Pith. sign in

REVIEW 3 major objections 4 minor 59 references

Active Learning via Vision-Language Model Adaptation with Open Data

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Retrieving a vision-language model's own pretraining data to augment active learning, then labeling the rarest classes first, yields large accuracy gains on five benchmarks.

desk verdict Useful CT-vs-PT comparison and a practical RDA result, but the headline +7% gain is confounded — TFS's own effect is well under a point when adaptation is held fixed. read the letter →

arxiv 2506.01724 v1 pith:UXJOZYWM submitted 2025-06-02 cs.CV

Tong Wang , Jiaqi Wang , Shu Kong This is my paper · ORCID
classification cs.CV
keywords activelearningvision-languagemodelsretrieval-baseddataaugmentationtail-firstsamplingcontrastivetuninglong-taileddistributionsopenLAION-400M
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

This paper seeks to show that active learning with a vision-language model is materially improved by using the model's own public pretraining data, not just the model itself. It retrieves task-relevant images from LAION-400M by string-matching class names, augments the labeled set with them, and observes that the retrieved data is long-tailed across classes. That observation motivates Tail First Sampling (TFS), which spends each labeling round on the current most-underrepresented class, choosing the most uncertain example among those the model predicts as that class. The paper also compares four VLM adaptation methods and finds that contrastive tuning of both encoders consistently beats prompt tuning, linear probing, and ordinary finetuning. Combined, retrieval augmentation, tail-first sampling, and contrastive tuning outperform prior active-learning methods by about 7.2 accuracy points on average across five benchmarks.

What carries the argument

The machinery has three parts. First, retrieval-based data augmentation (RDA) string-matches task class names against captions in LAION-400M, fetches the corresponding images, filters them by image-text feature similarity, and caps each class at 500 images to balance the set. Second, Tail First Sampling (TFS) runs the adapted VLM over unlabeled data, finds the class with the fewest currently labeled examples, and selects the highest-entropy example among those pseudo-labeled as that rare class for oracle labeling. Third, contrastive tuning (CT) updates both the visual and text encoders with a contrastive loss, the same objective used to pretrain the VLM. The load-bearing mechanism is the informational link from retrieved data's class frequencies to the unlabeled pool's class frequencies; TFS converts that link into a labeling policy.

What would settle it

Compute per-class image counts in the retrieved set and in the unlabeled task pool for a dataset such as Semi-Aves, then compare the two orderings; if the rank correlation is close to zero or negative, the premise that retrieved data reveals which classes are underrepresented in the task pool is false, and TFS's advantage should disappear.

Watch

Extended reading notes

Core claim

The central claim is that open data is a near-free source of signal for active learning: the retrieval-based augmentation pipeline (RDA) turns the vision-language model's pretraining corpus into a stream of task-relevant training examples, and the long-tailed frequencies of those retrieved examples reveal which classes the model handles poorly. The authors argue that this retrieved distribution mirrors the unlabeled task-specific pool, so TFS can identify rare classes before an oracle labels anything. In controlled comparisons, contrastive tuning, which updates both image and text encoders with the pretraining objective, outperforms prompt tuning, linear probing, and finetuning in every active-learning setting they test. The assembled method, ALOR, is claimed to exceed prior active-learning methods by about 7.2 accuracy points on average and similar macro-F1 gains, with per-class analysis showing the gain comes substantially from rapid improvement on underrepresented classes.

Load-bearing premise

The load-bearing premise is that the class distribution of images retrieved from LAION-400M by class-name matching matches the class distribution of the unlabeled task-specific pool, so the rare classes revealed by retrieval are the right ones for the active learner to label.

Editorial extensions

If this is right

  • All compared active-learning methods, including entropy, coreset, BADGE, PCB+BADGE, ALFA-Mix, and LoGo, improve when retrieval augmentation is added, with more than 8 accuracy points gained by the final round under prompt tuning alone.
  • Tail First Sampling outperforms existing active-learning methods under prompt tuning and improves further under contrastive tuning, suggesting that rare-class-first labeling is a strong default choice for VLM-based active learning.
  • Contrastive tuning beats prompt tuning, linear probing, and finetuning on the Semi-Aves dataset regardless of whether retrieved data is used and regardless of the selection strategy, challenging the current practice of defaulting to prompt tuning in active learning.
  • The final ALOR method, combining RDA, TFS, and CT, achieves roughly 7.2 accuracy points improvement over prior methods across five benchmark datasets, with the biggest per-class gains appearing on underrepresented classes.
  • Per-round analysis shows that RDA gives a 1.7x accuracy boost at round zero, before any active selection has happened, indicating that retrieved open data can substitute for a substantial amount of initial labeling effort.

