acceptodds
Under review as a conference paper at ICLR 2027

From Static Descriptions to State Dynamics: Multi-Scale Pretraining for EEG Transfer

Abstract

Electroencephalography (EEG) is a temporal signal whose task-relevant structure spans transient events and sustained rhythms. Conventional temporal models learn these features through task-specific supervision, while EEG foundation models seek broader transfer through large-scale unlabelled pretraining. Many use masked content recovery, which can exploit temporal continuity and spatial redundancy without requiring predictions of subsequent activity. Autoregressive EEG approaches introduce future prediction, but typically supervise the next token or chunk without explicitly coupling fine-resolution and longer-range forecasts from the same state. We introduce MSDE, which jointly predicts future teacher states at micro and macro scales through a causal encoder and a shared transition operator conditioned on elapsed time. With a 5.0M-parameter encoder pretrained on 2,115 hours of EEG, MSDE-10% achieves the second-best average fine-tuning rank across 20 OmniEEG-Bench tasks, behind only MSDE-Full among eight evaluated models. Matched controls favour future prediction over content recovery, and joint supervision exceeds both single-scale variants on 18 tasks. Transfer gains also extend to cross-dataset sleep staging without target adaptation. Frozen-state readouts reveal improved access to task-relevant physiological features, and component interventions link these features to classification gains. Spectrum-preserving shifts on FACED further show dependence on temporal alignment beyond whole-window power.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG foundation models, Self-supervised learning, Multi-scale temporal prediction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_RuoyOpenReview ↗
Rating
4(+2)
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper introduces MSDE, an EEG pretraining method that jointly predicts future teacher states at short and longer timescales using a causal encoder and a shared transition operator. Experiments across 20 tasks, cross-dataset evaluations, and physiological analyses support its transfer benefits.

Critical strengths

The motivation is sound. EEG tasks involve both brief events and longer-lasting rhythmic activity, making joint prediction across timescales practically relevant. Training a shared representation to support both scales may help preserve complementary information for downstream tasks with different temporal requirements.

Critical weaknesses

The variant without variance–covariance regularization is evaluated using the epoch-52 checkpoint, whereas the base protocol specifies 70 epochs. Although this is described as the last available checkpoint, the reason for using it remains unclear. The performance difference may therefore reflect both the removed regularization and the shorter training budget.

Questions for authors

Please compare MSDE with time-series foundation models used for classification, such as MOMENT, Mantis, TimEE, or Tic-FM. Consistent subject level splits, label budgets, and adaptation protocols would help establish its value beyond comparisons with EEG-specific models. If MSDE outperforms these models, please explain which aspects of its design account for the gains.

Please further clarify the motivation behind the loss design, including why each auxiliary term is needed and how it complements the multi-scale prediction objective.

6 comments
AuthorsOpenReview ↗

Response to W1/Q1

Thank you for identifying the training-duration confound and requesting comparisons beyond EEG-specific models.

W1. Equal-duration no-var-cov comparison

The regularization gain persists at equal training duration. The no-var-cov control completes the same 70 epochs as MSDE and reaches 50.74% FT BAcc versus 54.79%: +4.05 pp (95% CI [2.10, 6.12], p = 5.01e-04, 17/20 task wins). LP gives 40.42% versus 43.95%, a +3.52-pp gain (CI [1.59, 5.72], p = 0.0023). This replaces the original epoch-52 control.

Q1. Time-series baselines and sources of improvement

We evaluate all four named models on the 20-task benchmark. Mantis and MOMENT use the same FT protocol, subject-level splits, and labeled training/validation data as MSDE-Full. Hyperparameters are selected on seed-42 validation data and reused on seeds 5/10. TimEE and TIC-FM use in-context learning (ICL). MSDE-Full uses a 5.0M-parameter encoder pretrained on 21,195 source hours and reaches 56.06% mean FT BAcc.

ComparatorAdaptationBAcc (%)Full gain (pp)95% task-bootstrap CIW/T/LOne-sided exact p
MantisFT49.24+6.83[2.30, 11.66]17/0/30.0016
MOMENTFT40.44+15.63[10.36, 21.08]18/0/21.81e-05
TimEEICL44.48+11.58[5.64, 17.24]19/0/11.61e-04
TIC-FMICL42.67+13.39[8.01, 19.29]18/0/23.15e-05

Gains are MSDE-Full minus comparator. CIs use 100,000 task-bootstrap resamples; p values use one-sided exact paired Wilcoxon signed-rank tests over 20 task means. MSDE-Full significantly exceeds all four models under the adaptation settings shown, with 17-19 task wins. Mantis/MOMENT comparisons match FT and label access.

Mantis [1] aligns crop-and-resize views of the same sample while separating different samples. This is closely related to MSDE's identity anchoring, which aligns the two temporal halves of an EEG window while distinguishing different windows. A functionally similar consistency constraint is already part of MSDE, where it supports future-state learning. Anchoring promotes persistent, discriminative features; micro/macro prediction additionally requires the same states to encode EEG evolution across timescales.

The no-prediction control retains anchoring and variance-covariance regularization but removes future prediction. With the same encoder, 10% TUH data, and training budget, MSDE gains 4.62 pp (95% CI [2.83, 6.55], p = 2.38e-05, 18/20 wins), demonstrating the contribution beyond anchoring and regularization.

