Google Brings Federated Learning to Android
Personalization has become a key aspect in today’s world. From a business perspective, the level of personalization that a particular tool or technology offers determines its usability, and in many ways, its popularity too.
No one understands this better than Google. In fact, Google is only company that probably knows everything about you, next to the NSA of course!
But the good side of it is Google uses this information that it collects about you to enhance the level of personalization it offers through its many free services.
If you’re wondering how Google makes its money, it’s simple. It collects information about you, gives you personalized data when you search for something or even when you use its services like Gmail and Google docs. In all this, it also inserts ads that are relevant to you and this is how it makes it’s money.
While all that is good, there are many privacy enthusiasts who see this as a breach of their privacy. When someone collects information and stores it in their server, it clearly violates many laws and even common sense. This is why Google has been in the middle of many controversies and lawsuits, especially in Europe.
To circumvent this breach of privacy without compromising on your personalization, Google has come up with an idea to train artificial intelligence (AI) to give achieve this twin objective. Called Federated Learning, this new AI training procedure will take advantage of the computing power of your phone.
It starts with downloading the latest model from the cloud. This model is kept as a base and the AI system improves it by learning from your data on the phone. Finally, it sends an updated model back to Google through an encrypted communication. This model is then averaged with the model obtained from other users and all these together help to improve the shared model. The data that is used to train this model is still on your device though, and none of it reaches the Google servers at any point in time.
Let’s take a practical situation here. Say, you searched for pizza on your Gboard. The phone stores this information locally and also remembers the links you clicked. Federated learning processes this information to improve the suggestions that the Gboard query will make next time. So, when you search for pizza a few days later, the links that you clicked the previous time will be on top of your search results.
Though this may sound great, it can also bring up concerns about battery life and overuse of data. These are things that you don’t have to worry at all because Google will use your phone to update the model only when it is idling around when connected to a power source and it also uses only a free source of Internet like your Wi-Fi for uploading this model.
With such a Federated Learning, it looks like Google has covered everything this time. Your data never leaves your device, AI is used to improve the model, you phone’s battery and data usage is not affected, and yet you get high levels of customization.
Too good to be true? Time is the answer.