Distributed Event-Based Learning via ADMM

Authors: Guner Dilsad Er, Sebastian Trimpe, Michael Muehlebach

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

Reproducibility Variable Result LLM Response
Research Type Experimental The article concludes by presenting numerical results from distributed learning tasks on the MNIST and CIFAR-10 datasets. The experiments underline communication savings of 35% or more due to the event-based communication strategy, show resilience towards heterogeneous data-distributions, and highlight that our approach outperforms common baselines such as Fed Avg, Fed Prox, SCAFFOLD and Fed ADMM.
Researcher Affiliation Academia 1 Learning and Dynamical Systems Group, Max Planck Institute for Intelligent Systems, 72076 Tübingen, Germany 2 Institute for Data Science in Mechanical Engineering, RWTH Aachen University, 52068 Aachen, Germany. Correspondence to: Guener Dilsad ER <EMAIL>.
Pseudocode Yes Algorithm 1 Event-Based Distributed Learning with Over Relaxed ADMM Algorithm 2 Event-Based Distributed Optimization with Over-Relaxed ADMM Algorithm 3 Event-Based Distributed Optimization with Over-Relaxed ADMM
Open Source Code No The paper does not contain an explicit statement about releasing code or a link to a code repository.
Open Datasets Yes The article concludes by presenting numerical results from distributed learning tasks on the MNIST and CIFAR-10 datasets. ... We start by evaluating the performance of Alg. 1 on MNIST (Deng, 2012) and CIFAR-10 (Krizhevsky, 2009).
Dataset Splits Yes Our setup included N =10 agents for MNIST, each storing data for a single digit, resulting in the most extreme non-i.i.d. distribution of data among agents. For a CIFAR-10 classifier, the data are distributed among N =100 agents according to a Dirichlet distribution, i.e., we sample pa Dir N(β), where N is the number of agents and β =0.5. We then assign a pa,j proportion of the training data of class a to agent j.
Hardware Specification No The paper does not provide specific hardware details (GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes Table 3: The table summarizes the hyperparameters used for distributed training of MNIST classifier (Fig. 8, Tab. 1) Table 4: The table summarizes the hyperparameters used for the distributed training of CIFAR-10 classifier (Fig. 8, Tab. 1) Table 5: The table summarizes the hyperparameters used for the distributed linear regression and LASSO experiments (Fig. 9). Table 6: The table summarizes the hyperparameters used for the distributed LASSO experiment against communication drops (Fig. 10). Table 7: The table summarizes the hyperparameters used for the distributed training of MNIST classifier over a graph (Fig. 11). Table 8: The table summarizes the hyperparameters used for the distributed linear regression experiment over a graph (Fig. 12).