acceptodds
Under review as a conference paper at ICLR 2027

Continuous Neural Fields for Functional Connectivity

Abstract

Brain Functional connectivity (FC) is typically represented after discretizing the brain into a predefined set of regions, making the resulting representation dependent on the chosen parcellation and spatial scale. We introduce a continuous representation of resting-state functional organization in which each cortical location is associated with a low-dimensional functional embedding, and connectivity is induced by similarity between embeddings. We parameterize this representation with implicit neural representations (INRs), first learning a population functional field that captures shared organization and then adapting it to each individual through constrained subject-specific deviations. The resulting continuous fields accurately represent functional connectivity while capturing reproducible individual variation across independent acquisitions. Beyond reconstruction, the differentiable representation allows cortical regions to serve as spatial readouts of the field, whose location and geometry can be optimized for downstream analyses. Across three cohorts, optimized regions yield stronger held-out associations with age and sex than the best connections of coarse and fine parcellations in every setting, and the advantage extends to behavioral and cognitive phenotypes. Together, these results establish continuous functional fields as an alternative to parcellation-defined FC representations, supporting individualized modeling and spatially flexible analysis within a common framework.

Primary area:
applications to neuroscience & cognitive science
Keywords:
implicit neural representation, functional MRI, functional connectivity, neuroscience

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

The paper presents a way to represent brain functional connectivity without first dividing the brain into fixed atlas regions. It learns a continuous map that assigns a functional embedding to each cortical location. A shared population map captures common patterns, while constrained individual adjustments capture differences between people. The method then adjusts the locations, sizes, and shapes of regions to find connections associated with age, sex, or behavioral scores. Experiments on HCP-D, PNC, and HBN show stronger reported held-out associations than selected region pairs from Schaefer-100 and Schaefer-1000 atlases.

Critical strengths
  1. The continuous representation allows brain regions to change location, size, and shape for different questions, while capturing both shared population patterns and individual differences.

  2. Experiments on three datasets show stronger reported held-out associations than selected Schaefer atlas connections. Repeat-scan experiments and ablation studies help assess reliability and explain the effects of model choices.

Critical weaknesses
  1. Comparisons are limited to fixed Schaefer atlases. Ward and k-means also learn regions from data rather than relying on predefined atlas boundaries. Comparisons with these and supervised parcellation methods are needed to establish the benefit of the continuous representation.
  2. The best region configuration is reported based on held-out performance, which may overestimate generalization. A separate validation set or nested cross-validation would provide a stronger evaluation.
Questions for authors

Q1 How does the method compare with Ward, k-means, and supervised parcellation methods, which also learn regions from data rather than using predefined atlas boundaries?

Q2 How could this method be extended to clinical datasets such as ADHD-200, ABIDE, and ADNI, particularly when the available sample is small? Would a cohort-matched population field, stronger regularization, or simpler region geometry be needed?

Q3 How are the initial radius r_0 and contour complexity K selected without using held-out results? Please clarify the validation procedure, split seeds, and number of region-optimization steps.

Q4 The proposed method receives geometry optimization and a sweep over radius and contour complexity, while the Schaefer baselines only select the best parcel pair. How do the authors account for this difference in tuning effort? Would the gains remain under comparable tuning budgets?

5 comments
AuthorsOpenReview ↗

Response to Reviewer ubRR

We thank the reviewer for the constructive review and detailed questions. We appreciate that the reviewer recognized the flexibility of our regions in location, size and shape, and valued our experiments across three datasets, including repeat-scan and ablation analyses. The main concerns can be summarized as: (1) comparisons limited to fixed Schaefer atlases, (2) selection on held-out data and unequal tuning budgets, and (3) extension to small clinical datasets. We address each point in detail below:

Comparisons with these and supervised parcellation methods are needed to establish the benefit of the continuous representation.

