Pith. sign in

REVIEW 3 major objections 5 minor 87 references

Solving Semi-Supervised Few-Shot Learning from an Auto-Annotation Perspective

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Vision-language models output near-uniform softmax scores, so SSL finetuning ignores unlabeled data; two temperatures restore it, gaining ~5 points across five benchmarks.

desk verdict Genuinely useful diagnosis and a simple fix that mostly works, but the universal temperature story does not survive EuroSAT; still deserves a real review. read the letter →

arxiv 2512.10244 v3 pith:E6OJEEFU submitted 2025-12-11 cs.CV cs.LG

Tian Liu , Anwesha Basu , James Caverlee , Shu Kong This is my paper · ORCID
classification cs.CVcs.LG
keywords semi-supervisedfew-shotlearningvision-languagemodelstemperaturetuningpseudo-labelingretrievalaugmentationfine-grainedclassificationstage-wisefinetuningauto-annotation
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

Vision-language models (VLMs) produce unusually flat softmax probability vectors when asked to classify fine-grained images. The paper argues that this flatness — a side effect of contrastive pretraining — is why established semi-supervised learning (SSL) methods fail to finetune a VLM from a few labeled images plus many unlabeled ones: pseudo-label confidence scores almost never pass the 0.8 threshold, so the unlabeled data is never used, and the gradient signal from labeled data is too weak to move the encoder. The proposed fix is two temperatures: a fixed confidence temperature Tconf=0.01 that sharpens pseudo-label scores, and a learnable loss temperature Tloss (initialized to 0.07) that magnifies the cross-entropy gradients. Combined with text-embedding classifier initialization, retrieval of task-relevant images from the VLM's public pretraining set, and a stage-wise schedule (classifier learning, SSL finetuning, final few-shot finetuning), the SWIFT pipeline reports a ~5 accuracy-point gain over prior FSL and SSL baselines across five fine-grained benchmarks, and rivals a fully supervised reference that assumes ground-truth labels on the unlabeled data.

What carries the argument

Temperature Tuning (TT): a fixed confidence temperature Tconf=0.01 applied to the softmax of weakly augmented unlabeled images, lifting pseudo-label confidence above the selection threshold, plus a learnable loss temperature Tloss initialized to 0.07 that scales logits in all cross-entropy losses, amplifying the otherwise weak supervision. The supporting devices are classifier initialization (seeding the classifier head with the VLM's text embeddings of class names) and a stage-wise schedule: finetune the classifier on labeled data, run SSL finetuning with labeled, unlabeled, and retrieved data, then finetune again on labeled data only to correct domain gaps and noise.

What would settle it

Run the full SWIFT pipeline on five to ten additional fine-grained datasets spanning class counts from 2 to 1000, comparing fixed Tconf=0.01 against a per-task tuned Tconf; if the fixed value fails to beat or match the tuned value on datasets with very few or very many classes, the universal-temperature claim collapses. A cheaper check: rerun FixMatch+TT on EuroSAT at 4-shot with several seeds and confirm the 11.9-point drop is stable.

Watch

Extended reading notes

Core claim

Working from an auto-annotation perspective, the paper contends that semi-supervised few-shot learning should exploit open-source vision-language models and their pretraining data, and that the reason SSL methods cannot finetune these models is the VLMs' flat softmax outputs. Two temperatures address this: Tconf=0.01 sharpens the confidence distribution so pseudo-labels exceed the threshold, and a learnable Tloss initialized to 0.07 amplifies the supervision signal in the loss. On top of this, classifier initialization with text embeddings and retrieval augmentation from the VLM's own pretraining set contribute further gains, and the three-stage schedule yields the final SWIFT method, which

Load-bearing premise

The load-bearing premise is that one fixed confidence temperature (0.01) and a learnable loss temperature initialized to 0.07, tuned once on semi-Aves and then frozen, sharpen VLM softmax outputs uniformly across class counts and datasets — yet the paper's own ablation shows the same Tconf hurts EuroSAT, the only 10-class dataset, by 11.9 points at 4-shot.

Editorial extensions

If this is right

  • Adding TT to standard SSL methods such as FixMatch or DebiasPL improves their accuracy by 14–20 points when finetuning a VLM, showing the failure is in the VLM's output distribution, not the SSL recipe.
  • Retrieving task-relevant images from the VLM's public pretraining set adds 6–11 points on top of TT, and is the component that lets SWIFT surpass the strongest FSL baseline (SWAT).
  • SWIFT rivals a fully supervised reference that uses ground-truth labels on the unlabeled data, implying that, in the low-label regime, unlabeled data plus a VLM can substitute for labels.
  • Because the temperature terms add negligible compute, the recipe can be applied to future, stronger SSL methods without architectural changes.

