Pith. sign in

REVIEW 3 major objections 5 minor 60 references

Calibratable Disambiguation Loss for Multi-Instance Partial-Label Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read A margin-modulated disambiguation loss improves both accuracy and calibration in multi-instance partial-label learning, cutting expected calibration error by more than half in many settings.

desk verdict First calibration-targeted loss for MIPL with compelling ECE reductions, but the theory as written doesn't support the 'calibratable' claim and the Theorem 1 proof is invalid for the CC variant. read the letter →

arxiv 2512.17788 v2 pith:3LHNKDIZ submitted 2025-12-19 cs.LG

classification cs.LG
keywords multi-instancepartial-labellearningmodelcalibrationexpectederrordisambiguationlossfocalmarginmodulationattentionmechanismweaklysupervised
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-instance partial-label learning (MIPL) handles data where both the instance and the label are only approximately supervised: each training bag contains a true label inside a candidate set with false positives, and the positive instances are unidentified. Existing MIPL models disambiguate reasonably well but are poorly calibrated, so their confidence scores do not track actual correctness. This paper proposes a plug-and-play loss, the Calibratable Disambiguation Loss (CDL), which weights each candidate label's focal loss by a top-vs-competitor prediction margin, with two competitor choices: the second-strongest candidate label (CC) or the strongest non-candidate label (CN). The paper reports that adding CDL to three existing attention mechanisms improves accuracy in 105 of 110 comparisons and calibration in 93 of 95 comparisons across benchmark and real-world pathology datasets, often halving the expected calibration error, and derives a theorem showing CDL is a confidence-scaled version of the momentum-based disambiguation loss.

What carries the argument

The key object is the Calibratable Disambiguation Loss (CDL), a margin-modulated focal loss defined in Eq. (14). It starts from the momentum-based disambiguation loss (MDL), L_MDL = −∑_{c∈S_i} w_{i,c} log p_{i,c} = KL(w_i ∥ p_i) + H[w_i], and multiplies each log-probability by (1 − M_i + Φ(p_i))^γ, where M_i is the maximum candidate probability. The two instantiations set Φ to the second-highest candidate probability (CDL-CC) or the highest non-candidate probability (CDL-CN), giving respectively 'separate candidate labels from each other' and 'suppress non-candidates'. The margin β_i = M_i − Φ(p_i) measures prediction confidence, and Theorem 1 shows CDL is bounded below by (1 − γβ_i) L_MDL,

What would settle it

Run the six CDL variants against their base attention models on a MIPL benchmark not used in the paper; if average accuracy or expected calibration error is not consistently better across ten random splits, the central claim fails. Alternatively, on a trained model, directly compute L_CDL-CC and L_MDL for each bag: if any bag satisfies L_CDL-CC < (1 − γβ_i) L_MDL, then Theorem 1 is false for that instantiation.

Watch

Extended reading notes

Core claim

The central claim is that a single scalar — the margin between the highest candidate-label probability and a competitor probability — can modulate a disambiguation loss to fix both the under-confidence and over-confidence that arise when focal losses are naively adapted to MIPL. Formally, CDL is L_CDL = −∑_{c∈S_i} w_{i,c} (1 − max_{c′∈S_i} p_{i,c′} + Φ(p_i))^γ log p_{i,c}, where Φ(p_i) is either the second-highest candidate probability (CDL-CC) or the strongest non-candidate probability (CDL-CN). Theorem 1 states L_CDL ≥ (1 − γβ_i) L_MDL, where β_i is that margin and L_MDL is the momentum-based disambiguation loss (KL divergence plus entropy of the candidate-weight distribution), making CDL

Load-bearing premise

The theoretical guarantee in Theorem 1 treats the confidence margin as a single number that factors out of the loss sum, which is only valid for the second CDL instantiation; for the first instantiation the margin shifts per candidate label, so the lower-bound proof does not apply as written.

Editorial extensions

If this is right

  • Applied to the three existing attention mechanisms (DAM, SAM, MAM), CDL improves accuracy in 105 of 110 comparisons and calibration in 93 of 95 comparisons on benchmark and real-world datasets.
  • The largest accuracy gains appear at the highest ambiguity level (three false positive labels), with a mean improvement of 8.93% on benchmark datasets, indicating CDL is most valuable when disambiguation is hardest.
  • Expected calibration error drops by more than half in 26 of 60 benchmark cases, with a mean reduction of 44.76%, aligning predicted confidence with true correctness.
  • CDL also plugs into partial-label learning methods: augmenting the POP method with CDL raises accuracy and lowers ECE on colorectal-cancer datasets with ResNet-34 features.
  • The paper's visualization of bag-level features shows CDL produces more compact and separated clusters, which it ties to the improved classification performance.

