acceptodds
Under review as a conference paper at ICLR 2027

Geometric Feature Learning for Functional Data Valued on the Symmetric Positive Definite Manifold

Abstract

We here develop a functional neural network, termed MatFAE, for learning trajectories on the Riemannian manifold of symmetric positive definite (SPD) matrices. MatFAE features intrinsic layers that map manifold-valued functions to Euclidean vector-valued functions, followed by a functional layer that projects them into a finite-dimensional Euclidean space. Unlike most neural networks for discrete-time sequences, MatFAE treats each sequence as a continuous function and can therefore encode trajectory dynamics (e.g., first-order derivatives) in its latent representations. Additionally, the morphology of the functional weights in the functional layer offers interpretability by revealing the regions of the input functional data that contribute most to the latent representations. We justify the design principles and properties of each intrinsic layer and detail how matrix factorization is handled during backpropagation. We apply MatFAE to a range of fMRI datasets, demonstrating its ability to efficiently learn informative representations from high-dimensional SPD trajectories and its practical value for real-world neuroimaging analysis.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation learning, geometric deep learning, Riemannian geometry, SPD manifold, functional data analysis, interpretability, functional MRI

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_2YVTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The author introduces a framework for learning representations from trajectories of SPD matrices called MatFAE. Instead of treating the data as discrete sequences, the authors model them as continuous functions. MatFAE combines geometry-aware layers, including multi-head BiMap, manifold pooling, and geodesic shrinkage, with a functional autoencoder to learn low-dimensional representations and reconstruct SPD trajectories. The authors also provide theoretical analysis of the proposed layers. The method is evaluated on ten simulation scenarios and six resting-state fMRI datasets, using clustering metrics such as AMI and ARI. MatFAE generally outperforms methods including SPDNet, TSRVF, GeoAtt, and PGA. The authors also test a supervised classification variant and conduct ablation experiments. However, the clustering scores on real fMRI datasets remain relatively low.

Critical strengths
  1. I think the combination of SPD geometry and functional data learning is an interesting direction. MatFAE provides a structured way to learn from SPD trajectories while preserving their geometric properties. The theoretical analysis of the proposed layers is also a useful part of the work.
  2. The experimental evaluation is fairly extensive. The authors test MatFAE on ten simulation scenarios and six real fMRI datasets, and compare it with several existing geometric learning methods. I particularly appreciate the simulation experiments, as they examine different aspects of the trajectories, such as smoothness, temporal ordering, and transition dynamics. The additional ablation and classification experiments also provide some evidence that the learned representations are useful beyond reconstruction.
Critical weaknesses
  1. My main concern is the relatively low clustering performance on the real fMRI datasets. Although MatFAE outperforms the baselines, the AMI scores are only between 0.030 and 0.137. This makes me wonder what the learned clusters actually represent. The paper discusses identifying meaningful patient subgroups, but I am not sure the current results are enough to support this. It would be interesting to see whether the clusters are related to clinical symptoms, cognitive measures, or other meaningful brain features, rather than just diagnostic labels.
  2. I also wonder how much the results depend on the sliding-window settings. The authors use 30-second windows with a 4-second step, which creates substantial overlap. Since temporal dynamics are central to this work, I think the authors should check whether different window sizes or steps lead to similar results.
Questions for authors
  1. The clustering results on real fMRI datasets are relatively low in terms of AMI and ARI. Beyond matching diagnostic labels, have the authors examined what the learned clusters actually represent? For example, are they associated with symptom severity, cognitive performance, or other clinical information?
  2. The dynamic FC matrices are estimated using 30-second sliding windows with a 4-second step. How sensitive are the results to these choices? I am curious whether longer windows or less overlap would change the clustering performance.
  3. Since the paper focuses on learning temporal information from SPD trajectories, have the authors compared MatFAE with a static FC baseline using one covariance matrix per subject? It would be useful to know how much the temporal information actually contributes beyond static functional connectivity.
  4. The paper reports improvements in AMI and ARI, but are these differences statistically significant across subjects or repeated runs? I also wonder how stable the clustering assignments are when using different random seeds or slightly different preprocessing settings.
6 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1: what the learned clusters represent

We thank the reviewer for this important question.

Data and setup. We performed this analysis on CNP, the cohort with the most comprehensive clinical characterisation: clinician-rated symptom scales, self-report questionnaires and a neuropsychological battery, across four diagnostic groups. We applied the same pipeline as detailed in the manuscript plus motion correction (global signal regression, exclusion of the 47 subjects with mean head motion above 0.25 mm (257 → 210), and motion regressed out of the connectivity features). MatFAE never had access to diagnosis, symptoms or cognition, so all associations below are post hoc.

Cluster composition. With stricter motion control, MatFAE's agreement with diagnosis is similar to that under the manuscript's pipeline (ARI 0.136 vs. 0.124 ± 0.019; Table R6).

Table R9. Diagnostic composition of the four MatFAE clusters (n=210n = 210). Enrichment odds ratios are from Fisher's exact test (Holm-corrected p<0.001p < 0.001).

ClusternnControlADHDBipolarSchizophreniaEnriched for (odds ratio)
C167373126Schizophrenia: 79% of all patients with schizophrenia (12.3)
C261406132Controls: 66% of the cluster
C357249222Bipolar disorder: 59% of all patients with bipolar disorder (5.8)
C42531813ADHD: 50% of all patients with ADHD (23.9)

The clusters are strongly associated with diagnosis (χ2(9)=125.1\chi^2(9) = 125.1, p<0.001p < 0.001). The association is considerably stronger among patients alone (n=106n = 106; ARI 0.334, χ2(6)=90.5\chi^2(6) = 90.5, p<0.001p < 0.001). C1, C3 and C4 each concentrate one patient group. The overall ARI is low mainly because controls do not form a cluster of their own: 97% of controls are spread over C1-C3. The low AMI and ARI therefore reflect how controls are distributed across clusters, not clusters unrelated to psychopathology.

Symptoms and cognition.

Table R10. Cluster means of clinical and cognitive composites, in SD units. Clinician-rated composites are standardised over the 106 patients; the other composites are standardised to controls (n=210n = 210). η2\eta^2 is the variance explained; pp is from the Kruskal-Wallis test. The last column is the FF-test of adding clusters to diagnosis.

