Reproduce “Domain Generalization for Zero-calibration BCIs with Knowledge Distillation-based Phase Invariant Feature Extraction” on OpenResearch

About this paper

Zilin Liang, Zheng Zheng, Weihai Chen, Xinzhi Ma, Zhongcai Pei, Xiantao Sun

Researchers developed KnIFE, a framework for zero-calibration Brain-Computer Interfaces that extracts both intra-domain phase-invariant features using knowledge distillation and inter-domain invariant features through correlation alignment. The method achieved state-of-the-art average accuracies of 71.37% on BCICIV-2a, 82.53% on BCICIV-2b, and 81.51% on OpenBMI datasets, enhancing generalization performance across unseen subjects and sessions.

Key insights

• Fourier phase information in EEG signals is more robust to distribution shifts across subjects and sessions than amplitude information, serving as a valuable intra-domain invariant feature.
• Combining the learning of both intra-domain (phase-specific) and inter-domain invariant features is crucial for achieving superior generalization performance in zero-calibration BCIs.
• Knowledge distillation effectively transfers high-level, phase-invariant semantic information from noisy and dynamic EEG signals to a student model.

Summary by alphaXiv.