Global Update Tracking: A Decentralized Learning Algorithm for Heterogeneous Data

Decentralized learning enables the training of deep learning models over large distributed datasets generated at different locations, without the need for a central server. However, in practical scenarios, the data distribution across these devices can be significantly different, leading to a degradation in model performance. In this paper, we focus on designing a decentralized learning algorithm that is less susceptible to variations in data distribution across devices. We propose Global Update Tracking (GUT), a novel tracking-based method that aims to mitigate the impact of heterogeneous data in decentralized learning without introducing any communication overhead. We demonstrate the effectiveness of the proposed technique through an exhaustive set of experiments on various Computer Vision datasets (CIFAR-10, CIFAR-100, Fashion MNIST, and ImageNette), model architectures, and network topologies. Our experiments show that the proposed method achieves state-of-the-art performance for decentralized learning on heterogeneous data via a $1-6\%$ improvement in test accuracy compared to other existing techniques.

Paper

References (29)

Scroll for more · 17 remaining

Similar papers

Peer review

Reviewer b7eV5/10 · confidence 3/52023-07-04

Summary

This work proposes a novel decentralized learning algorithm based on gradient tracking mechanism, called Global Update Tracking (GUT), which aims to mitigate the impact of heterogeneous data distribution. The proposed GUT algorithm overcomes the bottleneck of communication overhead by allowing agents to store a copy of their neighbors’ model parameters and then tracking the model updates instead of the gradients. Numerous experiments have proven that the proposed Global Update Tracking with Quasi-Global momentum (QG-GUTm) outperforms the current state-of-the-art decentralized learning algorithm on a spectrum of heterogeneous data.

Strengths

1. This paper proposes a powerful global update tracking method and achieves good performance on multiple benchmarks. 2. The authors provide a detailed sensitivity study of the method's performance with respect to model architecture and hyper-parameters. 3. The paper is well written and easy for the reader to read and understand.

Weaknesses

1. In the experimental part, the quantitative experiments about communication parameters(cost) are not clear enough. Can the authors provide specific quantitative results for comparison so that readers can easily understand them?

Questions

Please refer to the weaknesses section.

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

2 fair

Limitations

This paper does not have any specific societal impact, other than the ones associated with other algorithms.

Reviewer tQhr5/10 · confidence 1/52023-07-05

Summary

The performance of decentralized learning is limited to the different distribution over devices. To address this issue, this paper proposes a method that is less susceptible to variations in data distribution, named GUT. The proposed GUT tracks the global/average model updates. Experiments show GUT achieves good performance on some benchmarks with various models. The paper also analyzes the convergence of the proposed method.

Strengths

1) The issue of heterogeneous data on devices in decentralized learning is critical, the paper shows a good motivation to address this. 2) On the given benchmarks and models, the proposed method seems to perform well. 3) The analysis of convergence of the method is detailed.

Weaknesses

1) There are too many descriptions of related work in the introduction, making it difficult for readers to find the gist. 2) The analysis of method strengths is not clear enough. Why the proposed method is better than previous methods? Why tracking the model updates instead of the gradients can save twice the communication overhead? Can differences in the data distribution of all agents be mitigated by communicating with neighboring agents? 3) References to employed benchmarks (CIFAR-10, CIFAR-100, Fashion MNIST, and Imagenette) and models (VGG-11, ResNet-20, LeNet-5 and, MobileNet-V2) are missed. 4) The evaluated benchmarks are relatively small. More large-scale benchmarks should be supplemented, such as ImageNet. 5) Some larger and state-of-the-art models should be evaluated to verify the effectiveness of the proposed method, such as ResNet-50, ResNet-100, ViT. 6) The evaluations on Fashion MNIST, CIFAR-100, Imagenette employ different model. The models should be unified. 7) Since one of the main contributions of the method is to reduce overhead, the article lacks a quantitative analysis of overhead.

Questions

In rebuttal, the reviewer hope to see the response to 2),4)-7) in weaknesses.

Rating

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

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

The paper provides a compelling discussion about the limitation.

Reviewer NsAw6/10 · confidence 3/52023-07-06

Summary

The paper proposes a new decentralized learning algorithm called Global Update Tracking (GUT) that can mitigate the impact of heterogeneous data in decentralized learning without introducing any communication overhead. The proposal was demonstrated 4 computer vision datasets.

Strengths

The paper proposes an important and useful application for training deep learning models in decentralized way. There are many strengths of the proposal including the following: 1. GUT is less susceptible to variations in data distribution across devices as shown by different experiments and ablation study 2. The performance is better than DSGD and DSGDm. However, it is shown only for CIFAR-10.

Weaknesses

There are several questions: 1. The paper claims that GUT does not introduce any communication overhead but no results are shown to justify that. However using Figure 2, authors have claimed that their method converges faster than quasi global gossip. The comparison/improvement is not clear from the figure. 2. The author claim that agents in GUT store copy of their neighbors. How often this copy is updated? What is the communication overhead in that? 3. Performance comparison with related methods is not shown on all 4 datasets. 4. The data partition is fixed, non-overlapping and never shuffled across agents during training. How does this partition is defined?

Questions

Please see weaknesses.

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

2 fair

Contribution

3 good

Limitations

There are several limitations that the authors should address: 1. The data partition across agents is not clearly explained. Necessary details for all datasets is required. For example, number of samples per class per agent. Implementation details are also missing from the main paper. 2. Detailed comparison with the related methods is missing. For example, Table 3 compare performance on 3 datasets but no related method is shown.

Reviewer LUfd6/10 · confidence 1/52023-07-07

Summary

The authors propose an algorithm for decentralized learning, where training datasets with heterogenous distributions are collected across different devices. They propose a global update tracking (GUT)-based approach where the IID assumption for the data is removed, and aim to generalize their method to non-IID data by tracking the consensus model while reducing the communication overhead by their GUT algorithm, at the expense of additional memory for storing the parameters for the neighboring model. They validate their method by experiments on various datasets.

Strengths

- Proposed method is applicable to more realistic settings for decentralized learning, where non-IID heterogeneous distribution is more common for the real world data. - Their GUT-based approach adds minimal computational overhead on the overall agents, and they show the performance of their algorithm on multiple datasets in the experimental results. - They provide ablation studies for further analysis on the hyperparameter sensitivity, different levels of heterogeneity, and various implementations of momentum. - Authors provide the source code for their implementation in the supplementary materials.

Weaknesses

- Since one of the main claims of this approach is its generalizability to non-IID heterogeneous data distributions for each agent, what aspect in the proposed formulation corrects the gap between the individual distributions? - Compared to previous decentralized learning approaches where the agent tracks gradients, momentum of gradients, or gloabl gradient obtained from the adjacent agents, proposed GUT tracks model updates. What statistical characteristic/effect makes the model update ($x^{t+1}_i - x^{t}_i$) beneficial for tracking? A simple toy (e.g. Gaussian) example experiment with update with/without model update tracking would suffice. - Are there any visualizations for non-IID characteristic on the datasets under varying degree of $\alpha$? Also, are there experimetal results of additional memory consumption and computational overhead due to the additional update terms?

Questions

Please refer to the questions in the weaknesses section. Since I am not an expert in this field, my positive and negative opinions are mixed. Additionally, repeating the full paper in the supplementary materials seems unnecessary.

Rating

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

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

The authors included a separate discussions and limitations section in their paper, with adequate explanations and descriptions on the weaknesses and possible future directions.

Reviewer LUfd2023-08-16

After reading the rebuttal

Thanks for the feedback for my comments. After reading the response from the authors, I think my concerns were adequately addressed and thus I am inclined to keep my original rating of weak accept.

Reviewer tQhr2023-08-16

Re: Official Review of Submission597 by Reviewer tQhr and Rebuttal

1. Thanks for the authors' rebuttal, I have updated my rating due to their rebuttal and other reviewers' comments; 2. I must say that this submission is out of my scope and encourage the AC consider this issue when they make decision.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC