Wednesday, August 17

Google explains its text selection technology: your mobile sends data at night to feed the algorithm

Google has published in its artificial intelligence blog how they are improving the automatic prediction function in the selection of text in Android thanks to federated learning. This term refers to machine learning that trains an algorithm across multiple devices, from which the model can obtain data.

With the gigantic network of devices that authorize the sending of data, Google’s neural network feeds on the information necessary to constantly improve. Specifically, Google claims that since using federated learning they have improved model accuracy by up to 20%.

Your keyboard is getting better thanks to data from other phones

Google has taken out its chest about the operation of its automatic prediction for text selection. If we use the Google keyboard and we are going to select a piece of text, the keyboard is able to anticipate the words that we are going to want to select, making it even faster and autocompleting said selection.

Google’s model parses words that correspond to addresses or numbers, to be clear about where the text selection prediction should cut

Smart Text Selection, the name given to this function, focuses on keywords very well defined, such as addresses, phone numbers and so on, to be clear about where the limit of the selected phrase will be.

Tg Image 4068415457

This feature has been around for a while on Android, but now Google has explained how they are improving it thanks to federated learning. The model trains with real user interactions, data that is shared with Google’s servers when we connect the phone at night (although if we have other habits, you can share it when you are “ready”).

google gboard

According to Google the data are shared completely anonymously. Raw device data is not shared, just minor updates to the model. In addition to not sending information itself on our phone, Google ensures that “the data on this network is protected by policies that restrict its use.”

Gboard: 38 tricks and functions to get the most out of Google's mobile keyboard

Google summarizes the operation of its model in a simple abc scheme:

  • a) Devices are selected.
  • b) The selected devices send their local data to the server.
  • c) The improved model is sent back to the device (not the data used for training that comes from other phones).

Thanks to federated learning, Google ensures improvements of between 5 and 7% for the selection of several words, and between 8 and 20% improvement in the specific case of the selection of addresses.