{"id":"c883de05-f142-4d3b-83a9-8e504f5fb155","arxiv_id":"2505.04412","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The paper integrates a manifold fitting layer with topological and geometric regularizers in an autoencoder, reporting improved manifold reconstruction and dimensionality reduction on noisy point clouds.","lead":"A new autoencoder pipeline adds a manifold reconstruction layer that contracts noisy points toward an estimated latent surface, then applies topological and geometric regularizers during dimensionality reduction. For noisy 3D and high-dimensional point clouds, the authors report better preservation of global shape and local detail than t-SNE, UMAP, and Topological AutoEncoders.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 contradicts the central outperformance claim: across 24 metric-dataset cells the proposed method wins 9, t-SNE wins 10, and Topological AutoEncoder wins 6; the abstract's quantitative validation is therefore not supported by the paper's own data.","rationale":"The reader's weakest assumption concerned the differentiability of the Manifold Reconstruction Layer radii in Section 3.3; that is a real issue for the 'mutual promotion' narrative, and the paper does not describe any relaxation or straight-through gradient. However, even if the radii were fully trainable, the abstract's broader and more central claim of outperforming baselines would still be contradicted by Table 1. The paper's own quantitative results show that the proposed method wins fewer metric cells than t-SNE and only modestly more than Topological AutoEncoder, with zero wins on the Spheres dataset. Since the abstract explicitly invokes quantitative metrics as validation, this is an internal inconsistency in the central claim, not merely a missing derivation. I therefore treat the empirical claim as the single most load-bearing concern. The concern is not about author intent; it is about the conclusion not following from the reported evidence. A revised version could either weaken the claim to 'balances local and global metrics' or provide a pre-specified aggregation under which the method wins; the current text does neither.","tokens_in":21231,"tokens_out":6775,"duration_ms":59922,"concrete_test":"Recompute the win matrix from Table 1 (and optionally Table S1) using the stated metric directions: KL and RMSE lower is better; kNN, Trust, and Spear higher is better. Count, per dataset and in total, how many of the six metric cells each method wins. If Ours does not win a clear plurality under a pre-specified aggregation, the abstract's 'outperforms' claim is not supported. From the published numbers, Ours wins 9/24 cells, t-SNE wins 10/24, and Topological AutoEncoder wins 6/24, so the claim already fails this test without any rerunning.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim, stated in the abstract and repeated in Section 4.4, is that the method 'outperforms baselines like t-SNE, UMAP, and Topological AutoEncoders in discovering manifold structures from noisy data and preserving them through dimensionality reduction, as validated by visualization and quantitative metrics.' Section 4.5 and Table 1 are the quantitative support, and they do not support that claim. Using the standard direction of each metric (KL and RMSE lower is better; kNN, Trust, and Spear higher is better), the win counts from Table 1 are: Swiss roll Ours 3/6, Mammoth Ours 3/6, PartNet Ours 3/6, and Spheres Ours 0/6, for 9 wins out of 24 total metric-dataset cells. By comparison, t-SNE wins 10 cells and Topological AutoEncoder wins 6 cells. No pre-specified aggregation, weighting, or significance procedure is given that would turn this partial record into 'outperforms baselines.' Thus the abstract's quantitative-validation clause is internally contradicted by the paper's own evaluation. This is load-bearing: if the Table 1 numbers stand, the central claim must be weakened, regardless of whether the MRL radii are trainable. The visualizations are qualitative and cannot by themselves support the comparative claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an autoencoder-based manifold learning method that prepends a Manifold Reconstruction Layer (MRL) to the encoder and adds topological and geometric regularizers computed between the reconstructed manifold and the low-dimensional embedding. The MRL contracts noisy points toward an estimated latent manifold following Yao et al. [8]; the topological regularizer is the topological signature loss of Moor et al. [6]; the geometric regularizer is a relaxed distortion measure. The total loss is a weighted combination of an autoencoder reconstruction term, the topological loss, and the geometric loss. The authors claim that the manifold reconstruction and representation components promote each other during end-to-end training, and that experiments on Swiss roll, mammoth, PartNet, and