acceptodds
Under review as a conference paper at ICLR 2027

FedC²: From Client Compatibility to Class Complementarity for Federated Learning from Label Proportions

Abstract

Federated learning from label proportions (Fed-LLP) aims to learn instance-level predictors from distributed clients whose local bags are supervised only by class proportions. Existing LLP methods mainly focus on extracting supervision from aggregate proportions, while federated learning methods typically address client heterogeneity under standard instance-level supervision. However, in Fed-LLP, aggregate supervision is further complicated by substantial client heterogeneity, while the absence of instance-level labels makes such heterogeneity harder to characterize and exploit. Towards this end, we propose FedC², a hierarchical collaboration framework that organizes cross-client knowledge sharing from client compatibility to class-wise complementarity. FedC² first constructs a target-specific collaboration base from the geometry of clients’ observed bag-proportion sets, and then conservatively refines class-wise collaborator contributions according to target-relative coverage and local proportion consistency. We further provide a theoretical analysis characterizing when the resulting hierarchical reweighting tightens class-specific aggregation-error bounds. Extensive experiments across four benchmarks and three levels of client heterogeneity show that FedC² achieves the best accuracy in 10 of 12 settings against representative LLP-based, generic FL, and personalized FL methods. Further analyses support the effectiveness, generality, and transfer relevance of the proposed hierarchical collaboration strategy. A quick visual overview of this work is available at [https://fdletf7qxe.github.io](https://fdletf7qxe.github.io).

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Federated Learning, Learning from Label Proportions, Weak Supervision

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bwJPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies federated learning from label proportions, where clients hold bag-level label proportions rather than instance labels. The authors propose a two-stage aggregation scheme. The first stage measures client compatibility in proportion space using a symmetric nearest-neighbor Jensen-Shannon set discrepancy, converts it to a similarity with an exponential kernel, keeps the top H=25 collaborators, and forms weights proportional to similarity times sample count. The second stage refines these weights with a class-wise complementarity term capped at 0.25, blended into the first-stage weights with mixing coefficient eta=0.05. Theoretically, Proposition 3.1 gives a covariance sign condition for improvement, Theorem 3.2 decomposes the improvement into a covariance term and a quality transfer term, and the blending is shown to satisfy a total-variation bound of eta. Experiments report best results on 10 of 12 settings.

Critical strengths
  1. Using bag-proportion geometry as a compatibility proxy is a reasonable answer to the problem, and the symmetric nearest-neighbor construction handles unaligned bags.
  2. The leave-one-collaborator-out experiment is the right check on whether the compatibility score measures what it claims to.
Critical weaknesses
  1. The class-complementarity refinement, adds only 1.0 point in Table 2 and is negative on two of the three Fashion-MNIST groups in Table 6 .
  2. The theoretical improvement conditions are stated in terms of covariances that the paper never measures, so the theory and the experiments do not connect.
Questions for authors
  1. Have the covariance quantities in Proposition 3.1 been measured on any of the benchmarks?
  2. Why is eta set to 0.05, and what does the accuracy curve over eta look like?
4 comments
AuthorsOpenReview ↗

Response to Reviewer bwJP

We sincerely thank the reviewer for the constructive feedback. Below, we address the concerns in the Weaknesses and Questions.

W1 (1): The class-complementarity refinement, adds only 1.0 point in Table 2.

RW1(1): We thank the reviewer for this comment. Although the average improvement in Table 2 is 1.0 percentage point, it represents a 2.83% relative gain over the strong V1 baseline, which we consider meaningful. To further substantiate this improvement, we conducted additional three-seed experiments under β=0.25\beta=0.25, as reported below.

DatasetV1FedC²Relative Gain
CIFAR-1019.41 ± 2.0819.88 ± 2.05+2.42%
SVHN18.56 ± 0.9721.69 ± 1.67+16.86%

Furthermore, we performed two-sided paired t-tests with Holm correction to assess the statistical significance of these improvements.

DatasetHolm-adjusted p-valueSignificance
CIFAR-100.0459Significant
SVHN0.0478Significant

These results demonstrate that class-wise refinement delivers substantial relative improvements, particularly on SVHN (+16.86%), and that the gains are statistically significant on both datasets (p<0.05p<0.05). Therefore, its contribution is both practically meaningful and statistically supported.

