Ah well, yes, actually there is plenty of scope for audio white space in the DarwinTunes GP representation. Loops do tend to be rather busy though - could be a consequence of either selection or biases in the representation. Thanks for your interest!
exactly, and one more thing - the selection is pretty close natural selection because multiple raters in isolation provide the feedback.
Obviously the music stays fairly saccharine (but now less so than I had imagined) but with enough people you could speciate/split into sub-populations and get more edgy (literally!) music.
But there's more than one listener - they are generally not in contact with each other - so the selection is not really directional in the sense of one person breeding dogs or roses. It's pretty close to a natural selection environment.
Well yes, it's easy to knock out some evo-music toy and move on. We were less interested in the toy and more interested in what happens when you let music evolve in an environment of as many human listeners as we can muster.
Yeah that 2009 post was my doing - this new one is "organic". The main thing we've added since then is the re-rating of the loops to assess the increase or otherwise of musical appeal through the generations, and to investigate why it slowed down.
No way are we billing it as machine generated music - the PNAS paper title and website tagline are pretty clear about the role of the consumer/listener.
We thought it would be interesting to test just how far listener-selection can get. Seems like quite far, but in its current state it's obviously not music that will provoke a particularly profound response. This tallies with your comments about the music industry.
This one (the first image in the Wired article) seems to be exactly the same dimensions as the image tiles - zoom out until you see different "vintage" images and you'll see what I mean. Could just be an artifact. The others look real though.
While on the subject of CSI, don't you love the way they just happen to have a custom spinny-roundy 3D database for just about anything you can care to think of.
That, and the rather optimistic natural language queries they type into the aforementioned databases.
I switched from MS Excel to OOcalc for some analysis I'm doing at the moment. Excel was slow, made huge files and the bar charts were hideously shaded by default.
We really need a video of the shared "retina" thing, which is of course not in the original critterding video. However, this sounds cool enough for me to try it at home later.
On second thoughts, maybe they didn't want to "produce hallucinations" in millions of Slashdotters...
Thanks for the comments. I'll take them one by one (while I wait for the algorithm to tick over to 250 generations).
1. to give a fair comparison with the hand-picked better sounding loops given in all the subsequent "tasters", the time=zero loops are also hand-picked. Rest assured that most of them sounded pretty horrific. Yes we did set a minimum amount of complexity (I think it was at least 8 different "tracks") in the initial Adam and Eve, but then let them evolve under no selection for a long time.
2. yes we have to keep it short so that rating can happen in this lifetime:-) I have put several tracks together for my own projects (just me doing the ratings, and using pre-recorded samples as well as evolved synths) - here's the best example. I'll probably put something from DarwinTunes together over the holidays (prob using consistently highly rated loops from the slashdot surge)
3. there are no constraints on harmonies or anything, however the "palette" of notes is defined once (all evolved from random) and then the notes are picked from the palette. Mutations to the palette are going to be rare (because defining it takes many fewer "genes" than defining all the music) - hence the good agreement between loops.
4. yep, there have been scientific studies showing herd behaviour in music "selection". The rest of what you say can't be denied, and that's what we're interested in and why we're doing the experiment.
5. no it's real - but I didn't know about the tenori-on, so thanks for the heads up on that:-)
darwintunes.org has no ads at all - it's an academic experiment website, it would be inappropriate. Even my slightly more commercial evolectronica.com doesn't have ads. From my experience that would just be a colossal waste of time and turn people away.
What newscientist.com (or perhaps DNS hackers) do with their ads is nothing to do with me!
Our goal here is to look in detail at the evolutionary dynamics and mechanisms, as well as just answering the basic question "does it still work if loads of people provide the fitness ratings?"
Slashdot Mirror
Ah well, yes, actually there is plenty of scope for audio white space in the DarwinTunes GP representation. Loops do tend to be rather busy though - could be a consequence of either selection or biases in the representation. Thanks for your interest!
exactly, and one more thing - the selection is pretty close natural selection because multiple raters in isolation provide the feedback.
Obviously the music stays fairly saccharine (but now less so than I had imagined) but with enough people you could speciate/split into sub-populations and get more edgy (literally!) music.
But there's more than one listener - they are generally not in contact with each other - so the selection is not really directional in the sense of one person breeding dogs or roses. It's pretty close to a natural selection environment.
Thanks, I had a stab at it (you meant this comment?)
Well yes, it's easy to knock out some evo-music toy and move on. We were less interested in the toy and more interested in what happens when you let music evolve in an environment of as many human listeners as we can muster.
