Sound: Advanced Topics I (GR6610, Fall 2023) week 5

Short history of music AI

video of this class

This starts our "AI unit" of the class. I gave a brief history of the development of different AI approaches along with a discussion of how neural-nets/machine learning works.

Links

These are the web pages and videos I showed in class. There's many, many more web-page resources if you look. Just google terms like "genetic algorithm", "neural networks" (Rumelhart and McClelland, Geoffrey Hinton), "connectionism", "symbolic AI" (Marvin Minsky). You will find a lot.

on-line verion of the ELIZA program -- developed at MIT by Joseph Weizenbaum, this was one of the first 'natural language' uses of computers. Very simple, but people enjoyed fooling around with it.
Gödel, Escher, Bach: an Eternal Golden Braid -- the award-winning book written by Douglas Hofstadter that captured the spirit (and techniques) of the symbolic approach to AI.
IBM researcher Kemal Ebcioglu's symbolic-AI program CHORALE -- this created four part contrapuntal harmonies in the style of Bach. There's a weird young Columbia professor commenting on the program part-way through. The video was produced by IBM to showcase some of the research being done at the IBM-Watson Research Center.
EMI Vivaldi and EMI Bach -- two pieces created by David Cope's "Experiments in Musical Intelligence" project. Symbolic AI approach.
The Elthar Program -- an article about my doctoral dissertation work published in Perspectives of New Music (Winter, 1989)
Rough Raga Riffs -- the "heavy metal raga guitar" rule-system piece I played in class.
Almost Real -- the fake Irish folk ensemble rule-system piece.
various neural-net topologies -- different ways of connecting the nodes of a neural net.
Google's tensor-flow/magenta system -- the 'continuation' of Debussy's Claire de Lune (massive neural net).
tensor-flow/magenta sound synthesis-combination project -- promotion video showing the Google system creating 'hybrid' sounds (massive neural net).

Class Downloads

iglesia-net-js.zip -- this is the simple neural-net maxpatch I showed that can learn to disambiguate "low" sounds from "high" sounds (if the data inputs are imagined as FFT frequency bins). It's in a folder called "iglesia-net" because it's a revised version of a revision that former CMC student Daniel Iglesia made of an original maxpatch I wrote. Dan's version was much better than mine.
trump-o-matic.zip -- this is the silly "Donald Trump Quote Generator" I did to show a really basic approach to "AI".