Reading between the lines

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

  • Testable extension: if the margin modulation is the active ingredient, the same margin-weighted focal term could be applied to other weakly supervised losses (e.g., partial-label learning without bags, or noisy-label learning) where miscalibrated confidence is also a problem.
  • The proof of Theorem 1 appears to assume the margin β_i is a single scalar that factors out of the sum over candidate labels; for CDL-CC, where Φ depends on c, the factorization in Eq. (A8) does not go through. A repaired proof or a counterexample would settle whether the lower bound holds for both instantiations.
  • A natural stress test is to vary γ per-sample based on the margin rather than using a fixed γ; the paper's sensitivity analysis (γ ∈ {1,...,5}) shows diminishing returns, suggesting an adaptive schedule might squeeze out further calibration gains.
  • If the empirical gains persist across a wider range of MIPL datasets (e.g., with more classes or larger candidate sets), CDL could become a default loss component for weakly supervised medical image classification, where calibration is critical.
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 a plug-and-play Calibratable Disambiguation Loss (CDL) for multi-instance partial-label learning (MIPL). CDL replaces the standard momentum-based disambiguation loss (MDL) with a margin-modulated variant that down-weights confident bags, with two instantiations: CDL-CC, which uses the second-highest candidate probability, and CDL-CN, which uses the highest non-candidate probability. The loss is integrated with three attention mechanisms (DAM, SAM, MAM), yielding six model variants. The authors claim that CDL improves both classification accuracy and expected calibration error, support this with a theorem lower-bounding CDL by a scaled MDL, and present extensive experiments on four benchmark datasets, seven real-world CRC pathology datasets, and adapted PLL baselines. The empirical results show large accuracy and ECE improvements in most of the 110 accuracy and 95 ECE comparisons, with ECE reductions often exceeding 50%.

Significance. If the main claims hold, this would be a practically useful contribution: it introduces a simple, architecture-agnostic loss that improves both accuracy and calibration in a weakly supervised setting where calibration has received little attention. The experimental evaluation is unusually thorough: it covers multiple MIPL algorithms, three attention mechanisms, benchmark and real-world pathology datasets, a newly introduced C-R34-9 dataset, and PLL adaptation experiments. The reported gains are large and mostly consistent across tables and reliability diagrams. The main weakness is the theoretical component: the advertised calibration-alignment and gradient/momentum analyses are not present, and the proof of the stated lower-bound theorem is invalid for the CDL-CC variant. The empirical evidence is extensive enough to stand on its own, but the current manuscript substantially overstates the theoretical support for the 'calibratable' property.

major comments (3)
  1. [Section 5 / Appendix A, Eq. (A8)] The proof of Theorem 1 factors the scalar term (max_{c'∈S_i} p_{i,c'} − Φ(p_i)) out of the sum over candidate labels c. This factorization is valid only if Φ is independent of c. For L_CDL-CC in Eq. (16), Φ(p_i) = max_{c'∈S_i, c'≠c} p_{i,c'} depends on c, so Eq. (A8) is invalid for the first CDL variant. For every non-top candidate c, Φ equals the top candidate probability, so the modulating factor is 1 and the margin is concentrated in the top term. A termwise repair may be possible — the top term gives margin top−second and all other terms give margin 0 — but that proof is not what the manuscript provides, and Theorem 1 as stated (with β_i = max p − Φ) is ambiguous for CDL-CC. This is a load-bearing flaw because Theorem 1 is the paper's only theoretical justification for CDL-CC.
  2. [Abstract / Section 5] The abstract and introduction promise a 'pseudo-label confidence-alignment bound for calibration' and 'gradient and momentum analyses' showing how margin shaping affects weight updates. Neither appears in Section 5 or Appendix A. Section 5 contains only the MDL lower bound (Theorem 1), and Appendix A contains only its proof. No result in the paper bounds ECE, relates CDL to any calibration metric, or analyzes gradients or momentum dynamics. The theoretical component of the central 'calibratable' claim is therefore missing, and the paper should either add such analyses or revise the claims to describe the contribution as an empirical one.
  3. [Section 5, Eq. (18) and Section 6.1.3] For L_CDL-CN, the condition Φ(p_i) ∈ [max_{c'∈S_i} p_{i,c'} − 1, max_{c'∈S_i} p_{i,c'}] reduces to the assumption that the maximum non-candidate probability never exceeds the maximum candidate probability. This is asserted based on Fig. 7, which shows mean probabilities at the last epoch on one training set; it does not establish the inequality for all bags and all epochs, particularly early in training when the model is not yet disambiguating. Additionally, all main experiments set γ=1, and in that case L_CDL-CN is exactly (1−β_i)L_MDL, not merely bounded below by it. The theorem's inequality is then an equality for CDL-CN, and the paper should state that the contribution of Theorem 1 for this variant is a margin-scaling view of MDL, not a calibration guarantee.
minor comments (5)
  1. [Eq. (16)] The first factor in L_CDL-CC should be the candidate weight w_{i,c}, not the probability p_{i,c}; as written, Eq. (16) drops the momentum weights that are part of the CDL definition in Eq. (14).
  2. [Section 5, Theorem 1] The notation β_i = max_{c'∈S_i} p_{i,c'} − Φ(p_i) is ambiguous for CDL-CC because Φ depends on c inside the sum. If a repaired theorem is given, β should be defined per term or the instantiation should be restricted to constant Φ.
  3. [Section 6.5.2] The sentence 'where the weights w^{(t)}_{c,t} of candidate labels...' contains a typo: the subscript should be (i,c), not (c,t).
  4. [Section 6.2.1] The paper reports improvements relative to DEMIPL, ELIMIPL, or MIPLMA but does not report any statistical significance tests (e.g., paired t-tests or Wilcoxon). Given the large standard deviations on some benchmark rows (e.g., MNIST-MIPL r=3), a significance analysis would strengthen the claims.
  5. [Section 6.5.1] The γ-sensitivity analysis is performed only on C-R34-25. Since the main experiments all use γ=1 and the theorem requires γ ∈ [1, 1/β_max), it would be useful to report sensitivity on at least one benchmark dataset as well.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CDL is a newly defined loss whose theorem is an algebraic consequence of the definition, and the accuracy/ECE claims rest on independent held-out measurements.

