acceptodds
Under review as a conference paper at ICLR 2027

Learning Individual Representations from Aggregate Supervision

Abstract

Aggregate-supervised learning observes a label for a bag of instances while the instance labels remain hidden. When that label is a sum of finite-support contributions, an instance predictor can be trained by maximizing the probability of the observed total. We present FS-Conv, a likelihood layer that separates the instance predictor, the map from predictions to contribution probabilities, and the convolution backend. Under conditional factorization, discrete convolution computes the exact model-implied aggregate probability mass function. With Bernoulli contributions, the objective reduces to established count likelihoods; categorical contributions support integer-valued sums, and known signs encode cancellation. We compare this objective with mean matching and a Gaussian approximate likelihood. In three-seed digit-sum experiments on MNIST and on SVHN with an ImageNet-pretrained ResNet-18, FS-Conv has lower mean aggregate error and discrete aggregate negative log-likelihood, and higher hidden-digit accuracy, than both alternatives. Random-sign MNIST also favors FS-Conv on the reported means. On feature-defined Criteo click-count bags, FS-Conv is competitive with the evaluated label-proportion baselines. These results favor exact discrete likelihood training on the digit-sum and known-sign tasks; instance recovery still depends on the representation and the adequacy of the factorized model.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
aggregate supervision, learning from label proportions, weakly supervised learning, representation learning, multiple-instance learning, dependence modeling

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_dGCYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper learns an instance classifier when each bag only has the sum of hidden integer labels, like counts, signed counts or digit sums. FS-Conv turns each instance prediction into a distribution over its contribution. It convolves these into a distribution over the bag total, and trains on the observed total’s likelihood. The convolution is not new, and the authors say so. The pitch is a shared interface. The binary case reduces to known count losses. A toy example shows two models with the same mean and variance that give the observed total different probabilities. Moment-based losses miss this. Over three seeds, FS-Conv is best on MNIST and SVHN digit sums and on signed MNIST counts.

Critical strengths
  1. The paper gives a clean, modular formulation that separates the instance predictor, the contribution-PMF map Φ\Phi, and the convolution backend.
  2. The analysis explains why the discrete likelihood should behave differently from moment objectives.
Critical weaknesses
  1. Methodological novelty is limited. The likelihood is the classical convolution of per-instance PMFs. Its Bernoulli case coincides with the CountLoss and LLP-PVC count likelihoods, the known-sign case reduces to a shifted count, and DeepProbLog already showed that digits can be learned from sums, so the contribution rests mainly on the unifying interface and the empirical comparison.
  2. The vision experiments compare only against two moment-based objectives, with no comparison to DeepProbLog-style exact inference or to the full CountLoss or LLP-PVC pipelines.
  3. Neither the central factorization assumption nor the broader scope is tested. Robustness to within-bag dependence is not measured, and the only real-world task shows no advantage.
Questions for authors
  1. How stable are the results with independently resampled image splits and bag sets, ideally with more seeds and paired tests?
  2. For digit sums, is this the same likelihood DeepProbLog computes exactly? If so, what does FS-Conv add beyond organizing the computation? A direct comparison would help.
2 comments

Thank you for the detailed review.

Q1. Stability with resampled splits and bags, more seeds, paired tests. Done: 20 seeds on MNIST and 10 on SVHN, each with its own image split and bags.

Hidden-label accuracySeedsMean-MSEGaussian-AMLEFS-ConvFS-Conv − Gaussian-AMLE [95% CI]paired t / Wilcoxon pSeeds won
MNIST digit sums200.208 ± 0.0210.736 ± 0.0800.889 ± 0.087+0.153 [+0.103, +0.204]4e-6 / 1e-519/20
SVHN digit sums100.777 ± 0.1150.762 ± 0.1010.941 ± 0.007+0.179 [+0.108, +0.250]3e-4 / 0.00210/10
Random-sign MNIST200.940 ± 0.0460.982 ± 0.0040.991 ± 0.001+0.010 [+0.008, +0.012]2e-9 / 2e-620/20

FS-Conv also has lower expected-sum MAE and discrete NLL than both baselines in every seed of all three tasks (Table 12). On MNIST, FS-Conv itself is bimodal across seeds (at least 0.95 in ten, 0.70–0.86 in the other ten). These runs use each method's own checkpoint rule and 50 epochs. With one rule for all methods (a second script, 20 seeds) the gap is the same: 0.861 vs 0.706, +0.155 [+0.102, +0.208], 18 of 20 seeds. With 16 times longer training both likelihoods improve and the gap narrows but persists: 0.932 vs 0.849, +0.084, paired t p=0.02, 15 of 20 seeds (Tables 17–18).

