Opportunistic Model Ensembling for Decentralized Learning over Intermittent Edge Networks

Abstract

Reliable training and deployment of machine learning models in remote, resource-constrained environments is hindered by intermittent connectivity, limited computation, and decentralized data. These challenges limit the potential of applications like plant disease detection, where the ability of edge devices to collaboratively learn from local observations could greatly enhance agricultural resilience. We propose a simple yet efficient approach for decentralized model training that does not rely on continuous connectivity. Edge nodes train lightweight base models on local data and exchange them opportunistically during transient peer-to-peer encounters, enabling incremental refinement across diverse datasets. Rather than synchronized updates, the system ensembles these independently trained models to support robust predictions under sparse, asynchronous communication. Experiments show that our method achieves strong classification performance across varying levels of connectivity, heterogeneity, and node participation, offering a scalable solution for distributed learning in real-world agricultural and infrastructure-limited settings.

Links

code:references.bib

@article{

}

https://scrapbox.io/files/6921a7ddfb4755da61af7386.png

Category

Conference

Keywords

Collaborating Organization