Probing EEG Foundation Model Attention Maps for Lightweight Interpretable Emotion Recognition

Using REVE’s first-layer self-attention on SEED-IV, a single-head attention-derived weighted FC matrix enables competitive emotion recognition with balanced accuracy and F1 while cutting computation by at least two orders of magnitude.

• Why it matters: Foundation models for EEG emotion recognition are costly and have not reliably beaten task-specific networks, and their internal representations remain hard to interpret. This work targets lightweight, interpretable probing without fine-tuning.
• What they did: They aggregate raw first-layer attention weights from one head across full SEED-IV trials into weighted channel-to-channel adjacency matrices, then train classical classifiers on the most informative head selected by a label-free symmetry criterion. No fine-tuning of REVE is used.
• The result: The resulting linear probing achieves balanced accuracy and F1 scores competitive with several fully fine-tuned EEG foundation models and uses at least two orders of magnitude fewer calculations than standard encoder use. Frontal connections are over-represented among top discriminative edges, aligning with affective-computing findings.

AI summary of the abstract · may contain errors · not medical advice · read the paper.