acceptodds
Under review as a conference paper at ICLR 2027

Certifying Robustness via Topological Representations

Abstract

Exploring the shape of data spaces is providing new insights in data analysis and deep learning tasks within a variety of application domains. A common approach in Topological Data Analysis to extract multi-scale intrinsic geometric properties of data is persistent homology. This method enjoys theoretical stability results (i.e. Lipschitz continuity with respect to appropriate metrics), however the significance of this robustness when persistent homology is used in machine learning is underexplored. We propose a neural network architecture that can learn discriminative geometric representations from persistence with a controllable Lipschitz constant. In adversarial learning, this end-to-end stability can be used to certify -robustness for samples in a dataset, which we demonstrate on the ORBIT5K data set representing the orbits of a discrete dynamical system.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Topological Data Analysis, Persistent Homology, Certified Robustness, Gromov-Hausdorff distances

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_ew4aOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper composes a stable-rank vectorization of persistence diagrams with an ℓ∞\ell_\infty-Lipschitz network, yielding a classifier with a known Lipschitz constant and a margin-based robustness certificate (Prop.~2). Experiments on ORBIT5K and a synthetic cycle-counting task show reasonable accuracy and resistance to small bottleneck perturbations.

Critical strengths
  • Addresses a real gap that stability of persistent homology does not automatically survive the learned classifier.
  • Prop.2 is clear and correct; the fixed test-time centering is an ℓ∞\ell_\infty -isometry, so the guarantee holds.
  • Certified lower bounds and attack-based upper bounds on robust accuracy are correctly distinguished.
  • Prop.3 is a neat observation on local sharpness.
Critical weaknesses
  1. The main baseline has a known structural vulnerability, so the comparison is uninformative. C.2 implements ``PersLay'' as an unweighted sum of unconstrained MLP point embeddings. Adding a bar (a,a+δ)(a,a+\delta) changes the sum by ≈φ(a,a)\approx \varphi(a,a), which does not vanish as δ→0\delta\to 0 although the bottleneck distance does. This is well known from prior work. Hofer et al. (2019) and Divol & Lacombe (2021) characterize when sum-aggregation is continuous, and the PersLay paper discusses it and uses a weight function. PersLay's actual configuration also uses a prominence filter and a top-kk aggregation rather than a sum. The reimplementation drops all of this. The collapse therefore demonstrates a property of this aggregator, not an advantage over stable persistence-based learners.

  2. The learnable reparameterization is never evaluated, and with F=idF=\mathrm{id} the representation discards birth times. C.4 fixes F=idF=\mathrm{id}; for p=∞p=\infty the stable-rank vector is then the descending half-lifetimes, zero-padded. Birth enters only through ℓF=F(b)−F(a)\ell_F=F(b)-F(a), so it is discarded entirely. The attack in C.3 adds points that move freely along the diagonal. Table~2 thus partly measures invariance from information loss rather than Lipschitz control. D.1 shows the same mechanism: the added bar is born near zero, so the PersLay sum shifts by an almost constant vector, consistent with the drop to exactly chance (20%20\%) at both radii. An ablation separating representation, Lipschitz head, and margin training is needed, together with a discussion of what lifetimes alone lose.

  3. Comparisons and certificate scales are misaligned. Table 1 mixes published and new numbers: PersLay used Alpha-complex H0+H1H_0{+}H_1 diagrams averaged over 100 random 70/30 splits; SRN uses Vietoris--Rips H1H_1 and ten seeds on one split. The authors' own H1H_1-only VR PersLay (81.681.6) is the like-for-like reference and belongs in Table 1. C.1 scales coordinates by 100100, a certified radius rr corresponds to a Gromov--Hausdorff radius of r/200r/200 in raw units. Conversely, adversarial diagrams need not be realizable by any point cloud, so attack-based numbers in diagram space may understate robustness in the point-cloud threat model.

Questions for authors
  • Does the advantage persist against a diagonal-weighted PersLay, and against SRN's vectorization followed by an unconstrained MLP, under the same attack?
  • What changes when FF is learned, and how much of the result is due to margin training versus the representation?
  • How are diagrams with more than NN points handled under point-addition attacks? Truncating the sorted vector appears to preserve 11-Lipschitzness, but this should be stated.
3 comments

1. Structural vulnerability of the baseline

We have revised the PersLay baseline to use the actual architecture employed in the original PersLay paper for ORBIT5K, rather than the generic DeepSets implementation used in our initial submission. The main differences are the use of top-k aggregation instead of sum aggregation, and learned grid weights. These weights, however, are not based on persistence and do not vanish on the diagonal, unlike the weighting considered in the theoretical stability analysis of PersLay (see also our response to Reviewer 2).

