Pith. sign in

REVIEW 3 major objections 5 minor 36 references

The Impact of Semantic Pairs on Self-Supervised Representation Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Pretraining on curated semantic pairs—two different images from the same class—consistently beats pretraining on augmented views of the same image, improving transfer learning and object detection across contrastive and non-contrastive SSL

desk verdict A useful controlled comparison of semantic vs. augmented positive pairs in SSL, but the central matching claim is unverified as written because the dataset description never specifies unique-image counts and the 'statistical significance' has no statistics behind it. read the letter →

arxiv 2510.08722 v3 pith:5367LWDS submitted 2025-10-09 cs.LG cs.AI

Mohammad Alkhalefi , Georgios Leontidis , Mingjun Zhong This is my paper
classification cs.LGcs.AI
keywords self-supervisedlearningsemanticpositivepairsinstancediscriminationcontrastivedataaugmentationtransferobjectdetectionImageNetsubset
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 asks whether self-supervised positive pairs should be two augmented views of one image or two distinct images from the same semantic class. It builds a manually curated dataset of 29,359 semantic pairs from 187 ImageNet classes and a matched augmented-pair baseline with the same class composition and pair count. Across four SSL methods, transfer learning on STL-10/CIFAR, and PASCAL VOC detection, semantic-pair pretraining consistently outperforms augmented-pair pretraining. The improvement is largest and most consistent for contrastive learning, with SimCLR gaining 3.83 points on STL-10 at 200 epochs. The paper also finds that semantic pairs reduce reliance on handcrafted transformations, making models more robust when the transformation pipeline is incomplete.

What carries the argument

The semantic-pair dataset itself is the central object: 187 ImageNet classes, each with 157 manually curated pairs, totaling 29,359 positive pairs, where each pair contains two distinct images from the same class chosen to differ in contextual attributes. The experiment replaces the standard SSL positive pair (two augmented views of the same image) with this same-class different-image pair while keeping the training objective and augmentation pipeline otherwise unchanged. The paired-image construction is what forces the encoder to share information across genuinely different contexts, thereby learning invariance to nuisance factors that same-image augmentation cannot remove.

What would settle it

Train SimCLR on an augmented-pair dataset that uses 58,718 unique images (e.g., two augmented views of each distinct image) to match the likely unique-image count of the semantic-pair set, while keeping the 29,359-pair count, and evaluate transfer accuracy on STL-10. If the 3.8-point improvement disappears or shrinks substantially, the reported gain is attributable to data quantity rather than the semantic-pair mechanism. Conversely, if the gap persists, the semantic-pair claim is supported.

Watch

Extended reading notes

Core claim

The central claim is that semantic positive pairs—two different instances belonging to the same class, placed in different real-world contexts—provide a stronger and more generalizable supervisory signal than augmented views of a single image. Because the model must pull together representations of two distinct images, it is forced to retain only the object-related information shared across contexts and to discard nuisance correlations such as background, texture, illumination, and object-specific details. The paper validates this by controlled comparison: a semantic-pair dataset and an augmented-pair dataset matched in class composition and pair count, trained with the same SSL objectives,

Load-bearing premise

The datasets are assumed to be matched in scale, but the paper never states how many unique images the semantic-pair set contains—if each pair has two distinct images, the semantic model sees twice as many images as the augmented model, so the improvement could be due to data volume rather than semantic pairing.

