{"id":"caa57dcc-db21-4e00-891e-c8f6b152b87f","arxiv_id":"2508.04200","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"BootSC is an end-to-end deep spectral clustering method that learns affinity, spectral embedding, and k-means jointly with optimal-transport-derived supervision.","lead":"This paper proposes BootSC, a deep learning model that performs spectral clustering end-to-end by jointly learning the affinity matrix, spectral embeddings, and k-means assignments in one network. It uses optimal transport to generate supervision signals and reports large accuracy gains on image clustering benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"OT-derived self-supervision is circular: jointly learned affinity and assignments can converge to balanced but semantically meaningless partitions; with method/experiments absent, the 16% NMI claim is unverified.","rationale":"I agree with the reader's weakest assumption. The OT-derived supervision is the natural place where the central claim is least secure. A deep clustering method bootstrapping from its own predictions can satisfy its training loss by discovering a balanced but semantically arbitrary partition; the spectral affinity and k-means components can then make that partition self-consistent. Sinkhorn's uniform marginal is a constraint on balance, not on semantics. Because the manuscript supplied to reviewers contains only the abstract, introduction, and references—no equations or experimental tables—the concrete mechanism that would prevent degenerate solutions (e.g., stop-gradient, a queue, augmentation, an auxiliary loss, or the details of the orthogonal re-parameterization) is not available. The reported 16% NMI gain is therefore a claim to be checked against the released code, which is a positive and makes the proposed control experiment practical. The reader's UNVERDICTED status remains appropriate; my concern would move toward REJECT only if the control experiment shows no semantic signal, which is not established.","tokens_in":6968,"tokens_out":5470,"duration_ms":67190,"concrete_test":"Run the released BootSC code (github.com/spdj2271/BootSC) on ImageNet-Dogs with the paper's hyperparameters, recording per-epoch: (i) NMI against ground truth on a held-out subset, (ii) mean entropy of predicted cluster assignments, and (iii) distance between consecutive OT target matrices. Then run a control in which the OT targets are replaced by one fixed random balanced assignment (same marginals, drawn once) while keeping all other components unchanged. If the control matches BootSC's final NMI, or if BootSC's assignment entropy collapses to near-uniform over few clusters while the affinity block-diagonalizes over arbitrary neighbor groups, the OT bootstrap is not semantically grounded; if BootSC clearly beats the control and assignments are diverse, the concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—state-of-the-art clustering and a 16% NMI gain on ImageNet-Dogs—requires that the optimal-transport-derived targets described in the abstract provide a reliable, non-degenerate training signal. But those targets are computed from the model's own mini-batch predictions, and the same network outputs both the affinity matrix and the cluster assignment matrix. With no external labels or pretrained features, the loss can be minimized by any balanced partition: Sinkhorn's equipartition constraint rules out the all-one-cluster solution, but it does not favor a semantically correct grouping. A self-consistent fixed point exists where the affinity is block-diagonal according to an arbitrary but balanced assignment and the spectral-embedding/k-means stage simply re-discovers that assignment. In the available text, the only empirical support is the Fig. 1 caption asserting a diagonal block structure; no loss equations, architecture details, or experimental protocol are present to show how this circularity is broken (e.g., stop-gradient, queue, data augmentation, or an auxiliary loss). Thus the claimed improvement over the runner-up is currently unverified, and the main risk is exactly the collapse/empty-solution scenario the reader flagged.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BootSC, a deep spectral clustering model that aims to jointly learn affinity matrix construction, spectral embedding, and k-means clustering in a single end-to-end network. Supervision is derived from optimal transport (OT) in a bootstrapped manner, and a \"semantically-consistent orthogonal re-parameterization\" is said to improve embedding discrimination. The abstract claims state-of-the-art clustering performance, including a 16% NMI improvement over the runner-up on ImageNet-Dogs. The submitted text, however, contains only the abstract, an introductory section, one figure caption (Fig. 1), and references; the technical method, training objective, architecture details, and experimental results are absent.","tokens_in":7275,"tokens_out":5532,"duration_ms":61108,"significance":"If the claimed method works as described, it would address two recognized limitations of deep spectral clustering: the disjoint pipeline of affinity construction, spectral embedding, and k-means, and the poor scalability of full affinity matrices. The use of OT-based self-supervision is timely and could be a meaningful contribution. The paper also promises publicly available code, which would