Can a Computer Write a Hit Musical? Thanks to Machine Learning, We’re About to Find Out

Computers can drive, create recipes, even compose rap songs. London audiences will soon find out whether they can write a hit musical, too.

The new show, “Beyond the Fence,” is the world’s first musical conceived and substantially crafted by computer. Opening next month in the West End, it’s an experiment to determine how technology affects art and the creative process.

“Obviously one of the hardest things for a machine to do would be to create, experience and understand art,” said Archie Baron, the show’s executive producer, who helped devise the experiment. “So, we thought, ‘Why not try?’”

Recipe for Success

“Beyond the Fence” started with an advantage human-written shows lack: a clear recipe for success. University of Cambridge researchers used machine learning to analyze what separates a hit musical from a flop. The team looked at factors like cast size, setting and themes explored. They asked whether successful shows always had romance or death, or both.

Two of the researchers, Alex Davis and James Robert Lloyd, then created a machine learning-based lyricist named Claudia. In machine learning, computers are trained with vast amounts of data, and then learn on their own, correcting mistakes and improving over time.

Claudia had a less-than-traditional musical education: She learned English from Wikipedia. Her knowledge of lyrics came from a collection from Wingspan Productions, a TV production company. Researchers trained her on a machine learning system powered by four GPUs hosted in the Amazon Web Services cloud. GPUs are ideal for machine learning because they have the computing power to dramatically speed up the learning process.

Once trained, Claudia generated lyrics for the show’s human creators, composer Benjamin Till and writer and actor Nathan Taylor.