Distributionally Robust Federated Averaging

Authors: Yuyang Deng, Mohammad Mahdi Kamani, Mehrdad Mahdavi

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We give corroborating experimental evidence for our theoretical results in federated learning settings.
Researcher Affiliation Academia Yuyang Deng Mohammad Mahdi Kamani Mehrdad Mahdavi The Pennsylvania State University {yzd82,mqk5591,mzm616}@psu.edu
Pseudocode Yes Algorithm 1: Distributionally Robust Federated Averaging (DRFA)
Open Source Code Yes The code repository used for these experiments can be found at: https://github.com/MLOPTPSU/TorchFed/
Open Datasets Yes We use three datasets, namely, Fashion MNIST [48], Adult [1], and Shakespeare [4] datasets.
Dataset Splits No The paper mentions using 'test accuracies' and 'training' but does not provide specific details on train/validation/test dataset splits, percentages, or methodology for partitioning the data.
Hardware Specification Yes We implement our algorithm based on Distributed API of Py Torch [41] using MPI as our main communication interface, and on an Intel Xeon E5-2695 CPU with 28 cores.
Software Dependencies No We implement our algorithm based on Distributed API of Py Torch [41] using MPI as our main communication interface (PyTorch is mentioned but its specific version number is not provided, nor are specific versions for MPI or Python).
Experiment Setup Yes We use different synchronization gaps of τ {5, 10, 15}, and set η = 0.1 and γ = 8 10 3. [...] The batch size is 50 and synchronization gap is τ = 10. We set η = 0.1 for all algorithms, γ = 8 10 3 for DRFA and AFL, and q = 0.2 for q-Fed Avg.