acceptodds
Under review as a conference paper at ICLR 2027

DENO: Signed Determinantal Evidence Pooling for Correlated Multiple-Instance Learning

Abstract

Multiple-instance learning (MIL) often involves strongly correlated or redundant instances, yet standard aggregation does not explicitly separate redundant evidence from prevalence information. We introduce DENO, a signed determinantal evidence pooling operator that combines signed instance evidence with a degree-normalized similarity kernel and contrasts log-determinants under opposite evidence tilts. A separate mean-evidence branch retains frequency information. We show that, for exact replicas, the full instance-level determinant reduces exactly to an effective latent-source determinant: uniform replication leaves the determinantal term unchanged, while selective replication can alter the effective source geometry. For near-redundant groups, we bound the approximation error using residual spectral structure and within-group evidence variation. We further develop a boundary-aware ordinal extension and evaluate DENO across five heterogeneous MIL benchmarks, where it attains the highest primary classification point estimate on three tasks and remains competitive on the other two. Controlled duplication experiments show targeted robustness to nuisance multiplicity, and on post-stroke FAC prediction the ordinal variant achieves 75% accuracy and 0.30 MAE versus 45% and 0.70 for mean pooling.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Multiple-Instance Learning, Set Aggregation, Determinantal Pooling, Correlated Instances, Redundancy-Aware Aggregation, Weak Supervision, Ordinal Regression

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tTuDOpenReview ↗
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 O(N3)\mathcal{O}(N^3) instead of O(N2⋅D)\mathcal{O}(N^2 \cdot D) 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.
1 comment
AuthorsOpenReview ↗

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.

  1. 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.

  2. 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.

  3. 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.

  4. 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.

  5. 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.

  6. 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.

  7. 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.

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

This article studies a common problem in multi instance learning: the same set of data may contain many highly similar or even repeated instances. If the model treats these repeated instances as new independent evidence, the final prediction may be influenced by repeated information; On the other hand, a certain pattern that appears frequently may also be a useful signal in itself. The author proposes DENO, hoping to handle both situations simultaneously. The method learns two types of information for each instance: one represents how similar it is to other instances, and the other represents its direction and degree of support for the prediction results. Subsequently, DENO used determinant based calculations to reduce the repetitive effects of similar instances and retained information on the frequency of pattern occurrence using a separate mean branch. The paper also provides theoretical analysis in the case of repeated instances, evaluates five multi instance learning tasks, and designs controllable MNIST repeated experiments. Overall, the connection between method design, theoretical analysis, and experimental objectives is relatively clear, but there are still some issues that need to be strengthened.

Critical strengths

The paper presents a theoretical analysis of repeated instances, evaluates five multi instance learning tasks, and designs controllable MNIST repeated experiments. Overall, the connection between method design, theoretical analysis, and experimental objectives is relatively clear.

Critical weaknesses
  1. It needs to be compared with more direct "de duplication" methods. Controlled repeated experiments mainly indicate that DENO is more resistant to interference from repeated instances than mean pooling and attention pooling. But this still cannot prove that the performance improvement necessarily comes from DENO's unique determinant calculation. A more direct question is: Can we also achieve similar results by reducing duplicate information based solely on instance similarity?

  2. Theoretical stability still needs to be tested under more realistic changing conditions. The paper proves that when the entire set of instances is proportionally repeated, the results of DENO can remain unchanged. However, mean pooling and common normalized attention often have similar properties, so this is not a unique advantage of DENO. More importantly, whether DENO remains stable when faced with instances being repeated more frequently or when repeated instances are not exactly the same. The current controlled MNIST experiment uses the same type of repetition settings during training and testing, which is not sufficient to demonstrate that the model can handle unseen repetition patterns.

