ajb@itivax.iti.org (Al Boehnlein) (02/08/90)
Has any one ever though about using a neural net to program a neural net. It would seem that a neural net might be just the thing to use to determine what weights to change, and how much to change them, when adding a new example to a network. ajb
andrew@dtg.nsc.com (Lord Snooty @ The Giant Poisoned Electric Head ) (02/08/90)
In article <4914@itivax.iti.org>, ajb@itivax.iti.org (Al Boehnlein) writes: > Has any one ever though about using a neural net to > program a neural net. It would seem that a neural net > might be just the thing to use to determine what weights > to change, and how much to change them, when adding > a new example to a network. I had this thought three years ago, and have done nothing about it. Formally, of course, it's "turtles all the way down", isn't it? :-) -- ........................................................................... Andrew Palfreyman andrew@dtg.nsc.com Albania before April!
krulwich@ils.nwu.edu (Bruce Krulwich) (02/09/90)
In article <4914@itivax.iti.org>, ajb@itivax (Al Boehnlein) writes: >Has any one ever though about using a neural net to >program a neural net. It would seem that a neural net >might be just the thing to use to determine what weights >to change, and how much to change them, when adding >a new example to a network. This was discussed by Lapedes and Farber under the title "master-slave networks," in an article in the 1986 Snowbird Conference. It was also discussed by Jordan Pollack under the title "cascaded neural nets" in the '87 Cog Sci conference, and probably also in his thesis. Bruce Krulwich Institute for the Learning Sciences
aam9n@hudson.acc.virginia.edu (minai ali ahmad) (02/10/90)
In article <620@berlioz.nsc.com> andrew@dtg.nsc.com (Lord Snooty @ The Giant Poisoned Electric Head ) writes: >In article <4914@itivax.iti.org>, ajb@itivax.iti.org (Al Boehnlein) writes: >> Has any one ever though about using a neural net to >> program a neural net. It would seem that a neural net >> might be just the thing to use to determine what weights >> to change, and how much to change them, when adding >> a new example to a network. > >I had this thought three years ago, and have done nothing about it. >Formally, of course, it's "turtles all the way down", isn't it? :-) >-- About 2 years ago, Lapedes and Farber reported using a neural net to program an 2 2 asymmetric Hopfield net. They basically used an N X N symmetric net to 2 optimize the N connections of an N neuron net. Of course, the connections had no symmetry limitations, which was sort of neat then, but I thought it was a huge case of overkill. The paper appeared in Physica 22D. >........................................................................... >Andrew Palfreyman andrew@dtg.nsc.com Albania before April! Ali Minai