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

Researchers introduce B[FM]², a brain foundation model using flow matching on raw EEG signals without discretization, paired with SplitUNet architecture to handle the asymmetry between time and electrode dimensions. The approach achieves state-of-the-art results on 7 of 9 EEG classification tasks while requiring 30x less pretraining data than existing models and generates synthetic EEGs indistinguishable from real brain data.

B[FM]² represents a significant methodological shift in how foundation models approach neuroscience data. Rather than following the dominant paradigm of discretizing continuous EEG signals into patches or tokens before transformer processing, this work operates directly on raw waveforms using continuous-time flow matching. This architectural choice reflects a deeper understanding of EEG's fundamental properties—continuous brain rhythms cannot be faithfully represented through discrete tokens without information loss.

The innovation addresses a critical technical challenge in EEG processing: the extreme asymmetry between the temporal axis (thousands of densely sampled timepoints with high autocorrelation) and the spatial axis (tens of electrodes at fixed scalp positions). SplitUNet's factorized approach—separating temporal and electrode convolutions while downsampling only along time—directly solves this asymmetry rather than forcing EEG into architectures designed for other modalities.

The practical implications are substantial. Achieving superior performance with merely 307 hours of pretraining data (versus typical requirements of 9,000+ hours) dramatically reduces computational barriers to entry for EEG foundation model development. This democratization could accelerate clinical and brain-computer interface applications where labeled data remains scarce.

The generation of neurologically indistinguishable synthetic EEGs suggests the model captures genuine brain dynamics rather than surface-level statistical patterns. This capability opens possibilities for data augmentation in clinical settings where patient data is limited. The work establishes that continuous-signal processing, when paired with architecturally appropriate inductive biases, outperforms discretization-based approaches in capturing biological data.

• →B[FM]² achieves state-of-the-art performance on 7 of 9 EEG benchmarks using 30x less pretraining data than existing foundation models
• →Flow matching on raw signals without tokenization preserves continuous brain rhythms better than discretization approaches
• →SplitUNet architecture factorizes temporal and electrode dimensions separately, solving the time-electrode asymmetry problem
• →Generated synthetic EEGs achieve neurologist-level indistinguishability from real brain data, validating model quality
• →Reduced computational requirements lower barriers to developing specialized EEG foundation models for clinical and BCI applications