W1 (2): The class-complementarity refinement is negative on two of the three Fashion-MNIST groups in Table 6.

RW1(2): We thank the reviewer for this comment. To further validate class-wise refinement, we conducted three-seed experiments and report the results under β=0.25\beta=0.25 below.

DatasetLowMediumHigh
Fashion-MNIST+2.66 ± 3.21+1.75 ± 3.10+1.87 ± 3.36
CIFAR-10+0.77 ± 0.47+0.77 ± 0.21+0.36 ± 0.12
SVHN+1.40 ± 1.23+1.00 ± 1.82+1.05 ± 1.64

All nine dataset–coverage combinations show positive mean gains, including the previously negative groups on Fashion-MNIST. Despite seed-level variability, these results support the overall effectiveness of class-wise refinement.

W2: The theoretical covariance conditions are not empirically verified, leaving a gap between theory and experiments.

RW2: We thank the reviewer for this valuable comment. To bridge the gap between theory and experiments, we conducted additional analyses to empirically verify the covariance-based improvement conditions in Proposition 3.1 and Theorem 3.2. Since the exact reference classifier is unavailable, we approximate it by optimizing the classifier-only LLP objective from three different head initializations with the backbone fixed.

DatasetRoundProp. 3.1 Holds (%)Thm. 3.2 Holds (%)Prop. 3.1 Sign Agreement (%)Thm. 3.2 Sign Agreement (%)
CIFAR-10558.07 ± 3.3667.28 ± 1.6493.07 ± 1.6093.61 ± 1.53
2561.67 ± 1.8969.73 ± 1.6593.87 ± 2.5393.74 ± 2.66
5062.00 ± 1.7869.86 ± 2.7996.67 ± 0.1296.39 ± 0.51
Fashion-MNIST563.53 ± 2.5866.12 ± 1.4790.00 ± 2.6990.14 ± 1.01
2569.47 ± 0.3173.13 ± 1.1296.40 ± 0.8097.01 ± 0.82
5063.53 ± 0.4274.76 ± 0.7297.87 ± 0.6198.16 ± 0.82
Overall—63.04 ± 0.6970.15 ± 0.6394.64 ± 0.6994.84 ± 0.76

The results show that the theoretical improvement conditions hold in 63.04% and 70.15% of cases, respectively, with over 94% sign agreement across reference-head initializations. These findings establish a direct connection between the theoretical analysis and observed behavior.

Q1: Are the covariance-based improvement conditions in Proposition 3.1 satisfied in practice?

RQ1: As demonstrated in our RW2, the theoretical conditions hold in 63.04% and 70.15% of cases, respectively, with over 94% sign agreement, empirically supporting our theoretical analysis.

Q2: Why is η\eta set to 0.05, and how does it affect accuracy?

RQ2: As shown in Fig. 3(d) and our additional experiments below, η=0.05\eta=0.05 achieves the best or near-best accuracy across datasets. Therefore, we empirically select η=0.05\eta=0.05 as a suitable refinement strength.

η\etaCIFAR-10SVHNFashion-MNIST
0.0021.1919.6771.23
0.0521.7121.8271.05
0.1021.9521.7370.89
0.1521.0219.4465.97
0.2021.5520.7667.22
0.5014.8820.1162.77
0.7516.4120.3152.08
1.0016.2915.9146.44

The results show that moderate class-wise refinement improves accuracy over η=0\eta=0, validating the effectiveness of our proposed component.


Dear Reviewer bwJP,

Your questions helped us better validate class-wise refinement 💡. We hope the new results address your concerns. Thank you for your thoughtful feedback, and best wishes ✨.

AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

Dear Reviewer bwJP,

Thank you for your valuable feedback. We have provided our responses and would be happy to address any additional questions.

Best regards,
The Authors

Dear Reviewer bwJP,

Thanks for your time and effort in reviewing this paper. As the author-reviewer discussion period ends in about 24 hours, we would be grateful if you could take a look at our rebuttal and let us know whether your concerns have been addressed. We are happy to clarify any remaining questions before the deadline.

Regards from the authors.

