{"id":"9797aa58-90dc-4bca-b69b-ee08608e93e8","arxiv_id":"2509.07184","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A no-labels pipeline built from known parts (DINOv2, normalization, UMAP/t-SNE, K-means) reports strong clustering and novel-class accuracies on four benchmarks.","lead":"DROWCULA is an unsupervised image-clustering pipeline: DINOv2 embeddings are normalized, reduced with t-SNE or UMAP, then clustered with K-means. The authors report high accuracies on CIFAR-10, CIFAR-100, ImageNet-100, and Tiny ImageNet, including when the number of clusters must be estimated.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"NCD SoTA claim rests on a mismatched protocol: DROWCULA is fully unsupervised at 100% novel while NCD baselines use labels and 50% novel, so Table 4 does not establish superiority on Novel Class Discovery.","rationale":"I read the paper in good faith. The proposed pipeline—DINOv2 embeddings, L2 normalization, UMAP/t-SNE reduction, K-means—is coherent, and the known-k clustering results, modulo missing variance and benchmark tuning, are plausible. However, the central claim has two parts: single-modal clustering SoTA and NCD SoTA. The reader's weakest_assumption focuses on whether semantic structure survives dimensionality reduction. That is a real concern, but I think the more load-bearing issue is the NCD evaluation protocol. The abstract explicitly claims NCD SoTA, and Table 4 is the only evidence. The baseline rows are semi-supervised NCD methods evaluated at 50% novel with labeled known classes; DROWCULA is fully unsupervised at 100% novel. The table's 'All' accuracy is not comparable because the data composition and supervision differ. This is not an internal inconsistency in the method, but it is a correctness risk in the empirical argument: the headline claim is not established by the presented comparison. A matched-protocol experiment would settle it. If DROWCULA wins under the same split, the NCD claim stands; if not, the claim should be weakened. This does not change the reader's conditional verdict, since the necessary fix is additional experiments, not a wholesale rejection of the method. I therefore mark verdict_should_be as UNCHANGED and agreement as partial: I agree with the reader's overall caution, but identify a different primary weak point.","tokens_in":16725,"tokens_out":6864,"duration_ms":85387,"concrete_test":"Re-run DROWCULA on the exact OpenLDN-UDA evaluation splits used in Table 4 (Novel%=50): take the unlabeled pool containing both known and novel test images, apply DROWCULA with no labels, map cluster IDs to classes by Hungarian matching, and report ACC on the novel subset and on all images. Compare directly against the corresponding OpenLDN rows under the identical split and metric. If DROWCULA still beats OpenLDN-UDA on the novel subset, the NCD SoTA claim survives; if not, Table 4 does not support the claim and the abstract's NCD statement must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims new State-of-the-Art results on Novel Class Discovery on CIFAR-10, CIFAR-100, ImageNet-100, and Tiny ImageNet. The only support is Table 4, but the comparison is not same-protocol. Every NCD baseline (FixMatch, DTC, RankStats, UNO, ORCA, OpenLDN) is semi-supervised: it receives labeled images from known classes and is evaluated on a pool with Novel%=50, meaning half the evaluation data are from known classes and half are novel. DROWCULA rows use Novel%=100: no labels, no known classes, and clustering is performed on the entire test set, with only an 'All' accuracy reported. These are different tasks with different class composition, different label access, and different evaluation subsets. The 74%, 30%, and 207% improvement statements in Section 6 are all computed from this mismatched comparison. Even if DROWCULA's clustering numbers are correct and reproducible, the NCD SoTA claim does not follow from Table 4. A fully unsupervised method could be better or worse than a semi-supervised NCD method on the actual NCD task; the table as presented cannot tell. This is a load-bearing flaw in the central claim, independent of whether the dimension-reduction pipeline preserves class structure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DROWCULA, a fully unsupervised clustering pipeline for image data. Images are embedded with a pretrained DINOv2 Vision