acceptodds
Under review as a conference paper at ICLR 2027

Sensor Geometry as a Flow-Matching Prior for Multi-Channel Brain Signals

Abstract

Flow-matching models start from an isotropic Gaussian source, the standard choice when the correlation structure of the data is unknown in advance. For multi-channel brain recordings, however, part of this structure is known in advance. Electrodes sit at fixed positions on the head, and volume conduction through the skull and scalp makes nearby electrodes co-vary in a way that is shared across subjects. Existing EEG generative models nonetheless leave the network to learn this from scratch. We put this structure into the source instead. From the sensor coordinates alone, we build a -nearest-neighbor graph and take a graph-Mat\'ern function of its Laplacian as the source covariance, so the flow starts from spatially coherent patterns rather than channel-independent noise. The change adds no learned parameters, works with any coupling and any drift network, and uses the same three hyperparameters on every dataset. Across eight EEG datasets and four flow-matching methods, the graph-Mat\'ern source lowers the spectral discrepancy between generated and real signals in the five clinical bands (PSD-KL) on most datasets. PSD-KL falls by % to % in geometric mean over datasets depending on the method and by up to % on PhysioNet-MI, the densest montage. We show that the improvement stems from the spatial eigenvectors of the local graph of sensor positions, since randomizing the eigenvectors while preserving the eigenvalue spectrum eliminates the gain. Furthermore, a prior fitted directly to the empirical data covariance performs worse than isotropic noise. The same construction applies unchanged to MEG, intracranial EEG with patient-specific grids, and a traffic-sensor network, lowering PSD-KL for every method on each.

Primary area:
applications to neuroscience & cognitive science
Keywords:
neuroscience, Graph-Matérn Prior, flow matching, EEG

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_cH53OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a geometry-informed Gaussian source distribution for flow-matching models of multichannel brain signals. It constructs a nearest-neighbor graph from sensor coordinates and defines the source covariance through a Matérn function of the normalized graph Laplacian, while retaining the original training objective, coupling procedure, and drift architecture. Across eight EEG datasets and four flow-matching formulations, the method reports a 12 to 17% geometric-mean reduction in PSD-KL, with additional evaluations on MEG, intracranial EEG, traffic signals, and EEG classification augmentation. Ablations examine the contributions of the spatial eigenbasis, covariance spectrum, and graph construction, suggesting that alignment with local sensor geometry is an important factor in the observed improvements.

Critical strengths
  1. The authors introduce a physically motivated spatial bias through the source covariance without adding learned parameters or changing the flow-matching objective.
  2. Constructing this prior from coordinates alone and normalizing its total variance makes the intervention straightforward to implement and reasonably well controlled.
  3. The authors conducted comprehensive experiments comparing to multiple flow matching methods to showcase the good performance.
Critical weaknesses
  1. The comparison of this paper mainly focused on the flow matching baselines, but I think as a work about Multi-Channel Brain Signals generation, the authors should conduct comparison directly with the related works like JET, GPT2MEG, MEG-GPT, EEG-GAN etc. It can better showcase the improvement of generated signals.
  2. In the paper the authors claim that All methods share a 1D U-Net drift, but recent works show that transformer based backbone can have better performance. Can authors conduct simple experiments based on transformer backbone?
  3. The downstream evaluation is restricted to Mumtaz-MDD, the authors should provide more donstream evaluation results across different datasets.
  4. It seems that the major improvement of downstream task comes from adding iso-source, the graph-prior only gives limited improvement. The author should have further analysis to show that this prior is pratically usefull.
Questions for authors

Please refer to the weakness.

4 comments
AuthorsOpenReview ↗

Response to Reviewer cH53 (1/2)

We appreciate the reviewer's constructive feedback and address each concern below.

Re: The comparison of this paper mainly focused on the flow matching baselines, but I think as a work about Multi-Channel Brain Signals generation, the authors should conduct comparison directly with the related works like JET, GPT2MEG, MEG-GPT, EEG-GAN etc. It can better showcase the improvement of generated signals.

We trained EEGDM and JET on each of the eight datasets with their official codebases and default settings, with five random seeds, and evaluated them with the metrics of Tables 2 and 3 (Table cH53-1). EEGDM, which denoises one channel at a time, reaches a PSD-KL between 1.43 (PhysioNet-MI) and 17.94 (BCI-IV 2a), with standard deviations below 1 on every dataset. JET reaches 3.29 on TUAB and 6.59 on TUEV, the two datasets its recipe was developed on. On the other six its PSD-KL is between 23.11 and 1662.54, and its seeds differ widely (for example 43 to 546 on SEED-V). Samples of both models carry almost no coupling between channels. Their wPLI correlation is between -0.12 and 0.05 on every dataset.