Dear Reviewer bwJP,

Thanks for your time and effort in reviewing this paper. As the author-reviewer discussion period ends in about 4 hours, we would be grateful if you could take a look at our rebuttal and let us know whether your concerns have been addressed. We are happy to clarify any remaining questions before the deadline.

Regards from the authors.

Reviewer_73kTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies federated learning from label proportions under heterogeneous clients and proposes FedC2, a hierarchical collaboration framework that first selects compatible clients based on bag-proportion geometry and then performs class-wise refinement using target-relative coverage and local proportion-consistency signals. The paper also provides a theoretical analysis of the aggregation mechanism and evaluates the method on four image-classification benchmarks under multiple heterogeneity settings.

Critical strengths
  1. The paper addresses an underexplored Fed-LLP setting and clearly motivates the need for collaboration under heterogeneous proportion-supervised clients.
  2. FedC2 provides an intuitive coarse-to-fine design that combines client-level compatibility with class-wise complementarity.
  3. The paper provides detailed implementation settings, benchmark construction procedures, baseline configurations, and publicly available code.
Critical weaknesses
  1. The analysis depends on favorable relationships between the proposed weighting signals and unobservable classifier-parameter discrepancies. It also does not provide an end-to-end guarantee for the final global model.
  2. All datasets are converted into Fed-LLP benchmarks through Dirichlet partitioning and artificial bag construction. Evaluation on a naturally federated LLP dataset would better demonstrate practical applicability.
  3. The experiments appear to use a single random seed and mainly report point estimates without variance. Multi-seed results with standard deviations or confidence intervals would make the conclusions more convincing.
  4. Several FL and personalized-FL baselines are adapted to Fed-LLP using DLLP, which introduces additional implementation choices. A clearer justification of these adaptations and comparisons with more closely related class-aware collaboration methods would improve fairness.
Questions for authors

Please see Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 73kT (part 1/2)

We sincerely thank the reviewer for the constructive comments. Below, we address the concerns raised in the Weaknesses and Questions.

W1: The analysis depends on favorable relationships between the proposed weighting signals and unobservable classifier-parameter discrepancies. It also does not provide an end-to-end guarantee for the final global model.

RW1: We thank the reviewer for the valuable comments. To bridge the gap between theory and experiments, we conducted additional analyses to empirically examine the improvement conditions in Proposition 3.1 and Theorem 3.2. Since the exact reference classifier is unavailable, we approximate it by optimizing the classifier-only LLP objective from three different head initializations with the backbone fixed. The results are averaged over three seeds:

DatasetRoundProp. 3.1 Holds (%)Thm. 3.2 Holds (%)Prop. 3.1 Sign Agreement (%)Thm. 3.2 Sign Agreement (%)
CIFAR-10558.07 ± 3.3667.28 ± 1.6493.07 ± 1.6093.61 ± 1.53
2561.67 ± 1.8969.73 ± 1.6593.87 ± 2.5393.74 ± 2.66
5062.00 ± 1.7869.86 ± 2.7996.67 ± 0.1296.39 ± 0.51
Fashion-MNIST563.53 ± 2.5866.12 ± 1.4790.00 ± 2.6990.14 ± 1.01
2569.47 ± 0.3173.13 ± 1.1296.40 ± 0.8097.01 ± 0.82
5063.53 ± 0.4274.76 ± 0.7297.87 ± 0.6198.16 ± 0.82
Overall—63.04 ± 0.6970.15 ± 0.6394.64 ± 0.6994.84 ± 0.76

The improvement conditions hold in 63.04% and 70.15% of cases, respectively, with over 94% sign agreement across reference-head initializations. These findings provide empirical support for our theoretical analysis.

While our theory assumes a fixed representation and does not establish an end-to-end guarantee, our experiments validate the effectiveness of the complete FedC² framework across multiple datasets and heterogeneous settings.

W2: All datasets are converted into Fed-LLP benchmarks through Dirichlet partitioning and artificial bag construction. Evaluation on a naturally federated LLP dataset would better demonstrate practical applicability.

RW2: We thank the reviewer for the valuable suggestion. We conducted additional experiments on two complementary real-world datasets, covering naturally federated clients and real-world LLP bags.

