SuperCM: Improving Semi-Supervised Learning and Domain Adaptation through differentiable clustering

Ahc\`ene Boubekki; Durgesh Singh; Michael Kampffmeyer; Robert Jenssen

REVIEW 2 major objections 5 minor 49 references

SuperCM: Improving Semi-Supervised Learning and Domain Adaptation through differentiable clustering

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read SuperCM shows that a differentiable clustering loss with class-conditional centroids estimated from labeled data raises semi-supervised accuracy by up to 15 percentage points and DANN domain-adaptation accuracy by up to 10 percentage…

desk verdict A useful plug-in clustering regularizer with real low-supervision gains, but the centroid update equation needs a careful second look before the mechanism claims can be trusted. read the letter →

arxiv 2507.13779 v1 pith:4DV47JEV submitted 2025-07-18 cs.CV

Durgesh Singh , Ahcène Boubekki , Robert Jenssen , Michael Kampffmeyer This is my paper · ORCID

classification cs.CV

keywords semi-supervisedlearningunsuperviseddomainadaptationdifferentiableclusteringGaussianmixturemodelregularizerclass-conditionalcentroidsDANNlow-supervisionregimes

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

SuperCM is a training strategy that makes the clustering assumption explicit instead of relying on implicit consistency or entropy losses. It attaches a differentiable clustering module to a classifier's feature extractor and computes the module's cluster centers as moving averages of labeled examples per class. The paper claims this raises SSL accuracy by up to 15 percentage points and DANN's UDA accuracy by up to 10 points, with the largest gains in low-supervision regimes. The same plug-in term can regularize several established SSL methods and DANN-based UDA variants.

What carries the argument

The central object is the Clustering Module (CM), a one-layer auto-encoder that outputs soft cluster assignments and is trained with a GMM-derived loss containing reconstruction, sparsity/cluster-merging, and Dirichlet-prior terms. SuperCM sets K equal to the number of classes, computes centroids as exponentially moving averages of labeled-source features according to Eq. (3), and sets the Dirichlet coefficient alpha = 1 to disable the prior. The total training loss is cross-entropy on labeled class responsibilities plus beta times the CM loss on labeled and unlabeled data plus delta times the base SSL/UDA loss, so the clustering term acts as a differentiable regularizer in an end-to-end pipeline.

What would settle it

Train SuperCM on a benchmark where classes are deliberately multimodal, such as a modified CIFAR-10 where each object class is split into two visually distinct subgroups, using 100 to 250 labels per dataset, and compare against the plain cross-entropy baseline; if SuperCM fails to beat or matches that baseline, the single-cluster-per-class assumption is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a GMM-inspired, end-to-end differentiable clustering loss becomes a broadly useful regularizer once its centroids are anchored to supervised information: per-class moving averages of labeled features replace freely learned cluster centers, preventing collapsed or misaligned clusters. On MNIST with 100 labels, SuperCM alone goes from 82.28% to 97.45% accuracy over the cross-entropy baseline; on Office-31 A to W it raises DANN from 80.80% to 91.03%. Similar improvements appear across CIFAR-10, SVHN, CIFAR-100, STL-10 and four UDA benchmarks, and the learned features become visibly more compact and separated.

Load-bearing premise

The load-bearing premise is that each class is well described by a single compact cluster in the learned feature space, and that the centroids estimated from the few labeled examples are reliable enough to anchor that cluster.

Editorial extensions

If this is right

In low-label SSL settings, e.g. MNIST with 100 labels, SuperCM brings standalone accuracy from 82.28% to 97.45%.
DANN's accuracy improves by up to 10 points on Office-31, and average accuracy increases on Office-31, Office-Home, Office-Caltech, and ImageClef.
The regularizer plugs into five existing SSL methods (Pi-model, Mean Teacher, VAT, Pseudo-labels, ICT) and improves most of their tested configurations.
Combining SuperCM with MCC or BNM regularizers further improves several Office-31 and Office-Home transfers.
Training with SuperCM reduces Proxy-A distance between source and target features, indicating better domain alignment.

