SmartIdx: Reducing Communication Cost in Federated Learning by Exploiting the CNNs Structures

Authors: Donglei Wu, Xiangyu Zou, Shuyu Zhang, Haoyu Jin, Wen Xia, Binxing Fang4254-4262

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on mainstream CNNs and datasets show that our proposed Smart Idx performs 2.5 69.2 higher compression ratio than the state-of-the-art FL compression algorithms without degrading training performance.
Researcher Affiliation Academia Donglei Wu1, Xiangyu Zou1, Shuyu Zhang1, Haoyu Jin1, Wen Xia1,2*, Binxing Fang1 1School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen, China 2Peng Cheng Laboratory, China
Pseudocode Yes Algorithm 1: Packaging the Selected Parameters. Input: The selected parameters & their properties: wi, {Si, Ni}; // Si: Size of wi located unit; Ni: Number of wi located unit; ki is the convolution kernel containing wi. Output: Individual Package(IP), Kernel Package(k P), Pattern Package(PP) 1: for wi and {Si, Ni}, i = 0, 1, 2, . . . do 2: if Si = 1 then 3: Individual Package wi 4: else if Si > 1, Ni < 2Si then 5: Kernel Package ki 6: else if Si > 1, Ni > 2Si then 7: Pattern Package ki 8: end if 9: end for 10: return IP,KP,PP
Open Source Code Yes More related work about FL compression approaches are introduced in supplementary1. 1https://github.com/wudonglei99/smartidx
Open Datasets Yes The datasets used to training are IID (independent identically distributed) partitioned Tiny Image Net (Tmg Net) (Krizhevsky, Sutskever, and Hinton 2012) / Cifar10 (Krizhevsky and Hinton 2009), and Non-IID partitioned Mnist (Le Cun 1998) / Fashion Mnist (FMnist) (Xiao, Rasul, and Vollgraf 2017).
Dataset Splits No The paper mentions training on datasets and evaluating test accuracy but does not provide specific details on validation dataset splits (e.g., percentages or sample counts for validation).
Hardware Specification No No specific hardware details (e.g., exact GPU/CPU models, memory) used for running the experiments are explicitly provided in the paper.
Software Dependencies No The paper mentions using 'Adam optimizer' but does not provide specific version numbers for any software, libraries, or frameworks used in the experiments.
Experiment Setup Yes In this paper, the base FL configurations are set by referring to the existed studies (Konecn y et al. 2016a; Mc Mahan et al. 2017), there are total of 100 clients, 10 clients will be randomly selected to participate in each FL round. Each client conducts their local model training by using Adam optimizer (Kingma and Ba 2015) for 1 epoch, the local batch size is set to 10.