CompositeC1C2C3C4η2\eta^2 (clusters)ppη2\eta^2 (diagnosis)pp (clusters beyond diagnosis)
Psychosis (BPRS positive/negative, SAPS, SANS)0.55-0.46-0.16-0.480.286< 0.0010.4180.115
Affective (BPRS mania, depression/anxiety; HAM-D; YMRS)-0.16-0.070.26-0.100.0570.1580.0530.153
Self-report (12 questionnaires)0.360.410.840.940.0620.0030.3620.373
Cognition (18 tests)-0.48-0.06-0.22-0.440.0690.0080.3020.497

The clusters differ significantly in psychosis severity, self-reported symptoms and cognition, and all three differences remain significant after Holm correction across the four composites.

  • Psychosis severity explains 29% of the variance among patients and is highest in C1, the schizophrenia-enriched cluster.
  • Self-reported symptoms are highest in C3 and C4.
  • Cognition is lowest in C1 and C4, nearly half an SD below controls.
  • Affective symptoms do not differ between clusters (p=0.158p = 0.158).

These differences largely follow diagnosis: adding the clusters to diagnosis does not significantly improve the fit for any composite (p≥0.115p \ge 0.115).

Connectivity beyond diagnosis. We compared the clusters on 28 static connectivity features: the mean Fisher-zz connectivity within and between the seven Yeo networks (Schaefer-100 parcellation), with Benjamini-Hochberg FDR correction. Nine features differ between clusters (q<0.05q < 0.05), and five of these involve the default-mode network. C4 has the highest connectivity within the default-mode network (η2=0.109\eta^2 = 0.109) and the lowest connectivity between the default-mode and dorsal-attention networks (η2=0.090\eta^2 = 0.090), making its default-mode network the most segregated. Diagnosis explains at most 5% of the variance in any of these features, and adjusting for diagnosis changes little (partial η2\eta^2 = 0.112 and 0.098; eight of the nine features remain at q<0.05q < 0.05). Because the clusters are learned from connectivity, these tests are descriptive rather than independent validation.

Summary. The clusters correspond to recognisable clinical profiles:

  • C1 is enriched for schizophrenia, with the highest psychosis severity and the lowest cognition.
  • C2 is dominated by controls and has low symptom levels.
  • C3 is enriched for bipolar disorder, with elevated self-reported symptoms.
  • C4 is enriched for ADHD, with the highest self-reported symptoms, reduced cognition and the most segregated default-mode network.

We will add this analysis to the appendix.

AuthorsOpenReview ↗

Response to Weakness 2 and Question 2: sensitivity analysis of the sliding window

We thank the reviewer for this suggestion. We agree that robustness to the sliding-window settings is important, because temporal dynamics are central to this work.

We note first that MatFAE treats each trajectory X(t)X(t) as a continuous function of time, so the windowed matrices are samples of a continuous trajectory. Overlapping windows are therefore the natural way to sample such a trajectory, and the stride mainly determines the sampling density.

To assess sensitivity, we repeated the COBRE analysis with window lengths from 30 s to 150 s (Table R11). Each window has 150 volumes, so longer windows also yield fewer windows per subject. We increased the stride for the longest windows accordingly.

Table R11. Sensitivity of MatFAE to window length and stride on COBRE (TR = 2 s). The number of clusters is fixed at their known value.

Window (TR / s)Stride (TR)Windows per subjectMatFAE ARI
15 / 30 (default)2680.164
30 / 602610.134
45 / 903360.072
60 / 1203310.047
75 / 1505160.040

It can be seen that performance is stable across the commonly used range. Between 30 s and 60 s windows, ARI changes only from 0.137 to 0.134, so the results do not depend on the particular choice within this range. However, performance declines for long windows, as expected when temporal dynamics carry the signal. The drop in ARI shows that MatFAE exploits temporal structure rather than static connectivity. (As pointed out by Reviewer 2YVT, in our conclusion, we mistakenly compared with the AMI value (0.137), while the ARI value is 0.164. Table R11 shows that MatFAE performs best with the shortest tested window, and its ARI decreases consistently as window length increases.)

Choosing the window remains a modelling decision that depends on the data and the scientific question, for example the repetition time, the frequency band and the expected time scale of connectivity changes. Optimising it is beyond the scope of this work. Our Python package therefore takes the window length and stride as user-specified parameters.

AuthorsOpenReview ↗

Response to Question 3: comparison with static FC baseline

We thank the reviewer for this question. To isolate the contribution of temporal information, we compare full MatFAE with two types of controls:

  • Static FC baselines, which use one covariance matrix per subject computed over the entire run: tangent PCA followed by kk-means, and static SPDNet autoencoder.

  • Time-shuffled MatFAE, in which each subject's windows are randomly permuted before training. This keeps the set of visited connectivity states but destroys their order and rate of change.

Table R12. Mean ± standard deviation of ARI over 100 repetitions, with the number of clusters fixed at the true value. Full MatFAE values are from Table 3.

RungStatic PCAStatic SPDNetMatFAE, time-shuffledMatFAE (full)
A (temporal ordering)-0.005 ± 0.0080.000 ± 0.018-0.004 ± 0.0051.000 ± 0.000
B (static identity)0.231 ± 0.200-0.003 ± 0.0090.425 ± 0.1960.980 ± 0.009
C (dwell time)0.384 ± 0.0780.359 ± 0.0670.432 ± 0.0680.541 ± 0.032
D (transition frequency)-0.001 ± 0.008-0.004 ± 0.005-0.004 ± 0.0070.286 ± 0.052
E (smoothness)0.056 ± 0.0440.157 ± 0.0560.157 ± 0.0770.808 ± 0.078
F (Wishart concentration)0.056 ± 0.0340.008 ± 0.0220.999 ± 0.0051.000 ± 0.000
G (real HRF)0.040 ± 0.040-0.003 ± 0.0090.042 ± 0.0410.960 ± 0.018
H (trajectory direction)0.004 ± 0.0210.002 ± 0.0150.008 ± 0.0271.000 ± 0.000
I (multi-scale dynamics)0.011 ± 0.0360.011 ± 0.0370.089 ± 0.0640.085 ± 0.064
J (Neurosynth)0.579 ± 0.2580.632 ± 0.3110.760 ± 0.1681.000 ± 0.000

