Pith. sign in

REVIEW 4 major objections 4 minor 50 references

An Inclusive Theoretical Framework of Robust Supervised Contrastive Loss against Label Noise

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper identifies a single condition that decides whether a supervised contrastive loss survives symmetric label noise, proves InfoNCE fails it, and constructs SymNCE that passes it.

desk verdict A genuinely useful framework with a clever SymNCE construction, but the central robustness proof has a log-K normalization error and only holds in an infinite limit, so 'provably robust' overreaches. read the letter →

arxiv 2501.01130 v1 pith:FQUECXHS submitted 2025-01-02 cs.LG

Jingyi Cui , Yi-Ge Zhang , Hengyu Liu , Yisen Wang This is my paper · ORCID
classification cs.LG
keywords supervisedcontrastivelearninglabelnoiserobustlossfunctionInfoNCEsymmetricadditionalrisktolerancerepresentation
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 tries to turn robust supervised contrastive learning under label noise from a collection of heuristics into a checkable criterion. It shows the risk of any contrastive loss under symmetric label noise can be split into the clean risk plus an additional risk term, and that if this extra term is a constant independent of the learned representation, minimizing the noisy risk also minimizes the clean risk. The popular InfoNCE loss fails this test, so the paper builds Symmetric InfoNCE (SymNCE) by adding a reverse term, RevNCE, that makes the additional risk constant. If the theory is right, practitioners can screen any pairwise contrastive loss for label-noise robustness before training, and previously heuristic fixes such as nearest-neighbour sample selection receive a principled explanation.

What carries the argument

The load-bearing object is the additional risk ΔR(L; f): the contrastive risk evaluated when anchor and positive samples are drawn independently from the data distribution rather than from the same latent class. Theorem III.4 proves that, under symmetric label noise and class balance, the noisy contrastive risk equals (1 − Cγ/(C−1))² R(L; f) plus Cγ/(C−1)(2 − Cγ/(C−1)) ΔR(L; f), so robustness reduces to checking whether the additional risk is a constant independent of f. The proposed SymNCE loss pairs InfoNCE with RevNCE, a loss that swaps the roles of the positive and negative sums; its defining property is that ΔRlim(RevNCE; f) = −ΔRlim(InfoNCE; f) + log K, making the combined loss satisfy ΔRlim(SymNCE; f) = log K.

What would settle it

Run SymNCE on a synthetic two-class problem with class proportions 0.9/0.1 under symmetric 40% label noise and check whether the minimizer of the noisy contrastive risk still equals the minimizer of the clean risk; if the learned representation or downstream accuracy shifts measurably when class imbalance is introduced, the decomposition carrying the robustness proof has broken. A direct check would be to compute ΔRlim(SymNCE; f) on that imbalanced distribution for two different representation functions and show that it is not constant.

Watch

Extended reading notes

Core claim

The paper's central claim is that, for class-balanced data under symmetric label noise with noise rate γ < (C−1)/C, a contrastive loss is noise tolerant whenever its additional risk ΔR(L; f) is a constant independent of the representation function f. The additional risk is defined by drawing the anchor, positive, and negative samples independently from the data distribution, so it measures how the loss behaves when the positive pair carries no real class information. Under these assumptions the noisy risk decomposes as an affine combination of the clean risk R(L; f) and ΔR(L; f), so a constant ΔR makes minimizing the noisy risk equivalent to minimizing the clean risk. The paper then shows InfoNCE is not robust because its limiting additional risk depends on f, and constructs SymNCE by adding a reverse loss RevNCE, whose limiting additional risk cancels the non-constant part and leaves ΔR = log K.

Load-bearing premise

The whole guarantee assumes the training classes are balanced, πi = 1/C for every class; if class proportions are unequal, the noisy risk no longer decomposes as a constant-weighted sum of the clean risk and the additional risk, and the theorem's proof no longer applies.