We added three suggested dataset-specific region-learning baselines to the updated Table 3 below:

TraitCohortSchaeferWardk-meansSBPOurs, fixedOurs, nested
SexHCP-D6.5 [4.7, 9.9]6.4 [3.1, 12.2]8.8 [3.3, 15.4]7.5 [4.4, 10.9]12.6 [7.5, 15.4]11.9 [7.6, 16.0]
PNC12.5 [7.7, 25.7]15.1 [9.7, 19.3]16.0 [7.1, 20.0]5.2 [1.9, 8.6]15.9 [13.6, 20.3]20.4 [12.9, 23.0]
HBN14.3 [8.5, 18.6]12.9 [8.9, 14.8]10.7 [9.3, 12.3]5.7 [3.4, 8.2]17.3 [11.0, 19.4]15.1 [10.8, 17.3]
AgeHCP-D74.3 [64.6, 83.3]41.5 [31.2, 55.9]42.4 [35.4, 46.5]69.1 [47.1, 80.4]80.3 [70.1, 98.7]81.6 [68.8, 95.6]
PNC73.7 [56.0, 81.1]61.7 [54.5, 68.1]65.5 [60.2, 72.4]76.7 [67.7, 90.7]80.2 [74.8, 90.0]79.5 [75.1, 89.7]
HBN81.6 [75.4, 96.1]92.9 [89.1, 103.3]95.3 [78.9, 101.3]99.5 [93.5, 121.8]112.5 [95.2, 126.0]114.3 [95.4, 122.3]

Revised Table 3: held-out t2t^2, median [IQR] over 12 splits (35% of subjects held out). All choices are made on training subjects by 5-fold cross-validation within each training split: the number of parcels for Schaefer, Ward and k-means (100 or 1000), the number of parcels and penalty λ\lambda for SBP (100 or 1000, λ∈0,0.1,0.3,1,3\lambda \in \\{0, 0.1, 0.3, 1, 3\\}), and r0∈4,8,12∘r_0 \in \\{4, 8, 12\\}^\circ and K∈2,4,6K \in \\{2, 4, 6\\} for Ours, nested. Ours, fixed: r0=8∘r_0 = 8^\circ, K=4K = 4. Bold: best per row.

In all these methods:

  • Each subject's individual field is pooled within each parcel, and the connection between two parcels is the cosine of their pooled fields. Each hemisphere is divided into K/2K/2 parcels, as in the Schaefer atlases, and every parcel pair is a candidate. After regressing out the covariates (coefficients fitted on training subjects), the pair with the largest training ∣t∣|t| is selected and scored once on held-out subjects.

  • We used nested cross-validation, the number of parcels K∈100,1000K \in \\{100, 1000\\} (and λ\lambda for SBP) is chosen by mean validation t2t^2 over 5 inner folds of each training split; the parcel pair is re-selected inside every inner fold.

Results indicate our continuously optimized regions still achieve stronger associations on held-out subjects. Nevertheless, we would emphasize that brain parcellation methods serve a different purpose. Our framework focuses on localization: identifying where in the brain FC is most strongly associated with a phenotype. This is distinct from the goal of partitioning the entire brain into discrete parcels. Therefore, when interpreting these results, we would like to remind the reviewer that our central scientific question is whether continuous, gradient-based optimization enabled by our INR can identify region boundaries that yield stronger brain–phenotype associations. This represents a fundamentally new perspective that has not been systematically explored in previous work.

AuthorsOpenReview ↗

Response to Reviewer ubRR

The best region configuration is reported based on held-out performance. A separate validation set or nested cross-validation would provide a stronger evaluation. How are the initial radius r0r_0 and contour complexity KK selected without using held-out results? Would the gains remain under comparable tuning budgets?

We apologize for the ambiguity in our original wording. While Table 3 reported the best-performing region configuration based on the held-out t2t^2 results in Table 9, we did so primarily because of the page limit, to avoid displaying the full grid in the main text. More importantly, the choice among configurations turns out to matter little: in Table 9, the median held-out t2t^2 of every configuration of our method is higher than that of the Schaefer baselines. Following the reviewer's suggestion, we now report both alternatives the reviewer suggests:

  1. A configuration fixed in advance: r0=8∘r_0 = 8^\circ, K=4K = 4, with center, radius and contour all fitted.

  2. Nested cross-validation: within each training split, r0∈4,8,12∘r_0 \in \\{4, 8, 12\\}^\circ and K∈2,4,6K \in \\{2, 4, 6\\} are chosen by 5-fold cross-validation on training subjects; the model is refit on the full training split and scored once on the held-out split.