Table R13. ARI over 20 seeds

CohortStatic PCAStatic SPDNetMatFAE, time-shuffledMatFAE (full)
CNP0.005 ± 0.0060.021 ± 0.0070.002 ± 0.0060.124 ± 0.019
COBRE0.043 ± 0.0050.063 ± 0.0000.075 ± 0.0130.164 ± 0.012
ADHD-2000.037 ± 0.013-0.007 ± 0.0050.026 ± 0.0240.103 ± 0.021
ABIDE-0.001 ± 0.001-0.002 ± 0.001-0.001 ± 0.0010.039 ± 0.010
TCP0.009 ± 0.002-0.010 ± 0.0030.020 ± 0.0130.055 ± 0.018
CAT-D-0.006 ± 0.000-0.007 ± 0.0040.007 ± 0.0130.079 ± 0.011

Full MatFAE outperforms both static baselines and the time-shuffled control on nine of the ten rungs. On rungs whose group differences are purely temporal (A, D, E, G, H), the static baselines and the time-shuffled control stay near chance (ARI ≤ 0.16). Where static connectivity carries part of the signal (C, J), temporal information still adds substantially (0.541 vs. ≤ 0.432 on C; 1.000 vs. ≤ 0.760 on J). On B and F, averaging over the whole run also removes non-temporal differences that time-resolved trajectories preserve: a briefly visited state on B, and connectivity concentration on F. The only exception is rung I, which remains difficult for all methods.

Reviewer_AyS8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes MatFAE, which combines multi-head BiMaps, SPD-manifold pooling, geodesic-shrinkage activation, and a functional autoencoder to map SPD matrix trajectories into low-dimensional Euclidean functions for clustering and classification. MatFAE combines SPD geometry with functional neural networks, and its manifold constraints, layer definitions, and gradient derivations are presented in a technically structured manner. The experiments cover synthetic data, real fMRI, classification, ablations, and hyperparameter sensitivity, with results supporting advantages on several dynamic structures. However, the novelty boundary is not established through a systematic comparison with the closest functional and manifold models, and the empirical evidence for the multi-head mechanism and continuous-time/derivative modeling remains limited.

Critical strengths
  1. Geometry-preserving architecture. Figure 1 and Eqs. (1), (6), and (10) describe SPD-preserving BiMap, manifold averaging, activation, and logarithmic mapping operations; this provides a technically principled construction. Section 2 and Propositions 1–3 analyze pooling equivariance, Lipschitz behavior, and contraction properties; this connects architectural choices to explicit geometric properties.
  2. Modular functional modeling. Section 3 and Eqs. (2)–(4) combine functional weight layers with a geometry-aware decoder, allowing the model to process complete trajectories rather than isolated snapshots. Figure 2 distinguishes the functional layers from the MLP autoencoder, while Section 1 motivates continuous-function, morphological, and dynamic representations.
  3. Broad empirical coverage. Tables 1 and 3 evaluate ten structured simulation scenarios with means and standard deviations over 100 repetitions, covering temporal ordering, dwell time, smoothness, and multi-scale dynamics. Tables 2 and 4 evaluate six real fMRI cohorts against TSRVF, GeoAtt, SPDNet, PGA, DC, and PDM.
Critical weaknesses
  1. Insufficiently established novelty boundary. Sections 1 and Appendix A state that MatFAE is the first network for this setting, but do not provide a systematic taxonomy or itemized comparison with the closest alternatives. No direct evidence found in the manuscript.
  2. Limited methodological novelty beyond component integration. MatFAE appears to be largely constructed by integrating several existing ideas: the BiMap operation is inherited from SPDNet and extended to multiple heads (Sec. 2.1), while the functional encoder–decoder framework is explicitly adapted from FAEclust (Sec. 3). The Log-Euclidean/Log-Cholesky operations and Fréchet pooling are also based on established SPD-manifold constructions (Sec. 2). Although the geodesic-shrinkage activation and row-space diversity regularization may constitute new design elements (Sec. 2.3), the manuscript does not demonstrate that the overall contribution is more than a task-specific combination of prior modules. No direct evidence found in the manuscript that the proposed integration yields a qualitatively new representation principle rather than an effective architectural composition.
  3. The current research gap is not introduced clearly in the paper. Especially I did not see it in the abstract, which makes me confusing. The abstract does not explicitly identify the concrete limitation of existing approaches or explain why the proposed new method is needed.
  4. “In words: we shrink the coordinate vector toward zero” is not proper academic writing. Same problem goes to “To keep overhead negligible, rather than “pool, then activate,” we combine the Pooling layer and the Activation layer into one layer, namely, the “Activated Pooling” layer in Figure 1.”
  5. The empirical improvements are heterogeneous and often modest, particularly on the real-data tasks. The current evidence does not yet establish that the observed gains are consistently large or practically meaningful across tasks.
Questions for authors
  1. What fundamental methodological insight distinguishes MatFAE from a composition of SPDNet-style BiMaps, FAEclust-style functional layers, and established Fréchet mean operations? The overall architecture relies substantially on existing components. The methodological novelty is not yet convincingly established for an ICLR-level contribution.
  2. How does the geodesic-shrinkage activation differ mathematically and empirically from existing SPD nonlinearities or regularization mechanisms, and what specific benefit does it provide (Sec. 2.3)?
  3. What quantitative evidence supports the statement that combining pooling and activation introduces “negligible” overhead, and how does its runtime, memory use, and asymptotic complexity compare with implementing the two operations separately?
  4. The baseline methods are outdated. The latest method is in 2022. Can the authors consider more recent functional autoencoders, manifold neural networks, or SPD sequence models to show the ability of the method?
8 comments
AuthorsOpenReview ↗

Response to methodological novelty (Weakness 1& 2 and Question 1)

We apologize that our presentation was difficult to follow. Space constraints limited our explanation of the technical contributions. We here would like to clarify that MatFAE is not a simple combination of existing components:

