acceptodds
Under review as a conference paper at ICLR 2027

Quantile-Domain Multiple Instance Learning for Whole-Slide Image Classification

Abstract

Whole-slide images (WSIs) contain billions of pixels, making end-to-end modeling costly in computation and storage. Their classification is therefore typically formulated as Multiple Instance Learning (MIL) over patch features. However, some methods construct bag-level context with positional structures, neighborhoods, or approximations that depend on patch indices. Their predictions can therefore change with patch order. We show that patch interactions need not cause this sensitivity and that it can instead arise from index-addressed operators. Based on this insight, we introduce Quantile-Domain Multiple Instance Learning (QDMIL), a three-stage framework. First, it forms a fixed-size, permutation-invariant quantile grid over multiple learnable projections. This grid captures the slide-wide distribution of morphological features. Second, QDMIL uses the grid to derive soft quantile coordinates for each patch and bag-shared affine modulation. Together, these signals contextualize the patch representations. Finally, QDMIL fuses attention-pooled local discriminative evidence with a separate grid encoding for slide classification. We establish sufficient operator-level conditions for permutation invariance and prove that QDMIL satisfies them under deterministic inference and exact arithmetic. Against 15 baselines across four datasets, QDMIL achieves the highest mean accuracy and F1-Score in all eight dataset–metric comparisons. Paired order-control experiments across seven methods further support distribution-relative patch positioning as a general principle for permutation-invariant bag-level context.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
MIL, WSI, Quantile, Domain, Classification

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2UfNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper analyzes patch-order sensitivity by reordering feature indices. In terms of design, the paper presents K learnable projection directions and samples M empirical quantiles in each direction to achieve permutation-invariant learning for WSI classification. The proposed method outperforms SOTA methods on both binary and multi-class classification datasets.

Critical strengths
  1. I appreciate that the authors highlight that order sensitivity is not inherent to contextual interactions at the WSI level.

  2. The novelty of the paper lies in projecting patch features along K learnable directions and integrating them into gated attention pooling for classification.

  3. At the decision level, the separation between attention-pooled local evidence and distributional information is sensible for heterogeneous WSIs.

Critical weaknesses
  1. Equation (4) computes soft quantile coordinates using distances between projected scores and quantile values. However, these coordinates are sensitive to score scales, even when patches have the same quantile ranks. Why do the authors use this approach instead of exact ranks or normalized scores? A comparison with these alternatives would help justify the proposed design.

  2. The distinctive patch-positioning mechanism has a limited isolated gain. Removing FiLM leads to a larger accuracy drop than removing quantile coordinates on both ablation datasets. This suggests that global bag conditioning may explain much of the improvement. The current ablation does not fully distinguish quantile-relative positioning from a simpler globally conditioned attention or FiLM model (Table 2).

  3. Although QDMIL avoids quadratic attention, its N×K×M soft-assignment tensor introduces a memory bottleneck for large bags. For 10,000 patches, QDMIL uses 210 MiB of memory and 1.68M parameters, compared with 71 MiB and 0.33M parameters for ABMIL.

Questions for authors

See the weakness

3 comments
AuthorsOpenReview ↗

We added coordinate and global conditioning comparisons and a chunked implementation to clarify QDMIL’s benefits and memory cost.

Q1. Scale sensitivity of soft quantile coordinates

Thank you for noting the scale dependence. Scaling projected scores changes the effective temperature in Eq. (4). We compared the original coordinates with normalized soft coordinates and exact empirical ranks while keeping the rest of the model fixed. The original form leads both metrics on KIRP and PANDA in the table below. The pairwise differences are not significant. A direct standardized score control on KIRP also trails at 0.8729 ACC and 0.8127 F1-Score. We retain soft coordinates because they are differentiable and preserve distances to slide quantiles, whereas exact ranks discard those distances. The normalized soft form provides a scale invariant alternative, which we will include in the revision.

