acceptodds
Under review as a conference paper at ICLR 2027

Zero-Shot Generation of the Next Moment of Brain Activity by Predicting Its Distribution

Abstract

EEG is a continuous, highly stochastic signal dominated by broadband noise and non-stationary fluctuations, so a pointwise autoregressive model chases noise it cannot reliably estimate. We therefore regard brain activity as intrinsically stochastic: the next moment of EEG is governed by a probability distribution. We propose GenBrain, a generative autoregressive model of EEG that predicts this distribution directly and generates brain activity by sampling from it. Grounded in the computational-neuroscience account of EEG as rhythmic oscillations riding on an aperiodic background, we parameterize the distribution with Rayleigh amplitudes, von Mises phases, phase-amplitude coupling, and a spectral slope, so that every predicted parameter carries a neurophysiological meaning. To capture transient bursts and spikes beyond this parameterization, a dual-head design pairs distribution prediction with zero-initialized raw-signal reconstruction; both heads share a block-causal Transformer backbone and train jointly, while neurophysiological priors enter as differentiable soft regularizers. Pretrained on 14 public EEG corpora and evaluated zero-shot on 10 held-out datasets spanning six paradigms, GenBrain attains lower teacher-forced negative log-likelihood than marginal, persistence, and autoregressive baselines on all 10 datasets (32–62% relative improvement, up to 7.0 nats below the strongest baseline); probability-integral-transform residuals localize where the parameterized family's tails fall short; and autoregressive generation improves spectral fidelity on 9 of 10 datasets (PSD correlation up to 0.94) while holding amplitude calibration at or below the real–real floor on 8 of 10. GenBrain recasts the modeling of brain activity as distribution prediction, offering a novel solution to the brain's intrinsic stochasticity.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, generative autoregressive model, neuro-mechanistic distribution prediction, block-causal attention, brain activity generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_EimrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper present GenBrain. It investigating how to probabilistically predict and generate EEGs with strong randomness. It predicts the distribution parameters for each channel over the next second based on historical EEG data, constructing a parametric generative model using Rayleigh amplitudes, von Mises phases, 1/f background components, and phase-amplitude coupling. The model employs a block-causal Transformer architecture, integrating a distribution prediction head, an auxiliary waveform prediction head, and physiological prior constraints; generation involves sampling followed by an inverse FFT, with the option to incorporate residuals provided by the waveform head.Experiments involved pre-training on 14 public EEG datasets, followed by zero-shot generation evaluation on 10 datasets not included in the pre-training phase.

Critical strengths

Board evaluation on mutiple datasets and tasks: The model pre-trains on 14 EEG datasets and performs zero-shot evaluation on 10 datasets covering six scenarios. It also examines single-step prediction likelihood, distribution calibration, and autoregressive generation quality. This design tests the method's applicability across diverse EEG data conditions and mitigates the risk of conclusions relying on a single dataset or task. Such comprehensive experimental evidence makes the proposed method more convincing.

Critical weaknesses

1. The advantages over latent prediction have not been fully substantiated. Directly reconstructing highly random EEG waveforms may not be an ideal pre-training target, and this motivation is reasonable, but it can still support predictions in the learned latent space. The paper lacks a direct comparison to justify the choice of these manually designed distribution parameters, particularly against probabilistic latent prediction methods that also account for uncertainty.

2. The validation of distribution parameter dynamics and their physiological significance is insufficiently systematic. Since the manually defined distribution forms constrain the model's representational space, the ability of these parameters to effectively capture actual EEG variations is crucial to the method's validity. Although the appendix provides local analyses such as spectral slope correlation, the paper has not yet systematically shown the time variation curves of various parameters, as well as the correspondence between the predicted parameters and the actual EEG estimated parameters. Consequently, there is insufficient evidence to confirm that these parameters possess the claimed physiological interpretability.

Questions for authors
  1. Latent prediction offers an alternative to directly regressing highly stochastic EEG waveforms. Could the authors further elaborate on the specific advantages of predicting parameters of a predefined physiological distribution versus predicting a learned latent representation? Would it be possible to provide a simple comparative experiment against probabilistic latent prediction.

  2. Could the authors demonstrate the trajectories of predicted parameters as pre-training progresses, or report the parameter values ​​at different training stages? Illustrating parameter changes surrounding brain state transitions would help substantiate claims regarding physiological interpretability.

  3. The model is restricted to adjusting parameters within a predefined family of distributions. Have the authors considered or experimented with more flexible distribution families (i.e. Weibull distribution)?

