Research map / Probabilistic and scientific modeling / Molecules and proteins

B[FM]$^2$: Brain Foundation Model via Flow Matching with SplitUNet

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Flow matching generative modeling

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 a flow-matching velocity-network UNet (SplitUNet, factorized temporal and electrode 1D convs) directly on raw 19-channel TUEG EEG, achieving SOTA on 7 of 9 downstream EEG tasks with ~307 h of pretraining data and generating EEG that neurologists could not distinguish from real.

Problem

Existing EEG foundation models discretize continuous multi-channel EEG into patches or codebook tokens and use masked self-supervision, which fragments continuous brain rhythms; flow matching on raw EEG requires an architecture that handles the long time axis versus the short electrode axis.

Approach

Pretrain directly on the raw multi-channel waveform with OT-CFM flow matching, using a UNet velocity network whose 2D convolutions are factorized into a temporal 1D conv followed by an electrode 1D conv, with downsampling only along time so electrode topology is preserved at every layer.

Key results

• B[FM]^2 reaches suite-averaged balanced accuracy 0.754 and sets SOTA on 7 of 9 downstream EEG tasks, e.g. 0.840 on MAT vs 0.766 for REVE.
• On BCIC-IV-2a, B[FM]^2 gets 0.570 vs REVE's 0.640, attributed to REVE pretraining on motor imagery datasets.
• Two board-certified neurologists rated B[FM]^2 samples as slightly more real than real TUEG segments (2.46 vs 3.02 mean Likert), with Cohen's kappa -0.096.

Surprising findings

• Blinded neurologists rated generated EEG as slightly more realistic than real held-out segments.
• UNet-2D did not improve average downstream accuracy over the 1D joint-electrode variant, despite full spatial cross-mixing.

Limitations

• Pretraining uses only TUEG, a clinical corpus, which likely explains the gap on motor imagery (BCIC-IV-2a).
• Reported baseline numbers are taken from prior papers rather than re-run; B[FM]^2 is worse than CBraMod/CSBrain-level on the TUAB and AUC-PR for some tasks (e.g., Siena AUC-PR 0.202 vs 0.487 for CSBrain).

Experimental setup

Datasets: TUEG Mumtaz MAT Siena ISRUC HMC TUEV TUAB BCIC-IV-2a SEED-V

Benchmarks: Mumtaz MAT Siena ISRUC HMC TUEV TUAB BCIC-IV-2a SEED-V

Metrics: balanced accuracy Cohen's kappa weighted F1 AUC-PR AUROC Wasserstein-1 band-power distance PSNR SSIM Likert realness rating

Baselines: EEGNet EEGConformer SPaRCNet ContraWR CNN-Transformer FFCL ST-Transformer BIOT LaBraM-Base CBraMod REVE CSBrain

Builds on: flow matching OT-CFM DDPM UNet R(2+1)D CBraMod LaBraM

Closest papers by meaning

• Standing on the Shoulders of Giants: Rethinking EEG Foundation Model Pretraining via Multi-Teacher Distillation (NeurIPS 2026 · similarity 0.89)
• BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.88)
• NeurIPT: Foundation Model for Neural Interfaces (NeurIPS 2025 · similarity 0.88)
• What do EEG Foundation Models Capture from Human Brain Signals? (NeurIPS 2026 · similarity 0.88)
• CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding (ICLR 2025 · similarity 0.87)
• NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.87)