Pith. sign in

REVIEW 5 major objections 4 minor 87 references

The best SSL training paradigm depends on the method, domain, and label count

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 05:54 UTC pith:2CNCMRME

load-bearing objection Broad, useful benchmark, but the PFT–JT comparison is confounded: PFT freezes the backbone for classification/IQA while JT trains it end-to-end, so the headline claims overstate the joint-objective effect. the 5 major comments →

arxiv 2607.13192 v1 pith:2CNCMRME submitted 2026-07-14 cs.CV

Self-Supervised Visual Representation Learning: Pretrain-Finetuning or Joint Training?

Nusrat Munia , Tyler Ward , Nishat Nayla , Matthew A. Massey , Abdullah-Al-Zubaer Imran This is my paper
classification cs.CV
keywords self-supervised learningjoint trainingpretrain-finetunesemi-supervised learningrepresentation learningtransfer learningcomputer visionempirical benchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper asks whether self-supervised visual learning should follow the standard two-stage pipeline (pretrain on unlabeled data, then finetune on labeled data) or instead train both objectives together in one stage. Across eight self-supervised methods, eleven datasets, and four vision tasks, it finds that neither paradigm is universally better: the winner depends on the SSL objective, the domain, and how much labeled data is available. Joint training is consistently more data- and compute-efficient and remains robust when labels are scarce, while pretrain-finetune is more reliable in specialized domains such as medical imaging and for contrastive methods. The paper provides a broad empirical benchmark and practical guidance for choosing between the two strategies.

Core claim

The paper establishes that the relative effectiveness of pretrain-finetune (PFT) and joint training (JT) is governed by the SSL objective, the downstream task, the domain, and the labeled-data fraction. Reconstruction-oriented and auxiliary-task methods (Colorization, Rotation, MAE) generally benefit from JT, sometimes with large gains — for example, Colorization on CIFAR-10 rises from 0.5090 accuracy under PFT to 0.9099 under JT at full labels, and MAE from 0.5920 to 0.8341. Contrastive and redundancy-reduction methods (SimCLR, MoCo, DINO, Barlow Twins) often transfer better under PFT, especially in specialized domains like crisis response, medical imaging, and image quality assessment. JT

What carries the argument

The central comparison is between two training protocols instantiated across eight SSL methods: PFT, which pretrains a backbone (ResNet-18, or ViT for MAE) with a self-supervised loss and then trains only a downstream head on labeled data; and JT, which optimizes L_total = L_SSL + L_sup simultaneously on the shared encoder. The benchmark spans classification (CIFAR-10, CrisisMMD, ISIC, JSRT), detection (COCO, PASCAL VOC), segmentation (ISIC, JSRT, EarthScape), and image quality assessment (KADID-10k, KonIQ-10k, LDCTIQA) at 10%, 20%, 50%, and 100% labeled fractions. The fixed equal weighting of the two losses in JT is a key design choice, as the paper does not tune a loss weight.

Load-bearing premise

The paper assumes that adding the self-supervised and supervised losses with equal, fixed weights is a fair and representative implementation of joint training, and that single-seed runs are sufficient to establish consistent trends.

What would settle it

Re-run a representative subset (e.g., CIFAR-10 with Colorization and SimCLR) over a grid of loss weights (λ from 0.1 to 10) and check whether the JT-vs-PFT ranking flips for any λ; if it does, the equal-weighting choice is load-bearing. Also repeat the main comparisons with multiple random seeds: if the JT advantages (e.g., Colorization 0.9099 vs 0.5090) are not consistent across seeds, the 'consistent trends' claim is weakened.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners with scarce labels can adopt joint training for reconstruction-based SSL methods to gain both accuracy and a large reduction in training time.
  • In specialized domains (medical, crisis, remote sensing) and for contrastive SSL methods, the conventional two-stage pretrain-finetune pipeline remains a safer default.
  • JT's computational savings of up to 7–8× make it attractive for rapid iteration and resource-constrained deployment when task performance is competitive.
  • The benchmark offers a reference map of which SSL objectives are compatible with joint optimization, helping practitioners avoid known failure cases such as Rotation-on-JSRT segmentation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The equal weighting of L_SSL and L_sup is a strong simplifying choice; tuning the loss weight or scheduling it over training could change which methods prefer JT, so the practical guidance may need re-qualification.
  • The paper's single-seed runs leave open whether the 'consistent' trends are statistically robust; multi-seed re-runs would strengthen or qualify the claims.
  • A testable hypothesis suggested by the results: SSL objectives that preserve spatial structure (colorization, masked reconstruction) align with supervised gradients and thus benefit from JT, while objectives that enforce invariance (contrastive) compete with them; measuring gradient alignment could predict JT benefit.
  • The CLIP extension on CIFAR-10 shows JT underperforming PFT at 10% labels, hinting that the interaction may differ for contrastive-language objectives; this deserves exploration on larger multimodal datasets.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper compares two paradigms for combining self-supervised and supervised learning: pretrain-then-finetune (PFT) and joint training (JT). Across eight SSL methods (Colorization, Rotation, SimCLR, BYOL, MoCo, DINO, MAE, Barlow Twins) and eleven datasets spanning natural, medical, crisis, remote-sensing, detection, and IQA tasks, the authors report performance under 10–100% labeled data, plus analyses of robustness, efficiency, interpretability, and cross-domain transfer. The stated conclusion is that JT is often more data- and training-efficient and robust in low-label settings, while PFT tends to be more reliable in specialized domains and for contrastive methods. The paper is primarily an empirical benchmark and presents a large set of tables and figures in support of this conclusion.