Reading between the lines

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

The single-centroid-per-class assumption suggests the method is most suited to balanced, unimodal class structure; on long-tail or intra-class-multimodal data, the same recipe may need multiple centroids per class.
Because the centroids are moving averages over labeled examples, their reliability should depend on feature quality early in training, so warm-starting with self-supervised pretraining might amplify the gains.
An ablation separating the reconstruction term from the pure clustering terms would test whether part of the benefit comes from representation smoothness rather than clustering per se.
If the mechanism transfers beyond images, the same labeled-centroid clustering regularizer could be applied to medical or textual data with few annotations, where class structure is less clean.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

Quick read on arXiv:2507.13779 (SuperCM). The paper extends the authors' earlier clustering module to SSL and UDA by computing class-conditional centroids from labeled data via a moving average, then using the CM loss as a regularizer. The headline results are real in the tables: e.g., MNIST with 100 labels jumps from 82.3 to 97.5, and DANN on Office-31 A to W from 80.8 to 91.0. Those are the kind of gains people in label-scarce domains would care about, and the code is available.

What's genuinely new here is the UDA application and the broader SSL benchmark coverage (CIFAR-100, STL-10, etc.). The implementation is simple, the overhead is negligible, and the ablations on centroid strategies and hyperparameter sensitivity are useful. The t-SNE and Proxy-A distance analysis give some supporting evidence for the alignment story.

Now the soft spots.

First, Eq. (2) claims a simplification of E2+E3 that doesn't follow from Eq. (1) with the printed signs. I worked through the two-cluster case; the expression in Eq. (2) is not what you get from the loss as written. This is probably a typo or an unstated rearrangement, but it makes the derivation hard to trust.

Second, and more seriously, Eq. (3) as written is ambiguous. The update uses a global iteration counter tau, and for classes absent from the current labeled batch the data term is zero, so the centroid is multiplied by (tau-1)/tau and shrinks. Over hundreds of thousands of iterations, classes that appear rarely get their norms pulled toward zero. In the low-supervision and many-class settings where the paper claims its largest gains, absent classes are common. If the code implements Eq. (3) literally, the clustering terms for those classes vanish and the mechanism is not what the text claims. If the implementation instead uses per-class update counters or skips absent classes, the equation needs to say so. This is easily checked against the released code, but it's currently a load-bearing ambiguity.

Third, the text overclaims. "Consistently outperforms" is contradicted by some of its own tables: Pseudo-Label on SVHN 250 drops from 81.2 to 79.9, and on STL-10 600 from 61.7 to 55.2. Several improvements are within standard deviation, and the significance tests are only mentioned for bold numbers, not systematically reported.

The baselines are also dated (mostly pre-2020), but that's somewhat acceptable for a regularizer study.

Overall: the core idea is plausible, the empirical evidence is substantial, and the code availability is a plus. The paper deserves a serious referee, but the referee should ask for a code check of the centroid update, a corrected Eq. (2), and a more careful wording of the claims. If the centroid update is indeed as ambiguous as it looks, the method's mechanism needs to be re-stated before the results can be taken at face value. I'd still cite it as an example of differentiable clustering regularization, but not before the ambiguity is resolved.

Referee Report

2 major / 5 minor

Summary. The paper proposes SuperCM, a regularization method for semi-supervised learning (SSL) and unsupervised domain adaptation (UDA) that appends a differentiable clustering module (CM) to a feature extractor. The module's centroids are updated as class-wise moving averages of labeled features rather than learned by gradient descent, and the CM loss is combined with a standard cross-entropy loss and optionally with an existing SSL/UDA objective. The method is evaluated as a standalone SSL objective and as a regularizer for five SSL baselines and for DANN, on MNIST, SVHN, CIFAR-10/100, STL-10, Office-31, Office-Home, Office-Caltech, and ImageClef. The authors report gains of up to 15 percentage points in SSL accuracy and up to 10 percentage points in UDA performance, particularly in low-supervision settings, and provide an ablation of centroid estimation strategies.

