Pith. sign in

REVIEW 5 major objections 4 minor 39 references

Multi-level Supervised Contrastive Learning

T0 review · 5 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Multi-level contrastive learning, with one projection head per label level, outperforms single-head contrastive learning and delivers its largest gains when training data are scarce.

desk verdict Simple multi-head SupCon extension with solid math and eye-catching low-data gains, but the empirical reporting (no error bars, test-set tuning, missing baselines) keeps the SOTA claim from being verified. read the letter →

arxiv 2502.02202 v2 pith:DUZ5DK77 submitted 2025-02-04 cs.LG

Naghmeh Ghanooni , Barbod Pajoum , Harshit Rawal , Sophie Fellenz , Vo Nguyen Le Duy , Marius Kloft This is my paper · ORCID
classification cs.LG
keywords supervisedcontrastivelearningmulti-levelrepresentationprojectionheadshierarchicalclassificationmulti-labeltemperaturescalinglow-data
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

The paper proposes MLCL, a supervised contrastive learning framework that replaces the single projection head of standard contrastive learning with several heads, one per level of label structure (subclass and superclass, or each aspect label plus a global head). Each head runs the standard supervised contrastive loss with its own temperature, and the final loss is a weighted sum. The central claim is that this lets one encoder represent multiple notions of similarity at once, and that this improves accuracy over single-head contrastive learning, especially when training data are scarce. The authors report gains of roughly 1 percentage point on full CIFAR-100 and DeepFashion, 9 to 10 points when CIFAR-100 is cut to 5K or 10K training samples, and consistent gains on two aspect-based sentiment datasets. If true, this means contrastive learning can exploit label hierarchies and multi-label structure without task-specific architectures, with the largest benefits in data-limited settings.

What carries the argument

The load-bearing object is the set of $H$ projection heads $\{g_h\}$, each producing a projection $z^h_i = g_h(f(\tilde{x}_i))$ and contributing a supervised contrastive loss $\mathcal{L}_h$ with its own temperature $\tau_h$; the total loss is $\sum_h \alpha_h \mathcal{L}_h$ (plus cross-entropy for text). Positive pairs are defined separately per head by the relevant similarity criterion, so one head sees same-subclass pairs, another sees same-superclass pairs, and a global head sees pairs with high Jaccard label overlap. The temperature analysis is what carries the argument: as $\tau \to 0^+$ the loss concentrates on the hardest negative and the hardest positive, while as $\tau \to +\infty$ all negatives contribute uniformly, so the paper sets low temperature for fine subclass separation and high temperature for coarse superclass grouping. This lets one encoder integrate multiple similarity notions through a single weighted objective.

What would settle it

Train MLCL on the 5K and 10K CIFAR-100 subsets with the superclass temperature chosen by a validation split, for example a grid search over $\tau_2 \in \{0.1, 0.3, 0.5, 0.7, 1.0\}$, and compare against SupCon under identical epochs and batch size; if the gap shrinks to near zero or reverses, the low-data advantage is driven by test-set selection rather than the multi-level objective.

Watch

Extended reading notes

Core claim

The central claim is that multiple projection heads, each specialized to one level of a label hierarchy or one aspect label, produce better encoders than a single supervised contrastive head. In the hierarchical setting, a low-temperature head on subclass labels separates fine classes while a high-temperature head on superclass labels keeps coarse groupings together; in the multi-label text setting, one head per aspect plus a global head that weights pairs by Jaccard similarity acts as a regularizer. After training, the heads are discarded and a linear classifier is trained on the frozen encoder, so the gains come from the representation itself. The paper presents this as a generalization of supervised contrastive learning that subsumes the single-head case and is applicable to both hierarchical and multi-label classification.

Load-bearing premise

The reported gains rest on the assumption that the manually chosen temperature for the superclass head, and the similarity threshold for the global head in the text experiments, would still be good choices on a separate validation set rather than having been tuned to the test data.