  4. In the limitations section, the authors note that the phase learning branch is weak and that there is a discrepancy between the predicted phase concentration and the observed data. Given that phase modeling is one of the core components of the proposed method, please explain why this situation was not further fixed.

9 comments
AuthorsOpenReview ↗

Response to Reviewer Eimr [1/N]

Weakness 1

The advantages over latent prediction have not been fully substantiated. Directly reconstructing highly random EEG waveforms may not be an ideal pre-training target, and this motivation is reasonable, but it can still support predictions in the learned latent space. The paper lacks a direct comparison to justify the choice of these manually designed distribution parameters, particularly against probabilistic latent prediction methods that also account for uncertainty.

Response: We agree a direct comparison is the right way to settle this, and we ran the strongest form of this baseline: a conditional-VAE baseline sharing the same trunk architecture, data pipeline, and the full 10-epoch training budget as our distribution head (64-d Gaussian latent, Gaussian waveform decoder, ELBO-trained; 6.15M vs 5.81M parameters), whose reconstruction target is the raw waveform patch — it never routes through our 19-dimensional physiological parameterization — trained jointly from scratch. Design: trunk states hth_t are block-causal (hth_t encodes x≤tx_{\le t}); a posterior head q(zt ∣ h≤t)q(z_t\,|\,h_{\le t}) over a 64-dim latent (amortized: it sees the current target, as standard), a prior head p(zt ∣ h<t)p(z_t\,|\,h_{<t}) (used for sampling and rollout), and a decoder mapping [zt;ht−1][z_t; h_{t-1}] to a Gaussian likelihood over the 200 raw waveform dimensions (μ\mu, σ\sigma per dim). Training is teacher-forced ELBO (Gaussian waveform NLL + KL) on the identical data pipeline, batch size, learning rate and 10-epoch budget as the paper's dual model. Capacity: trunk 5.81M (identical architecture), variational heads +0.34M, total 6.15M parameters.

Our priors, stated before the numbers: (a) the physiological parameterization gives every prediction error a mechanistic reading (band power, phase locking, aperiodic slope), whereas latent errors are unattributable; (b) it yields a closed-form likelihood and exact sampling, enabling the zero-shot cross-dataset protocol; (c) our signal-only ablation — a soft version of "learn the representation implicitly" — trails the dual model by 1.6–7.0 nats (Table 1). The complete results are given in our response to Question 1 below.

AuthorsOpenReview ↗

Response to Reviewer Eimr [2/N]

Weakness 2

The validation of distribution parameter dynamics and their physiological significance is insufficiently systematic. Since the manually defined distribution forms constrain the model's representational space, the ability of these parameters to effectively capture actual EEG variations is crucial to the method's validity. Although the appendix provides local analyses such as spectral slope correlation, the paper has not yet systematically shown the time variation curves of various parameters, as well as the correspondence between the predicted parameters and the actual EEG estimated parameters. Consequently, there is insufficient evidence to confirm that these parameters possess the claimed physiological interpretability.

Response: Agreed — the appendix's local analyses (e.g., the spectral-slope correlation) were a start, but a systematic validation was needed, and we built one: a trajectory analysis of the predicted physiological parameters (predicted vs. observed band amplitudes, phases, and 1/ff slope over 30-s epochs) — on held-out ISRUC and CHB-MIT records we compare the predicted parameter trajectories θ^(t)\hat\theta(t) against moment-matched references computed from the observed signal, with unsupervised brain-state segments (kk-means, k=3k=3, on per-patch band relative power; no external labels). Pooled Pearson correlations (per-record and pooled):

DatasetParameterrr (record 0)rr (record 1)rr (pooled)nn
ISRUCσδ\sigma_\delta0.740.530.761160
ISRUCσθ\sigma_\theta0.750.680.751160
ISRUCσβ\sigma_\beta0.820.710.841160
ISRUCσα\sigma_\alpha0.610.640.621160
ISRUCσγ\sigma_\gamma0.860.580.871160
ISRUCα\alpha (1/ff slope)0.710.530.681160
ISRUCκθ\kappa_\theta0.120.050.141160
CHB-MITσδ\sigma_\delta0.850.730.7618
CHB-MITσθ\sigma_\theta0.340.780.5618
CHB-MITσβ\sigma_\beta0.390.790.7018
CHB-MITσα\sigma_\alpha−0.140.630.5218
CHB-MITσγ\sigma_\gamma−0.330.960.5518
CHB-MITα\alpha (1/ff slope)0.100.040.1818
CHB-MITκθ\kappa_\theta−0.130.230.0718