Significance. If the comparison were clean, this would be a useful benchmark for practitioners choosing between two-stage SSL and joint training. The breadth of methods, domains, and tasks is a real strength, as is the consistent reporting of training time alongside accuracy. However, the central comparison is confounded by an asymmetric evaluation protocol: for classification, IQA, and most segmentation experiments, PFT uses a frozen backbone and trains only a head, while JT trains the entire encoder. In addition, nearly every quantitative claim rests on a single fixed-random-seed run with no error bars. The inconsistencies between the abstract's sweeping claims and the appendix's own results (e.g., Barlow Twins collapsing under JT on crisis data, DINO degrading sharply on EarthScape) further weaken the conclusions as stated. The paper's breadth makes the topic significant, but the present evidence does not cleanly support the headline claims.

major comments (5)
  1. [Sec. 4.1, Fig. 1, Table 2] The PFT vs JT comparison is confounded by unequal trainable capacity. In Sec. 4.1 the authors state that for classification and IQA 'the encoder is frozen and only a linear layer is trained' in PFT, and Fig. 1 says 'During fine-tuning, the pretrained backbone is kept frozen.' In JT, by contrast, 'the encoder is shared and optimized jointly through both supervised and self-supervised losses.' Thus any JT advantage could stem from adapting the backbone to the task, not from the joint SSL+supervised objective. A concrete example is Colorization on CIFAR-10 (Table 2): PFT reaches only 0.5090 accuracy while JT reaches 0.9099, yet a conventionally fine-tuned ResNet-18 with only supervised loss typically exceeds 0.90 on CIFAR-10. Without a PFT control that also fine-tunes the backbone (supervised loss only), the paper's central comparison is not cleanly interpretable. This is the load-bearing p
  2. [Abstract, Sec. 4.2, Tables 4 and 13] The abstract's claim that 'JT consistently improves data and training efficiency while being robust in low-label settings' is contradicted by the paper's own appendix results. On CrisisMMD (Table 4), Barlow Twins JT drops to 0.6258 accuracy / 0.6130 F1 at 100% labels versus 0.7169 / 0.6885 under PFT, and cross-domain DMD F1 falls from 0.6845 to 0.4661. On EarthScape (Table 13), DINO JT at 100% labels drops from 0.6230/0.8487 to 0.5126/0.5552 accuracy/F1, and cross-domain F1 drops from 0.8038 to 0.4964. The body of Sec. 4.2 does acknowledge these instabilities, but the abstract and conclusion overstate consistency. The wording should be aligned with the actual, heavily conditional findings.
  3. [Tables 2–14, Sec. H.1] Every quantitative result appears to be a single point estimate with a fixed random seed ('a fixed random seed of 1' in Sec. H.1), with no error bars, multiple seeds, or significance tests. Many reported differences are small — for example, Table 7 shows COCO detection mAP differences as tiny as 0.001–0.003 between PFT and JT, and Table 8 shows almost identical cross-domain VOC mAP values. Without repeated runs or variance estimates, claims of 'consistent' improvement or degradation cannot be supported, especially for the training-time and 'comparable performance' statements. At minimum, the authors should report mean±std over at least three seeds for a subset of key experiments, or explicitly caveat all qualitative conclusions.
  4. [Sec. 4.1, Sec. H.1] The PFT protocol is not defined consistently across tasks. For classification and IQA, PFT is linear probing with a frozen backbone. For object detection, the authors write that they 'fine-tune YOLOv12' after injecting pretrained weights — i.e., the backbone is trained. For segmentation, the encoder is frozen and only the U-Net decoder is optimized (Sec. H.1: 'the SSL encoder is frozen, and only the U-Net decoder is optimized'). This means the comparison across tasks simultaneously changes the definition of PFT, making the headline conclusion 'relative effectiveness depends on task' partly an artifact of protocol inconsistency. The authors should adopt a uniform PFT definition (e.g., supervised fine-tuning of the full model, or both frozen- and fine-tuned variants) across all task families.
  5. [Sec. 3, Eq. (1)–(2)] The JT objective is defined as L_total = L_SSL + L_sup with equal weighting, and no loss-weight tuning or sensitivity analysis is reported. The relative benefit of JT is likely sensitive to this weighting; a fixed equal weight may favor some SSL methods (e.g., reconstruction losses with small magnitude) and disfavor others. The paper's practical guidance about which methods benefit from JT is therefore conditional on an arbitrary weighting choice. The authors should either justify the equal-weight choice with a sensitivity analysis or explicitly restrict the claims to this protocol.
minor comments (4)
  1. [Table 8] The MAE PFT 100% row in Table 8 reports the same COCO metrics as the 10% row (0.494 P, 0.375 R, 0.376 mAP, 0.252 mAP50–95), which appears to be a copy/paste error; Table 7 lists MAE PFT 100% as 0.689 P / 0.560 R / 0.608 mAP / 0.445 mAP50–95. Please correct.
  2. [Fig. 2] The color legend says 'warmer colors indicate lower performance, cooler colors indicate higher performance,' which is the reverse of common intuition and of many other heatmaps in the literature. Consider inverting the colormap or relabeling the legend for clarity.
  3. [Appendix F.1, F.3] Figures 17 and 18 appear to be duplicates with different captions, both showing SROCC/PLCC under clean and adversarial settings for MoCo and DINO. Please remove or differentiate them.
  4. [Sec. H.1, Appendix I] The description of hyperparameters is spread over the appendix and is not complete enough for exact reproduction; for example, the loss weighting, augmentation strengths, and early-stopping criteria for EarthScape differ from other datasets and are described only qualitatively. A code release would substantially strengthen the reproducibility of this benchmark.

