Local Composite Saddle Point Optimization
Authors: Site Bai, Brian Bullins
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Furthermore, the empirical evaluation demonstrates the effectiveness of Fe Dual Ex for inducing structure in these challenging settings. To complement our largely theoretical results, we verify in this section the effectiveness of Fe Dual Ex by numerical evaluation. Additional experiments and detailed settings are deferred to Appendix A. |
| Researcher Affiliation | Academia | Site Bai Department of Computer Science Purdue University bai123@purdue.edu Brian Bullins Department of Computer Science Purdue University bbullins@purdue.edu |
| Pseudocode | Yes | Algorithm 1 FEDERATED-DUAL-EXTRAPOLATION (Fe Dual Ex) for Composite SPP |
| Open Source Code | No | The paper does not provide explicit access to source code for the methodology described, such as a repository link or a clear statement of code release in supplementary materials. |
| Open Datasets | Yes | The training data for MNIST is evenly distributed across M = 100 clients, each possessing 600. For the CIFAR-10 experiments, each of the 100 clients holds 500 of the training data. |
| Dataset Splits | No | The paper describes the datasets used (MNIST and CIFAR-10) and mentions training and testing, but does not provide explicit train/validation/test split percentages or sample counts to reproduce the data partitioning. It mentions 'Validation is done on the whole validation dataset on the server with unattacked data' but does not specify the size or how this validation set was created/split from the main datasets. |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details, such as library or solver names with version numbers, needed to replicate the experiment. |
| Experiment Setup | Yes | We only tune the global step size ηs and the local step size ηc. For all experiments, the parameters are searched from the combination of ηs {1, 3e-1, 1e-1, 3e-2, 1e-2} and ηc {1, 3e-1, 1e-1, 3e-2, 1e-2, 3e-3, 1e-3}. The client makes K = 5 local updates and communicates for R = 20 rounds. For the CIFAR-10 experiments, each of the 100 clients holds 500 of the training data. The client makes K = 5 local updates and communicates for R = 40 rounds. D = 0.05 for data normalized between 0 and λ = 0.1. |