full rationale

The paper's derivation chain is not circular. The proposed CDL (Eqs. 14, 16, 17) is a new loss defined by the authors, and Theorem 1 states an algebraic inequality between this defined loss and the momentum-based disambiguation loss L_MDL = -sum_c w_c log p_c. The proof obtains the bound by applying Bernoulli's inequality to the factor (1 - max_c' p_c' + Phi(p))^gamma present in the definition of CDL, so the theorem is a mathematical property of the definition rather than a prediction extracted from data. No parameter is fitted to the ECE or accuracy values that are later reported as results, and no quantity is defined in terms of the target outcome and then renamed as a prediction. The empirical ECE and accuracy comparisons in Tables 2-6 are independent measurements on held-out test splits, and the paper includes ablations against focal-loss variants and PLL adaptations. Self-citations to the authors' prior MIPL papers supply baselines and notation, but the central comparison is against reported baseline numbers rather than an unverified uniqueness theorem. The paper does contain a substantive completeness gap: the abstract promises a pseudo-label confidence-alignment bound and gradient/momentum analyses, while Section 5 and Appendix A contain only the lower-bound theorem, and the proof as written appears to factor a c-dependent Phi for L_CDL-CC as a scalar in Eq. (A8). These are correctness/completeness concerns, not circularity, because the empirical claims do not reduce to the theorem or to the self-citations.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method uses only hyperparameters and assumptions; no new physical entities. The most consequential items are the unverified margin condition for CDL-CN and the invalid factorization assumption for CDL-CC.

free parameters (3)
  • Exponential factor γ = γ=1 in main experiments; sensitivity sweep over 2–5
    Controls the margin-modulation strength. Set constant rather than learned; theorem requires γ∈[1,1/β_max), but that condition is not verified in experiments.
  • Momentum coefficient schedule α(t)=(T−t)/T = T-dependent schedule
    Inherited from momentum-based disambiguation loss and used to update candidate-label weights in Eq. (15). Chosen by hand, not data-driven.
  • MAM temperature annealing parameters τ(0), τ_m, decay 0.95 = not reported here; taken from MIPLMA
    Used by the MAM-based CDL variants; not tuned in this paper but affects training dynamics.
assumptions (4)
  • standard math Bernoulli's inequality: (1+x)^n ≥ 1+nx for integer n≥1 and real x≥−1.
    Central to the proof of Theorem 1 (Appendix A).
  • domain assumption Each MIPL bag contains at least one positive instance for the true label and no positive instance for false-positive labels.
    Standard MIPL assumption stated in Section 3.1.
  • ad hoc to paper For L_CDL-CN, max non-candidate probability ≤ max candidate probability throughout training, so Φ(p)∈[max p −1, max p].
    Needed for Theorem 1's condition; justified only by Fig. 7 on one dataset rather than by proof.
  • ad hoc to paper The condition γβ_i < 1 holds for every bag.
    Required for a positive stable scaling in Theorem 1 but never checked during training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Calibratable Disambiguation Loss for Multi-Instance Partial-Label Learning." pith.science (2026). https://pith.science/paper/3LHNKDIZ