Both are in the revised Table 3 above. For both settings, regions are optimized with Adam (learning rate 0.02) for 450 steps: 150 at each of three mask temperatures, τ=1.5∘\tau = 1.5^\circ, 0.9∘0.9^\circ and 0.5∘0.5^\circ. Performance is reported over 12 random train–test splits using random seed 101.

Regarding the tuning budget, all hyperparameter selection is performed exclusively on training subjects, and all methods are evaluated on the same held-out splits. Under this evaluation protocol, our method with nested cross-validation achieves a higher median held-out t2t^2 than the tuned atlas baselines in all six settings. More importantly, even with a configuration fixed in advance, without any hyperparameter tuning, our method outperforms every baseline in five of the six settings. In the remaining setting (PNC sex), its performance is nearly identical to that of the best baseline (15.9 versus 16.0 for k-means). These results demonstrate that our improvements are robust to configuration selection and cannot be explained by a larger hyperparameter tuning budget.

How could this method be extended to clinical datasets such as ADHD-200, ABIDE, and ADNI, particularly when the available sample is small

For applications in clinical datasets such as ADHD-200, the population field need not be fitted on the clinical cohort. It can come from a large normative dataset, matched in age, with the clinical subjects' individual fields fitted against it. Our PNC and HBN results in Table 3 already work this way: both use the population field trained on HCP-D, only the individual fields are fitted to their own subjects.

When scans are short or noisy, the residual penalty λ\lambda can be increased to pull each individual field toward the population field. Table 2 shows this trade-off between in-sample fit and test–retest reliability on HCP-D subjects. Because the best λ\lambda depends on scan length and quality, it can be re-chosen for new small datasets.

For region geometry in small samples, we recommend a fixed, relatively large radius and a low contour order KK. This limits the number of fitted parameters and reduces the risk of overfitting.

More broadly, our results already show that the framework transfers across developmental cohorts that differ in population, scanner, scan length and phenotypes, using a population field learned on a single cohort. Extending these experiments to a wider range of clinical datasets and tasks is the next step we plan.

AuthorsOpenReview ↗

To Reviewer ubRR

Thank you again for your constructive feedback. In our response, we have clarified the configurations of all methods, explained how our framework can be applied to clinical cohorts, and added new results with nested cross-validation (revised Table 3).

As the discussion period is coming to an end, we would greatly appreciate it if you could let us know whether our responses have addressed your concerns. We would be glad to answer any further questions or provide additional clarification.

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

In this paper, continuous implicit neural representation (INR) is proposed to model the resting-state functional connection: cortical positions are mapped into low-dimensional functional embedding, arbitrary positions are induced by inner product, the embedded targets are obtained by low-rank spectral decomposition of BOLD time series, and shared and individualized functional organizations are captured by population INR and subject-specific residual INR. Furthermore, the author pools the continuous field with the Microsoft regional membership to obtain the regional FC, and makes the regional center, radius and contour optimized for the downstream brain-phenotype association.

Critical strengths
  1. The conceptual design of the presentation level is clear. Decoupling FC from fixed parcellation into a continuous functional field, and using Population Field+Subject Residual to deal with the tension between comparability and individualization across subjects is a meaningful problem reconstruction.
  2. Empirical coverage has a certain breadth.
Critical weaknesses
  1. The held-out evaluation of the main result is suspected of serious selection bias, which may directly invalidate the core empirical conclusion.
  2. The baseline is unfair, and the gain cannot be attributed to the continuous field representation itself.
Questions for authors
  1. Is the held-out grid of Table 9 used to select the configuration of Table 3? Please provide the held-out result under fixed parameter or nested selection.
  2. Is the rs-fMRI of the test subjects used for fitting the individual functional field?
2 comments
AuthorsOpenReview ↗

