Pith. sign in

REVIEW 4 major objections 3 minor 50 references

Robust Losses from Univariate Base Functions for Noisy-Label Learning

T0 review · 4 major / 3 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Robust losses for noisy labels can be built from univariate base functions, with noise tolerance reduced to derivative inequalities on that base function.

desk verdict The framework is a genuinely useful way to design robust losses from univariate base functions, but the missing appendix and unverified sufficient conditions for BEF/BQF leave the main theoretical claim uncheckable as submitted. read the letter →

arxiv 2607.16768 v1 pith:GZTWIL7Q submitted 2026-07-18 cs.LG

Peng Hu , Jianwei Ma This is my paper · ORCID
classification cs.LG
keywords noisy-labellearningrobustlossfunctionsnoiserobustnesssymmetriclossesasymmetrictargetseparationbinaryreductionunivariatebase
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

This paper claims that designing a multiclass loss that tolerates noisy labels can be reduced to choosing a univariate base function, because the noise robustness of the induced loss is controlled by simple inequalities on the base function's derivative. It proposes two construction schemes—Target Separation, which promotes the target class while suppressing non-target classes independently, and Binary Reduction, which compares the target against each non-target class through pairwise conditional probabilities. For both schemes it proves sufficient conditions (Theorems 4.2 and 4.5) under which the induced loss is symmetric or asymmetric, and therefore noise-tolerant at the population-risk level. Following those conditions, the paper constructs two losses, BEF and BQF, and reports that they rank first or second on most CIFAR-100, WebVision, and Clothing1M settings. If correct, the framework turns robust-loss design from crafting a final objective into checking derivative conditions on one function.

What carries the argument

The central object is the univariate base function h (or binary loss ℓ) together with a mapping operator that lifts it to a multiclass loss. For Target Separation the operator is L(u,t)=h(u_t)-a∑_{i≠t}h(u_i), and the load-bearing quantity is the derivative h' controlled by the separation coefficient a; Theorem 4.2's inequality turns base-function shape into a certificate of asymmetry. For Binary Reduction the operator averages pairwise losses on normalized conditional probabilities, and the load-bearing object is the centered function g(p)=ℓ(p,1-p)-ℓ(0.5,0.5), whose oddness determines symmetry and whose weighted derivative ratio determines asymmetry. Because both conditions are derivative co

What would settle it

Take a univariate base function whose derivative satisfies the inequality in Theorem 4.2, integrate it into a target-separated loss, and train a standard deep network on CIFAR-100 with 0.8 symmetric noise under the exact schedule used for BEF; if the induced loss lands near plain cross-entropy rather than near BEF, the derivative conditions do not by themselves transfer to practical robustness.

Watch

Extended reading notes

Core claim

The paper's central claim is that robust multiclass losses can be induced from univariate base functions by two mappings. In Target Separation, the loss is written as L(u,t)=h(u_t)-a∑_{i≠t}h(u_i); Theorem 4.2 states that the loss is asymmetric—hence noise-tolerant when clean labels dominate—if the derivative h' satisfies ((1-w_t)a - w_t)h'(u_t) ≥ ((1-w_s)a - w_s)h'(u_s) for all classes s≠t. In Binary Reduction, the loss averages pairwise losses on normalized conditional probabilities, L(u,t)=1/(K-1)∑_{i≠t}ℓ(u_t/(u_t+u_i), u_i/(u_t+u_i)); Theorem 4.5 states asymmetry follows from the derivative relation w_i g'(p) ≤ w_j g'(1-p) on the centered binary function g(p)=ℓ(p,1-p)-ℓ(0.5,0.5), with sym

Load-bearing premise

The load-bearing premise is that the population-level noise-tolerance guarantees—derived under the assumption that clean-risk and noisy-risk minimizers are unique and that clean labels dominate—carry over unchanged to finite-sample stochastic-gradient training of deep networks.