Editorial extensions

If this is right

  • Semantic-pair pretraining consistently outperforms augmented-pair pretraining on transfer learning (STL-10, CIFAR-10, CIFAR-100) and object detection (PASCAL VOC), across contrastive (SimCLR) and non-contrastive (BYOL, DINO, VICReg) methods.
  • Among the tested methods, contrastive learning benefits most from semantic pairs, with SimCLR showing the largest relative improvement (e.g., +3.83% on STL-10 at 200 epochs).
  • Models trained on semantic pairs degrade far less when data transformations are removed (e.g., grayscale or cropping-only pipelines), indicating that semantic pairs induce invariances beyond the standard transformation suite.
  • A model pretrained on the 187-class semantic-pair dataset (≈29K images) outperforms a model pretrained on Tiny-ImageNet (200 classes, 100K images) on downstream transfer, while using roughly one-third of the training time.
  • Semantic-pair pretraining improves object detection across AP, AP@50, and AP@75, with the largest gain in AP@75, suggesting better object localization precision.
  • The benefit of semantic pairs holds with a different backbone (ViT-S/8) and task (fine-tuning), as shown by DINO (SP) outperforming DINO (AP).
  • Smaller semantic-pair sets (50 images per class) still outperform larger augmented-pair sets (157 images per class), indicating a data-efficiency advantage.

Reading between the lines

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

  • The paper's matched-comparison premise is unverified: the semantic-pair dataset has 29,359 pairs but likely 58,718 unique images (two per pair), while the augmented baseline has 29,359 unique images. If so, the reported improvements could partly reflect twice the data volume rather than the semantic-pair mechanism; a control experiment holding unique-image count fixed would settle this.
  • The manual curation cost (6 months of full-time work) means practical adoption of this idea would require automated semantic pairing. The clean, isolated gains demonstrated here may shrink when automatic pairing introduces false positives, so a study with noisy automatic pairs is a natural next test.
  • Semantic pairs derived from ImageNet classes implicitly use class-level supervision at pretraining time, placing the method between self-supervised and supervised learning. The improvement might be a form of label leakage; comparing against randomly paired same-cluster (pseudo-labeled) images would clarify the source of the gain.
  • The dataset and code release will let the community probe whether the advantages extend to larger backbones, more classes, and other downstream tasks such as segmentation or video, which the paper does not test.
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. The paper presents a curated semantic-pair dataset from ImageNet-1K (187 classes, 157 pairs per class) and compares self-supervised models pretrained on these semantic pairs against an augmented-pair baseline using the same class composition and reported pair count. Across transfer learning (STL-10, CIFAR-10/100) and PASCAL VOC object detection, semantic-pair pretraining is reported to consistently outperform augmented-pair pretraining, with SimCLR showing the largest relative gain (+3.83% on STL-10 at 200 epochs). Ablations remove transformations, vary the backbone, and vary dataset size. The authors conclude that semantic pairs induce invariances beyond standard augmentation and that contrastive methods benefit most from semantic-pair supervision.

Significance. If the experimental matching and class-selection concerns are resolved, the study would provide a useful controlled comparison of semantic positive pairs versus standard augmented pairs in self-supervised learning. Its strengths are the isolation of semantic pairs from hybrid training schemes, the use of multiple SSL methods, the inclusion of object detection and transfer tasks, and a manually curated, publicly promised dataset of 29,359 pairs. The dataset alone could be a community resource. At present, however, the central claim is weakened by an unresolved unique-image-count ambiguity, target-informed class selection, and the absence of repeated-seed statistics.

