Smarter Android: AI-powered Google services will get better as you use them
The phone personalizes the machine-learning model locally, based on how it is used (A). Many users' updates are aggregated (B) to form a consensus change (C) to the shared model, after which the procedure is repeated.
Google is introducing a new way for its AI-powered services to improve as people use them.
The tech giant is testing whether its mobile services could use an approach called Federated Learning to refine their underlying machine-learning models.
For each Google service, a machine-learning model is downloaded to a mobile device. Federated Learning allows these models to improve by learning from data on the phone, and then to summarize any local changes as a small update. This update is then encrypted and sent back to the Google cloud, where it is averaged with other user updates to improve the shared backend model.
The continual refinement of the machine learning model stored on the phone benefits the end user, as improvements no longer depend solely on the improved machine learning models being downloaded to their phone.
Google says the approach also has the advantage of improving privacy, as all the training data remains on the device, and no individual updates are stored in the cloud. Updates will only be decrypted and averaged with those from other phones once hundreds or thousands of similar updates have been gathered.
"Federated Learning allows for smarter models, lower latency, and less power consumption, all while ensuring privacy," Google research scientists Brendan McMahan and Daniel Ramage said in a blog post.
"And this approach has another immediate benefit: in addition to providing an update to the shared model, the improved model on your phone can also be used immediately, powering experiences personalized by the way you use your phone."
Google is testing the Federated Learning approach in Gboard, a keyboard for Android handsets. In this instance, the machine learning model will remember which suggested inputs and information the user clicked on and use that data to improve future suggestions.
The blog post goes into some detail about the complexity of introducing the Federated Learning approach, including mentioning that the on-device training uses a miniature version of TensorFlow, Google's open-source software library for machine learning.
Google says that the Federated Learning approach can't be used to help solve every machine learning challenge, with exceptions including using labelled images to teach a machine to recognize the breed of dog in a photo.
In a somewhat similar move, last year Apple said it would approach machine learning in a way that respects personal data, by using what it called 'differential privacy'. This approach allows it to analyze customer data for trends without being able to identify any particular individuals: for example, to be able to spot trending words that need to be added to the QuickType keyboard suggestions.
VIDEO: Google uses smartphone data to tweak AI-powered services
Read more on Google and mobile
- Google mobile app, website redesign adds search shortcuts
- Google to crack down on intrusive mobile ads
- Google to desktop owners: You're about to get staler search results than mobile users
- Google mobile Search: Will new 'rich cards' really help you sift results
- Google's mobile app can figure out what's wrong with you (CNET)
- How to more effectively manage mobile devices in Google's G Suite (TechRepublic)