PaperlayerView exact revision on arXiv

Source-linked AI summary

MANAS-2: Constrained Reconstruction for EEG Foundation Models

Arvasu Kulkarni, Aditya Ray Mishra, Jeet Bandhu Lahiri, Mahir Jain, Parshva Runwal, Lakshya Saini, Siddharth Panwar, Sandeep Singh

arXiv:2609.13717v2cs.AIcs.LG

TL;DR

Masked reconstruction can underuse EEG’s oscillatory and continuous structure. MANAS-2 combines temporal and spectral reconstruction with a temporal energy constraint, improving frozen spectral recovery and downstream transfer across held-out datasets.

  • Problem

    Masked EEG reconstruction commonly focuses on raw patch values, despite EEG’s low signal-to-noise ratio, oscillatory structure, and continuity across adjacent patches.

    Source: §1 · p. 2

  • Method

    MANAS-2 combines temporal waveform and spectral-band reconstruction in RBH with ConRec, which penalizes short-window RMS-energy jumps in the temporal decoder output.

    Source: §1 · p. 2

  • Results

    Across seven held-out EEG datasets, mean six-band recovery rises from R2 = 0.860 for RBH to 0.906, alongside improved band-energy dynamics and downstream transfer.

    Source: §1 · p. 2, §5 · p. 6

  • Takeaways & Limitations

    A local energy constraint on temporal reconstruction can strengthen spectral organization without relying only on explicit spectral reconstruction targets.

    Source: §5 · p. 6

    Motor-MV was in EEGPT’s pretraining corpus, so EEGPT’s Motor results should be interpreted with this overlap in mind.

    Source: §4.4 · p. 5

Abstract

from arXiv · show

Masked reconstruction is widely used for EEG foundation models, but optimizing reconstruction on low-SNR waveforms does not necessarily produce the most useful latent representation. We introduce MANAS-2, a new EEG foundation model that combines a Raw-Band Hybrid (RBH) masked autoencoder with Constrained Reconstruction (ConRec), a physics-motivated regularizer. RBH jointly reconstructs temporal waveform patches and compact spectral-band targets, while ConRec acts only on the temporal decoder output, penalizing differences in RMS energy between adjacent short windows of the reconstructed waveform. ConRec is intended to shape the encoder by biasing it toward the organization of oscillatory-envelope information. Across seven held-out EEG datasets, adding ConRec to an otherwise identical RBH model increases 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. Applied to a temporal-only masked autoencoder, ConRec also improves frozen downstream transfer and frequency-dependent latent geometry despite receiving no spectral targets: i.e., the effects of ConRec are architecture-independent. MANAS-2 also outperforms leading EEG Foundation Models on most downstream knowledge-transfer tasks. From the effects of ConRec, we see that a physically motivated constraint imposed through the decoder can make for a more spectrally organized and transferable latent space. MANAS-2 therefore provides a new EEG foundation model built around constrained reconstruction as a mechanism for shaping representation--rather than reconstruction--quality.

1 Introduction

EEG masked reconstruction can overlook oscillatory structure and continuity, motivating MANAS-2’s hybrid targets and temporal constraint. Across held-out datasets, ConRec improves spectral information recoverable from frozen latents while MANAS-2 also performs better than most leading EEG foundation models on downstream tasks.

1 Introduction

  • Generic masked reconstruction treats raw EEG as local fragments, although EEG is low-SNR, oscillatory, and continuous across adjacent patches.Source: §1 · p. 2
  • ConRec penalizes short-window RMS-energy jumps in reconstructed waveforms, while RBH jointly reconstructs temporal patches and compact spectral-band targets.Source: §1 · p. 2
  • Mean six-band recovery rises from R2 = 0.860 for RBH to 0.906, and bandflow from 0.283 to 0.354, while temporal waveform recoverability stays high.ConRec acts on the reconstructed temporal waveform without adding a spectral target.Source: §1 · p. 2
  • MANAS-2 produces better downstream performance across seven held-out datasets than most leading EEG foundation models.Source: §1 · p. 2

2 Related work

Related work traces EEG self-supervision from contrastive learning to masked and tokenization-based approaches, and reviews spectral objectives, frozen probes, and constrained learning. MANAS-2 is positioned as studying how reconstruction constraints shape frozen representations.

