FedALA: Adaptive Local Aggregation for Personalized Federated Learning
Authors: Jianqing Zhang, Yang Hua, Hao Wang, Tao Song, Zhengui Xue, Ruhui Ma, Haibing Guan
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To evaluate the effectiveness of Fed ALA, we conduct extensive experiments with five benchmark datasets in computer vision and natural language processing domains. Fed ALA outperforms eleven state-of-the-art baselines by up to 3.27% in test accuracy. |
| Researcher Affiliation | Academia | 1Shanghai Jiao Tong University 2Queen s University Belfast 3Louisiana State University |
| Pseudocode | Yes | Algorithm 1: Fed ALA |
| Open Source Code | Yes | Code is available at https://github.com/Tsing Z0/Fed ALA. |
| Open Datasets | Yes | We conduct extensive experiments in computer vision (CV) and natural language processing (NLP) domains. For the CV domain, we study the image classification tasks with four widely used datasets including MNIST (Le Cun et al. 1998), Cifar10/100 (Krizhevsky and Geoffrey 2009) and Tiny-Image Net (Chrabaszcz, Loshchilov, and Hutter 2017) (100K images with 200 classes) using the 4layer CNN (Mc Mahan et al. 2017). For the NLP domain, we study the text classification tasks with AG News (Zhang, Zhao, and Le Cun 2015) and fast Text (Joulin et al. 2017). |
| Dataset Splits | No | The paper states '25% of the local data forms the test dataset, and the remaining 75% data is used for training.' which defines train/test split, but does not explicitly define a separate validation split. |
| Hardware Specification | Yes | We implement Fed ALA using Py Torch-1.8 and run all experiments on a server with two Intel Xeon Gold 6140 CPUs (36 cores), 128G memory, and eight NVIDIA 2080 Ti GPUs, running Cent OS 7.8. |
| Software Dependencies | Yes | We implement Fed ALA using Py Torch-1.8 and run all experiments on a server with two Intel Xeon Gold 6140 CPUs (36 cores), 128G memory, and eight NVIDIA 2080 Ti GPUs, running Cent OS 7.8. |
| Experiment Setup | Yes | We set the local learning rate to 0.005 for the 4-layer CNN (0.1 on MNIST following Fed Avg) and 0.1 for both fast Text and Res Net-18. We set the batch size to 10 and the number of local model training epochs to 1, following Fed Avg. We run all the tasks for 2000 iterations to make all the methods converge empirically. |