MANAS-2: Constrained Reconstruction for EEG Foundation Models

Details

EEG foundation models typically use masked autoencoder pretraining on raw waveforms. The problem: minimizing reconstruction loss on low signal-to-noise EEG does not produce the most useful latent representations for downstream tasks like seizure detection or sleep staging. MANAS-2 introduces two contributions. First, a Raw-Band Hybrid (RBH) masked autoencoder that reconstructs both temporal waveform patches and compact spectral-band targets jointly. Second, ConRec (Constrained Reconstruction), a physics-motivated regularizer that penalizes differences in RMS energy between adjacent short windows of the reconstructed temporal decoder output — biasing the encoder toward capturing oscillatory envelope information. The empirical results are clear: adding ConRec to an otherwise identical RBH model improves frozen ridge recovery of six-band spectral power from mean R²=0.860 to 0.906, and improves inter-patch band-energy dynamics recovery from R²=0.283 to 0.354, while temporal waveform recovery from frozen latents remains high. Critically, ConRec also improves a temporal-only masked autoencoder (no spectral targets), confirming the effect is architecture-independent. MANAS-2 outperforms leading EEG foundation models on most of seven held-out downstream transfer datasets. The technique generalizes: physics-motivated constraints imposed through the decoder can shape representation quality even when the encoder receives no explicit spectral supervision.