acceptodds
Under review as a conference paper at ICLR 2027

Spherical Parametrized Manifold Convolution

Abstract

In geometric deep learning, manifold convolution aims to extend the convolution operator from CNNs to non-Euclidean domains. In this field, equivariance plays an important role as the target property that allows convolution filters to be applied at different locations on the manifold; for CNNs, this corresponds to translation symmetry. In curved spaces such as smooth 2D surfaces, this symmetry group is no longer valid. Therefore, the goal of equivariant convolution theory is to bridge this gap. Current approaches typically either impose restrictive topological or symmetry assumptions to obtain exact equivariance, or abandon global symmetry in favor of local geometric constructions. The former limits applicability to real-world objects with diverse shapes, while the latter must address gauge equivariance and often requires expensive differential-geometric computations. In this work, we propose to combine both worlds: equivariant convolution on general manifolds and the canonical geometry of the sphere, which offers closed-form formulas for computing connections, exponential maps, and geodesics. We use a conformal spherical parametrization, which introduces minimal distortion while allowing to precisely understand the approximation errors. The kernels are polar splines applied across rotational orientation banks, which satisfy equivariance under discrete rotations, while the pullback to the manifold accounts for the local distortion induced by the spherical parametrization. We study the behavior of our method through targeted experiments on a custom synthetic dataset, and show that the convolution is approximately equivariant under rotations. Further, our method maintains competitive performance under deformation, compared to other methods, while offering greater flexibility.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
equivariant convolution, manifold convolution, geometric deep 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_gMN7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

A convolution for genus 0 surfaces is introduced. It combines convolutions defined on the unit sphere with conformal spherical parametrizations. This allows for computing convolutions in a simpler setting (on the sphere as opposed to the general curved surface) using closed from formulas. Moreover, for parametrizations with low conformal distortion, symmetry properties of the convolutions on the sphere approximately transfer to the surface. Experiments on synthetic data indicate robustness and approximate equivariance of the proposed convolutions.

Critical strengths

(S1) The method allows for using closed from formulas for computing convolutions on surfaces at the cost of a one time computation of a spherical parametrization. This may make the method attractive in practice

(S2) The suggested approximate equivariance is open to an error analysis that combines explicit error sources such as the conformal distortion of the parametrization and quadrature errors.

Critical weaknesses

(W1) The proposed framework is restricted to surfaces of genus 0 and a conformal parametrization with low distortion. In many shape collections, not all surfaces satisfy these requirements. This may severely limit the practical applicability of the method. In addition, it may be hard to tell in advance whether or not the method is applicable to the data set, as even if all surfaces are of genus zero and do not have boundary, the conformal distortions may not be small enough.

(W2) The experimental evaluation is limited to synthetic data. In combination with (W1), I find that not enough experimental evaluation demonstrating that the proposed convolution is competitive with alternative convolutions for learning on surfaces is given. Results and comparisons to prior work on benchmarks could help to demonstrate this.

Questions for authors

I am wondering what happens when the conformal distortion of the spherical parametrization is high? In the theory, small conformal distortion is considered. In practice, bounding the conformal distortion for a whole collection of shapes may be difficult. Is there a distortion value at which distortion begins to significantly degrade equivariance or accuracy?

For which data sets, such as FAUST, would the method be applicable and what is the maximum conformal distortion for these data sets?

3 comments

We thank the reviewer for highlighting the closed-form computation and the error analysis. Two issues found since submission limit the submitted claims.

  • Sampler mismatch. Appendix A analysed q(w)=Expp(Apw)q(w)=\mathrm{Exp}_p(A_pw), Ap=(dηp)−1A_p=(d\eta_p)^{-1}, while the code samples πp(w)=η−1(ExpuS2(w))\pi_p(w)=\eta^{-1}\big(\mathrm{Exp}^{S^2}_u(w)\big), u=η(p)u=\eta(p). The revised PDF treats πp\pi_p: the two points differ at second order in the kernel radius ρ\rho (leading term 12∣∇ln⁡s∣ sp−2ρ2\tfrac12|\nabla\ln s|\,s_p^{-2}\rho^2 for a conformal map with factor ss, Lemma A.4), and, on a mesh, the bound is of first order in ρ\rho at vertex queries, because incident faces generally have different Jacobians (Lemma A.7). This term is not controlled by the conformal distortion εp\varepsilon_p, so the bounds do not guarantee small error at our radii.
  • Gauge defect. A per-vertex gauge test found a residual of 0.14 from the second layer, because features were interpolated before transport; transporting first gives about 2⋅10−72\cdot10^{-7}. Group normalisation, which the paper did not describe and the revised PDF adds, keeps a commutator of about 0.03.