Editorial extensions

If this is right

  • If MLCL is right, supervised contrastive learning can exploit class hierarchies and multi-label structure with no change to the encoder, only additional projection heads and per-head temperatures.
  • The reported 9 to 10 point gains on 5K and 10K CIFAR-100 samples suggest the method's main value is in low-data regimes, where a single-head loss overfits or fails to structure the space.
  • Because the heads are removed at inference, MLCL adds no inference cost and yields a single representation usable by any downstream linear classifier.
  • The global projection head's stronger tolerance to label noise, up to 7 points in the appendix, implies the method can act as a regularizer in noisy-label settings.
  • The framework is stated to generalize standard contrastive learning through multiple projection heads, so the same recipe applies whenever label structure has more than one level.

Reading between the lines

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

  • A test the paper leaves implicit: whether the superclass temperature of 0.5 remains optimal when chosen on a held-out validation split rather than the test set; if not, part of the low-data gain may be selection noise.
  • The multi-head objective can be read as a form of loss ensembling, where each head supplies a different hardness profile over negatives and the weighted sum reshapes the gradient contributions of easy versus hard pairs; this could be tested by ablating the head weights $\alpha_h$.
  • The global Jaccard-weighted head suggests a general recipe for ordinal or multi-label data: define a soft positive weight from label overlap and use it to modulate the contrastive loss, which could transfer to other multi-label domains such as medical coding or scene understanding.
  • The full-data improvement is modest, so the practically decisive claim is the low-data improvement; re-running the low-data experiments with multiple seeds and reported variance would clarify how much of the gap is stable.
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

5 major / 4 minor

Summary. The paper proposes Multi-level Supervised Contrastive Learning (MLCL), a supervised contrastive learning method that attaches multiple projection heads to an encoder, each trained with a temperature-scaled SupCon loss using a different similarity criterion: one per hierarchy level for hierarchical classification (CIFAR-100, DeepFashion) and one per label plus a global Jaccard-based head for multi-label classification (TripAdvisor, BeerAdvocate). The final loss is a weighted sum of per-head losses, optionally combined with cross-entropy. The authors report accuracy gains over SupCon and cross-entropy baselines, especially in low-data regimes, and provide temperature-limit and gradient analyses in the main text and appendix.

Significance. If the empirical claims hold, MLCL is a simple and general extension of SupCon that could be useful for hierarchical and multi-label tasks, with no change to the encoder and only a modest increase in parameters from the extra projection heads. The paper also provides a derived gradient formula and temperature-limit analyses, which are useful pedagogical contributions even if the τ→0 derivation has a technical flaw. However, the empirical case is currently under-supported because of test-set-based hyperparameter selection, missing error bars, absent multi-label contrastive baselines, and unreported text hyperparameters; these issues are load-bearing for the stated 'outperforms state-of-the-art' claim.

