Pith. sign in

REVIEW 3 major objections 5 minor 53 references

A Survey on Data Curation for Visual Contrastive Learning: Why Crafting Effective Positive and Negative Pairs Matters

T0 review · 3 major / 5 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 survey argues that how contrastive pairs are curated determines representation quality, training speed, and compute cost.

desk verdict A useful but under-built survey: the pair-curation taxonomy is a good starting map, but category boundaries are messy and the coverage claim outruns the method. read the letter →

arxiv 2502.08134 v1 pith:FCUMCXKB submitted 2025-02-12 cs.CV

Shasvat Desai , Debasmita Ghose , Deep Chakraborty This is my paper · ORCID
classification cs.CV
keywords contrastivelearningdatacurationpositivepairsnegativeself-supervisedrepresentationtaxonomyhardnegatives
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

The paper's central claim is that in visual contrastive learning, the choice of which samples form positive and negative pairs—what the authors call data curation—directly controls representation quality, training efficiency, and computational cost. It offers a taxonomy that splits positive-pair creation into single-instance and multi-instance techniques, and negative-pair creation into hard-negative selection, false-negative elimination, and synthetic negatives. A sympathetic reader would care because most surveys focus on architectures or loss functions, while this one isolates the data-side lever, arguing that well-curated pairs yield stronger representations and faster convergence. The paper is a survey, so its contribution is organizational: it groups existing methods, describes their trade-offs, and poses open questions about balancing diversity and semantic relevance.

What carries the argument

The central object is a taxonomy of pair-curation techniques, summarized in Figures 1–3. It maps every surveyed method onto one of two positive-pair branches (single-instance vs. multi-instance, the latter subdivided into embedding-based, synthetic, supervised, attribute-based, and cross-modal) and one of three negative-pair branches (hard-negative selection, false-negative elimination, synthetic negatives). The carrying mechanism is the InfoNCE loss, whose softmax over similarity scores is where pair composition enters the learning objective; the taxonomy turns 'data curation' from a vague concern into a structured design space.

What would settle it

Run a standard contrastive baseline (e.g., SimCLR on ImageNet) with random positive/negative pairing against each curated strategy described, holding architecture, batch size, and training steps fixed; if curated pairs show no consistent gain in linear-evaluation accuracy, the central claim that curation significantly impacts representation quality would be undermined. Alternatively, finding a published pair-curation method that cannot be placed in the taxonomy would refute its claimed completeness.

Watch

Extended reading notes

Core claim

The discovery is that data curation deserves to be treated as a first-class design axis in contrastive learning, separate from architecture and loss design. The authors organize the field into two positive-pair families—single-instance (augmentations of one sample) and multi-instance (embedding-based neighbors, synthetic generation, supervised labels, attribute grouping, and cross-modal alignment)—and three negative-pair strategies (hard-negative selection, false-negative elimination, and synthetic negatives). Each strategy trades off diversity, semantic alignment, computational cost, and risk of overfitting or noise. The paper argues that these trade-offs mean no single curation strategy dominates, and that future work should find principled ways to balance diversity and relevance, and to handle emerging modalities.

Load-bearing premise

The taxonomy is asserted to be a complete and accurate organization of existing pair-curation methods, but the paper does not document a systematic literature search, inclusion criteria, or comparison with prior survey taxonomies.

Editorial extensions

If this is right

  • Practitioners should treat pair selection as a hyperparameter: choosing single-instance versus multi-instance positives and curated versus random negatives should depend on the downstream task and data budget.
  • Multi-instance and synthetic pairs can mitigate rare modalities and category imbalance, so curation is a route to broader coverage without acquiring new raw data.
  • Hard negatives improve discrimination but risk overfitting, while false-negative elimination reduces noise but risks removing valid hard examples; the correct balance is dataset- and task-dependent.
  • Effective curation promises faster convergence and lower training cost, which matters as contrastive pre-training scales to larger datasets and models.

Reading between the lines

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

  • The taxonomy suggests a natural extension: an automated curator that learns which pair-creation strategy to apply per sample or per training phase, something the paper does not propose.
  • The diversity-versus-relevance trade-off could be framed as a curriculum question: early training may favor diverse pairs and later training task-aligned ones; the authors leave this open, and a dynamic schedule is a testable next step.
  • The same curation axes may extend to modalities the paper only touches, such as LiDAR, hyperspectral imaging, or haptic data, where synthetic or attribute-based pairing could compensate for scarce data.
  • A quantitative meta-analysis of the cited methods, controlling for architecture and compute, would be needed to confirm that curation rather than scale drives the reported gains; the paper provides no such comparison.
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

3 major / 5 minor

Summary. This survey organizes data curation techniques for visual contrastive learning into a taxonomy of positive-pair creation (single-instance vs. multi-instance, with multi-instance subdivided into embedding-based, synthetic, supervised, attribute-based, and cross-modal) and negative-pair creation (hard negative selection, false negative elimination, and synthetic negatives). It describes representative methods for each category, discusses trade-offs between them, and closes with open research questions. The paper makes no new empirical claims; its stated contribution is the taxonomy itself as a framework for understanding how positive and negative pair construction affects representation quality, training efficiency, and computational cost.

Significance. If the taxonomy were reliable and complete, it would fill a complementary niche relative to existing surveys that focus on architectures or loss objectives in contrastive learning. The paper has several strengths: it covers many recent works, gives concrete examples in each category, and articulates practical trade-offs and open questions. The discussion of synthetic positives/negatives and of attribute-based pairing is particularly timely. However, the central contribution is the taxonomy, and the paper does not yet demonstrate that the taxonomy is a faithful and complete map of the literature. The lack of a documented literature-selection protocol, the absence of comparison with prior survey taxonomies, and the presence of category placements that conflict with the paper's own definitions undermine the main claim. These issues are fixable, but they bear directly on the value of the survey as a reference.

major comments (3)
  1. [Sec. 4, Figs. 1 and 3] The negative-pair taxonomy is not mutually exclusive, and some method placements contradict the paper's own definitions. MoCHi is discussed under Hard Negative Selection (Sec. 4.1) even though the text describes it as 'creat[ing] synthetic negatives' by combining hard-negative features, which is the same operation that defines Synthetic Hard Negatives in Sec. 4.3. Similarly, Chuang et al. is placed under False Negative Elimination (Sec. 4.2) even though the description says it 'introduces a re-weighted loss function' rather than eliminating or reclassifying false negatives. This makes it difficult for a reader to map methods onto the taxonomy. Please specify a primary-category assignment rule (e.g., by the main algorithmic mechanism versus the training objective) and re-place methods accordingly, or explicitly discuss intentional overlaps.
  2. [Sec. 2 and Introduction] The survey does not document its literature-selection methodology. There is no search strategy, inclusion/exclusion criteria, or statement of the period and venues covered, nor is there a comparison with taxonomies from prior surveys (e.g., [Gui et al., 2024; Jaiswal et al., 2020]). Since the paper's stated aim is to 'create a taxonomy of existing techniques,' completeness is a load-bearing property. Please add a methodology paragraph (or at least a scope statement) and a table or appendix listing methods per category with references; also position the proposed taxonomy relative to existing ones to make clear what is new.
  3. [Sec. 4.1] The description of [Robinson et al., 2020] is inaccurate: that work proposes a hard-negative sampling distribution and a debiased contrastive objective, not 'adversarial[ly] generated or synthesized through feature interpolation' negatives as stated. Since this entry is used as an example of hard negative selection, the misdescription reduces confidence in the survey's reliability. Please correct this description and check the other brief method summaries for similar precision.
minor comments (5)
  1. [Sec. 3.1] There are typos in method names: 'SwA V' should be 'SwAV' and 'VicReg' should be 'VICReg.' Please also introduce the abbreviations consistently at first use.
  2. [Sec. 2, Eq. (1)] In the InfoNCE equation, the denominator includes both the positive term and a sum over j=1 to N of negative terms, but the text does not clarify whether the positive sample is excluded from the negative set. Please add a note that the sum is over negatives only, or write the denominator with the positive term separated explicitly.
  3. [Fig. 2 caption] The caption contains the typo 'Pre-de/f_ined Attributes'; this should read 'Pre-defined Attributes.' Also, panel (e) is referenced in the text as 'Attributed-based,' which should be 'Attribute-based.'
  4. [Sec. 3.1] The survey includes methods that do not use negative pairs (SimSiam, BYOL, Barlow Twins, VICReg, SwAV) under 'Single Instance Positives.' This is reasonable as context, but it may confuse readers because the paper's framing is about positive and negative pair curation. Please add a sentence explaining that these non-contrastive methods are included as single-instance positive baselines even though they lack explicit negative-pair curation.
  5. [Sec. 3.2, 'Synthetic Data Generation for Positive Pairs'] The phrase 'By interpolating Gaussian noise with diffusion-based features' is vague; it is unclear whether the interpolation happens in latent space, pixel space, or noise space. Please clarify the mechanism of CLSP to give readers an accurate picture of the method.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the survey's claims are organizational and rest on external citations, not on fitted parameters or self-referential derivations.

