Pith. sign in

REVIEW 5 major objections 6 minor 40 references

Adaptive Dataset Quantization

T0 review · 5 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Scoring data bins by texture and diversity beats uniform sampling

desk verdict The adaptive sampling idea is plausible, but Eq. 9–10 don't enforce a fixed keep ratio, so the reported gains over DQ are uncontrolled until the sampling step is clarified. read the letter →

arxiv 2412.16895 v1 pith:2OM7QB5Z submitted 2024-12-22 cs.CV

Muquan Li , Dongyang Zhang , Qiang Dong , Xiurui Xie , Ke Qin This is my paper · ORCID
classification cs.CV
keywords datasetquantizationadaptivesamplingcompressioncoresetselectiontexturelevelcontrastivelearningimportancescorecross-architecturegeneralization
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 Quantization (DQ) compresses a training set by carving it into non-overlapping bins and then drawing an equal fraction from every bin. This paper argues that uniform bin sampling is the weak point: bins generated early are better at representing the whole dataset, later bins add more diversity, and their actual importance varies unevenly. Adaptive Dataset Quantization (ADQ) scores each bin for representativeness (by the texture level of its patches) and diversity (by a contrastive discriminator), sums the normalized scores into an importance score, and samples from each bin in proportion to that score and the bin's size. The paper reports consistent gains over DQ across CNNs and transformers, averaging roughly 3% on CIFAR-10 at 10–30% keep ratios, and lossless compression at 60% data on CIFAR-10 and ImageNet-1K. The practical point is that a cheap per-bin scoring step can replace uniform sampling without touching the rest of the DQ pipeline.

What carries the argument

The load-bearing object is the per-bin importance score $\hat{I}_n = \mathrm{Norm}(\mathrm{Rep}_n) + \mathrm{Norm}(\mathrm{Div}_n)$, where $\mathrm{Rep}_n$ is the mean gradient magnitude of a bin's $L\times L$ patches, $T(P) = \frac{1}{L^2}\sum_{i,j} G(P_{i,j})$, and $\mathrm{Div}_n$ is the negative contrastive ratio from a discriminator. The score feeds the adaptive sampling ratio $r_n = \alpha \hat{I}_n + (1-\alpha) N(n)/\sum_n N(n)$, replacing uniform sampling. The paper's argument stands on this scalar correctly ranking bins by their contribution to training; the texture-level identity $\mathrm{Rep} = T$ is justified by a single trajectory-distance experiment and is never derived from first principles.

What would settle it

Train a model on a compressed set made with ADQ's importance scores, and on a second set where the same scores are randomly permuted across bins before sampling; if accuracy does not drop, the scores are not the active ingredient. A direct check is to compute, on ImageNet-1K with the ViT feature extractor, the rank correlation between each bin's texture level and the actual parameter-space distance between a student trained on that bin and an expert trained on the full dataset; a non-monotonic or near-zero correlation would falsify the $\mathrm{Rep}=T$ mapping.

Watch

Extended reading notes

Core claim

The central claim is that the bins produced by DQ's recursive submodular selection differ systematically in how much they contribute to training, so the uniform sampling used by naive DQ is a spurious balance rather than an optimal one. ADQ makes this difference measurable by assigning each bin a representativeness score $\mathrm{Rep}_n$ equal to its texture level $T(P)$, a diversity score $\mathrm{Div}_n$ from a contrastive instance-discrimination discriminator, and an importance score $\hat{I}_n = \mathrm{Norm}(\mathrm{Rep}_n) + \mathrm{Norm}(\mathrm{Div}_n)$. Sampling then uses the proportion $r_n = \alpha \hat{I}_n + (1-\alpha)\, N(n)/\sum_{n} N(n)$ for each bin. The authors show that this adaptive sampling outperforms uniform sampling on every architecture they test, with the largest relative gains at higher keep ratios, and that the extra computation is near negligible.

Load-bearing premise

The load-bearing premise is that the texture level of a bin's patches measures how representative that bin is of the full dataset; this is supported by one trajectory-distance curve on ResNet-18/CIFAR-10, with no verification on the other architectures, datasets, or bin orders used in the main experiments.