Significance. If the proposed mechanism is sound, the paper offers a simple, plug-in clustering regularizer with substantial empirical gains in low-label and domain-shift regimes, supported by extensive benchmarks and released code. The main conceptual contribution is the injection of label information into a differentiable clustering objective through class-conditional centroids, which is a plausible and potentially useful idea. However, the load-bearing derivations and the centroid update rule contain unresolved technical issues that must be addressed before the reported gains can be confidently attributed to the claimed mechanism.

major comments (2)

[§2.3.1, Eq. (2)] The claimed simplification of E2+E3 is algebraically incorrect. For two clusters, writing γ_i1=γ and γ_i2=1−γ, Eq. (1) gives E2 = −γ(1−γ)(‖μ1‖²+‖μ2‖²) and E3 = γ²‖μ1‖² + (1−γ)²‖μ2‖² + 2γ(1−γ)μ1ᵀμ2, whose sum is not γ(1−γ)‖μ1−μ2‖². A concrete counterexample is γ=0.5 with orthogonal unit centroids, for which Eq. (1) yields 0 while Eq. (2) yields 0.5. Because Eq. (2) is the paper's only stated justification for the 'sparsity and cluster merging' effect of E2+E3, the explanation of the clustering mechanism is unsupported as printed and should be corrected or replaced with the actual expression.
[§3.2, Eq. (3)] The centroid update rule is either underspecified or incompatible with the claimed mechanism. If τ is the global iteration counter, then for a class absent from the labeled mini-batch the data term is zero and the centroid is multiplied by (τ−1)/τ on every such iteration, so the norms of rarely seen classes collapse toward zero over training. The paper does not state that τ is a per-class counter or that the update is skipped when class k has no labeled sample in the batch. Additionally, the data term is normalized by the total labeled batch size n_B^(l) rather than by the number of class-k samples, which further dilutes updates for rare classes. In the low-supervision and many-class settings where the largest gains are claimed (100 MNIST labels, Office-31/Office-Home with 31/65 classes), most classes are absent from most batches, so the literal Eq. (3) would destroy the label-injected centroids and the E2/E3 terms would not receive the class-conditional signal that is the paper's novelty. The authors must specify the update precisely and, ideally, verify against the released code; otherwise the reported results cannot be attributed to the described mechanism.

minor comments (5)

[Footnote on page 2] The sentence 'This work was supported by by the Research Council of Norway' contains a duplicated 'by' that should be removed.
[Table 1] The column headings '10δ' and '10β' are ambiguous: it is unclear whether the entries are δ and β themselves or ten times those values, and if a scaling is intended it should be stated explicitly in the caption or text.
[Table 4] The baseline column label 'DANN [?]' contains an unresolved citation placeholder that should be replaced with the proper reference.
[§7.3.3, Eq. (5)] The symbol γ is used both for the ramp-up steepness in Eq. (5) and for the soft assignment responsibilities γ_ik in Eq. (1); these should be distinguished to avoid confusion.
[§5, STL-10 paragraph] The text states that experiments use '400 and 4000 labeled data' for STL-10, but Table 2 and the surrounding discussion report '600 labels' and '4000 labels'; please correct the inconsistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the reported SSL and UDA gains are empirical comparisons against independently trained baselines, and the reuse of the authors' earlier Clustering Module is a component choice rather than a self-referential derivation.

full rationale

The paper's central claim is empirical: adding the CM regularizer with labeled-data centroids improves accuracy over CE and DANN baselines. The CM loss (Eq. 1) is stated explicitly and is not defined in terms of the reported accuracies; no reported number is a fitted parameter renamed as a prediction. The centroids in Eq. (3) are computed from labeled data, but the test accuracies are measured on held-out test sets after training, so the comparison is not forced by construction. The reuse of the authors' prior CM [7] and the preliminary ICASSP version [8] is self-citation, but it is not load-bearing in the sense of substituting for evidence: the paper's own tables compare SuperCM against external baselines, and the CM equations are reproduced in the text. No uniqueness theorem or ansatz is imported from the authors' prior work to forbid alternatives. The only notable concern, an ambiguity in Eq. (3) about whether centroids of classes absent from a mini-batch are decayed, is a potential implementation or correctness issue rather than a circularity; it does not make any reported result equivalent to its input by definition.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. The moving-average class centroids are parameters, not invented objects. The main load-bearing inputs are the clustering assumption, the prior CM loss, and the tuned weights beta and delta.

