What masking geometry works best for EEG foundation models?
A controlled masking ablation found the same best configuration across MAE and JEPA EEG pre-training.
The paper trains 58 models under one pipeline, varying spatio-temporal masking while holding the broader setup fixed. It tests them on 12 OpenEEGBench datasets with linear probes. Many settings near the top were statistically tied with the best, but both frameworks shared clear failure modes. The authors also report a JEPA-only “bias-inflation collapse” that standard detectors miss. HF Daily Papers' note
The paper trains 58 models under one pipeline, varying spatio-temporal masking while holding the broader setup fixed. It tests them on 12 OpenEEGBench datasets with linear probes. Many settings near the top were statistically tied with the best, but both frameworks shared clear failure modes. The authors also report a JEPA-only “bias-inflation collapse” that standard detectors miss. HF Daily Papers' note
score 4