Editorial extensions

If this is right

  • Any DQ-based pipeline can adopt ADQ as a drop-in replacement for uniform sampling, gaining 2–4 percentage points on CIFAR-10 without changing bin generation or downstream training.
  • At a 60% data keep ratio, the compressed set matches full-data accuracy, so storage and bandwidth can be cut by 40% with no accuracy loss.
  • The improvement transfers across CNNs and transformers, so the method is not tied to a single architecture family.
  • The full generation cost stays close to DQ's, making the accuracy gain available at roughly the same compute budget.

Reading between the lines

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

  • Not tested in the paper: the same adaptive sampling could be layered on top of any bin-generation scheme, not only DQ's GraphCut bins, since the scores are computed after bins exist.
  • The equal-weight sum of normalized RS and DS is a design choice; a multiplicative combination or a learned weight might do better, especially at high keep ratios where the paper observes the optimal $\alpha$ shifting.
  • The texture-level representativeness proxy is a natural-image heuristic; extending ADQ to non-image data would require a different cheap representativeness measure, which the paper does not propose.
  • A harder stress test, not in the paper, would be full-scale ImageNet-21k, where the single trajectory-distance justification for texture level is under less direct control.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes Adaptive Dataset Quantization (ADQ), an extension of Dataset Quantization (DQ) that replaces uniform sampling across bins with an adaptive sampling strategy. Each bin is scored by a representativeness score based on texture level, a diversity score based on contrastive learning, and an importance score defined as the sum of the normalized RS and DS. Adaptive sampling then draws more images from bins with higher importance. The authors report consistent accuracy gains over DQ on CIFAR-10 across five architectures (2.6%, 2.8%, and 3.3% average improvements at 10%, 20%, and 30% keep ratios), similar improvements on ImageNet-1K, and lossless compression at 60% data, with negligible additional generation time.

Significance. If the reported results hold, ADQ would be a practical, low-cost improvement over DQ that preserves cross-architecture generalization. The paper ships code, reports five-run averages with error bars, and identifies a genuine limitation of DQ's uniform sampling. However, the central quantitative claim is not yet established because the method as written does not control the size of the compressed set, and several components of the scoring mechanism are underspecified. The strengths are real but contingent on fixing the sampling-size issue.

major comments (5)
  1. [Adaptive Sampling, Eqs. (9)–(10); Table 1] The method does not enforce the target keep ratio ρ. With m equal bins of size K=M/m, Eq. (9) gives r_n = α Î_n + (1−α)/m and Eq. (10) gives q_n = floor(r_n K), so the total selected size is approximately (M/m)[α Σ_n Î_n + (1−α)]. If Norm in Eqs. (6)–(7) is z-score normalization, ΣÎ_n = 0 and the total is about (1−α)M/m, which for m=10 and α=0.6 is 4% of the dataset rather than the nominal 10%. If Norm is min-max normalization, the total varies with the data and can range far above or below ρM. The floor function only bounds each bin's contribution and cannot make Σ q_n equal the intended ρM. Therefore the reported gains over DQ may be an artifact of training on a different number of images rather than of adaptive sampling. The authors must specify the normalization, add an explicit rescaling or completion step to match the target size, and rerun the comparisons at matched dataset sizes.
  2. [Representativeness Score, Eq. (3) and Fig. 4] The identification Rep(·)=T(·) is supported only by a single experiment comparing three global texture-level batches on ResNet-18/CIFAR-10. This does not establish that mean gradient magnitude ranks individual DQ bins by their trajectory closeness to the expert model, nor that the relationship transfers to other architectures, datasets, or bin orders. Since this ordering directly determines sampling proportions, the mechanism remains a heuristic. A per-dataset validation at the bin level, or a derivation, is needed before the representativeness score can be considered a principled component of the method.
  3. [Diversity Score, Eq. (5)] Eq. (5) is mathematically incomplete. The term N(x−) is not defined as a function of x_i even though it appears inside an expectation over x_i, and the leading negative sign is unexplained. As written, Div(S_n) is not a standard contrastive objective, and it is unclear whether larger or smaller values correspond to greater diversity. This ambiguity propagates into the importance score in Eq. (8) and into the adaptive sampling in Eq. (9).
  4. [Ablation Study, Fig. 6 and Comparisons with Previous Methods] The paper states that 'During the actual experiment, we adjust corresponding values of α according to different architectures,' but it does not report the chosen α values or the selection criterion. If α is selected based on validation accuracy for each architecture and each keep ratio, while the DQ baseline has no analogous tuned parameter, the comparison in Table 1 and Fig. 5 is not apples-to-apples. The authors should report the α values used for every architecture and dataset, and ideally show results with a fixed α in addition to tuned α.
  5. [Comparisons with Previous Methods] The paper claims that 'our ADQ also achieves lossless results with only 60% of the data' on CIFAR-10 and ImageNet-1K, but no 60% results are tabulated and no criterion for 'lossless' is given. Please provide the numerical results with error bars at ρ=60% and specify the statistical threshold used to declare lossless compression.
minor comments (6)
  1. [Importance Evaluation, Eqs. (6)–(7)] The normalization operator Norm is never specified (z-score, min-max, or other). The citation to Ioffe and Szegedy (2015) is for batch normalization, which is not a generic normalization method; please define the operation explicitly.
  2. [Related Works, Remark] There is a typo: 'excessively comrplex' should be 'excessively complex'.
  3. [Experimental Setup, Datasets] The ImageNet-1K training size is given as '128,1126 samples', which appears to be a typo; the standard value is 1,281,167.
  4. [Figure 2] The subplot axes are unlabeled and the legend text is partially garbled, which makes the claimed trends in RS, DS, and IS difficult to read. Please provide clear axis labels and legends.
  5. [Algorithm 1] The input list 'Required:' is unconventional; the hyperparameters L, τ, and α should be listed as algorithm parameters rather than required inputs, and the output should note that the patch-dropping stage is applied after the initial compressed dataset is formed.
  6. [Adaptive Sampling, Eq. (9)] The equal weighting of RS and DS in Eq. (8) is asserted rather than derived. The ablation study reports that DS contributes slightly more than RS, so a short discussion of why equal weights are preferred across datasets would improve the presentation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ADQ's importance score is a hand-defined heuristic, the accuracy claims are empirical comparisons against external baselines, and there are no load-bearing self-citations.