free parameters (4)

beta (CM loss weight) = varies; examples from Table 1 (10*beta): 0.18 for None, 0.12 for MT, 0.63 for VAT, 0.17 for PI, 3.01 for PL, 2.23 for…
Tuned via Bayesian optimization on validation for each SSL base model; fixed by hand for UDA. Controls the influence of the clustering loss in Eq. (4).
delta (base loss weight) = varies; examples from Table 1 (10*delta): 0.00, 0.15, 5.62, 0.19, 0.18, 0.16; in UDA, delta follows the ramp-up…
Tuned per SSL setup; in UDA it ramps from 0 to 1 as training progresses.
gamma (UDA ramp-up steepness) = 10
Chosen by hand following DANN's original schedule in [30]; not tuned.
alpha (Dirichlet prior coefficient) = 1
Set to 1 to discard the cluster-prior term E4, as stated in Section 3.2.

assumptions (4)

domain assumption Clustering assumption: points in the same feature-space cluster belong to the same class
Invoked in Section 1 and formalized through the CM loss; the method relies on this to transfer clustering to classification.
domain assumption Latent features are drawn from K isotropic Gaussian clusters, one per class
Inherited from the Clustering Module in [7], Section 2.3.1; SuperCM sets K equal to the number of classes.
domain assumption The CM loss from [7] is a valid differentiable surrogate for GMM clustering
The method reuses LCM from Eq. (1) as its clustering objective without re-deriving it.
standard math DANN's adversarial domain alignment provides a sound baseline objective
The UDA experiments use DANN [30] as the base model with its standard gradient-reversal training.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SuperCM: Improving Semi-Supervised Learning and Domain Adaptation through differentiable clustering." pith.science (2026). https://pith.science/paper/4DV47JEV

@misc{pith2026250713779,
  author       = {Pith},
  title        = {Pith review of: SuperCM: Improving Semi-Supervised Learning and Domain Adaptation through differentiable clustering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4DV47JEV}},
  note         = {Machine review of arXiv:2507.13779}
}

read the original abstract

Semi-Supervised Learning (SSL) and Unsupervised Domain Adaptation (UDA) enhance the model performance by exploiting information from labeled and unlabeled data. The clustering assumption has proven advantageous for learning with limited supervision and states that data points belonging to the same cluster in a high-dimensional space should be assigned to the same category. Recent works have utilized different training mechanisms to implicitly enforce this assumption for the SSL and UDA. In this work, we take a different approach by explicitly involving a differentiable clustering module which is extended to leverage the supervised data to compute its centroids. We demonstrate the effectiveness of our straightforward end-to-end training strategy for SSL and UDA over extensive experiments and highlight its benefits, especially in low supervision regimes, both as a standalone model and as a regularizer for existing approaches.

Figures

Figures reproduced from arXiv: 2507.13779 by the authors.

**Figure 1.** Architecture of our proposed method classification responsibilities of the labeled data and the CM loss applied to both types of data. SuperCM may also be used as a regularizer for existing SSL and UDA models in which case the combined loss can be stated as follows: L SSL/UDA SuperCM = CE(Γ (l) B , Y (l) B ) + β · LCM(X (l+u) B , Γ (l+u) B , X¯ (l+u) B ) + δ · L(SSL/UDA) (4) Where β ≥ 0 and δ ≥ 0 weights the CM loss… view at source ↗

**Figure 2.** Training with CE and SuperCM for different supervision on CIFAR-10 [9] dataset. 6. SSL Analysis 6.1. Amount of Supervision We evaluate the performance of our method on CIFAR-10 for different training amounts of label data ranging from 250 to 4000 instances, with and without VAT as the SSL base model. The results are summarized in [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗

**Figure 3.** Top-1 accuracy of SuperCM trained with 600 labels and di [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: t-SNE [43] plots for backbone features of the models trained with CE and SuperCM. The visual [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]

**Figure 5.** Figure 5: Ablation on different DA configurations for determining centroids. The visualization shows different domain pairs for the Office-Caltech [46] and Office-31 [44] dataset. tation training, resulting in better backbone features, and consequently improving the overall per…

**Figure 6.** Figure 6: t-SNE [43] visualization of backbone features during the DANN training. The visualization shows [PITH_FULL_IMAGE:figures/full_fig_p028_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 38 canonical work pages

[1]

J. Gu, Z. Wang, J. Kuen, L. Ma, A. Shahroudy, B. Shuai, T. Liu, X. Wang, G. Wang, J. Cai, T. Chen, Recent advances in convolutional neural networks, Pat- tern Recognition 77 (2018) 354–377. doi:10.1016/j.patcog.2017.10.013

doi:10.1016/j.patcog.2017.10.013 2018
[2]

C. Tang, X. Zeng, L. Zhou, Q. Zhou, P. Wang, X. Wu, H. Ren, J. Zhou, Y . Wang, Semi-supervised medical image segmentation via hard positives oriented con- trastive learning, Pattern Recognition 146 (2024) 110020. doi:10.1016/j. patcog.2023.110020

arXiv 2024
[3]

Sheikhpour, M

R. Sheikhpour, M. A. Sarram, S. Gharaghani, M. A. Z. Chahooki, A survey on semi-supervised feature selection methods, Pattern Recognition 64 (2017) 141 – 158, cited by: 375. doi:10.1016/j.patcog.2016.11.003

work page doi:10.1016/j.patcog.2016.11.003 2017
[4]

X. Liu, C. Yoo, F. Xing, H. Oh, G. E. Fakhri, J.-W. Kang, J. Woo, Deep unsupervised domain adaptation: A review of recent advances and perspec- tives, APSIPA Transactions on Signal and Information Processing 11 (2022). doi:10.1561/116.00000192

doi:10.1561/116.00000192 2022
[5]

Ben-David, J

S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza, F. Pereira, J. W. Vaughan, 30 A theory of learning from di fferent domains, Machine Learning 79 (1) (2010) 151–175. doi:10.1007/s10994-009-5152-4

doi:10.1007/s10994-009-5152-4 2010
[6]

A. E. Ezugwu, A. M. Ikotun, O. O. Oyelade, L. Abualigah, J. O. Agushaka, C. I. Eke, A. A. Akinyelu, A comprehensive survey of clustering algorithms: State-of- the-art machine learning applications, taxonomy, challenges, and future research prospects, Engineering Applications of Artificial Intelligence 110 (2022) 104743. doi:10.1016/j.engappai.2022.104743

work page arXiv 2022
[7]

Boubekki, M

A. Boubekki, M. C. Kamp ffmeyer, U. Brefeld, R. Jenssen, Joint optimization of an autoencoder for clustering and embedding, Machine Learning 110 (2021) 1901–1937

work page 2021
[8]

Singh, A

D. Singh, A. Boubekki, R. Jenssen, M. C. Kamp ffmeyer, Supercm: Revisiting clustering for semi-supervised learning, in: ICASSP 2023 - 2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5. doi:10.1109/ICASSP49357.2023.10095856

arXiv 2023

Show all 49 references

[9]

Krizhevsky, G

A. Krizhevsky, G. Hinton, Learning multiple layers of features from tiny images, Tech. rep., Technical Report, University of Toronto (2009)

2009
[10]

Laine, T

S. Laine, T. Aila, Temporal ensembling for semi-supervised learning, 5th Inter- national Conference on Learning Representations (ICLR) (2017)

2017
[11]

Tarvainen, H

A. Tarvainen, H. Valpola, Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results, in: Proceed- ings of the 31st International Conference on Neural Information Processing Sys- tems (NeurIPS 2017), Curran Associate...

