Pith. sign in

REVIEW 4 major objections 4 minor 60 references

The paper claims that cleaning the latent features with a low-rank plus sparse decomposition, on top of loss correction, keeps out-of-distribution detection accurate even when up to half of the training labels are noisy.

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 ↗

NOODLE corrects noisy labels with a transition matrix and cleans features via low-rank sparse decomposition, improving OOD detection under label noise.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A promising combination of loss correction and low-rank decomposition, but the test-time protocol is internally inconsistent—queries are not projected while references are—so the central claim is not yet established. the 4 major comments →

arxiv 2509.06918 v1 pith:UFNBQ7PB submitted 2025-09-08 cs.LG cs.AI

Tackling the Noisy Elephant in the Room: Label Noise-robust Out-of-Distribution Detection via Loss Correction and Low-rank Decomposition

Tarhib Al Azad , Shahana Ibrahim This is my paper
classification cs.LG cs.AI
keywords out-of-distribution detectionlabel noiseloss correctionlow-rank decompositionsparse residualkNN distanceMahalanobis distancefeature cleansing
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

Label noise destroys the cluster structure of learned feature representations, and that breaks distance-based out-of-distribution (OOD) detectors: samples from unseen classes mix with in-distribution ones in feature space. The paper argues that existing noise-robust training methods are not enough, because loss correction mostly adjusts softmax probabilities rather than repairing the feature embeddings that OOD scoring relies on. It proposes NOODLE, which combines a loss-correction term with a low-rank plus column-sparse decomposition of each batch's feature matrix, treating the projected low-rank part as in-distribution signal and the residual as outlier content. Training with this joint objective yields cleaner ID clusters, and at test time a kNN or Mahalanobis score computed on the cleaned features outperforms prior OOD detectors, especially under severe label noise, e.g., an average FPR95 of 27.34% at 50% synthetic noise versus 31.35% for the best baseline. If correct, this turns the overlooked problem of noisy training labels into a tractable feature-cleaning procedure for safety-critical deployment.

Core claim

The paper's central claim is that the right place to counter label noise for OOD detection is the latent feature space, not just the loss. It assumes a batch of features H decomposes as H = QQ^T H + (H - QQ^T H), where Q is the top-K left singular subspace with K equal to the number of classes; the first term is the low-rank ID component, the second is a column-sparse residual carrying non-ID/outlier information. NOODLE learns a classifier with a loss-correction loss plus a column-sparsity penalty on the residual, and then performs OOD detection with kNN or Mahalanobis distances on the cleaned, projection-retained features. The paper reports consistent gains over OOD detection baselines and

What carries the argument

The central object is the low-rank plus column-sparse decomposition of the latent feature matrix H, written H ≈ QQ^T H + (H - QQ^T H), with Q estimated by power iteration as the top-K left singular subspace and K set to the number of classes. The projected part QQ^T H is treated as the in-distribution feature used for OOD scoring, while the residual H - QQ^T H is penalized to be column-sparse via the mixed norm ||H - QQ^T H||_{2,1}. This forces any deviation from the class-subspace structure to be concentrated on few samples, interpreted as potential OOD or outlier content. Combined with a loss-correction term in the training objective, the decomposition produces cleaner ID clusters and bett

Load-bearing premise

The method assumes that the top-K subspace of features trained on noisy labels still captures true in-distribution class structure, so that any deviation left in the residual is non-ID noise or outliers, not genuine class information.

What would settle it

Train NOODLE under symmetric noise and inspect the residual norm ||h(x) - QQ^T h(x)||_2 for samples whose ground-truth labels are known. If a substantial fraction of true in-distribution samples have residual norms as large as those of held-out OOD samples, the projection is discarding genuine ID signal. A sharper variant: vary K from well below to well above the number of classes and check whether OOD AUROC changes monotonically; a large change would show that the K-equals-classes assumption, not the decomposition idea, is carrying the result.

