Google develops new machine learning model for phones

Font Size

GOOGLE, INC. has introduced federated learning, a new approach for machine learning through user interaction.

Invented by Google in 2017, federated learning allows machine training to take place within the mobile phones of users without collecting any personal data or information.

“Rather than thinking about how to reduce the amount of data needed for the algorithm to learn, we think instead how to distribute the learning algorithm across devices, in such a way that Google does not have to see any of the data,” Google Distinguished Scientist Blaise Aguëra y Arcas said in a roundtable discussion.

“With federated learning, what if you were to do something, like a giant distributed supercomputer out of every device, out of everybody’s phone? Your own data stays your own and your phone learns from those data,” Mr. Aguëra said.

Mr. Aguëra added that the learning process only takes place when the device is idle or in a charging state.

The model will run locally on the phone, improving and updating the model from data currently in the device. The data is then summarized or compressed. After summarizing the data, the updated model is encrypted and sent to the cloud where it will be averaged with other model results from other users.

The averaged data will be used to update the current shared model phones have through additional training. An improved model will be sent and integrated to the existing model of the user’s mobile phone.

The new approach will allow the user’s model to learn from the interaction of the user and from other users as well.

“When you plug it in, and it does not have anything else to do at night, it adjusts its own neural net weights and then the adjustments to the weights can be sent back to the cloud and combined with everybody else’s adjustments in order to generate a better neural net and that is then sent back out to all the devices and the cycle repeats,” Mr. Aguëra said.

Currently, Mr. Aguëra said the learning model has been implemented in Gboard or the Google Keyboard application on Android. The model learns when users interact or select the suggested words in the application’s search bar.

“We wanted an application that intelligence could really benefit, that works across all apps not just with Google services and where it was important for us to preserve the sovereignty and privacy of whatever you type,” Mr. Aguëra said when asked why the new learning model was first used in Gboard.

“I think the opportunities for federated learning are just as big, if not bigger, in IoT or Internet of Things devices, as in phones,” he added.

Meanwhile, he said federated learning assures tighter security in the collection and processing of data from updated models sent from phones.

“In this way, we don’t have to make a trade-off between privacy and functionality with AI (artificial intelligence). You can have both. The federated learning can apply AI to a huge volume of information that is accessible on phones. [Data] can be learned from but in a way that Google cannot see any of those data,” Mr. Aguëra said.

The new learning technique only allows the shared model to gather the summarized changes made in the user’s device, leaving out any of the user’s direct interactions with the model, he said.

Likewise, the encrypted data summary is only decrypted when there are available summaries from other users that are available for averaging. Once averaged, the user’s summarized data are deleted.

Data privacy leaks struck Google in 2018 after the application Google+ contained a bug that allowed third-party app developers to access data of users and their friends.

However, Google said a blog that they found no evidence that any developer was aware of the bug and found no profile data misused.

Google later shut down consumer access to the application after failing to disclose the leak to the public and sought to improve data privacy against third-party apps. — Marc Wyxzel C. Dela Paz