Taming Cross-Domain Representation Variance in Federated Prototype Learning with Heterogeneous Data Domains

Federated learning (FL) allows collaborative machine learning training without sharing private data. While most FL methods assume identical data domains across clients, real-world scenarios often involve heterogeneous data domains. Federated Prototype Learning (FedPL) addresses this issue, using mean feature vectors as prototypes to enhance model generalization. However, existing FedPL methods create the same number of prototypes for each client, leading to cross-domain performance gaps and disparities for clients with varied data distributions. To mitigate cross-domain feature representation variance, we introduce FedPLVM, which establishes variance-aware dual-level prototypes clustering and employs a novel $α$-sparsity prototype loss. The dual-level prototypes clustering strategy creates local clustered prototypes based on private data features, then performs global prototypes clustering to reduce communication complexity and preserve local data privacy. The $α$-sparsity prototype loss aligns samples from underrepresented domains, enhancing intra-class similarity and reducing inter-class similarity. Evaluations on Digit-5, Office-10, and DomainNet datasets demonstrate our method's superiority over existing approaches.

Paper

Similar papers

Peer review

Reviewer gr7B6/10 · confidence 5/52024-06-14

Summary

This paper studies federated learning under heterogeneous data domains and introduces a federated prototype learning strategy, denoted as FedPLVM, to mitigate the problem. A dual-level prototype generation method is proposed to address domain variance between hard and easy domains, reducing the communication burden between clients and the server. Moreover, an \alpha-sparsity prototype loss is proposed to enhance the sparsity of the prototype representations. The experiment results demonstrate the efficacy of the proposed method and the necessity of each proposed module.

Strengths

1. The paper is overall well-written and easy to follow. 2. The authors explain why models exhibit varying performance across domains. 3. The experiment results show the superiority of the proposed method compared to SOTA methods. The ablation results also indicate the effectiveness of each proposed module.

Weaknesses

1. The proposed dual-level clustering strategy appears similar to FPL. What is the key difference between them? The primary distinction lies in the proposed \alpha-sparsity loss, which introduces incremental innovation to the paper. 2. The proposed dual-level prototype strategy utilizes FINCH for clustering. However, the clustering results from FINCH can vary at different steps. Which step's result is selected for use? Do results from different steps vary? 3. The impact of \alpha is only evaluated on small-scale datasets. Does the trend consist with big-scale datasets like DomainNet? 4. In Sec. G, the authors claim that the evaluation protocol used in FPL is unfair. However, the protocol used in FPL is sensitive to both easy and hard datasets.

Questions

This paper is well-organized and the experiment results significantly surpass existing methods. But the level of innovation is questionable: the dual-level prototype generation is similar to FPL, and the only key difference is the \alpha-sparsity prototype loss. Moreover, I also have some concerns on this paper. For more details, please check the weakness part.

Rating

6

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes. The authors addressed limitations of the paper.

Reviewer 8nBN4/10 · confidence 5/52024-06-15

Summary

This paper focuses on Federated Prototypes Learning and reveals that existing methods create the same prototypes for different clients, which neglects the distribution diversity. In this work, the authors introduce the variance-aware dual-level prototype clustering and alpha sparsity prototype loss. Various experiments demonstrate the effectiveness of the proposed method.

Strengths

Authors conduct comprehensive experiments to demonstrate the effectiveness.

Weaknesses

This paper has several drawbacks. 1. Motivation for the α-Sparsity Prototype Loss. It is a little strange that "This multiplicity of prototypes could potentially lead to overlapping feature representations among different classes, especially in challenging client scenarios". As the authors claim the feature overlapping, the naive solution is to use other ways to cover the local feature behavior rather than multiple local prototypes. Furthermore, I assume that authors could employ the OT (optimal transport) to require them to concentrate on different parts. I do not think the current operation is a suitable way to deal with the feature overlapping. 2. The paper architecture is not suitable. I wonder why authors spend a lot of space to write meaningless or complicated words and even leave no space for DomainNet and Office-10. Furthermore, I encourage to consider the label skew effect, ie, with different label skew degree. Besides, the experiments on large scale of clients are also important. 3. The Figure 1 is confusing. Why monitor would capture the digitis figure? It is a not soundable problem figure.