Response to Reviewer kmUM

We thank the reviewer for the careful and detailed review. We appreciate that the reviewer found the conceptual design clear and recognized the population field + subject residual formulation as a meaningful reformulation of the problem. The main concerns can be summarized as: (1) potential selection bias in the held-out evaluation, (2) the fairness of the baselines, and (3) whether test subjects' rs-fMRI is used to fit individual fields. We address each point in detail below:

Is the held-out grid of Table 9 used to select the configuration of Table 3? Does this lead to selection bias? Please provide the held-out result under fixed parameter or nested selection.

We apologize for the ambiguity in our original wording. While Table 3 reported the best-performing region configuration based on the held-out t2t^2 results in Table 9, we did so simply because of the page limit, to avoid displaying the full grid in the main text. More importantly, the choice among configurations turns out to matter little: in Table 9, the median held-out t2t^2 of every configuration of our method is higher than that of the Schaefer baselines. Following the reviewer's suggestion, we now report two settings of our method that involve no selection on held-out data:

  1. A configuration fixed in advance: r0=8∘r_0 = 8^\circ, K=4K = 4, with center, radius and contour all fitted.

  2. Nested cross-validation: within each training split, r0∈4,8,12∘r_0 \in \\{4, 8, 12\\}^\circ and K∈2,4,6K \in \\{2, 4, 6\\} are chosen by 5-fold cross-validation on training subjects; the model is refit on the full training split and scored once on the held-out split.

TraitCohortSchaeferWardk-meansSBPOurs, fixedOurs, nested
SexHCP-D6.5 [4.7, 9.9]6.4 [3.1, 12.2]8.8 [3.3, 15.4]7.5 [4.4, 10.9]12.6 [7.5, 15.4]11.9 [7.6, 16.0]
PNC12.5 [7.7, 25.7]15.1 [9.7, 19.3]16.0 [7.1, 20.0]5.2 [1.9, 8.6]15.9 [13.6, 20.3]20.4 [12.9, 23.0]
HBN14.3 [8.5, 18.6]12.9 [8.9, 14.8]10.7 [9.3, 12.3]5.7 [3.4, 8.2]17.3 [11.0, 19.4]15.1 [10.8, 17.3]
AgeHCP-D74.3 [64.6, 83.3]41.5 [31.2, 55.9]42.4 [35.4, 46.5]69.1 [47.1, 80.4]80.3 [70.1, 98.7]81.6 [68.8, 95.6]
PNC73.7 [56.0, 81.1]61.7 [54.5, 68.1]65.5 [60.2, 72.4]76.7 [67.7, 90.7]80.2 [74.8, 90.0]79.5 [75.1, 89.7]
HBN81.6 [75.4, 96.1]92.9 [89.1, 103.3]95.3 [78.9, 101.3]99.5 [93.5, 121.8]112.5 [95.2, 126.0]114.3 [95.4, 122.3]

Revised Table 3: held-out t2t^2, median [IQR] over 12 splits (35% of subjects held out). All choices are made on training subjects by 5-fold cross-validation within each training split: the number of parcels for Schaefer, Ward and k-means (100 or 1000), the number of parcels and penalty λ\lambda for SBP (100 or 1000, λ∈0,0.1,0.3,1,3\lambda \in \\{0, 0.1, 0.3, 1, 3\\}), and r0∈4,8,12∘r_0 \in \\{4, 8, 12\\}^\circ and K∈2,4,6K \in \\{2, 4, 6\\} for Ours, nested. Ours, fixed: r0=8∘r_0 = 8^\circ, K=4K = 4. Bold: best per row.

The baseline is unfair

We would appreciate further clarification from the reviewer regarding this concern. Our proposed method integrates a continuous FC representation with differentiable optimization of region geometry, both of which are methodological contributions of this work that have rarely been explored in previous studies. We wonder whether the additional baselines reported in the table above, Ward, k-means, and SBP, would be considered fair comparisons. Note that these are data-driven parcellation methods, so we had to adapt them to our problem setting by first deriving their respective parcellations and then identifying the FC connection most strongly associated with the phenotype using the training cohort.