Q2 / W2. Is this DeepProbLog's likelihood? Comparison with CountLoss, LLP-PVC. Yes, it is the same likelihood. Against DeepProbLog 2.1 with its exact engine, probabilities and gradients agree to within 2e-16 for sums of two to four digits. On DeepProbLog's single-digit addition task (configuration of its released example, same initial weights and data order) the two reach 0.9738 ± 0.0035 and 0.9758 ± 0.0027 test digit accuracy, in 681 s and 160 s on one CPU thread (Tables 25–26). With the two encodings we tried, DeepProbLog needs minutes to compile four digits; for five, the faster one exceeded a 64 GB memory limit and the slower was not run. Our bags have ten; our layer evaluates the same likelihood in polynomial time on a GPU. That computation is not new either: Section 2 now cites De Smet & Zuidberg Dos Martires (NeurIPS 2024), whom we had missed.

For binary labels our likelihood is the count likelihood of CountLoss and LLP-PVC (Eq. 11); the full LLP-PVC method needs class histograms, which a digit sum does not provide. On the binary tasks we did not run their full pipelines; Section 3.4 limits the claim to the likelihood components.

W1. Limited novelty. We agree that neither the likelihood nor its computation is new, and Section 2 now says so. The main contribution is empirical: with the instance model and the data protocol held fixed, the paper indicates when the exact likelihood improves instance recovery and when it does not. On digit sums the mean gap to Gaussian-AMLE grows with the number of digit values (none for K ≤ 3; 0.853 vs 0.680 at K=10, confirmed with 20 seeds) and is largest for small bags (Section 4.4). FS-Conv has no advantage for large bags with few training bags, on binary Criteo bags, or under the dependence below.

W3. Factorization untested; no advantage on the real task.

Dependence. New stress test (Appendix C.4): a synthetic count task in which a shared bag-level random effect of scale τ is added to every instance logit; 10 seeds.

τ (within-bag label correlation)0 (0.000)0.5 (0.040)1 (0.132)2 (0.334)
Hidden accuracy, Mean-MSE0.6860.6730.6410.585
Hidden accuracy, Gaussian-AMLE0.6860.6530.5960.569
Hidden accuracy, FS-Conv0.6860.6510.5940.569
Best single-instance classifier (from the generator)0.6910.6840.6660.628

A larger τ also adds label noise that the features cannot explain (last row; majority-class rate 0.60, 0.60, 0.59, 0.57). Relative to that ceiling, Mean-MSE loses 1–4 points and the two likelihoods 3–7 for τ > 0: the mean of the total is a sum of marginal means under any dependence, while both likelihoods assume the factorized variance. The result does not favor our method; the abstract and Section 5 now say that the exact likelihood is not robust to this misspecification. The coverage of the model's 90% intervals for held-out totals falls from 0.93 to 0.31, which flags a miscalibrated aggregate model without instance labels but is not a test for dependence.

Real-world tasks. We added two (Section 4.5). On review star totals (K=5) FS-Conv is ahead of Gaussian-AMLE by 1.4, 1.2 and 0.6 points of rating accuracy at n=4, 8 and 16 (all five seeds, paired t p < 0.003), with indistinguishable accuracy at n=32 and 64. On PanNuke nuclei counts (K=16) it lowers the image-total MAE from 3.68 to 3.09 (all five seeds). On Criteo (binary labels, bags of 128) FS-Conv is slightly below proportion matching (ROC-AUC 0.749 vs 0.752), and we claim no advantage there.

Thank you to the authors for their responses and additional experiments. While I appreciate their efforts to address my concerns, I believe that the novelty of the proposed method remains questionable. In my view, the contributions and experimental validation are still insufficient to meet the standards of a top-tier conference.

Reviewer_KCW5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper proposes a method for learning individual instance representations within a Multiple Instance Learning (MIL), where only aggregated labels are provided. While existing method suffer from information loss due to the use of approximations, the paper proposes FS-Conv, which defines a probability mass function for each individual instance and aggregates these using an discrete convolution. This allows for the exact likelihood calculation for an bag of instances, and the model trained by minimizing the negative log-likelihood (NLL). In the experiment, FS-Conv achieves better performance on various discrete aggregate applications compared to existing approximation base approaches.

