Eliminating Domain Bias for Federated Learning in Representation Space

Authors: Jianqing Zhang, Yang Hua, Jian Cao, Hao Wang, Tao Song, Zhengui XUE, Ruhui Ma, Haibing Guan

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

Reproducibility Variable Result LLM Response
Research Type Experimental Besides, extensive experiments on four datasets show that DBE can greatly improve existing FL methods in both generalization and personalization abilities.
Researcher Affiliation Academia 1Shanghai Jiao Tong University 2Queen s University Belfast 3Louisiana State University
Pseudocode Yes Algorithm 1 The Learning Process in Fed Avg+DBE
Open Source Code Yes Our code is public at https://github.com/TsingZ0/DBE.
Open Datasets Yes we use four public datasets for classification problems in FL, including three CV datasets: Fashion-MNIST (FMNIST) [77], Cifar100 [42], and Tiny-Image Net (100K images with 200 labels) [19], as well as one NLP dataset: AG News [91].
Dataset Splits No To simulate the common FL scenario where data only exists on clients, we split the data among each client into two parts: a training set (75% data) and a test set (25% data).
Hardware Specification Yes We run all experiments on a machine with two Intel Xeon Gold 6140 CPUs (36 cores), 128G memory, eight NVIDIA 2080 Ti GPUs, and Cent OS 7.8.
Software Dependencies No The paper mentions 'Cent OS 7.8' as the operating system, but does not provide specific version numbers for software dependencies such as deep learning frameworks (e.g., PyTorch, TensorFlow) or other relevant libraries.
Experiment Setup Yes Following Fed Avg, we set the batch size to 10 and the number of local epochs to 1, so the number of local SGD steps is ni /10 for client i. We run three trials for all methods until empirical convergence on each task and report the mean value. For hyperparameter tuning, we use grid search to find optimal hyperparameters, including κ and µ. In this paper, we set κ = 50, µ = 1.0 for the 4-layer CNN, κ = 1, µ = 0.1 for the Res Net-18, and κ = 0.1, µ = 1.0 for the fast Text.