LBI-FL: Low-Bit Integerized Federated Learning with Temporally Dynamic Bit-Width Allocation
Authors: Li Ding, Hao Zhang, Wenrui Dai, Chenglin Li, Weijia Lu, Zhifei Yang, Xiaodong Zhang, Xiaofeng Ma, Junni Zou, Hongkai Xiong
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that the proposed LBI-FL achieves a reduction of more than 50% Bit OPs per client on average for FL with less than 2% accuracy loss compared to low-bit training with INT8 precision. Section 5 is titled "Experiments" and includes "Image Classification" and "Ablation Studies" evaluating performance on various models and datasets. |
| Researcher Affiliation | Collaboration | 1Department of Electronic Engineering, Shanghai Jiao Tong University, Shanghai, China. 2Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China. 3United Automotive Electronic Systems, Shanghai, China. |
| Pseudocode | Yes | Algorithm 1 Low-bit Integerized Federated Learning. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing the code for the methodology described, nor does it provide a link to a code repository. |
| Open Datasets | Yes | Experiment setting. We evaluate with Le Net (Le Cun et al., 1998), Res Net-18/50/101 (He et al., 2016), Mobile Net V2 (Sandler et al., 2018), and Vi T-S (Dosovitskiy et al., 2021) on the CIFAR-10/100 dataset. |
| Dataset Splits | Yes | We collect the teacher samples by training Res Net-20 (He et al., 2016) for 100 epochs on a subset of the CIFAR-10 dataset randomly sampled with 10% of the original data. ... For Le Net, the number of training epochs is set at 2000 and the client number is 100. 10% of the clients are selected to update at every epoch. For other larger networks, the number of training epochs is set at 200 and the client number is 10. All the clients are updated at every epoch. The local update epoch is 2 and the learning rate decay is 1. ... We consider both iid (independent and identically distributed) and non-iid data distributions. The parameters of non-iid Dirichlet distribution are 0.25 and 0.5 in our experiments. |
| Hardware Specification | Yes | The experiments in this paper are conducted on a single NVIDIA 3090 GPU. |
| Software Dependencies | No | The paper mentions the use of an "agent" and "Q-network" but does not specify any software libraries or frameworks with version numbers (e.g., PyTorch, TensorFlow, Python versions) that were used for implementation. |
| Experiment Setup | Yes | In the reward function, θ is set as 0.25 and δ as 0.5. ... For Le Net, the number of training epochs is set at 2000 and the client number is 100. 10% of the clients are selected to update at every epoch. For other larger networks, the number of training epochs is set at 200 and the client number is 10. All the clients are updated at every epoch. The local update epoch is 2 and the learning rate decay is 1. |