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

If this is right

  • At 50% symmetric label noise on CIFAR-10, NOODLE's average FPR95 is 27.34%, a 12.5% relative reduction over the strongest baseline, with corresponding AUROC gains.
  • On real crowd-annotated noise, the cleaned-feature variants beat every baseline: roughly 9.5% FPR95 reduction on CIFAR-10N worst-case, about 14% on Animal-10N, and about 15% on CIFAR-100N.
  • Loss correction alone is insufficient: CM, VolMinNet, SCE, GCE, DivideMix, and Co-Teaching improve classification but leave ID/OOD feature mixing that the low-rank projection specifically removes.
  • The best loss-correction choice depends on the label space: transition-matrix correction works best for 10-class problems, while on CIFAR-100N the SCE-based NOODLE variant is better because estimating a 100x100 transition matrix is harder.
  • The gains hold for both kNN and Mahalanobis scoring, indicating the mechanism is the feature-cleaning step rather than the particular distance metric.
  • The method is modular: NOODLE with CM loss correction and NOODLE with SCE loss correction both outperform their corresponding base losses, so the feature-cleansing idea composes with different noise-robust losses.
  • Cleaned features also preserve competitive ID accuracy, meaning the projection does not simply throw away class information while improving OOD separation.
  • On CIFAR-100N, NOODLE improves FPR95 by about 15% over the best baseline while keeping ID accuracy close to the best baselines, showing the benefit is not achieved by sacrificing classification.

Where Pith is reading between the lines

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

  • Beyond the paper's benchmarks, the decomposition predicts that NOODLE's advantage widens when OOD data is well outside the subspace spanned by ID class structure and shrinks for near-OOD samples; a reader could test this by measuring performance against datasets ordered by feature similarity to the ID data.
  • A natural extension the paper leaves implicit is tuning or learning the rank K instead of fixing it to the number of classes; on large-label-space datasets or with unknown class counts, an adaptive rank chosen from singular-value decay might behave differently.
  • The modular design suggests that any improved loss-correction or sample-selection scheme could be dropped into NOODLE; the SCE variant already shows the framework is not tied to transition matrices.
  • The core mechanism — project features onto a class-subspace and score distances on the residual-cleaned part — transfers to other distance-based scores or anomaly detection settings, though the paper only demonstrates kNN and Mahalanobis.
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

4 major / 4 minor

Summary. The paper addresses OOD detection when the training set has noisy labels. It proposes NOODLE, which combines a loss-correction module (transition-matrix correction or robust loss) with a low-rank plus column-sparse decomposition of the latent feature matrix. During training, features are projected onto the top-K principal subspace (K equal to the number of classes), and the residual is encouraged to be column-sparse. At test time, distance-based OOD scores (kNN and Mahalanobis) are computed against the projected training reference features. Experiments on CIFAR-10 with synthetic symmetric noise and on CIFAR-10N, CIFAR-100N, and Animal-10N report consistent improvements in FPR95 and AUROC over a broad set of baselines.

Significance. If the reported gains are real, the paper addresses a genuine and underexplored problem: most OOD detectors assume clean training labels, and the recent empirical study by Humblot-Renaux et al. shows that label noise strongly degrades them. The proposed low-rank projection idea is intuitive and could be practically useful. The paper also includes a useful set of experiments across multiple noise types and OOD benchmarks, and the ablation studies show the pipeline is internally coherent. However, the central empirical claim is weakened by a test-time protocol asymmetry in Section 3.3 and by baseline fairness issues acknowledged in the supplementary. The contribution is potentially publishable, but the current evidence does not establish that the low-rank decomposition itself is what drives the improvement.

