Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Connecting Federated ADMM to Bayes
Authors: Siddharth Swaroop, Mohammad Emtiyaz Khan, Finale Doshi-Velez
ICLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through numerical experiments, we validate the improvements obtained in performance. The work shows connection between two fields that are believed to be fundamentally different and combines them to improve federated learning. Our empirical results in Sec. 4 show that the new algorithms lead to improved performance (improved convergence) across a range of different models, datasets, numbers of clients, and dataset heterogeneity levels. |
| Researcher Affiliation | Academia | Siddharth Swaroop Harvard University, US Mohammad Emtiyaz Khan RIKEN Center for AI Project, Japan Finale Doshi-Velez Harvard University, US |
| Pseudocode | No | The paper describes methods through mathematical equations (e.g., equations 9, 12, 14, 21 outline update steps), but it does not include a distinct 'Pseudocode' or 'Algorithm' block with structured, code-like formatting. |
| Open Source Code | No | The paper does not contain any explicit statement regarding the release of source code for the described methodology, nor does it provide links to any code repositories. |
| Open Datasets | Yes | We learn a logistic regression model on two (binary classification) datasets: UCI Credit (Quinlan, 1987) and FLamby-Heart (Janosi et al., 1988; du Terrail et al., 2022). We also train a 2-hidden layer perceptron... on MNIST (Le Cun et al., 1998) and Fashion MNIST (FMNIST) (Xiao et al., 2017)... Lastly, we train a convolutional neural network on CIFAR10 (Krizhevsky and Hinton, 2009). |
| Dataset Splits | Yes | For all heterogeneous splits, we sample Dirichlet distributions that decide how many points per class go into each client (details in App. E.1). Our sampling procedure usually gives 2 clients 50% of all the data, and 6 clients have 90% of the data. Within the clients, 60-95% of client data is within 4 classes out of 10. We use α1 = 1 and α2 = 0.5 in our heterogeneous splits on MNIST, Fashion MNIST and CIFAR10. |
| Hardware Specification | Yes | Each run took up to 10 hours on an A100 GPU. |
| Software Dependencies | No | We fix the local batch size and local learning rate (using Adam (Kingma and Ba, 2015)) to be equal for every algorithm... The paper mentions the 'Adam' optimizer and learning rates, but it does not specify version numbers for any software dependencies or libraries used. |
| Experiment Setup | Yes | We fix the local batch size and local learning rate (using Adam (Kingma and Ba, 2015)) to be equal for every algorithm, and sweep over number of local epochs, the δ and α hyperparameters, for Fed Dyn the additional local weight-decay hyperparameter, and for Fed Lap-Func the additional τ hyperparameter. Hyperparameters and further details are in App. E. For example, in Appendix E.2, for UCI Credit: 'We set local Adam learning rate at 10^-3, and minibatch size to 4 (ensuring there is gradient noise always). All methods have a hyperparameter sweep over number of local epochs = [5, 10, 20]. We perform a sweep over the α or δ hyperparameter for Fed Prox, Fed Lap, Fed Lap-Cov and Fed Lap-Func over [10, 1, 10^-1].' |