Band-amplitude and slope trajectories track the observed signal (ISRUC σβ\sigma_\beta: r=0.84r=0.84); the CHB-MIT α\alpha/κ\kappa rows are limited by the 10-s test windows (18 scored patches), the σα\sigma_\alpha/σγ\sigma_\gamma rows show the same record-level instability there (record-0 rr −0.14-0.14/−0.33-0.33, pooled 0.520.52/0.550.55), and the κ\kappa rows use a moment estimator with a known downward concentration bias (quantified in our response to Question 4 below). Figure 3 (App. M, p. 29) shows the predicted trajectories following sleep-state transitions of the observed signal. The figure and this table — which now covers all five band amplitudes — are included in the revision, highlighted in blue (App. M: Figure 3, p. 29; Table 14, p. 29; the CHB-MIT panel as Figure 5, p. 31).

AuthorsOpenReview ↗

Response to Reviewer Eimr [3/N]

Question 1

Latent prediction offers an alternative to directly regressing highly stochastic EEG waveforms. Could the authors further elaborate on the specific advantages of predicting parameters of a predefined physiological distribution versus predicting a learned latent representation? Would it be possible to provide a simple comparative experiment against probabilistic latent prediction.

Response: Yes — the conditional-VAE baseline described in our response to Weakness 1 above is exactly this comparative experiment, and it is now complete. The evaluation protocol is identical to the paper's dual model. (a) One-step likelihood on the paper's scored split (full test set): to make the comparison direct, we score the CVAE through the same family lens the paper uses for the signal-only ablation — the paper already scores the dual head (parameters predicted directly) and signal-only (parameters induced by moment-fitting the predicted waveform) with one fixed closed-form family likelihood and Energy Score — so we moment-fit the CVAE's predictive samples into physiological parameters θ^\hat\theta and apply the identical scoring; the lens columns below are therefore directly comparable with the dual column, exactly as dual vs. signal-only is in Table 1. The CVAE's own waveform-space likelihood (importance sampling, K=500K{=}500; and −-ELBO) is reported as a supplementary column in its native units. (b) Rollout on the full test split under the paper's per-dataset context/rollout lengths: zz sampled from the prior, the decoded waveform fed back into the history with block-causal re-encoding at every step, plus an RMS-rescaled variant replicating the paper's rescaling row. (c) Family-lens calibration: 200 prior samples per position moment-fit into θ^\hat\theta, on which we run the paper's PIT machinery (per-band KS), compared against the dual model's published calibration numbers.

One-step predictive likelihood (full test set, paper's scored split; NLL / Energy Score):

DatasetDual (family) NLL / ESCVAE (pooled) NLL / ESCVAE (point) NLLSignal-only NLL / ESCVAE waveform NLL (IS / −-ELBO)
ISRUC4.308 / 1.2096.669 / 1.206155.33711.324 / 1.286−345.1 / −389.9
CHB-MIT5.523 / 4.5076.853 / 4.311209.17310.893 / 4.771−111.9 / −159.0
SHU-MI2.932 / 0.2353.105 / 0.24431.7094.928 / 0.478−568.8 / −598.3
FACED4.777 / 0.7975.463 / 0.78587.3858.040 / 0.877−342.6 / −381.7
SEED-V4.057 / 1.2774.340 / 1.25389.7535.759 / 1.340−277.9 / −317.0
SEED-VII4.365 / 1.7954.568 / 1.745116.3075.938 / 1.893−233.3 / −270.5
PhysioNet-MI5.050 / 2.8785.742 / 2.846177.8397.727 / 3.070−107.3 / −146.5
BCIC-IV-2a4.400 / 0.7365.475 / 0.84558.9517.393 / 0.800−334.7 / −380.4
MODMA3.971 / 0.8664.403 / 0.84457.3246.108 / 0.932−359.7 / −392.3
MentalArithmetic4.511 / 0.7676.680 / 0.81775.1259.473 / 0.834−331.3 / −376.1
Mean4.390 / 1.5075.330 / 1.490105.8907.758 / 1.628−301.3 / −341.2

(Bold = lower family NLL per row. CVAE-point = the prior-mean latent decoded to a point waveform and moment-fit, the verbatim analogue of the signal-only treatment; CVAE-pooled = 200 prior samples moment-fit per position. The last column is the CVAE's NLL under its own Gaussian waveform likelihood — strictly proper but in different units, shown only as supplementary and never differenced.)

The dual head attains the lower family NLL on 10/10 datasets (mean 4.390 vs 5.330), and the CVAE in turn beats the signal-only ablation on 10/10 (mean 5.330 vs 7.758): a learned continuous bottleneck does help over raw waveform prediction, but the physiological parameterization helps more. The sample-based Energy Score is close (CVAE 1.490 vs dual 1.507, CVAE better on 7/10), and the point-decoded variant collapses (NLL 31.709–209.173), confirming the CVAE's value is distributional rather than point-predictive.