full rationale

This is a survey paper; its central claim is that pair design affects representation quality and that a taxonomy of positive and negative pair curation techniques is useful. There is no fitted parameter, no predicted quantity, and no equation whose output is defined by its input. The two self-citations (Ghose et al. 2023 and Panambur et al. 2022) appear only as example methods within the taxonomy (Sections 3.2 and 4), not as evidence for the taxonomy's completeness or for the survey's central thesis; they are not load-bearing and do not constitute circular support. The taxonomy itself is asserted without a documented literature-selection protocol, and category boundaries are arguable (for example, MoCHi's feature mixing appears under Hard Negative Selection while a similar operation is discussed under Synthetic Negatives, and Chuang et al. reweights negatives rather than eliminating them), but that is a correctness or completeness concern, not circularity under the stated criteria. No step in the paper reduces by construction to its own inputs, so the circularity score is 0.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

No free parameters or invented entities are introduced. The survey assumes the InfoNCE contrastive framework as its organizing lens (Section 2) and assumes its taxonomy is complete and accurate (Section 2, Figures 1-3). These are background assumptions typical of a review, not fitted quantities.

assumptions (2)
  • domain assumption The InfoNCE loss shown in Section 2 is an adequate canonical framework for discussing contrastive pair curation.
    All taxonomy categories are described as ways of choosing pairs within this loss; alternative objectives such as Barlow Twins or VICReg are mentioned but not integrated into the taxonomy.
  • domain assumption The taxonomy in Section 2 and Figures 1-3 faithfully and exhaustively covers existing positive and negative pair-curation techniques.
    The paper asserts these categories without a systematic literature search or external validation, despite presenting the taxonomy as comprehensive.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Survey on Data Curation for Visual Contrastive Learning: Why Crafting Effective Positive and Negative Pairs Matters." pith.science (2026). https://pith.science/paper/FCUMCXKB

@misc{pith2026250208134,
  author       = {Pith},
  title        = {Pith review of: A Survey on Data Curation for Visual Contrastive Learning: Why Crafting Effective Positive and Negative Pairs Matters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FCUMCXKB}},
  note         = {Machine review of arXiv:2502.08134}
}
read the original abstract

Visual contrastive learning aims to learn representations by contrasting similar (positive) and dissimilar (negative) pairs of data samples. The design of these pairs significantly impacts representation quality, training efficiency, and computational cost. A well-curated set of pairs leads to stronger representations and faster convergence. As contrastive pre-training sees wider adoption for solving downstream tasks, data curation becomes essential for optimizing its effectiveness. In this survey, we attempt to create a taxonomy of existing techniques for positive and negative pair curation in contrastive learning, and describe them in detail.

Figures

Figures reproduced from arXiv: 2502.08134 by the authors.

Figure 1
Figure 1. Taxonomy for crafting positive and negative pairs. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Positive Pair Curation Techniques: Positive pair selection can utilize single-instance and multi-instance techniques. (a) Single￾instance curation applies augmentations to a single sample. On the other hand, multi-instance positive pair generation can be further classified into several category of techniques. (b) Embedding-based retrieves the top-K nearest neighbors of the anchor sample’s augmentation in the embeddi… view at source ↗
Figure 3
Figure 3. Negative Pair Curation Techniques: This figure shows three categories of techniques for negative pair curation. (a). Hard Negative Selection prioritizes negatives that are semantically similar to the anchor sample, such as a different cat breed, instead of an unrelated category like an airplane. The negatives are then augmented and fed into the encoder. (b). False Negative Elimination removes or reclassifies negativ… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 37 canonical work pages

  1. [1]

    Flamingo: a visual language model for few-shot learning

    [Alayrac et al., 2022] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS,

    work page 2022

  2. [5]

    Learning from positive and unlabeled data: A survey

    [Bekker and Davis, 2020] Jessa Bekker and Jesse Davis. Learning from positive and unlabeled data: A survey. Ma- chine Learning, 109(4),

    work page 2020

  3. [6]

    Unsupervised learning of visual features by con- trasting cluster assignments

    [Caron et al., 2020] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by con- trasting cluster assignments. NeurIPS,

    work page 2020

  4. [7]

    Emerging properties in self-supervised vi- sion transformers

    [Caron et al., 2021] Mathilde Caron, Hugo Touvron, Ishan Misra, Herv´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vi- sion transformers. In ICCV,

    work page 2021

  5. [8]

    Chen and K

    [Chen and He, 2021] X. Chen and K. He. Exploring simple siamese representation learning. In CVPR,

    work page 2021

  6. [9]

    [Chen et al., 2020] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple framework for contrastive learning of visual representations. In ICML. PMLR,

    work page 2020

  7. [10]

    Incremental false negative detection for contrastive learning

    [Chen et al., 2021] Tsai-Shien Chen, Wei-Chih Hung, Hung-Yu Tseng, Shao-Yi Chien, and Ming-Hsuan Yang. Incremental false negative detection for contrastive learning. arXiv preprint arXiv:2106.03719,

    arXiv 2021

  8. [11]

    Debiased contrastive learning

    [Chuang et al., 2020] Ching-Yao Chuang, Joshua Robinson, Yen-Chen Lin, Antonio Torralba, and Stefanie Jegelka. Debiased contrastive learning. NeurIPS,

    work page 2020

Show all 53 references
  1. [12]

    Ad- vancing multi-modal sensing through expandable modal- ity alignment

    [Dai et al., 2024] Shenghong Dai, Shiqi Jiang, Yifan Yang, Ting Cao, Mo Li, Suman Banerjee, and Lili Qiu. Ad- vancing multi-modal sensing through expandable modal- ity alignment. arXiv preprint arXiv:2407.17777,

    2024 arXiv

  2. [14]

    Dwibedi, Y

    [Dwibedi et al., 2021] D. Dwibedi, Y . Aytar, J. Tompson, P. Sermanet, and A. Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of vi- sual representations. In ICCV,

    2021

  3. [15]

    All4one: Symbiotic neighbour contrastive learning via self-attention and re- dundancy reduction

    [Estepa et al., 2023] Imanol G Estepa, Ignacio Saras´ua, Bha- laji Nagarajan, and Petia Radeva. All4one: Symbiotic neighbour contrastive learning via self-attention and re- dundancy reduction. In ICCV,

    2023

  4. [17]

    Seed: Self-supervised distillation for visual representation.arXiv preprint arXiv:2101.04731,

    [Fang et al., 2021] Zhiyuan Fang, Jianfeng Wang, Lijuan Wang, Lei Zhang, Yezhou Yang, and Zicheng Liu. Seed: Self-supervised distillation for visual representation.arXiv preprint arXiv:2101.04731,

    2021 arXiv

  5. [18]

    Cloob: Modern hopfield networks with in- foloob outperform clip

    [F¨urst et al., 2022] Andreas F ¨urst, Elisabeth Rumetshofer, Johannes Lehner, Viet T Tran, Fei Tang, Hubert Ramsauer, David Kreil, Michael Kopp, G ¨unter Klambauer, Angela Bitto, et al. Cloob: Modern hopfield networks with in- foloob outperform clip. NeurIPS,

    2022

  6. [19]

    Tailoring visual object representations to human requirements: A case study with a recycling robot

    [Ghose et al., 2023] Debasmita Ghose, Michal Adam Lewkowicz, Kaleb Gezahegn, Julian Lee, Timothy Adam- son, Marynel V ´azquez, and Brian Scassellati. Tailoring visual object representations to human requirements: A case study with a recycling robot. In Conference on Robot Lear...

    2023

  7. [20]

    A review on discriminative self-supervised learning methods

    [Giakoumoglou and Stathaki, 2024a] Nikolaos Giak- oumoglou and Tania Stathaki. A review on discriminative self-supervised learning methods. arXiv preprint arXiv:2405.04969,

    arXiv

  8. [21]

    Synco: Synthetic hard negatives in contrastive learning for better unsupervised visual representations

    [Giakoumoglou and Stathaki, 2024b] Nikolaos Giak- oumoglou and Tania Stathaki. Synco: Synthetic hard negatives in contrastive learning for better unsupervised visual representations. arXiv:2410.02401,

    arXiv

  9. [22]

    Strub, F

    [Grill et al., 2020] JB Grill, F. Strub, F. Altch ´e, et al. Boot- strap your own latent-a new approach to self-supervised learning. NeurIPS,

    2020

  10. [23]

    A survey on self-supervised learning: Algorithms, applications, and future trends

    [Gui et al., 2024] Jie Gui, Tuo Chen, Jing Zhang, Qiong Cao, Zhenan Sun, Hao Luo, and Dacheng Tao. A survey on self-supervised learning: Algorithms, applications, and future trends. TPAMI,

    2024

  11. [24]

    Audioclip: Extending clip to image, text and audio

    [Guzhov et al., 2022] Andrey Guzhov, Federico Raue, J ¨orn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP. IEEE,

    2022

  12. [25]

    Momentum contrast for unsu- pervised visual representation learning

    [He et al., 2020] Kaiming He, Haoqi Fan, Yuxin Wu, Sain- ing Xie, and Ross Girshick. Momentum contrast for unsu- pervised visual representation learning. In CVPR,

    2020

  13. [26]

    Adco: Adversarial contrast for efficient learn- ing of unsupervised representations from self-trained neg- ative adversaries

    [Hu et al., 2021] Qianjiang Hu, Xiao Wang, Wei Hu, and Guo-Jun Qi. Adco: Adversarial contrast for efficient learn- ing of unsupervised representations from self-trained neg- ative adversaries. In CVPR,

    2021

  14. [27]

    Boosting contrastive self-supervised learning with false negative cancellation

    [Huynh et al., 2022] Tri Huynh, Simon Kornblith, Matthew R Walter, Michael Maire, and Maryam Khademi. Boosting contrastive self-supervised learning with false negative cancellation. In WACV,

    2022

  15. [28]

    A survey on contrastive self-supervised learn- ing

    [Jaiswal et al., 2020] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learn- ing. Technologies, 9(1),

    2020

  16. [29]

    Scaling up visual and vision-language representation learning with noisy text su- pervision

    [Jia et al., 2021] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text su- pervision. In ICML. PMLR,

    2021

  17. [30]

    Kalantidis, M.B

    [Kalantidis et al., 2020] Y . Kalantidis, M.B. Sariyildiz, N. Pion, P. Weinzaepfel, and D. Larlus. Hard negative mixing for contrastive learning. NeurIPS,

    2020

  18. [31]

    Supervised con- trastive learning

    [Khosla et al., 2020] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised con- trastive learning. NeurIPS,

    2020

  19. [32]

    Mean shift for self-supervised learning

    [Koohpayegani et al., 2021] Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Pirsiavash. Mean shift for self-supervised learning. In ICCV,

    2021

  20. [33]

    Visualbert: A simple and performant baseline for vision and language

    [Li et al., 2019] Liunian Harold Li, Mark Yatskar, D Yin, CJ Hsieh, and KW Chang. Visualbert: A simple and performant baseline for vision and language. arxiv

    2019

  21. [35]

    Mapl: Parameter-efficient adaptation of uni- modal pre-trained models for vision-language few-shot prompting

    [Ma˜nas et al., 2022] Oscar Ma˜nas, Pau Rodriguez, Saba Ah- madi, Aida Nematzadeh, Yash Goyal, and Aishwarya Agrawal. Mapl: Parameter-efficient adaptation of uni- modal pre-trained models for vision-language few-shot prompting. arXiv preprint arXiv:2210.07179,

    2022 arXiv

  22. [36]

    Audio-visual instance discrimination with cross-modal agreement

    [Morgado et al., 2021] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audio-visual instance discrimination with cross-modal agreement. In CVPR,

    2021

  23. [37]

    Representation learning with contrastive predic- tive coding

    [Oord et al., 2018] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predic- tive coding. arXiv preprint arXiv:1807.03748,

    2018 arXiv

  24. [39]

    Focus on the pos- itives: Self-supervised learning for biodiversity monitor- ing

    [Pantazis et al., 2021] Omiros Pantazis, Gabriel J Brostow, Kate E Jones, and Oisin Mac Aodha. Focus on the pos- itives: Self-supervised learning for biodiversity monitor- ing. In ICCV,

    2021

  25. [40]

    Learning transferable visual models from nat- ural language supervision

    [Radford et al., 2021] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from nat- ural language supervision. In ICML. PMLR,

    2021

  26. [41]

    Contrastive learning with hard negative samples

    [Robinson et al., 2020] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. arXiv preprint arXiv:2010.04592,

    2020 arXiv

  27. [42]

    Singh, R

    [Singh et al., 2022] A. Singh, R. Hu, V . Goswami, G. Coua- iron, W. Galuba, M. Rohrbach, and D. Kiela. Flava: A foundational language and vision alignment model. In CVPR,

    2022

  28. [43]

    Hard negative sampling strategies for contrastive representation learning

    [Tabassum et al., 2022] Afrina Tabassum, Muntasir Wahed, Hoda Eldardiry, and Ismini Lourentzou. Hard negative sampling strategies for contrastive representation learning. arXiv preprint arXiv:2206.01197,

    2022 arXiv

  29. [44]

    Lxmert: Learning cross-modality encoder representations from transformers

    [Tan and Bansal, 2019] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490,

    2019 arXiv

  30. [45]

    Udandarao, N

    [Udandarao et al., 2024] V . Udandarao, N. Parthasarathy, M.F. Naeem, T. Evans, S. Albanie, F. Tombari, Y . Xian, A. Tonioni, and O.J. H ´enaff. Active data curation ef- fectively distills large-scale multimodal models. arXiv preprint arXiv:2411.18674,

    2024 arXiv

  31. [46]

    Oracle-guided contrastive clustering

    [Wang et al., 2022a] Mengdie Wang, Liyuan Shang, Suyun Zhao, Yiming Wang, Hong Chen, Cuiping Li, and Xizhao Wang. Oracle-guided contrastive clustering. arXiv preprint arXiv:2211.00409,

    arXiv

  32. [47]