Is the rs-fMRI of the test subjects used for fitting the individual functional field?

Yes, but we would like to clarify that fitting an individual functional field is an unsupervised step that uses only the subject's own rs-fMRI data, without any phenotype information. This is analogous to computing an FC matrix for a test subject before applying a trained model. Each test subject's field is fitted independently, without using information from other test subjects or their phenotypes.

Importantly, all phenotype-dependent steps, including seed selection, covariate regression, region geometry optimization, readout weights, and hyperparameter selection, are performed exclusively using training subjects. The resulting region definitions are then applied unchanged to the test subjects' individual fields to evaluate held-out brain–phenotype associations.

AuthorsOpenReview ↗

To Reviewer kmUM

Thank you again for your thoughtful review. We have posted a detailed response, including new experiments with nested cross-validation for all methods and additional learned-parcellation baselines (revised Table 3).

As the discussion period is coming to an end, we would greatly appreciate it if you could let us know whether our responses have addressed your concerns. If there are any remaining questions or points that need further clarification, we would be happy to discuss them.

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

This paper proposes to use a learned continuous representation of cortical fMRI-based functional connectivity (FC) or functional organisation across the human cortex. This is a learned follow-up of approaches to go beyond parcellations that have existed for a while in the literature, like the gradients proposed by Margulies et al. (PNAS 2016); the Moyer et al. (medIA 2017) proposal based on S2×S2S^2\times S^2 kernels; the connectopic proposal (Haak et al. NImg 2017), which, as this paper uses, a joint spatial-functional SVD; or eigenmode/harmonics-based representations (Mansour et al. 2024, Pang et al. Nature, Glomb et al. Cell Reports 2021).

The main conceptual advancement of this proposal is their proposal of what could be seen as a random effects approach in which the authors propose to have a perturbation model of population and individual contributions to the individual FC signal. Furthermore, by casting the problem into current autodiff technologies, they can produce an approach which is fully differentiable to adapt their representation through objective training. This is not to state that previous approaches couldn't have been made differentiable, but the specific approach in this contribution takes specific advantage of current technologies for this.

Critical strengths

The paper is concisely presented. It goes all the way from the presentation and justification of the proposed model to evaluation with out-of-scanner markers. Furthermore they are versed in the general literature of neuroscience as well as methods, such that they can compare their take with the Margulies gradients approach which is an unsupervised continuous representation of the FC signal on the cortex with notable presence in the neuroscience field.

The method the authors propose poses in a clean manner questions that have been around for a while in neuroscience and neuroimaging. Wether the cortex is actually divided in discrete regions, wether the synthetic representation should be continuous and if this representation should be goal specific (for gender classification, age prediction, or cognitive ability prediction). The authors propose a method with several moving parts but also propose alation studies (e.g. table 6).

Critical weaknesses

I find two main weaknesses which I can present succinctly:

  1. First and foremost, why did you choose the t2t^2 statistic for the evaluation ? As the current literature uses generally balanced accuracy (for gender prediction) and Pearson’s r for age or cognitive score performance, the choice of showing t2t^2 makes the overall comparison with current literature hard for no discernible reason.
  2. As a follow-up, the differentiable spatial readouts (presented in Supp A) seem quite complex, if clever, in terms of the region representations. Per Table 6, these seem to be important, even key, to be relegated as a supplementary detail.
  3. In the same vein as 1, it would be great to have a good comparison with simpler methods proposed in the literature such as the Margulies gradients.
  4. The value of the individual + population model is hard to grasp wrt to a simple individual model.
Questions for authors
  1. Could you re-cast the results in a way that these can be compared easily with current literature ?
  2. Can you clarify more specifically the importance of the differentiable spatial readout with respect to the full continuous representation and a simpler readout? Table 6 is interesting but it's only shown in terms of age.
  3. How does this method compare in performance with other, simpler, continuous representations of the FC ?
7 comments
AuthorsOpenReview ↗

Response to Reviewer zkdh

We thank the reviewer for the thoughtful and encouraging review. We appreciate that the reviewer found the paper concise, recognized its clean framing of whether cortical representations should be discrete or continuous. The main concerns can be summarized as: (1) the use of t², (2) the importance and placement of the spatial readout, (3) comparison with simpler continuous representations, and (4) the value of the population anchor. We address each point in detail below:

Could you re-cast the results in a way that these can be compared easily with current literature, which generally assesses prediction performance (e.g. balanced accuracy for gender prediction)?

