A Bias Correction Mechanism for Distributed Asynchronous Optimization

Authors: Yuan Gao, Yuki Takezawa, Sebastian U Stich

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

Reproducibility Variable Result LLM Response
Research Type Experimental We also conduct numerical experiments to corroborate our theoretical findings. ... 5 EXPERIMENTS In this section, we conduct numerical experiments to validate the theoretical results of our algorithms.
Researcher Affiliation Academia Yuan Gao EMAIL CISPA Helmholtz Center for Information Security Universität des Saarlandes Yuki Takezawa EMAIL Kyoto University OIST Sebastian Stich EMAIL CISPA Helmholtz Center for Information Security
Pseudocode Yes Algorithm 1 Async BC-GD ... Algorithm 2 Async BC-SGD
Open Source Code Yes All our code can be accessed at here and here.
Open Datasets Yes Regularized Logistic Regression For Fashion MNIST Classification Now we consider a regularized logistic regression problem for the opensourced Fashion MNIST dataset (Xiao et al., 2017).
Dataset Splits Yes We use 20% of the training dataset for the validation dataset.
Hardware Specification Yes All experiments were run on an Intel(R) Xeon(R) CPU E7-8890 v4 @ 2.20GHz chip.
Software Dependencies No The paper mentions 'All our code can be accessed at here and here.' but does not specify any software versions for frameworks or libraries used.
Experiment Setup Yes For the synthetic least squares problem, we set the number of clients n = 4, problem dimension d = 10, and target error (the average of the last 20 iterations) at 0.01. η is searched over {1.0 10 10, 5.0 10 10, , 1.0 10 1}. ... We set the number of clients n = 64, set the batch size to 32, and simulate the different computation speeds as in the previous section, where we set τ = 50. We perform a grid search over {0.5, 0.1, 0.05, 0.01, 0.005} for the best η parameter, and select the step size with the best average accuracy at the last 10 iterations.