Federated Wasserstein Distance
Authors: Alain Rakotomamonjy, Kimia Nadjahi, Liva Ralaivola
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In addition to establishing the convergence properties of Fed Wa D, we provide empirical results on federated coresets and federate optimal transport dataset distance, that we respectively exploit for building a novel federated model and for boosting performance of popular federated learning algorithms. |
| Researcher Affiliation | Collaboration | Alain Rakotomamonjy Criteo AI Lab Paris, France alain.rakoto@insa-rouen.fr Kimia Nadjahi CSAIL, MIT Boston, MA knadjahi@mit.edu Liva Ralaivola Criteo AI Lab Paris, France l.ralaivola@criteo.com |
| Pseudocode | Yes | Algorithm 1 Fed Wa D |
| Open Source Code | Yes | The code for reproducing part of the results is available at https://github.com/arakotom/fedwad and is built on top of the Python Optimal Transport library (Flamary et al., 2021). |
| Open Datasets | Yes | We sampled 20000 examples randomly from the MNIST dataset... We considered four real datasets, namely MNIST, KMNIST, USPS and Fashion MNIST... We have run experiments on MNIST and CIFAR10 |
| Dataset Splits | Yes | Half of the training samples have been used for learning the autoencoder and the other half for the classification task. |
| Hardware Specification | No | No specific hardware details such as GPU models, CPU types, or memory specifications are provided for the experimental setup. |
| Software Dependencies | No | The code... is built on top of the Python Optimal Transport library (Flamary et al., 2021). |
| Experiment Setup | Yes | We learn 10 coresets over 1000 epochs and at each epoch, we assume that only 10 random clients are available and can be used for computing Fed Wa D. For Fed Wa D, the support size of the interpolating measure has been set to either 10 or 100 and the number of iteration in Fed Wa D to 20. ... The number of epochs has a very small impact on the distance and using 10 epochs suffices to get a reasonably accurate approximation of the distance. On the other hand, the number of support point seems more critical, and we need at least 5000 support points to obtain a very accurate approximation, although we have a nice linear convergence of the distance with respect to support size. |