Machine learning has already enabled Google's artificial intelligence (AI) to beat a human champion at the complex game of Go and create psychedelic images in the style of Vincent Van Gogh. So it's no surprise that the company has unveiled a new project to explore machine learning's artistic and musical potential.
Google research scientist Douglas Eck announced the public launch of Magenta yesterday in a post on the TensorFlow Web site. Developed by the Google Brain Team, TensorFlow is the company's open source software library for machine intelligence.
An MP3 sample of Magenta's song-writing ability demonstrates a simple piano melody that, while not exactly hummable, clearly has structure and themes. It also illustrates one of the biggest challenges facing researchers trying to get machines to create art: how to teach AI to tell stories that humans will find compelling.
AI Art 'Good in Small Chunks'
"So much machine-generated music and art is good in small chunks, but lacks any sort of long-term narrative arc," Eck wrote in his post. "Alternately, some machine generated content does have long-term structure, but that structure is provided 'to' rather than learned 'by' the algorithm."
Eck said Google has two goals for Magenta. The first is to advance current machine intelligence capabilities in music and art generation. The second is to create a community of artists, coders and researchers who can help Google with the first goal, which is why the Magenta team is open-sourcing its codes and tools. Eventually, Google plans to open up its Magenta site on the GitHub code repository, Eck said.
"Once we have a stable set of tools and models, we'll invite external contributors to check in code to our GitHub," he said. "If you're a musician or an artist (or aspire to be one -- it's easier than you might think), we hope you'll try using these tools to make some noise or images or videos . . . or whatever you like."
Tackling the Age-Old Question: 'Is It Art?'
Humans have spent millennia struggling with the question: "But is it art?" The first challenge with machine learning is creating "generative models" that can essentially create something out of nothing without being given any starting material on which to build. Then the next problem becomes finding ways to determine whether the AI's "artwork" is any good, Eck noted.
"In the end, to answer the evaluation question we need to get Magenta tools in the hands of artists and musicians, and Magenta media in front of viewers and listeners," Eck said. "As Magenta evolves, we'll be working on good ways to achieve this."
Last year Google demonstrated how it could train artificial neural networks to "learn" different artistic styles, and then apply those to random images. The technique, called "Inceptionism," resulted in a collection of 29 paintings that were later sold in a charity auction.
Earlier this year, Google's intelligent machine AlphaGo bested human Go champion Lee Sedol in a five-game match, winning four games to Sedol's one.
MATT - benton, illinois:
Posted: 2016-06-04 @ 2:57am PT
The DATATRON 2000 mainframe computer generated 'music' from random #'s in 1957. Results were printed on a 'flexowriter' for interpretation.
Posted: 2016-06-04 @ 2:33am PT
Why would an "aspiring" musician want to make his life even harder by helping to create an AI that competes with him? The argument that AI will just be a tool that the musician can use to generate music is total BS, such tools already exist and true AI would actually replace and not assist. Do not compare this to Chess AI that can already beat humans, since the competition between humans is what makes it compelling enough for humans to still participate in chess. There is not a competitive element in music, thus when AI makes better music than humans, human composers and creators of music will be substantially replaced.
Posted: 2016-06-04 @ 1:23am PT
I thought soulless machines already made pop music.