Remember watching that kid at the mall beat round after round of Dance Dance Revolution, and thinking to yourself, "I could do that!", only to Left Shark yourself into oblivion? This is kind of like that, except instead of that kid, it's a deep-learning algorithm, because the dystopian future is now: Turns out a team of researchers out of UC San Diego have trained a machine to choreograph Dance Dance Revolution's pop tracks by feeding it audio, and getting some slick moves in return.
In a newly-published study, Researcher Chris Donahue and his crew tapped DDR's massive database of user-designed dances—nearly 20 years into the game's launch, there are quite a few—to create a system that choreographs music based off of raw audio files from the game.
The resulting machine is called Dance Dance Convolution (we see what they did there), which trains via a two-tiered approach to learn when to place steps (based on rhythms and melodies), and which steps to take (based on the 256 combos a player can make on the dance pad at any given moment).
But even machines need to practice. The system learned when and which moves to bust by running through a data set of more than 35 hours of music and 350,000 steps developed from online banks of user-created sets. It's similar to the kind of model testing used to develop machine learning tasks like speech recognition on your phone.
Plus, the study is more than a great excuse to make some excellent "dancing machine" puns. As the MIT Technology Review points out, music research is often handicapped by copyright issues that can limit or prevent songs from being used in research. DDR skirts that because all of its songs—over 100,000—have been cleared or are created specifically for the game.
Considering computers are already helping us determine which dance moves are the sexiest dance moves, we may as well just leave dancing to the machines altogether. That always goes well.
Andrea Domanick is Noisey's West Coast Editor. Follow her on Twitter.