Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
Convergence Analysis of Sequential Federated Learning on Heterogeneous Data
Authors: Yipeng Li, Xinchen Lyu
NeurIPS 2023 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results validate the counterintuitive analysis result that SFL outperforms PFL on extremely heterogeneous data in cross-device settings. |
| Researcher Affiliation | Academia | Yipeng Li and Xinchen Lyu National Engineering Research Center for Mobile Network Technologies Beijing University of Posts and Telecommunications Beijing, 100876, China EMAIL |
| Pseudocode | Yes | Algorithm 1: Sequential FL; Algorithm 2: Parallel FL |
| Open Source Code | Yes | Our code is partly from Gao et al. (2021); Zeng et al. (2021); Jhunjhunwala et al. (2023) (more references are included in the code), and it is available at https://github.com/liyipeng00/convergence. |
| Open Datasets | Yes | We partition the training sets of CIFAR-10 (Krizhevsky et al., 2009) and CINIC-10 (Darlow et al., 2018). |
| Dataset Splits | No | The paper describes partitioning 'training sets' and sparing 'test sets' but does not explicitly mention or detail a validation set or its split. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments were mentioned in the paper. |
| Software Dependencies | No | The paper mentions using VGGs and Res Nets models, and SGD as a local solver, but does not provide specific version numbers for software frameworks, libraries, or dependencies. |
| Experiment Setup | Yes | We fix the number of participating clients to 10 and the mini-batch size to 20. The local solver is SGD with learning rate being constant, momentem being 0 and weight decay being 1e-4. We apply gradient clipping to both algorithms (Appendix G.2) and tune the learning rate by grid search (Appendix G.3). |