Pith. sign in

REVIEW 2 major objections 5 minor 57 references

Any-Class Presence Likelihood for Robust Multi-Label Classification with Abundant Negative Data

T0 review · 2 major / 5 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 This paper claims that augmenting standard multi-label losses with an any-class presence likelihood—derived as a normalized weighted geometric mean of class probabilities—improves F1, F2, and mAP on datasets dominated by negative…

desk verdict Useful loss tweak with consistent gains; the 'any-class likelihood' name oversells a geometric-mean heuristic, but the empirics are solid enough to referee. read the letter →

arxiv 2506.05721 v1 pith:EELOVIPY submitted 2025-06-06 cs.LG cs.AIcs.CV

Dumindu Tissera , Omar Awadallah , Muhammad Umair Danish , Ayan Sadhu , Katarina Grolinger This is my paper
classification cs.LGcs.AIcs.CV
keywords multi-labelclassificationnegativedataany-classpresencelikelihoodgeometricmeanfocallossclass-balancedimbalancedlearningfunctiondesign
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

Multi-label classifiers trained on datasets where most instances carry no positive label can be overwhelmed by those negative examples, suppressing the learning signal for the positive ones. This paper proposes adding an explicit auxiliary objective: the likelihood that at least one class is present, computed as a normalized weighted geometric mean of per-class sigmoid probabilities, with weight 1 for present classes and a small weight λ for absent ones. Combining this term with Binary Cross-Entropy and Focal Loss, and extending class-balanced reweighting so that negative instances form an explicit category, consistently improves F1, F2, and mAP across SewerML, modified COCO, and ChestX-ray14. The reported gains reach 6.01 percentage points in F1, 8.06 in F2, and 3.11 in mean average precision, all without extra parameters or added inference complexity. A reader should care because this is a drop-in loss modification that claims to work across different architectures and application domains.

What carries the argument

The load-bearing object is the normalized weighted geometric mean of Eq. (4), which the paper calls the any-class presence probability $p_a$. It is a bounded, differentiable aggregation of per-class sigmoid outputs that stands in for the probability that at least one class is present, and the paper shows it equals the sigmoid of the weighted average of the logits. The hyperparameter $\lambda$ (set to 0.02 in most experiments) controls how much absent-class probabilities contribute to $p_a$ for positive instances, and the class-balanced extension in Eqs. (11)–(14) treats negative instances as an additional category with its own effective sample count. This construction carries the argument by converting a missing objective—collective presence awareness—into a closed-form loss term that requires no network changes and no inference-time cost.

What would settle it

On a synthetic multi-label dataset with known label correlations and a controllable fraction of all-negative instances, train the same backbone with the proposed any-class loss and with an auxiliary loss based on the exact union probability $1 - \prod_j (1-p_j)$; if the geometric-mean variant yields worse positive-instance recall or higher false positives than the exact-union variant, the proxy itself, not the presence signal, is the limiting factor.

Watch

Extended reading notes

Core claim

The central claim is that optimizing an any-class presence likelihood alongside standard per-class objectives improves multi-label classification when negative data dominate. The any-class presence probability is defined in Eq. (4) as $p_a = \frac{(\prod_j p_j^{w_j})^{1/\sum_j w_j}}{(\prod_j p_j^{w_j})^{1/\sum_j w_j} + (\prod_j (1-p_j)^{w_j})^{1/\sum_j w_j}}$, with $w_j=1$ for present classes and $w_j=\lambda$ for absent classes. The paper shows that this aggregate simplifies to the sigmoid of the weighted mean of the logits, so it behaves like a proper probability, and the redesigned losses in Eqs. (8) and (10) add only an $\alpha\log(p_t^a)$ term (with a focal re-weighting in the focal variant). Appendix A.1 derives the resulting gradient: each output neuron receives an extra error $(p_a - y_a) \cdot w_j / \sum_j w_j$, which pushes present-class neurons up and absent-class neurons down less strongly on positive instances, while forcing all neurons down on negative instances. The empirical contribution is that this mechanism consistently improves F-scores and mAP on three large-scale datasets with roughly half of all instances negative, while largely preserving the F1 of negative-instance detection.

