Solving a Class of Non-Convex Minimax Optimization in Federated Learning

Authors: Xidong Wu, Jianhui Sun, Zhengmian Hu, Aidong Zhang, Heng Huang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results on fair classification and AUROC maximization show the efficiency of our algorithms.
Researcher Affiliation Academia Xidong Wu Electrical and Computer Engineering University of Pittsburgh Pittsburgh, PA 15213 xidong_wu@outlook.com; Jianhui Sun Computer Science University of Virginia Charlottesville, VA 22903 js9gu@virginia.edu; Zhengmian Hu Computer Science University of Maryland College Park, MD 20742 huzhengmian@gmail.com; Aidong Zhang Computer Science University of Virginia Charlottesville, VA 22903 aidong@virginia.edu; Heng Huang Computer Science University of Maryland College Park, MD 20742 henghuanghh@gmail.com
Pseudocode Yes Algorithm 1 Fed SGDA+ Algorithm; Algorithm 2 Fed SGDA-M Algorithm
Open Source Code Yes The code is available https://github.com/xidongwu/Federated-Minimax-and-Conditional-Stochastic-Optimization
Open Datasets Yes Fashion MNIST dataset has 70, 000 28 28 gray images (10 categories, 60, 000 training images and 10, 000 testing images). CIFAR-10 dataset consists of 50, 000 training images and 10, 000 testing images. Tiny-Image Net dataset has 200 classes of (64 64) colored images and each class has 500 training images, 50 validation images, and 50 test images.
Dataset Splits Yes Tiny-Image Net dataset has 200 classes of (64 64) colored images and each class has 500 training images, 50 validation images, and 50 test images.
Hardware Specification Yes Experiments are completed on the computer cluster with AMD EPYC 7513 Processors and NVIDIA RTX A6000.
Software Dependencies No The paper mentions using 'Py Torch' but does not specify version numbers for PyTorch or any other software dependencies, which would be necessary for reproducibility.
Experiment Setup Yes The network has 20 clients. In experiments, we run grid search for step size, and choose the step size for primal variable in the set {0.01, 0.03, 0.05, 0.1, 0.3} and that for dual variable in the set {0.001, 0.01, 0.1}. We choose the global step size in the set {0.1, 0.5, 1, 1.5, 2}. The batch-size b is in 50 and the inner loop number Q is seleted from {20, 50, 100}, The outer loop number S is selected from {1, 5, 10} for Fed SGDA+ and {1, 5, 10, Q} for local SGDA+.