Pith. sign in

REVIEW 4 major objections 6 minor 31 references

Contrastive Learning-Enhanced Trajectory Matching for Small-Scale Dataset Distillation

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Adding a supervised SimCLR-style contrastive loss to trajectory matching lets one-image-per-class synthetic datasets train models to 53.0% on CIFAR-10, up 6.1 points over the previous state of the art.

desk verdict Plausible but unproven: the gains over DATM may be pipeline effects, not the contrastive loss, and the paper's internal inconsistencies need fixing. read the letter →

arxiv 2505.15267 v2 pith:UTRDLSEL submitted 2025-05-21 cs.CV

classification cs.CV
keywords datasetdistillationcontrastivelearningtrajectorymatchinglow-dataregimesyntheticdataSimCLRCIFAR-10Tiny-ImageNet
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 show that the extreme low-data regime of dataset distillation—one or a few synthetic images per class—can be rescued by adding contrastive learning to trajectory matching. It proposes DATM-CLR, which takes the state-of-the-art DATM method and adds a supervised SimCLR-style loss that pulls augmented views of the same synthetic image together and pushes views from other classes apart. On CIFAR-10 with one image per class, the resulting synthetic dataset trains a ConvNet to 53.0% accuracy, up from DATM's 46.9%, with smaller but consistent gains on CIFAR-100 and Tiny-ImageNet. If the attribution holds, the contrastive objective is a cheap and transferable addition to existing trajectory-matching pipelines.

What carries the argument

The machinery is a two-term objective: the existing trajectory matching loss $L_{\mathrm{tm}}$ keeps the student's parameter path close to the teacher's, and a supervised SimCLR-style contrastive loss $L_{\mathrm{contrast}}$ on augmented synthetic images, defined by Eq. (2), forces instance-level feature discrimination in a projection-head embedding. Positive pairs are two different augmentations of the same synthetic image, and negatives are samples from other classes, so the loss explicitly uses label information. The load-bearing update is the joint gradient on the synthetic dataset through $\alpha L_{\mathrm{contrast}} + \beta L_{\mathrm{tm}}$, which is what makes synthetic images both trajectory-faithful and class-discriminative.

What would settle it

Run the paper's exact pipeline with the contrastive loss weight $\alpha$ set to zero while keeping the same augmentations, soft-label handling, inner steps, and trajectory sampling; if CIFAR-10 IPC=1 accuracy stays near 53.0% rather than dropping toward DATM's reported 46.9%, the central attribution to contrastive learning fails, whereas a drop to the DATM level would confirm the claim.

Watch

Extended reading notes

Core claim

The central claim is that the semantic impoverishment of synthetic datasets at very low IPC is not a fixed cost of trajectory matching; it can be offset by a carefully placed contrastive objective. The paper's DATM-CLR optimizes synthetic images so that a student's parameter trajectory on them matches a teacher's real-data trajectory, while simultaneously a projection head computes a normalized temperature-scaled cross-entropy loss over two augmented views of each synthetic image, with negatives drawn from other classes. The two losses are combined as $L_{\mathrm{total}} = \alpha L_{\mathrm{contrast}} + \beta L_{\mathrm{tm}}$; the contrastive loss is also optionally folded into the student's inner-loop updates (the 'Contrastive Update' strategy), though the ablation shows the synthetic-data-level 'Contrastive Fusion' works better. The result claimed is consistent accuracy gains over DATM and other trajectory-matching baselines at IPC=1, 10, and 50, with the largest absolute gain (6.1 points) at CIFAR-10 IPC=1.

Load-bearing premise

The reported gains come from the contrastive loss itself, not from the many implementation differences between the paper's pipeline and the DATM baseline it compares against—a premise the paper never directly tests because its ablation lacks a no-contrastive row under the identical pipeline.

