Reducing Variability of Multiple Instance Learning Methods for Digital Pathology

Ali Mammadov; Guillaume Hocquet; Lo\"ic le Folgoc; Pietro Gori

REVIEW 4 major objections 5 minor 29 references

Reducing Variability of Multiple Instance Learning Methods for Digital Pathology

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Training several MIL models for a few epochs, selecting the best by early validation AUC, and averaging their weights reduces run-to-run performance variance in whole-slide classification.

desk verdict A cheap, practical recipe for stabilizing MIL training variance, with a real empirical payoff, but the paper's central mechanism—early validation selection—is not actually isolated from plain weight averaging. read the letter →

arxiv 2507.00292 v2 pith:VMKGMYJW submitted 2025-06-30 cs.CV cs.AI

Ali Mammadov , Loïc Le Folgoc , Guillaume Hocquet , Pietro Gori This is my paper · ORCID

classification cs.CVcs.AI

keywords MultipleInstanceLearningWholeSlideImageVariabilityreductionModelmergingsoupsTIES-mergingDigitalpathologyReproducibility

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 takes on a reproducibility problem in digital pathology: multiple instance learning (MIL) classifiers for whole-slide images can differ by 10 to 15 AUC points between runs with the same data, making it hard to tell whether one method genuinely beats another. The authors propose a simple fix: train M copies of the same model for a small number of epochs (M=10, K=5), rank them by validation AUC, keep the top T=3, and merge their weights before full training. Averaging the weights directly (uniform soup) or with sign-consistent TIES-merging both give lower standard deviation, higher minimum AUC, and mean AUC at least as good as fully trained baselines with tuned learning rates, while using far fewer total epochs than ensembles. If this holds, MIL comparisons become more trustworthy and expensive hyperparameter tuning becomes less necessary.

What carries the argument

The core mechanism is weight averaging of partially trained models, combined with a cheap multi-fidelity selection step. Uniform SOUP averages the weights of the top T models; TIES-Merging first trims small weight differences from the initialization, determines the dominant sign of each parameter across models, and averages only values agreeing with that sign. The selection step uses validation AUC after only K=5 epochs as a proxy for final quality, so that the merging focuses on the most promising models. This is what turns ten expensive full runs into one short warm-up phase followed by a single full training.

What would settle it

Compute the Spearman correlation between validation AUC at epoch 5 and final test AUC across many initialization and shuffle seeds on a third dataset; if it is near zero or negative, the selection step is not choosing the better models and the method's gains would come from averaging alone.

Watch

Extended reading notes

Core claim

The central claim is that a multi-fidelity, model-fusion strategy reduces variability in MIL methods for WSI classification. Instead of fully training one model, the authors partially train ten models from the same initialization with different shuffle seeds for five epochs, select the three with the highest validation AUC, and average their weights (using uniform Model Soups or TIES-Merging with sign alignment). The resulting merged model is then fully trained. Across five MIL methods and two datasets (Camelyon16 and BRACS), the proposed Soup3 and Ties3 methods achieve better or similar mean test AUC than a baseline model and a learning-rate-tuned model, while consistently having the smallest standard deviation across ten initialization seeds and raising the minimum test AUC. The method also approaches or matches ensemble and best-on-validation performance at roughly one seventh of the training epochs and one tenth of the inference cost.

Load-bearing premise

The load-bearing premise is that after just five epochs, validation AUC already ranks the partially trained models the same way a full training would, so the three models chosen to merge are genuinely the best.

Editorial extensions

If this is right

Any existing MIL method can be wrapped with this procedure and inherit lower run-to-run variability without architectural changes.
Comparisons between MIL methods become more reliable, since reported gaps larger than the now smaller standard deviation can be taken seriously.
Total training cost stays low: 150 epochs for Soup3/Ties3 versus 600 for learning-rate tuning and 1000 for ensembles.
The same principle may make hyperparameter tuning cheaper, since early validation scores from the warm-up phase give a fast signal for ranking model candidates.

