Skip to main content

Creepy Next Generation AI

Posted

As I'm going deeper into exploring the next generation of AI for music applications, I come across more and more really creepy stuff. Will machines eventually take over the world?

Check out this: (http://deepdreamgenerator.com)

It is difficult to understand for anyone not familiar with neural networks, but these nets can be trained with memories and recognize these later. The pictures show what's going on in the computer when it looks at a picture. This is an example where it looks at a lemon tree (leaves, lemons):

Now I can't stop thinking what might happen if we trained such a net with 1 million songs and look into it what happens when it listens to one or two bars of music you give it ...


Wed, 2016-05-25 - 14:30 Permalink

Yeah. If only we had a database with 1 million songs in MIDI format ;-)

The networks are extremely useful, provided you have enough material to train them.

Tue, 2016-06-07 - 15:06 Permalink

just google "midi collection reddit" and you should find enough songs to train it ;) 

Tue, 2016-06-07 - 15:09 Permalink

deep dream is more than a toy and there are already tools based on this you should seriously give it a try. some of the most interesting tools will come from this tech in the next years

Tue, 2016-06-07 - 17:47 Permalink

Yep. It's definitely not a toy. It's serious technology. No idea yet however, if it will work with MIDI the same way. Experimentation is the only way to find out.

Thanks for the reddit link. I'm afraid Cognitone wouldn't have the permission to use this MIDI material for commercial purposes. There are other collections created for research purposes that we might use.

Wed, 2016-06-08 - 10:55 Permalink

Thanks for the link. Coincidently, I got my TensorFlow cluster up and running already ;-)

The philosophy of Synfire is to intelligently assist a composer, rather than compose some random song on its own. The latter is probably a lot of fun at first, but canned music recipes get predictable and boring quickly. If anything, the goal is to help the composer come up with new figures and textures more easily, by deriving them from his/her own examples.

Also keep in mind this kind of fundamental research takes a lot of trial and error. We can only devote so much time each year to research. Whether this will eventually make it into a feature is not clear.

Thu, 2016-06-16 - 15:20 Permalink

So lets get this google tpu chip in Synfire! I trust you and it not to try to tale over the world and exterminate humanity. 

"I'll be back"!