Understanding How Consistency Works in Federated Learning via Stage-wise Relaxed Initialization

Federated learning (FL) is a distributed paradigm that coordinates massive local clients to collaboratively train a global model via stage-wise local training processes on the heterogeneous dataset. Previous works have implicitly studied that FL suffers from the ``client-drift'' problem, which is caused by the inconsistent optimum across local clients. However, till now it still lacks solid theoretical analysis to explain the impact of this local inconsistency. To alleviate the negative impact of the ``client drift'' and explore its substance in FL, in this paper, we first design an efficient FL algorithm \textit{FedInit}, which allows employing the personalized relaxed initialization state at the beginning of each local training stage. Specifically, \textit{FedInit} initializes the local state by moving away from the current global state towards the reverse direction of the latest local state. This relaxed initialization helps to revise the local divergence and enhance the local consistency level. Moreover, to further understand how inconsistency disrupts performance in FL, we introduce the excess risk analysis and study the divergence term to investigate the test error of the proposed \textit{FedInit} method. Our studies show that optimization error is not sensitive to this local inconsistency, while it mainly affects the generalization error bound in \textit{FedInit}. Extensive experiments are conducted to validate this conclusion. Our proposed \textit{FedInit} could achieve state-of-the-art~(SOTA) results compared to several advanced benchmarks without any additional costs. Meanwhile, stage-wise relaxed initialization could also be incorporated into the current advanced algorithms to achieve higher performance in the FL paradigm.

Paper

Similar papers

Peer review

Reviewer g4w55/10 · confidence 3/52023-06-29

Summary

The manuscript proposes a novel federated learning method to alleviate the negative impact of the "client drift" problem and enhance consistency in the FL paradigm. The manuscript also analyzes the intrinsic impact of local consistency on optimization error, test error, and generalization error. Several experiments are conducted to validate the efficiency of the proposed method.

Strengths

1. The proposed method outperforms baseline methods without increasing communication costs and can be easily incorporated into other methods. 2. The manuscript provides ample theoretical analysis to demonstrate the theoretical bounds of the proposed method.

Weaknesses

1. The hyperparameters were not tuned for different methods. Since the experiments are highly sensitive to hyperparameter settings, the superiority of the proposed method is likely due to the selection of hyperparameters. Given that the personalized process proposed in the manuscript bears resemblance to the momentum of parameter updates, I suspect that when tuning the learning rate for the baseline method, the baseline method may outperform the proposed method. 2. The results shown in Figure 1 suggest that the proposed method is highly sensitive to the choice of beta.

Questions

1. The manuscript thinks that inconsistent local optima cause "client drift," which results in the inferiority of federated learning (FL) algorithms. The proposed method aims to alleviate this divergence by employing personalized relaxed initialization at the beginning of each round. I have a question: based on my understanding, personalized initialization in each round would lead to greater differences across different clients. So, how does this strategy help alleviate the divergence? 2. After a brief review of the theoretical analysis, I am uncertain if I correctly understood the proof. In formulas (9) and (14), the upper bounds achieve their minimum when beta=0. If that's the case, why does the personalized relaxation contribute to improving the performance of FL algorithms?

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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

The experiments in the manuscript are not rigorous enough as the hyperparameters for different methods were not tuned. This could potentially lead to an overestimation of the performance of the proposed method.

Reviewer sALG6/10 · confidence 3/52023-06-30

Summary

This paper proposes to initialize the local state by moving away from the current global state toward the reverse direction of the latest local state. They demonstrate theoretically and empirically that this revision can help consistency for better performance. The method is also a practical plug-in that could easily to incorporated into other methods.

Strengths

- This paper goes deeper into how consistency work in FL systems, and provides a simple and effective initialization-based solution. - This paper gives a comprehensive theoretical analysis of the problem, which I think is a good contribution. - The paper is well-written and easy to follow. - Experiment shows good performance.

Weaknesses

- I understand there might not be enough space for experiments and the focus looks to be theoretical analysis, but it would be better if the authors can give more experimental results for different FL settings, e.g., more clients/datasets, which may help evaluate the method better. For example, 100 clients/10% participation rate and 200 clients/ 5% participation rate can be relatively limited. Maybe can we have a curve for participation rate with the same amount of clients and perhaps a curve for number of clients to see how these FL settings influence the performance?

Questions

- Does the method also work well for a large number of clients, e.g., 1000? - How does participation rate influence performance?

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors have stated the limitation and future work as pFL.

Reviewer 9rh86/10 · confidence 4/52023-07-05

Summary

This paper aims to solve the “client-drift” problem in Federated Learning, which is caused by the NonIID data. Specifically, this paper proposes initializing the local model of each client with its personalized model to alleviate the problem. Further, the paper theoretically analyzes the impact of inconsistency on the convergence of FL. Besides, extensive experiments also demonstrate the effectiveness of the proposed method.

