BrainMoE: Towards Universal EEG Foundation Models with Channel-wise Mixture-of-Experts

Fan Ma, Mingyang Jiang, Lingfei Qian, Zhiling Gu, Hua Xu

Keywords: EEG, Foundation Models, MoE

Abstract:

Electroencephalography (EEG) is pivotal for brain-computer interface (BCI) and healthcare applications, with foundation models offering a promising paradigm for generalized decoding. However, current models typically apply a uniform strategy across all channels, overlooking the brain's inherent functional heterogeneity. This one-size-fits-all approach limits model adaptability and performance, as different EEG tasks activate distinct neural pathways.
To address this, we introduce BrainMoE, a novel universal EEG foundation model featuring a channel-wise Mixture-of-Experts (MoE) architecture. The core innovation is the dynamic assignment of specialized experts to different channels for more nuanced decoding. BrainMoE first employs a ChannelFormer to capture distinct representations for each channel. A router then uses these representations to intelligently select and weight the most relevant experts. This mechanism allows BrainMoE to tailor its decoding strategy on a per-channel basis, aligning its computation with the varied spatial demands of diverse EEG tasks.
Comprehensive experiments on seven downstream tasks across nine benchmarks demonstrate that BrainMoE achieves significant performance gains, setting a new state-of-the-art on six datasets. These results validate that channel-wise specialization is a critical step towards more powerful and truly universal EEG foundation models, showcasing the robust capability and generalizability of our approach.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 3132