2 Related work

  • EEG self-supervision spans contrastive learning (Kostas et al., 2021) through masked autoencoding and tokenization approaches, while surveys note evaluation-protocol instability (Kuruppu et al., 2026).Source: §2 · p. 2
  • Time-frequency objectives model EEG’s spectral structure, while the paper argues temporal energy-plausibility constraints can also encourage spectral-envelope information.Source: §2 · pp. 2–3
  • Frozen probes assess information in representations; ConRec is related to physics-informed learning because it adds structural conditions beyond target matching.Source: §2 · p. 3

3 Method

MANAS-2 adapts masked autoencoding to EEG with a Raw-Band Hybrid architecture and Constrained Reconstruction. RBH retains temporal waveform targets while adding spectral-band supervision, and ConRec regularizes energy changes in the temporal decoder output.

3 Method

  • MANAS-2 uses an EEG-adapted masked-autoencoding scaffold with overlapping temporal patches, spatiotemporal encodings, and temporal and spectral decoding.Source: §3 · p. 3, Figure 1 · p. 3
  • The pretraining objective combines temporal waveform reconstruction, the RBH band objective, and the ConRec RMS regularizer.Source: §3 · p. 3
  • RBH retains temporal waveform patches and adds a spectral decoder trained on masked-token log-power targets across six canonical EEG bands.The spectral targets are generated by a frozen STFT frontend and trained with an L1 loss.Source: §3 · pp. 3–4
  • ConRec compares RMS energy in windows immediately before and after regular boundaries in the temporal decoder’s reconstructed waveform.MANAS-2 places boundaries every 200 samples (1 s) and uses windows of length q = 32.Source: §3 · p. 4

4 Experiments

Across seven held-out datasets, MANAS-2 improves spectral information recoverable from frozen latents while maintaining high waveform recoverability. ConRec also strengthens frequency-distance ordering in both architectural families, and MANAS-2 gives the best balanced accuracy on most downstream datasets.

4.1 Setup and probes

  • With frozen encoders, the evaluation probes six-band power, bandflow, PAF, periodic and aperiodic spectral properties, and waveform recoverability.The spectral probes include properties not directly supervised during pretraining; waveform recovery checks whether spectral gains accompany loss of accessible time-domain information.Source: §4.1 · p. 4

4.2 Spectral information recoverability

  • Mean six-band recovery rises from R2 = 0.860 for RBH to 0.906 for MANAS-2, while bandflow rises from 0.283 to 0.354.These are mean test R2 values across seven held-out datasets.Source: Table 1 · p. 4, §4.2 · p. 4
  • PAF recovery rises from R2 = 0.135 to 0.208 and periodic PSD recovery from 0.203 to 0.254, while waveform recoverability remains 0.920 to 0.917.The aperiodic exponent changes little, from 0.788 to 0.792.Source: §4.2 · p. 4

4.3 Latent geometry

  • MANAS-2 raises frequency-distance ordering ρf from 0.792 to 0.822 for single-channel inputs and from 0.767 to 0.860 for 18-channel inputs.ConRec also improves frequency-distance ordering in the temporal-only family, with the strongest ordering observed in MANAS-2.Source: Table 2 · p. 5, §4.3 · p. 5

4.4 Frozen downstream transfer

  • MANAS-2 achieves the best balanced accuracy on most downstream datasets under frozen LP-AVG transfer, compared with six leading EEG foundation models.The protocol uses average pooling over encoder tokens and a lightweight classifier head, without encoder updates.Source: Table 3 · p. 5, §4.4 · pp. 5–6

5 Conclusion

MANAS-2 combines RBH with ConRec to improve knowledge transfer, spectral recovery, band-energy dynamics, and latent geometry while retaining high waveform recoverability. Its pretraining pool draws on unlabeled internal and public clinical EEG sources.

5 Conclusion

  • ConRec with RBH improves knowledge transfer, spectral-power recovery, band-energy dynamics, and synthetic spectral latent geometry while preserving high waveform recoverability.The conclusion identifies local energy constraints on temporal reconstruction as a means of strengthening spectral organization, not only explicit spectral targets.Source: §5 · p. 6

A Dataset description

  • Pretraining uses unlabeled EEG, and downstream evaluation follows EEG-FM-Bench tasks and splits with a frozen encoder and lightweight heads.No diagnosis, event, sleep-stage, seizure, or downstream task labels are used in the self-supervised objective.Source: A Dataset description · p. 8, §A.1 · p. 8

A.1 Pretraining datasets

  • The pretraining pool combines Internal-A and Internal-B with public clinical EEG sources, including TUH EEG and I-CARE.The two internal sources together contain 5,589 subjects and 2,981.62 hours; TUH EEG and I-CARE figures are corpus-level references, and effective training hours may differ after processing and exclusions.Source: §A.1 · p. 8, Table 4 · p. 8

A.2 Downstream datasets

The seven-dataset EEG-FM-Bench suite spans diverse clinical and behavioral tasks, with signals standardized through a common preprocessing pipeline.

