Beyond Accuracy: Robustness, Interpretability and Expressiveness of EEG Foundation Models

In three linesComparative study of 6 EEG foundation models across 8 datasets beyond clean accuracy. Robustness analysis (noise, channel dropout), interpretability via Attention-Aware Layer-Wise Relevance Propagation, and expressiveness through block-wise probing. Findings: no single model dominates all failure modes; models focus on task-appropriate brain regions but decode corrupted content poorly.

Summary generated by Claude — human-verified