Google is using federated learning to improve Assistant’s “Hey Google” accuracy
Google is now using “federated learning” on Android in an effort to reduce “Hey Google” misactivations and misses. Federated learning is a privacy-enhancing technique that allows Google to improve machine learning models without sending any raw data to Google servers.
Select users now have access (via 9to5Google) to a new setting in Google Assistant that allows them to improve the artificial intelligence. Google said it will save the audio of those who opt-in so Assistant can learn over time.
“When Google Assistant activates or nearly activates, federated learning temporarily stores short bits of voice recordings on your device,” Google said in a support document. “With federated learning, we use these recordings to learn how to adjust Google Assistant’s triggering logic.”
Audio recordings will stay on your device while a privacy-preserving technology combines information from you and many other participants to help Assistant learn over time and develop better smart features. Your device may store up to 20 recordings per day, including information about your phone configuration and how successful an interaction is.
Images via 9to5Google
Google said that with federated learning, its goal is to cut down on instances when Google Assistant randomly misactivates. It also wants to avoid scenarios when you say “Hey Google” and Assistant doesn’t trigger.
“Federated learning uses the voice recordings stored on users’ devices to refine models like ‘Hey Google,’” Google said. “It learns how to adjust the model from the voice data, and sends a summary of the model changes to Google servers. To provide a better model for everyone, these summaries are aggregated across users.”
According to Google, federated learning will perform computations on your device only when the device is idle, plugged in, and connected to WiFi. Recordings used to improve the “Hey Google” model are kept by Google for no longer than 63 days unless you delete them before then.
If federated learning sounds familiar, that’s because Google has used the feature for its new Health Studies app.