Reading between the lines

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

  • If the flat-softmax diagnosis is correct, the same two-temperature fix should transfer to other contrastively pretrained encoders and to other pseudo-labeling losses, though optimal temperature values may need re-tuning per model family and class count.
  • The paper fixes Tconf=0.01 for all datasets; an adaptive Tconf that scales with the number of classes could potentially recover the EuroSAT loss (11.9 points at 4-shot when Tconf is applied) while keeping gains elsewhere — a testable variant not explored in the paper.
  • A natural stress test is to run the same pipeline on a VLM whose logits are already sharp (e.g., after a small amount of supervised finetuning); if gains persist, the root-cause story is incomplete, whereas if they vanish, the flatness explanation is strengthened.
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 studies semi-supervised few-shot learning (SSFSL) in a realistic auto-annotation setup where the learner has access to a pretrained VLM, a small labeled set, an abundant unlabeled set, and task-relevant data retrieved from the VLM's pretraining corpus. The authors first show that directly applying standard SSL methods (FixMatch, DebiasPL) to finetune an OpenCLIP VLM substantially underperforms few-shot finetuning. They diagnose the cause as the 'flat' softmax distributions produced by contrastively pretrained VLMs, which lead to zero selection of pseudo-labels at the default confidence threshold and weak supervision for the CE loss. To address this, they propose classifier initialization from text embeddings and two temperature mechanisms: a fixed confidence temperature Tconf for pseudo-label selection, and a learnable loss temperature Tloss for strengthening supervision. Under the name SWIFT, they combine these with retrieval augmentation from LAION-400M and a three-stage training procedure (classifier init, semi-supervised finetuning with SSL loss, and final few-shot finetuning). Across five fine-grained benchmarks they report gains of about 5 accuracy points over the prior FSL state of the art (SWAT) and results rivaling fully supervised finetuning. The supplementary material contains per-dataset tables, ablations, additional backbone experiments (DINOv2, ImageNet-pretrained models), and pseudo-code.

Significance. If the claims hold, the paper makes a useful empirical and conceptual contribution: it identifies a simple, actionable cause for why conventional SSL training fails when finetuning contrastively pretrained VLMs, and it demonstrates a lightweight fix that delivers large gains. The strengths of the manuscript include extensive and mostly honest ablations (Table 3, Table 9), replication of the mechanism with two SSL methods and two backbones, per-dataset reporting, release of code, and a clear stage-wise pipeline. The central claim, however, is broader than the evidence, and the paper contains a protocol inconsistency regarding validation-set usage. Because the claimed 'root cause' and the universal temperature prescription are load-bearing for the paper's framing, these issues need to be addressed before the contribution can be fully accepted.

major comments (3)
  1. [Sec. 4.1 / Tab. 2] The paper's 'no validation set' claim is contradicted by its own tuning protocol. In Sec. 4.1, after stating that the setup 'eschews a validation set', the authors write: 'we follow [56] to carry out cross-dataset tuning. Specifically, we tune the temperature on semi-Aves (Fig. 6) and apply it to other datasets.' semi-Aves is one of the five benchmark datasets, and Fig. 6 reports test accuracy on it. Therefore the headline average in Tab. 2 includes the dataset used to select Tconf, Tloss init, and the confidence threshold. The claimed advantage of being validation-free and ready for real-world auto-annotation is not established by the current experiments. The limitation in Sec. 5 mentions this only as a possible future extension. Please retune on a genuine held-out validation split, or at minimum report results with semi-Aves excluded and acknowledge that the current protocol involves t
  2. [Tab. 9 / Tab. 8] EuroSAT is a direct counterexample to the central claim that Tconf=0.01 fixes the flat-softmax failure. In Table 9, 4-shot: adding Tconf to the retrieved-data run drops EuroSAT from 83.9 to 72.0 (-11.9), and the final SWIFT (75.0) is below the no-Tconf configuration (83.9). At 8-shot, SWIFT (87.1) is again below the no-Tconf row (88.0). EuroSAT is the only 10-class dataset in the benchmark, suggesting that class count or initial logit sharpness modulates whether Tconf helps. Because the protocol eschews a validation set, a practitioner applying the prescribed Tconf to a new low-class dataset cannot detect this harmful regime. The root-cause claim is therefore not uniform across the paper's own benchmarks. The paper should either provide a rule for setting Tconf (e.g., based on class count or logit statistics) or substantially temper the universality of the temperature prescription; Tab.
  3. [Sec. 3.2 / Eq. (3)] The causal attribution of Tloss to 'strengthening supervision' is confounded with loss scaling. Eq. (3) divides logits by Tloss before the cross-entropy, which is mathematically equivalent to scaling the CE loss by 1/Tloss (up to a constant shift in the log-sum-exp term). The paper compares Tloss=1.0 against Tloss=0.07 or a learnable Tloss, but does not compare against simply increasing the learning rate or the loss weight at Tloss=1.0, nor against a fixed small Tloss with a re-tuned LR. Figure 6's convergence plots are consistent with a learning-rate/loss-magnitude effect. Since the central claim is that the flat softmax weakens supervision in a way that temperature specifically fixes, please add a control that decouples loss temperature from loss magnitude, or at least acknowledge this alternative explanation.
minor comments (5)
  1. [Sec. 3.2 / Eq. (3) text] Typo: 'incorporating both Tconf and Tconf' should read 'Tconf and Tloss'.
  2. [Sec. 3.3] Typo: 'all the data in L, U, and U' should be 'L, U, and R'.
  3. [Tab. 2 caption / text] The fully supervised reference row is labeled 'w/ RAFT' in the main text; this looks like a typo for 'w/ RA'.
  4. [Supp. Tab. 9] In the 4-shot block, the SWIFT row reads '63.571.5' with no delimiter; please insert a space.
  5. [General] Main tables report no variance or number of seeds. Fig. 11 gives standard deviations only for semi-Aves. Adding per-dataset standard deviations over at least three seeds would strengthen the benchmarking claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: evaluation is held-out and temperatures are anchored to CLIP pretraining conventions, not fitted to the reported benchmarks.