(1) Biophysically-Motivated Regularized Multi-head BiMap

The choice of BiMap is not purely geometry, we here further provide a biophysical motivation. Let Z=(Z1,…,Zm)⊤\mathbf{Z} = (Z_1, \ldots, Z_m)^\top denote a high-dimensional random vector, where each ZiZ_i represents the standardized fMRI signal from one ROI, and let W∈Rm1×m\mathbf{W} \in \mathbb{R}^{m_1 \times m}, with m1<mm_1<m, be a full-row-rank matrix with orthonormal rows. The transformed signal WZ\mathbf{W}\mathbf{Z} provides a lower-dimensional representation obtained by combining ROI-level signals. In this sense, W\mathbf{W} can be interpreted as a learnable analogue of parcellation in its dimensionality-reduction role: rather than relying on a predefined atlas, the transformation learns combinations of ROI-level signals that produce more stable low-dimensional representations. Correspondingly, the BiMap operation $ \mathbf{X} \mapsto \mathbf{W}\mathbf{X}\mathbf{W}^{\top}: \operatorname{cov}(\mathbf{Z}, \mathbf{Z}) \mapsto \operatorname{cov}(\mathbf{W}\mathbf{Z}, \mathbf{W}\mathbf{Z}) $ preserves second-order dependence structures while transforming high-dimensional ROI-level functional connectivity into a lower-dimensional representation.

For functional data, a single BiMap cannot capture the heterogeneous patterns that evolve over the functional domain. We therefore introduce multiple BiMap heads to learn complementary representations. However, having multiple heads is not guaranteed to capture complementary information. We further develop a regularization mechanism that explicitly promotes diversity among the heads (Appendix F). The regularized multi-head BiMap constitutes a key contribution of MatFAE.

(2) A New Principled Activation Mechanism

Pooling is required to aggregate the complementary representations learned by the multiple BiMap heads. Our theoretical analysis establishes that Fréchet-mean pooling is nonlinear and non-contractive (Propositions 2-4), and these properties directly motivate the design of the activated pooling layer: (1). The nonlinearity eliminates the need for a conventional activation such as ReLU. (2). Due to the non-contractiveness, we introduce geodesic shrinkage to control the magnitude and dispersion of the pooled representation.

Geodesic shrinkage is a new, principled activation mechanism applicable to general manifolds, not only SPD manifolds. Unlike existing methods, it is scale-invariant, computationally efficient, and provides a direct geometric analogue of weight decay in Euclidean neural networks.

(3) Intrinsic Interpretation of SPD Trajectories

After mapping the low-dimensional SPD trajectories to a tangent space, MatFAE processes them using FAEclust, motivated by its universal approximation property and interpretable functional weights. Importantly, MatFAE extends this interpretability to the original SPD-valued inputs. This is non-trivial because the functional layer contains thousands of weight functions whose contributions must be systematically aggregated to obtain a meaningful interpretation of the input SPD trajectory. Appendix G.6 provides a step-by-step procedure for performing this interpretation (Figs. 4&5, which are validated against ground-truth neuroscientific findings). The corresponding visualization and interpretation tools are included in our Python package.

(4) A Unified, Neuroscience-Ready Framework

  • To the best of our knowledge, MatFAE is the first structurally lightweight neural architecture designed specifically for SPD-valued functional data, a data structure commonly encountered in functional neuroimaging. We provide MatFAE as a fully documented Python package (anonymous GitHub link in the paper), with sensible default hyperparameter settings and built-in tools for visualization and interpretation.
  • Existing relevant neural models are all for supervised classification. In contrast, MatFAE supports supervised classification, semi-supervised learning, and unsupervised clustering within a unified framework. This capability is particularly important in neuroscience, where many neurological and psychiatric conditions are heterogeneous and may contain clinically meaningful subtypes that are not known in advance. MatFAE can therefore be used to identify latent subgroups and support symptom-driven, transdiagnostic characterizations of psychiatric conditions.
AuthorsOpenReview ↗

Response to methodological novelty (cont.)

Our ablation and sensitivity study in Appendix G.5 (and additional results in Tables R1, R2, R5) confirmed all the structural components are essential to MatFAE's performance.

As required in Weakness 1, we here provide an itemized comparison.

MethodInput objectIntrinsic SPD geometryLearnedUses temporal orderTime-resolution independentFirst-order sensitivityUnsupervised objectiveTemporal interpretability
SPDNet (2017), ManifoldNet (2022)single SPD matrixyesyesno (needs external pooling)nonono (supervised)no
LieBN (2024), MMNN++ (2024)single SPD matrixyesyesno (needs external pooling)nonono (supervised)no
SPD-SRU (2018)SPD sequenceyesyesyes (recurrence)noimplicitno (supervised)no
GeoAtt (2022)SPD sequenceyesyesyes (GRU)noimplicityes (hidden state)no
STAGIN (2021)graph sequencenoyesyes (attention)noimplicitno (supervised)attention maps
TSRVF (2020)SPD trajectoryyes (AI metric)noyes (warping)yesyes (SRVF)distance onlyno
PGA / PDMSPD trajectoryyesnoPGA: via fPCA; PDM: noPGA: yesnodistance onlyno
DC (MLP/1-D CNN AE)vectorised entriesnoyesyesnoimplicityes (Euclidean)no
FAEclust (2025)Euclidean functionnoyesyesyesyes (functional)yes (Euclidean)yes (W(t)\mathbf W(t))
MatFAESPD-valued functionyes (LE/LC)yesyesyesyesyes (Riemannian reconstruction)yes (W(1)(t)\mathbf W^{(1)}(t))
AuthorsOpenReview ↗

Response to Weakness 3-5

Response to writing (Weakness 3 & 4)

We thank the reviewer for the suggestions. We will explain the research gap in the abstract, and refine the two sentences mentioned by the reviewer.

Response to empirical improvements (Weakness 5)

On the synthetic ladder, MatFAE achieves the best or joint-best performance.

