ZIPBrain Accelerates EEG Foundation Models for Local Deployment

Electroencephalograph (EEG) foundation models (EFMs) are gaining traction for their general-purpose representations, but their computational demands, which grow quadratically with input length, hinder real-time deployment on resource-constrained devices, particularly for clinical monitoring. Recognizing that EEG's low signal-to-noise ratio (SNR) suggests much of its data is redundant, this research introduces ZIPBrain.ZIPBrain is a novel, redundancy-aware EEG token pooling module. It operates by partitioning token sequences into redundant and unique groups, then merging each redundant token with its most similar counterpart in the unique group. This process effectively reduces the token count without significant loss of information.Crucially, ZIPBrain is a training-free, plug-and-play module that seamlessly integrates into standard Transformer encoders with negligible computational overhead. Extensive experiments across multiple EEG foundation models demonstrate its versatility, achieving average accuracy improvements of 1.3% to 10.5% over baselines, while reducing wall-clock inference time by 32.7% (up to 41.8% with CUDA Graph) compared to original EFMs.