full rationale

SWIFT's headline numbers are measured on held-out test sets against standard baselines; no equation in the paper defines a predicted quantity in terms of a fitted parameter. The two temperatures are anchored to CLIP's pretraining conventions (Tconf=0.01, 'the final temperature used in CLIP's pretraining'; Tloss initialized to 0.07, 'consistent with the setting in CLIP's pretraining'), and the confidence threshold 0.8 comes from [56], not from the target datasets. The cross-dataset tuning on semi-Aves is disclosed ('we tune the temperature on semi-Aves (Fig. 6) and apply it to other datasets'), and the paper explicitly lists as a limitation that unlabeled/retrieved data 'could potentially be leveraged for tuning hyperparameters and temperature.' The self-citations to [41] provide baselines, retrieval procedure, and hyperparameters; they are not invoked as a uniqueness theorem or as proof of the flat-softmax mechanism, which is demonstrated directly in Figs. 3, 6 and 7. Tab. 9's EuroSAT degradation (83.9→72.0 with Tconf; final SWIFT 75.0 below the no-Tconf 83.9) indicates the temperature prescription is not universally safe, and Tab. 4's DINOv2 result limits the universality of the 'root cause' story, but these are empirical robustness/correctness concerns, not circular derivation. No circular step is identifiable.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper's measurable contribution is a recipe of existing ingredients; its gains are driven by two temperature hyperparameters (Tconf fixed at 0.01, Tloss learnable with init 0.07) anchored to CLIP pretraining conventions, the FixMatch threshold 0.8, and 500 retrieved images per class from LAION-400M. The causal diagnosis assumes normalized cosine-similarity logits, leakage-free retrieval, and pseudo-label accuracy. No new entities are postulated.

free parameters (4)
  • Tconf (confidence temperature) = 0.01
    Fixed confidence temperature for pseudo-label softmax (Sec. 3.2); anchored to CLIP pretraining's final temperature and cross-dataset tuned on semi-Aves (Sec. 4.1). Load-bearing: it is what makes pseudo-labels pass the 0.8 threshold.
  • Tloss (learnable loss temperature, init) = 0.07
    Learnable temperature scaling the CE loss for labeled and pseudo-labeled data; initialized to 0.07 following CLIP/SimCLR pretraining conventions (Sec. 3.2, Fig. 6). Ablations show gains from learning it.
  • sigma (confidence threshold) = 0.8
    Pseudo-label selection threshold inherited from FixMatch [56]; interacts with Tconf to determine unlabeled-data utilization (Fig. 7).
  • Retrieved data size per class = 500
    Number of retrieved LAION-400M images per class, following [41]; the largest ablation jump comes from adding R (Tab. 3), so this quantity materially affects the headline gains.
assumptions (5)
  • domain assumption VLM logits are cosine similarities of normalized embeddings, so with tau=1 softmax is flat for large C (e.g., max prob ~ 1/200)
    The core diagnosis (Sec. 3.1, Eq. 1, Fig. 3) assumes W^T V(I) yields bounded cosine-similarity logits; otherwise the flatness claim would not hold.
  • domain assumption String-matching retrieval from LAION-400M yields task-relevant examples with useful (noisy) labels and no test-set leakage
    Sec. 4.1 asserts leakage-free retrieval; the largest single performance jump in the ablation (stage-2 RA) depends on this assumption.
  • domain assumption Text-embedding classifier initialization transfers task knowledge from the VLM
    Stage-1 and the FixMatch/DebiasPL remedies rely on W = [T(1),...,T(C)] (Sec. 3.2), standard CLIP practice following [48, 68].
  • domain assumption High-confidence pseudo-labels (above 0.8 after Tconf=0.01) are accurate enough to improve finetuning
    The mechanism only helps if sharpening selects correct rather than confidently-wrong labels; the EuroSAT drop in Tab. 9 shows this does not always hold.
  • domain assumption Hyperparameters from prior work transfer across datasets without a validation set
    Sec. 4.1 adopts learning rates, batch sizes, and epochs from [41, 56]; the no-validation-set protocol is load-bearing for the 'realistic setup' claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Solving Semi-Supervised Few-Shot Learning from an Auto-Annotation Perspective." pith.science (2026). https://pith.science/paper/E6OJEEFU

@misc{pith2026251210244,
  author       = {Pith},
  title        = {Pith review of: Solving Semi-Supervised Few-Shot Learning from an Auto-Annotation Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E6OJEEFU}},
  note         = {Machine review of arXiv:2512.10244}
}
abstract

