Multimodal-Guided Dynamic Dataset Pruning for Robust and Efficient Data-Centric Learning

Dongzhan Zhou; Furao Shen; Peijia Li; Peng Ye; Suorong Yang; Wanli Ouyang; Yujie Liu; Zhiming Xu

REVIEW 4 major objections 5 minor 21 references

Multimodal-Guided Dynamic Dataset Pruning for Robust and Efficient Data-Centric Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a dynamic, dual-supervision pruning score can remove redundant and noisy samples during training while beating full-dataset accuracy at reduced compute.

desk verdict Sensible incremental idea, but the central score optimization in Eq. (4) is ill-posed as written and the empirical margins over InfoBatch are within noise; needs major revision before the claims can be trusted. read the letter →

arxiv 2507.12750 v1 pith:6PSCJXXB submitted 2025-07-17 cs.LG cs.CV

Suorong Yang , Peijia Li , Yujie Liu , Zhiming Xu , Peng Ye , Wanli Ouyang , Furao Shen , Dongzhan Zhou This is my paper

classification cs.LGcs.CV

keywords datasetpruningdynamicdataselectionmultimodalsupervisionCLIPnoisylabelsdata-centriclearningtrainingefficiencysemanticconsistency

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 tries to establish that dataset pruning can be both dynamic and multimodal: instead of fixing a subset before training, the training loop selects a fresh subset each epoch using two signals together. The first signal is the model's own loss on each sample, which marks difficulty; the second is the cosine similarity between the image embedding and the text embedding of "A photo of [CLASS]", computed by the pretrained multimodal model CLIP, which marks semantic correctness. A per-sample score is optimized to favor samples that are hard but semantically consistent, and the actual mini-batches are drawn from samples whose scores sit near the median. The claim is that this dual supervision beats static selection baselines and even full-dataset training: on CIFAR-10/100, 70% of the data gives higher accuracy than the whole dataset, and on ImageNet-1k at a 60% selection ratio ResNet-50 reaches 76.8% versus 76.4% for the full dataset while using roughly 40% fewer forward-backward operations. If true, this matters because it offers a cheap, online recipe for turning redundant and noisy data into faster training without a performance penalty.

What carries the argument

The load-bearing object is a per-sample learnable score vector $s$, updated each epoch by a lightweight numerical optimization against the multi-objective loss $L_s = \frac{1}{\|s\|_0}\, s \cdot (\lambda s_C - s_T)$. The two supervisory signals are $s_T$, the current model's task loss on each sample, and $s_C$, the scaled cosine similarity between CLIP's image embedding and CLIP's text embedding of the prompt "A photo of [CLASS]", with dataset-specific linear adapters fine-tuned on the training set using the contrastive InfoNCE loss. The optimizer promotes samples with high semantic consistency and low task loss, and the selected subset is formed from samples whose optimized scores fall near the median, a choice the paper justifies as a robust proxy for distributional balance. Because CLIP features are precomputed once, the online overhead is small: on ImageNet-1k the extra wall-clock cost is listed as 0.003 hours against 10.5 hours of training.

What would settle it

Inject known label noise by swapping labels between visually similar classes (for example, cat and dog in CIFAR-10), run the method, and compute precision and recall of the samples it prunes against the known swap set; if the pruned set does not contain most of the swapped samples, or if replacing the CLIP score with random values leaves final accuracy unchanged, the claimed noise-filtering mechanism is not doing the work.

Watch

Extended reading notes

Core claim

The central discovery claimed is that combining task-driven difficulty with cross-modality semantic consistency in a single learnable score makes sample selection more robust than either signal alone. Samples with high task loss but low CLIP consistency are treated as likely noise and suppressed, while samples that are both hard and semantically aligned are kept. Optimizing the score vector with $L_s = \frac{1}{\|s\|_0}\, s \cdot (\lambda s_C - s_T)$ and then selecting samples near the median produces stable selection trajectories that smooth out transient loss spikes. In the noisy-label visualization with 20% injected label noise, the method is said to consistently retain clean samples and filter out mislabeled ones. The headline quantitative claim is that at a 60% selection ratio on ImageNet-1k, the method beats full-data training (76.8% versus 76.4%) while cutting wall-clock training time roughly in half (10.5 hours versus 17.5 hours).

