Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Trust-Aware Diversion for Data-Effective Distillation

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper proposes Trust-Aware Diversion (TAD), a dual-loop wrapper that lets dataset distillation survive mislabeled data by recalibrating untrusted samples instead of discarding them.

desk verdict A novel problem framing and extensive experiments, but the inner-loop formula as written inverts the reliability weighting, so the method description cannot be reproduced until it is corrected. read the letter →

arxiv 2502.05027 v1 pith:ZS3NSQL3 submitted 2025-02-07 cs.CV

Zhuojie Wu , Yanbin Liu , Xin Shen , Xiaofeng Cao , Xin Yu This is my paper · ORCID
classification cs.CV
keywords datasetdistillationnoisylabelstrust-awarediversiontrajectorymatchinglabelnoiserobustnessdual-loopoptimizationMahalanobisdistancepseudo-labelrecalibration
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

Dataset distillation compresses a large training set into a small synthetic one by matching the training trajectories of models trained on each. This paper argues that the usual clean-label assumption breaks in real use, because mislabeled samples distort the expert trajectory that the synthetic data is matched to. The proposed Trust-Aware Diversion (TAD) wraps trajectory-matching distillation in two alternating loops: an outer loop that uses per-sample loss to separate trusted from untrusted samples and trains only on trusted ones, and an inner loop that scores trusted samples by their Mahalanobis distance to synthetic anchors and uses the most reliable ones to pseudo-label and recover untrusted samples. The paper claims that TAD consistently improves four existing distillation methods on CIFAR-10, CIFAR-100, and Tiny ImageNet under symmetric, asymmetric, and real-world label noise, with the largest gains at higher noise ratios and larger images-per-class settings.

What carries the argument

The load-bearing object is the dual-loop trust diversion. The outer loop is a two-mode Gaussian mixture model on per-sample cross-entropy losses, with per-class dynamic thresholds equal to the mean trusted posterior for each class; it reroutes training and distillation toward trusted samples. The inner loop is a reliability score $M_i^c = D_M(x_i, \mu_c) = \sqrt{(x_i-\mu_c)^T \Sigma_c^{-1} (x_i-\mu_c)}$, the Mahalanobis distance from a trusted sample to the mean and covariance of synthetic anchors in its class, followed by a pseudo-label score $p_c(x'_j) = \sum_{i=1}^{k} M_i^c\,\mathrm{sim}(x'_j, x_i)\,\delta(y_i=c)$ that uses the top-$k$ reliable trusted samples to recalibrate untrusted ones. The mechanism makes the small distilled set itself the proxy for clean class structure, so that samples the outer loop misclassifies can still be recovered rather than wasted.

What would settle it

Run TAD at IPC=1 on CIFAR-100 with 40% symmetric noise and inspect the inner loop: Eq. (7) needs an invertible per-class covariance, but one anchor per class gives no covariance estimate at all. If the reported IPC=1 gains come entirely from the outer loop or from an unstated fallback, the claim that inner-loop recalibration drives the improvement is false.

Watch

Extended reading notes

Core claim

The paper's central claim is that noisy labels break trajectory-matching dataset distillation by dragging the expert trajectory away from the clean-data path, and that the damage can be repaired by a dual-loop trust-aware diversion rather than by a separate denoising preprocessing stage. In the outer loop, a two-component Gaussian mixture model fit to per-sample cross-entropy losses produces class-wise dynamic thresholds that divide the data into trusted and untrusted spaces; the expert model is trained only on trusted samples with a consistency regularizer, keeping the trajectory clean. In the inner loop, synthetic images distilled from early training phases serve as class anchors, and each trusted sample gets a reliability score from its Mahalanobis distance to its class's anchor distribution; the top-$k$ reliable samples then assign pseudo-labels to untrusted samples by weighted cosine similarity. Iterating the two loops expands the trusted space and shrinks the untrusted space. On the paper's own numbers, this lifts e.g. CIFAR-100 with 40% symmetric noise at IPC=10 from 32.6% accuracy for ATT to 41.5% with TAD, and improves MTT, FTD, and DATM similarly.