CoordinatesKIRP ACCKIRP F1-ScorePANDA ACCPANDA F1-Score
Soft, QDMIL0.88780.83020.76280.7248
Normalized soft0.86910.80580.76160.7205
Exact ranks0.87290.81700.75810.7191

Q2. Quantile positioning and global conditioning

The component ablation shows a larger effect from FiLM than from quantile coordinates. To examine the role of global conditioning, we replaced the quantile grid with patch mean and standard deviation while retaining the FiLM, bag encoder, attention pooling, and classifier. We also tested this simpler model with quantile coordinates. QDMIL leads both variants on KIRP, Brain Mets, and BRACS, as the table below shows. The cohortwise differences are not significant, but the direction is consistent across these three large bag datasets. PANDA favors Global FiLM by 1.44 ACC points, with Holm corrected p=0.0011p=0.0011. The quantile design also improves pathology localization. On 1,000 randomly sampled PANDA slides with tumor annotations, QDMIL raises the tumor fraction among the top 10% attended patches from 0.547 to 0.566 relative to Global FiLM. Tumor patch AUROC rises from 0.856 to 0.871, with the coordinate variant between them on both measures. We will revise the contribution claim to distinguish the classification benefit of global conditioning from the localization benefit of quantile positioning.

ModelParams.KIRP ACCBrain Mets ACCBRACS ACCPANDA ACC
Global FiLM1.45M0.86920.91980.82750.7772
Global FiLM with quantile coordinates1.63M0.86540.92850.83010.7756
QDMIL1.68M0.88780.93720.84530.7628

Q3. Memory and parameter cost

Thank you for identifying the dense assignment tensor as a memory bottleneck. Each patch's assignment is independent, which lets us compute it in chunks with activation recomputation. This preserves the output without storing the full N×K×MN\times K\times M tensor. At 100,000 patches, peak memory falls from 1,760 to 1,077 MiB for inference and from 3,822 to 2,182 MiB for training. With chunks of 4,096 patches, inference takes 12.9 ms rather than 14.2 ms for the dense version, while training takes 34.3 ms rather than 32.7 ms. Chunking leaves 1.68M parameters unchanged, mostly in the grid encoding networks, whose size does not depend on bag size. On KIRP, a 1.72M parameter gated attention control reaches 0.8614 ACC versus 0.8878 for QDMIL. We will use the chunked implementation in the revision. The table below reports peak memory in MiB under the same profiling setup.

PatchesGated attention inference / trainingDense QDMIL inference / trainingChunked QDMIL inference / training
8,00082 / 138165 / 340116 / 229
16,000145 / 258302 / 637195 / 400
100,000803 / 1,4891,760 / 3,8221,077 / 2,182
Reviewer_2UfNOpenReview ↗

Responses to Authors

Thank you for the detailed response and the additional experiments. I appreciate that you also reported results that do not favor your method.

W3 (memory): The chunked implementation and the parameter-matched control resolve my concern.

W1 (coordinates): Thank you for the comparison. The three coordinate forms perform similarly, so please state in the revision that the specific form of Eq. (4) is not critical, and include the normalized soft variant as you proposed.

W2 (global conditioning): The new comparison is helpful. I have a few remaining questions:

On PANDA, Global FiLM with mean and standard deviation outperforms QDMIL (0.7772 vs. 0.7628, p = 0.0011). You explain this by the small bag size, but Appendix E.3.4 reports that QDMIL’s gains are largest in the smallest-bag quartile. Could you clarify how these fit together? The BRACS value for QDMIL in the new tables (0.8453) seems to be the M=65 setting rather than the shared configuration (0.8427). Which configuration was used for these comparisons?

I also suggest revising the abstract and conclusion so that the classification benefit of global conditioning and the benefit of quantile positioning are stated separately, as you already proposed. However, I remain positive about the paper.

AuthorsOpenReview ↗

We clarified the PANDA quartile results and BRACS configurations, and revised the manuscript to distinguish global conditioning from quantile positioning.

1. Main text revisions

Thank you for your positive assessment and for suggesting a clearer separation of the two contributions. We revised the abstract to state separately that global conditioning contributes strongly to classification and that quantile positioning improves attention alignment with annotated tumor tissue. The conclusion makes the same distinction and specifies that the annotation evidence comes from PANDA. Section 4.3 now introduces the controlled comparisons and guides readers to the appendix. Most new experimental results and configuration details appear in Appendices E.3.7, E.3.10, and E.3.11. These revisions are marked in blue.

2. PANDA quartiles

Thank you for pointing out the apparent tension in the PANDA quartile results. The quartile gains compare each model with DyHG. The 0.7772 versus 0.7628 result instead compares Global FiLM directly with QDMIL. In the smallest quartile, both methods have their largest gains over DyHG, with Global FiLM gaining more. Global FiLM also exceeds QDMIL in every quartile. The revised Appendix E.3.11 therefore treats the PANDA preference for the moment summary as a dataset specific result. It does not attribute that difference to bag size, since the quartiles may also differ in class composition.

PANDA bag quartileQDMIL over DyHG, ACC pointsGlobal FiLM over DyHG, ACC pointsGlobal FiLM over QDMIL, ACC points
Q1, smallest+2.93+4.66+1.73
Q2+1.66+3.25+1.59
Q3+1.25+1.82+0.57
Q4, largest+0.38+2.26+1.88

3. BRACS configuration and matched comparisons

Thank you for noticing the differing BRACS values. The 0.8453 BRACS ACC comes from the additional matched comparison with K=64K=64, M=65M=65, and T0=0.2T_0=0.2. The shared configuration in the main comparison uses K=64K=64, M=17M=17, and T0=0.2T_0=0.2 and gives 0.8427. Global FiLM with coordinates matches QDMIL's grid settings in each additional comparison. Global FiLM itself has no grid settings. For PANDA, the additional QDMIL run uses T0=0.05T_0=0.05 and gives 0.7628, while the shared configuration gives 0.7592. Appendices E.3.7 and E.3.10 and their captions now identify these runs explicitly.

The matched results below show QDMIL leading both moment controls on TCGA-KIRP, Brain-Mets, and BRACS. PANDA is the exception. There, Global FiLM reaches 0.7772 versus 0.7628 for QDMIL, with a Holm adjusted paired McNemar p=0.0011p=0.0011.

DatasetQDMILGlobal FiLMGlobal FiLM with coordinatesQDMIL K,M,T0K,M,T_0
TCGA-KIRP0.88780.86920.865464, 17, 0.2
Brain-Mets0.93720.91980.928564, 17, 0.2
BRACS0.84530.82750.830164, 65, 0.2
PANDA0.76280.77720.775664, 17, 0.05
Reviewer_CNzTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

QDMIL is a MIL aggregation head that summarizes each slide by the distribution of its patch features and uses that summary to modulate each patch. It projects the patch features onto a set of learned directions and sorts the projections to get a fixed-size grid of quantiles for each slide. Each patch then gets a soft position inside its own slide's quantile grid, i.e., where it falls relative to the other patches. The grid is also used to rescale and shift the patch features, so each patch is read relative to the slide it comes from. The final prediction combines gated-attention pooling with a separate encoding of the quantile grid. The paper also gives sufficient conditions for permutation invariance and a bound on how far the quantile grid can be from the true per-direction distribution. Evaluation covers TCGA-KIRP, PANDA, BRACS and Brain-Mets against 15 baselines, using CONCH features, five folds and a single seed.

Critical strengths
  1. The idea is clean and cheap: O(Nlog⁡N)O(N \log N), about 1.37 ms for N=104N = 10^4, and no N×NN \times N matrix. Using the within-bag distribution as a reference frame for each patch is intuitive for pathology ("is this pattern rare or common in this slide?").

  2. App. C is a nice, careful experiment. It compares random permutations against repeat controls and separates TransMIL/DGR-MIL index dependence from DyHG's inference-time stochasticity. GVLMIL shows that interaction alone isn't the problem.

  3. The PANDA analysis is thorough: per-grade F1, per-fold paired gains, bag-size quartiles, and ordinal errors / QWK.

  4. The KIRP replacement study (mean, or mean+max+SD, with the same shape) is the right control for showing that the quantile grid itself matters.

Critical weaknesses
  1. The main table appears to be tuned on test. There's no validation split, and Table 1 reports the best point of the Table 8 sweep rather than the reference config (T0=0.2T_0 = 0.2 for KIRP, T0=0.05T_0 = 0.05 for PANDA, M=65M = 65 for BRACS), while each baseline gets a single default run.

  2. The margins on small cohorts are within noise. Results come from a single seed, and the fold std is about 0.040.04–0.070.07 on KIRP and BRACS. Brain-Mets has only 111 slides, so the margin over CLAM-MB amounts to about 2 slides. BRACS is also split at slide level, which may leak patients across folds.