Editorial extensions

If this is right

  • At CIFAR-10 IPC=1, the synthetic dataset reaches 53.0% test accuracy, 6.1 points above DATM, meaningfully narrowing the gap between one-image-per-class training and full-dataset training.
  • The gains extend to CIFAR-100 IPC=1 (+0.8 points) and Tiny-ImageNet IPC=1 (+1.2 points), and remain positive at IPC=10 and IPC=50, so the benefit is not confined to the single-image setting.
  • The distilled datasets transfer across architectures: on ConvNet, ResNet18, VGG, and AlexNet the paper reports accuracy above the DATM baseline in every case.
  • The ablation shows that Contrastive Fusion, which adds the contrastive loss to the synthetic-image update, beats Contrastive Update, which inserts the loss into the student's inner loop, at every dataset and IPC setting tested.
  • Hyperparameter sweeps show that increasing the contrastive fusion weight $\alpha$ up to 0.1 helps, while increasing the inner-loop weight $\lambda$ hurts, giving a concrete tuning rule for users of the method.

Reading between the lines

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

  • Editorial inference: the large IPC=1 gain suggests a specific mechanism—with a single image per class, trajectory matching alone may let synthetic images collapse toward one prototype, and the contrastive term adds discriminative pressure that the trajectory loss does not provide.
  • Editorial inference: the same supervised contrastive loss could be bolted onto other trajectory-matching baselines such as MTT, TESLA, or FTD, and if the gains reappear, the mechanism is general rather than specific to DATM's difficulty-aligned trajectory selection.
  • Editorial inference: a low-cost test of the claimed representation quality is to train students on DATM-CLR and DATM synthetic data and compare linear-probe or k-nearest-neighbor accuracy on their features; the paper's story predicts the contrastive-trained features are more linearly separable.
  • Editorial inference: the abstract's assertion that synthetic images become visually more faithful is not backed by any qualitative figure or perceptual metric in the paper, so a direct image-quality comparison between DATM and DATM-CLR outputs at IPC=1 would be a natural way to test that part of the claim.
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 DATM-CLR, a dataset distillation method that adds SimCLR-style supervised contrastive learning to the DATM trajectory-matching framework. The method combines a trajectory matching loss with a contrastive loss in two proposed variants, “Contrastive Fusion” and “Contrastive Update,” and is evaluated on CIFAR-10, CIFAR-100, and Tiny-ImageNet at low images-per-class settings. The main empirical claim is that the contrastive objective yields consistent gains over DATM, with the headline result being 53.0% versus 46.9% on CIFAR-10 at IPC=1, and smaller gains on CIFAR-100 and Tiny-ImageNet at IPC=1.

Significance. If the causal attribution were established, the reported improvement on CIFAR-10 IPC=1 (6.1 percentage points over DATM) would be a meaningful advance for low-data dataset distillation, and the method is simple enough to be useful as a baseline for future work. The paper's strengths are its direct head-to-head comparisons on standard benchmarks and its ablation of the two proposed integration strategies. However, the central causal claim is not yet supported because the ablation lacks a same-code no-contrastive control, and there are several reporting inconsistencies in the experimental section. The paper does not mention code or data release, which further limits reproducibility given that some hyperparameters are undisclosed.