W1 / Q1 — High distortion: can one tell in advance?

This is a good question, and we explored it further for this response.

Short answer. Yes, the distortion can be measured before training, and the theory says how it enters. What we cannot give is a validated value at which accuracy or equivariance breaks down.

  • Theory. With the per-face distortion q=σ1/σ2q=\sigma_1/\sigma_2, the ratio of the singular values of the face Jacobian, the best similarity fit in Appendix A gives εp=(q−1)/(q+1)\varepsilon_p=(q-1)/(q+1). The distortion term of (A.18) is then proportional to 2εp/(1−εp)=q−12\varepsilon_p/(1-\varepsilon_p)=q-1: it grows linearly, without a threshold. The sampler term SpS_p does not depend on qq; it grows with ∣∇ln⁡s∣|\nabla\ln s| and 1/sp21/s_p^2, that is, with how strongly and how unevenly the map compresses area (Lemma A.4).
  • Data. qq is cheap to compute, and we report it below for every dataset. In the range we tested, angle distortion did not limit the method: up to qq p99 ≈2.1\approx2.1 on MNIST, pmConv-ResNet accuracy changed by about one point (85.5 to 84.3), while GEM-CNN-ResNet dropped from 91.0 to 65.2. Where pmConv failed (FAUST, fixed stencil), qq was moderate (p99 1.5), but the maps compress area extremely (34–43% of faces near zero area), as the sampler term would lead one to expect. A variant that takes its support from the mesh works on the same maps. We have not shown that the area compression is the cause.
  • Equivariance. We have not measured the equivariance error as a function of distortion.

In more detail:

  1. Diagnostics before training: folded or degenerate spherical faces, and qq. Measured statistics:

    Datasetmapq medianq p99folded facesmeshes with a fold
    MNIST, ARAP onlyCEM1.031.0800 / 1000
    MNIST, noise σ = 2.0CEM1.031.09<0.001%1 / 1000
    MNIST, noise σ = 3.0CEM1.031.12<0.001%8 / 1000
    MNIST, strongest noiseCEM1.092.090.06%955 / 1000
    ADNI (853 accepted subjects)CMCF1.111.3400 / 853
    FAUST (full-precision maps)CMCF1.111.50≤0.034%, but 34–43% of faces near-zero area—

    "q median" and "q p99" are the per-mesh median and 99th percentile over faces, each summarised by its median over meshes.

  2. Distortion vs. error on ADNI (exploratory), on 514 test subjects, within each diagnosis. Errors correlate with distortion for pmConv and for GEM-CNN, which does not use the sphere map, so the association may reflect the cohort. With curvature input, the pmConv−GEM error difference is unrelated to distortion (Spearman ρs=0.02\rho_s=0.02 in HC, 0.040.04 in AD); with coordinate input it is related (+0.25+0.25, −0.28-0.28). No degradation threshold is identified.

  3. Synthetic stress. The MNIST accuracies in the short answer span Table 1 from σ=2.0\sigma=2.0 to 4.54.5. The strongest setting has folds in almost every mesh, so this is evidence outside the injective-map setting.

Q2 — Applicable data sets, e.g. FAUST.