Reading between the lines

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

  • Extension beyond the paper: a direct stress test of the load-bearing assumption is to compute the rank correlation between per-class retrieved counts from LAION-400M and per-class frequencies in each benchmark's unlabeled pool; if the orderings barely agree, TFS's mechanism is not what drives the gains.
  • Extension beyond the paper: since contrastive tuning wins even without RDA, a minimal recipe for VLM-based active learning is contrastive tuning plus any standard selector; an ablation separating the extra-data effect of RDA from its class-rarity signal would clarify where the improvement comes from.
  • Extension beyond the paper: the tail-first principle does not depend on LAION specifically, so any cheap signal of class prevalence, such as caption frequencies, search-engine hit counts, or label-name counts in a large corpus, should predict a similar labeling policy helps on imbalanced pools.
  • Extension beyond the paper: the paper's own limitation note suggests TFS can get trapped on one rare class when the budget is tiny, so combining rarity-first selection with a diversity constraint is a natural next step that would make the method more robust.
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 / 4 minor

Summary. The paper proposes ALOR, an active learning method for vision-language models that combines three components: retrieval-based data augmentation (RDA) from the VLM's pretraining corpus (LAION-400M), a new acquisition rule called Tail First Sampling (TFS) that prioritizes labeling examples from under-represented classes, and contrastive tuning (CT) as the adaptation strategy. The authors report that RDA improves existing AL methods, that CT outperforms prompt tuning, linear probing, and finetuning for AL, and that the full ALOR pipeline achieves roughly +7.2 average accuracy over prior methods on five fine-grained benchmarks. The paper includes per-seed results, an open-source code release, and a candid limitations section.

Significance. If the central claims are cleanly supported, the paper makes a useful contribution: it is one of the first to bring open web-scale data into the AL loop, it provides a systematic comparison of adaptation approaches for VLM-based AL, and it proposes an extremely simple acquisition rule (TFS) that could be valuable in practice. The empirical effort is substantial: five benchmarks, three seeds, per-round and per-class analyses, and a supplement with full seed-level tables. The authors also honestly flag the domain-gap and pseudo-label bias risks. However, the headline claim that TFS is responsible for the large gains is not currently isolated from the choice of adaptation method and the use of retrieved data, and the algorithmic description of TFS has an ambiguity that affects reproducibility.

