What masking geometry works best for EEG foundation models?
Pierre Guetschel, Bruno Aristimunha, Yassine El Ouahidi, Arnaud Delorme, Thomas Moreau, Michael Tangermann
Radboud University · Yneuro · University of California San Diego · IMT Atlantique · CNRS · Université Paris-Saclay
Published September 27, 2026 · on the feed September 29, 2026 · ▲ 3 · arXiv 2609.33487
The abstract
EEG foundation models hold promise for scalable brain-signal decoding across clinical and cognitive neuroscience applications, yet their pre-training pipelines remain poorly understood. Among design choices, the masking strategy is particularly critical: it determines what the network must predict and from which context. Yet it has never been ablated in isolation, as each new model bundles a new masking strategy with a new backbone and objective. In this paper, we formalize the design choices for spatio-temporal masking strategies and train various models with a single pipeline under varying masking configurations across two SSL frameworks (MAE and JEPA). We then systematically evaluate the resulting 58 pre-trained models on the 12 datasets of OpenEEGBench under a linear probe. Both frameworks agree on an optimal masking configuration and on shared failure modes. Outside these, performance is robust: 11 MAE and 9 JEPA configurations are statistically indistinguishable from the best. We further identify a novel JEPA-specific failure mode, tagged bias-inflation collapse, invisible to standard detectors. With a well-chosen mask, our pipeline reaches REVE-level downstream performance at a fraction of REVE's pre-training compute.
The Brief
2:16A short podcast summary of the paper. Written and voiced by AI.
Transcript
What masking geometry works best for EEG foundation models?. By Pierre Guetschel and five others, from Radboud University, Yneuro, and six other institutions. Published on the twenty-seventh of September. Its code repository is public and real: about thirty-five source files.
EEG foundation models promise brain-signal decoding that scales across clinical and cognitive neuroscience, but how their pre-training should be set up is still poorly understood. One choice matters especially: the masking strategy, which decides what the network has to predict and what context it gets to predict it from. Until now that choice has never been tested on its own, because each new model arrives with a new mask, a new backbone and a new objective all bundled together. The paper's claim, once the bundle is untied, is that the answer is more settled than it looks. Two quite different self-supervised frameworks agree on which masking configuration is best, and agree on how masking goes wrong.
To get there, the authors first formalize the design space of spatio-temporal masking. They then train models through one single pipeline, varying only the masking configuration, under two frameworks, MAE and JEPA. That yields fifty-eight pre-trained models, each evaluated with a linear probe on the twelve datasets of OpenEEGBench. Both frameworks land on the same optimal configuration and share the same failure modes.
Away from those failure modes, performance is robust. Eleven MAE configurations and nine JEPA configurations are statistically indistinguishable from the best one. The study also turns up something new: a failure mode specific to JEPA, which the authors tag bias-inflation collapse, and which standard detectors do not see.
The conclusions are drawn under a linear probe, across these two frameworks and the benchmarks in OpenEEGBench. Within that, there is a practical payoff. With a well-chosen mask, this single pipeline reaches downstream performance at the level of REVE, while spending only a fraction of the pre-training compute that REVE used.
There are two messages for anyone building such a model. The broad one is reassuring: many masks work about equally well, so the choice is less fragile than a new mask with every new model might suggest. The sharp one is a warning. A collapse that ordinary checks miss can sit inside a JEPA run, and only a controlled comparison like this one brings it into view.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every new model changes everything at once. This quest is about changing one thing.
You are pre-training a model on brain signals recorded from electrodes on the scalp. Like most such models, it hides part of the signal and learns to predict it from the rest.
The road ahead
- 1 · The problem ~6 min EEG models learn by hiding part of the signal, but nobody knows which way of hiding works best, because every new model changes several things at once.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Describe every EEG mask with three numbers, vary only the mask's shape in one fixed pipeline, and repeat the sweep under two frameworks.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min First put all masks on one parameter grid, then pre-train a model at each grid point, then score every model the same cheap way.
- 6 · The experiment ~6 min You’ll guess: When EEG foundation models were pre-trained under many masking configurations with both MAE and JEPA, did the two frameworks agree on the best mask?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.