Reading emotion from EEG works reasonably well until you change something. A different person, a different dataset, a different recording setup, and performance falls away. The usual fix is domain adaptation, which nudges a model trained on one population toward a new one.

Conventional methods for this need the original training data on hand, and in this field that is often impossible. EEG recordings are medical data and privacy rules mean the source dataset frequently cannot be shared, only the model trained on it. Source-free adaptation exists for exactly that situation, adapting with the model alone.

It brings its own difficulties, which the authors name plainly: large domain gaps, noisy pseudo-labels, and unstable adaptation. Pseudo-labels are the fragile part, since the model is being asked to teach itself using its own guesses about unfamiliar data, and early mistakes reinforce themselves. This work targets that instability directly.