spheres datasets show the method outperforms baselines such as t-SNE, UMAP, and Topological AutoEncoders. The supplement contains ablations, noise robustness experiments, scaling experiments, and a hyperparameter grid search.","tokens_in":21467,"tokens_out":9188,"duration_ms":82578,"significance":"If the central claims were fully supported, the contribution would be useful: an end-to-end model that balances global topology and local geometry in dimensionality reduction while also denoising the input through manifold reconstruction. The paper ships a public code repository, reports metrics averaged over 10 embeddings, and includes ablation and robustness studies, which are strengths. However, the quantitative evidence in Table 1 does not support the abstract's outperformance claim, the trainability of the MRL radii is not established, and the main formula for the geometric regularizer is asserted without derivation. The idea of coupling manifold fitting with representation learning is promising, but the validation and the theoretical presentation need substantial work before the central claims can be accepted.","major_comments":[{"comment":"Table 1 contradicts the abstract's claim that the method 'outperforms baselines like t-SNE, UMAP, and Topological AutoEncoders.' Taking the standard direction of each metric (lower is better for KL and RMSE; higher is better for kNN, Trust, and Spearman), the proposed method wins only 9 of 24 metric-dataset cells; t-SNE wins 10 cells and Topological AutoEncoder wins 6 cells, and on the Spheres dataset the proposed method wins none of the six metrics. No aggregation rule, weighting scheme, or significance test is specified that would convert this partial record into an overall win. This is load-bearing: the abstract and Section 4.4 present quantitative validation as support for the central claim, but the paper's own table does not support that claim.","section":"Section 4.5, Table 1"},{"comment":"The statement that r0, r1, and r2 are 'passed through back-propagation and optimized in the training process' is not supported by the definitions. The weights in Eqs. (4) and (5) contain hard cutoffs such as '0, otherwise,' and the neighbor sets I_x and J_x depend on the radii through ball-membership conditions, so the mapping from the radii to the MRL output is not differentiable with respect to the radii in the standard sense. The paper does not describe any relaxation, smoothing, or straight-through estimator. If the radii are in fact fixed, the claimed mutual promotion between the MRL and the autoencoder loses its stated mechanism; if they are optimized, the optimization must be specified.","section":"Section 3.3, Eqs. (4)-(5)"},{"comment":"The closed form L_geom = D/2 * E[Tr(H^2)] / E[Tr(H)]^2 - D is asserted without derivation. Section 3 explicitly states that most mathematical derivations are omitted, and Definition 2 says the formula follows from choosing a specific convex function h and symmetric function S in Eq. (9), but no such choice is given. Since the geometric regularizer is one of the two core regularizers, the equality needs to be derived or a precise pointer to the reference (e.g., [26] or [10]) with the exact h and S must be stated. As written, the formula cannot be verified from the manuscript.","section":"Section 3.5, Eq. (10)"},{"comment":"The 'mutual promotion' conclusion is not cleanly identified by the ablation. The comparison of 'Final model' versus 'MR AE' in the 'Point Cloud vs Manifold' group shows improvement, but the Final model differs from MR AE by both the topological regularizer and the geometric regularizer in addition to the autoencoder's reconstruction feedback, so the improvement cannot be attributed specifically to 'the representation component (the AutoEncoder).' The additional comparison between 'Manifold vs Embedding' for the Final model and 'Point Cloud vs Embedding' for Topo-geom AE compares different quantities and lacks a controlled baseline; it is suggestive but not conclusive. The paper should either add an ablation that isolates the autoencoder's feedback to the MRL or soften the mutual-promotion claim.","section":"Section 5, Table S2"}],"minor_comments":[{"comment":"The unrelated aphorism about 'clearer answers' and 'clearer writing' interrupts the ablation discussion and should be removed; it is not a scientific statement and is not previously claimed in the paper.","section":"Section 5"},{"comment":"Definition C1 states 'Positivity: |v| ≤ 0,' which should be '|v| ≥ 0.'","section":"Appendix C.1, Definition C1"},{"comment":"The support of a function is written with an equality sign on the right-hand side; the closure should be indicated with an overline or explicitly stated, since the text says it is the closure.","section":"Appendix C.1, Definition C2"},{"comment":"Table 1 contains apparent notation typos, such as '503 × 10^{-1}' for t-SNE RMSE on Swiss roll and '114 × 10^{-1}' for UMAP on the same row, which would be implausible values; these entries should be corrected and the table formatting checked.","section":"Table 1"},{"comment":"The paragraph on scaling to higher-dimensional data says 'the input sample sizes of the Swiss roll dataset' in the context of the Spheres dataset; it should refer to the input dimensions of the Spheres dataset.","section":"Appendix E.3.