Strengths

1. The idea of initializing the local model for solving the NonIID problem is interesting. 2. The theoretical analysis for the proposed method is solid. 3. The experimental results demonstrate the effectiveness of the proposed method.

Weaknesses

1. Although the theoretical results are sufficient to verify the convergence of the proposed method, they cannot present the effectiveness of the initialization strategy in principle. More specifically, how does the proposed method reduce the divergence term compared to the vanilla FedAvg? 2. The workflow of the proposed method seems not correct. Are the locations of line 10 and line 12 placed correctly? 3. How to obtain the value of $w_{i,k}^{t-1}$ is not clear. The client may not participate in the previous round $t-1$ under the setting of random client selection. If the local model $w_{i,k}^{t’-1}$ is obtained in many previous rounds, the motivation of bias correction using $ w_{i,k}^{t’-1} - w^{t-1}$ is not reasonable due to the significant gap between $t’$ and $t$. 4. The key hyperparameter K is set inappropriately. In typical FL literature, the number of local epochs is usually set to range from $5$, $10$, and up to $20$, which may contain many local iterations (mini-batches). However, the experiment of this paper only adopts a small number of local iterations instead of local epochs, e.g., 5 local iterations in Table 1, which seems manually adjusted. 5. It would be better if there is a figure to illustrate the principles of the proposed method.

Questions

See above.

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

3 good

Limitations

Yes.

Reviewer qyTU7/10 · confidence 4/52023-07-10

Summary

This paper proposes an efficient stage-wise initialization for the federated learning paradigm, named FedInit, which could be extended as a plug-in to several existing methods. It provides the theoretical analysis on both the convergence and generalization to illustrate how consistency term affects the FL. Experiments also show its efficiency in practical scenarios.

Strengths

1. FedInit is a light and nimble technique that does not introduce extra communication costs in the FL framework. Compared with other algorithms, FedInit can almost achieve the same performance as SOTA algorithms while maintaining the same communication cost as the vanilla FedAvg. And, Relaxed initialization (RI) could be easily extended into other advanced algorithms as a plug-in and efficiently improve their performance. 2. This paper explores the impacts of global consistency constraints in FL, which is an essential and interesting problem in federated learning. It also indicates the relationships between generalization and optimization. 3. This paper provides a theoretical analysis of the excess risk to comprehensively explain how the dominant factors affect the final performance (test accuracy) throughout the training process. Ablation studies are conducted to validate the conclusion of the theoretical analysis on the hyperparameters, i.e. local interval $K$, coefficient of RI $\beta$.

Weaknesses

1. The relaxed initialization (RI) looks like an extrapolation algorithm between the global model and the local models as $w+\beta(w-w_i)$. Several methods including the benchmarks of this paper, i.e. FedProx, and FedCM, are accustomed to applying interpolation methods to adjust local training to make the entire heterogeneous training smoother. So what is their essential difference? I think the author should add a paragraph to discuss it. 2. This paper involves a lot of analytical proofs, so I suggest that the author write a simple proof sketch for each sub-sections to facilitate reading. 3. In Table.2, RI helps to largely improve the performance of FedSAM and SCAFFOLD but shows limited help for FedCM and FedDyn. What is the main reason for this phenomenon? Does it imply that RI is subject to some existing algorithms? I think the author should add a paragraph to discuss it. 4. Some typos (for example): (1) Line.150 "... ... its generality" (2) Line.79 "client drift" and Line.151 "client-drift" (3) Line.192 the sentence is difficult to read

Questions

Please see the weaknesses.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed 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

3 good

Contribution

3 good

Limitations

Please see the weaknesses.

Authorsrebuttal2023-08-11

Additional Experiments on Tuning Learning Rate

Dear reviewer g4w5, to address your concerns on the finetuning learning rates, we re-tuned the learning rate to show that our hyperparameters selection is correct. Due to the rebuttal time limitation, we test the experiments on the following setup. CIFAR-10 Dirichlet-0.1 split, 10 active clients of total 100 clients, other hyperparameters are selected as we introduced in our paper (Appendix) Generally, the global learning rate is set as 1.0 in FL except for some specific methods. Global lr=1.0 means to average the local models. We follow this to fix the global learning rate and finetune the local learning rate as follows. | Method | lr=0.05 | lr=0.1 | lr=0.2 | lr=0.5 | | :----: | :----: | :----: | :----: | :----: | | SCAFFOLD | 73.65% | **75.19%** | 74.93% | 53.16% | | FedCM | 69.56% | **74.14%** | 73.73% | 60.32% | | FedDyn | 71.74% | 75.10% | **75.27%** | - | "-": can not converge As shown in our paper (Appendix), the best selection of the local learning rate is 0.1, which could fairly compare the performance of the whole baselines. We finetune the hyperparameters in a wide range and report the fair and best selection. We also learn from the classical paper [1,2,3] to finish the experiments. Moreover, relaxed initialization could further improve the performance of some current advanced methods. Both theoretical analysis and experiments verify its efficiency. **Relaxed initialization is not a trick that relies on finetuning the learning rate.** [1] Federated Learning Based on Dynamic Regularization (ICLR 2021) [2] Generalized Federated Learning via Sharpness Aware Minimization (ICML 2022) [3] FedCM: Federated Learning with Client-level Momentum

Reviewer g4w52023-08-11

About hyper-parameter

Thanks for the response. Could you please add some experiments with both different local intervals and learning rates for baseline methods? Thanks.

Authorsrebuttal2023-08-11

Re: About hyper-parameter

Thank you very much for your quick reply. We currently use 5 local epochs in our experimental setups. We will test the experiments on 10 local epochs. The search range of the local learning rate is between 0.01 and 0.5 and the global learning rate will still be fixed as 1.0. We will report the results when they are finished ASAP. Thanks!

Authorsrebuttal2023-08-14

Additional Experiments on Tuning Learning Rate on Different Local Epochs

Dear reviewer g4w5, we extend the local epochs to 10 and tune the learning rate for the baselines. Results are shown in the following table. CIFAR-10 Dirichlet-0.1 split, 10 active clients of total 100 clients, local epochs=10, other hyperparameters are selected as we introduced in our paper (Appendix) | Method | lr=0.01 | lr=0.05 | lr=0.1 | lr=0.2 | lr=0.5 | | :----: | :----: | :----: | :----: | :----: | :----: | | SCAFFOLD | 67.47% | 74.34% | **75.74%** | 75.01% | 49.03% | | FedCM | 74.06% | **74.94%** | 74.23% | 68.60% | 59.26% | | FedDyn | 69.37% | 73.05% | **75.29%** | 74.10% | - | | FedInit | 62.82% | 73.50% | **75.48%** | 73.82% | 63.43% | "-": can not converge When local epochs are extended to 10, the results are similar to the selection of 5. As claimed in [1], local clients could be optimized well enough for FL by adopting 5 local epochs. Similar experiments also could be seen in [2] (figure 2 in their paper). FedInit still could achieve comparable performance with the current advanced methods. **And different from the baselines, FedInit only communicates one vector and does not require any additional communication costs, which makes it very useful in practical scenarios.** Moreover, relaxed initialization still could help the advanced methods to improve their performance. As shown in the following Table, hyperparameters are selected as the best group above, | Method | vanilla | + FedInit | | :----: | :----: | :----: | | SCAFFOLD | 75.74% | **76.34%** | | FedCM | 74.94% | **75.10%** | | FedDyn | 75.29% | **75.42%** | The coefficient of relaxed initialization (RI) for SCAFFOLD is selected as 0.1, and 0.02 for the others. As we claimed before, the theoretical analysis indicates that RI benefits from a positive coefficient instead of 0 in practical training, which could minimize the excess risk in the FL paradigm. Experimental results also validate its efficiency. [1] Federated Learning Based on Dynamic Regularization (ICLR 2021) [2] Generalized Federated Learning via Sharpness Aware Minimization (ICML 2022)

Reviewer g4w52023-08-14

Thanks for the response. The updated results correspond with those presented in Table 1 and Table 2 of the original manuscript, correct? In Table 1 of the original manuscript, for the cifar-10 dataset, under the conditions of 10%-100 clients and Dir-0.1, the optimal performance achieved by SCAFFOLD is reported as 75.06±0.16, while the optimal performance achieved by FedInit is reported as 75.95±0.19. However, in the revised table provided by the author, following hyper-parameter fine-tuning (local epochs=10 and learning rate=0.1), the best performance of SCAFFOLD is reported as 75.74%, which is quite comparable to FedInit's best performance presented in Table 1. As such, I am inclined to disagree with the author's assertion that "the results are similar to the selection of 5". It is plausible that the performance of baseline methods could be further enhanced by extending the number of local epochs (perhaps the author could perform experiments with SCAFFOLD alone for efficiency). If this hypothesis holds true, particularly for the outcomes showcased in Table 1, the claimed superiority of FedInit may not be as compelling. Actually, the proposed method is likely to enhance model performance by leveraging an increased volume of updates in each round. Consequently, when the baseline methods fall short of full convergence within each round, it is reasonable to expect that the proposed approach will outperform them.

Authorsrebuttal2023-08-14

Additional Experiments on Local Epochs=15 and Rebuttal

Dear reviewer g4w5, thanks for your quick reply. We finish the experiments on the local epochs=15 as shown in the following. ### About the question of "It is plausible that the performance of baseline methods could be further enhanced by extending the number of local epochs (perhaps the author could perform experiments with SCAFFOLD alone for efficiency). If this hypothesis holds true, particularly for the outcomes showcased in Table 1, the claimed superiority of FedInit may not be as compelling." **This hypothesis is not correct.** If we extend the local epochs to a very large number, the local client will be optimized well and almost achieve the local minimal. This is fatal for federated learning. All local models will overfit the local dataset and are far away from the global optimal. This is the "client-drift" problem proposed in [1], which could be formulated as $\frac{1}{m}\sum_i x_i^\star \neq x^\star$. This is also one of the major issues in FL. Therefore, the local epochs can not be extended too much. As shown in the following table, under large local epochs, SCAFFOLD will also be affected. | Method | E=1 | E=5 | E=10 | E=15 | | :----: | :----:| :----:| :----: | :----: | | SCAFFOLD | 71.51% | 75.06%| **75.74%** | 75.25%| Actually, many previous works [1,2,3] have been devoted to preventing local clients from fully converging. [1] uses the variance reduction to correct the local gradient as the global gradient estimation to avoid overfitting. [2] changes the local objective from $F_i(w_i)$ to $F_i(w_i)+\lambda\Vert w_i-w\Vert^2$ to force the local model to be close to the last global state. [3] uses the ADMM method to change the local objective. However, even with these advanced methods, the local interval still cannot be set too large. Training with large intervals will eventually degrade performance and fall into the average of the local minimal $\frac{1}{m}\sum_i x_i^\star$ instead of the global minimal $x^\star$. Our proposed FedInit also could work as a light plug-in and help them to further improve the performance without conflicts. [1] SCAFFOLD: Stochastic Controlled Averaging for Federated Learning [2] Federated Optimization in Heterogeneous Networks [3] Federated Learning Based on Dynamic Regularization ### About the fair comparison: As we mentioned above, due to the "client-drift" problem in FL, increasing the local interval does not always improve the results of the algorithm. So about your question about SCAFFOLD outperforming FedInit on local epochs=10, that is not fair. Here we make a comprehensive table comparing the performance of SCAFFOLD and FedInit and their training costs. Under the setup in our paper, 1 epochs=10 iterations. After 500 communication rounds, the total iterations is 5000 if local epoch=1. | | E=1 | E=5 | E=10 | E=15 | | :----: | :----:| :----:| :----: | :----: | | SCAFFOLD | 71.51% | 75.06%| **75.74%** | 75.25% | | FedInit | 73.11% | **75.95%** | 75.48% | 73.92% | | Total Iterations | 5000 | 25000| 50000 | 75000 | As stated above, **to achieve the best results, FedInit requires fewer iterations and less training time**. Furthermore, SCAFFOLD requires to communicate 2 vectors per communication round and FedInit only requires to communicate 1 vector. Another main issue in FL is the communication bottleneck in the practical training. Therefore, the proposed FedInit is better for its fewer communication costs and running wall-clock time costs. SCAFFOLD applies a variance reduction technique, which does not conflict with our relaxed initialization. What we would like to claim is that SCAFFOLD adopts the variance reduction and double communication costs to improve performance, while FedInit can achieve comparable performance with fewer training costs.

Reviewer g4w52023-08-15

Thanks for all the explanation. I have updated my score to 5.

Authorsrebuttal2023-08-15

Thank you very much for this discussion

Dear reviewer g4w5, thank you very much for your affirmation of our work. The discussion also helps us to further improve this work. We will fix all the typos and misunderstandings above in the next version. Experiments will be added in the appendix due to the page limitation. Thank you again for this detailed discussion!

Reviewer 9rh82023-08-15

Responses to answers

I have read all the reviews from other reviewers and responses from the author. I would like to improve my score.

Authorsrebuttal2023-08-15

Thank you very much for the reviews

Dear reviewer 9rh8, we appreciate you for finding the mistakes in the algorithm box and suggesting drawing the schematic to illustrate the principles, which helps us further complete this work. We will fix the mistakes and misunderstandings mentioned above in the next version. Thank you again for the review!

Reviewer sALG2023-08-18

Thanks for the response

I have read all the reviews from other reviewers and responses from the author. I would like to keep the score.

Authorsrebuttal2023-08-18

Thank you very much for the reviews

Dear reviewer sALG, thank you very much for your affirmation of our work. We will add the experiments of total 1000 clients into the Appendix in the next version. Thank you again for this review!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC