El-Ouazzane did not elaborate on how the device may take its acquired data and send it back to Google so it could be added to the database, but it must, somehow. Without that exchange of information, neither the neural net at large nor the model on-device could grow.
The communication will probably follow the existing model of Android updates. Google will probably collect data, as an independent activity, then train their models & improve their algorithms on their servers. The resulting code & data would be pushed out to devices, periodically.
Maybe Google will surprise us withe some kind of distributed learning that harnesses the compute power and data from the devices to update their centralized model, but that's not how this sort of thing is usually done.
Of course, devices can learn things about their owner and the environment they inhabit, but I see that as a sort of user-specific customization and not something that would get propagated back to Google.