Pith. sign in

REVIEW 3 major objections 5 minor 49 references

ULFine: Unbiased Lightweight Fine-tuning for Foundation-Model-Assisted Long-Tailed Semi-Supervised Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read ULFine shows that a frozen CLIP with confidence-adapted text prototypes and dual-logit fusion lifts long-tailed semi-supervised learning far above prior methods while cutting training cost by over ten times.

desk verdict Useful failure-mode analysis for CLIP in LTSSL, but the headline 10x cost claim is contradicted by Table VIII and the main SOTA comparisons use a different backbone. read the letter →

arxiv 2505.05062 v1 pith:NID3LXWV submitted 2025-05-08 cs.CV

classification cs.CV
keywords long-tailedsemi-supervisedlearningfoundationmodelCLIPlightweightfine-tuningpseudo-labelsduallogitfusionprototypeadaptivefittingclassimbalance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper asks whether a pretrained vision-language foundation model (CLIP) can improve long-tailed semi-supervised learning, a setting with few labeled samples in a skewed class distribution and many unlabeled samples. It reports that the obvious strategies fail in opposite ways: linear probing helps head classes but leaves tail classes stuck, and lightweight fine-tuning reduces false pseudo-labels but becomes overconfident about the ones it still gets wrong. The paper proposes ULFine, which adapts text prototypes to the data through confidence-aware updates and then fuses the text-prototype logits with the linear-probe logits at a fixed ratio. If the claims hold, practitioners could obtain much higher accuracy in this realistic setting with roughly a tenth of the training cost of current scratch-trained methods.

What carries the argument

The load-bearing mechanism is the pair (PAF, DLF). PAF maintains visual prototypes as an exponential moving average of per-class feature means and moves each CLIP text prototype toward the corresponding visual prototype with a confidence-aware coefficient $\alpha_k = \mu \cdot P_u^k / \max_i P_u^i$, where $P_u^i$ is the pseudo-labeled predictive distribution for class $i$; it also applies an orthogonality MSE loss to the visual-prototype similarity matrix. DLF aligns the text-similarity logits $p_i^t$ to the linear-probe logits $p_i^v$ via the ratio $\beta = (\max p_i^v - \min p_i^v)/(\max p_i^t - \min p_i^t)$, rescales them to $\hat{p}_i^t$, and fuses $p_i = \eta p_i^v + (1-\eta)\hat{p}_i^t$ with $\eta=0.7$; pseudo-labels are drawn from the fused logits under logit adjustment. This carries the argument by combining two biased estimators into one designed to be less biased: the linear probe leans head-ward, the semantic prototype leans tail-ward, and their fusion generates both pseudo-labels and test predictions.

What would settle it

Construct an LTSSL dataset whose class names have text embeddings that encode the same head-class bias as the linear classifier, then check whether the fused pseudo-labels still beat both individual sources; if the fusion merely averages two correlated biased signals, the claimed unbiasedness fails.

Watch

Extended reading notes

Core claim

The central claim is that the two pathologies of foundation-model-assisted long-tailed semi-supervised learning, underlearning under linear probing and biased overconfidence under lightweight fine-tuning, can be corrected without full fine-tuning. ULFine's Prototype Adaptive Fitting (PAF) moves CLIP text prototypes toward per-class visual prototypes with a confidence-aware update rate and pushes the prototypes toward mutual orthogonality, shrinking the influence of head-dominated pseudo-label distributions. Dual Logit Fusion (DLF) then rescales the text-similarity logits to the range of the linear-probe logits and combines them with a fixed weight, exploiting the observation that linear-probe logits favor head classes while semantic-prototype logits favor tail classes. On CIFAR10-LT, CIFAR100-LT, STL10-LT, and ImageNet-127, the paper reports top-1 accuracies far above existing long-tailed semi-supervised baselines while training only 15,000 iterations with about 0.1 million learnable parameters.

Load-bearing premise

The load-bearing premise is that after lightweight adaptation, the linear-probe logits and the text-prototype similarity logits stay complementary across datasets and distribution shifts, so a single fixed blend weight of 0.7 in the fusion equation always yields less biased pseudo-labels than either source alone.

