ZIPBrain: Can EEG Foundation Models Be Faster, Locally Deployable, but Accurate?

arXiv:2608.07033v1 Announce Type: new Abstract: This work investigates whether Electroencephalograph (EEG) foundation models (EFMs) can be made faster and locally deployable without sacrificing accuracy. EEG foundation models are a major trend, offering strong general-purpose representations. However, their computational burden grows quadratically with input length, hindering deployment on resource-constrained scenario, particularly for real-time clinical monitoring. EEG's low SNR further suggests many of these tokens are redundant and compressible with little accuracy cost. We propose ZIPBrain, a novel redundancy-aware EEG token pooling module that leverages this low-SNR characteristic to reduce token count. Given a token sequence, ZIPBrain partitions tokens into redundant and unique groups, then merges each redundant token with its most similar counterpart in the unique group. Furthermore, ZIPBrain serves as 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 show ZIPBrain's strong versatility, achieving 1.3%-10.5% average improvement over baselines, while reducing wall-clock inference time by 32.7% (up to 41.8% with CUDA Graph) compared to the original EEG foundation models.

Related

• TAP: Two-Stage Adaptive Personalization of Multi-Task and Multi-Modal Foundation Models in Federated Learning
• When NPUs Are Not Always Faster: A Stage-Level Analysis of Mobile LLM Inference
• Self-supervision drives representational convergence in medical foundation models more than clinical supervision

Source: arXiv cs.AI | 2026-08-10