The goal of a brain–phenotype analysis is often not prediction alone but localization: identifying where in the brain a phenotype is expressed, so that the finding can be interpreted. We therefore would like to clarify that the scientific objective of Section 3.3 is to identify where in the brain FC is most strongly associated with a phenotype, which is distinct from predicting phenotypes using whole-brain FC. Accordingly, our evaluation is an association test, not a prediction task.

Specifically, we investigate whether continuous, gradient-based optimization enabled by our INR can identify region boundaries that yield stronger brain–phenotype associations. This represents a fundamentally new perspective that has not been systematically explored in previous work. Existing methods, including those suggested by Reviewer ubRR, primarily focus on how to optimally parcellate the whole brain. For this reason, we did not initially consider a direct comparison particularly appropriate.

Nevertheless, to address the reviewers' concerns, we adapted these methods to our evaluation setting by first deriving their respective parcellations and then identifying the most strongly associated FC connection using the training cohort. The following table, reporting Pearson's rr for age and Pearson's rr and Cohen's dd for sex, shows that our continuously optimized regions still achieve stronger associations on held-out subjects. However, we emphasize again that this evaluation procedure was not part of their original methods and represents our adaptation for comparison. If the reviewers consider this comparison appropriate and informative, we would be happy to include it in the final manuscript.

Held-out effect sizes, median [IQR] over 12 splits. All choices are made on training subjects by 5-fold cross-validation within each training split: the number of parcels for Schaefer, Ward and k-means (100 or 1000), the number of parcels and penalty λ\lambda for SBP (100 or 1000, λ∈0,0.1,0.3,1,3\lambda \in \\{0, 0.1, 0.3, 1, 3\\}), and r0∈4,8,12∘r_0 \in \\{4, 8, 12\\}^\circ and K∈2,4,6K \in \\{2, 4, 6\\} for Ours, nested. Ours, fixed: r0=8∘r_0 = 8^\circ, K=4K = 4. Bold: highest median per row (ties broken at the third decimal).

Effect sizeCohortSchaefer, tunedWardk-meansSBPOurs, fixedOurs, nested
Age, rrHCP-D0.51 [0.48, 0.53]0.40 [0.35, 0.45]0.40 [0.38, 0.42]0.49 [0.42, 0.52]0.52 [0.49, 0.56]0.52 [0.49, 0.55]
PNC0.36 [0.32, 0.38]0.33 [0.32, 0.35]0.34 [0.33, 0.36]0.37 [0.35, 0.39]0.37 [0.36, 0.39]0.37 [0.36, 0.39]
HBN0.32 [0.30, 0.34]0.33 [0.33, 0.35]0.34 [0.31, 0.35]0.34 [0.34, 0.38]0.36 [0.34, 0.38]0.37 [0.34, 0.38]
Sex, rrHCP-D0.17 [0.15, 0.21]0.17 [0.12, 0.23]0.20 [0.12, 0.26]0.18 [0.14, 0.22]0.23 [0.18, 0.26]0.23 [0.18, 0.26]
PNC0.16 [0.12, 0.22]0.17 [0.14, 0.19]0.18 [0.12, 0.20]0.10 [0.06, 0.13]0.18 [0.16, 0.20]0.20 [0.16, 0.21]
HBN0.14 [0.11, 0.16]0.13 [0.11, 0.14]0.12 [0.11, 0.13]0.09 [0.07, 0.11]0.15 [0.12, 0.16]0.14 [0.12, 0.15]
Sex, ddHCP-D0.35 [0.30, 0.43]0.35 [0.24, 0.48]0.40 [0.25, 0.53]0.37 [0.29, 0.46]0.49 [0.37, 0.53]0.47 [0.37, 0.54]
PNC0.32 [0.25, 0.46]0.35 [0.28, 0.40]0.36 [0.24, 0.40]0.21 [0.12, 0.26]0.36 [0.33, 0.40]0.41 [0.32, 0.43]
HBN0.29 [0.22, 0.33]0.28 [0.23, 0.29]0.25 [0.23, 0.27]0.18 [0.14, 0.22]0.32 [0.25, 0.34]0.30 [0.25, 0.32]