Table cH53-1. EEGDM and JET, trained on our training splits with their released code and default settings and scored on Table 2's test windows. PSD-KL (lower is better) as in Table 2 and wPLI correlation (higher is better) as in Table 3, mean ± std over five random seeds.

Dataset(ch)EEGDM,PSD-KLJET,PSD-KLEEGDM,wPLIJET,wPLI
TUAB(16)5.50±0.323.29±0.84-0.02±0.020.03±0.04
TUEV(16)2.75±0.166.59±0.320.02±0.03-0.02±0.05
Mumtaz-MDD(19)2.46±0.36828.10±1063.290.02±0.02-0.01±0.02
BCI-IV 2a(22)17.94±0.22318.76±201.420.01±0.020.01±0.03
FACED(32)8.17±0.62121.34±43.31-0.01±0.01-0.01±0.03
SHU(32)9.49±0.311662.54±1297.400.01±0.01-0.00±0.01
SEED-V(62)1.88±0.02160.30±216.500.05±0.00-0.00±0.01
PhysioNet-MI(64)1.43±0.0623.11±21.57-0.12±0.010.01±0.02

Re: In the paper the authors claim that All methods share a 1D U-Net drift, but recent works show that transformer based backbone can have better performance. Can authors conduct simple experiments based on transformer backbone?

We appreciate the reviewer's suggestion. We tested a DiT-style transformer (patch 8, width 256, depth 8) with four flow-matching methods used in Table 2 (Table cH53-2). On the dense montages the pattern is the same as with the U-Net. The graph source lowers PSD-KL by 21 to 23% on SHU, by 19% on SEED-V and by 30 to 36% on PhysioNet-MI for SF2M, SI, OT-CFM and RF. On the five sparser montages the differences are mostly within seed variance, again as with the U-Net. At this size the transformer is the weaker generator, with a higher PSD-KL than the U-Net in 61 of the 64 cells (eight datasets, four methods, two sources; for SI on FACED, for example, 822 against 33), so the U-Net remains our main model. Because the gain does not appear on every dataset with either network, we will change "works with any coupling and any drift network" in the abstract to say that the prior requires no change to the coupling or the drift network.

Table cH53-2. Transformer drift in place of the U-Net, everything else as in Table 2, in Table 2's layout. PSD-KL, mean ± sample SD over five seeds (lower is better), for the four methods with the isotropic source and with the graph-Matérn source (GP-). Bold denotes the better source within each method.

Dataset(ch)SF2MGP-SF2MSIGP-SIOT-CFMGP-OT-CFMRFGP-RF
TUAB(16)1.57±0.561.90±0.291.34±0.171.90±0.671.71±0.341.87±0.463.89±1.562.96±0.96
TUEV(16)7.67±2.546.44±3.168.19±6.609.66±8.1710.71±7.595.34±1.803.69±0.866.45±5.80
Mumtaz-MDD(19)22.03±14.026.69±3.9218.83±13.327.92±7.3431.65±32.208.45±7.5036.42±59.787.41±3.76
BCI-IV 2a(22)6.01±0.305.73±0.536.19±0.436.52±0.576.24±0.285.80±0.395.92±0.155.72±0.75
FACED(32)668.63±335.94395.30±351.90821.91±13.52423.30±213.46791.31±45.99570.17±223.55822.39±5.78609.69±175.08
SHU(32)289.27±1.92225.76±3.84288.14±2.65227.66±3.19288.31±2.30220.74±12.84289.78±2.71227.06±1.84
SEED-V(62)51.88±0.2242.12±0.1551.88±0.1942.14±0.0951.89±0.1942.12±0.0551.95±0.1542.23±0.25
PhysioNet-MI(64)66.95±0.3544.03±2.2667.99±0.6245.56±0.9566.19±1.9842.42±3.7668.12±0.5647.41±1.46
AuthorsOpenReview ↗

Response to Reviewer cH53 (2/2)

Re: The downstream evaluation is restricted to Mumtaz-MDD, the authors should provide more donstream evaluation results across different datasets. Re: It seems that the major improvement of downstream task comes from adding iso-source, the graph-prior only gives limited improvement. The author should have further analysis to show that this prior is pratically usefull.

We agree with the reviewer's point that the improvement in Table 5 comes from adding generated data. The graph prior is slightly higher than the isotropic prior (0.833±0.0540.833 \pm 0.054 vs. 0.819±0.0370.819 \pm 0.037).