Transformer, L2-normalized, projected to 2–3 dimensions with UMAP or t-SNE, and clustered with K-means; when the number of clusters is unknown, it is selected by maximizing the Silhouette score, optionally accelerated by Bayesian optimization. Experiments on CIFAR-10, CIFAR-100, ImageNet-100, and Tiny ImageNet report high accuracies (e.g., 99.1 on CIFAR-10, 81.8 on CIFAR-100) and the authors claim new state-of-the-art results for single-modal clustering and Novel Class Discovery, both with known and unknown cluster counts. The paper also contains extensive ablations over distance metrics, dimension-reduction methods, feature extractors, and cluster validity indices.","tokens_in":17136,"tokens_out":7208,"duration_ms":77965,"significance":"If the claims are taken at face value, the main contribution is a simple and effective recipe: nonlinear manifold projection of self-supervised ViT embeddings makes K-means substantially more accurate and makes internal validity indices usable for cluster-count selection. The ablation study is unusually thorough, and the code is promised for reproducibility. However, the headline claim of state-of-the-art Novel Class Discovery is not established because the comparison protocol in Table 4 is mismatched, and the 'single-modal' qualifier for the clustering SOTA is not reflected in the abstract. The core empirical observation about dimension-reduction-improved clustering is credible and useful, but the paper's broader claims need reframing and additional validation.","major_comments":[{"comment":"The NCD comparison is not same-protocol. All NCD baselines (FixMatch, DTC, RankStats, UNO, ORCA, OpenLDN) are semi-supervised: they use labeled known classes and evaluate on a pool with 50% novel data, reporting Known/Novel/All. DROWCULA rows use Novel%=100, no labels, and the 'All' column is accuracy on the entire test set. The relative gains in Section 6 ('74%...207%') are computed by comparing DROWCULA's All to baselines' Novel column, which is not a like-for-like comparison. Therefore the abstract's claim of new SOTA on Novel Class Discovery does not follow from the table. Please either run DROWCULA under the standard NCD protocol, compare with fully unsupervised clustering methods under the same 100%-novel protocol, or remove the NCD SOTA claim and position OWUL as a new setting.","section":"Section 6, Table 4"},{"comment":"The abstract claims 'new State-of-the-Art results on single-modal clustering'; in Section 6 the claim is restricted to 'single-modal vision settings without relying on any language supervision' and excludes VLM-based methods. As printed, Table 3 shows TURTLE (CLIP) at 99.5 on CIFAR-10 and 89.9 on CIFAR-100, both above DROWCULA's 99.1 and 81.8. If the intended claim is 'best among methods using only vision features', that is a different, narrower claim and should be stated in the abstract and conclusion. Otherwise the SOTA claim is false as written.","section":"Abstract and Section 6, Table 3"},{"comment":"All accuracy numbers are point estimates without error bars or multiple seeds, and key hyperparameters appear to be selected using the evaluation benchmarks. Table 12 fixes UMAP n_neighbors=10 and n_components=3, t-SNE perplexity=30, etc.; Figure 12 sweeps n_neighbors and reports accuracy on CIFAR-100, and the text recommends keeping this hyperparameter between 10 and 20. This is a form of circularity for the advertised 'completely remove the labeling process' and 'generalizes without any prior information' claims. Please provide variance over K-means/UMAP/t-SNE runs and a validation protocol that does not use test labels, or explicitly state which hyperparameters are tuned and how.","section":"Section 5, Appendix H.2, Figure 12"},{"comment":"The unknown-k comparisons repeat the same protocol mismatch as Table 4 (DTC/RankStats/ORCA use 50% novel and labels). Moreover, Algorithm 1 requires an upper bound kmax, which is not specified in Table 12; the reported unknown-k results are therefore not fully 'without any prior information.' The evidence that Silhouette selects the correct k is limited to CIFAR-10 and one CIFAR-25 subset (Figures 4 and 5). Please report cluster-count estimation accuracy and the kmax used for all datasets, and compare against fully unsupervised cluster-count baselines.","section":"Algorithm 1, Tables 5/9"}],"minor_comments":[{"comment":"The Silhouette