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

Zero-Shot Generation of the Next Moment of Brain Activity by Predicting Its Distribution proposes a model that predicts neural data, EEG in this case, based on an interesting parametrization. While the parametrization itself is not novel, applying it to the large Transformer setting is. The baselines are simple but show very strong evidence of the effectiveness of the model. The main issue is that the manuscript is written with excessive slight-of-hand, hiding unpleasant results and putting fundamental sections like related works in the Appendix to gain space in the main sections. The work is strong enough not to need this.

Critical strengths
  1. The model is simple but effective, and is very easy to understand
  2. The parametrization is not novel, but the application is clever and works well
  3. The breadth of evaluation datasets is sufficient to make the case
  4. The end-to-end discriminative fine-tuning setup is good validation, though it’s hidden in the Appendix
Critical weaknesses
  1. Too much material is hidden in the Appendices, including related works and limitations: overall the manuscript is poorly organized and this hinders the message
  2. The two-head architecture, advertised as a central contribution, is shown to be mostly ineffective in Table 10; again, the structure of the paper is disingenuous in this, as this result is hidden from Table 1: there is no “dist-only” ablation which would clearly show what I mentioned above
Questions for authors
  1. Stronger ablations are necessary to establish the effectiveness of the two-head architecture
  2. There are bib entries with x and xx (I understand it might be difficult to cite some datasets, but this is not the right way)
3 comments
AuthorsOpenReview ↗

Response to Reviewer VZRV [1/N]

Weakness 1

Too much material is hidden in the Appendices, including related works and limitations: overall the manuscript is poorly organized and this hinders the message

Response: We apologize for the organizational choices that hindered readability; the placement of Related Work and Limitations in the appendix was a page-budget decision, and we agree it damaged the presentation. We restructured the revised submission accordingly, with all changes highlighted in blue: (a) Related Work is condensed and moved into the main text before the Method section (Sec. 2, p. 3); (b) a Limitations subsection now closes the main-text Conclusion (Sec. 5), with a pointer to the full discussion (App. R, p. 37); (c) an explicit positioning paragraph against the closest recent systems (Neuro-GPT, CodeBrain) is included in the appendix discussion (App. R, p. 37).

Weakness 2

The two-head architecture, advertised as a central contribution, is shown to be mostly ineffective in Table 10; again, the structure of the paper is disingenuous in this, as this result is hidden from Table 1: there is no “dist-only” ablation which would clearly show what I mentioned above

Response: We believe Table 10 supports the design rather than undermining it, and we take seriously that our presentation suggested otherwise — the key ablation living in an appendix table with no dist-only column in the main table was misleading, and we fixed both in the revision (dist-only column added to Table 1, p. 8; the head-ablation table and a new paired-bootstrap significance test are reported in App. L, Tables 11–12, p. 27). The reading we intend — stated at the end of Sec. 4.1 — is that dist-only vs. dual NLL parity is the success criterion of the design: the reconstruction head must come at zero cost to the distributional likelihood while adding capabilities that a single head cannot provide. To substantiate "zero cost" beyond a point estimate, we ran a paired bootstrap over per-sequence one-step NLL of the distribution-only head vs. the dual-head model on held-out sequences (2,000 resamples, seed 0; $\Delta = $ NLL(dual) −- NLL(dist) per sequence, negative = dual better; dist/dual = the paper's 10-epoch distribution-only / dual-head checkpoints):

DatasetnnNLL distNLL dualΔ\Delta (dual$-$dist)95% CIppsig.
ISRUC1603.9453.914−0.031[−0.039, −0.023]<0.001yes
CHB-MIT84.7234.744+0.021[+0.010, +0.032]0.001yes
SHU-MI82.8562.855−0.001[−0.019, +0.014]0.923no
FACED85.1445.101−0.043[−0.087, −0.009]0.006yes
SEED-V474.8074.807−0.000[−0.011, +0.011]0.978no
SEED-VII83.5603.565+0.005[−0.002, +0.012]0.175no
PhysioNet-MI85.5375.527−0.009[−0.033, +0.010]0.382no
BCIC-IV-2a84.0764.062−0.014[−0.023, −0.004]0.006yes
MODMA84.2904.290+0.000[−0.004, +0.005]0.928no
MentalArithmetic85.0625.045−0.017[−0.031, +0.001]0.063no