We ran the same experiment on the remaining seven datasets (Table cH53-3). Augmentation improves accuracy only on Mumtaz-MDD. It makes no difference on TUAB, TUEV, and SEED-V, and it decreases accuracy on BCI-IV 2a, FACED, SHU, and PhysioNet-MI, for both priors. When generated data from the isotropic prior does not help, generated data from the graph prior does not help either.

The graph prior is designed to improve generation quality. On the datasets with more channels, it lowers PSD-KL for all four methods (15%15\% on SEED-V, 33–41%33\text{–}41\% on PhysioNet-MI) and improves TS-FID and the autocorrelation error (Tables cH53-4 and cH53-5).

Table cH53-3. Downstream augmentation on each dataset with an EEGNet-8,2 classifier.

Dataset(ch)Real only+iso-source+graph-prior
TUAB(16)0.769±0.0300.764±0.0230.751±0.026
TUEV(16)0.462±0.0250.462±0.0490.453±0.031
Mumtaz-MDD(19)0.608±0.1460.819±0.0370.833±0.054
BCI-IV 2a(22)0.563±0.0320.374±0.0280.388±0.031
FACED(32)0.182±0.0270.132±0.0270.140±0.035
SHU(32)0.831±0.0090.816±0.0070.807±0.014
SEED-V(62)0.370±0.0230.356±0.0270.363±0.032
PhysioNet-MI(64)0.586±0.0130.536±0.0220.502±0.013

Table cH53-4. TS-FID (lower is better) for the four methods with the isotropic source and with the graph-Matérn source (GP-). Bold denotes the better source within each method.

Dataset(ch)SF2MGP-SF2MSIGP-SIOT-CFMGP-OT-CFMRFGP-RF
TUAB(16)473.17±11.66512.52±34.27473.09±22.31514.74±19.93468.35±25.85510.91±19.22590.37±79.91624.72±93.63
TUEV(16)658.04±22.40641.44±37.19651.20±19.27647.42±40.13652.32±29.64658.51±32.15664.17±59.77647.29±27.79
Mumtaz-MDD(19)514.88±22.17530.41±16.29497.56±47.23509.19±42.41492.32±30.10517.56±13.95522.95±28.13548.31±23.80
BCI-IV 2a(22)158.68±15.31180.03±28.93178.24±8.54176.22±14.87170.25±29.03176.08±23.68180.68±23.19183.23±33.82
FACED(32)313.52±10.06284.54±25.52307.77±17.40301.14±46.48302.73±17.05310.85±58.38314.23±26.85293.09±15.81
SHU(32)728.66±99.82520.43±62.73619.76±86.87531.10±53.68701.91±69.70550.46±101.70767.51±75.19553.10±45.42
SEED-V(62)171.01±22.9181.01±9.85155.52±15.2187.40±4.52173.85±8.2876.19±8.93152.03±34.8484.86±15.62
PhysioNet-MI(64)570.35±58.08277.21±54.57585.62±69.31287.35±62.64537.23±103.35300.79±57.54612.13±59.51323.20±62.61

Table cH53-5. Autocorrelation error, multiplied by 100 (lower is better).

Dataset(ch)SF2MGP-SF2MSIGP-SIOT-CFMGP-OT-CFMRFGP-RF
TUAB(16)0.68±0.230.81±0.180.56±0.180.84±0.320.73±0.310.81±0.261.27±0.291.21±0.31
TUEV(16)9.30±6.847.47±2.127.14±2.726.18±1.637.36±1.427.76±2.0410.64±1.978.80±1.87
Mumtaz-MDD(19)0.95±0.211.12±0.350.88±0.340.95±0.440.92±0.421.10±0.240.95±0.211.01±0.38
BCI-IV 2a(22)3.97±0.694.49±1.034.46±0.314.42±0.534.30±0.814.42±0.654.51±0.524.52±1.08
FACED(32)5.59±0.543.54±1.745.35±1.123.85±1.544.86±1.004.18±0.875.02±0.854.56±0.60
SHU(32)24.68±4.4622.80±1.6723.13±2.7921.62±2.6725.14±1.7720.87±7.2526.55±1.6523.02±1.29
SEED-V(62)7.35±0.096.46±0.277.29±0.146.67±0.127.39±0.026.37±0.207.25±0.206.51±0.36
PhysioNet-MI(64)5.84±0.574.25±1.125.98±0.764.46±1.335.06±1.494.39±1.356.75±0.525.11±0.88

I sincerely thank the authors for the additional experiments clarification; they address my concerns about baselines and backbones. However, since the gains are largely confined to high-channel montages and the downstream benefit does not generalize, I will maintain my current score.

