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

2026-08-07

AI Summary

This work addresses the quadratic computational overhead and redundancy in EEG foundation models caused by long input sequences and low signal-to-noise ratios. To this end, the authors propose ZIPBrain, a training-free, plug-and-play perceptual redundancy-aware token pooling module. ZIPBrain partitions EEG tokens into redundant and distinctive groups, then merges each redundant token with its most similar distinctive counterpart, substantially compressing sequence length. The method seamlessly integrates into standard Transformer encoders without requiring fine-tuning and leverages CUDA Graph acceleration for efficient inference. Evaluated across multiple EEG foundation models, ZIPBrain consistently improves average accuracy by 1.3%–10.5% while reducing inference time by 32.7% on average—up to 41.8%—achieving a favorable trade-off between efficiency and performance.

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.

Problem

• EEG foundation models
• local deployment
• computational efficiency
• token redundancy
• real-time monitoring

Innovation

• EEG foundation models
• token compression
• redundancy-aware pooling
• efficient inference
• plug-and-play module