Editorial extensions

If this is right

  • Long-tailed semi-supervised learning no longer needs training from scratch: a frozen or lightly tuned CLIP can reach higher accuracy at roughly one-tenth the training cost, making large-scale LTSSL feasible on modest hardware.
  • The minority-bottleneck and majority-overconfidence analysis gives a diagnostic: methods that raise overall accuracy without raising tail accuracy may still be injecting confident false pseudo-labels, and ULFine's confidence statistics show both the count and confidence of false pseudo-labels dropping.
  • Dual logit fusion implies that when linear-probe and semantic-similarity signals disagree, their weighted combination is a better pseudo-label source than either alone, and this same fused logit is kept at inference time.
  • ULFine is a general wrapper over lightweight fine-tuning: the paper reports consistent gains across BitFit, visual prompt tuning variants, Adapters, LoRA, and AdaptFormer, so the recipe attaches to the adapter rather than to a single architecture.

Reading between the lines

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

  • The paper tests only CLIP, but the complementarity it exploits is structural, between a linear classifier and a semantic similarity classifier, so a natural extension is to test whether the same dual-logit pattern recurs with other vision-language encoders.
  • The fixed fusion weight $\eta=0.7$ is a weak point: the reported sensitivity curve shows overall accuracy is stable while head and tail accuracy trade off, so an adaptive or per-class $\eta$ could extract further tail gains.
  • The diagnosis that lightweight fine-tuning becomes overconfident on false pseudo-labels suggests that explicit confidence-calibration losses could be combined with ULFine to push the overconfidence curve down further, a direction the paper does not explore.
  • If the tenfold cost reduction holds, the practical bottleneck for LTSSL shifts from compute to the availability of a pretrained foundation model whose label vocabulary matches the task, reframing LTSSL as an adaptation rather than a representation-learning problem.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 investigates how a pre-trained vision-language foundation model (CLIP) can be used for long-tailed semi-supervised learning (LTSSL). It first compares three strategies for employing CLIP, namely linear probing (LP), lightweight fine-tuning (LFT), and full fine-tuning (FFT), and identifies two failure modes: a "minority bottleneck" where tail classes are neglected, and "majority overconfidence" where false pseudo-labels receive high confidence. The proposed method, ULFine, combines Prototype Adaptive Fitting (PAF), which updates text prototypes in a confidence-aware manner and enforces orthogonality between prototypes, with Dual Logit Fusion (DLF), which fuses linear-probing logits and text-prototype similarity logits. Experiments on CIFAR10-LT, CIFAR100-LT, STL10-LT, and ImageNet-127 report large accuracy gains over prior LTSSL methods, and the abstract claims that ULFine "markedly decreases training costs by over ten times" while substantially improving accuracy.

Significance. If the empirical claims held as stated, the paper would establish a practically important direction: replacing scratch training with foundation-model initialization plus parameter-efficient fine-tuning can yield large accuracy gains and much shorter training schedules in LTSSL. The paper has real strengths: the analysis of LP/LFT/FFT is clearly presented; the ablations in Table IV show that PAF and DLF improve over LP and LFT with the same CLIP backbone; Table VI demonstrates that ULFine works across several lightweight fine-tuning strategies; and the balanced-CIFAR100 results in Table V show competitiveness with an existing CLIP-based SSL method. However, the headline cost claim is contradicted by the paper's own wall-clock timings in Table VIII, and the main accuracy comparisons against prior state-of-the-art methods use a different, scratch-trained backbone, so the significance of the cross-method gains is currently overstated and needs recalibration.

major comments (3)
  1. [Section VI-I, Tables VII and VIII]
  2. [Section V-C and Tables I-III, VII]
  3. [Section IV-B and VI-B]
minor comments (5)
  1. [Table VII and Section VI-I]
  2. [Eq. (1) and Section III-A]
  3. [Section VI-G and Fig. 6(a)]
  4. [Section IV-A, Eq. (4)]
  5. [Section I and Fig. 1]