major comments (3)
  1. [Algorithm 1] The headline comparison in Table 2 confounds the acquisition rule with the adaptation method and the use of retrieved data. ALOR ('TFS w/ CT' with RDA) is compared against prior methods that all use prompt tuning (PT). Tables 3 and 4 show that CT alone gives large gains over PT: +5.65 accuracy on Semi-Aves Round-0 with RDA (Table 3) and +2.74 on average at Round-6 without RDA (Table 4). Therefore the reported +7.2-point gain over PCB+BADGE and LoGo is not attributable to the proposed TFS acquisition rule; it could come entirely from switching to CT and adding RDA. The only experiment that holds adaptation and RDA fixed is Table 1, where TFS achieves 63.52 average accuracy versus 63.26 for the best prior method (PCB+BADGE), a difference of 0.26 points, and Table 1 does not report standard deviations. The paper needs an experiment that varies only the acquisition rule (e.g., Entropy/PCB/BADGE/LoGo and TFS all under CT with RDA, using identical retrieved data and identical adaptation budgets) before claiming TFS significantly outperforms prior AL methods.
  2. [Section 3.2] Algorithm 1 is ambiguous about the model used for selection within a round. The entropy e_i and pseudo-labels yhat_i are computed once over U before the inner loop, but after each single-sample acquisition the model is adapted via 'M_r = adapt(M_{r-1}, L_r)' while the remaining e_i and yhat_i are not recomputed. As a result, for the second and later samples in a round, the acquisition scores come from a stale model, not from the adapted model that the pseudocode appears to imply. This matters because TFS's entire mechanism depends on which model produces the pseudo-labels that identify the tail class. The authors should either recompute e_i and yhat_i after each adaptation, or state explicitly that all K samples in a round are selected using the model from the end of the previous round (with the inner-loop adaptation being only for the next round). The current pseudocode is not reproducible as written.
  3. [Section 4.1, Table 2] The core assumption behind TFS — that the class distribution of retrieved LAION-400M data reflects the class distribution of the task-specific unlabeled pool — is stated but not tested. The paper argues that 'as VLM's pretraining data is sampled from the real world, the unlabeled task-specific data likely follows similar imbalanced distributions.' This is plausible but not self-evident, especially for benchmarks like Semi-Aves or Aircraft where the task-specific pool is a curated fine-grained dataset. If the retrieved distribution is not indicative of the unlabeled pool, TFS could systematically sample from classes that are actually well-represented in the target data, and the modest 0.26-point gain in Table 1 might become negative on other benchmarks. The authors should directly compare the per-class frequency of the retrieved data with the per-class frequency of the unlabeled pool (e.g., using the VLM's own predictions or a small labeled probe) to validate this assumption, or at least discuss what happens when the assumption fails.
minor comments (4)
  1. [Section 4] Table 2 reports accuracy and F1 after Round-6 but does not include standard deviations, despite the protocol statement that results are averaged over three random runs with standard deviations. Since the differences among PT-based methods in Table 1 are small (e.g., TFS 63.52 vs. PCB+BADGE 63.26), the absence of variance information makes it impossible to judge whether even the 0.26-point difference is meaningful.
  2. [Section 4.1] The main text states that Round-0 provides one labeled example per class, but earlier in the same section it says the initial labeled data is 'randomly sampled' from the training set. The supplement clarifies that initialization samples one image per class with fixed seeds (666, 777, 888). Please align the main-text wording with the supplement.
  3. [References] There are several typos and minor inconsistencies: 'depeneding' (Section 4), 'reconrfirm' (Section 4.1), 'bencnhmark' (supplement Section A), 'Linear Probing (PL)' should be 'LP' (Section 4.1), and reference [9] appears twice in the bibliography. These do not affect the technical content but should be fixed in revision.
  4. [Section 4, Implementation Details] The limitations paragraph mentions that TFS 'can be trapped into tail classes if the budget is too small and imbalance ratio is too high.' This is a valid concern, and it would strengthen the paper to show a sensitivity analysis on the per-round budget or on the imbalance ratio of the retrieved data, since the current experiments use a fixed budget equal to the number of classes.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the empirical gains are measured on held-out test splits, and the proposed TFS heuristic is not fitted to those labels.

full rationale

The paper's derivation chain is empirically self-contained rather than circular. The central comparisons in Tables 1 and 2 are evaluated on held-out validation/test splits after active selection from training pools, so no quantity is 'predicted' from data used to fit it. TFS is a heuristic that selects the unlabeled example with largest entropy among those pseudo-labeled as the current rarest class; using model pseudo-labels within an active-learning loop is standard self-referential training, not a reduction of the claimed result to its inputs. The with/without RDA comparison in Table 1 independently demonstrates the retrieval augmentation benefit, and the CT-versus-PT comparison in Tables 3 and 4 is a separate ablation. The only author-overlapping citations are the adoption of string-matching RDA from prior work by the same group ([36, 32]); that is an implementation choice, not a load-bearing justification of the target result, and its contribution is verified by the paper's own controlled experiments. No uniqueness theorem, forbidden alternative, or fitted parameter is imported from self-citation. The skeptic's concern that the +7.2% gain is attributable to CT and RDA rather than TFS is a confounding/attribution concern about experimental isolation, not a circularity of the derivation chain.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that open-data distributions mirror task-specific distributions, plus several chosen hyperparameters; no fitted free parameters in a derivation, but the empirical results depend on these choices.

