acceptodds
Under review as a conference paper at ICLR 2027

BrainSaber: Reconstruction-Based Information Decomposition for Functional Brain Atlas Construction

Abstract

Brain atlases define the regional units through which we study brain function. Yet parcellation based on signal or representational similarity does not explicitly distinguish how locations share information or jointly predict activity elsewhere. We introduce BrainSaber, a whole-brain atlas framework built on our proposed reconstruction-based information decomposition RID. RID uses controlled masking of a frozen, pretrained fMRI autoencoder to compare reconstructions of common targets from individual and paired sources. It distinguishes redundant, unique, and synergistic predictive contributions through shared, source-specific, and joint-only reconstruction successes. BrainSaber combines these components with brain-wide predictive flow and spatial constraints to construct contiguous parcels across cortical and non-cortical structures. Regional analyses reveal contrasting associations of redundancy and synergy with conventional functional connectivity. Exploratory cortical and whole-brain evaluations show favorable parcel homogeneity and separability, while prediction across five datasets and cross-window and cross-session analyses support downstream utility and group-atlas repeatability. These findings motivate information composition as a complementary basis for functional parcellation, extending similarity-based grouping to account for individual and joint predictive contributions.

Primary area:
applications to neuroscience & cognitive science
Keywords:
fMRI, Functional Brain Parcellation, Brain Atlas, Information Decomposition, Masked Reconstruction, Self-Supervised Learning

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_f99uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce BrainSaber which is a novel framework for the generation of fMRI brain atlases. It relies on a pretrained autoencoder enabling "reconstruction-based information decomposition" (RID). Through a structured method of masking, reconstructions of the autoencoder are compared between single and double patches which aims to decompose reduntant, unique, and synergistic contributions. These in turn appropriately inform a graph which finally is used for clustering into parcels. This constitutes a novel approach to the creation of brain atlases and the authors evaluate their atlases in a broad set of assessments, which includes downstream utility. Specifically, parcels are shown to have strong properties using standard metrics (such as separability, homogeneity, repeatability) as well as enabling improved predictive performnace.

Critical strengths

The paper takes a fresh and novel approach to brain atlases, which are very widely used but (to my knowledge) their generation does not enjoy nearly the same amount of attention as many other branches. Second, the authors provide a broad set of evaluations.

Critical weaknesses
  1. The contribution of RID itself does not seem to be validated, e.g. by running the entire pipeline exactly as is but instead slotting in something like standard FC instead of RID
  2. For some evaluations the paper does not make it clear where the large improvements come from. Somewhat similar to 1, but I can't really tell in the comparison of BrainSaber to Schaefer atlas what is being compared; is it truly only the parcels which are different? If so, deltas of +7% are in my experience fairly unheard of; after all, only the boundaries are moved around here. Similarly for backbone performance, while the paper has interesting ideas for the atlas generation, the autoencoder does not seem special - are general setup differences (eg number of frames) driving this effect? I cant tell.
Questions for authors
  1. Could the authors try to isolate the contribution of RID itself (see weakness 1)
  2. Could the authors add appropriate controls for the downstream evaluations? How controlled is the comparison to schaefer and to the other backbones?
  3. Fig 11 shows the gain from a single patch is tiny. But this seems to form the basis for their atlas creation. I would expect this gain to be mostly driven by spatially local information; would it be possible to see the gain based on proximity? And if long-range gain on this metric is low, what does that imply for the captured "global and long range information flow" (fig 1) of the atlas
10 comments
AuthorsOpenReview ↗

Response to Q1–Q2: RID Controls and Evaluation Protocols

We thank the reviewer for the constructive comments. We address the contribution of RID, the evaluation protocols, and the spatial extent of single-patch reconstruction gains below.

Q1. Isolating the contribution of RID with FC controls

We added FC embedding and Signal FC, which construct graphs using correlations between patch representations from the same frozen backbone and between patch-averaged BOLD time series, respectively. Both retain the final spectral embedding and spatial Ward clustering steps.

For cortical atlas-quality evaluation, all three methods use the same 300 construction participants and a separate cohort of 707 evaluation participants. At K=200, Ours achieves a lower within-parcel residual and a higher boundary merge cost than both controls on the original signals (Table R3-1).

Table R3-1. Cortical atlas quality at K=200.

MethodWithin-parcel residual ↓Boundary merge cost ↑
Ours0.41696\textcolor{red}{\mathbf{0.41696}}0.07736\textcolor{red}{\mathbf{0.07736}}
FC embedding0.433470.06681
Signal FC0.469220.06511

These results support the benefit of reconstruction-derived connectivity over these two controls on the reported original-signal metrics.