major comments (5)
  1. [Section 4.4, Figure 4] The superclass temperature τ2=0.5 is selected based on accuracy that appears to be test-set accuracy. Section 4.4 says the figure uses 'a subset of training samples,' but the y-axis is reported as accuracy and all other accuracy numbers in the paper are test accuracy. If the sweep is evaluated on the test set, the choice of τ2 is made by peeking at the test set, which biases the subsequently reported results. Please either perform the sweep on a validation split or carefully describe the split used; if the sweep is on a training subset, clarify and confirm the same value was chosen without test-set information.
  2. [Tables 2 and 3] No error bars, standard deviations, or number of seeds are reported for the CIFAR-100 and DeepFashion experiments. The full-data gains over SupCon are about 1 percentage point, and even the large small-data gains (9-10 points at 5K/10K) could be affected by run-to-run variance and by the test-set-based hyperparameter choice. Please report mean ± standard deviation over at least five seeds for all image experiments, including the baseline numbers, so the reader can assess whether the improvements are significant.
  3. [Table 4 and Section 2] The multi-label text experiments compare only against cross-entropy fine-tuning and an ablation without the global head. However, Section 2 cites multiple multi-label contrastive baselines (Zaigrajew and Zieba 2022; Zhang and Wu 2024; Sajedi et al. 2023; Małkiński and Mandziuk 2022), and the abstract and introduction claim outperformance over state-of-the-art contrastive methods. Without including at least one of these baselines under the same BERT encoder and training protocol, the multi-label claim is unsupported. Please add the most directly comparable multi-label contrastive baseline(s).
  4. [Section 4.2] The hyperparameters for the text experiments are incompletely reported. The Jaccard threshold t for the global head, the temperature for each per-label head and the global head, and the exact per-head α values are not all given; for TripAdvisor the stated α_i values (7×0.03 + 0.1 = 0.31) imply a cross-entropy weight of 0.69, not the claimed 0.7. Without these values, the TripAdvisor and BeerAdvocate results cannot be reproduced or checked for tuning on the test set. Please report all hyperparameters and any validation procedure used.
  5. [Section 3.5] The derivation of the τ→0+ limit in Equation (1) is mathematically incorrect. The step that drops the term log(1 + Σ_{a∈A(i)\p} exp((z_i·z_a - z_i·z_max)/τ)) is not valid: for any p that is not the maximizer, this term tends to log(1 + m) where m is the multiplicity of the maximum, not to zero. The qualitative conclusion that the loss becomes dominated by the hardest negative is consistent with the prior literature (Wang and Liu 2021), but the written derivation does not establish it. Please correct the derivation or cite a valid one.
minor comments (4)
  1. [Appendix, Sections 8.1-8.2 and Tables 5-6] The method is referred to as 'MSCL' in the appendix, while the main text uses 'MLCL'. Please unify the name throughout.
  2. [Equation (2) and (3)] Equation (2) states that Σ_h α_h = 1, but Equation (3) relaxes this by adding a cross-entropy term with weight (1 - Σ_h α_h). The constraint should be stated consistently, e.g., by writing Σ_h α_h ≤ 1 for the combined loss.
  3. [Section 3.4, global head gradient] The gradient analysis in the appendix assumes |P(i)| = |P(t)| for t ∈ P(i), which is true for class-based positive sets but not necessarily for the Jaccard-threshold global head, where the positive set is not guaranteed to be symmetric. The appendix should state this assumption and its limitation.
  4. [Figure 4 caption] The caption says 'Accuracy of MLCL as a function of the superclass projection head temperature' but does not specify whether this accuracy is on the training, validation, or test split. Please make this explicit.

Circularity Check

1 steps flagged · score 4.0 of 10

Test-set-based selection of τ2 makes the reported accuracy gains partially fitted; the multi-head architecture itself is an independent contribution.

  1. fitted input called prediction [Section 4.2 Implementation Details; Section 4.4 Ablation Study (Figure 4)]
    "The parameters for the first head are set as τ1 = 0.1 and α1 = 0 .5, and for the second head as τ2 = 0 .5, α2 = 0.5 ... Figure 4 illustrates the impact of the superclass projection head’s temperature on the final accuracy using a subset of training samples. Both excessively high and low temperatures negatively affect performance, while τ = 0.5 yields the best results."

    The value τ2=0.5 is not derived from the temperature analysis in Section 3.5; it is chosen because Figure 4 reports that this value yields the best final accuracy on a subset of training samples (evaluated, per the paper's protocol, on the test set). The same test accuracy is then reported in Tables 2 and 3 as evidence that MLCL outperforms SupCon. The reported gains—particularly the 9–10% improvements at 5K/10K samples—are therefore partly the result of selecting the hyperparameter on the evaluation metric, not an independent prediction of it. The theoretical temperature discussion is invoked post hoc to explain the selected value, but it does not predict τ2=0.5 a priori.

full rationale

The core MLCL formulation is a straightforward multi-head extension of SupCon: multiple projection heads each apply the standard supervised contrastive loss (Eq. 1) with a weighted sum (Eq. 2), plus cross-entropy for text (Eq. 3). This is not defined in terms of the target result, and no load-bearing self-citation or imported uniqueness theorem appears; the cited temperature behavior (Wang and Liu 2021) and Jaccard positive-pair idea (Zaigrajew and Zieba 2022) are external. The one genuinely circular element is the empirical validation loop: the superclass temperature is set to the value that maximizes final accuracy in Figure 4, and that same accuracy is then offered as evidence of improvement. This makes the reported margins partially fitted rather than parameter-free predictions. The text experiments additionally omit the Jaccard threshold t and per-head temperatures, preventing reconstruction, but that is a completeness issue rather than a circular reduction. On balance, the central architectural claim retains independent content, so the circularity score is moderate, not extreme.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The core contribution is a loss composed of per-head SupCon terms; the only mathematical content beyond standard contrastive limits is the Taylor-expansion analysis in the appendix, which rests on standard assumptions. The empirical claims depend on a large set of hand-chosen or unreported hyperparameters, notably temperatures, head weights, and the Jaccard threshold t. The method introduces no new entities.

