Yacht fed their old music to the machine and got a killer new album

Yacht fed their old music to the machine and got a killer new album(arstechnica.com)

71 points by pimienta 6 years ago | 11 comments

yowlingcat 6 years ago |

This is both very cool and goes along the lines of the way I imagined deep learning would advance the state of the art in songwriting. YACHT's got a special place in my heart for "Psychic City" which sounds like a lost child of Talking Heads, not to mention all the other great music they've made.

Secondly, it's interesting how they used the tech to stretch their existing material: they basically notated their existing corpus of work (82 songs) into MIDI, chopped it by part into loops, fed that into the MusicVae deep learning model, got out thousands of loops, and reintegrated the output material into a full album. This is probably the way things can and should go.

Exciting times! I want to have a go with this tech with my own music.

amelius 6 years ago | |

How about a system where music is generated by a computer and a person listens to it while a functional MRI scanner provides feedback to the algorithm?

yowlingcat 6 years ago | | |

I remember thinking about such a system a long time ago. In a theoretical sense, that seems really cool, but I'm not sure if the hardware is quite there yet. From what I can recall, practical brain-computer interfaces are still in their infancy, and not reliable enough to use even for simple binary interfaces. Otherwise, that would basically make a closed circuit for the palette generation part of this -- just let the deep learning net come up with material and approve/deny it in realtime until you've got the material you need.

Fjolsvith 6 years ago | | |

An echo chamber?

sharkweek 6 years ago |

I love when things related to my non-tech fandom of some random entity show up on HN.

YACHT is fantastic and explores a lot of fascinating ideas both lyrically and musically (try “I thought the future would be cooler” for a catchy dance tune with a depressingly great dystopian theme, never has our impending doom sounded so good. For the less cynical, try “Shanghai-LA”).

dejawu 6 years ago |

My understanding is that even humans don't fully know why neural nets are doing exactly what they're doing. If we were indeed able to just press a button and have an AI give us a finished track (which is not what YACHT did here), who should take credit for the track? I certainly don't think the human who trained the net deserves all the credit, right? That feels like if an artist's otherwise absent father gave them Kid Pix in their childhood and then tried to take credit for their creative output thirty years later.

andreyk 6 years ago | |

This has been an open question for a while, and is in the process of being figures out. Some fun examples:

- "Generative art and UK copyright law - good news", http://mcld.co.uk/blog/2011/generative-art-and-uk-copyright-...

- "Template License and Collaboration Agreements for AI Art" https://clinic.cyber.harvard.edu/2019/02/04/template-license...

TheOtherHobbes 6 years ago | | |

There's a more complicated problem. Generative art explores a space of possible outputs, and in some cases some of those outputs will be exactly like existing work.

If a melody generator produces a tune that's exactly like $famous_tune, is that plagiarism? What if it produces $famous_tune among thousands of other tunes? Does it make a difference if $famous_tune was in the training set, but there isn't any evidence of over-fitting?

What if you argue that the generative space includes $famous_tune, and therefore in some sense all output is influenced by it, making all output a related or derivative work - even if the space is astronomically huge?

And so on. I suspect this is going to keep lawyers very busy.

tortasaur 6 years ago | |

The optimist in me wants to believe these quandaries will force us to reevaluate the assumptions behind our copyright system.