major comments (4)
  1. [Section 3.3, Eq. (u_ID and u(x*))] The test-time scoring is asymmetric. Training reference embeddings are stored as u_ID(x_n) = (QQ^T h(x_n))/||QQ^T h(x_n)||, while test queries are embedded as u(x*) = h(x*)/||h(x*)|| with no projection onto Q. Every kNN and Mahalanobis distance in Tables 1 and 2 is therefore computed between vectors living in different spaces: references have the residual H - QQ^T H removed, queries do not. This asymmetry is not derived from the low-rank plus sparse model of Section 3.2 and can artificially inflate ID-OOD separability if OOD queries have substantial residual content. Since all headline results use this protocol, the paper does not establish that the proposed feature-cleaning mechanism is responsible for the gains; a simple mismatch artifact could be. The authors should either apply the same projection to test queries or justify why the unprojected query is correct, and rerun the experime
  2. [Supplementary, 'More Experiment Results'] The supplementary states that for CIDER and SSD+ the authors 'replace their default ResNet-18 encoder with DenseNet-101 for consistency' and that 'these methods may require additional fine-tuning to fully realize their potential.' In Tables 1 and 2, CIDER and SSD+ perform near random (e.g., FPR95 around 98-99%, AUROC around 30-43%), which is far below their reported performance in the original papers. This is an acknowledged disadvantage to two state-of-the-art distance-based OOD baselines. The claim in Section 4 of 'consistently outperforming all other approaches' is therefore not supported by a fair comparison. The authors should tune and validate CIDER and SSD+ with the same protocol used for their own method, or restrict the comparison to baselines that are not knowingly handicapped.
  3. [Section 4, Implementation Settings and Tables 1-2] Model selection is circular as reported. The text says 'We present the best performing variants of the NOODLE approach in the main result tables' and hyperparameter lambda is selected from {0.0001, 0.0005, 0.001, 0.005, 0.1}, with a choice between CM and SCE loss correction and between kNN and Mahalanobis metrics. No validation-based selection procedure, standard deviation, or number of seeds is given. Choosing the best variant per dataset and reporting only that result inflates performance estimates and makes the 'consistent outperformance' claim difficult to evaluate. The authors should report the full selection protocol, all variant results, error bars over multiple runs, and ideally fix the variant and lambda by validation before evaluating on test OOD sets.
  4. [Section 3.2] The central modeling assumption is that the low-rank component L = QQ^T H captures ID class structure and the residual H - QQ^T H is non-ID/outlier content. This assumption is load-bearing but not tested. Since Q is estimated from features trained on noisy labels and K is set to the number of classes, label noise can distort the top-K subspace, and the projection may discard genuine ID signal or retain noise. The paper provides no diagnostic evidence that the residual is actually non-ID content; in fact, the test-time asymmetry described above means the residual is never examined at test time. A concrete test would be to compare residual norms for ID versus OOD queries, or to compare performance using a projection estimated from clean labels as an oracle.
minor comments (4)
  1. [Section 3.1 and 3.2] The notation for the feature dimension is inconsistent: Eq. (1) uses D for input dimension, while Eq. (4) also uses D for the latent feature dimension. Please distinguish them.
  2. [Supplementary, Algorithm 2] The algorithm description is mostly clear, but the phrase 'right singular vectors' should be checked: with H in R^{N x D} and Q in R^{D x k}, Q spans the top-k right singular vectors, which is correct, but the variable name n_iter is used both as iteration count and as input to the algorithm. Minor clarity fix.
  3. [Throughout] There are several typos and inconsistent names: 'arobust' in the abstract, 'DenseNet-100' in supplementary tables versus 'DenseNet-101' in the main text, and 'CIDAR' instead of 'CIDER'. Please proofread.
  4. [Section 4, Tables 1 and 2] The tables are dense and hard to read because the top-two boldfaces are not explained clearly enough; also for real-noise datasets the formatting is cramped. Consider splitting tables or using a clearer visual distinction.

