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

The article presents B[FM]$^2$, a novel EEG foundation model that utilizes continuous-time flow matching to learn from raw EEG signals without discretization, tokenization, or masking. It introduces SplitUNet, an architecture that separates 1D temporal and electrode convolutions to address the challenges posed by the dense sampling of time and limited electrode channels. B[FM]$^2$ achieves state-of-the-art performance on 7 out of 9 standard EEG classification tasks with significantly reduced pretraining requirements, demonstrating its potential for efficient transfer learning in clinical and brain-computer interface applications.