@misc{pith2026251217788,
  author       = {Pith},
  title        = {Pith review of: Calibratable Disambiguation Loss for Multi-Instance Partial-Label Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3LHNKDIZ}},
  note         = {Machine review of arXiv:2512.17788}
}
read the original abstract

Multi-instance partial-label learning (MIPL) is a weakly supervised framework that extends the principles of multi-instance learning (MIL) and partial-label learning (PLL) to address the challenges of inexact supervision in both instance and label spaces. However, existing MIPL approaches often suffer from poor calibration, undermining classifier reliability. In this work, we propose a plug-and-play calibratable disambiguation loss (CDL) for classification and calibration, which modulates a disambiguation objective by a top-vs-competitor prediction margin. The competitor is instantiated either as the second strongest candidate label or as the strongest non-candidate label, yielding two variants that respectively emphasize candidate-level separation and candidate-vs-non-candidate suppression. Theoretically, we analyze CDL as a margin-modulated momentum-based disambiguation loss (MDL) objective, derive a lower-bound and a pseudo-label confidence-alignment bound for calibration, and show through gradient and momentum analyses how margin shaping affects weight updates. Experimental results on benchmark and real-world MIPL datasets, together with representative PLL adaptation, confirm that our CDL significantly improves both classification accuracy and expected calibration error.

Figures

Figures reproduced from arXiv: 2512.17788 by the authors.

Figure 1
Figure 1. Pathology image classification with crowd-sourced candidate [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Reliability diagrams of (a) DEMIPL [18], (b) ELIMIPL [22], (c) MI￾PLMA [23], (d) DAMCC, (e) SAMCC, (f) MAMCC, (g) DAMCN, (h) SAMCN, and (i) MAMCN on the C-KMeans test set. The diagrams display mean accuracy (ACC) and expected calibration error (ECE) from ten runs, with (d)-(i) representing our methods. The bar color intensity reflects the number of samples assigned to the corresponding confidence intervals. Despite … view at source ↗
Figure 3
Figure 3. Reliability diagrams of SAM [22] with FL or IFL on the FMNIST￾MIPL dataset with one false positive label (r = 1). Xi = {xi,1, xi,2, . . . , xi,ni } contains ni instances within a d-dimensional space. The candidate label set Si and the non-candidate label set S¯ i are subsets of Y, satisfying the constraints Si ∪ S¯ i = Y and Si ∩ S¯ i = ∅. Notably, each bag contains at least one instance corre￾sponding to the true l… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Framework of MIPL approaches within the embedded-space paradigm. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Reliability diagrams of DEMIPL [18], ELIMIPL [22], and MIPLMA [23], and our methods on the Birdsong-MIPL datasets with varying numbers of false positive labels (r ∈ {1, 2, 3}). The bar color intensity indicates that more samples are assigned with the corresponding conf…
Figure 6
Figure 6. Figure 6: t-SNE visualization of aggregated bag-level feature representations produced by the attention mechanisms in D [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Probabilities at the last epoch on the training set of the C-KMeans [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Classification accuracy and expected calibration error of D [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Classification accuracy and expected calibration error (mean and [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 2 linked inside Pith

  1. [1]

    A brief introduction to weakly supervised learning,

    Z.-H. Zhou, “A brief introduction to weakly supervised learning,” National Science Review, vol. 5, no. 1, pp. 44–53, 2018

  2. [2]

    Multiple instance classification: Review, taxonomy and comparative study,

    J. Amores, “Multiple instance classification: Review, taxonomy and comparative study,”Artificial Intelligence, vol. 201, pp. 81–105, 2013

  3. [3]

    Multiple instance learning: A survey of problem characteristics and applications,

    M.-A. Carbonneau, V . Cheplygina, E. Granger, and G. Gagnon, “Multiple instance learning: A survey of problem characteristics and applications,”Pattern Recognition, vol. 77, pp. 329–353, 2018. 14

  4. [4]

    Attention-based deep multiple instance learning,

    M. Ilse, J. M. Tomczak, and M. Welling, “Attention-based deep multiple instance learning,” inProceedings of the 35th International Conference on Machine Learning, Stockholmsm¨ assan, Stockholm, Swe- den, 2018, pp. 2132–2141

  5. [5]

    Revisiting multiple instance neural networks,

    X. Wang, Y. Yan, P . Tang, X. Bai, and W. Liu, “Revisiting multiple instance neural networks,”Pattern Recognition, vol. 74, pp. 15–24, 2018