Load-bearing premise

The normalized weighted geometric mean of per-class sigmoid probabilities (Eq. 4) is a faithful and useful proxy for the probability that at least one class is present; if this proxy is poorly aligned with true presence semantics under correlated or noisy labels, the auxiliary signal could bias learning rather than help it.

Editorial extensions

If this is right

  • Any multi-label network with sigmoid outputs can adopt the proposed losses by adding a single scalar term, with no change to architecture, parameters, or inference threshold.
  • The class-balanced extension gives negative instances an explicit weight, so the method can handle negative dominance and label imbalance at the same time, a combination the paper argues is missing from prior loss-based approaches.
  • The $\lambda$ ablation shows the optimal absent-class contribution is small ($\lambda \approx 0.01$–$0.05$) for typical negative rates, while an 80%-negative COCO variant favors larger $\lambda$, giving practitioners a knob to trade positive recall against negative-detection performance.
  • Because the auxiliary term is orthogonal to the base loss, it can be composed with future losses that model label correlations or partial labels without retraining the base method.

Reading between the lines

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

  • A natural comparison the paper does not run is against an auxiliary loss using the true union probability $1-\prod_j(1-p_j)$; the geometric mean is smoother and less prone to saturation, so such an ablation would isolate whether the proxy or the mere presence signal drives the gains.
  • Since $y_a$ is a coarser label than the full multi-label vector, the auxiliary loss may tolerate individual label noise better than per-class losses, suggesting a testable application to weakly supervised or noisy-label multi-label settings.
  • The gradient in Eq. (23) shows the auxiliary term acts as a collective bias that pushes all logits of positive instances upward, a mechanism reminiscent of contrastive or energy-based objectives; connecting the two could explain why the method helps even when negatives are not extreme.
  • All experiments are on image datasets; applying the loss to multi-label text, audio, or sensor data would show whether the benefit is visual-data-specific or modality-general.
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

2 major / 5 minor

Summary. The paper proposes an auxiliary 'any-class presence likelihood' term for multi-label classification (MLC) under abundant negative data (instances with no positive labels). The any-class probability is defined in Eq. (4) as a normalized weighted geometric mean of per-class predicted probabilities, with weights 1 for present classes and λ for absent classes. This term is added to BCE and focal losses, and combined with a class-balanced reweighting scheme that assigns a negative-instance weight. Experiments on SewerML, modified COCO, and ChestX-ray14 across three architectures (TResNet-L, ViT-B16, MaxViT-S) report consistent improvements in F1, F2, and mAP over the corresponding standard losses, at no additional parameter cost. The paper also provides a derivation of the gradient in Appendix A.1, ablations over λ, and releases code.

Significance. If the proposed loss's mechanism is valid, it offers a simple, architecture-agnostic improvement for an important practical problem: multi-label classification with many fully negative instances. The empirical comparisons are controlled (same network, same base loss, only the auxiliary term changes) and show consistent directional gains, which is a meaningful and reproducible contribution. The class-balanced treatment of negative instances (Eq. 13-14) is also a useful addition. The main weakness is that the paper's central theoretical framing—that Eq. (4) is the 'any-class presence likelihood'—is not supported by the derivation in Appendix A.1; the quantity is a sigmoid of a weighted average of logits, not the model-implied union probability. The paper also lacks statistical uncertainty estimates. With appropriate rework and additional experiments, the empirical results could stand as a valuable finding, but the current manuscript overclaims the likelihood interpretation.