Semi-supervised few-shot learning (SSFSL) resembles real-world applications such as auto-annotation, as it aims to learn a model from a few labeled and abundant unlabeled task-specific examples to annotate the unlabeled ones. Despite the availability of powerful open-source Vision-Language Models (VLMs) and open-world data, existing SSFSL literature largely neglects these resources. In contrast, the related area few-shot learning (FSL) has already exploited them to boost performance. Arguably, to solve real-world auto-annotation, SSFSL should leverage such open resources. To bridge this gap, we explore established SSL methods to finetune a VLM. Unexpectedly, they significantly underperform FSL baselines that do not use unlabeled data. Our in-depth analysis reveals the root cause of failure: VLMs produce flat distributions of softmax probabilities, resulting in zero utilization of unlabeled data and weak supervision signals. To address this challenge, we propose an embarrassingly simple solution that uses temperatures to sharpen the softmax output, which not only increases the confidence scores of pseudo-labels to improve the utilization of unlabeled data, but also strengthens training supervision for effective finetuning. Furthermore, we exploit task-relevant open data, e.g., those retrieved from VLMs' publicly available pretraining set. To mitigate the imbalance and domain gaps in retrieved data, we employ a stage-wise training strategy. Building on the successful finetuning of VLMs and the exploitation of open data, we present a simple yet effective SSFSL method, Stage-Wise Finetuning with Temperatures (SWIFT). Across five benchmarks, SWIFT outperforms recent FSL and SSL methods by $\sim$5 accuracy points. SWIFT even rivals supervised learning, which finetunes a VLM assuming unlabeled data having ground-truth labels!

Figures

Figures reproduced from arXiv: 2512.10244 by the authors.

Figure 1
Figure 1. Summary of SSFSL benchmarking results across five datasets. Influential SSL methods such as FixMatch [54] and DebiasPL [63], which finetune an ImageNet-pretrained ResNet50 backbone [23], underperform the state-of-the-art (SOTA) zero-shot learning (ZSL) methods like REAL [48]. By exploiting a VLM, the SOTA SSL method FineSSL [19] achieves notable gains, but it only learns prompts with a frozen VLM. Intuitively, finet… view at source ↗
Figure 2
Figure 2. Illustration of our SSFSL setup motivated by the real￾world data auto-annotation application. It embraces open-source VLMs and their publicly available pretraining data. The latter en￾ables retrieving task-relevant data to facilitate learning. Importantly, the setup simulates the data scarcity issue and eschews a validation set. This helps develop real-world applicable SSFSL approaches. pretraining data. Temperature… view at source ↗
Figure 3
Figure 3. Challenges in finetuning VLMs and the effect of temperature.. We run OpenCLIP ViT-B/32 on the unlabeled images of the semi-Aves dataset by zero-shot prompting its 200 class names. (A) VLMs produce a “flat” distribution of softmax probabilities (200-dim column vector per unlabeled example), resulting in low confidence scores and weak supervision signals that prevent effective finetuning. (B) Applying a temperature sh… view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: SWIFT workflow. SWIFT finetunes VLM’s visual encoder using few-shot labeled data, abundant unlabeled data, and retrieved data. It incorporates our Temperature Tuning (TT) to enhance pseudo-label selection and strengthen learning signals. pseudo-labeled data ( [PITH_FU…
Figure 6
Figure 6. Figure 6: Demonstration of the importance of loss temperature Tloss through few-shot finetuning. We finetune the visual encoder of the VLM OpenCLIP ViT-B/32 [12] on 16-shot samples from semi-Aves [55], where the classifier weights are initialized via linear probing on the 16-sho…
Figure 8
Figure 8. Figure 8: Visual examples of few-shot labeled data, unlabeled data, and retrieved data for representative classes from each dataset. Notably, the unlabeled data follows the same distribution as the few-shot labeled data, exhibiting similar visual appearances. In contrast, the re…
Figure 10
Figure 10. Figure 10: Comparison of confidence scores between ImageNet￾pretrained and CLIP ResNet-50 models. Following the same settings in [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Impact of initial value for learning loss tempera￾ture Tloss on linear probing, FSL (e.g., using Few-shot Finetuning [41]), and SSFSL (e.g., using FixMatch [54]) with VLM. We run experiments using OpenCLIP ViT-B/32 model [12] with 16-shot labeled data sampled with thr…
Figure 12
Figure 12. Figure 12: Example PyTorch-style pseudo-code for Temperature Tuning in stage-2 semi-supervised finetuning of SWIFT. We illustrate our simple TT with FixMatch. Specifically, we apply (1) a learnable loss temperature Tloss_x initialized to 0.07 to calculate the cross-entropy loss …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

87 extracted references · 6 linked inside Pith

  1. [81]

    Revisiting semi-supervised learning in the era of foun- dation models.Advances in Neural Information Processing Systems (NeurIPS), 2025

    Ping Zhang, Zheda Mai, Quang-Huy Nguyen, and Wei-Lun Chao. Revisiting semi-supervised learning in the era of foun- dation models.Advances in Neural Information Processing Systems (NeurIPS), 2025. 2, 3, 4, 5, 7

    2025

  2. [56]

    A real- istic evaluation of semi-supervised learning for fine-grained classification

    Jong-Chyi Su, Zezhou Cheng, and Subhransu Maji. A real- istic evaluation of semi-supervised learning for fine-grained classification. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 4, 6, 7, 18

    2021

  3. [1]