Our revised implementation reproduces the published clean accuracy (87.6% versus 87.7%). However, the change does not substantially improve robustness: at ϵ=10−5\epsilon=10^{-5}, attack-based accuracy increases from 6.4% to only 11.8%. The large robustness gap therefore persists when using the original PersLay architecture.

We agree that the published PersLay architecture is particularly vulnerable to near-diagonal insertions. However, the theoretical stability results for PersLay do not provide the guarantees considered in our paper. There are two important distinctions:

  1. Bottleneck stability. The stability results for weighted sum aggregation in PersLay concern W1W_1, as does the analysis of Divol & Lacombe. Such results do not generally extend to W∞W_\infty with a cardinality-independent Lipschitz constant. In particular, a sum of nonzero point contributions can grow with the number of inserted points, even when their bottleneck distance from the diagonal remains fixed.

  2. End-to-end stability. Stability of the vectorization alone does not guarantee robustness of the full classifier. Subsequent neural network layers may amplify perturbations unless their Lipschitz constants are controlled. PersLay studies conditions for stability of the aggregation layer, but does not impose the end-to-end Lipschitz constraints required for the robustness certificates considered here.

Developing an architecture with controlled end-to-end stability with respect to WpW_p, for arbitrary p∈[1,∞]p\in[1,\infty], is precisely one of the objectives of our work. We use stable ranks as the basis for this construction, while PersLay serves as a representative persistence-based neural architecture for comparison.

Empirically, the diagonal-weighted PersLay variant could still be more robust than the published configuration. Our experiments establish the comparison with the original PersLay architecture, rather than with all possible modifications designed to improve its stability.

2. Learnable reparameterization and ablations

Learnable reparameterization. We agree that the original evaluation should have included the learnable reparameterizations introduced in the methods section. Although certified robustness can already be demonstrated using F=idF=\mathrm{id}, we have now revised the experiments to evaluate the complete architecture.

Specifically, we concatenate the identity representation with four representations using jointly learned, increasing, piecewise-linear reparameterizations. This improves clean accuracy from 77.5% to 82.1%, while maintaining comparable certified robustness (40.4% versus 39.3% at ϵ=1\epsilon=1). The results and implementation details are presented in Tables 1–3 and Appendix C.4.

Information loss versus Lipschitz control. We agree that these are distinct properties. With F=idF=\mathrm{id}, the stable-rank representation depends only on lifetimes and therefore discards the absolute birth times. However, the robustness guarantees do not rely on this particular loss of information.

With learned reparameterizations, the representation depends on F(b)−F(a)F(b)-F(a), which generally depends on both birth and death positions. The resulting representation therefore captures information about where features occur along the filtration scale, while retaining the same Lipschitz guarantees. The revised experiments demonstrate that this additional information improves classification accuracy without substantially compromising certified robustness.

More generally, we consider the possibility of inserting points arbitrarily close to the diagonal a fundamental feature of Wasserstein and bottleneck geometry, rather than an artifact of our particular vectorization. Such insertions can move a persistence diagram only a small distance while taking it outside the region represented by the training data. One of the motivations of our work is precisely that conventional persistence-based neural networks may be highly sensitive to these perturbations, despite the stability of persistent homology itself.

Ablations. We have also added an ablation removing the margin-training term in Appendix C.4.3. We have not included an ablation replacing the Lipschitz layers with unconstrained layers, since the resulting architecture would no longer admit the same margin-based robustness certificates. Its robustness would instead need to be estimated through adversarial attacks, making a direct comparison of certified robustness impossible.

Truncation of stable-rank vectors. The practical implementation uses a fixed number NN of coordinates. If an adversarial insertion produces a diagram with more points, the representation is truncated to the first NN coordinates. This truncation is a coordinate projection and is 1-Lipschitz with respect to the ℓ∞\ell_\infty norm. Consequently, it preserves the stability bound and does not invalidate the robustness certificates.

3. Comparisons and interpretation of robustness

PersLay comparison. As discussed above, our revised PersLay implementation reproduces the published clean accuracy. The remaining differences concern the persistence diagrams: we use Vietoris–Rips rather than alpha-complex filtrations, since the former allow us to apply the Gromov–Hausdorff stability results used in our theoretical framework. Both PersLay and SRN are evaluated on the same Vietoris–Rips H1H_1 diagrams, using the same data split.

