dataset, demonstrating good improvements in classification accuracy compared to the baseline DNNs. Our findings highlight the potential of improved normalization techniques to enhance the reliability and efficiency of SSVEP-based BCI systems.