major comments (2)
  1. [§3.1, Eq. (4); Appendix A.1] The quantity p_a defined in Eq. (4) is called the 'any-class presence likelihood', but it is not the probability of y_a=1 implied by the per-class sigmoid model. The true model probability is u = 1 - ∏_j (1-p_j), whereas Appendix A.1 shows p_a = σ(∑(w_j z_j)/∑w_j), i.e., a sigmoid of a weighted average of logits. For many classes this can be far from u: with M=100, p_1=0.99, p_j=0.01 for j>1, and λ=0.02, u≈0.996 but p_a≈0.18. The auxiliary loss then penalizes a confident correct prediction, and its gradient (Eq. 23) pushes all logits upward, including absent classes, which can reverse the normal per-class BCE update. The paper does not check for calibration of p_a nor compare against a loss based on the true union probability. This is load-bearing because the title and abstract attribute the gains to modeling any-class presence. I recommend either replacing p_a with the true union probability, or adding concrete evidence that the geometric-mean aggregation is a suitable proxy—for example, reliability diagrams, a comparison of the proposed loss against the true-union loss on the same datasets, or a theoretical argument with testable conditions.
  2. [Tables 1-3] All results are reported from a single training run without error bars, confidence intervals, or significance tests. Some improvements are small (e.g., Table 1, MaxViT-S, F1: 66.99→67.05; Table 2, MaxViT-S, mAP often decreases while F-scores increase), and the reported F1-Neg drops, although called 'minimal', are not quantified statistically. Since the central claim is consistent improvement across datasets and networks, the paper should provide at least three seeds for the main configurations, or bootstrap confidence intervals, to ensure that the observed gains are not within run-to-run noise.
minor comments (5)
  1. [§4.4 vs Appendix A.4, Table 7] The recommendation that λ should be set in [0.01, 0.05] is not consistent with the COCO extreme-negative experiment in Appendix A.4, where the best F1 and F2 are obtained at λ=0.5. The authors should explicitly state that the optimal λ depends on the negative-to-positive ratio and the class distribution, or adjust the recommendation accordingly.
  2. [Table 3 header] The header contains a typo: 'COCO perforamnce metrics' should be 'COCO performance metrics'.
  3. [Reference [47]] Reference [47] has typographical errors in the author list and title: 'Richard M. V ogel and. The geometric mean?' should be corrected to the proper citation, e.g., Vogel, R.M. (2022), 'The geometric mean?', Communications in Statistics - Theory and Methods.
  4. [Appendix A.3] The main text should state more prominently that the COCO results are obtained on a synthetic re-labeled subset with the 'person' class and several correlated classes removed, not on the original COCO benchmark. While this is described in the appendix, the abstract and Section 4.1 could be misread as reporting results on standard COCO.
  5. [Eq. (17)] Equation (17) is the expanded form of Equation (1) and is redundant; consider removing it or clarifying that it is the same loss, to avoid confusion.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the any-class term is an explicit design choice rather than a derived likelihood, and the reported gains are evaluated on held-out splits with validation-based hyperparameter selection.

full rationale

The paper's central chain is: y_a is defined as the OR of the per-class labels (Eq. 3); p_a is synthesized as a normalized weighted geometric mean of per-class probabilities (Eq. 4); the model likelihood for y_a is then written as p_a^t (Eq. 5); and this is added to BCE or focal loss (Eqs. 8, 10, 15, 16). The step from Eq. 3 to Eq. 4 is not a mathematical derivation: under the per-class sigmoid model, the probability that at least one class is present is the union probability 1 - prod(1 - p_j), not the normalized geometric mean. So calling p_a the "any-class presence likelihood" is an ansatz, or at most a definition, rather than a derived likelihood. This is a validity or mechanism concern, not a circularity concern for the paper's empirical claim. The proposed losses are not fitted to the reported F1/F2/mAP test results: alpha, beta, gamma are fixed standard choices, and lambda is selected on validation data (Section 4.4) with the full lambda sweep reported, including values where performance degrades. All headline comparisons are on held-out test splits or, for SewerML, on a test benchmark queried from the dataset authors. There are no load-bearing self-citations, no imported uniqueness theorems, and no fitted parameter that is later renamed as a prediction. The observed gains could have gone the other way, so the evaluation is self-contained. The paper therefore shows no significant circularity; score 1 reflects only the minor overclaim that an assumed functional form is a derived likelihood.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on two introduced hyperparameters (λ and α) and on the heuristic geometric-mean proxy for any-class presence. No new physical entities are posited. The main load-bearing assumption is that the geometric-mean aggregation of per-class sigmoid probabilities is a useful training signal for detecting presence versus absence.