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: ULFine is an empirical training scheme validated against external benchmarks; the self-referential prototype update is a self-training feedback loop, not a derived prediction.

full rationale

ULFine does not present a formal derivation whose outputs reduce to its inputs. Its contributions are an empirical analysis of CLIP adaptation strategies and a training scheme (PAF + DLF) evaluated on CIFAR10/100-LT, STL10-LT, and ImageNet-127. PAF updates textual prototypes using the model's own pseudo-label distribution (Eq. 4), and DLF fuses those prototypes' logits with linear-probing logits (Eqs. 6-7); this is a standard self-training feedback loop, and the paper does not claim to predict the resulting pseudo-labels or accuracies from the update rule by construction—the update rule is the method, and its success is measured externally against baselines. The complementarity assumption behind DLF is attributed to DASO [21], an external prior work, not to a same-author uniqueness theorem, and is empirically probed with the eta sensitivity curve (Fig. 6a). The ablation in Table IV isolates each component's contribution rather than renaming a fitted value as a prediction. The abstract's 'over ten times' training-cost claim is based on iteration counts in Table VII and appears inconsistent with the wall-clock times in Table VIII (ULFine 7470 s vs FixMatch 15500 s), but that is an internal consistency/correctness concern, not a circular reduction. No equation equals its input by definition, no prediction is statistically forced by a fitted parameter, and no load-bearing premise rests on a self-citation. Therefore no significant circularity is present.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central method rests on several unstated or lightly justified choices: the fusion weight eta=0.7, the prototype momentum mu, the temperature T, and the assumption that pseudo-label distributions approximate true class distributions. None are derived, and some values are not reported.

free parameters (6)
  • eta (fusion weight in Eq. 7) = 0.7
    Fixed at 0.7 in Sec IV-B; sensitivity analysis on one dataset (Fig 6a) shows stable overall accuracy but head/tail tradeoff.
  • mu (weighting in confidence-aware coefficient alpha_k) = not specified
    Controls prototype update speed in Eq. 4; value not reported in the paper.
  • Temperature T in similarity logit = not specified
    Scales text-prototype logits in Sec IV-B (pt_i = sim(zw_i, Ct)/T); value not reported.
  • FixMatch masking threshold tau = not restated for ULFine
    Standard SSL threshold used in the mask of Eq. 1; the paper does not state the value used.
  • Orthogonality loss weight for Lo = not specified
    Added to total loss in Eq. 9 with no explicit weighting coefficient.
  • EMA decay for visual prototypes Cv = not specified
    Used to compute visual prototype matrix in Sec IV-A; decay rate is not given.
assumptions (5)
  • domain assumption CLIP's visual and text encoders share a comparable embedding space, so linear min-max alignment of logits (Eq. 6) is meaningful.
    DLF relies on aligning the ranges of linear-probing and text-prototype logits; if the spaces are not calibrated, the fusion is ad hoc.
  • domain assumption The pseudo-labeled predictive distribution P_u approximates the true unlabeled class distribution.
    Used in PAF's alpha_k (Eq. 4) and in post-hoc logit adjustment Pl; if pseudo-labels are biased, the confidence-aware updates are biased.
  • ad hoc to paper The complementary relationship between linear-classifier and semantic-similarity pseudo-labels, established for DASO [21] with a different backbone, transfers to CLIP-based features.
    DLF is explicitly inspired by DASO [21]; no same-backbone DASO baseline is provided to validate this transfer.
  • domain assumption FixMatch's consistency-training assumptions (weak/strong augmentation, confidence masking) remain valid with a frozen or lightly tuned CLIP encoder.
    The whole framework is built on FixMatch (Eq. 1); if CLIP features violate the augmentation invariance, the training signal is unreliable.
  • standard math Standard mathematical operations (EMA, MSE, softmax, logit adjustment) are used as stated.
    No new mathematics is introduced; the paper uses standard losses and probability operations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ULFine: Unbiased Lightweight Fine-tuning for Foundation-Model-Assisted Long-Tailed Semi-Supervised Learning." pith.science (2026). https://pith.science/paper/NID3LXWV