The mean difference is at most 0.05 nats in absolute value on every dataset and is not statistically significant on 6/10 (p>0.05p>0.05); where it is significant, dual is better on three (ISRUC −0.031, FACED −0.043, BCIC-IV-2a −0.014) and dist is better on one (CHB-MIT +0.021). Most per-dataset splits contain n=8n=8 held-out sequences (ISRUC 160, SEED-V 47), so the non-significant CIs are correspondingly wide — we report them as they are. The reconstruction head therefore does not degrade the likelihood (and slightly improves it on 3/10), while the union provides: (i) signal-reconstruction quality on par with a dedicated signal-only checkpoint (Table 11, SNR ≫\gg persistence); (ii) tail rescue that the dist head cannot express (SHU-MI excess kurtosis +369.4 → +117.8, −68%; Table 3 documents the over-light tails that motivate this); (iii) transient injection that lifts rollout spectral correlation on 9/10 datasets (+0.006 to +0.053) and repairs geometric attenuation (ISRUC overshoot 2.45 → 1.87; Tables 3 and 13). No single-head variant provides this combination.

AuthorsOpenReview ↗

Response to Reviewer VZRV [2/N]

Question 1

Stronger ablations are necessary to establish the effectiveness of the two-head architecture

Response: The stronger statistical test you ask for is exactly the paired bootstrap over held-out sequences reported in our response to Weakness 2 above (2,000 resamples, per-dataset 95% CIs and pp-values). We also surfaced the ablation in the revised paper as described there: the dist-only column is added to Table 1 (p. 8) and the head-ablation table with the new paired-bootstrap test is reported in App. L (Tables 11–12, p. 27), so the parity result and the union's added capabilities are easy to find.

Question 2

There are bib entries with x and xx (I understand it might be difficult to cite some datasets, but this is not the right way)

Response: Thank you for catching this. The two OpenNeuro dataset entries (ds006367, ds006370) carried placeholder author fields; we have retrieved the correct contributor metadata and replaced them in the revised bibliography.

The Authors have mostly solved the presentation issues in the new revision, given the now wider page limit. Still, fitting everything in 9 pages is part of the work and should not be sidestepped in this way. Moreover, the overly positive framing of some contributions has remained even in spite of evidence pointing to, at best, no effect.

Indeed, regarding the dual-head contribution, the Authors now provide strong evidence that it does not hinder NLL, and none that it actually improves reconstruction in a meaningful way. Concretely:

  1. Table 2 shows equivalent kurtosis on 6/10 datasets, worse on 3/10, and better on 1/10.
  2. The dual-head reconstruction reaches an SNR comparable to the dedicated signal-only checkpoint while substantially exceeding a persistence reconstruction: I couldn't find any SNR results.
  3. Table 11 shows again that dual-head is inconsistent, better on 4/10 in psd_corr and 3/10 in rmsR.
  4. Table 12 should show the statistical analysis of Table 1, but the numbers do not match, e.g. −0.066 gap for SHU-MI (the largest) becomes −0.001. It might be that the splits are different, but then the comparison is not correct.

Therefore, there is still no strong evidence that the dual-head design, discussed multiple times as a major contribution, actually provides meaningful improvements.

Reviewer_S638OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes GenBrain which is an autoregressive model of EEG generation. Instead of predicting the next brain state / EEG activity chasing noise fluctuations, GenBrain uses distribution prediction based on a suite of neuroscientific parameters e.g. 1/f

Critical strengths

The paper is well-written. The motivation of the architecture is well-grounded to the observation that continuous EEG (which is dominated by noise) forces generation models to chase high frequency components to predict the next time point. The authors describe the methodology and the theoretical principles in great detail.

Critical weaknesses

The current hypothesis is that high noise levels is the bottleneck for EEG prediction. The authors do not discuss or take into account the non-stationarity of EEG signal. Unlike other domains (e.g. language) or other more periodic time-series (e.g. energy consumption), EEG signals are dominated by also unpredictable changes due to subject’s task, mental state etc.

In addition, the model is mostly based on metrics that are influenced by the noise levels for example FFT, bands, etc.

Questions for authors

How could the authors include the non-stationarity into their model training?

If the prediction performance comes from providing a bottleneck (predict distributions rather than signal), how about a comparison with EEG tokenizers that use a bottleneck (e.g. codebook-based tokenizers)?

Downstream performance in Appendix is low compared to other models. If GenBrain predicts future signal activity better (as claimed), wouldn’t we expect that to translate to better downstream task performance?

