FL-NTK: A Neural Tangent Kernel-based Framework for Federated Learning Analysis
Authors: Baihe Huang, Xiaoxiao Li, Zhao Song, Xin Yang
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Theoretically, FL-NTK converges to a global-optimal solution at a linear rate with properly tuned learning parameters. Furthermore, with proper distributional assumptions, FL-NTK can also achieve good generalization. In Section 7, we conduct experiments that affirmatively support our theoretical results. |
| Researcher Affiliation | Academia | 1Peking University, Beijing, China 2The University of British Colombia, Vancouver, BC, Canada 3Institute for Advanced Study, Princeton, NJ, United States 4The University of Washington, Seattle, WA, United States. |
| Pseudocode | Yes | Algorithm 1 Training Neural Network with Fed Avg under NTK setting. |
| Open Source Code | No | The paper states the full version is available at arXiv and mentions using Fed ML as an implementation base, but it does not provide a link to the authors' own source code for the methodology described. |
| Open Datasets | Yes | We examine our theoretical results on a benchmark dataset Cifar10 in FL study. |
| Dataset Splits | No | The paper mentions the total number of training instances (50,000 for Cifar10) and describes non-iid splits using Dirichlet distribution, but it does not specify explicit train/validation/test dataset splits (e.g., percentages or counts for each split). |
| Hardware Specification | Yes | The experiment is conducted on one Ti2080 Nvidia GPU. |
| Software Dependencies | No | The paper mentions using SGD as the local optimizer and refers to Fed ML (He et al., 2020b) for implementation, but it does not provide specific version numbers for these or other software dependencies like Python, PyTorch, or TensorFlow. |
| Experiment Setup | Yes | We use SGD the the local optimizer with a learning rate of 0.03. We set batch size as 128. We set the local update epoch K = 1, 2, 5 and 10. We set Dirichlet distribution parameter for non-iid data as α = 0.5. For all local update epoch settings, we set the client and sever communication round as 200. |