SpaFL: Communication-Efficient Federated Learning With Sparse Models And Low Computational Overhead

Authors: Minsu Kim, Walid Saad, Merouane DEBBAH, Choong Hong

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that Spa FL improves accuracy while requiring much less communication and computing resources compared to sparse baselines.
Researcher Affiliation Academia Minsu Kim Virginia Tech Walid Saad Virginia Tech Merouane Debbah Khalifa University Choong Seon Hong Kyung Hee University
Pseudocode Yes The overall algorithm is illustrated in Fig 1. and summarized in Algorithm 1.
Open Source Code Yes The code is available at https://github.com/news-vt/Spa FL_Neru IPS_2024
Open Datasets Yes We conduct experiments on three image classification datasets: FMNIST [37], CIFAR-10, and CIFAR-100 [38] datasets with NVIDA A100 GPUs. To distribute datasets in a non-iid fashion, we use Dirichlet (0.2) for FMNIST and Dirichlet (0.1) for CIFAR-10 and CIFAR-100 datasets as done in [39] with N = 100 clients.
Dataset Splits No The paper specifies training data distribution and explicitly mentions a 'test dataset' for evaluation, but it does not specify a separate 'validation' dataset split.
Hardware Specification Yes We conduct experiments on three image classification datasets: FMNIST [37], CIFAR-10, and CIFAR-100 [38] datasets with NVIDA A100 GPUs.
Software Dependencies No The paper mentions using 'Py Torch' in the appendix, but it does not specify a version number for PyTorch or any other software library.
Experiment Setup Yes For FMNIST dataset, we use the Lenet-5-Caffe. For the Lenet model, we set η(t) = 0.001, E = 5, α = 0.002, and a batch size to be 64. For CIFAR-10 dataset, we use a convolutional neural network (CNN) model with seven layers used in [40] with η(t) = 0.01, E = 5, α = 0.00015, and a batch size of 16. We adopt the Res Net-18 model for CIFAR-100 dataset with η(t) = 0.01, E = 7, α = 0.0007, and a batch size of 64. The learning rate of CIFAR-100 is decayed by 0.993 at each communication round.