I spent a reasonable portion of last year digging into neural networks and machine learning, wondering if and when computers were going to take my job. So when news broke of the "computer-generated musical" Beyond the Fence hitting London's West End, I was intrigued. After waiting for a couple of months, I headed to a performance of the show's limited run, but I left the theater unmoved. Computers can help write a musical, it seems, but they can't yet write a good one.
"What if there was a wounded soldier who had to learn how to understand a child in order to find true love?"
That's the core, machine-derived premise for Beyond the Fence. But before you can understand why it doesn't work and how the musical falls short, you need to understand how and why it was made. Luckily, the story of how it came to be is more interesting than the show itself. Local TV channel Sky Arts commissioned the production, and it has also produced a two-part "making-of" documentary, which is currently airing in the UK. The elevator pitch was to produce a musical "conceived by computer and substantially crafted by computer." Technically, though, it used several computers and more than a few human beings.
Programs from multiple educational institutions were put to work. It started at Cambridge University, which analyzed thousands of musicals to ascertain what makes a hit or a flop. Its conclusions, if you've seen a few musicals, are unlikely to surprise you. The University's analysis asked for a love story set in '80s Europe with a female lead, a death, the overcoming of loss and a happy ending. For music, multiple styles are apparently key, as is the pacing of tunes: It's best to end the first and second acts with vitality. Note to aspiring composers: Cambridge's software says not to set your musical in present-day America and give it a male lead.
The next step was story. This is where Goldsmiths, the University of London's "What-If Machine" (WHIM), stepped in. While Cambridge analyzed musicals based on raw data, WHIM is a much more random endeavor. Its purpose is to spark creativity by combining topics in surprising, subversive ways. WHIM fired out hundreds of concepts, which the people helming the project sifted through. After discarding many ideas based on Cambridge's hit/flop research, they landed on the premise of a single, wounded soldier.
With the core idea in the bag, the creative process moved to the University of Madrid. There, software named PropperWryter, which works in a similar data-driven way as Cambridge's hit/miss analysis, was tasked with refining the plot. The first step was to feed it thousands of musicals. From there, the software analyzed lyrics to chart the emotional course of each musical's acts. It then averaged this out to show when a story should hit beats of love, danger, hate, happiness and so on. This provided a definitive structure for the plot and set the tone for the music.
And here's where the computer-generated claim starts to unravel. There's no software that can put all of these elements together and turn them into a musical. That requires a human. Or a pair of humans, in this case. Writer/composer duo Benjamin Till and Nathan Taylor -- best known in the UK for broadcasting their wedding as a musical -- took all of the programs' output and planned a musical around it.
They fell upon the location of Greenham Common Women's Peace Camp, home for over a decade to hundreds of women protesting NATO's nuclear missile program. This was, despite the duo's claims (made in the accompanying documentary), an entirely human decision. Cambridge's analysis may have helped narrow the search by restricting the time period to the '80s, but the only word from a computer that led to Greenham Common was "soldier." And even that was picked by human hand from a list of WHIM's suggestions.
Music and lyrics are perhaps the best place for computers to truly get involved in the creative process. Computers have been composing interesting music for some time now, and machine learning has led to a number of viable writing projects. Till and Taylor tapped Durham University's Dr. Nick Collins, a composer and musician who uses machine learning to create songs, and adapted some tunes for the musical.
For lyrics, Beyond the Fence looked to a "cloud lyricist," which, upon closer inspection, turns out is just Andrej Karpathy's Char-RNN (a character-based recurrent neural network) trained to write songs. I'm well versed in Char-RNN -- I trained it to write like Engadget last year -- and also intimately familiar with its weaknesses. There's no rhyme or reason behind its words: It guesses what the next letter should be based on what came before. Because of this, Char-RNN was just used for inspiration. The musical has expressions and visuals from the network, but the bulk of lyrics were written and/or arranged by Taylor.
Everything about Beyond the Fence, it turns out, was meticulously managed by Till and Taylor. And that's why it doesn't work. By coming up with the perfect formula for a musical, you invariably create something that's formulaic. When watching, every song felt calculated, each plot point carefully measured. Neither act shocked: There's a chance meeting here, a romantic entanglement there, a death and redemption thrown in for good measure. Nothing moved the needle. Nothing felt fresh.
This is a game that advertisers and brands play every day, and there's nothing creative about it.
The musical's venue, Arts Theatre West End, has a history of showing experimental and "edgy" productions, and this should have followed in that rich tradition. But the Beyond the Fence approach has more in common with marketing than with experimenting. Averaging out trends, using numbers to decipher what does and doesn't sell -- this is a game that advertisers and brands play every day, and there's nothing creative about it.
In the areas where computers can create things no human ever would -- music with bizarre structure and cadences, nonsensical rhymes, and random themes -- the computers were largely ignored. Rather than believing in their tools, Till and Taylor injected their own ideas into Beyond the Fence, only trusting computers to guide them toward a guaranteed hit. In the process, they removed any personality or vitality it could have had.
So maybe I was wrong. Computers might be able to write a good musical. It's just that no one has trusted them to do it yet.