Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Sanitizing Manufacturing Dataset Labels Using Vision-Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read AI merges 6,426 noisy manufacturing labels into 408

desk verdict A sensible CLIP+DBSCAN combination with a few good qualitative hits, but the evaluation is circular and the whole-image embeddings don't match the box-level labels, so the main claim doesn't hold. read the letter →

arxiv 2506.23465 v1 pith:NTN34IMB submitted 2025-06-30 cs.CV cs.AI

classification cs.CVcs.AI
keywords vision-languagemodelslabelnoisedatasetsanitizationCLIPDBSCANmulti-labelclassificationmanufacturingclustering
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 proposes VLSR, a framework that uses CLIP's shared image-text embedding space to clean noisy multi-label manufacturing image datasets. It computes cosine similarity between image and label embeddings to detect misspelled, irrelevant, or semantically weak labels and to suggest better labels from the dataset. It then clusters label embeddings with DBSCAN and merges small clusters, replacing each cluster with its most frequent label. Applied to the Factorynet dataset, the framework claims to have reduced the label vocabulary from 6,426 to 408 distinct labels and to have improved label consistency. The stated benefit is higher-quality training data for industrial machine learning with minimal human intervention.

What carries the argument

The load-bearing object is the CLIP vision-language model's shared embedding space, which maps images and textual labels to vectors so that cosine similarity $\text{sim}(a,b)=\frac{a\cdot b}{\|a\\|\|b\|}$ measures semantic alignment. Three uses of this similarity carry the argument: image-to-assigned-label comparison validates or rejects each original label; image-to-dataset comparison retrieves the best matching label from all labels in the dataset; and label-to-label cosine distance feeds DBSCAN, a density-based clustering algorithm that groups closely packed label embeddings and marks outliers. Within each cluster the most frequent label is chosen as representative, and clusters with few labels are merged into the nearest larger cluster by cosine distance.

What would settle it

Take a random sample of several hundred Factorynet images with both original and VLSR-refined labels, ask domain experts to judge which label better describes the object, and compute the proportion of refinements judged correct or improved. If the refined label is not judged at least as good as the original in the majority of cases, the framework's central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that cosine similarity between CLIP embeddings of images and their textual labels is a sufficient diagnostic for label quality in manufacturing datasets, and that density-based clustering of label embeddings can consolidate semantically equivalent labels. The paper supports this with qualitative examples on Factorynet showing nonsense labels, incorrect labels, partial labels, and misspellings replaced by more appropriate alternatives. It also reports a reduction from 6,426 distinct labels to 408 as evidence of improved consistency, achieved by grouping casing variants, synonyms, misspellings, and overly specific phrases into unified clusters with a frequency-based representative label.

Load-bearing premise

The entire pipeline assumes that CLIP cosine similarity between an image and a label accurately reflects whether that label correctly describes the object in the image for manufacturing items.

Editorial extensions

If this is right

  • If the framework works as claimed, noisy web-scraped and crowdsourced manufacturing datasets can be cleaned without per-image human review, lowering dataset curation cost.
  • The reported vocabulary reduction from 6,426 to 408 distinct labels means downstream classifiers face far fewer classes and more examples per class, which should improve training stability and performance.
  • The same pipeline extends to other multi-label domains where label noise is prevalent, since it depends only on CLIP embeddings and cosine similarity.
  • Clustering and representative selection standardize inconsistent naming, including casing, misspellings, and synonyms, making label semantics uniform across the dataset.
  • The framework can also surface labels that are more specific or more accurate than the original assigned label, not merely remove bad ones.

