Pith. sign in

REVIEW 4 major objections 5 minor 56 references

Dimensionally Reduced Open-World Clustering: DROWCULA

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict Viable clustering recipe, but the NCD SoTA claim rests on a mismatched comparison and should be redone. read the letter →

arxiv 2509.07184 v1 pith:JCR5HVPZ submitted 2025-09-08 cs.CV cs.LG

Erencem Ozbey , Dimitrios I. Diochnos This is my paper · ORCID
classification cs.CVcs.LG
keywords open-worldclusteringnovelclassdiscoveryfullyunsupervisedlearningdimensionreductionUMAPt-SNEVisionTransformersclustercountestimation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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

What would settle it

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.

Watch

Extended reading notes

Core claim

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

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

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.

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 (4)
  1. [Section 6, Table 4] 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.
  2. [Abstract and Section 6, Table 3] 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.
  3. [Section 5, Appendix H.2, Figure 12] 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.
  4. [Algorithm 1, Tables 5/9] 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.
minor comments (5)
  1. [Equation (8)] 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.
  2. [Equation (6)] 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.
  3. [Table 2 vs Figure 11] 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.
  4. [Figure 3] 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?).
  5. [Table 9] '85.279.4' is missing a separator; it should read '85.2 / 79.4' for the t-SNE row.

Circularity Check

2 steps flagged · score 6.0 of 10

DROWCULA's headline SoTA numbers on CIFAR-10 and CIFAR-100 are partially selected on those same benchmarks; the rest of the pipeline is empirical and not circular.

  1. fitted input called prediction [Section 4 (Feature Extractors) and Section 5 (Methodology), Table 2, Figure 3]
    "Since DINOv2-giant performs better than the other feature extractors among all dimension reduction techniques, we choose to present our proposed methodology using this model. ... In light of our preliminary investigations, we construct our base methodology DROWCULA in the following way."

    The preliminary investigation on CIFAR-10 (Tables 1-2, Figure 3) uses ground-truth labels to compare distance measures, dimension-reduction methods, and feature extractors. The best-performing combination on CIFAR-10 is then fixed as DROWCULA. The later SoTA claim on CIFAR-10 is therefore not an independent test of the pipeline: it is the accuracy of a pipeline explicitly chosen to maximize accuracy on that very dataset. The reported number is forced to be no worse than all discarded alternatives on CIFAR-10, so the headline result partly reduces to the model-selection fit rather than being a standalone prediction.

  2. fitted input called prediction [Appendix H.2 (Hyperparameters), Figure 12, Table 12; Table 4 CIFAR-100 row]
    "Our experiments show that having a number of neighbors smaller than 8 performs significantly worse. We recommend keeping this hyperparameter between 10 and 20 to take advantage of local structures in dimension reduction. ... UMAP n_components = 3, n_neighbors = 10, min_dist = 0.1"

    The UMAP hyperparameter n_neighbors is chosen from an accuracy sweep on CIFAR-100 (Figure 12), and DROWCULA-UMAP then reports 80.4% on CIFAR-100 as a state-of-the-art result. Thus the CIFAR-100 accuracy is an in-sample selection score for that hyperparameter, not an out-of-sample prediction. While the overall approach has independent content, this specific SoTA number is partly an artifact of fitting the hyperparameter to the same benchmark on which the claim is made.

full rationale

The paper is primarily empirical: it applies a pretrained DINOv2 model, L2 normalization, UMAP/t-SNE reduction, and K-means, and reports measured clustering accuracies. There is no equation-level derivation that is circular by definition, and the only self-citation ([26], a survey by one of the present authors) is used for background taxonomy, not as load-bearing evidence. The Silhouette-based cluster-count estimation is supported by correlation experiments, not by circular reasoning. The main circularity concern is pattern 2: the pipeline and its hyperparameters are selected using ground-truth accuracy on CIFAR-10 and CIFAR-100, and the same benchmarks are then used to claim new state-of-the-art results. This makes the CIFAR-10 and CIFAR-100 headline numbers partially in-sample rather than independent predictions. The separate Table 4 protocol mismatch (comparing fully unsupervised 100%-novel rows with semi-supervised 50%-novel NCD baselines) is a serious benchmark-validity issue, but it is not a circularity, so it does not affect the circularity score. Overall, the central claim of new SoTA on CIFAR-10 and CIFAR-100 is partially circular due to selection on those exact benchmarks, while the ImageNet-100 and Tiny ImageNet results and the general pipeline retain independent empirical content.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The ledger contains 7 free choices (reduction method, UMAP/t-SNE hyperparameters, K-means init, kmax), 5 domain assumptions, and no invented entities. The absence of invented entities is expected since this is a benchmark systems paper.

