FedLPA: One-shot Federated Learning with Layer-Wise Posterior Aggregation

Authors: Xiang Liu, Liangxi Liu, Feiyang Ye, Yunheng Shen, Xia Li, Linshan Jiang, Jialin Li

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results demonstrate that Fed LPA significantly improves learning performance over state-of-the-art methods across several metrics.
Researcher Affiliation Academia Xiang Liu1, , Liangxi Liu2, , Feiyang Ye3, Yunheng Shen4, Xia Li5, Linshan Jiang1, , Jialin Li1, 1National University of Singapore, 2Northeastern University, 3University of Technology Sydney, 4Tsinghua University, 5ETH Zurich
Pseudocode Yes Algorithm 1 Fed LPA Global Aggregation
Open Source Code Yes Our Fed LPA is available in https://github.com/lebronlambert/Fed LPA_Neur IPS2024.
Open Datasets Yes We conduct experiments on MNIST [63], Fashion-MNIST [64], CIFAR-10 [65], and SVHN [66] datasets.
Dataset Splits Yes We use the data partitioning methods for non-IID settings of the benchmark 1 to simulate different label skews. Specifically, we try two different kinds of partition: 1) #C = k: each client only has data from k classes. ... 2) pk Dir(β): for each class, we sample from Dirichlet distribution pk and distribute pk,j portion of class k samples to client j.
Hardware Specification Yes We conduct experiments on CIFAR-10 on a single 2080Ti GPU to estimate the overall communication and computation overhead.
Software Dependencies No The paper mentions 'Pytorch' in the context of floating point precision ('The default floating point precision is 32 bits in Pytorch.'), but does not specify a version number or list other software dependencies with version numbers.
Experiment Setup Yes We set the batch size to 64, the learning rate to 0.001, and the λ = 0.001 for Fed LPA. By default, we set 10 clients and run 200 local epochs for each client.