Circularity Check

0 steps flagged

No significant circularity: the PFT-vs-JT comparison is an empirical benchmark, and the self-citations are not load-bearing.

full rationale

The paper is an empirical comparison rather than a derivation chain. The only formal object, L_total = L_SSL + L_sup (Sec. 3), is a definition of the JT protocol, not a fitted model whose output is later renamed a prediction. The reported accuracy, F1, Dice, mIoU, SROCC, and wall-clock numbers are measured outcomes on external datasets; no parameter is tuned to a subset of the target result and then presented as a forecast. The self-citations [36] and [49] are used as prior-work background and as one benchmark dataset respectively; neither supplies the PFT-vs-JT comparisons, which are run in this paper against CIFAR-10, COCO, ISIC, JSRT, CrisisMMD, KADID, and other datasets. Appendix I explicitly states 'our work is primarily empirical in nature' and that a deeper theoretical understanding remains open, confirming that no first-principles derivation is claimed. The protocol asymmetry noted by a skeptical reader—PFT freezes the backbone while JT trains the full encoder (Sec. 4.1: 'the encoder is frozen and only a linear layer is trained')—is a possible confound for the comparison's validity, but it is not a circular step: the observed accuracy differences are not constructed to equal the input losses. Similarly, JT's wall-clock speed advantage is a measured consequence of using a single training phase, not an algebraic consequence of the loss equation. No step in the paper reduces to its own inputs, so the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

This is an empirical benchmark paper, so the ledger contains no fitted free parameters in the sense of a derivation. The relevant choices are experimental design points, per-dataset hyperparameters, the equal weighting of the joint loss, and assumptions about representativeness and single-seed reliability. No new theoretical entities are postulated.

free parameters (4)
  • Fixed random seed = 1
    All experiments use a single seed (Appendix H.1), so no variance or sensitivity information is available.
  • Label fractions = 10%, 20%, 50%, 100%
    These are chosen experimental design points, not fitted, but the central trends are defined relative to these fractions.
  • Per-dataset hyperparameters = various (learning rates, batch sizes, epochs)
    Each dataset uses hand-selected hyperparameters; no sensitivity or ablation of these choices is reported.
  • JT loss weighting = 1.0 for both losses
    The joint objective is L_SSL + L_sup with equal weights and no reported tuning; conclusions about JT assume this weighting.
axioms (3)
  • domain assumption Frozen-backbone linear probing is a valid measure of representation quality for PFT.
    The paper uses linear evaluation on a frozen encoder (Section 4.1), a standard SSL protocol, but it may not capture the full finetuning picture.
  • domain assumption ResNet-18 and the selected datasets are representative of practical SSL use.
    The limitations section acknowledges lightweight backbones and a finite set of architectures; the practical guidance extrapolates beyond these.
  • ad hoc to paper Single-run results without error bars are sufficient to establish qualitative trends.
    Appendix H.1 fixes the random seed to 1, and all tables report single values; this assumption is load-bearing for claims of consistency.