A.2 Downstream datasets

  • Seven held-out EEG-FM-Bench datasets span clinical, motor, sleep, seizure, cognitive, and upper-extremity movement tasks with heterogeneous recording and class characteristics.The suite is intended to test transfer beyond the pretraining objective.Source: §A.2 · p. 8
  • ADFTD tests three-class discrimination among Alzheimer’s disease, frontotemporal dementia, and cognitively normal controls using resting-state EEG from 88 subjects.The dataset uses a 19-channel clinical montage.Source: §A.2 · p. 8
  • Motor tasks include four-class cue-based imagery on BCIC-IV 2A, four-class imagery on PhysioNet MI, and intuitive arm and hand movements in MIMUL-11.BCIC-IV 2A has 9 subjects and 22 channels; PhysioNet MI includes 109 volunteers (Schalk et al., 2009, 2004); MIMUL-11 includes 25 participants.Source: §A.2 · p. 9
  • Other tasks cover five-class sleep staging in HMC, binary seizure detection in Siena, and binary workload classification in EEGMAT.HMC uses 30 s epochs (Alvarez-Estevez and Rijsman, 2022, 2021); Siena includes expert-annotated seizure events; EEGMAT uses mental-arithmetic recordings.Source: §A.2 · p. 9

A.3 Preprocessing

  • Preprocessing resamples and filters signals, segments fixed-length windows, maps available channel labels to electrode coordinates, and stores the resulting samples.Downstream splits and label definitions follow EEG-FM-Bench, with dataset-specific channel mapping following its protocol (Xiong et al., 2026).Source: §A.3 · p. 9

B Additional Experimental Details

Models and objective ablations use a shared pretraining setup with signals resampled to 200 Hz and segmented into 10-second windows.

B Additional Experimental Details

  • Unless stated otherwise, models and objective ablations share a pretraining setup using signals resampled to 200 Hz and segmented into 10 s windows.The passage notes that key architectural and optimization settings are summarized below.Source: B Additional Experimental Details · p. 9

C Additional Downstream Transfer Protocols

MANAS-2 remains competitive or strongest across most datasets under alternative probing and fine-tuning settings. The appendix also specifies the transfer protocols and reports their results alongside a caveat about EEGPT’s Motor-MV pretraining overlap.

C Additional Downstream Transfer Protocols

  • Evaluation includes LP-AVG, LP-FLAT, FT-AVG, and FT-FLAT, with models using the same downstream settings within each protocol.LP-FLAT passes frozen tokens to a flatten-MLP head; FT-AVG and FT-FLAT fine-tune the encoder with average pooling and a flatten-MLP head, respectively.Source: C Additional Downstream Transfer… · pp. 9–10
  • Table 6 identifies the key pretraining configuration for MANAS-2 and its ablations.The supplied passage gives the table’s scope, not its configuration values.Source: Table 6 · p. 10
  • The compared foundation models are EEGPT, CSBrain, CBraMod, BIOT, REVE, and LaBraM.The passages identify these models as the comparison set but provide no per-model outcome values here.Source: C Additional Downstream Transfer… · p. 10
  • Tables 7–9 report LP-FLAT, FT-AVG, and FT-FLAT mean test scores at the best validation epoch, with variation across seeds.The tables correspond to frozen transfer under LP-FLAT and full fine-tuning under FT-AVG and FT-FLAT.Source: Table 7 · p. 10, Table 8 · p. 11, Table 9 · p. 11
  • EEGPT’s Motor results require caution because Motor-MV was included in its pretraining corpus.This overlap qualifies interpretation of the Motor comparison.Source: C Additional Downstream Transfer… · p. 10
  • Across alternative probing and fine-tuning settings, MANAS-2 remains competitive or strongest on most datasets, so its performance is not specific to average pooling.This result is stated for LP-FLAT, FT-AVG, and FT-FLAT protocols.Source: C Additional Downstream Transfer… · p. 12

D Objective-family and architecture ablations

The ablations compare ways to add spectral structure to masked EEG pretraining and test how RBH and ConRec contribute. The evaluation favors combining temporal and spectral pathways with the RMS-energy regularizer.