free parameters (8)
  • subclass temperature tau1 = 0.1
    Chosen by hand in Section 4.2 for CIFAR-100 and DeepFashion; standard SupCon default.
  • superclass temperature tau2 = 0.5
    Selected via test-set sweep in Figure 4; no validation split is reported.
  • hierarchical head weights alpha1, alpha2 = 0.5, 0.5
    Chosen by hand in Section 4.2 with no sensitivity analysis.
  • per-label head weights alpha_l (TripAdvisor) = 0.03 per head, 0.1 for global head
    Chosen by hand in Section 4.2; no ablation on these values is reported.
  • per-label head weights alpha_l (BeerAdvocate) = 0.04 per head, 0.1 for global head
    Chosen by hand in Section 4.2; no sensitivity analysis.
  • cross-entropy weight = 0.7 for text experiments, 0.2 in noise ablation
    Hand chosen in Section 4.2 and Section 8.1.
  • Jaccard positive threshold t = not reported
    The global head loss Equation (4) depends on a threshold t for defining positive pairs, but no value is given anywhere in the paper.
  • temperatures for text heads = not reported
    No tau values are reported for the per-label or global heads in the TripAdvisor and BeerAdvocate experiments.
assumptions (5)
  • domain assumption Augmentations preserve the labels of the original sample.
    Section 3.2, Step 1 generates two augmentations of each sample and assumes both have the same label.
  • domain assumption Linear evaluation on frozen representations measures representation quality.
    The linear probing protocol follows SupCon and SimCLR and is used in Sections 3.3 and 4.2.
  • standard math Wang and Liu (2021) temperature analysis applies to each projection head's SupCon loss.
    The limits in Section 3.5 and the appendix use Taylor expansions and the hard-negative interpretation from prior work.
  • domain assumption Ground-truth hierarchy and multi-label annotations are correct.
    Superclass labels for CIFAR-100/DeepFashion and aspect ratings for TripAdvisor/BeerAdvocate are treated as reliable supervision.
  • domain assumption Pretrained BERT provides a suitable initialization for the text tasks.
    Section 4.2 fine-tunes a pretrained BERT encoder; no alternative initialization is tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-level Supervised Contrastive Learning." pith.science (2026). https://pith.science/paper/DUZ5DK77

@misc{pith2026250202202,
  author       = {Pith},
  title        = {Pith review of: Multi-level Supervised Contrastive Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DUZ5DK77}},
  note         = {Machine review of arXiv:2502.02202}
}
read the original abstract

Contrastive learning is a well-established paradigm in representation learning. The standard framework of contrastive learning minimizes the distance between "similar" instances and maximizes the distance between dissimilar ones in the projection space, disregarding the various aspects of similarity that can exist between two samples. Current methods rely on a single projection head, which fails to capture the full complexity of different aspects of a sample, leading to suboptimal performance, especially in scenarios with limited training data. In this paper, we present a novel supervised contrastive learning method in a unified framework called multilevel contrastive learning (MLCL), that can be applied to both multi-label and hierarchical classification tasks. The key strength of the proposed method is the ability to capture similarities between samples across different labels and/or hierarchies using multiple projection heads. Extensive experiments on text and image datasets demonstrate that the proposed approach outperforms state-of-the-art contrastive learning methods

Figures

Figures reproduced from arXiv: 2502.02202 by the authors.

