Видео

Be nice to Ruth!

Ha ha. You don't like the no no cat? I can remove from the mix. :)

It's funny, but I guess I just want to hear only what the AI has created. You obviously don't have to though. But it would be very cool if all the separate tracks were on a CD (:

Hey Bob Sturm can you please remove the noise from The Arian

Matlab... python?

MATLAB

Not this one! This is a real traditional tune. :)

Wonderful! That was real quick. :) I love the chord progression.

Thanks. No, it is all from synthetic instruments I mixed in Reaper. The sound you hear is from a particularly famous viral video. Artistic license. :) I didn't have to pitch shift it either.

Why did you put it there

Thanks for listening! Yes, the "no no no cat" is in there. I didn't even have to pitch shift it either.

Hi Kram. While this piece was produced by the original folk-rnn, we have constructed several other versions since then and have been studying their properties. Have you heard this? highnoongmt.wordpress.com/2016/12/24/taking-a-christmas-carol-toward-the-dodecaphonic-by-derp-learning/ We also have several new pieces coming up at this workshop: highnoongmt.wordpress.com/2017/02/28/mar-25-folk-music-composed-by-a-computer-inside-out-festival/ Then there is our concert coming up, which will feature new stuff too: highnoongmt.wordpress.com/2017/02/26/partnerships-tickets-tue-23-may-2017-at-1900-eventbrite/

Unfortunately it won't be easy for me to get to Great Britain but I'm glad to hear/read that you are also working on new stuff :) I like both of those compositions drawn from Christmas music although I gotta say neither seem particularly Christmassy. - Which isn't necessarily a bad thing. They actually sound a good deal more interesting than most Christmas music.

Thanks Kram. In this case, "best" means the posterior probability of the classification is at least 0.9999. The point of this experiment was to show how misleading "classification accuracy" is for music genre recognition. Since that time, things haven't changed. Most researchers are still using classification experiments to measure success in "genre recognition".

I see. I think I agree that perhaps researchers put a bit too much merit in such ideas. Though with better architectures, the model could probably be improved. Notably on the image generation side of things, huge improvements were made just recently, using things like priors that tell the system how natural images actually look like, or General Advisorial Networks (GANs) which learn to tell it whether what it produces is actually what it thinks it produces. I'd imagine such would greatly improve the genre classifier too, especially with respect to robustness against otherwise very easily generated match outliers. (By which I don't mean that they are rare but rather that they are far away from the actual examples they are trained on. - if the system you picked actually is trained like a NN anyway... Though I'd imagine such a task to be precisely where NNs shine, compared to hand-corded techniques)

Hello!

+Károly Zsolnai-Fehér (Two Minute Papers) Thanks! Keep in mind that a significant amount of magic comes from the orchestration, which is entirely due to me. Also, I selected these pieces from over 100,000 ABC outputs of the system because I like how much they "fail" to represent the training material. :)

Hey love both of your work Bob Sturm and Two Minute Papers. I have checked out most of your videos Károly they're all fantastic. Bob I have a question: I have recently graduated and have been self-teaching some ML for the past month I have a few of the necessary basics down and am now looking into neural networks. More specifically I am going over the CS224 course on RNNs. Would you recommend any books/literature on neural networks which tackle the intersection of neural networks and music generation? Many thanks to both of you for your hard work

Starts off sort of normally then goes full Zappa.