Can we use machine learning to create compelling music?

This page will be a place to discuss the frequent news stories about the efforts of scientists to teach Artificial Intelligence (AI) to compose music.

Here’s one to get us started: “Magenta is a Google Brain project to ask and answer the questions, Can we use machine learning to create compelling art and music? If so, how? If not, why not?” (See https://magenta.tensorflow.org)

I was intrigued by the question “if not, why not?” When we have answered that question, then we have learned something about the nature of music and perhaps of art in general.

In January 2017, I heard an interview with Douglas Eck, the lead research scientist for the project. In the interview Eck admits, “So far the music is not good. One of the things we struggle with: structure, the big chord changes, the story telling, the long-time scale structure.” (You can hear the interview by searching with keyword “magenta” at marketplacetech.org.)

Where did the idea, that structure and story-telling are what humans love about music, come from?

This idea started with the 1956 publication of Leonard Meyer’s Emotion and Meaning in Music. In the first sentence, Meyer states that composers, performers, theorists and music lovers “are all agreed that music has meaning and that this meaning is somehow communicated to both participants and listeners.”

In the years since, what has music theory figured out about how these meanings get communicated? Daniel Levitin sums up sixty years of progress in his book, This is Your Brain on Music. He says, “The appreciation that we have for music is intimately related to our ability to learn the underlying structure of the music we like.”

According to Levitin you have to learn the structures of the music you like to find out if you like the music. Apparently the appreciation of music does not come easily. This quote happens all the way on page 111 because it takes all of the first 110 pages of the book for Levitin to explain the structures.

Levitin is far from alone in his assertion that you have to learn the structures of the music you like to find out if you like the music. These are the structures that are taught in music theory and appreciation classes throughout the Western world. They are the same structures that Magenta is feeding into their machine learning programs. The interesting thing about Magenta is that their results are revealing the limits of this theory, which is the direct descendent of Meyer’s assumption that music has meaning.

But what if music has no meaning? What if music is just sound?