5 comments
AuthorsOpenReview ↗

Response to Reviewer S638 [1/N]

Weakness 1

The current hypothesis is that high noise levels is the bottleneck for EEG prediction. The authors do not discuss or take into account the non-stationarity of EEG signal. Unlike other domains (e.g. language) or other more periodic time-series (e.g. energy consumption), EEG signals are dominated by also unpredictable changes due to subject’s task, mental state etc.

Response: We agree. High noise levels are one bottleneck — they are the original motivation for predicting distributions rather than chasing waveform fluctuations — but they are not the only one: task- and state-driven non-stationarity is an equally central difficulty, and conditional distribution prediction is designed to handle exactly this. GenBrain never assumes stationarity: every parameter of the predictive density θ^t+1=fϕ(x≤t)\hat\theta_{t+1}=f_\phi(x_{\le t}) is a function of the history, so the predictive distribution adapts over time. The evidence is already in Table 1: the gap between the conditional model and the best stationary forecaster (marginal) is largest exactly on the datasets where brain state varies most — ISRUC (sleep stages, +7.0 nats), CHB-MIT (seizure dynamics, +5.4), MentalArithmetic (+5.4) — i.e., non-stationarity increases the advantage of conditional prediction rather than invalidating it. A new direct analysis of predicted parameter trajectories against the observed signal is presented in our response to Question 1 below.

Weakness 2

In addition, the model is mostly based on metrics that are influenced by the noise levels for example FFT, bands, etc.

Response: We take this point in both of its parts. On the model's basis: our parameters are indeed windowed spectral estimates (per-band amplitudes from the FFT, 1/ff slope), chosen deliberately because they are physically interpretable, moment-matched statistics of the observed signal rather than raw noisy samples; the dual head additionally carries a raw-waveform residual path with no spectral quantity in it. On the metrics: the primary metric, the log score, is a strictly proper scoring rule evaluated under one fixed likelihood for every forecaster: a method that mis-models the noise — including any noise influence that remains in our spectral targets — is penalized by it, not shielded from it. The PIT diagnostics in Table 2 (p. 8) go in the opposite direction of hiding noise effects — they are exactly the tool that exposes where our noise model is imperfect (over-light tails), which then motivates the dual-head design. To remove any dependence on spectral quantities entirely, we added a purely time-domain, sample-based evaluation to the revision, highlighted in blue: the Energy Score (ES-CRPS), a strictly proper scoring rule evaluated on waveform samples, on all 10 datasets (App. K.2, Table 10, p. 25). GenBrain attains the lowest ES under both checkpoints on 10/10 datasets (per-row minimum in bold):

DatasetGenBrain (dist)GenBrain (dual)MarginalPersistenceAR
ISRUC1.1621.1621.2771.3311.239
CHB-MIT1.8091.8062.0102.0361.917
SHU-MI0.2560.2550.4830.4990.515
FACED1.0651.0661.1611.2371.161
SEED-V2.7962.8043.0853.1023.037
SEED-VII0.7820.7830.8950.9190.912
PhysioNet-MI3.4503.4503.6733.9223.852
BCIC-IV-2a0.4970.4960.6200.6370.640
MODMA0.8510.8520.9310.9640.926
MentalArithmetic0.7930.7920.8940.9180.897

(dist/dual = the paper's 10-epoch distribution-only / dual-head checkpoints; AR is second-order on SHU-MI, PhysioNet-MI and BCIC-IV-2a, third-order elsewhere, matching the paper's baselines.) The NLL conclusions do not depend on spectral metrics.

AuthorsOpenReview ↗

Response to Reviewer S638 [2/N]

Question 1

How could the authors include the non-stationarity into their model training?

Response: The mechanism is the history conditioning described in our response to Weakness 1 above; to make its effect visible rather than inferential, we added a trajectory analysis of the predicted physiological parameters (predicted vs. observed band amplitudes, phases, and 1/ff slope over 30-s epochs): on held-out ISRUC and CHB-MIT records we compare the predicted parameter trajectories θ^(t)\hat\theta(t) against moment-matched references computed from the observed signal, with unsupervised brain-state segments (kk-means, k=3k=3, on per-patch band relative power; no external labels). Pooled Pearson correlations between predicted and observed trajectories:

