"[We] returned to the drawing board to design a completely new, lightweight, machine learning architecture -- not only to enable Smart Reply on Android Wear, but also to power a wealth of other on-device mobile applications," the team wrote.
It tried using current neural net tech and so-called graph learning, but the models didn't fit on a smartwatch and attempts to limit the number of replies "did not produce useful results," they wrote. In an attempt to make it more compact, the researchers built a simpler system that groups messages requiring a similar responses, like "Hey, how's it going?" and "How's it going buddy," rapidly and with a low memory hit.
From there, it uses "semi-supervised graph learning" that checks your replies to messages, word and phrase similarity and other factors to predict the best possible replies. The entire model, including the training, resides and performs "completely on device," the team notes. "The model can also be adapted to cater to the user's writing style and individual preferences to provide a personal experience."
The researchers were surprised at how well it works on Android Wear devices, which aren't renowned as computing powerhouses, and plans to use the AI algorithms behind it to "enable completely new applications in the months to come." As with Google's very similar Gmail-based smart replies, however, be sure to only use it when needed -- even with AI smarts, the person on the other end can tell it's not you.