aid reproducibility. However, because the submitted manuscript lacks all technical substance, the significance cannot be evaluated beyond the plausibility of the idea. The current version functions as an extended abstract, not a reviewable archival paper.","major_comments":[{"comment":"The technical content is missing. The introduction describes Figure 2 and then the text jumps directly to the reference list; there is no model definition, no loss functions, no training algorithm, no architecture description, and no pseudo-code. Consequently, the central claim of joint end-to-end learning of affinity, spectral embedding, and k-means is an assertion rather than a derivable or checkable result. This is a load-bearing omission that prevents any technical evaluation.","section":"Manuscript after Section I"},{"comment":"The claimed 16% NMI improvement over the runner-up on ImageNet-Dogs is unsupported by any table, protocol, baseline definition, metric definition, or hyperparameter setting. The only empirical evidence in the submitted text is the Fig. 1 caption, which qualitatively shows a diagonal block structure on ImageNet-10. No quantitative comparison, ablation, or statistical analysis is provided for any dataset.","section":"Abstract and experimental claims"},{"comment":"The described supervision is computed from the model's own mini-batch predictions, and the same network produces both the affinity matrix and the cluster assignment matrix. This raises a concrete circularity risk: a self-consistent fixed point could consist of arbitrary but balanced assignments, with the spectral-embedding/k-means stage simply rediscovering its own partition. The submitted text does not describe any mechanism (e.g., stop-gradient, target network, queue, data augmentation, or auxiliary loss) that would prevent such degenerate equilibria. This needs to be addressed explicitly with equations and a stability argument.","section":"Section I, OT-derived bootstrapped supervision"},{"comment":"The \"semantically-consistent orthogonal re-parameterization\" is introduced only by name. There is no definition of semantic consistency, no orthogonality constraint, no optimization step, and no algorithm. It is therefore impossible to assess the claim that this technique \"significantly enhances\" discrimination capability.","section":"Section I, orthogonal re-parameterization"},{"comment":"The introduction asserts that mini-batch training enables scalability, but no complexity analysis, memory analysis, or large-scale experiment is provided. Since the method is said to construct an affinity matrix within a mini-batch, the quadratic cost relative to batch size should be discussed and compared with existing full-graph spectral clustering approaches.","section":"Section I, scalability claim"}],"minor_comments":[{"comment":"The header still reads \"JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020\" though the submission is dated 2025; the template header should be updated.","section":"General formatting"},{"comment":"Figure 2 is referenced in the Introduction but is not present in the submitted text. The manuscript also appears to jump from page 1 to the reference list at page 12, suggesting missing pages.","section":"Figures"},{"comment":"The footnote distinguishing \"bootstrap\" from the statistical sense is useful, but the idiomatic sense still requires a precise description of what is being bootstrapped and how the bootstrapping is updated over training.","section":"Footnote 1"},{"comment":"The reference list is extensive, but with the technical sections missing it is not possible to verify that all cited works are relevant or that the related-work discussion is complete.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The submitted file appears to be incomplete: only the abstract, introduction, one figure caption, and references are present. If this is the full submission, it is not suitable for review and might be better handled as a desk reject or a request for the complete manuscript. If the complete paper exists, the authors should be asked to supply it, because the current version provides no technical substance to evaluate. The circularity concern raised by the stress-test is legitimate but cannot be resolved without the missing method details."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you a quick take on arXiv:2508.04200. The abstract and intro describe BootSC, a deep spectral clustering method that jointly learns affinity construction, spectral embedding, and k-means in one network, using optimal-transport-derived targets to bootstrap both the affinity and the assignment matrix, plus an orthogonal re-parameterization step. The idea makes sense: existing deep spectral clustering methods often decouple embedding from clustering and pay quadratic cost on the affinity matrix. The authors correctly identify those weaknesses, and the combination of OT bootstrapping with an orthogonal embedding step is not something I've seen in the papers they cite. Credit where due: they frame the problem well, they cite the relevant literature including their own earlier joint spectral embedding work and the OT assignment methods, and the promise of mini-batch, from-scratch training directly addresses the scalability complaint.