Questions

I encourage the author to take careful thinking for the federated prototype learning field. The existing solution shares a high similarity with the FPL (CVPR'23) and seems like an incremental work.

Rating

4

Confidence

5

Soundness

3

Presentation

2

Contribution

2

Limitations

Yes, the authors have discussed.

Reviewer 8nBN2024-08-09

Thank you for the AC tips. I have provided more details in my response to the review. As for the details of the review, 1. The naive solution is to utilize a Gaussian distribution to describe class information. For example, FedFA [1] is based on Gaussian modeling of feature statistic augmentation. I encourage the authors to compare the conceptual differences between local prototype clustering and class Gaussian construction. Additionally, the authors should clearly explain the advantages of utilizing a dual-level prototype approach. 2. With respect to paper organization, the authors should spend more space on concept comparison with existing methods. But the authors focus on preliminary discussion. Thank you for authors feedback. I still have the following questions: 1. You refer to a paper from 2019, but many papers related to Optimal Transport (OT) have been published in recent years. For instance, FedOTP [2] introduces the OT loss in regularization. 2. The authors did not address question 3: “Figure 1 is confusing. Why would a monitor capture the digits? It does not seem like a logical problem illustration.” 3. You mentioned that “This loss function reasonably utilizes the variance information to reduce feature similarity across different classes while increasing it within the same class, promoting more effective and stable learning.” Could the authors provide the corresponding theoretical or convergence analysis? Additionally, regarding “which mitigates the potential risk of feature representation overlapping,” could the authors offer visualizations to support this claim rather than relying on terms like “potential”? [1] FedFA: Federated Feature Augmentation [2] Global and Local Prompts Cooperation via Optimal Transport for Federated Learning. CVPR 2024

Authorsrebuttal2024-08-11

Thank you for your further comments. For the details of the review: 1. FedFA aims to augment the feature statistic by constructing a vicinity Gaussian distribution for each data point and change the empirical risk to the vicinal risk. Such an augmentation operation can enrich the feature information after the feature extractor layer. However, this approach has its limitations, especially in cases where data volume is small or distribution is uneven. In these scenarios, the constructed Gaussian distribution might not fully capture the true diversity of the data and can have bias. Also, it requires additional computation regarding the distribution parameters locally for each feature extractor layer with augmentation. In contrast, our proposed FedPLVM utilizes only the feature representation of the final feature extractor layer to cluster several local prototypes regardless of the even or uneven feature distribution and further performs global clustering and combines with the $\alpha$-sparsity loss in the local training. Lastly, FedFa is one of the baselines we compared in our paper and the result shows our method outperforms FedFA in Digit-5, Office-10 and DomainNet. 2. We appreciate your suggestions regarding the organization of our paper. However, we must respectfully disagree with the comment that excessive space is devoted to preliminary discussions at the expense of comparisons with existing methods. Firstly, the preliminary section occupies only half a page and concisely presents essential knowledge required for understanding our proposed method. This information is crucial and cannot be further condensed without loss of clarity. Secondly, we have integrated comparisons with existing methods throughout the introduction and methodology sections. Specifically, our comparison with FPL, which is most relevant to our work, includes dedicated paragraphs discussing the differences (see lines 232-247). For Questions: 1. We appreciate for suggesting the paper on FedOTP. While FedOTP is indeed a noteworthy contribution, we believe it addresses a different aspect from that of our paper. FedOTP focuses on developing efficient collaborative prompt learning strategies to integrate pre-trained visual-language models into federated learning frameworks. This topic, although valuable, is not closely related to the core focus of our research. 2. The icon in Figure 1 of our paper mentioned in your question is not that of a monitor but a laptop. In FL, the client device doesn’t necessarily need to capture the data itself. In our illustration, the laptop processes the stored digital data without needing to capture figures or photos directly. 3. We thank you for your advice on including a convergence speed analysis. We agree that theoretical convergence analysis is a crucial aspect of federated learning. Typically, such analyses in existing FL research depend heavily on strong assumptions about gradient bounds and the differences between global and local loss functions. Importantly, incorporating prototype learning alters the loss function itself. Since FL convergence analyses require the specific assumptions regarding the loss functions, the advantages of integrating prototype learning into the loss function might not be readily apparent. This limitation is also why existing FL studies that utilize prototype learning, such as FPL, do not provide theoretical convergence analysis. Instead, to address concerns regarding convergence speed, we have provided experimental results that clearly demonstrate the improvements our method achieves. In our Digit-5 experiment, we performed 100 global rounds as stated in our paper and compared the convergence speed with and without our proposed $\alpha$-sparsity loss. We reported the average test accuracy after each 10 rounds training. Obviously with the $\alpha$-sparsity loss we have a faster convergence speed (70 rounds compared to 80 rounds without the proposed loss function) and a higher converged accuracy: | Round | 10 | 20 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 | | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | w/o $\alpha$-sparsity loss | 35.14 | 56.44 | 57.58 | 62.20 | 64.90 | 65.72 | 65.02 | 66.66 | 66.32 | 66.06 | | w/ $\alpha$-sparsity loss | 54.12 | 62.44 | 65.32 | 66.12 | 67.32 | 68.20 | 69.62 | 69.22 | 69.86 | 69.26 | 4. We appreciate your suggestion to include visualizations to support our claims. However, during the author-reviewer discussion period, we are unable to add new figures. Should our paper be accepted, we will certainly include the recommended visualizations in the final version of the manuscript as advised.