Editorial extensions

If this is right

  • InfoNCE, the default supervised contrastive loss, is not robust to symmetric label noise, so using it directly can overfit corrupted positive pairs at high noise rates.
  • SymNCE inherits InfoNCE's alignment behaviour while adding a reverse term, so it can replace the loss in existing supervised contrastive pipelines without changing the rest of the training setup.
  • Nearest-neighbour positive selection works because restricting positives to high-similarity neighbours pushes the additional risk toward a constant, making the loss approximately noise tolerant.
  • RINCE is robust exactly when λ = 1/(K+1), a parameter choice the framework derives rather than tunes.
  • The criterion gives a general design rule: to make a contrastive loss robust, make its additional risk independent of the representation function.

Reading between the lines

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

  • The constant-ΔR condition is sufficient but may not be necessary; a loss whose additional risk varies only mildly with f could still be nearly noise tolerant, and the M,K→∞ limit hides finite-batch effects that matter in real training.
  • The class-balanced assumption enters the proof where a(i) is set to 1/C, so a weighted reformulation of the noisy risk might extend the same guarantee to imbalanced datasets.
  • The reverse-term construction is a template: any contrastive loss with non-constant ΔR could be made robust by adding a suitably reversed version, provided the reverse term preserves the positive-alignment and negative-repulsion behaviour.
  • A testable extension would be to replace the uniform πi in the theorem with empirical class frequencies on imbalanced noisy data and check whether the predicted residual noise bias matches the observed gap in accuracy.
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.

Referee Report

4 major / 4 minor

Summary. The paper develops a theoretical framework for supervised contrastive losses under symmetric label noise. Its main theoretical result is a decomposition of the noisy contrastive risk into a clean-risk term and an additional-risk term (Lemma III.3 and Theorem III.4), leading to a sufficient robustness condition: a contrastive loss is noise tolerant if the additional risk is constant over representation functions (Theorem III.5). On this basis the paper argues that InfoNCE is non-robust, proposes a symmetrized variant SymNCE obtained by adding a Reverse InfoNCE (RevNCE) term, and claims that SymNCE is provably robust. It further presents the framework as an inclusive explanation of nearest-neighbor sample selection and of the RINCE loss, and reports experiments on CIFAR, Tiny ImageNet, and Clothing1M showing favorable accuracy.

Significance. If the robustness guarantee for SymNCE were rigorously established, the paper would provide one of the first general criteria for verifying noise tolerance of arbitrary contrastive losses, together with a constructive recipe for turning a non-robust loss into a robust one. The algebraic decomposition in Theorem III.4 is a useful contribution on its own, and the derived parameter choice λ = 1/(K + 1) for RINCE is a crisp, falsifiable byproduct that the paper tests experimentally. The empirical results are encouraging and suggest that the SymNCE loss has practical value even if the theoretical claim needs repair. However, the advertised 'provably robust' statement is not currently backed by a correct proof, and the supporting NN-selection explanation is largely asserted rather than derived.

