Pith. sign in

REVIEW 3 major objections 3 minor 47 references

The paper claims that three leading self-supervised contrastive learning methods—BYOL, Barlow Twins, and SwAV—are special cases of a single two-part framework, and that a new label-free calibration component improves them all.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A single framework unifies BYOL, Barlow Twins, and SwAV, plus a plug-in calibration method, ADC, that improves learned representations by preserving input-space distances.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Coherent unification and a new plug-in, but the corrupt text makes this unverifiable; deserves peer review only in readable form. the 3 major comments →

arxiv 2508.13596 v1 pith:EUVWCECI submitted 2025-08-19 cs.LG

A Generalized Learning Framework for Self-Supervised Contrastive Learning

Lingyu Si , Jingyao Wang , Wenwen Qiang This is my paper
classification cs.LG
keywords self-supervised contrastive learninggeneralized learning frameworkBYOLBarlow TwinsSwAVAdaptive Distribution Calibrationrepresentation learningfeature-space geometry
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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's central claim is that three well-known self-supervised contrastive learning methods—BYOL, Barlow Twins, and SwAV—are not independent designs but specific instances of one two-part template. The template, called the Generalized Learning Framework (GLF), splits any such method into an aligning part, which pulls augmented views of the same sample together, and a constraining part, which shapes the remaining feature distribution. From theory and experiments the paper derives two properties the constraining part should produce: intra-class compactness and inter-class separability. Because self-supervised learning has no class labels, the paper introduces Adaptive Distribution Calibration (ADC), a plug-and-play component that makes samples closer to an anchor in the original input space also closer in feature space, and farther samples farther. If the claim holds, the field gains a unified vocabulary for comparing self-supervised methods and a drop-in improvement for their representations.

Core claim

The Generalized Learning Framework (GLF) decomposes a self-supervised contrastive objective into an aligning term and a constraining term. The paper shows that BYOL corresponds to one choice of constraining part, Barlow Twins to another, and SwAV to a third, so existing methods differ mainly in how they constrain the feature distribution rather than in the fundamental learning task. Its design analysis identifies intra-class compactness and inter-class separability as the two properties a good constraining part should enforce. Adaptive Distribution Calibration (ADC) implements these properties without labels by iteratively capturing the dynamic relationship between an anchor and other sample

What carries the argument

The central object is the Generalized Learning Framework (GLF), which formalizes self-supervised contrastive learning as aligning positive views plus a constraining term on the feature distribution. The load-bearing component inside it is Adaptive Distribution Calibration (ADC), which uses the distance of other samples from the anchor in the original input space as a label-free signal to reshape their positions in feature space, pulling near samples in and pushing far samples out. ADC is the mechanism that translates the two design insights—intra-class compactness and inter-class separability—into an actual training signal.

Load-bearing premise

The method assumes that distance from the anchor in the original input space is a reliable proxy for semantic or class similarity, so that pulling feature-space geometry into line with input-space geometry helps rather than hurts.

What would settle it

Take a dataset with class labels, train a self-supervised method with ADC, and measure, for many anchors, whether same-class probability rises with input-space proximity to the anchor; if near-in-input samples are often from different classes, the calibration premise is violated. Then rerun the method with ADC's distance ordering randomly permuted; if performance does not drop, the claimed mechanism, not mere added regularization, is responsible.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If GLF is correct, BYOL, Barlow Twins, and SwAV can be analyzed, compared, and improved through their constraining part instead of treated as separate algorithms.
  • ADC is plug-and-play: adding it to each of the three methods should improve their representations without changing their aligning objectives or requiring labels.
  • The two design targets, intra-class compactness and inter-class separability, give a quantitative way to diagnose why one self-supervised method beats another on downstream tasks.
  • Improvements in the constraining part can be shared across methods, so future work can focus on designing better constraining parts rather than new full pipelines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension of the paper's logic: the benefit of ADC should grow as input-space distance becomes more semantically informative, and should vanish when augmented views make near neighbors cross class boundaries; this could be checked by corrupting or permuting the distance ordering.
  • The same aligning/constraining split likely extends beyond BYOL, Barlow Twins, and SwAV to other self-supervised contrastive designs, including instance-discrimination methods that use negative samples; the paper does not make this extension.
  • The monotone-distance rule suggests a practical diagnostic for representation quality: rank correlation between input-space distance from an anchor and feature-space distance, measured across many anchors, should be high after training with ADC.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The manuscript (arXiv:2508.13596, cs.LG) proposes a Generalized Learning Framework (GLF) that splits self-supervised contrastive learning into an aligning part and a constraining part, claims that BYOL, Barlow Twins, and SwAV are special cases of GLF, and introduces Adaptive Distribution Calibration (ADC), a plug-and-play module that encourages feature-space distances to mirror input-space distances near anchor points. The abstract asserts that both theoretical analysis and empirical evaluation demonstrate ADC's superiority. However, the supplied full text is almost entirely unreadable mojibake: no equation, theorem, or experimental table can be verified. The document header also shows a different arXiv ID (2508.13593v2 [cs.IT]) than the paper under review.

