Hierarchical Federated Learning with Multi-Timescale Gradient Correction

Authors: Wenzhi Fang, Dong-Jun Han, Evan Chen, Shiqiang Wang, Christopher Brinton

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

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experiments on various datasets and models, we validate the effectiveness of MTGC in diverse HFL settings.
Researcher Affiliation Collaboration Wenzhi Fang Purdue University fang375@purdue.edu Dong-Jun Han Yonsei University djh@yonsei.ac.kr Evan Chen Purdue University chen4388@purdue.edu Shiqiang Wang IBM Research wangshiq@us.ibm.com Christopher G. Brinton Purdue University cgb@purdue.edu
Pseudocode Yes Algorithm 1: HFL with Multi-Timescale Gradient Correction (MTGC)
Open Source Code Yes The code for this project is available at https://github.com/wenzhifang/MTGC.
Open Datasets Yes In our experiments, we consider four widely used datasets: EMNIST-Letters (EMNIST-L) [7], Fashion-MNIST [53], CIFAR-10 [23], and CIFAR-100 [23].
Dataset Splits Yes The CINIC-10 dataset contains 90,000 training images, 90,000 validation images, and 90,000 test images, significantly larger than CIFAR-10 and CIFAR-100 with 60,000 images.
Hardware Specification Yes We conduct the experiments based on a cluster of 3 NVIDIA A100 GPUs with 40 GB memory.
Software Dependencies No The paper mentions "Our code is based on the framework of [1]" but does not specify particular software dependencies with version numbers (e.g., Python version, specific library versions like PyTorch, TensorFlow, etc.).
Experiment Setup Yes Across all algorithms considered, we maintain a consistent learning rate η = 0.1 and batch size 50.