MOMENT [2] reconstructs masked signal patches, the content-recovery paradigm whose EEG limitations motivate MSDE. EEG's temporal continuity and spatial redundancy can allow recovery from nearby observations without encoding subsequent evolution. Samplewise reconstruction also rewards waveform details regardless of their relevance to downstream decoding. MSDE instead predicts learned future states from the observed past at multiple offsets, explicitly requiring temporal evolution without samplewise waveform recovery. With the same architecture and pretraining budget, MSDE exceeds bidirectional latent infilling and bidirectional raw-signal MAE by 4.10 and 4.56 pp, respectively, with 19/20 wins in both comparisons.

TimEE [3] meta-trains on synthetic classification tasks, using temporal and cross-series attention to predict query labels from labeled support series. TIC-FM [4] combines a contrastively pretrained time-series encoder, projection adapter, and in-context classifier. Both exploit temporal patterns for classification. MSDE explicitly trains EEG states to predict future states at multiple time offsets. Before downstream labels are introduced, each state must retain information for short-term changes and longer-range persistence; the no-prediction and single-scale controls test these requirements.

The encoder produces subsecond causal states integrating information across electrodes. A shared, elapsed-time-conditioned transition predicts future teacher states at micro offsets of 0.5-1.0 s and macro offsets around 1, 2, and 4 s. Joint prediction encourages retention of local event progression and rhythmic persistence, improving mean FT BAcc by 3.0 pp over micro-only and 3.7 pp over macro-only. Together, these results support learning discriminative EEG features alongside their future evolution across complementary timescales.

AuthorsOpenReview ↗

Response to Q2 and References

Q2. Motivation for auxiliary losses

As developed in Section 3.4 and Appendices B.3 and J, the auxiliary losses address three complementary requirements: discouraging constant-feature collapse while discovering discriminative EEG features, preserving variation in the pooled representation used downstream, and maintaining consistent temporal state transitions. Because the student and EMA teacher co-evolve, prediction alone can be made trivial by constant states and a constant-preserving transition.

Identity anchoring discourages constant-feature collapse and promotes discriminative feature discovery. It uses the two temporal halves of the same EEG window as a positive pair and other windows as negatives. If every window produces the same feature vector, positive and negative pairs cannot be distinguished. Window discrimination therefore discourages this trivial solution during pretraining. Comparing normalized half-window means also encourages the encoder to discover features that persist across temporal halves yet differ between windows. Anchoring thus supplies predictive training with discriminative states, while micro/macro prediction trains those states to capture future evolution; the manuscript also shows that overweighting anchoring reduces transfer performance.

Variance–covariance regularization guards against collapse of the pooled representation used downstream. Anchoring constrains directions of half-window means, whereas the downstream classifier reads full-window means. Variation among token states can disappear under this averaging: a trajectory shared by all windows can have token variance while producing identical pooled vectors. We therefore penalize low feature variance at both the token and pooled-window levels and off-diagonal covariance to discourage redundant dimensions. Applied during pretraining, the pooled variance penalty directly preserves between-window variation in the representation that the downstream classifier will read. Appendix J.1 relates this penalty to separation between pooled states.

Semigroup consistency maintains temporal coherence between direct and composed state transitions. Its motivation comes from temporal composition: the learned transition should give compatible states when advancing by a+b directly or by a followed by b. The prediction losses train transitions from encoded states to future teacher states; the composed path additionally uses the first predicted state as the input to its second step. We encourage agreement between Φ(s_t, a+b) and Φ(Φ(s_t, a), b) using Smooth L1, training both paths through the shared operator for a+b ≤ 1 s. Within the directly supervised 0.5–1.0 s range, Appendix J.2 bounds mean absolute composed prediction error using direct prediction error and path discrepancy evaluated on the same triples. This supplies a temporal-consistency constraint on the learned dynamics in addition to micro/macro endpoint supervision.

The removal ablations quantify the contribution of each requirement:

TermMean FT drop when removed
Identity anchoring3.58 pp
Variance–covariance4.05 pp, equal 70 epochs
Semigroup consistency1.86 pp

All three removal comparisons significantly reduce mean FT BAcc (all one-sided exact p < 0.01).

Planned manuscript revisions. We will replace the epoch-52 no-var-cov control with the completed 70-epoch results in the ablation table and training details (W1). We will report Mantis, MOMENT, TimEE and TIC-FM with their FT/ICL settings and relate their learning objectives to MSDE's multiscale future-state prediction, supported by the no-prediction, bidirectional-reconstruction and single-scale controls (Q1). In Section 3.4 and the appendices, we will clarify identity anchoring's roles in discouraging constant-feature collapse and promoting feature discovery, variance–covariance regularization's preservation of pooled and token-level variation, and semigroup consistency's alignment of direct and composed transitions. Each role will be linked to its removal ablation (Q2).

References

[1] Vasilii Feofanov et al. Mantis: Lightweight Calibrated Foundation Model for User-Friendly Time Series Classification. arXiv preprint arXiv:2502.15637, 2025.

[2] Mononito Goswami et al. MOMENT: A Family of Open Time-series Foundation Models. arXiv preprint arXiv:2402.03885, 2024.

[3] Jaris Küken et al. TimEE: End-to-end Time Series Classification via In-Context Learning. arXiv preprint arXiv:2607.07500, 2026.

[4] Juntao Fang et al. Rethinking Zero-Shot Time Series Classification: From Task-specific Classifiers to In-Context Inference. arXiv preprint arXiv:2602.00620, 2026.

Thank you for the detailed response and additional experiments. The equal duration ablation resolves my concern about the training budget confound. The added time series baselines strengthen the empirical evaluation, and the explanations of the auxiliary losses clarify their motivation and complementary roles. These revisions address my main concerns, and I will raise my score accordingly.

Reviewer_omyAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MSDE pretrains a small causal EEG encoder (5.0M parameters, seven tokens per second) to predict an EMA teacher's future states. A shared transition operator Φ(st,h)\Phi(s_t,h), conditioned on elapsed time, predicts token-level states 0.5-1.0 s ahead (micro) and patch-end states 1-4 s ahead (macro). InfoNCE anchoring and variance-covariance regularisation discourage collapse, and a semigroup consistency term aligns direct and composed predictions (Eq. 2). Pretrained on 2,115 h of TUH and tested on 20 OmniEEG-Bench tasks against reproduced LaBraM, CBraMod, REVE-B/L and Neuro-GPT, it gets the second-best average FT rank (behind its full-corpus variant) and beats same-encoder content-recovery controls by 4.1-5.5 points. It transfers across sleep datasets. Readouts, interventions and circular shifts link it to physiological components.

Critical strengths
  1. The objective is tested against matched controls. No prediction, latent infilling, MAE at three mask ratios, micro-only and macro-only all use the same encoder, 10% corpus, schedule and pretraining seed (Appendix B.1). They trail MSDE by 3.0-5.5 FT points, with task-bootstrap intervals that exclude zero (Table 6), and joint prediction beats both single scales on 18 of 20 tasks. For a pretraining objective, this is stronger evidence than comparing released checkpoints trained on different data.

  2. The reporting is honest. The paper says MSDE-10% ties the best foundation model picked per task (+0.2, CI [-1.9, +2.4], Table 8) and REVE-L has the best LP rank (Table 10). It lists test subjects, seed SDs and compute (Tables 3, 13, 14). A 5.0M-parameter encoder that competes with a 69M one is useful in itself.

Critical weaknesses
  1. Small test sets, one pretraining run. Several headline claims rest on one or two test subjects and on a single pretraining run. Table 3 lists one test subject per split for ISRUC-S3, ISRUC-S2, SEED, SEED-IV, SEED-V, BCIC-IV-2a, BCIC-IV-1 and ErrP. The cross-dataset result in the abstract and Section 4.5 ("10.5 points above ... REVE-L") uses a target test split fixed at seed 42 (L1356-1358), and for ISRUC-S3 that split is one subject. In Table 18, HMC→ISRUC-S3 (74.2 vs. 31.0 for REVE-L) gives 43.2 of the 63.0 summed points. If I drop it, the mean margin over REVE-L is about 4.0 points, and MSDE loses two of the other five directions. Several case studies in Appendix I (ISRUC-S3 +10.5, ErrP +10.4, BCIC-IV-2a +8.1 over infilling) also rest on one test subject per seed. Every pretraining variant is also one run with pretraining seed 0 (Table 4). So the CIs in Tables 6-8 cover task and split variation but not run-to-run variation in pretraining, while the gaps are 1.9-5.5 points. I'd like a second pretraining seed for MSDE and at least the infilling and MAE controls, cross-dataset targets with more subjects (all ISRUC-S3 subjects, or leave-subject-out), and a note in Sections 4.4-4.5 on how many subjects each claim rests on.

  2. Possible stimulus shortcut on FACED. The largest gain, and the basis of Section 5.3, may come from recognising stimulus-locked clips rather than emotion. Appendix K.1 shows that FACED positions identify a window within a shared video. Below 3 Hz, same-position signals correlate across subjects at 0.23-0.29, against about -0.01 after a 0.25-0.5 s shift (Table 23). The split is by subject, so test subjects watch the same clips as training subjects, and the 9-class label is presumably tied to the clip. So a model can read the label from clip-specific, time-locked low-frequency responses without learning emotion. The paper's own interventions fit this. MSDE depends on frontal 0.5-3 Hz activity (-15.6 points, Table 31), and a 0.25 s circular shift costs it 21.5 points (Table 24). Section 5.3 and the abstract present this as "dependence on temporal alignment beyond whole-window power", a sign of good representations. But deployed windows are not stimulus-aligned, so this can also be read as fragility. FACED matters a lot in the summary: MSDE leads LaBraM there by 39.8 points, and it is the largest-gain task removed in Table 15 for three of the five baselines. A split that holds out clips (or stimulus positions), plus an analysis that separates ocular from neural low-frequency activity in the frontal channels, would show which reading is right. The SEED datasets show a weaker form of the same position structure in Table 23 and should get the same check.

Some smaller points. All reproduced baselines share one FT configuration and a six-candidate grid picked on split seed 42 (Table 5, L856-857). Appendix F compares only ranks with the published OmniEEG-Bench numbers, so I can't tell whether a baseline's score (for example LaBraM at 21.5 on FACED) matches its published value. The "macro" horizons reach 4.5 s within 16-s inputs (Table 4). That is shorter than the abstract's "longer-range forecasts" and "sustained rhythms" suggest. The no-var-cov control is taken at epoch 52 rather than 70 (Table 4).

Questions for authors
  1. How much does the 20-task FT mean vary from run to run across pretraining seeds, for MSDE and for the infilling and MAE controls?
  2. Without HMC→ISRUC-S3, or with an ISRUC-S3 target of more than one subject, what is the cross-dataset margin over REVE-L and over infilling?
  3. On FACED (and SEED), what do MSDE and the baselines score under a split that holds out video clips? Does the frontal 0.5-3 Hz dependence stay after ocular-artifact removal?
  4. For tasks shared with OmniEEG-Bench, how do your reproduced baseline scores compare with the published ones?

If the FACED advantage goes away, the physiological-alignment claims in Section 5.3 would need to be removed.

6 comments
AuthorsOpenReview ↗

Response to Q1/W1/Q2

Thank you for the specific requests. We added an independent pretraining run, expanded ISRUC-S3 target coverage, and evaluated video holdout and ocular sensitivity across FACED and the four SEED datasets. We retain your Q1–Q4 numbering, grouping related evidence under thematic subheadings where needed. The remaining smaller points are answered separately.

Q1 / W1. Variation across pretraining seeds

The second independent pretraining run preserves the transfer advantage. The 20-task FT means are:

ModelPretraining seed 0Pretraining seed 1Mean ± sample SD
MSDE-10%54.7955.3655.08 ± 0.40
Causal infilling50.7450.5050.62 ± 0.17
Causal MAE (ratio 0.5)50.5250.6550.58 ± 0.10

In seed 1, MSDE improves on causal infilling by +4.86 pp (p = 1.91e-06) and causal MAE by +4.71 pp (p = 7.16e-04). Averaging the two runs within each task gives gains of +4.46 and +4.49 pp, with wins on 20/20 and 18/20 tasks, respectively. These two runs directly answer the requested replication and provide an initial estimate of pretraining variability.

Q2 / W1. Cross-dataset sensitivity and target-subject coverage

Q2.1. Expanded target coverage and subject-count reporting

We expanded both ISRUC-S3 target evaluations to all target subjects. The updated BAcc (%) is:

ScopeMSDE-10%InfillingREVE-L
HMC→ISRUC-S365.7463.7228.54
SleepEDF→ISRUC-S345.7843.0737.76
Six-direction mean42.2737.3433.22

The six-direction gains are +4.93 pp over infilling and +9.05 pp over REVE-L. We will replace the original +10.5-pp headline with the expanded-target +9.05-pp result and show the exclusion analysis alongside it.

For the separate within-dataset case studies, we will explicitly retain the test-subject counts from Table 3.

Q2.2. Results without HMC→ISRUC-S3

Excluding HMC→ISRUC-S3 leaves +5.51 pp over infilling (5/5 direction wins) and +3.42 pp over REVE-L (3/5). The mean advantage therefore remains positive after removing the dominant direction.

AuthorsOpenReview ↗

Response to Q3.1

Q3 / W2. FACED/SEED video holdout and ocular sensitivity

Q3.1. Video holdout and stimulus generalization

We evaluated video holdout on FACED and all four SEED variants, with eight models including CBraMod and REVE-B. Held-video BAcc (%) is:

DatasetMSDEInfillMAEScratchLaBraMCBraModREVE-BREVE-L
FACED11.5711.9510.6410.739.8610.2610.7612.46
SEED54.1147.4953.2153.1151.0059.0256.6650.19
SEED-IV32.5532.4633.3326.1328.6624.9932.4232.37
SEED-V31.9830.5428.9722.2827.8123.8321.4323.65
SEED-VII16.2914.6914.3017.7315.2015.7719.6216.76

FACED averages nine evaluations (seeds 5/10/42 × folds 0/1/2) for each of the eight models; each SEED variant averages three split-seed runs. These are additional holdout evaluations, separate from the main-table protocol.

The near-chance FACED result is shared across all eight tested models. MSDE falls from 62.30% on seen videos to 11.57% on held videos, close to nine-class chance (11.11%); infilling, MAE, Scratch, LaBraM, CBraMod, REVE-B, and REVE-L also perform near chance.

FACED assigns fixed emotion labels to videos, which recur across subjects; subject-only splits therefore preserve a stimulus-to-label shortcut. Xu and Plataniotis [1] independently identify this confound in CLISA, MSHCL, and DAEST on FACED and SEED. With frozen encoders, their classifier-stage joint subject-and-stimulus holdout on FACED-9 yields permutation-corrected drops of 25.99, 2.17, and 44.43 pp, respectively. Our common near-chance outcome across eight models is consistent with this shared dataset/protocol limitation. The collapse is not specific to MSDE's predictive objective or architecture.

The large FACED drop is not reproduced at comparable magnitude in the SEED evaluations. MSDE's seen-minus-held differences are +4.31, -0.29, +0.20, +1.21 pp on SEED, SEED-IV, SEED-V, and SEED-VII, respectively. Rankings vary across datasets: MSDE has the highest held-video mean on SEED-V (31.98%). On SEED, MSDE reaches 54.11%, compared with 59.02% for CBraMod and 56.66% for REVE-B.

The transfer advantage is not driven solely by FACED. On the remaining 19 original tasks, MSDE-10% retains higher mean FT BAcc than each of the five pretrained baselines in the submitted paper:

Pretrained baselineMSDE-10% FT gain (pp)One-sided p
LaBraM+3.560.0115
CBraMod+7.074.83e-04
REVE-B+2.330.0616
REVE-L+2.790.0308
NeuroGPT+5.585.86e-04

The differences are significant for four of the five baselines at p < 0.05; the gain over REVE-B remains positive but is not significant (p = 0.0616). The matched objective comparisons also retain significant gains: MSDE exceeds bidirectional infilling by +4.21 pp (18/19 wins, p = 1.91e-05) and bidirectional MAE by +4.79 pp (18/19 wins, p = 2.61e-04). These results support the broader transfer benefit of MSDE beyond the FACED-specific physiological interpretation.

The spectrum-preserving FACED shifts establish that the classifier is sensitive to temporal alignment. We will present the shift, video-holdout, and ICA measurements together as diagnostics of temporal sensitivity, stimulus dependence, and low-frequency contributions, with the corresponding revision specified in Q3.2.

AuthorsOpenReview ↗

Response to Q3.2-1

Q3.2. Physiological interpretation and ocular sensitivity

We added ICA interventions to test whether the FACED low-frequency and temporal-alignment effects reported in Section 5.3 persist after removing components selected by frontal ocular proxies. We fitted 30 ICA components per test subject and evaluated selection thresholds of 0.5, 0.6, and 0.7 with the fine-tuned checkpoints fixed. The variance-matched control removes the same number of non-selected components with the closest variance contributions. Band-stop and circular-shift effects are the additional BAcc decreases relative to the input before each perturbation. Each threshold is reported separately in the tables: values average split seeds 5/10/42 within that threshold, with no averaging across thresholds.

Before component removal on FACED, MSDE reaches 61.32% BAcc, with a 15.56-pp frontal 0.5–3 Hz band-stop effect and a 41.88-pp circular-shift effect. The results at all three thresholds are:

ThresholdBAcc: ICA (%)BAcc: control (%)Band-stop: ICA (pp)Band-stop: control (pp)Circular shift: ICA (pp)
0.529.4657.084.4214.4515.69
0.631.6857.044.9114.0917.61
0.737.8856.917.8314.0723.16

The variance-matched controls preserve much more of the original classification performance and low-frequency effect, showing that the effects depend on which components are removed. Both low-frequency dependence and temporal-alignment sensitivity weaken but remain after ICA at all three thresholds.

The selected components therefore carry information used for classification and contribute to the original FACED sensitivities. Their removal can affect both ocular and neural activity because these sources overlap in frequency and may coexist within ICA components. The results are compatible with mixed contributions and show continued use of low-frequency and temporal structure after removal; they do not identify the neural origin of the residual effects. This refines the source interpretation of the original FACED observations while retaining the measured signal sensitivities.

We extended the same checks to all four SEED datasets. The large frontal low-frequency dependence observed on FACED is not reproduced at comparable magnitude across these datasets. The following table reports MSDE-10% separately at each threshold. Thresholds 0.5 and 0.6 select identical components on all four SEED datasets; SEED-V also has the same selection at 0.7, yielding identical results for these settings. No component is removed on SEED at 0.7 or on SEED-IV at any threshold, so their inputs and scores remain unchanged:

DatasetThresholdOriginal BAcc (%)After ICA (%)Band-stop: original (pp)Band-stop: after ICA (pp)
SEED0.557.2553.390.963.32
SEED0.657.2553.390.963.32
SEED0.757.2557.250.960.96
SEED-IV0.532.9032.90-0.01-0.01
SEED-IV0.632.9032.90-0.01-0.01
SEED-IV0.732.9032.90-0.01-0.01
SEED-V0.532.6332.61-0.05-0.16
SEED-V0.632.6332.61-0.05-0.16
SEED-V0.732.6332.61-0.05-0.16
SEED-VII0.519.4819.141.281.23
SEED-VII0.619.4819.141.281.23
SEED-VII0.719.4819.701.281.78

The SEED results show a different response. On SEED-V, removing a selected component accounting for 70.9% of variance and leaving 20.7% of frontal low-frequency power changes every model's BAcc by less than 0.25 pp. On SEED-VII, MSDE changes by −0.34 to +0.22 pp across thresholds. On SEED, thresholds 0.5/0.6 remove one component accounting for 4.7% of variance: MSDE changes from 57.25% to 53.39%, whereas the variance-matched control gives 58.33%. Its band-stop effect increases from 0.96 to 3.32 pp. Thus, the ICA-related score change does not track the elimination of frontal low-frequency dependence.

Reviewer_KGVQOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper tried to address the gap about the training objective of EEG Foundation Models. Typically EEG FMs relied heavily on masked reconstruction, without consideration about temporal dynamics at multiple time-scales. The authors hypothesized that this can be address by pre-training with multiple time-scales prediction loss. They validated this hypothesis by proposing MDSE, which uses a causal encoder, trained using their proposed loss with an EMA teacher in typical representation learning methods. The proposed method is evaluated on around 20 downstream datasets and showed some improvement against the baselines they chose.

Critical strengths
  • The hypothesis about multi time-scales prediction loss could improve the representation learning process is relevant, given the current literature of EEG FMs have shown that the objective losses now might not be suitable.
Critical weaknesses
  1. Honestly, I don't see any substantial novelty compared with existing multi time-scales and prediction-type losses for EEG FMs pre-training. The author misses comparison with strong and relevant baselines and including THD-BAR [1], PATCHCODE [2], CodeBrain [3], CSBrain [4]. For example, in CSBrain, they also use multi time-scales for representation learning. I know that they don't use the future-predictive types of loss, but if you hypothesize the multi time-scales are relevant, you can improve it by leverage multi time-scales with a single scale prediction loss, I think it also can learn from multiple time-scales. And the improvement with existing chosen baselines are not significant. For example, despite of MSDE improvement over individual baselines, Tab. 8 shows no significant advantage over the best existing FM selected per task (+0.2, p=0.478), while Table 9 shows no significant linear probe improvement over REVE. I do not find the novelty and significance convincing.

  2. I find this paper didn't prove where the improvement actually comes from. The proposed method is basically combining prediction losses at different horizons with a shared transition operator, but this objective is quite general and can be applied to different EEG backbones, the authors only evaluate it on their own architecture. I suggest that you can apply the same objective to other existing EEG FMs to show whether the improvement is generalizable. Without this, I don't think that this paper can introduce well a new learning principle.