Circularity Check

0 steps flagged

No significant circularity: the method combines standard loss correction with an explicit low-rank projection assumption; the only self-citation is a non-load-bearing general reference.

full rationale

The derivation chain is self-contained. The loss correction module (Sec. 3.1) is drawn from standard transition-matrix and robust-loss literature, and no step defines the OOD score in terms of the fitted transition matrix or low-rank projector. The low-rank plus sparse decomposition (Sec. 3.2) is a stated modeling assumption; the projector Q is estimated from the same noisy features, but that is joint training rather than a definitional reduction. The reported OOD scores (Sec. 3.3) compare projected training references to unprojected test queries, which is an inconsistency and a potential source of artifact, but it is not circular: the score is neither equal to the fitted parameter nor obtained by renaming an input. Hyperparameter/variant selection ('we present the best performing variants') is a reporting/selection concern, not a circular derivation. The only self-citation, [36] by a co-author, appears in a list of transition-matrix works and is not load-bearing for NOODLE's design or results. Hence no step reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on the low-rank plus sparse model of features, the estimability of the transition matrix from noisy data (standard), and the choice of hyperparameters. The low-rank subspace is fitted per batch; no standalone theoretical justification is provided beyond intuition and UMAP plots.

free parameters (4)
  • lambda (sparsity regularizer) = searched over {0.0001,0.0005,0.001,0.005,0.1}; selected values not reported
    Controls the strength of the column-sparsity penalty on H_OOD in Eq. (5); selected by tuning, with selected values not reported.
  • rank K of low-rank subspace = number of classes (10 for CIFAR-10/Animal-10N, 100 for CIFAR-100)
    Chosen as the number of classes, based on the assumption that clean features have K clusters. No sensitivity analysis is provided.
  • transition matrix T = learned, initialized to identity
    Part of the CM loss correction; estimated from noisy data via the optimization in Algorithm 1. The update rule is not fully specified.
  • NOODLE variant (loss correction and distance metric) = best among CM/SCE x kNN/Mahalanobis
    Main tables report the best-performing variant, which is selected after evaluation, a form of test-set model selection.
axioms (4)
  • domain assumption Latent features of clean ID data are approximately low-rank with rank equal to the number of classes
    Sec 3.2: 'in the absence of label noise, latent feature vectors naturally exhibit certain clustering patterns, reflecting their low-rank structure'. Used to justify the decomposition H = QQ^T H + residual.
  • domain assumption The residual H - QQ^T H is column-sparse and corresponds to non-ID or outlier content
    Sec 3.2: 'the sparse term can handle the outlier data items that does not strictly conform to the low-rank assumption'. This is assumed, not derived, and underpins the regularizer L_sparse.
  • domain assumption The noise transition matrix T is identifiable and estimable from the noisy dataset using established CM methods
    Sec 3.1 relies on transition-matrix approaches [22,35,23,36]; the paper adopts CM and initializes T as identity. The theory is borrowed from prior literature.
  • standard math Training samples are i.i.d. and the OOD test set is disjoint from the ID label space
    Sec 2 formalizes the setting, standard for OOD detection.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Tackling the Noisy Elephant in the Room: Label Noise-robust Out-of-Distribution Detection via Loss Correction and Low-rank Decomposition." pith.science (2026). https://pith.science/paper/UFNBQ7PB

@misc{pith2026250906918,
  author       = {Pith},
  title        = {Pith review of: Tackling the Noisy Elephant in the Room: Label Noise-robust Out-of-Distribution Detection via Loss Correction and Low-rank Decomposition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UFNBQ7PB}},
  note         = {Machine review of arXiv:2509.06918}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Robust out-of-distribution (OOD) detection is an indispensable component of modern artificial intelligence (AI) systems, especially in safety-critical applications where models must identify inputs from unfamiliar classes not seen during training. While OOD detection has been extensively studied in the machine learning literature--with both post hoc and training-based approaches--its effectiveness under noisy training labels remains underexplored. Recent studies suggest that label noise can significantly degrade OOD performance, yet principled solutions to this issue are lacking. In this work, we demonstrate that directly combining existing label noise-robust methods with OOD detection strategies is insufficient to address this critical challenge. To overcome this, we propose a robust OOD detection framework that integrates loss correction techniques from the noisy label learning literature with low-rank and sparse decomposition methods from signal processing. Extensive experiments on both synthetic and real-world datasets demonstrate that our method significantly outperforms the state-of-the-art OOD detection techniques, particularly under severe noisy label settings.