Load-bearing premise

The method assumes that the tiny set of synthetic images created in early training reliably represents each class, so that measuring how far real samples sit from those images' average and spread is meaningful; with only one synthetic image per class there is no spread to measure.

Editorial extensions

If this is right

  • Trajectory-matching distillation methods can inherit robustness to label noise without changing their matching loss, because TAD operates on the data split and label calibration before matching.
  • A separate denoising stage is not necessary; TAD's recalibration converts untrusted samples into useful training data, which the two-stage denoise-then-distill baseline cannot do.
  • The benefit grows with noise ratio and IPC: deeper trajectory matching exposes the student to later, noise-corrupted parts of the expert trajectory, where diversion matters most.
  • TAD transfers to human-annotated noisy benchmarks such as CIFAR-10N and CIFAR-100N, not just synthetic noise, improving all tested baselines at IPC=10 and IPC=50.

Reading between the lines

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

  • Beyond the paper: the same diversion loop should apply to gradient-matching and distribution-matching distillation, because noise corrupts their gradients and distribution estimates just as it corrupts trajectories.
  • Beyond the paper: the inner loop's accuracy depends on the self-supervised feature extractor, so a domain shift between pretraining and target data is a likely failure mode worth testing.
  • Beyond the paper: a shrinkage or pseudo-inverse estimator for $\Sigma_c$ would let TAD work at IPC=1, where a single anchor per class currently leaves the covariance in Eq. (7) undefined.
  • Beyond the paper: the dual loop resembles an EM-style alternation between sample reweighting and label recalibration, so the convergence of the trusted/untrusted partition could be studied directly rather than only through final accuracy.
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 / 6 minor

Summary. The paper proposes Trust-Aware Diversion (TAD), a dual-loop dataset distillation method designed for settings with noisy labels. The outer loop fits a two-component Gaussian Mixture Model to per-sample cross-entropy losses, uses class-wise dynamic thresholds to split data into trusted and untrusted spaces, and trains the expert model on trusted samples with an added consistency regularizer. The inner loop uses synthetic anchors produced by early-trajectory matching to compute Mahalanobis distances, selects top-k reliable trusted samples per class, and pseudo-labels untrusted samples via cosine similarity weighted by the reliability score. Experiments on CIFAR-10, CIFAR-100, and Tiny ImageNet under symmetric, asymmetric, and real-world noise report consistent improvements over MTT, FTD, DATM, and ATT, with additional ablations and comparisons against learning-with-noisy-labels baselines.

Significance. If the method is implemented as intended, the paper addresses a relevant gap—dataset distillation under label noise—and supports its claims with extensive experiments across three benchmarks and three noise types. The evaluation is external (test accuracy on clean test sets), so there is no logical circularity in the main comparisons. The breadth of experiments, the inclusion of real-world noisy label sets, and the ablations are clear strengths. However, the central inner-loop equation is written with an inverted reliability weighting, and the IPC=1 setting is under-specified; these issues must be resolved before the empirical claims can be fully credited.

major comments (4)
  1. [Section 4.2, Eq. (8)] The pseudo-label score p_c(x'_j) multiplies the cosine similarity by M_i^c, where M_i^c is defined immediately above as the Mahalanobis distance D_M(x_i, mu_c) from Eq. (7). Under this definition, a larger M_i^c corresponds to a trusted sample farther from the class anchor distribution, i.e., a less reliable sample. Equation (8) therefore assigns larger weight to less reliable trusted samples when voting on the pseudo-label of an untrusted sample, which is the reverse of the stated top-k-reliable selection and of the description of M_i^c as a reliability score. The min-max normalization to [0,1] does not change the direction of this weighting. Please replace M_i^c with a monotonically decreasing function of the normalized distance (e.g., 1 - normalized D_M, or exp(-D_M)) or explicitly redefine M_i^c as a similarity score. Because Eq. (8) is the only mechanism that recalibrates untrusted samples, the reported gains are not reproducible from the text as written.
  2. [Section 4.2, Eq. (7); Tables 1 and 2] The per-class covariance matrix Sigma_c is estimated from anchors of that class, with n anchors per class (N_A = n * C, n = IPC). At IPC=1, each class has a single anchor, so Sigma_c is a zero matrix or undefined, and the Mahalanobis distance in Eq. (7) is not computable without regularization, shrinkage, or a pseudo-inverse. The paper reports IPC=1 results for all datasets and noise settings but does not describe any such procedure. Please specify how Sigma_c is computed for n=1, or explicitly restrict the method's applicability to n >= 2.
  3. [Section 4.1, Eq. (6)] The consistent regularization term L_C contains log(y'_i), which is undefined whenever y'_i = 0. The statement 'log(0) is defined as a negative constant' is not a precise specification: the actual value of this constant changes the gradient of the entire objective and thus affects the outer-loop partition and the resulting distilled data. Please specify the exact constant (or an alternative smoothing such as label clipping) and confirm that the reported experiments use that choice.
  4. [Section 4.2] The anchor distribution is itself derived from synthetic images distilled on the current trusted set, so misclassified trusted samples can bias the anchor means and covariances used for Mahalanobis scoring. The paper acknowledges that noisy samples can remain in the trusted space, but it does not analyze how such contamination propagates through the reliability scoring. Please provide an experiment or analysis that varies anchor quality (e.g., anchors synthesized from the clean subset, or anchors from earlier versus later trajectories) to demonstrate that the reported gains do not depend on a favorable bootstrap.
minor comments (6)
  1. [Section 1] The phrase 'These highly-reliabile samples' contains a typo; it should be 'These highly reliable samples'.
  2. [Section 4.2] The notation M_i^c is introduced as 'M c i' and then used in Eq. (8); please define it consistently as a scalar reliability score per trusted sample and class, and distinguish it clearly from the raw Mahalanobis distance.
  3. [Section 5.2] The feature extractor is described as 'ResNet' pre-trained 'in a self-supervision manner (e.g., SimCLR, MoCo)'; SimCLR and MoCo are training objectives rather than architectures, so the paper should specify which extractor and which pretraining are used in each experiment.
  4. [Table 1] Several entries for Tiny ImageNet at IPC=50 (FTD and ATT) are missing and replaced by backslashes; please report these values or explain why they are absent.
  5. [Eq. (6)] The use of y'_i as both a one-hot label vector and a scalar index (e.g., 'For y'_i = 1' and 'For y'_i = 0') should be clarified to avoid ambiguity.
  6. [Section 4.1, Eq. (5)] The claim that the class-wise mean confidence 'eliminates the effect of outliers' is imprecise, since the mean is itself sensitive to outliers; consider saying 'reduces the effect' or justify why the mean is preferred.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: the reported accuracy gains are external benchmark results, and the method's internal self-training bootstrap does not make the evaluation circular.

full rationale

The paper's central claim is empirical: adding TAD to trajectory-matching distillation improves held-out test accuracy under label noise. Test accuracy is measured on fixed test sets that are not used in any optimization or threshold selection, so the claim cannot reduce to its inputs by construction. The iterative trusted/untrusted split and the use of synthetic anchors produced by distillation on the current trusted set form a self-referential bootstrap, but this is a standard self-training-style algorithm design rather than a logical circularity: the GMM parameters, thresholds, and anchors are not fit to test data, and the reported accuracies are externally evaluated. The paper's reliance on prior results (memorization effect, GMM-based loss modeling, trajectory matching, Mahalanobis distance) is external and does not form a self-citation chain that forces the conclusion. The possible inversion in Eq. 8, where a larger Mahalanobis distance appears to receive a larger pseudo-label weight despite being called a reliability score, is a correctness or description issue, not a case of a prediction reducing to its input by definition. Similarly, the IPC=1 singular-covariance omission is an implementation gap, not circularity. Overall, no load-bearing step equates an output with an input or renames a fit as a prediction.