full rationale

ADQ is an empirical method layered on DQ. The representativeness score is stipulated as texture level (Rep(·)=T(·), Eq. 3) and justified by the trajectory-distance experiment in Fig. 4; the diversity score is borrowed from contrastive learning; and the importance score is defined by Eq. 8 as the normalized sum of the two. Using that same IS in the adaptive sampling formula (Eq. 9) is method construction, not a derivation that makes a predicted quantity identical to an input. The paper does not invoke a self-citation chain: none of the references are authored by the present authors, and no uniqueness claim is imported from the authors' prior work. The most serious technical concern is that Eqs. 9-10 do not, as written, enforce a fixed total keep ratio (floor bounds each bin but cannot make Σq_n equal the target), which threatens the validity of the fixed-ρ comparisons; that is an experimental-control/correctness issue, not a circularity. Because the central accuracy claims are external benchmark comparisons against DM and DQ, not consequences of the definitions, the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several inherited or asserted premises. No new physical or conceptual entities beyond scoring metrics are introduced. Most parameters come from DQ or are left unspecified in the main text; alpha is the one parameter actively tuned to validation accuracy. The representativeness and importance-score mappings are the largest unproved inputs.

free parameters (6)
  • alpha = not reported per architecture or dataset
    Eq. 9; tuned on validation accuracy in the ablation study; values used for Table 1 and Fig. 5 omitted.
  • patch size L = not specified
    Algorithm 1 requires L x L patches; no value is given.
  • temperature tau = not specified
    Eq. 5 discriminator temperature; no value reported.
  • gradient operator G = not specified
    Eq. 3 uses G(.) with no concrete operator, so the texture level is not fully defined.
  • number of bins m = 10 (from DQ)
    Inherited from DQ's optimal setting; not re-optimized for ADQ.
  • drop ratio theta = 25% (from DQ)
    Inherited from DQ for patch dropping; appendix with details missing.
assumptions (5)
  • domain assumption DQ bin-generation theory: earlier bins are more representative, later bins are more diverse.
    Adopted from Zhou et al. 2023 as the starting point; not re-proven. It is the basis for choosing importance scores.
  • ad hoc to paper Texture level monotonically tracks trajectory closeness to the expert model.
    Only evidence is Fig. 4 from ResNet-18 on CIFAR-10; used to set Rep(.) = T(.).
  • domain assumption A contrastive instance-discrimination score measures bin diversity.
    Borrowed from contrastive learning; no formal mapping from instance discrimination to classification-relevant diversity.
  • ad hoc to paper Normalized RS and DS add with equal weight into a valid importance score.
    Eq. 8 fixes equal weights to avoid tuning, without theoretical or per-dataset justification.
  • domain assumption Validation accuracy is the right criterion for selecting alpha and reporting gains.
    Standard practice, but since alpha is per-architecture, part of the comparison advantage is tuned rather than fixed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Dataset Quantization." pith.science (2026). https://pith.science/paper/2OM7QB5Z