First, we evaluate Fed-ISIC2019 [1], retaining its six original medical-center partitions as clients and constructing LLP bags within each client:

MethodFed-ISIC2019
FedAvg-DLLP35.10
FedProx-DLLP35.51
Ditto-DLLP37.91
FedC²48.22

FedC² achieves the highest global accuracy, outperforming the strongest baseline by 10.31 percentage points, demonstrating its effectiveness under naturally defined client partitions.

We further evaluate LLP-Bench Criteo [2], preserving its original feature-based LLP bags while simulating federated client partitions:

MethodLLP-Bench Criteo
FedAvg-DLLP77.01
FedProx-DLLP75.76
Ditto-DLLP76.19
FedC²77.19

FedC² also achieves the highest accuracy on LLP-Bench Criteo, supporting its effectiveness with real-world LLP bags.

References:

[1]: FLamby: Datasets and Benchmarks for Cross-Silo Federated Learning in Realistic Healthcare Settings, NeurIPS 2022.

[2]: LLP-Bench: A Large Scale Tabular Benchmark for Learning from Label Proportions, CIKM 2024.

AuthorsOpenReview ↗

Response to Reviewer 73kT (part 2/2)

W3: The experiments appear to use a single random seed and mainly report point estimates without variance. Multi-seed results with standard deviations or confidence intervals would make the conclusions more convincing.

RW3: We thank the reviewer for the valuable suggestion. We conducted additional experiments using three random seeds and report the mean accuracy ± standard deviation below:

DatasetMethodβ=0.25β=0.50β=0.75
Fashion-MNISTFedAvg-DLLP49.04 ± 5.9451.86 ± 11.9760.56 ± 2.16
FedProx-DLLP52.02 ± 4.5756.62 ± 9.2959.42 ± 2.29
Ditto-DLLP49.03 ± 5.9651.85 ± 12.0060.57 ± 2.17
FedC²69.30 ± 6.8967.35 ± 4.6066.27 ± 3.66
CIFAR-10FedAvg-DLLP14.58 ± 2.4213.12 ± 1.8613.19 ± 1.56
FedProx-DLLP14.20 ± 1.3213.65 ± 2.0713.04 ± 2.88
Ditto-DLLP12.54 ± 3.9611.27 ± 1.4110.91 ± 2.41
FedC²19.88 ± 2.0516.65 ± 1.6613.53 ± 1.24
SVHNFedAvg-DLLP19.67 ± 1.5217.38 ± 0.3611.87 ± 6.56
FedProx-DLLP18.34 ± 2.6818.29 ± 0.3617.36 ± 1.19
Ditto-DLLP15.43 ± 4.0014.12 ± 1.2116.25 ± 0.61
FedC²21.69 ± 1.6718.72 ± 0.4617.60 ± 1.69
Mini-ImageNetFedAvg-DLLP30.69 ± 0.6032.33 ± 0.4532.48 ± 0.31
FedProx-DLLP31.06 ± 0.7032.32 ± 0.4432.54 ± 0.31
Ditto-DLLP30.54 ± 0.8132.25 ± 0.5132.25 ± 0.21
FedC²30.48 ± 0.7632.55 ± 0.8332.91 ± 0.69

FedC² achieves the highest mean accuracy in 11 of 12 settings, providing further evidence of its effectiveness across different random seeds.

W4: Several FL and personalized-FL baselines are adapted to Fed-LLP using DLLP, which introduces additional implementation choices. A clearer justification of these adaptations and comparisons with more closely related class-aware collaboration methods would improve fairness.

RW4: We thank the reviewer for the constructive comments. To examine the influence of the LLP solver, we further compare DLLP[1], PVC [2], and LLP-DC [3] under β=0.5\beta=0.5:

LLP SolverMethodFashion-MNISTMini-ImageNet
DLLPFedAvg52.5131.81
FedProx56.0731.43
Ditto52.5031.65
FedC²67.5632.60
PVCFedAvg10.0024.25
FedProx57.0824.65
Ditto61.6522.56
FedC²67.7226.70
LLP-DCFedAvg10.0027.34
FedProx58.3330.12
Ditto10.3328.89
FedC²67.0830.82

FedC² consistently outperforms the baselines across all three LLP solvers on both datasets, indicating that its advantage is not specific to DLLP.

We further compare FedC² with three class-aware collaboration methods adapted to Fed-LLP under β=0.5\beta=0.5: cwFedAvg [4], FedPAC [5], and FedIC [6].