On real data, all methods yield low absolute AMI/ARI scores, which may partly reflect the imperfect alignment between DSM-based diagnostic labels and the underlying data structure; the supervised accuracies cited in lines 421-427 also indicate that predicting these labels is challenging. Nevertheless, MatFAE’s AMI is 1.4-3.0 times that of the strongest baseline on every cohort, and MatFAE-C performs best across all six cohorts (Table 5), with performance above chance on each (Table 6).

In our response to Reviewer 9G2p (Q4), we also performed paired test against the strongest baseline (Tables R6 and R7). The results show that MatFAE outperforms the strongest clustering baseline on all six cohorts, and the gains are significant.

Reviewer_9G2pOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents MatFAE, a functional autoencoder for learning representations of trajectories on the SPD manifold, motivated by dynamic functional connectivity analysis. The architecture combines multi-head bilinear dimensionality reduction, Fréchet-mean pooling, and geodesic shrinkage under Log-Euclidean or Log-Cholesky geometry, before mapping the resulting trajectories to Euclidean vector-valued functions. These functions are processed by an adapted FAEclust encoder, while a geometry-aware decoder reconstructs the input SPD trajectories. The paper analyzes properties of the geometric layers and derives the gradients needed for training. It also proposes interpreting the temporal profiles of the functional encoder weights to identify influential trajectory intervals. Validation includes clustering experiments on ten simulation scenarios and six resting-state fMRI datasets, supervised classification using the encoder, and component ablations and sensitivity analyses. The reported results show generally competitive or improved clustering performance against the selected baselines and higher balanced classification accuracy across the six real datasets.

Critical strengths
  1. The paper addresses a useful representation-learning problem, i.e., modeling entire SPD-valued trajectories while accounting for both matrix geometry and temporal variation. Connecting SPD dimensionality reduction with functional encoding offers a practical approach to this problem, particularly for dynamic functional connectivity analysis.

  2. In this paper, the authors select ten simulated scenarios, each characterized by distinct trajectory features, alongside six real-world fMRI datasets. The high-quality clustering results, combined with evaluations of a supervised encoder, demonstrate that the representations learned by this integrated architecture are generalizable, extending well beyond the confines of any single dataset or narrowly defined synthetic environment.

Critical weaknesses
  1. The architecture largely combines established SPD operations with an adapted FAEclust network. The time-independent BiMap heads do not explicitly model temporal neighborhoods, and the functional encoder does not explicitly incorporate trajectory derivatives, leaving the claimed temporal specialization and first-order modeling benefits insufficiently justified. Ablations on three synthetic scenarios provide some support, but corresponding real-data ablations and comparisons against parameter-matched functional baselines would be needed to establish the source of the gains.

  2. Propositions 2–3 and their proofs establish Lipschitz upper bounds, but these bounds do not imply the stated failure of non-expansiveness or contractivity. Moreover, although geodesic shrinkage contracts distances under the chosen metric, its composition with the preceding mapping is guaranteed contractive only when the shrinkage factor times that mapping’s Lipschitz constant is below one. This condition is not established, weakening the stability argument used to motivate a central component.

Questions for authors
  1. What exactly enables a time-agnostic BiMap head to become specialized for different temporal regimes? Does the functional encoder explicitly capture first-order dynamics, or is this merely an empirical interpretation? Please distinguish these arguments from the capacity to "learn time-dependent weighted integration."

  2. The proofs of Propositions 2–3 establish Lipschitz upper bounds. What additional argument establishes the claimed failure of non-expansiveness? Under what conditions does the shrinkage factor make the composite activated-pooling mapping contractive?

  3. Under the LE metric, Eq. (4) constrains the reconstructed matrix logarithm to the form M2A(t)M2⊤M_2 A(t)M_2^\top. When p2<mp_2<m, this forces at least (m−p2m-p_2) eigenvalues of every reconstructed SPD matrix to equal one. How restrictive is this shared subspace constraint for the evaluated trajectories, and what reconstruction errors does it produce?

  4. On the real datasets, how does MatFAE compare with a similarly sized FAEclust model using vectorized log-mapped trajectories? This comparison would help determine whether the geometric modules explain the gains. Please also clarify whether Table 5 uses five-fold or (5\times3) cross-validation, whether test folds were excluded from hyperparameter selection, and whether the gains over the strongest baselines are statistically reliable (should provide the p-value\textit{p-value} results).

4 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1

We apologize for not clearly stating our contributions due to space constraints. Reviewer AyS8 also requested clarification on this point; please see our first response to Reviewer AyS8.

Time-independent BiMap head: We would like to clarify that, although each BiMap head is time-invariant, the multi-head BiMap layer is motivated directly by the functional nature of the input. Because X(t)X(t) at different time points generally does not share a common eigenbasis, no single projection W1W_1 can preserve the discriminative covariance structure over tt. Multiple BiMap heads, encouraged by the row-space decorrelation penalty (Appendix F) to project onto complementary subspaces, retain these different structures in parallel. Temporal selectivity is then provided by the functional weight W(1)(t)W^{(1)}(t), which can weight each head’s output differently over time. The heads therefore specialize in subspace, and W(1)(t)W^{(1)}(t) determines when each subspace matters. This division of labour is also supported empirically: reducing to a single head collapses ARI on Scenario E from 0.808 to 0.031 (Table 7), and performance increases with H until the relevant regimes are covered (Table 8).

We would also like to note that snapshot models cannot do this. Any symmetric pooling is invariant to permutations of the time grid, so a per-frame network with pooling cannot separate classes that differ only in temporal order (rung A), switching frequency (D), smoothness (E) or velocity profile (H). Table 1 confirms this (SPDNet AMI 0.000/0.000/0.009/0.029 against MatFAE 1.000/0.368/0.725/1.000).

First-order dynamics: In our experiments, the functional weights in the encoder layer are B-spline expansions that can concentrate on local regions of the domain. A weight with adjacent positive and negative lobes acts as a smoothed finite-difference filter: its inner product with y(t)y(t) approximates a local rate of change. More generally, integration by parts gives ∫t0t1W(t)y′(t)dt=W(t)y(t)∣t0t1−∫t0t1W′(t)y(t)dt\int_{t_0}^{t_1} W(t)y'(t)dt = \left.W(t)y(t)\right|_{t_0}^{t_1} - \int_{t_0}^{t_1} W'(t)y(t)dt. Any linear derivative-based feature is therefore (approximately) representable as an encoder feature on y(t)y(t) itself, without differentiating the noisy input. MatFAE thus does not hard-code derivatives, but it can learn first-order features alongside broader functional patterns.