Significance. If the unification claim holds, framing BYOL, Barlow Twins, and SwAV under a single aligning/constraining decomposition is a useful conceptual contribution, and ADC could be an appealingly simple add-on to existing methods. However, with the manuscript in its present form, nothing can be checked: the derivations, proofs, and numerical results are absent to the reader. Moreover, the core mechanism of ADC -- using input-space distance as a proxy for semantic similarity -- is not validated in any readable portion. The paper needs a clean, complete version and targeted experiments or theoretical conditions for that proxy before its claims can be assessed.

major comments (3)
  1. [Full text (all sections)] The body of this submission is an unreadable sequence of mojibake; no equation, theorem, or experimental table can be parsed. Consequently, the abstract's central assertion -- 'Both the theoretical analysis and the empirical evaluation demonstrate the superiority of ADC' -- is unsupported in the provided manuscript. This is a load-bearing defect, not a presentation issue: the reader cannot verify that BYOL/Barlow/SwAV reduce to GLF, nor that ADC improves representation quality. A clean, correctly encoded PDF with complete proofs, experimental settings, and tables is required before substantive review can proceed.
  2. [Abstract / ADC definition] ADC is defined by the rule that samples 'near or far from the anchor point in the original input space' are made correspondingly near or far in feature space. On standard vision data, raw pixel or augmented-view distance is frequently dominated by low-level factors (background, color, pose) rather than class identity. Enforcing monotonicity with respect to this proxy can therefore push apart same-class instances and pull together different-class instances with similar low-level statistics. The manuscript contains no readable analysis that states conditions under which input-space distance is a reliable semantic proxy. Reported intra-class compactness / inter-class separability improvements would be partially circular, because ADC is constructed to make those quantities move in the claimed direction. The non-circular evidence would be downstream classification/transfer performance, side-b
  3. [Document header / arXiv ID] The manuscript is titled and abstracted as arXiv:2508.13596 (cs.LG), but the only readable header in the full text is 'arXiv:2508.13593v2 [cs.IT] 8 Jan 2026', with an inconsistent version/date line. This mismatch prevents verification of the submission's provenance and indicates the uploaded file is defective. The authors must correct the paper ID, version metadata, and re-supply the file.
minor comments (3)
  1. [Abstract] The term 'plug-and-play' is not qualified in the readable text: ADC appears to introduce at least a calibration weight and a neighbor/scale threshold. These free parameters should be named and their sensitivity discussed explicitly.
  2. [Section 1 (Introduction, unreadable portion)] The three method names BYOL, Barlow Twins, and SwAV appear, but their loss definitions are not visible in the corrupted text. A clean version should include explicit equations and a table showing how each method instantiates GLF's constraining part.
  3. [General] No code, data, or reproducibility details are legible. The resubmission should state the experimental setups (datasets, architectures, epochs, batch size, evaluation protocol, number of seeds) and provide error bars for all reported comparisons.

Circularity Check

1 steps flagged

GLF unification is a post-hoc renaming of existing methods; ADC's superiority claim is not shown to be independent of its defining objective.

specific steps
  1. renaming known result [Abstract, contribution 1 (GLF)]
    "we generalize the standard SSCL methods to a Generalized Learning Framework (GLF) consisting of two parts: the aligning part and the constraining part. We analyze three existing SSCL methods: BYOL, Barlow Twins, and SwAV, and show that they can be unified under GLF with different choices of the constraining part."