Load-bearing premise

The whole method rests on the premise that CLIP's image-to-text similarity score, after a linear adapter is fine-tuned on the same training data, correctly tells which samples are semantically valid and which are mislabeled.

Editorial extensions

If this is right

On CIFAR-10/100, keeping only 70% of the samples per epoch yields higher test accuracy than training on the full dataset, implying roughly a 30% compute saving with no accuracy cost.
At a 60% selection ratio on ImageNet-1k, ResNet-50 reaches 76.8% accuracy versus 76.4% on the full dataset, while wall-clock training time drops from 17.5 to 10.5 hours.
The method outperforms strong static selectors such as EL2N, GraNd, Herding, and Moderate-DS, which pay large offline costs for surrogate training or Gram-matrix inversion.
In a 20% label-noise setting, the selection score is claimed to consistently filter out noisy or mislabeled images while keeping semantically meaningful samples.
Because the selection signal is computed online from precomputed CLIP features, the approach avoids the heavy one-shot scoring overhead of static pruning and scales to ImageNet-1k.

Reading between the lines

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

Editorial inference: the noise-filtering behavior inherits CLIP's biases, so a clean but atypical image whose embedding does not match the text prompt may be pruned as if it were noise, potentially sacrificing tail-class recall for average accuracy.
Editorial inference: selecting samples near the median score is an implicit anti-curriculum that excludes both the hardest and the easiest examples, and it should be directly testable against explicit curriculum schedules.
Editorial inference: the same dual-supervision objective should transfer to any domain with paired embeddings, such as audio-text or code-text learning, by replacing CLIP with a matching pretrained encoder; the paper's own limitations section points in this direction.

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.

Desk Editor's Note

read the letter

Short version: the idea is plausible and the writing is clear, but the central optimization is not a well-defined problem, and the experiments as reported cannot distinguish the method from InfoBatch plus run-to-run noise.

What's new: combining per-epoch dynamic selection with a CLIP cross-modal consistency score (with lightweight adapters) is not in the cited literature, although it is a direct extension of the authors' own static CLIP-based selector and InfoBatch's median-based dynamic selection. The adapters are a sensible way to address domain shift, and the reported overhead is tiny (0.003 GPU-hours on ImageNet). The empirical trend is consistently positive: small gains over InfoBatch on CIFAR-10/100 and ImageNet-1k.

The soft spots are serious. First, Eq. (4) defines Ls = (1/||s||0) s·(λsC − sT). As written, the gradient with respect to s does not depend on s (except through the constant norm), so unconstrained minimization has no finite minimizer: each coordinate blows up according to the sign of λsC_i − sT_i. No constraint, regularizer, normalization, or update rule is given, and no value of λ is reported. The selection rule "near the median of s" is therefore not executable. This is a load-bearing gap: even with perfect semantic and task signals, the combination mechanism is undefined. Second, all experiments lack error bars, seed counts, and significance tests; the gains over InfoBatch are 0.1–0.7 points, which is within plausible noise for ResNet-18/50. Third, the noisy-label evidence is only a figure—no quantitative filtering rate or accuracy under noise. Fourth, key hyperparameters (λ, adapter training details, median window) are missing.

The CLIP-bias concern you raised is real but secondary; the optimization gap matters before that. If the "optimization" is actually just a fixed composite score with median selection, the paper's central claim is much weaker than advertised.

Who should read it: people in data pruning who want to see what happens when you add a frozen multimodal model to dynamic selection. As presented, a competent group could not reproduce or falsify it. I'd still send it to a serious referee because the idea is worth engaging and the flaw may be a presentation issue—but the revision must specify the actual score update, add variance-aware experiments, and compare against their own static CLIP selector.

Referee Report

4 major / 5 minor

Summary. The paper introduces a dynamic dataset pruning framework that scores each training sample by combining the current task loss with a CLIP-based semantic consistency score between the image and the class-label text. A per-sample score vector s is said to be optimized through Eq. (4), and samples with scores near the median are retained each epoch. The experiments compare the method with static and dynamic pruning baselines on CIFAR-10/100 and ImageNet-1k, reporting accuracy close to or above full-data training at reduced compute, and Figure 1 is offered as evidence that noisy samples are filtered.

