Machine learning is the key technology of almost every instance of modern Artificial Intelligence. Enormous datasets are produced in digitized industrial processes and in the Internet of Things, which can well be exploited by learning in deep artificial neural networks. Standard machine learning algorithms require these datasets to be centralized before learning a model. Several good reasons - ranging from data privacy over latency to economic efficiency - favor learning at the edge so that reasoning is fast and no local data is transferred. The article shows how decentralized learning works and how to evaluate it. Moreover, we point to special resource-efficient learning algorithms and discuss small remaining risks of data reconstruction.
