Anarchic Federated Learning
Authors: Haibo Yang, Xin Zhang, Prashant Khanduri, Jia Liu
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate the proposed algorithms with extensive experiments on real-world datasets. |
| Researcher Affiliation | Academia | 1Department of Electrical and Computer Engineering, The Ohio State University, Columbus, OH 43210, USA; 2Department of Statistics, Iowa State University, Ames, IA 50011, USA; 3Department of Electrical and Computer Engineering, University of Minnesota, Minneapolis, MN 55455, USA. |
| Pseudocode | Yes | The general framework of AFL is illustrated in Algorithm 1. [...] Algorithm 2 AFA-CD Algorithm for Cross-Device AFL. [...] Algorithm 3 The AFA-CS Algorithm for Cross-Silo AFL. |
| Open Source Code | No | No concrete statement about releasing open-source code for the methodology or a link to a code repository is provided in the paper. |
| Open Datasets | Yes | We use i) logistic regression (LR) on manually partitioned non-i.i.d. MNIST dataset (Le Cun et al., 1998), ii) convolutional neural network (CNN) for manually partitioned CIFAR-10 (Krizhevsky, 2009), and iii) recurrent neural network (RNN) on natural non-i.i.d. dataset Shakespeare (Mc Mahan et al., 2016). |
| Dataset Splits | No | For MNIST and CIFAR-10, each dataset has ten classes of images. To impose statistical heterogeneity, we split the data based on the classes (p) of images each worker contains. We distribute the data to M = 10(or 100) workers such that each worker contains only certain classes with the same number of training/test samples. This describes a data partitioning strategy for heterogeneity and mentions training/test samples, but it does not specify explicit reproducible train/validation/test splits (e.g., 80/10/10% or absolute counts) for the overall dataset. |
| Hardware Specification | No | No specific hardware (GPU models, CPU models, cluster specs) used for running the experiments is detailed in the paper. |
| Software Dependencies | No | No specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions) are explicitly mentioned in the paper. |
| Experiment Setup | Yes | For MNIST and CIFAR-10, we use global learning rate η = 1.0 and local learning rate ηL = 0.1. For MNIST, the batch size is 64 and the total communication round is 150. For CIFAR-10, the batch size is 500 and the total communication round is 10000. For the Shakespeare dataset, the global learning rate is η = 50, the local learning rate is ηL = 0.8, batch size is b = 10, and the total communication round is 300. |