We also report prediction results across five downstream tasks, using multi-resolution fusion at K=100, 200, 400, and 500 (Table R3-2).

Table R3-2. Downstream prediction with the FC controls.

Dataset / MetricOursFC embeddingSignal FC
HCP: ACC / F1 (%) ↑92.57\textcolor{red}{\mathbf{92.57}} / 92.52\textcolor{red}{\mathbf{92.52}}86.63 / 86.5482.67 / 82.60
CHCP: ACC / F1 (%) ↑79.37\textcolor{red}{\mathbf{79.37}} / 79.28\textcolor{red}{\mathbf{79.28}}66.67 / 66.6760.32 / 60.28
NKI: ACC / F1 (%) ↑56.54\textcolor{red}{\mathbf{56.54}} / 47.25\textcolor{red}{\mathbf{47.25}}50.63 / 38.7446.84 / 37.40
SALD: MSE_z ↓ / r ↑0.401\textcolor{red}{\mathbf{0.401}} / 0.789\textcolor{red}{\mathbf{0.789}}0.465 / 0.6680.494 / 0.639
ABIDE: ACC / F1 (%) ↑67.82\textcolor{red}{\mathbf{67.82}} / 67.53\textcolor{red}{\mathbf{67.53}}63.79 / 63.6566.67 / 66.51

Ours uses the archived three-run Table 2 means. The new controls are point estimates from a separate fixed-split evaluation protocol.

Q2. Controls for the backbone and atlas comparisons

Backbone comparison in the manuscript's Table 1. All models were evaluated on the same participants using the same 300 fMRI frames per participant, with identical downstream data splits and a common linear-probe hyperparameter selection procedure. We retained each model's native input configuration: BrainSaber used three 100-frame windows, DCA one 300-frame input, and SwiFT and NeuroSTORM fifteen 20-frame windows each. The backbones remained frozen, and their feature outputs were mean-pooled into participant-level representations. Thus, BrainSaber received no additional frames or observations.

Atlas comparison in the manuscript's Table 2. FC features are computed from the mean time courses of ROIs defined by each atlas. In the current main-text Table 2, all methods share participant splits, common cortical support, the four-resolution ROI–FC pipeline at K=100, 200, 400, and 500, and predictor-selection rules. The comparison therefore evaluates atlas-defined parcellations and their resulting FC features under a common downstream evaluation protocol.

As supplementary evidence on the magnitude of the reported gains, Table R3-3 summarizes a post-hoc analysis of the archived three-run Table 2 predictions. We performed 10,000 subject-cluster bootstrap resamples, sharing resampling weights for the same subject across methods and runs.

Table R3-3. Supplementary uncertainty analysis of the archived three-run Table 2 comparison with Schaefer.

DatasetMetricMean: Schaefer → BrainSaberCohen's dRelative improvement [95% CI]
HCPACC ↑88.94% → \textcolor{red}{\mathbf{92.57\\%}}1.63+4.08% [+0.95%, +7.47%]
CHCPACC ↑71.96% → \textcolor{red}{\mathbf{79.37\\%}}1.59+10.29% [−2.27%, +25.53%]
NKI (BrainSaber-HCP)ACC ↑54.01% → \textcolor{red}{\mathbf{59.07\\%}}0.92+9.38% [−2.90%, +24.01%]
SALDMSE_z ↓0.4451 → 0.4005\textcolor{red}{\mathbf{0.4005}}0.81+10.03% [−12.34%, +27.26%]
SALDPearson r ↑0.7339 → 0.7888\textcolor{red}{\mathbf{0.7888}}2.00+7.47% [−2.00%, +17.71%]
ABIDEACC ↑60.54% → \textcolor{red}{\mathbf{67.82\\%}}1.59+12.03% [+4.06%, +20.97%]
AuthorsOpenReview ↗

Response to Q3: Distance-Resolved Reconstruction Gains

Q3. Single-patch reconstruction gain as a function of distance

We performed a distance-resolved analysis using 101 validation participants, 100 distinct source locations per participant, and 12 evaluation windows. Source and scored target patches both contained 64/64 valid voxels, and the visible source was excluded from evaluation. Distance was measured between patch centers.

Table R3-4 reports single-source normalized reconstruction error and BrainSaber's improvement over the all-masked condition. Gains are in percentage points, with 95% CIs from 10,000 participant-level bootstrap resamples.

Table R3-4. Single-patch reconstruction error and gain by source–target distance.

Distance groupDistanceBrainSaber error ↓Omni error ↓BrainSaber gain ↑, 95% CI
Near: immediate neighbors8–16 mm0.542\textcolor{red}{\mathbf{0.542}}0.9753.20 [1.83, 4.55]
Intermediate: beyond immediate neighbors16–24 mm0.483\textcolor{red}{\mathbf{0.483}}1.0345.43 [4.49, 6.36]
Intermediate: beyond immediate neighbors24–32 mm0.464\textcolor{red}{\mathbf{0.464}}1.0265.01 [4.43, 5.62]
Farther: beyond immediate neighbors32–40 mm0.467\textcolor{red}{\mathbf{0.467}}0.9953.15 [2.84, 3.49]
Farther: beyond immediate neighbors40–48 mm0.471\textcolor{red}{\mathbf{0.471}}0.9701.53 [1.38, 1.69]

The distance groups are descriptive labels within the displayed 8–48 mm range. BrainSaber retains positive gains beyond the 26-neighbor region: improvements at 16–48 mm range from 1.53 to 5.43 percentage points, with all 95% CIs above zero. Thus, one visible patch provides additional predictive information for non-neighbor targets.

BrainSaber also has lower mean reconstruction error than Omni in every displayed bin. This supplementary comparison matches the visible spatial source footprint while retaining each model's native temporal context.

These results show that BrainSaber's reconstruction-derived connectivity extends beyond immediate spatial neighbors and captures predictive links to more distant, non-neighbor regions (16–48 mm).

Thank you to the authors for their comments and additional analyses.

Overall, it seems plausible that RID adds value. However, as things stand, I am keeping my current score, as I still don't find the results clear and interpretable:

  1. RID/FC Controls: Thank you for the FC controls. For the dowstrean comparison (r3-2): likely due to small test sample sizes, errors on gains are really large (r3-3), which I assume to a significant degree relates to the splits used. Authors write "Ours uses the archived three-run Table 2 means. The new controls are point estimates from a separate fixed-split evaluation protocol." but consequently that just does not seem like a sufficiently robust evaluation for this setting. Also, how come the controls here are generally worse than schaefer in r3-3?

  2. Eval controls: Thank you. Trend is quite consistent, but the CIs are informative, with 4 CIs including zero. I find the method interesting, but again, I don't think the evaluation scheme is strong enough here to understand when and to what degree the method actually helps.

  3. Distance analysis: Could the authors please include the all-masked error per bin (or clarify the gain is absolute or relative %)? Why does error fall with distance, I'm not sure how to interpret this, combined with the Omni comparison having an error of approx 1.0 in every bin. The gain seems to shrink with distance - the authors claim global and long-information flow, but the results here only provide evidence of a small gain up to 40-48mm, so as it stands those claims need to be revised.

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

The paper builds a brain atlas from a frozen masked autoencoder trained on fMRI. The model is probed by hiding most of the brain and revealing one or two patches. If one patch helps reconstruct another, that is called flow. If two neighbouring patches are revealed, each target voxel is tested for whether it is reconstructed well enough. Targets that pass with either patch alone are called redundant. Targets that pass with only one patch are called unique. Targets that pass only when both are shown are called synergistic. These four quantities become edge weights on a patch graph. The graph is fused, embedded and clustered into parcels. The authors claim the atlas is more homogeneous than Schaefer and DCA, that it gives better downstream prediction, and that the information decomposition is what drives the parcellation.

Critical strengths

The idea of using a frozen foundation model as a probe rather than a feature extractor is interesting. The masking protocol is clearly described. The authors are careful to say what the decomposition is not. The backbone is strong (six pretraining datasets and is beating SWift and NeuroSTORM) and the atlas covers the whole brain including subcortex. The appendices are thorough. The released code is clean, tested, and matches the method description.

Critical weaknesses

The synergy definition does not measure interaction. A target is called synergistic when each patch alone fails the threshold and both together pass. This also happens when two small additive contributions simply add up. I checked this with a simple additive linear model with no interaction at all. A large fraction of targets came out as synergistic and none as redundant. The paper’s synthetic test does not cover this case. On real data the decomposition is also almost empty. Most targets are unresolved and synergy covers under one percent. The evaluation favours the method by construction. The homogeneity metric is the same quantity the clustering minimises. It is computed on signals that are not centred, so differences in mean intensity count as functional similarity. Schaefer is resampled from the surface to the volume, which hurts it. Also the gaps at the headline resolution are small compared with the spread across participants. The ablation does not support the main claim. Removing flow or synergy makes homogeneity slightly better, not worse. Only the uniqueness term matters, and it acts like spatial smoothing. There is no control with the information layers removed and only spatial adjacency kept. The downstream statistics are weak. Variability is reported over three refits of one split, not over subjects. The gains on the smaller datasets correspond to a handful of test subjects. Several confidence intervals include zero.