Reading between the lines

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

  • The 6,426-to-408 reduction is presented as success without quantitative validation that merged clusters preserve true object classes; a natural test is to measure downstream classifier accuracy or human agreement on a random sample of re-labeled images.
  • Because the final label per image is chosen by maximum image-label similarity, the framework may systematically prefer generic scene-level labels over the specific object in the image, depending on CLIP's biases; this could be checked by computing per-class precision and recall.
  • The approach could be adapted to an active-learning loop where only images with low confidence require human review, reducing manual effort further; the paper mentions adaptive thresholding as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes VLSR, a framework that uses CLIP to embed images and label texts, compute cosine similarities, and then (1) detect problematic labels and surface better labels from the dataset, and (2) cluster labels with DBSCAN and merge small clusters to reduce the label vocabulary. The method is evaluated on the Factorynet manufacturing dataset, reducing the distinct label count from 6,426 to 408. The paper reports qualitative examples of corrected labels and a table of example clusters, and concludes that VLSR successfully improves label consistency.

Significance. The problem of noisy labels in web-scraped and crowd-sourced manufacturing datasets is real and relevant. Using a vision-language model as a similarity oracle is a plausible direction, and the paper is among few attempts to apply CLIP to label cleaning specifically for multi-label datasets. The use of public models and data, and the clear description of the pipeline, are strengths. However, the current evidence is far from convincing: there is no quantitative evaluation, no baseline, and the whole-image versus box-level mismatch casts doubt on the correctness of the method itself. The paper does not ship code or annotated examples beyond a few figures, which limits reproducibility of the evaluation.

major comments (3)
  1. [§2.2, §3.1] The framework computes cosine similarity between whole-image embeddings and label embeddings (Eqs. 5 and 6), but the Factorynet dataset contains bounding-box annotations, as described in §3.1 and Figure 3. The method never crops or uses the box coordinates, so sim(e_i, e_label) measures whether the label describes the entire scene, not the object inside the box. This is directly visible in Figure 5(c): labels such as 'Saw', 'ratchet', and 'ear protector' are replaced by 'worker with circular saw', 'workshop with tools', and 'worker with handheld saw'. For a box-level label dataset, these replacements are category drift rather than refinement, and the central claim of successful label sanitization is therefore not supported.
  2. [§3.3, §3.4] The evaluation is entirely qualitative. There are no quantitative metrics such as error-detection precision/recall, no comparison against any baseline, and no validation on a labeled ground-truth subset. The headline reduction from 6,426 to 408 labels is not evidence of correctness: the paper states (§3.4) that the DBSCAN parameters (ε = 0.07, min_samples = 1) were chosen after evaluating multiple configurations and inspecting cluster outputs, and that parameter tuning requires human inspection of cluster outputs. Thus the 408-label result is an outcome of parameters selected on the test set, not a robust or principled finding.
  3. [§3.3, Figures 5 and 6] The evaluation is circular. The refined label L is chosen because it has the highest cosine similarity (Eq. 6), and the figures present as evidence that L has a higher similarity score than the original A. This is a direct consequence of the selection rule, so it cannot validate the correctness of the refinement. To establish that the framework improves label consistency, the authors would need an independent criterion, such as human agreement on a random sample of the 10,160 images, a downstream object-detection or classification benchmark using box-level labels, or a comparison with the original labels by a third-party annotator.
minor comments (5)
  1. [§3.1] The sentence 'The dataset consist of 10,160 images and 6,426 distinct labels' contains a subject-verb agreement error ('consist' should be 'consists').
  2. [§3.1] The phrase 'extracted form CSV files' appears to be a typo for 'extracted from CSV files'.
  3. [§3.3, Figure 6 caption] The caption of Figure 6 is incomplete: the phrase 'Improved Matches where A and (b) Best Matches' seems to be missing a verb or clause; the caption should be reworded for clarity.
  4. [§2.3] The merging threshold for small clusters (the 'predefined threshold' mentioned in the text) is never specified, nor is the definition of 'closest neighboring cluster' operationalized; this makes the clustering merging step hard to reproduce.
  5. [§3.4 vs Abstract] The abstract claims 'minimal human intervention', but §3.4 acknowledges that the clustering parameters require human inspection of cluster outputs; this tension should be clarified.