Figures

Figures reproduced from arXiv: 2509.06918 by Shahana Ibrahim, Tarhib Al Azad.

Figure 1
Figure 1. Figure 1: The effect of label noise for OOD detection. The figure shows the UMAP representations [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The effect of label noise on OOD detection for robust methods. The figure shows the UMAP [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of OOD detection performance (FPR95 [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of OOD detection performance (FPR95 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

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

60 extracted references · 37 canonical work pages · 1 internal anchor

  1. [1]

    Explaining and harnessing adversar- ial examples.arXiv preprint arXiv:1412.6572, 2014

    Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversar- ial examples.arXiv preprint arXiv:1412.6572, 2014

    Pith/arXiv arXiv 2014

  2. [2]

    Are we ready for autonomous driving? the kitti vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In2012 IEEE conference on computer vision and pattern recognition, pages 3354–3361. IEEE, 2012

    2012

  3. [3]

    Waldstein, Ursula Schmidt-Erfurth, and Georg Langs

    Thomas Schlegl, Philipp Seeböck, Sebastian M. Waldstein, Ursula Schmidt-Erfurth, and Georg Langs. Unsupervised anomaly detection with generative adversarial networks to guide marker discovery. InInformation Processing in Medical Imaging, pages 146–157. Springer International Publishing, 2017

    work page 2017

  4. [4]

    A baseline for detecting misclassified and out-of-distribution examples in neural networks.arXiv preprint arXiv:1610.02136, 2016

    Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks.arXiv preprint arXiv:1610.02136, 2016

    Pith/arXiv arXiv 2016

  5. [5]

    Generalized out-of-distribution detection: A survey.International Journal of Computer Vision, 132(12):5635–5662, 2024

    Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey.International Journal of Computer Vision, 132(12):5635–5662, 2024

    2024

  6. [6]

    Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. InInternational Conference on Learning Representations (ICLR), 2018

    work page 2018

  7. [7]

    Scaling out-of-distribution detection for real- world settings

    Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joseph Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real- world settings. InProceedings of the 39th International Conference on Machine Learning (ICML), pages 8759–8773. PMLR, 2022

    work page 2022

  8. [8]

    Dice: Leveraging sparsification for out-of-distribution detection

    Yiyou Sun and Yixuan Li. Dice: Leveraging sparsification for out-of-distribution detection. In Computer Vision – ECCV 2022, pages 691–708. Springer Nature Switzerland, 2022

    work page 2022

  9. [9]

    React: Out-of-distribution detection with rectified activations

    Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. InAdvances in Neural Information Processing Systems (NeurIPS), pages 144–157. Curran Associates, Inc., 2021

    work page 2021

  10. [10]

    Neural mean discrep- ancy for efficient out-of-distribution detection

    Xin Dong, Junfeng Guo, Ang Li, Wei-Te Ting, Cong Liu, and HT Kung. Neural mean discrep- ancy for efficient out-of-distribution detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19217–19227, 2022

    work page 2022

  11. [11]

    A simple unified framework for detecting out-of-distribution samples and adversarial attacks

    Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. InAdvances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., 2018

    work page 2018

  12. [12]

    Out-of-distribution detection with deep nearest neighbors

    Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. InProceedings of the 39th International Conference on Machine Learning (ICML), pages 20827–20840. PMLR, 2022

    work page 2022

  13. [13]

    How to exploit hyperspherical embed- dings for out-of-distribution detection?, 2023

    Yifei Ming, Yiyou Sun, Ousmane Dia, and Yixuan Li. How to exploit hyperspherical embed- dings for out-of-distribution detection?, 2023

    work page 2023

  14. [14]

