A Generalized Zero-Shot Learning Scheme for SSVEP-Based BCI System.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Steady-state visual evoked potential (SSVEP) based brain-computer interfaces (BCIs) are widely used.
• Current high-accuracy SSVEP systems require extensive calibration data for each target, limiting practical application.

Purpose of the Study:

• To develop an SSVEP classification method that achieves high accuracy using training data from only a subset of targets.
• To reduce the calibration time and data requirements for SSVEP-based BCIs.

Main Methods:

• Proposed a generalized zero-shot learning (GZSL) scheme for SSVEP classification.
• Utilized convolutional neural networks (CNNs) to embed electroencephalogram (EEG) data and sine waves into a shared latent space.
• Classification was performed using the correlation coefficient between embedded EEG data and target sine waves.

Main Results:

• The GZSL scheme achieved 89.9% of the accuracy of state-of-the-art (SOTA) data-driven methods that require all target training data.
• Demonstrated a multifold improvement compared to SOTA training-free methods.
• Successfully classified SSVEP signals from both seen and unseen targets during testing.

Conclusions:

• The proposed GZSL approach is a promising method for developing SSVEP classification systems with reduced calibration requirements.
• This technique enables high classification accuracy without needing training data for all potential targets.
• The findings suggest a more efficient and user-friendly approach to SSVEP-based BCIs.