formula is not the standard definition. With d(i) as average intra-cluster distance and c(i) as average distance to the nearest cluster, the standard Silhouette is (c(i)-d(i))/max(c(i),d(i)); the printed expression (max{d,c}-d)/c is incorrect or missing a divide-by-max. This matters because Silhouette is the criterion for cluster-count selection.","section":"Equation (8)"},{"comment":"The definition of RI is unclear. If k and l are the numbers of agreeing and disagreeing pairs, RI is k/(k+l), not (k+l)/C(n,2). Please correct and define the symbols clearly.","section":"Equation (6)"},{"comment":"Autoencoder+Normalization is reported as 99.2 on CIFAR-10 in Table 2, while Section G.2 and Figure 11 emphasize instability and fluctuation. Please specify the normalization and architecture used in Table 2 and reconcile the apparent discrepancy.","section":"Table 2 vs Figure 11"},{"comment":"The dashed 'Max t-SNE' and 'Max UMAP' lines are not explained in the caption. The text in Section 5 mentions this only briefly; please define explicitly how the maximum is taken (over k? over restarts?).","section":"Figure 3"},{"comment":"'85.279.4' is missing a separator; it should read '85.2 / 79.4' for the t-SNE row.","section":"Table 9"}],"recommendation":"major_revision","confidential_remarks":"The paper's core empirical observation (nonlinear dimension reduction of DINOv2 embeddings improves K-means and makes silhouette-based selection usable) is likely sound and the ablations are valuable. The main barrier is the NCD comparison protocol and the overbroad SOTA claims. This is fixable within the scope of a revision: the authors can rerun under a matched protocol or explicitly rename the setting and refrain from NCD SOTA claims. I would encourage the editor to request the promised code and the specification of kmax and tuning procedures. The self-cited survey [26] is not load-bearing and should not be a concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper's clustering recipe is real; the NCD SoTA claim is not. The systematic ablation shows that L2-normalized DINOv2 features, reduced to 2-3 dimensions with UMAP or t-SNE, then clustered with K-means, gives strong performance on these benchmarks. That's a useful empirical finding, with code and detailed tables. The silhouette-based cluster-count selection also works much better after reduction, which is a nice result.\n\nBut the central headline doesn't hold. Table 4 claims new SoTA on Novel Class Discovery, yet every NCD baseline is semi-supervised with Novel=50% and DROWCULA is fully unsupervised with Novel=100%. Different label access and different class composition make those numbers incomparable. The \"74%, 30%, 207%\" improvements are computed from this mismatched setup. A fully unsupervised method might be better or worse than semi-supervised NCD on the actual NCD task; the table cannot tell us. That's a load-bearing flaw, independent of whether the clustering pipeline works.\n\nOther issues are softer. No error bars or variance across seeds, and hyperparameters like UMAP neighbors and kmax are selected on the same benchmarks used for evaluation. That's a common issue, but it means the exact numbers should be treated with caution. The single-modal clustering comparison with TEMI, PRO-DSC, TURTLE, PRCut is fair, but those methods use CLIP or other language-supervised models, so \"SoTA in single-modal vision\" is a narrower claim than the abstract suggests at first read.\n\nThe paper is a workmanlike empirical study, not a conceptual breakthrough. Nothing here is unknown in mechanism, but the recipe is clean and likely reusable. It deserves a serious referee, and I'd want to see the NCD comparison redone with same-protocol baselines—or the claim narrowed to fully unsupervised clustering—before it's publishable. For a practitioner looking for a ready-made clustering baseline, the paper is worth reading despite the overreach.