    SSD: A unified framework for self- supervised outlier detection.CoRR, abs/2103.12051, 2021

    Vikash Sehwag, Mung Chiang, and Prateek Mittal. SSD: A unified framework for self- supervised outlier detection.CoRR, abs/2103.12051, 2021

    Pith/arXiv arXiv 2021

  15. [15]

    How to overcome curse-of-dimensionality for out-of-distribution detection?Proceedings of the AAAI Conference on Artificial Intelligence, 38(18):19849–19857, Mar

    Soumya Suvra Ghosal, Yiyou Sun, and Yixuan Li. How to overcome curse-of-dimensionality for out-of-distribution detection?Proceedings of the AAAI Conference on Artificial Intelligence, 38(18):19849–19857, Mar. 2024

    work page 2024

  16. [16]

    Amazon’s mechanical turk: A new source of inexpensive, yet high-quality data? 2016

    Michael Buhrmester, Tracy Kwang, and Samuel D Gosling. Amazon’s mechanical turk: A new source of inexpensive, yet high-quality data? 2016

    work page 2016

  17. [17]

    Galadrielle Humblot-Renaux, Sergio Escalera, and Thomas B Moeslund. A noisy elephant in the room: Is your out-of-distribution detector robust to label noise? InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22626–22636, 2024. 10

    work page 2024

  18. [18]

    Learning from noisy labels with deep neural networks: A survey.IEEE transactions on neural networks and learning systems, 34(11):8135–8153, 2022

    Hwanjun Song, Minseok Kim, Dongmin Park, Yooju Shin, and Jae-Gil Lee. Learning from noisy labels with deep neural networks: A survey.IEEE transactions on neural networks and learning systems, 34(11):8135–8153, 2022

    2022

  19. [19]

    Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, and Simon Lacoste-Julien

    Devansh Arpit, Stanisław Jastrzundefinedbski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S. Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, and Simon Lacoste-Julien. A closer look at memorization in deep networks. InProceedings of International Conference on Machine Learning, page 233–242, 2017

    work page 2017

  20. [20]

    Understanding deep learning requires rethinking generalization

    Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. InProceedings of International Conference on Learning Representations, 2016

    work page 2016

  21. [21]

    Classification with noisy labels by importance reweighting

    Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38:447–461, 2016

    work page 2016

  22. [22]

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

    Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen 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

    work page 2017

  23. [23]

    Provably end-to-end label-noise learning without anchor points

    Xuefeng Li, Tongliang Liu, Bo Han, Gang Niu, and Masashi Sugiyama. Provably end-to-end label-noise learning without anchor points. InProceedings of International Conference on Machine Learning, pages 6403–6413, 2021

    work page 2021

  24. [24]

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

    Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance- dependent label noise. InAdvances in Neural Information Processing Systems, volume 33, pages 7597–7610, 2020

    work page 2020

  25. [25]

    Estimating instance-dependent Bayes-label transition matrix using a deep neural network

    Shuo Yang, Erkun Yang, Bo Han, Yang Liu, Min Xu, Gang Niu, and Tongliang Liu. Estimating instance-dependent Bayes-label transition matrix using a deep neural network. InProceedings of International Conference on Machine Learning, 2021

    work page 2021

  26. [26]

    Learning with bounded instance and label-dependent label noise

    Jiacheng Cheng, Tongliang Liu, Kotagiri Ramamohanarao, and Dacheng Tao. Learning with bounded instance and label-dependent label noise. InProceedings of International Conference on Machine Learning, volume 119, pages 1789–1799, 2020

    work page 2020

  27. [27]

    Generalized cross entropy loss for training deep neural networks with noisy labels.Advances in neural information processing systems, 31, 2018

    Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels.Advances in neural information processing systems, 31, 2018

    2018

  28. [28]

    Curriculum loss: Robust learning and generalization against label corruption.arXiv preprint arXiv:1905.10045, 2019