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is labeled 'Preprint. Work in progress.' and contains an unrelated aphorism in Section 5 that is out of place in a scientific submission; the editor may wish to verify the provenance of the text. The technical components are largely drawn from prior work (Yao et al., Moor et al., Lee et al.), making the integration itself the main novelty, but the parts that are claimed to be new -- especially the end-to-end optimization of the MRL radii -- are exactly the parts that are not adequately supported. I recommend that the revision be evaluated against the requirement that the abstract's quantitative claim be justified by a pre-specified aggregation or by a weaker, accurately qualified claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: the integration is real but the main claim is not backed by the paper's own numbers. The new element is an end-to-end autoencoder that combines a manifold reconstruction layer (Yao et al.), a topological signature loss (Moor et al.), and a relaxed distortion measure (Lee et al.; Vankadara and von Luxburg). That assembly is a legitimate extension of the Topological AutoEncoder program, and the authors deserve credit for a clear pipeline, a thoughtful ablation, and a genuinely useful set of appendices. The mutual-promotion claim—reconstruction improves representation and vice versa—is at least plausible from the ablation, though the comparison is indirect.\n\nThe load-bearing problem is the abstract's quantitative claim. Table 1 gives 24 metric-dataset cells; by the metrics' own directions, the proposed method wins 9, t-SNE wins 10, and Topological AutoEncoder wins 6. No aggregation, weighting, or significance test is specified. So 'outperforms baselines like t-SNE, UMAP, and Topological AutoEncoders' is simply not supported by the evidence in the paper. The visualizations are appealing, but they cannot carry a comparative claim. This is the first thing a referee should push on.\n\nSecond, Section 3.3 says the MRL radii r0, r1, r2 are optimized through back-propagation, but the weighting functions in Eqs. (4) and (5) have hard cutoffs and depend on a discrete neighbor set that changes with the radius. There's no relaxation or straight-through gradient described. The mechanism might be implementable with some care, but as written it isn't demonstrated. Since the mutual-promotion story relies on this optimization path, the authors need to clarify or revise.\n\nThen there are smaller issues. Section 3 explicitly omits most derivations and Eq. (10) is asserted with a pointer to references. Table 1 reports averages over 10 runs but no variance. The Isomap exclusion feels post-hoc (it's relegated to Appendix E.1 with a rationale that's plausible but offered after the fact). And there's a stray sentence at the end of Section 5—'clearer answers are easier to be written down with the same pen'—that looks like accidental inserted text. A copyedit pass would catch it.\n\nWho's this for? People working on topological and geometric regularizers for dimensionality reduction. It's a useful pointer for what combination of pieces works together, and the ablation is informative. I'd send it to a serious referee, not desk-reject: the flaws are fixable and the integration deserves scrutiny. But I would not currently rely on the quantitative results, and the abstract needs major revision.","headline":"A credible integration of existing manifold-learning pieces with attractive visuals, but the abstract's outperformance claim is contradicted by the paper's own Table 1 and the trainable-radius mechanism is under-specified.","tokens_in":22048,"tokens_out":2641,"would_cite":false,"duration_ms":24574,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An autoencoder with a manifold reconstruction layer and topological and geometric regularizers produces low-dimensional embeddings of noisy point clouds that preserve both global shape and local detail, outperforming t-SNE, UMAP, and…","keywords":["manifold learning","dimensionality reduction","autoencoder","persistent homology","topological regularization","geometric regularization","point clouds","manifold reconstruction"],"falsifier":"Freeze the MRL radii $r_0, r_1, r_2$ at their initialization values, retrain on the Swiss roll dataset, and compare reconstruction error and embedding metrics with the version that back-propagates through the radii; if the two are indistinguishable, the claim