Reviewer dkgn6/10 · confidence 3/52024-07-08

Summary

This paper aims to investigate the federated prototype learning problem with data heterogeneity. To handle the cross-domain representation variance problem, a new method termed FedPLM is proposed, which includes a dual-level prototype clustering mechanism and an alpha-sparsity prototype loss. Experiments are conducted to show the effectiveness of the proposed method.

Strengths

1. The proposed method is well-designed for the specific research problem and has good novelty. 2. The motivation of the paper is solid. 3. The experiments are comprehensive to discuss the properties of the proposed method. 4. The paper is well-written and easy to follow.

Weaknesses

1. On several datasets (e.g. MNIST and USPS), the performance of the proposed method is not significant enough. Can the author give more explanation for this point? 2. The running efficiency of the proposed method is not given. Complexity analysis and efficiency-related experiments are expected to discuss the efficiency of the proposed method.

Questions

The authors are expected to address the concerns in the block of "Weaknesses". Also, I have one more question: How to balance the two sub-terms (contra and corr) in alpha loss? As they may be on different scales, is that feasible to directly add them together?

Rating

6

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

Limitations are discussed in the paper. For the negative societal impact, I didn't find any concern from my side.

Reviewer GDmh5/10 · confidence 4/52024-07-11

Summary

The paper introduces FedPLVM, a federated learning approach that improves federated prototype learning (FedPL) in heterogeneous data domain setting. Traditional FedPL methods create the same number of prototypes for each client, leading to performance disparities across clients with different data distributions. FedPLVM mitigates this by introducing variance-aware dual-level prototype clustering and an α-sparsity prototype loss. Specifically, each client first clusters the feature vectors of all same-class local samples to get local clustered prototypes. Then, the server conducts global prototype clustering to reduce the number of prototypes for each class. Experiments show that FedPLVM outperforms the other baselines on the data from multiple domains.

Strengths

1. The motivation is clear. It makes sense to get different prototypes under tasks with different difficulties. 2. The proposed \alpha-sparsity prototype loss is interesting. The ablation study about the loss and dual-level prototype generation is clear.

Weaknesses

1. The paper misses important baselines. Personalized federated learning studies are not studied, which should be suitable in the studied setting. 2. The computation and communication cost of FedPLVM compared with other studies is not presented. 3. The clustering method needs more details. It is not clear why clustering can address the heterogeneous domain issue. 4. The work is incremental work based on FedPL.