  • FAUST is a failure case. The CMCF sphere maps compress the extremities extremely (34–43% of faces below 10−610^{-6} of the uniform face area); with full-precision maps, 74 of 100 meshes pass a strict recovery test. On vertex correspondence, pmConv with the fixed polar stencil fails (0.33 top-1); a variant that takes its support from the two-ring mesh neighbourhood reaches 0.932, and GEM-CNN 0.994 (settings not matched). The diagnostics reveal the compression before training.
  • ADNI hippocampi: 853 of 928 maps pass the screen (rejected: 6.3% of HC, 11.1% of AD). Angle distortion is low (qq p99 median 1.34), but area distortion is strong.
  • MNIST: near-conformal up to σ=3.0\sigma=3.0 (at most 8 of 1000 meshes have a fold); folds are widespread only at the strongest setting.

W2 — Real data.

We add hippocampus surface classification on ADNI: Alzheimer's disease (AD) vs healthy controls (HC), five splits with disjoint test sets (641 of the 853 subjects), identical training rules for all models, and the pmConv operator as submitted. We use two inputs: unsigned mean curvature (the main input, fixed in advance) and vertex coordinates. The full tables for both inputs go to the appendix of the revised PDF.

  • Ranking, curvature input. On the pilot fold (AUC, 3 seeds), pmConv-ResNet 0.834 ± 0.017 and pmConv 0.833 lead, followed by GEM-CNN 0.817, shallow GEM-CNN 0.812, SplineCNN 0.809, GCN 0.801 and MLP 0.799. Over the four remaining folds, the pmConv variants average 0.818–0.826 and the GEM-CNN variants 0.812–0.818.
  • Locked comparison. We selected pmConv-ResNet vs GEM-CNN on the pilot and locked the comparison before running folds 2–5. The AUC difference averages +0.006 (positive in 3 of 4 folds), and all four DeLong intervals include zero.
  • Coordinate input. SplineCNN, GEM-CNN and pmConv improve, to 0.838–0.856, and pmConv-ResNet (0.852) is within 0.004 of the best (SplineCNN); GCN drops. The coordinate encoding was chosen using pmConv results only.
  • Caveats. The 75 maps that fail the screen are excluded for every model, the folds overlap in training and validation roles, and the intervals are unadjusted.

We read this as competitive performance among the evaluated surface models in the screened cohort; superiority is not established.

Summary. New in this response: diagnostics before training with distortion statistics per dataset, an exploratory error-versus-distortion analysis, FAUST as a failure case, the ADNI benchmark, and a revised Appendix A for the implemented sampler (revised PDF).

Thank you for the detailed answers. The additional experiments and analysis shed more light on the weaknesses. However, the weaknesses remain, and my rating has not changed.

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

The paper proposes pmConv, a convolutional operator for genus-zero surfaces that uses a spherical parametrization to construct sampling neighborhoods and approximate feature transport. It combines polar spline kernels with orientation channels and evaluates the method on projected MNIST classification and synthetic Gaussian orientation regression.

Critical strengths

The improved accuracy under the strongest synthetic deformation (sigma = 4.5) in Table 1 is promising and suggests that the spherical construction may provide some benefits in this setting.

Critical weaknesses
  1. The contribution and intended equivariance are unclear. The paper is dense, and it does not clearly explain the design difference and the clear advantage over existing gauge-equivariant mesh methods. It also moves between local gauge changes, global rotations, and permutations of orientation channels without specifying the precise transformation. The claimed computational advantage is also not supported by runtime comparison.
  2. The operator definition and theoretical analysis are inconsistent. In equation (3.11) x′x' is defined as ExpxM((dηx)−1w)Exp^M_{x}((d\eta_x)^{-1}w). This is generally different from πx(w)\pi_x(w) (the authors themselves pointed that out after equation (3.8)). Nevertheless, equation (3.11) evaluates the feature at one point and describes transport from the other. Moreover, there is a clear mismatch between the implementation and what is proved. In the implementation (described in section 4), the sampling rule is x′=πx(w)x'=\pi_x(w); however, in the appendix A, the proof uses ExpxM((dηx)−1w)Exp^M_{x}((d\eta_x)^{-1}w).
  3. The experimental results are not convincingly showing clear advantage of the proposed method. There is no runtime comparison (as mentioned above in point 1). Also, equivariance error should be reported to demonstrate equivariance is indeed achieved. The evaluation is also restricted to two synthetic surface-signal tasks.
  4. The method is limited to only genus-zero surfaces. It's not clear how to extend to arbitrary mesh topology. Also, what's the cost of computing η\eta for each deformed mesh?
Questions for authors

See the weakness part.