that trained radii drive the mutual-promotion effect is falsified.","tokens_in":20923,"feed_emoji":"🧩","tokens_out":13356,"duration_ms":110559,"temperature":0.7,"pith_summary":"This paper claims that adding a learned manifold reconstruction step to an autoencoder lets dimensionality reduction preserve both the global shape and the local geometry of noisy point clouds. The central idea is to first contract noisy points toward the latent manifold with a reconstruction layer, then train the encoder with two regularizers: one that matches the persistent-homology signatures of the reconstructed manifold and the embedding, and one that penalizes distortion measured against a scaled isometry. The authors report that this combined system beats standard baselines on noisy Swiss roll, mammoth, 100-dimensional spheres, and object part point clouds, and that the reconstruction and representation components improve each other during training. If the claim is right, a single end-to-end pipeline can denoise, reconstruct, and meaningfully embed noisy Euclidean data without losing the shape that makes the data interpretable.","feed_headline":"Autoencoder cleans noisy point clouds, then preserves shape in 2D","feed_subtitle":"A reconstruction layer removes noise; topology and geometry losses keep the 2D map from tearing or flattening.","key_machinery":"The central object is the Manifold Reconstruction Layer (MRL), a map $F_M$ from the ambient Euclidean space to itself that estimates the latent manifold by local contraction: for each point $x$, it computes a weighted mean direction toward nearby points and repositions $x$ inside a cylinder aligned with that direction, using the smooth weight functions of Eqs. (4) and (5). Its output $Y$ feeds both the autoencoder and the two regularizers. The topological regularizer uses persistent homology, comparing distance entries of $Y$ and $Z$ selected by the persistent pairings of the Vietoris-Rips filtration, via a topological signature loss. The geometric regularizer evaluates the eigenvalues of the pulled-back metric $J^T H J$ relative to the metric on $Y$ and penalizes deviations from a scaled isometry. Together the two regularizers carry the paper's claim that global topology and local geometry can be preserved simultaneously.","core_discovery":"The paper's central claim is that a low-dimensional embedding of a noisy Euclidean point cloud can simultaneously preserve the global topology and the local geometry of the latent manifold, provided the raw points are first contracted toward that manifold by a trainable Manifold Reconstruction Layer. The layer estimates, for each point, the direction toward the closest manifold point by a weighted average of neighbors in a ball of radius $r_0$, then moves the point along that direction using a cylinder-shaped neighborhood with radii $r_1$ and $r_2$; the three radii are claimed to be optimized through back-propagation. The reconstructed cloud $Y$ is then encoded to $Z$, and two regularizers are applied between $Y$ and $Z$: a topological signature loss comparing persistent diagrams of $Y$ and $Z$, and a geometric loss based on a relaxed distortion measure that encourages the encoder to be a scaled isometry. The total loss is a weighted sum of the autoencoder reconstruction loss, $L_{topo}$, and $L_{geom}$. The paper argues that this configuration makes the reconstruction and representation tasks mutually beneficial, and presents experiments on four datasets where its embeddings retain shapes that t-SNE, UMAP, and Topological AutoEncoders fracture or flatten.","pith_inferences":["If the radii do not actually receive gradients because of the hard cutoffs in Eqs. (4) and (5), the reported gain may come from the reconstruction layer acting as a fixed denoiser; freezing the radii in an ablation would separate the two explanations.","The topological regularizer's cost grows with the Vietoris-Rips complex size, so for very large point clouds a practical extension is to compute persistent homology on landmark subsamples and compare signatures on those landmarks.","Because the geometric regularizer enforces approximate scaled isometry, the method may need curvature-aware extensions on strongly curved or high-intrinsic-dimension manifolds; the paper does not test this regime."],"forward_implications":["If the claim is correct, noisy point clouds can be embedded without a separate denoising pass; the reconstruction layer performs denoising and manifold estimation in the same end-to-end training.","The results imply that a topological regularizer alone is not enough to prevent folded, self-overlapping embeddings; the geometric term is needed to preserve local proportions, and the two regularizers are complementary.","The ablation evidence supports the mutual-promotion claim: the reconstructed manifold is cleaner when the autoencoder and regularizers are present, and the