Pith. sign in

REVIEW 4 major objections 41 references

SciLT: Long-tailed Image Classification under Scientific Image Domains

T0 review · 4 major / 0 minor · reviewed 2026-07-13 · grok-4.5

Challenge this review Download PDF Read on arXiv ↗

Pith's one-line read Fine-tuning foundation models on long-tailed scientific images yields only limited gains, but fusing penultimate and final layer features with dual supervision balances head and tail accuracy.

desk verdict Useful domain-shift diagnosis for scientific long-tail PEFT, but the headline “foundation models barely help” rests on a confounded ResNet vs CLIP-ViT comparison; SciLT itself is a modest, clean dual-head recipe. read the letter →

arxiv 2604.03687 v2 submitted 2026-04-04 cs.CV

Jiahao Chen , Bing Su This is my paper
classification cs.CV
keywords long-tailedrecognitionscientificimagesfoundationmodelsparameter-efficientfine-tuningpenultimatelayerfeaturesfeaturefusiondomainshiftmedicalimaging
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

Long-tailed classification work has mostly assumed natural images, where pre-training and the target task live in similar visual worlds. Scientific images (cells, skin lesions, chest X-rays) look different and carry different semantic goals, so the usual fine-tuning recipe may not transfer. This paper shows, on three scientific benchmarks, that parameter-efficient fine-tuning of a vision foundation model often adds little over training from scratch and can even hurt under strong domain shift. It further finds that features from the penultimate layer are especially useful for rare classes and are distributionally different from the final layer. SciLT fuses those two layers with instance-dependent gates, trains one head with logit adjustment and one with ordinary cross-entropy, and averages their logits at test time, producing more balanced head–tail performance.

What carries the argument

SciLT: adaptive routing-calibrated fusion of penultimate- and final-layer features, dual heads supervised by logit adjustment (fused features) and cross-entropy (final features), with logit ensembling at inference.

What would settle it

Re-run the Table 2 / Finding 1 protocol with a convolutional backbone of similar capacity that is first pre-trained on natural images then fine-tuned the same way; if that setup closes the gap to training from scratch on NIH-Chest and erases SciLT’s BScore advantage, the claim that domain shift makes foundation PEFT ineffective would be undermined.

Watch

Extended reading notes

Core claim

Under a purely visual fine-tuning setup, foundation models give only limited—and sometimes negative—gains on scientific long-tailed data relative to training from scratch, while penultimate-layer features supply complementary signal that is particularly helpful for tail classes; jointly using penultimate and final features via adaptive fusion and dual LA/CE supervision (SciLT) improves balanced accuracy across Blood, ISIC, and NIH-Chest.

Load-bearing premise

That comparing an AdaptFormer-tuned CLIP ViT-B/16 to a ResNet-18 trained from scratch is a fair measure of whether foundation-model fine-tuning helps on scientific long-tailed data.

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

4 major / 0 minor

Summary. The paper studies long-tailed image classification on scientific domains (Blood, ISIC, NIH-Chest) under a purely visual PEFT setup with CLIP ViT-B/16 and AdaptFormer. Empirically, it reports that fine-tuning foundation models yields only limited or even negative gains relative to training from scratch, and that penultimate-layer features can match or exceed final-layer features, especially for tail classes. Motivated by a measured Wasserstein gap between penultimate and final features, the authors propose SciLT: adaptive fusion of the two layers, dual supervision (LA on the fused head, CE on the final-layer head), and logit ensembling at inference. On the three benchmarks SciLT improves a proposed BScore (harmonic mean of overall and macro accuracy) over CE and LA, with the largest gains on NIH-Chest medium/few classes.

Significance. Scientific long-tailed recognition under large domain shift is a genuine and underexplored setting; most foundation-model long-tail work is confined to natural-image benchmarks. The penultimate-vs-final analysis (Tables 3–5) and the purely visual design are useful contributions if the limited-gains narrative is cleaned up. SciLT itself is a simple, practical PEFT recipe with clear BScore gains on ISIC and NIH-Chest. The work would be a solid empirical baseline paper for adapting foundation models to scientific long-tail data, provided the architecture confound in Finding 1 is fixed and SciLT is compared against stronger same-backbone long-tail methods.

major comments (4)
  1. §4.2, Table 2 and Finding 1: the claim that “foundation models merely benefit” scientific long-tailed data rests on AdaptFormer-tuned CLIP ViT-B/16 versus ResNet-18 trained from scratch. Capacity, inductive bias, optimizer dynamics, and pre-training status all differ, so the gap cannot be attributed cleanly to domain shift. This comparison is load-bearing for the paper’s motivation. Please add architecture-matched controls (e.g., ViT-B/16 from scratch or ImageNet-pretrained ResNet/ViT under the same PEFT protocol) before stating Finding 1 in its current form.
  2. §6.1, Tables 6–8: SciLT is evaluated almost exclusively against CE and LA on the same backbone. Table 2 already lists CB, LDAM, Focal, and LADE for ResNet, but those methods are not re-run under AdaptFormer-CLIP. Without same-backbone long-tail baselines (e.g., LA, LDAM, PaCo-style, or classifier re-training on the fused features), the claim that SciLT “consistently outperforms existing methods” is overstated. At minimum, report LA/CE/LDAM/Focal with the identical AdaptFormer setup and the same dual-head budget.
  3. Appendix A / NIH-Chest protocol: multi-label ChestX-ray labels are collapsed to single-label by “selecting the first non-No Finding label.” This is an ad-hoc reduction that can systematically bias class frequencies and co-occurrence structure. Please justify or ablate this choice (e.g., random disease among positives, multi-label metrics, or official multi-label evaluation) and show that SciLT’s BScore gains are not an artifact of the conversion rule.
  4. §5.3, Theorem 5.3 and Remark: the Rademacher bound is a standard sum-space argument for ensembles; the claimed link from Wasserstein distance (Table 5) to a positive ensemble gain Δ_ens is informal and not used in the proof. Either tighten the theory to a concrete condition under which W-distance implies Δ_ens > complexity penalty, or reframe §5.3 as a brief complexity discussion rather than a formal justification of SciLT’s advantage.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical PEFT study with standard training/evaluation and non-load-bearing theory.

full rationale

This is a purely empirical methods paper. Observations (limited PEFT gains on scientific LT data vs. ResNet-from-scratch, penultimate-layer utility for tails, Wasserstein gap) are measured on held-out splits and motivate SciLT; they do not define or force the method's success. SciLT trains adaptive fusion gates and dual heads (LA on fused features, CE on final-layer) on labeled data in the ordinary supervised way, then ensembles logits at inference. BScore is an evaluation harmonic mean of OvAcc and Macro, not a fitted target or training objective. The Rademacher/generalization bounds (Lemmas 5.1–5.2, Thm. 5.3) are standard textbook applications that upper-bound complexity of the sum hypothesis class and do not reduce the empirical claims to their inputs by construction. No self-definitional loops, no fitted-parameter-as-prediction, no load-bearing self-citation uniqueness theorems, and no ansatz smuggled via overlapping-author citations. The confounded ResNet-vs-CLIP comparison is a validity concern, not circularity. The derivation chain is self-contained against external benchmarks.

Assumptions & free parameters 5 free parameters · 6 assumptions · 3 invented entities

The central claims rest on standard long-tail losses and PEFT machinery plus a few paper-specific modeling choices (which layers to fuse, dual LA/CE heads, BScore, and the ResNet-vs-ViT comparison as evidence of domain-shift failure). Free parameters are ordinary training hyperparameters and fusion coefficients; axioms are domain and evaluation assumptions rather than deep mathematical postulates. Invented entities are the SciLT modules and the BScore summary metric.

free parameters (5)
  • Per-dataset learning rates (Blood 0.012, ISIC 0.008, NIH-Chest 0.01)
    Chosen for training; performance of SciLT and baselines depends on these PEFT optimization settings.
  • Fusion interaction coefficient γ and gating parameters W1, W2
    Control how penultimate and final experts are mixed in Eq. (1); not fixed by theory.
  • Logit-adjustment temperature / class-prior strength in LA
    Standard LA hyperparameter that strongly affects tail vs head trade-off in all LA-based runs.
  • AdaptFormer bottleneck rank and which blocks receive adapters
    PEFT capacity choice that can change transfer behavior; not exhaustively ablated.
  • Training epochs (10/10/5) and batch size 128
    Short schedules may under- or over-adapt the foundation model relative to from-scratch ResNet training.
assumptions (6)
  • domain assumption CLIP ViT-B/16 pretraining plus AdaptFormer is a representative foundation-model fine-tuning paradigm for scientific long-tailed recognition.
    All main foundation-model results use this single PEFT setup (§4.1); conclusions about 'fine-tuning foundation models' inherit this choice.
  • ad hoc to paper ResNet-18 trained from scratch is a fair reference for whether foundation models 'merely benefit' scientific datasets.
    Table 2 contrasts different architectures; the limited-gain claim depends on this comparison being meaningful.
  • domain assumption Wasserstein distance between penultimate and final features indicates complementary, non-redundant information useful for long-tail learning.
    Proposition 1 and method motivation (§4.3–5) treat distributional discrepancy as evidence for fusion benefit.
  • ad hoc to paper NIH-Chest multi-label labels can be reduced to single-label by taking the first non-'No Finding' disease label.
    Appendix A defines this conversion; class frequencies and difficulty depend on it.
  • ad hoc to paper BScore = 2·OvAcc·Macro/(OvAcc+Macro) is an appropriate primary summary of balanced scientific long-tail performance on imbalanced test sets.
    Introduced in §6; SciLT’s 'consistent outperformance' is largely argued via this metric.
  • standard math Standard Rademacher complexity bounds for sum hypothesis classes apply to SciLT’s ensemble of two heads.
    Lemmas 5.1–5.2 and Theorem 5.3 cite classical complexity results; they bound complexity but do not prove empirical superiority.
invented entities (3)
  • SciLT adaptive multi-level fusion + dual-supervision framework
    purpose: Fuse penultimate and final ViT features with instance-dependent gates, train with LA on fused head and CE on final head, ensemble at inference.
    Core proposed method (§5.2); independent evidence is only the paper’s own benchmarks, not external replications.
  • BScore (harmonic mean of overall and macro accuracy)
    purpose: Summarize head/tail trade-off on naturally imbalanced scientific test sets.
    Paper-defined evaluation scalar (§6); not an external standard metric.
  • Routing-calibrated aggregation module (gates r and interaction Ψ)
    purpose: Instance-adaptive combination of layer-wise expert features.
    Architectural component introduced for SciLT; no external validation beyond ablations in Table 9.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SciLT: Long-tailed Image Classification under Scientific Image Domains." pith.science (2026). https://pith.science/paper/2604.03687

@misc{pith2026260403687,
  author       = {Pith},
  title        = {Pith review of: SciLT: Long-tailed Image Classification under Scientific Image Domains},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2604.03687}},
  note         = {Machine review of arXiv:2604.03687}
}
read the original abstract

Long-tailed recognition has benefited from foundation models and fine-tuning paradigms, yet existing studies and benchmarks are mainly confined to natural image domains, where pre-training and fine-tuning data share similar distributions. In contrast, scientific images exhibit distinct visual characteristics and supervision signals, raising questions about the effectiveness of fine-tuning foundation models in such settings. In this work, we investigate scientific long-tailed recognition under a purely visual and fine-tuning paradigm. Experiments on three scientific benchmarks show that fine-tuning foundation models yields limited gains, and reveal that penultimate-layer features play an important role, particularly for tail classes. Motivated by these findings, we propose SciLT, a framework that exploits multi-level representations through adaptive feature fusion and dual-supervision learning. By jointly leveraging penultimate- and final-layer features, SciLT achieves balanced performance across head and tail classes. Extensive experiments demonstrate that SciLT consistently outperforms existing methods, establishing a strong and practical baseline for scientific long-tailed recognition and providing valuable guidance for adapting foundation models to scientific data with substantial domain shifts.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 6 linked inside Pith

  1. [1]

    Large-scale long-tailed recognition in an open world

    Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2537–2546, 2019

    2019

  2. [2]

    Class-balanced loss based on effective number of samples

    Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9268–9277, 2019

    2019

  3. [3]

    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

    arXiv 2010

  4. [4]

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

    Jiang-Xin Shi, Tong Wei, Zhi Zhou, Jie-Jing Shao, Xin-Yan Han, and Yu-Feng Li. Long-tail learning with foundation model: Heavy fine-tuning hurts. InForty-first International Conference on Machine Learning, 2024

    2024

  5. [5]

    Lpt: Long-tailed prompt tuning for image classification.arXiv preprint arXiv:2210.01033, 2022

    Bowen Dong, Pan Zhou, Shuicheng Yan, and Wangmeng Zuo. Lpt: Long-tailed prompt tuning for image classification.arXiv preprint arXiv:2210.01033, 2022

    arXiv 2022

  6. [6]

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

    Changyao Tian, Wenhai Wang, Xizhou Zhu, Jifeng Dai, and Yu Qiao. Vl-ltr: Learning class- wise visual-linguistic representation for long-tailed visual recognition. InEuropean Conference on Computer Vision, pages 73–91. Springer, 2022

    2022

  7. [7]

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

    Qihao Zhao, Yalun Dai, Hao Li, Wei Hu, Fan Zhang, and Jun Liu. Ltgc: Long-tail recognition via leveraging llms-driven generated content. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19510–19520, 2024

    2024

  8. [8]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. InInternational conference on machine learning, pages 8821–8831. Pmlr, 2021

    2021

Show all 41 references
  1. [9]

    Long-tail learning via logit adjustment.arXiv preprint arXiv:2007.07314, 2020

    Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment.arXiv preprint arXiv:2007.07314, 2020

    2007 arXiv

  2. [10]

    Balanced meta-softmax for long- tailed visual recognition.Advances in Neural Information Processing Systems, 33:4175–4186, 2020

    Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long- tailed visual recognition.Advances in Neural Information Processing Systems, 33:4175–4186, 2020

    2020

  3. [11]

    Long-tailed multi-label visual recognition by collaborative training on uniform and re-balanced samplings

    Hao Guo and Song Wang. Long-tailed multi-label visual recognition by collaborative training on uniform and re-balanced samplings. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15089–15098, 2021

    2021

  4. [12]

    Imbalanced continual learning with partitioning reservoir sampling

    Chris Dongjoo Kim, Jinseo Jeong, and Gunhee Kim. Imbalanced continual learning with partitioning reservoir sampling. InEuropean Conference on Computer Vision, pages 411–428. Springer, 2020

    2020

  5. [13]

    A simple long-tailed recognition baseline via vision-language model.arXiv preprint arXiv:2111.14745, 2021

    Teli Ma, Shijie Geng, Mengmeng Wang, Jing Shao, Jiasen Lu, Hongsheng Li, Peng Gao, and Yu Qiao. A simple long-tailed recognition baseline via vision-language model.arXiv preprint arXiv:2111.14745, 2021

    2021 arXiv

  6. [14]

    Learning from neighbors: Category extrapolation for long-tail learning

    Shizhen Zhao, Xin Wen, Jiahui Liu, Chuofan Ma, Chunfeng Yuan, and Xiaojuan Qi. Learning from neighbors: Category extrapolation for long-tail learning. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 30483–30492, 2025

    2025

  7. [15]

    Matthew Ragoza, Julien Hochuli, Ernesto Idrobo, Jocelyn Sunseri, and David R. Koes. Protein– ligand scoring with convolutional neural networks.Journal of Chemical Information and Modeling, 57(4):942–957, 2017

    2017

  8. [16]

    Kdeep: Protein– ligand absolute binding affinity prediction via 3d convolutional neural networks.Journal of Chemical Information and Modeling, 58(2):287–296, 2018

    Javier Jiménez, Miha Skalic, Gerard Martinez-Rosell, and Gianni De Fabritiis. Kdeep: Protein– ligand absolute binding affinity prediction via 3d convolutional neural networks.Journal of Chemical Information and Modeling, 58(2):287–296, 2018. 10

    2018

  9. [17]

    Transfusion: Understanding transfer learning for medical imaging.Advances in neural information processing systems, 32, 2019

    Maithra Raghu, Chiyuan Zhang, Jon Kleinberg, and Samy Bengio. Transfusion: Understanding transfer learning for medical imaging.Advances in neural information processing systems, 32, 2019

    2019

  10. [18]

    Big self-supervised models advance medical image classification

    Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, et al. Big self-supervised models advance medical image classification. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021

    2021

  11. [19]

    Climd: A curriculum learning framework for imbalanced multimodal diagnosis

    Kai Han, Chongwen Lyu, Lele Ma, Chengxuan Qian, Siqi Ma, Zheng Pang, Jun Chen, and Zhe Liu. Climd: A curriculum learning framework for imbalanced multimodal diagnosis. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 65–74. Sprin...