3 comments

We thank the reviewer. Two points are errors, the transport source x′x' in Eq. 3.11 and the sampling rule in Appendix A; both are corrected in the revised PDF (W2). We also add an equivariance test, runtime and a real dataset.

W1 — Which symmetry, and what advantage over gauge-equivariant mesh methods?

  • Intended symmetry. Local SO(2) gauge transformations: an independent rotation of the tangent frame at each vertex. With KK orientation channels, a gauge change by a multiple of 2π/K2\pi/K should shift the output channels accordingly.
  • What Tables 1–2 tested (correction). The "SO(3)" and "C12" columns rotate the mesh and its sphere image rigidly in the ambient space, so they measure ambient invariance (invariance to a rotation of the embedding, GEM-CNN App. F.2.2), not gauge equivariance. pmConv's changes there come from frames built from a fixed axis; co-rotating these frames makes the models invariant to within 5⋅10−35\cdot10^{-3}. We relabel the columns accordingly.
  • Gauge test. As in GEM-CNN, we rotate each vertex frame by its own angle βp∈2π12Z\beta_p\in\frac{2\pi}{12}\mathbb Z, keep geometry and signal fixed, and measure per layer e=∥Φβ(ρin(−β)f)−ρout(−β)Φ0(f)∥/∥Φ0(f)∥e=\|\Phi_\beta(\rho_{\mathrm{in}}(-\beta)f)-\rho_{\mathrm{out}}(-\beta)\Phi_0(f)\|/\|\Phi_0(f)\|, with Φβ\Phi_\beta the layer in the transformed frames. A must-fail control gives e=1.37e=1.37.
  • Result: a defect and a fix. The first convolution passes (e≈10−7e\approx10^{-7}); from the second on, e=0.14e=0.14. With ρ(s)\rho(s) the cyclic channel shift of Eq. 4.4, the submitted code interpolated the corner features of a face σ\sigma and then applied one shift, h=ρ(sσ→x)∑v∈σλvfv,h=\rho(s_{\sigma\to x})\sum_{v\in\sigma}\lambda_v f_v, whereas each corner must first be shifted into the frame at xx: h=∑v∈σλv ρ(sv→x) fv.h=\sum_{v\in\sigma}\lambda_v\,\boldsymbol{\rho(s_{v\to x})}\,f_v . Layer 1 passes because its scalar input is equal across channels. With the fix, e≈2⋅10−7e\approx2\cdot10^{-7}.
  • Normalization. The paper did not describe the normalization layers, an omission. Group normalization (used in Table 1) averages each orientation channel over all vertices and has a commutator of about 0.03; LayerNorm commutes exactly. The revised PDF describes the normalization in the main text and analyses it in the appendix.
  • Retrained with the fix (3 seeds), orientation accuracy does not change. Orientation MAE, fixed vs submitted (both with group normalization): 0.171 vs 0.176 (ResNet) and 0.236 vs 0.225 (plain). On MNIST the fixed models use LayerNorm: the plain model loses 2.7–4.7 points of accuracy and the ResNet changes by −4.6 to +2.5 points; with group normalization (Case 5c, plain) the loss is 2.4 points. The revised PDF reports both rows.
  • Difference from GEM-CNN. GEM-CNN computes the log map and transport on the mesh over the one-ring. pmConv uses the closed-form exponential map and transport of the sphere, pulled back through η\eta, with a prescribed polar stencil not selected by mesh adjacency. Our claimed advantages are this connectivity-independent support and robustness under strong deformation (Table 1, σ=4.5\sigma=4.5: 84.3 vs 65.2 for the ResNets). We do not claim speed.

W2 — Eq. 3.11 and Appendix A.

Both observations are correct.

