- 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 -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 -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
-
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 changes the sum by , which does not vanish as 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- 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.
-
The learnable reparameterization is never evaluated, and with the representation discards birth times. C.4 fixes ; for the stable-rank vector is then the descending half-lifetimes, zero-padded. Birth enters only through , 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 () at both radii. An ablation separating representation, Lipschitz head, and margin training is needed, together with a discussion of what lifetimes alone lose.
-
Comparisons and certificate scales are misaligned. Table 1 mixes published and new numbers: PersLay used Alpha-complex diagrams averaged over 100 random 70/30 splits; SRN uses Vietoris--Rips and ten seeds on one split. The authors' own -only VR PersLay () is the like-for-like reference and belongs in Table 1. C.1 scales coordinates by , a certified radius corresponds to a Gromov--Hausdorff radius of 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 is learned, and how much of the result is due to margin training versus the representation?
- How are diagrams with more than points handled under point-addition attacks? Truncating the sorted vector appears to preserve -Lipschitzness, but this should be stated.
3 comments
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 , 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 ). 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 , 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 , 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 of coordinates. If an adversarial insertion produces a diagram with more points, the representation is truncated to the first coordinates. This truncation is a coordinate projection and is 1-Lipschitz with respect to the 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 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 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 in the reported diagram coordinates implies robustness for point-cloud perturbations of Gromov–Hausdorff distance less than 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.
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, 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:
Bottleneck stability. The stability results for weighted sum aggregation in PersLay concern W1, as does the analysis of Divol & Lacombe. Such results do not generally extend to W∞ 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.
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 Wp, for arbitrary p∈[1,∞], 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.