Questions for authors

Q1. The code release notes record different fusion weights and a much smaller construction cohort than the paper states, and say the match is unverified. Which settings produced the reported atlases? Will the pretrained weights be released? Q2. What fraction of targets does the rule call synergistic when the two sources are independent and additive? And what does the atlas look like with all information layers removed?

7 comments
AuthorsOpenReview ↗

Response to (1/4): Atlas configurations

We thank the reviewer for the careful assessment and constructive questions. We clarify the construction configurations, the interpretation of RID, and the additional controls and uncertainty analyses below.

Q1 Atlas configurations and pretrained weights

We thank the reviewer for identifying this inconsistency. The reported cortical and whole-brain atlases used different construction configurations (Table R2-1):

Table R2-1. Construction configurations of the reported atlases.

AtlasConstruction participantsFlow / R / S weights
Cortical300(0.30, 0.25, 0.45)
Whole-brain comparison10(0.0375, 0.95, 0.0125)

For the whole-brain comparison, the DCA resources we used provided pretrained weights but no ready-to-use whole-brain group atlas. Generating this baseline required eight rounds of participant-specific adaptation and spectral clustering, followed by group aggregation, with substantial computational cost. We therefore used the same 10 construction participants for BrainSaber and DCA, controlling construction-cohort size and composition in the comparison.

The construction cohort and the evaluation cohort serve different purposes. Figure 4's whole-brain maps show the resulting 10-participant group atlases, whereas its quality curves evaluate these fixed atlases on 707 participants, reporting mean ± SD. This keeps the comparison matched in its construction cohort and assesses the fixed atlases across a substantially larger evaluation sample. We will correct the manuscript and release documentation to distinguish these configurations and provide all pretrained weights with their corresponding configurations and checksums.

AuthorsOpenReview ↗

Response to (2/4): Synergy and nonlinear reconstruction

Q2a Additive sources and RID Synergy

We thank the reviewer for this helpful comment and for providing the additive counterexample. We would first like to distinguish information-theoretic synergy from statistical interaction. These concepts are not equivalent: certain Gaussian PID formulations allow positive synergy in linear systems (Barrett, 2015). RID uses an operational reconstruction criterion—joint reconstruction succeeds where neither single-source reconstruction meets the criterion—and we will clarify its distinction from formal information-theoretic PID.

We agree that additive contributions can produce such joint benefits. In our independent, equal-strength Gaussian additive control, the RID-S fraction was 100% with noise SD 0.25 and the original threshold of 0.5. RID can also record joint reconstruction success arising from non-additive interactions. Table R2-2 reports the three-seed semi-synthetic controls using real BOLD sources and the same noise SD of 0.25 and threshold of 0.5. The additive-only predictor models each source separately and sums their predictions, whereas the joint GRU models both sources together.

Table R2-2. Semi-synthetic controls on an independent test set, averaged over three seeds.

TargetAdditive-only predictor MSE ↓Joint GRU MSE ↓Joint GRU RID-S
Additive0.06269\textcolor{red}{\mathbf{0.06269}}0.06270100%
Product interaction1.056980.06296\textcolor{red}{\mathbf{0.06296}}99.994%

For the additive target, the two predictors have nearly identical test MSEs (0.06269 and 0.06270). For the product target, the joint GRU achieves 0.06296, compared with 1.05698 for the additive-only predictor, demonstrating its ability to learn the known multiplicative interaction. The RID-S proportions of 100% and 99.994% show reconstruction complementarity under both generating mechanisms.

The brain is a complex system involving collective information processing and nonlinear, state-dependent interactions. Human neuroimaging links synergistic information to integration across higher-order brain networks (Luppi et al., 2022). Nonlinear DCM also models neuronal gating, where one population modulates the interaction between two others, and has been tested with fMRI (Stephan et al., 2008). These findings motivate nonlinear reconstruction and joint-source probing in atlas construction, allowing dependencies beyond a purely additive description to inform the atlas alongside additive contributions.