Circularity Check

3 steps flagged · score 6.0 of 10

Refined labels are selected as the argmax of the same CLIP cosine similarity used to declare success; the Factorynet sanitization results are therefore partly self-confirming.

  1. self definitional [Section 2.3 (final label selection) and Section 3.3 / Figures 5-6]
    "In this step, the cosine similarity between each representative label and its image was calculated using Eq. (6). Among the possible labels for an image, the one with the highest similarity score was used as the final label. ... Experimental results demonstrate that the VLSR framework successfully identifies problematic labels and improves label consistency."

    The refined label is defined as the label maximizing sim(e_i, e_j) via Eq. (6). Therefore every displayed 'refinement' L is selected by construction to have at least as high a cosine similarity as the original label A (and typically strictly higher). When Figures 5 and 6 present L with a higher similarity score than A as evidence of successful sanitization, they are restating the selection rule, not independently validating semantic correctness. The abstract's claim of 'successfully identifies problematic labels' is supported only by this same metric, with no held-out ground truth, human study, or box-level verification.

  2. fitted input called prediction [Section 3.4, DBSCAN hyperparameter selection and label-count reduction]
    "After evaluating multiple configurations, the most coherent groupings of labels that shared a clear semantic relationship were achieved with the ε = 0.07 and minimum number of samples set to 1. ... The method successfully reduced the total number of labels in the whole dataset from 6,426 to 408 distinct labels."

    The headline reduction is the output of ε and min_samples that the authors chose by inspecting the cluster outputs for 'coherent groupings.' No external benchmark or held-out label set is used to decide whether the merged clusters are correct, and the paper explicitly acknowledges that 'tuning these hyperparameters requires human inspection of cluster outputs.' Reporting the resulting reduction as evidence of improved consistency is therefore a fitted output, not a prediction: the same inspected clusters are used both to set the parameters and to demonstrate success.

1 more flagged steps
  1. other [Section 2.2.1 Eq. (5) vs. Section 3.1 Figure 3 (box-level annotations)]
    "In the first comparison, the embedding of each image ei was compared to each label embedding eij in the set Li. ... This comparison evaluates how well each label corresponds to the content of the image. ... Figure 3... showcases bounding box annotations along with their associated labels and sources."

    The dataset's labels are bounding-box annotations, but Eq. (5) embeds the entire uncropped image I_i. Thus 'how well each label corresponds to the content of the image' is actually 'how well the label describes the whole scene.' The claimed sanitization of object-level labels is therefore, by construction, a scene-level scoring operation. Its own examples confirm this: in Figure 5(c), 'ratchet' is replaced by 'workshop with tools' and 'ear protector' by 'worker with handheld saw,' i.e., the scoring function labels correct box-level annotations as wrong and promotes background/scene descriptions. The reported 'success' is then the success of the scene-level objective, not of object-label cleaning, so the evaluation does not address the task it claims to validate.

full rationale

The central evaluation is self-confirming. VLSR's refined labels are, by Eq. (6) and the selection rule in Section 2.3, the labels with maximum CLIP cosine similarity to the whole image. Showing that the selected label has a higher similarity than the original is a tautology: it merely restates the argmax operation. The clustering result (6,426 to 408 labels) is also a fitted output, since the DBSCAN parameters were chosen by inspecting cluster coherence and no held-out validation is reported. Additionally, because Factorynet labels are box-level while Eq. (5) uses whole-image embeddings, the sanitization objective is not even the object-label alignment the paper claims; the examples in Figure 5(c) show scene-level replacements such as 'workshop with tools' for 'ratchet.' There is no self-citation chain here, and CLIP itself is an external pretrained model, so the circularity is not about citation; it is that the reported improvement is measured with the same scoring function that generates the output. This warrants a moderate-to-high circularity score rather than a claim that no independent content exists. The method's pipeline is real and the clustering may reduce vocabulary, but the paper's evidence of 'successfully identifies problematic labels and improves label consistency' is not independent of the algorithm's own objective.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The framework relies on CLIP embeddings as a trusted semantic oracle, and the reported label reduction is a direct output of hand-tuned DBSCAN parameters. No external ground truth is used to validate either sanitization or clustering, so the main evidence is self-referential.