Ablation study on COBRE: For the source of the gains on real data, Table R1 repeats the Table 7 ablation on COBRE with 20 seeds. Full MatFAE achieves the highest ARI on COBRE, with performance declining under every ablation and most sharply when removing the module F\mathcal F (reduced to FAEclust), supporting the contribution of each component.

Table R1. Component ablation on the COBRE dataset (ARI with number of clusters fixed at the ture value).

VariantCOBRE ARI
Full MatFAE0.164 ±\pm 0.012
multi-head H→1H\to10.032 ±\pm 0.020
λdiv=0\lambda_{\mathrm{div}}=00.072 ±\pm 0.011
functional layers0.034 ±\pm 0.023
module F\mathcal F (direct half-vectorization)-0.003 ±\pm 0.008
module F\mathcal F (Riemannian logarithm vectorization)0.021 ±\pm 0.008
second BiMap layer0.075 ±\pm 0.017
geodesic shrinkage (α→1\alpha\to1)0.116 ±\pm 0.010

Ablation study on BiMap heads: Table R2 reports respectively the ARI of the full model, no-decorrelation-penalty model and single-head model. The full model achieves the highest ARI across all datasets, supporting the benefits of multiple heads and the decorrelation penalty, with removing either causing a pronounced drop (especially on COBRE).

Table R2. Head-specialisation diagnostics.

DataARI: full / λdiv=0\lambda_{\mathrm{div}}=0 / H=1H=1
A (temporal ordering)1.000 / 0.992 / 0.888
C (dwell time)0.536 / 0.490 / 0.310
D (transition frequency)0.280 / 0.251 / 0.135
F (Wishart concentration)1.000 / 0.991 / 0.863
COBRE0.164 / 0.072 / 0.032

Comparisons against parameter-matched FAEclust is provided in Table R5, in response to Question 4.

AuthorsOpenReview ↗

Response to Weakness 2 and Question 2

We thank the reviewer for these thoughtful comments. To avoid a notational clash with the shrinkage factor α\alpha, we write the spectral band in Propositions 2-3 as Xαs,β\mathcal{X}_{\alpha_s,\beta} = {X:αsI⪯X⪯βIX : \alpha_s I \preceq X \preceq \beta I} and the pooling map as FF.

Q2(a): What additional argument establishes the failure of non-expansiveness?

We agree that the Lipschitz upper bounds in Propositions 2-3 do not, by themselves, establish that FF fails to be non-expansive. However, input pairs whose geodesic distance increases under FF are easy to construct: they arise whenever the matrix congruence W(⋅)W⊤W(\cdot)W^\top mixes a large eigenvalue of an ill-conditioned input into a small output eigenvalue. For example, under the LE metric, take m=2m = 2, m1=1m_1 = 1, H=1H = 1. Let W=e1⊤W = e_1^\top, Z=diag(0.01,1)Z = \mathrm{diag}(0.01, 1), and X=RZR⊤X = RZR^\top, where RR is the rotation with sin⁡(δ)=0.2\sin(\delta) = 0.2. Then X11=0.01cos⁡2(δ)+sin⁡2(δ)=0.0496X_{11} = 0.01\cos^2(\delta) + \sin^2(\delta) = 0.0496, so dLE(F(X),F(Z))=log⁡(0.04960.01)=1.60, dLE(X,Z)=∥Rlog⁡(Z)R⊤−log⁡(Z)∥F=2sin⁡(δ)log⁡(100)=1.30,d_{LE}(F(X), F(Z)) = \log(\frac{0.0496}{0.01}) = 1.60, ~d_{LE}(X, Z) = \|R\log (Z) R^\top - \log (Z)\|_F = \sqrt{2}\sin(\delta)\log (100) = 1.30, a ratio of 1.23>11.23 > 1.

The mechanism is that matrix congruence does not commute with the matrix logarithm: log⁡(WXW⊤)≠Wlog⁡(X)W⊤\log(WXW^\top) \neq W\log(X)W^\top, so a small change in log⁡(X)\log (X) can produce a larger change in log⁡(WXW⊤)\log(WXW^\top) when XX is ill-conditioned. Analogous examples under the LC metric also exist. Since FF can strictly increase distances, it is neither non-expansive nor contractive in general.

In the revision, we will clarify that the upper bound is a worst-case statement, and that for particular weights and inputs (e.g., with many heads), the pooling map may well be contractive.

Q2(b): Under what conditions is the activated-pooling map contractive?

Following Section 2.3, we write the pooling map, the geodesic-shrinkage activation and their composite as F(X)=ϕ−1(1H∑h=1Hϕ(W1hX(W1h)⊤)),S(Y)=ϕ−1(αϕ(Y)),F(X) = \phi^{-1}\Big(\frac{1}{H}\sum_{h=1}^H \phi\big(W_1^h X (W_1^h)^\top\big)\Big), \qquad S(Y) = \phi^{-1}\big(\alpha\phi(Y)\big), (S∘F)(X)=ϕ−1(αH∑h=1Hϕ(W1hX(W1h)⊤)),(S \circ F)(X) = \phi^{-1}\Big(\frac{\alpha}{H}\sum_{h=1}^H \phi\big(W_1^h X (W_1^h)^\top\big)\Big), where the last expression is the activated-pooling layer used in Eq. (1). Under both the LE and LC metrics, ϕ\phi is a global isometry onto a Euclidean space, and SS is linear scaling in ϕ\phi-coordinates. Hence $ d_g(S(A), S(B)) = \alpha d_g(A, B) \text{ for all } A, B, $ not only for distances from II. Consequently, the activated-pooling map S∘FS \circ F satisfies $ \mathrm{Lip}(S \circ F) = \alpha \cdot \mathrm{Lip}(F) $ exactly. This gives a necessary and sufficient condition: on Xαs,β\mathcal{X}_{\alpha_s,\beta}, S∘FS \circ F is contractive if and only if α⋅Lip(F)<1\alpha \cdot \mathrm{Lip}(F) < 1. Combined with Propositions 2-3, this yields:

  • under the LE metric, S∘FS \circ F is contractive on Xαs,β\mathcal{X}_{\alpha_s,\beta} whenever α<αs/β\alpha < \alpha_s / \beta;
  • under the LC metric, S∘FS \circ F is contractive on Xαs,β\mathcal{X}_{\alpha_s,\beta} whenever α<αs/(2β3/2max⁡(1,αs−1))\alpha < \alpha_s / (2\beta^{3/2}\max (1, \alpha_s^{-1})).