major comments (3)
  1. [§3.1.1, §3.1.2, Table 5] The matched-comparison premise is unverified because the number of unique images in the semantic-pair (SP) dataset is never stated. §3.1.1 gives 187 classes × 157 pairs = 29,359 semantic pairs; §3.1.2 gives 187 classes × 157 images = 29,359 augmented-pair (AP) images. If each SP pair consists of two distinct images, SP contains ≈58,718 unique images—twice the AP set. Table 5 labels both SP and AP columns as '157 Images,' which would imply 157 images/class for SP, but that conflicts with the '157 pairs/class' definition unless images are heavily reused. The paper only matches pair count, not unique-image count; the central +3.83% result (Table 1) could therefore be a data-quantity effect rather than a semantic-pair effect. Please specify unique-image counts for both datasets and either enforce equality or explicitly control for dataset size.
  2. [§3.1.1, §4] The claim that transfer is to 'unseen datasets' is weakened by the construction: the selected 187 ImageNet classes were sampled 'based on their semantic overlap with established benchmark datasets (e.g., STL-10, CIFAR),' and Appendix A says class selection was 'strategically aligned with standard benchmarks.' Those same benchmarks are then used for linear evaluation in §4.1. This target-informed selection can inflate transfer accuracy relative to a random class subset, so the observed SP–AP gaps are not necessarily evidence of a general semantic-pair benefit. Please evaluate on class-disjoint benchmarks or re-state the claim as transfer to overlapping benchmarks.
  3. [§4.1, Table 1, Figure 4] The text calls the accuracy gap 'statistically significant' and Figure 4 reports delta values, but no seed repetitions, standard deviations, confidence intervals, or significance tests are reported anywhere in the manuscript. With dataset construction and hyperparameters fixed, a single run cannot support a claim of statistical significance. Report multiple seeds (at least 3–5) for the key comparisons and provide error bars or an appropriate statistical test, especially for the small gaps (+0.5 to +4.3 percentage points) in Figure 4.
minor comments (5)
  1. [§4] The text says 'we trained a suite of eight SOTA approaches,' but only SimCLR, VICReg, BYOL, and DINO are described and plotted in Figure 4. Either the other four methods should be reported, or the number should be corrected.
  2. [Figure 4] The x-axis labels contain 'STL100,' which appears to be a typo for 'STL-10.' Also, the caption uses 'STL10' while the main text uses 'STL-10'; please standardize.
  3. [Table 2] The comparison with Tiny-ImageNet is not controlled: it differs in class composition (187 vs 200), number of samples (29.4K vs 100K), and pretraining time. The conclusion that 'the Semantic Pairs dataset requires much less pre-training time' is expected from the smaller size and should be framed as a resource/efficiency observation, not a matched comparison.
  4. [§4.4.1, Figure 5] The transformation-ablation labels 'Gray_S', 'Color Jitter', and 'Only_C' are not defined in the main text. The caption defines some, but the figure itself would benefit from clearer names and an explicit list of which transformations remain in each condition.
  5. [Table 5] The 'Improvement' row omits the plus sign for the 50-image and 100-image rows (4.20%, 4.04%), while the 157-image row has '+3.83%'. This is minor but should be made consistent.

Circularity Check

1 steps flagged · score 4.0 of 10

Target-informed class selection weakens the 'unseen data' claim, but the central SP-vs-AP comparison remains internally controlled.

  1. other [Section 3.1.1 (Dataset Construction) and Section 4.1 (Transfer Learning)]
    "The selected classes in the Semantic Pair dataset were sampled from ImageNet-1K based on their semantic overlap with established benchmark datasets (e.g., STL-10, CIFAR). This targeted sampling strategy enhances transferability to downstream tasks ... models pre-trained on semantic pairs outperform those using augmented pairs on unseen data."

    The pretraining classes are selected because they overlap with the exact benchmarks (STL-10, CIFAR-10/100) later used to measure 'generalization to unseen data.' Thus the transfer performance is partly an artifact of target-informed dataset construction rather than an independent test on truly unseen categories. However, both the SP and AP arms use the same 187 classes, so this leakage affects the absolute 'unseen data' framing more than the relative SP-vs-AP comparison; the paper provides no control with class selection independent of the evaluation benchmarks.

full rationale

The paper's central comparison—semantic pairs (SP) vs augmented pairs (AP) under matched class composition and training-pair count—is a real training/evaluation protocol, not a fitted parameter renamed as a prediction. The reported deltas (e.g., +3.83% for SimCLR on STL-10) come from experiments, not from an equation that reduces to its inputs. The self-citations (Alkhalefi et al. 2024a,b) are supporting references and are not load-bearing for the main claim. The main circularity concern is target-informed data construction: Section 3.1.1 states the 187 ImageNet classes were selected for semantic overlap with STL-10/CIFAR, and Section 4.1 then evaluates transfer on those same benchmarks and calls it 'unseen data.' This weakens the absolute generalization claim, but because both SP and AP use the same class selection, the relative benefit of semantic pairing is not an artifact of that leakage. A separate validity concern—not circularity per se—is that the number of unique images in the SP dataset is never stated: 157 pairs/class versus 157 images/class for AP. If each semantic pair uses two distinct images, the SP arm sees roughly twice as many unique images, which could confound the matched comparison. That ambiguity is a missing control rather than a derivation-equivalence. Score 4 reflects the partial, evaluation-level circularity while acknowledging the core SP-vs-AP comparison retains independent content.

