Controlled Evaluation of EEG Foundation Models: Random Initialisation, Band-Power and Validation-Selected Probes
In brief
A controlled evaluation of five publicly released EEG foundation models shows that, with the probed layer selected on validation data, random initialisation matches or exceeds pretraining for half of the models on sleep and emotion tasks. The study covers five benchmarks, four added task families and several random seeds, including an in-house JEPA-style model; no model beats a five-band log-power baseline on emotion or motor imagery.
Why it matters
Untrained encoders, spectral baselines and validation-selected probes let readers separate pretraining gains from layer selection and baseline choice.