We will include the above condition in the revised version. Independently of contractivity, the activation always provides spectral control, since dg(I,X3)=αdg(I,X2)d_g(I, X_3) = \alpha d_g(I, X_2) shrinks the log-spectrum toward zero. Empirically, the activation also shows benefits. Removing it (α→1\alpha \to 1) reduces the mean ARI from 0.808 to 0.624 on rung E, from 0.960 to 0.804 on rung G, and from 1.000 to 0.960 on rung J (Table 7).

AuthorsOpenReview ↗

Response to Question 3

We thank the reviewer for this precise question.

The constraint is restrictive for reconstruction but not for the clustering objective. MatFAE is trained for representation learning: the decoder supplies a training signal that requires the latent trajectory to retain the dominant structure, and it is not intended as a generative model of every eigen-direction.

How restrictive is the constraint? The constraint has three components, which we address in turn.

(i) Rank. Each time varying functional connectivity (TVFC) trajectory is estimated from a sliding window of ww time points with w≪mw \ll m (w=15w = 15 for COBRE and CNP, w=12w = 12 for CAT-D, m=100m = 100). Each window's sample matrix C(t)C(t) therefore has rank at most w−1w - 1, and the Ledoit-Wolf estimate X(t)=(1−δ)C(t)+δμIX(t) = (1-\delta)C(t) + \delta\mu I assigns the same eigenvalue δμ\delta\mu to the entire null space of C(t)C(t). Hence every input already satisfies log⁡(X(t))=log⁡(δμ)I+(a term of rank≤w−1),\log (X(t)) = \log(\delta\mu) I + (\text{a term of rank} \le w - 1), so at least m−(w−1)m - (w - 1) of its eigenvalues are identical. The low-rank-plus-identity form imposed by Eq. (4) is therefore already present in the data, and with p2≥w−1p_2 \ge w - 1 the rank of the decoder is not the binding restriction.

(ii) Offset. The repeated eigenvalue of the data is δμ\delta\mu rather than one. This mismatch is a single scalar per window and can be removed exactly by adding a learnable term c(t)Ic(t) I to the decoder.

(iii) Shared subspace. The genuinely restrictive component is that M2M_2 is shared, whereas the informative subspace of each window varies across time and subjects. We quantify its cost below.

What reconstruction errors does it produce? Write Si(t)=log⁡(Xi(t))S_i(t) = \log(X_i(t)). All sums run over subjects ii and windows tt, and values are medians over 20 seeds. We report in Table R3:

  • the achieved decoder error: ∑i,t∣Si(t)−S^i(t)∣F2/∑i,t∣Si(t)∣F2\sum_{i,t}\\|S_i(t)-\hat{S}_i(t)\\|_F^2 / \sum_{i,t}\\|S_i(t)\\|_F^2;
  • the floor of the trained subspace: ∑i,t∣Si(t)−PSi(t)P∣F2/∑i,t∣Si(t)∣F2\sum_{i,t}\\|S_i(t)-PS_i(t)P\\|_F^2 / \sum_{i,t}\\|S_i(t)\\|_F^2, where P=M2M2⊤P = M_2M_2^\top is the orthogonal projector onto col(M2)\mathrm{col}(M_2);
  • the floor of the best shared subspace: the same expression evaluated at P⋆=arg⁡max⁡rank(P)=p2∑i,t∣PSi(t)P∣F2P^\star = \arg\max_{\text{rank}(P)=p_2}\sum_{i,t}\\|PS_i(t)P\\|_F^2, computed by HOSVD followed by HOOI refinement. The floor is the lowest error any decoder of the form M2AM2⊤M_2AM_2^\top can reach. Since ∣S−PSP∣F2=∣S∣F2−∣PSP∣F2\\|S-PSP\\|_F^2 = \\|S\\|_F^2 - \\|PSP\\|_F^2, the best shared subspace is the one that retains the most energy.

Table R3. Relative LE reconstruction error and subspace floors.

Datasetp2p_2floor, best shared subspacefloor, trained M2M_2achieved decoder error
CNP280.2470.2720.483
COBRE280.2290.2510.348
ADHD-200160.2770.3440.497
ABIDE-I360.3830.3830.402
TCP280.3000.3750.408
CAT-D400.2990.3440.573

Table R3 shows that a single rank-p2p_2 subspace shared by all subjects and windows retains 62-77% of the total log-domain energy. The trained M2M_2 is within 0.00-0.08 of the best achievable floor. Much of the remaining error is not caused by the constraint. The gap between the achieved error and the trained floor (0.02-0.23) arises within the subspace, from compressing each trajectory via matrix congruence. This compression is however intended: the latent embedding is meant to retain cohort-level structure rather than every window-specific detail.

Relaxing the constraint improves reconstruction, not clustering. Table R4 varies the decoder width p2p_2 on COBRE.

Table R4. Subspace floor, reconstruction error, clustering score and cost against p2p_2 (COBRE).

p2p_2floor, trained M2M_2achieved decoder errorAMItrainable parameterss / epoch
200.3090.4380.13611.3 M7.3
400.1990.3870.14119.6 M10.3
600.1540.3230.14454.6 M24.4
800.1180.2750.134147.9 M67.0

Increasing p2p_2 from 20 to 80 lowers the subspace floor by 62% and the achieved error by 37%, but ARI remains within 0.134-0.144, with no consistent trend. Meanwhile, the parameter count grows 13-fold and training time 9-fold. The shared-subspace constraint therefore limits reconstruction but not the clustering objective, and a moderate p2p_2 gives the best trade-off between accuracy and cost.

AuthorsOpenReview ↗

Response to Question 4

(1) Compare with parameter-matched FAEclust. Table R5 runs FAEclust on vl(log⁡(X(t)))\text{vl}(\log (X(t))) on all six cohorts with default widths and with widths matched to MatFAE's parameter count within 5%, under the same training objective. Final partitions are obtained by k-means, over the same 20 seeds as MatFAE.

Table R5. FAEclust on vl(log⁡(X(t)))\text{vl}(\log (X(t))) versus MatFAE; cells give ARI at known K, mean ± sd over 20 seeds.

CohortFAEclust-defaultFAEclust-matchedMatFAE
CNP0.034 ±\pm 0.0100.058 ±\pm 0.0170.124 ±\pm 0.019
COBRE0.046 ±\pm 0.0240.053 ±\pm 0.0100.164 ±\pm 0.012
ADHD-2000.040 ±\pm 0.0080.074 ±\pm 0.0120.103 ±\pm 0.021
ABIDE-I0.009 ±\pm 0.0070.015 ±\pm 0.0050.039 ±\pm 0.010
TCP0.020 ±\pm 0.0180.026 ±\pm 0.0160.055 ±\pm 0.018
CAT-D0.016 ±\pm 0.0070.034 ±\pm 0.0170.079 ±\pm 0.011

MatFAE outperforms FAEclust variants on every cohort, achieving 1.4-3.1 times the ARI of the best variant even when FAEclust's parameter count is matched to MatFAE's, so the gain comes from MatFAE's geometry-aware architecture rather than from model capacity.

(2) Clarification on Table 5 and hyperparameter selection. All classification results, on both the synthetic and the real data, use stratified 5-fold cross-validation repeated three times (15 outer folds); "five-fold" in the Table 5 caption and Appendix G.2 is shorthand. Early stopping used an internal 80/20 split of each training fold (Appendix G.2). The test folds were excluded from hyperparameter selection.

(3) Significance on real datasets. Table R6 compares MatFAE with the strongest clustering baseline on each cohort over the same 20 seeds. The columns give each method’s ARI with diagnosis and the paired difference, tested with paired t and Wilcoxon tests (Holm-corrected across cohorts). Pairwise ARI measures the reproducibility of the clusters: the mean ARI over all 20×19/2 = 190 pairs of seed partitions. The seed-level CI is the 95% interval of the mean paired difference over the 20 seeds. The subject-level CI is a paired percentile bootstrap over subjects (2000 resamples, without retraining). MatFAE outperforms the strongest baseline on all six cohorts, and all seed-level intervals exclude zero. The subject-level intervals are wider and exclude zero only on CNP and ABIDE-I.

Table R6. Paired comparison with the strongest clustering baseline per cohort.

CohortMatFAE ARI (mean ± sd, 20 seeds)strongest baseline ARI (mean ± sd)methodpaired differenceWilcoxon p (Holm)paired-t (Holm p)MatFAE pairwise ARI (mean ± sd)strongest baseline pairwise ARI (mean ± sd)seed-level 95% CI of the ARI differencesubject-level 95% CI of the ARI difference
CNP0.124 ± 0.0190.058 ± 0.017FAEclust-matched0.066 ± 0.023<0.00112.743 (<0.001)0.457 ± 0.0460.210 ± 0.027[0.056, 0.076][0.010, 0.125]
COBRE0.164 ± 0.0120.085 ± 0.012TSRVF0.079 ± 0.018<0.00119.298 (<0.001)0.543 ± 0.1340.155 ± 0.049[0.071, 0.087][-0.037, 0.208]
ADHD-2000.103 ± 0.0210.078 ± 0.024PDM0.025 ± 0.021<0.0015.324 (<0.001)0.398 ± 0.0760.275 ± 0.062[0.016, 0.034][-0.071, 0.127]
ABIDE-I0.039 ± 0.0100.015 ± 0.005FAEclust-matched0.024 ± 0.010<0.00110.402 (<0.001)0.039 ± 0.0110.025 ± 0.010[0.020, 0.028][0.010, 0.038]
TCP0.055 ± 0.0180.038 ± 0.014TSRVF0.017 ± 0.0260.0152.981 (0.008)0.057 ± 0.0280.021 ± 0.016[0.007, 0.029][-0.017, 0.052]
CAT-D0.079 ± 0.0110.055 ± 0.011PGA0.024 ± 0.014<0.0017.684 (<0.001)0.354 ± 0.0900.146 ± 0.071[0.018, 0.030][-0.057, 0.112]

Table R7 compares MatFAE-C with the best classification baseline on each cohort, using balanced accuracy (mean ± sd over the 15 outer folds of 5×3 stratified cross-validation).

Table R7. Paired comparison with the best classification baseline per cohort.

CohortMatFAE-Cbest baselinemethodpaired differencepaired-t (p), 15 foldsBayesian correlated t: P(MatFAE-C better)
CNP (0.25)0.453 ±\pm 0.0480.338 ±\pm 0.032BW-norm0.115 ±\pm 0.0577.857 (<0.001)0.9986
COBRE (0.5)0.741 ±\pm 0.0470.638 ±\pm 0.036STAGIN0.103 ±\pm 0.0646.238 (<0.001)0.9937
ADHD-200 (0.5)0.630 ±\pm 0.0720.571 ±\pm 0.041BW-norm0.059 ±\pm 0.0723.184 (0.007)0.9170
ABIDE-I (0.5)0.628 ±\pm 0.0630.587 ±\pm 0.030STAGIN0.041 ±\pm 0.0602.635 (0.019)0.8767
TCP (0.5)0.609 ±\pm 0.0670.505 ±\pm 0.030SPD-SRU0.104 ±\pm 0.0695.816 (<0.001)0.9908
CAT-D (0.5)0.642 ±\pm 0.0560.542 ±\pm 0.033STAGIN0.100 ±\pm 0.0646.047 (<0.001)0.9925

Comments on the paper

1 comment

We thank the reviewers again for their time and constructive feedback. We hope we have clarified the contributions of MatFAE, and that the additional experiments demonstrating SOTA performance on fMRI data address the main concerns raised. We would welcome any further questions or discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.