2017
[12]

Miyato, S

T. Miyato, S. ichi Maeda, M. Koyama, S. Ishii, Virtual adversarial training: A regularization method for supervised and semi-supervised learning, IEEE Trans- actions on Pattern Analysis and Machine Intelligence (IEEE TPAMI) 41 (2019) 1979–1993. 31

2019
[13]

Grandvalet, Y

Y . Grandvalet, Y . Bengio, Semi-supervised learning by entropy minimization, in: Advances in Neural Information Processing Systems (NeurIPS 2005), 2005

2005
[14]

D.-H. Lee, Pseudo-label : The simple and e fficient semi-supervised learning method for deep neural networks, ICML 2013 Workshop : Challenges in Rep- resentation Learning (WREPL) (07 2013)

2013
[15]

Z. Feng, Q. Zhou, Q. Gu, X. Tan, G. Cheng, X. Lu, J. Shi, L. Ma, Dmt: Dynamic mutual training for semi-supervised learning, Pattern Recognition 130 (2022) 108777. doi:https://doi.org/10.1016/j.patcog.2022.108777

2022
[16]

Y . Wang, J. Zhang, M. Kan, S. Shan, Learning pseudo labels for semi-and-weakly supervised semantic segmentation, Pattern Recognition 132 (2022) 108925.doi: https://doi.org/10.1016/j.patcog.2022.108925

2022
[17]

Y . Wang, K. Yuan, G. Schaefer, X. Liu, L. Jing, K. Guo, J. Z. Wang, H. Fang, Re- fining pseudo-labels through iterative mix-up for weakly supervised semantic seg- mentation, Pattern Recognition 159 (2025) 111975. doi:10.1016/j.patcog. 2025.111975

2025
[18]

D. Guan, Y . Xing, J. Huang, A. Xiao, A. El Saddik, S. Lu, S2match: Self- paced sampling for data-limited semi-supervised learning, Pattern Recognition 159 (2025) 111121. doi:doi.org/10.1016/j.patcog.2024.111121

2025
[19]

Huang, S

Y . Huang, S. Li, Z. Guo, Q. Mei, Z. Han, X. Wang, H. Wang, Boundary feature alignment for semi-supervised medical image segmentation, Pattern Recogni- tion 170 (2026) 111946. doi:https://doi.org/10.1016/j.patcog.2025. 111946

2026 doi
[20]

Verma, A

V . Verma, A. Lamb, J. Kannala, Y . Bengio, D. Lopez-Paz, Interpolation consis- tency training for semi-supervised learning, in: Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI 2019), Interna- tional Joint Conferences on Artifi...

2019
[21]

Berthelot, N

D. Berthelot, N. Carlini, I. Goodfellow, N. Papernot, A. Oliver, C. Ra ffel, Mix- match: A holistic approach to semi-supervised learning, in: Proceedings of the 33rd Conference on Neural Information Processing Systems (NeurIPS), 2019, pp. 5050–5060

2019
[22]

Zhang, M

H. Zhang, M. Cisse, Y . N. Dauphin, D. Lopez-Paz, mixup: Beyond empirical risk minimization, in: International Conference on Learning Representations (ICLR), 2018

2018
[23]

K. Sohn, D. Berthelot, C. Li, Z. Zhang, N. Carlini, E. D. Cubuk, A. Kurakin, H. Zhang, C. Ra ffel, Fixmatch: Simplifying semi-supervised learning with con- sistency and confidence, in: Proceedings of the 34th Conference on Neural Infor- mation Processing Systems (NeurIPS), 202...

2020
[24]

K.-B. Nguyen, Sequencematch: Revisiting the design of weak–strong augmen- tations for semi-supervised learning, in: Proceedings of the IEEE /CVF Winter Conference on Applications of Computer Vision (W ACV), 2024, pp. 96–106. doi:10.1109/WACV56614.2024.00428

2024
[25]

F. Xu, F. Zhu, F. Chen, Graphixmatch: Improving semi-supervised learning for graph classification with fixmatch, Knowledge-Based SystemsIn press (2024). doi:10.1016/j.neucom.2024.128356

2024
[26]

Gretton, K

A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Sch ¨olkopf, A. Smola, A kernel two-sample test, Journal of Machine Learning Research 13 (2012) 723–773

