Xref: utzoo comp.music:928 comp.ai.neural-nets:1512
Path: utzoo!utgpu!news-server.csri.toronto.edu!clyde.concordia.ca!uunet!cs.utexas.edu!uwm.edu!csd4.csd.uwm.edu!markh
From: markh@csd4.csd.uwm.edu (Mark William Hopkins)
Newsgroups: comp.music,comp.ai.neural-nets
Subject: Re: Music by Kohonen's NN
Keywords: neural nets, music
Message-ID: <2940@uwm.edu>
Date: 18 Mar 90 02:46:25 GMT
References: <1990Mar6.200147.21195@cec1.wustl.edu> <76@nrl-cmf.UUCP>
Sender: news@uwm.edu
Reply-To: markh@csd4.csd.uwm.edu (Mark William Hopkins)
Followup-To: comp.music
Organization: University of Wisconsin-Milwaukee
Lines: 25

In article <76@nrl-cmf.UUCP> tedwards@cmsun.UUCP (Thomas Edwards) writes:
>
>The exact algorithm is not expected to be published exactly because there is
>much commercial interest in his technique....
>...had the typical computer generated lack of overall theme, and there were
>some times when I just couldn't help thinking "That was the wrong note for
>this song."  However, it did seem to create often wonderfully powerful
>repetitons of strong note sequences.
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Sounds to me like the net encoded music using something analogous to
"Wickelphones".  If you look for combinations in limited contexts (such as
all 3 note sequences), you can abstract out the explicit time ordering
while still preserving much of the identity of the music.  A musical
piece would be identified by which "Wickelpitches" occurred in it, but
all the global structural information would be lost.  Hence the behavior
you describe.

It is interesting to note that simply by compiling a table of probabilities
for, say, 5 character sequences in English text, you can automatically
generate text that is remarkably English-like using those probabilities
and a random number generator.

If you have time (and storage space and generous amounts of text), you may
want to try it.