Would the Algorithm of Fugue end with A B C?

April 11, 2006 4:35 AM

Douglas Hofstadter says, "What troubles me is the notion that things that touch me at my deepest core -- pieces of music most of all, which I have always taken as direct soul-to-soul messages -- might be effectively produced by mechanisms thousands if not millions of times simpler than the intricate biological machinery that gives rise to a human soul.". That was prompted by his reception to the output of David Cope's project Experiments in Musical Intelligence.
posted by Gyan (22 comments total)

A few responses to this:

1) In order for music to be expressive, your own imaginative imput is required to link that sound to emotional feelings/landscapes etc... So any piece of music cannot by itself capture the human soul.

2) If a human performs a piece he will imbue it with all the tiny details of human expression not specified in the score.

3) This kind of thing is why originality/non-cliche is so important to composers. Who wants to hear another piece by Bach? We've got enough already.

4) That said, there's nothing so wrong about human-machine collaboration. It's just like exploring any other medium.
posted by leibniz at 5:05 AM on April 11, 2006

leibniz : "your own imaginative imput is required to link that sound to emotional feelings"

That's implicitly given. The point presumably is that most efforts so far, by artificial engines, that engaged the imaginative input of listeners have been found wanting. This gave rise to a conviction in Hofstadter that the sophistication involved requires something as complex as the biological substrate of the brain.

"If a human performs a piece he will imbue it with all the tiny details of human expression not specified in the score."

Any reason that this also can't be demystified?
posted by Gyan at 5:19 AM on April 11, 2006

Can anyone provide a mirror of the MP3s, preferably coral cached? I'm keen to hear them.
posted by wilberforce at 5:22 AM on April 11, 2006

Right click, save as... worked for me. The in-browser quicktime plug-in kept breaking.
posted by Roger Dodger at 5:31 AM on April 11, 2006

A few things occur to me. One is that every age has had human forgers and copiers who could imitate, say, Bach, Beethoven, or Mozart with great success; but still, their memory fades away in a relatively short time. History and memory aren't kind to imitators.

Another is that we're talking about a program that's been in development, according to Hofstadter, since early 80s; but what we get to hear (or would, if AI guy could master the idea of a torrent) is a few tens of pieces. Now, this is NOT because EMI lingers and suffers over the composition of a piece for some weeks or months at a time. From what I understand about it, once you set it up, it runs, and spits out a "work" in pretty short order. So I have kind of a suspicion that what we're hearing is the very greatest successes of EMI, carefully culled by Cope. And that is where the human element has crept in again. Again, I don't know, because I can't get my hands on it, but I suspect that EMI is capable of turning out completely insipid trash at one turn, and Chopin-like dances at another and never knows the difference.

I'm not mystical about human consciousness, though it's still mysterious and wonderful to me; I do believe we'll eventually see machines and algorithms that think like we do and have comparable depth. I don't think we're anywhere close yet, and I guess that what we have here is nothing more than a "typing monkeys" phenomenon.

I still think it's cool, and I'd love to be working on a project like that. I just hope I wouldn't get sucked into making the kind of hyperbolic claims that usually accompany such a project.
posted by Wolfdog at 5:38 AM on April 11, 2006

Yeah, that does it - clicking on them in the browser kept giving me a Too Many Users error. Thanks.

Interesting.
posted by wilberforce at 5:38 AM on April 11, 2006

When I read Godel, Escher, Bach I felt from the outset that Hofstadter was overstating his case. He drags everything but the kitchen sink into his argument for the specialness of intelligence while quietly tiptoeing around information theory, which speaks to the problem much more directly.