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

The method depends on several unvalidated modeling assumptions, especially the bimodal loss distribution and the reliability of synthetic anchors for covariance estimation. There are at least five free hyperparameters that are not reported. No new physical or conceptual entities are introduced.

free parameters (5)
  • lambda (regularization coefficient in Eq. 6) = not reported
    Balances the cross-entropy loss and the reverse cross-entropy regularizer. Not specified in the paper; likely tuned per dataset, affecting the trusted/untrusted dynamics.
  • top-k reliable samples = not reported
    Number of the most reliable trusted samples used to assign pseudo-labels in Eq. 8. Not disclosed; directly controls the pseudo-label quality.
  • number of outer/inner loop iterations = not reported
    The iterative dual-loop is claimed to expand trusted space, but no schedule or stopping criterion is given.
  • early trajectory anchor length = first four epochs (stated, but not justified)
    The paper uses only the initial training phases to synthesize anchors; this choice is stated but not supported by an ablation.
  • feature extractor choice (SimCLR vs MoCo) = not fixed
    The Mahalanobis distance computations depend on the self-supervised features; the paper does not specify which one is used or whether results vary.
assumptions (5)
  • domain assumption The per-sample loss distribution is a two-mode Gaussian mixture, with one mode for trusted (clean) and one for untrusted (noisy) samples.
    Invoked in Eq. 4 of Section 4.1, motivated by Fig. 2, but not validated under asymmetric noise or at high noise ratios.
  • domain assumption Deep networks memorize clean samples before noisy ones, so early training losses separate clean from noisy data.
    Cited from prior work and used as the basis of the outer loop, but the separation quality is not measured.
  • domain assumption Synthetic anchors from early-trajectory matching capture class distributions accurately enough for Mahalanobis distance computations.
    Used in Section 4.2 before Eq. 7; with IPC=1 the class covariance is not estimable.
  • domain assumption Cosine similarity in a self-supervised feature space corresponds to label agreement.
    Used to assign pseudo-labels in Eq. 8; no evidence is given that the feature space separates classes under noisy labels.
  • ad hoc to paper log(0) is treated as a negative constant in the regularization term of Eq. 6.
    The paper explicitly defines log(0) as a negative constant for numerical stability; this alters the loss landscape and is not a standard operation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Trust-Aware Diversion for Data-Effective Distillation." pith.science (2026). https://pith.science/paper/ZS3NSQL3

@misc{pith2026250205027,
  author       = {Pith},
  title        = {Pith review of: Trust-Aware Diversion for Data-Effective Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZS3NSQL3}},
  note         = {Machine review of arXiv:2502.05027}
}
read the original abstract

Dataset distillation compresses a large dataset into a small synthetic subset that retains essential information. Existing methods assume that all samples are perfectly labeled, limiting their real-world applications where incorrect labels are ubiquitous. These mislabeled samples introduce untrustworthy information into the dataset, which misleads model optimization in dataset distillation. To tackle this issue, we propose a Trust-Aware Diversion (TAD) dataset distillation method. Our proposed TAD introduces an iterative dual-loop optimization framework for data-effective distillation. Specifically, the outer loop divides data into trusted and untrusted spaces, redirecting distillation toward trusted samples to guarantee trust in the distillation process. This step minimizes the impact of mislabeled samples on dataset distillation. The inner loop maximizes the distillation objective by recalibrating untrusted samples, thus transforming them into valuable ones for distillation. This dual-loop iteratively refines and compensates for each other, gradually expanding the trusted space and shrinking the untrusted space. Experiments demonstrate that our method can significantly improve the performance of existing dataset distillation methods on three widely used benchmarks (CIFAR10, CIFAR100, and Tiny ImageNet) in three challenging mislabeled settings (symmetric, asymmetric, and real-world).

Figures

Figures reproduced from arXiv: 2502.05027 by the authors.