pith-pipeline@v1.3.0-alltime-deepseek · 38299 in / 7565 out tokens · 78257 ms · 2026-08-02T05:54:07.743248+00:00 · methodology

0 comments
read the original abstract

Self-supervision is a powerful technique for learning visual representations from unlabeled data. Existing techniques primarily adopt a two-stage approach for self-supervised learning (SSL): a pretraining stage on unlabeled data followed by a finetuning stage on labeled data. While this pipeline has demonstrated extreme effectiveness, the interaction between self-supervised and supervised learning objectives remains insufficiently understood. In this work, we systematically investigate whether jointly optimizing the self-supervised and supervised objectives during training provides a better alternative. We compare two training paradigms: (1) the aforementioned pretraining followed by finetuning (PFT) and (2) joint training (JT), where self-supervised and supervised losses are optimized simultaneously in the same network. Across eight representative SSL methods and diverse computer vision tasks on natural, medical, crisis response, and remote sensing data, we evaluate performance under varying percentages of labeled data. Our results reveal that the relative effectiveness of PFT and JT depends strongly on the task at hand, the availability of labeled data, and the complexity of the domain. We find that JT consistently improves data and training efficiency while being robust in low-label settings, while PFT is more reliable in more specialized domains. We further analyze representation quality, robustness, and cross-domain generalization, providing new insights into how self-supervised and supervised objectives interact during optimization. We establish a comprehensive empirical benchmark for hybrid SSL-based semi-supervised learning and offer practical guidance for selecting appropriate training strategies across diverse vision applications.

Figures

Figures reproduced from arXiv: 2607.13192 by Abdullah-Al-Zubaer Imran, Matthew A. Massey, Nishat Nayla, Nusrat Munia, Tyler Ward.

Figure 1
Figure 1. Figure 1: Illustration of the two competing training paradigms in self-supervised learning. The [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall performance comparison of self-supervised learning methods across multiple [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: CIFAR-10 classification accuracy under PFT and JT settings. Left: standard training and [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Crisis analysis performance. (a) Standard CrisisMMD [ [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: IQA performance (SROCC) for the eight SSL techniques with PFT and JT evaluated [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Change in downstream performance under joint training ( [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: t-SNE embeddings on CIFAR-10 for various SSL methods (PFT vs JT). Embedding color [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Radar plots depicting image classification accuracy of the eight SSL techniques on the [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Grad-CAM visualizations for different SSL methods under PFT and JT settings on CIFAR [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Radar plots depicting PFT and JT performance across each of the explored SSL techniques [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Radar plots depicting the accuracy of skin lesion classification on the ISIC dataset using [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Radar plots depicting PFT and JT performance across each of the explored SSL techniques [PITH_FULL_IMAGE:figures/full_fig_p031_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Predicted segmentation masks vs. the ground truth masks on the JSRT dataset across each [PITH_FULL_IMAGE:figures/full_fig_p032_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Radar plots depicting the Dice of skin lesion segmentation on the ISIC dataset using both [PITH_FULL_IMAGE:figures/full_fig_p034_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Comparison of SROCC performance of self-supervised learning (SSL) methods across [PITH_FULL_IMAGE:figures/full_fig_p036_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: KonIQ SROCC comparison of SSL methods across different label regimes under PFT and [PITH_FULL_IMAGE:figures/full_fig_p037_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Comparison of SROCC and PLCC performance under clean and adversarial settings for [PITH_FULL_IMAGE:figures/full_fig_p038_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Comparison of SROCC and PLCC performance under clean and adversarial conditions for [PITH_FULL_IMAGE:figures/full_fig_p038_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Cross-Dataset PLCC Performance Analysis of SSL Methods for Image Quality Assessment [PITH_FULL_IMAGE:figures/full_fig_p039_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Cross-Dataset Image Quality Assessment PLCC Comparison of SSL Methods Under 10% [PITH_FULL_IMAGE:figures/full_fig_p039_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Overall SSL Benchmark Landscape Across IQA Datasets [PITH_FULL_IMAGE:figures/full_fig_p039_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Change in geographic domain-shift retention under joint training ( [PITH_FULL_IMAGE:figures/full_fig_p040_22.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

87 extracted references · 12 linked inside Pith

  1. [1]

    GPT-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  2. [2]

    JGCL: Joint self-supervised and supervised graph contrastive learning

    Selahattin Akkas and Ariful Azad. JGCL: Joint self-supervised and supervised graph contrastive learning. InCompanion Proceedings of the Web Conference, 2022