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

No numeric free parameters are fitted in a theoretical derivation; the experimental hyperparameters (200 epochs, batch 256, 64x64 resolution) are fixed choices, not optimized to produce the headline gap. The main assumptions are domain assumptions about what semantic pairing does and about the validity of evaluation protocols.

assumptions (4)
  • domain assumption Two augmented views of the same image share nuisance correlations (background, texture, illumination) that limit generalization to unseen data.
    Introduction, Section 1; this premise motivates the study but is not established by the paper.
  • domain assumption The standard handcrafted transformation pipeline is insufficient to cover the full spectrum of real-world variations.
    Section 1 and ablation 4.4.1; assumed in the framing of representation learning.
  • domain assumption Manually curated same-class image pairs from ImageNet-1K class labels constitute 'ground-truth' semantic positive pairs.
    Section 3.1.1 and Appendix A.1; this uses human labels as supervision for pretext task construction.
  • domain assumption Linear evaluation on CIFAR/STL-10 and fine-tuning on PASCAL VOC measure representation generalization.
    Section 3.4; standard SSL evaluation protocol assumed valid.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Impact of Semantic Pairs on Self-Supervised Representation Learning." pith.science (2026). https://pith.science/paper/5367LWDS

@misc{pith2026251008722,
  author       = {Pith},
  title        = {Pith review of: The Impact of Semantic Pairs on Self-Supervised Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5367LWDS}},
  note         = {Machine review of arXiv:2510.08722}
}
read the original abstract

Instance discrimination learns visual representations by treating different augmented views of the same image as positive pairs. While this encourages invariance to handcrafted transformations, same-image positives can preserve nuisance correlations such as background, texture, illumination, and object-specific details. Semantic positive pairs, i.e., different same-class instances, may reduce these correlations by presenting objects across diverse contexts. However, previous studies often combine semantic pairs with augmented positives or false neighbors (i.e., incorrectly mapped semantic pairs), making it difficult to isolate the effect of semantic pairing. We present a controlled empirical study of semantic positive pairs for self-supervised representation learning. From ImageNet-1K, we construct two matched subsets: an augmented-pair baseline and a manually curated semantic-pair dataset with the same class composition and training-pair count. We use these datasets to compare representative contrastive and non-contrastive SSL methods under matched training conditions. Across transfer learning and object detection evaluations, semantic-pair pretraining consistently improves generalisation over augmented-pair pretraining. Additional ablations show that semantic pairs induce invariances beyond the standard transformation pipeline. Among the evaluated methods, contrastive learning benefits most strongly from semantic pairs, with SimCLR showing the largest relative improvement. These results clarify the role of semantic positive pairs in SSL and provide guidance for selecting and designing frameworks that can exploit semantic pair information effectively

Figures

Figures reproduced from arXiv: 2510.08722 by the authors.

