Oh Yes I Kan

homeblogtwitterthingiverse



I have just read a book that made sense of a great many things: Architectural Ornament: Banishment & Return, by Brent C. Brolin. I now know the name of my enemy, and knowing the name of one's enemy is a very good thing.

The name of my enemy is Kant's concept of genius.

Genius, according to Kant, is the ability to create new things. Crucially, this ability can not be described scientifically. No rule can be given that produces the things that genius produces. It can not be learnt.

This is a very comforting idea, and has proven popular. People thought to have this quality of genius are given high social status. It fills a much needed gap in people's minds with the idea that worthwhile artistic or even scientific expression is something that one brings forth from within. This is a gap that might otherwise be filled with skill acquired by study, and by trying things out and mostly failing.

Furthermore, geniuses don't fully understand their own abilities. Pat pat. Don't try to explain your process. We'll love you more if you say it is magic.


My counter-concept is that the key ingredient of "genius" is just to be able to recognize success. To select accurately. Beyond this, it is simply a matter of generate and test. You get better at generating with practice. Anyone with the ability to appreciate good art also has the ability to create it by this method, given sufficient time.

Evolution by natural selection is a good example of this. God is the ultimate Kantian genius, and evolution shows that He is unnecessary. Many people find this discomforting, it at once destroys God and denies their own access to transcendent genius.


If we turn this concept in on itself there is a most wonderful idea to arrive at. It might at first seem that Kant's concept of genius precludes the construction of an intelligent machine. However we can get around this simply by constructing a machine that we can not understand. Genius can beget genius. Thus: the Artificial Neural Network. Computer scientists are easy prey for this. Their time could be better spent getting a haircut and learning some real statistics.

This explains the strange feeling I have had that my experiments in music generation are a little bit monstrous. They are monstrous because I understand them. If I were to use something more mysterious, such as a cellular automaton, neural network, or genetic algorithm, the problem would not arise. You will find a lot of this kind of thing, the results are usually terrible as music.

David Cope's approach to music generation by pastiche is similarly monstrous.


I expect I shall have more to say on this in future.


10/9/09: Jiri notes the "neats vs scruffies" debate in AI. In my experience, scruffy code motivated by neat concepts often beats pure neatness or pure scruff. Scruff gets the wrong answer, and neatness is too slow. Computational efficiency is an issue I haven't addressed above.




[æ]