@misc{pith2026241216895,
  author       = {Pith},
  title        = {Pith review of: Adaptive Dataset Quantization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2OM7QB5Z}},
  note         = {Machine review of arXiv:2412.16895}
}
read the original abstract

Contemporary deep learning, characterized by the training of cumbersome neural networks on massive datasets, confronts substantial computational hurdles. To alleviate heavy data storage burdens on limited hardware resources, numerous dataset compression methods such as dataset distillation (DD) and coreset selection have emerged to obtain a compact but informative dataset through synthesis or selection for efficient training. However, DD involves an expensive optimization procedure and exhibits limited generalization across unseen architectures, while coreset selection is limited by its low data keep ratio and reliance on heuristics, hindering its practicality and feasibility. To address these limitations, we introduce a newly versatile framework for dataset compression, namely Adaptive Dataset Quantization (ADQ). Specifically, we first identify the sub-optimal performance of naive Dataset Quantization (DQ), which relies on uniform sampling and overlooks the varying importance of each generated bin. Subsequently, we propose a novel adaptive sampling strategy through the evaluation of generated bins' representativeness score, diversity score and importance score, where the former two scores are quantified by the texture level and contrastive learning-based techniques, respectively. Extensive experiments demonstrate that our method not only exhibits superior generalization capability across different architectures, but also attains state-of-the-art results, surpassing DQ by average 3\% on various datasets.

Figures

Figures reproduced from arXiv: 2412.16895 by the authors.

Figure 1
Figure 1. The paradigm of three types of dataset condensa [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The evaluation of normalized representativeness score, diversity score and importance score on CIFAR-10 [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The overall framework of the proposed Adaptive Dataset Quantization (ADQ). Following Dataset Quantization (DQ), [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The illustration of three types of texture level (TL) [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: The performance of DM (Zhao and Bilen 2023), DQ (Zhou et al. 2023) and ADQ on [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Ablation study on the weighting coefficient [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 37 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    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 word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Agarwal, S.; Arora, H.; Anand, S.; and Arora, C. 2020. Contextual Diversity for Active Learning. In Computer Vision - ECCV 2020 - 16th European Conference ECCV , volume 12361, 137--153. Springer

    work page 2020

  4. [4]

    A.; and Zhu, J

    Cazenavette, G.; Wang, T.; Torralba, A.; Efros, A. A.; and Zhu, J. 2022. Dataset Distillation by Matching Training Trajectories. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR , 10708--10717. IEEE

    work page 2022

  5. [5]

    A.; and Zhu, J

    Cazenavette, G.; Wang, T.; Torralba, A.; Efros, A. A.; and Zhu, J. 2023. Generalizing Dataset Distillation via Deep Generative Prior. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR , 3739--3748. IEEE

    work page 2023

  6. [6]

    Ceccarello, M.; Pietracaprina, A.; and Pucci, G. 2018. Fast Coreset-based Diversity Maximization under Matroid Constraints. In Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, WSDM , 81--89. ACM

    work page 2018

  7. [7]

    Coleman, C.; Yeh, C.; Mussmann, S.; Mirzasoleiman, B.; Bailis, P.; Liang, P.; Leskovec, J.; and Zaharia, M. 2020. Selection via Proxy: Efficient Data Selection for Deep Learning. In 8th International Conference on Learning Representations, ICLR . OpenReview.net

    work page 2020

  8. [8]

    Cui, J.; Wang, R.; Si, S.; and Hsieh, C. 2023. Scaling Up Dataset Distillation to ImageNet-1K with Constant Memory. In International Conference on Machine Learning, ICML , volume 202, 6565--6590. PMLR

    work page 2023

Show all 40 references
  1. [9]

    Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; Uszkoreit, J.; and Houlsby, N. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In 9th International Con...