Why reporting t2t^2 instead Pearson's rr for age or cognitive score performance

Our reported t2t^2 statistics are directly related to Pearson's rr = t2/(t2+n−2)\sqrt{t^2 / (t^2 + n - 2)}, where nn is the number of held-out subjects. We only reported t2t^2 because it is a unified test statistic applicable to all analyses: sex binary comparison, age correlation, and multi-region readout (Hotelling's T2T^2). In this response, we further showed Pearon's rr and Cohen's dd in the above table, which lead to the same conclusion.

AuthorsOpenReview ↗

Response to Reviewer zkdh

Can you clarify more specifically the importance of the differentiable spatial readout with respect to the full continuous representation and a simpler readout?

The differentiable spatial readout allows us to directly optimize the location, size, and shape of cortical regions to maximize their association with a phenotype. This capability is naturally enabled by our INR, which provides a continuous, differentiable function of cortical position, allowing computationally tractable gradient-based optimization of region geometry. To our knowledge, no existing FC representation has been demonstrated for this purpose.

We agree that spatial optimization could, in principle, also be performed using discrete vertex-wise FC representations. However, this remains largely unexplored and would require additional methodological development, such as discrete combinatorial optimization or differentiable approximations of region boundaries. Our framework provides a direct and unified solution to this problem.

We would also like to clarify that all region parameterizations evaluated in Table 6 are differentiable spatial readouts, including simpler alternatives that optimize only the region center or radius. Once a region's geometry is parameterized, its connectivity readout can be differentiated with respect to those parameters. Table 6 therefore does not compare differentiable versus non-differentiable readouts. Instead, it examines how different levels of geometric flexibility affect the strength of brain–phenotype associations, providing empirical justification for our chosen parameterization.

Because the key conceptual innovation lies in making region geometry differentiable and optimizable, rather than in any particular geometric parameterization, we prioritized explaining the former in the main text and placed the technical details of the latter in the supplement to accommodate the page limit.

It would be great to have a good comparison with simpler methods such as the Margulies gradients?

We agree that the functional gradients of Margulies et al. (2016) and our population field are related, but we believe a direct performance comparison would be difficult because the two approaches serve different purposes.

Our work seeks to establish a continuous, individualized representation of FC that enables differentiable optimization of spatial regions for brain–phenotype associations. Our numerical fitting procedure anchors the population field to a group spectral embedding, whose spatial patterns correspond to the Margulies gradients, lending interpretability to our learned representation. However, the Margulies gradients themselves are defined at discrete cortical locations and do not directly provide a mechanism for optimizing spatial regions based on phenotype associations.

One possible comparison would be to develop a procedure that identifies clusters of vertices in the Margulies gradient maps whose average gradient values are maximally associated with a phenotype. If this is what the reviewer has in mind, we would be happy to conduct such an experiment. However, we emphasize that this would require substantial methodological development beyond the original Margulies framework.

What about other, simpler, continuous representations of the FC?

The most closely related work on continuous connectivity representations is Moyer et al., which focused on structural rather than functional connectivity. Importantly, their framework is not capable of continuous optimization of region geometry for phenotype associations. Instead, their Figure 3 demonstrates conventional statistical testing at discrete spatial locations, despite the underlying continuous representation.

AuthorsOpenReview ↗

Response to Reviewer zkdh

Value of the population + individual model over an individual-only model.