Figure 1
Figure 1. Illustration of our proposed Trust-Aware Diversion (TAD) dataset distillation method. The outer loop separates data into trusted and untrusted spaces, rerouting distillation toward trusted samples. The inner loop recalibrates untrusted samples and transposes them to the trusted space. Through iterative interac￾tions, the trusted space expands while the untrusted space shrinks, improving dataset distillation under no… view at source ↗
Figure 2
Figure 2. Training loss and test accuracy curves of an expert model on CIFAR-10 with clean and 40% symmetric noisy labels. Noise impedes convergence and significantly degrades performance. spaces to ensure trustworthy distillation, while the inner loop recalibrates untrusted samples, refining them into valu￾able data for distillation. Through this iterative interaction, TAD can reliably refine the distinction between trusted … view at source ↗
Figure 3
Figure 3. Overview of the proposed Trust-Aware Diversion (TAD) dataset distillation method. TAD introduces a dual-loop optimization framework for trustworthy dataset distillation. The outer loop divides data into trusted and untrusted spaces, rerouting distillation toward reliable samples, while the inner loop refines untrusted samples for potential reuse. Through iterative interaction, the two loops progressively expand the … view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. The Evolution of Dataset Distillation: Toward Scalable and Generalizable Solutions

    cs.CV 2025-02 conditional novelty 3.0 of 10

    A 2023-2025 survey of dataset distillation that organizes matching, generative, decoupling, and selective methods and tabulates ImageNet-scale accuracy comparisons.

Reference graph

Works this paper leans on

61 extracted references · 46 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    E., and McGuinness, K

    Albert, P., Ortego, D., Arazo, E., O'Connor, N. E., and McGuinness, K. Addressing out-of-distribution label noise in webly-labelled data. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pp.\ 392--401, 2022

    work page 2022

  3. [3]

    Unsupervised label noise modeling and loss correction

    Arazo, E., Ortego, D., Albert, P., O’Connor, N., and McGuinness, K. Unsupervised label noise modeling and loss correction. In International conference on machine learning, pp.\ 312--321. PMLR, 2019

    work page 2019

  4. [4]

    S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., et al

    Arpit, D., Jastrz e bski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M. S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., et al. A closer look at memorization in deep networks. In International conference on machine learning, pp.\ 233--242. PMLR, 2017

    work page 2017

  5. [5]

    A., and Zhu, J.-Y

    Cazenavette, G., Wang, T., Torralba, A., Efros, A. A., and Zhu, J.-Y. Dataset distillation by matching training trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

    work page 2022

  6. [6]

    A simple framework for contrastive learning of visual representations

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

    2020

  7. [7]

    Rethinking backdoor attacks on dataset distillation: A kernel method perspective

    Chung, M.-Y., Chou, S.-Y., Yu, C.-M., Chen, P.-Y., Kuo, S.-Y., and Ho, T.-Y. Rethinking backdoor attacks on dataset distillation: A kernel method perspective. In Proceedings of the International Conference on Learning Representations (ICLR), 2024

    work page 2024

  8. [8]

    Beyond mahalanobis distance for textual ood detection

    Colombo, P., Dadalto, E., Staerman, G., Noiry, N., and Piantanida, P. Beyond mahalanobis distance for textual ood detection. Advances in Neural Information Processing Systems, 35: 0 17744--17759, 2022

    work page 2022

Show all 61 references
  1. [9]

    Dc-bench: Dataset condensation benchmark

    Cui, J., Wang, R., Si, S., and Hsieh, C.-J. Dc-bench: Dataset condensation benchmark. Advances in Neural Information Processing Systems, 35: 0 810--822, 2022

    2022

  2. [10]

    De Maesschalck, R., Jouan-Rimbaud, D., and Massart, D. L. The mahalanobis distance. Chemometrics and intelligent laboratory systems, 50 0 (1): 0 1--18, 2000

    2000

  3. [11]

    Exploiting inter-sample and inter-feature relations in dataset distillation

    Deng, W., Li, W., Ding, T., Wang, L., Zhang, H., Huang, K., Huo, J., and Gao, Y. Exploiting inter-sample and inter-feature relations in dataset distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 17057--17066, 2024

