Tech titan Google has introduced a new approach to machine learning called “Federated Learning” which enables mobile devices to collaborate with one another in learning a prediction model that is shared while storing the training data on-device.
With the goal of addressing privacy concerns and at the same time improving functionality, Federated Learning utilizes on-device data in training a smarter central model for enhanced customer experience without the data ever reaching the server. The only elements that will touch the server would be the training results that also undergo a secure aggregation protocol which adds zero-sum masks to scramble the results, like encryption with a key that the server does not have.
The mobile device first downloads the current model and enhances it based on the pre-existing data on the device. The changes will come in a form of a small focused update that is sent to the server and is averaged along with other user updates to produce a shared model that does not rely on user data. Instead, it will be results-driven.
With privacy already addressed, Federated Learning also accommodates less power consumption with lower latency, with the shared model readily available to the users. A well-known use case is the Gboard on Android that learns based on the present data in delivering better suggestions that contribute to a better personalized user experience.
When the concept is applied in different industries, processes can drastically improve. For example, if Google Maps displayed ETAs based on aggregated data on Filipino drivers’ behavior and traffic patterns, the results would most likely be very accurate. Federated Learning is a way to increase efficiency in data processing while still making sure that privacy remains intact.
“In these years, it has really gone from just an idea to an entire sub-discipline of AI. There’s more than one paper a day being published now about Federated Learning. There are companies basing their business plans on Federated Learning,” said Blaise Aguëra y Arcas, distinguished scientist for Google AI.
In any case, Google’s current work has barely scratched the surface of the capabilities empowered by cloud-based and mobile machine learning, but in time, even identifying the breed of a dog in a photograph done entirely by mobile processes and visual filters/labels on the cloud will be a piece of cake.
Federated Learning is positioned in the right spot with the emerging IoT and smart city integrations as well as 5G capabilities and better mobile devices, to produce user-centric models without the raw data, heralding as an example of the unison between security and efficiency.