  3. I also have concerns about the experimental design. In Appendix B.1, the masked latent infilling baseline uses causal attention, meaning that masked positions cannot access future patches. However, I find this is an irrelevant comparison with masked modeling, which can use both past and future context for reconstruction. Therefore, I don't think the comparison in Table 6 is meaningful, because it cannot conclude that future prediction is better than masked reconstruction in general.

References:

[1] THD-BAR: Topology Hierarchical Derived Brain Autoregressive Modeling for EEG Generic Representations. NeurIPS 2025.

[2] PATCHCODE: Discrete Latent Predictive Learning for EEG Foundation Model. ICML 2026.

[3] CodeBrain: Bridging Decoupled Tokenizer and Multi-Scale Architecture for EEG Foundation Model. ICLR 2026.

[4] CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding. NeurIPS 2025.

Questions for authors

See Critical Weaknesses.

10 comments
AuthorsOpenReview ↗

Response to W1-1

Thank you for the questions about the learning objective, related models, and experimental controls. We address your three weaknesses in order, explaining the requirement imposed by future-state supervision and testing it through matched controls and expanded EEG-baseline comparisons.

W1. Novelty, EEG-baseline comparisons, and statistical significance

We understand the reviewer's concern about novelty. MSDE is conceptually simple: our aim is to align pretraining with EEG's temporal structure, so that representations capture both transient event progression and sustained rhythmic activity. Widely used masked-recovery objectives can exploit local signal redundancy without requiring prediction beyond the observed context, while forecasting at one granularity does not explicitly couple these complementary temporal demands. MSDE addresses this gap through joint micro- and macro-scale future-state supervision.

We apologize for not including CodeBrain and CSBrain results in the original submission due to time constraints. We now report complete 20-task FT and LP evaluations for CodeBrain and CSBrain, together with THD-BAR, PATCHCODE, EEGMamba, FEMBA Base, LUNA Base, and BrainOmni.

The learning objective and its relation to existing EEG models.

CSBrain [1] combines cross-scale spatiotemporal tokenization with structured sparse attention to reconstruct masked EEG waveforms. Its multiscale architecture enriches the context available for reconstruction, while the learning target remains the masked signal.

EEG's temporal continuity and spatial redundancy allow nearby observations to support local interpolation or continuation of correlated patterns. Masked reconstruction can therefore succeed without requiring the representation to anticipate activity beyond the observed context. Adding multiscale context does not remove this objective-level limitation.

MSDE trains one causal EEG state to predict future teacher states through a shared, elapsed-time-conditioned transition, jointly covering micro offsets of 0.5–1.0 s and macro offsets around 1, 2, and 4 s. This imposes a more demanding temporal modeling requirement: the same state must infer multiple future targets from history alone, retaining information useful for both event progression and rhythmic persistence. This predictive demand encourages the encoder to capture EEG patterns underlying temporal evolution, with supervision in learned state space rather than samplewise waveform recovery.

Among the other baselines, LaBraM [2] predicts masked discrete codes produced by a spectral tokenizer, while CodeBrain [3] recovers masked temporal/frequency code indices with a multiscale EEGSSM encoder. CBraMod [4] and REVE [5] reconstruct masked signal patches; EEGMamba [6] and FEMBA [7] use bidirectional Mamba encoders for masked-signal reconstruction, and LUNA [8] reconstructs waveform patches after learned-query channel unification. BrainOmni [9] predicts masked BrainTokenizer code indices non-autoregressively, whereas PATCHCODE [10] combines masked latent-patch prediction with discrete-code supervision. NeuroGPT [11] predicts the next continuous chunk embedding from growing prefixes, while THD-BAR [12] autoregressively predicts hierarchical discrete tokens. MSDE's distinction is the joint prediction of continuous future teacher states at explicit temporal offsets from one shared causal state.

AuthorsOpenReview ↗

Response to W1-2

EEG-baseline comparisons and statistical significance.

The comparisons below cover all eight EEG baselines across the same 20 tasks, with MSDE-Full as the reference. Because neither THD-BAR [13] nor PATCHCODE [14] had publicly released pretrained checkpoints, we pretrained both models using 10% of the TUH data. For PATCHCODE, the paper and repository did not fully specify the channel configuration and base-model hyperparameters; we completed these settings while keeping the total parameter count comparable to the base model reported in its paper. For THD-BAR's FT evaluation, we report the instruction-fine-tuning results. Evaluation settings are specified in the reporting conventions.

MSDE-Full reaches 56.06% mean FT BAcc:

ComparatorFT BAcc (%)MSDE-Full − comparator (pp)95% task-bootstrap CIFull W/T/LOne-sided exact p
CodeBrain48.64+7.42[3.67, 11.22]16/0/47.16e-04
CSBrain49.14+6.93[3.00, 11.51]17/0/34.25e-04
THD-BAR42.86+13.20[8.50, 18.06]19/0/11.34e-05
PATCHCODE48.38+7.68[3.92, 11.14]18/0/24.25e-04
EEGMamba45.19+10.87[6.31, 15.68]18/0/28.39e-05
FEMBA Base48.92+7.14[3.49, 11.16]19/0/11.61e-04
LUNA Base49.46+6.60[3.44, 10.01]18/0/21.97e-04
BrainOmni47.54+8.53[4.21, 13.23]17/0/33.54e-04

MSDE-Full significantly exceeds all eight baselines in FT, with gains of 6.60–13.20 pp (all p < 0.001). At the same 10% TUH source-data scale used for PATCHCODE and THD-BAR, MSDE-10% also exceeds them by +6.41 and +11.93 pp, respectively (both p < 0.001).

