Understanding the Stability-based Generalization of Personalized Federated Learning

Authors: Yingqi Liu, Qinglun Li, Jie Tan, Yifan Shi, Li Shen, Xiaochun Cao

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Promising experiments on CIFAR datasets also corroborate our theoretical insights. Our code can be seen in https://github.com/Yingqi Liu1999/Understanding-the Stability-based-Generalization-of-Personalized-Federated-Learning. [...] Massive experiments verify theoretical findings. Our experiments on CIFAR datasets with different models under non-convex conditions strongly support our theoretical insights.
Researcher Affiliation Academia Yingqi Liu1,2 Qinglun Li3 Jie Tan4 Yifan Shi Li Shen1,2 Xiaochun Cao1 1School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, China 2Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), China 3College of Systems Engineering, National University of Defense Technology, China 4Intelligent Game and Decision Lab, China EMAIL; EMAIL; EMAIL EMAIL; EMAIL
Pseudocode Yes Algorithm 1: Local updating for PFL. Input :Local steps K, local learning rate ηu and ηv, initialize ut i,0 = ut, and vt i,0 = vt i. Output :For each client, locally update ut+1 i , vt+1 i . Algorithm 2: C-PFL and D-PFL. Input :Total communication rounds T, number of selected clients n, initial the shared and personal variables u0, v0 = {v0 i }n i=0. Output :Personal solution u T and v T = {v T i }n i=0.
Open Source Code Yes Our code can be seen in https://github.com/Yingqi Liu1999/Understanding-the Stability-based-Generalization-of-Personalized-Federated-Learning.
Open Datasets Yes We conduct the experiments on CIFAR-10 datasets (Krizhevsky et al., 2009) in the Dirichlet distribution (Non-IID α = 0.3) with Res Net-18 (He et al., 2016) and CIFAR-100 datasets in the Pathological distribution (Non-IID c = 20) with VGG-11 (Simonyan & Zisserman, 2014) for C-PFL and D-PFL.
Dataset Splits No The paper mentions 'training and testing error' but does not explicitly state the specific split percentages or sample counts for these splits. It describes a mechanism for generating neighboring datasets for stability analysis, but this is not standard train/test/validation split information.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. It only mentions models (ResNet-18, VGG-11) and datasets (CIFAR-10, CIFAR-100).
Software Dependencies No The paper mentions using SGD as the base local optimizer but does not specify any software libraries, frameworks, or their version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes We keep the same experiment setting for all methods and perform 300 communication rounds. The number of client sizes is 20. The client sampling radio is 0.2 in C-PFL, while each client communicates with 4 neighbors in D-PFL accordingly. The batch size is 128 and the number of local epochs is 5. We set SGD (Robbins & Monro, 1951) as the base local optimizer with a learning rate η = 0.1.