free parameters (2)
  • lambda (λ) = 0.02
    Controls the weight of absent-class probabilities in the any-class presence probability for positive instances; selected from validation-set ablation (Section 4.4).
  • alpha (α) = 1
    Scales the any-class presence likelihood term in the total loss; set to 1 without a dedicated ablation study, described as reflecting equal contribution (Section 4.2).
assumptions (3)
  • ad hoc to paper The normalized weighted geometric mean of per-class predicted probabilities provides a meaningful probability of any class being present.
    Eq. (4) defines p_a as a heuristic aggregation, not derived from the joint label distribution. This is the load-bearing modeling choice for the auxiliary loss.
  • domain assumption All-zero target vectors correctly identify true negative instances.
    The method assumes negative examples are genuinely label-free. The paper acknowledges ChestX-ray14's known label noise in Section 4.3, which could bias the auxiliary signal.
  • domain assumption Class-balanced weighting with the effective number of samples adequately rebalances label imbalance.
    The paper adopts the class-balanced formulation from prior work (Eq. 11-14) and extends it to negative instances, assuming this weighting scheme improves learning under imbalance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Any-Class Presence Likelihood for Robust Multi-Label Classification with Abundant Negative Data." pith.science (2026). https://pith.science/paper/EELOVIPY

@misc{pith2026250605721,
  author       = {Pith},
  title        = {Pith review of: Any-Class Presence Likelihood for Robust Multi-Label Classification with Abundant Negative Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EELOVIPY}},
  note         = {Machine review of arXiv:2506.05721}
}
read the original abstract

Multi-label Classification (MLC) assigns an instance to one or more non-exclusive classes. A challenge arises when the dataset contains a large proportion of instances with no assigned class, referred to as negative data, which can overwhelm the learning process and hinder the accurate identification and classification of positive instances. Nevertheless, it is common in MLC applications such as industrial defect detection, agricultural disease identification, and healthcare diagnosis to encounter large amounts of negative data. Assigning a separate negative class to these instances further complicates the learning objective and introduces unnecessary redundancies. To address this challenge, we redesign standard MLC loss functions by deriving a likelihood of any class being present, formulated by a normalized weighted geometric mean of the predicted class probabilities. We introduce a regularization parameter that controls the relative contribution of the absent class probabilities to the any-class presence likelihood in positive instances. The any-class presence likelihood complements the multi-label learning by encouraging the network to become more aware of implicit positive instances and improve the label classification within those positive instances. Experiments on large-scale datasets with negative data: SewerML, modified COCO, and ChestX-ray14, across various networks and base loss functions show that our loss functions consistently improve MLC performance of their standard loss counterparts, achieving gains of up to 6.01 percentage points in F1, 8.06 in F2, and 3.11 in mean average precision, all without additional parameters or computational complexity. Code available at: https://github.com/ML-for-Sensor-Data-Western/gmean-mlc

Figures

Figures reproduced from arXiv: 2506.05721 by the authors.

Figure 1
Figure 1. Performance across varying λ for TresNet-L in SewerML and COCO validation splits with redesigned BCE loss. The Y-axis is disjoint to remove idle plotting space between the F1-Neg curve and other curves. The metric values for standard BCE loss are plotted on the left of each plot. Across both datasets, F1-Neg is slightly compromised with redesigned loss for smaller λ (0 to 0.02) and improves with moderate values (λ =… view at source ↗
Figure 2
Figure 2. Surface plots of product and geometric mean between two probabilities, and their normalized [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Likelihood surface plots for two probabilities, the standard BCE likelihood, our any class [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: COCO: 20 most occurring class distribution, abstract category distribution, and category [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Performance across varying λ values for TresNet-L network for setting in [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 54 canonical work pages

