acceptodds
Withdrawn from ICLR 2027

TopologyUnderstandingNet (TUN): Geometry-Conditioned Classification of Persistence Diagram Points

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Persistence diagrams (PDs) summarize topological changes across a filtration, but deciding which individual entries reflect source structure rather than sampling artifacts remains difficult because persistence alone does not encode the geometry that generated the diagram. We formulate this problem as point-wise significance classification and construct a supervised benchmark spanning planar shapes, CAD models, triply periodic minimal surfaces, and zeolite frameworks. Its labels combine known topological counts with instance-specific geometric review, supported by a conditional analysis of an early-birth criterion. We then propose TopologyUnderstandingNet (TUN), which models interactions among diagram entries and conditionally refines each prediction using a point-cloud representation of the source geometry. Experiments show that TUN achieves the highest average classification performance and improves over its matched PD-only counterpart across all four categories. In aggregate CAD source-identity retrieval, equal-size subdiagrams selected by TUN also outperform persistence ranking, indicating that geometry-conditioned predictions preserve source information not captured by persistence alone.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Topological Data Analysis, Persistent Homology, Persistence Diagram, Topology Understanding

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_yV6hOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a supervised way for understanding Persistence Diagram (PD), i.e. separating noise and true topologiccal features. This is important for inferring the underlaying structure of objects.The authors propose Topology Understanding Net (TUN), which uses deep feature of each PD point for an intial label. And it lates uses deep point cloud geometric feature for correction. The corrected $s_{base}} are then used to identify PD point label. TUN is then shown to outperform old unsupervsied 2-meana, confidence set methods and supervised baselines.

Critical strengths
  1. Unlike traditional confidence set method, it is novel that this paper propose a \textbf{supervised} way to idenitify topological features and nosie points in persistence diagram. And the proposed method TUN use geometric feature in addition to topological feature for correction.
Critical weaknesses
  1. The intuition of the ambiguity of point labeling is questionable. The ambiguity of point labeling shown in Figure 2 can be avoided by simply useing Fermat distance [1] to build Persistence Diagram instead of plain alpha complex, even with the addition of outliers. For figure 1, supervised filtration leanring [2] can be used to find the correct topology. If the noise point can be eliminated with suitable filtration choices, then what is the point of study the persistence diagram from alpha filtration?

  2. A minor issue: insufficient demonstration of the ground truth labels. The benchmark labels are initialized by selecting the top-β1\beta_1 persistence entrioes and then manually revised by one author. The criteria for these revisionsa are not sufficiantly specificed. No systematic geometric validation is reported. This weakens the benchmark contribution.

  3. The contribution of geometry conditioning is limited. The PD-encoder only already achieces near-perfect performance in Table 1. And both deepsets and set transformer, i.e., pure geometric and non-topological baselinses, also have near-perfect performance. The performance gain from geometry conditioning is very limited.

[1] Fernández X, Borghini E, Mindlin G, et al. Intrinsic persistent homology via density-based metric learning[J]. Journal of Machine Learning Research, 2023, 24(75): 1-42.

[2] Nishikawa N, Ike Y, Yamanishi K. Adaptive topological feature via persistent homology: Filtration learning for point clouds[J]. Advances in Neural Information Processing Systems, 2023, 36: 9131-9143.

Questions for authors
  1. What motivates the specific form of the Conditional Point-Cloud Correction, si,1=si,1base+qiri−ts_{i,1}=s^{\mathrm{base}}_{i,1}+q_i r_i-t?
Reviewer_F6MdOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses binary classification of individual retained entries (at most 100) in H1 alpha-filtration persistence diagrams as “significant” or “noise”. The labels are defined operationally: the top-β1 entries are first selected by persistence, after which a single author applies an “early-birth” review that replaces some late-born entries with earlier-born ones. This review changes 25% of diagrams (Table 3). The proposed benchmark contains 4,990 persistence-diagram–point-cloud–label records spanning planar shapes, CAD models from ABC/Thingi10K, TPMS, and zeolites. The method uses a set-attention persistence-diagram encoder to produce base logits, while a global PointNet-style descriptor of the source point cloud adjusts only the positive logit via a gated residual and a sample-specific threshold.

