Heterogeneous Personalized Federated Learning by Local-Global Updates Mixing via Convergence Rate

Authors: Meirui Jiang, Anjie Le, Xiaoxiao Li, Qi Dou

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We have theoretically analyzed the convergence of our method in the over-parameterize regime, and experimentally evaluated our method on five datasets. These datasets present heterogeneous data features in natural and medical images. With comprehensive comparison to existing state-of-the-art approaches, our LG-Mix has consistently outperformed them across all datasets (largest accuracy improvement of 5.01%), demonstrating the outstanding efficacy of our method for model personalization.
Researcher Affiliation Academia Meirui Jiang Department of Computer Science and Engineering The Chinese University of Hong Kong mrjiang@cse.cuhk.edu.hk Anjie Le Department of Computer Science and Engineering The Chinese University of Hong Kong ajle@cuhk.edu.hk Xiaoxiao Li Department of Electrical and Computer Engineering The University of British Columbia xiaoxiao.li@ece.ubc.ca Qi Dou Department of Computer Science and Engineering The Chinese University of Hong Kong qidou@cuhk.edu.hk
Pseudocode Yes Algorithm 1: LG-Mix: Local-global updates mixing algorithm
Open Source Code Yes Code is available at https://github.com/med-air/Hetero PFL.
Open Datasets Yes Digits-5 (Zhou et al., 2020; Li et al., 2021c) ... Office-Caltech10 (Gong et al., 2012) ... Domain Net (Peng et al., 2019) ... Camelyon17 (Bandi et al., 2018) ... Retinal dataset which contains retinal fundus images from 6 different sources (Fumero et al., 2011; Sivaswamy et al., 2015; Almazroa et al., 2018; Orlando et al., 2020).
Dataset Splits Yes For all datasets, we take each data source as one client and split the data of each client into train, validation, and testing sets with a ratio of 0.6, 0.2, and 0.2.
Hardware Specification Yes The GPU we used for training is Ge Force RTX 2080 Ti.
Software Dependencies No The paper mentions software like "Py Torch" but does not provide specific version numbers for software dependencies.
Experiment Setup Yes We use the SGD optimizer with a learning rate of 0.01 and Cross Entropy loss for classification tasks and use Adam optimizer with learning rate of 1e 3 with β = (0.9, 0.99), dice loss (Milletari et al., 2016) for the segmentation task. ... The total number of training rounds is 100 with a local update epoch of 1. ... All input images are resized to 28 28.