major comments (4)
  1. The normalization of ΔRlim is internally inconsistent. Eq. (22) defines ΔRlim(LInfoNCE; f) as lim_{M,K→∞}(ΔR(LInfoNCE; f) − log K), while Theorem IV.1 defines ΔRlim(LInfoNCE; f) as lim_{M,K→∞} ΔR(LInfoNCE; f) without the log K shift, and then concludes ΔRlim(LRevNCE; f) = −ΔRlim(LInfoNCE; f) + log K. The two definitions cannot both be correct, and the conclusion is false under either reading. For the constant encoder f ≡ 1, direct computation gives ΔR(LInfoNCE) = log(K + 1), ΔR(LRevNCE) = 0, and hence ΔR(LSymNCE) = log(K + 1); therefore lim_{K→∞}(ΔR(LSymNCE) − log K) = 0, not log K as claimed in Eq. (27). With the normalized definition of ΔRlim used in Eq. (22), the correct identity is ΔRlim(LRevNCE) = −ΔRlim(LInfoNCE); the extra log K term should not appear. This error directly undermines the proof that SymNCE satisfies the constant-additional-risk condition.
  2. Theorem III.5 requires the exact finite-batch additional risk ΔR(L; f) to be constant in f, but the derivation preceding the noise-tolerance claim for SymNCE only establishes a statement about the M,K → ∞ limit. No finite-sample bound, uniform-convergence argument, or limit version of Theorem III.5 is provided, so the gap is load-bearing. For finite batches the constancy actually fails: when K = M = 1, ΔR(LSymNCE; f) = E[log(1 + exp(f(x)⊤(f(x⁻) − f(x⁺))))], which generally depends on f. Consequently, the conclusion that 'LSymNCE is noise tolerant' is not justified for the finite-batch loss used in the experiments, and the phrase 'provably robust' overstates what the paper proves.
  3. The claim that InfoNCE is non-robust is also based on the limit object rather than on the exact condition of Theorem III.5. The paper shows that ΔRlim(LInfoNCE; f) is non-constant in f, but Theorem III.5 concerns exact constancy of the finite additional risk. To prove non-robustness under the paper's own criterion, one should show that ΔR(LInfoNCE; f) is not constant for finite M and K; non-constancy of the limit does not logically imply non-constancy of the finite quantities. This is probably true and could be shown by a simple counterexample, but the manuscript currently asserts it rather than proves it.
  4. The NN-selection explanation is asserted rather than proved. The text states that for a given f one can select a threshold t so that lim_{M,K→∞}(ΔR(LInfoNCE-NN; f, t) − log K) = 0, and concludes that LInfoNCE-NN is noise tolerant. No construction of t is given, t would depend on the unknown encoder f, and the exact equality to zero is not shown to be achievable. As in the SymNCE argument, the limit statement is not sufficient for Theorem III.5, which requires exact constancy. This section should be reframed as a heuristic explanation or supported by a rigorous argument with explicit conditions on t and f.
minor comments (4)
  1. The proof writes 'if γ ≤ (C − 1)/C' where the theorem states the strict inequality γ < (C − 1)/C; strictness is needed for the coefficient of R(L; f) to be positive, so the proof's inequality should be corrected.
  2. There is a typo, 'ecall' should be 'recall', and the intuitive approximations in Eqs. (28) and (29) drop log M or log K constants without comment; the approximations are useful for intuition but the constants matter for the exact robustness argument.
  3. The verification of the RINCE parameter λ = 1/(K + 1) does not state which K (number of negatives) is used in Figure 2c; since K determines the theoretical value, this missing detail makes the empirical confirmation difficult to check.
  4. The display in Eq. (23) is garbled: the term f(x)⊤f(x⁻) appears to be placed inside the expectation with the logarithm in an ambiguous way. The expression should be rewritten as E_{x,x⁻}[f(x)⊤f(x⁻)] − E_x[log E_{x⁺} exp(f(x)⊤f(x⁺))] or an equivalent unambiguous form.

Circularity Check

1 steps flagged · score 3.0 of 10

Central SymNCE derivation is constructive and not circular; the NN-selection explanation is tautological because the threshold is chosen to force the robust condition, so circularity is limited to a secondary contribution.

  1. self definitional [Section V-A, 'NN Selection under the Unified Theoretical Framework']
    "Because E_{x′∈N (x;t)}f(x)⊤f(x′) ≥ E_{x′∼PX} f(x)⊤f(x′), for a given f, we can select a proper threshold parameter t to make lim M,K→∞ ∆R(LInfoNCE−NN; f, t) − log K = 0, and thus LInfoNCE−NN is noise tolerant."

    The robustness of NN-selected contrastive learning is asserted after choosing t so that the defining expression of the additional risk is zero, which is exactly the sufficient condition of Theorem III.5. The framework does not derive NN robustness from an independent property; it tunes the threshold to satisfy the condition, making the explanation tautological. Moreover, the choice is made 'for a given f', so the loss is not fixed when minimizing over f, and Theorem III.5's premise of constant ΔR for a fixed loss is not shown for any concrete NN selection rule. Any loss whose positive set can be adjusted to zero out ΔR would be declared noise tolerant by the same argument.