  1. Eq. 3.11 evaluates the feature at πx(w)=η−1(ExpuS2(w))\pi_x(w)=\eta^{-1}\big(\mathrm{Exp}^{S^2}_u(w)\big), u=η(x)u=\eta(x), but transports from x′=ExpxM((dηx)−1w)x'=\mathrm{Exp}^M_x\big((d\eta_x)^{-1}w\big). The discrete operator and the code use the same point for both; the revised PDF sets x′:=πx(w)x':=\pi_x(w).
  2. Appendix A analyzed q(w)=Expp(Apw) and Ap=(dηp)−1q(w)=\mathrm{Exp}_p(A_pw)\,\text{and}\,A_p=(d\eta_p)^{-1}, instead of the implemented πp\pi_p. The two agree to first order, and Lemma A.4 of the corrected appendix (revised PDF) bounds the gap:

dξ(πp(w),Expp(Apw))d_\xi\big(\pi_p(w),\mathrm{Exp}_p(A_pw)\big)

≤LExp,p(12∥Υp∥∥Apw∥2+16c3,p∣∣w∣∣3),\qquad\le L_{\mathrm{Exp},p}\Big(\tfrac12\parallel\Upsilon_p\parallel \parallel A_pw \parallel^2+\tfrac16 c_{3,p}|| w||^3\Big),

with Υ\Upsilon the difference between the Levi-Civita connections of η∗gS2\eta^*g_{S^2} and ξ\xi. The estimate (A.18) gains a sampler term SpS_p of second order in the kernel radius ρ\rho. It is not controlled by the conformal distortion, so the limit statement now also needs ρ→0\rho\to0; on a mesh the gap is generically of first order at vertex queries (Lemma A.7). Transport through the sphere differs from Levi-Civita transport at first order in ρ\rho for vector features. The measure and the bank formula are also corrected, and we disclose that the code reads the kernel at the mesh-frame chord rather than at ww. The bounds are conditional and guarantee neither small error at our radii nor gauge equivariance.

W3 — Runtime, equivariance error, real data.

  • Runtime (one A100, batch 1, width 16, 800-vertex meshes, the paper's fixed polar stencil):

    ModelParamsInference (ms/graph)Peak GPU memory
    GEM-CNN (shallow)5.8k3.823 MB
    GEM-CNN114k12.231 MB
    pmConv (scalar)5.2k10.554 MB
    pmConv, orientation bank5.3k29.50.47 GB
    pmConv-ResNet, orientation bank25k88.80.47 GB

    The cost comes from materializing kernel values per vertex, support point and orientation channel. We withdraw the "efficient" and "fast inference" wording; the trade-off is deformation robustness, paid for in runtime and memory.

  • Equivariance error: per layer, under W1.

  • Real data. We add Alzheimer's disease vs healthy controls on ADNI hippocampus surfaces (five splits, pmConv operator as submitted), with two inputs. With mean curvature, the main input fixed in advance, the pmConv models rank first (pilot AUC: pmConv-ResNet 0.834 vs GEM-CNN 0.817, other baselines 0.799–0.812). In a comparison selected on the pilot and locked before the remaining four folds, the mean AUC difference to GEM-CNN is +0.006 and all intervals include zero. With vertex coordinates, SplineCNN, GEM-CNN and pmConv improve (0.838–0.856), and pmConv-ResNet is within 0.004 of the best. Maps failing a geometry screen (75 of 928) are excluded for all models, the folds overlap in training and validation roles, and the coordinate encoding was chosen using pmConv results only. We consider this results as competitive; tables presented in the appendix.

W4 — Genus zero and per-mesh cost.

Genus zero is a real limitation; the flat torus and hyperbolic domains are natural candidates for higher genus, left to future work. Each deformed mesh needs one spherical parametrization (0.27 s mean for 3000 vertices, offline) and a SphereMap and geometry cache (74–224 ms for 800–3000 vertices).

I thank the authors for the detailed (and very candid) response. I really appreciate the substantial efforts the authors have made in correcting the mathematical and implementation error and conducting additional experiments. These comments and revision has addressed several of my original concerns. For that reason, I am raising my score.

However, two major concern still remain:

  1. Theoretical justification: the revised analysis acknowledges the difference between the proposed geometric algorithm and its implementation.
  2. Empirical advantage: the robustness under severe deformation is indeed very promising. However, the experiment on Alzheimer's disease show marginal improvment over GEM-CNN, while the runtime and memory costs are substantially larger.