To examine whether dependencies between real fMRI signals are better described by nonlinear modeling than by a linear mapping, we compared a GRU with a linear predictor on a fixed 64-subject training subset comprising 768 windows of 100 frames. Inputs, targets and normalization were identical, and each model contained 133,505 parameters (Table R2-3).

Table R2-3. Training fit to real fMRI signals with 133,505 parameters per model.

ModelTraining MSE ↓
GRU, base learning rate 0.0030.00641\textcolor{red}{\mathbf{0.00641}}
Linear model, converged0.73049

The GRU was continued from its original checkpoint for 20,000 epochs. The linear control had already converged: its training MSE of 0.73048618 was only 7.54 × 10⁻⁷ above the stable SVD least-squares reference of 0.73048543, satisfying the preset 10⁻⁶ convergence tolerance. Thus, under the stated numerical protocol, further optimization within this affine function class has negligible remaining room to reduce training error. The continued GRU reaches a training MSE of 0.00641, compared with 0.73049 for the converged linear control. These results show that, for the real fMRI training samples examined, joint nonlinear modeling captures source–target variation that a linear mapping does not fit adequately. Together with the product-target control, this supports using joint reconstruction models capable of representing nonlinear and non-additive dependencies in atlas construction.

AuthorsOpenReview ↗

Response to (3/4): Spatial controls, coverage and metrics

Q2b Spatial only control and component ablations

We added a spatial-only control that removes all information-derived relations and assigns equal weights to the same spatial-neighbor graph, while retaining the cortical support, spectral-embedding settings, and spatially constrained Ward clustering. At K=200, both atlases contain 200 parcels and are evaluated on the same 707 HCP participants who did not contribute to construction of the cortical Full atlas (Table R2-4).

Table R2-4. Spatial-only control at K=200, with temporal means retained.

MethodWithin-ROI residual variance ↓Boundary merge cost ↑
Spatial-only0.457170.06543
Full0.41696\textcolor{red}{\mathbf{0.41696}}0.07736\textcolor{red}{\mathbf{0.07736}}

With voxelwise temporal means retained, Full reduces within-ROI residual variance by 8.80% and increases boundary merge cost by 18.22% relative to Spatial-only. Thus, spatial adjacency alone does not match Full on these two original-signal endpoints.

We also agree that the original ablations do not show every component improving every endpoint: removing Flow or Synergy need not worsen within-parcel residuals. The U term modulates local weights using source-specific reconstruction outcomes, and its ablation shows a contribution within this configuration.

Reconstruction coverage and unresolved targets

Thank you for raising this concern. RID decomposes targets that can be successfully reconstructed from sparse sources; it does not require each source pair to reconstruct the entire brain. A single patch or two patches provide limited observations, so only a subset of targets meets the reconstruction criterion. After excluding the source patches and targets already reconstructed successfully under the all-masked baseline, we define a target as solved if it succeeds under at least one singleton or joint-source condition. These solved targets are then partitioned into redundant, unique, and synergistic contributions.

Using the aggregate component proportions, Synergy accounts for approximately 8.55% of the solved portion. These are targets for which neither singleton reconstruction succeeds, but joint-source reconstruction does. This percentage describes the composition of successful predictions. Unresolved targets fail the criterion under all three source conditions and characterize reconstruction coverage for the given source pair, model, and threshold. Overall coverage and the unresolved fraction are reported separately in the supplementary material.

Evaluation objectives and temporal demeaning

Thank you for this comment. Ward clustering minimizes dispersion in graph-embedding space, whereas Figure 4 evaluates voxel-level BOLD residuals. These are distinct numerical objectives.

As defined in the Appendix, within-ROI residual measures the mean-squared error incurred when an ROI-mean time course represents its constituent voxel signals. It captures both mean-level and temporal differences, providing an explicit measure of original-signal representation fidelity. We agree that mean-level similarity alone does not establish greater dynamic functional homogeneity, and therefore supplement this endpoint with voxelwise demeaning.

We report both original-signal and demeaned results at all three original resolutions: K=100, 200, and 500 (Table R2-5). Paired 95% CIs use 10,000 participant bootstrap resamples; negative differences favor BrainSaber.

Table R2-5. Whole-brain residuals before and after temporal demeaning, with paired 95% CIs.

