NeuralDraw
Getting from Google’s Quick, Draw! to an AI doodle game tailored for a fun mobile experience
What it takes to make a smartphone doodle smart
We mixed and matched social gaming components we saw work for mobile games with Quick, Draw!’s doodling experience to create a full game. For the underlying AI tech, we started with an open-sourced dataset. We had to build an AI layer that was accurate enough to guess the user’s doodlings, but also instantly responsive to not compromise the game experience.
On the user side, gameplay success means rising in the ranks
We wanted to design a game that was easily played by any kind of user. They would start with simple levels and move up towards more complex doodling challenges. All the while, the app tracked their performance in a tournament-style point system, so the user could check their position in the leaderboard and see how they’re compared to players all-over the world.
The Neural Network that could get the job done
Instead of the Recurrent Neural Network the ‘Quick, Draw!’ team used, we switched to an on-device Convolutional Neural Network model which is 400kB in size. It does inference in between 14-24ms, depending on the type of device, and works in real time on iPhone 6 or newer generation smartphones. Moreso, the build supports an unlimited number of people playing at the same time, with little need to scale the backend of the app.
“All in all, Neural Draw was a great AI-product design and implementation exercise for the team. We proved that you could take cutting edge AI and combine it with mature mobile technology to create an enjoyable mobile experience.”
LEVENTE SZABO – HALCYON MOBILE CO-FOUNDER, HALCYON LABS LEAD