\n\nHere's the problem. The text we have stops after the first page of the introduction. No method section, no loss functions, no architecture details, no experiments. So the central claim—state-of-the-art performance, with a 16% NMI improvement on ImageNet-Dogs—is currently an assertion from the abstract. That's not a flaw in the idea, but it is a gap in what we can evaluate.\n\nThe specific concern I'd want answered is the circularity we both see: the network produces both the affinity and the assignments, and the OT targets are computed from its own mini-batch predictions. Sinkhorn's equipartition constraint prevents the trivial all-one-cluster solution, but it doesn't prevent a self-consistent balanced partition that has nothing to do with real classes. In principle, the loss can be minimized by any balanced block-diagonal affinity that the spectral embedding and k-means stages then rediscover. Without seeing stop-gradient tricks, a queue, data augmentation, or an auxiliary loss, I can't tell how they break that degeneracy. The orthogonal re-parameterization may help—it enforces orthogonality of the embeddings, which is a non-trivial constraint—but we don't have the derivation.\n\nAll that said, the paper is not a dismissal. It is a plausible and timely approach to a real problem. If the full manuscript shows a non-degenerate training signal and solid experiments, this could be a useful contribution. A serious referee should look at it, but only with the complete text. For now, my recommendation is to ask the authors for the full version before any decision.","headline":"BootSC attacks a real weakness in deep spectral clustering and the OT-bootstrap idea is plausible, but with only the abstract and intro available we can't verify the 16% NMI claim or the method's soundness.","tokens_in":7703,"tokens_out":2821,"would_cite":false,"duration_ms":32526,"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":"BootSC trains affinity construction, spectral embedding, and k-means in one end-to-end network, using optimal-transport-derived targets as self-supervision.","keywords":["deep clustering","spectral clustering","optimal transport","end-to-end learning","affinity matrix","unsupervised learning","self-supervised clustering","image clustering"],"falsifier":"Train BootSC on a dataset with known ground truth and heavily imbalanced cluster sizes, then measure NMI. Because the Sinkhorn targets tend to force balanced assignments, the model should be pushed toward equal-sized clusters; a clear drop in NMI as imbalance grows would show that the optimal-transport targets, rather than semantic signal, are steering the partition.","tokens_in":6929,"feed_emoji":"🗂️","tokens_out":9272,"duration_ms":105724,"temperature":0.7,"pith_summary":"This paper claims that spectral clustering no longer needs to be split into separately optimized stages or depend on pre-trained features. The proposed BootSC trains a single network to predict pairwise affinities, spectral embeddings, and k-means assignments at once, using optimal-transport-derived target matrices to supervise its own affinity and assignment predictions in each mini-batch. The authors argue that this bootstrapped supervision plus orthogonalization of the embeddings lets the network learn a clustering-specific affinity from raw data without a pre-trained network, while staying scalable because only mini-batches are needed. If correct, the method would make spectral clustering practical for large, high-dimensional image datasets and better at separating nonconvex clusters, with a reported 16% NMI improvement over the runner-up on ImageNet-Dogs.","feed_headline":"One network learns all of spectral clustering, from affinity to k-means","feed_subtitle":"BootSC learns affinity, embedding and k-means together, bootstrapped by optimal transport—a 16% NMI jump on ImageNet-Dogs.","key_machinery":"Optimal-transport bootstrapping: in each mini-batch the model's own predicted affinity and assignment matrices are matched to doubly stochastic target matrices computed with a Sinkhorn optimal-transport solver, and these targets provide the self-supervision that steers training away from meaningless partitions. A second mechanism is a semantically-consistent orthogonal re-parameterization that constrains the embeddings to be orthogonal, playing the role of the eigenvector constraint in classical spectral clustering while remaining end-to-end differentiable.","core_discovery":"On its own terms, the paper's discovery is that a single end-to-end network can learn the three separate pieces of spectral clustering—affinity construction, spectral embedding, and k-means grouping—by bootstrapping itself with optimal transport. In each mini-batch the network emits an affinity matrix and a cluster-assignment matrix; an optimal-transport step converts these raw predictions into cleaner target matrices, and the network is trained to move toward those targets. A semantically-consistent orthogonal re-parameterization keeps the spectral embeddings orthogonal, which sharpens discrimination. The learned affinity matrix develops a clear diagonal block structure as training proceeds","pith_inferences":["If the learned affinity matrix genuinely captures semantic similarity, it could transfer to other tasks such as semi-supervised label propagation or graph-cut segmentation—something the paper does not explore.","Because Sinkhorn-based targets typically favor balanced clusters, applying BootSC unchanged to heavily imbalanced data may force equal-sized partitions; an unbalanced or cost-augmented optimal-transport variant would be a natural testable extension.","The k-means head assumes a fixed cluster number k; a nonparametric clustering head could let the same bootstrap signal also discover the number of clusters, extending the method beyond the paper's setting."],"forward_implications":["If BootSC works as claimed, deep spectral clustering scales to datasets too large for a full affinity matrix, since training uses only mini-batches.","A clustering-specific affinity can be learned from raw data without pre-trained features, removing a major source of representation bias in earlier deep spectral methods.","Jointly optimizing affinity, embedding, and k-means avoids the error accumulation of two-stage pipelines where one stage is frozen while the next is tuned.","Optimal-transport-generated targets can bootstrap more than one output matrix in the same network, suggesting a general recipe for self-supervised clustering.","The reported 16% NMI gain over the runner-up on ImageNet-Dogs indicates the approach is competitive with current deep clustering baselines on challenging image data."],"supporting_citations":[{"why":"Defines the classical spectral clustering pipeline that BootSC re-implements as a differentiable, jointly learned network.","marker":"[12]"},{"why":"Supplies the efficient optimal-transport solver used to compute the bootstrap target matrices from predicted affinities and assignments.","marker":"[22]"},{"why":"Provides the orthogonal Procrustes solution used in the semantically-consistent orthogonal re-parameterization of spectral embeddings.","marker":"[23]"},{"why":"Prior method that jointly learns spectral embedding and k-means; BootSC extends this joint-optimization idea to affinity learning and optimal-transport bootstrapping.","marker":"[31]"},{"why":"Shows how optimal transport can generate cluster pseudo-labels for simultaneous clustering and representation learning, the self-labelling idea BootSC adapts.","marker":"[76]"},{"why":"Demonstrates contrastive learning with cluster assignments produced by optimal transport, supporting the bootstrap-signal mechanism.","marker":"[77]"},{"why":"Transfers optimal-transport assignment into a neural training objective, a pattern BootSC applies to affinity and assignment matrices.","marker":"[78]"}],"fun_headline_variants":["One network, all of spectral clustering: affinity, embedding, k-means","End-to-end spectral clustering: BootSC learns all steps together","Optimal transport bootstraps deep spectral clustering","Deep spectral clustering with self-taught affinity and embeddings","BootSC: one network for affinity, embedding, and k-means"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The whole method relies on the optimal-transport targets generated from the network's own current predictions being informative; if those self-generated targets are degenerate or meaningless, the network can converge to a self-consistent but semantically wrong partition.","fun_headline_variants_meta":{"raw":{"variants":["One network, all of spectral clustering: affinity, embedding, k-means","End-to-end spectral clustering: BootSC learns all steps together","Optimal transport bootstraps deep spectral clustering","Deep spectral clustering with self-taught affinity and embeddings","BootSC: one network for affinity, embedding, and k-means"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1067,"prompt_tokens":674,"completion_tokens":393,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":418,"completion_tokens_details":{"reasoning_tokens":307}},"tokens_in":418,"tokens_out":393,"duration_ms":4272,"temperature":1.0,"reasoning_tokens":307,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:47:23.202213+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train BootSC on a dataset with known ground truth and heavily imbalanced cluster sizes, then measure NMI. Because the Sinkhorn targets tend to force balanced assignments, the model should be pushed toward equal-sized clusters; a clear drop in NMI as imbalance grows would show that the optimal-transport targets, rather than semantic signal, are steering the partition.","supporting_citations":[{"cited_title":"Deep spectral clustering via joint spectral embedding and kmeans,","cited_arxiv_id":null,"evidence_quote":"Prior method that jointly learns spectral embedding and k-means; BootSC extends this joint-optimization idea to affinity learning and optimal-transport bootstrapping."},{"cited_title":"Self-labelling via simultane- ous clustering and representation learning,","cited_arxiv_id":null,"evidence_quote":"Shows how optimal transport can generate cluster pseudo-labels for simultaneous clustering and representation learning, the self-labelling idea BootSC adapts."},{"cited_title":"Unsupervised learning of visual features by contrasting cluster assign- ments,","cited_arxiv_id":null,"evidence_quote":"Demonstrates contrastive learning with cluster assignments produced by optimal transport, supporting the bootstrap-signal mechanism."},{"cited_title":"Ota: Optimal transport as- signment for object detection,","cited_arxiv_id":null,"evidence_quote":"Transfers optimal-transport assignment into a neural training objective, a pattern BootSC applies to affinity and assignment matrices."}],"review_version":1}