New brain foundation model B[FM]^2 outpaces EEG benchmarks with 30x less data

B[FM]^2 (Brain Foundation Model via Flow Matching) represents a paradigm shift in EEG pretraining. Unlike existing models that fragment continuous multi-channel signals into patches or tokens and rely on masked self-supervision, B[FM]^2 operates directly on raw waveforms using flow matching—a continuous-time generative framework. This preserves fine-grained temporal dynamics and brain rhythms that discretization typically obscures. To tackle the architectural asymmetry between the densely sampled time axis (thousands of timepoints) and the short electrode axis (tens of channels), the team introduced SplitUNet, a velocity network that factorizes each block into separate 1D temporal and 1D electrode convolutions, downsampling only along time to maintain electrode topology throughout the hierarchy.

The results are striking: B[FM]^2 sets new state-of-the-art performance on 7 out of 9 standard EEG classification tasks (e.g., sleep staging, motor imagery) using a pretraining budget of just 36,895 segments (~307 hours)—roughly 30 times less data than required by leading EEG foundation models. In a further validation of its generative capabilities, synthetic EEGs produced by B[FM]^2 were indistinguishable from real brain recordings by two board-certified neurologists (Cohen’s κ = -0.096, indicating no agreement on which data was synthetic). This breakthrough suggests that the flow matching approach, combined with the SplitUNet design, could dramatically lower the barrier for building generalizable EEG models across clinical and brain-computer interface applications, without requiring massive proprietary datasets.