Reading between the lines

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

A direct testable prediction is that the early-validation ranking at epoch 5 correlates strongly with final test AUC; if that correlation is weak on a new dataset, the selection step would add little over merging any T models.
The variance reduction likely comes mostly from weight averaging, with selection acting as a guardrail; a reader could test this by comparing Soup3 with averaging three randomly chosen warm-up models.
The approach should transfer to other high-variance medical image classification settings where bag-of-patches MIL is used, and also to non-medical multiple-instance problems, because the mechanism is independent of the aggregator architecture.
Because the warm-up models are all initialized identically and differ only in shuffle seed, the method isolates ordering noise rather than initialization noise; a variant that also varies initialization could reduce variance further.

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.

Desk Editor's Note

read the letter

Colleague,

Here's the short version: this paper gives computational pathology labs a genuinely useful, cheap trick. Train M=10 copies of your MIL aggregator for K=5 epochs, keep the top 3 by validation AUC, average their weights (either uniform soup or TIES), then fully train the merged model. Across 5 MIL methods and 2 datasets the reported variance drops are large—MaxMIL on BRACS goes from STD 7.3 to 1.7—and minimum AUCs improve substantially. That is a real result for a field where 1–2 AUC point differences get buried under 10–15 points of run-to-run noise.

What's new is the specific protocol: multi-fidelity early validation selection on top of model soups, applied to the MIL aggregator, with a broad empirical sweep. It is not a restatement of Wortsman et al. or TIES. The ablation on K, T, and initialization type is useful, and the anonymous code is a plus.

Now the soft spots, in order of seriousness. First, the causal claim is under-supported. The paper attributes the variance reduction to selecting the 'most stable and promising' models by early validation, but it never runs the obvious control: random T=3, or T=10 at the same budget. At K=5 the validation ranking may be mostly optimization noise; if random T=3 gives the same mean/min/STD, then the multi-fidelity selection isn't load-bearing and the observed gain is just weight averaging plus extra warm-up epochs. The authors should add this control; it is cheap and would settle the matter.

Second, overclaims in the text. The Discussion says the proposed methods have 'the smallest STD across all MIL methods and datasets,' but the table shows Ensemble with STD 0.9 on TransMIL/Camelyon versus Soup3's 1.4. That is a factual contradiction and should be fixed. Similarly, the intro says averaging 'smooth(s) out differences caused by random initialization,' but all M models share the same initialization seed; the averaging only smooths shuffle noise. Minor but worth correcting.

Third, the choice of K=5 and T=3 is selected after ablations on the same datasets without significance tests. The K=10 values look better in Table 2; the authors justify K=5 by compute, which is fair, but they should say every headline result is conditional on that choice.

Who gets value: any lab working on MIL for pathology that wants reproducible benchmarks without spending 10x compute on grid search. It deserves a serious referee, but with the missing control and the overclaim corrected. Send it to review, and ask for the random-T3 baseline before acceptance.

Best, [You]

Referee Report

4 major / 5 minor

Summary. The paper proposes a multi-fidelity model fusion strategy for MIL-based WSI classification, called Soup3/Ties3: train M models for K epochs (typically M=10, K=5), select the top T models (T=3) by validation AUC, merge their weights with uniform Soup or TIES-Merging, then fully train the merged model. The authors validate on Camelyon16 and BRACS with five MIL methods, reporting that the method reduces run-to-run test AUC variability compared to a single baseline and to per-run learning-rate tuning, while using fewer total epochs than ensembles or LR grid search. The paper includes ablations over K, T, and initialization.

Significance. If the claim holds, the method is practically useful: it is model-agnostic, simple, and appears to reduce the well-documented high variability of MIL results in digital pathology, which currently hampers reliable method comparison. The strengths of the paper are the large experimental campaign (more than 2000 experiments), the use of two datasets and five MIL methods, and the release of code (anonymous link). The central mechanism, however, rests on an untested assumption: that validation AUC after only K=5 epochs reliably selects the models that will yield the best fully-trained merged model. The paper does not provide the control experiment that would isolate this selection effect from plain weight averaging. The variability-reduction results in Table 1 are encouraging, but the headline claim is weakened by an overstatement in Section 3.3 that is directly contradicted by the same table (Ensemble has lower STD than Soup3/Ties3 for TransMIL on Camelyon16).

major comments (4)

[Section 3.2 / Table 1] The paper claims in Section 3.3 that 'our proposed methods obtain more stable results across all MIL methods and datasets, having the smallest STD.' This is contradicted by Table 1: for TransMIL on Camelyon16, Ensemble has STD 0.9 while Soup3 has 1.4 and Ties3 1.5; on BRACS, Ensemble has 1.2 versus 1.5 for both proposed methods. The overclaim should be corrected to say that the proposed methods generally reduce STD compared to Baseline and LR tuned, not that they achieve the smallest STD overall.
[Section 2 / Section 3.2] The core mechanism is the selection of the top T models based on validation AUC after K=5 epochs, but no experiment isolates this selection step. There is no comparison against a randomly chosen T=3 subset, against merging all M=10 models at K=5, or against a single model trained for the same total budget of 150 epochs. The ablations in Table 2 vary K and T but always use validation-based selection, so they do not indicate whether the observed variance reduction comes from the early-validation ranking or simply from weight averaging plus extra warm-up epochs. This control is necessary to support the claimed multi-fidelity selection mechanism.
[Section 2 / Table 2] The premise that validation AUC after K=5 epochs reliably ranks the quality of fully-trained merged models is untested. At K=5, models are in an early, high-loss regime, and rankings across shuffling seeds may be noisy. The paper should provide evidence on the correlation between early validation AUC and final test AUC (or final merged-model test AUC) for the five MIL methods. Without this, the choice of K=5 is an assumption rather than a validated design decision.
[Section 3.2 / Table 1] The comparison against Baseline and LR tuned is confounded by unequal training budgets: Baseline uses 100 epochs, LR tuned 600, and Soup3/Ties3 150. Since the proposed method also benefits from an effective warm-start through partial training, it is unclear how much of the improvement is due to the fusion/selection mechanism versus the larger total number of optimization steps. A same-budget control (e.g., a single model trained for 150 or 600 epochs) would clarify the source of the improvement.

minor comments (5)

[Keywords] The keyword 'Variadion Reduction' contains a typo; it should read 'Variance Reduction'.
[Section 3.1] The phrase 'F eature Extraction' in the section heading has an unusual space; it should be 'Feature Extraction'.
[Section 2] The notation for the dataset S and its elements is inconsistently typeset (e.g., 'S= X1, X2, . . . XN' and 'Y = {y1, y2, . . . , yN }'); please use consistent set notation and fix the missing braces around the bag definition.
[Section 3.2] The ablation study in Table 2 reports single AUC values without any variance or repeated-run information, which makes it impossible to judge whether differences between K=3, K=5, and K=10 are significant.
[Section 3.2] The text says 'For LR tuned we perform a supplementary grid search over 6 learning rates' but does not list which learning rates were searched; this information is needed for reproducibility.

Circularity Check

1 steps flagged · score 2.0 of 10

No derivation-level circularity; the only self-referential step is choosing K=5/T=3 on the same Val/Test data that is later reported as the headline configuration.

fitted input called prediction [Section 3.3 Discussion; Table 2 ablation; Fig. 3; Table 1 main results]
"To keep the computational burden low, while preserving a good performance, we chose the combination K = 5 and T = 3, which gives almost always the best or second best performances on the Val and Test set (using M = 10). This means that aggregating 3 models out of 10 seems to be a good compromise between stability and performance. This is why we chose T = 3 in Table 1."

The paper first uses Table 2 (Val and Test columns) and Fig. 3 (test AUC vs T) to select the hyperparameters K=5 and T=3, explicitly choosing the values that give the best or second best performances on the Val and Test set. It then presents Table 1, which reports Soup3/Ties3 results using exactly K=5 and T=3, as evidence that the proposed configuration improves mean/min/STD over baselines. The reported comparison is therefore not an independent confirmation: the headline configuration was selected on the same evaluation data that is later quoted as the result. This is a mild form of fitting the evaluation, though it does not make the weight-averaging mechanism itself circular because that mechanism is external to the measured outcome.

full rationale

The paper contains no equation-level derivation whose output equals its input, no self-citations (the cited model-soup and TIES methods are external prior work), and no imported uniqueness theorem. The central claim that partial training followed by top-T weight averaging reduces run-to-run variability is an empirical result: the variance reduction is measured on test AUC and is not defined in terms of the selection criterion. The only self-referential step is the hyperparameter choice: K=5 and T=3 are chosen after inspecting Val and Test AUCs in Table 2 and Fig. 3 on the same datasets, then the same configuration is used to produce the headline Table 1 results. This is a test-set selection issue that partially weakens the evidential force of the comparison, but it does not reduce the method's mechanism to its inputs. The missing control (validation-selected T=3 vs random T=3 vs full averaging at equal budget) concerns whether the multi-fidelity selection is load-bearing; that is an empirical validity question, not circularity, so it does not raise the score beyond 2.

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

The paper introduces no new physical or conceptual entities. Its contribution is a training protocol, not a new latent variable, force, or conserved quantity.

free parameters (3)

M (number of partially trained models) = 10
Chosen by the authors as a default; no ablation on M is reported, so it is a hand-set cost and sample tradeoff.
K (partial training epochs before selection) = 5
Selected from Table 2 ablation, where K=10 is sometimes better on test but costs more epochs; the authors adopt K=5 as a compromise.
T (number of top models merged) = 3
Selected from Fig. 3 and Table 2 as the best compromise; the paper states T=3 seems a good choice, based on validation and test scores on the same datasets.

assumptions (4)

domain assumption Early validation AUC after K epochs is a reliable predictor of final model quality.
The whole selection step depends on this; introduced in Section 2 and validated only empirically through the ablations.
ad hoc to paper Averaging weights of partially trained models and then continuing full training is a valid and beneficial operation.
Model soups and TIES were designed for fully trained models; this paper applies them after K=5 epochs without a formal justification that gradient descent on the averaged weights inherits the soup benefit.
domain assumption The dominant sources of run-to-run variability in MIL are initialization seed, shuffle seed, and learning rate.
Used to justify the experimental design in Section 1 and Section 3.1; other factors such as feature extractor, data splits, and optimizer choice are fixed.
domain assumption The frozen pre-trained feature encoder f_phi removes encoder variability by assumption.
Section 2 states the authors ignore encoder variability and focus only on g_theta and c_theta; this restricts the scope of the conclusion.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Reducing Variability of Multiple Instance Learning Methods for Digital Pathology." pith.science (2026). https://pith.science/paper/VMKGMYJW

@misc{pith2026250700292,
  author       = {Pith},
  title        = {Pith review of: Reducing Variability of Multiple Instance Learning Methods for Digital Pathology},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VMKGMYJW}},
  note         = {Machine review of arXiv:2507.00292}
}

read the original abstract

Digital pathology has revolutionized the field by enabling the digitization of tissue samples into whole slide images (WSIs). However, the high resolution and large size of WSIs present significant challenges when it comes to applying Deep Learning models. As a solution, WSIs are often divided into smaller patches with a global label (\textit{i.e., diagnostic}) per slide, instead of a (too) costly pixel-wise annotation. By treating each slide as a bag of patches, Multiple Instance Learning (MIL) methods have emerged as a suitable solution for WSI classification. A major drawback of MIL methods is their high variability in performance across different runs, which can reach up to 10-15 AUC points on the test set, making it difficult to compare different MIL methods reliably. This variability mainly comes from three factors: i) weight initialization, ii) batch (shuffling) ordering, iii) and learning rate. To address that, we introduce a Multi-Fidelity, Model Fusion strategy for MIL methods. We first train multiple models for a few epochs and average the most stable and promising ones based on validation scores. This approach can be applied to any existing MIL model to reduce performance variability. It also simplifies hyperparameter tuning and improves reproducibility while maintaining computational efficiency. We extensively validate our approach on WSI classification tasks using 2 different datasets, 3 initialization strategies and 5 MIL methods, for a total of more than 2000 experiments.

Figures

Figures reproduced from arXiv: 2507.00292 by the authors.

**Figure 1.** Top: Violin plots on 2 different datasets. Each dot represents the test AUC of a model trained with a random shuffle and initialization seed and with learning rate tuned on the validation set. Bottom: we apply the proposed method (using Soup for MaxMIL and ABMIL and Ties for DSMIL, CLAM and TransMIL) using the same initialization seeds as in the Top figure and M = 10, K = 5 and T = 3. The proposed method clearly red… view at source ↗

**Figure 2.** Pipeline of the proposed Multi-Fidelity Model Averaging method for Whole Slide Image Classification. whose most important hyper-parameters are: initialization seed, shuffling seed and learning rate. By changing one of them, results may drastically vary. Model Overview In our work, we propose combining Multi-Fidelity with Model Averaging to mitigate variability and increase robustness (see Fig.2). Each slide is first… view at source ↗

**Figure 3.** Ablation Study on the effect of the T to the average AUC score across 5 MILs on Camelyon16 and BRACS datasets It’s also important to notice that the proposed methods have a similar or better performance than Ensemble and Best on VAL models, but they require almost 7 times less training epochs and they are 10 times faster at inference. Eventually, it’s worth mentioning that Soup3/Ties3 improve the minimum results, … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 23 canonical work pages

[1]

In: Merlo, P., Tiedemann, J., Tsarfaty, R

Belz, A., Agarwal, S., Shimorina, A., Reiter, E.: A systematic review of repro- ducibility research in natural language processing. In: Merlo, P., Tiedemann, J., Tsarfaty, R. (eds.) Proceedings of the 16th Conference of the European Chap- ter of the Association for Computational Linguistics: Main Volume. pp. 381–393. Association for Computational Linguist...

work page 2021
[2]

In: International Conference on Machine Learning

Bouthillier, X., Laurent, C., Vincent, P.: Unreproducible research is reproducible. In: International Conference on Machine Learning. pp. 725–734. PMLR (2019)

work page 2019
[3]

Database: The Journal of Biological Databases and Curation 2022, baac093 (Oct 2022)

Brancati, N., Anniciello, A.M., Pati, P., Riccio, D., Scognamiglio, G., Jaume, G., De Pietro, G., Di Bonito, M., Foncubierta, A., Botti, G., Gabrani, M., Feroce, F., Frucci, M.: BRACS: A Dataset for BReAst Carcinoma Subtyping in H&E Histology Images. Database: The Journal of Biological Databases and Curation 2022, baac093 (Oct 2022)

work page 2022
[4]

In: Proceedings of the IEEE/CVF international conference on computer vision

Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650–9660 (2021)

2021
[5]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)

Chan, T.H., Cendra, F.J., Ma, L., Yin, G., Yu, L.: Histopathology whole slide image analysis with heterogeneous graph representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)

work page 2023
[6]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16144–16155 (2022)

work page 2022
[7]

Journal of Machine Learning Research 23(56), 1–6 (2022)

Costa, V.G.T.d., Fini, E., Nabi, M., Sebe, N., Ricci, E.: solo-learn: A Library of Self-supervised Methods for Visual Representation Learning. Journal of Machine Learning Research 23(56), 1–6 (2022)

work page 2022
[8]

Egele, R., Guyon, I., Sun, Y., Balaprakash, P.: Is one epoch all you need for multi- fidelity hyperparameter optimization? In: 31st European Symposium on Artifi- cial Neural Networks, Computational Intelligence and Machine Learning, ESANN (2023)

work page 2023

Show all 29 references

[9]

JAMA318(22) (2017)

Ehteshami Bejnordi, B., Veta, M., Johannes van Diest, P., van Ginneken, B., Karssemeijer,N.,Litjens,G.,vanderLaak,J.A.W.M.,theCAMELYON16Consor- tium: Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer. JAMA318(2...

2017
[10]

Advances in Computa- tional Science and Engineering1(4), 351–400 (2023)

Fernández-Godino, M.G.: Review of multi-fidelity models. Advances in Computa- tional Science and Engineering1(4), 351–400 (2023)

2023
[11]

arXiv preprint arXiv:2204.07610 (2022)

Gundersen, O.E., Coakley, K., Kirkpatrick, C., Gil, Y.: Sources of irreproducibility in machine learning: A review. arXiv preprint arXiv:2204.07610 (2022)

2022 arXiv
[12]

In: Proceedings of the AAAI conference on artificial intelligence (2018)

Henderson, P., Islam, R., Bachman, P., Pineau, J., Precup, D., Meger, D.: Deep reinforcement learning that matters. In: Proceedings of the AAAI conference on artificial intelligence (2018)

2018
[13]

In: Proceedings of the 35th International Conference on Machine Learning

Ilse, M., Tomczak, J., Welling, M.: Attention-based Deep Multiple Instance Learn- ing. In: Proceedings of the 35th International Conference on Machine Learning. pp. 2127–2136. PMLR (Jul 2018), iSSN: 2640-3498

2018
[14]

Song, A., J

Jaume, G., Vaidya, A., Zhang, A., H. Song, A., J. Chen, R., Sahai, S., Mo, D., Madrigal, E., Phi Le, L., Mahmood, F.: Multistain pretraining for slide represen- tation learning in pathology. In: European Conference on Computer Vision. pp. 19–37. Springer (2024) 10 A. Mammadov et al

2024
[15]

In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

Kang, M., Song, H., Park, S., Yoo, D., Pereira, S.: Benchmarking Self-Supervised Learning on Diverse Pathology Datasets. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 3344–3354 (2023)

2023
[16]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Lazard, T., Lerousseau, M., Decencière, E., Walter, T.: Giga-ssl: Self-supervised learning for gigapixel images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4305–4314 (2023)

2023
[17]

In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition

Li, B., Li, Y., Eliceiri, K.W.: Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14318–14328 (2021)

2021
[18]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Li, J., Chen, Y., Chu, H., Sun, Q., Guan, T., Han, A., He, Y.: Dynamic graph rep- resentation with knowledge-aware attention for histopathology whole slide image analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11323–11332 (2024)

2024
[19]

Nature biomedical engineering 5(6), 555–570 (2021), publisher: Nature Publishing Group UK London

Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering 5(6), 555–570 (2021), publisher: Nature Publishing Group UK London

2021
[20]

Advances in neural information processing systems31 (2018)

Lucic, M., Kurach, K., Michalski, M., Gelly, S., Bousquet, O.: Are gans created equal? a large-scale study. Advances in neural information processing systems31 (2018)

2018
[21]

Journal of ma- chine learning research22(164), 1–20 (2021)

Pineau, J., Vincent-Lamarre, P., Sinha, K., Larivière, V., Beygelzimer, A., d’Alché Buc, F., Fox, E., Larochelle, H.: Improving reproducibility in machine learning research (a report from the neurips 2019 reproducibility program). Journal of ma- chine learning research22(164),...

2021
[22]

Advances in neural information processing systems34, 2136–2147 (2021)

Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., others: Transmil: Trans- former based correlated multiple instance learning for whole slide image classifica- tion. Advances in neural information processing systems34, 2136–2147 (2021)

2021
[23]

Medical Image Analysis97, 103252 (2024)

Wang, Z., Ma, J., Gao, Q., Bain, C., Imoto, S., Liò, P., Cai, H., Chen, H., Song, J.: Dual-stream multi-dependency graph neural network enables precise cancer survival analysis. Medical Image Analysis97, 103252 (2024)

2024
[24]

In: International conference on machine learning

Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., et al.: Model soups: averagingweightsofmultiplefine-tunedmodelsimprovesaccuracywithoutincreas- ing inference time. In: International conf...

2022
[25]

Advances in Neural Information Processing Systems 36 (2024)

Yadav, P., Tam, D., Choshen, L., Raffel, C.A., Bansal, M.: Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems 36 (2024)

2024
[26]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Yang, S., Wang, Y., Chen, H.: Mambamil: Enhancing long sequence modeling with sequence reordering in computational pathology. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 296–306. Springer (2024)

2024
[27]

In: International Conference on Machine Learn- ing

Zbontar, J., Jing, L., Misra, I., LeCun, Y., Deny, S.: Barlow twins: Self-supervised learning via redundancy reduction. In: International Conference on Machine Learn- ing. pp. 12310–12320. PMLR (2021)

2021
[28]

In: European Conference on Computer Vision

Zhang, Y., Li, H., Sun, Y., Zheng, S., Zhu, C., Yang, L.: Attention-challenging mul- tiple instance learning for whole slide image classification. In: European Conference on Computer Vision. pp. 125–143. Springer (2024) Reducing Variability of MIL Methods 11

2024
[29]

IEEE transactions on medical imaging41(11), 3003–3015 (2022)

Zheng, Y., Gindra, R.H., Green, E.J., Burks, E.J., Betke, M., Beane, J.E., Ko- lachalama, V.B.: A graph-transformer for whole slide image classification. IEEE transactions on medical imaging41(11), 3003–3015 (2022)

2022

Pith tools

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

[1] [1]

In: Merlo, P., Tiedemann, J., Tsarfaty, R

Belz, A., Agarwal, S., Shimorina, A., Reiter, E.: A systematic review of repro- ducibility research in natural language processing. In: Merlo, P., Tiedemann, J., Tsarfaty, R. (eds.) Proceedings of the 16th Conference of the European Chap- ter of the Association for Computational Linguistics: Main Volume. pp. 381–393. Association for Computational Linguist...

work page 2021

[2] [2]

In: International Conference on Machine Learning

Bouthillier, X., Laurent, C., Vincent, P.: Unreproducible research is reproducible. In: International Conference on Machine Learning. pp. 725–734. PMLR (2019)

work page 2019

[3] [3]

Database: The Journal of Biological Databases and Curation 2022, baac093 (Oct 2022)

Brancati, N., Anniciello, A.M., Pati, P., Riccio, D., Scognamiglio, G., Jaume, G., De Pietro, G., Di Bonito, M., Foncubierta, A., Botti, G., Gabrani, M., Feroce, F., Frucci, M.: BRACS: A Dataset for BReAst Carcinoma Subtyping in H&E Histology Images. Database: The Journal of Biological Databases and Curation 2022, baac093 (Oct 2022)

work page 2022

[4] [4]

In: Proceedings of the IEEE/CVF international conference on computer vision

Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650–9660 (2021)

2021

[5] [5]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)

Chan, T.H., Cendra, F.J., Ma, L., Yin, G., Yu, L.: Histopathology whole slide image analysis with heterogeneous graph representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)

work page 2023

[6] [6]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16144–16155 (2022)

work page 2022

[7] [7]

Journal of Machine Learning Research 23(56), 1–6 (2022)

Costa, V.G.T.d., Fini, E., Nabi, M., Sebe, N., Ricci, E.: solo-learn: A Library of Self-supervised Methods for Visual Representation Learning. Journal of Machine Learning Research 23(56), 1–6 (2022)

work page 2022

[8] [8]

Egele, R., Guyon, I., Sun, Y., Balaprakash, P.: Is one epoch all you need for multi- fidelity hyperparameter optimization? In: 31st European Symposium on Artifi- cial Neural Networks, Computational Intelligence and Machine Learning, ESANN (2023)

work page 2023