embedding improves when the reconstruction layer is present.","The same architecture can in principle be applied to any Euclidean point cloud input, and the supplementary robustness experiments suggest it tolerates a range of noise levels, sample sizes, and input dimensions without changing the training setup."],"supporting_citations":[{"why":"Supplies the two-step manifold fitting procedure (contraction direction then contracted point) that the paper's reconstruction layer implements.","marker":"[8]"},{"why":"Supplies the topological signature loss used as the topological regularizer and serves as the Topological AutoEncoder baseline.","marker":"[6]"},{"why":"Defines persistent homology, the framework from which the persistent diagrams and pairings used in the topological regularizer are computed.","marker":"[9]"},{"why":"Supplies the relaxed distortion measure rooted in scaled isometry that underlies the geometric regularizer.","marker":"[10]"},{"why":"Supplies the geometric autoencoder baseline and the local/global evaluation metrics adopted for the comparisons.","marker":"[25]"},{"why":"Provides the coordinate-free relaxed distortion formulation related to the geometric regularizer's pullback metric.","marker":"[26]"},{"why":"t-SNE is one of the baselines that the paper's method is compared against.","marker":"[4]"},{"why":"UMAP is one of the baselines that the paper's method is compared against.","marker":"[5]"}],"fun_headline_variants":["No tear, no flatten: autoencoder keeps shape from noisy clouds","Denoise, then preserve topology and geometry in 2D","Autoencoder cleans noise, then embeds without tearing or flattening","Reconstruct manifold first, then embed with topological and geometric fidelity","Two regularizers: topology and geometry for faithful low-dim embeddings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the three radius parameters controlling how far the reconstruction layer looks are actually trainable: the weight functions in the layer contain hard cutoffs, and the paper does not say how gradients pass through those cutoffs, so if the radii stay at their initial values the claimed mutual promotion of reconstruction and representation lacks its mechanism.","fun_headline_variants_meta":{"raw":{"variants":["No tear, no flatten: autoencoder keeps shape from noisy clouds","Denoise, then preserve topology and geometry in 2D","Autoencoder cleans noise, then embeds without tearing or flattening","Reconstruct manifold first, then embed with topological and geometric fidelity","Two regularizers: topology and geometry for faithful low-dim embeddings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00099,"raw_usage":{"total_tokens":4207,"prompt_tokens":968,"completion_tokens":3239,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":3151}},"tokens_in":584,"tokens_out":3239,"duration_ms":24810,"temperature":1.0,"reasoning_tokens":3151,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:29:15.433859+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Freeze the MRL radii $r_0, r_1, r_2$ at their initialization values, retrain on the Swiss roll dataset, and compare reconstruction error and embedding metrics with the version that back-propagates through the radii; if the two are indistinguishable, the claim that trained radii drive the mutual-promotion effect is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the two-step manifold fitting procedure (contraction direction then contracted point) that the paper's reconstruction layer implements."},{"cited_title":"Topologicalautoencoders","cited_arxiv_id":null,"evidence_quote":"Supplies the topological signature loss used as the topological regularizer and serves as the Topological AutoEncoder baseline."},{"cited_title":"Persistenthomology—asurvey.Discrete&ComputationalGeom- etry, 453, 01 2008","cited_arxiv_id":null,"evidence_quote":"Defines persistent homology, the framework from which the persistent diagrams and pairings used in the topological regularizer are computed."},{"cited_title":"Measuresofdistortionformachinelearning","cited_arxiv_id":null,"evidence_quote":"Supplies the relaxed distortion measure rooted in scaled isometry that underlies the geometric regularizer."},{"cited_title":"Geometricautoencoders-whatyouseeiswhat you decode","cited_arxiv_id":null,"evidence_quote":"Supplies the geometric autoencoder baseline and the local/global evaluation metrics adopted for the comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the coordinate-free relaxed distortion formulation related to the geometric regularizer's pullback metric."},{"cited_title":"van der Maaten and G","cited_arxiv_id":null,"evidence_quote":"t-SNE is one of the baselines that the paper's method is compared against."},{"cited_title":"McInnes, J","cited_arxiv_id":null,"evidence_quote":"UMAP is one of the baselines that the paper's method is compared against."}],"review_version":1}