full rationale

Score 3: no central circularity. The decomposition in Lemma III.3 and Theorem III.4 is an algebraic identity under the stated class-balanced and symmetric-noise assumptions, and Theorem III.5 is a direct consequence rather than a restatement of the conclusion. SymNCE is deliberately constructed by adding RevNCE whose ΔRlim cancels InfoNCE's nonconstant part, so its provable robustness is a verification of a designed property, not a fitted prediction. The main concern for the central claim is a correctness gap, not circularity: Eq. (27) establishes only lim_{M,K} ΔR(LSymNCE)=logK, and Theorem IV.1 redefines ΔRlim(LInfoNCE) inconsistently with Eq. (22), while Theorem III.5 requires exact constancy of the finite ΔR; for finite batches ΔR(LSymNCE;f) generally depends on f. The NN-selection explanation in Section V-A is the one genuinely circular step: the threshold t is selected to force the additional-risk expression to zero, so the explanation reduces to imposing the robust condition by definition. RINCE's λ=1/(K+1) follows validly from setting ΔR=0; its experimental check in Section VI-D is under-specified because K is not identified, but it is not circular. Self-citations such as [18] are contextual and not load-bearing.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The framework introduces no new physical entities; SymNCE and RevNCE are loss functions, and the additional risk ΔR is a mathematical construct. The load-bearing assumptions are class balance, symmetric feature-independent noise, and the infinite-sample limit.

free parameters (4)
  • β (SymNCE weight) = 0.2 to 1.0 depending on noise rate
    Balances InfoNCE and RevNCE; selected via validation, not derived from the theory.
  • temperature τ = 0.5 (CIFAR-10), 0.07 (CIFAR-100), 0.1 (TinyImageNet/Clothing1M)
    Chosen by hand per dataset; standard practice, but not theoretically grounded.
  • threshold t in NN-selection explanation = unspecified
    The proof that NN selection makes ΔR constant relies on choosing t to force the expression to zero for each f; t is a free parameter with no practical selection rule.
  • K in RINCE λ verification = 10 (implied)
    The theoretical λ=1/(K+1)≈0.09 is matched to experiments, but the actual number of negatives K in the RINCE implementation is not reported.
assumptions (6)
  • domain assumption Class-balanced data: πi = 1/C for all classes
    Used in Theorems III.4 and III.5; without it the risk decomposition into R and ΔR with constant coefficients is not proven.
  • domain assumption Symmetric label noise: qi(i)=1-γ, qi(j)=γ/(C-1)
    Central assumption of the main theorem; real-world noise is often instance-dependent.
  • domain assumption Label corruption independent of features
    Assumed in the label noise generation model at the start of Section III-A.
  • ad hoc to paper Limit M,K→∞ for SymNCE robustness
    The constant-ΔR proof for SymNCE uses the limiting risk; finite-sample behavior is not theoretically characterized.
  • domain assumption Negatives drawn i.i.d. from marginal PX
    Used in Lemma III.3 and throughout; in practice negatives come from a batch and may include same-class samples.
  • standard math Standard probability tools (Bayes, law of total probability, Jensen)
    Used in Lemma III.2 and Section V-A.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Inclusive Theoretical Framework of Robust Supervised Contrastive Loss against Label Noise." pith.science (2026). https://pith.science/paper/FQUECXHS

@misc{pith2026250101130,
  author       = {Pith},
  title        = {Pith review of: An Inclusive Theoretical Framework of Robust Supervised Contrastive Loss against Label Noise},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FQUECXHS}},
  note         = {Machine review of arXiv:2501.01130}
}
read the original abstract

Learning from noisy labels is a critical challenge in machine learning, with vast implications for numerous real-world scenarios. While supervised contrastive learning has recently emerged as a powerful tool for navigating label noise, many existing solutions remain heuristic, often devoid of a systematic theoretical foundation for crafting robust supervised contrastive losses. To address the gap, in this paper, we propose a unified theoretical framework for robust losses under the pairwise contrastive paradigm. In particular, we for the first time derive a general robust condition for arbitrary contrastive losses, which serves as a criterion to verify the theoretical robustness of a supervised contrastive loss against label noise. The theory indicates that the popular InfoNCE loss is in fact non-robust, and accordingly inspires us to develop a robust version of InfoNCE, termed Symmetric InfoNCE (SymNCE). Moreover, we highlight that our theory is an inclusive framework that provides explanations to prior robust techniques such as nearest-neighbor (NN) sample selection and robust contrastive loss. Validation experiments on benchmark datasets demonstrate the superiority of SymNCE against label noise.

Figures

Figures reproduced from arXiv: 2501.01130 by the authors.