2012
[27]

C. Cao, D. Phukan, C. Wu, J. Wang, Deep joint subdomain alignment for un- supervised domain adaptation, Pattern Recognition 151 (2024) 109408. doi: 10.1016/j.patcog.2024.109408

2024
[28]

Y . Jin, X. Wang, M. Long, J. Wang, Minimum class confusion for versatile do- main adaptation, in: A. Vedaldi, H. Bischof, T. Brox, J.-M. Frahm (Eds.), Com- puter Vision ECCV 2020, Springer International Publishing, Cham, 2020, pp. 464–480. 33

2020
[29]

S. Cui, S. Wang, J. Zhuo, L. Li, Q. Huang, Q. Tian, Towards discriminability and diversity: Batch nuclear-norm maximization under label insu fficient situations, in: 2020 IEEE /CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 3940–3949. doi:10.1109/C...

2020
[30]

Ganin, E

Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. Marchand, V . Lempitsky, Domain-Adversarial Training of Neural Networks, Springer International Publishing, Cham, 2017, pp. 189–209. doi:10.1007/ 978-3-319-58347-1_10

2017
[31]

Y . Shi, K. Wu, Y . Han, Y . Shao, B. Li, F. Wu, Source-free and black-box do- main adaptation via distributionally adversarial training, Pattern Recognition 145 (2023) 109750. doi:10.1016/j.patcog.2023.109750

2023
[32]

Y . Fang, C. Chen, W. Zhang, J. Wu, Z. Zhang, S. Xie, Prototype learning for adversarial domain adaptation, Pattern Recognition 158 (2024) 110653. doi: 10.1016/j.patcog.2024.110653

2024
[33]

Caron, P

M. Caron, P. Bojanowski, A. Joulin, M. Douze, Deep clustering for unsuper- vised learning of visual features, in: European Conference on Computer Vision (ECCV), 2018

2018
[34]

J. Li, P. Zhou, C. Xiong, S. Hoi, Prototypical contrastive learning of unsuper- vised representations, in: International Conference on Learning Representations (ICLR), 2021

2021
[35]

LeCun, B

Y . LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, L. D. Jackel, Backpropagation Applied to Handwritten Zip Code Recognition, Neural Computation 1 (4) (1989) 541–551. doi:10.1162/neco.1989.1.4.541

1989 doi
[36]

Coates, A

A. Coates, A. Ng, H. Lee, An analysis of single-layer networks in unsupervised feature learning, Journal of Machine Learning Research - Proceedings Track 15 (2011) 215–223

2011
[37]

Oliver, A

A. Oliver, A. Odena, C. Raffel, E. D. Cubuk, I. J. Goodfellow, Realistic evaluation of deep semi-supervised learning algorithms, in: Proceedings of the 32nd Interna- 34 tional Conference on Neural Information Processing Systems, Curran Associates Inc., Red Hook, NY , USA, 2018...

2018
[38]

D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, in: 3rd Inter- national Conference on Learning Representations (ICLR), San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015

2015
[39]

Nesterov, A method for solving the convex programming problem with conver- gence rate o(1/k2), Proceedings of the USSR Academy of Sciences 269 (1983) 543–547

Y . Nesterov, A method for solving the convex programming problem with conver- gence rate o(1/k2), Proceedings of the USSR Academy of Sciences 269 (1983) 543–547

1983
[40]

Loshchilov, F

I. Loshchilov, F. Hutter, SGDR: stochastic gradient descent with warm restarts, in: 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings, 2017

2017
[41]

Biewald, Experiment tracking with weights and biases, software available from wandb.com (2020)

L. Biewald, Experiment tracking with weights and biases, software available from wandb.com (2020)

2020
[42]

Netzer, T

Y . Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, A. Y . Ng, Reading digits in natural images with unsupervised feature learning (2011)

2011
[43]

van der Maaten, G

L. van der Maaten, G. Hinton, Visualizing data using t-sne, Journal of Machine Learning Research 9 (86) (2008) 2579–2605

2008
[44]

