Personalized Federated Learning via Variational Bayesian Inference

Authors: Xu Zhang, Yinchuan Li, Wenpeng Li, Kaiyang Guo, Yunfeng Shao

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that the proposed method outperforms other advanced personalized methods on personalized models, e.g., p Fed Bayes respectively outperforms other SOTA algorithms by 1.25%, 0.42% and 11.71% on MNIST, FMNIST and CIFAR-10 under non-i.i.d. limited data.
Researcher Affiliation Collaboration 1LSEC, Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing, China 2Noah s Ark Lab, Huawei, Beijing, China.
Pseudocode Yes Algorithm 1 p Fed Bayes: Personalized Federated Learning via Bayesian Inference Algorithm
Open Source Code No The paper does not provide any specific links or explicit statements about the release of source code.
Open Datasets Yes We generate the non-i.i.d. datasets based on three public benchmark datasets, MNIST (Le Cun et al., 2010; 1998), FMNIST (Fashion MNIST)(Xiao et al., 2017) and CIFAR-10 (Krizhevsky, 2009).
Dataset Splits Yes For small, medium and large datasets of MNIST/FMNIST, there were 50, 200, 900 training samples and 950, 800, 300 test samples for each class, respectively. For the small, medium and large datasets of CIFAR-10, there were 25, 100, 450 training samples and 475, 400, 150 test samples in each class, respectively.
Hardware Specification Yes We did all experiments in this paper using servers with two GPUs (NVIDIA Tesla P100 with 16GB memory), two CPUs (each with 22 cores, Intel(R) Xeon(R) Gold 6152 CPU @ 2.10GHz), and 192 GB memory.
Software Dependencies No The paper mentions 'We use Py Torch (Paszke et al., 2019) for all experiments.', but it does not specify a version number for PyTorch or any other software dependency.
Experiment Setup Yes Based on the experimental results, we set the learning rate of Fed Avg and Per-Fed Avg to 0.01. The learning rate and regularization weight of Fedprox are respectively set as 0.01 and λ = 0.001. ... For the proposed p Fed Bayes, we set the initialization of weight parameters ρ = 2.5, the tradeoff parameter ζ = 10, and the learning rates of the personalized model and global model η1 = η2 = 0.001.