Figure 1
Figure 1. An illustration of two projection spaces using different examples: a) TripAdvisor reviews and b) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Our proposed architecture with multiple projection heads for a) hierarchical classification and b) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The t-SNE visualization of samples from five different classes in the representation space. The SupCon representation indicates a lack of meaningful structure, as samples from class camel are positioned far from chimpanzee, kangaroo, and cattle compared to the bottle class. In contrast, our proposed method, MLCL, groups animal classes more closely in the representation space. roses, sunflowers, and tulips. In this p… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Accuracy of MLCL as a function of the superclass projection head temperature. Increasing the temperature to 0.5 reduces the network’s focus on hard negatives, leading to improved accuracy. method to capture label-specific similarities, exploit class hierarchies, and ac…
Figure 5
Figure 5. Figure 5: SupCon positions samples from the cattle class close to pear and sweet pepper. In contrast, MLCL clusters the fruit and vegetable classes more closely in the representation space while separating them from the cattle class [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: MLCL groups the flower classes more tightly in the representation space while separating them from [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: In SupCon, samples from the can class are positioned farther from bottle, cup, and bowl compared to the mouse class. In contrast, MLCL clusters the container classes more closely in the representation space. (a) SupCon (b) MLCL [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: MLCL aligns the furniture classes more closely in the representation space and distinct from the class [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 29 canonical work pages

  1. [1]

    D., and Buchwalter, W

    Bachman, P., Hjelm, R. D., and Buchwalter, W. (2019). Learning representations by maximizing mutual information across views. Advances in neural information processing systems , 32

    work page 2019

  2. [2]

    A., Dufumier, B., Tartaglione, E., Grangetto, M., and Gori, P

    Barbano, C. A., Dufumier, B., Tartaglione, E., Grangetto, M., and Gori, P. (2023). Unbiased supervised contrastive learning. In The Eleventh International Conference on Learning Representations

    work page 2023

  3. [3]

    Y., Narayan, A., Zhang, M., Song, Z., Fatahalian, K., and R \'e , C

    Chen, M., Fu, D. Y., Narayan, A., Zhang, M., Song, Z., Fatahalian, K., and R \'e , C. (2022). Perfectly balanced: Improving transfer and robustness of supervised contrastive learning. In International Conference on Machine Learning , pages 3090--3122. PMLR

    work page 2022

  4. [4]

    Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. (2020a). A simple framework for contrastive learning of visual representations. In International conference on machine learning , pages 1597--1607. PMLR

    work page 2020

  5. [5]

    Chen, X., Fan, H., Girshick, R., and He, K. (2020b). Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297

    arXiv 2020

  6. [6]

    Chuang, C.-Y., Robinson, J., Lin, Y.-C., Torralba, A., and Jegelka, S. (2020). Debiased contrastive learning. Advances in neural information processing systems , 33:8765--8775

    work page 2020

  7. [7]

    Ciortan, M., Dupuis, R., and Peel, T. (2021). A framework using contrastive learning for classification with noisy labels. Data , 6(6):61

    work page 2021

  8. [9]

    D., Zhao, E., Phung, D., and Cai, J

    Dao, S. D., Zhao, E., Phung, D., and Cai, J. (2021b). Multi-label image classification with contrastive learning. arXiv preprint arXiv:2107.11626

    arXiv 2021

Show all 39 references
  1. [10]

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. (2009). Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition , pages 248--255. Ieee

    2009

  2. [11]

    Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. (2018). Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805

    2018 arXiv

  3. [12]

    and Lan, A

    Ghosh, A. and Lan, A. (2021). Contrastive learning improves model robustness under label noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 2703--2708

    2021

  4. [13]

    Grill, J.-B., Strub, F., Altch \'e , F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al. (2020). Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing syste...

    2020

  5. [14]

    Gunel, B., Du, J., Conneau, A., and Stoyanov, V. (2020). Supervised contrastive learning for pre-trained language model fine-tuning. In International Conference on Learning Representations

    2020

  6. [15]

    He, K., Fan, H., Wu, Y., Xie, S., and Girshick, R. (2020). Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 9729--9738

    2020

  7. [16]

    He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 770--778

    2016

  8. [17]

    Henaff, O. (2020). Data-efficient image recognition with contrastive predictive coding. In International conference on machine learning , pages 4182--4192. PMLR

    2020

  9. [18]