Saenko, B

K. Saenko, B. Kulis, M. Fritz, T. Darrell, Adapting visual category models to new domains, in: European Conference on Computer Vision (ECCV), 2010

2010
[45]

Venkateswara, J

H. Venkateswara, J. Eusebio, S. Chakraborty, S. Panchanathan, Deep hashing network for unsupervised domain adaptation, in: IEEE International Conference on Computer Vision (ICCV) Workshops, 2017

2017
[46]

B. Gong, Y . Shi, F. Sha, K. Grauman, Geodesic flow kernel for unsupervised domain adaptation, in: 2012 IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2012), 2012, pp. 2066–2073. doi:10.1109/CVPR.2012. 6247911. 35

2012 doi
[47]

Caputo, H

B. Caputo, H. M ¨uller, J. Martinez-Gomez, M. Villegas, B. Acar, N. Patricia, N. Marvasti, S. ¨Usk¨udarlı, R. Paredes, M. Cazorla, I. Garcia-Varea, V . Morell, Imageclef 2014: Overview and analysis of the results, in: E. Kanoulas, M. Lupu, P. Clough, M. Sanderson, M. Hall, A. ...

2014
[48]

K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recogni- tion, 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR

2016
[3641]

doi:10.24963/ijcai.2019/504. 32

2019 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

J. Gu, Z. Wang, J. Kuen, L. Ma, A. Shahroudy, B. Shuai, T. Liu, X. Wang, G. Wang, J. Cai, T. Chen, Recent advances in convolutional neural networks, Pat- tern Recognition 77 (2018) 354–377. doi:10.1016/j.patcog.2017.10.013

doi:10.1016/j.patcog.2017.10.013 2018

[2] [2]

C. Tang, X. Zeng, L. Zhou, Q. Zhou, P. Wang, X. Wu, H. Ren, J. Zhou, Y . Wang, Semi-supervised medical image segmentation via hard positives oriented con- trastive learning, Pattern Recognition 146 (2024) 110020. doi:10.1016/j. patcog.2023.110020

arXiv 2024

[3] [3]

Sheikhpour, M

R. Sheikhpour, M. A. Sarram, S. Gharaghani, M. A. Z. Chahooki, A survey on semi-supervised feature selection methods, Pattern Recognition 64 (2017) 141 – 158, cited by: 375. doi:10.1016/j.patcog.2016.11.003

work page doi:10.1016/j.patcog.2016.11.003 2017

[4] [4]

X. Liu, C. Yoo, F. Xing, H. Oh, G. E. Fakhri, J.-W. Kang, J. Woo, Deep unsupervised domain adaptation: A review of recent advances and perspec- tives, APSIPA Transactions on Signal and Information Processing 11 (2022). doi:10.1561/116.00000192

doi:10.1561/116.00000192 2022

[5] [5]

Ben-David, J

S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza, F. Pereira, J. W. Vaughan, 30 A theory of learning from di fferent domains, Machine Learning 79 (1) (2010) 151–175. doi:10.1007/s10994-009-5152-4

doi:10.1007/s10994-009-5152-4 2010

[6] [6]

A. E. Ezugwu, A. M. Ikotun, O. O. Oyelade, L. Abualigah, J. O. Agushaka, C. I. Eke, A. A. Akinyelu, A comprehensive survey of clustering algorithms: State-of- the-art machine learning applications, taxonomy, challenges, and future research prospects, Engineering Applications of Artificial Intelligence 110 (2022) 104743. doi:10.1016/j.engappai.2022.104743

work page arXiv 2022

[7] [7]

Boubekki, M

A. Boubekki, M. C. Kamp ffmeyer, U. Brefeld, R. Jenssen, Joint optimization of an autoencoder for clustering and embedding, Machine Learning 110 (2021) 1901–1937

work page 2021

[8] [8]

Singh, A

D. Singh, A. Boubekki, R. Jenssen, M. C. Kamp ffmeyer, Supercm: Revisiting clustering for semi-supervised learning, in: ICASSP 2023 - 2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5. doi:10.1109/ICASSP49357.2023.10095856

arXiv 2023