Ranking-based Client Imitation Selection for Efficient Federated Learning

Authors: Chunlin Tian, Zhan Shi, Xinpeng Qin, Li Li, Cheng-Zhong Xu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results reveal that Fed Rank boosts model accuracy by 5.2% to 56.9%, accelerates the training convergence up to 2.01 and saves the energy consumption up to 40.1%.
Researcher Affiliation Academia 1University of Macau 2University of Texas at Austin 3University of Electronic Science and Technology of China.
Pseudocode Yes Algorithm 1 summarizes the offline pre-train algorithm for Fed Rank policy πθ.
Open Source Code No No explicit statement or link providing access to the open-source code for the described methodology was found.
Open Datasets Yes Specifically, MNIST (Lecun et al., 1998) on Le Net5 (Le Cun et al., 1998) was adopted as the ID dataset. While for OOD datasets, Res Net18 (He et al., 2016) on CIFAR10 (Krizhevsky et al., 2009), VGG16 (Simonyan., 2014) on CINIC10 (Darlow et al., 2018), Shuffle Net (Zhang et al., 2018) on Tiny Image Net (Deng et al., 2009) for image classification were utilized.
Dataset Splits No For four computer vision datasets, we generate IID data splits by randomly assigning training examples to each client without replacement. For Non-IID splits, we simulate data heterogeneity by sampling label ratios from a Dirichlet distribution pk Dir N(σ) with the symmetric parameter σ.
Hardware Specification No Monsoon Power Monitor (Monsoon, 2023) was utilized to monitor energy consumption during the training process.
Software Dependencies No Specifically, we first built a simulator following the server/client architecture based on Py Torch (Paszke et al., 2019)
Experiment Setup Yes Specifically, we set the reward hyperparameters α = β = 2 in Eq. 1. This is for the reason that Fed Rank jointly considers energy consumption and training time in each round in the reward function. We set the device pool N = 100, and K = 10 devices were selected to participate according to various selection policies, with r = 50 training rounds and lep = 5 local training epochs per round.