Federated Learning is a distributed machine learning training approach in which a shared model is trained collaboratively across multiple decentralized devices or sites, each using its own local data, with only the resulting model parameter updates transmitted to a central server for aggregation, rather than transmitting the underlying raw data itself.

Key Points

Federated learning is a training method, not an inference method — it changes how a model learns, not how a trained model is subsequently deployed or served.
Only model weight updates, not raw data, cross the network in federated learning, substantially reducing the bandwidth burden relative to centralized training on aggregated raw data.
Federated learning is a natural fit for fleets of connectivity-constrained assets (vessels, remote pipeline sensors, distributed mine sites) where each node holds locally relevant training data that cannot economically be centralized.
Because raw data never leaves its originating site, federated learning reduces data-sovereignty and confidentiality exposure relative to centralized training pipelines.
Federated learning requires periodic, but not continuous, connectivity — each participating node needs an intermittent window to exchange model updates with the aggregator, making it tolerant of scheduled or opportunistic links.

Concept

In a conventional centralized training pipeline, raw data from every source is transmitted to a central location before a model is trained on the aggregate. Federated learning inverts this: the model, not the data, travels. Each participating node — a vessel, a remote sensor cluster, a mine site — trains a local copy of the model on its own data, then sends only the resulting weight updates to a central aggregator, which combines updates from all participants into an improved global model that is then redistributed. This dramatically reduces the volume of data that must cross constrained links, since model updates are typically far smaller than the raw sensor, image, or telemetry data used to produce them, and avoids the need to centralize potentially sensitive operational data at all.

Explainer

For ConnectedEarth's audience, federated learning is best understood as a response to a specific connectivity economics problem: many industrial fleets generate more raw data at the edge than their available backhaul can economically transport. A shipping company operating dozens of vessels, each generating gigabytes of engine and navigation sensor data daily, cannot practically backhaul all of that over VSAT or LEO links to train a single central fleet-wide model. Federated learning lets each vessel train locally during normal operation and exchange only compact model updates during scheduled connectivity windows — port calls, satellite passes, or off-peak link availability — making fleet-wide model improvement achievable within realistic bandwidth budgets. The same logic applies to distributed energy assets and remote mining operations, where the volume asymmetry between raw sensor data and model updates is the deciding factor in whether centralized or federated training is viable.