    2021

  2. [10]

    Du, J.; Jiang, Y.; Tan, V. Y. F.; Zhou, J. T.; and Li, H. 2023. Minimizing the Accumulated Trajectory Error to Improve Dataset Distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR , 3749--3758. IEEE

    2023

  3. [11]

    Fang, G.; Song, J.; Wang, X.; Shen, C.; Wang, X.; and Song, M. 2021. Contrastive Model Invertion for Data-Free Knolwedge Distillation. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI , 2374--2380

    2021

  4. [12]

    Feldman, V.; and Zhang, C. 2020. What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation. In A Annual Conference on Neural Information Processing Systems, NeurIPS

    2020

  5. [13]

    Guo, C.; Zhao, B.; and Bai, Y. 2022. DeepCore: A Comprehensive Library for Coreset Selection in Deep Learning. In Database and Expert Systems Applications - 33rd International Conference, DEXA , volume 13426, 181--195. Springer

    2022

  6. [14]

    He, K.; Chen, X.; Xie, S.; Li, Y.; Doll \' a r, P.; and Girshick, R. B. 2022. Masked Autoencoders Are Scalable Vision Learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR , 15979--15988. IEEE

    2022

  7. [15]

    He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR , 770--778. IEEE Computer Society

    2016

  8. [16]

    Ioffe, S.; and Szegedy, C. 2015. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. In Bach, F. R.; and Blei, D. M., eds., Proceedings of the 32nd International Conference on Machine Learning, ICML , volume 37, 448--456

    2015

  9. [17]

    K.; Khargoankar, N.; Bilmes, J

    Iyer, R. K.; Khargoankar, N.; Bilmes, J. A.; and Asanani, H. 2021. Submodular combinatorial information measures with applications in machine learning. In Algorithmic Learning Theory, volume 132, 722--754. PMLR

    2021

  10. [18]

    Killamsetty, K.; Sivasubramanian, D.; Ramakrishnan, G.; De, A.; and Iyer, R. K. 2021. GRAD-MATCH: Gradient Matching based Data Subset Selection for Efficient Deep Model Training. In Proceedings of the 38th International Conference on Machine Learning, ICML , volume 139, 5464--...

    2021

  11. [19]

    J.; Yun, S.; Song, H.; Jeong, J.; Ha, J.; and Song, H

    Kim, J.; Kim, J.; Oh, S. J.; Yun, S.; Song, H.; Jeong, J.; Ha, J.; and Song, H. O. 2022. Dataset Condensation via Efficient Synthetic-Data Parameterization. In International Conference on Machine Learning, ICML , volume 162, 11102--11118. PMLR

    2022

  12. [20]

    Krizhevsky, A.; Hinton, G.; et al. 2009. Learning multiple layers of features from tiny images. Technical report Citeseer

    2009

  13. [21]

    Le, Y.; and Yang, X. 2015. Tiny imagenet visual recognition challenge. Technical Report, 7(7): 3

    2015

  14. [22]

    Lei, S.; and Tao, D. 2024. A Comprehensive Survey of Dataset Distillation. IEEE Trans. Pattern Anal. Mach. Intell. , 46(1): 17--32

    2024

  15. [23]

    P.; Tang, J.; and Liu, H

    Li, J.; Cheng, K.; Wang, S.; Morstatter, F.; Trevino, R. P.; Tang, J.; and Liu, H. 2018. Feature Selection: A Data Perspective. ACM Comput. Surv. , 50(6): 94:1--94:45

    2018

  16. [24]

    Liu, Z.; Hu, H.; Lin, Y.; Yao, Z.; Xie, Z.; Wei, Y.; Ning, J.; Cao, Y.; Zhang, Z.; Dong, L.; Wei, F.; and Guo, B. 2022 a . Swin Transformer V2: Scaling Up Capacity and Resolution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR , 11999--12009. IEEE

    2022

  17. [25]

