Glossary · AI engineering and governance
Federated learning
German: Föderiertes Lernen
In machine learning, federated learning is a training approach in which a shared model is trained across several sites or devices that keep their data locally; only model updates, not raw data, are sent to a coordinating server and combined into a global model.
- Industrial AI
- AI
In one sentence
Federated learning trains a shared model across several sites that keep their data locally and share only model updates.
Example
A machine builder improves a wear model with data from ten customer plants, each training locally on its own machines and sending only parameter updates.
How it applies
- Engineering: Federated learning suits cases where data cannot leave a site for confidentiality, legal or bandwidth reasons. It requires compatible data formats and features at all sites.
- Security: Sharing model updates instead of raw data reduces but does not eliminate the risk of leaking information; manipulated updates from one site can also degrade the global model.
- Documentation: Document participating sites, the aggregation method, validation of the global model and the agreements on data use (Data sovereignty).
Federated learning vs. central training
In central training, all data is collected in one place. In federated learning, data stays where it was generated. Data from different plants often differs, which makes validation of the combined model more demanding.