  1. [1]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InEuropean Conference of Computer Vision, pages 740–755, 2014

    work page 2014

  2. [2]

    Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale.International Journal of Computer Vision, 128(7):1956–1981, 2020

    work page 1956

  3. [3]

    Improving transfer learning for movie trailer genre classification using a dual image and video transformer.Information Processing & Management, 60(3):103343, 2023

    Ricardo Montalvo-Lezama, Berenice Montalvo-Lezama, and Gibran Fuentes-Pineda. Improving transfer learning for movie trailer genre classification using a dual image and video transformer.Information Processing & Management, 60(3):103343, 2023

    work page 2023

  4. [4]

    Neural legal judgment prediction in English

    Ilias Chalkidis, Ion Androutsopoulos, and Nikolaos Aletras. Neural legal judgment prediction in English. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4317–4323, 2019

    work page 2019

  5. [5]

    Efficient few-shot learning for multi-label classification of scientific documents with many classes

    Tim Schopf, Alexander Blatzheim, Nektarios Machner, and Florian Matthes. Efficient few-shot learning for multi-label classification of scientific documents with many classes. InInternational Conference on Natural Language and Speech Processing, pages 186–198, 2024

    work page 2024

  6. [6]

    Toward purifying defect feature for multilabel sewer defect classification.IEEE Transactions on Instrumentation and Measurement, 72: 1–11, 2023

    Chuanfei Hu, Bo Dong, Hang Shao, Jiapeng Zhang, and Yongxiong Wang. Toward purifying defect feature for multilabel sewer defect classification.IEEE Transactions on Instrumentation and Measurement, 72: 1–11, 2023

    work page 2023

  7. [7]

    Defecttr: End-to-end defect detection for sewage networks using a transformer.Construction and Building Materials, 325: 126584, 2022

    L Minh Dang, Hanxiang Wang, Yanfen Li, Tan N Nguyen, and Hyeonjoon Moon. Defecttr: End-to-end defect detection for sewage networks using a transformer.Construction and Building Materials, 325: 126584, 2022

    work page 2022

  8. [8]

    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. InIEEE Conference on Computer Vision and Pattern Recognition, pages 2097–2106, 2017

    work page 2017

Show all 57 references
  1. [9]

    Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison

    Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. InProceedings of the AAAI c...

    2019

  2. [10]

    Deep-learning-assisted diagnosis for knee magnetic resonance imaging: development and retrospective validation of mrnet.PLoS medicine, 15(11): e1002699, 2018

    Nicholas Bien, Pranav Rajpurkar, Robyn L Ball, Jeremy Irvin, Allison Park, Erik Jones, Michael Bereket, Bhavik N Patel, Kristen W Yeom, Katie Shpanskaya, et al. Deep-learning-assisted diagnosis for knee magnetic resonance imaging: development and retrospective validation of mr...

    2018

  3. [11]

    Sewer-ml: A multi-label sewer defect classification dataset and benchmark

    Joakim Bruslund Haurum and Thomas B Moeslund. Sewer-ml: A multi-label sewer defect classification dataset and benchmark. InIEEE Conference on Computer Vision and Pattern Recognition, pages 13456– 13467, 2021

    2021

  4. [12]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. InIEEE Conference on Computer Vision and Pattern Recognition, pages 2980–2988, 2017

    2017

  5. [13]

    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. InIEEE Conference on Computer Vision and Pattern Recognition, pages 9268–9277, 2019. 10

    2019

  6. [14]

    On active learning in multi-label classification

    Klaus Brinker. On active learning in multi-label classification. InFrom Data and Information Analysis to Knowledge Engineering: Proceedings of the 29th Annual Conference of the Gesellschaft für Klassifikation eV University of Magdeburg, March 9–11, 2005, 2006

    2005

  7. [15]

    Comprehensive comparative study of multi-label classification methods.Expert Systems with Applications, 203:117215, 2022