Figure 1
Figure 1. Comparisons in linear probing accuracy between supervised contrastive [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a)(b) Parameter analysis of weight parameter [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 41 canonical work pages

  1. [1]

    Imagenet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Communications of the ACM , vol. 60, no. 6, pp. 84–90, 2017

    2017

  2. [2]

    You only look once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” in CVPR, 2016

    2016

  3. [3]

    Deep learning over multi-field categorical data: –a case study on user response prediction,

    W. Zhang, T. Du, and J. Wang, “Deep learning over multi-field categorical data: –a case study on user response prediction,” in ECIR, 2016

    work page 2016

  4. [4]

    Neural information retrieval: at the end of the early years,

    K. D. Onal, Y . Zhang, I. S. Altingovde, M. M. Rahman, P. Karagoz, A. Braylan, B. Dang, H.-L. Chang, H. Kim, Q. McNamara et al., “Neural information retrieval: at the end of the early years,” Information Retrieval Journal, vol. 21, pp. 111–182, 2018

    work page 2018

  5. [5]

    Universal language model fine-tuning for text classification,

    J. Howard and S. Ruder, “Universal language model fine-tuning for text classification,” in ACL, 2018

    work page 2018

  6. [6]

    Twitter sentiment analysis with deep convolutional neural networks,

    A. Severyn and A. Moschitti, “Twitter sentiment analysis with deep convolutional neural networks,” in Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval, 2015, pp. 959–962

    work page 2015

  7. [7]

    Supervised contrastive learning,

    P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learning,” in NeurIPS, 2020

    work page 2020

  8. [8]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” in ICML, 2020

    2020

Show all 50 references
  1. [9]

    Multi-objective interpolation training for robustness to label noise,

    D. Ortego, E. Arazo, P. Albert, N. E. O’Connor, and K. McGuinness, “Multi-objective interpolation training for robustness to label noise,” in CVPR, 2021

    2021

  2. [10]

    Unicon: Combating label noise through uniform selection and contrastive learning,

    N. Karim, M. N. Rizve, N. Rahnavard, A. Mian, and M. Shah, “Unicon: Combating label noise through uniform selection and contrastive learning,” in CVPR, 2022

    2022

  3. [11]

    Selective-supervised contrastive learning with noisy labels,

    S. Li, X. Xia, S. Ge, and T. Liu, “Selective-supervised contrastive learning with noisy labels,” in CVPR, 2022

    2022

  4. [12]

    Twin contrastive learning with noisy labels,

    Z. Huang, J. Zhang, and H. Shan, “Twin contrastive learning with noisy labels,” in ICCV, 2023

    2023

  5. [13]

    Pico: Contrastive label disambiguation for partial label learning,

    H. Wang, R. Xiao, Y . Li, L. Feng, G. Niu, G. Chen, and J. Zhao, “Pico: Contrastive label disambiguation for partial label learning,” in International conference on learning representations , 2022

    2022

  6. [14]

    Pico+: Contrastive label disambiguation for robust partial label learning,

    ——, “Pico+: Contrastive label disambiguation for robust partial label learning,” IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 2023

    2023

  7. [15]

    Leveraged weighted loss for partial label learning,

    H. Wen, J. Cui, H. Hang, J. Liu, Y . Wang, and Z. Lin, “Leveraged weighted loss for partial label learning,” in ICML, 2021

    2021

  8. [16]

    Class-aware contrastive semi-supervised learning,

    F. Yang, K. Wu, S. Zhang, G. Jiang, Y . Liu, F. Zheng, W. Zhang, C. Wang, and L. Zeng, “Class-aware contrastive semi-supervised learning,” in CVPR, 2022

    2022

  9. [17]

    A graph-theoretic framework for understanding open-world semi-supervised learning,

    Y . Sun, Z. Shi, and Y . Li, “A graph-theoretic framework for understanding open-world semi-supervised learning,” Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  10. [18]

    Rethinking weak supervision in helping contrastive learning,

    J. Cui, W. Huang, Y . Wang, and Y . Wang, “Rethinking weak supervision in helping contrastive learning,” in ICML, 2023

    2023

  11. [19]

    Constrained mean shift using distant yet related neighbors for representation learning,