    Pseudo-labeling and confirmation bias in deep semi-supervised learning

    Eric Arazo, Diego Ortego, Paul Albert, Noel E O’Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In2020 International Joint Conference on Neural Networks (IJCNN), 2020. 2

    2020

  4. [2]

    Mixmatch: A holistic approach to semi-supervised learning.Advances in Neural Information Processing Systems (NeurIPS), 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 (NeurIPS), 32, 2019. 1

    2019

  5. [3]

    Cubuk, Alex Ku- rakin, Kihyuk Sohn, Han Zhang, and Colin Raffel

    David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Ku- rakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix- match: Semi-supervised learning with distribution matching and augmentation anchoring. InInternational Conference on Learning Representations (ICLR), 2020. 1

    2020

  6. [4]

    Adamatch: A unified approach to semi-supervised learning and domain adaptation

    David Berthelot, Rebecca Roelofs, Kihyuk Sohn, Nicholas Carlini, and Alex Kurakin. Adamatch: A unified approach to semi-supervised learning and domain adaptation. 2022. 2

    2022

  7. [5]

    Exponential moving average normalization for self-supervised and semi- supervised learning

    Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Char- less Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi- supervised learning. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR),

  8. [6]

    Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning

    Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2

    2021

Show all 87 references
  1. [7]

    Semi-supervised learning (chapelle, o

    Olivier Chapelle, Bernhard Scholkopf, and Alexander Zien. Semi-supervised learning (chapelle, o. et al., eds.; 2006)[book reviews].IEEE Transactions on Neural Networks, 20(3):542– 542, 2009. 2

    2006

  2. [8]

    Plot: Prompt learning with optimal transport for vision-language models

    Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. InInternational Conference on Learning Representations (ICLR), 2023. 6, 15

    2023

  3. [9]

    Softmatch: Addressing the quantity-quality tradeoff in semi- supervised learning

    Hao Chen, Ran Tao, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Bhiksha Raj, and Marios Savvides. Softmatch: Addressing the quantity-quality tradeoff in semi- supervised learning. InThe Eleventh International Confer- ence on Learning Representations (ICLR), 2023. 1, 2

    2023

  4. [10]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge- offrey Hinton. A simple framework for contrastive learning of visual representations. InInternational Conference on Machine Learning (ICML), 2020. 3, 4

    2020

  5. [11]

    Big self-supervised models are strong semi-supervised learners.Advances in Neural Information Processing Systems (NeurIPS), 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 (NeurIPS), 2020. 2

    2020

  6. [12]

    Reproducible scaling laws for contrastive language-image learning

    Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuh- mann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In IEEE/CVF Conference on Computer Vision and Pattern Reco...

    2023

  7. [13]

    Enhancing semi-supervised learning with zero-shot pseudolabels.arXiv preprint arXiv:2502.12584, 2025

    Jichan Chung and Irene Y Chen. Enhancing semi-supervised learning with zero-shot pseudolabels.arXiv preprint arXiv:2502.12584, 2025. 2

    2025 arXiv

  8. [14]

    Describing textures in the wild

    Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 5, 13

    2014

  9. [15]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. InProceedings of the IEEE/CVF Computer Society Confer- ence on Computer Vision and Pattern Recognition (CVPR),

  10. [16]

    Pseudo-labeling based practical semi-supervised meta-training for few-shot learning.IEEE Transactions on Image Processing (TIP), 2024

    Xingping Dong, Tianran Ouyang, Shengcai Liao, Bo Du, and Ling Shao. Pseudo-labeling based practical semi-supervised meta-training for few-shot learning.IEEE Transactions on Image Processing (TIP), 2024. 1, 2

    2024

  11. [17]

    An image is worth 16x16 words: Transform- ers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Transform- ers for image recognition at scale. InThe Eleventh I...

    2021

  12. [18]

    Simpro: A simple probabilistic framework towards realistic long-tailed semi- supervised learning.arXiv preprint arXiv:2402.13505, 2024

    Chaoqun Du, Yizeng Han, and Gao Huang. Simpro: A simple probabilistic framework towards realistic long-tailed semi- supervised learning.arXiv preprint arXiv:2402.13505, 2024. 2

    2024 arXiv

  13. [19]

    Erasing the bias: Fine-tuning founda- tion models for semi-supervised learning.Forty-first Interna- tional Conference on Machine Learning (ICML), 2024

    Kai Gan and Tong Wei. Erasing the bias: Fine-tuning founda- tion models for semi-supervised learning.Forty-first Interna- tional Conference on Machine Learning (ICML), 2024. 1, 2, 6, 13, 15, 16

    2024

  14. [20]

    Clip- adapter: Better vision-language models with feature adapters

    Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip- adapter: Better vision-language models with feature adapters. International Journal of Computer Vision (IJCV), 132(2): 581–595, 2024. 2, 3, 6, 15

    2024

  15. [21]

    On calibration of modern neural networks

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. InInternational Conference on Machine Learning (ICML), 2017. 3

    2017

  16. [22]

    Class-imbalanced semi- supervised learning with adaptive thresholding

    Lan-Zhe Guo and Yu-Feng Li. Class-imbalanced semi- supervised learning with adaptive thresholding. InInterna- tional Conference on Machine Learning (ICML), 2022. 2