    Jasmin Bogatinovski, Ljupˇco Todorovski, Sašo Džeroski, and Dragi Kocev. Comprehensive comparative study of multi-label classification methods.Expert Systems with Applications, 203:117215, 2022

    2022

  8. [16]

    Bingzhi Chen, Jinxing Li, Guangming Lu, Hongbing Yu, and David Zhang. Label co-occurrence learning with graph convolutional networks for multi-label chest x-ray image classification.IEEE Journal of Biomedical and Health Informatics, 24(8):2292–2302, 2020

    2020

  9. [17]

    Learning a deep convnet for multi-label classification with partial labels

    Thibaut Durand, Nazanin Mehrasa, and Greg Mori. Learning a deep convnet for multi-label classification with partial labels. InIEEE Conference on Computer Vision and Pattern Recognition, 2019

    2019

  10. [18]

    Binary relevance for multi-label learning: an overview.Frontiers of Computer Science, 12:191–202, 2018

    Min-Ling Zhang, Yu-Kun Li, Xu-Ying Liu, and Xin Geng. Binary relevance for multi-label learning: an overview.Frontiers of Computer Science, 12:191–202, 2018

    2018

  11. [19]

    Multi-label learning with stronger consistency guarantees

    Anqi Mao, Mehryar Mohri, and Yutao Zhong. Multi-label learning with stronger consistency guarantees. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https: //openreview.net/forum?id=zAuerb1KGx

    2024

  12. [20]

    Multilabel classification via calibrated label ranking.Machine learning, 73:133–153, 2008

    Johannes Fürnkranz, Eyke Hüllermeier, Eneldo Loza Mencía, and Klaus Brinker. Multilabel classification via calibrated label ranking.Machine learning, 73:133–153, 2008

    2008

  13. [21]

    Classifier chains for multi-label classification.Machine learning, 85:333–359, 2011

    Jesse Read, Bernhard Pfahringer, Geoff Holmes, and Eibe Frank. Classifier chains for multi-label classification.Machine learning, 85:333–359, 2011

    2011

  14. [22]

    Multi-label learning from single positive labels

    Elijah Cole, Oisin Mac Aodha, Titouan Lorieul, Pietro Perona, Dan Morris, and Nebojsa Jojic. Multi-label learning from single positive labels. InIEEE Conference on Computer Vision and Pattern Recognition, 2021

    2021

  15. [23]

    Deep long-tailed learning: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10795–10816, 2023

    Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10795–10816, 2023

    2023

  16. [24]

    When noisy labels meet long tail dilemmas: A representation calibration method

    Manyi Zhang, Xuyang Zhao, Jun Yao, Chun Yuan, and Weiran Huang. When noisy labels meet long tail dilemmas: A representation calibration method. InIEEE International Conference on Computer Vision, pages 15890–15900, 2023

    2023

  17. [25]

    Long tail multi-label learning

    Mengqi Yuan, Jinke Xu, and Zhongnian Li. Long tail multi-label learning. InIEEE Conference on Artificial Intelligence and Knowledge Engineering, pages 28–31, 2019

    2019

  18. [26]

    Distribution-balanced loss for multi-label classification in long-tailed datasets

    Tong Wu, Qingqiu Huang, Ziwei Liu, Yu Wang, and Dahua Lin. Distribution-balanced loss for multi-label classification in long-tailed datasets. InEuropean Conference on Computer Vision, pages 162–178, 2020

    2020

  19. [27]

    Edcloc: a prediction model for mrna subcellular localization using improved focal loss to address multi-label class imbalance.BMC genomics, 25(1):1252, 2024

    Yu Deng, Jianhua Jia, and Mengyue Yi. Edcloc: a prediction model for mrna subcellular localization using improved focal loss to address multi-label class imbalance.BMC genomics, 25(1):1252, 2024

    2024

  20. [28]

    Asymmetric loss for multi-label classification

    Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. InIEEE International Conference on Computer Vision, pages 82–91, 2021

    2021

  21. [29]