    B., Pion, N., Weinzaepfel, P., and Larlus, D

    Kalantidis, Y., Sariyildiz, M. B., Pion, N., Weinzaepfel, P., and Larlus, D. (2020). Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems , 33:21798--21809

    2020

  10. [19]

    Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., and Krishnan, D. (2020). Supervised contrastive learning. Advances in neural information processing systems , 33:18661--18673

    2020

  11. [20]

    Kingma, D. P. and Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    2014 arXiv

  12. [21]

    Krizhevsky, A. (2009). Learning multiple layers of features from tiny images. Technical Report

    2009

  13. [22]

    and Garnot, V

    Landrieu, L. and Garnot, V. S. F. (2021). Leveraging class hierarchies with metric-guided prototype learning. In British Machine Vision Conference (BMVC)

    2021

  14. [23]

    Li, J., Xiong, C., and Hoi, S. C. (2021a). Learning from noisy data with robust representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 9485--9494

    2021

  15. [24]

    Li, J., Xiong, C., and Hoi, S. C. (2021b). Mopro: Webly supervised learning with momentum prototypes. Proceedings of the International Conference on Learning Representations

    2021

  16. [25]

    S., Indyk, P., and Katabi, D

    Li, T., Cao, P., Yuan, Y., Fan, L., Yang, Y., Feris, R. S., Indyk, P., and Katabi, D. (2022). Targeted supervised contrastive learning for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 6918--6928

    2022

  17. [26]

    Liu, Z., Luo, P., Qiu, S., Wang, X., and Tang, X. (2016). Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2016

  18. [27]

    and Ma \'n dziuk, J

    Ma ki \'n ski, M. and Ma \'n dziuk, J. (2022). Multi-label contrastive learning for abstract visual reasoning. IEEE Transactions on Neural Networks and Learning Systems

    2022

  19. [28]

    McAuley, J., Leskovec, J., and Jurafsky, D. (2012). Learning attitudes and attributes from multi-aspect reviews. In 2012 IEEE 12th International Conference on Data Mining , pages 1020--1025. IEEE

    2012

  20. [29]

    Ruder, S. (2016). An overview of gradient descent optimization algorithms. arXiv preprint arXiv:1609.04747

    2016 arXiv

  21. [30]

    N., and Hosseini, M

    Sajedi, A., Khaki, S., Plataniotis, K. N., and Hosseini, M. S. (2023). End-to-end supervised multilabel contrastive learning. arXiv preprint arXiv:2307.03967

    2023 arXiv

  22. [31]

    and Hinton, G

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

    2008

  23. [32]

    and Liu, H

    Wang, F. and Liu, H. (2021). Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 2495--2504

    2021

  24. [33]

    Wang, H., Lu, Y., and Zhai, C. (2010). Latent aspect rating analysis on review text data: a rating regression approach. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining , pages 783--792

    2010

  25. [34]

    Wang, L., Koniusz, P., Gedeon, T., and Zheng, L. (2023). Adaptive multi-head contrastive learning. arXiv preprint arXiv:2310.05615

    2023 arXiv

  26. [35]

    X., and Lin, D

    Wu, Z., Xiong, Y., Yu, S. X., and Lin, D. (2018). Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 3733--3742

    2018

  27. [36]

    Xue, Y., Whitecross, K., and Mirzasoleiman, B. (2022). Investigating why contrastive learning benefits robustness against label noise. In International Conference on Machine Learning , pages 24851--24871. PMLR

    2022

  28. [37]

    I., and Wang, B

    Yi, L., Liu, S., She, Q., McLeod, A. I., and Wang, B. (2022). On learning contrastive representations for learning with noisy labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 16682--16691

    2022

  29. [38]

    and Zieba, M

    Zaigrajew, V. and Zieba, M. (2022). Contrastive learning for multi-label classification. NeurIPS 2022 Workshop: Self-Supervised Learning - Theory and Practice

    2022

  30. [39]

    and Wu, M

    Zhang, P. and Wu, M. (2024). Multi-label supervised contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 16786--16793

    2024

  31. [40]

    Zhang, S., Xu, R., Xiong, C., and Ramaiah, C. (2022). Use all the labels: A hierarchical multi-label contrastive learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 16660--16669

    2022

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.