    2022

  3. [3]

    CrisisMMD: Multimodal twitter datasets from natural disasters

    Firoj Alam, Ferda Ofli, and Muhammad Imran. CrisisMMD: Multimodal twitter datasets from natural disasters. InProceedings of the international AAAI conference on web and social media, volume 12, 2018

    2018

  4. [4]

    Flamingo: A visual language model for few-shot learning.Advances in neural information processing systems, 35: 23716–23736, 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.Advances in neural information processing systems, 35: 23716–23736, 2022

    2022

  5. [5]

    Hyperspectral target detection using self-supervised background learning.Advances in Space Research, 74(2):628–646, 2024

    Muhammad Khizer Ali, Benish Amin, Abdur Rahman Maud, Farrukh Aziz Bhatti, Komal Nain Sukhia, and Khurram Khurshid. Hyperspectral target detection using self-supervised background learning.Advances in Space Research, 74(2):628–646, 2024

    2024

  6. [6]

    Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging.Nature Biomedical Engineering, 7(6):756–779, 2023

    Shekoofeh Azizi, Laura Culp, Jan Freyberg, Basil Mustafa, Sebastien Baur, Simon Kornblith, Ting Chen, Nenad Tomasev, Jovana Mitrovi´c, Patricia Strachan, et al. Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging.Nature Biomedical Engineering, 7(6):756–779, 2023

    2023

  7. [7]

    BEiT: BERT pre-training of image transformers.arXiv preprint arXiv:2106.08254, 2021

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEiT: BERT pre-training of image transformers.arXiv preprint arXiv:2106.08254, 2021

    Pith/arXiv arXiv 2021

  8. [8]

    VICReg: Variance-invariance-covariance regular- ization for self-supervised learning.arXiv preprint arXiv:2105.04906, 2021

    Adrien Bardes, Jean Ponce, and Yann LeCun. VICReg: Variance-invariance-covariance regular- ization for self-supervised learning.arXiv preprint arXiv:2105.04906, 2021

    Pith/arXiv arXiv 2021

  9. [9]

    Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring.arXiv preprint arXiv:1911.09785, 2019

    David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring.arXiv preprint arXiv:1911.09785, 2019

    Pith/arXiv arXiv 1911

  10. [10]

    Mixmatch: A holistic approach to semi-supervised learning.Advances in neural information processing systems, 32, 2019

    David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning.Advances in neural information processing systems, 32, 2019

    2019

  11. [11]

    Unsupervised learning of visual features by contrasting cluster assignments.Advances in neural information processing systems, 33:9912–9924, 2020

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments.Advances in neural information processing systems, 33:9912–9924, 2020

    2020

  12. [12]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 9650–9660, 2021

    2021

  13. [13]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. InInternational conference on machine learning, pages 1597–1607. PmLR, 2020

    2020

  14. [14]

    Big self-supervised models are strong semi-supervised learners.Advances in neural information processing systems, 33:22243–22255, 2020

    Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners.Advances in neural information processing systems, 33:22243–22255, 2020. 10

    2020

  15. [15]

    Exploring simple siamese representation learning

    Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750–15758, 2021

    2021

  16. [16]

    Not-so-supervised: a survey of semi-supervised, multi-instance, and transfer learning in medical image analysis.Medical image analysis, 54:280–296, 2019

    Veronika Cheplygina, Marleen De Bruijne, and Josien PW Pluim. Not-so-supervised: a survey of semi-supervised, multi-instance, and transfer learning in medical image analysis.Medical image analysis, 54:280–296, 2019

    2019

  17. [17]

    Orienting novel 3D objects using self-supervised learning of rotation transforms

    Shivin Devgon, Jeffrey Ichnowski, Ashwin Balakrishna, Harry Zhang, and Ken Goldberg. Orienting novel 3D objects using self-supervised learning of rotation transforms. In2020 IEEE 16th International Conference on Automation Science and Engineering (CASE), pages 1453–1460. IEEE, 2020

    2020

  18. [18]

    BERT: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short papers), pages 4171–4186, 2019

    2019

  19. [19]

    Unsupervised visual representation learning by context prediction

    Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. InProceedings of the IEEE international conference on computer vision, pages 1422–1430, 2015

    2015

  20. [20]

    Self- supervised colorization towards monochrome-color camera systems using cycle cnn.IEEE Transactions on Image Processing, 30:6609–6622, 2021

    Xuan Dong, Chang Liu, Weixin Li, Xiaoyan Hu, Xiaojie Wang, and Yunhong Wang. Self- supervised colorization towards monochrome-color camera systems using cycle cnn.IEEE Transactions on Image Processing, 30:6609–6622, 2021