major comments (4)
  1. [Section 4.3, Table 3] The ablation varies the two proposed integration strategies and single versus multiple augmentations, but it contains no no-contrastive row that disables L_contrast while keeping the same augmentation schedule, soft-label initialization and optimization, and inner-loop procedure. Since the only DATM comparison is the externally published number, the paper's central claim that the contrastive loss causes the 6.1/0.8/1.2 point gains over DATM is not causally verified; any implementation difference could account for part or all of the gain. Add a same-code DATM baseline (no contrastive loss) to Table 3 for at least CIFAR-10 IPC=1 and CIFAR-100 IPC=1.
  2. [Section 4.2, Table 1] The text reports “as IPC increases to 50, our method still achieves an accuracy of 70.2±0.3%, compared to 66.8±0.3%”, but Table 1 lists these values under CIFAR-10 IPC=10, while the IPC=50 entries are 77.6 and 76.1. For CIFAR-100, the text reports “At IPC=50, our method reaches 48.7±0.3%… DATM (47.2±0.4%)”, but Table 1 has no CIFAR-100 IPC=50 column; these numbers are the IPC=10 entries. This is a load-bearing reporting inconsistency and must be corrected.
  3. [Section 3.2, Section 3.3, Algorithm 1] Algorithm 1 line 11 and the surrounding text in Section 3.3 state that only the cross-entropy loss updates the student model, whereas the Contrastive Update strategy described by Eq. (4) and evaluated in Table 3 adds L_contrast to the inner-loop gradient. The algorithm as written implements only the Contrastive Fusion strategy, so it does not match the described method space; the authors should either update Algorithm 1 to include the Contrastive Update branch or explicitly state that the final method uses Fusion and relegate Contrastive Update to an ablation variant.
  4. [Section 3.2 and Section 4.3] The final values of the loss weights alpha and beta used in the main results are never reported. The text says lambda is fixed to 1 and Fig. 3a sweeps alpha, but no chosen alpha (or beta) value is given, and the figure is not readable enough to recover it. Because the method and its comparison depend on these weights and they appear to be tuned on the same benchmark sets, the missing reporting is a reproducibility and selection-risk issue. Report the exact weights and state whether the Table 1 numbers correspond to the best value found in the sweep.
minor comments (6)
  1. [Abstract and Section 1] The phrasing “6.1% improvement” should be “6.1 percentage points”, since it describes the absolute difference between 46.9% and 53.0%; the same applies to the 0.8% and 1.2% gains.
  2. [Table 2] Table 2 has no header row specifying the dataset and IPC; the text should state that these are CIFAR-10 IPC=1 results, and it should also explain why DenseNet, listed in Section 4.1, is missing from the table.
  3. [Section 4.2] The claim that the CIFAR-100 IPC=1 improvement “remains statistically significant” is unsupported because no significance test is reported; with five trials and overlapping error bars, this is not established.
  4. [Table 3] The caption of Table 3 is inconsistent with the main text: the text says “Single” versus “Multiple” refers to single versus multiple data augmentation, while the caption describes “Multiple” as “an alternative strategy using multiple synthetic trajectories.” Align the caption with the method description.
  5. [Throughout] Typographical errors should be cleaned, including “Datset” (Fig. 1), “belive” (Section 3.2), “syntetic” (Section 3.1), “knowledege” (Fig. 1), and “Puls” (Section 3.2).
  6. [Section 1] The introduction claims the method “exceeds both” DCC and MIM4DD, but neither method appears in Table 1 or anywhere in the experiments; either add these comparisons or weaken the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central improvement claim is an empirical head-to-head against an external baseline, not an identity or self-citation.

full rationale

This paper is an empirical method comparison rather than a derivation chain. Section 3.2 defines a combined objective L_total = α L_contrast + β L_tm (Eq. 3) and a student-update variant (Eq. 4); neither equation is defined in terms of the test accuracies reported in Table 1. The contrastive loss (Eq. 2) is a standard SimCLR InfoNCE objective, and the trajectory-matching loss (Eq. 1) is the external MTT/DATM formulation; no term encodes the final benchmark numbers. The claimed gains over DATM are comparisons against an external method (Ref. [5]) and are not forced by construction. There are no load-bearing self-citations: the authors of this paper are not authors of Refs. [1,2,5], and the cited prior work is independently published. The ablation in Table 3 compares two strategies that both contain the contrastive loss but it does not contain a no-contrastive row; that absence, together with undisclosed α values selected on the same benchmarks, is a genuine experimental-control and evaluation-leakage concern, but it is a correctness risk rather than a definitional reduction. The paper's own conclusion that balancing α and λ is 'dataset-dependent' is a limitation statement, not an admission that the result is equivalent to its input. No prediction in the paper reduces by construction to a fitted parameter or to a self-citation chain, so the circularity score is 0.

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

The method inherits its entire theoretical load from prior trajectory matching and contrastive learning literature. No new theoretical derivation is offered; the contribution is a specific loss combination. The main uncharged premises are that trajectory matching is a valid generalization proxy, that supervised contrastive pairs on synthetic images add useful signal, and that the DATM baseline is reproduced fairly.

free parameters (4)
  • alpha (Contrastive Fusion weight) = Not reported; Fig. 3a tunes alpha in [0, 0.1] and suggests higher alpha helps
    Eq. 3 weight on L_contrast when updating synthetic data. The final value is missing, so the exact recipe is incomplete.
  • beta (Trajectory matching weight) = Not reported
    Eq. 3 weight on L_tm; never specified in Section 4, though the headline results depend on it.
  • lambda (Contrastive Update weight) = 0.1 recommended (Section 4.3)
    Eq. 4 inner-loop contrastive weight. The ablation shows accuracy decreasing as lambda increases, so 0.1 is chosen, but the main method does not clearly use this strategy.
  • Contrastive temperature tau = 0.1
    Set in Section 4.1. Standard SimCLR temperature but it affects the scale of the contrastive loss and is tuned by hand.
assumptions (4)
  • domain assumption Trajectory matching objective Eq. (1) is a valid proxy for downstream generalization, per MTT [1] and DATM [5].
    The paper inherits this objective without independent justification; the whole method optimizes L_tm plus a contrastive term.
  • ad hoc to paper Supervised contrastive pairs built from synthetic images and their class labels provide useful feature-level signal.
    This is the paper's core design bet; no proof or prior validation specific to trajectory matching is offered.
  • domain assumption Soft labels from a pretrained teacher, continuously optimized as in DATM, remain valid when a contrastive loss is added.
    Section 4.1 says soft labels are jointly optimized; the interaction between the contrastive regularizer and label learning is not analyzed.
  • domain assumption Baseline DATM results in Table 1 are reproduced fairly with comparable compute and hyperparameters.
    The claimed gains are differences against DATM; if the baseline is under-tuned, the improvement is inflated. No code or logs are shipped.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Contrastive Learning-Enhanced Trajectory Matching for Small-Scale Dataset Distillation." pith.science (2026). https://pith.science/paper/UTRDLSEL

@misc{pith2026250515267,
  author       = {Pith},
  title        = {Pith review of: Contrastive Learning-Enhanced Trajectory Matching for Small-Scale Dataset Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UTRDLSEL}},
  note         = {Machine review of arXiv:2505.15267}
}
read the original abstract

Deploying machine learning models in resource-constrained environments, such as edge devices or rapid prototyping scenarios, increasingly demands distillation of large datasets into significantly smaller yet informative synthetic datasets. Current dataset distillation techniques, particularly Trajectory Matching methods, optimize synthetic data so that the model's training trajectory on synthetic samples mirrors that on real data. While demonstrating efficacy on medium-scale synthetic datasets, these methods fail to adequately preserve semantic richness under extreme sample scarcity. To address this limitation, we propose a novel dataset distillation method integrating contrastive learning during image synthesis. By explicitly maximizing instance-level feature discrimination, our approach produces more informative and diverse synthetic samples, even when dataset sizes are significantly constrained. Experimental results demonstrate that incorporating contrastive learning substantially enhances the performance of models trained on very small-scale synthetic datasets. This integration not only guides more effective feature representation but also significantly improves the visual fidelity of the synthesized images. Experimental results demonstrate that our method achieves notable performance improvements over existing distillation techniques, especially in scenarios with extremely limited synthetic data.

Figures

Figures reproduced from arXiv: 2505.15267 by the authors.

Figure 1
Figure 1. (a) The teacher model transfers knowledege to the student model through [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the synthetic data optimization process integrating Tra [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Variation of different effects of hyperparameters in different strategies [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 20 canonical work pages

  1. [1]

    In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (2022)

    Cazenavette, G., Wang, T., Torralba, A., Efros, A.A., Zhu, J.Y.: Dataset distilla- tion by matching training trajectories. In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (2022)

  2. [2]

    In: Proceedings of the 37th International Conference on Machine Learning

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con- trastive learning of visual representations. In: Proceedings of the 37th International Conference on Machine Learning. ICML’20, JMLR.org (2020)

  3. [3]

    In: Proceedings of the 40th International Conference on Machine Learning

    Cui, J., Wang, R., Si, S., Hsieh, C.J.: Scaling up dataset distillation to imagenet- 1k with constant memory. In: Proceedings of the 40th International Conference on Machine Learning. ICML’23, JMLR.org (2023)