    The framework is defined after the fact: 'aligning' and 'constraining' are broad placeholders, and the 'different choices' are selected to match each existing method's loss. The unification is therefore true by construction of the category labels; it does not derive any property of BYOL/Barlow/SwAV from shared first principles. Presenting this decomposition as a 'Generalized Learning Framework' is a renaming of known losses, not a result that could fail.

full rationale

The only clearly identifiable circular step is the GLF unification, which is a post-hoc decomposition of existing methods into aligning and constraining components. This is a renaming/unification-by-construction rather than a derivation that could be falsified. The ADC method's definition explicitly enforces input-space distance order in feature space; however, the abstract does not specify the evaluation metric, so I cannot conclude without speculation that the 'superiority' claim reduces to that objective. The reliance on input-space distance as a semantic proxy is an unvalidated assumption and a correctness risk, but not itself a demonstrated circularity. No load-bearing self-citation is visible in the readable portions of the manuscript. The corrupted full text prevents verification of the theorems and experiments, which is a support gap rather than a circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim depends on ADC's input-space distance proxy and on calibration hyperparameters. No new physical entities are introduced.

free parameters (2)
  • ADC calibration weight = not stated in abstract
    Any combination of the aligning loss, constraining loss, and ADC term requires a weighting hyperparameter; its choice affects the reported trade-off between compactness and separability.
  • Input-space distance threshold or neighbor scale = not stated in abstract
    Deciding what counts as near versus far from the anchor requires a scale or neighbor count; the abstract does not report how this is set.
axioms (2)
  • domain assumption Input-space proximity after augmentation is a valid proxy for semantic or class similarity
    ADC moves samples based on near/far relations in the original input space. If this proxy fails, the calibration cannot deliver true intra-class compactness or inter-class separability.
  • domain assumption The theoretical analysis relies on smoothness or concentration assumptions about the feature map
    The abstract claims a theoretical analysis, but the assumptions are not legible in the corrupted full text. Such analyses typically require Lipschitz, bounded-variance, or augmentation-robustness conditions.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of A Generalized Learning Framework for Self-Supervised Contrastive Learning." pith.science (2026). https://pith.science/paper/EUVWCECI

@misc{pith2026250813596,
  author       = {Pith},
  title        = {Pith review of: A Generalized Learning Framework for Self-Supervised Contrastive Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EUVWCECI}},
  note         = {Machine review of arXiv:2508.13596}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Self-supervised contrastive learning (SSCL) has recently demonstrated superiority in multiple downstream tasks. In this paper, we generalize the standard SSCL methods to a Generalized Learning Framework (GLF) consisting of two parts: the aligning part and the constraining part. We analyze three existing SSCL methods: BYOL, Barlow Twins, and SwAV, and show that they can be unified under GLF with different choices of the constraining part. We further propose empirical and theoretical analyses providing two insights into designing the constraining part of GLF: intra-class compactness and inter-class separability, which measure how well the feature space preserves the class information of the inputs. However, since SSCL can not use labels, it is challenging to design a constraining part that satisfies these properties. To address this issue, we consider inducing intra-class compactness and inter-class separability by iteratively capturing the dynamic relationship between anchor and other samples and propose a plug-and-play method called Adaptive Distribution Calibration (ADC) to ensure that samples that are near or far from the anchor point in the original input space are closer or further away from the anchor point in the feature space. Both the theoretical analysis and the empirical evaluation demonstrate the superiority of ADC.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 29 canonical work pages · 1 internal anchor

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Arora, S.; Khandeparkar, H.; Khodak, M.; Plevrakis, O.; and Saunshi, N. 2019. A theoretical analysis of contrastive unsupervised representation learning. arXiv preprint arXiv:1902.09229

    Pith/arXiv arXiv 2019

  4. [4]

    V.; Hardin, D

    Borodachov, S. V.; Hardin, D. P.; and Saff, E. B. 2019. Discrete energy on rectifiable sets. Springer

    work page 2019

  5. [5]

    Caron, M.; Misra, I.; Mairal, J.; Goyal, P.; Bojanowski, P.; and Joulin, A. 2020. Unsupervised Learning of Visual Features by Contrasting Cluster Assignments. CoRR, abs/2006.09882