    2021

  21. [21]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  22. [22]

    Everingham, L

    M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascal- network.org/challenges/VOC/voc2012/workshop/index.html

    2012

  23. [23]

    The Pascal Visual Object Classes (VOC) Challenge.International Journal of Computer Vision, 88(2):303–338, 2010

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge.International Journal of Computer Vision, 88(2):303–338, 2010

    2010

  24. [24]

    Self-supervised representation learning by rotation feature decoupling

    Zeyu Feng, Chang Xu, and Dacheng Tao. Self-supervised representation learning by rotation feature decoupling. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10364–10374, 2019

    2019

  25. [25]

    Bootstrap your own latent-a new approach to self-supervised learning.Advances in neural information processing systems, 33:21271–21284, 2020

    Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning.Advances in neural information processing systems, 33:21271–21284, 2020

    2020

  26. [26]

    A survey on self-supervised learning: Algorithms, applications, and future trends.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):9052–9071, 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.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):9052–9071, 2024

    2024

  27. [27]

    CrisisKAN: Knowledge- infused and explainable multimodal attention network for crisis event classification

    Shubham Gupta, Nandini Saini, Suman Kundu, and Debasis Das. CrisisKAN: Knowledge- infused and explainable multimodal attention network for crisis event classification. InEuropean Conference on Information Retrieval, pages 18–33. Springer, 2024

    2024

  28. [28]

    David Gutman, Noel CF Codella, Emre Celebi, Brian Helba, Michael Marchetti, Nabin Mishra, and Allan Halpern. Skin lesion analysis toward melanoma detection: A challenge at the international symposium on biomedical imaging (isbi) 2016, hosted by the international skin imaging collaboration (isic).arXiv preprint arXiv:1605.01397, 2016

    Pith/arXiv arXiv 2016

  29. [29]

    Deep residual learning for im- age recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for im- age recognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016. 11

    2016

  30. [30]

    Momentum contrast for unsupervised visual representation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729–9738, 2020

    2020

  31. [31]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000–16009, 2022

    2022

  32. [32]

    Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment.IEEE Transactions on Image Processing, 29:4041–4056, 2020

    Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment.IEEE Transactions on Image Processing, 29:4041–4056, 2020

    2020

  33. [33]

    Self-supervised learning by image colorization

    Jie Hou. Self-supervised learning by image colorization. Master’s thesis, Universitat Politècnica de Catalunya, 2021

    2021

  34. [34]

    Adco: Adversarial contrast for efficient learning of unsupervised representations from self-trained negative adversaries

    Qianjiang Hu, Xiao Wang, Wei Hu, and Guo-Jun Qi. Adco: Adversarial contrast for efficient learning of unsupervised representations from self-trained negative adversaries. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1074–1083, 2021

    2021

  35. [35]

    Self-supervised color- concept association via image colorization.IEEE Transactions on Visualization and Computer Graphics, 29(1):247–256, 2022

    Ruizhen Hu, Ziqi Ye, Bin Chen, Oliver van Kaick, and Hui Huang. Self-supervised color- concept association via image colorization.IEEE Transactions on Visualization and Computer Graphics, 29(1):247–256, 2022

    2022

  36. [36]

    Self-supervised, semi-supervised, multi-context learning for the combined classification and segmentation of medical images

    Abdullah-Al-Zubaer Imran, Chao Huang, Hui Tang, Wei Fan, Yuan Xiao, Dingjun Hao, Zhen Qian, and Demetri Terzopoulos. Self-supervised, semi-supervised, multi-context learning for the combined classification and segmentation of medical images. InProceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13815–13816, 2020

    2020

  37. [37]

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

    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 supervision. InInternational conference on machine learning, pages 4904–4916. PMLR, 2021

    2021

  38. [38]

    Self-supervised visual feature learning with deep neural networks: A survey.IEEE transactions on pattern analysis and machine intelligence, 43(11): 4037–4058, 2020

    Longlong Jing and Yingli Tian. Self-supervised visual feature learning with deep neural networks: A survey.IEEE transactions on pattern analysis and machine intelligence, 43(11): 4037–4058, 2020

    2020

  39. [39]

    Self-supervised spatiotemporal feature learning via video rotation prediction.arXiv preprint arXiv:1811.11387, 2018

    Longlong Jing, Xiaodong Yang, Jingen Liu, and Yingli Tian. Self-supervised spatiotemporal feature learning via video rotation prediction.arXiv preprint arXiv:1811.11387, 2018

    Pith/arXiv arXiv 2018

  40. [40]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

    2009

  41. [41]