free parameters (3)
  • DBSCAN epsilon (eps) = 0.07
    Chosen by evaluating multiple configurations and picking the one with the most coherent clusters; directly determines cluster boundaries and the 6,426 to 408 reduction.
  • DBSCAN min_samples = 1
    Set to 1 after inspecting outputs; with min_samples=1 the clustering reduces to connected components at the epsilon threshold, so all labels get clustered and no noise remains.
  • Cluster merging threshold = not specified
    Small clusters are merged into the closest neighboring cluster based on a threshold that is described but never given, so the final label count depends on an unreported choice.
assumptions (5)
  • domain assumption Cosine similarity between CLIP image and text embeddings is a valid measure of semantic alignment between an image and a label.
    The entire sanitization pipeline (Eqs. 5 and 6) treats high similarity as correct and low similarity as erroneous, with no calibration against human ground truth.
  • domain assumption CLIP text embeddings place semantically equivalent labels close enough for DBSCAN to recover true label groups.
    Section 2.3 assumes that noun phrases with the same meaning cluster together in embedding space; the chosen epsilon then defines what 'same meaning' means.
  • domain assumption The most frequent label in a cluster is the best representative label.
    Section 2.3 selects the highest-frequency label as representative, assuming frequency implies correctness or appropriate granularity, which is untested.
  • domain assumption The Factorynet dataset's original labels are noisy enough that replacement by CLIP predictions improves quality.
    The paper states labels exhibit high error rates and uses examples to support this, but no error-rate measurement or downstream validation is given.
  • standard math DBSCAN and standard linear algebra can be applied as described.
    The paper uses standard definitions of cosine similarity and DBSCAN; these are unproblematic.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sanitizing Manufacturing Dataset Labels Using Vision-Language Models." pith.science (2026). https://pith.science/paper/NTN34IMB

@misc{pith2026250623465,
  author       = {Pith},
  title        = {Pith review of: Sanitizing Manufacturing Dataset Labels Using Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NTN34IMB}},
  note         = {Machine review of arXiv:2506.23465}
}
read the original abstract

The success of machine learning models in industrial applications is heavily dependent on the quality of the datasets used to train the models. However, large-scale datasets, specially those constructed from crowd-sourcing and web-scraping, often suffer from label noise, inconsistencies, and errors. This problem is particularly pronounced in manufacturing domains, where obtaining high-quality labels is costly and time-consuming. This paper introduces Vision-Language Sanitization and Refinement (VLSR), which is a vision-language-based framework for label sanitization and refinement in multi-label manufacturing image datasets. This method embeds both images and their associated textual labels into a shared semantic space leveraging the CLIP vision-language model. Then two key tasks are addressed in this process by computing the cosine similarity between embeddings. First, label sanitization is performed to identify irrelevant, misspelled, or semantically weak labels, and surface the most semantically aligned label for each image by comparing image-label pairs using cosine similarity between image and label embeddings. Second, the method applies density-based clustering on text embeddings, followed by iterative cluster merging, to group semantically similar labels into unified label groups. The Factorynet dataset, which includes noisy labels from both human annotations and web-scraped sources, is employed to evaluate the effectiveness of the proposed framework. Experimental results demonstrate that the VLSR framework successfully identifies problematic labels and improves label consistency. This method enables a significant reduction in label vocabulary through clustering, which ultimately enhances the dataset's quality for training robust machine learning models in industrial applications with minimal human intervention.

Figures

Figures reproduced from arXiv: 2506.23465 by the authors.