    2024

  4. [12]

    Y., Zhou, J

    Du, J., Jiang, Y., Tan, V. Y., Zhou, J. T., and Li, H. Minimizing the accumulated trajectory error to improve dataset distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 3749--3758, 2023

    2023

  5. [13]

    Ghosh, A., Kumar, H., and Sastry, P. S. Robust loss functions under label noise for deep neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 31, 2017

    2017

  6. [14]

    Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning

    Goswami, D., Liu, Y., Twardowski, B., and van de Weijer, J. Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning. Advances in Neural Information Processing Systems, 36, 2024

    2024

  7. [15]

    Summarizing stream data for memory-restricted online continual learning

    Gu, J., Wang, K., Jiang, W., and You, Y. Summarizing stream data for memory-restricted online continual learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2024

    2024

  8. [16]

    Towards lossless dataset distillation via difficulty-aligned trajectory matching

    Guo, Z., Wang, K., Cazenavette, G., Li, H., Zhang, K., and You, Y. Towards lossless dataset distillation via difficulty-aligned trajectory matching. In The Twelfth International Conference on Learning Representations, 2024

    2024

  9. [17]

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

    Han, B., Yao, Q., Yu, X., Niu, G., Xu, M., Hu, W., Tsang, I., and Sugiyama, M. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018

    2018

  10. [18]

    Deep residual learning for image recognition

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

    2016

  11. [19]

    Momentum contrast for unsupervised visual representation learning

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

    2020

  12. [20]

    Distilling the knowledge in a neural network

    Hinton, G. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015

    2015 arXiv

  13. [21]

    N., Rahnavard, N., Mian, A., and Shah, M

    Karim, N., Rizve, M. N., Rahnavard, N., Mian, A., and Shah, M. Unicon: Combating label noise through uniform selection and contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 9676--9686, 2022

    2022

  14. [22]

    J., Yun, S., Song, H., Jeong, J., Ha, J.-W., and Song, H

    Kim, J.-H., Kim, J., Oh, S. J., Yun, S., Song, H., Jeong, J., Ha, J.-W., and Song, H. O. Dataset condensation via efficient synthetic-data parameterization. In Proceedings of the International Conference on Machine Learning (ICML), pp.\ 11102--11118, 2022

    2022

  15. [23]

    Dataset condensation with contrastive signals

    Lee, S., Chun, S., Jung, S., Yun, S., and Yoon, S. Dataset condensation with contrastive signals. In International Conference on Machine Learning, pp.\ 12352--12364. PMLR, 2022

    2022

  16. [24]

    Dataset distillation for medical dataset sharing

    Li, G., Togo, R., Ogawa, T., and Haseyama, M. Dataset distillation for medical dataset sharing. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), Workshop, pp.\ 1--6, 2023

    2023

  17. [25]

    Li, J., Socher, R., and Hoi, S. C. Dividemix: Learning with noisy labels as semi-supervised learning. In International Conference on Learning Representations, 2020

    2020

  18. [26]

    Neighborhood collective estimation for noisy label identification and correction

    Li, J., Li, G., Liu, F., and Yu, Y. Neighborhood collective estimation for noisy label identification and correction. In European Conference on Computer Vision, pp.\ 128--145. Springer, 2022

    2022

  19. [27]

    and Kainz, B

    Li, Z. and Kainz, B. Image distillation for safe data sharing in histopathology. In International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2024

    2024

  20. [28]

    N., Wang, K., et al

    Li, Z., Guo, Z., Zhao, W., Zhang, T., Cheng, Z.-Q., Khaki, S., Zhang, K., Sajed, A., Plataniotis, K. N., Wang, K., et al. Prioritize alignment in dataset distillation. arXiv preprint arXiv:2408.03360, 2024

    2024 arXiv

  21. [29]

    Dataset distillation by automatic training trajectories

    Liu, D., Gu, J., Cao, H., Trinitis, C., and Schulz, M. Dataset distillation by automatic training trajectories. In European Conference on Computer Vision. Springer, 2024

    2024

  22. [30]

    Dataset distillation via the wasserstein metric

    Liu, H., Li, Y., Xing, T., Dalal, V., Li, L., He, J., and Wang, H. Dataset distillation via the wasserstein metric. arXiv preprint arXiv:2311.18531, 2023

    2023 arXiv

  23. [31]

    Early-learning regularization prevents memorization of noisy labels

    Liu, S., Niles-Weed, J., Razavian, N., and Fernandez-Granda, C. Early-learning regularization prevents memorization of noisy labels. Advances in neural information processing systems, 33: 0 20331--20342, 2020

    2020

  24. [32]

    and D'yakonov, A

    Medvedev, D. and D'yakonov, A. Learning to generate synthetic training data using gradient matching and implicit differentiation. In Proceedings of the International Conference on Analysis of Images, Social Networks and Texts (AIST), pp.\ 138--150, 2021

    2021

  25. [33]

    Confident learning: Estimating uncertainty in dataset labels

    Northcutt, C., Jiang, L., and Chuang, I. Confident learning: Estimating uncertainty in dataset labels. Journal of Artificial Intelligence Research, 70: 0 1373--1411, 2021

    2021

  26. [34]

    A label is worth a thousand images in dataset distillation

    Qin, T., Deng, Z., and Alvarez-Melis, D. A label is worth a thousand images in dataset distillation. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  27. [35]

    Z., Lawryshyn, Y

    Sajedi, A., Khaki, S., Amjadian, E., Liu, L. Z., Lawryshyn, Y. A., and Plataniotis, K. N. DataDAM : Efficient dataset distillation with attention matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 17097--17107, 2023

    2023

  28. [36]

    Sample condensation in online continual learning

    Sangermano, M., Carta, A., Cossu, A., and Bacciu, D. Sample condensation in online continual learning. In 2022 International Joint Conference on Neural Networks (IJCNN), pp.\ 01--08. IEEE, 2022

    2022

  29. [37]

    Generalized large-scale data condensation via various backbone and statistical matching

    Shao, S., Yin, Z., Zhou, M., Zhang, X., and Shen, Z. Generalized large-scale data condensation via various backbone and statistical matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 16709--16718, 2024

    2024

  30. [38]

    FYI : Flip your images for dataset distillation

    Son, B., Oh, Y., Baek, D., and Ham, B. FYI : Flip your images for dataset distillation. In Proceedings of the European Conference on Computer Vision (ECCV), 2024

    2024

  31. [39]

    and Schonlau, M

    Sucholutsky, I. and Schonlau, M. Soft-label dataset distillation and text dataset distillation. In Proceedings of the International Joint Conference on Neural Networks (IJCNN), pp.\ 1--8, 2021

    2021

  32. [40]

    On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm

    Sun, P., Shi, B., Yu, D., and Lin, T. On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 9390--9399, 2024

    2024

  33. [41]

    Joint optimization framework for learning with noisy labels

    Tanaka, D., Ikami, D., Yamasaki, T., and Aizawa, K. Joint optimization framework for learning with noisy labels. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 5552--5560, 2018

    2018

  34. [42]

    Wang, T., Zhu, J.-Y., Torralba, A., and Efros, A. A. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018

    2018 arXiv

  35. [43]

    Combating noisy labels by agreement: A joint training method with co-regularization

    Wei, H., Feng, L., Chen, X., and An, B. Combating noisy labels by agreement: A joint training method with co-regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 13726--13735, 2020

    2020

  36. [44]

    Mitigating memorization of noisy labels by clipping the model prediction

    Wei, H., Zhuang, H., Xie, R., Feng, L., Niu, G., An, B., and Li, Y. Mitigating memorization of noisy labels by clipping the model prediction. In International Conference on Machine Learning, pp.\ 36868--36886. PMLR, 2023 a

    2023

  37. [45]

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

    Wei, J., Zhu, Z., Cheng, H., Liu, T., Niu, G., and Liu, Y. Learning with noisy labels revisited: A study using real-world human annotations. In International Conference on Learning Representations, 2022

    2022

  38. [46]

    Sparse parameterization for epitomic dataset distillation

    Wei, X., Cao, A., Yang, F., and Ma, Z. Sparse parameterization for epitomic dataset distillation. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2023 b

    2023

  39. [47]

    Robust early-learning: Hindering the memorization of noisy labels

    Xia, X., Liu, T., Han, B., Gong, C., Wang, N., Ge, Z., and Chang, Y. Robust early-learning: Hindering the memorization of noisy labels. In International conference on learning representations, 2020

    2020

  40. [48]

    and He, Y

    Xiao, L. and He, Y. Are large-scale soft labels necessary for large-scale dataset distillation? In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  41. [49]

    An efficient dataset condensation plugin and its application to continual learning

    Yang, E., Shen, L., Wang, Z., Liu, T., and Guo, G. An efficient dataset condensation plugin and its application to continual learning. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  42. [50]

    Learning with twin noisy labels for visible-infrared person re-identification

    Yang, M., Huang, Z., Hu, P., Li, T., Lv, J., and Peng, X. Learning with twin noisy labels for visible-infrared person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 14308--14317, 2022

    2022

  43. [51]

    Neural spectral decomposition for dataset distillation

    Yang, S., Cheng, S., Hong, M., Fan, H., Wei, X., and Liu, S. Neural spectral decomposition for dataset distillation. In Proceedings of the European Conference on Computer Vision (ECCV), 2024

    2024

  44. [52]

    Yao, Q., Yang, H., Han, B., Niu, G., and Kwok, J. T.-Y. Searching to exploit memorization effect in learning with noisy labels. In International Conference on Machine Learning, pp.\ 10789--10798. PMLR, 2020

    2020

  45. [53]

    and Wu, J

    Yi, K. and Wu, J. Probabilistic end-to-end noise correction for learning with noisy labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 7017--7025, 2019

    2019

  46. [54]

    Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective

    Yin, Z., Xing, E., and Shen, Z. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  47. [55]

    and Bilen, H

    Zhao, B. and Bilen, H. Dataset condensation with differentiable siamese augmentation. In International Conference on Machine Learning, pp.\ 12674--12685. PMLR, 2021 a

    2021

  48. [56]

    and Bilen, H

    Zhao, B. and Bilen, H. Dataset condensation with gradient matching. In Proceedings of the International Conference on Learning Representations (ICLR), 2021 b

    2021

  49. [57]

    Centrality and consistency: two-stage clean samples identification for learning with instance-dependent noisy labels

    Zhao, G., Li, G., Qin, Y., Liu, F., and Yu, Y. Centrality and consistency: two-stage clean samples identification for learning with instance-dependent noisy labels. In European Conference on Computer Vision, pp.\ 21--37. Springer, 2022

    2022

  50. [58]

    Improved distribution matching for dataset condensation

    Zhao, G., Li, G., Qin, Y., and Yu, Y. Improved distribution matching for dataset condensation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 7856--7865, 2023

    2023

  51. [59]

    M., and Litany, O

    Zheltonozhskii, E., Baskin, C., Mendelson, A., Bronstein, A. M., and Litany, O. Contrast to divide: Self-supervised pre-training for learning with noisy labels. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp.\ 1657--1667, 2022

    2022

  52. [60]

    Robust curriculum learning: from clean label detection to noisy label self-correction

    Zhou, T., Wang, S., and Bilmes, J. Robust curriculum learning: from clean label detection to noisy label self-correction. In International Conference on Learning Representations, 2020

    2020

  53. [61]

    P., and Xing, L

    Zhou, Y., Li, X., Liu, F., Wei, Q., Chen, X., Yu, L., Xie, C., Lungren, M. P., and Xing, L. L2b: Learning to bootstrap robust models for combating label noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 23523--23533, 2024

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.