Beyond ADMM: A Unified Client-Variance-Reduced Adaptive Federated Learning Framework

Authors: Shuai Wang, Yanqing Xu, Zhiguo Wang, Tsung-Hui Chang, Tony Q. S. Quek, Defeng Sun

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments based on (semi-)supervised image classification tasks demonstrate superiority of Fed VRA over the existing schemes in learning scenarios with massive heterogeneous clients and client sampling.
Researcher Affiliation Academia 1 Information Systems Technology and Design, Singapore University of Technology and Design, 487372 Singapore 2 School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen 518172, China 3 College of Mathematics, Sichuan University, Chengdu, Sichuan 610064, China 4 Department of Applied Mathematics, The Hong Kong Polytechnic University, Hong Kong
Pseudocode Yes Algorithm 1: Proposed Fed VRA algorithm
Open Source Code No No explicit statement providing access to open-source code (e.g., a specific repository link or a clear statement about code availability in supplementary materials) was found.
Open Datasets Yes The CIFAR-10 (Krizhevsky and Hinton 2009) and MNIST (Le Cun, Cortes, and Burges 2010) datasets are considered for evaluation.
Dataset Splits No The paper describes the distribution of training samples to clients (N=100) and the method for Non-IID data partitioning, and states that testing samples are used for test accuracy. However, it does not explicitly mention a validation split.
Hardware Specification No The paper does not provide specific hardware details such as CPU/GPU models, processor types, or memory used for running the experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions) that are needed to replicate the experiment.
Experiment Setup Yes The learning rate η is set to be 0.01 for all algorithms. The mini-batch size S is set to be 50. In each round, we uniformly sample 10 clients (|Ar| = 10), and randomly choose the number of local epochs from [1, 5] if HLU is considered, and otherwise set it to 2 for each client. Other algorithm specific parameters are tuned individually. Fed Prox uses µ = 0.1 and SCAFFOLD takes ηg = 1. Both the parameter α of Fed Dyn and γi of Fed VRA are 0.1 by default. All algorithms stop when 500 rounds are achieved.