Overall, I think the revised paper has materialized improved, but I am still unconvinced that the paper demonstrates a sufficiently strong theoretical or empirical contribution.

Reviewer_y2nROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2(+1)
Presentation
3(+1)
Soundness
3(+2)
Read review
Paper summary

The paper proposes the method pmConv to perform convolution on a surface M of genus 0. Given a conformal map η:M→S2\eta: M \to S^2 the method defines B-spline kernel on a fixed stencil PP with a fixed gauge in the tangent space of the sphere, using the exponential map the stencil is pushed to the sphere and then using the inverse of the conformal map η\eta the stencil is then pushed to the surface M. P⊂TuS2→S2→MP \subset T_uS^2 \to S^2 \to M In practice, pmConv performs manifold convlution using a proxy-manifold (the sphere S2S^2), where the surface is replaced by a mesh and the conformal η\eta becomes an affine map on triangles of the mesh. In comparison to GEM-CNN, pmConv perform convolution on the sphere instead of on the mesh directly and by that it trades exactness of equivariance to efficient computation by using the closed-form geometry of the sphere and deformation robustness.

Critical strengths
  • The idea is simple and practical because the computations on the sphere have a closed form.

  • The proposed convolution is robust under deformation.

Critical weaknesses
  • Theory does not match implementation. In the implementation the stencil points are first sent to the sphere, by the sphere exponential, then to the M by η−1\eta^{-1}. However in Appendix A (paragraph starting at line: 682) , it seems that another direction is followed in which the stencil is sent to tangent space of the surface first, using the inverse differential dη−1d\eta^{-1}, then to M using the exponential map. These two paths are not equivalent by lines: 226-234.

  • Although the paper cites Maron et. al 2017, however, it does not mention how similar the main ideas in the two papers are. Maron etal 2017 main idea is to compute convolution over a sphere-like surfaces by using the torus as a proxy-manifold. I think this is the closest predecessor of this work and therefore it deserves a relevant discussion of their method in the related work section and a baseline model.

  • Ablation study contradicts the design. line 1146: the model with orientation-bank order 1 performs better that the base line model which has orientation-bank order 12 for SO(3) and C12C_{12} which according the the design principle adding orientation bank should improve equivariance.

Questions for authors
  • Would the analysis in Appendix A hold for the implemented sampling?

  • How do you explain the ablation study results especially Q3a line:1146?

  • How is your method different from Maron etal 2017; Is it more efficient, easier to implement for deformed surfaces, or does it preserve equivariance better?

4 comments

We thank the reviewer for a careful reading, which found a real error in Appendix A. Q1 is answered here, Q2 and Q3 in the next comments.

W1 / Q1 — Does Appendix A hold for the implemented sampling?

Not as submitted; it holds with one additional term and a weaker limit statement. The revised PDF contains a corrected Appendix A written for the implemented sampler.

  • Mismatch. The code samples πp(w)=η−1(ExpuS2(w))\pi_p(w)=\eta^{-1}\big(\mathrm{Exp}^{S^2}_u(w)\big), u=η(p)u=\eta(p) (Eq. 3.6), whereas Appendix A analysed q(w)=Expp(Apw)q(w)=\mathrm{Exp}_p(A_pw), Ap=(dηp)−1A_p=(d\eta_p)^{-1}. The text after Eq. 3.11 makes the same mix-up for the transport source; the revised PDF sets x′:=πx(w)x':=\pi_x(w), as in the code.
  • Repair. Since ApA_p is the differential of πp\pi_p at w=0w=0, the two samplers agree to first order, and the rotation estimate (A.5)–(A.13) is unchanged. Lemma A.4 of the corrected appendix bounds the second-order gap:

dξ(πp(w),Expp(Apw))≤d_\xi\big(\pi_p(w),\mathrm{Exp}_p(A_pw)\big)\le

$ \qquad L_{\mathrm{Exp},p}\Big(\tfrac12|\Upsilon_p|,|A_pw|^2+\tfrac16,c_{3,p}|w|^3\Big),$

where Υ\Upsilon is the difference between the Levi-Civita connections of η∗gS2\eta^*g_{S^2} and ξ\xi; for a conformal η\eta with factor ss, ∥Υp∥=∣∇ln⁡s(p)∣\|\Upsilon_p\|=|\nabla\ln s(p)|. The estimate (A.18) gains a sampler term SpS_p of second order in the kernel radius ρ\rho.

  • What we no longer claim. SpS_p is not controlled by the conformal distortion εp\varepsilon_p; it vanishes on an open set only where MM is locally a round sphere (Remark A.6). The limit statement of Prop. A.1 now needs ρ→0\rho\to0, and on a mesh the bound at vertex queries is of first order (Lemma A.7). The bound compares the implemented operator with the intrinsic reference, so it need not vanish at Q=IQ=I. It is conditional: it shows neither small error at our radii nor gauge equivariance of the network.
  • Further corrections. Transport through the sphere differs from Levi-Civita transport at first order in ρ\rho for vector features, about kmax⁡∥Ap∥ ρ ∣∇ln⁡s(p)∣k_{\max}\|A_p\|\,\rho\,|\nabla\ln s(p)| (zero for scalar features). The measure and the bank formula are also corrected. The code reads the kernel at the mesh-frame chord to the sample rather than at ww; this differs from ApwA_pw by a per-vertex normalisation and the gap above.

W3 / Q2 — Why does K=1K=1 beat K=12K=12 in the rotated column?

The completed ablation does not support an improvement with bank order. Investigating why led to two corrections and a new test, described in the order we found them.

  1. Seeds. The bolded K=12K=12 row used one of three seeds. With all three, its rotated MAE is 0.266±0.1160.266\pm0.116, and across K∈{1,4,10,12,32}K\in\{1,4,10,12,32\} the error rises: 0.159 / 0.188 / 0.245 / 0.266 / 0.307. Bolding a one-seed row was a mistake.
  2. The column measures ambient invariance. The rotated evaluation turns each mesh and its sphere image in the ambient space about the blob centre cc, and keeps the label of the unrotated sample. It therefore measures invariance to a rotation of the embedding (what GEM-CNN calls an ambient transformation, App. F.2.2), not equivariance; we mislabelled it. GEM-CNN builds its frames from the mesh itself, so it is unaffected and scores the same in both columns.
  3. Why larger banks lose. Our tangent frames, used for the stencil layout and for transport, are built from a fixed world axis, so they do not turn with the mesh. Co-rotating all of them at evaluation makes every model invariant to within 5⋅10−35\cdot10^{-3}. Co-rotating only the transport frames lowers the error of K=10/12/32K=10/12/32 from 0.231 / 0.242 / 0.284 to 0.097 / 0.109 / 0.091 (1−cos⁡1-\cos, kept label), below K=1K=1 (0.115), which has no bank to transport.
  4. From there to a direct gauge test. In exact theory the choice of frame is a gauge and cannot change the output, so a frame-dependent result means the discretisation breaks gauge equivariance somewhere. We therefore implemented a direct test, as in GEM-CNN: rotate each vertex frame by its own angle, keep geometry and signal fixed, and measure the error layer by layer. It showed that the submitted operator passes in layer 1 but not from layer 2 on (relative residual 0.14), because bank features were interpolated before being transported; transporting first reduces it to about 2⋅10−72\cdot10^{-7}. The cause, the fix, the role of normalisation and the retrained results are in our reply to Reviewer fvXN (W1). Whether this defect explains the KK trend is not established: rounding to 2π/K2\pi/K and the rotation projection of the transport are other candidates.
  5. The test measures invariance because the readout averages over the angles. Our readout averages the bank over its orientation channels, so its output does not change when the mesh turns. If we recomputed the orientation of the anisotropic signal for the rotated mesh, every such model would be at chance (1−cos⁡≈11-\cos\approx1), GEM-CNN included. Hence we test a pooling layer that keeps the orientation information instead of averaging it. Retrained end to end with the corrected operator (ResNet, 3 seeds), this new pooling (readout) gives 1−cos⁡1-\cos 0.0025 at the identity and 0.0025 on meshes rotated about the blob centre: the prediction turns with the mesh.
  6. Control on MNIST. There is no KK trend. K=1K=1 is the weakest at the identity (85.5% vs 87.5–88.7% for K≥4K\ge4 on Case 2; 78.3% vs 82.5–84.2% on Case 5c), and the drop under rotation is not monotone in KK.

