What has four arms, eight sticks, a brain suffused with two million musical motifs, and creates some smooth tunes? As it turns out, Shimon the marimba-playing robot.
Developed by Mason Bretan at the Georgia Institute of Technology, Shimon has actually been playing music for several years. But in the past it’s only known how to accompany other musicians, by playing notes that complement precomposed pieces. Now, Bretan has made use of neural networks to help the robot ingest 5,000 songs and over two million riffs—drawn from artists from Beethoven to Lady Gaga—in order to develop an ability to compose its own pieces.
Shimon listens to four measures of music provided by a human, and then writes its own tune based on them. The robot is “coming up with higher-level musical semantics,” Bretan says in a statement. “Rather than thinking note by note, it has a larger idea of what it wants to play as a whole.” You can listen to one of its compositions in the video above.
Particularly interesting, though, is that because it’s currently impossible to work out exactly what’s going on inside a deep neural network, there’s no way of knowing which artist Schimon is channeling when it composes a tune. “They sound like a fusion of jazz and classical,” says Bretan. “I definitely hear more classical, especially in the harmony. But then I hear chromatic moving steps ... that’s definitely something you hear in jazz.”
It’s not the first time that artificial intelligence has composed music. Google has been working to use neural networks to the same end as part of a project called Magenta, and its piano compositions are surprisingly catchy. But Douglas Eck, a researcher behind some of that work, says that the really fun part of creating these kinds of machines will come when they “help us make a new kind of art.”