The robustness comparison is also conservative with respect to SRN: its reported robust accuracy is a certified lower bound, whereas PersLay's attack-based robust accuracy is an upper bound on its true robust accuracy. We also emphasize that SRN uses only H1H_1 information.

Scale and realizability of perturbations. The ORBIT5K experiment primarily concerns robustness in the metric space of persistence diagrams, equipped with the bottleneck distance. The coordinates of the persistence diagrams are multiplied by 100, as stated in Appendix C.1. Consequently, a certified radius rr in the reported diagram coordinates implies robustness for point-cloud perturbations of Gromov–Hausdorff distance less than r/200r/200 in the original coordinates, by the Vietoris–Rips stability theorem.

We agree that adversarial persistence diagrams found by the attack need not be realizable by perturbations of the original point cloud. Thus, the attack-based robustness results should be understood as measurements in persistence-diagram space, not necessarily as attacks realizable in the original point-cloud space.

To complement this analysis, Appendix D.1 provides an additional experiment in which perturbations are applied directly to point clouds, producing corresponding changes in their persistence diagrams. This illustrates the relevance of the phenomenon beyond adversarial perturbations constructed directly in diagram space.

HI, thanks for going over the questions I raised one by one. The rebuttal had a more detailed explanation but didn't close the questions, so I keep the score. Despite this, the question studied in this work is quite interesting, and with an improved writing, more careful and substantial experiments, and careful comparison, it has the potential to be a very nice work.

Reviewer_2rkbOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes the Stable Rank Network (SRN), which combines stable-rank representations of persistence diagrams with Lipschitz neural-network layers to obtain an end-to-end classifier with an explicit Lipschitz constant and sample-wise robustness certificates. The paper also studies Lipschitz reparameterizations of filtration scales and proves local sharpness of the stable-rank Lipschitz constant. Experiments on ORBIT5K and a synthetic point-cloud example show that the resulting certified model remains practically useful.

Critical strengths
  1. A meaningful end-to-end integration of topological stability and certified learning.

Although the individual ingredients---stable-rank representations, Lipschitz neural networks, and margin-based certification---are established, I view their principled integration into a persistence-based learning pipeline with an explicit end-to-end robustness guarantee as a substantive methodological contribution. The paper identifies a real gap between the stability of persistent homology itself and the potential loss of that stability after a learned downstream map, and addresses this gap in a mathematically coherent way.

  1. The theoretical development is complete and practically non-vacuous.

The stability results compose cleanly into a sample-wise certification guarantee, and I did not identify a material gap in the main theoretical arguments. The experiments further show that the resulting constrained model remains trainable, achieves useful classification accuracy, and yields nontrivial certified radii. For a theory-driven contribution, I consider this sufficient evidence that the construction is not merely formal.

Critical weaknesses
  1. The main comparative robustness result requires clarification regarding the PersLay baseline.

The very large robustness gap in Table 2 is obtained against an H1-only DeepSet-style reimplementation of PersLay, while the attack explicitly allows insertion of persistence points near the diagonal. If the pointwise representation of this baseline does not suppress the contribution of points as they approach the diagonal, then the baseline may be structurally highly sensitive to precisely this type of perturbation. In that case, the magnitude of the reported gap could partly reflect the particular implementation rather than a general robustness advantage over competitive persistence-based neural representations.

I do not view the limited number of datasets, the absence of extensive ablations, or the fact that the experiments use F=idF=\mathrm{id} as critical weaknesses for this theory-driven paper. However, because Table 2 is the main comparative robustness evidence, I think the interpretation of this comparison should be clarified.

Questions for authors
  1. Robustness baseline --- material to my assessment.

In the PersLay robustness experiment, does the pointwise representation suppress the contribution of persistence points as they approach the diagonal? If not, can the authors clarify whether the qualitative robustness gap persists for a persistence-based baseline that is well behaved under near-diagonal insertions, or otherwise qualify the conclusion drawn from Table 2? I do not think additional datasets or a broad new experimental study are necessary; this question is only intended to determine whether the main robustness comparison reflects the proposed stability mechanism rather than a particularly vulnerable baseline.

