Research map / Vision-language and multimodal / Human and embodied modeling
REVE: A Foundation Model for EEG - Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects
Read the paperCheck your idea against 108,967 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.
Pretrains REVE, a transformer EEG foundation model with 4D Fourier positional encoding over electrode coordinates and time, using block-masked autoencoding on ~60k hours from 92 datasets.
Problem
EEG foundation models generalize poorly across heterogeneous electrode montages, recording setups and durations, and often need full fine-tuning, especially under linear probing.
Approach
A 4D (x,y,z,time) Fourier positional encoding computed directly from electrode coordinates and patch index, so any montage and sequence length can be processed without learned position tables; combined with spatio-temporal block masking and a secondary global-token reconstruction loss in an MAE setup.
Key results
- REVE-Base reaches 0.7150 average balanced accuracy across 10 downstream tasks versus 0.6898 for CBraMod (about +2.5%).
- Under linear probing, REVE-Large averages 0.654 balanced accuracy versus 0.501 for CBraMod (no pooling), i.e., up to ~17% gains.
- Pretraining improves REVE-Base on PhysioNet-MI by about 11 points (0.5409 to 0.6480), while CBraMod gains only about 2 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add secondary loss (ablation) | REVE · Mumtaz (linear probing) | balanced accuracy: 0.818 -> 0.920 | Helped |
| add secondary loss (ablation) | REVE · Mumtaz (fine-tuning) | balanced accuracy: 0.818 -> 0.922 | Helped |
| add secondary loss (ablation) | REVE · TUAB (linear probing) | balanced accuracy: 0.797 -> 0.802 | Helped |
| add secondary loss (ablation) | REVE · TUAB (fine-tuning) | balanced accuracy: 0.803 -> 0.810 | Helped |
| add secondary loss (ablation) | REVE · ISRUC (linear probing) | balanced accuracy: 0.699 -> 0.625 | Hurt |
| add secondary loss (ablation) | REVE · ISRUC (fine-tuning) | balanced accuracy: 0.777 -> 0.770 | Hurt |
| add secondary loss (ablation) | REVE · HMC (linear probing) | balanced accuracy: 0.598 -> 0.591 | Hurt |
| add secondary loss (ablation) | REVE · HMC (fine-tuning) | balanced accuracy: 0.713 -> 0.723 | Helped |
| pretrain REVE-Base instead of training without pretraining | REVE-Base · PhysioNet-MI | balanced accuracy: 0.5409 -> 0.6480 | Helped |
| use REVE-Base rather than CBraMod | EEG classification benchmarks · Average across 9 EEG classification tasks | balanced accuracy: +2.5% | Helped |
Surprising findings
- Without pretraining, CBraMod outperforms REVE-Base by at least 8% on PhysioNet-MI.
- Model souping gave limited benefit for small models and sometimes hurt performance.
Limitations
- Requires signals of at least one second and multiples of one second.
- Pretraining corpus is dominated by North American and European data, limiting demographic diversity.
- Gains over baselines are modest on some tasks (e.g., ISRUC where CBraMod is higher on balanced accuracy).
Experimental setup
Datasets: OpenNeuro MOABB TUH Physionet PhysioNet-MI BCIC-IV-2a TUEV TUAB HMC ISRUC FACED Mumtaz
Benchmarks: PhysioNet-MI BCIC-IV-2a TUEV TUAB HMC ISRUC FACED Mumtaz MAT BCIC2020-3
Metrics: balanced accuracy Cohen's kappa weighted F1 AUROC AUC-PR
Baselines: EEGNet EEGConformer SPaRCNet ContraWR CNN-Transformer FFCL ST-Transformer BIOT LaBraM-Base CBraMod
Builds on: masked autoencoder BIOT LaBraM CBraMod Flash Attention v2
Closest papers by meaning
- Standing on the Shoulders of Giants: Rethinking EEG Foundation Model Pretraining via Multi-Teacher Distillation (NeurIPS 2026 · similarity 0.89)
- What do EEG Foundation Models Capture from Human Brain Signals? (NeurIPS 2026 · similarity 0.88)
- BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.87)
- NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.87)
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.87)
- CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding (ICLR 2025 · similarity 0.87)
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.