MethodCIFAR-10SVHNFashion-MNIST
cwFedAvg13.3814.5952.83
FedPAC9.9416.7510.94
FedIC12.4911.9731.15
FedC²17.5619.8167.56

FedC² achieves the highest accuracy on all three datasets, further demonstrating its effectiveness against closely related class-aware methods.

References:

[1]: Co-Training for Demographic Classification Using Deep Learning from Label Proportions, ICDMW 2017.

[2]: Learning from Label Proportions via Proportional Value Classification, ICLR 2026.

[3]: Learning from Label Proportion with Dual-Proportion Constraints, CVPR Findings 2026.

[4]: Class-Wise Federated Averaging for Efficient Personalization, ICCV 2025.

[5]: Personalized Federated Learning with Feature Alignment and Classifier Collaboration, ICLR 2023.

[6]: Intra-Class Unbiased Prototype Aggregation and Classifier Collaboration for Personalized Federated Learning, AAAI 2026.


Dear Reviewer 73kT,

Your questions helped us strengthen both our theoretical analysis and empirical evaluation 💡. We hope the new results address your concerns. Thank you for your thoughtful feedback, and we wish you continued success in your research ✨.

Thanks for the authors' responses. The authors have addressed my concerns adequately, so I will keep my positive score.

AuthorsOpenReview ↗

Thank You for Your Positive Feedback

We sincerely thank you for your encouraging feedback and for taking the time to consider our responses. We greatly appreciate your positive assessment of our work!

Reviewer_mJr6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the Fed-LLP problem, proposed FedC², a two-stage collaborative framework that transitions from client compatibility to class complementarity. First, FedC² measures client compatibility using a symmetric nearest-neighbor set discrepancy based on the JS divergence between bag-level label-proportion vectors across clients; this guides the selection of suitable collaborators and the construction of compatibility-guided aggregation weights.. Then, it estimates class-level complementarity by considering both the class coverage gaps of the target client and the consistency of bag-proportion predictions among collaborators, enabling a weighted aggregation for the classification head. Experimental results demonstrate that this approach outperforms existing methods across several benchmarks.

Critical strengths
  1. Clear writing and presentation.

The paper is well-written and clearly organized , making the method easy to follow.

  1. Meaningful and interesting research problem.

The paper addresses an important and relatively underexplored problem: cross-client collaboration in federated learning from label proportions under data heterogeneity. The proposed perspective of combining client compatibility with class-wise complementarity is interesting.

  1. Appropriate experimental evaluation and theoretical analysis.

The paper provides a theoretical analysis to characterize the proposed aggregation strategy and evaluates its effectiveness across four benchmarks and three levels of client heterogeneity. The ablation studies, sensitivity analyses, and mechanism analyses further support the result.

Critical weaknesses
  1. Lack of evaluation under different client participation rates.

The paper assumes full client participation in every communication round. However, partial client participation is a common setting in federated learning. Why does the paper not consider different client participation rates? It remains unclear how FedC² performs when only a subset of clients is active, particularly when important collaborators are unavailable.

  1. Reliability estimation based on softmax predictions.

The paper measures class-wise reliability by comparing the average softmax predictions within each bag with its observed label proportions. However, matching bag-level proportions does not necessarily indicate accurate instance-level classification. Why is softmax-based proportion consistency considered a suitable measure of a collaborator's reliability? Would using the proportions of hard predictions obtained through Argmax provide a more meaningful signal?

Questions for authors

As mentioned in weakness

  1. Why does the paper only consider full client participation? How would FedC² perform under different client participation rates, especially when important collaborators are unavailable?

  2. Why does the proposed method use averaged softmax predictions rather than Argmax-based class proportions to estimate collaborator reliability? Could the authors justify this choice and provide an experimental comparison between these two approaches?

4 comments
AuthorsOpenReview ↗

Response to Reviewer mJr6 (part 1/2)

We sincerely thank the reviewer for the positive assessment of our work and the constructive suggestions. Below are our responses to the comments in Weaknesses and Questions.

W1: Lack of evaluation under different client participation rates. It remains unclear how FedC² performs when only a subset of clients is active, particularly when important collaborators are unavailable.