    Liu, Z.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; and Guo, B. 2021. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV , 9992--10002. IEEE

    2021

  18. [26]

    Liu, Z.; Mao, H.; Wu, C.; Feichtenhofer, C.; Darrell, T.; and Xie, S. 2022 b . A ConvNet for the 2020s. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR , 11966--11976. IEEE

    2022

  19. [27]

    Paul, M.; Ganguli, S.; and Dziugaite, G. K. 2021. Deep Learning on a Data Diet: Finding Important Examples Early in Training. In Annual Conference on Neural Information Processing Systems, NeurIPS , 20596--20607

    2021

  20. [28]

    S.; Berg, A

    Russakovsky, O.; Deng, J.; Su, H.; Krause, J.; Satheesh, S.; Ma, S.; Huang, Z.; Karpathy, A.; Khosla, A.; Bernstein, M. S.; Berg, A. C.; and Fei - Fei, L. 2015. ImageNet Large Scale Visual Recognition Challenge. Int. J. Comput. Vis., 115(3): 211--252

    2015

  21. [29]

    G.; Zhu, M.; Zhmoginov, A.; and Chen, L

    Sandler, M.; Howard, A. G.; Zhu, M.; Zhmoginov, A.; and Chen, L. 2018. MobileNetV2: Inverted Residuals and Linear Bottlenecks. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR , 4510--4520

    2018

  22. [30]

    T.; Trischler, A.; Bengio, Y.; and Gordon, G

    Toneva, M.; Sordoni, A.; des Combes, R. T.; Trischler, A.; Bengio, Y.; and Gordon, G. J. 2019. An Empirical Study of Example Forgetting during Deep Neural Network Learning. In 7th International Conference on Learning Representations, ICLR . OpenReview.net

    2019

  23. [31]

    Wan, Z.; Wang, Z.; Wang, Y.; Wang, Z.; Zhu, H.; and Satoh, S. 2024. Contributing Dimension Structure of Deep Feature for Coreset Selection. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI , 9080--9088. AAAI Press

    2024

  24. [32]

    Wang, K.; Zhao, B.; Peng, X.; Zhu, Z.; Yang, S.; Wang, S.; Huang, G.; Bilen, H.; Wang, X.; and You, Y. 2022. CAFE: Learning to Condense Dataset by Aligning Features. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022 , 12186--12195. IEEE

    2022

  25. [33]

    Wang, T.; Zhu, J.; Torralba, A.; and Efros, A. A. 2018. Dataset Distillation. CoRR, abs/1811.10959

    2018 arXiv

  26. [34]

    Zhang, H.; Li, S.; Wang, P.; Zeng, D.; and Ge, S. 2024. M3D: Dataset Condensation by Minimizing Maximum Mean Discrepancy. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI , 9314--9322. AAAI Press

    2024

  27. [35]

    Zhang, L.; Zhang, J.; Lei, B.; Mukherjee, S.; Pan, X.; Zhao, B.; Ding, C.; Li, Y.; and Xu, D. 2023. Accelerating Dataset Distillation via Model Augmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023 , 11950--11959. IEEE

    2023

  28. [36]

    Zhao, B.; and Bilen, H. 2023. Dataset Condensation with Distribution Matching. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV , 6503--6512. IEEE

    2023

  29. [37]

    R.; and Bilen, H

    Zhao, B.; Mopuri, K. R.; and Bilen, H. 2021. Dataset Condensation with Gradient Matching. In 9th International Conference on Learning Representations, ICLR . OpenReview.net

    2021

  30. [38]

    Zhao, G.; Li, G.; Qin, Y.; and Yu, Y. 2023. Improved Distribution Matching for Dataset Condensation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR , 7856--7865. IEEE

    2023

  31. [39]

    Zhou, D.; Wang, K.; Gu, J.; Peng, X.; Lian, D.; Zhang, Y.; You, Y.; and Feng, J. 2023. Dataset Quantization. In IEEE/CVF International Conference on Computer Vision, ICCV , 17159--17170. IEEE

    2023

  32. [40]

    Zhou, Y.; Nezhadarya, E.; and Ba, J. 2022. Dataset Distillation using Neural Feature Regression. In Annual Conference on Neural Information Processing Systems NeurIPS

    2022

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.