CCA-Net: Zero-Shot SSVEP Classification via an Integration of Canonical Correlation Analysis and Deep Neural Network

The accurate decoding of electroencephalographic (EEG) signals is a crucial foundation for brain-computer interface (BCI) applications. Among various decoding approaches, those that do not require calibration are particularly significant for advancing BCI technologies into everyday life, as they alleviate user fatigue by eliminating the need for user-specific training. In this study, we integrate a deep neural network (DNN) with canonical correlation analysis (CCA) to form the CCA-Net approach for decoding steady-state visually evoked potential (SSVEP) based BCI without user-specific calibration, i.e., at the zero-shot scenario. The CCA-Net aims to reduce the cross-subject domain gap for better transfer learning and make full use of the testing signal itself. Specifically, it adapts the cross-subject dual-domain fusion network (CSDuDoFN) by transferring the global multireference least-squares transformation (GMLST) coefficient from source subjects to both source and target data to reduce the cross-subject domain gap. In addition, CCA-Net enhances the transfer template-based CCA (tt-CCA) by introducing a subject-specific template to replace the sine-cosine reference template and adding an extra CCA coefficient to make full use of the testing signal itself, resulting in the Modified tt-CCA. Finally, the features extracted by CSDuDoFN and Modified tt-CCA are integrated to produce the final decoding result. Our approach leverages the strengths of both data-driven and model-driven methods and achieves state-of-the-art (SOTA) performance on three publicly available datasets, thus holding the potential to facilitate everyday BCI applications based on SSVEP. The reproducibility code is available at: https://github.com/Sungden/Zero-shot-SSVEP-classification