Critical strengths
  • Unlike existing MIL approaches that rely on approximate objectives (e.g., mean matching, Gaussian likelihood approximations), the paper proposes a framework for directly optimizing the exact Likelihood of the aggregated labels.
  • The proposed method is not limited to a specific type of aggregate supervision. The method provides a general framework applicable to a variety of discrete aggregate functions like binary count, known-sign count, integer sum and etc.
  • Instead of relying solely on matching aggregate statistics, the proposed approach explicitly considers the probabilistic contribution of individual instances. This could provide more informative supervision for recovering latent semantic information at the instance level.
Critical weaknesses
  • W1. A major concern is the potential high computational cost of computing the exact likelihood through convolution. Especially, for large bags or instance label, the computational and memory requirements may become substantial. This could limit the applicability of the proposed method to large-scale datasets. It would be helpful to report the actual computational costs, memory usage, and training time compared with approximation-based methods.

  • W2. In the Digit Sums and Known-Sign Counts experiments, the method compares only with baseline approximation methods rather than existing approaches from prior studies. This makes difficult to assess the relative effectiveness of the proposed method.

  • W3. The paper often introduces concepts or technical details without giving enough background, which makes some sections hard to understand when first read. Some sentences feel abrupt and require referencing later sections to fully understand earlier paragraphs. Providing clearer explanations and smoother transitions would improve readability.

    • e.g. L107, our fix-size ... / L111, FS-Conv specializes ... / L112 Our comparison... / L118, Our interface ...
Questions for authors
  • Although the proposed method appears applicable to arbitrary bag sizes, the experiments primarily use fixed sizes. How does performance vary with bag size? And can the method handle datasets containing bags of different sizes (regarding training efficiency)?
2 comments

Thank you for the review. We address the three weaknesses and the question in turn.

W1. Computational cost, memory and training time. We measured all three (Appendix C.5). Time per epoch relative to Mean-MSE:

Task (instance model)nKFS-Conv, sequentialFS-Conv, FFTPeak GPU memory (same for all objectives)
MNIST sums (LeNet)≈10101.68×–292 MB
MNIST sums, fixed n (LeNet)50103.79×1.11×1533 MB
SVHN sums (ResNet-18)≈10101.18×–4121 MB
Review star sums (embedding bag)64514.01×1.89×2622 MB
Nuclei counts (CNN)16162.06×1.04×6386 MB

The same peak memory (PyTorch's peak allocated GPU memory, within 1 MB) was measured for the three objectives in every task. Gaussian-AMLE takes 1.0–1.3× the time of Mean-MSE. Most of FS-Conv's overhead comes from the sequential backend of the submission. An FFT product-tree backend evaluates the same likelihood (float64, trained up to n=64) with comparable accuracy (1.5 points lower in one setting; Appendix C.5). In a float32 benchmark of the loss layer alone, it stays under 5 ms per batch and 0.25 GB for bags of up to 1,024 instances (Table 24).

W2. Comparison with approaches from prior studies. The best-known prior method for digit sums is DeepProbLog, and we now compare with it directly (Appendix C.6). It computes the same likelihood: its exact engine returns the same probabilities and gradients as our layer to within 2e-16, and on its own single-digit addition task the two reach the same digit accuracy within seed variation (0.974 and 0.976), with our layer four times faster. With the two encodings we tried, DeepProbLog needs minutes to compile a sum of four digits, and the faster one exceeded a 64 GB memory limit on five (the slower was not run), so we could not run it on our bags of ten.

The two baselines are prior approximate objectives that apply to a sum label: mean matching from deep LLP, and the approximate maximum likelihood of Baručić & Kybic (2022) extended to integer contributions. Most other LLP methods need the class histogram of each bag, which a digit sum does not determine; where the label is a binary count they apply, and our Criteo experiment compares with proportion matching and a GenBags-style objective. For known-sign counts the likelihood reduces to the count likelihood of CountLoss (Section 3.4), so the FS-Conv row there is that prior likelihood. We also added an instance-supervised oracle and two real-data tasks (Section 4.5).

