On the Convergence of FedAvg on Non-IID Data

Authors: Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, Zhihua Zhang

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically verify our results through numerical experiments.
Researcher Affiliation Academia Xiang Li School of Mathematical Sciences Peking University Beijing, 100871, China smslixiang@pku.edu.cn Kaixuan Huang School of Mathematical Sciences Peking University Beijing, 100871, China hackyhuang@pku.edu.cn Wenhao Yang Center for Data Science Peking University Beijing, 100871, China yangwenhaosms@pku.edu.cn Shusen Wang Department of Computer Science Stevens Institute of Technology Hoboken, NJ 07030, USA shusen.wang@stevens.edu Zhihua Zhang School of Mathematical Sciences Peking University Beijing, 100871, China zhzhang@math.pku.edu.cn
Pseudocode No The paper describes the algorithm steps in a descriptive paragraph under 'Algorithm description' in Section 2, but does not include a formal pseudocode block or algorithm listing.
Open Source Code No The paper does not provide an explicit statement about open-sourcing its code or a link to a code repository.
Open Datasets Yes We distribute MNIST dataset (Le Cun et al., 1998) among N = 100 workers in a non-iid fashion such that each device contains samples of only two digits.
Dataset Splits No The paper describes how data is distributed across devices and how the model is evaluated during training, but it does not specify explicit training/validation/test dataset splits for reproducibility.
Hardware Specification No The paper does not provide any specific details regarding the hardware used for the experiments.
Software Dependencies No The paper does not specify any software dependencies with version numbers.
Experiment Setup Yes The regularization parameter is set to λ = 10 4. In each round, all selected devices run E steps of SGD in parallel. We decay the learning rate at the end of each round by the following scheme ηt = η0 1+t, where η0 is chosen from the set {1, 0.1, 0.01}. For unbalanced MNIST, we use batch size b = 64. The hyperparameters are the same for all schemes: E = 20, K = 10 and b = 64.