ReCIL: Rehearsal-Based Class Incremental Learning for Cross-Subject Motor Imagery Classification

Electroencephalography (EEG) based motor imagery (MI) is a classic brain-computer interface (BCI) paradigm, which can be used in neuro-rehabilitation and device control. Existing MI classification approaches only consider the scenario that the number of MI classes is pre-determined and fixed; when new MI classes arrive, data for all classes must be recollected and the model retrained, which is inefficient, and sometimes impossible, as the training data for previous classes may be lost or inaccessible due to privacy concern.

This paper proposes rehearsal-based class incremental learning (ReCIL) for cross-subject MI classification, where the model learns new MI classes sequentially, and the test subjects are different from the training subjects. ReCIL uses Euclidean Alignment to reduce the inter-subject EEG data distribution shift, global-local replay to preserve the knowledge of old tasks, and dimensionality reduction to facilitate reliable similarity computation among EEG samples.

Experiments on three public MI datasets showed that ReCIL achieved a good balance between plasticity and stability.

To our knowledge, this is the first study on cross-subject class incremental learning for MI classification, offering a practical solution for incremental class expansion without retraining from scratch.