    Yueming Lyu and Ivor W Tsang. Curriculum loss: Robust learning and generalization against label corruption.arXiv preprint arXiv:1905.10045, 2019

    Pith/arXiv arXiv 1905

  29. [29]

    Symmetric cross entropy for robust learning with noisy labels

    Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. InProceedings of the IEEE/CVF international conference on computer vision, pages 322–330, 2019

    2019

  30. [30]

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

    Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. InInternational conference on machine learning, pages 2304–2313. PMLR, 2018

    work page 2018

  31. [31]

    How does disagreement help generalization against label corruption? InInternational conference on machine learning, pages 7164–7173

    Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does disagreement help generalization against label corruption? InInternational conference on machine learning, pages 7164–7173. PMLR, 2019

    2019

  32. [32]

    Self: Learning to filter noisy labels with self- ensembling.arXiv preprint arXiv:1910.01842, 2019

    Duc Tam Nguyen, Chaithanya Kumar Mummadi, Thi Phuong Nhung Ngo, Thi Hoai Phuong Nguyen, Laura Beggel, and Thomas Brox. Self: Learning to filter noisy labels with self- ensembling.arXiv preprint arXiv:1910.01842, 2019

    Pith/arXiv arXiv 1910

  33. [33]

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

    Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 11

    work page 2018

  34. [34]

    Dividemix: Learning with noisy labels as semi-supervised learning.arXiv preprint arXiv:2002.07394, 2020

    Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning.arXiv preprint arXiv:2002.07394, 2020

    Pith/arXiv arXiv 2002

  35. [35]

    Learning from noisy labels by regularized estimation of annotator confusion

    Ryutaro Tanno, Ardavan Saeedi, Swami Sankaranarayanan, Daniel C Alexander, and Nathan Silberman. Learning from noisy labels by regularized estimation of annotator confusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11244–11253, 2019

    work page 2019

  36. [36]

    Deep learning from crowdsourced labels: Coupled cross-entropy minimization, identifiability, and regularization

    Shahana Ibrahim, Tri Nguyen, and Xiao Fu. Deep learning from crowdsourced labels: Coupled cross-entropy minimization, identifiability, and regularization. InProceedings of International Conference on Learning Representations, 2023

    work page 2023

  37. [37]

    Robust principal component analysis?Journal of the ACM (JACM), 58(3):1–37, 2011

    Emmanuel J Candès, Xiaodong Li, Yi Ma, and John Wright. Robust principal component analysis?Journal of the ACM (JACM), 58(3):1–37, 2011

    2011

  38. [38]

    Image classification by non-negative sparse coding, low-rank and sparse decomposition

    Chunjie Zhang, Jing Liu, Qi Tian, Changsheng Xu, Hanqing Lu, and Songde Ma. Image classification by non-negative sparse coding, low-rank and sparse decomposition. InCVPR 2011, pages 1673–1680. IEEE, 2011

    work page 2011

  39. [39]

    Candes, Xiaodong Li, Yi Ma, and John Wright

    Emmanuel J. Candes, Xiaodong Li, Yi Ma, and John Wright. Robust principal component analysis?, 2009

    work page 2009

  40. [40]

    A new alternating minimization algorithm for total variation image reconstruction.SIAM J

    Yilun Wang, Junfeng Yang, Wotao Yin, and Yin Zhang. A new alternating minimization algorithm for total variation image reconstruction.SIAM J. Imaging Sciences, 1:248–272, 01 2008

    work page 2008

  41. [41]

    A randomized algorithm for principal component analysis.SIAM Journal on Matrix Analysis and Applications, 31(3):1100–1124, 2010

    Vladimir Rokhlin, Arthur Szlam, and Mark Tygert. A randomized algorithm for principal component analysis.SIAM Journal on Matrix Analysis and Applications, 31(3):1100–1124, 2010

    work page 2010

  42. [42]