  3. The computational cost on large-scale instance collections needs to be more thoroughly evaluated. DENO needs to calculate the similarity between instances and perform matrix determinant operations. As the number of instances in a single set increases, computation time and memory overhead may rapidly increase; The paper also used double precision calculations, further increasing the cost. The current experimental set is relatively limited in scale, so it is not yet clear whether the method is suitable for larger pathological images or other large-scale MIL scenarios.

The conclusion of the paper should more accurately reflect the performance on different datasets. DENO achieved the best main indicator results in three of the five main tasks, with overall competitive performance, but did not lead in MUSK2 and Mutagenesis, and the improvement on regular MNIST Bags was relatively limited. Therefore, a more appropriate conclusion is that DENO is particularly suitable for tasks where instances are highly correlated or prone to duplicate information, rather than having been proven to replace existing aggregation methods in all MIL tasks

Questions for authors
  1. Can reduce the weight of duplicate instances based on similarity, group similar instances first and then aggregate them, or compress similar features first and then make predictions
  2. Can fix the trained model and only change the number of repetitions, noise level, or repetition ratio during testing to test whether the approximation error discussed in theory can explain the changes in real predictions
  3. It is recommended to report the training time, inference time, and video memory usage under different instance numbers, and discuss whether performance and stability can still be maintained after using low rank approximation or other acceleration methods
3 comments

My native language is Chinese. After reading this manuscript, I provided my opinion in Chinese and had it translated into English by AI. The editor seems to believe that this is an AI generated opinion, and they are requesting me to make further comments. But I think their viewpoint is obviously inappropriate.

No one is requesting you to make additional points. However, if you used AI for translation purposes, you are supposed to declare it in the review form.

AuthorsOpenReview ↗

Response to Reviewer hXxX - New Results and Clarifications

We thank the reviewer for the constructive feedback. To further strengthen our evaluation, we conducted 70 additional end-to-end experiments on SICAP-MIL prostate histopathology, comprising 349 whole-slide images. We compared seven pooling methods across two prediction tasks (ISUP Grade Group classification and Gleason-pattern presence prediction), using five seeds and jointly trained ImageNet-initialized ResNet-18 encoders.

  1. Why determinantal pooling instead of simpler de-duplication? Our objective is not merely to remove duplicate instances. DENO is designed for tasks where both rare discriminative patterns and the prevalence of recurring patterns matter simultaneously. Simple de-duplication can suppress redundant evidence but may discard informative frequency information. DENO instead combines a signed determinantal branch that attenuates redundant evidence while retaining distinct evidence directions with a mean branch that preserves relative pattern prevalence. In our additional SICAP-MIL experiments, DENO achieved 0.7517 QWK, compared with 0.6944 for similarity-weighted pooling, 0.3792 for cluster pooling, and 0.6960 for determinant-only pooling. For Gleason-pattern prediction, DENO achieved 0.8433 macro-AUROC, versus 0.8222, 0.6982, and 0.8169, respectively. These results support the complementary two-branch design and demonstrate improvements over the evaluated similarity-based alternatives.

  2. Stability under changing duplication patterns. We agree that uniform whole-bag replication invariance is not unique to DENO; normalized mean and attention pooling can share this property. Our theoretical contributions additionally include an exact replicated-source reduction and quantitative near-redundancy stability bounds. Numerical verification found zero bound violations across 2,500 near-redundancy cases. Binary and ordinal stability certificates covered 56.2% and 74.4% of their respective cases, with no prediction changes among certified cases. Our controlled MNIST experiments also use independently varying source multiplicities, rather than uniform whole-bag replication. DENO achieved 98.61 ± 0.44% accuracy under exact duplication, compared with 91.65 ± 0.84% for mean pooling and 91.57 ± 0.72% for attention. Under near duplication, DENO achieved 97.43 ± 0.59%, versus 92.71 ± 0.57% and 92.80 ± 0.74%, respectively. These experiments demonstrate robustness under the evaluated conditions. We agree that fixing a trained model and varying repetition ratios or noise levels at test time would provide stronger evidence of generalization. I will add those results to the paper.