2 comments

PersLay implementation

We thank the reviewer for pointing out this issue. The original submission used a generic DeepSets architecture rather than the specific PersLay architecture used for ORBIT5K. In the revised manuscript, we have replaced it with a faithful PyTorch implementation of the architecture described in the PersLay paper and provided in the authors' repository. In particular, this includes the learned grid weighting and top-k aggregation rather than simple sum pooling. We also verified that our implementation reproduces the outputs of the original model to within 10−610^{-6} when loaded with the authors' trained weights.

The main remaining differences concern the input persistence diagrams: PersLay uses alpha-complex persistence, while we use Vietoris–Rips persistence to align the experiments with the Gromov–Hausdorff stability results underlying our framework. We also use only H1H_1 diagrams, ensuring that PersLay and SRN receive the same inputs. These differences do not appear to affect standard classification accuracy substantially: our implementation achieves 87.6%, compared with the 87.7% reported in the PersLay paper.

Using the original PersLay architecture improves robustness only marginally compared with our previous DeepSets implementation: for example, attack-based robust accuracy at ϵ=10−5\epsilon=10^{-5} increases from 6.4% to 11.8%. The substantial robustness gap observed in the original submission therefore remains.

Robustness baseline

Near-diagonal sensitivity. The reviewer is correct that the pointwise representation does not explicitly suppress contributions from points approaching the diagonal. This is true both of our original implementation and of the authors' own ORBIT5K architecture, which is now used in Table 2.

In particular, the pointwise representation does not vanish on the diagonal. Although PersLay includes a learned grid weighting, this is not a persistence-dependent weighting. This is consistent with the design choices discussed in the PersLay paper, which does not assume that points close to the diagonal should necessarily be considered less important. The revised Table 2 thus evaluates the published PersLay architecture rather than a simplified implementation.

Comparison with diagonally weighted representations. A natural alternative is the weighted sum representation considered in the PersLay paper's stability analysis,

Φ(D)=∑p∈Dw(p)ϕ(p)\Phi(D)=\sum_{p\in D} w(p)\phi(p),

with a weight such as w(p)=d−bw(p)=d-b, followed by a neural network ρ\rho. Such a construction can be Lipschitz with respect to W1W_1, under suitable assumptions. However, two important differences remain compared with SRN:

  1. Stability with respect to W∞W_\infty. A nontrivial weighted sum does not, in general, provide a uniform Lipschitz bound with respect to the bottleneck distance on diagrams of unrestricted cardinality. Adding nn copies of a point at bottleneck distance ϵ\epsilon from the diagonal costs only ϵ\epsilon in W∞W_\infty, while the change in the summed representation grows linearly with nn whenever the point contribution is nonzero. Diagonal weighting alone therefore does not provide the W∞W_\infty stability required for the Gromov–Hausdorff guarantees considered in our paper.

  2. End-to-end stability. Even when the vectorization is stable, this does not automatically provide a useful robustness certificate for the complete classifier. The downstream network must also have a controlled Lipschitz constant. A small contribution from near-diagonal points can be amplified by subsequent layers, and PersLay does not impose the corresponding constraints on the downstream network. What matters for certification is the Lipschitz constant of the complete pipeline relative to the classification margin, rather than the stability of the vectorization in isolation.

We therefore agree that a diagonally weighted baseline could exhibit better empirical robustness than the published PersLay configuration, but this would not by itself establish the end-to-end W∞W_\infty stability and certified robustness provided by SRN. Our comparison demonstrates the difference between the published PersLay architecture and a pipeline designed to preserve these guarantees, rather than claiming superiority over every possible stable persistence-based architecture.

Reviewer_2rkbOpenReview ↗

Response to Authors' Rebuttal

Thank you for the detailed response and the additional experiments. I appreciate the effort to reproduce the original PersLay architecture and to clarify the role of near-diagonal sensitivity. The updated comparison adequately addresses my main concern, and I find the explanation of the distinction between representation-level stability and end-to-end certified robustness convincing.

I also appreciate the additional experiments with learnable reparameterizations, which further strengthen the empirical support for the proposed framework.

Overall, I am satisfied with the authors' response and have no further major concerns. I maintain my rating of 3 (Weak Accept).

Reviewer_8KgFOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors study adverserial learning and robustness at test-time, particularly \epsilon robustness which is challenging to compute. To this end, they proposes a method entitled Stable Rank Network (SRN). They key contribution involves turning a persistence diagram into a vector using stable ranks. To control the lipschitz constent for analysis, the authors adopt n ℓ∞-distance networks (Zhang et al., 2021). Because each stage has a known Lipschitz constant, a margin boundcan be derived to yield a certified radius for each sample in bottleneck distance. The authors evaluate their proposed approach on ORBIT5K, where SRN reaches about 40% certified accuracy for ε=1.