Significance. If the method worked as described, combining a pretrained multimodal prior with an online difficulty signal would be a useful contribution to data-centric learning, and the ImageNet result at 60% selection (76.8% vs 76.4% for the full dataset) is promising. The paper also makes a good practical point that dynamic selection avoids the heavy offline overhead of static coreset methods. However, the current manuscript does not make the central mechanism well-defined, provides no code or repeated trials, and supports the noise-filtering claim only by an illustrative figure. As a result, the significance is presently a promising direction rather than an established result.

major comments (4)

[3.5, Eq. (4)] The optimization defining s is not well-posed. Ls is linear in s: the partial derivative with respect to s_i is (lambda*sC_i - sT_i)/||s||0, which is independent of s. Unconstrained minimization therefore drives each coordinate to +∞ or -∞ depending on the sign of the coefficient, so no finite minimizer exists. The initialization s=1 and the 'near the median' selection rule cannot be executed as stated. The paper does not specify a constraint (e.g., fixed norm), a regularizer, an optimizer, a schedule, or a value of lambda. Please either define a constrained/regularized objective and report its solver, or remove the claim that s is optimized and specify the selection rule directly as a function of sT and sC.
[4, Tables 1 and 2] All accuracy numbers appear to be single runs with no error bars, seeds, or significance tests. Several margins over InfoBatch are within typical run-to-run variation (e.g., 94.9 vs 94.7 at 30% CIFAR-10; 95.1 vs 95.1 at 50%; 76.8 vs 76.5 on ImageNet). Without repeated trials, the claim that the method 'consistently outperforms' baselines is not established. Please report means with standard deviations over at least 3-5 seeds and a significance test where appropriate.
[4.3, Figure 1] The claim that the method 'consistently filters out noisy or mislabeled data' is supported only by an illustrative figure, without quantitative filtering rates, precision/recall, or comparison to baselines. Since the semantic score is computed against the label text, a mislabeled image is expected to have low similarity by construction; the experiment should measure how many noisy samples are actually removed and compare to random or loss-only selection.
[3.4] The cross-modal semantic signal is obtained by fine-tuning dataset-specific adapters on the same target training data with InfoNCE loss. Because the adapters are fit to the very data being pruned, the semantic consistency score is not an independent source of supervision; the noise-filtering result in Figure 1 may partly reflect the adapter's ability to fit the training set rather than a pretrained multimodal prior. Please report the filtering behavior with frozen zero-shot CLIP versus adapter-tuned CLIP, and quantify how much of the filtering effect survives without adapter fine-tuning.

minor comments (5)

[2] There are typos and duplicated words: 'categoried' should be 'categorized', and 'the the' appears twice in the score-based static selection paragraph.
[3.5] The symbol ||s||0 is used without definition; if it denotes the number of nonzero entries, the behavior of the objective when some entries are zero should be clarified, since division by zero can occur.
[3.5] The claim that parameter complexity is 'less than 5.5%' on ImageNet with ResNet-50 should state the storage assumption explicitly (e.g., 1.2M float32 scores, or the number of trainable parameters in the adapters); as written the comparison is not verifiable.
[4.3, Figure 1] The figure caption should identify which panels correspond to selected clean and selected noisy data and state the exact selection ratio and noise ratio; the current caption only says 20%.
[4.2, Table 2] The overhead of Ours (0.0030 h) is slightly higher than InfoBatch (0.0028 h); the text should not imply that our overhead is the smallest, and the wall-clock times should be clarified as being identical because both methods use the same training schedule.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the selection mechanism is an explicit heuristic and the reported gains are external benchmarks, not the method's own definitions.

full rationale

The paper does not derive its central quantity from the quantity it predicts. The score s in Eq. (4) is defined directly from the task loss sT and the CLIP consistency sC; sC is computed by Eq. (3) as cosine similarity to the prompt text, and no fitted parameter is later relabeled as a prediction. The noise-filtering claim in Sec. 4.3 is evaluated against injected label noise and is an empirical consequence of choosing CLIP similarity as a signal, not a definitional equivalence: a randomly relabeled image need not have low similarity to the wrong label text, so the connection is an assumption about CLIP rather than a tautology. The adapters are fine-tuned on the target data, which weakens the independence of the semantic signal but does not make the accuracy numbers in Tables 1 and 2 follow by construction. The paper cites several prior works by the same authors, but only as related work; no uniqueness theorem, ansatz, or fitting procedure is imported from those citations. The optimization in Eq. (4) is not well-posed as written—the unconstrained linear objective has no finite minimizer and no update rule is given—but that is a correctness and reproducibility defect, not a circular reduction of the claimed result to its inputs. Since no load-bearing step is equivalent to its own input, no circular step is identified.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The method depends on several unstated or hand-picked quantities: the weighting lambda, the selection ratio, adapter training details, and the median-bandwidth. The semantic signal from CLIP is treated as reliable, but the adapters are fine-tuned on the same target data, which weakens its independence. No external or formal evidence is provided for these premises.