Critical strengths

A key strength is the transparency of the evaluation. The authors report results over ten seeds, include a matched PD-only ablation and full confusion counts, provide per-corruption retrieval results that also disclose cases where TUN performs worse, and audit the labeling process in Table 3 to quantify deviations from a pure persistence ranking. The paper also raises a relevant question for applied topological data analysis: whether information about source geometry can resolve ambiguity at the level of individual persistence-diagram entries that persistence ranking alone cannot. The multi-domain benchmark could become a useful community resource, provided that the labels and the manual-review protocol are released.

Critical weaknesses

My first major concern is that the reported improvement has not been shown to originate from geometry. The labels are largely determined by persistence-diagram coordinates, β1, and the early-birth decisions of a single annotator. PD-only set models (DeepSets, Set Transformer, PD-Encoder only) already reach mean F1 of 0.9926–0.9933, leaving less than one percentage point of room for improvement. Moreover, the point-cloud branch produces a single global, RMS-normalized descriptor from 4,096 points. It does not link an individual diagram entry to any particular region of the shape and introduces no absolute scale beyond what is already encoded by the diagram; candidate-specific alignment is explicitly left to future work (l. 468). Relative to the PD-only baseline, TUN sharply decreases false positives (113.7 to 25.0) while increasing false negatives (20.0 to 33.7), including an FN increase in three of four categories (Tables 4–5). This pattern is consistent with a shifted decision threshold. Global count calibration through the point-cloud count head, or simply greater capacity and training time due to the warm start and subsequent epochs, could explain the gains. The current experiments do not isolate these possibilities from the effect of geometry. My second major concern is evaluation validity and the scope of the retrieval claim. The splits are stated to be disjoint only at the observation level (l. 345). If the same underlying shapes or shape families appear across splits, the nearly perfect F1 could partly reflect memorization rather than generalization. The retrieval improvement of approximately 0.0325 mAP is entirely attributable to the two outlier-corruption settings. Also, across the remaining seven settings, top-persistence selection is only slightly better (Table 6).

Questions for authors
  1. Are the train/test splits disjoint at the source-shape level, including meshes used for retrieval? If not, could you report shape-disjoint results?
  2. Can you isolate the contribution of geometry by testing mismatched point clouds and comparing against PD-only models with a tuned threshold and equal training budget?
  3. How accurate is the point-cloud count head alone? How does selecting the predicted number of top-persistence points compare with TUN and a simple early-birth baseline?
  4. How were the early-birth labels validated? Were multiple annotators involved, and will the dataset and annotation protocol be released?
  5. Is TUN trained separately for each category? What is its end-to-end inference time, including PD computation?
Reviewer_KWGEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper casts the interpretation of a one-dimensional alpha persistence diagram as pointwise binary classification: each retained entry is either inherited from the source shape or a sampling artifact. It builds a benchmark of 4,990 point clouds in four families (planar shapes, CAD, TPMS, zeolites) and proposes TUN, a set encoder over diagram entries whose positive logit is corrected by a gated residual and a per-sample threshold computed from a PointNet embedding of the source cloud. Labels are the top-β₁ entries by persistence, revised by one author under an "early-birth" rule motivated by Proposition 3.1. TUN improves F1 over its diagram-only ablation by 0.0015–0.0063 and fixed-budget CAD retrieval mAP from 0.3215 to 0.3540.

Critical strengths
  • The question is a real one: which diagram entries are inherited from the source and which come from sampling is a longstanding problem, and a public labeled benchmark across these four shape families would be useful if its labels were reproducible.
  • The experimental protocol is careful: ten seeds, a matched ablation with paired comparison, confusion counts in the appendix, a shown failure case (Fig. 4), code in the supplement.
  • The architecture is simple and clearly described; Appendix B suffices to reimplement it.
Critical weaknesses
  1. The labels are not reproducible, and the paper's own numbers show they are nearly a function of the diagram. Ground truth is persistence ranking plus manual swaps by one author, with no written rule, no second annotator, and no agreement statistic. Table 3 reports 5,284 swaps changing 25% of diagrams (48% on TPMS). Yet diagram-only models recover the final labels at F1 ≥ 0.9965, so the manual decisions are predictable from (b, d) alone. This undercuts the paper's thesis that the diagram lacks the needed information, and it shows the rule could have been written down and applied algorithmically.