Critical strengths
  • S1. The problem studied is genuinly interesting. Persistent homology are known to be stable, but what is less well studied are perturbations of persistence diagrams under neural network transformations.
  • S2. The certificate is cheap to compute, costing only one forward pass through the SRN network, yielding results superior to Perslay in H_1.
Critical weaknesses
  • W1. Writing and clarity. In the current form the paper is difficult to digest. Consider adding some additional motivation for the problem setting, and re-organizing the preliminaries and methods section, which currently seem very disjoint. Connecting statements between subsections highlighting why these formulations / concepts are needed for the proposed framework are very helpful. It is not necessarily appropriate to have experimental/introductory material (e.g., experiments on pointcloud perturbations) at the beginning of the methods section. These would better help position and clarify the intent of the work. There are also a number of typos throughout the work (adverserial, "They key", "adopt n", "boundcan", etc). W2. Overly simplified empirical evaluation. The authors limit evaluation to a single benchmark (ORBIT5K) and small synthetic evaluation in the appendix, which is too limited to thoroughly study the use cases of their proposed network. Moreover, using p=inf⁡\inf and F=id in the experiments results in Equations 1-2 simply yielding lifetimes sorted in decreasing order (and potentially zero-padded). Therefore, what the authors are evaluating are merely sorted life-times, passed through l_$\inf$ net, and then calculating a margin certificate. Merely combining these individual components limits the paper's contributions. Moreover, the “learnable representation” (as mentioned in the abstract and Sec. 3.3/3.4) is never learned, leaving a crucial component of the paper untested.
  • W3. Limited theoretical results. While appropriately attributed, most of the results in the paper come from previously known works, e.g. Prop 2 from Tsuzuku et al, Eq (4) a known stable-rank result. Moreover, Prop. 1/4 and Cor. 5 are simple calculations based on the definition of a Lipschitz constant, and Prop. 3 (local Lipschitz constant in this case gives you a global one).
Questions for authors

The paper requires substantial rewriting and validation of the method on additional datasets, as well as comparisons to potentially stronger baselines.

2 comments

W1. Writing and clarity

Typos. We could not locate any of the specific typos mentioned by the reviewer in our manuscript. The examples provided (“adverserial”, “They key”, “adopt n”, “boundcan”) do not appear in the paper. We kindly ask the reviewer to verify that these observations or other mentioned in the review have not inadvertently slipped in from other reviews.

Structure. Section 3 follows a deliberate progression. Section 3.1 introduces bottleneck-based distances between point clouds and motivates their use through an illustrative example (Figure 1, which is not part of the experimental evaluation). Section 3.2 then formulates adversarial attacks in the space of persistence diagrams. Sections 3.3–3.4 introduce the proposed defense, namely the SRN architecture, while Sections 3.5–3.6 establish its theoretical properties, including certified robustness and local Lipschitz continuity. We consider this a natural progression, but recognize that the underlying structure may not have been sufficiently explicit. We are therefore adding a short outline at the beginning of Section 3.

Motivation. The motivation for the work is primarily practical. Stability is arguably the defining theoretical strength of persistent homology, yet its implications for machine learning pipelines can easily be misunderstood. In particular, the stability of persistent homology does not automatically translate into robustness of a downstream neural network. As persistence-based neural architectures become increasingly common, we see a need to clarify what these stability guarantees mean in practice and how they can be preserved throughout a learning pipeline. This also leads to mathematical questions concerning the geometry of bottleneck and Wasserstein spaces, some of which we address in the paper.

W2. Empirical evaluation

Scope of the evaluation. We agree that experiments on additional datasets would be valuable. The purpose of the present evaluation, however, is to demonstrate that the proposed architecture can achieve useful classification performance while providing nontrivial certified robustness guarantees, rather than to conduct an exhaustive empirical benchmark. We consider a broader experimental evaluation an interesting direction for future work.

Baselines. The reviewer suggests comparisons with "potentially stronger baselines", but does not specify which methods are intended. PersLay was selected as a relevant persistence-based neural architecture for comparison. We would appreciate clarification as to which additional baselines the reviewer considers appropriate, in particular whether the suggestion concerns other TDA-based architectures or methods operating on different input representations.

