Research map / Video / Video segmentation and tracking
MAGE: Towards Generalizable Multi-timescale EEG Representations
Tuan-Kiet Doan, Van Chien Nguyen, Trung-Hieu Tran, Quang Hung Pham, Ngoc-Son Vu, Huy Phan, Phi Le Nguyen, Duc-Han LE, Nikola Simidjievski, Van-Tam Nguyen
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MAGE, an EEG representation approach that captures multi-timescale neural dynamics instead of relying on a fixed temporal scale in EEG foundation models.
Problem
Current EEG foundation models rely on a fixed temporal scale, which limits their ability to capture neural dynamics that span multiple timescales.
Approach
Moves beyond fixed-scale EEG encoding to learn multi-timescale representations intended to generalize across datasets and tasks (details beyond the abstract not available).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use multi-timescale tokens with adaptive aggregator instead of fixed-patch foundation models | MAGE · BCIC-IV-2a, PhysioNet-MI, FACED, TUAB, KaggleERN | balanced accuracy: e.g. PhysioNet-MI 59.96 -> 62.31 vs CSBrain | Improved |
| use multi-timescale model on Sleep-EDFx and SHU-MI | MAGE · Sleep-EDFx, SHU-MI | balanced accuracy: Sleep-EDFx 70.39 (CBraMod) vs 69.98; SHU-MI 63.22 (LaBram) vs 62.31 | No clear effect |
| MAGE with fewer parameters | MAGE · seven EEG datasets | average rank: table: 1.81 vs 2.71 (CSBrain); text states 1.43 vs 2.81 | Mixed |
| combine all three timescales instead of a single timescale (ablation) | MAGE · FACED | balanced accuracy: 46.65 (1s only) -> 51.30 | Improved |
| combine all three timescales instead of a single timescale (ablation) | MAGE · Sleep-EDFx | balanced accuracy: 67.99 (1s only) -> 70.24 | Improved |
| route to only the top-weighted timescale instead of weighted sum (ablation) | MAGE-router · FACED, PhysioNet-MI, Sleep-EDFx, SHU-MI | balanced accuracy: FACED 51.30 -> 49.87; PhysioNet-MI 61.15 -> 58.83 | Hurt |
| train aggregator downstream instead of freezing (ablation) | MAGE · Sleep-EDFx | balanced accuracy: 66.16 -> 70.24 | Improved |
| train aggregator downstream instead of freezing (ablation) | MAGE · PhysioNet-MI | balanced accuracy: 61.61 (frozen) vs 61.15 (trained) | Hurt |
| add Gaussian noise to aggregator logits during pretraining (ablation) | MAGE · Sleep-EDFx | balanced accuracy: 68.79 -> 70.24 | Improved |
| add Gaussian noise to aggregator logits during pretraining (ablation) | MAGE · PhysioNet-MI | balanced accuracy: 61.57 -> 61.15 | Hurt |
| learned aggregator instead of uniform weights (ablation) | MAGE · Sleep-EDFx | balanced accuracy: 70.00 -> 70.24 | Improved |
| aggregator instead of Medformer inter attention (ablation) | MAGE · SHU-MI | balanced accuracy: 63.52 -> 62.43 | Hurt |
Closest papers by meaning
- BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.90)
- Brain-OF: An Omnifunctional Foundation Model for fMRI, EEG and MEG (NeurIPS 2026 · similarity 0.88)
- NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.87)
- EEG2Video: Towards Decoding Dynamic Visual Perception from EEG Signals (NeurIPS 2024 · similarity 0.87)
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.87)
- REVE: A Foundation Model for EEG - Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
