- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes a signed determinantal evidence pooling operator (DENO) for correlated MIL. In DENO, each instance gives a signed evidence and a similarity representation that describes its relationships to other instances. The authors show that for exact replicas the bag level evidence can be reduced to a computation over distinct instances. The paper includes experiments on 5 datasets, experiments with controlled nuisance multicity, and several ablations. DENO performs quite well in these experiments.
Critical strengths
- The paper is clearly written.
- The authors propose a principled MIL method for learning with redundant instances
- The authors include theoretical guarantees for exact replicas and near-redundant sources.
- Results are quite good for the proposed method.
Critical weaknesses
- No discussion of computational complexity. The proposed signed determinantal evidence pooling is instead of for the multi-head attention mechanism. How much does this impact training? Whole slide images are typically quite large (~10,000 patches) which is why Shao et al. (NeurIPS 2021) and following works used a nyström-based algorithm for approximating self-attention.
- The authors only compare to one other method that models interactions between instances (Set Transformer). All other methods do not model interactions between instances. The authors should compare to other methods that explicitly handle redundancy (Yao et al., 2026; Sun et al., 2026).
- Experimental setup could be clearer. Do all methods use the same pre-trained encoder and loss function? Is the only difference between methods the pooling operation? It is unclear if improved results on FAC are because of DENO or additional loss terms.
Questions for authors
- Could the authors clarify what they mean by the phrase correlated MIL. This phrase is inherited from Shao et al. (NeurIPS 2021) but is unclear without a generative model that describes correlation.
- When discussing the signed determinantal evidence pooling the author note that “The complete predictor is not claimed to be invariant to arbitrary changes in multiplicity” (L209-210). Can the author comment on the implications of this?
- Are differences of 0.0002 AUROC statistically significant?
- It would be nice to see how methods perform as a function of the percentage of nuisance multiplicity.
Response to Reviewer tTuD - New Results and Clarifications
We thank the reviewer for the constructive feedback. To address concerns, we conducted 70 additional end-to-end experiments on SICAP-MIL dataset. evaluating seven pooling methods (mean, max, attention, similarity-weighted, cluster, determinant-only, and DENO) across two prediction tasks and five seeds. We evaluated 349 whole-slide images using original train/validation/test partitions (187/64/98), with 2–56 patches per slide (median 27). All methods used the same ImageNet-initialized ResNet-18 architecture, jointly trained with pooling heads.
New experimental results. For ISUP Grade Group classification (0–5), DENO achieved 0.7517 QWK, 54.90% accuracy, 0.3854 macro-F1, and 0.8102 MAE, outperforming the strongest competing results of 0.6962, 50.00%, 0.3587, and 0.8837, respectively. For Gleason-pattern presence prediction (G3/G4/G5), DENO achieved 0.8433 ± 0.0231 macro-AUROC, compared with 0.8222 ± 0.0124 for the strongest baseline, similarity-weighted pooling.
Computational complexity and practical overhead. Exact DENO has (O(N^3)) determinant complexity, compared with linear-complexity normalized attention pooling. Our SICAP-MIL experiments quantify its practical cost using the jointly trained ResNet-18 model. DENO required 45.74 ms/slide for synchronized GPU forward inference versus 25.80 ms for attention. However, measured training wall time was 22.33 minutes versus 30.56 minutes, and effective test time was 199.68 versus 248.94 ms/slide, respectively. These wall times include concurrent-execution and early-stopping effects. Peak allocated inference memory was approximately 78 MiB at the evaluated bag sizes.
Comparison with interaction-aware and redundancy-aware methods. DENO outperformed Set Transformer on FAC (75% vs. 45% ensemble accuracy; 0.30 vs. 0.60 MAE), Stanford FoG (73.64% vs. 71.61% balanced accuracy; 0.78 vs. 0.76 AUROC), MNIST-Bags (98.14% vs. 97.68% accuracy), and MUSK2 (83.04% vs. 80.57% balanced accuracy). To address explicit redundancy handling, we additionally evaluated similarity-weighted and cluster-based pooling on SICAP-MIL. DENO achieved 0.7517 QWK vs. 0.6944 and 0.3792, respectively, and 0.8433 macro-AUROC vs. 0.8222 and 0.6982.
Encoder, loss, and pooling attribution. Within each benchmark, all methods use the same encoder architecture, preprocessing, splits, and seeds. For FAC, the pretrained BiGRU gait encoder uses auxiliary supervision for seven gait quantities, with earlier layers frozen and final layers fine-tuned. DENO uses a boundary-balanced ordinal loss with auxiliary gait supervision and regularization. the architecture is shared. Table 4 directly evaluates DENO's components under the FAC protocol. Full DENO improves ensemble accuracy by 30 percentage points over mean pooling and 10 points over determinant-only pooling. Additionally, SICAP-MIL uses the same ResNet-18 architecture across seven methods, with jointly trained encoders. DENO achieves 0.7517 QWK, compared with 0.6962 for mean and 0.6960 for determinant-only pooling.
Meaning of correlated MIL. We use correlated MIL to describe bags containing statistically dependent or representation-similar instances, such as neighboring histopathology patches or overlapping gait windows. DENO models these relationships using a Gaussian similarity kernel, without requiring one specific generative correlation model.
Why is DENO not invariant to arbitrary multiplicity changes? This is intentional. Uniformly replicating an entire bag preserves its relative composition and leaves the DENO score unchanged under the theorem's assumptions. Selective replication can change informative pattern prevalence. DENO therefore separates redundancy attenuation from prevalence preservation rather than eliminating all multiplicity information. On SICAP-MIL, full DENO exceeded determinant-only pooling by 0.0557 QWK and 0.0264 AUROC.
Is MNIST-Bags AUROC advantage significant? MNIST-Bags AUROC is not main improvement observed. it needs statistical test to be sure but as both numbers are very close to one, seems approach perform well enough for AUROC metric which might not be improved significantly. Some improvements observed on other datasets and metrics: +6.96 and +4.72 percentage points over mean pooling under exact and near duplication, respectively; 75% vs. 45% FAC ensemble accuracy; and 73.64% vs. 72.06% balanced accuracy on Stanford FoG. five-seed SICAP-MIL evaluation shows +0.0555 QWK, +4.90 percentage points in accuracy, and +0.0211 macro-AUROC over the strongest respective baseline point estimates.
Performance versus nuisance multiplicity. Our controlled experiments show that DENO maintains 98.61% accuracy under exact duplication and 97.43% under near duplication, compared with 91.65% and 92.71% for mean pooling. I agree that test across increasing, unseen nuisance-multiplicity ratios would provide a stronger quantitative characterization of distribution-shift robustness. We will add those results to the paper.