    Image as a foreign lan- guage: Beit pretraining for all vision and vision-language tasks

    [Wang et al., 2022b] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, et al. Image as a foreign lan- guage: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442,

    arXiv

  33. [48]

    Wav2clip: Learning robust audio representations from clip

    [Wu et al., 2022] Ho-Hsiang Wu, Prem Seetharaman, Kun- dan Kumar, and Juan Pablo Bello. Wav2clip: Learning robust audio representations from clip. In ICASSP,

    2022

  34. [49]

    Cit: Curation in training for effective vision-language data

    [Xu et al., 2023] Hu Xu, Saining Xie, Po-Yao Huang, Licheng Yu, Russell Howes, Gargi Ghosh, Luke Zettle- moyer, and Christoph Feichtenhofer. Cit: Curation in training for effective vision-language data. InCVPR,

    2023

  35. [50]

    Audiotoken: Adaptation of text- conditioned diffusion models for audio-to-image genera- tion

    [Yariv et al., 2023] Guy Yariv, Itai Gat, Lior Wolf, Yossi Adi, and Idan Schwartz. Audiotoken: Adaptation of text- conditioned diffusion models for audio-to-image genera- tion. arXiv preprint arXiv:2305.13050,

    2023 arXiv

  36. [51]

    Zbontar, L

    [Zbontar et al., 2021] J. Zbontar, L. Jing, I. Misra, Y . LeCun, and S. Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML. PMLR,

    2021

  37. [52]

    Contrastive learning with syn- thetic positives

    [Zeng et al., 2024] Dewen Zeng, Yawen Wu, Xinrong Hu, Xiaowei Xu, and Yiyu Shi. Contrastive learning with syn- thetic positives. arXiv preprint arXiv:2408.16965,

    2024 arXiv

  38. [53]

    Cupid: Adaptive curation of pre- training data for video-and-language representation learn- ing

    [Zhou et al., 2021] Luowei Zhou, Jingjing Liu, Yu Cheng, Zhe Gan, and Lei Zhang. Cupid: Adaptive curation of pre- training data for video-and-language representation learn- ing. arXiv preprint arXiv:2104.00285, 2021

    2021 arXiv

  39. [2018]

    Panambur, D

    [Panambur et al., 2022] T. Panambur, D. Chakraborty, M. Meyer, R. Milliken, E. Learned-Miller, and M. Par- ente. Self-supervised learning to guide scientifically relevant categorization of martian terrain images. In CVPR,

    2022

  40. [2019]

    arXiv preprint arXiv:1908.03557, 2,

    1908 arXiv

  41. [2020]

    Geography-aware self-supervised learn- ing

    [Ayush et al., 2021] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learn- ing. In ICCV,

    2021

  42. [2021]

    Vicreg: Variance-invariance-covariance regu- larization for self-supervised learning

    [Bardes et al., 2021] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regu- larization for self-supervised learning. arXiv preprint arXiv:2105.04906,

    2021 arXiv

  43. [2022]

    Self-supervised learning by cross-modal audio- video clustering

    [Alwassel et al., 2020] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio- video clustering. NeurIPS,

    2020

  44. [2023]

    Data curation via joint example selection further accelerates multimodal learning

    [Evans et al., 2024] Talfan Evans, Nikhil Parthasarathy, Hamza Merzic, and Olivier J Henaff. Data curation via joint example selection further accelerates multimodal learning. arXiv preprint arXiv:2406.17711,

    2024 arXiv

  45. [2024]

    Synthetic hard negative samples for contrastive learning

    [Dong et al., 2024] Hengkui Dong, Xianzhong Long, and Yun Li. Synthetic hard negative samples for contrastive learning. Neural Processing Letters, 56(1),

    2024

Pith tools

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