free parameters (7)
  • UMAP n_neighbors = 10
    Selected from the sweep in Figure 12 by accuracy on CIFAR-100.
  • UMAP n_components = 3
    Chosen from CIFAR-10 experiments; more components did not help.
  • t-SNE n_components = 2
    Chosen from CIFAR-10 experiments; standard for visualization.
  • t-SNE perplexity = 30
    Default value, not independently justified for each dataset.
  • K-means n_init = 50
    Chosen for stability; no analysis of its effect on reported numbers.
  • Dimension reduction method (UMAP/t-SNE vs PCA/MDS/Isomap/AE) = UMAP/t-SNE
    Method selected on CIFAR-10 accuracy in Table 2, then used for SOTA claims on the same datasets.
  • kmax upper bound for unknown-k search = unspecified
    Algorithm 1 loops from k_min to kmax, but the paper never states how kmax is set for each dataset; this affects the silhouette-based estimate.
assumptions (5)
  • domain assumption DINOv2-giant embeddings separate benchmark classes after dimension reduction
    Section 4 and Section 5 rely on pretrained ViT features being class-discriminative; no proof beyond benchmark results.
  • domain assumption t-SNE/UMAP preserve enough local geometry to improve K-means and Silhouette validity
    Section 4, Table 2 and Figure 4 demonstrate this on CIFAR-10, then the paper extrapolates to all datasets.
  • domain assumption Silhouette score on reduced space is a reliable estimator of true cluster count
    Section 5 and Appendix E.1; shown for CIFAR-10 and one CIFAR-25 subset only.
  • domain assumption Benchmark class labels correspond to clusters in the embedding space
    Evaluation uses ACC/NMI/ARI against human labels; this assumes semantic alignment between embeddings and labels.
  • domain assumption Pretrained DINOv2 weights are treated as fixed inputs from prior literature
    Section 5 uses DINOv2-giant as a fixed feature extractor, inheriting whatever data and supervision went into those weights.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dimensionally Reduced Open-World Clustering: DROWCULA." pith.science (2026). https://pith.science/paper/JCR5HVPZ

@misc{pith2026250907184,
  author       = {Pith},
  title        = {Pith review of: Dimensionally Reduced Open-World Clustering: DROWCULA},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JCR5HVPZ}},
  note         = {Machine review of arXiv:2509.07184}
}
read the original abstract

Working with annotated data is the cornerstone of supervised learning. Nevertheless, providing labels to instances is a task that requires significant human effort. Several critical real-world applications make things more complicated because no matter how many labels may have been identified in a task of interest, it could be the case that examples corresponding to novel classes may appear in the future. Not unsurprisingly, prior work in this, so-called, `open-world' context has focused a lot on semi-supervised approaches. Focusing on image classification, somehow paradoxically, we propose a fully unsupervised approach to the problem of determining the novel categories in a particular dataset. Our approach relies on estimating the number of clusters using Vision Transformers, which utilize attention mechanisms to generate vector embeddings. Furthermore, we incorporate manifold learning techniques to refine these embeddings by exploiting the intrinsic geometry of the data, thereby enhancing the overall image clustering performance. Overall, we establish new State-of-the-Art results on single-modal clustering and Novel Class Discovery on CIFAR-10, CIFAR-100, ImageNet-100, and Tiny ImageNet. We do so, both when the number of clusters is known or unknown ahead of time. The code is available at: https://github.com/DROWCULA/DROWCULA.

Figures

Figures reproduced from arXiv: 2509.07184 by the authors.