Figure 1
Figure 1. Overview of dataset sanitization method by the VLSR framework. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Overview of clustering and merging labels by the VLSR framework. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. An example image from the Factorynet dataset. This example showcases bounding box annotations [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Example of cosine similarity comparison, where embedding of each image is compared with the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Examples of label sanitization results. Each example compares the originally assigned label (A) with [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Examples of label sanitization results. Each example compares the originally assigned label (A) with [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Enhancing Automatic Modulation Recognition With a Reconstruction-Driven Vision Transformer Under Limited Labels

    cs.CV 2025-08 reject novelty 4.0 of 10

    A reconstruction-pretrained Vision Transformer reportedly recognizes modulation types with 15% labeled data at 71% accuracy, surpassing its fully supervised counterpart, though data handling errors undercut the claim.

Reference graph

Works this paper leans on

43 extracted references · 32 canonical work pages · cited by 1 Pith paper

  1. [1]

    Mohammadi, V

    A. Mohammadi, V. Hemmati, R. Ahmari, F. Owusu-Ambrose, M. Mahmoud, A. Homaifar, Gps spoofing attack detection on autonomous vehicles using modified dbscan with dynamic threshold, in: Proceedings of the 5th IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0/5.0 (ARCI’2025), 2025, pp. 74–76. 13

  2. [2]

    E. Ziad, Z. Yang, Y. Lu, F. Ju, Knowledge constrained deep clustering for melt pool anomaly detection in laser powder bed fusion, in: 2024 IEEE 20th International Conference on Automation Science and Engineering (CASE), IEEE, 2024, pp. 670–675

  3. [3]

    Jandaghi, X

    E. Jandaghi, X. Chen, C. Yuan, Motion dynamics modeling and fault detection of a soft trunk robot, in: 2023 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM), IEEE, 2023, pp. 1324–1329

  4. [4]

    Borazjani, P

    K. Borazjani, P. Abdisarabshali, F. Nadimi, N. Khosravan, M. Liwang, X. Wang, Y. Hong, S. Hosseinalipour, Multi-modal multi-task (m3t) federated foundation models for embodied ai: Potentials and challenges for edge integration, arXiv preprint arXiv:2505.11191 (2025)

  5. [5]

    Hosseinzadeh, M

    A. Hosseinzadeh, M. Shahin, M. Maghanaki, H. Mehrzadi, F. F. Chen, Minimizing waste- via novel fuzzy hybrid stacked ensembleof vision transformers and cnns to detect defects in metal surfaces, The International Journal of Advanced Manufacturing Technology (2024) 1–26

  6. [6]

    Safari, S

    K. Safari, S. Khalfalla, F. Imani, Dependency evaluation of defect formation and print- ing location in additive manufacturing, in: ASME International Mechanical Engineering Congress and Exposition, Vol. 86632, American Society of Mechanical Engineers, 2022, p. V02AT02A016

  7. [7]

    Rezvani Boroujeni, H

    S. Rezvani Boroujeni, H. Abedi, T. Bush, Enhancing glass defect detection with diffusion models: Addressing imbalanced datasets in manufacturing quality control, Computer and Decision Making – An International Journal (COMDEM) 2 (2025) 687–707. doi:10. 59543/comdem.v2i.14391. URL https://doi.org/10.59543/comdem.v2i.14391

  8. [8]

    E. Ziad, F. Ju, Z. Yang, Y. Lu, Pyramid learning based part-to-part consistency analysis in laser powder bed fusion, in: International Manufacturing Science and Engineering Con- ference, Vol. 88100, American Society of Mechanical Engineers, 2024, p. V001T01A024

Show all 43 references
  1. [9]

    Drost, M

    B. Drost, M. Ulrich, P. Bergmann, P. Hartinger, C. Steger, Introducing mvtec itodd - a dataset for 3d object recognition in industry, in: Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshops, 2017

  2. [10]

    G. Y. Lee, L. Alzamil, B. Doskenov, A. Termehchy, A survey on data cleaning methods for improved machine learning model performance (2021). arXiv:2109.07127. URL https://arxiv.org/abs/2109.07127