free parameters (4)

lambda (weighting coefficient in Ls)
Controls the balance between CLIP consistency sC and task loss sT; its value is never reported in the paper.
Selection ratio k = 30%, 50%, 70% (CIFAR); 60% (ImageNet)
Chosen per experiment and not derived from any principle; directly controls the size of the pruned subset.
CLIP adapter linear-layer parameters
Image and text adapters are fine-tuned with InfoNCE on the target data; no training details (epochs, learning rate, data split) are given.
Near-median selection threshold
The method selects samples with scores near the median of s, but the threshold or bandwidth used to meet the target selection ratio is not specified.

assumptions (4)

domain assumption CLIP image-text similarity is a reliable semantic consistency signal for CIFAR and ImageNet images.
Invoked in Sec. 3.4 without validation on the target domains or discussion of CLIP's failure modes on fine-grained classes.
ad hoc to paper The score vector s can be optimized with the stated loss to produce a stable and meaningful ranking.
Sec. 3.5 presents Ls but provides no optimizer, step size, convergence argument, or demonstration that the optimized scores improve selection.
ad hoc to paper Selecting samples near the median of the combined score yields a distributionally balanced and informative subset.
The paper cites James et al. (2013) for the median as a robust proxy, but does not prove or empirically show that this holds for the specific score distribution.
domain assumption Fine-tuning the CLIP adapters on the target dataset does not introduce label leakage or bias into the semantic consistency scores.
The adapters are trained on the same data being pruned; the paper does not analyze how this affects the independence of the sC signal.

invented entities (1)

Per-sample learnable selection score vector s
purpose: Ranks samples for retention or pruning in each training epoch.
This is an internal optimization variable with no falsifiable handle outside the method; its behavior is only demonstrated through the final accuracy numbers.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multimodal-Guided Dynamic Dataset Pruning for Robust and Efficient Data-Centric Learning." pith.science (2026). https://pith.science/paper/6PSCJXXB

@misc{pith2026250712750,
  author       = {Pith},
  title        = {Pith review of: Multimodal-Guided Dynamic Dataset Pruning for Robust and Efficient Data-Centric Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6PSCJXXB}},
  note         = {Machine review of arXiv:2507.12750}
}

read the original abstract

Modern deep models are trained on large real-world datasets, where data quality varies and redundancy is common. Data-centric approaches such as dataset pruning have shown promise in improving training efficiency and model performance. However, most existing methods rely on static heuristics or task-specific metrics, limiting their robustness and generalizability across domains. In this work, we introduce a dynamic dataset pruning framework that adaptively selects training samples based on both task-driven difficulty and cross-modality semantic consistency. By incorporating supervision from pretrained multimodal foundation models, our approach captures training dynamics while effectively filtering out uninformative samples. Our work highlights the potential of integrating cross-modality alignment for robust sample selection, advancing data-centric learning toward more efficient and robust practices across application domains.

Figures

Figures reproduced from arXiv: 2507.12750 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 5 canonical work pages

[4]

DONOD: Efficient and Generalizable Instruction Fine-Tuning for LLMs via Model-Intrinsic Dataset Pruning

Hu, J., Yang, S., Zhou, D., and Wu, L. Donod: Ro- bust and generalizable instruction fine-tuning for llms via model-intrinsic dataset pruning. arXiv preprint arXiv:2504.14810,

work page Pith review arXiv
[5]

and Tao, D

Lei, S. and Tao, D. A comprehensive survey to dataset distillation. arXiv preprint arXiv:2301.05603,

arXiv
[7]

Confident learning: Estimating uncertainty in dataset labels

Northcutt, C., Jiang, L., and Chuang, I. Confident learning: Estimating uncertainty in dataset labels. volume 70, pp. 1373–1411, 2021a. Northcutt, C. G., Athalye, A., and Mueller, J. Pervasive la- bel errors in test sets destabilize machine learning bench- marks. arXiv preprint arXiv:2103.14749, 2021b. Oord, A. v. d., Li, Y ., and Vinyals, O. Representati...

arXiv
[9]

Infobatch: Lossless training speed up by unbiased dynamic data pruning

Qin, Z., Wang, K., Zheng, Z., Gu, J., Peng, X., Xu, Z., Zhou, D., Shang, L., Sun, B., Xie, X., et al. Infobatch: Lossless training speed up by unbiased dynamic data pruning. arXiv preprint arXiv:2303.04947,

arXiv
[11]

A weighted k- center algorithm for data subset selection

Ramalingam, S., Awasthi, P., and Kumar, S. A weighted k- center algorithm for data subset selection. arXiv preprint arXiv:2312.10602,

arXiv
[12]

Shen, H., Zhao, Z.-Q., and Zhang, W

URL https://arxiv.org/abs/ 2204.06125. Shen, H., Zhao, Z.-Q., and Zhang, W. Adaptive dynamic fil- tering network for image denoising. InProceedings of the AAAI Conference on Artificial Intelligence, volume 37, pp. 2227–2235,

arXiv
[13]

Toneva, M., Sordoni, A., Combes, R. T. d., Trischler, A., Bengio, Y ., and Gordon, G. J. An empirical study of example forgetting during deep neural network learning. arXiv preprint arXiv:1812.05159,

arXiv
[14]

Dataset distillation with neural character- istic function: A minmax perspective

Wang, S., Yang, Y ., Liu, Z., Sun, C., Hu, X., He, C., and Zhang, L. Dataset distillation with neural character- istic function: A minmax perspective. arXiv preprint arXiv:2502.20653, 2025a. Wang, S., Yang, Y ., Liu, Z., Sun, C., Hu, X., He, C., and Zhang, L. Dataset distillation with neural characteristic function: A minmax perspective. In Proceedings of...

arXiv

Show all 21 references

[15]

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. arXiv preprint arXiv:2110.12088,

arXiv
[16]

Submodularity in data subset selection and active learning

Wei, K., Iyer, R., and Bilmes, J. Submodularity in data subset selection and active learning. In International conference on machine learning, pp. 1954–1963. PMLR,

1954
[18]

Combating noisy labels with sample selection by mining high-discrepancy examples

Xia, X., Han, B., Zhan, Y ., Yu, J., Gong, M., Gong, C., and Liu, T. Combating noisy labels with sample selection by mining high-discrepancy examples. In Proceedings of the IEEE/CVF international conference on computer vision, pp. 1833–1843, 2023a. Xia, X., Liu, J., Yu, J., Sh...

arXiv
[20]

Zhang, X., Du, J., Li, Y ., Xie, W., and Zhou, J

URL https://arxiv.org/abs/2111.03930. Zhang, X., Du, J., Li, Y ., Xie, W., and Zhou, J. T. Spanning training progress: Temporal dual-depth scoring (tdds) for enhanced dataset pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 262...

arXiv
[21]

C., and Liu, Z

Zhou, K., Yang, J., Loy, C. C., and Liu, Z. Learn- ing to prompt for vision-language models. Interna- tional Journal of Computer Vision , 130(9):2337–2348, July 2022a. ISSN 1573-1405. doi: 10.1007/ s11263-022-01653-1. URL http://dx.doi.org/ 10.1007/s11263-022-01653-1 . Zhou, Y...

arXiv
[2009]

Wu, S., Zhang, W., Xu, L., Jin, S., Li, X., Liu, W., and Loy, C. C. Clipself: Vision transformer distills itself for open-vocabulary dense prediction. arXiv preprint arXiv:2310.01403,

arXiv
[2018]

Infonce loss provably learns cluster- preserving representations

Parulekar, A., Collins, L., Shanmugam, K., Mokhtari, A., and Shakkottai, S. Infonce loss provably learns cluster- preserving representations. In The Thirty Sixth Annual Conference on Learning Theory, pp. 1914–1961. PMLR,

1914
[2020]

Clip-adapter: Better vision- 5 language models with feature adapters

Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y ., Li, H., and Qiao, Y . Clip-adapter: Better vision- 5 language models with feature adapters. arXiv preprint arXiv:2110.04544,

arXiv
[2021]