We appreciate the reviewer's response.

Re: The gains are largely confined to high-channel montages

We agree that the gain is concentrated on the datasets with more channels. When the sensors are dense, neighboring channels are strongly correlated. The isotropic source ignores this correlation, while the graph prior starts from noise that already has it. Dense montages are therefore the setting in which the choice of source matters, and there the gain is consistent. On SEED-V and PhysioNet-MI the graph prior lowers PSD-KL for all four methods, by 15% and 33 to 41% with the U-Net (Table 2) and by 19% and 30 to 36% with the transformer (Table cH53-2).

Re: The downstream benefit does not generalize

Downstream accuracy is a useful proxy for sample quality, but more realistic samples do not always give better augmentation. On ImageNet, adding nearly photorealistic generated images to the training set improves top-5 accuracy by about 0.2% in the best case and otherwise degrades it, and FID does not predict classification accuracy [R1]. Samples from a good diffusion model can even harm contrastive learning [R2]. We observe the same. Generated data do not raise accuracy on seven of the eight datasets with either source (Table cH53-3). Our claim is about generation quality (PSD-KL, wPLI, TS-FID, and autocorrelation error).

References

[R1] Ravuri and Vinyals. Classification Accuracy Score for Conditional Generative Models. NeurIPS. 2019.

[R2] Wang, Zhang, and Wang. Do Generated Data Always Help Contrastive Learning? ICLR. 2024.

Reviewer_kE8SOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This article proposes a flow matching source distribution prior based on sensor geometry. This method constructs a k-nearest neighbor graph based on sensor coordinates, performs eigendecomposition of the graph Laplacian matrix, and uses spectral density to assign variances to different eigenmodes, thereby constructing a source distribution with spatial coherence.

Critical strengths
  • By utilizing sensor geometric priors to construct source distributions, this method does not require additional learnable parameters or changes to the training objectives, making it a plug-and-play feature.

  • By using methods such as random permutation, the authors systematically distinguished the three factors of eigenvectors, spectral shape, and graph structure, which convincingly demonstrates that the performance improvement comes from the physical geometric structure.

Critical weaknesses
  • The statistical significance after correction is limited, with only SI and RF being close to significance, while the rest are not significant.

  • The evaluation indicators used in the paper are limited. The paper avoids distance-based indicators, citing the argument that "FID is not applicable to EEG," but there is a problem of selective avoidance.

  • The prior only introduces structure in the spatial dimension, while maintaining isotropy in the temporal dimension.

Questions for authors

Please refer to weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer kE8S (1/3)

We appreciate the reviewer's constructive feedback and address each concern below.

Re: The statistical significance after correction is limited, with only SI and RF being close to significance, while the rest are not significant.

Thank you for pointing this out. While checking the results in Table 2, we found that the submission uses incorrect scalp positions for TUEV, and we apologize for this error. We reran TUEV with each bipolar channel at the midpoint of its two electrodes (Table kE8S-1), and SHU with a subject-disjoint split (Table kE8S-2). With these rows, the same one-sided Wilcoxon test across the eight datasets gives p=0.004p = 0.004 for SF2M, 0.0160.016 for SI, 0.0200.020 for RF, and 0.0980.098 for OT-CFM, or 0.0160.016, 0.0470.047, 0.0470.047, and 0.0980.098 after Holm correction. The graph prior has the lower mean in 27 of the 32 cells, with one tie.

Regarding the two methods that were not close to significance, the pp-value of SF2M in the submission (0.0980.098) came from a single dataset, TUEV. With the corrected row, SF2M has the lower mean on all eight datasets. For OT-CFM, the graph prior has the lower mean on six of the eight datasets. The two exceptions are Mumtaz-MDD (4.97→5.514.97 \to 5.51) and BCI-IV 2a (3.06→3.613.06 \to 3.61), which have 19 and 22 channels, and the gain of the graph prior depends on the number of channels. On SEED-V and PhysioNet-MI, OT-CFM and SF2M improve as much as SI and RF (15%15\% and 40–41%40\text{–}41\%).

We agree that a test over eight datasets has limited resolution and that the improvement is not uniform. It is concentrated on the datasets with more channels, where all four methods improve (15%15\% on SEED-V, 33–41%33\text{–}41\% on PhysioNet-MI). We will state this limitation in Section 3.2.

Table kE8S-1. TUEV (16 channels) as submitted and rerun with corrected channel positions. PSD-KL, mean ± sample SD over five seeds (lower is better). Bold denotes the better source within each method.

TUEV (16)SF2MGP-SF2MSIGP-SIOT-CFMGP-OT-CFMRFGP-RF
Submitted2.03 ± 0.333.85 ± 2.472.55 ± 0.782.15 ± 1.105.30 ± 2.303.80 ± 0.593.19 ± 1.213.09 ± 0.29
Corrected positions5.52 ± 2.473.32 ± 0.398.08 ± 7.466.16 ± 3.784.45 ± 1.214.14 ± 2.234.09 ± 2.153.26 ± 1.89

Table kE8S-2. SHU as submitted and rerun with the subject-disjoint split. PSD-KL, mean ± std over five seeds (lower is better). Bold denotes the better source within each method.

SHUSF2MGP-SF2MSIGP-SIOT-CFMGP-OT-CFMRFGP-RF
Submitted113.08 ± 3.9081.36 ± 8.01114.36 ± 1.7382.76 ± 4.57111.27 ± 5.1985.36 ± 0.71115.72 ± 2.2682.88 ± 10.33
Subject-disjoint194.55 ± 4.70134.02 ± 43.38182.22 ± 14.88136.96 ± 35.88169.20 ± 51.50139.93 ± 34.56167.27 ± 49.66151.56 ± 11.13
AuthorsOpenReview ↗

Response to Reviewer kE8S (2/3)

Re: The evaluation indicators used in the paper are limited. The paper avoids distance-based indicators, citing the argument that "FID is not applicable to EEG," but there is a problem of selective avoidance.

As we mentioned in L237 and L689–690, we avoided network-based FID because it relies on a pretrained feature extractor and because Hartmann et al. (2018) showed that learned-extractor FID on EEG can disagree with spectral and spatial signal fidelity. However, we agree that feature-based Fréchet distance and time-domain metrics provide valuable complementary evaluation without requiring a pretrained EEG network. To evaluate overall sample quality and temporal dynamics directly, we add two metrics across all four methods, TS-FID (the Fréchet distance on statistical time-series features used by JET, computed with JET's codebase) and autocorrelation error in Tables kE8S-3 and kE8S-4, respectively.

On the three dense montages, the graph prior improves both metrics for every method. It lowers TS-FID by 14–29% on SHU and by 44–56% on SEED-V and PhysioNet-MI, and lowers autocorrelation error by 7–17% on SHU, 9–14% on SEED-V, and 13–27% on PhysioNet-MI. On FACED, it also lowers autocorrelation error across all four methods (by 9–37%) and TS-FID for three of four. On the sparser 16–22 channel montages, the graph prior gives no improvement (e.g., TS-FID is 2–9% higher on TUAB and Mumtaz-MDD), mirroring the montage-density pattern observed on PSD-KL. We will clarify the scope of our FID discussion in Section 6.

Table kE8S-3. TS-FID (lower is better) for the four methods with the isotropic source and with the graph-Matérn source (GP-). Bold denotes the better source within each method.

Dataset(ch)SF2MGP-SF2MSIGP-SIOT-CFMGP-OT-CFMRFGP-RF
TUAB(16)473.17±11.66512.52±34.27473.09±22.31514.74±19.93468.35±25.85510.91±19.22590.37±79.91624.72±93.63
TUEV(16)658.04±22.40641.44±37.19651.20±19.27647.42±40.13652.32±29.64658.51±32.15664.17±59.77647.29±27.79
Mumtaz-MDD(19)514.88±22.17530.41±16.29497.56±47.23509.19±42.41492.32±30.10517.56±13.95522.95±28.13548.31±23.80
BCI-IV 2a(22)158.68±15.31180.03±28.93178.24±8.54176.22±14.87170.25±29.03176.08±23.68180.68±23.19183.23±33.82
FACED(32)313.52±10.06284.54±25.52307.77±17.40301.14±46.48302.73±17.05310.85±58.38314.23±26.85293.09±15.81
SHU(32)728.66±99.82520.43±62.73619.76±86.87531.10±53.68701.91±69.70550.46±101.70767.51±75.19553.10±45.42
SEED-V(62)171.01±22.9181.01±9.85155.52±15.2187.40±4.52173.85±8.2876.19±8.93152.03±34.8484.86±15.62
PhysioNet-MI(64)570.35±58.08277.21±54.57585.62±69.31287.35±62.64537.23±103.35300.79±57.54612.13±59.51323.20±62.61

Table kE8S-4. Autocorrelation error, multiplied by 100 (lower is better).

Dataset(ch)SF2MGP-SF2MSIGP-SIOT-CFMGP-OT-CFMRFGP-RF
TUAB(16)0.68±0.230.81±0.180.56±0.180.84±0.320.73±0.310.81±0.261.27±0.291.21±0.31
TUEV(16)9.30±6.847.47±2.127.14±2.726.18±1.637.36±1.427.76±2.0410.64±1.978.80±1.87
Mumtaz-MDD(19)0.95±0.211.12±0.350.88±0.340.95±0.440.92±0.421.10±0.240.95±0.211.01±0.38
BCI-IV 2a(22)3.97±0.694.49±1.034.46±0.314.42±0.534.30±0.814.42±0.654.51±0.524.52±1.08
FACED(32)5.59±0.543.54±1.745.35±1.123.85±1.544.86±1.004.18±0.875.02±0.854.56±0.60
SHU(32)24.68±4.4622.80±1.6723.13±2.7921.62±2.6725.14±1.7720.87±7.2526.55±1.6523.02±1.29
SEED-V(62)7.35±0.096.46±0.277.29±0.146.67±0.127.39±0.026.37±0.207.25±0.206.51±0.36
PhysioNet-MI(64)5.84±0.574.25±1.125.98±0.764.46±1.335.06±1.494.39±1.356.75±0.525.11±0.88
AuthorsOpenReview ↗

Response to Reviewer kE8S (3/3)

Re: The prior only introduces structure in the spatial dimension, while maintaining isotropy in the temporal dimension.

We agree that the prior is structured only in the spatial dimension, and we state this as a limitation in Section 6. We tested a temporal Matérn prior in Appendix B (Table 8), and its Kronecker product with the spatial prior on FACED (L726-729), where it has a lower PSD-KL than the spatial prior alone (27.80±8.0027.80 \pm 8.00 vs. 29.0329.03).

Table kE8S-5 compares spatial, temporal and spatio-temporal priors on all eight datasets with SI. The spatio-temporal prior achieves the best performance on FACED and Mumtaz-MDD and comes close on TUAB, where the temporal source already improves over isotropic noise on its own. However, on SHU, SEED-V, and PhysioNet-MI, as well as BCI-IV 2a, the spatial-only graph prior outperforms both the temporal and Kronecker priors. On SEED-V and PhysioNet-MI, the temporal prior alone is worse than isotropic noise and on SHU it is level with isotropic noise.

Table kE8S-5. Comparison of isotropic, spatial (graph-Matérn), temporal (Matérn), and spatio-temporal Kronecker priors in PSD-KL (lower is better; mean ± std over five seeds).

Dataset (ch)SI, isotropicSI, spatial (graph)SI, temporalSI, Kronecker
TUAB (16)1.43 ± 0.371.16 ± 0.200.90 ± 0.190.95 ± 0.22
TUEV (16)8.08 ± 7.466.16 ± 3.782.69 ± 0.923.64 ± 1.44
Mumtaz-MDD (19)3.91 ± 2.423.91 ± 1.112.27 ± 0.612.14 ± 0.74
BCI-IV 2a (22)3.30 ± 0.733.50 ± 0.593.69 ± 0.384.55 ± 1.76
FACED (32)33.34 ± 3.5429.03 ± 4.5632.86 ± 5.9327.80 ± 8.00
SHU (32)182.22 ± 14.88136.96 ± 35.88182.66 ± 43.70139.57 ± 38.73
SEED-V (62)32.93 ± 0.7128.02 ± 0.1336.80 ± 0.8031.48 ± 0.56
PhysioNet-MI (64)33.42 ± 0.8819.60 ± 2.3736.13 ± 1.8621.87 ± 4.85
Reviewer_bWwaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper starts from a simple observation: the source prior for flow-matching models is isotropic Gaussian noise, which is reasonable when nothing is known about the structure of the data, but there may be a better prior for multi-channel brain recordings, where the structure is partially known. Authors use the sensor coordinates, build a graph based on k-nearest neighbors, and use the structure of this graph through its Laplacian to define a source prior. The idea is to take a graph Matern function of the graph Laplacian as the source covariance. The approach is evaluated across eight EEG datasets and four flow-matching methods.

Critical strengths
  • Simple idea and well-motivated. I really like that the paper revisits a very basic assumption which is rarely questioned.
  • Substantial improvements on most datasets and improvements are robust to the choice of the flow matching model.
  • Algorithmic and experimental details are generally clear and the discussions are insightful.
  • Some ablations are very insightful, including failure cases.
Critical weaknesses
  • My major concern is that the comparison is limited to four flow matching models: SF2M, SI, OT-CFM, and RF. None of these baselines are graph-aware. Given the motivation of the paper is on structured sources and geometric priors in generative models, I expect comparisons with graph-aware methods, which are actually reviewed by the authors in the paper. While I understand that this paper does not propose any architectural changes, and differs in that respect from these works, the goal is closely related in my understanding. It would also be helpful to identify whether this particular method has any advantages over other graph-aware methods. Appendix F.5 has some graph mixer ablations in this direction but they are limited to TUAB and appear inconclusive to me.
  • Empirically, in some datasets, the source prior actually hurts performance. Authors are transparent about this and discuss likely explanations, but the performance degradation is quite substantial. I think a deeper analysis and follow up may be necessary.
Questions for authors

Please refer to weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer bWwa (1/2)

We thank the reviewer for their insightful comments and provide responses to each concern below.

Re: The comparison is limited to four flow matching models. I expect comparisons with graph-aware methods. Appendix F.5 has some graph mixer ablations in this direction but only TUAB

We extended the graph-mixer ablation (isotropic source) of Appendix F.5 from TUAB to all eight datasets (Table bWwa-1). The mixer places the same graph Laplacian inside the drift network, so the comparison shows where the graph should enter, in the source (GP-SI) or in the network.

On the three dense montages the graph helps in the source and not in the drift. GP-SI lowers PSD-KL by 10% on SHU, 15% on SEED-V and 38% on PhysioNet-MI, while the mixer changes it by +11%, +1% and -2%. On SHU every GP-SI seed lies below every mixer seed. On the five sparser montages neither placement is consistently better than plain SI. The mixer has the lowest mean on TUEV, GP-SI on Mumtaz-MDD, and plain SI on TUAB, BCI-IV 2a and FACED.

Table bWwa-1. PSD-KL on eight EEG datasets comparing SI, graph-mixer, and GP-SI.

Dataset(ch)SISI,graph-mixer driftGP-SI
TUAB(16)1.23±0.361.46±0.541.54±0.91
TUEV(16)3.97±1.093.23±1.433.43±1.88
Mumtaz-MDD(19)5.22±1.745.26±1.632.88±1.02
BCI-IV 2a(22)3.56±0.903.66±0.284.46±1.47
FACED(32)31.90±3.1732.99±6.4132.47±6.43
SHU(32)174.71±41.96194.23±3.97157.66±3.55
SEED-V(62)33.01±0.7733.42±0.8528.06±0.31
PhysioNet-MI(64)33.41±2.9332.60±2.3020.66±3.12
AuthorsOpenReview ↗

Response to Reviewer bWwa (2/2)

Re: Deeper Analysis for performance degradaution in some datasets

We agree that the graph prior does not improve every method on every dataset, and that "dataset-specific" (L308) was not an analysis. With more channels the graph prior consistently improves on the isotropic prior, while on the sparser montages the two priors are within seed spread of each other. We address this in three points.

  1. The gain depends on montage density (CC), not on sample count, σ\sigma, or kk. The graph prior has the lower mean for all four methods on the dense montages (SHU, SEED-V, PhysioNet-MI, and 102-sensor MEG), and makes no consistent difference on the sparser ones (16–3216\text{–}32ch). With few channels, each sensor's k=4k = 4 nearest neighbors already span a large part of the montage, so the graph carries little spatial structure beyond what the network learns by itself. We tested three alternative explanations.
  • Sample count: when SEED-V, PhysioNet-MI, and SHU are trained on only 1,4401{,}440 windows, the training-set size of BCI-IV 2a, they keep their gains (SI PSD-KL ratios 0.820.82, 0.580.58, and 0.630.63, with every graph-prior seed below every isotropic seed).
  • Graph connectivity: no k∈{2,…,8}k \in \{2,\dots,8\} gives a consistent gain on the sparse montages (Table bWwa-3).
  • Noise scale: on BCI-IV 2a with SI, the isotropic and the graph prior give 4.00±0.794.00 \pm 0.79 vs. 3.57±0.553.57 \pm 0.55 at σ=0.005\sigma = 0.005, 3.69±0.723.69 \pm 0.72 vs. 3.64±0.353.64 \pm 0.35 at 0.010.01, 3.51±0.553.51 \pm 0.55 vs. 4.01±0.874.01 \pm 0.87 at 0.020.02, and 3.77±0.773.77 \pm 0.77 vs. 3.46±0.483.46 \pm 0.48 at 0.040.04, all within standard deviation.

