On Mon, 7 Jan 2002, Bob Miller wrote:
> > 3. run the music files through a feature extraction engine
> > to get a coded representation of the song (this is the hard part)
> 
> This is the hard part because it's a hard problem, not because the CPU
> power costs money.
>
Yes it is a hard problem, however it's also a fairly popular one.

When i whipped out my email this afternoon I was thinking of something
along the lines of a kohonen map for each input vector with a genetic
algorithm to generate new candidates. I then realized that it would be
fairly difficult to run the mapping backward, which suggested to me that
the mapping should be made using the bidirectional associative memory.

bidirectional associative memory is a generalization of backpropagation
networks which are used in many contexts to extract structure from time
series data. But for a practical application it would probably be
necessary to do a fair amount of preprocessing on the input files.

An anecdotal example with radical oversimplification:

1.Take a bunch of relatively similar tunes (same key, same time signature)
transcribe each one as sheet music on a clear plastic sheet. (assume they
are all one page long)
lay all the transparencies on top of each other, you should have something
that looks like several rows of fairly jagged blotches.

2Using some random method(this is a gross oversimplifcation)
produce new transparencies that look sort of like the others.
3Compare them to the transparencies you produced in the first place.
4Throw away ones that are too different from the average of the inputs.
do 2,3,4 again and again, slowly modifying your production method until
you are usually producing  something that looks sort of like one of the
original inputs, and pretty much fits inside the boundaries of the sum of
inputs.

take some of the new transparencies you've produced, and have someone
sight read them.

This should produce something almost but not quite entirely unlike tea.
--share and enjoy
 
> If/when you have an algorithm that works, getting the CPUs will be
> easy.
> 
Define works.

It's kind of a fun idea, but the cost of implementation probably outweighs
the rewards of doing so. Unless of course I can get K-tel to sponsor a
research project.

Reply via email to