S., Daruwalla, K., and Lipasti, M

6 Raju, R. S., Daruwalla, K., and Lipasti, M. Accelerating deep learning with dynamic data pruning. arXiv preprint arXiv:2111.12621,

arXiv
[2022]

Imagenet: A large-scale hierarchical image database

Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255. Ieee,

2009
[2023]

D2 pruning: Mes- sage passing for balancing diversity and difficulty in data pruning

Maharana, A., Yadav, P., and Bansal, M. D2 pruning: Mes- sage passing for balancing diversity and difficulty in data pruning. arXiv preprint arXiv:2310.07931,

arXiv
[2024]

Guo, Z., Wang, K., Cazenavette, G., Li, H., Zhang, K., and You, Y

URL https://arxiv.org/ abs/2406.06973. Guo, Z., Wang, K., Cazenavette, G., Li, H., Zhang, K., and You, Y . Towards lossless dataset distillation via difficulty-aligned trajectory matching. arXiv preprint arXiv:2310.05773,

arXiv
[2025]

7 Zhang, R., Fang, R., Zhang, W., Gao, P., Li, K., Dai, J., Qiao, Y ., and Li, H

URL https: //arxiv.org/abs/2505.03809. 7 Zhang, R., Fang, R., Zhang, W., Gao, P., Li, K., Dai, J., Qiao, Y ., and Li, H. Tip-adapter: Training-free clip- adapter for better vision-language modeling,

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [4]

DONOD: Efficient and Generalizable Instruction Fine-Tuning for LLMs via Model-Intrinsic Dataset Pruning

Hu, J., Yang, S., Zhou, D., and Wu, L. Donod: Ro- bust and generalizable instruction fine-tuning for llms via model-intrinsic dataset pruning. arXiv preprint arXiv:2504.14810,

work page Pith review arXiv

[2] [5]

and Tao, D

Lei, S. and Tao, D. A comprehensive survey to dataset distillation. arXiv preprint arXiv:2301.05603,

arXiv

[3] [7]

Confident learning: Estimating uncertainty in dataset labels

Northcutt, C., Jiang, L., and Chuang, I. Confident learning: Estimating uncertainty in dataset labels. volume 70, pp. 1373–1411, 2021a. Northcutt, C. G., Athalye, A., and Mueller, J. Pervasive la- bel errors in test sets destabilize machine learning bench- marks. arXiv preprint arXiv:2103.14749, 2021b. Oord, A. v. d., Li, Y ., and Vinyals, O. Representati...

arXiv

[4] [9]

Infobatch: Lossless training speed up by unbiased dynamic data pruning

Qin, Z., Wang, K., Zheng, Z., Gu, J., Peng, X., Xu, Z., Zhou, D., Shang, L., Sun, B., Xie, X., et al. Infobatch: Lossless training speed up by unbiased dynamic data pruning. arXiv preprint arXiv:2303.04947,

arXiv

[5] [11]

A weighted k- center algorithm for data subset selection

Ramalingam, S., Awasthi, P., and Kumar, S. A weighted k- center algorithm for data subset selection. arXiv preprint arXiv:2312.10602,

arXiv

[6] [12]

Shen, H., Zhao, Z.-Q., and Zhang, W

URL https://arxiv.org/abs/ 2204.06125. Shen, H., Zhao, Z.-Q., and Zhang, W. Adaptive dynamic fil- tering network for image denoising. InProceedings of the AAAI Conference on Artificial Intelligence, volume 37, pp. 2227–2235,

arXiv

[7] [13]

Toneva, M., Sordoni, A., Combes, R. T. d., Trischler, A., Bengio, Y ., and Gordon, G. J. An empirical study of example forgetting during deep neural network learning. arXiv preprint arXiv:1812.05159,

arXiv

[8] [14]

Dataset distillation with neural character- istic function: A minmax perspective

Wang, S., Yang, Y ., Liu, Z., Sun, C., Hu, X., He, C., and Zhang, L. Dataset distillation with neural character- istic function: A minmax perspective. arXiv preprint arXiv:2502.20653, 2025a. Wang, S., Yang, Y ., Liu, Z., Sun, C., Hu, X., He, C., and Zhang, L. Dataset distillation with neural characteristic function: A minmax perspective. In Proceedings of...

arXiv