Unsupervised Adaptive Deep Learning Method For BCI Motor Imagery Decoding

TL;DR

This work tackles EEG non-stationarity in motor imagery BCIs by introducing an unsupervised online adaptation framework that keeps a pre-trained deep learning backbone frozen while continuously realigning data in both input and latent spaces. The approach combines online Euclidean Alignment with adaptive batch normalization, augmented by a warm-up data buffer, to achieve near-offline performance in cross-subject MI decoding without requiring labels or retraining. Across BNCI IIa, BNCI2, and Large datasets, the method matches or surpasses online baselines and reaches offline accuracy within roughly 10–20 trials, demonstrating strong cross-subject transfer capabilities. The work provides reproducible code and highlights practical considerations such as buffer sizing and fatigue effects, underscoring the potential for rapid, unsupervised calibration in real-time BCI applications.

Abstract

In the context of Brain-Computer Interfaces, we propose an adaptive method that reaches offline performance level while being usable online without requiring supervision. Interestingly, our method does not require retraining the model, as it consists in using a frozen efficient deep learning backbone while continuously realigning data, both at input and latent spaces, based on streaming observations. We demonstrate its efficiency for Motor Imagery brain decoding from electroencephalography data, considering challenging cross-subject scenarios. For reproducibility, we share the code of our experiments.

Figures (2)

• Figure 1: Adaptive classifier performance compared to online and offline baselines in Cross-Subject (CS) and Cross-Subject Fine-Tuning (CS+FT) evaluations.
• Figure 2: Impact of shuffling the data to test the hypothesis on user fatigue, on the BNCI dataset.