Sorted lifetimes. It is correct that, for p=∞p=\infty and F=idF=\mathrm{id}, the stable-rank vectorization reduces to sorted lifetimes. We do not, however, regard this algorithmic simplicity as a weakness. The purpose of the paper is not to introduce a more elaborate vectorization, but to develop a mathematically grounded architecture that preserves stability throughout the learning pipeline and makes these guarantees useful in adversarial learning.

The individual components may be relatively simple, but their compatibility is essential. In particular, obtaining robustness guarantees with respect to the Gromov–Hausdorff distance requires a vectorization that is Lipschitz with respect to the bottleneck distance. Simply combining an arbitrary persistence vectorization with Lipschitz neural network layers would not, in general, provide the same guarantee, since different vectorizations have different stability properties.

Moreover, the framework extends beyond the specific case of sorted lifetimes: it accommodates other Wasserstein distances and learnable reparameterizations. For p=∞p=\infty, the latter yield sorted transformed lifetimes F(b)−F(a)F(b)-F(a), which can encode information about the position of features along the filtration scale, unlike ordinary lifetimes. We regard the simplicity of these representations, together with their explicit stability guarantees, as a strength of the approach.

Learnable reparameterization. We agree with the reviewer that the original experiments did not adequately evaluate this component of the proposed architecture. Although the main objective of the paper is certified robustness, which can already be demonstrated with the identity reparameterization, the evaluation should also reflect the full architecture described in the methods section.

We have therefore revised the experiments to include four jointly learned reparameterizations alongside the identity. This improves standard classification accuracy from 77.5% to 82.1%, an increase of 4.6 percentage points, while retaining comparable certified robustness at ϵ=1\epsilon=1. The updated results are reported in Tables 1–3, with implementation details and ablations in Appendix C.4.

Dear authors,

Thank you for your response. While I believe the authors are pursuing a worthwhile research question, I do not see the revised manuscript crossing the bar for an accepted ICLR paper. I see that a new revision was uploaded, but the changes were not marked, so I am not able to easily judge what was changed.

However, many of my original comments stand. The paper still seems hastily put together, is inadequately motivated and insufficiently experimentally substantiated, and lacks positioning within the literature, as mentioned in my earlier comments. To name a few examples: the abstract does not clearly state what the paper does, instead opening with vague statements about TDA. The intro suffers from the same problems. Up until and into the methods section, it is not easy for me to understand what the authors are trying to achieve. The method proposed is overly simplified, merely ingesting a sorted persistence diagram as mentioned above. Studying one dataset along with a simulated example is simply not sufficient to convincingly justify their argument. For example, the authors could attempt, in good faith, to define a benchmark for persistence-based classifier certification. Alternatively, numerous other benchmarks and data would need to be explored to make these arguments convincing. For instance, WGTL (AAAI 2025) does node classification on Cora-ML, Citeseer, PubMed, etc. TU graph datasets (MUTAG, COX2, etc.) would be another option, with diagrams built from heat-kernel-signature filtrations, as done by PersLay. Finally, there are many approaches and methods that the authors are not contextualizing: a contributing factor to the poor presentation is the lack of a related works section and numerous missing references/comparisons (see a few below).

The list of typos was included in error and does not refer to your manuscript. Please disregard these, and apologies for the confusion.

Bubenik (2015), “Statistical Topological Data Analysis using Persistence Landscapes”, JMLR. Landscapes are 1-Lipschitz from bottleneck distance to the sup norm. Kim et al. (2020), “PLLay: Efficient Topological Layer based on Persistence Landscapes”, NeurIPS. Carrière et al (2015), “Stable Topological Signatures for Points on 3D Shapes”, Computer Graphics Forum (SGP). Liu, Jia & Gong (2021), “PointGuard: Provably Robust 3D Point Cloud Classification”, CVPR. Zhang et al. (2023), “PointCert: Point Cloud Classification with Deterministic Certified Robustness Guarantees” CVPR. Lorenz et al. (2021), “Robustness Certification for Point Cloud Models” (3DCertify), ICCV. Madry et al. (2018), “Towards Deep Learning Models Resistant to Adversarial Attacks”, ICLR.

Comments on the paper

1 comment

Dear Reviewers,

As the discussion period draws to a close, please read and acknowledge the authors’ rebuttal within the next 24 hours. I also encourage you to engage in the discussion and share any remaining questions or concerns.

Thank you for your time and thoughtful reviews.

Best,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.