    Semi-supervised multi-label learning with balanced binary angular margin loss.Advances in Neural Information Processing Systems, 37: 97884–97906, 2024

    Ximing Li, Silong Liang, Changchun Li, Fangming Gu, et al. Semi-supervised multi-label learning with balanced binary angular margin loss.Advances in Neural Information Processing Systems, 37: 97884–97906, 2024

    2024

  22. [30]

    Long-tail learning via logit adjustment

    Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. InInternational Conference on Learning Representations,

  23. [31]

    Duc-Quang Vu, Trang T. T. Phung, Jia-Ching Wang, and Son T. Mai. Lcsl: Long-tailed classification via self-labeling.IEEE Transactions on Circuits and Systems for Video Technology, 34(11):12048–12058, 2024

    2024

  24. [32]

    Revisiting deep learning models for tabular data.Advances in Neural Information Processing Systems, 34:18932–18943, 2021

    Yury Gorishniy, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Revisiting deep learning models for tabular data.Advances in Neural Information Processing Systems, 34:18932–18943, 2021

    2021

  25. [33]

    The emerging trends of multi-label learning

    Weiwei Liu, Haobo Wang, Xiaobo Shen, and Ivor W Tsang. The emerging trends of multi-label learning. IEEE transactions on pattern analysis and machine intelligence, 44(11):7955–7974, 2021. 11

    2021

  26. [34]

    Multi-label local awareness and global co-occurrence priori learning improve chest x-ray classification.Multimedia Systems, 30(3):132, 2024

    Guoli Wang, Pingping Wang, and Benzheng Wei. Multi-label local awareness and global co-occurrence priori learning improve chest x-ray classification.Multimedia Systems, 30(3):132, 2024

    2024

  27. [35]

    Improving multi-label recognition using class co-occurrence probabilities

    Samyak Rawlekar, Shubhang Bhatnagar, Vishnuvardhan Pogunulu Srinivasulu, and Narendra Ahuja. Improving multi-label recognition using class co-occurrence probabilities. InInternational Conference on Pattern Recognition, pages 424–439, 2025

    2025

  28. [36]

    Multi-label out-of-distribution detection via exploiting sparsity and co-occurrence of labels.Image and Vision Computing, 126:104548, 2022

    Lei Wang, Sheng Huang, Luwen Huangfu, Bo Liu, and Xiaohong Zhang. Multi-label out-of-distribution detection via exploiting sparsity and co-occurrence of labels.Image and Vision Computing, 126:104548, 2022

    2022

  29. [37]

    Evidential mixture machines: Deciphering multi-label correlations for active learning sensitivity

    Dayou Yu, Minghao Li, Weishi Shi, and Qi Yu. Evidential mixture machines: Deciphering multi-label correlations for active learning sensitivity. Inhe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URLhttps://openreview.net/forum?id=n5lLSskwtu

    2024

  30. [38]

    In pursuit of causal label correlations for multi-label image recognition

    Zhao-Min Chen, Xin Jin, YisuGe, and Sixian Chan. In pursuit of causal label correlations for multi-label image recognition. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems,

  31. [39]

    Ml-decoder: Scalable and versatile classification head

    Tal Ridnik, Gilad Sharir, Avi Ben-Cohen, Emanuel Ben-Baruch, and Asaf Noy. Ml-decoder: Scalable and versatile classification head. InIEEE Winter Conference on Applications of Computer Vision, pages 32–41, January 2023

    2023

  32. [40]

    Coocnet: a novel approach to multi-label text classification with improved label co-occurrence modeling.Applied Intelligence, 54(17):8702–8718, 2024

    Yi Li, Junge Shen, and Zhaoyong Mao. Coocnet: a novel approach to multi-label text classification with improved label co-occurrence modeling.Applied Intelligence, 54(17):8702–8718, 2024

    2024

  33. [41]

    Dao, Ethan Zhao, Dinh Phung, and Jianfei Cai

    Son D. Dao, Ethan Zhao, Dinh Phung, and Jianfei Cai. Multi-label image classification with contrastive learning.arXiv, 2021

    2021

  34. [42]

    A review of methods for imbalanced multi-label classification.Pattern Recognition, 118:107965, 2021

    Adane Nega Tarekegn, Mario Giacobini, and Krzysztof Michalak. A review of methods for imbalanced multi-label classification.Pattern Recognition, 118:107965, 2021

    2021

  35. [43]

    Multi-label learning with weak label

    Yu-Yin Sun, Yin Zhang, and Zhi-Hua Zhou. Multi-label learning with weak label. InProceedings of the AAAI conference on artificial intelligence, volume 24, pages 593–598, 2010

    2010

  36. [44]

    Using deep learning for image-based plant disease detection.Frontiers in plant science, 7:215232, 2016

    Sharada P Mohanty, David P Hughes, and Marcel Salathé. Using deep learning for image-based plant disease detection.Frontiers in plant science, 7:215232, 2016

    2016

  37. [45]

    Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection. InIEEE Conference on Computer Vision and Pattern Recognition, pages 9584–9592, 2019

    2019

  38. [46]

    Multi-label classification by exploiting local positive and negative pairwise label correlation.Neurocomputing, 257:164–174, 2017

    Jun Huang, Guorong Li, Shuhui Wang, Zhe Xue, and Qingming Huang. Multi-label classification by exploiting local positive and negative pairwise label correlation.Neurocomputing, 257:164–174, 2017

    2017

  39. [47]

    V ogel and

    Richard M. V ogel and. The geometric mean?Communications in Statistics - Theory and Methods, 51(1): 82–94, 2022

    2022

  40. [48]

    Multi-label classification of chest x-ray abnormalities using transfer learning techniques.Journal of Personalized Medicine, 13(10): 1426, 2023

    Jakub Kufel, Michał Bielówka, Marcin Rojek, Adam Mitr˛ ega, Piotr Lewandowski, Maciej Cebula, Dariusz Krawczyk, Marta Bielówka, Dominika Kondoł, Katarzyna Bargieł-Ł ˛ aczek, et al. Multi-label classification of chest x-ray abnormalities using transfer learning techniques.Journ...

    2023

  41. [49]

    Tresnet: High performance gpu-dedicated architecture

    Tal Ridnik, Hussam Lawen, Asaf Noy, Emanuel Ben Baruch, Gilad Sharir, and Itamar Friedman. Tresnet: High performance gpu-dedicated architecture. InIEEE Winter Conference on Applications of Computer Vision, pages 1400–1409, 2021

    2021

  42. [50]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

    2021

  43. [51]

    Maxvit: Multi-axis vision transformer

    Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. InEuropean Conference on Computer Vision, pages 459–479, 2022

    2022

  44. [52]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InInternational Conference on Learning Representations, 2019. URLhttps://openreview.net/forum?id=Bkg6RiCqY7. 12

    2019

  45. [53]

    SGDR: Stochastic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. InInterna- tional Conference on Learning Representations, 2017. URL https://openreview.net/forum?id= Skq89Scxx

    2017

  46. [54]

    Multi-scale hybrid vision transformer and sinkhorn tokenizer for sewer defect classification.Automation in Construction, 144: 104614, 2022

    Joakim Bruslund Haurum, Meysam Madadi, Sergio Escalera, and Thomas B Moeslund. Multi-scale hybrid vision transformer and sinkhorn tokenizer for sewer defect classification.Automation in Construction, 144: 104614, 2022

    2022

  47. [55]

    Any-Class

    Luke Oakden-Rayner. Exploring the chestxray14 dataset: problems, 2017. URL https:// laurenoakdenrayner.com/2017/12/18/the-chestxray14-dataset-problems. A Appendix / supplemental material A.1 Learning signal on final layer neurons In this Appendix, we explain how the redesigned...

    2017

  48. [2021]

    URLhttps://openreview.net/forum?id=37nvvqkCo5

  49. [2024]

    URLhttps://openreview.net/forum?id=yBHbeSpwYS

Pith tools

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