D.1 Naive Temporal+STFT supervision

  • Direct multi-resolution STFT supervision underperforms the temporal MAE on six of seven datasets, motivating a dedicated spectral decoder instead.TEMPORAL+STFT adds magnitude-and-phase losses at resolutions 32, 64, and 128 without a dedicated spectral decoder.Source: Table 11 · p. 12, §D.1 · p. 12

D.2 BandCube spectral tokenization

  • BANDCUBE tokenizes, masks, and reconstructs a channel × time × band spectral cube, contrasting with RBH’s temporal tokens and dedicated spectral pathway.Its macro-tokens span 12 STFT frames and one frequency band; masking and reconstruction retain the cube’s three-dimensional organization.Source: §D.2 · pp. 12–13

D.3 BandCube temporal hybrid

  • 56.9% mean balanced accuracy for BANDCUBE and 54.1% for BANDCUBE-T trail TEMPORAL-MAE’s 60.8%; adding temporal reconstruction does not recover the loss.The authors suggest pooled spectral macro-tokens may less faithfully preserve time-varying motor-imagery ERD/ERS, motivating temporal tokens with a dedicated spectral decoder.Source: Table 12 · p. 13, §D.3 · p. 13

D.4 From temporal reconstruction to RBH to MANAS-2

  • 54.5% balanced accuracy, 57.9% F1/AUROC, and 37.0% κ/AUC-PR make MANAS-2 strongest; ConRec-Temporal and RBH reach 53.2% and 53.7% mean balanced accuracy.ConRec-Temporal improves on the 51.2% temporal baseline; the combined result supports complementary contributions from RBH spectral supervision and ConRec.Source: §D.4 · p. 13, Table 13 · p. 14

E.1 Probe Methodology and Dataset Details

  • Probe targets are aligned to latent patches; bandflow uses adjacent-patch band-power differences, while PAF, periodic, and aperiodic probes use pooled four-patch windows.The six-band target comes from aligned log-magnitude STFT frames; PAF is the PSD maximum within the alpha band.Source: §E.1 · p. 14
  • BCI-IV-2a’s four-second windows yield fewer samples for estimating PAF and periodic components, making absolute recovery noisier even when relative model differences remain consistent.Each example contains only one short motor-imagery interval for these estimates.Source: §E.1 · p. 14
  • Across seven datasets, identical subject-balanced splits are paired across models and repeated over five data seeds; ridge regularization is selected by validation mean R2.The sampling caps are up to 480 training, 192 validation, and 192 test windows per dataset and seed.Source: §E.1 · p. 15

E.2 Additional methods

The appendix tests whether ConRec’s spectral-recovery effect extends beyond ridge probes. It reports that the effect persists with nonlinear recovery, with one target-level exception to statistical significance.

E.2 Additional methods

  • The analysis tests whether ConRec’s spectral-recovery effect holds across multiple objectives and probe types.The nonlinear probe complements the linear ridge probes.Source: §E.2 · p. 15
  • The MLP probe uses standardized frozen latents, one 256-unit GELU hidden layer, and fixed training settings without per-model or per-dataset tuning.Validation MSE with patience 3 selects among at most 20 epochs.Source: §E.2 · p. 15
  • ConRec’s spectral-recovery effect holds with nonlinear probes; paired dataset-level tests were below 0.05 for every target except the aperiodic exponent.The significance tests compare MANAS-2 with RBH for both ridge and MLP probes.Source: Table 15 · p. 16, §E.2 · p. 16

F Additional Latent Geometry Details

The synthetic analysis examines whether latent geometry tracks frequency when each input has a constant frequency. PCA trajectories and pairwise centroid distances provide complementary views.

F Additional Latent Geometry Details

  • Sinusoids span 0.5–40 Hz in 0.5 Hz steps with five repetitions; single-channel Cz is a complementary diagnostic, not an in-distribution pretraining input.The 18-channel condition randomizes phase across channels, and constant frequency isolates frequency-dependent organization from evolving spectral dynamics.Source: F Additional Latent Geometry Details · p. 16
  • ConRec variants show more extended, frequency-dependent PCA trajectories in both temporal-only and RBH models.The authors describe the trajectories as larger and more visibly ordered than those without ConRec.Source: Figure 2 · p. 16, F Additional Latent Geometry Details · p. 17
  • ConRec variants show a more regular progression of pairwise frequency-centroid distances, consistent with improved frequency-distance ordering.The heatmap provides a complementary view to the PCA trajectories; the supplied passages do not give the Spearman values.Source: Figure 3 · p. 17, F Additional Latent Geometry Details · p. 17
Loading 2609.13717v2…