Google has announced a new machine learning solution that doesn’t require all related data to leave a device, which it dubbed as “federated learning.” In a somewhat radical change in its machine learning strategy, Google has moved away from its norm of gathering information, storing it in the cloud, and processing the data on its servers. A solution based federated learning will download a model developed centrally by Google onto your device, modify this model within the device using the data stored in it, and send a summary of the changes to the model to Google’s servers. The major difference between this solution and its predecessors is the location where data is primarily stored and processed, which happens on a client device in this case.
Google is trying out this machine learning strategy on the query suggestion model of Gboard, a popular Google-developed keyboard for Android devices. Locally stored data includes the timing and context of suggestions, the Mountain View-based company revealed. Once stored, the data is processed in the smartphone as it builds an update for the query suggestion model. In the machine learning perspective, federated learning setups like the one in Gboard have some major issues to deal with. The key issues Google mentioned on its Research Blog are higher latency and slower connections, as well as uneven distribution of data. In order to solve those issues, algorithms used for machine learning have also been modified to form the “Federated Averaging algorithm” that reduces upload times of new model updates and energy consumption of smartphones through creating larger, more substantial updates to Google’s models which are then compressed into smaller packages before uploading. In order for the user experience to not be substantially affected by the processing of model updates, Google has scheduled the model updates to only be processed when the phones are idling, charging, and connected to Wi-Fi.
Google sees several advantages to federated learning. One of them improved privacy, as Google does not have access to the processed data but rather only the small updates detailing changes in the models delivered to its servers. Additionally, users can immediately experience the benefits of updated models as they no longer need to wait for Google to release new updates.