  2. The empirical evidence is weak. The task is near saturation: DeepSets, Set Transformer, and the PD-only ablation are all at F1 ≥ 0.99. TUN's gain over its matched ablation is 0.0015–0.0063 F1, i.e., roughly 15–30 fewer false positives out of 25k–50k entries per category, within one to two standard deviations of the diagram-only set models, and DeepSets is better on CAD. The gain is precision-driven, and the point-cloud branch contains a learned per-sample threshold (Eq. 8), so a comparison at argmax does not separate a modeling gain from a threshold shift. No PR curves, AUPRC, or significance tests accompany the bold entries in Table 1.

  3. "Geometry" in the model is a global vector trained to predict the Betti number. The only geometric input is one PointNet descriptor g_pc broadcast to every entry; it is supervised with its own count loss toward β₁ (§3.3, App. B) and drives the threshold head. There is no per-entry geometric link (no representative cycle, no localization), which the authors acknowledge as future work (§4.4). The ablations needed to separate geometry from count are absent: threshold head only versus residual head only, PD-only with oracle β₁, and a probe of g_pc for β₁.

  4. The theory is standard, applies to one of the four categories, and is never used. Proposition 3.1 is the usual reach-plus-stability argument and is correct. But (a) polyhedral CAD meshes, planar shapes with corners, and zeolite frameworks have reach zero; the weak feature size would suffice, is in the authors' own citation (Chazal, Cohen-Steiner, Lieutier 2009), and bottleneck stability itself requires no reach at all. (b) ε and τ are never estimated and no instance is audited against ε < τ/4; Fig. 4 and the TPMS swap rate indicate the hypothesis fails exactly where the manual review is most active. (c) In the proposition's own regime, a late-born persistent entry is an H₁ class of the offset X^r for r ≥ τ − ε, i.e., geometry of the thickened shape, not a sampling artifact; the "noise" label on such entries (Fig. 1d, f) is a task choice, not a consequence of the theory. Minor: finite β₁ is implied by compactness and positive reach; the coefficient field is stated only for the retrieval set (App. E.2).

  5. The retrieval gain comes from one corruption type. The +0.0325 mAP average over nine corruptions in Table 6 is produced by the two outlier rows (+0.233, +0.096); over the other seven, TUN averages −0.005 against top-persistence at the same budget. "Across all conditions" (L456) and the conclusion overstate this.

  6. Missing comparisons. The statistical literature on the significance of individual diagram points is absent: Bobrowski and Skraba's universal null distribution of death/birth ratios gives a training-free per-point test (the model already consumes log(d/b)); Reani and Bobrowski's cycle registration and the subsampling bands of Chazal et al. (ICML 2015) are the natural references. The deterministic baselines (2-means, CS(0.5)) are straw men without access to β₁; the obvious baseline, top-β₁ by persistence with a predicted count, is not reported.

Questions for authors
  1. State the labeling rule precisely, release the pre-review labels, and report inter-annotator agreement on a subsample. Why was the early-birth rule not applied algorithmically?
  2. For each diagram, is there a persistence gap between the β₁-th and (β₁+1)-th most persistent early-born entries? Report the fraction of the benchmark where such a gap exists; that is the part Proposition 3.1 covers.
  3. Ablate the threshold head against the residual head; report PD-only with oracle β₁; probe g_pc linearly for β₁, sampling density, and category.
  4. Report F1 at validation-optimal thresholds, or AUPRC, for every model in Table 1.
  5. Restate Proposition 3.1 with the weak feature size (or μ-reach), and report the Poisson-disk radius δ, the perturbation bound ν, and sampling density relative to feature size per category.
  6. Table 6: discuss the seven non-outlier corruptions, and compare with a DTM-based filtration for the outlier case.
  7. Compare against the Bobrowski–Skraba per-point test.
  8. Give the category composition of the 3,500-record development pool; how β₁ is obtained for ABC/Thingi10k and for zeolites; and whether diagram coordinates are normalized per instance (the point cloud is RMS-normalized, the diagram is clipped at 100 squared-radius units).

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

Related papers

Discussion (0)

Sign in to comment.