Improving Generalized Zero-Shot Learning SSVEP Classification Performance From Data-Efficient Perspective.

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Generalized zero-shot learning (GZSL) reduces training data needs for steady-state visual evoked potential (SSVEP) brain-computer interfaces (BCIs).
• Existing GZSL methods suffer from inefficient SSVEP data utilization, limiting accuracy and information transfer rate (ITR).
• Traditional BCIs often overlook inter-subject variance in visual latency, using fixed sampling starting times.

Purpose of the Study:

• To propose a novel framework for more effective SSVEP data utilization in GZSL-BCIs.
• To enhance accuracy and ITR by optimizing data acquisition, feature extraction, and decision-making.
• To address limitations of current GZSL approaches in SSVEP-based BCIs.

Main Methods:

• Introduced a dynamic sampling starting time (DSST) strategy using subject-specific optimal sampling starting times (OSST).
• Developed a Transformer-based network to capture global input data information and compensate for limited receptive fields.
• Designed a classifier selection strategy for optimal classifier choice for seen and unseen classes, integrated via a proposed training procedure.

Main Results:

• The proposed framework demonstrated superior performance compared to state-of-the-art (SOTA) methods on three public datasets.
• Outperformed representative methods that require complete training data for all classes.
• Achieved significant improvements in accuracy and ITR for GZSL-BCIs.

Conclusions:

• The developed framework effectively enhances SSVEP data utilization across acquisition, feature extraction, and decision-making levels.
• This approach significantly improves GZSL performance in SSVEP-based BCIs, even outperforming methods with full training data.
• The study offers a promising direction for developing more efficient and accurate brain-computer interfaces.