    2022

  17. [23]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 16, 17, 18 9

    2016

  18. [24]

    Momentum contrast for unsupervised visual repre- sentation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual repre- sentation learning. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR),

  19. [25]

    Introducing eurosat: A novel dataset and deep learning benchmark for land use and land cover clas- sification

    Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Introducing eurosat: A novel dataset and deep learning benchmark for land use and land cover clas- sification. InIEEE International Geoscience and Remote Sensing Symposium (IGARSS). IEEE, 2018. 5, 13

    2018

  20. [26]

    Deep anomaly detection with outlier exposure

    Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. InICLR,

  21. [27]

    Distill- ing the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distill- ing the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015. 3

    2015 arXiv

  22. [28]

    The curious case of neural text degeneration

    Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. InThe Eleventh International Conference on Learning Representa- tions (ICLR), 2020. 3

    2020

  23. [29]

    Pseudo-loss confidence metric for semi-supervised few- shot learning

    Kai Huang, Jie Geng, Wen Jiang, Xinyang Deng, and Zhe Xu. Pseudo-loss confidence metric for semi-supervised few- shot learning. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2

    2021

  24. [30]

    Retrieval-enhanced contrastive vision-text models

    Ahmet Iscen, Mathilde Caron, Alireza Fathi, and Cordelia Schmid. Retrieval-enhanced contrastive vision-text models. InInternational Conference on Learning Representations (ICLR), 2024. 2, 5

    2024

  25. [31]

    Categorical reparam- eterization with gumbel-softmax

    Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparam- eterization with gumbel-softmax. InICLR, 2017. 3

    2017

  26. [32]

    Scaling up visual and vision-language representa- tion 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 representa- tion learning with noisy text supervision. InInternational Conference on Machine Learning (ICML), 2021. 2

    2021

  27. [33]

    Opengan: Open-set recog- nition via open data generation

    Shu Kong and Deva Ramanan. Opengan: Open-set recog- nition via open data generation. InProceedings of the IEEE/CVF international conference on computer vision, pages 813–822, 2021. 2

    2021

  28. [34]

    3d object representations for fine-grained categorization

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE International Conference on Computer Vision (ICCV) Workshops, 2013. 5, 13

    2013

  29. [35]

    Temporal ensembling for semi- supervised learning

    Samuli Laine and Timo Aila. Temporal ensembling for semi- supervised learning. 2017. 2

    2017

  30. [36]

    Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks

    Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. InWorkshop on challenges in representation learning, In- ternational Conference on Machine Learning (ICML), 2013. 2

    2013

  31. [37]

    Learning to self-train for semi-supervised few-shot classification.Advances in Neural Information Processing Systems (NeurIPS), 2019

    Xinzhe Li, Qianru Sun, Yaoyao Liu, Qin Zhou, Shibao Zheng, Tat-Seng Chua, and Bernt Schiele. Learning to self-train for semi-supervised few-shot classification.Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 2, 6

    2019

  32. [38]

    Multimodality helps unimodality: Cross- modal few-shot learning with multimodal models

    Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross- modal few-shot learning with multimodal models. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 4, 6, 8, 15

    2023

  33. [39]

    Semi- supervised few-shot learning via multi-factor clustering

    Jie Ling, Lei Liao, Meng Yang, and Jia Shuai. Semi- supervised few-shot learning via multi-factor clustering. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2022. 1, 2

    2022

  34. [40]

    Learning customized visual models with retrieval-augmented knowledge

    Haotian Liu, Kilho Son, Jianwei Yang, Ce Liu, Jianfeng Gao, Yong Jae Lee, and Chunyuan Li. Learning customized visual models with retrieval-augmented knowledge. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5

    2023

  35. [41]

    Few-shot recognition via stage-wise retrieval-augmented fine- tuning

    Tian Liu, Huixin Zhang, Shubham Parashar, and Shu Kong. Few-shot recognition via stage-wise retrieval-augmented fine- tuning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 3, 4, 5, 6, 13, 14, 15, 16, 17, 18

    2025

  36. [42]

    Tenet: Beyond pseudo-labeling for semi-supervised few-shot learn- ing.Machine Intelligence Research, pages 1–13, 2025

    Chengcheng Ma, Weiming Dong, and Changsheng Xu. Tenet: Beyond pseudo-labeling for semi-supervised few-shot learn- ing.Machine Intelligence Research, pages 1–13, 2025. 2

    2025

  37. [43]

    Revisiting few-shot object detection with vision-language models

    Anish Madan, Neehar Peri, Shu Kong, and Deva Ramanan. Revisiting few-shot object detection with vision-language models. InAdvances in Neural Information Processing Sys- tems (NeurIPS) Datasets & Benchmark Track, 2024. 1, 2

    2024

  38. [44]

    Fine-grained visual clas- sification of aircraft.arXiv:1306.5151, 2013

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual clas- sification of aircraft.arXiv:1306.5151, 2013. 5, 13

    2013 arXiv

  39. [45]

    En- hancing clip with clip: Exploring pseudolabeling for limited- label prompt tuning.Advances in Neural Information Pro- cessing Systems (NeurIPS), 2023

    Cristina Menghini, Andrew Delworth, and Stephen Bach. En- hancing clip with clip: Exploring pseudolabeling for limited- label prompt tuning.Advances in Neural Information Pro- cessing Systems (NeurIPS), 2023. 2

    2023

  40. [46]

    Long-tail learning via logit adjustment

    Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. InInternational Con- ference on Learning Representations (ICLR), 2021. 2

    2021

  41. [47]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rab...