@misc{pith2026250505062,
  author       = {Pith},
  title        = {Pith review of: ULFine: Unbiased Lightweight Fine-tuning for Foundation-Model-Assisted Long-Tailed Semi-Supervised Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NID3LXWV}},
  note         = {Machine review of arXiv:2505.05062}
}
read the original abstract

Based on the success of large-scale visual foundation models like CLIP in various downstream tasks, this paper initially attempts to explore their impact on Long-Tailed Semi-Supervised Learning (LTSSL) by employing the foundation model with three strategies: Linear Probing (LP), Lightweight Fine-Tuning (LFT), and Full Fine-Tuning (FFT). Our analysis presents the following insights: i) Compared to LTSSL algorithms trained from scratch, FFT results in a decline in model performance, whereas LP and LFT, although boosting overall model performance, exhibit negligible benefits to tail classes. ii) LP produces numerous false pseudo-labels due to \textit{underlearned} training data, while LFT can reduce the number of these false labels but becomes overconfident about them owing to \textit{biased fitting} training data. This exacerbates the pseudo-labeled and classifier biases inherent in LTSSL, limiting performance improvement in the tail classes. With these insights, we propose a Unbiased Lightweight Fine-tuning strategy, \textbf{ULFine}, which mitigates the overconfidence via confidence-aware adaptive fitting of textual prototypes and counteracts the pseudo-labeled and classifier biases via complementary fusion of dual logits. Extensive experiments demonstrate that ULFine markedly decreases training costs by over ten times and substantially increases prediction accuracies compared to state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2505.05062 by the authors.

Figure 1
Figure 1. On the CIFAR100-LT dataset, (a): Comparison of top-1 accuracy of Linear Probing (LP), Lightweight Fine-Tuning (LFT), and Full Fine-Tuning (FFT) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Performance comparison of various methods on the CIFAR10-LT with [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Statistics of relevant results on CIFAR10-LT. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Top-1 classification accuracy per class under different distribution settings of CIFAR100-LT, comparing various strategies. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The vertical axes (left and right) indicate the confidence level (area plot) and the sample size (dashed line) of false pseudo-labeling. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: (a) Effect of different a on model performance on the CIFAR10-LT dataset. (b) Comparison of similarity matrices between textual prototypes before [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 30 canonical work pages

  1. [1]

    Dash: Semi-supervised learning with dynamic thresholding,

    Y . Xu, L. Shang, J. Ye, Q. Qian, Y .-F. Li, B. Sun, H. Li, and R. Jin, “Dash: Semi-supervised learning with dynamic thresholding,” in International conference on machine learning . PMLR, 2021, pp. 11 525–11 536

  2. [2]

    Cossl: Co-learning of representation and classifier for imbalanced semi-supervised learning,

    Y . Fan, D. Dai, A. Kukleva, and B. Schiele, “Cossl: Co-learning of representation and classifier for imbalanced semi-supervised learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 14 574–14 584

  3. [3]

    Freematch: Self-adaptive thresholding for semi-supervised learning,

    Y . Wang, H. Chen, Q. Heng, W. Hou, Y . Fan, Z. Wu, J. Wang, M. Sav- vides, T. Shinozaki, B. Raj et al., “Freematch: Self-adaptive thresholding for semi-supervised learning,” arXiv preprint arXiv:2205.07246 , 2022

  4. [4]

    Allmatch: Exploiting all unlabeled data for semi- supervised learning,

    Z. Wu and J. Cui, “Allmatch: Exploiting all unlabeled data for semi- supervised learning,” arXiv preprint arXiv:2406.15763 , 2024

  5. [5]

    Fixmatch: Simplifying semi- supervised learning with consistency and confidence,

    K. Sohn, D. Berthelot, N. Carlini, Z. Zhang, H. Zhang, C. A. Raffel, E. D. Cubuk, A. Kurakin, and C.-L. Li, “Fixmatch: Simplifying semi- supervised learning with consistency and confidence,” Advances in neural information processing systems , vol. 33, pp. 596–608, 2020

  6. [6]