DatasetParameterrr (record 0)rr (record 1)rr (pooled)nn
ISRUCσδ\sigma_\delta0.740.530.761160
ISRUCσθ\sigma_\theta0.750.680.751160
ISRUCσβ\sigma_\beta0.820.710.841160
ISRUCσα\sigma_\alpha0.610.640.621160
ISRUCσγ\sigma_\gamma0.860.580.871160
ISRUCα\alpha (1/ff slope)0.710.530.681160
ISRUCκθ\kappa_\theta0.120.050.141160
CHB-MITσδ\sigma_\delta0.850.730.7618
CHB-MITσθ\sigma_\theta0.340.780.5618
CHB-MITσβ\sigma_\beta0.390.790.7018
CHB-MITσα\sigma_\alpha−0.140.630.5218
CHB-MITσγ\sigma_\gamma−0.330.960.5518
CHB-MITα\alpha (1/ff slope)0.100.040.1818
CHB-MITκθ\kappa_\theta−0.130.230.0718

Band-amplitude and slope trajectories follow the observed signal closely (ISRUC σβ\sigma_\beta: r=0.84r=0.84); the CHB-MIT α\alpha/κ\kappa rows are limited by the 10-s test windows (18 scored patches), the newly added σα\sigma_\alpha/σγ\sigma_\gamma rows show the same record-level instability there (record-0 rr −0.14-0.14/−0.33-0.33, pooled 0.520.52/0.550.55), and the κ\kappa rows use a moment estimator with a known downward concentration bias (quantified in our phase-concentration reference diagnostics). On transition alignment, predicted-parameter steps at detected state changes exceed the within-state baseline for most parameters in ISRUC record 0 but not in record 1:

ISRUCrecord 0: step@change / baseline (ratio)record 1: step@change / baseline (ratio)
σδ\sigma_\delta1.05 / 0.90 (1.17)0.20 / 0.28 (0.69)
σθ\sigma_\theta0.37 / 0.33 (1.12)0.07 / 0.11 (0.63)
σβ\sigma_\beta0.27 / 0.28 (0.96)0.04 / 0.06 (0.72)
α\alpha0.07 / 0.07 (1.09)0.04 / 0.05 (0.85)
κθ\kappa_\theta0.06 / 0.05 (1.18)0.02 / 0.03 (0.79)

We therefore read the transition alignment as suggestive only and rest the case on the trajectory correlations. Figure 3 (App. M, p. 29) shows the predicted trajectories following sleep-state transitions of the observed signal; the figure and both tables are included in the revision, highlighted in blue (App. M, Tables 14–15, pp. 29–30; the CHB-MIT panel is Figure 5 in App. M, p. 31), and the trajectory table now covers all five band amplitudes. An explicit latent-state (e.g., switching) mechanism is a natural next step and is noted in Future Work.

AuthorsOpenReview ↗

Response to Reviewer S638 [3/N]

Question 2

If the prediction performance comes from providing a bottleneck (predict distributions rather than signal), how about a comparison with EEG tokenizers that use a bottleneck (e.g. codebook-based tokenizers)?

Response: We appreciate the suggestion. The directly comparable baseline for this question is a forecaster built on a learned bottleneck, and we report it here (included in the revision as App. N, Tables 17–18, p. 33):

We trained a conditional-VAE baseline sharing the same trunk architecture, data pipeline, and the full 10-epoch training budget as our distribution head (64-d Gaussian latent, Gaussian waveform decoder, ELBO-trained; 6.15M vs 5.81M parameters). To make the comparison direct we scored it through the same family lens the paper already uses for the signal-only ablation: Table 1 compares the dual head (parameters predicted directly) with signal-only (parameters induced by moment-fitting the predicted waveform) under one fixed closed-form family likelihood — we apply the identical treatment to the CVAE, moment-fitting its predictive samples into physiological parameters and scoring them with the same likelihood and the same Energy Score, so the numbers below are directly comparable exactly as dual vs. signal-only is. One-step predictive likelihood on the paper's scored split (full test set; NLL / Energy Score):

DatasetDual (family) NLL / ESCVAE (pooled) NLL / ESCVAE (point) NLLSignal-only NLL / ESCVAE waveform NLL (IS / −-ELBO)
ISRUC4.308 / 1.2096.669 / 1.206155.33711.324 / 1.286−345.1 / −389.9
CHB-MIT5.523 / 4.5076.853 / 4.311209.17310.893 / 4.771−111.9 / −159.0
SHU-MI2.932 / 0.2353.105 / 0.24431.7094.928 / 0.478−568.8 / −598.3
FACED4.777 / 0.7975.463 / 0.78587.3858.040 / 0.877−342.6 / −381.7
SEED-V4.057 / 1.2774.340 / 1.25389.7535.759 / 1.340−277.9 / −317.0
SEED-VII4.365 / 1.7954.568 / 1.745116.3075.938 / 1.893−233.3 / −270.5
PhysioNet-MI5.050 / 2.8785.742 / 2.846177.8397.727 / 3.070−107.3 / −146.5
BCIC-IV-2a4.400 / 0.7365.475 / 0.84558.9517.393 / 0.800−334.7 / −380.4
MODMA3.971 / 0.8664.403 / 0.84457.3246.108 / 0.932−359.7 / −392.3
MentalArithmetic4.511 / 0.7676.680 / 0.81775.1259.473 / 0.834−331.3 / −376.1
Mean4.390 / 1.5075.330 / 1.490105.8907.758 / 1.628−301.3 / −341.2

