Compact convolutional neural networks for classification of asynchronous steady-state visual evoked potentials.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Steady-state visual evoked potentials (SSVEPs) are brain signals measured via EEG, commonly used in brain-computer interfaces (BCIs).
• Current SSVEP decoding methods often require domain-specific knowledge and user calibration, limiting their application in asynchronous BCIs.

Purpose of the Study:

• To introduce a compact convolutional neural network (Compact-CNN) for automatic feature extraction and decoding of SSVEP signals.
• To demonstrate the efficacy of the Compact-CNN in a 12-class SSVEP dataset without user-specific calibration.

Main Methods:

• Utilized raw electroencephalogram (EEG) signals as input for the Compact-CNN.
• Trained and evaluated the Compact-CNN on a 12-class SSVEP dataset.
• Compared the Compact-CNN's performance against state-of-the-art methods like canonical correlation analysis (CCA).

Main Results:

• The Compact-CNN achieved an average accuracy of approximately 80% across subjects.
• Outperformed traditional hand-crafted approaches, including CCA and Combined-CCA.
• The model revealed extraction of phase- and amplitude-related features inherent to the SSVEP signals.

Conclusions:

• The Compact-CNN offers a robust and calibration-free method for SSVEP decoding, suitable for asynchronous BCI applications.
• This deep learning approach enhances understanding of visual cortex processing and has potential for advanced BCI development.