federated learning
/ FED-er-ay-tid LER-ning /
Imagine hospitals that all want a better diagnostic model but are legally and ethically forbidden from sharing patient records with each other. Federated learning offers a way: instead of pooling everyone's raw data in one place, you send a copy of the model to where each dataset lives, train locally, and combine only what was learned — never the data itself.
Here is the loop. A central server holds a shared model and sends it to many participants — phones, hospitals, banks. Each one trains the model on its own private data and sends back just the resulting parameter updates (numbers describing how the model changed), not any raw examples. The server averages all those updates into an improved global model and sends it out again. Round after round, the model learns from everyone's data while that data stays put on each device or institution.
The appeal is privacy and practicality: data that is sensitive, regulated, or simply too big to move can still contribute to a shared model. But it is not automatically private or easy. The shared updates can sometimes leak information about the underlying data, so federated learning is often paired with extra protections like encryption or added noise. It is also harder to train — participants' data differs, devices drop out, and communication is a bottleneck — so it usually trades some accuracy and a lot of engineering complexity for the privacy it buys.
A phone keyboard's next-word predictor improves by learning from how you type — but your messages never leave the phone. Each night your phone trains the shared model on your typing and uploads only the parameter changes. The server averages updates from millions of phones into a better model. Your private texts stayed on your device the whole time.
Move the model to the data, not the data to the model.
Keeping raw data local is not the same as guaranteed privacy — the shared parameter updates can leak information, so federated learning usually needs extra protections like encryption or added noise to be truly safe.