Wiki · Translation · concept
Federated learning
Train one shared model across many sites without moving the raw data — the model travels to the data.
Federated learning trains a single shared model across many sites without ever pooling their data. Instead of shipping every hospital’s records to one central server, you ship the model out to each site, let it learn locally, and bring back only what it learned. The data stays put; the model does the traveling.
How it works
A round looks like this: a central coordinator sends the current model to every participating site. Each site trains it on its own private data and sends back only the resulting update — the changed weights or gradients, never the raw records. The coordinator aggregates those updates (often just averaging them) into a new shared model, then starts the next round. Repeat until the model converges. Because the updates themselves can still hint at the underlying data, federated learning is frequently paired with differential privacy or secure aggregation, which lets the server combine updates without reading any single one.
Why it matters (for bio × AI)
Much of the most valuable biomedical data — patient scans, genomes, clinical notes — legally and ethically cannot leave the institution that holds it. Federated learning is how you build a model that has effectively seen many hospitals’ patients while each hospital’s records never move, and without creating one central honeypot to breach. That unlocks the scale medical AI needs. The honest caveats: sites are non-IID — each hospital’s population and instruments differ, which can destabilize training; passing models back and forth is communication-heavy; and, as noted, the shared updates can still leak information unless you add privacy protections on top.
Related
See differential privacy for the guarantee that keeps shared updates from leaking, and fine-tuning for adapting a shared model to one site’s local data.