    Temporal ensembling for semi-supervised learning.arXiv preprint arXiv:1610.02242, 2016

    Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning.arXiv preprint arXiv:1610.02242, 2016

    Pith/arXiv arXiv 2016

  42. [42]

    Hsieh, and Jang-Hwan Choi

    Wonkyeong Lee, Fabian Wagner, Andreas Maier, Adam Wang, Jongduk Baek, Scott S. Hsieh, and Jang-Hwan Choi. Low-dose computed tomography perceptual image quality assessment grand challenge dataset (miccai 2023), 2023. URL https://doi.org/10.5281/zenodo. 7833096

    doi:10.5281/zenodo 2023

  43. [43]

    BLIP: Bootstrapping language- image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language- image pre-training for unified vision-language understanding and generation. InInternational conference on machine learning, pages 12888–12900. PMLR, 2022

    2022

  44. [44]

    BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InInternational conference on machine learning, pages 19730–19742. PMLR, 2023

    2023

  45. [45]

    KADID-10k: A large-scale artificially distorted IQA database

    Hanhe Lin, Vlad Hosu, and Dietmar Saupe. KADID-10k: A large-scale artificially distorted IQA database. In2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1–3. IEEE, 2019. 12

    2019

  46. [46]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InEuropean conference on computer vision, pages 740–755. Springer, 2014

    2014

  47. [47]

    Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    2023

  48. [48]

    Self-supervised learning: Generative or contrastive.IEEE transactions on knowledge and data engineering, 35(1):857–876, 2021

    Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Generative or contrastive.IEEE transactions on knowledge and data engineering, 35(1):857–876, 2021

    2021

  49. [49]

    Earthscape: A multi- modal dataset for surficial geologic mapping and earth surface analysis.arXiv preprint arXiv:2503.15625, 2025

    Matthew Massey, Nusrat Munia, and Abdullah-Al-Zubaer Imran. Earthscape: A multi- modal dataset for surficial geologic mapping and earth surface analysis.arXiv preprint arXiv:2503.15625, 2025

    arXiv 2025

  50. [50]

    Damage identification in social media posts using multimodal deep learning

    Hussein Mouzannar, Yara Rizk, and Mariette Awad. Damage identification in social media posts using multimodal deep learning. InISCRAM. Rochester, NY , USA, 2018

    2018

  51. [51]

    Self-supervision for medical image classification: State-of-the-art performance with ˜100 labeled training samples per class.Bioengineering, 10(8):895, 2023

    Maximilian Nielsen, Laura Wenderoth, Thilo Sentker, and René Werner. Self-supervision for medical image classification: State-of-the-art performance with ˜100 labeled training samples per class.Bioengineering, 10(8):895, 2023

    2023

  52. [52]

    Unsupervised learning of visual representations by solving jigsaw puzzles

    Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. InEuropean conference on computer vision, pages 69–84. Springer, 2016

    2016

  53. [53]

    Real- istic evaluation of deep semi-supervised learning algorithms.Advances in neural information processing systems, 31, 2018

    Avital Oliver, Augustus Odena, Colin A Raffel, Ekin Dogus Cubuk, and Ian Goodfellow. Real- istic evaluation of deep semi-supervised learning algorithms.Advances in neural information processing systems, 31, 2018

    2018

  54. [54]

    Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

    Pith/arXiv arXiv 2018

  55. [55]

    DINOv2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Pith/arXiv arXiv 2023

  56. [56]

    Self-supervised learning through colorization for microscopy images

    Vaidehi Pandey, Christoph Brune, and Nicola Strisciuglio. Self-supervised learning through colorization for microscopy images. InInternational Conference on Image Analysis and Processing, pages 621–632. Springer, 2022

    2022

  57. [57]

    Context encoders: Feature learning by inpainting

    Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2536–2544, 2016

    2016

  58. [58]

    Kosmos-2: Grounding multimodal large language models to the world.arXiv preprint arXiv:2306.14824, 2023

    Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world.arXiv preprint arXiv:2306.14824, 2023

    Pith/arXiv arXiv 2023

  59. [59]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021

    2021

  60. [60]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  61. [61]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015. 13

    2015

  62. [62]

    Junji Shiraishi, Shigehiko Katsuragawa, Junpei Ikezoe, Tsuneo Matsumoto, Takeshi Kobayashi, Ken-ichi Komatsu, Mitate Matsui, Hiroshi Fujita, Yoshie Kodera, and Kunio Doi. Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists’ detection of pulmonary nodules.A...