  6. [6]

    Multi- instance causal representation learning for instance label predic- tion and out-of-distribution generalization,

    W. Zhang, X. Zhang, H.-W. Deng, and M.-L. Zhang, “Multi- instance causal representation learning for instance label predic- tion and out-of-distribution generalization,” inAdvances in Neural Information Processing Systems 35, New Orleans, LA, USA, 2022, pp. 34 940–34 953

  7. [7]

    DTFD-MIL: Double-tier feature distillation multiple instance learning for histopathology whole slide image classifi- cation,

    H. Zhang, Y. Meng, Y. Zhao, Y. Qiao, X. Yang, S. E. Coupland, and Y. Zheng, “DTFD-MIL: Double-tier feature distillation multiple instance learning for histopathology whole slide image classifi- cation,” inProceedings of the 35th IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA, 2022, pp. 18 802– 18 812

  8. [8]

    Incorporating probabilistic domain knowledge into deep multiple instance learning,

    G. S. A. Hajj, A. Hubin, C. Kanduri, M. Pavlovic, K. D. Rand, M. Widrich, A. S. Solberg, V . Greiff, J. Pensar, G. Klambauer, and G. K. Sandve, “Incorporating probabilistic domain knowledge into deep multiple instance learning,” inProceedings of the 41st International Conference on Machine Learning, Vienna, Austria, 2024, pp. 17 279–17 297

Show all 60 references
  1. [9]

    Data- driven knowledge fusion for deep multi-instance learning,

    Y. Zhang, Z. Zhou, X. He, A. R. Adhikary, and B. Dutta, “Data- driven knowledge fusion for deep multi-instance learning,”IEEE Transactions on Neural Networks and Learning Systems, pp. 1–15, 2024

  2. [10]

    Learning from partial labels,

    T. Cour, B. Sapp, and B. Taskar, “Learning from partial labels,”The Journal of Machine Learning Research, vol. 12, pp. 1501–1536, 2011

  3. [11]

    Exploiting class activation value for partial-label learning,

    F. Zhang, L. Feng, B. Han, T. Liu, G. Niu, T. Qin, and M. Sugiyama, “Exploiting class activation value for partial-label learning,” in Proceedings of the 10th International Conference on Learning Repre- sentations, Virtual Event, 2022, pp. 1–17

  4. [12]

    Partial label learning with semantic label representations,

    S. He, L. Feng, F. Lv, W. Li, and G. Yang, “Partial label learning with semantic label representations,” inProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, 2022, pp. 545–553

  5. [13]

    A unifying probabilistic framework for partially labeled data learning,

    X. Gong, D. Yuan, W. Bao, and F. Luo, “A unifying probabilistic framework for partially labeled data learning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 7, pp. 8036– 8048, 2023

  6. [14]

    Learning with partial labels from semi-supervised perspective,

    X. Li, Y. Jiang, C. Li, Y. Wang, and J. Ouyang, “Learning with partial labels from semi-supervised perspective,” inProceedings of the 37th AAAI Conference on Artificial Intelligence, Washington, DC, USA, 2023, pp. 8666–8674

  7. [15]

    Progressive purification for instance-dependent partial label learning,

    N. Xu, B. Liu, J. Lv, C. Qiao, and X. Geng, “Progressive purification for instance-dependent partial label learning,” inProceedings of the 40th International Conference on Machine Learning, Honolulu, HI, USA, ser. Proceedings of Machine Learning Research, vol. 202, 2023, pp. 3...

  8. [16]

    Distilling reliable knowledge for instance-dependent partial label learning,

    D.-D. Wu, D.-B. Wang, and M.-L. Zhang, “Distilling reliable knowledge for instance-dependent partial label learning,” inPro- ceedings of the 38th AAAI Conference on Artificial Intelligence, Van- couver, Canada, 2024, pp. 15 888–15 896

  9. [17]

    Partial label causal representation learning for instance-dependent supervision and domain generalization,

    Y.-Z. Wang, W. Zhang, and M.-L. Zhang, “Partial label causal representation learning for instance-dependent supervision and domain generalization,” inProceedings of the 39th AAAI Conference on Artificial Intelligence, Philadelphia, P A, USA, Philadelphia, Penn- sylvania, 2025, pp. 1–9

  10. [18]

    Disambiguated attention embedding for multi-instance partial-label learning,

    W. Tang, W. Zhang, and M.-L. Zhang, “Disambiguated attention embedding for multi-instance partial-label learning,” inAdvances in Neural Information Processing Systems 36, New Orleans, LA, USA, 2023, pp. 56 756–56 771

  11. [19]

    Multi-instance partial-label learning: Towards exploiting dual inexact supervision,

    W. Tang, W. Zhang, and M.-L. Zhang, “Multi-instance partial-label learning: Towards exploiting dual inexact supervision,”Science China Information Sciences, vol. 67, no. 3, pp. 132 103:1–132 103:14, 2024

  12. [20]

    Clinical-grade computational pathology using weakly supervised deep learning on whole slide images,

    G. Campanella, M. G. Hanna, L. Geneslaw, A. Miraflor, V . Werneck Krauss Silva, K. J. Busam, E. Brogi, V . E. Reuter, D. S. Klimstra, and T. J. Fuchs, “Clinical-grade computational pathology using weakly supervised deep learning on whole slide images,”Nature Medicine, vol. 25,...

  13. [21]

    Crowdsourcing of histological image labeling and object delineation by medical students,

    A. Grote, N. S. Schaadt, G. Forestier, C. Wemmert, and F. Feuer- hake, “Crowdsourcing of histological image labeling and object delineation by medical students,”IEEE Transactions Medical Imag- ing, vol. 38, no. 5, pp. 1284–1294, 2019

  14. [22]

    Exploiting conjugate label information for multi-instance partial-label learning,

    W. Tang, W. Zhang, and M.-L. Zhang, “Exploiting conjugate label information for multi-instance partial-label learning,” inProceed- ings of the 33rd International Joint Conference on Artificial Intelligence, Jeju, South Korea, 2024, pp. 4973–4981

  15. [23]

    Multi-instance partial-label learning with margin adjustment,

    W. Tang, Y.-F. Yang, Z. Wang, W. Zhang, and M.-L. Zhang, “Multi-instance partial-label learning with margin adjustment,” inAdvances in Neural Information Processing Systems 37, Vancouver, Canada, 2024, pp. 26 331–26 354

  16. [24]

    Pro- gressive identification of true labels for partial-label learning,

    J. Lv, M. Xu, L. Feng, G. Niu, X. Geng, and M. Sugiyama, “Pro- gressive identification of true labels for partial-label learning,” in Proceedings of the 37th International Conference on Machine Learning, Virtual Event, 2020, pp. 6500–6510

  17. [25]

    Solving the multiple instance problem with axis-parallel rectangles,

    T. G. Dietterich, R. H. Lathrop, and T. Lozano-P ´erez, “Solving the multiple instance problem with axis-parallel rectangles,”Artificial intelligence, vol. 89, no. 1-2, pp. 31–71, 1997

  18. [26]

    Multi-instance learning by treating instances as non-i.i.d. samples,

    Z.-H. Zhou, Y.-Y. Sun, and Y.-F. Li, “Multi-instance learning by treating instances as non-i.i.d. samples,” inProceedings of the 26th International Conference on Machine Learning, Montreal, Quebec, Canada, 2009, pp. 1249–1256

  19. [27]

    Multiple instance active learning for object detection,

    T. Yuan, F. Wan, M. Fu, J. Liu, S. Xu, X. Ji, and Q. Ye, “Multiple instance active learning for object detection,” inProceedings of the 34th IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, Virtual Event, 2021, pp. 5330–5339

  20. [28]

    Unbiased multiple instance learning for weakly supervised video anomaly detection,

    H. Lv, Z. Yue, Q. Sun, B. Luo, Z. Cui, and H. Zhang, “Unbiased multiple instance learning for weakly supervised video anomaly detection,” inProceedings of the 34th IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, Vancouver, Canada, 2023, pp. 8022–8031

  21. [29]

    Loss-based attention for deep multiple instance learning,

    X. Shi, F. Xing, Y. Xie, Z. Zhang, L. Cui, and L. Yang, “Loss-based attention for deep multiple instance learning,” inProceedings of the 34th AAAI Conference on Artificial Intelligence, New York, NY, USA, 2020, pp. 5742–5749

  22. [30]

    Bayes-MIL: A new probabilistic perspective on attention-based multiple instance learning for whole slide im- ages,

    Y. Cui, Z. Liu, X. Liu, X. Liu, C. Wang, T.-W. Kuo, C. J. Xue, and A. B. Chan, “Bayes-MIL: A new probabilistic perspective on attention-based multiple instance learning for whole slide im- ages,” inProceedings of the 11th International Conference on Learning Representations, K...

  23. [31]

    CAMIL: context-aware multiple instance learning for cancer detection and subtyping in whole slide images,

    O. Fourkioti, M. D. Vries, and C. Bakal, “CAMIL: context-aware multiple instance learning for cancer detection and subtyping in whole slide images,” inProceedings of the 12th International Conference on Learning Representations, Vienna, Austria, 2024, pp. 1–16

  24. [32]

    Inherently interpretable time series classification via multiple instance learning,

    J. Early, G. K. C. Cheung, K. Cutajar, H. Xie, J. Kandola, and N. Twomey, “Inherently interpretable time series classification via multiple instance learning,” inProceedings of the 12th International Conference on Learning Representations, Vienna, Austria, 2024, paper 1–29

  25. [33]

    GM-PLL: Graph matching based partial label learning,

    G. Lyu, S. Feng, T. Wang, C. Lang, and Y. Li, “GM-PLL: Graph matching based partial label learning,”IEEE Transactions on Knowl- edge and Data Engineering, vol. 33, no. 2, pp. 521–535, 2019

  26. [34]

    Deep discriminative CNN with temporal ensembling for ambiguously- labeled image classification,

    Y. Yao, J. Deng, X. Chen, C. Gong, J. Wu, and J. Yang, “Deep discriminative CNN with temporal ensembling for ambiguously- labeled image classification,” inProceedings of the 34th AAAI Con- ference on Artificial Intelligence, New York, NY, USA, 2020, pp. 12 669– 12 676

  27. [35]

    A conditional multinomial mixture model for superset label learning,

    L. Liu and T. G. Dietterich, “A conditional multinomial mixture model for superset label learning,” inAdvances in Neural Informa- tion Processing Systems 25, Cambridge, MA, USA, 2012, pp. 548–556

  28. [36]

    Tuning the right foun- dation models is what you need for partial label learning,

    K. He, W. Tang, T. Wei, and M. Zhang, “Tuning the right foun- dation models is what you need for partial label learning,”CoRR, vol. abs/2506.05027, 2025

  29. [37]

    Rank-loss support instance machines for MIML instance annotation,

    F. Briggs, X. Z. Fern, and R. Raich, “Rank-loss support instance machines for MIML instance annotation,” inProceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Beijing, China, 2012, pp. 534–542

  30. [38]

    Semi-supervised partial label learning via confidence-rated margin maximization,

    W. Wang and M. Zhang, “Semi-supervised partial label learning via confidence-rated margin maximization,” inAdvances in Neural Information Processing Systems 33, Virtual Event, 2020, pp. 6982– 6993

  31. [39]

    Adaptive graph guided disambiguation for partial label learning,

    D.-B. Wang, M.-L. Zhang, and L. Li, “Adaptive graph guided disambiguation for partial label learning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 12, pp. 8796– 8811, 2022

  32. [40]

    Provably consistent partial-label learning,

    L. Feng, J. Lv, B. Han, M. Xu, G. Niu, X. Geng, B. An, and M. Sugiyama, “Provably consistent partial-label learning,” inAd- vances in Neural Information Processing Systems 33, Virtual Event, 2020, pp. 10 948–10 960. 15

  33. [41]

    Leveraged weighted loss for partial label learning,

    H. Wen, J. Cui, H. Hang, J. Liu, Y. Wang, and Z. Lin, “Leveraged weighted loss for partial label learning,” inProceedings of the 38th International Conference on Machine Learning, Virtual Event, 2021, pp. 11 091–11 100

  34. [42]

    Realistic evaluation of deep partial-label learning algorithms,

    W. Wang, D.-D. Wu, J. Wang, G. Niu, M.-L. Zhang, and M. Sugiyama, “Realistic evaluation of deep partial-label learning algorithms,” inProceedings of the 13th International Conference on Learning Representations, Singapore, 2025, pp. 1–25

  35. [43]

    ProMIPL: A probabilistic generative model for multi-instance partial-label learning,

    Y.-F. Yang, W. Tang, and M.-L. Zhang, “ProMIPL: A probabilistic generative model for multi-instance partial-label learning,” in Proceedings of the 24th IEEE International Conference on Data Mining, Abu Dhabi, UAE, 2024, pp. 560–569

  36. [44]

    Fast multi-instance partial- label learning,

    Y.-F. Yang, W. Tang, and M.-L. Zhang, “Fast multi-instance partial- label learning,” inProceedings of the 39th AAAI Conference on Artifi- cial Intelligence, Philadelphia, P A, USA, Philadelphia, Pennsylvania, 2025, pp. 1–9

  37. [45]

    On learning latent models with multi-instance weak supervision,

    K. Wang, E. Tsamoura, and D. Roth, “On learning latent models with multi-instance weak supervision,” inAdvances in Neural Information Processing Systems 36, New Orleans, LA, USA, 2023, pp. 9661–9694

  38. [46]

    On calibration of modern neural networks,

    C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” inProceedings of the 34th International Conference on Machine Learning, Sydney, NSW, Australia, 2017, pp. 1321–1330

  39. [47]

    When does label smoothing help?

