A Flexible Dilated Causal Convolution Model for Steady-State Visual Evoked Potential Signal Classification

As one of the most rapidly developing research directions in the field of brain-computer interface (BCI), steady-state visual evoked potential signal (SSVEP) received great attention from researchers. However, due to the difficulty of collecting sufficient electroencephalogram (EEG) data and the difference between SSVEP-based BCI system experimental paradigms, the current proposed deep learning models for SSVEP decoding did not perform stably in different experimental scenarios and had problems in balancing pursuing high accuracy and low training cost. In this study, we proposed a flexible multi-task SSVEP model and proved its adaptive ability on three publicly available SSVEP datasets with different experimental paradigms which included experimental data from the elders. The model reached an average accuracy of 85.22%, 91.67% and 73.98% in each dataset, higher than the advanced SSVEP decoding methods FBCCA and TRCA. Benefitted by its adaptive ability in adjusting the hyperparameters of the models, it successfully reduced the training cost while ensuring the accuracy. These results indicated the proposed model can be applied to new experimental paradigms, and has strong potential in building online SSVEP-based BCI systems thus achieving variable applications in the future.