Google trains Gboard algorithms using federated learning, which sends a tiny neural network to your smartphone to study typos locally and only uploads mathematical improvements…

A closer look

Nobody wants the tech giants literally reading their late-night text messages just to make autocomplete slightly more accurate. Privacy-preserving federated analytics flips the traditional AI model on its head. Instead of sending your raw data to a central cloud to train the brain, the cloud sends a copy of the brain to your phone. It learns your unique slang entirely on your local device, and then only uploads the generalized 'lessons learned'—weight updates—back to the central server. It’s a genius way to harness big data without actually looking at the data.