The cases on BCI-IV 2a and Mumtaz-MDD are within seed spread. On BCI-IV 2a the two sources differ by at most 0.550.55 against standard deviations of 0.2–1.50.2\text{–}1.5, and on Mumtaz-MDD OT-CFM is 4.97±2.064.97 \pm 2.06 vs. 5.51±1.665.51 \pm 1.66 while SF2M and RF have lower means with the graph prior (4.42→3.894.42 \to 3.89, 6.07→3.796.07 \to 3.79).

  1. TUEV: the SF2M degradation does not reproduce. While checking this case we found that the submission uses incorrect scalp positions for TUEV. We apologize for this error. We reran the whole row with each bipolar channel at the midpoint of its two electrodes (Table bWwa-2). In the rerun SF2M is 5.52±2.475.52 \pm 2.47 against 3.32±0.393.32 \pm 0.39 for GP-SF2M, lower in four of five seeds, and the graph prior has the lower mean for all four methods. TUEV varies strongly across seeds with either source (for SI, 8.08±7.468.08 \pm 7.46 against 6.16±3.786.16 \pm 3.78), and all four differences are within that spread, so we claim no gain on TUEV.

Table bWwa-2. TUEV (16 channels) as submitted and rerun with corrected channel positions. PSD-KL.

TUEV(16)SF2MGP-SF2MSIGP-SIOT-CFMGP-OT-CFMRFGP-RF
Submitted2.03±0.333.85±2.472.55±0.782.15±1.105.30±2.303.80±0.593.19±1.213.09±0.29
Corrected positions5.52±2.473.32±0.398.08±7.466.16±3.784.45±1.214.14±2.234.09±2.153.26±1.89
  1. FACED: the β/γ\beta/\gamma degradation appears only at the high noise scale of Figure 5. At the paper's setting (σ=0.02\sigma = 0.02, Table 2) the graph prior has the lower mean for all four methods (3%3\%, 13%13\%, 6%6\%, 18%18\%). Figure 5 reports FACED at σ=0.10\sigma = 0.10, five times the setting of Table 2. At that noise scale the graph prior matches or improves the lower bands and is worse in β\beta and γ\gamma.

Table bWwa-3. Ablation study of kk with SI on eight datasets in terms of PSD-KL.

Dataset(channels)SI,isotropicGP-SI,k=2k=3k=4k=5k=6k=8
TUAB(16)1.17±0.200.99±0.211.00±0.361.06±0.261.01±0.301.07±0.180.96±0.18
TUEV(16)8.08±7.464.05±1.045.81±4.676.27±3.357.14±5.223.70±1.305.96±5.26
Mumtaz-MDD(19)4.08±2.962.82±0.423.81±1.883.91±2.113.34±1.463.03±0.783.62±1.30
BCI-IV 2a(22)3.54±0.573.96±0.653.52±0.524.01±0.873.45±0.783.21±0.983.51±0.59
FACED(32)32.0±4.426.5±5.530.6±3.227.1±2.430.6±5.729.2±3.826.9±5.2
SHU(32)182.2±14.9126.5±47.5146.5±8.2151.8±7.1151.6±7.8155.3±5.9165.3±4.2
SEED-V(62)32.8±0.625.8±0.627.0±0.328.0±0.329.3±0.230.2±0.230.8±0.4
PhysioNet-MI(64)33.5±1.718.1±2.319.7±4.920.4±4.320.7±2.722.0±2.920.4±5.1

I thank the authors for extending their analysis to include a graph-aware baseline and also their explanations regarding the performance degradation on some datasets. I am satisfied with the results regarding graph-aware baselines. I am still unsure about the performance degradation. I thank the authors for being transparent in not being able to reproduce their initial result on TUEV. This is not going to affect my assessment. I am going to take the additional insights into account in my final assessment.

We really appreciate the reviewer's response, and we are glad that the graph-aware baseline concern is resolved.

Re: I am still unsure about the performance degradation

The gain depends on the number of channels. In Table 2, with the corrected TUEV row (Table bWwa-2), the graph prior has the lower PSD-KL in all 16 cells on the datasets with 32 to 64 channels. On the datasets with 16 to 22 channels, it is lower in 11 cells, equal in 1, and higher in 4 (RF on TUAB, OT-CFM on Mumtaz-MDD, SI and OT-CFM on BCI-IV 2a). We also tested that the training-set size, kk (Table bWwa-3), and σ\sigma, do not explain this difference.

When the sensors are dense, neighboring channels are strongly correlated. The isotropic source ignores this correlation, while the graph prior starts from noise that already has it. Dense montages are therefore the setting in which the choice of source matters, and there the gain is consistent.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.