\n\nRecommendation: send to peer review, but expect major revision.","headline":"Viable clustering recipe, but the NCD SoTA claim rests on a mismatched comparison and should be redone.","tokens_in":17549,"tokens_out":2242,"would_cite":true,"duration_ms":25876,"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":"DROWCULA claims fully unsupervised clustering and novel-class discovery, reaching 99.1% accuracy on CIFAR-10 and beating semi-supervised NCD baselines on four benchmarks without any labels.","keywords":["open-world clustering","novel class discovery","fully unsupervised learning","dimension reduction","UMAP","t-SNE","Vision Transformers","cluster count estimation"],"falsifier":"Take a labeled fine-grained visual dataset (classes that look similar to a non-expert), run DROWCULA with the number of clusters hidden, and record k at the maximum Silhouette score on the UMAP-reduced space. If that k does not match the true class count and the corresponding accuracy is far below the reported range, the central cluster-count and accuracy claims are falsified.","tokens_in":16672,"feed_emoji":"🧩","tokens_out":9056,"duration_ms":93907,"temperature":0.7,"pith_summary":"DROWCULA is a fully unsupervised recipe for image clustering and novel-class discovery: take embeddings from a pretrained Vision Transformer (DINOv2-giant), L2-normalize them, compress them to two or three dimensions with t-SNE or UMAP, and run K-means. The paper argues that this pipeline reaches state-of-the-art accuracy on CIFAR-10, CIFAR-100, ImageNet-100, and Tiny ImageNet while also estimating the number of clusters from the data via Silhouette scores, optionally searched by Bayesian optimization. If correct, it shows that the labeled bootstrap used by semi-supervised novel-class discovery is unnecessary: an open-world method with zero labels can outperform methods that are given labels for half the classes.","feed_headline":"No labels, no cluster count: DROWCULA hits 99.1% on CIFAR-10","feed_subtitle":"DINOv2 plus UMAP/t-SNE and K-means beats semi-supervised novel-class discovery on four benchmarks.","key_machinery":"The key machinery is the dimension-reduction pipeline. UMAP and t-SNE are manifold-learning methods that build a neighborhood graph and place points in a low-dimensional space so that nearby points stay nearby; in DROWCULA they are applied after L2 normalization of DINOv2-giant embeddings, before K-means. This does two jobs: it makes Euclidean distance meaningful (high-dimensional embeddings can be deceptive, as a toy 2D-manifold-in-3D illustration in the paper shows) and it makes Silhouette a reliable guide for choosing the number of clusters, because cluster validity computed on the reduced space correlates strongly with external accuracy. The cluster-count step is a sweep over k (or Bayes","core_discovery":"On the paper's own terms, the central discovery is that high-dimensional ViT embeddings are not directly clusterable with internal validity indices, but a normalization-plus-manifold-reduction step changes that. DINOv2-giant's 1536-dimensional vectors, normalized and then embedded in 3D by UMAP (or 2D by t-SNE), yield K-means clusters whose Silhouette score tracks true accuracy almost perfectly (r=0.99 on CIFAR-10) and whose accuracy exceeds prior work. With the true number of clusters known, DROWCULA reports 99.1% accuracy on CIFAR-10 and 81.8% on CIFAR-100; with the number of clusters unknown and chosen by maximizing Silhouette, it reports 95.4% and 80.0% respectively, and it outperforms s","pith_inferences":["The paper's numbers rest on DINOv2's pretrained semantics; on domains far from its training distribution the same pipeline would likely need a domain-adapted feature extractor, so the general claim is 'given a good self-supervised ViT, dimensionality reduction and K-means suffice,' not 'any ViT works.'","Because UMAP's locality depends on n_neighbors (the paper recommends 10-20), the unknown-k estimator could be made adaptive by searching over neighborhood size jointly with k; the paper tests n_neighbors only on CIFAR-100.","A hybrid of t-SNE's robustness and UMAP's speed, such as an ensemble or consensus clustering over both reductions, is a natural next step not explored in the paper."],"forward_implications":["Fully unsupervised novel class discovery becomes possible: DROWCULA operates with zero labels and yet reports 100%-novel accuracy that beats semi-supervised NCD methods that see labels for half the classes on three of four benchmarks.","Unknown-k clustering becomes practical: on the reduced space, Silhouette maxima align with true class counts, so an open-world method can output both the number of classes and the assignment.","Memory and speed improve: compressing 1536-D embeddings to 3D or 2D makes K-means cheaper and more accurate; UMAP offers speed while t-SNE