Unlocking the Potential of Model Calibration in Federated Learning
Authors: Yun-Wei Chu, Dong-Jun Han, Seyyedali Hosseinalipour, Christopher Brinton
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that NUCFL offers flexibility and effectiveness across various FL algorithms, enhancing accuracy as well as model calibration. |
| Researcher Affiliation | Academia | 1Purdue University, 2Yonsei University, 3University at Buffalo-SUNY EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 General FL Framework Algorithm 2 Applying NUCFL to FL |
| Open Source Code | No | The paper does not explicitly provide a link to open-source code for the methodology described, nor does it state that the code is available in supplementary materials or upon publication. |
| Open Datasets | Yes | We conduct experiments using four image classification datasets commonly utilized in FL research (Caldas et al., 2018; Mc Mahan et al., 2017; Mohri et al., 2019): MNIST (Le Cun et al., 1998), FEMNIST (Cohen et al., 2017), CIFAR-10 (Krizhevsky, 2009), and CIFAR-100 (Krizhevsky, 2009). |
| Dataset Splits | Yes | In the IID setup, data samples from each class are distributed equally to M = 50 clients. To simulate non-IID conditions across clients, we follow (Hsu et al., 2019; Nguyen et al., 2023; Chen et al., 2023) to partition the training set into M = 50 clients using a Dirichlet distribution with α = 0.5. |
| Hardware Specification | Yes | We run all experiments on a 3-GPU cluster of Tesla V100 GPUs, with each GPU having 32GB of memory. |
| Software Dependencies | No | The paper mentions using the SGD optimizer but does not specify versions for any key software components like programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow). |
| Experiment Setup | Yes | We run each FL algorithm for 100 rounds, evaluating the final global model, with 5 epochs for each local training. We use the SGD optimizer with a learning rate of 10-3, weight decay of 10-4, and momentum of 0.9. For additional details on the training specifics of each algorithm, please see Appendix A.2. |