    Pith/arXiv arXiv 2020

  6. [6]

    Chen, S.; Niu, G.; Gong, C.; Li, J.; Yang, J.; and Sugiyama, M. 2021. Large-margin contrastive learning with distance polarization regularizer. In International Conference on Machine Learning, 1673--1683. PMLR

    work page 2021

  7. [7]

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

    work page 2020

  8. [8]

    Chen, T.; Luo, C.; and Li, L. 2021. Intriguing properties of contrastive losses. Advances in Neural Information Processing Systems, 34: 11834--11845

    work page 2021

  9. [9]

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

    Pith/arXiv arXiv 2020

  10. [10]

    Chen, X.; and He, K. 2021. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 15750--15758

    work page 2021

  11. [11]

    Chen, X.; Xie, S.; and He, K. 2021. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 9640--9649

    work page 2021

  12. [12]

    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

  13. [13]

    Coates, A.; Ng, A.; and Lee, H. 2011. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, 215--223. JMLR Workshop and Conference Proceedings

    work page 2011

  14. [14]

    Cohn, H.; and Kumar, A. 2007. Universally optimal distribution of points on spheres. Journal of the American Mathematical Society, 20(1): 99--148

    work page 2007

  15. [15]

    Ermolov, A.; Siarohin, A.; Sangineto, E.; and Sebe, N. 2021. Whitening for self-supervised representation learning. In International Conference on Machine Learning, 3015--3024. PMLR

    work page 2021

  16. [16]

    K.; Winn, J.; and Zisserman, A

    Everingham, M.; Van Gool, L.; Williams, C. K.; Winn, J.; and Zisserman, A. 2010. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2): 303--338

    work page 2010

  17. [17]

    Fang, Y.; Sun, Q.; Wang, X.; Huang, T.; Wang, X.; and Cao, Y. 2023. EVA-02: A Visual Representation for Neon Genesis. CoRR, abs/2303.11331

    Pith/arXiv arXiv 2023

  18. [18]

    Ferguson, T. S. 1962. A representation of the symmetric bivariate Cauchy distribution. The Annals of Mathematical Statistics, 33(4): 1256--1266

    work page 1962

  19. [19]

    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 a . Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33: 21271--21284

    work page 2020

  20. [20]

    H.; Buchatskaya, E.; Doersch, C.; Pires, B

    Grill, J.-B.; Strub, F.; Altch \'e , F.; Tallec, C.; Richemond, P. H.; Buchatskaya, E.; Doersch, C.; Pires, B. A.; Guo, Z. D.; Azar, M. G.; et al. 2020 b . Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733

    Pith/arXiv arXiv 2020

  21. [21]

    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, 9729--9738

    2020

  22. [22]

    D.; Fedorov, A.; Lavoie-Marchildon, S.; Grewal, K.; Bachman, P.; Trischler, A.; and Bengio, Y

    Hjelm, R. D.; Fedorov, A.; Lavoie-Marchildon, S.; Grewal, K.; Bachman, P.; Trischler, A.; and Bengio, Y. 2018. Learning deep representations by mutual information estimation and maximization. arXiv preprint arXiv:1808.06670

    Pith/arXiv arXiv 2018

  23. [23]

    R.; Zadeh, M

    Jaiswal, A.; Babu, A. R.; Zadeh, M. Z.; Banerjee, D.; and Makedon, F. 2020. A survey on contrastive self-supervised learning. Technologies, 9(1): 2

    work page 2020

  24. [24]

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

    2009

  25. [25]

    Krizhevsky, A.; Sutskever, I.; and Hinton, G. E. 2012. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25: 1097--1105

    work page 2012

  26. [26]

    Le, Y.; and Yang, X. 2015. Tiny Imagenet Visual Recognition Challenge. CS 231N, 7(7): 3

    work page 2015

  27. [27]

    Li, J.; Qiang, W.; Zheng, C.; Su, B.; and Xiong, H. 2022. MetAug: Contrastive Learning via Meta Feature Augmentation. International Conference on Machine Learning

    work page 2022

  28. [28]

    Li, J.; Zhou, P.; Xiong, C.; and Hoi, S. C. 2020. Prototypical contrastive learning of unsupervised representations. arXiv preprint arXiv:2005.04966

