The Flow has not summarised this story yet — read it at arXiv Machine Learning.
The paper evaluates five publicly released EEG foundation models and an in‑house JEPA‑style model using a protocol that includes three key controls: an untrained encoder, spectral band‑power features, and a probe layer selected without test‑set look‑ahead. The study finds that selecting the probe layer by test score can modestly improve peak performance and alter the peak‑versus‑output relationship, that random initialization can match or surpass pretraining on certain tasks, and that full fine‑tuning rarely yields gains beyond seed variability. These findings lead the authors to propose four new reporting practices and to release all experimental results for re‑analysis.
MANAS-2 is a new EEG foundation model that integrates a Raw‑Band Hybrid masked autoencoder with a physics‑motivated Constrained Reconstruction (ConRec) regularizer. ConRec penalizes RMS energy differences in short temporal windows, guiding the encoder toward oscillatory‑envelope organization. Across seven held‑out EEG datasets, adding ConRec improves spectral‑power recovery (R² from 0.860 to 0.906) and band‑energy dynamics (R² from 0.283 to 0.354), while maintaining strong temporal waveform recoverability and outperforming leading EEG models on downstream tasks.
Neural State Prediction (NSP) is a latent‑predictive framework designed to curb shortcut learning in EEG foundation models. By using a target encoder updated with an exponential moving average, identity residualization, and topology‑separated context, NSP constrains both the prediction target and the available context. Trained on 2.2 million EEG segments, NSP outperforms baselines on 14 datasets in the EEG‑FM‑Bench, achieving 63.94 % macro balanced accuracy.
arXiv:2608. 01898v1 Announce Type: new Abstract: Increasing EEG pretraining data scale or model capacity does not consistently improve downstream performance.
arXiv:2607. 04851v1 Announce Type: cross Abstract: Generalizing sleep staging models to unseen datasets is challenging, and typical domain generalization (DG) methods often rely on multiple source domains or domain labels that are rarely available in practice.
BRIDGE-EEG is an efficient multi‑task EEG classification pipeline that leverages self‑supervised pretraining while dramatically reducing model size. It maps heterogeneous EEG recordings to a unified 62‑channel time‑frequency representation, pretrains an SE‑ResNet18 teacher with SimCLR, and distills it into smaller SE‑ResNet8 and SE‑ResNet4 students. The compact models achieve accuracy comparable to or better than larger foundation models on abnormality detection and emotion recognition, and they consume up to three times less energy on edge devices, enabling deployment on wearable hardware.