    K. Navaneet, S. Abbasi Koohpayegani, A. Tejankar, K. Pourahmadi, A. Subramanya, and H. Pirsiavash, “Constrained mean shift using distant yet related neighbors for representation learning,” in ECCV 2022, 2022

    2022

  12. [20]

    Robust contrastive learning against noisy views,

    C.-Y . Chuang, R. D. Hjelm, X. Wang, V . Vineet, N. Joshi, A. Torralba, S. Jegelka, and Y . Song, “Robust contrastive learning against noisy views,” in CVPR, 2022

    2022

  13. [21]

    Training deep neural-networks using a noise adaptation layer,

    J. Goldberger and E. Ben-Reuven, “Training deep neural-networks using a noise adaptation layer,” in ICLR, 2017

    2017

  14. [22]

    Masking: A new perspective of noisy supervision,

    B. Han, J. Yao, G. Niu, M. Zhou, I. Tsang, Y . Zhang, and M. Sugiyama, “Masking: A new perspective of noisy supervision,” 2018

    2018

  15. [23]

    Deep learning from noisy image labels with quality embedding,

    J. Yao, J. Wang, I. W. Tsang, Y . Zhang, J. Sun, C. Zhang, and R. Zhang, “Deep learning from noisy image labels with quality embedding,” IEEE Transactions on Image Processing , vol. 28, no. 4, pp. 1909–1922, 2018

    1909

  16. [24]

    Does label smoothing mitigate label noise?

    M. Lukasik, S. Bhojanapalli, A. Menon, and S. Kumar, “Does label smoothing mitigate label noise?” in ICML, 2020

    2020

  17. [25]

    Reg- ularizing neural networks by penalizing confident output distributions,

    G. Pereyra, G. Tucker, J. Chorowski, Ł. Kaiser, and G. Hinton, “Reg- ularizing neural networks by penalizing confident output distributions,” arXiv preprint arXiv:1701.06548 , 2017

    2017 arXiv

  18. [26]

    Open-set label noise can improve robustness against inherent label noise,

    H. Wei, L. Tao, R. Xie, and B. An, “Open-set label noise can improve robustness against inherent label noise,” 2021

    2021

  19. [27]

    Co-teaching: Robust training of deep neural networks with extremely noisy labels,

    B. Han, Q. Yao, X. Yu, G. Niu, M. Xu, W. Hu, I. Tsang, and M. Sugiyama, “Co-teaching: Robust training of deep neural networks with extremely noisy labels,” 2018

    2018

  20. [28]

    Selfie: Refurbishing unclean samples for robust deep learning,

    H. Song, M. Kim, and J.-G. Lee, “Selfie: Refurbishing unclean samples for robust deep learning,” in ICML, 2019

    2019

  21. [29]

    How does disagreement help generalization against label corruption?

    X. Yu, B. Han, J. Yao, G. Niu, I. Tsang, and M. Sugiyama, “How does disagreement help generalization against label corruption?” in ICML, 2019

    2019

  22. [30]

    Iterative learning with open-set noisy labels,

    Y . Wang, W. Liu, X. Ma, J. Bailey, H. Zha, L. Song, and S.-T. Xia, “Iterative learning with open-set noisy labels,” in CVPR, 2018

    2018

  23. [31]

    Using trusted data to train deep networks on labels corrupted by severe noise,

    D. Hendrycks, M. Mazeika, D. Wilson, and K. Gimpel, “Using trusted data to train deep networks on labels corrupted by severe noise,” 2018

    2018

  24. [32]

    Dimensionality-driven learning with noisy labels,

    X. Ma, Y . Wang, M. E. Houle, S. Zhou, S. Erfani, S. Xia, S. Wijewick- rema, and J. Bailey, “Dimensionality-driven learning with noisy labels,” in ICML, 2018

    2018

  25. [33]

    Making deep neural networks robust to label noise: A loss correction approach,

    G. Patrini, A. Rozza, A. Krishna Menon, R. Nock, and L. Qu, “Making deep neural networks robust to label noise: A loss correction approach,” in CVPR, 2017

    2017

  26. [34]