  4. [4]

    Du, J., Jiang, Y., Tan, V.Y.F., Zhou, J.T., 0001, H.L.: Minimizing the accumulated trajectoryerrortoimprovedatasetdistillation.In:IEEE/CVFConferenceonCom- puter Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023. pp. 3749–3758. IEEE (2023). https://doi.org/10.1109/CVPR52729. 2023.00365, https://doi.org/10.1109/CVPR52729.2023.00365

  5. [5]

    In: The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Aus- tria, May 7-11, 2024

    Guo, Z., Wang, K., Cazenavette, G., Li, H., Zhang, K., You, Y.: Towards loss- less dataset distillation via difficulty-aligned trajectory matching. In: The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Aus- tria, May 7-11, 2024. OpenReview.net (2024), https://openreview.net/forum?id= rTBL8OhdhH

  6. [6]

    2020 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) pp

    He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.B.: Momentum contrast for un- supervised visual representation learning. 2020 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) pp. 9726–9735 (2019), https: //api.semanticscholar.org/CorpusID:207930212

  7. [7]

    In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770–778 (2016). https://doi.org/10.1109/CVPR.2016.90

  8. [8]

    CoRRabs/1503.02531 (2015), http://dblp.uni-trier.de/db/journals/corr/ corr1503.html#HintonVD15 14 F

    Hinton, G.E., Vinyals, O., Dean, J.: Distilling the knowledge in a neural net- work. CoRRabs/1503.02531 (2015), http://dblp.uni-trier.de/db/journals/corr/ corr1503.html#HintonVD15 14 F. Wenmin Li et al

Show all 31 references
  1. [9]

    In: 7th International Conference on Learning Rep- resentations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019

    Hjelm, R.D., Fedorov, A., Lavoie-Marchildon, S., Grewal, K., Bachman, P., Trischler, A., Bengio, Y.: Learning deep representations by mutual information estimation and maximization. In: 7th International Conference on Learning Rep- resentations, ICLR 2019, New Orleans, LA, USA...

  2. [10]

    Huang, G., Liu, Z., van der Maaten, L., Weinberger, K.Q.: Densely connected convolutionalnetworks.In:2017IEEEConferenceonComputerVisionandPattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 2261–2269. IEEE Computer Society (2017). https://doi.org/10.1109/C...

  3. [11]

    Technologies9(1), 2 (2021)

    Jaiswal, A., Babu, A.R., Zadeh, M.Z., Banerjee, D., Makedon, F.: A survey on contrastive self-supervised learning. Technologies9(1), 2 (2021)

  4. [12]

    Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning (2020)

  5. [13]

    IEEE Transactions on Circuits and Systems for Video Tech- nology 34(1), 140–152 (Jan 2024)

    Kosugi, S., Yamasaki, T.: Personalized image enhancement featuring masked style modeling. IEEE Transactions on Circuits and Systems for Video Tech- nology 34(1), 140–152 (Jan 2024). https://doi.org/10.1109/tcsvt.2023.3285765, http://dx.doi.org/10.1109/TCSVT.2023.3285765

  6. [14]

    32–33 (2009), https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf

    Krizhevsky, A.: Learning multiple layers of features from tiny images pp. 32–33 (2009), https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf

  7. [15]

    Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Commun. ACM60(6), 84–90 (May 2017). https: //doi.org/10.1145/3065386, https://doi.org/10.1145/3065386

  8. [16]

    semanticscholar.org/CorpusID:16664790

    Le, Y., Yang, X.S.: Tiny imagenet visual recognition challenge (2015), https://api. semanticscholar.org/CorpusID:16664790

  9. [17]

    Journal of Statistical Mechanics: Theory and Experi- ment 2020(12), 124002 (Dec 2020)

    Lee, J., Xiao, L., Schoenholz, S.S., Bahri, Y., Novak, R., Sohl-Dickstein, J., Pen- nington, J.: Wide neural networks of any depth evolve as linear models un- der gradient descent *. Journal of Statistical Mechanics: Theory and Experi- ment 2020(12), 124002 (Dec 2020). https:/...