    R. M ¨uller, S. Kornblith, and G. E. Hinton, “When does label smoothing help?” inAdvances in Neural Information Processing Systems 32, Vancouver, BC, Canada, 2019, pp. 4696–4705

  40. [48]

    Rethinking calibration of deep neural networks: Do not be afraid of overconfidence,

    D.-B. Wang, L. Feng, and M.-L. Zhang, “Rethinking calibration of deep neural networks: Do not be afraid of overconfidence,” in Advances in Neural Information Processing Systems 34, Virtual Event, 2021, pp. 11 809–11 820

  41. [49]

    On mixup training: Improved calibration and predictive uncertainty for deep neural networks,

    S. Thulasidasan, G. Chennupati, J. A. Bilmes, T. Bhattacharya, and S. Michalak, “On mixup training: Improved calibration and predictive uncertainty for deep neural networks,” inAdvances in Neural Information Processing Systems 32, Vancouver, BC, Canada, 2019, pp. 13 888–13 899

  42. [50]

    When and how mixup improves calibration,

    L. Zhang, Z. Deng, K. Kawaguchi, and J. Zou, “When and how mixup improves calibration,” inProceedings of the 39th International Conference on Machine Learning, Baltimore, Maryland, USA, vol. 162, 2022, pp. 26 135–26 160

  43. [51]

    Calibrating deep neural networks using focal loss,

    J. Mukhoti, V . Kulharia, A. Sanyal, S. Golodetz, P . H. S. Torr, and P . K. Dokania, “Calibrating deep neural networks using focal loss,” inAdvances in Neural Information Processing Systems 33, Virtual Event, 2020, pp. 15 288–15 299

  44. [52]

    Dual focal loss for calibration,

    L. Tao, M. Dong, and C. Xu, “Dual focal loss for calibration,” in Proceedings of the 40th International Conference on Machine Learning, Honolulu, HI, USA, 2023, pp. 33 833–33 849

  45. [53]

    Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods,

    J. Plattet al., “Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods,”Advances in Large Margin Classifiers, vol. 10, no. 3, pp. 61–74, 1999

  46. [54]

    Post-hoc uncertainty calibration for domain drift scenarios,

    C. Tomani, S. Gruber, M. E. Erdem, D. Cremers, and F. Buettner, “Post-hoc uncertainty calibration for domain drift scenarios,” in Proceedings of the 32th IEEE/CVF Conference on Computer Vision and Pattern Recognition, Virtual Event, June 19-25, 2021, 2021, pp. 10 124– 10 132

  47. [55]

    On the pitfall of mixup for uncertainty calibration,

    D.-B. Wang, L. Li, P . Zhao, P .-A. Heng, and M.-L. Zhang, “On the pitfall of mixup for uncertainty calibration,” inProceedings of the 34th IEEE/CVF Conference on Computer Vision and Pattern Recognition, Vancouver, BC, Canada, 2023, pp. 7609–7618

  48. [56]

    Calibration bottleneck: Over- compressed representations are less calibratable,

    D.-B. Wang and M.-L. Zhang, “Calibration bottleneck: Over- compressed representations are less calibratable,” inProceedings of the 41st International Conference on Machine Learning, Vienna, Austria, 2024, pp. 52 156–52 170

  49. [57]

    Focal loss for dense object detection,

    T. Lin, P . Goyal, R. B. Girshick, K. He, and P . Doll ´ar, “Focal loss for dense object detection,”IEEE transactions on pattern analysis and machine intelligence, no. 2, pp. 318–327, 2020

  50. [58]

    Gradient-based learning applied to document recognition,

    Y. LeCun, L. Bottou, Y. Bengio, and P . Haffner, “Gradient-based learning applied to document recognition,”Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998

  51. [59]

    Fashion-MNIST: A novel image dataset for benchmarking machine learning algorithms,

    H. Xiao, K. Rasul, and R. Vollgraf, “Fashion-MNIST: A novel image dataset for benchmarking machine learning algorithms,”CoRR, vol. abs/1708.07747, 2017. [Online]. Available: http://arxiv.org/abs/1708.07747

  52. [60]

    Multiple-instance active learn- ing,

    B. Settles, M. Craven, and S. Ray, “Multiple-instance active learn- ing,” inAdvances in Neural Information Processing Systems 20, Van- couver, British Columbia, Canada, 2007, pp. 1289–1296. APPENDIXA PROOF OFLOWERBOUND ANDREGULARIZATION PROPERTIES OFCDL (THEOREM1) Theorem 1(Lo...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.