In the revised PDF we relabel the column as ambient invariance, add a recomputed-label column, the gauge test and the orientation readout, and complete the K=12K=12 row.

W2 / Q3 — Relation to Maron et al. (2017).

This is the closest predecessor in spirit, and the revised PDF discusses it:

  • Maron et al. cut the surface, build a four-copy cover and map it conformally to the flat torus. pmConv maps to the sphere itself: no cut, cover or triplet choice, and the output lives on the mesh vertices.
  • Their Tutte construction is bijective by construction; our sphere maps need a validity check (75 of 928 ADNI maps fail it) and distort area strongly.
  • The flat torus has a global frame, so ordinary CNNs apply, but it offers no local SO(2) gauge action. The curved sphere needs local frames, transport and the orientation bank.

On the three options in Q3: we have no direct comparison with Maron et al., so we claim none of them.

  • Efficiency. Against GEM-CNN, the orientation-bank models are 2.4–7.3× slower and use about 15× more memory. We remove the efficiency wording.
  • Equivariance. No claim; our own gauge test is described under Q2. Group normalisation, which the paper did not describe, keeps a commutator of about 0.03; the revised PDF describes and analyses it.

Changes. Revised PDF: the corrected Eq. 3.11 and Appendix A, the Q2 changes above, the Maron et al. discussion, runtime with the efficiency claim removed, and the normalisation layers.

Thanks a lot to the authors for putting the effort of carefully addressing my concerns and revising the paper accordingly. Because of this I will raise my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Submission 56198: summary of our author response and revision

Dear Area Chair and Reviewers,

First of all, I want to thank the reviewers for their time, comments, and engagement. They made the rebuttal process very productive for this project and greatly improved the paper.

We summarize our responses to the three reviews of "Spherical Parametrized Manifold Convolution" and the revised PDF. We answered each reviewer to the best of our ability within the rebuttal period.

What we conceded and corrected

  • Two genuine errors: the transport source in Eq. 3.11, and Appendix A analyzing a sampler that differs from the implemented one. Both are corrected. The corrected appendix has a second-order sampler term, and its limit statement now also requires the kernel radius to go to zero.
  • The rotated column of the orientation task measures invariance to a rotation of the embedding about the blob center, not equivariance. This changes how Reviewer y2nR's K=1 versus K=12 observation should be read. We relabeled the columns, completed the K=12 row with all three seeds, and analyzed where the pose dependence enters (Appendix F).
  • We added a direct gauge-equivariance test, as in GEM-CNN. It found a real defect in the submitted operator (bank features were interpolated before transport; the residual was 0.14 from the second convolution on). A covariant interpolation removes it (2e-7). We retrained with it and report both variants.
  • We removed our efficiency wording: the orientation-bank models are 2.4 to 7.3 times slower and use about 15 times more memory than GEM-CNN.

New evidence

  • Alzheimer's disease versus healthy controls on ADNI hippocampus surfaces (five splits, same training rules for all models). pmConv is competitive with GEM-CNN: the mean AUC difference in a pre-specified comparison is +0.006, and all intervals include zero. We do not claim superiority.
  • A readout that keeps the orientation information reaches a 1-cos of 0.0025 at the identity and 0.0025 under rotation about the blob center.

Limits stated in the paper

  • The equivariance analysis covers one task family and rotations about the blob center. We have no comparison with Maron et al. (2017). FAUST was tested but is outside the scope of the paper; collapsed triangles in its sphere maps appear to hinder our model.

The revised PDF has no change marks. The changes are in Sec. 3 (Eq. 3.11), Sec. 4, and Appendices A and E to J. We thank the reviewers and you for your time.

Sincerely, The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.