Federated Multi-Objective Learning

Authors: Haibo Yang, Zhuqing Liu, Jia Liu, Chaosheng Dong, Michinari Momma

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experiments also corroborate the efficacy of our proposed FMOO algorithms. We present the numerical results in Section 5 and conclude the work in Section 6. In this section, we show the main numerical experiments of our FMGDA and FSMGDA algorithms in different datasets, while relegating the experimental settings and details to the appendix.
Researcher Affiliation Collaboration Haibo Yang Dept. of Comput. & Info. Sci. Rochester Institute of Technology Rochester, NY 14623 hbycis@rit.edu Zhuqing Liu Dept. of ECE The Ohio State University Columbus,OH 43210 liu.9384@osu.edu Jia Liu Dept. of ECE The Ohio State University Columbus,OH 43210 liu@ece.osu.edu Chaosheng Dong Amazon.com Inc. Seattle, WA 98109 chaosd@amazon.com Michinari Momma Amazon.com Inc. Seattle, WA 98109 michi@amazon.com
Pseudocode Yes Algorithm 1 Federated (Stochastic) Multiple Gradient Descent Averaging (FMGDA/FSMGDA).
Open Source Code No The paper does not provide any explicit statements about releasing source code or links to a code repository.
Open Datasets Yes First, we compare the convergence results in terms of the number of communication rounds using the Multi MNIST dataset [53] with two tasks (L and R) as objectives. In this experiment, we use the River Flow dataset[54], which contains eight tasks in this problem. We utilize the Celeb A dataset [55], consisting of 200K facial images annotated with 40 attributes.
Dataset Splits No The paper mentions data partitions (non-i.i.d. and i.i.d.) and the dataset size for each client, but it does not specify explicit train, validation, and test splits with percentages or absolute counts for reproducibility of the splits themselves.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments (e.g., CPU/GPU models, memory, or cloud instances).
Software Dependencies No The paper mentions models like ResNet-18 but does not provide specific software dependencies with version numbers (e.g., PyTorch, TensorFlow, scikit-learn, etc.).
Experiment Setup Yes We test our algorithms with four different cases with batch sizes being [16, 64, 128, 256]. To reduce computational costs in this experiment, the dataset size for each client is limited to 256. We set the local update rounds K = 10. The learning rates are chosen as ηL = 0.1 and ηt = 0.1, t. In this experiment, we set ηL = 0.001, ηt = 0.1, M = 10, and keep the batch size = 256 while comparing K, and keep K = 30 while comparing the batch size. In this experiment, we set ηL = 0.0005, ηt = 0.1, M = 10, and K = 10.