free parameters (6)
  • per-class retrieved data cap = 500
    Selected via a study on Semi-Aves (Table 6); controls balance of retrieved data and affects all experiments.
  • training epochs = 50 with RDA, 200 without
    Set to match prior work; may change relative ranking of adaptation approaches.
  • learning rate for backbone = 1e-6
    Used for AdamW during CT/FT/LP; affects performance.
  • learning rate for classifier = 1e-4
    Classifier head learning rate.
  • temperature learning rate = 1e-4
    For contrastive tuning; not standard in all adapters.
  • retrieval similarity threshold = not specified
    Used to filter outliers after string matching; the concrete threshold is not reported.
assumptions (4)
  • domain assumption String-matched LAION-400M images are relevant to the downstream task classes.
    Adopted from prior RDA work; used without per-task validation.
  • domain assumption The unlabeled task-specific pool has similar class imbalance to the retrieved data.
    Stated in Section 3.2 as the basis for TFS; no direct measurement on the task pools is provided.
  • domain assumption Pseudo-labels from the adapted VLM are sufficiently reliable to identify tail-class candidates.
    Acknowledged in the Remark and Limitations; the paper offers an empirical conjecture that even wrong pseudo-labels select useful hard examples.
  • domain assumption The five benchmark datasets are representative of AL use cases.
    Used for evaluation; standard in the literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Active Learning via Vision-Language Model Adaptation with Open Data." pith.science (2026). https://pith.science/paper/UXJOZYWM

@misc{pith2026250601724,
  author       = {Pith},
  title        = {Pith review of: Active Learning via Vision-Language Model Adaptation with Open Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UXJOZYWM}},
  note         = {Machine review of arXiv:2506.01724}
}
read the original abstract

Pretrained on web-scale open data, VLMs offer powerful capabilities for solving downstream tasks after being adapted to task-specific labeled data. Yet, data labeling can be expensive and may demand domain expertise. Active Learning (AL) aims to reduce this expense by strategically selecting the most informative data for labeling and model training. Recent AL methods have explored VLMs but have not leveraged publicly available open data, such as VLM's pretraining data. In this work, we leverage such data by retrieving task-relevant examples to augment the task-specific examples. As expected, incorporating them significantly improves AL. Given that our method exploits open-source VLM and open data, we refer to it as Active Learning with Open Resources (ALOR). Additionally, most VLM-based AL methods use prompt tuning (PT) for model adaptation, likely due to its ability to directly utilize pretrained parameters and the assumption that doing so reduces the risk of overfitting to limited labeled data. We rigorously compare popular adaptation approaches, including linear probing (LP), finetuning (FT), and contrastive tuning (CT). We reveal two key findings: (1) All adaptation approaches benefit from incorporating retrieved data, and (2) CT resoundingly outperforms other approaches across AL methods. Further analysis of retrieved data reveals a naturally imbalanced distribution of task-relevant classes, exposing inherent biases within the VLM. This motivates our novel Tail First Sampling (TFS) strategy for AL, an embarrassingly simple yet effective method that prioritizes sampling data from underrepresented classes to label. Extensive experiments demonstrate that our final method, contrastively finetuning VLM on both retrieved and TFS-selected labeled data, significantly outperforms existing methods.

Figures

Figures reproduced from arXiv: 2506.01724 by the authors.

Figure 2
Figure 2. Retrieved data from a VLM’s pretraining dataset (i.e., LAION-400M [44]) follows long-tailed distributions w.r.t classes concerned by specific tasks. They imply how the VLM is potentially biased as being trained on such data. They also indicate how unlabeled task-specific data is similarly imbalanced distributed. This motivates our simple and novel Tail-First Sampling (TFS) strategy that prioritizes sampling data for… view at source ↗
Figure 4
Figure 4. Conceptual comparison between four different adaptation approaches. (a) Linear Probing (LP) learns a linear classifier on top of the frozen pretrained visual encoder of a VLM. (b) Typical finetuning (FT) learns to update parameters of the pretrained visual encoder. (c) Prompt Tuning (PT) learns parameters in the input space which are concatenated with text prompt over pretrained visual and text encoders. Note that P… view at source ↗
Figure 5
Figure 5. RDA significantly boosts active learning methods. For each method in each round, we report the averaged accuracy and standard deviation over three random runs across five datasets. In Round-0, methods with RDA achieve 1.7× higher accuracy than without! In the last round (Round-6), RDA helps each method obtain >8% accuracy gains. reader to the supplement (Section A) for details of these datasets. In each dataset, we … view at source ↗
Figures from the paper (5 more)
Figure 6
Figure 6. Figure 6: Comparison of different AL methods w.r.t per-round accuracy. For each method in each round, we report its averaged accuracy with standard devia￾tion (reflected by the shallow) over three random runs across five different datasets. All methods start from the same pretra…
Figure 7
Figure 7. Figure 7: Visualization of per-round (x-axis) per-class (y-axis) accuracies on semi-Aves benchmark. For each method, we sort its per-class accuracies in round-0 and track the accuracies over time. Results of different methods with and without RDA are shown in the top and bottom …
Figure 6
Figure 6. Figure 6: Results show that our TFS consistently outperforms the compared methods, demonstrating [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: CT consistently outperforms other adap￾tation approaches (PT, LP, and FT). To demonstrate this, we apply each adaptation approach to the com￾pared AL methods listed in [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Examples of task-specific data and retrieved data. We showcase task-specific data from our benchmark datasets alongside their retrieved data from LAION-400M [44]. Comparative analysis reveals significant domain discrepancies between task-specific data and retrieved dat…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 54 canonical work pages

  1. [1]

    Gone fishing: Neural active learning with fisher embeddings

    Jordan Ash, Surbhi Goel, Akshay Krishnamurthy, and Sham Kakade. Gone fishing: Neural active learning with fisher embeddings. In Advances in Neural Information Processing Systems (NeurIPS), volume 34, pages 8927–8939, 2021

    work page 2021

  2. [2]

    Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal

    Jordan T. Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. In International Conference on Learning Representations (ICLR), 2020

    work page 2020

  3. [3]

    Active prompt learning in vision language models

    Jihwan Bang, Sumyeong Ahn, and Jae-Gil Lee. Active prompt learning in vision language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27004–27014, 2024

    work page 2024

  4. [4]

    Retrieval-augmented diffusion models

    Andreas Blattmann, Robin Rombach, Kaan Oktay, Jonas Müller, and Björn Ommer. Retrieval-augmented diffusion models. In Advances in Neural Information Processing Systems (NeurIPS), volume 35, 2022

    work page 2022

  5. [5]

    Food-101 - mining discriminative components with random forests

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), 2014

    work page 2014

  6. [6]

    Visualgpt: Data-efficient adaptation of pretrained language models for image captioning

    Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18030–18040, 2022

    work page 2022

  7. [7]

    Re-imagen: Retrieval-augmented text-to-image generator

    Wenhu Chen, Hexiang Hu, Chitwan Saharia, and William W Cohen. Re-imagen: Retrieval-augmented text-to-image generator. In International Conference on Learning Representations (ICLR), 2023

    work page 2023

  8. [9]

    Reproducible scaling laws for contrastive language-image learning

    Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818–2829, 2023

    work page 2023

Show all 59 references
  1. [10]

    Jiwoong Choi, Ismail Elezi, Hyuk-Jae Lee, Clement Farabet, and Jose M. Alvarez. Active learning for deep object detection via probabilistic modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10264–10273, October 2021

    2021

  2. [11]

    Deep bayesian active learning with image data

    Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In International Conference on Machine Learning (ICML), pages 1183–1192. PMLR, 2017

    2017

  3. [12]

    Finetune like you pretrain: Improved finetuning of zero-shot vision models

    Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19338–19347, 2023

    2023

  4. [13]

    From images to textual prompts: Zero-shot visual question answering with frozen large language models

    Jiaxian Guo, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Boyang Li, Dacheng Tao, and Steven Hoi. From images to textual prompts: Zero-shot visual question answering with frozen large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern R...