Figure 1
Figure 1. Overview of DROWCULA (our method). An image dataset is fed into a feature extraction mechanism and in sequence the dimension of the resulting embeddings is reduced, yielding the low-dimensional instances that we ultimately cluster. Along these lines, we propose a methodology to completely re￾move the labeling process for NCD and approach this problem as arXiv:2509.07184v1 [cs.CV] 8 Sep 2025 [PITH_FULL_IMAGE:figures… view at source ↗
Figure 2
Figure 2. A sample 2D manifold in 3D space. Points that were originally distant in the 2D manifold (e.g., red and blue) end up being close to each other in the 3D space. Illustration created with artificial data. Dimension Reduction. In [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Effect of Dimension Reduction on the Correlation Coefficients between Silhouette Score and Clustering Accuracy on CIFAR-10 Dataset. Additionally to what we present in [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Effect of the UMAP (3D) method on clustering accuracy and silhouette score on a random CIFAR-25 dataset. Acc is marked at the left, and the Sil is marked at the right axis, respectively. Both plots are aligned according to the Acc values, but the Sil is scaled to match…
Figure 6
Figure 6. Figure 6: shows the correlation between CHI and ACC on the CIFAR-10 dataset. The correlation coefficient is -0.69, where there is no dimension reduction technique applied. We observe that Isomap reduction increases this value up to 0.95 ,indicating ACC is highly correlated with …
Figure 7
Figure 7. Figure 7: Comparison of the UMAP (3D) and Isomap (8D) methods for Calinski-Harabasz Index (CHI) on a randomly chosen CIFAR-25 dataset. Acc is marked at the left, and CHI is marked at the right axis, respectively. Both plots are aligned according to the Acc values, but CHI is sca…
Figure 8
Figure 8. Figure 8: Effect of the UMAP (3D) method on clustering accuracy and Davies-Bouldin Index (DBI) on a random CIFAR-25 dataset. In DBI, smaller values are better. Acc is marked at the left, and DBI is marked at the right axis, respectively. Both plots are aligned according to the A…
Figure 9
Figure 9. Figure 9: Performance of FasterMSC and K-Means over Normalization + UMAP (3D) applied CIFAR-10 when initialized with different numbers of clusters. G.1 Bayesian Optimization Since we increase the correlation between Sil and ACC as it can be seen in [PITH_FULL_IMAGE:figures/full…
Figure 10
Figure 10. Figure 10 [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Accuracy and Silhouette Score of DROWCULA-Autoencoder (1536 - 1024 - 256 - 16D) with respect to the number of clusters on CIFAR-10 [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Performance of DROWCULA-UMAP (3D) with respect to the number of clusters and number of neighbors on CIFAR-100 dataset [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 48 canonical work pages

  1. [1]

    Adaloglou, F

    N. Adaloglou, F. Michels, H. Kalisch, and M. Kollmann. Exploring the limits of deep image clustering using pretrained models.arXiv preprint arXiv:2303.17896, 2023

    arXiv 2023

  2. [2]

    Ahmed, R

    M. Ahmed, R. Seraj, and S. M. S. Islam. The k-means algorithm: A comprehensive survey and performance evaluation.Electronics, 9(8): 1295, 2020

    work page 2020

  3. [3]

    Arbelaitz, I

    O. Arbelaitz, I. Gurrutxaga, J. Muguerza, J. M. Pérez, and I. Perona. An extensive comparative study of cluster validity indices.Pattern recog- nition, 46(1):243–256, 2013

    work page 2013

  4. [4]

    Bendale and T

    A. Bendale and T. E. Boult. Towards Open World Recognition. In CVPR, pages 1893–1902. IEEE Computer Society, 2015

    work page 1902

  5. [5]

    Berthelot, N

    D. Berthelot, N. Carlini, I. J. Goodfellow, N. Papernot, A. Oliver, and C. Raffel. MixMatch: A Holistic Approach to Semi-Supervised Learn- ing. InNeurIPS, pages 5050–5060, 2019

    work page 2019

  6. [6]

    Berthelot, N

    D. Berthelot, N. Carlini, E. D. Cubuk, A. Kurakin, K. Sohn, H. Zhang, and C. Raffel. ReMixMatch: Semi-Supervised Learning with Distribu- tion Matching and Augmentation Anchoring. InICLR, 2020

    work page 2020

  7. [7]

    Cali ´nski and J

    T. Cali ´nski and J. Harabasz. A dendrite method for cluster analysis. Communications in Statistics-theory and Methods, 3(1):1–27, 1974

    work page 1974

  8. [8]

    K. Cao, M. Brbic, and J. Leskovec. Open-world supervised learning. In ICLR, 2022

    work page 2022

Show all 56 references
  1. [9]

    Caron, P

    M. Caron, P. Bojanowski, A. Joulin, and M. Douze. Deep clustering for unsupervised learning of visual features. InECCV, 2018

    2018

  2. [10]

    Chapelle, B

    O. Chapelle, B. Schölkopf, and A. Zien, editors.Semi-Supervised Learning. The MIT Press, 2006. ISBN 9780262033589

    2006

  3. [11]

    Y . Chen, X. Zhu, W. Li, and S. Gong. Semi-Supervised Learning under Class Distribution Mismatch. InAAAI, 2020

