|
Where did it say in the "let's create software" contract that the only acceptable goal was "revolutions or fundamental changes in our relationship to the world" ? As I said in another comment in this thread, I understood the role of computers to be helping humans with tasks they wanted to do by doing things that computers were good at. If that happens to include revolutions or fundamental changes, fine, but it definitely includes a lot of other things too. You are right that the particular examples of audio software capabilities do not in and of themselves bring anything in particular to music. But the timestretching stuff has totally changed how huge numbers of people now make music, because they can work with audio that is in the wrong key and/or at the wrong tempo, without really thinking about it. Do I think that this results in an aesthetic leap forward for music itself? I do not. In fact, probably the opposite in many senses. But that is true of so many human technologies, not just software. Some people would even argue that the advent/invention of well tempered tuning (and the concomittant move away from just intonation) hurt music in the west, and that was just as much the result of "sufficient engineering hours" as anything in the software world. Also, just to correct you, 20 years ago, I guarantee you that nobody, absolutely nobody, believed that you could ever do polyphonic note editing. When Melodyne first demonstrated it, most people who knew anything about this just had their jaws hit the floor. It was absolutely not an "utterly conceivable consequence", even though in retrospect of course it now all seems quite obvious. |
https://www.jstor.org/stable/3679550?seq=1
"The real power of a neural net is its ability to compute solutions for distributed representations. In most cases, the solutions for these complex cases are not obvious. The pitch class representation of pitch is a local rather than a distributed one. In this case a possible solution for the chord classification problem is apparent without the use of a learning algorithm. A net containing 36 hidden units, one representing each of the possible major, minor, and diminished triads, could be constructed so as to map chords to chord types. Thus our interest in using a pitch class representation was not to find this obvious solution, but to find a solution which used a minimum number of hidden units. We hypothesized that three hidden units would be adequate and that the hidden units would form concepts of the intervals found in triads: i.e., major third, minor third, perfect fifth, and diminished fifth.
Each pitch-class net used 12 input units to represent the 12 pitches of the chromatic scale and 3 Output units to represent chord type. The number of hidden units and the values of the learning parameters are summarized in Table 1 for each of the eight pitch class nets discussed. Net 1 had an adjacent layer architecture as shown in Fig. 2 and three hidden units. It identified 25 percent of the chords after more than 11,000 learning epochs. When a fully connected architecture was used in conjunction with three hidden units in Net2, 72 percent of the chords were identified after 2,800 learning epochs. "
https://secure.aes.org/forum/pubs/conventions/?elib=11400