W3. Concepts introduced without background; abrupt sentences. We agree. Section 2 now opens by defining the three terms it uses (contribution, contribution PMF, backend). L107 now states that our Criteo experiment cuts feature-defined groups into bags of exactly 128 impressions, so its numbers are not comparable with published benchmarks. L111–L112 are replaced by plain statements of what is shared with DeepProbLog and what differs, and L118 by the statement that the backend can be exchanged without changing the loss. If other passages remain hard to follow, please point us to them and we will revise them.

Q. Performance by bag size, and bags of different sizes. The method handles bags of different sizes, and the original experiments already used them (lengths drawn from a rounded normal with mean 10 and standard deviation 2). "Fixed" referred to bag sets that are sampled once and reused in every epoch; our wording was ambiguous, and Section 4.1 now says this. Shorter bags are padded with a point mass at zero, which leaves the likelihood unchanged, so the cost of a batch is set by its longest bag.

Hidden-digit accuracy (K=10, 1,000 bags)n≈5n≈10n≈20n≈50
Mean-MSE0.1960.2000.2340.213
Gaussian-AMLE0.8370.7420.6630.564
FS-Conv0.9720.9190.6510.564
FS-Conv time per epoch vs Mean-MSE (sequential backend)1.4×1.7×2.0×2.2×

FS-Conv's advantage is largest for small bags. On MNIST the gap returns for larger bags with more training: with 16,000 bags FS-Conv reaches 0.99 at n=20 and n=50, where Gaussian-AMLE reaches 0.91–0.93; training 16 times longer on 1,000 bags does not produce this gap (Tables 15–16). On real review data the accuracy gain is present for bags of up to 16 reviews and absent at 32 and 64 (Table 19).

Thank you for the detailed responses and the additional experiments. However, given that FS-Conv computes the same likelihood as DeepProbLog, I agree with reviewer dGCY’s assessment that the methodological novelty is limited. While the improvements in computational efficiency and the empirical analysis are valuable, this raises concerns about the novelty of the proposed method. Therefore, I will maintain my original rating.

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

This paper addresses the problem of training an instance prediction model using bag-level sum labels instead of individual instance labels. Learning based on the mean or a Gaussian approximation may overlook differences in the discrete probability assigned to the observed sum. The authors aim to account for these differences through a common likelihood-based learning framework for different types of supervision expressed as sums of discrete variables with finite support. To this end, they propose FS-Conv, which combines instance-level contribution distributions through convolution under a conditional independence assumption, and explain its relationship to existing count likelihoods.

Critical strengths
  1. To address instance-level learning from sum labels, the paper separates instance prediction, contribution distribution construction, and sum distribution computation into distinct components, explaining their roles and how they connect. It also formulates counts, integer sums, and signed counts within the same framework and uses these formulations in training, showing that different types of sum supervision can be handled in a common way.
  2. The mathematical derivations are clear, so I could follow how the steps connect and how sum labels provide learning signals for individual instances. The distinct roles of the components also helped me understand which parts would need to change when extending the framework to other tasks or computational methods. I think this structure can serve as a useful starting point for future research on various types of sum supervision.
Critical weaknesses
  1. The paper shows that different types of sum supervision can be formulated in a common way, but it is difficult to adequately assess the value of this generalization for real-world problems from the current experimental settings alone. The binary-count likelihood corresponds to existing count likelihoods, and the evaluated signed-count task can also be reduced to an ordinary count problem. Therefore, evaluation beyond these reductions is limited to digit sums on MNIST and SVHN. Since the experiment on Criteo, a real-world dataset, also involves counting, the current experiments provide limited evidence for assessing the practical scope of the common framework across different types of sum supervision.
  2. Directly computing the sum distribution incurs computational costs, but the relationship between predictive performance and computational cost across problem scales has not been sufficiently demonstrated. As explained in the paper, the costs of computing and storing the distribution increase with the number of instances per bag and the range of contribution values, but comparisons that vary these conditions within the same task were not provided. The number of training bags also differs across experiments, but it remains unclear how instance prediction performance varies with the number of bags within the same task. Without comparisons with Mean-MSE and Gaussian-AMLE in terms of predictive performance, runtime, and memory usage under these varying conditions, it is difficult to judge when the benefits of using the exact sum distribution justify its computational cost.
Questions for authors
  1. Could the authors compare how FS-Conv’s instance prediction performance, runtime, and memory usage change relative to Mean-MSE and Gaussian-AMLE as the number of training bags, the number of instances per bag, and the range of contribution values vary? In particular, I am interested in whether the performance advantages persist at larger scales and whether there are conditions under which computational costs make the method difficult to apply.
  2. The proposed framework seems applicable to various sum-supervision scenarios. Are there specific real-world tasks that the authors have in mind? I am particularly interested in tasks that do not reduce to ordinary counting and for which individual instance labels are difficult to obtain but sum labels can be obtained. It would also be helpful to explain what information or assumptions would be needed to apply the framework to these tasks.
1 comment