It's amusing that Hofstadter is figuring it out in this way at this late date. What's interesting is that he actually does get the implication -- that if he was wrong about music (and I thoroughly believe he is) then he was most likely wrong about intelligence too. As much as he claims to be saddened by the result, it's a breath of fresh air that he is admitting his mistake rather than holding his hands over his ears and shouting "LA LA LA" as so many other people do.
posted by localroger at 5:41 AM on April 11, 2006

Does Cope say whether his Mahler/Bach/Beethoven/Chopin/Joplin-like works are:
a) typical examples of complete works given only a random seed (i.e. his Bach software produces reliable Bach-like works), or
b) the best sounding examples from a large number of runs of his software, or
c) the successive output of an interactive program where the user requests, say, the next N measures, and elects to keep or discard each suggestion?
posted by Silki at 5:46 AM on April 11, 2006

Who wants to hear another piece by Bach?

me.
posted by grumblebee at 5:52 AM on April 11, 2006

I've been reading a book on creativity and it discusses Cope in a chapter on artificial creativity and points out:

(1) Obviously the machine doesn't generate its own code, which is created by Cope, a trained classical musician. While most artists have a sense of autonomy and develoment, here it might be more accurate to say that Cope has designed a tool, rather than to say a machine has created classical music. The book talks about this in the same chapter that it talks about the artificial artist, also created by a trained artist-programmer, which many people here likened to a photoshop filter.

(2) Selection and evaluation are parts of creativity, but the program spits out a large volume of songs and Cope goes and picks the ones that sound the most accurate. The program doesn't "know" what it's going for, i.e. produces many songs that are not Bach-like, beautiful, etc.

(3) Obviously Bach produced the code and the program just tries to derive something from that. For example, while it has an easy time with Bach (classicism, with its emphasis on geometry), the program only gets one in 80 songs that sound remotely like beethoven (romantic music). I think this program is less about AI and more about our western myths about formalism: the idea that, say, the way someone plays a song or the social context don't matter to the "inherent beauty" of a piece. Consider a program created to make hip hop or jazz.
posted by kensanway at 6:05 AM on April 11, 2006

I just finished reading On the Seeming Paradox of Mechanizing Creativity by Hofstadter. I couldn't find a link to the article, so here's (IMO) a salient clip:

"Each of us--even the Mozarts among us--exhibits a 'cognitive style' that in essence defines the ruts we are permanently caught in.

"Far from being a tragic flaw, this is what makes us interesting to each other. If we limit ourselves to thinking about music, for instance, each composer exhibits a 'cognitive style' in that domain--a musical style. Do we take it as a sign of weakness that Mozart did not have the power to break out of his 'Mozart rut' and anticipate the patterns of Chopin? And is it because he lacked spark that Chopin could not see his way to inventing the subtle harmonic ploys of Maurice Ravel? And from the fact that in 'Bolero' Ravel does not carry the idea of pseudo-sphexish music to the intoxicating extreme that Steve Reich has, should we conclude that Ravel was less than magical?

"On the contrary, we celebrate individual styles, rather than seeing them negatively, as proofs of inner limits. What in fact is curious is that those people who are able to put on or take off styles in the manner of a chameleon seem to have no style of their own and are simply saloon performers, amusing imitators. We accord greatness to those people whose 'limitations', if that is how you want to look at it, are the most apparent, the most blatant. If you are familiar with his style, you can recognize music by Maurice Ravel any time. He is powerful because he is so recognizable, because he trapped in that inimitable 'Ravel rut'. Even if Mozart had jumped that far out of his Mozart system, he still would have been trapped inside the Ravel system. You simply can't jump infinitely far!

"The point is that Mozart and Ravel, and you and I, are all highly antisphexish, but not perfectly so, and it is at that fuzzy boundary where we can no longer quite maintain the self-watching to a high degree of reliability that our own individual styles, characters, begin to emerge to the world."

Hofstadter set out two goals for a program attempting to mechanize creativity: 1) Flexible perception and 2) Self-watching. Cope seems like he recognizes that his program falls short of those lofty goals; from the Experiments link: "Ultimately, the computer is just a tool with which we extend our minds."
posted by carsonb at 6:11 AM on April 11, 2006

If this program, as the piece says, is simply chopping up and rearranging compositions, then real creativity (from which these derived works come) still belongs to the human soul, don't you think?
posted by nofundy at 6:12 AM on April 11, 2006

information theory, which speaks to the problem much more directly.

Hell yes! It seems so clear that this direction will be more fruitful than attempting to figure out general principles of intelligence through introspection, as so many General AIists do, or expecting "emergence" to birth some qualitative change without any model that predicts such a change.

With an info-theoretic framework, you inherently employ some quantitative measure of the effectiveness of the algorithm. You have some direction. It forces you to specify what you want out of the algorithm, but is a general enough theory to avoid special-casing everything.

This isn't to say that more 'philsophical' work might not be fruitful, I believe it will, but any such work will still be addressable by info-theoretic analysis.
posted by sonofsamiam at 6:47 AM on April 11, 2006

Cope's work, according to an AI researcher I work with, has passed the "musical Turing Test," insofar as one can play a Bach fuge and one produced by the system and a "lay" music listener (i.e., not someone who knows every extant Bach work) will claim that they're both by Bach (and therefore, created by humans).

This is not a unrestricted turing test however, as people up-thread have pointed out: Cope *selects* good works (works that match his understanding of Bach, and things like "good" melody and harmony), so therefore this is not an un-restricted Turing Test, but something much more akin to Eliza. If one could get at the full stream of system outputs, we'd see, I'm afraid, that many many pieces would not pass such a test.
posted by zpousman at 6:48 AM on April 11, 2006

I did a Cope test along with a room full of fellow composers (i.e. classically trained musicians who know the repertoire but obviously don't know every work by Bach etc.) We found that we second guessed some of the examples somewhat, but found that the computer generated Bach fooled us better than the computer generated Chopin etc. We were fooled some of the time. We were also unimpressed that we were fooled, and I think that generally we all felt that this was an interesting whimsy rather than the profound statement about creativity that Hofstatter seems to think that it is. Mind you this was a room full of composers, a group of people who know only too well that the inspirational side and the 'computational' (or numerical) side of musical creativity are one and the same thing.
posted by ob at 7:19 AM on April 11, 2006

Hofstadter's writing about this seems out of character to me, considering how he has talked about Racter. In fact, upon looking it up in Fluid Concepts and Creative Analogies, I see what he wrote there about Racter is nearly identical to what I wrote above; I think I plagiarized.
posted by Wolfdog at 7:39 AM on April 11, 2006

Interesting.
posted by Smedleyman at 9:15 AM on April 11, 2006

Great thought-provoking link, Gyan.
posted by storybored at 10:51 AM on April 11, 2006

The Beethoven sonata movement was both amazing and a recitation of a string of Beethoven cliches.

I agree Hofstadter is overstating. The programs might be simple in comparison to intelligence. Because they required input from human intelligence on two planes: the creativity they mimic and the ability to codify the mimicry.
posted by 3.2.3 at 11:22 AM on April 11, 2006

Now if we could write a generative music program that uses the Oblique Strategies...
posted by geekhorde at 12:53 PM on April 11, 2006

sonofsamiam & localroger: information theory, which speaks to the problem much more directly.

Can you elaborate on this? What is the argument that information theory would make here?
posted by jouke at 8:16 PM on April 11, 2006

but I suspect that EMI is capable of turning out completely insipid trash at one turn, and Chopin-like dances at another and never knows the difference.

Er, I'm a musician, and some days I can turn out completely insipid trash, and other days unusually inspired delights -- and I usually can't tell the difference either, at least not right away.

Heck, once I sat down to write an "80's style" song, got all the chords set up, recorded 'em, then realized it all sounded familiar -- and when I played the chords back with the melody from Quiet Riot's "We're not gonna take it", it fit perfectly. Oops.

Left as an exercise to the reader: is "we're not gonna take it" completely insipid trash, or an unusually inspired delight?
posted by davejay at 11:08 AM on April 12, 2006