    Subspace iteration randomization and singular value problems.SIAM Journal on Scientific Computing, 37(3):A1139–A1173, 2015

    Ming Gu. Subspace iteration randomization and singular value problems.SIAM Journal on Scientific Computing, 37(3):A1139–A1173, 2015

    work page 2015

  43. [43]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009

    2009

  44. [44]

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

    Jiaheng Wei, Zhaowei Zhu, Hao Cheng, Tongliang Liu, Gang Niu, and Yang Liu. Learning with noisy labels revisited: A study using real-world human annotations. InInternational Conference on Learning Representations, 2022

    2022

  45. [45]

    SELFIE: Refurbishing unclean samples for robust deep learning

    Hwanjun Song, Minseok Kim, and Jae-Gil Lee. SELFIE: Refurbishing unclean samples for robust deep learning. InICML, 2019

    work page 2019

  46. [46]

    Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y . Ng. Reading digits in natural images with unsupervised feature learning. InNIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011

    2011

  47. [47]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms.CoRR, abs/1708.07747, 2017

    Han Xiao, Kashif Rasul, and Roland V ollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms.CoRR, abs/1708.07747, 2017

    Pith/arXiv arXiv 2017

  48. [48]

    Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop, 2016

    Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop, 2016

    work page 2016

  49. [49]

    End-to-end Convolutional Network for Saliency Prediction

    Junting Pan and Xavier Giró-i-Nieto. End-to-end convolutional network for saliency prediction. CoRR, abs/1507.01422, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  50. [50]

    Describing textures in the wild.CoRR, abs/1311.3618, 2013

    Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild.CoRR, abs/1311.3618, 2013

    Pith/arXiv arXiv 2013

  51. [51]

    Places: An image database for deep scene understanding.CoRR, abs/1610.02055, 2016

    Bolei Zhou, Aditya Khosla, Àgata Lapedriza, Antonio Torralba, and Aude Oliva. Places: An image database for deep scene understanding.CoRR, abs/1610.02055, 2016

    Pith/arXiv arXiv 2016

  52. [52]

    A baseline for detecting misclassified and out-of-distribution examples in neural networks

    Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 12

    work page 2017

  53. [53]

    Energy-based out-of- distribution detection.ArXiv, abs/2010.03759, 2020

    Weitang Liu, Xiaoyun Wang, John Douglas Owens, and Yixuan Li. Energy-based out-of- distribution detection.ArXiv, abs/2010.03759, 2020

    Pith/arXiv arXiv 2010

  54. [54]

    Supervised contrastive learning.Advances in neural information processing systems, 33:18661–18673, 2020

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning.Advances in neural information processing systems, 33:18661–18673, 2020

    2020

  55. [55]

    Provably end-to-end label-noise learning without anchor points

    Xuefeng Li, Tongliang Liu, Bo Han, Gang Niu, and Masashi Sugiyama. Provably end-to-end label-noise learning without anchor points. InInternational Conference on Machine Learning, 2021

    work page 2021

  56. [56]

    Symmetric cross entropy for robust learning with noisy labels, 2019

    Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels, 2019

    2019

  57. [57]

    Zhilu Zhang and Mert R. Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels, 2018

    work page 2018

  58. [58]

    Junnan Li, Richard Socher, and Steven C. H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning, 2020

    work page 2020

  59. [59]

    Co-teaching: Robust training of deep neural networks with extremely noisy labels, 2018

    Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels, 2018

    work page 2018

  60. [60]

    Tackling the Noisy Elephant in the Room: Label Noise-robust Out-of-Distribution Detection via Loss Correction and Low-rank Decomposition

    Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks. CoRR, abs/1608.06993, 2016. 13 Supplementary Material of “Tackling the Noisy Elephant in the Room: Label Noise-robust Out-of-Distribution Detection via Loss Correction and Low-rank Decomposition” A Notation We use the following notation throughout the paper: x, x, X...

    Pith/arXiv arXiv 2016

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.