The advantage also appears in frozen-encoder linear probing. MSDE-Full reaches 45.26% mean LP BAcc across the same 20 tasks:

ComparatorLP BAcc (%)MSDE-Full − comparator (pp)95% task-bootstrap CIFull W/T/LTwo-sided exact p
CodeBrain42.03+3.23[0.31, 6.66]15/0/50.0637
CSBrain38.19+7.07[3.67, 11.50]17/0/31.05e-04
THD-BAR38.63+6.63[3.18, 10.32]17/0/30.0014
PATCHCODE41.76+3.50[1.16, 5.78]14/0/60.0083
EEGMamba37.30+7.96[4.17, 12.77]17/0/32.61e-04
FEMBA Base37.61+7.65[4.21, 11.30]17/0/34.83e-04
LUNA Base40.15+5.11[2.61, 7.75]17/0/37.08e-04
BrainOmni39.13+6.13[4.04, 8.32]19/0/11.34e-05

MSDE-Full improves mean LP BAcc over all eight baselines by 3.23–7.96 pp. Seven of the eight LP gains are significant under two-sided tests (all p < 0.01). The +3.23-pp gain over CodeBrain has p = 0.0637. These results support the quality of the frozen pretrained representations alongside the FT gains.

AuthorsOpenReview ↗

Response to W1-3

All eight added EEG baselines are included in the expanded ranking comparison. MSDE-Full and MSDE-10% achieve the best and second-best average FT ranks, 3.25 and 4.25, respectively, among 16 EEG models; MSDE-Full also achieves the best average LP rank (4.30) among 15 models. MSDE-10% achieves its FT rank with a 5.0M-parameter encoder pretrained on 2,115 source EEG hours. Scratch has no pretrained-encoder LP setting. Lower average rank is better:

ModelAverage FT rankAverage LP rank
MSDE-Full3.254.30
MSDE-10%4.255.25
REVE-B6.535.45
REVE-L6.884.85
LaBraM-Base8.406.83
CodeBrain8.557.50
CSBrain8.6010.05
LUNA Base8.638.28
BrainOmni9.009.90
FEMBA Base9.4010.13
MSDE-Scratch9.50—
NeuroGPT9.539.55
PATCHCODE10.007.18
EEGMamba10.5010.48
CBraMod10.8510.58
THD-BAR12.159.70

We agree that MSDE has not demonstrated a statistically significant FT improvement over the original five-baseline task-wise-best reference: the mean gains are +0.20 pp for MSDE-10% (p = 0.4782) and +1.47 pp for MSDE-Full (p = 0.1012). This is a particularly demanding criterion because the reference selects a different model whenever it performs best on a task. To put this criterion in context, we apply it symmetrically in a leave-one-model-out comparison using the existing FT scores of MSDE-10% and the original five external pretrained models. For each of these six models, we compare its score on every task with the highest score among the other five models, then average the differences over the same 20 tasks. Thus, the reference for each external model includes MSDE-10%.

ModelFT BAcc (%)Task-wise best of other five (%)Model − reference (pp)95% task-bootstrap CIW/T/L
MSDE-10%54.7954.59+0.20[-1.94, 2.39]10/0/10
REVE-B52.0956.49-4.40[-6.38, -2.50]3/0/17
REVE-L51.4456.65-5.20[-7.17, -3.39]1/0/19
LaBraM-Base49.4256.54-7.11[-11.47, -3.84]4/0/16
NeuroGPT48.5956.43-7.84[-10.64, -5.10]2/0/18
CBraMod46.7556.69-9.94[-13.51, -6.71]0/0/20

MSDE-10% is the only model with a positive mean difference against the task-wise best of the other five models (+0.20 pp); every external model falls short of its corresponding reference by 4.40–9.94 pp. MSDE-10% also leads the remaining models on 10/20 tasks, compared with 0–4/20 for each external model. A single MSDE pretrained encoder therefore attains mean FT performance close to selecting the best existing model separately for every task. Together with its significant FT gains over each of the five fixed external baselines in Table 8, this establishes a meaningful improvement in cross-task transfer with MSDE pretraining.

Regarding the reviewer's observation that the LP gains over REVE are not statistically significant, the original Table 9 compares MSDE-10% with models trained at markedly larger scales. MSDE uses a 5.0M-parameter encoder, whereas REVE-B/L use 69.2M/390.0M parameters under the manuscript's retained-backbone accounting—13.8×/78× larger. Both REVE models are pretrained on 61,415 source EEG hours, approximately 29.0× MSDE-10%'s 2,115 hours and 2.9× MSDE-Full's 21,195 hours. MSDE therefore operates with a substantial disadvantage in model capacity and pretraining-data budget.

Despite this resource gap, MSDE-10% reaches 43.95% mean LP BAcc, compared with 43.55%/43.78% for REVE-B/L; MSDE-Full reaches 45.26%. In FT, MSDE-10% and MSDE-Full achieve 54.79% and 56.06%, respectively, exceeding REVE-B/L at 52.09%/51.44%. These results show that MSDE pretraining learns transferable EEG representations with comparable LP accuracy and higher mean FT performance than much larger REVE models pretrained on substantially more data.

