FedBE: Making Bayesian Model Ensemble Applicable to Federated Learning
Authors: Hong-You Chen, Wei-Lun Chao
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our empirical studies validate FEDBE s superior performance, especially when users data are not i.i.d. and when the neural networks go deeper. |
| Researcher Affiliation | Academia | Hong-You Chen The Ohio State University, USA chen.9301@osu.edu Wei-Lun Chao The Ohio State University, USA chao.209@osu.edu |
| Pseudocode | Yes | Algorithm 1: FEDBE (Federated Bayesian Ensemble) |
| Open Source Code | No | The paper does not provide an explicit statement or link indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | We use CIFAR-10/100 (Krizhevsky et al., 2009), both contain 50K training and 10K test images, from 10 and 100 classes. We also use Tiny-Image Net (Le & Yang, 2015), which has 500 training and 50 test images per class for 200 classes. |
| Dataset Splits | Yes | We split part of the training data to the server as the unlabeled data, distribute the rest to the clients, and evaluate on the test set. We search the weight decay hyper-parameter for each network and each method in [1e 3, 1e 4] with a validation set. |
| Hardware Specification | Yes | Using a 2080 Ti GPU on CIFAR10 (Conv Net), building distributions and sampling takes 0.2s, inference of a model takes 2.4s, and distillation takes 10.4s. |
| Software Dependencies | No | The paper mentions 'TensorFlow' in a citation, but does not specify version numbers for any software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | We set the initial ηl as 0.01 and decay it by 0.1 at 30% and 60% of total rounds, respectively. Within each round of local training, we use SGD optimizer with weight decay and a 0.9 momentum and impose no further decay on local step sizes. |