Distributed Stochastic Bilevel Optimization: Improved Complexity and Heterogeneity Analysis

Authors: Youcheng Niu, Jinming Xu, Ying Sun, Yan Huang, Li Chai

JMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we present two numerical experiments to test the performance of the proposed Lo PA algorithm and verify the theoretical findings. Specifically, the first experiment considers classification problems for a 10-class classification task, while the second focuses on hyperparameter optimization in l2-regularized binary logistic regression problems for a two-class classification task. [...] The experiment results for loss, training accuracy, and testing accuracy are presented in Figures 1 and 2.
Researcher Affiliation Academia College of Control Science and Engineering, Zhejiang University, China School of Electrical Engineering and Computer Science, The Pennsylvania State University, USA
Pseudocode Yes Algorithm 1 Lo PA
Open Source Code No The paper does not provide a concrete link to source code, an explicit statement of code release, or mention of code in supplementary materials for the methodology described.
Open Datasets Yes We employ MNIST datasets to train m personalized classifiers for a 10-class classification task. [...] We conduct the experiment across various datasets including MNIST (784 features, 12000 samples for digits 0 and 1 ), covtype (54 features, 90000 samples for the Lodgepole and Ponderosa pine classes), and cifar10 (3072 features, 6000 samples for the dog and horse classes).
Dataset Splits No The paper describes how samples are distributed among nodes (e.g., 'each node having 14000 samples', 'validation and training sets for each node are randomly assigned with a uniform number of samples') but does not provide specific train/test/validation split percentages or absolute counts for the overall dataset or for individual nodes, which are necessary for full reproducibility of data partitioning.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., library names like PyTorch 1.9, or solver versions).
Experiment Setup Yes The step-sizes are set as α = 0.01, β = 0.01, λ = 0.008, γ = 0.4, τ = 0.4 both for Lo PA-LG and Lo PA-GT. [...] mini-batch sizes are set to 50 for all algorithms. [...] The first layer of the classifier contains 28 neurons, while the second layer contains 10 neurons.