offers slightly more robustness on CIFAR-100 subsets.","Pseudo-labeling becomes available for free: DROWCULA cluster centroids can label the inner 50% of each cluster at 99.3% accuracy on CIFAR-10 and 86.3% on CIFAR-100, giving a label-free bootstrap for semi-supervised pipelines."],"supporting_citations":[{"why":"Supplies the transformer backbone family (ViT) used for feature extraction.","marker":"[15]"},{"why":"Supplies UMAP, the main manifold-reduction method used on normalized embeddings.","marker":"[34]"},{"why":"Supplies t-SNE, the alternative reduction method in DROWCULA t-SNE.","marker":"[51]"},{"why":"Supplies the K-means algorithm applied to the reduced low-dimensional embeddings.","marker":"[2]"},{"why":"Provides the Silhouette cluster validity index used for estimating the number of clusters.","marker":"[3]"},{"why":"Defines the open-world supervised learning baseline (ORCA) compared in the NCD tables.","marker":"[8]"},{"why":"Provides OpenLDN, the strongest semi-supervised NCD baseline that DROWCULA outperforms.","marker":"[40]"},{"why":"Provides the DTC baseline used in known- and unknown-k NCD comparisons.","marker":"[23]"},{"why":"Provides the RankStats baseline used in known- and unknown-k NCD comparisons.","marker":"[24]"},{"why":"Provides TURTLE, the strongest recent clustering baseline, which DROWCULA qualifies by noting it uses VLM/text supervision.","marker":"[19]"}],"fun_headline_variants":["DROWCULA tops CIFAR-10 with no labels, no cluster count","Zero labels, zero cluster numbers: DROWCULA hits 99.1%","Unsupervised DROWCULA outperforms semi-supervised novel class discovery","No label, no cluster count: DROWCULA sets SOTA on four benchmarks","DROWCULA: ViT + UMAP + K-means, 99.1% on CIFAR-10"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that human-meaningful class structure survives both DINOv2's embedding and the compression to 2–3 dimensions by UMAP/t-SNE; the paper's evidence for this is correlation on CIFAR-10 and a 25-class subset.","fun_headline_variants_meta":{"raw":{"variants":["DROWCULA tops CIFAR-10 with no labels, no cluster count","Zero labels, zero cluster numbers: DROWCULA hits 99.1%","Unsupervised DROWCULA outperforms semi-supervised novel class discovery","No label, no cluster count: DROWCULA sets SOTA on four benchmarks","DROWCULA: ViT + UMAP + K-means, 99.1% on CIFAR-10"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001419,"raw_usage":{"total_tokens":5597,"prompt_tokens":806,"completion_tokens":4791,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":4673}},"tokens_in":550,"tokens_out":4791,"duration_ms":39034,"temperature":1.0,"reasoning_tokens":4673,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T22:41:07.477315+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a labeled fine-grained visual dataset (classes that look similar to a non-expert), run DROWCULA with the number of clusters hidden, and record k at the maximum Silhouette score on the UMAP-reduced space. If that k does not match the true class count and the corresponding accuracy is far below the reported range, the central cluster-count and accuracy claims are falsified.","supporting_citations":[{"cited_title":"Dosovitskiy and et al","cited_arxiv_id":null,"evidence_quote":"Supplies the transformer backbone family (ViT) used for feature extraction."},{"cited_title":"Ahmed, R","cited_arxiv_id":null,"evidence_quote":"Supplies the K-means algorithm applied to the reduced low-dimensional embeddings."},{"cited_title":"Arbelaitz, I","cited_arxiv_id":null,"evidence_quote":"Provides the Silhouette cluster validity index used for estimating the number of clusters."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the open-world supervised learning baseline (ORCA) compared in the NCD tables."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides OpenLDN, the strongest semi-supervised NCD baseline that DROWCULA outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DTC baseline used in known- and unknown-k NCD comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the RankStats baseline used in known- and unknown-k NCD comparisons."},{"cited_title":"Let Go of Your Labels with Unsupervised Transfer","cited_arxiv_id":"2406.07236","evidence_quote":"Provides TURTLE, the strongest recent clustering baseline, which DROWCULA qualifies by noting it uses VLM/text supervision."}],"review_version":1}