Masked reconstruction is widely used in EEG foundation models, yet optimizing reconstruction on low‑SNR waveforms does not guarantee the most useful latent representation. We introduce MANAS‑2, which couples a Raw‑Band Hybrid (RBH) masked autoencoder with a physics‑motivated Constrained Reconstruction (ConRec).
RBH jointly reconstructs temporal waveform patches and compact spectral‑band targets. ConRec operates only on the temporal decoder output, penalizing the RMS energy difference between adjacent short windows, i.e., $\Delta E_{\text{RMS}} = |E_{\text{RMS}}^{(i)} - E_{\text{RMS}}^{(i+1)}|$, thereby biasing the encoder toward organizing oscillatory‑envelope information.
Across seven held‑out EEG datasets, adding ConRec to an otherwise identical RBH model raises frozen ridge recovery of six‑band spectral power from mean $R^2=0.860$ to $0.906$ and recovery of inter‑patch band‑energy dynamics from $R^2=0.283$ to $0.354$, while temporal waveform information remains highly recoverable from the frozen latents. When applied to a temporal‑only masked autoencoder, ConRec also improves frozen downstream transfer and frequency‑dependent latent geometry, demonstrating that its benefits are architecture‑independent.
MANAS‑2 outperforms leading EEG foundation models on most downstream knowledge‑transfer tasks. The effects of ConRec show that a physically motivated constraint imposed through the decoder can produce a more spectrally organized and transferable latent space, making constrained reconstruction a mechanism for shaping representation rather than merely improving reconstruction quality.
Review