    2025

  12. [20]

    Deciphering the extremes: A novel approach for pathological long-tailed recognition in scientific discovery

    Zhe Zhao, HaiBin Wen, Xianfu Liu, Rui Mao, Pengkun Wang, Liheng Yu, Linjiang Chen, Bo An, Qingfu Zhang, and Yang Wang. Deciphering the extremes: A novel approach for pathological long-tailed recognition in scientific discovery. InThe Thirty-ninth Annual Conference on Neural In...

    2025

  13. [21]

    Decoupling representation and classifier for long-tailed recognition.arXiv preprint arXiv:1910.09217, 2019

    Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition.arXiv preprint arXiv:1910.09217, 2019

    1910 arXiv

  14. [22]

    Improving calibration for long-tailed recognition

    Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16489–16498, 2021

    2021

  15. [23]

    Parametric contrastive learning

    Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 715–724, 2021

    2021

  16. [24]

    Learning imbalanced data with vision transformers

    Zhengzhuo Xu, Ruikang Liu, Shuo Yang, Zenghao Chai, and Chun Yuan. Learning imbalanced data with vision transformers. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15793–15803, 2023

    2023

  17. [25]

    The inaturalist species classification and detection dataset

    Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 87...

    2018

  18. [26]

    Wbcatt: a white blood cell dataset annotated with detailed morphological attributes.Advances in Neural Information Processing Systems, 36:50796–50824, 2023

    Satoshi Tsutsui, Winnie Pang, and Bihan Wen. Wbcatt: a white blood cell dataset annotated with detailed morphological attributes.Advances in Neural Information Processing Systems, 36:50796–50824, 2023

    2023

  19. [27]

    Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly- supervised classification and localization of common thorax diseases

    Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly- supervised classification and localization of common thorax diseases. InProceedings of the IEEE conference on co...

  20. [28]

    Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic).arXiv preprint arXiv:1902.03368, 2019

    Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging ...

    2018 arXiv

  21. [29]

    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, pag...

    2021

  22. [30]

    Adaptformer: Adapting vision transformers for scalable visual recognition.Advances in Neural Information Processing Systems, 35:16664–16678, 2022

    Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition.Advances in Neural Information Processing Systems, 35:16664–16678, 2022. 11

    2022

  23. [31]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  24. [32]

    Learning imbal- anced datasets with label-distribution-aware margin loss

    Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbal- anced datasets with label-distribution-aware margin loss. InAdvances in Neural Information Processing Systems, pages 1567–1578, 2019

    2019

  25. [33]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. InProceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017

    2017

  26. [34]

    Disentangling label distribution for long-tailed visual recognition

    Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6626–6636, 2021

    2021

  27. [35]

    Dualprompt: Complementary prompting for rehearsal-free continual learning

    Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. InEuropean Conference on Computer Vision, pages 631–648. Springer, 2022

    2022

  28. [36]

    Resclip: Residual attention for training- free dense vision-language inference

    Yuhang Yang, Jinhong Deng, Wen Li, and Lixin Duan. Resclip: Residual attention for training- free dense vision-language inference. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 29968–29978, 2025

    2025

  29. [37]

    Clearclip: Decomposing clip representations for dense vision-language inference

    Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Clearclip: Decomposing clip representations for dense vision-language inference. InEuropean Conference on Computer Vision, pages 143–160. Springer, 2024

    2024

  30. [38]

    Sinkhorn distances: Lightspeed computation of optimal transport.Advances in neural information processing systems, 26, 2013

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport.Advances in neural information processing systems, 26, 2013

    2013

  31. [39]

    Rademacher and gaussian complexities: Risk bounds and structural results.Journal of machine learning research, 3(Nov):463–482, 2002

    Peter L Bartlett and Shahar Mendelson. Rademacher and gaussian complexities: Risk bounds and structural results.Journal of machine learning research, 3(Nov):463–482, 2002

    2002

  32. [40]

    Analysis of representations for domain adaptation.Advances in neural information processing systems, 19, 2006

    Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation.Advances in neural information processing systems, 19, 2006

    2006

  33. [41]

    No Finding

    Alexander Long, Wei Yin, Thalaiyasingam Ajanthan, Vu Nguyen, Pulak Purkait, Ravi Garg, Alan Blair, Chunhua Shen, and Anton van den Hengel. Retrieval augmented classification for long-tail visual recognition. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2022

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.