  3. The closest prior work is missing. The method is essentially sliced-Wasserstein / sort-based set pooling (PSWE, FSPool), but it isn't positioned against or compared with these. The permutation-invariance result is standard.

  4. The ablations are thin. They cover only KIRP and PANDA, and most of the KIRP gaps are within one std. The attention maps are hand-picked examples.

Questions for authors
  1. What are the Table 1 numbers at a single config (K=64K = 64, M=17M = 17, one T0T_0) chosen without test data? How is the reported checkpoint chosen?
  2. How were the baseline hyperparameters set? Were any of them tuned?
  3. Can you run multiple seeds, or report paired tests, for KIRP, BRACS and Brain-Mets?
  4. How does QDMIL compare with sliced-Wasserstein / sort-based pooling (PSWE, FSPool) used as the slide readout?
  5. Why are BRACS folds at slide level, and why 394 slides rather than the full release?
1 comment
AuthorsOpenReview ↗

We clarified the evaluation protocol and BRACS split, and added shared configuration results, paired tests, and PSWE and FSPool comparisons.

Q1. Single configuration and checkpoint selection

Thank you for identifying this issue. The original QDMIL row in Table 1 selected the best ACC point from the parameter sweep on each dataset. We will replace it with one configuration across all datasets, K=64K=64, M=17M=17, and T0=0.2T_0=0.2. Under this setting, QDMIL leads seven of eight comparisons with the strongest existing baselines shown below. With no separate validation split, all methods selected their best epoch on the evaluated fold. We will state this checkpoint rule and its test dependence explicitly.

DatasetQDMIL ACC / F1-ScoreBest baseline ACC / F1-ScoreQDMIL ranks
KIRP0.8878 / 0.8302DGR MIL 0.8693 / CLAM MB 0.82571 / 1
PANDA0.7592 / 0.7197DyHG 0.7471 / 0.70461 / 1
BRACS0.8427 / 0.7674RankMix 0.8428 / CLAM MB 0.75372 / 1
Brain Mets0.9372 / 0.9147GVLMIL 0.9285 / 0.88521 / 1

Q2. Baseline hyperparameters

All baselines used the same CONCH features and five folds. Their structural settings were fixed across datasets, with no dataset specific tuning. We used cross entropy loss, Adam with learning rate 2×10−42\times10^{-4} and weight decay 10−510^{-5}, and a maximum of 100 epochs. Most baselines used the framework StepLR schedule, while GVLMIL and QDMIL used its cosine schedule. As described in Q1, the revised main comparison will use one QDMIL configuration. The KK, MM, and T0T_0 sweep will serve only to assess sensitivity.

Q3. Multiple seeds or paired tests

Thank you. Using the original Table 1 predictions, we paired each slide across methods and applied exact McNemar tests and a fold stratified bootstrap with 5,000 resamples, followed by Holm correction. On PANDA, QDMIL significantly exceeds all 14 baselines with available predictions in both metrics. Its ACC advantage over DyHG remains significant under the shared Q1 configuration at 1.25 points with p<0.001p<0.001. On KIRP, BRACS, and Brain Mets, the confidence intervals against the strongest baseline include zero, so we describe those results as comparable. The tests quantify uncertainty from test slides, not training randomness.

DatasetBest ACC baselineΔ\DeltaACC in points, 95% CIΔ\DeltaF1-Score in points, 95% CIHolm corrected pp
PANDADyHG+1.65 [0.8, 2.5]+1.98 [1.0, 3.0]<0.001
KIRPDGR MIL+1.86 [-1.9, 5.6]+0.49 [-4.9, 5.7]1.00
BRACSRankMix+0.25 [-2.5, 3.0]+2.50 [-2.9, 8.1]1.00
Brain MetsGVLMIL+0.87 [-2.7, 4.5]+2.94 [-5.1, 11.2]1.00

Q4. PSWE and FSPool slide readouts

Thank you for pointing out PSWE and FSPool. We implemented them as slide readouts using the same patch embedding, folds, and training protocol, and reran QDMIL for this comparison. QDMIL gains 2.2 and 2.6 ACC points over FSPool and PSWE on KIRP. On PANDA, the gains are 2.5 and 2.4 points across all five folds, with p<0.001p<0.001. PSWE and FSPool use sorting for the slide readout. QDMIL also uses its grid to condition patches through coordinates and FiLM before attention pooling. We will cite both methods and clarify that permutation invariance itself is established.

ModelParams.KIRP ACC / F1-ScorePANDA ACC / F1-Score
FSPool MIL0.27M0.8655 / 0.79880.7383 / 0.6943
PSWE MIL0.55M0.8617 / 0.78940.7390 / 0.6941
QDMIL1.68M0.8878 / 0.83020.7628 / 0.7248

Q5. BRACS folds and dataset size

We used the official BRACS training partition for the same five fold evaluation protocol as the other datasets. The official validation and test partitions were not part of this study. Folds were stratified by slide label rather than patient ID. Our audit found that 19 patients span folds, affecting 121 of 394 test slides. We will disclose this limitation and describe BRACS performance as comparable to the strongest baselines. The official training partition contains 395 slides, but BRACS_1003731 repeatedly failed local image processing and yielded no CONCH features. We therefore evaluated the remaining 394 slides.

Reviewer_uVmWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Quantile-Domain Multiple Instance Learning (QDMIL), a permutation-invariant framework for whole-slide image classification. QDMIL constructs a fixed-size quantile grid using multiple learnable projections of patch embeddings. This grid serves two purposes: it provides distribution-relative coordinates for individual patches and a global representation of the slide's feature distribution. The patch representations are updated using soft quantile coordinates and FiLM-style modulation, while the final classification combines attention-pooled patch features with a separate quantile-grid encoding. The authors provide theoretical guarantees of permutation equivariance and invariance and evaluate the method on four pathology datasets.

Critical strengths
  • The paper identifies an important distinction between patch interactions and index-dependent contextual operations. It correctly emphasizes that interactions among patches do not inherently violate permutation invariance.

  • QDMIL achieves the highest mean accuracy and F1-score on all four evaluated datasets under a shared feature encoder and evaluation protocol. The experiments include a broad collection of MIL baselines, component ablations, parameter sensitivity analyses, and computational comparisons.

Critical weaknesses
  • The main experimental comparison does not sufficiently isolate the benefits of quantile-domain conditioning against alternative learned set representations with comparable model capacity. For example, a controlled comparison against feature-wise sorting, moment-based distribution summaries, or a learned Deep Sets representation used with the same conditioning and fusion architecture would be informative. Such experiments would help establish whether the improvements originate specifically from the proposed quantile representation or more generally from incorporating a learnable global distributional context.

  • The paper positions patch-order robustness as a central motivation. However, permutation invariance is already a well-established property of many MIL and set-learning architectures. The authors' own experiments demonstrate that ABMIL, ACMIL, and GVLMIL also maintain effectively invariant predictions under patch reordering. In particular, GVLMIL shows that full patch-to-patch interactions can be implemented without introducing order sensitivity. Therefore, the observed order sensitivity of TransMIL and DGR-MIL appears to arise from particular implementation choices rather than a general limitation of contextual MIL. The authors should more clearly distinguish between identifying order-dependent implementations and introducing a fundamentally new approach to permutation-invariant contextual modeling.

Questions for authors
  • Could the authors compare QDMIL with a model that replaces the quantile grid with learned Deep Sets pooling, feature-wise sorting, or mean/variance-based distribution summaries while retaining the same patch-conditioning and fusion modules?

  • Could the authors provide parameter-matched comparisons with simpler global-context conditioning architectures, such as attention MIL augmented with a global summary and FiLM modulation?

  • Although permutation invariance is demonstrated, it is unclear how robust the quantile representation is when the same WSI is represented by different subsets of patches. Since the quantile grid is estimated from the empirical patch distribution, systematic patch removal or sampling changes may alter both the global representation and patch-level coordinates.

1 comment
AuthorsOpenReview ↗

We added controlled global representation comparisons and patch subset experiments, which support QDMIL’s advantages on large bags and clarify its dependence on bag size.

Q1. Alternative set representations

Thank you for proposing this comparison. We replaced only the summary sent to FiLM and the bag encoder. All else remained fixed. Deep Sets used MLP mean pooling, FSPool used 16 segment weights, and the moment variant used patch mean and standard deviation.

Global summaryParams.KIRP ACCKIRP F1-Score
Mean and standard deviation1.63M0.8654 ± 0.04750.8038 ± 0.0592
Deep Sets1.76M0.8728 ± 0.05170.7991 ± 0.1057
Featurewise sorting1.25M0.8691 ± 0.04280.8131 ± 0.0552
Quantile grid, QDMIL1.68M0.8878 ± 0.03890.8302 ± 0.0576

QDMIL is the only model that leads in both metrics. It exceeds every replacement by 1.5 to 2.2 points in ACC and 1.7 to 3.1 points in F1-Score. This matters because KIRP represents the intended large bag setting, with thousands of patches per WSI. The contrasts are not significant after Holm correction on 267 slides, but their direction is consistent across all replacements and both metrics. Deep Sets even has slightly more parameters than QDMIL but does not match its performance. This pattern is therefore not explained by capacity or by an arbitrary invariant summary. Instead, it supports using the quantile grid within the shared conditioning and fusion architecture. The suggested alternatives also motivate future work on hybrid distribution summaries.

Q2. Global conditioning with comparable capacity

Thank you for suggesting this comparison. Global FiLM uses patch mean and standard deviation with the same FiLM and bag encoder as QDMIL. A second variant adds quantile coordinates to test their contribution. Both have parameter counts close to QDMIL and were evaluated on all four datasets.

ModelParams.KIRP ACCBrain Mets ACCBRACS ACCPANDA ACC
Global FiLM1.45M0.86920.91980.82750.7772
Global FiLM with quantile coordinates1.63M0.86540.92850.83010.7756
QDMIL1.68M0.88780.93720.84530.7628

QDMIL leads both Global FiLM variants on all three large bag cohorts by 0.9 to 2.2 ACC points. The individual differences are not significant, but their consistent direction favors the full model in this setting. PANDA instead favors the moment summary by 1.3 to 1.4 points, with p≤0.006p\leq0.006 after Holm correction. Although PANDA has 10,611 slides, its median bag contains only 144 patches. This contrast suggests that the usefulness of empirical quantiles may depend on the number of instances in each slide. Global conditioning accounts for much of the classification gain. Quantiles add a distribution relative reference whose classification benefit may be greater for large bags.

Q3. Robustness to patch subsets

Thank you for raising this practical question. Without retraining, we retained 90%, 70%, 50%, or 30% of each test bag, using five shared subsets per ratio. Each cell reports QDMIL ACC and agreement with its full bag prediction. The last column gives the strongest baseline at 50% and 30% retention.

Dataset100%90%70%50%30%Best baseline, 50% and 30%
KIRP0.88780.8856 (99.8%)0.8856 (99.8%)0.8863 (99.9%)0.8826 (99.5%)DGR MIL, 0.8535 and 0.8483
BRACS0.84530.8423 (99.5%)0.8392 (98.9%)0.8306 (97.6%)0.8291 (97.5%)RankMix, 0.8275 and 0.8189
Brain Mets0.93720.9319 (99.5%)0.9301 (99.3%)0.9300 (99.3%)0.9159 (97.8%)GVLMIL, 0.8820 and 0.8765
PANDA0.76360.7568 (95.9%)0.7432 (91.1%)0.7227 (86.2%)0.6889 (79.2%)DyHG, 0.7103 and 0.6786

Subsampling changes the empirical bag, so exact invariance is neither expected nor claimed. Even so, QDMIL retains the highest ACC among all tested methods at every retention ratio. At 50%, ACC decreases by only 0.15 points on KIRP and 0.72 points on Brain Mets, with more than 99% prediction agreement. The decrease is 1.47 points on BRACS. PANDA decreases by 4.09 points because only about 72 patches remain, yet QDMIL still ranks first.

This pattern is consistent with the stability of the quantile grid. At 30% retention, normalized grid change remains below 0.8% on the three large bag cohorts but reaches 3.6% on PANDA. Empirical quantiles are therefore stable when thousands of patches remain and less stable for small bags. We will report this dependence and clarify that permutation invariance concerns reordering a fixed bag, not changes in its composition.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.