Figure 1
Figure 1. This diagram illustrates how the instance discrimination approaches treat (a) aug￾mented pairs (i.e., two views for the same instance) and (b) semantic pairs (i.e., two instances belonging to the same category) during representation learning. resentations towards greater invariance to factors such as occlusion, background, pattern, and illumination, supporting the goal of learning general representations. Occlusion … view at source ↗
Figure 2
Figure 2. Examples from the semantic pairs dataset illustrate placing objects in various real-world context scenes. • Create a curated semantic pairs dataset that can be used in SSL approaches. • Compare different SOTA approaches and determine which approach is more capable of learning useful representations from semantic pairs. 2 Related Work SSL has emerged as a powerful paradigm for learning representations from unlabeled … view at source ↗
Figure 3
Figure 3. This diagram provides an overview of the four-stage framework for fair model comparison between models trained on augmented pairs (baseline) versus seman￾tic data pairs. In other words, by selecting a subset of ImageNet-1K classes that align with downstream task categories (e.g., selecting 10 ImageNet classes that also appear in STL-10), the model can achieve competitive transfer performance despite being pretrained… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Transfer learning performance of SOTA approaches pre-trained on Semantic Pairs (SP, striped) and Augmented Pairs (AP, solid), evaluated on CIFAR10, CI￾FAR100, and STL10.(SP) pre-training consistently improves downstream accu￾racy over (AP). (∆) values denote SP-AP diff…
Figure 5
Figure 5. Figure 5: Performance comparison of SimCLR (SP) and SimCLR (AP) pre-trained for 200 epochs with transformation ablations: (1) Gray s (grayscale removed), (2) color jitter (both color jitter and grayscale removed), and (3) only c (only random crop retained). representations that …
Figure 6
Figure 6. Figure 6: Examples from the proposed semantic pairs dataset [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: Additional examples from the proposed semantic pairs dataset [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 5 linked inside Pith

  1. [1]

    Alkhalefi, G

    M. Alkhalefi, G. Leontidis, and M. Zhong. Semantic positive pairs for enhancing visual representation learning of instance discrimination methods. Transactions on Machine Learning Research, 2024 a

    2024

  2. [2]

    Alkhalefi, G

    M. Alkhalefi, G. Leontidis, and M. Zhong. Leoclr: Leveraging original images for contrastive learning of visual representations. Transactions on Machine Learning Research, 2024 b

    2024

  3. [3]

    J. Auh, C. Cho, and S.-T. Kim. Contrastive learning for reducing false negatives with global and local views in augmented data. In 2023 Innovations in Intelligent Systems and Applications Conference (ASYU), pages 1--5, 2023. doi:10.1109/ASYU58738.2023.10296635

    arXiv 2023

  4. [4]

    Bardes, J

    A. Bardes, J. Ponce, and Y. LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021

    arXiv 2021

  5. [5]

    Caron, P

    M. Caron, P. Bojanowski, A. Joulin, and M. Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132--149, 2018

    2018

  6. [6]

    Caron, I

    M. Caron, I. Misra, J. Mairal, P. Goyal, P. Bojanowski, and A. Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33: 0 9912--9924, 2020

    2020

  7. [7]

    Caron, H

    M. Caron, H. Touvron, I. Misra, H. J \'e gou, J. Mairal, P. Bojanowski, and A. Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650--9660, 2021

    2021

  8. [8]

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

    2020

Show all 36 references
  1. [9]

    Chen and K

    X. Chen and K. He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750--15758, 2021

    2021

  2. [10]

    X. Chen, H. Fan, R. Girshick, and K. He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020 b

    2003 arXiv

  3. [11]

    Chuang, J

    C.-Y. Chuang, J. Robinson, Y.-C. Lin, A. Torralba, and S. Jegelka. Debiased contrastive learning. Advances in neural information processing systems, 33: 0 8765--8775, 2020

    2020

  4. [12]

    Coates, A

    A. Coates, A. Ng, and H. Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215--223. JMLR Workshop and Conference Proceedings, 2011

    2011

  5. [13]

    Dwibedi, Y

    D. Dwibedi, Y. Aytar, J. Tompson, P. Sermanet, and A. Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9588--9597, 2021

    2021

  6. [14]

    Everingham, L

    M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88: 0 303--338, 2010

    2010

  7. [15]

    Gidaris, A

    S. Gidaris, A. Bursuc, N. Komodakis, P. P \'e rez, and M. Cord. Learning representations by predicting bags of visual words. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6928--6938, 2020

    2020

  8. [16]

    Gidaris, A

    S. Gidaris, A. Bursuc, G. Puy, N. Komodakis, M. Cord, and P. Perez. Obow: Online bag-of-visual-words generation for self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830--6840, 2021

    2021

  9. [17]

    Grill, F

    J.-B. Grill, F. Strub, F. Altch \'e , C. Tallec, P. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Guo, M. Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33: 0 21271--2...