The gains are substantial against models at a closer resource scale. LaBraM (5.8M parameters; 2,535 source EEG hours) is close to MSDE-10% in both model and pretraining-data size, yet MSDE-10% improves FT BAcc by +5.37 pp. CBraMod and CSBrain [1] have similarly sized encoders (approximately 4.9M) but use more source EEG (9,246 hours); MSDE-10% exceeds them by +8.04 and +5.66 pp. It also exceeds CodeBrain [3], which uses a larger 15.17M-parameter backbone and 9,246 source hours, by +6.15 pp. All four 95% task-bootstrap intervals exclude zero, with one-sided p < 0.01. Together with the same-encoder objective controls, component ablations, and Scratch comparison in W2–W3, these results demonstrate the effectiveness of MSDE's pretraining method: joint micro- and macro-scale future-state supervision improves the transferability of EEG representations across downstream tasks.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank all three reviewers for their careful reading and constructive suggestions. In response to the reviews, we added 12 foundation-model baselines, bidirectional reconstruction and equal-duration ablations, a second backbone and independent pretraining seed, and full-cohort cross-dataset and baseline-reproduction audits. Together, these additions clarify why multi-scale future-state prediction transfers more effectively across EEG tasks than masked content recovery.

The principal 20-task fine-tuning (FT) comparisons are summarized below (balanced accuracy, %):

Configuration20-Task FT BAcc (%)
Strongest of 8 added EEG foundation models (LUNA Base)49.46
Strongest of 4 added time-series models (Mantis)49.24
Bidirectional raw-signal MAE (matched 10% TUH, 70 ep)50.24
Bidirectional latent infilling (matched 10% TUH, 70 ep)50.69
w/o variance-covariance (matched 10% TUH, 70 ep)50.74
MSDE-10% on adapted criss-cross backbone (vs. CBraMod 46.75%)50.74
MSDE-10%54.79
MSDE-Full56.06

1. Consistent gains across expanded EEG and time-series benchmarks (Section 4.2, Appendix D).
Across all 20 tasks, we evaluated eight additional EEG foundation models (CodeBrain, CSBrain, PATCHCODE, LUNA Base, FEMBA Base, BrainOmni, EEGMamba, and THD-BAR) and four general time-series models (Mantis and MOMENT under FT; TimEE and TIC-FM under ICL). MSDE-Full (56.06% FT, 45.26% LP) exceeds all eight added EEG models by 6.60-13.20 pp in FT (all p < 0.001) and 3.23-7.96 pp in LP (7/8 with p < 0.01), ranking first in both FT (3.25) and LP (4.30) among 16 EEG models (MSDE-10% ranks second in FT at 4.25), and exceeds all four time-series models by 6.83-15.63 pp (all p < 0.002). In a symmetric leave-one-model-out task-wise-best test across the original six models, MSDE-10% is the only model with a positive mean margin (+0.20 pp, vs. -4.40 to -9.94 pp for external baselines), showing with a 5.0M-parameter encoder that MSDE transfers competitively across EEG tasks even against much larger backbones and pretraining corpora.

2. Predictive dynamics beyond bidirectional reconstruction and training budget (Sections 3.4 and 4.3, Appendices C.1 and J).
Under the same encoder, 10% TUH subset, and 70-epoch budget, we added bidirectional latent infilling and bidirectional raw-signal MAE controls, and completed the 70-epoch no-variance-covariance ablation. Bidirectional context does not improve transfer over causal reconstruction (50.69% vs. 50.74% for infilling; 50.24% vs. 50.52% for MAE), and MSDE-10% still exceeds bidirectional infilling by 4.10 pp and bidirectional MAE by 4.56 pp (both 19/20 task wins, p < 2e-4). At equal 70 epochs, variance-covariance regularization adds +4.05 pp in FT (p = 5.01e-4) and +3.52 pp in LP (p = 0.0023), alongside identity anchoring (+3.58 pp) and semigroup consistency (+1.86 pp), confirming that MSDE's gains stem from joint micro/macro future-state prediction rather than causal masking limits or training duration, with each auxiliary term fulfilling a distinct non-collapse or consistency role.

3. Generalization to a second backbone and pretraining reproducibility (Appendices G.1-G.2).
We adapted CBraMod's criss-cross architecture to MSDE's causal state interface (2.86M parameters, omitting FFT and ACPE). Under the same 10% TUH, 70-epoch setup, it reaches 50.74% FT and 43.10% LP, improving over original CBraMod (4.90M) by 3.99 pp in FT and 5.09 pp in LP and surpassing LaBraM and Neuro-GPT. An independent second pretraining seed also reproduces MSDE-10%'s 20-task advantage (55.08 ± 0.40% FT, vs. 50.62 ± 0.17% for infilling and 50.58 ± 0.10% for MAE), confirming that multi-scale state-dynamics pretraining generalizes to another spatial-temporal backbone and remains stable across independent pretraining runs.

4. Full-subject cross-dataset transfer and baseline-reproduction audit (Section 4.5, Appendices B, F.1, and H).
Expanding ISRUC-S3 evaluation to all target subjects yields six-direction mean gains of +4.93 pp over infilling and +9.05 pp over REVE-L (+5.51 and +3.42 pp without HMC->ISRUC-S3). Across a complete 122-entry (48 FT, 74 LP) published-vs.-reproduced audit, our reproduced FT means exceed published OmniEEG-Bench scores for all four baselines (+1.33 to +11.73 pp), confirming that cross-dataset generalization holds across full target cohorts and that our downstream gains are established against strong, fairly tuned baselines on shared splits.

The completed experiments, protocol clarifications, and diagnostic analyses have been incorporated into the revised manuscript (marked in blue).

We hope these additional experiments and clarifications help you better understand and evaluate our work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.