    2024

  42. [48]

    The neglected tails in vision-language models

    Shubham Parashar, Zhiqiu Lin, Tian Liu, Xiangjue Dong, Yanan Li, Deva Ramanan, James Caverlee, and Shu Kong. The neglected tails in vision-language models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 6, 8, 13, 16

    2024

  43. [49]

    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 (ICM...

    2021

  44. [50]

    Im- proved zero-shot classification by adapting vlms with text 10 descriptions

    Oindrila Saha, Grant Van Horn, and Subhransu Maji. Im- proved zero-shot classification by adapting vlms with text 10 descriptions. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 5

    2024

  45. [51]

    Laion-400m: Open dataset of clip-filtered 400 million image-text pairs

    Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 5, 6, 13, 16

    2021 arXiv

  46. [52]

    Laion-5b: An open large-scale dataset for training next gen- eration image-text models.Advances in Neural Information Processing Systems (NeurIPS), 2022

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next gen- eration image-text models.Advances in Neural Inform...

    2022

  47. [53]

    A closer look at the few-shot adaptation of large vision-language models

    Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 6, 13, 15

    2024

  48. [54]

    Fixmatch: Simpli- fying semi-supervised learning with consistency and confi- dence.Advances in Neural Information Processing Systems (NeurIPS), 2020

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simpli- fying semi-supervised learning with consistency and confi- dence.Advances in Neural Information Processing Systems (N...

    2020

  49. [55]

    The semi- supervised inaturalist-aves challenge at fgvc7 workshop

    Jong-Chyi Su and Subhransu Maji. The semi- supervised inaturalist-aves challenge at fgvc7 workshop. arXiv:2103.06937, 2021. 5, 6, 7, 13, 16, 17, 18

    2021 arXiv

  50. [57]

    Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results.Advances in Neural Information Processing Systems (NeurIPS), 2017. 2

    2017

  51. [58]

    A closer look at the robustness of contrastive language-image pre-training (clip)

    Weijie Tu, Weijian Deng, and Tom Gedeon. A closer look at the robustness of contrastive language-image pre-training (clip). InAdvances in Neural Information Processing Systems (NeurIPS), 2023. 3

    2023

  52. [59]

    An empirical study into what matters for calibrating vision-language models

    Weijie Tu, Weijian Deng, Dylan Campbell, Stephen Gould, and Tom Gedeon. An empirical study into what matters for calibrating vision-language models. InInternational Confer- ence on Machine Learning (ICML), 2024. 3

    2024

  53. [60]

    Neural priming for sample-efficient adaptation

    Matthew Wallingford, Vivek Ramanujan, Alex Fang, Aditya Kusupati, Roozbeh Mottaghi, Aniruddha Kembhavi, Ludwig Schmidt, and Ali Farhadi. Neural priming for sample-efficient adaptation. InAdvances in Neural Information Processing Systems (NeurIPS), 2023. 2, 5

    2023

  54. [61]

    Normface: L2 hypersphere embedding for face verifi- cation

    Feng Wang, Xiang Xiang, Jian Cheng, and Alan Loddon Yuille. Normface: L2 hypersphere embedding for face verifi- cation. InProceedings of the 25th ACM International Confer- ence on Multimedia, 2017. 3

    2017

  55. [62]

    Enabling val- idation for robust few-shot recognition.arXiv preprint arXiv:2506.04713, 2025

    Hanxin Wang, Tian Liu, and Shu Kong. Enabling val- idation for robust few-shot recognition.arXiv preprint arXiv:2506.04713, 2025. 2

    2025

  56. [63]

    Debiased learning from naturally imbalanced pseudo-labels

    Xudong Wang, Zhirong Wu, Long Lian, and Stella X Yu. Debiased learning from naturally imbalanced pseudo-labels. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 6, 7, 8, 13, 15, 16

    2022

  57. [64]

    USB: A unified semi-supervised learning benchmark for classification

    Yidong Wang, Hao Chen, Yue Fan, Wang SUN, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, Heli Qi, Zhen Wu, Yu-Feng Li, Satoshi Nakamura, Wei Ye, Marios Savvides, Bhiksha Raj, Takahiro Shinozaki, Bernt Schiele, Jindong Wang, Xing Xie, and Yue Zhang. USB: A...

    2022

  58. [65]

    Freematch: Self-adaptive thresholding for semi-supervised learning

    Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shi- nozaki, Bhiksha Raj, Bernt Schiele, and Xing Xie. Freematch: Self-adaptive thresholding for semi-supervised learning. In The Eleventh International Conference on Learn...

