Domain Generalization for Zero-calibration BCIs with Knowledge Distillation-based Phase Invariant Feature Extraction

TL;DR

The paper tackles EEG-based BCI generalization under distribution shifts by introducing KnIFE, a zero-calibration domain generalization framework that learns both intra-domain and inter-domain invariant features. It leverages knowledge distillation to extract Fourier phase-invariant features and employs Fourier-based spectral transfer to augment phase information, while using correlation alignment to bridge domain gaps. The approach yields state-of-the-art performance on OpenBMI and BCICIV-2 datasets, with ablations showing the necessity of intra-domain phase distillation, spectral transfer, and cross-domain alignment. This work suggests phase information as a robust intra-domain invariant feature source, with practical potential for real-world zero-calibration BCIs. The methodology offers a versatile blueprint for EEG DG that could extend to other time-series modalities facing domain shifts.

Abstract

The distribution shift of electroencephalography (EEG) data causes poor generalization of braincomputer interfaces (BCIs) in unseen domains. Some methods try to tackle this challenge by collecting a portion of user data for calibration. However, it is time-consuming, mentally fatiguing, and user-unfriendly. To achieve zerocalibration BCIs, most studies employ domain generalization (DG) techniques to learn invariant features across different domains in the training set. However, they fail to fully explore invariant features within the same domain, leading to limited performance. In this paper, we present an novel method to learn domain-invariant features from both interdomain and intra-domain perspectives. For intra-domain invariant features, we propose a knowledge distillation framework to extract EEG phase-invariant features within one domain. As for inter-domain invariant features, correlation alignment is used to bridge distribution gaps across multiple domains. Experimental results on three public datasets validate the effectiveness of our method, showcasing stateof-the-art performance. To the best of our knowledge, this is the first domain generalization study that exploit Fourier phase information as an intra-domain invariant feature to facilitate EEG generalization. More importantly, the zerocalibration BCI based on inter- and intra-domain invariant features has significant potential to advance the practical applications of BCIs in real world.

Figures (9)

• Figure 1: Comparison of conventional method, domain generalization and our proposed method. (a) Conventional methods train on source domain data and calibrate with a portion of target domain data to bridge domain distribution gaps. (b) Domain generalization exclusively relies on source domain data to learn domain-invariant features and generalize to unseen domains. (c) Our approach extends this by considering diverse domain-invariant features, learning them from both intra- and inter-domain viewpoints.
• Figure 2: The detailed framework of KnIFE. Firstly, the teacher network is responsible for learning the phase information of the raw data and transferring this knowledge to the student network. This process enables the student network to effectively extract intra-domain invariant features. To facilitate this, Fourier-based spectral transfer techniques are employed to emphasize phase information. Additionally, the student network learns inter-domain invariant features by aligning the distribution between any pair of domains. During the model inference stage, the new subjects' data can be directly fed into the pre-trained student network model for prediction.
• Figure 3: Diagram of spectral transfer. $\mathcal{D}_v$ and $\mathcal{D}_u$ represent the EEG data of two subjects. FFT and iFFT represent the fast Fourier transform and the inverse transform, respectively.
• Figure 4: Loss and accuracy vary with epoch during training. All losses gradually decrease and converge during the training process.
• Figure 5: Ablation study of different components on KnIFE.