Loading…
EEG foundation model whose masked autoencoder adds an RMS-energy constraint on the decoder, sharpening spectral structure in frozen latents.
No providers recorded yet. Browse all providers
EEG is a low signal-to-noise measurement, which makes the masked-autoencoder recipe that EEG foundation models borrowed from vision an awkward fit. Mask a patch of raw waveform, score the reconstruction pointwise, and the objective rewards predicting whatever varies most — muscle artifact, electrode drift, line noise. What clinicians and brain-computer interface engineers actually read is oscillatory: power in the canonical delta through gamma bands, and how that power's envelope evolves second to second. A model can reconstruct waveforms well while organizing its latent space around something else.
MANAS-2, introduced by Mannas AI in a September 2026 preprint, addresses that mismatch in the loss function rather than the encoder. Its central idea, Constrained Reconstruction (ConRec), adds no new prediction target. It penalizes the reconstructed time-domain waveform: at each one-second boundary, the root-mean-square energy over the 32 samples immediately before must not jump relative to the 32 samples immediately after. Real EEG envelopes do not change energy discontinuously at arbitrary boundaries, so this is a physical-plausibility constraint, and acting through the decoder it back-propagates into the shared encoder — shaping how information is organized rather than what is predicted.
ConRec sits on a Raw-Band Hybrid (RBH) autoencoder, which keeps temporal waveform patches as the primary target and adds a second decoder predicting compact log-power targets over six bands on the same patch grid. MANAS-2 is the second EEG model to carry the MANAS name, following MANAS-1, which the company released on HuggingFace in February 2026 under its NeuroDX brand; the preprint benchmarks against six other EEG foundation models — CBraMod, CSBrain, LaBraM, EEGPT, BIOT and REVE — rather than against its namesake.
The encoder is a 22-layer transformer of width 512 with 8 attention heads; each decoder is 4 layers. EEG is resampled to 200 Hz and cut into 10-second windows tokenized as one-second patches with 0.1 s overlap, carrying spatiotemporal positional encodings adapted from REVE, with 55% of tokens block-masked. Training uses AdamW at learning rate 2.4e-4 with an effective batch of 2,048 in bfloat16; ConRec places boundaries every 200 samples, compares 32-sample RMS windows under a Smooth-L1 penalty, and is weighted at 3.0 against a band-loss weight of 0.5. Pretraining is label-free, mixing an internal clinical corpus of 5,589 subjects and roughly 2,981 hours with the Temple University Hospital EEG Corpus (over 15,000 subjects, about 25,000 hours) and I-CARE (600 post-cardiac-arrest patients, about 33,000 hours).
Evaluation uses the seven EEG-FM-Bench datasets, spanning dementia, motor imagery, sleep staging, seizure detection, workload and movement. Adding ConRec to an otherwise identical RBH model lifts frozen ridge recovery of six-band power from mean R² = 0.860 to 0.906 and of inter-patch band-energy dynamics from 0.283 to 0.354, with peak-alpha-frequency recovery rising from 0.135 to 0.208. The aperiodic exponent barely moves (0.788 to 0.792) and waveform recoverability holds at 0.917, so the spectral gains do not come from discarding time-domain information. Under a frozen average-pooling probe, MANAS-2 records the best balanced accuracy on five of the seven datasets; EEGPT leads on the two exceptions, one of which overlaps its pretraining corpus.
The model targets settings where labeled EEG is scarce but raw recordings are plentiful: screening resting-state clinical EEG for dementia, scoring overnight polysomnography, flagging seizures in long-term monitoring, and decoding motor imagery for brain-computer interfaces. Because the protocol freezes the encoder and trains only a ridge probe or a small head, a group with a few dozen labeled subjects can fit a decoder without the compute that fine-tuning a transformer demands.
ConRec is the contribution most likely to outlast the benchmark table. It reframes EEG pretraining around a distinction — what the decoder is asked to reconstruct versus what constraints that reconstruction must satisfy — and shows that an auxiliary physical constraint carrying no spectral target can reorganize spectral information in the encoder. That the effect survives transplanting into a temporal-only autoencoder suggests the mechanism is portable to other EEG architectures. The work is an unreviewed preprint, and no code, weights, model card or data card have been released, while the internal clinical corpus supplying its subject diversity is not public, so the frozen-probe results cannot currently be reproduced outside the group.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.