  3. Computational complexity, inference time, and memory. Dense DENO requires O(N3)O(N^3) determinant computation and O(N2)O(N^2) kernel storage, with additional costs for ordinal boundaries and double-precision calculations. Our SICAP-MIL experiments evaluate complete ResNet-18 encoder-plus-pooling models. DENO required 45.74 ms/slide for synchronized GPU forward inference, compared with 25.80 ms for attention (approximately 1.77× slower). Measured training time was 22.33 vs. 30.56 minutes, and effective test time was 199.68 vs. 248.94 ms/slide, respectively. These wall times include early-stopping and concurrent-execution effects. Peak allocated inference GPU memory was approximately 78 MiB for both methods. These measurements establish feasibility at the evaluated scale. Computational scaling, memory requirements, and stability under low-rank approximations for substantially larger bags was not evaluated. we will add them to the paper.

  4. Scope of our conclusions. We agree that DENO is not universally superior. Our original evaluation shows the highest primary classification point estimates on three of five benchmarks, including FAC and Stanford FoG, while remaining competitive on MUSK2 and Mutagenesis. Improvements on standard MNIST-Bags are relatively small because all performance values are high. We therefore emphasize DENO's suitability for correlated MIL tasks where rare discriminative evidence and informative pattern prevalence must both be considered simultaneously.

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

The paper introduces DENO, a multiple-instance learning (MIL) pooling method designed for bags containing correlated or redundant instances. It combines a signed determinantal evidence term, which discounts repeated evidence along similar directions, with a mean-evidence term that preserves prevalence information.

The authors provide theoretical results showing that exact replicas can be reduced to their underlying sources, uniform replication of an entire bag leaves the score unchanged, and near-redundant instances admit bounded approximation errors and prediction-stability conditions. They validate DENO on five MIL benchmarks, controlled MNIST duplication experiments, numerical tests of the theoretical results, and ablations. DENO performs best on three of the five benchmarks and is notably robust to nuisance duplication in controlled MNIST. The real-world validation is mixed, however, and the post-stroke gait result is based on only ten participants.

Critical strengths

The core idea is natural and interesting: similar or repeated instances may carry redundant evidence, but their frequency can still be informative. DENO addresses both aspects by separating redundancy-aware aggregation from prevalence-sensitive aggregation. This is a clear and potentially useful design choice for multiple-instance learning with correlated data.

The paper also gives a theoretical account of when duplication leaves the score unchanged and how near-duplicates affect it. This helps explain the method’s intended behavior and supports the controlled experiments.

Critical weaknesses

The experimental validation is not convincing enough to establish that DENO is broadly useful. The controlled MNIST duplication results are promising, but they come from a deliberately constructed setting. Results on the five benchmarks are mixed, and the strongest real-world result, on post-stroke gait prediction, is based on only 10 participants. I would like to see broader and more substantial evaluation before accepting the paper’s claims about practical value.

Questions for authors

The idea is interesting, but I am not yet convinced that the paper demonstrates a sufficiently important real-world problem or that DENO is broadly useful. The controlled MNIST experiment shows robustness to duplication, but MNIST is a relatively simple setting. Could the authors provide evidence that nuisance multiplicity is common in real datasets, and evaluate DENO on more challenging tasks—for example, SVHN or histopathology, where bags contain many correlated image patches? How much does DENO improve over strong baselines on such tasks, and under what conditions does the improvement persist? These results could materially change my assessment of the method’s practical relevance.

2 comments
AuthorsOpenReview ↗

Response to Reviewer FRGU — Additional Experiments

We thank the reviewer for emphasizing the need for stronger real-world validation. To address this concern, we conducted 70 additional experiments on SICAP-MIL prostate histopathology, evaluating seven pooling methods across two prediction tasks and five seeds using jointly trained ImageNet-initialized ResNet-18 encoders.

Real-world histopathology results. We evaluated 349 whole-slide images (187 training, 64 validation, 98 testing). For ISUP Grade Group classification, DENO achieved 0.7517 QWK, 54.90% accuracy, 0.3854 macro-F1, and 0.8102 MAE, outperforming the strongest baseline point estimates 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 similarity-weighted pooling.

Evidence across other real-world tasks. Our original evaluation covered five heterogeneous MIL benchmarks. DENO achieved the highest primary classification point estimate on three, including 73.64% balanced accuracy and 0.78 AUROC on Stanford FoG, compared with 72.06% and 0.76 for gated attention, and 75% ensemble accuracy and 0.30 MAE on post-stroke FAC, compared with 45% and 0.70 for mean pooling. It remained competitive on MUSK2 and Mutagenesis.

When is DENO particularly beneficial? DENO is designed for tasks where both rare discriminative patterns and the prevalence of recurring patterns matter simultaneously. Its signed determinantal branch discounts redundant evidence while retaining distinct evidence directions, and its mean branch preserves relative pattern prevalence. This distinction is particularly relevant to prostate histopathology, where identifying important Gleason patterns and accounting for their tissue composition are both essential. Full DENO outperformed determinant-only pooling on both SICAP-MIL tasks (0.7517 vs. 0.6960 QWK; 0.8433 vs. 0.8169 AUROC), supporting the value of combining these complementary components.

Together with our controlled duplication results (98.61% vs. 91.65% accuracy under exact duplication; 97.43% vs. 92.71% under near duplication), these findings substantially strengthen the evidence for DENO beyond synthetic redundancy.

AuthorsOpenReview ↗

Response to Reviewer FRGU — Complete SICAP-MIL Experimental Results and Computational Efficiency

SICAP-MIL prostate histopathology experiments for additional test, discussed in our previous response. The following tables report the performance of all seven pooling methods on both histopathology prediction tasks (mean ± standard deviation).

Table 1. ISUP Grade Group classification (0–5).

MethodQWK ↑Accuracy (%) ↑Macro-F1 ↑MAE ↓
Mean0.6962 ± 0.081648.37 ± 6.190.3295 ± 0.06990.8918 ± 0.1579
Max0.6767 ± 0.088047.55 ± 6.150.3179 ± 0.02070.9673 ± 0.1463
Attention0.6702 ± 0.076350.00 ± 4.330.3459 ± 0.03370.9510 ± 0.1167
Similarity-weighted0.6944 ± 0.095348.37 ± 5.620.3364 ± 0.06180.9082 ± 0.1609
Cluster0.3792 ± 0.091433.06 ± 4.180.2082 ± 0.03561.4122 ± 0.2109
Determinant-only0.6960 ± 0.061049.18 ± 5.170.3587 ± 0.06230.8837 ± 0.1255
DENO (ours)0.7517 ± 0.072254.90 ± 5.120.3854 ± 0.04360.8102 ± 0.1366

Table 2. Gleason-pattern presence prediction (G3/G4/G5).

MethodMacro-AUROC ↑
Mean0.7971 ± 0.0100
Max0.7993 ± 0.0194
Attention0.8054 ± 0.0177
Similarity-weighted0.8222 ± 0.0124
Cluster0.6982 ± 0.0688
Determinant-only0.8169 ± 0.0224
DENO (ours)0.8433 ± 0.0231

Table 3. Computational efficiency on SICAP-MIL Grade Group classification.

MethodTrain (s) ↓Test (s) ↓Test (ms/slide) ↓GPU forward (ms/slide) ↓Peak GPU (MiB) ↓
Mean1628.2623.632241.1425.11378.186
Max1740.0125.631261.5425.82578.072
Attention1833.4424.396248.9425.80478.106
Similarity-weighted1767.3924.160246.5425.79878.089
Cluster1277.7323.646241.2933.31178.089
Determinant-only1558.5021.162215.9446.71778.090
DENO (ours)1340.0719.569199.6845.73978.090

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.