Reproduce “Towards Brain Passage Retrieval -- An Investigation of EEG Query Representations” on OpenResearch

About this paper

Niall McGuire, Yashar Moshfeghi

The BPR (Brain Passage Retrieval) framework enables direct retrieval of text passages using electroencephalography (EEG) signals, bypassing the need for intermediate textual query translation. This approach achieved a Precision@5 of 12.39% for EEG-based queries, outperforming existing EEG-to-text baselines by 8.81% while demonstrating robust cross-subject generalization across 30 participants.

Key insights

• Direct, end-to-end learning between EEG signals and text passage representations is substantially more effective for brain-based information retrieval than intermediate EEG-to-text translation approaches.
• Subject-aware negative sampling is critical for robust cross-modal alignment, as it helps the model learn generalizable semantic relationships rather than subject-specific neural patterns.
• The `[CLS]` token within the Transformer architecture effectively captures sequence-level semantic information from EEG signals, outperforming other pooling strategies.

Summary by alphaXiv.