Questions

1. From Algorithm 1, the goal of FedPLVM is to train a single global model. From the experiments, the model is tested on the local datasets. Given data from different domains, why don’t the parties train personalized local models with personalized federated learning? 2. The paper mentioned that more prototypes are needed for harder domains. For the clustering method, is there a guarantee that the harder domain will also get more local prototypes? The paper may add more analysis or experiments to demonstrate it. 3. Will FedPL have significant higher computation and communication cost than FedAVg?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Reviewer e95T5/10 · confidence 5/52024-07-12

Summary

The domain gap among multiple clients impedes the generalization of federated learning. To mitigate cross-domain feature representation variance, the authors introduce FedPLVM, which establishes variance-aware dual-level prototypes clustering and employs a novel α-sparsity prototype loss. To verify the effectiveness of the proposed method, extensive experiments are conducted.

Strengths

1. This paper is well organized and written in a way that is easy to understand. 2. The experimental design is reasonable and a large number of experiments have also proved the effectiveness of the proposed methodology.

Weaknesses

1. It makes sense to perform global level prototype clustering on the server side, while performing intra-class prototype clustering on the local side makes unclear sense, why it is advantageous to get more prototypes for hard domain samples? 2. In Eq. (8), what is meaning of C_y? 3. From Fig. 4, \alpha has a significant impact on the results, which is not conducive to the robustness of the algorithm. 4. The latest compared method in the paper is from 2023 and lacks the latest comparison method.

Questions

Please refer to the weaknesses.

Rating

5

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

Please refer to the weaknesses.

Reviewer GDmh2024-08-10

Thanks for the author's response. Some of my concerns have been addressed. I'll increase my score to 5.

Authorsrebuttal2024-08-11

Dear Reviewers, We sincerely appreciate your extensive efforts in reviewing and commenting on our work. We hope that our rebuttal has effectively addressed your comments and concerns. As the author-reviewer discussion period is approaching to its end on August 13th, we kindly encourage you to reach out if you have any further questions or points of discussion. Your feedback is highly valued, and we are eager to address any additional remaining concerns you may have. Thank you very much for your attention to this matter. Anonymous Authors

Reviewer gr7B2024-08-11

Thank you for your response, some of my concerns are solved. However, I still do not find a distinction between dual-cluster and FPL. And for R4, the weighted average could be another valuable evaluation protocol but I do not think the protocol used in FPL and other methods is unfair. Thus, I would like to maintain my score.

Authorsrebuttal2024-08-11

Thank you for your continued feedback. We appreciate the opportunity to further clarify our approach. 1. We would like to further clarify the distinctions between our method and FPL. Our dual-level prototype clustering significantly differs from the single-level global prototype clustering FPL, both in purpose and effectiveness, as we detailed in our previous response and demonstrated in the visualization results of our paper. The local prototype clustering is crucial because it captures essential variance information, rather than just representing the average feature as in FPL. This is particularly important in challenging domains. For instance, as shown in Figure 1 of the attached PDF, the easier domain (MNIST) generates noticeably fewer prototypes compared to the harder domain (SVHN), which underscores the necessity of our local prototype clustering approach. Moreover, our proposed $\alpha$-sparsity loss introduces an innovative and effective enhancement to our dual-level prototype clustering. We conducted additional experiments to highlight the advantages of the $\alpha$--sparsity loss in terms of both convergence speed and final accuracy. In the Digit-5 experiment, we performed 100 global rounds as stated in our paper and compared the convergence speed with and without the $\alpha$-sparsity loss. The results, which report the average test accuracy after every 10 rounds, clearly show that the $\alpha$--sparsity loss leads to faster convergence (achieving it in 70 rounds compared to 80 rounds without the proposed loss) and higher final accuracy. | Rounds | 10 | 20 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 | |:-----------------:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:| | w/o \(\alpha\)-sparsity loss | 35.14 | 56.44 | 57.58 | 62.20 | 64.90 | 65.72 | 65.02 | 66.66 | 66.32 | 66.06 | | w/ \(\alpha\)-sparsity loss | 54.12 | 62.44 | 65.32 | 66.12 | 67.32 | 68.20 | 69.62 | 69.22 | 69.86 | 69.26 | 2. We appreciate your insightful comment. Following your suggestion, we have calculated the additional average accuracy using the protocol in FPL. Our updated results show that the average accuracy of our proposed method is 77.07%, compared to 75.54% for FPL. This further demonstrates the superior performance of our method.

Reviewer e95T2024-08-12

Thank you for your responses. In light of the novelty and other reviewers' comments, I tend to maintain my score.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC