FedBAT: Communication-Efficient Federated Learning via Learnable Binarization
Authors: Shiwei Li, Wenchao Xu, Haozhao Wang, Xing Tang, Yining Qi, Shijie Xu, Weihong Luo, Yuhua Li, Xiuqiang He, Ruixuan Li
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments are conducted on four popular datasets. The results show that Fed BAT significantly accelerates the convergence and exceeds the accuracy of baselines by up to 9%, even surpassing that of Fed Avg in some cases. |
| Researcher Affiliation | Collaboration | 1Huazhong University of Science and Technology, Wuhan, China 2The Hong Kong Polytechnic University, Hong Kong, China 3Fi T, Tencent, Shenzhen, China. |
| Pseudocode | Yes | Algorithm 1 Federated Binarization-Aware Training |
| Open Source Code | No | The paper does not provide an explicit statement or link to open-source code for the described methodology. |
| Open Datasets | Yes | In this section, we evaluate Fed BAT using four widely recognized datasets: FMNIST (Xiao et al., 2017), SVHN (Netzer et al., 2011), CIFAR-10 and CIFAR-100 (Krizhevsky & Hinton, 2009). |
| Dataset Splits | No | The paper describes data partitioning strategies (IID, Non-IID-1, Non-IID-2) and mentions using a 'test accuracy', but it does not specify explicit train/validation/test dataset splits with percentages, absolute counts, or references to predefined validation splits for reproduction. |
| Hardware Specification | Yes | Each experiment is run five times on Nvidia 3090 GPUs with Intel Xeon E5-2673 CPUs. |
| Software Dependencies | No | The paper states: 'All experiments are conducted on Flower (Beutel et al., 2020)', but it does not provide specific version numbers for Flower or any other ancillary software/libraries. |
| Experiment Setup | Yes | The number of clients is set to 30 and 100, respectively. 10 clients will participate in every round. The local epoch is set to 10 and the batch size is set to 64. SGD (Bottou, 2010) is used as the local optimizer. The learning rate is tuned from (1.0, 0.1, 0.01) and set to 0.1. The number of rounds are set to 100 for CNN and 200 for Res Net-10. For the baselines, each hyperparameter is carefully tuned among (1.0, 0.1, 0.01, 0.001), including the step size and the coefficient of noise. ... In Fed BAT, the coefficient ρ is set to 6 and the warm-up ratio ϕ is set to 0.5 by default. |