FedL2P: Federated Learning to Personalize

Federated learning (FL) research has made progress in developing algorithms for distributed learning of global models, as well as algorithms for local personalization of those common models to the specifics of each client's local data distribution. However, different FL problems may require different personalization strategies, and it may not even be possible to define an effective one-size-fits-all personalization strategy for all clients: depending on how similar each client's optimal predictor is to that of the global model, different personalization strategies may be preferred. In this paper, we consider the federated meta-learning problem of learning personalization strategies. Specifically, we consider meta-nets that induce the batch-norm and learning rate parameters for each client given local data statistics. By learning these meta-nets through FL, we allow the whole FL network to collaborate in learning a customized personalization strategy for each client. Empirical results show that this framework improves on a range of standard hand-crafted personalization baselines in both label and feature shift situations.

Paper

References (65)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Jjav5/10 · confidence 3/52023-06-18

Summary

This study primarily concentrates on refining hyperparameters in federated learning, especially concerning non-IID data. The researchers highlight two key hyperparameter groups—Batch Normalization (BN) and selective update hyperparameters—that yield impressive outcomes. BN hyperparameters help mitigate feature shifts, a prevalent problem in federated learning. FedBN, a cited method, aids in managing feature shift more effectively by localizing the BN layers while keeping other layers global. Details of the BN layer formation and its constituents are also given. Moreover, the study explores a global model's application, initially trained through existing federated algorithms, which is then fine-tuned to the client set. The CIFAR-10 dataset split among various clients is utilized for this process, demonstrating that their approach maintains global performance while enhancing personalized performance. The researchers also incorporate the Jacobian product to approximate the Hessian inverse and calculate the hyper gradient, forming a part of their suggested algorithm. For experimental purposes, ResNet-18 and SGD are consistently employed for all tests. Batch size, local epoch number, learning rate, hyper gradient clipping, maximum communication rounds, and other parameters are also predefined. This paper contributes to federated learning by introducing a method to optimize hyperparameters, especially in non-IID data scenarios, and backs its effectiveness through empirical results.

Strengths

**Originality:** The paper stands out due to its unique approach to handling non-IID data within federated learning. Implementing Batch Normalization (BN) hyperparameters to tackle feature shift is an inventive fusion of existing deep learning knowledge and federated learning's distinct hurdles. Moreover, the employment of the Jacobian product to estimate the Hessian inverse and compute the hyper gradient is a distinct application within the federated learning sphere, marking an inventive merge of established concepts to address a recurring federated learning issue. **Quality:** The paper's quality appears robust. It reflects the authors' thoughtful research methodology and the theoretical integrity of their approach to hyperparameter optimization. Leveraging the CIFAR-10 dataset for validation is a norm in this field, and the authors thoroughly outline their experimental setup. Nevertheless, a broader evaluation across diverse datasets and juxtaposition with other methodologies could further enrich the paper's quality. **Clarity:** The paper is well-crafted and clearly depicts the research. The authors present exhaustive descriptions of their methodology and the theoretical underpinnings of their strategy. They also lucidly detail their experimental setup and findings. Such explicitness in communication renders the paper comprehensible to a broad audience, from novices in federated learning to seasoned researchers. **Significance:** The paper holds substantial significance. It addresses pivotal challenges, such as dealing with non-IID data and optimizing hyperparameters, vital in federated learning. The authors' solutions to these challenges could potentially lead to the more efficient and practical training of federated learning models, making a notable contribution to the field. Additionally, this paper's insights could stimulate new research pathways and methodologies within the broader realm of machine learning.

Weaknesses

In summary, while the paper is fundamentally and conceptually solid, there are a few weaknesses that weaken its contribution and soundness. 1. **Dataset and Model Constraints:** The study principally uses the CIFAR-10, Office-Caltech-10, and DomainNet datasets in its experiments. Although these are recognized standards, the findings may not extend to different datasets and models. Testing the proposed method across diverse datasets (such as images, text, and tabular data) and various model architectures could better showcase the method's adaptability and resilience. 2. **Comparison to Other Techniques:** A broader comparison with cutting-edge methods could enhance the paper. While some algorithms are referenced, the relative performance of the authors' method remains unclear. A side-by-side comparison could clarify the comparative advantages and disadvantages of their approach. 3. **Computational Overhead:** The authors suggest the Jacobian product for approximating the Hessian inverse and computing the hypergradient, a fresh perspective. Appendix C mentions that the computational cost compared to FO / HF-MAML is still relatively large. 4. **Applicability:** The paper's significant contribution to federated learning cannot be overlooked, yet the universal applicability of their method warrants examination. Federated learning is applied across various sectors, each with unique data distributions, privacy stipulations, and computational resources. Discussing the limitations and potential solutions of their method would fortify the paper. 5. **Reproduction**: The codes are not available yet.

