Incentive-Aware Federated Learning with Training-Time Model Rewards
Authors: Zhaoxuan Wu, Mohammad Mohammadi Amiri, Ramesh Raskar, Bryan Kian Hsiang Low
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive experiments to demonstrate the superior incentivizing performance of our method compared to existing baselines. |
| Researcher Affiliation | Academia | Institute of Data Science, National University of Singapore, Republic of Singapore Department of Computer Science, Rensselaer Polytechnic Institute, USA Media Lab, Massachusetts Institute of Technology, USA Department of Computer Science, National University of Singapore, Republic of Singapore |
| Pseudocode | Yes | Algorithm 1: Incentive-Aware Federated Learning |
| Open Source Code | Yes | The code has been submitted as supplementary material. |
| Open Datasets | Yes | We use the datasets and heterogeneous data partition pipelines in the non-IID FL Benchmark (Li et al., 2022). We simulate 50 clients for all experiments. We perform extensive experiments on various vision datasets like MNIST (Le Cun et al., 1989), FMNIST (Xiao et al., 2017), SVHN (Netzer et al., 2011), CIFAR-10/100 (Krizhevsky, 2009), Tiny-Image Net (Deng et al., 2009) and language datasets Stanford Sentiment Treebank (SST) (Socher et al., 2013), Sentiment140 (Go et al., 2009). |
| Dataset Splits | No | The paper mentions using datasets from the non-IID FL Benchmark and discusses 'validation accuracy' in figures. However, it does not explicitly provide details about the specific training/validation/test dataset splits (e.g., percentages or sample counts) used for its experiments within the paper's text. |
| Hardware Specification | Yes | All experiments were carried out on a server with Intel(R) Xeon(R)@2.70GHz processors and 1.5TB RAM. We utilized one Tesla V100 GPU for the experiments. |
| Software Dependencies | No | The paper describes model architectures (CNN, ResNet18, LSTM) and training hyperparameters but does not provide specific software dependencies with version numbers (e.g., 'Python 3.x', 'PyTorch 1.x'). |
| Experiment Setup | Yes | We used a default initial learning rate of η0 = 0.001, unless otherwise specified. We used an exponential learning rate decay with a rate of 0.977, causing the learning rate to decrease by 10 folds in 100 iterations. For MNIST, we used η0 = 0.01. For CIFAR-100 with Res Net18, we used η0 = 0.005 with a decay rate of 0.988. The total number of FL training iterations used was 50 for MNIST, FMNIST, SVHN and 100 for CIFAR-10, CIFAR-100 and SST. Each FL iteration involved a client training for 1 local epoch. For all classification tasks, we employed the cross-entropy loss. |