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 and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Explicit Personalization and Local Training: Double Communication Acceleration in Federated Learning
Authors: Kai Yi, Laurent Condat, Peter Richtárik
TMLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show by extensive experiments that Scafflix is efficient in real-world learning setups and outperforms existing algorithms. |
| Researcher Affiliation | Academia | Kai Yi EMAIL Department of Computer Science King Abdullah University of Science and Technology (KAUST) Laurent Condat EMAIL Department of Computer Science King Abdullah University of Science and Technology (KAUST) SDAIA-KAUST Center of Excellence in Data Science and Artificial Intelligence (SDAIA-KAUST AI) Peter Richtárik EMAIL Department of Computer Science King Abdullah University of Science and Technology (KAUST) SDAIA-KAUST Center of Excellence in Data Science and Artificial Intelligence (SDAIA-KAUST AI) |
| Pseudocode | Yes | Algorithm 1 Scafflix for (FLIX) Algorithm 2 i-Scaffnew for (ERM) |
| Open Source Code | Yes | The code is publicly available at https://github.com/William Yi96/Scafflix. |
| Open Datasets | Yes | We employ the mushrooms, a6a, and w6a datasets from the Lib SVM library (Chang & Lin, 2011) to conduct these tests. Our selection comprises two notable large-scale FL datasets: Federated Extended MNIST (FEMNIST) (Caldas et al., 2018), and Shakespeare (Mc Mahan et al., 2017). |
| Dataset Splits | Yes | We consider several non-iid splits and present the results on feature-wise non-iid in Figure 1. We discuss the difference among non-iid settings and complementary results in Appendix E.1. In line with the methodology described in Fed Jax (Ro et al., 2021), we distributed these samples across 3,400 devices, with each device exhibiting a naturally non-iid characteristic. For the Shakespeare dataset, we distribute randomly across 1,129 devices. Appendix E.1: IID: Data is uniformly distributed across all clients with identical weighting factors, denoted as αi. Label-wise Non-IID: We induce imbalances in label distribution among clients. Feature-wise Non-IID: Variations in feature distribution across clients are introduced by segmenting the features into clusters with the k-means algorithm. Quantity-wise Non-IID: Data volume variance among clients is realized. The distribution of data samples per client follows a Dirichlet distribution, with a default setting of α = 0.5. |
| Hardware Specification | Yes | All the experiments were conducted on a single NVIDIA A100 GPU with 80GB of memory. |
| Software Dependencies | No | The paper mentions using 'Fed Jax (Ro et al., 2021)' but does not provide specific version numbers for this or any other software libraries used. |
| Experiment Setup | Yes | For both FLIX and Scafflix, local training is required to achieve the local minima for each client. By default, we set the local training batch size at 100 and employ SGD with a learning rate selected from the set Cs := {10 5, 10 4, , 1}. Upon obtaining the local optimum, we execute each algorithm with a batch size of 20 for 1000 communication rounds. The model s learning rate is also selected from the set Cs. We set the communication probability to p = 0.2. To align with Figure 2, we set α = 0.5 for FEMNIST and α = 0.3 for Shakespeare. We set the batch size to 128 and determine the most suitable learning rate through a hyperparameter search. We conduct extensive experiments with different client numbers per round, choosing τ from {1, 5, 10, 20}. We select p from {0.1, 0.2, 0.5}. |