On Machine Learning and Music

The buzzword of “Deep Learning”: that even students and professionals of music may not have memorized every piece by Bach and that computers can also copy and paste familiar harmonic structures to mislead said ‘experts.’

To find out, the team asked more than 1,600 people to listen two different harmonies of the same melody. More than 400 of them were professional musicians or music students. Each had to determine which of the two harmonies sounded more like Bach. The team also included harmonies produced by other algorithms in this test. — Technology Review

I recall taking a long drive with co-workers, all of whom had little amiability with each other. The driver and passenger seat hadn’t spoken as a classical piece was played over the Puerto Rican radio. The passenger broke the silence: “Is this Bach? It’s probably Bach.” The driver — who tends to be more of a European techno listener than an avid classical music fan — did not reply.

And then I also remember the“procedurally-generated” gaming craze that overtook the indie horror and top-down RPG / rogue-like scene just a couple years ago: preset rooms, tunnels and pathways that have been designated specific properties to allow engine algorithms to fit together compatible combinations, allowing the developer to focus on unique game mechanics than worry about map design. The practice essentially took the Minecraft “chunk” to the next level (or perhaps it was a regression, which I briefly touch upon further down).

Of course this example was (and is) not as dynamic or exciting as the potential of “#MachineLearning” — a set of words that want to signify an anthropomorphic artificial intelligence — but the point still stands: scan the music, tag each bar or segment with metadata and allow the algorithm to fit together correct sequences. And here the buzzword may transform into the mystical — that we are not told of the fairly simple software machinations that add up to the complex, thereby leading to the conclusion of a type of “creativity” creates the familiar.

It will be exciting times when algorithms will have moved past replication, past reinforcement-based learning, past making listenable — if not yet enjoyable — music. But for the algorithm to be satisfied with itself — with tastes beyond human. Music we must adapt ourselves to. The unlistenable that becomes so decades later.

Back to Minecraft and those beautiful areas where “chunks” did not correspond with accepted notions of terrain. Where the player had to act creatively in order to navigate the map, because the ground - defied gravity!