Questions

1. Could you elaborate on how generalizable your method is to different federated learning scenarios? For instance, how would your method perform with different data distributions, privacy requirements, and computational resources? 2. Could you provide a more comprehensive comparison with other state-of-the-art methods? It would be helpful to understand how your method performs relative to these other approaches in terms of accuracy and computational efficiency. I am sill unclear about whether your baselines cover enough range of comparable works. 3. What particular `pre-trained` models did you use in your experiments? 4. Additional Datasets and Models: Have you considered testing your method on various datasets with different characteristics (e.g., image, text, tabular data) and different model architectures? 5. Could you discuss the limitations of your approach and potential ways to overcome these limitations? This would provide a more balanced view of your work and could help guide future research in this area.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

In terms of broader societal impacts, it's not clear from the provided excerpts whether the authors have considered this aspect.

Reviewer wFN96/10 · confidence 3/52023-07-06

Summary

This paper proposes a new personalization strategy for the Federated Learning setting. Their key idea is to train a meta-network that takes as input the client's data and outputs optimal hyperparameters for that data. These optimal hyperparameters can then be used to personalize the pretrained model. The hyperparameters they choose to output control the batch-norm mean, std-dev and per layer learning rate. They evaluate their technique on 4 datasets, and show that their technique outperforms other personalization strategies on these datasets.

Strengths

- The core idea of using a meta-net proposed by the authors is quite novel and I believe it is a very promising direction. - The hypergradient technique they propose to train this metanet is also quite novel and interesting. - For the most part their algorithm and technique is clearly presented.

Weaknesses

- The experiment setup is not clearly defined and confusing. Examples - In their CIFAR10 experiments, they say 20% of the data is used for test, it's not clear if that means 20% of the clients are not used in the FL step or 20% of the data in each client is not used in the FL step. - It is not clear what fraction of clients are used for training the FL model in each expt - Figures 2 and 3 are poorly labeled and not clearly explained. Figure 3 is completely unreadable. - Experimental results are not very strong - FedL2P seems very computationally expensive (total compute used is not compared anywhere) and the results are only marginally better than other techniques. - No evaluation on actual non-IID datasets. There are several non-IID datasets available publicly for FL - EMNIST, Shakespeare, StackOverflow etc. Evaluating their method on these datasets would make the results much more compelling. - No evaluation on non-image datasets. - Hyperparameters limited to continuous values - many hyperparameters like # epochs, choice of optimizer etc. are not continuous.

Questions

N/A

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

3 good

Limitations

N/A

Reviewer aauE5/10 · confidence 3/52023-07-10

Summary

The paper introduces a novel approach that utilizes meta-learning to learn hypernets that aim to determine the optimal fine-tuning hyperparameters for individual clients. The proposed idea is intriguing and holds promise. The results presented in the paper are compelling. However, the paper would benefit from improvements in terms of writing and clarity. Overall, it is a borderline paper. --- post rebuttal -- In the rebuttal, the author address my questions about clarity and positioning their work. So I increase my score from 4 to 5.

Strengths

- The paper proposes a novel idea of using meta-learning to learn hypernet that can estimate the best fine-tuning hyperparameters for personalizing local models. - The results show the proposed FedL2P can effectively improve performance on multiple standard FL methods and remain robust when there is a data distribution shift across clients.

Weaknesses

- The **clarity** needs to be improved. Especially, some details of the proposed method are missing. Please see the question section below. - The proposed method is **limited** to one certain type of personalized federated learning, i.e., fine-tuning local models. There are other strategies of personalization that are not applicable to this work. For example, one strategy is decoupled model (global shared encoder + local predictive head) [1,2]. Another strategy could be using Hypernet to directly optimize different local models [3]. [1] Zhong, A., He, H., Ren, Z., Li, N. and Li, Q., 2022, September. FedDAR: Federated Domain-Aware Representation Learning. In The Eleventh International Conference on Learning Representations. [2] Collins, L., Hassani, H., Mokhtari, A. and Shakkottai, S., 2021, July. Exploiting shared representations for personalized federated learning. In International conference on machine learning (pp. 2089-2099). PMLR. [3] Shamsian, A., Navon, A., Fetaya, E. and Chechik, G., 2021, July. Personalized federated learning using hypernetworks. In International Conference on Machine Learning (pp. 9489-9502). PMLR.

Questions

- (line 184) Could you explain what is "the degree of data heterogeneity of the client’s local data with respect to the pretrained model"? The data heterogeneity comes from local data and the pretrained model? - (line 205) Should $BNNet(w_{bn}; ξ_1, ξ_2, . . . , ξ_{B−1}, ξ_B)$ be $BNNet(ξ_1, ξ_2, . . . , ξ_{B−1}, ξ_B;w_{bn})$? It seems that $ξ_1, ξ_2, . . . , ξ_{B−1}, ξ_B$ are BNNet's input. - (line 206) Could you explain why there is a Hadamard product for the LRNet's output? - title of section 3.3 is a bit weird. why you are Define Training/Validation Losses for IFT (implicit function theorem). Isn't it that training/validation losses should defined by your tasks?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Not discussed.

Reviewer nqFj6/10 · confidence 4/52023-07-26

Summary

The paper introduces a method for fine-tuning a client's model based on a pretrained model using its local data. The proposed personalization approach involves fine-tuning the learnable weights and biases (assume, there are $M$ such layers) as well as the means and standard deviations (SDs) of batch normalization (BN) layers (assume, there are $B$ BN layers). This fine-tuning process is facilitated by two meta neural networks. The first meta neural network predicts $2M$ learning rates, which correspond to each weight and bias layer, enabling the local model to perform several gradient steps from the pretrained model. The input for this first meta neural network consists of the mean and standard deviation values for each of the $M$ layers. On the other hand, the second meta neural network directly predicts the means and SDs. Its input comprises the distance between the pretrained means and the running means of the model with the pretrained weights on local data, as well as a similar distance measure between the SDs. Both meta neural networks are trained using FedAvg on hyper gradients. -------- Post-rebuttal update: I did not have significant concerns about this paper. I acknowledge the authors' rebuttal and extend my gratitude to them, particularly for adding the new experiment. I have opted not to modify my favorable rating of 6.

Strengths

1. The paper demonstrates sufficient originality by introducing a novel approach using meta neural networks for computing personalized hyperparameters, as outlined in the summary above. Notably, the concept of trainable meta nets holds promise and appears to have practical applicability. 2. The authors effectively elaborate on the algorithm's intricacies, providing comprehensive descriptions of various components of the main method. Additionally, Figure 1 illustrates a schematic depiction of the main approach, enhancing the clarity of their explanations. 3. Extensive computational experiments are incorporated within the paper to substantiate the efficacy of the proposed method. 4. The paper is a valuable contribution to the fields of personalized Federated Learning and Meta Learning.

Weaknesses

1. Despite the authors' efforts in providing extensive explanations for their ideas, regrettably, the text still contains several notable inconsistencies, as detailed in the "Questions" section. 2. A comparative analysis with the FOMAML method [1] is notably absent in the work. It is worth mentioning that FOMAML shares significant similarities with the authors' proposed method, as both utilize the MAML approach for fine-tuning the model using the training dataset and making optimization steps for the meta parameter based on the validation dataset, as observed in the current study. Additionally, in another related work [2], which builds upon MAML, the hyper gradient is explicitly addressed (see lines 14 and 16 of Algorithm 1 in [2]). [1] "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks", Finn et al. [2] "Unsupervised Meta-Learning for Few-Shot Image Classification", Khodadadeh et al.

Questions

1. Line 225: If I understand correctly, the text implies that $\partial_{{\omega}_{lr}} {\cal L}_V = 0$. If my interpretation is correct, could you please explain why this is the case? 2. Equation 3 & Line 199: What is the range of the index $i$? Minor comments & suggestions: 1. Line 6: Depending -> depending 2. Line 35: Even -> even 4. Inaccuracy on lines 87-91: Although it is evident what the authors intended to convey, the wording is imprecise. I would recommend specifying the subset of chosen clients as $\tilde{C}$ on line 87, as done in Algorithm 1, and then using $i \in \tilde{C}$ instead of $i \in C$ on line 88. The statement $P_i = \cal{X} \times \cal{Y}$ on line 90 is also misleading: $\cal{X} \times \cal{Y}$ represents the domain of the distribution and does not define the distribution itself. Accordingly, the summation on line 91 should be taken over $i \in \tilde{C}$. 5. Line 93: The space before the word "for" is missing. 6. The first part of the statement on lines 93-95, in my opinion, is incorrect. From a statistical perspective, the aim of Federated Learning is to minimize the generalized local loss function $\mathbb{E}_{(x, y) \in P_i} \cal{L}_i (\theta, x, y)$. However, many Federated Learning frameworks minimize the finite sum because they make an implicit assumption that the distributions are similar. Otherwise, one could construct two distributions $P_1$ and $P_2$ such that the global solution for $1/2(P_1 + P_2)$ is ineffective for both clients. 7. The abbreviation SD is used on line 109 but only defined later on line 180. 8. Line 121: "are" -> "is" 10. Line 191: For clarity, I would suggest spelling out the abbreviation MLP. 11. Line 190: Please provide explicit definitions and domains for $\eta_i$ and $\beta_j$ as they are not scalars. 12. Equation 5 is misleading. The loss function on both lines is taken over the data distribution $P_i$, but there are also indices $T$ and $V$, which denote train and test datasets. $theta_i(\lambda)$ is not a minimizer of the generalized loss function but rather a fine-tuning of the global solution $\theta_g$. It takes some time to understand the authors' intended meaning here. 13. Line 226: There is a redundant full stop after the first word "Algorithm." 14. Algorithm 1, the input section: The phrases "Total no. of clients C. Num of update iterations K" use different notation to express "number," leading to inconsistency within the text. 15. Lines 9-10 in Algorithm 1 do not clearly depict the intended FedAvg iterates. 16. The document concludes on page 8 with Section "4.4 Personalizing to Different Domains."

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

Authors adequately address the limitations of the work.

Reviewer P9q95/10 · confidence 3/52023-08-07

Summary

The paper proposes FedL2P, a federated learning framework for learning personalized strategies in the federated meta-learning problem. The authors introduce meta-nets to estimate hyper-parameters for personalized fine-tuning, such as batch normalization parameters and learning rates, based on client-specific data statistics. The empirical results show that FedL2P outperforms standard hand-crafted personalization baselines in both label and feature shift situations.

Strengths

- The proposed method demonstrates soundness in its approach. Utilizing hyperparameter inference for federated learning shows promise and effectiveness in achieving personalization, as it tries to customize only a few crucial hyper-parameters. - The experiments are meticulously evaluated across diverse distribution shifts, encompassing feature shift, label shift, and domain shift.

Weaknesses

- Calculating meta-gradients involves time-consuming computations of second-order derivatives. - The comparison is limited to variants of finetuning, neglecting the majority of personalized federated learning baselines such as [1,2]. Additionally, recent works [3] addressing distribution shifts by optimizing different blocks of the network should be thoroughly discussed in the paper. - The sufficiency and effectiveness of communicating the selected hyper-parameters for personalized federated learning lack rigorous verification. [1] Towards personalized federated learning. TNNLS [2] Layer-wised model aggregation for personalized federated learning. CVPR 2022 [3] Surgical fine-tuning improves adaptation to distribution shifts. ICLR 2023

Questions

- What is the computational cost associated with the proposed method? - How many parameters will be personalized in this study? - Could you provide insights into the learning rate distribution for different blocks/layers in the paper and elaborate on what the model has learned? This would greatly enhance the understanding of the results.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

I'm not aware of any potential negative societal impact of the work.

Authorsrebuttal2023-08-14

Thanks for the feedback. If you are satisfied by our responses, please consider increasing your score. Otherwise, please let us know any remaining concerns you have so we can try to answer them.

Reviewer aauE2023-08-14

Thanks for your reply

The author did a greate job in the rebuttal. Most of my questions are answered. I will increase my score. I like the table in appendix B where the author positioned their paper in the area. Please consider adding the new related works I mentioned in the review. Maybe you can also update the related work section in the main paper. For BNNet notation, I am okay with your notations. Just please make sure they are consistent in your paper.

Reviewer nqFj2023-08-15

Thanks for the reply

I extend my gratitude to the authors for their insightful rebuttal and for providing the accompanying table. Their comprehensive response is duly noted. I remain optimistic that the authors will fulfill their commitment to refining the manuscript in accordance with their assurances.

Reviewer wFN92023-08-16

Thanks for rebuttal

Thanks for the detailed rebuttal. I have updated my score, please do ensure that the final version has the changes.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC