Personalized Federated Learning with Moreau Envelopes

Authors: Canh T. Dinh, Nguyen Tran, Josh Nguyen

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we empirically evaluate the performance of pFedMe using both real and synthetic datasets that capture the statistical diversity of clients data. We show that pFedMe outperforms the vanilla FedAvg and a meta-learning based personalized FL algorithm Per-FedAvg in terms of convergence rate and local accuracy.
Researcher Affiliation Academia 1The University of Sydney, Australia tdin6081@uni.sydney.edu.au, nguyen.tran@sydney.edu.au 2The University of Melbourne, Australia tuandungn@unimelb.edu.au
Pseudocode Yes Algorithm 1 pFedMe: Personalized Federated Learning using Moreau Envelope Algorithm
Open Source Code Yes The code and datasets are available online1. 1https://github.com/Charlie Dinh/pFedMe
Open Datasets Yes We consider a classification problem using both real (MNIST) and synthetic datasets. MNIST [51] is a handwritten digit dataset containing 10 labels and 70,000 instances.
Dataset Splits Yes All datasets are split randomly with 75% and 25% for training and testing, respectively.
Hardware Specification No The paper mentions that
Software Dependencies Yes All experiments were conducted using PyTorch [52] version 1.4.0.
Experiment Setup Yes We fix the subset of clients S = 5 for MNIST, and S = 10 for Synthetic. We compare the algorithms using both cases of the same and fine-tuned learning rates, batch sizes, and number of local and global iterations. ... We fix |D| = 20, R = 20, K = 5, and T = 800 for MNIST, and T = 600 for Synthetic, β = 2 for pFedMe (ˆα and ˆβ are learning rates of Per-FedAvg).