Federated Learning under Partially Disjoint Data via Manifold Reshaping

Authors: Ziqing Fan, Jiangchao Yao, Ruipeng Zhang, Lingjuan Lyu, Yanfeng Wang, Ya Zhang

TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments on a range of datasets to demonstrate that our Fed MR achieves much higher accuracy and better communication efficiency.
Researcher Affiliation Collaboration Ziqing Fan, Jiangchao Yao B, Ruipeng Zhang EMAIL Cooperative Medianet Innovation Center, Shanghai Jiao Tong University Shanghai AI Laboratory Lingjuan Lyu EMAIL Sony AI Ya Zhang, Yanfeng Wang B EMAIL Cooperative Medianet Innovation Center, Shanghai Jiao Tong University Shanghai AI Laboratory
Pseudocode Yes Algorithm 1 Fed MR Input: a set of K clients that participate in each round, the initial model weights w0, the maximal round T, the learning rate η, the local training epochs E.
Open Source Code Yes Source code is available at: https://github.com/MediaBrain-SJTU/FedMR.
Open Datasets Yes We adopt four popular benchmark datasets SVHN (Netzer et al. (2011)), FMNIST (Xiao et al. (2017)), CIFAR10 and CIFAR100 (Le Cun et al. (1998)) in federated learning and a real-world PCDD medical dataset ISIC2019 (Codella et al. (2018); Tschandl et al. (2018); Combalia et al. (2019)) to conduct experiments.
Dataset Splits Yes In order to better study pure PCDD, for the former four benchmarks, we split each dataset into ϱ clients, each with ς categories, abbreviated as PϱCς. For example, P10C10 in CIFAR100 means that we split CIFAR100 into 10 clients, each with 10 classes.
Hardware Specification No The paper discusses hardware in a hypothetical context of local clients being mobile phones or other small devices, but does not specify any hardware used for the experiments themselves.
Software Dependencies No The paper mentions model architectures (Res Net18, wide Res Net, Efficient Net) and an optimizer (SGD), but does not provide specific software dependencies with version numbers.
Experiment Setup Yes The optimizer is SGD with a learning rate 0.01, the weight decay 10 5 and momentum 0.9. The batch size is set to 128 and the local updates are set to 10 epochs for all approaches.