Thank you for the constructive review. Both weaknesses asked for experiments, and we ran them.

W1. The value of the generalization beyond counting. We added two real-data tasks in which each instance contributes one of several values (Section 4.5).

Review star totals (Amazon Reviews 2023; a bag is n reviews of one product and the label is the sum of their 1–5 star ratings; five seeds):

Hidden rating accuracyn=4n=8n=16n=32n=64
Mean-MSE0.7370.7240.7120.7040.697
Gaussian-AMLE0.7430.7350.7260.7030.695
FS-Conv0.7570.7470.7320.7060.696
Oracle (instance labels)0.7660.7670.7680.7660.769
Seeds in which FS-Conv is ahead of Gaussian-AMLE5/55/55/54/54/5

Nuclei counts (PanNuke; a bag is an image of 16 tiles, the label is its total nuclei count, each tile's count is modeled on 0–15): image-total MAE is 3.09 for FS-Conv against 3.68 (Gaussian-AMLE) and 3.64 (Mean-MSE); tile accuracy is 0.575 against 0.540 and 0.546 (oracle 0.659). FS-Conv is ahead in all five seeds (Appendix C.3; Table 21).

What does not work: at n=32 and 64 rating accuracy is indistinguishable across objectives (other metrics differ; Table 19), and an oracle gap remains in both tasks.

W2 / Q1. Performance and cost as the number of bags, the bag size and the contribution range vary. All three are varied on MNIST digit sums, five seeds each (Section 4.4, Appendix C.2).

Hidden-digit accuracy (n=10, 1,000 bags)K=2K=3K=5K=7K=10
Mean-MSE0.9990.9950.4730.4860.201
Gaussian-AMLE0.9990.9950.9920.9570.680
FS-Conv0.9990.9970.9950.9890.853
Oracle (instance labels)0.9990.9960.9950.9890.977
Hidden-digit accuracy (K=10, 1,000 bags)n≈5n≈10n≈20n≈50
Mean-MSE0.1960.2000.2340.213
Gaussian-AMLE0.8370.7420.6630.564
FS-Conv0.9720.9190.6510.564
FS-Conv time per epoch vs Mean-MSE (sequential backend)1.4×1.7×2.0×2.2×

(Separate runs; the shared setting n=10, K=10 differs within FS-Conv's seed spread.)

Digit values. Restricting the digits to 0..K−1 changes the number of values and the classes together, so this is not a pure range effect. There is no gap for K ≤ 3 and a mean gap of 17 points at K=10 (five seeds, paired p=0.10), which the same script confirms with 20 seeds (0.861 vs 0.706, 18 of 20 seeds; Table 17); with 16 times longer training this gap narrows to 8 points (Table 18).

Bag size and training bags. With 1,000 bags the two likelihoods reach similar accuracy from n=20 on, also when trained 16 times longer (0.72–0.79). With 16,000 bags FS-Conv reaches 0.99 at n=20 and n=50, where Gaussian-AMLE reaches 0.91–0.93; training longer on 1,000 bags does not produce this gap (Tables 15–16). With 250 or 500 bags no objective exceeds 0.58 (Table 13).

Cost. Time per epoch relative to Mean-MSE is 1.2× (ResNet-18) to 14× (a very small text model) with the sequential backend, and 1.0–1.9× with an FFT backend (float64, trained up to n=64) that gives comparable accuracy. The same peak memory, within 1 MB, was measured for all objectives (Table 23).

When is it hard to apply? With the sequential backend, bags of 1,024 instances cost 1.5–2.4 s per batch and 1.2–2.0 GB for K=10–16; the FFT backend stays under 5 ms and 0.25 GB there (float32 benchmark of the layer alone, Table 24). Values that are far apart on the integer grid, for example {0, 1, 1000}, make the support wide. The more important limit is statistical: with large bags and few training bags the exact likelihood gives no accuracy gain over the Gaussian approximation.