Editorial extensions

If this is right

  • Noise-robust loss design reduces to a checklist: pick a differentiable decreasing univariate h and verify the derivative inequality in Theorem 4.2 (or the centered derivative inequality in Theorem 4.5); any h passing the check induces an asymmetric, noise-tolerant multiclass loss.
  • Symmetric losses can be obtained without normalization: choosing a=1/(K-1) in Target Separation makes the induced loss symmetric for any base loss (Theorem 4.1), and Binary Reduction preserves symmetry exactly when the centered binary function is odd about 0.5.
  • Base functions with no elementary closed-form antiderivative are usable, because the construction works at the derivative level and recovers the loss by integration.
  • Combining a target-separation term with a binary-reduction term is claimed to be synergistic: on CIFAR-10N and CIFAR-100N the balanced combination outperforms either component alone.
  • If the theory transfers to deep-network training, the two constructed losses BEF and BQF should remain competitive under symmetric, asymmetric, instance-dependent, and human-annotation noise; the paper reports exactly this, including best or second-best results on most CIFAR-100, WebVision, and Clothing1M settings.

Reading between the lines

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

  • The paper leaves implicit that the derivative-only characterization makes the loss family searchable: one could numerically optimize over base functions whose sampled derivative profile satisfies the inequalities, rather than hand-picking the two examples.
  • The authors motivate Binary Reduction by late-training overfitting of shared noisy-class features but do not directly test that mechanism; an inference is that the pairwise term's benefit should grow as noise becomes more class-correlated and shrink under purely uniform noise.
  • Because the theorems are population-level, the strongest untested implication is about transfer: if the derivative conditions are checked and the loss still overfits in stochastic-gradient training, the discrepancy would point to optimization dynamics rather than the risk-minimization theory.
  • The framework's conditions are sufficient, not necessary; other existing robust losses might also satisfy them under appropriate weights, so the criterion could serve as a unifying test for whether a proposed loss is noise-robust.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 3 minor

Summary. The paper proposes a framework for constructing multiclass robust losses from univariate base functions through two mappings: Target Separation and Binary Reduction. It states sufficient conditions (Theorems 4.1–4.5 and Corollaries 4.1–4.3) that characterize symmetry and asymmetry of the induced losses in terms of the derivative profile of the base function, and instantiates two new losses, BEF and BQF, by combining the two constructions. Experiments on CIFAR-10/100, CIFAR-10N/100N, WebVision, ILSVRC12, and Clothing1M report strong performance, with BEF/BQF usually in the top two under several noisy-label settings. The central theoretical claims are deferred to an appendix that is absent from the submitted text, and the paper does not report the parameter values used to instantiate BEF/BQF nor verify that they satisfy the stated sufficient conditions.

Significance. If the theoretical results are correct, the framework provides a useful design principle: robust multiclass losses can be engineered by choosing a univariate base function whose derivative satisfies simple inequalities. This would unify and extend existing symmetric and asymmetric loss constructions, and the proposed BEF/BQF losses show competitive or superior empirical accuracy across a wide range of synthetic and real-world noise benchmarks. The explicit sufficient conditions and the broad experimental comparison are strengths. However, the current manuscript is not independently verifiable: the proofs of the main theorems are missing, and the link between the theory and the proposed losses is not demonstrated, since the actual parameters and the satisfaction of the derived conditions are not reported. The central claim is therefore currently unsubstantiated.

major comments (4)
  1. [Sections 4.2–4.3, Theorems 4.1–4.5] The proofs of all theoretical results are stated to be in an appendix, but the submitted manuscript contains no appendix. Theorems 4.1–4.5 and Lemmas 4.1–4.3 constitute the main theoretical contribution and the basis for the claim that robustness is characterized through properties of the base function. Without the proofs, the sufficiency chain cannot be checked. This is a load-bearing gap and must be fixed by including complete proofs in the manuscript.
  2. [Section 4.4, Tables 1–5] The paper never reports the actual values of α1, β1, γ1, α2, β2, γ2, λ, µ, or the separation coefficient a used for BEF and BQF in the experiments. It also does not verify that these choices satisfy the sufficient conditions of Theorem 4.2, Theorem 4.5, or Corollary 4.3. The statement that BEF/BQF are asymmetric because "the symmetric loss is asymmetric and the sum of asymmetric loss functions remains asymmetric" is not a substitute: the components must themselves be asymmetric under the chosen parameters, and the paper only says this holds "under certain parameter settings." Without the parameter values or a verification, the theoretical characterization does not explain the observed empirical robustness.
  3. [Section 4.4, Corollary 4.3 and the numerical example] The derivation of the admissible region for ξ is garbled. The displayed inequality "0.8ξ/9−0.2 / 8.2ξ/9−0.8 ≤ 1/30" is syntactically ambiguous, and the conclusion "1≤ξ≤234/79" is not justified as written. In addition, Corollary 4.3 uses ξ in the context of Target Separation, but ξ was introduced in Definition 4.2 for Binary Reduction; the meaning of ξ here and the formulas for f1/f2's ξ values are not defined or derived. This obscures whether the actual losses used in the experiments fall in the admissible region.
  4. [Section 3.2 and Fig. 1] The noise-tolerance theory is stated for unique minimizers of the expected clean and noisy risks, but the experiments train deep networks with SGD on finite samples. This transfer is not automatic, and the paper's own Fig. 1 shows that theoretically robust losses such as GCE and NCE+AUL can still deteriorate late in training. The manuscript should state explicitly what the theoretical results do and do not claim about finite-sample SGD training, or provide a more direct connection between the derived conditions and optimization dynamics.
minor comments (3)
  1. [Definition 4.1] The symbol L is used both for the base loss and for the induced target-separated loss: L(u,t)=L(u,t)−τ/(K−1)∑_{i≠t} L(u,i). This is confusing; use a different symbol, e.g., ℓ, for the base function.
  2. [Throughout] The text repeatedly misspells "target-separated" as "target-separatd" (Sections 4.4, Corollary 4.3). Typos should be corrected.
  3. [Section 4.4] The parameters of f1 and f2 are listed, but no guidance is given for choosing them beyond the existence of an admissible region. Given that task-specific tuning (Table 4) improves performance, the paper should either report the shared and tuned parameter settings or state where they will be documented.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the framework's sufficient conditions are genuinely derived; the BEF/BQF parameter verification gap is a rigor issue, not a self-referential reduction.

full rationale

I walked the derivation chain from Definitions 3.2/3.3 through Lemmas 4.1/4.3 and Theorems 4.2/4.5. The lemmas prove equivalences between finite-shift weighted-loss inequalities and derivative inequalities, and the theorems use those equivalences to state sufficient conditions for the paper's own asymmetry definition. That is mathematical derivation, not circular reduction: the sufficient conditions are not identical to the claimed robustness property, and the base functions f1/f2 are then instantiated in an attempt to satisfy them. The empirical evaluations in Tables 1-5 are external benchmarks, not fitted inputs to the theory, so there is no fitted-input-called-prediction step. The cited rule 'the sum of asymmetric loss functions remains asymmetric [22]' is prior work by Zhou et al., not a self-citation of the present authors. The load-bearing weakness is a verification gap, not circularity. Section 4.4 says 'Both f1 and f2 are strictly negative over [0,1]. Corollary 4.1 indicates that the target-separatd loss is asymmetric under certain parameter settings,' but it never reports the alpha, beta, gamma, a, lambda, mu values used in the experiments nor checks the Theorem 4.2/4.5 inequalities for those values; the only numerical check is the illustrative 'Consider a ten-class dataset with a symmetric noise rate of 0.8 ... 1 <= xi <= 234/79,' which is not tied to Tables 1-5. The manuscript also states 'The proofs of all theoretical results are provided in the Appendix,' but the appendix is absent from the provided text, so the sufficiency chain cannot be independently checked here. None of these omissions makes a prediction equal to its input or makes the central claim reduce to a self-citation, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on: (i) the standard noise-tolerance theory for symmetric/asymmetric losses treated as a domain axiom, (ii) an explicit uniqueness assumption on risk minimizers, (iii) differentiability assumptions, and (iv) an unverified behavioral hypothesis about overfitted feature sharing. The concrete BEF/BQF losses carry at least five hand-tuned parameters whose values are not fully reported. There are no invented entities.

free parameters (5)
  • separation coefficient a (τ/(K-1)) for Target Separation = a=1/30 used in the worked example (K=10); other values not reported
    Controls the non-target suppression term in Definition 4.1; Theorem 4.2/Corollary 4.1 give asymmetry conditions in terms of a; numeric settings for experiments are not given in the text.
  • f1 shape parameters (α1, β1, γ1)
    Define the Gaussian-like base derivative function for BEF; only sign constraints are given in Section 4.4; actual values omitted.
  • f2 shape parameters (α2, β2, γ2)
    Define the shifted-quadratic base derivative function for BQF; only sign constraints are given; actual values omitted.
  • combination weights (λ, µ) = (0.5,0.5), (0.8,0.2), (0.2,0.8) considered in Fig. 10; reported BEF/BQF values not stated
    Balance L_TS and L_BR in L_TS+BR; chosen after inspecting accuracy curves on the evaluation benchmarks.
  • degenerate binary value ξ = ξ=l(0.5,0.5) for symmetry (Theorem 4.3); boundary constraints l(1,0)=inf, l(0,1)=sup in Lemma 4.2
    Resolves the undefined u_t+u_i=0 case in Definition 4.2; the ξ constraints enter Corollary 4.3's parameter bounds.
assumptions (4)
  • domain assumption Symmetric and asymmetric losses are noise-tolerant under the conditions of Ghosh et al. [17] and Zhou et al. [18,22] (clean-label dominance for asymmetric losses)
    The paper adopts these results as the backbone of its robustness criteria (Section 3.3). If they fail for the considered distributions, the constructed losses inherit no guarantee.
  • domain assumption The clean-risk and noisy-risk minimizers f* and f*_η are unique
    Section 3.2 states this explicitly; Definition 3.1 and the theory rely on a well-defined arg-min comparison.
  • standard math Base functions h and g are differentiable and monotone on [0,1] as required
    Lemma 4.1 and 4.3 convert finite probability-shift inequalities into pointwise derivative inequalities; non-differentiable base functions (e.g., -log near 0) are outside the stated conditions.
  • ad hoc to paper The 'shared overfitted features' behavioral hypothesis
    Section 1 hypothesizes that samples of one true class with different noisy labels develop similar overfitted features; this motivates Binary Reduction but is not verified except by qualitative visual reasoning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Losses from Univariate Base Functions for Noisy-Label Learning." pith.science (2026). https://pith.science/paper/GZTWIL7Q

@misc{pith2026260716768,
  author       = {Pith},
  title        = {Pith review of: Robust Losses from Univariate Base Functions for Noisy-Label Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GZTWIL7Q}},
  note         = {Machine review of arXiv:2607.16768}
}
read the original abstract

Learning with noisy labels is a fundamental problem in training reliable deep neural networks. Robust loss functions provide a direct and effective way to mitigate the adverse effects of label noise. However, most existing robust losses are designed directly at the level of the final multiclass objective, which makes it difficult to systematically characterize and extend their robustness properties. In this paper, we propose a general framework that constructs robust multiclass losses from univariate base functions. By defining mapping operators from base functions to multiclass losses, the robustness of the induced losses can be characterized through simple properties of the base functions. We develop two complementary construction schemes, Target Separation and Binary Reduction, corresponding to inter-class independent and inter-class dependent formulations, respectively. For both schemes, we analyze their symmetry and asymmetry properties and derive corresponding sufficient conditions, which provide theoretical criteria for noise-robust loss design. The proposed framework also provides a new route to constructing symmetric losses, serving as a complement to normalization-based symmetric loss designs. Extensive experiments on synthetic and real-world noisy-label benchmarks demonstrate that the proposed losses achieve competitive or superior performance under various noise settings.

Figures

Figures reproduced from arXiv: 2607.16768 by the authors.

Figure 1
Figure 1. Test accuracy curves of GCE (left) and NCE+AUL [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Noisy-Label Learning. Left: original data with noisy labels. Middle: during early training, clean-label samples [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed framework. Starting from univariate base functions, we derive two complementary robust [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Training curves of the target-separation term only, the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: K-class classification via (K-1) binary comparisons. We still consider a multiclass classification problem with K classes. And the probability vector output by the network [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Label transition matrices for representative noisy-label settings on CIFAR-10 and CIFAR-10N. The first three [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: 2D t-SNE [43] visualization of learned representations on the CIFAR-10 test set under 0.4 instance-dependent noise. BEF and BQF yield more discriminative feature distribu￾tions with improved class separation and clearer bound￾aries. The BEF and BQF perform particularly…
Figure 8
Figure 8. Figure 8: Test accuracy curves of different methods under representative noisy settings on CIFAR-10. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Test accuracy curves of different methods under representative noisy settings on CIFAR-100. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Test accuracies of BEF and BQF with different weighting coefficients under “Worst” noise types on CIFAR-10N and [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 3 linked inside Pith

  1. [1]

    Deep learning,

    Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015

    2015

  2. [2]

    A survey of label-noise representa- tion learning: Past, present and future,

    B. Han, Q. Yao, T. Liu, G. Niu, I. W. Tsang, J. T. Kwok, and M. Sugiyama, “A survey of label-noise representa- tion learning: Past, present and future,”arXiv preprint arXiv:2011.04406, 2021

    arXiv 2011

  3. [3]

    A closer look at memorization in deep networks,

    D. Arpit, S. Jastrz˛ ebski, N. Ballas, D. Krueger, E. Bengio, M. S. Kanwal, T. Maharaj, A. Fischer, A. Courville, Y. Bengio, and S. Lacoste-Julien, “A closer look at memorization in deep networks,” inProceedings of the 34th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 70, 06– 11 Aug 2017, pp. 233–242

    2017

  4. [4]

    Weak-to-strong gen- eralization: Eliciting strong capabilities with weak su- pervision,

    C. Burns, P . Izmailov, J. H. Kirchner, B. Baker, L. Gao, L. Aschenbrenner, Y. Chen, A. Ecoffet, M. Joglekar, J. Leike, I. Sutskever, and J. Wu, “Weak-to-strong gen- eralization: Eliciting strong capabilities with weak su- pervision,” inProceedings of the 41st International Confer- ence on Machine Learning, ser. Proceedings of Machine Learning Research, v...

    2024

  5. [5]

    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,” inAdvances in Neural Information Processing Systems, vol. 31, 2018

    2018

  6. [6]

    Decoupling

    E. Malach and S. Shalev-Shwartz, “Decoupling "when to update" from "how to update",” inAdvances in Neural Information Processing Systems, vol. 30, 2017

    2017

  7. [7]

    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?” inProceedings of the 36th International Conference on Machine Learning, ser. Proceedings of Ma- chine Learning Research, vol. 97, 09–15 Jun 2019, pp. 7164–7173

    2019

  8. [8]

    L DMI: A novel information-theoretic loss function for training deep nets robust to label noise,

    Y. Xu, P . Cao, Y. Kong, and Y. Wang, “L DMI: A novel information-theoretic loss function for training deep nets robust to label noise,” inAdvances in Neural In- formation Processing Systems, vol. 32, 2019. 14

    2019

Show all 50 references
  1. [9]

    Cleannet: Transfer learning for scalable image classifier training with label noise,

    K.-H. Lee, X. He, L. Zhang, and L. Yang, “Cleannet: Transfer learning for scalable image classifier training with label noise,” in2018 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2018, pp. 5447–5456

    2018

  2. [10]

    Toward robustness against label noise in training deep discriminative neural networks,

    A. Vahdat, “Toward robustness against label noise in training deep discriminative neural networks,” inAd- vances in Neural Information Processing Systems, vol. 30. Curran Associates, Inc., 2017

    2017

  3. [11]

    Learning from noisy labels with distillation,

    Y. Li, J. Yang, Y. Song, L. Cao, J. Luo, and L.-J. Li, “Learning from noisy labels with distillation,” in2017 IEEE International Conference on Computer Vision (ICCV), 2017, pp. 1928–1936

    2017

  4. [12]

    Learning from noisy large-scale datasets with minimal supervision,

    A. Veit, N. Alldrin, G. Chechik, I. Krasin, A. Gupta, and S. Belongie, “Learning from noisy large-scale datasets with minimal supervision,” in2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 6575–6583

    2017

  5. [13]

    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,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017

    2017

  6. [14]

    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,” inAdvances in Neural Information Processing Systems, vol. 31, 2018

    2018

  7. [15]

    Rethinking the inception architecture for computer vision,

    C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the inception architecture for computer vision,” in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 2818– 2826

    2016

  8. [16]

    Training deep neural-networks using a noise adaptation layer,

    J. Goldberger and E. Ben-Reuven, “Training deep neural-networks using a noise adaptation layer,” inIn- ternational Conference on Learning Representations, 2017

    2017

  9. [17]

    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,” Proceedings of the AAAI Conference on Artificial Intelli- gence, vol. 31, no. 1, Feb. 2017

    2017

  10. [18]

    Asymmet- ric loss functions for noise-tolerant learning: Theory and applications,

    X. Zhou, X. Liu, D. Zhai, J. Jiang, and X. Ji, “Asymmet- ric loss functions for noise-tolerant learning: Theory and applications,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 7, pp. 8094–8109, 2023

