Communication-Efficient Federated Group Distributionally Robust Optimization

Authors: Zhishuai Guo, Tianbao Yang

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

Reproducibility Variable Result LLM Response
Research Type Experimental The effectiveness of our algorithms has been demonstrated on a variety of real-world tasks, including natural language processing and computer vision.
Researcher Affiliation Academia Zhishuai Guo, Tianbao Yang Department of Computer Science and Engineering Texas A&M University zhishguo@tamu.edu,tianbao-yang@tamu.edu
Pseudocode Yes Algorithm 1 FGDRO-CVa R
Open Source Code No All used data are publicly available. Code will be released later.
Open Datasets Yes We use Pile [15], Civil Comments [5], Camelyon17 [1], i Wild Cam2020 [3], and Poverty [74].
Dataset Splits Yes Camelyon17 focuses on tumor detection from lymph node images [1], with data from five hospitals split into training (3), validation (1), and testing (1) sets.
Hardware Specification Yes Each algorithm was run on a high performance cluster where each machine uses a NVIDIA A100 GPU.
Software Dependencies No The paper mentions specific models and frameworks (e.g., GPT2, Distil BERT, Huggingface, DenseNet, ResNet) but does not provide specific version numbers for programming languages, libraries, or other software dependencies.
Experiment Setup Yes We tune the initial step size in [1e-4, 1e-3, 1e-2, 1e-1]. All algorithms set the communication interval I = 32 unless otherwise specified. The local mini-batch sizes are set to 32. Experiments are run for 20K local iterations except for Pile, which runs for 200K iterations. The β parameters of FGDRO-KL and FGDRO-KL-Adam are tuned in [0.01, 0.1, 0.2, 0.5].