  3. [11]

    Fr´ enay, A

    B. Fr´ enay, A. Kab´ an, et al., A comprehensive introduction to label noise., in: ESANN, Citeseer, 2014

  4. [12]

    Frenay, M

    B. Frenay, M. Verleysen, Classification in the presence of label noise: A survey, IEEE Transactions on Neural Networks and Learning Systems 25 (5) (2014) 845–869. doi: 10.1109/TNNLS.2013.2292894

  5. [13]

    V. S. Sheng, F. Provost, P. G. Ipeirotis, Get another label? improving data quality and data mining using multiple, noisy labelers, in: Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’08, Association for Computing Machin...

  6. [14]

    W. Tang, M. Lease, Semi-supervised consensus labeling for crowdsourcing, in: SIGIR 2011 workshop on crowdsourcing for information retrieval (CIR), 2011, pp. 1–6

  7. [15]

    Khetan, Z

    A. Khetan, Z. C. Lipton, A. Anandkumar, Learning from noisy singly-labeled data, arXiv preprint arXiv:1712.04577 (2017)

  8. [16]

    X. Chen, A. Gupta, Webly supervised learning of convolutional networks, in: Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015

  9. [17]

    Schroff, A

    F. Schroff, A. Criminisi, A. Zisserman, Harvesting image databases from the web, IEEE Transactions on Pattern Analysis and Machine Intelligence 33 (4) (2011) 754–766. doi: 10.1109/TPAMI.2010.133

  10. [18]

    S. Li, J. Zhu, B. Chang, H. Wu, F. Xu, S. Zhong, Multi-label and evolvable dataset preparation for web-based object detection, ACM Transactions on Knowledge Discovery from Data 18 (9) (2024) 1–21

  11. [19]

    X. Chen, A. Gupta, Webly supervised learning of convolutional networks, in: Proceedings of the IEEE international conference on computer vision, 2015, pp. 1431–1439

  12. [20]

    Y. Yao, J. Zhang, F. Shen, L. Liu, F. Zhu, D. Zhang, H. T. Shen, Towards automatic construction of diverse, high-quality image datasets, IEEE Transactions on Knowledge and Data Engineering 32 (6) (2019) 1199–1211

  13. [21]

    S. K. Divvala, A. Farhadi, C. Guestrin, Learning everything about anything: Webly- supervised visual concept learning, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2014, pp. 3270–3277

  14. [22]

    Verma, A

    A. Verma, A. Sharbafchi, B. Touri, S. Mohajer, Distributed fact checking, in: 2023 IEEE International Symposium on Information Theory (ISIT), 2023, pp. 2649–2654. doi:10. 1109/ISIT54713.2023.10206599

  15. [23]

    Bernhardt, D

    M. Bernhardt, D. C. Castro, R. Tanno, A. Schwaighofer, K. C. Tezcan, M. Monteiro, S. Bannur, M. P. Lungren, A. Nori, B. Glocker, et al., Active label cleaning for improved dataset quality under resource constraints, Nature communications 13 (1) (2022) 1161

  16. [24]

    B. Han, Q. Yao, X. Yu, G. Niu, M. Xu, W. Hu, I. Tsang, M. Sugiyama, Co-teaching: Robust training of deep neural networks with extremely noisy labels, in: S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, R. Garnett (Eds.), Advances in Neural Information Proces...