    Dynamic learnable logit adjustment for long-tailed visual recognition,

    E. Zhang, C. Geng, C. Li, and S. Chen, “Dynamic learnable logit adjustment for long-tailed visual recognition,” IEEE Transactions on Circuits and Systems for Video Technology , 2024

  7. [7]

    Learning label shift correction for test-agnostic long-tailed recognition,

    T. Wei, Z. Mao, Z.-H. Zhou, Y . Wan, and M.-L. Zhang, “Learning label shift correction for test-agnostic long-tailed recognition,” 2024, p. 52611–52631

  8. [8]

    Large- scale long-tailed recognition in an open world,

    Z. Liu, Z. Miao, X. Zhan, J. Wang, B. Gong, and S. X. Yu, “Large- scale long-tailed recognition in an open world,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 2537–2546

Show all 49 references
  1. [9]

    Continuous contrastive learning for long-tailed semi-supervised recog- nition,

    Z.-H. Zhou, S. Fang, Z.-J. Zhou, T. Wei, Y . Wan, and M.-L. Zhang, “Continuous contrastive learning for long-tailed semi-supervised recog- nition,” arXiv preprint arXiv:2410.06109 , 2024

  2. [10]

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

    T. Wei and K. Gan, “Towards realistic long-tailed semi-supervised learning: Consistency is all you need,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 3469–3478

  3. [11]

    Three heads are better than one: Complementary experts for long-tailed semi-supervised learn- ing,

    C. Ma, I. Elezi, J. Deng, W. Dong, and C. Xu, “Three heads are better than one: Complementary experts for long-tailed semi-supervised learn- ing,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 13, 2024, pp. 14 229–14 237

  4. [12]

    Distribu- tion aligning refinery of pseudo-label for imbalanced semi-supervised learning,

    J. Kim, Y . Hur, S. Park, E. Yang, S. J. Hwang, and J. Shin, “Distribu- tion aligning refinery of pseudo-label for imbalanced semi-supervised learning,” Advances in neural information processing systems , vol. 33, pp. 14 567–14 579, 2020. JOURNAL OF LATEX CLASS FILES, VOL. 14,...

  5. [13]

    Crest: A class- rebalancing self-training framework for imbalanced semi-supervised learning,

    C. Wei, K. Sohn, C. Mellina, A. Yuille, and F. Yang, “Crest: A class- rebalancing self-training framework for imbalanced semi-supervised learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 10 857–10 866

  6. [14]

    Class-imbalanced semi-supervised learning with adaptive thresholding,

    L.-Z. Guo and Y .-F. Li, “Class-imbalanced semi-supervised learning with adaptive thresholding,” in International conference on machine learning. PMLR, 2022, pp. 8082–8094

  7. [15]

    Smoothed adaptive weighting for imbalanced semi-supervised learning: Improve reliability against unknown distribution data,

    Z. Lai, C. Wang, H. Gunawan, S.-C. S. Cheung, and C.-N. Chuah, “Smoothed adaptive weighting for imbalanced semi-supervised learning: Improve reliability against unknown distribution data,” in International Conference on Machine Learning . PMLR, 2022, pp. 11 828–11 843

  8. [16]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning . PMLR, 2021, pp. 8748–8763

  9. [17]

    What makes clip more robust to long-tailed pre-training data? a controlled study for transferable insights,

    X. Wen, B. Zhao, Y . Chen, J. Pang, and X. Qi, “What makes clip more robust to long-tailed pre-training data? a controlled study for transferable insights,” in The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

  10. [18]

    Long-tail learning with foundation model: Heavy fine-tuning hurts,

    J.-X. Shi, T. Wei, Z. Zhou, J.-J. Shao, X.-Y . Han, and Y .-F. Li, “Long-tail learning with foundation model: Heavy fine-tuning hurts,” in Forty-first International Conference on Machine Learning , 2024

  11. [19]

    Learning transferable negative prompts for out-of-distribution detection,

    T. Li, G. Pang, X. Bai, W. Miao, and J. Zheng, “Learning transferable negative prompts for out-of-distribution detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 17 584–17 594

  12. [20]

    Locoop: Few-shot out- of-distribution detection via prompt learning,

    A. Miyai, Q. Yu, G. Irie, and K. Aizawa, “Locoop: Few-shot out- of-distribution detection via prompt learning,” Advances in Neural Information Processing Systems , vol. 36, 2024

  13. [21]

    Daso: Distribution-aware semantics- oriented pseudo-label for imbalanced semi-supervised learning,

    Y . Oh, D.-J. Kim, and I. S. Kweon, “Daso: Distribution-aware semantics- oriented pseudo-label for imbalanced semi-supervised learning,” in Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 9786–9796

  14. [22]

    Bacon: Boosting imbalanced semi- supervised learning via balanced feature-level contrastive learning,

    Q. Feng, L. Xie, S. Fang, and T. Lin, “Bacon: Boosting imbalanced semi- supervised learning via balanced feature-level contrastive learning,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 11, 2024, pp. 11 970–11 978

  15. [23]

    Abc: Auxiliary balanced classifier for class-imbalanced semi-supervised learning,

    H. Lee, S. Shin, and H. Kim, “Abc: Auxiliary balanced classifier for class-imbalanced semi-supervised learning,” Advances in Neural Information Processing Systems , vol. 34, pp. 7082–7094, 2021

  16. [24]

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

    C. Jia, Y . Yang, Y . Xia, Y .-T. Chen, Z. Parekh, H. Pham, Q. Le, Y .-H. Sung, Z. Li, and T. Duerig, “Scaling up visual and vision-language representation learning with noisy text supervision,” in International conference on machine learning . PMLR, 2021, pp. 4904–4916

  17. [25]

    Slip: Self-supervision meets language-image pre-training,

    N. Mu, A. Kirillov, D. Wagner, and S. Xie, “Slip: Self-supervision meets language-image pre-training,” in European conference on computer vision. Springer, 2022, pp. 529–544

  18. [26]

    Coca: Contrastive captioners are image-text foundation models,

    J. Yu, Z. Wang, V . Vasudevan, L. Yeung, M. Seyedhosseini, and Y . Wu, “Coca: Contrastive captioners are image-text foundation models,” Transactions on Machine Learning Research , 2022

  19. [27]

    Simvlm: Simple visual language model pretraining with weak supervision,

    Z. Wang, J. Yu, A. W. Yu, Z. Dai, Y . Tsvetkov, and Y . Cao, “Simvlm: Simple visual language model pretraining with weak supervision,” arXiv preprint arXiv:2108.10904, 2021

  20. [28]

    Uniformly distributed cat- egory prototype-guided vision-language framework for long-tail recog- nition,

    X. He, S. Fu, X. Ding, Y . Cao, and H. Wang, “Uniformly distributed cat- egory prototype-guided vision-language framework for long-tail recog- nition,” in Proceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 5027–5037

  21. [29]

    Vl-ltr: Learning class- wise visual-linguistic representation for long-tailed visual recognition,

    C. Tian, W. Wang, X. Zhu, J. Dai, and Y . Qiao, “Vl-ltr: Learning class- wise visual-linguistic representation for long-tailed visual recognition,” in European conference on computer vision. Springer, 2022, pp. 73–91

  22. [30]

    Ltgc: Long- tail recognition via leveraging llms-driven generated content,

    Q. Zhao, Y . Dai, H. Li, W. Hu, F. Zhang, and J. Liu, “Ltgc: Long- tail recognition via leveraging llms-driven generated content,” in Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 19 510–19 520

  23. [31]

    A simple long-tailed recognition baseline via vision-language model,

    T. Ma, S. Geng, M. Wang, J. Shao, J. Lu, H. Li, P. Gao, and Y . Qiao, “A simple long-tailed recognition baseline via vision-language model,” arXiv preprint arXiv:2111.14745 , 2021

  24. [32]

    Lpt: Long-tailed prompt tuning for image classification,

    B. Dong, P. Zhou, S. Yan, and W. Zuo, “Lpt: Long-tailed prompt tuning for image classification,” in The Eleventh International Conference on Learning Representations, 2022

  25. [33]

    Improving visual prompt tuning by gaussian neighborhood minimization for long- tailed visual recognition,

    M. Li, Y . Liu, Y . Lu, Y . Zhang, Y .-m. Cheung, and H. Huang, “Improving visual prompt tuning by gaussian neighborhood minimization for long- tailed visual recognition,” in The Thirty-eighth Annual Conference on Neural Information Processing Systems , 2024

  26. [34]

    Imbalanced semi- supervised learning with bias adaptive classifier,

    R. Wang, X. Jia, Q. Wang, Y . Wu, and D. Meng, “Imbalanced semi- supervised learning with bias adaptive classifier,” 2023

  27. [35]

    Bem: Balanced and entropy-based mix for long-tailed semi-supervised learning,

    H. Zheng, L. Zhou, H. Li, J. Su, X. Wei, and X. Xu, “Bem: Balanced and entropy-based mix for long-tailed semi-supervised learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 22 893–22 903

  28. [36]

    Twice class bias correction for imbalanced semi-supervised learning,

    L. Li, B. Tao, L. Han, D.-c. Zhan, and H.-j. Ye, “Twice class bias correction for imbalanced semi-supervised learning,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 12, 2024, pp. 13 563–13 571

  29. [37]

    Continuous contrastive learning for long-tailed semi-supervised recog- nition,

    Z.-H. Zhou, S. Fang, Z.-J. Zhou, T. Wei, Y . Wan, and M.-L. Zhang, “Continuous contrastive learning for long-tailed semi-supervised recog- nition,” in The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

  30. [38]

    Long-tail learning via logit adjustment,

    A. K. Menon, S. Jayasumana, A. S. Rawat, H. Jain, A. Veit, and S. Kumar, “Long-tail learning via logit adjustment,” arXiv preprint arXiv:2007.07314, 2020

  31. [39]

    Learning multiple layers of features from tiny images,

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

  32. [40]

    An analysis of single-layer networks in unsupervised feature learning,

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

  33. [41]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al. , “Pytorch: An imperative style, high-performance deep learning library,” Advances in neural information processing systems , vol. 32, 2019

  34. [42]

    Erasing the bias: Fine-tuning foundation models for semi-supervised learning,

    K. Gan and T. Wei, “Erasing the bias: Fine-tuning foundation models for semi-supervised learning,” in Forty-first International Conference on Machine Learning, 2024

  35. [43]

    Adaptformer: Adapting vision transformers for scalable visual recogni- tion,

    S. Chen, C. Ge, Z. Tong, J. Wang, Y . Song, J. Wang, and P. Luo, “Adaptformer: Adapting vision transformers for scalable visual recogni- tion,” Advances in Neural Information Processing Systems , vol. 35, pp. 16 664–16 678, 2022

  36. [44]

    Transfer and share: semi-supervised learning from long-tailed data,

    T. Wei, Q.-Y . Liu, J.-X. Shi, W.-W. Tu, and L.-Z. Guo, “Transfer and share: semi-supervised learning from long-tailed data,” Machine Learning, vol. 113, no. 4, pp. 1725–1742, 2024

  37. [45]

    Debiased learning from naturally imbalanced pseudo-labels,

    X. Wang, Z. Wu, L. Lian, and S. X. Yu, “Debiased learning from naturally imbalanced pseudo-labels,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 14 647–14 657

  38. [46]

    Bitfit: Simple parameter- efficient fine-tuning for transformer-based masked language-models,

    E. B. Zaken, S. Ravfogel, and Y . Goldberg, “Bitfit: Simple parameter- efficient fine-tuning for transformer-based masked language-models,” arXiv preprint arXiv:2106.10199 , 2021

  39. [47]

    Visual prompt tuning,

    M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim, “Visual prompt tuning,” in European Conference on Computer Vision. Springer, 2022, pp. 709–727

  40. [48]

    Parameter-efficient transfer learning for nlp,

    N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly, “Parameter-efficient transfer learning for nlp,” in International conference on machine learning . PMLR, 2019, pp. 2790–2799

  41. [49]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen et al., “Lora: Low-rank adaptation of large language models,” in Inter- national Conference on Learning Representations

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.