    2023

  59. [66]

    Towards realistic long-tailed semi- supervised learning: Consistency is all you need

    Tong Wei and Kai Gan. Towards realistic long-tailed semi- supervised learning: Consistency is all you need. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2

    2023

  60. [67]

    An embarrassingly simple approach to semi-supervised few-shot learning.Advances in Neural Information Process- ing Systems (NeurIPS), 2022

    Xiu-Shen Wei, H-Y Xu, Faen Zhang, Yuxin Peng, and Wei Zhou. An embarrassingly simple approach to semi-supervised few-shot learning.Advances in Neural Information Process- ing Systems (NeurIPS), 2022. 1, 2

    2022

  61. [68]

    Robust fine-tuning of zero-shot models

    Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gon- tijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. InIEEE/CVF Conference on Computer Vision and ...

    2022

  62. [69]

    Unsupervised feature learning via non-parametric instance discrimination

    Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3, 7

    2018

  63. [70]

    Self-training with noisy student improves imagenet clas- sification

    Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet clas- sification. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2

    2020

  64. [71]

    Demysti- fying clip data

    Hu Xu, Saining Xie, Xiaoqing Ellen Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demysti- fying clip data. InInternational Conference on Learning Representations (ICLR), 2024. 2, 3, 7

    2024

  65. [72]

    Dash: Semi-supervised learning with dynamic thresholding

    Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. InInternational Conference on Machine Learning (ICML). PMLR, 2021. 2

    2021

  66. [73]

    Unimatch v2: Pushing the limit of semi-supervised semantic segmenta- tion

    Lihe Yang, Zhen Zhao, and Hengshuang Zhao. Unimatch v2: Pushing the limit of semi-supervised semantic segmenta- tion. InIEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI). IEEE, 2025. 2

    2025

  67. [74]

    Task residual for tuning vision-language models

    Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6

    2023

  68. [75]

    Task residual for tuning vision-language models

    Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. 11 InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6, 15

    2023

  69. [76]

    Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling.Advances in Neural Information Processing Systems (NeurIPS), 2021

    Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling.Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 2

    2021

  70. [77]

    En- hancing vision-language few-shot adaptation with negative learning

    Ce Zhang, Simon Stepputtis, Katia Sycara, and Yaqi Xie. En- hancing vision-language few-shot adaptation with negative learning. In2025 IEEE/CVF Winter Conference on Applica- tions of Computer Vision (WACV), 2025. 2

    2025

  71. [78]

    Candidate pseudolabel learning: Enhancing vision-language models by prompt tuning with unlabeled data.Forty-first International Conference on Machine Learning (ICML), 2024

    Jiahan Zhang, Qi Wei, Feng Liu, and Lei Feng. Candidate pseudolabel learning: Enhancing vision-language models by prompt tuning with unlabeled data.Forty-first International Conference on Machine Learning (ICML), 2024. 2, 3

    2024

  72. [79]

    Revisiting semi-supervised learning in the era of foundation models

    Ping Zhang, Zheda Mai, Quang-Huy Nguyen, and Wei-Lun Chao. Revisiting semi-supervised learning in the era of foundation models. https : / / github . com / OSU - MLB / SSL - Foundation - Models / blob / 3b4eb48d4986456b6f3c6c5576ff25aff781026f/ semilearn/core/criterions/cross_e...

    2025

  73. [80]

    Revisiting semi-supervised learning in the era of foundation models

    Ping Zhang, Zheda Mai, Quang-Huy Nguyen, and Wei-Lun Chao. Revisiting semi-supervised learning in the era of foundation models. https : / / github . com / OSU - MLB / SSL - Foundation - Models / blob / 3b4eb48d4986456b6f3c6c5576ff25aff781026f/ semilearn/nets/vit_petl/vit.py#L2...

    2025

  74. [82]

    Tip- adapter: Training-free adaptation of clip for few-shot classifi- cation

    Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kun- chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip- adapter: Training-free adaptation of clip for few-shot classifi- cation. InEuropean Conference on Computer Vision (ECCV),

  75. [83]

    Simmatch: Semi-supervised learning with similarity matching

    Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2

    2022

  76. [84]

    string- matching

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models.In- ternational Journal of Computer Vision (IJCV), 2022. 3, 6, 15 12 Solving Semi-Supervised Few-Shot Learning from an Auto-Annotation Perspective Supplementary Material ...

    2022

  77. [85]

    flat softmax probabilities

    to use a learning rate of 1e-4 to update the classifier, and a smaller learning rate of 1e-6 to update the visual encoder to preserve the pretrained features, with a cosine annealing learning rate scheduler. The weight decay is set to 1e-2 with theAdamWoptimizer. For FixMatch ...

    2012

  78. [86]

    Specifically, for both FSL and SSL, we initialize the classifier via linear probing on few-shot data per Tab

    models and the VLM CLIP [12, 49] of ResNet-50 [23] and ViT-B/32 [17] architectures under different temperature settings. Specifically, for both FSL and SSL, we initialize the classifier via linear probing on few-shot data per Tab. 10. Then, for FSL, we finetune a pretrained vi...

  79. [87]

    We compare the performance with finetuning DINOv2 directly using few-shot data (FS-FT) following [41]

    with 4-, 8-, and 16-shot labeled data. We compare the performance with finetuning DINOv2 directly using few-shot data (FS-FT) following [41]. Results show that each component yields significant gains, especially the retrieval augmentation (RA), validating that SWIFT effectivel...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.