Instance-Based Transfer Learning With Similarity-Aware Subject Selection for Cross-Subject SSVEP-Based BCIs.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCIs) require substantial training data for high accuracy.
• Transfer learning can reduce data demands by using data from other subjects, but individual variability poses a challenge.

Purpose of the Study:

• To propose and evaluate a novel transfer learning framework (iTRCA) to address individual variability in SSVEP-BCIs.
• To develop an enhanced framework (SS-iTRCA) with subject selection to mitigate negative transfer.

Main Methods:

• Instance-based task-related component analysis (iTRCA) extracts subject-general and subject-specific features.
• Subject selection-based iTRCA (SS-iTRCA) uses similarity-based subject selection to identify optimal source subjects.
• Frameworks were evaluated on Benchmark, BETA, and self-collected datasets.

Main Results:

• iTRCA effectively leverages knowledge from source subjects while preserving target subject uniqueness.
• SS-iTRCA demonstrated improved performance by selecting appropriate source subjects.
• Both frameworks showed effectiveness in comparative evaluations.

Conclusions:

• The proposed iTRCA and SS-iTRCA frameworks offer a potential solution for high-performance SSVEP-BCIs with reduced data requirements.
• These methods effectively address individual variability in transfer learning for BCIs.