If Mozart were alive today (and if he was feeling a bit uninspired) he might well sit down and produce a piece of music like this:
The tune is actually (you guessed it) the work of a machine-learning algorithm that was fed thousands of pieces of MIDI music as training data.
The algorithm, called MuseNet, was developed by researchers at OpenAI, a research company in San Francisco that's focused on researching intelligence and studying its potential impact.
The researchers trained a very large neural network known as a transformer. This type of network learns to predict the next few notes in a piece of music. You can then give the network a few notes, and have it conjure up something new. It makes it possible to mix different genres and styles, and even to add and remove specific instruments.
The work shows how effectively such a model can capture and reproduce statistical patterns that reflect the character of something like a piece of music.
The same researchers previous used similar technology to auto-generate text from a starting sentence. These results were sometimes remarkably realistic—prompting the researchers to fret (a little dramatically) about the risk that such a tool could be used to mass-produce fake news.
The MuseNet project is interesting from a music-history perspective, as it points to some interesting connections (statistically speaking) between different artists across genres and centuries. Who’d have thought that Richard Wagner and Britney Spears shared so much musical taste?
The tool is also quite fun to play with. If you’ve ever wondered what it might sound like if the Beatles jammed with Lady Gaga, the algorithm offers an answer of sorts:
Some people see great potential for this sort of technology to inspire new music. Sageev Oere, a machine learning researcher at the University of Toronto who's interested in AI-generated music, was wowed by the tool's ability to riff on a famous piece of Mozart's.
It's true that tools like MuseNet may inspire new ways of making music. But how does it compare to human musical creativity? I asked Zach Lipton, an assistant professor at CMU and an accomplished jazz musician, what he thought of MuseNet’s jazz improvisations.
It is uninteresting in precisely the same way as every generic “we trained an LSTM to generate ____”. I don’t think there is anything here that a musician should find interesting.— Zachary Lipton (@zacharylipton) April 25, 2019
Lipton’s skepticism isn’t just musical snobbery. MuseNet, like any AI system that generates music, or art, or text, isn’t being creative or inventive in the same way as a human musician. It’s learning the patterns in existing pieces of work and then regurgitating some statistical variation.
As we’ve noted before, it’s unclear how artistically creative AI can be at all. Unlike MuseNet’s creations, human music is rooted in culture, history, and language. It has remarkable capacity to surprise, shock, and inspire. The algorithms have some ways to go yet.
Updated April 27 with additional comment.