At Google’s I/O 2019 developer conference in Mountain View, California, the tech giant announced a new approach to machine learning called ‘Federated Learning’ that aims to protect the privacy of Google users.
Google has already integrated Federated Learning into Gboard to make next word and emoji predictions across millions of devices. Federated Learning allows Gboard to learn new words such as: ‘BTS’ and ‘YOLO’ as their usage increases organically.
Federated Learning is able to learn these words after thousands of people start using these words, without Google ever seeing what you type, according to the company.
This new approach does not send data to the cloud. When the phone computes an update to the Google model, only the updates are sent, not the data. The updates are secured and aggregated in batches to improve the Global Model, according to Google.
The updated Global Model is then sent back to each user’s device.
Although Federated Learning is fairly new, Google is looking forward to the potential of the approach in other Google products.