Nominal KSignalBrainSaber residual ↓DCA residual ↓Paired 95% CI for BrainSaber − DCA
100Original0.39407\textcolor{red}{\mathbf{0.39407}}0.43222[−0.03893, −0.03736]
100Demeaned0.018452570.01844914\textcolor{red}{\mathbf{0.01844914}}[3.222, 3.661] × 10⁻⁶
200Original0.35715\textcolor{red}{\mathbf{0.35715}}0.38491[−0.02829, −0.02721]
200Demeaned0.018422480.01841960\textcolor{red}{\mathbf{0.01841960}}[2.706, 3.045] × 10⁻⁶
500Original0.32142\textcolor{red}{\mathbf{0.32142}}0.33712[−0.01605, −0.01534]
500Demeaned0.01835264\textcolor{red}{\mathbf{0.01835264}}0.01836000[−7.639, −7.089] × 10⁻⁶

The original-signal differences favor BrainSaber at all three resolutions, with paired CIs excluding zero. Comparing the mean gap only with between-participant SD does not replace this within-participant comparison. After demeaning, DCA has lower residuals at K=100 and 200, whereas BrainSaber has a lower residual at K=500. Relative residual differences are below 0.05%, despite the CIs excluding zero. In the original cortical evaluation, BrainSaber also has higher demeaned residuals than DCA and Schaefer at K=100, 200, 400, and 500. We therefore retain the original-signal fidelity interpretation and qualify claims of consistently superior centered temporal homogeneity.

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

This work proposes a whole-brain atlas framework that combines reconstruction-based information decomposition from a frozen fMRI autoencoder with predictive flow and spatial constraints.

Critical strengths

The motivation is clear, and the method is supported by synthetic validation, atlas-quality analyses, and downstream evaluations across five datasets.

Critical weaknesses

The validation of RID neurofunctional interpretation relies on an auxiliary ROI model. The added value of the proposed graph construction beyond backbone representations is not fully isolated. Fusion-weight sensitivity and comparative computational efficiency require further evaluation.

Questions for authors
  1. Can you compare RID with a representational-similarity graph derived from the same frozen backbone to isolate the contribution of decomposition?
  2. How were the fusion weights selected, and how sensitive are atlas quality and downstream performance to these choices, beyond the reported component-removal ablations?
  3. How do runtime and peak memory usage compare with competing methods under consistent conditions, and can approximations reduce probing costs while preserving atlas quality?
  4. Are the downstream gains in Table 2 statistically reliable under paired evaluation, and can you provide confidence intervals for these differences?
2 comments
AuthorsOpenReview ↗

Response to: FC controls, fusion-weight selection and sensitivity

Response to Reviewer PPkY (1/2): Q1–Q2

We sincerely thank the reviewer for the thoughtful assessment and constructive suggestions. We appreciate the opportunity to clarify the contribution of reconstruction-derived connectivity and the selection and sensitivity of the fusion weights.

Q1 Contribution beyond backbone representations

Thank you for suggesting controls based on the same frozen backbone. We added FC embedding and Signal FC, using correlations between patch representations from that backbone and between patch-averaged BOLD time series, respectively. Both retain spectral embedding and spatial Ward clustering.

All three atlases use the same 300 construction participants and are evaluated on 707 separate participants using their initial 300 original-signal frames. At K=200, Ours has lower within-parcel residual variance and higher boundary merge cost than both controls (Table R1.1). These results support the value of reconstruction-derived connectivity for the two atlas-quality measures evaluated here.

Table R1.1. Atlas quality at K=200.

MethodWithin-parcel residual variance ↓Boundary merge cost ↑
Ours0.41696\color{red}{\mathbf{0.41696}}0.07736\color{red}{\mathbf{0.07736}}
FC embedding0.433470.06681
Signal FC0.469220.06511

We also provide the prediction results for five downstream tasks, using fusion across K=100/200/400/500 (Table R1.2). We hope these additional results help provide a broader view of the utility of the resulting atlases across datasets.

Table R1.2. Downstream prediction.

Dataset and metricOursFC embeddingSignal FC
HCP: ACC / macro-F1 (%) ↑92.57/92.52\color{red}{\mathbf{92.57 / 92.52}}86.63 / 86.5482.67 / 82.60
CHCP: ACC / macro-F1 (%) ↑79.37/79.28\color{red}{\mathbf{79.37 / 79.28}}66.67 / 66.6760.32 / 60.28
NKI: ACC / macro-F1 (%) ↑56.54/47.25\color{red}{\mathbf{56.54 / 47.25}}50.63 / 38.7446.84 / 37.40
SALD: MSE_z ↓ / Pearson r ↑0.401/0.789\color{red}{\mathbf{0.401 / 0.789}}0.465 / 0.6680.494 / 0.639
ABIDE: ACC / macro-F1 (%) ↑67.82/67.53\color{red}{\mathbf{67.82 / 67.53}}63.79 / 63.6566.67 / 66.51

Q2 Fusion-weight selection and sensitivity

We thank the reviewer for requesting further clarification of weight selection and sensitivity beyond component-removal ablations.

Grid search on 300 HCP construction participants jointly tuned fusion weights, the uniqueness exponent, and embedding dimension. We maximized the minimum relative gain over the strongest matched baselines across two backbones, five resolutions, and two quality metrics, selecting Flow/Redundancy/Synergy weights (0.30,0.25,0.45).

To examine sensitivity around this configuration, we tested six transfers of 0.05 between weights, keeping other settings fixed at each resolution. Evaluation used 707 HCP participants outside the construction cohort and 11 resolutions from K=100 to 1000. Residual variance and boundary merge cost changed by at most 1.11% and 6.20%, respectively (Table R1.3). We report the full resolution-wise results below to make the magnitude of these changes explicit.

Table R1.3. Maximum absolute relative changes in participant means over six perturbations.

KResidual variance maximum changeBoundary merge cost maximum change
1001.08%6.20%
2001.11%2.41%
3000.93%3.11%
3601.07%3.15%
4001.08%5.09%
5001.03%3.07%
6000.59%2.17%
7000.63%2.21%
8000.54%2.07%
9000.49%1.96%
10000.42%1.06%

Table R1.4 also presents the corresponding atlas-quality and downstream prediction results for the selected configuration and all six nearby weight settings.

Table R1.4. Results for seven fusion-weight settings.

Flow / Redundancy / Synergy weightsResidual variance (K=200) ↓Boundary merge cost (K=200) ↑HCP ACC / macro-F1 (%) ↑CHCP ACC / macro-F1 (%) ↑NKI ACC / macro-F1 (%) ↑SALD MSE_z ↓ / r ↑ABIDE ACC / macro-F1 (%) ↑
(0.30, 0.25, 0.45)0.416960.0773692.57/92.52\color{red}{\mathbf{92.57 / 92.52}}79.37/79.28\color{red}{\mathbf{79.37 / 79.28}}56.54/47.25\color{red}{\mathbf{56.54 / 47.25}}0.401/0.789\color{red}{\mathbf{0.401 / 0.789}}67.82 / 67.53
(0.35, 0.20, 0.45)0.416330.0759487.62 / 87.5766.67 / 66.6343.04 / 37.610.458 / 0.67766.67 / 66.45
(0.25, 0.30, 0.45)0.41233\color{red}{\mathbf{0.41233}}0.0760886.14 / 86.1066.67 / 66.6746.84 / 39.980.431 / 0.69768.39 / 68.31
(0.35, 0.25, 0.40)0.414540.0754987.13 / 87.0571.43 / 71.3646.84 / 41.090.464 / 0.67066.09 / 65.96
(0.25, 0.25, 0.50)0.414750.07814\color{red}{\mathbf{0.07814}}87.13 / 87.1161.90 / 61.9048.10 / 42.440.434 / 0.69365.52 / 65.35
(0.30, 0.30, 0.40)0.418970.0755184.65 / 84.5963.49 / 63.4646.84 / 42.730.423 / 0.70570.11/69.92\color{red}{\mathbf{70.11 / 69.92}}
(0.30, 0.20, 0.50)0.419680.0761888.61 / 88.5968.25 / 68.2544.30 / 37.190.468 / 0.66767.82 / 67.75
AuthorsOpenReview ↗

Response to: Computational cost, downstream uncertainty and ROI validation

Response to Reviewer PPkY (2/2): Q3–Q4 and ROI validation

Q3 Runtime, peak memory, and reduced probing

Thank you for suggesting a comparison of computational costs and an assessment of reduced probing.

We compared individual-stage costs using the same seven HCP construction participants and first 100 observed frames per participant. Each job used one NVIDIA L40 on a server with two 64-core AMD EPYC 9554 CPUs. Both methods used FP32, batch size one, and four PyTorch/OpenMP/BLAS threads; pretraining was excluded. DCA retained its native 300-channel input (Table R1.5).

Table R1.5. Individual-stage resources.