    2020

  4. [12]

    Darcet, M

    T. Darcet, M. Oquab, J. Mairal, and P. Bojanowski. Vision Transformers Need Registers. InICLR, 2024

    2024

  5. [13]

    D. L. Davies and D. W. Bouldin. A Cluster Separation Measure.IEEE Trans. Pattern Anal. Mach. Intell., 1(2):224–227, 1979

    1979

  6. [14]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. InCVPR, 2009

    2009

  7. [15]

    Dosovitskiy and et al

    A. Dosovitskiy and et al. An Image is Worth 16x16 Words: Transform- ers for Image Recognition at Scale. InICLR, 2021

    2021

  8. [16]

    P. Du, S. Zhao, Z. Sheng, C. Li, and H. Chen. Semi-Supervised Learn- ing via Weight-aware Distillation under Class Distribution Mismatch. InICCV, 2023

    2023

  9. [17]

    E. Fini, E. Sangineto, S. Lathuilière, Z. Zhong, M. Nabi, and E. Ricci. A Unified Objective for Novel Class Discovery. InICCV, 2021

    2021

  10. [18]

    P. I. Frazier. A tutorial on bayesian optimization.arXiv preprint arXiv:1807.02811, 2018

    2018 arXiv

  11. [19]

    Gadetsky, Y

    A. Gadetsky, Y . Jiang, and M. Brbic. Let go of your labels with unsu- pervised transfer.arXiv preprint arXiv:2406.07236, 2024

    2024 arXiv

  12. [20]

    Ghriss and C

    A. Ghriss and C. Monteleoni. Deep clustering via probabilistic ratio-cut optimization.arXiv preprint arXiv:2502.03405, 2025

    2025 arXiv

  13. [21]

    L. Guo, Z. Zhang, Y . Jiang, Y . Li, and Z. Zhou. Safe Deep Semi- Supervised Learning for Unseen-Class Unlabeled Data. InICML, 2020

    2020

  14. [22]

    Gupta, S

    A. Gupta, S. Narayan, K. J. Joseph, S. Khan, F. S. Khan, and M. Shah. OW-DETR: Open-world Detection Transformer. InCVPR, 2022

    2022

  15. [23]

    K. Han, A. Vedaldi, and A. Zisserman. Learning to Discover Novel Visual Categories via Deep Transfer Clustering. InICCV, 2019

    2019

  16. [24]

    K. Han, S. Rebuffi, S. Ehrhardt, A. Vedaldi, and A. Zisserman. Auto- matically Discovering and Learning New Visual Categories with Rank- ing Statistics. InICLR, 2020

    2020

  17. [25]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. InCVPR, 2016

    2016

  18. [26]

    P. Kage, J. C. Rothenberger, P. Andreadis, and D. I. Diochnos. A Review of Pseudo-Labeling for Computer Vision.arXiv preprint arXiv:2408.07221v3, 2025

    2025

  19. [27]

    Kardan and K

    N. Kardan and K. O. Stanley. Mitigating fooling with competitive over- complete output layer neural networks. InIJCNN, pages 518–525, 2017

    2017

  20. [28]

    Kaufman and P

    L. Kaufman and P. J. Rousseeuw.Finding groups in data: an introduc- tion to cluster analysis. John Wiley & Sons, 2009

    2009

  21. [29]

    Krizhevsky, G

    A. Krizhevsky, G. Hinton, et al. Learning multiple layers of features from tiny images. Toronto, ON, Canada, 2009

    2009

  22. [30]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks.NeurIPS, 2012

    2012

  23. [31]

    Le and X

    Y . Le and X. Yang. Tiny imagenet visual recognition challenge.CS 231N, 7(7):3, 2015

    2015

  24. [32]

    D.-H. Lee. Pseudo-Label : The Simple and Efficient Semi-Supervised Learning Method for Deep Neural Networks.Workshop on Challenges in Representation Learning, ICML, 3(2):896, 2013

    2013

  25. [33]

    Lenssen and E

    L. Lenssen and E. Schubert. Medoid silhouette clustering with au- tomatic cluster number selection.Information Systems, 120:102290, 2024

    2024

  26. [34]

    McInnes, J

    L. McInnes, J. Healy, and J. Melville. Umap: Uniform manifold ap- proximation and projection for dimension reduction.arXiv preprint arXiv:1802.03426, 2018

    2018 arXiv

  27. [35]

    X. Meng, Z. Huang, W. He, X. Qi, R. Xiao, and C.-G. Li. Exploring a principled framework for deep subspace clustering.arXiv preprint arXiv:2503.17288, 2025