RW1: We thank the reviewer for the valuable comments. We have conducted additional experiments under different client participation rates as shown below.

DatasetMethod20%40%60%80%100%
Fashion-MNISTFedAvg39.6945.2155.2751.4052.51
FedProx37.0446.2249.0252.3856.07
Ditto37.3448.5753.0151.2852.50
FedC²20.0534.4353.3459.1367.56
CIFAR-10FedAvg12.8812.3811.9112.7412.33
FedProx13.0412.4112.8012.0313.38
Ditto10.458.288.4311.6810.55
FedC²11.8113.2313.9916.0717.56
SVHNFedAvg18.5716.3016.5919.8719.59
FedProx19.4018.0418.5918.7919.41
Ditto14.8716.9016.0015.4316.25
FedC²14.3416.8717.9818.9019.81

As participation increases from 20% to 100%, FedC² improves by 47.51, 5.75, and 5.47 points on Fashion-MNIST, CIFAR-10, and SVHN, respectively, showing larger gains than the baselines. Notably, at 80% participation, FedC² surpasses the strongest baselines by 6.75 and 3.33 points on Fashion-MNIST and CIFAR-10.

We further evaluate the unavailability rate (%) of each client's most compatible collaborator under partial participation:

Dataset20%40%60%80%
Fashion-MNIST81.4060.5040.2720.50
CIFAR-1081.4061.4043.0720.85
SVHN84.0059.0039.6020.60

The unavailability rate decreases from approximately 82% to 21% as participation increases from 20% to 80%. The corresponding performance improvements suggest that FedC² benefits from greater access to compatible and complementary collaborators.

Q1: Why does the paper only consider full client participation? How would FedC² perform under different client participation rates, especially when important collaborators are unavailable?

RQ1: As demonstrated in our RW1, FedC² benefits substantially from increased client participation, achieving the highest accuracy on two of three datasets at 80% participation and all three at 100%.

AuthorsOpenReview ↗

Response to Reviewer mJr6 (part 2/2)

W2: Reliability estimation based on softmax predictions. The paper measures class-wise reliability by comparing the average softmax predictions within each bag with its observed label proportions. However, matching bag-level proportions does not necessarily indicate accurate instance-level classification. Why is softmax-based proportion consistency considered a suitable measure of a collaborator's reliability? Would using the proportions of hard predictions obtained through Argmax provide a more meaningful signal?

RW2: We thank the reviewer for this insightful comment. We clarify that our reliability estimation measures bag-level proportion consistency rather than instance-level classification accuracy. Since LLP provides bag-level supervision, we adopt averaged softmax predictions to preserve predictive uncertainty and align with the probabilistic LLP objective [1], whereas Argmax discards confidence information through hard assignments [2].

We further compare Softmax and Argmax-based reliability estimation under different Dirichlet concentration parameters (β) below:

DatasetReliabilityβ=0.25β=0.50β=0.75
CIFAR-10Softmax21.7117.5613.16
Argmax22.0316.5911.69
SVHNSoftmax21.8219.8117.18
Argmax21.9318.4818.75
Fashion-MNISTSoftmax71.0567.5666.24
Argmax71.8165.0170.82

Softmax performs better across all three datasets at β=0.5, while Argmax performs better in some other settings.

Q2: Why does the proposed method use averaged softmax predictions rather than Argmax-based class proportions to estimate collaborator reliability? Could the authors justify this choice and provide an experimental comparison between these two approaches?

RQ2: As demonstrated in our RW2, Softmax provides a confidence-aware reliability proxy aligned with LLP supervision, while Argmax offers an alternative based on hard predictions. Their relative effectiveness varies across settings.

References:

[1]: Learning from Label Proportion with Dual-Proportion Constraints.

[2]: Distilling the Knowledge in a Neural Network.


Dear Reviewer mJr6,

Thank you again for taking the time to consider our response. We hope the clarifications and additional results have answered your questions 🙏. Wishing you all the best in your research 🌟.

Thanks for the authors' responses. The authors have addressed my concerns adequately, so I will keep my positive score.

AuthorsOpenReview ↗

Thank You for Your Encouraging Feedback

We are glad that our additional experiments and clarifications addressed your concerns. We sincerely thank you for your encouraging response and positive assessment!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.