(Bold = lower family NLL per row. CVAE-point = the prior-mean latent decoded to a point waveform and moment-fit, the verbatim analogue of the signal-only treatment; CVAE-pooled = 200 prior samples moment-fit per position. The last column is the CVAE's NLL under its own Gaussian waveform likelihood (importance sampling, K=500K{=}500; and −-ELBO) — strictly proper but in different units, shown only as supplementary and never differenced.)

Three readings: (i) an equally-capacious learned latent bottleneck trained on the identical budget yields a strictly worse predictive density than the physiological parameterization (lower family NLL on 10/10 datasets) — the advantage does not come merely from "having a bottleneck," but from which structure the bottleneck preserves (amplitude precision and phase continuity); (ii) the CVAE still beats the signal-only ablation on 10/10 (mean 5.330 vs 7.758), so a learned continuous bottleneck does help over raw waveform prediction — the parameterization simply helps more; (iii) the point-decoded variant collapses (NLL 31.709–209.173), confirming the CVAE's value is distributional rather than point-predictive, which is also why the sample-based Energy Score is close between CVAE and dual (mean 1.490 vs 1.507). The full-test rollout comparison shows the same ordering (the dual model attains higher spectral fidelity and stable amplitudes where the CVAE's prior drifts), and the calibration comparison is mixed (the CVAE better on the α\alpha band, the dual model better on θ\theta/β\beta); both tables are included in the revision (Tables 17–18, App. N, p. 33; calibration comparison in App. N, Table 19, p. 34).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank all reviewers for recognizing the motivation and neurophysiological grounding of the work (S638: "well-grounded"; VZRV: "simple but effective"), the breadth of the evaluation (VZRV: "sufficient to make the case"; Eimr: "comprehensive experimental evidence makes the proposed method more convincing"), and the value of validating generative pretraining by discriminative fine-tuning (S638).

During the discussion period we ran new analyses and experiments, summarized below; all details are given inline in the itemized responses, and every addition is highlighted in blue in the revised PDF at the locations noted in the right-hand column.

New evidenceAnswersScopeHeadlineIn the revision (revised PDF)
Predicted-parameter trajectory analysis (predicted vs. observed band amplitudes / phases / 1/ff slope)non-stationarity & parameter dynamicsISRUC, CHB-MIT (held-out); training-progress grid over all 10 datasetspooled rr up to 0.84 (σβ\sigma_\beta), above a persistence reference on every dataset; the margin emerges within the first three epochs and is stable thereafterApp. M (Figure 3, p. 29; Tables 14–15, pp. 29–30; CHB-MIT panel Figure 5, p. 31); training-stage analysis App. M (Figure 4, Table 16, p. 30; grid Figure 6, p. 32)
Amplitude distribution-family fit (Rayleigh / Weibull / Rice)distribution-family flexibility (Weibull)SHU, ISRUC, FACED (4 records each)Weibull cuts envelope KS by 26–59%, but gains sit in the body (core 27–60%), not the tail (8–31%)App. O (Tables 20–21, p. 35)
Energy Score (time-domain, sample-based, strictly proper)noise-robust, non-spectral evaluationall 10 datasetslowest ES on 10/10 under both checkpointsApp. K.2 (Table 10, p. 25)
Paired bootstrap over per-sequence one-step NLL (dist-only vs. dual head)dual-head ablationall 10 datasets∥Δ∥≤0.05\|\Delta\| \le 0.05 nats; non-significant on 6/10App. L (Table 12, p. 27)
Phase-concentration (κ\kappa) reference diagnostics (moment vs. short-window Hilbert estimates)phase-concentration referenceSHU, ISRUC (8 records each)the two standard κ\kappa references differ by two orders of magnitudeApp. P (Table 22, p. 35)
Conditional-VAE latent-prediction baseline (same trunk, 64-d Gaussian latent, ELBO-trained)probabilistic latent prediction & learned-bottleneck baselinesall 10 datasetsdual head wins one-step family NLL on 10/10; rollout gap dominated by amplitude calibrationApp. N (Tables 17–19, pp. 33–34)

We respond to each weakness and each question individually below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.