    2023

  5. [14]

    Retrieval augmented language model pre-training

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. Retrieval augmented language model pre-training. In International Conference on Machine Learning (ICML), volume 119, pages 3929–3938, 2020

    2020

  6. [15]

    Using trusted data to train deep networks on labels corrupted by severe noise

    Dan Hendrycks, Mantas Mazeika, Duncan Wilson, and Kevin Gimpel. Using trusted data to train deep networks on labels corrupted by severe noise. In Advances in Neural Information Processing Systems (NeurIPS), 2018

    2018

  7. [16]

    Entropy-based active learning for object recognition

    Alex Holub, Pietro Perona, and Michael C Burl. Entropy-based active learning for object recognition. In 2008 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1–8. IEEE, 2008

    2008

  8. [17]

    Active learning by learning

    Wei-Ning Hsu and Hsuan-Tien Lin. Active learning by learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 29, 2015. 10

    2015

  9. [18]

    Active learning by querying informative and representative examples

    Sheng-Jun Huang, Rong Jin, and Zhi-Hua Zhou. Active learning by querying informative and representative examples. In Advances in Neural Information Processing Systems (NeurIPS), volume 23, 2010

    2010

  10. [19]

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

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning (ICML), 2021

    2021

  11. [20]

    Maple: Multi-modal prompt learning

    Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19113–19122, 2023

    2023

  12. [21]

    Re-thinking federated active learning based on inter-class diversity

    SangMook Kim, Sangmin Bae, Hwanjun Song, and Se-Young Yun. Re-thinking federated active learning based on inter-class diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3944–3953, 2023

    2023

  13. [22]

    Batchbald: Efficient and diverse batch acquisition for deep bayesian active learning

    Andreas Kirsch, Joost Van Amersfoort, and Yarin Gal. Batchbald: Efficient and diverse batch acquisition for deep bayesian active learning. In Advances in Neural Information Processing Systems (NeurIPS) , volume 32, 2019

    2019

  14. [23]

    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 Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2013

    2013

  15. [24]

    Fine-tuning can distort pretrained features and underperform out-of-distribution

    Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In International Conference on Learning Representations (ICLR), 2022

    2022

  16. [25]

    Retrieval- augmented generation for knowledge-intensive nlp tasks

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval- augmented generation for knowledge-intensive nlp tasks. In Advances in Neura...

    2020

  17. [26]

    Internet explorer: Targeted representation learning on the open web

    Alexander Cong Li, Ellis Langham Brown, Alexei A Efros, and Deepak Pathak. Internet explorer: Targeted representation learning on the open web. In International Conference on Machine Learning (ICML) , volume 202, pages 19385–19406, 2023

    2023

  18. [27]

    Align before fuse: Vision and language representation learning with momentum distillation

    Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural Information Processing Systems (NeurIPS), volume 34, pages ...

    2021

  19. [28]

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

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning (ICML), pages 12888–12900. PMLR, 2022

    2022

  20. [29]

    Active learning with multi-label svm classification

    Xin Li and Yuhong Guo. Active learning with multi-label svm classification. In International Joint Conference on Artificial Intelligence (IJCAI), volume 13, pages 1479–1485, 2013

    2013

  21. [30]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19325–19337, 2023

    2023

  22. [31]

    Learn- ing customized visual models with retrieval-augmented knowledge

    Haotian Liu, Kilho Son, Jianwei Yang, Ce Liu, Jianfeng Gao, Yong Jae Lee, and Chunyuan Li. Learn- ing customized visual models with retrieval-augmented knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15148–15158, June 2023

    2023

  23. [32]

    Few-shot recognition via stage-wise retrieval- augmented finetuning

    Tian Liu, Huixin Zhang, Shubham Parashar, and Shu Kong. Few-shot recognition via stage-wise retrieval- augmented finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  24. [33]

    Influence selection for active learning

    Zhuoming Liu, Hao Ding, Huaping Zhong, Weijia Li, Jifeng Dai, and Conghui He. Influence selection for active learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9274–9283, 2021

    2021

  25. [34]

    Fine-grained visual classification of aircraft

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013

    2013 arXiv

  26. [35]

    A study of the effect of different types of noise on the precision of supervised learning techniques

    David F Nettleton, Albert Orriols-Puig, and Albert Fornells. A study of the effect of different types of noise on the precision of supervised learning techniques. Artificial intelligence review, 33:275–306, 2010. 11

    2010

  27. [36]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12988–12997, 2024

    2024

  28. [37]

    Cats and dogs

    Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3498–3505. IEEE, 2012

    2012

  29. [38]

    Active learning by feature mixing

    Amin Parvaneh, Ehsan Abbasnejad, Damien Teney, Gholamreza Reza Haffari, Anton Van Den Hengel, and Javen Qinfeng Shi. Active learning by feature mixing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12237–12246, 2022

    2022

  30. [39]

    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. In International Conference on Machine Learning (IC...

    2021

  31. [40]

    A survey of deep active learning

    Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning. ACM computing surveys (CSUR), 54(9):1–40, 2021

    2021

  32. [41]

    Margin-based active learning for structured output spaces

    Dan Roth and Kevin Small. Margin-based active learning for structured output spaces. In In European Conference on Machine Learning (ECML), pages 413–424. Springer, 2006

    2006

  33. [42]

    Consistency-guided prompt learning for vision-language models

    Shuvendu Roy and Ali Etemad. Consistency-guided prompt learning for vision-language models. In International Conference on Learning Representations (ICLR), 2024

    2024

  34. [43]

    Bardia Safaei and Vishal M. Patel. Active learning for vision language models. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 4902–4912, February 2025

    2025

  35. [44]

    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

    2021 arXiv

  36. [45]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    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 generation image-text models. In Advances in Neural Info...

    2022

  37. [46]

    Active learning for convolutional neural networks: A core-set approach

    Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations (ICLR), 2018

    2018

  38. [47]

    Active learning literature survey

    Burr Settles. Active learning literature survey. University of Wisconsin-Madison Department of Computer Sciences, 2009

    2009

  39. [48]

    Get another label? improving data quality and data mining using multiple, noisy labelers

    Victor S Sheng, Foster Provost, and Panagiotis G Ipeirotis. Get another label? improving data quality and data mining using multiple, noisy labelers. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining (SIGKDD), pages 614–622, 2008

    2008

  40. [49]

    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 preprint arXiv:2103.06937, 2021

    2021 arXiv

  41. [50]

    Alpha-clip: A clip model focusing on wherever you want

    Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alpha-clip: A clip model focusing on wherever you want. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13019–13029, 2024

    2024

  42. [51]

    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. In Advances in Neural Information Processing Systems (NeurIPS), volume 36, pages 65566–65584, 2023

    2023

  43. [52]

    A new active labeling method for deep learning

    Dan Wang and Yi Shang. A new active labeling method for deep learning. InInternational Joint Conference on Neural Networks (IJCNN), pages 112–119. IEEE, 2014

    2014

  44. [53]

    Cogvlm: Visual expert for pretrained language models

    Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Song XiXuan, et al. Cogvlm: Visual expert for pretrained language models. In Advances in Neural Information Processing Systems (NeurIPS), volume 37, pages 121475–121499, 2024

    2024

  45. [54]

    Debiased learning from naturally imbalanced pseudo-labels

    Xudong Wang, Zhirong Wu, Long Lian, and Stella X Yu. Debiased learning from naturally imbalanced pseudo-labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14647–14657, 2022. 12

    2022

  46. [55]

    Aligning pretraining for detection via object-level contrastive learning

    Fangyun Wei, Yue Gao, Zhirong Wu, Han Hu, and Stephen Lin. Aligning pretraining for detection via object-level contrastive learning. Advances in Neural Information Processing Systems (NeurIPS) , 34: 22682–22694, 2021

    2021

  47. [56]

    Using active learning to expand training data for implicit discourse relation recognition

    Yang Xu, Yu Hong, Huibin Ruan, Jianmin Yao, Min Zhang, and Guodong Zhou. Using active learning to expand training data for implicit discourse relation recognition. In Proceedings of Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 725–731, 2018

    2018

  48. [57]

    Multi-class active learning by uncertainty sampling with diversity maximization

    Yi Yang, Zhigang Ma, Feiping Nie, Xiaojun Chang, and Alexander G Hauptmann. Multi-class active learning by uncertainty sampling with diversity maximization. International Journal of Computer Vision (IJCV), 113:113–127, 2015

    2015

  49. [58]

    Learning loss for active learning

    Donggeun Yoo and In So Kweon. Learning loss for active learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

    2019

  50. [59]

    Conditional prompt learning for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16816–16825, 2022

    2022

  51. [60]

    727-200" (Aircraft dataset) retrieves an eBay product listing titled

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 130(9):2337–2348, 2022. 13 Active Learning via Vision-Language Model Adaptation with Open Data (Supplemental Document) T...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.