  10. [18]

    In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S

    Lee,S.,Chun,S.,Jung,S.,Yun,S.,Yoon,S.:Datasetcondensationwithcontrastive signals. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research...

  11. [19]

    Loo, N., Hasani, R., Lechner, M., Rus, D.: Dataset distillation with convexified implicit gradients (2023), https://arxiv.org/abs/2302.06755

  12. [20]

    In: ICLR 2021 (2021), https://openreview.net/forum?id= l-PrrQrK0QR

    Nguyen, T.C., Chen, Z., Lee, J.: Dataset meta-learning from kernel ridge-regression. In: ICLR 2021 (2021), https://openreview.net/forum?id= l-PrrQrK0QR

  13. [21]

    CoRR abs/1807.03748 (2018), http://dblp.uni-trier.de/db/ journals/corr/corr1807.html#abs-1807-03748

    van den Oord, A., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. CoRR abs/1807.03748 (2018), http://dblp.uni-trier.de/db/ journals/corr/corr1807.html#abs-1807-03748

  14. [22]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)

    Radosavovic, I., Dollár, P., Girshick, R., Gkioxari, G., He, K.: Data distillation: Towards omni-supervised learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)

  15. [23]

    In: Thirty-seventh Conference on Neural Information Process- ing Systems (2023), https://openreview.net/forum?id=TZtw5YgxTE Dataset Distillation with Contrastive Learning 15

    Shang, Y., Yuan, Z., Yan, Y.: MIM4DD: Mutual information maximization for dataset distillation. In: Thirty-seventh Conference on Neural Information Process- ing Systems (2023), https://openreview.net/forum?id=TZtw5YgxTE Dataset Distillation with Contrastive Learning 15

  16. [24]

    In: Bengio, Y., LeCun, Y

    Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: Bengio, Y., LeCun, Y. (eds.) ICLR (2015), http://dblp. uni-trier.de/db/conf/iclr/iclr2015.html#SimonyanZ14a

  17. [25]

    In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR)

    Wang, K., Zhao, B., Peng, X., Zhu, Z., Yang, S., Wang, S., Huang, G., Bilen, H., Wang, X., You, Y.: Cafe: Learning to condense dataset by aligning fea- tures. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR). pp. 12186–12195 (2022). https://doi....

  18. [26]

    Wang, T., Zhu, J.Y., Torralba, A., Efros, A.A.: Dataset distillation (2020), https: //arxiv.org/abs/1811.10959

  19. [27]

    Zhao, B., Bilen, H.: Dataset condensation with differentiable siamese augmentation (2021), https://arxiv.org/abs/2102.08259

  20. [28]

    In: Pro- ceedings of the IEEE/CVF Winter Conference on Applications of Computer Vi- sion 2023 (WACV)

    Zhao, B., Bilen, H.: Dataset condensation with distribution matching. In: Pro- ceedings of the IEEE/CVF Winter Conference on Applications of Computer Vi- sion 2023 (WACV). pp. 6503–6512. IEEE Winter Conference on Applications of Computer Vision, Institute of Electrical and Ele...

  21. [29]

    CoRR abs/2006.05929 (2020), https://arxiv.org/abs/2006.05929

    Zhao, B., Mopuri, K.R., Bilen, H.: Dataset condensation with gradient matching. CoRR abs/2006.05929 (2020), https://arxiv.org/abs/2006.05929

  22. [30]

    In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR).pp.7856–7865(2023).https://doi.org/10.1109/CVPR52729

    Zhao, G., Li, G., Qin, Y., Yu, Y.: Improved distribution matching for dataset condensation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR).pp.7856–7865(2023).https://doi.org/10.1109/CVPR52729. 2023.00759

  23. [31]

    In: Proceedings of the 36th International Conference on Neural Information Processing Systems

    Zhou, Y., Nezhadarya, E., Ba, J.: Dataset distillation using neural feature regres- sion. In: Proceedings of the 36th International Conference on Neural Information Processing Systems. NIPS ’22, Curran Associates Inc., Red Hook, NY, USA (2022)

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.