MetricOptimized BrainSaberDCA
Peak allocated GPU memory ↓3.54\color{red}{\mathbf{3.54}} GiB10.38 GiB
Peak process-tree RSS, including subprocesses ↓2.21\color{red}{\mathbf{2.21}} GiB7.57 GiB
Peak main-process RSS ↓2.21\color{red}{\mathbf{2.21}} GiB2.90 GiB
Cumulative individual-stage runtime across five resolutions (estimated) ↓5.13\color{red}{\mathbf{5.13}} h/participant8.77 h/participant

Memory peaks are maxima over the seven K=200 jobs. BrainSaber reduced GPU allocation, process-tree RSS, and main-process RSS by 65.9%, 70.9%, and 23.9%, respectively. Flow/RID is computed once and reused at K=100/200/400/500/1000, followed by resolution-specific embedding and clustering; DCA repeats individual optimization for each K.

We also rebuilt K=200 atlases from two complementary 150-person subsets of the 300 constructors and evaluated them on the same 707 participants (Table R1.6).

Table R1.6. Reduced-probing approximation at K=200.

Construction settingProbe contextsResidual variance ↓Boundary merge cost ↑ARI versus full atlas
Full: 300 participants16,705,0240.416960.077361.000
Half A: 150 participants8,351,8730.41591\color{red}{\mathbf{0.41591}}0.07812\color{red}{\mathbf{0.07812}}0.780
Half B: 150 participants8,353,1510.417290.077220.809

Probe counts approximately halved, while both original-signal quality metrics changed by less than 1%.

Q4 Uncertainty in downstream gains

We appreciate your suggestion to quantify uncertainty in the downstream gains.

Using the original Table 2 predictions, we performed 10,000 subject-cluster bootstrap resamples, sharing weights for each subject across methods and runs. Tables R1.7–R1.8 report the original Cohen's d and relative improvements with conditional 95% bootstrap CIs.

Table R1.7. Comparison with DCA.

DatasetMetricMean: DCA → BrainSaberCohen's dRelative improvement [95% CI]
HCPACC ↑90.10% → 92.57\color{red}{\mathbf{92.57}}%1.06+2.75% [+0.11%, +5.53%]
CHCPACC ↑68.78% → 79.37\color{red}{\mathbf{79.37}}%1.87+15.38% [+2.09%, +31.58%]
NKI (BrainSaber-HCP)ACC ↑55.27% → 59.07\color{red}{\mathbf{59.07}}%0.60+6.87% [−3.95%, +19.58%]
SALDMSE_z ↓0.5026 → 0.4005\color{red}{\mathbf{0.4005}}1.56+20.31% [−0.57%, +36.06%]
SALDPearson r ↑0.6903 → 0.7888\color{red}{\mathbf{0.7888}}3.85+14.27% [+3.20%, +27.14%]
ABIDEACC ↑65.13% → 67.82\color{red}{\mathbf{67.82}}%0.98+4.12% [−2.13%, +10.56%]

Table R1.8. Comparison with Schaefer.

DatasetMetricMean: Schaefer → BrainSaberCohen's dRelative improvement [95% CI]
HCPACC ↑88.94% → 92.57\color{red}{\mathbf{92.57}}%1.63+4.08% [+0.95%, +7.47%]
CHCPACC ↑71.96% → 79.37\color{red}{\mathbf{79.37}}%1.59+10.29% [−2.27%, +25.53%]
NKI (BrainSaber-HCP)ACC ↑54.01% → 59.07\color{red}{\mathbf{59.07}}%0.92+9.38% [−2.90%, +24.01%]
SALDMSE_z ↓0.4451 → 0.4005\color{red}{\mathbf{0.4005}}0.81+10.03% [−12.34%, +27.26%]
SALDPearson r ↑0.7339 → 0.7888\color{red}{\mathbf{0.7888}}2.00+7.47% [−2.00%, +17.71%]
ABIDEACC ↑60.54% → 67.82\color{red}{\mathbf{67.82}}%1.59+12.03% [+4.06%, +20.97%]

Positive improvements favor BrainSaber: relative increases in ACC or Pearson r, or decreases in MSE_z. Pearson percentages describe relative changes in the correlation coefficient.

Additional concern Auxiliary ROI model

We thank the reviewer for the opportunity to clarify the role of the ROI analyses.

Regional and network-level physiological interpretation is harder at the voxel scale. Standard ROI annotations allow us to relate RID-derived redundancy and synergy to sensorimotor and association networks and their hierarchy. ROI signals aggregate voxel BOLD signals and share their measurement and temporal basis. Their interpretable RID results therefore provide empirical support for applying RID to voxel-level atlas construction.

Synthetic tests further assess recovery of predefined redundant, unique, and synergistic predictive contributions under known mechanisms; atlas-quality and downstream evaluations directly assess the resulting voxel atlases.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.