  17. [25]

    X. Xia, T. Liu, B. Han, M. Gong, J. Yu, G. Niu, M. Sugiyama, Sample selection with uncertainty of losses for learning with noisy labels, arXiv preprint arXiv:2106.00445 (2021)

  18. [26]

    Y. Wang, X. Ma, Z. Chen, Y. Luo, J. Yi, J. Bailey, Symmetric cross entropy for robust learning with noisy labels, in: Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 322–330

  19. [27]

    S. Liu, J. Niles-Weed, N. Razavian, C. Fernandez-Granda, Early-learning regularization prevents memorization of noisy labels, Advances in neural information processing systems 33 (2020) 20331–20342. 15

  20. [28]

    W. Ni, K. Zhang, X. Miao, X. Zhao, Y. Wu, J. Yin, Iterclean: An iterative data clean- ing framework with large language models, in: Proceedings of the ACM Turing Award Celebration Conference-China 2024, 2024, pp. 100–105

  21. [29]

    X. Wang, H. Kim, S. Rahman, K. Mitra, Z. Miao, Human-llm collaborative annotation through effective verification of llm labels, in: Proceedings of the CHI Conference on Human Factors in Computing Systems, 2024, pp. 1–21

  22. [30]

    J. Li, A comparative study on annotation quality of crowdsourcing and llm via label aggregation, in: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2024, pp. 6525–6529

  23. [31]

    Wei, H.-T

    T. Wei, H.-T. Li, C.-S. Li, J.-X. Shi, Y.-F. Li, M.-L. Zhang, Vision-language models are strong noisy label detectors, arXiv preprint arXiv:2409.19696 (2024)

  24. [32]

    C. Feng, G. Tzimiropoulos, I. Patras, Clipcleaner: Cleaning noisy labels with clip, in: Proceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 876– 885

  25. [33]

    Radford, J

    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

  26. [34]

    Wolf, Transformers: State-of-the-art natural language processing, arXiv preprint arXiv:1910.03771 (2020)

    T. Wolf, Transformers: State-of-the-art natural language processing, arXiv preprint arXiv:1910.03771 (2020). URL https://github.com/huggingface/transformers

  27. [35]

    Ghiasvand, H

    S. Ghiasvand, H. E. Oskouie, M. Alizadeh, R. Pedarsani, Few-shot adversarial low-rank fine-tuning of vision-language models, arXiv preprint arXiv:2505.15130 (2025)

  28. [36]

    Ahmadi, S

    H. Ahmadi, S. E. Mahdimahalleh, A. Farahat, B. Saffari, Unsupervised time-series sig- nal analysis with autoencoders and vision transformers: A review of architectures and applications, arXiv preprint arXiv:2504.16972 (2025)

  29. [37]

    Khaniki, M

    M. Khaniki, M. Mirzaeibonehkhater, S. Fard, Class imbalance-aware active learning with vision transformers in federated histopathological imaging, JM Med Stu 2025 1 (2) (2025) 65–73

  30. [38]

    Adami, N

    B. Adami, N. Karimian, Gru-aunet: A domain adaptation framework for contactless fin- gerprint presentation attack detection, arXiv preprint arXiv:2504.01213 (2025)

  31. [39]

    Kermani, V

    A. Kermani, V. Perez-Rosas, V. Metsis, A systematic evaluation of llm strategies for mental health text analysis: Fine-tuning vs. prompt engineering vs. rag, arXiv preprint arXiv:2503.24307 (2025)

  32. [40]

    Ester, H.-P

    M. Ester, H.-P. Kriegel, J. Sander, X. Xu, et al., A density-based algorithm for discovering clusters in large spatial databases with noise, in: kdd, Vol. 96, 1996, pp. 226–231

  33. [41]

    Steinhaus, et al., Sur la division des corps mat´ eriels en parties, Bull

    H. Steinhaus, et al., Sur la division des corps mat´ eriels en parties, Bull. Acad. Polon. Sci 1 (804) (1956) 801

  34. [42]

    Lloyd, Least squares quantization in pcm, IEEE transactions on information theory 28 (2) (1982) 129–137

    S. Lloyd, Least squares quantization in pcm, IEEE transactions on information theory 28 (2) (1982) 129–137. 16

  35. [43]

    Bowman, E

    A. Bowman, E. Braham, Factorynet hackathon dataset 2024 (Aug. 2024). doi:10.5281/ zenodo.13332887. URL https://doi.org/10.5281/zenodo.13332887 17

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.