    2020

  10. [18]

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

    2020

  11. [19]

    Huynh, S

    T. Huynh, S. Kornblith, M. R. Walter, M. Maire, and M. Khademi. Boosting contrastive self-supervised learning with false negative cancellation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2785--2795, 2022

    2022

  12. [20]

    Kalantidis, M

    Y. Kalantidis, M. B. Sariyildiz, N. Pion, P. Weinzaepfel, and D. Larlus. Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems, 33: 0 21798--21809, 2020

    2020

  13. [21]

    Krizhevsky, G

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

    2009

  14. [22]

    Y. Le, X. Yang, et al. Tiny imagenet challenge. 2015

    2015

  15. [23]

    Meehan, F

    C. Meehan, F. Bordes, P. Vincent, K. Chaudhuri, and C. Guo. Do ssl models have d \'e j \`a vu? a case of unintended memorization in self-supervised learning. Advances in Neural Information Processing Systems, 36: 0 42775--42798, 2023

    2023

  16. [24]

    Misra and L

    I. Misra and L. v. d. Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6707--6717, 2020

    2020

  17. [25]

    Purushwalkam and A

    S. Purushwalkam and A. Gupta. Demystifying contrastive self-supervised learning: Invariances, augmentations and dataset biases. Advances in Neural Information Processing Systems, 33: 0 3407--3418, 2020

    2020

  18. [26]

    Robinson, C.-Y

    J. Robinson, C.-Y. Chuang, S. Sra, and S. Jegelka. Contrastive learning with hard negative samples. arXiv preprint arXiv:2010.04592, 2020

    2010 arXiv

  19. [27]

    Russakovsky, J

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115: 0 211--252, 2015

    2015

  20. [28]

    Y. Tian, C. Sun, B. Poole, D. Krishnan, C. Schmid, and P. Isola. What makes for good views for contrastive learning? Advances in Neural Information Processing Systems, 33: 0 6827--6839, 2020

    2020

  21. [29]

    Wang and G.-J

    X. Wang and G.-J. Qi. Contrastive learning with stronger augmentations. IEEE transactions on pattern analysis and machine intelligence, 45 0 (5): 0 5549--5560, 2022

    2022

  22. [30]

    M. Wu, M. Mosse, C. Zhuang, D. Yamins, and N. Goodman. Conditional negative sampling for contrastive learning of visual representations. arXiv preprint arXiv:2010.02037, 2020

    2010 arXiv

  23. [31]

    J. Xiao, W. Guo, J. Liu, and M. Li. Generalization gap in data augmentation: Insights from illumination. In International Conference on Pattern Recognition, pages 190--205. Springer, 2024

    2024

  24. [32]

    T. Xiao, X. Wang, A. A. Efros, and T. Darrell. What should not be contrastive in contrastive learning. arXiv preprint arXiv:2008.05659, 2020

    2008 arXiv

  25. [33]

    Yosinski, J

    J. Yosinski, J. Clune, Y. Bengio, and H. Lipson. How transferable are features in deep neural networks? Advances in neural information processing systems, 27, 2014

    2014

  26. [34]

    Zbontar, L

    J. Zbontar, L. Jing, I. Misra, Y. LeCun, and S. Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310--12320. PMLR, 2021

    2021

  27. [35]

    D. Zeng, Y. Wu, X. Hu, X. Xu, and Y. Shi. Contrastive learning with synthetic positives. In European Conference on Computer Vision, pages 430--447. Springer, 2024

    2024

  28. [36]

    Zhang, C

    T. Zhang, C. Qiu, W. Ke, S. S \"u sstrunk, and M. Salzmann. Leverage your local and global representations: A new self-supervised learning strategy. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16580--16589, 2022

    2022

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.