Q2. Real tasks that do not reduce to counting, and what is needed. Ratings published as an average with a count (our review task emulates this); region-level cell counts, where a tile contributes a multi-valued count (our PanNuke task emulates this); per-group sums of small integer conversion values in privacy-preserving advertising reports; survey or census totals of ordinal answers. The last two usually involve large groups, where by our sweeps the exact likelihood needs many training groups to pay off; we have not tested it there.

One needs (i) a known finite set of integer values per instance, (ii) the bag membership and the bag total (reported noise enters as one more convolution term), and (iii) conditional independence of the contributions given the inputs. When (iii) fails through a shared bag effect, our stress test shows that the likelihoods lose more accuracy than Mean-MSE; poor coverage of the model's intervals for held-out totals flags such a misfit without instance labels, although it does not identify the cause (Appendix C.4).

Comments on the paper

2 comments
AuthorsOpenReview ↗

General comment

We thank the reviewers for their careful reading. We ran new experiments for each empirical point and revised the paper (new text is blue). Details are in the reply under each review; this is the summary.

1. Replication (dGCY Q1). 20 seeds on MNIST and 10 on SVHN, each with its own image split and bags. FS-Conv is ahead of Gaussian-AMLE in hidden-digit accuracy in 19 of 20 and 10 of 10 seeds (0.889 vs 0.736; 0.941 vs 0.762). With one checkpoint rule for all methods (a second script, 20 seeds) the MNIST gap is the same (0.861 vs 0.706, 18 of 20 seeds); with 16 times longer training it narrows but persists (0.932 vs 0.849, 15 of 20 seeds).

2. Where the exact likelihood matters (KeES Q1, KCW5 Q). New sweeps over bag size, number of digit values K and training bags. On digit sums the mean gain grows with K (none for K ≤ 3; 0.853 vs 0.680 at K=10) and is largest for small bags; larger bags need more training bags before it appears; training 16 times longer on 1,000 bags does not produce it.

3. Real data (KeES W1, dGCY W3). Review star totals (K=5) and nuclei counts per image (K=16). FS-Conv is the most accurate aggregate-supervised objective for bags of up to 16 reviews (+0.6 to +1.4 points, all five seeds) and on nuclei counts (image-total MAE 3.09 vs 3.64–3.68). With 32 or 64 reviews per bag, rating accuracy is indistinguishable across objectives.

4. Cost (KeES W2, KCW5 W1). With the sequential backend of the submission FS-Conv takes 1.2× to 14× the time of Mean-MSE per epoch, depending on the instance model. An FFT backend evaluates the same likelihood at 1.0–1.9× with comparable accuracy (trained up to n=64). Measured peak memory is the same for all objectives (within 1 MB).

5. Dependence (dGCY W3). In a new stress test with a shared bag-level effect, both likelihoods lose 2–5 points more instance accuracy than Mean-MSE. The exact likelihood is not robust to this misspecification, and the paper now says so.

6. DeepProbLog (dGCY W2, Q2). Same likelihood: probabilities and gradients agree to within 2e-16, and both reach 97% digit accuracy on DeepProbLog's single-digit addition task. With the two encodings we tried, its exact engine needs minutes for a sum of four digits, and the faster one exceeded a 64 GB memory limit on five; our bags have ten.

7. Writing (KCW5 W3). Section 2 now defines its terms before using them; the four flagged sentences are rewritten.

On novelty: the revised Section 2 states that neither the likelihood nor its computation is new. The main contribution is a controlled comparison that indicates when the exact likelihood improves instance recovery and when it does not.

We would be glad to run further checks during the remaining discussion period.

Area_Chair_pyyeOpenReview ↗

Reminder – Please engage with author rebuttal

Hi, Reviewers,

We are now approaching the end of the discussion phase. The authors have posted their rebuttal, and I encourage all reviewers to engage actively in the discussion.

Please take the time to: i) read the authors' responses carefully; ii) check whether your concerns and questions have been adequately addressed; and iii) provide the authors with further feedback where needed.

Your active participation at this stage is important for reaching a well-informed final decision.

Thank you for your time and effort.

Best,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.