Yeah that 2009 post was my doing - this new one is "organic". The main thing we've added since then is the re-rating of the loops to assess the increase or otherwise of musical appeal through the generations, and to investigate why it slowed down.
No way are we billing it as machine generated music - the PNAS paper title and website tagline are pretty clear about the role of the consumer/listener.
We thought it would be interesting to test just how far listener-selection can get. Seems like quite far, but in its current state it's obviously not music that will provoke a particularly profound response. This tallies with your comments about the music industry.
Yeah, you just follow the roads and find all kinds of weird stuff: http://g.co/maps/xwk8j
This one (the first image in the Wired article) seems to be exactly the same dimensions as the image tiles - zoom out until you see different "vintage" images and you'll see what I mean. Could just be an artifact. The others look real though.
This is a nice tool for viewing the cross section (altitude) of an arbitrary path drawn on a google map:
http://www.geocontext.org/publ/2010/04/profiler/en/
Dennis Ritchie
Totally agree, apart from the beating to death bit!
C'mon gman003. You can watch the whole video, surely, before posting?
The heating effect is important here too.
While on the subject of CSI, don't you love the way they just happen to have a custom spinny-roundy 3D database for just about anything you can care to think of.
That, and the rather optimistic natural language queries they type into the aforementioned databases.
Well personally I'd wait and see if they "do no evil" with regard to their blatantly obvious software patent for using geolocation info to target ads.
think he means bibtex (a LaTeX bibliography tool/format)
I switched from MS Excel to OOcalc for some analysis I'm doing at the moment. Excel was slow, made huge files and the bar charts were hideously shaded by default.
We really need a video of the shared "retina" thing, which is of course not in the original critterding video. However, this sounds cool enough for me to try it at home later.
On second thoughts, maybe they didn't want to "produce hallucinations" in millions of Slashdotters...
Thanks for the comments. I'll take them one by one (while I wait for the algorithm to tick over to 250 generations).
1. to give a fair comparison with the hand-picked better sounding loops given in all the subsequent "tasters", the time=zero loops are also hand-picked. Rest assured that most of them sounded pretty horrific. Yes we did set a minimum amount of complexity (I think it was at least 8 different "tracks") in the initial Adam and Eve, but then let them evolve under no selection for a long time.
2. yes we have to keep it short so that rating can happen in this lifetime :-) I have put several tracks together for my own projects (just me doing the ratings, and using pre-recorded samples as well as evolved synths) - here's the best example. I'll probably put something from DarwinTunes together over the holidays (prob using consistently highly rated loops from the slashdot surge)
3. there are no constraints on harmonies or anything, however the "palette" of notes is defined once (all evolved from random) and then the notes are picked from the palette. Mutations to the palette are going to be rare (because defining it takes many fewer "genes" than defining all the music) - hence the good agreement between loops.
4. yep, there have been scientific studies showing herd behaviour in music "selection". The rest of what you say can't be denied, and that's what we're interested in and why we're doing the experiment.
5. no it's real - but I didn't know about the tenori-on, so thanks for the heads up on that :-)
nearly at 250 generations now...
For the last time...
darwintunes.org has no ads at all - it's an academic experiment website, it would be inappropriate. Even my slightly more commercial evolectronica.com doesn't have ads. From my experience that would just be a colossal waste of time and turn people away.
What newscientist.com (or perhaps DNS hackers) do with their ads is nothing to do with me!
And finally, it was Slashdot who made the New Scientist link look like the main link for the article, not me! See my original submission:
http://slashdot.org/submission/1136438/Music-by-natural-selection
Some people are having problems with the New Scientist link. AardvarkCelery has some info in the post currently below this
http://entertainment.slashdot.org/comments.pl?sid=1485740&cid=30521010
It's not completely novel, no. Google weren't the first to do web search either ;-)
An incomplete list of related work is at
http://en.wikipedia.org/wiki/Evolutionary_music
Our goal here is to look in detail at the evolutionary dynamics and mechanisms, as well as just answering the basic question "does it still work if loads of people provide the fitness ratings?"
Sorry about that.
Although note that my original submission had the New Scientist link in a more obvious place:
http://slashdot.org/submission/1136438/Music-by-natural-selection
It did seem odd to me why the editor changed the links that way. Conspiracy?
Thanks for the info. I'll mention that to the journalist at New Scientist. I've got adblock so I don't see it.
It must be a DNS hack because I'm really not seeing it! Can you give some more info?