    2023

  11. [19]

    Student loss: Towards the probability assumption in inaccurate su- pervision,

    S. Zhang, J.-Q. Li, H. Fujita, Y.-W. Li, D.-B. Wang, T.- T. Zhu, M.-L. Zhang, and C.-Y. Liu, “Student loss: Towards the probability assumption in inaccurate su- pervision,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 6, pp. 4460–4475, 2024

    2024

  12. [20]

    Joint Asymmetric Loss for Learning with Noisy Labels,

    J. Wang, X. Liu, X. Zhou, G. Hu, D. Zhai, J. Jiang, and X. Ji, “Joint Asymmetric Loss for Learning with Noisy Labels,” inInternational Conference on Computer Vision, 2025, pp. 1947–1956

    2025

  13. [21]

    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,” inProceedings of the 37th International Conference on Machine Learning, ser. Proceedings of Ma- chine Learning Research, vol. 119, 13–18 Jul 2020, pp...

    2020

  14. [22]

    Asym- metric loss functions for learning with noisy labels,

    X. Zhou, X. Liu, J. Jiang, X. Gao, and X. Ji, “Asym- metric loss functions for learning with noisy labels,” inProceedings of the 38th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 139, 18–24 Jul 2021, pp. 12 846–12 856

    2021

  15. [23]

    Variation-bounded loss for noise-tolerant learning,

    J. Wang, X. Zhou, X. Liu, G. Hu, D. Zhai, J. Jiang, and H. Li, “Variation-bounded loss for noise-tolerant learning,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 31, pp. 26 251–26 259, 2026

    2026

  16. [24]

    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,” inAdvances in Neural Information Processing Systems, vol. 31, 2018

    2018

  17. [25]

    Secost: Sequential co- supervision for weakly labeled audio event detection,

    A. Kumar and V . K. Ithapu, “Secost: Sequential co- supervision for weakly labeled audio event detection,” arXiv preprint arXiv:1910.11789, 2019

    1910 arXiv

  18. [26]

    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,” in2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 2691– 2699

    2015

  19. [27]

    Mitigating memorization of noisy labels by clipping the model prediction,

    H. Wei, H. Zhuang, R. Xie, L. Feng, G. Niu, B. An, and Y. Li, “Mitigating memorization of noisy labels by clipping the model prediction,” inProceedings of the 40th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 202, 23–29 Jul ...

    2023

  20. [28]

    Noise tolerance under risk minimization,

    N. Manwani and P . S. Sastry, “Noise tolerance under risk minimization,”IEEE Transactions on Cybernetics, vol. 43, no. 3, pp. 1146–1151, 2013

    2013

  21. [29]

    Learning with symmetric label noise: The importance of being unhinged,

    B. van Rooyen, A. Menon, and R. C. Williamson, “Learning with symmetric label noise: The importance of being unhinged,” inAdvances in Neural Information Processing Systems, vol. 28, 2015

    2015

  22. [30]

    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,” in2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 322–330

    2019

  23. [31]

    Can cross entropy loss be robust to label noise?

    L. Feng, S. Shu, Z. Lin, F. Lv, L. Li, and B. An, “Can cross entropy loss be robust to label noise?” inProceedings of the Twenty-Ninth International Joint Conference on Artifi- cial Intelligence, IJCAI-20, C. Bessiere, Ed. International Joint Conferences on Artificial Intellig...

    2020

  24. [32]

    Generalized jensen- shannon divergence loss for learning with noisy la- bels,

    E. Englesson and H. Azizpour, “Generalized jensen- shannon divergence loss for learning with noisy la- bels,” inAdvances in Neural Information Processing Sys- tems, vol. 34, 2021, pp. 30 284–30 297

    2021

  25. [33]

    Learning with noisy labels via sparse regularization,

    X. Zhou, X. Liu, C. Wang, D. Zhai, J. Jiang, and X. Ji, “Learning with noisy labels via sparse regularization,” in2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 72–81

    2021

  26. [34]

    ϵ- softmax: Approximating one-hot vectors for mitigating label noise,

    J. Wang, X. Zhou, D. Zhai, J. Jiang, X. Ji, and X. Liu, “ϵ- softmax: Approximating one-hot vectors for mitigating label noise,” inAdvances in Neural Information Processing Systems, vol. 37, 2024, pp. 32 012–32 038

    2024

  27. [35]

    NLNL: Negative learning for noisy labels,

    Y. Kim, J. Yim, J. Yun, and J. Kim, “NLNL: Negative learning for noisy labels,” in2019 IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), 2019, pp. 101–110

    2019

  28. [36]

    Joint negative and positive learning for noisy labels,

    Y. Kim, J. Yun, H. Shon, and J. Kim, “Joint negative and positive learning for noisy labels,” in2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 9437–9446. 15

    2021

  29. [37]

    Active negative loss functions for learning with noisy labels,

    X. Ye, X. Li, s. dai, T. Liu, Y. Sun, and W. Tong, “Active negative loss functions for learning with noisy labels,” inAdvances in Neural Information Processing Systems, vol. 36, 2023, pp. 6917–6940

    2023

  30. [38]

    To smooth or not? When label smoothing meets noisy labels,

    J. Wei, H. Liu, T. Liu, G. Niu, M. Sugiyama, and Y. Liu, “To smooth or not? When label smoothing meets noisy labels,” inProceedings of the 39th International Confer- ence on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 162, 17–23 Jul 2022, pp. 23 589– 23 614

    2022

  31. [39]

    Can gradient clipping mitigate label noise?

    A. K. Menon, A. S. Rawat, S. J. Reddi, and S. Kumar, “Can gradient clipping mitigate label noise?” inInter- national Conference on Learning Representations, 2020

    2020

  32. [40]

    On symmetric losses for learning from corrupted labels,

    N. Charoenphakdee, J. Lee, and M. Sugiyama, “On symmetric losses for learning from corrupted labels,” inProceedings of the 36th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 97, 09–15 Jun 2019, pp. 961–970

    2019

  33. [41]

    Part-dependent label noise: Towards instance-dependent label noise,

    X. Xia, T. Liu, B. Han, N. Wang, M. Gong, H. Liu, G. Niu, D. Tao, and M. Sugiyama, “Part-dependent label noise: Towards instance-dependent label noise,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 7597–7610

    2020

  34. [42]

    Con- vexity, classification, and risk bounds,

    P . L. Bartlett, M. I. Jordan, and J. D. McAuliffe, “Con- vexity, classification, and risk bounds,”Journal of the American Statistical Association, vol. 101, no. 473, pp. 138–156, 2006

    2006

  35. [43]

    Visualizing data using t-sne,

    L. van der Maaten and G. Hinton, “Visualizing data using t-sne,”Journal of Machine Learning Research, vol. 9, no. 86, pp. 2579–2605, 2008

    2008

  36. [44]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, “Learning multiple layers of features from tiny images,” 2009

    2009

  37. [45]

    Learning with noisy labels revisited: A study using real-world human annotations,

    J. Wei, Z. Zhu, H. Cheng, T. Liu, G. Niu, and Y. Liu, “Learning with noisy labels revisited: A study using real-world human annotations,” inInternational Confer- ence on Learning Representations, 2022

    2022

  38. [46]

    Webvision database: Visual learning and understand- ing from web data,

    W. Li, L. Wang, W. Li, E. Agustsson, and L. V . Gool, “Webvision database: Visual learning and understand- ing from web data,”arXiv preprint arXiv:1708.02862, 2017

    2017 arXiv

  39. [47]

    Backpropa- gation applied to handwritten zip code recognition,

    Y. LeCun, B. Boser, J. S. D. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel, “Backpropa- gation applied to handwritten zip code recognition,” Neural Computation, vol. 1, no. 4, pp. 541–551, 1989

    1989

  40. [48]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in2016 IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

    2016

  41. [49]

    MentorNet: Learning data-driven curriculum for very deep neural networks on corrupted labels,

    L. Jiang, Z. Zhou, T. Leung, L.-J. Li, and L. Fei-Fei, “MentorNet: Learning data-driven curriculum for very deep neural networks on corrupted labels,” inProceed- ings of the 35th International Conference on Machine Learn- ing, ser. Proceedings of Machine Learning Research, vol...

    2018

  42. [50]

    Focal loss for dense object detection,

    T.-Y. Lin, P . Goyal, R. Girshick, K. He, and P . Dollár, “Focal loss for dense object detection,” in2017 IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2999–3007

    2017

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.