    2025 arXiv

  28. [36]

    H. Pham, Z. Dai, Q. Xie, and Q. V . Le. Meta Pseudo Labels. InCVPR, pages 11557–11568, 2021

    2021

  29. [37]

    A. P. Reynolds, G. Richards, and V . J. Rayward-Smith. The application of k-medoids and pam to the clustering of rules. InIDEAL, 2004

    2004

  30. [38]

    D. A. Reynolds et al. Gaussian mixture models.Encyclopedia of bio- metrics, 741(659-663):3, 2009

    2009

  31. [39]

    M. N. Rizve, K. Duarte, Y . S. Rawat, and M. Shah. In Defense of Pseudo-Labeling: An Uncertainty-Aware Pseudo-label Selection Framework for Semi-Supervised Learning. InICLR, 2021

    2021

  32. [40]

    M. N. Rizve, N. Kardan, S. Khan, F. S. Khan, and M. Shah. OpenLDN: Learning to Discover Novel Classes for Open-World Semi-Supervised Learning. InECCV, pages 382–401, 2022

    2022

  33. [41]

    Russakovsky and et al

    O. Russakovsky and et al. ImageNet Large Scale Visual Recognition Challenge.IJCV, 115(3):211–252, 2015

    2015

  34. [42]

    Sajjadi, M

    M. Sajjadi, M. Javanmardi, and T. Tasdizen. Regularization With Stochastic Transformations and Perturbations for Deep Semi- Supervised Learning. InNeurIPS, 2016

    2016

  35. [43]

    Schubert and P

    E. Schubert and P. J. Rousseeuw. Fast and eager k-medoids clustering: O (k) runtime improvement of the pam, clara, and clarans algorithms. Information Systems, 101:101804, 2021

    2021

  36. [44]

    Schubert, J

    E. Schubert, J. Sander, M. Ester, H. P. Kriegel, and X. Xu. Dbscan revisited, revisited: why and how you should (still) use dbscan.ACM Transactions on Database Systems (TODS), 42(3):1–21, 2017

    2017

  37. [45]

    Simonyan and A

    K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556, 2014

    2014 arXiv

  38. [46]

    Sohn and et al

    K. Sohn and et al. FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence. InNeurIPS, 2020

    2020

  39. [47]

    Tarvainen and H

    A. Tarvainen and H. Valpola. Mean teachers are better role mod- els: Weight-averaged consistency targets improve semi-supervised deep learning results. InNeurIPS, 2017

    2017

  40. [48]

    J. B. Tenenbaum, V . d. Silva, and J. C. Langford. A global geometric framework for nonlinear dimensionality reduction.Science, 290(5500): 2319–2323, 2000

    2000

  41. [49]

    Y . Tian, D. Krishnan, and P. Isola. Contrastive multiview coding. In ECCV, 2020

    2020

  42. [50]

    Van der Laan, K

    M. Van der Laan, K. Pollard, and J. Bryan. A new partitioning around medoids algorithm.Journal of Statistical Computation and Simulation, 73(8):575–584, 2003

    2003

  43. [51]

    Van der Maaten and G

    L. Van der Maaten and G. Hinton. Visualizing data using t-sne.Journal of machine learning research, 9(11), 2008

    2008

  44. [52]

    J. E. van Engelen and H. H. Hoos. A survey on semi-supervised learn- ing.Machine Learning, 109(2):373–440, 2020

    2020

  45. [53]

    Vaswani and et al

    A. Vaswani and et al. Attention is All you Need. InNeurIPS, 2017

    2017

  46. [54]

    Y . Yang, N. Jiang, Y . Xu, and D. Zhan. Robust Semi-Supervised Learn- ing by Wisely Leveraging Open-Set Data.IEEE Trans. Pattern Anal. Mach. Intell., 46(12):8334–8347, 2024

    2024

  47. [55]

    Zhao and K

    B. Zhao and K. Han. Novel Visual Category Discovery with Dual Rank- ing Statistics and Mutual Knowledge Distillation. InNeurIPS, 2021

    2021

  48. [56]

    seen” class. We obtain these

    X. Zhu and A. B. Goldberg.Introduction to Semi-Supervised Learning. Morgan & Claypool, 2009. A Distances: Metrics and Pseudo-Metrics Notation.We useXto denote the set of instances andYthe set of labels. Throughout the paper we usento denote the dimension of an instance. For tw...

    2009

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.