How We Used AI in a Video That Tells the Story of AI

by Brian Caulfield

I am AI. With those words, AI revealed itself Wednesday at our GPU Technology Conference as a technology that can not only change the world, but create a soundtrack for its own technology revolution.

For thousands of years, of course, storytellers have relied on music to turn moments into magic. So what happens when, in a twist, the music that sends your imagination soaring is itself part of the story?

We found out Wednesday, when we revealed that the inspiring music for the video introducing the keynote at GTC was generated by the very technology we had gathered to talk about.

Unleashed by the parallel processing power of GPUs, a new generation of neural networks are giving machines superhuman capabilities. AI can now classify galaxies, help the blind see, drive our cars and even help us find cures for diseases that have long plagued mankind.

How Far We’ve Come

None of this is news to our audience, many of whom helped build the technology powering these miracles. So to keep these marvels from seeming mundane, Eric, our senior video producer, realized we needed a stirring soundtrack — and a surprise — to put the magic of our shared endeavors into perspective.

So Eric’s team turned to Aiva Technologies, a Luxembourg startup led by Pierre Barreau that’s using GPUs to compose music. An amateur pianist, Barreau — who grew up watching his father work in music and movie production — knows, intuitively, how key music is to storytelling.

We couldn’t have been in better hands — virtual or not. Aiva’s deep learning system is the first to officially acquire the legal recognition as a composer. All of its work has a copyright to its own name. Just like any human music maker.

Classically Trained

We commissioned a 52-piece orchestra to bring our video's AI-generated score to life.
We commissioned a 52-piece orchestra to bring our video’s AI-generated score to life.

To make music — human musicians have already recorded an entire album of Aiva’s work — the four-person startup relies on CUDA, TITAN X Pascal GPUs and cuDNN with the TensorFlow deep learning framework.

That framework is classically trained, literally. Aiva’s been fed a database of work that includes Bach, Beethoven and Mozart.

For our score, Barreau set Aiva to work generating melodies in A minor — a key that sounds either sad, or, with skilfull pacing, soaring — then curating the result. In the end, we liked their collaboration so much we commissioned Barreau to assemble a 45-piece orchestra to record his AI’s creation last week.

The result, we hope, gives those gathered at our conference the ability to wonder, anew, at what they have wrought.