    Approximating instance-dependent noise via instance-confidence embedding,

    Y . Zhang and M. Sugiyama, “Approximating instance-dependent noise via instance-confidence embedding,” arXiv preprint arXiv:2103.13569 , 2021. 12

    2021 arXiv

  27. [35]

    Robust loss functions under label noise for deep neural networks,

    A. Ghosh, H. Kumar, and P. S. Sastry, “Robust loss functions under label noise for deep neural networks,” in AAAI, 2017

    2017

  28. [36]

    Normalized loss functions for deep learning with noisy labels,

    X. Ma, H. Huang, Y . Wang, S. Romano, S. Erfani, and J. Bailey, “Normalized loss functions for deep learning with noisy labels,” in ICML, 2020

    2020

  29. [37]

    Symmetric cross entropy for robust learning with noisy labels,

    Y . Wang, X. Ma, Z. Chen, Y . Luo, J. Yi, and J. Bailey, “Symmetric cross entropy for robust learning with noisy labels,” in ICCV, 2019

    2019

  30. [38]

    Generalized cross entropy loss for training deep neural networks with noisy labels,

    Z. Zhang and M. Sabuncu, “Generalized cross entropy loss for training deep neural networks with noisy labels,” 2018

    2018

  31. [39]

    Momentum contrast for unsupervised visual representation learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” in CVPR, 2020

    2020

  32. [40]

    Residual relaxation for multi-view representation learning,

    Y . Wang, Z. Geng, F. Jiang, C. Li, Y . Wang, J. Yang, and Z. Lin, “Residual relaxation for multi-view representation learning,” in NeurIPS, 2021

    2021

  33. [41]

    Noise is also useful: Negative correlation-steered latent contrastive learning,

    J. Yan, L. Luo, C. Xu, C. Deng, and H. Huang, “Noise is also useful: Negative correlation-steered latent contrastive learning,” in CVPR, 2022

    2022

  34. [42]

    Demystifying how self- supervised features improve training from noisy labels,

    H. Cheng, Z. Zhu, X. Sun, and Y . Liu, “Demystifying how self- supervised features improve training from noisy labels,” arXiv preprint arXiv:2110.09022, 2021

    2021 arXiv

  35. [43]

    Investigating why contrastive learning benefits robustness against label noise,

    Y . Xue, K. Whitecross, and B. Mirzasoleiman, “Investigating why contrastive learning benefits robustness against label noise,” in ICML, 2022

    2022

  36. [44]

    A theoretical analysis of contrastive unsupervised representation learning,

    S. Arora, H. Khandeparkar, M. Khodak, O. Plevrakis, and N. Saunshi, “A theoretical analysis of contrastive unsupervised representation learning,” in ICML, 2019

    2019

  37. [45]

    Learning with noisy labels,

    N. Natarajan, I. S. Dhillon, P. K. Ravikumar, and A. Tewari, “Learning with noisy labels,” 2013

    2013

  38. [46]

    Understanding contrastive representation learning through alignment and uniformity on the hypersphere,

    T. Wang and P. Isola, “Understanding contrastive representation learning through alignment and uniformity on the hypersphere,” in ICML. PMLR, 2020

    2020

  39. [47]

    Learning multiple layers of features from tiny images,

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

    2009

  40. [48]

    Learning from massive noisy labeled data for image classification,

    T. Xiao, T. Xia, Y . Yang, C. Huang, and X. Wang, “Learning from massive noisy labeled data for image classification,” in CVPR, 2015

    2015

  41. [49]

    Dividemix: Learning with noisy labels as semi-supervised learning,

    J. Li, R. Socher, and S. C. Hoi, “Dividemix: Learning with noisy labels as semi-supervised learning,” arXiv preprint arXiv:2002.07394 , 2020. 13 APPENDIX In this appendix, we present the proofs related to Sections III and IV in Appendices A and B respectively. The additional t...

    2002 arXiv

  42. [50]

    Suppose f ∗ is the minimizer of eR(L; f ), and thus we have when 1 − γu > γui, R(L; f ) − R(L; f ∗) = 1 c1(γ) − c2(γ) h eR(L; f ∗) − eR(L; f ) i < 0, that is, f ∗ is also the minimizer of R(L; f )

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.