As we explained in lines 185-192, a field fitted to one subject's FC alone is defined only up to a rotation of its embedding space, because FC depends only on inner products between embeddings. Therefore, when fitted separately, embeddings at the same location cannot be compared across subjects without a separate alignment step. An analogous example is the Margulies gradients mentioned by the reviewer: gradients estimated independently for different subjects are not inherently aligned.

Our model builds this alignment in. Each individual field is a regularized deviation from a shared population field, so every subject is expressed in the same basis. This brings two benefits:

  • Comparability across subjects. Because every subject's field is expressed in the same coordinates, a region means the same thing in every brain. That is what lets us optimize regions on the training subjects and apply them unchanged to held-out subjects.

  • Reliability. The population field acts as a prior that shrinks noisy, scan-specific variation. On 104 HCP-D subjects with two runs (Table 2), the anchored field (λ=1\lambda = 1) predicts the subject's other run at r=0.583r = 0.583. This compares with 0.525 for an unconstrained individual fit (λ=0\lambda = 0), 0.504 for the population field alone, and 0.524 for the measured FC of the first run itself. The combined model is thus a better estimate of a subject's connectivity than either component alone, and better than the subject's own measured FC.

Comments on the paper

1 comment
AuthorsOpenReview ↗

To all reviewers and AC

We sincerely thank all reviewers for their thoughtful and constructive feedback and for engaging with us during the discussion period. Below, we summarize the scope of our work, the new experiments conducted during the rebuttal, and key clarifications.

Scope and contribution. Our work addresses localization: identifying where in the cortex functional connectivity (FC) is most strongly associated with a phenotype. Instead of discretizing the brain into predefined regions, we propose (i) a continuous, individualized representation of FC as an implicit neural representation (INR) anchored to a shared population field and (ii) differentiable optimization of region geometry (location, size and boundary) on this field, which the continuous INR naturally enables. Our evaluation is therefore a univariate association test of a single optimized connection, rather than multivariate whole-brain prediction or whole-brain parcellation.

New experiments and results

  1. Nested cross-validation and a fixed configuration (to reviewer kmUM, ubRR). We no longer report the best setting on held-out data. With r0r_0 and KK chosen by 5-fold cross-validation on training subjects, our method achieves a higher median held-out t2t^2 than all baselines in all six settings. With a configuration fixed in advance (r0=8∘r_0 = 8^\circ, K=4K = 4, no tuning), it is best in five of six settings and nearly tied in the sixth (15.9 vs. 16.0).

  2. Data-driven and supervised parcellation baselines (to reviewer kmUM, ubRR). We added Ward clustering, k-means and Supervised Brain Parcellation (SBP), each with its number of parcels (and penalty, for SBP) selected by nested cross-validation and evaluated on the same splits (revised Table 3). Our optimized regions remain stronger across cohorts.

  3. Effect sizes comparable with the literature (to reviewer zkdh). We report Pearson's rr and Cohen's dd alongside t2t^2. In PNC, the same cohort as Satterthwaite et al. (2013, 2015), the typical significant single edge reaches age r≈0.13r \approx 0.13–0.150.15 and sex d=0.32d = 0.32 in-sample. Our single optimized connection reaches age r=0.45r = 0.45 and sex d=0.56d = 0.56 in-sample, and still r=0.37r = 0.37 and d=0.41d = 0.41 on held-out subjects. Our effects are smaller than the multivariate benchmarks cited by Reviewer zkdh because they come from a single connection rather than whole-brain prediction. When used for whole-brain prediction with an identical predictor on HCP-D, our fields reach age r=0.87r = 0.87 and sex balanced accuracy 0.830.83, matching Schaefer-400 FC and outperforming diffusion-map gradients (Margulies et al., 2016) for age.

Clarifications. We clarified that fitting a test subject’s field is unsupervised and uses no phenotype information (to reviewer kmUM); and unlike spectral embeddings, INR fields are comparable across subjects and more reproducible (to reviewer ubRR).

We thank the reviewers again for their time and valuable suggestions, which have substantially strengthened our work, and the AC for considering this summary in the final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.