    Pith/arXiv arXiv 2020

  29. [29]

    Lin, T.-Y.; Maire, M.; Belongie, S.; Hays, J.; Perona, P.; Ramanan, D.; Doll \'a r, P.; and Zitnick, C. L. 2014. Microsoft coco: Common objects in context. In European conference on computer vision, 740--755. Springer

    2014

  30. [30]

    Liu, X.; Wang, Z.; Li, Y.; and Wang, S. 2022. Self-Supervised Learning via Maximum Entropy Coding. arXiv preprint arXiv:2210.11464

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  31. [31]

    Molenberghs, G.; and Lesaffre, E. 1997. Non-linear integral equations to approximate bivariate densities with given marginals and dependence function. Statistica Sinica, 713--738

    work page 1997

  32. [32]

    W.; Tian, G.-L.; and Tang, M.-L

    Ng, K. W.; Tian, G.-L.; and Tang, M.-L. 2011. Dirichlet and related distributions: Theory, methods and applications

    work page 2011

  33. [33]

    Oord, A. v. d.; Li, Y.; and Vinyals, O. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748

    Pith/arXiv arXiv 2018

  34. [34]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I. 2021 a . Learning Transferable Visual Models From Natural Language Supervision. In Meila, M.; and Zhang, T., eds., Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July...

    work page 2021

  35. [35]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021 b . Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 8748--8763. PMLR

    work page 2021

  36. [36]

    S.; and Kumar, M

    Ram, A.; Jalal, S.; Jalal, A. S.; and Kumar, M. 2010. A density based algorithm for discovering density varied clusters in large spatial databases. International Journal of Computer Applications, 3(6): 1--4

    work page 2010

  37. [37]

    Robinson, J.; Chuang, C.-Y.; Sra, S.; and Jegelka, S. 2020. Contrastive learning with hard negative samples. arXiv preprint arXiv:2010.04592

    Pith/arXiv arXiv 2020

  38. [38]

    Si, L.; Qiang, W.; Li, J.; Xu, F.; and Sun, F. 2022. Multi-view representation learning from local consistency and global alignment. Neurocomputing, 501: 727--740

    work page 2022

  39. [39]

    Tian, Y.; Krishnan, D.; and Isola, P. 2020 a . Contrastive multiview coding. In European conference on computer vision, 776--794. Springer

    work page 2020

  40. [40]

    Tian, Y.; Krishnan, D.; and Isola, P. 2020 b . Contrastive Multiview Coding . arXiv:1906.05849

    Pith/arXiv arXiv 2020

  41. [41]

    Tomasev, N.; Bica, I.; McWilliams, B.; Buesing, L.; Pascanu, R.; Blundell, C.; and Mitrovic, J. 2022. Pushing the limits of self-supervised ResNets: Can we outperform supervised learning without labels on ImageNet? arXiv preprint arXiv:2201.05119

    Pith/arXiv arXiv 2022

  42. [42]

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

    2008

  43. [43]

    Wang, J.; Mou, L.; Ma, L.; Huang, T.; and Gao, W. 2023. AMSA: Adaptive multimodal learning for sentiment analysis. ACM Transactions on Multimedia Computing, Communications and Applications, 19(3s): 1--21

    work page 2023

  44. [44]

    Wang, J.; Mou, L.; Zheng, C.; and Gao, W. 2024. Image-based freeform handwriting authentication with energy-oriented self-supervised learning. IEEE Transactions on Multimedia

    work page 2024

  45. [45]

    Wang, T.; and Isola, P. 2020. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, 9929--9939. PMLR

    work page 2020

  46. [46]

    Zbontar, J.; Jing, L.; Misra, I.; LeCun, Y.; and Deny, S. 2021. Barlow twins: Self-supervised learning via redundancy reduction. arXiv preprint arXiv:2103.03230

    Pith/arXiv arXiv 2021

  47. [47]

    Zheng, M.; You, S.; Wang, F.; Qian, C.; Zhang, C.; Wang, X.; and Xu, C. 2021. Ressl: Relational self-supervised learning with weak augmentation. Advances in Neural Information Processing Systems, 34: 2543--2555

    work page 2021

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.