    2000

  63. [63]

    Fixmatch: Simplifying semi- supervised learning with consistency and confidence.Advances in neural information processing systems, 33:596–608, 2020

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raf- fel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi- supervised learning with consistency and confidence.Advances in neural information processing systems, 33:596–608, 2020

    2020

  64. [64]

    Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged con- sistency targets improve semi-supervised deep learning results.Advances in neural information processing systems, 30, 2017

    2017

  65. [65]

    What makes for good views for contrastive learning?Advances in neural information processing systems, 33:6827–6839, 2020

    Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning?Advances in neural information processing systems, 33:6827–6839, 2020

    2020

  66. [66]

    YOLOv12: Attention-centric real-time object detectors.Advances in neural information processing systems, 38:78433–78457, 2025

    Yunjie Tian, Qixiang Ye, and David Doermann. YOLOv12: Attention-centric real-time object detectors.Advances in neural information processing systems, 38:78433–78457, 2025

    2025

  67. [67]

    Real-time self-supervised achromatic face colorization.The Visual Computer, 39(12):6521–6536, 2023

    Hitika Tiwari, Venkatesh K Subramanian, and Yong-Sheng Chen. Real-time self-supervised achromatic face colorization.The Visual Computer, 39(12):6521–6536, 2023

    2023

  68. [68]

    Toward a collective agenda on ai for earth science data analysis.IEEE Geoscience and Remote Sensing Magazine, 9(2):88–104, 2021

    Devis Tuia, Ribana Roscher, Jan Dirk Wegner, Nathan Jacobs, Xiaoxiang Zhu, and Gustau Camps-Valls. Toward a collective agenda on ai for earth science data analysis.IEEE Geoscience and Remote Sensing Magazine, 9(2):88–104, 2021

    2021

  69. [69]

    A survey on semi-supervised learning.Machine learning, 109(2):373–440, 2020

    Jesper E Van Engelen and Holger H Hoos. A survey on semi-supervised learning.Machine learning, 109(2):373–440, 2020

    2020

  70. [70]

    Optimizing area under the roc curve using semi-supervised learning

    Shijun Wang, Diana Li, Nicholas Petrick, Berkman Sahiner, Marius George Linguraru, and Ronald M Summers. Optimizing area under the roc curve using semi-supervised learning. Pattern Recognition, 48(1):276–287, 2015

    2015

  71. [71]

    Unsupervised learning of visual representations using videos

    Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. InProceedings of the IEEE international conference on computer vision, pages 2794– 2802, 2015

    2015

  72. [72]

    Unsupervised data aug- mentation for consistency training.Advances in neural information processing systems, 33: 6256–6268, 2020

    Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data aug- mentation for consistency training.Advances in neural information processing systems, 33: 6256–6268, 2020

    2020

  73. [73]

    Simmim: A simple framework for masked image modeling

    Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653–9663, 2022

    2022

  74. [74]

    Image enhanced rotation prediction for self-supervised learning

    Shin’ya Yamaguchi, Sekitoshi Kanai, Tetsuya Shioda, and Shoichiro Takeda. Image enhanced rotation prediction for self-supervised learning. In2021 IEEE International Conference on Image Processing (ICIP), pages 489–493. IEEE, 2021

    2021

  75. [75]

    Self-supervised 3d action representation learning with skeleton cloud colorization.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):509–524, 2023

    Siyuan Yang, Jun Liu, Shijian Lu, Er Meng Hwa, Yongjian Hu, and Alex C Kot. Self-supervised 3d action representation learning with skeleton cloud colorization.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):509–524, 2023

    2023

  76. [76]

    Barlow twins: Self- supervised learning via redundancy reduction

    Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stephane Deny. Barlow twins: Self- supervised learning via redundancy reduction. InInternational conference on machine learning, pages 12310–12320. PMLR, 2021

    2021

  77. [77]

    S4L: Self-supervised semi-supervised learning

    Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4L: Self-supervised semi-supervised learning. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1476–1485, 2019. 14

    2019

  78. [78]

    Colorful image colorization

    Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. InEuropean conference on computer vision, pages 649–666. Springer, 2016

    2016

  79. [79]

    Combining self-supervised and supervised learning with noisy labels

    Yongqi Zhang, Hui Zhang, Quanming Yao, and Jun Wan. Combining self-supervised and supervised learning with noisy labels. In2023 IEEE International Conference on Image Processing (ICIP), pages 605–609. IEEE, 2023

    2023

  80. [80]

    iBOT: Image BERT pre-training with online tokenizer.arXiv preprint arXiv:2111.07832, 2021

    Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. iBOT: Image BERT pre-training with online tokenizer.arXiv preprint arXiv:2111.07832, 2021. 15 Appendices A Datasets 17 A.1 General Domain Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 A.2 Domain-Specific Datasets . . . . . . . . . . . . . . . ....

    Pith/arXiv arXiv 2021

Showing first 80 references.