gary@desi.ucsd.edu (Gary Cottrell) (09/11/89)
My mail is not getting through, so I'll post this:
To: chambers
Subject: eigenvalues
I have found that a linear encoder network
(e.g., 64x16x64, use a low learning rate of
about .01 with back prop) will converge on a
solution that spans the principal subspace,
even if it doesn't line up on the principal
components. See also Sanger's article in NIPS
88 (the book is 89), and Baldi's article in the
same volume, as well as Linsker's for an alternative
(ed Touretzy, Morgan Kaufmann).
gary cottrell 619-534-6640 Secretary: 619-534-5288
Computer Science and Engineering C-014
UCSD, La Jolla, Ca. 92093
gary@cs.ucsd.edu (ARPA)
{ucbvax,decvax,akgua,dcdwest}!sdcsvax!gary (USENET)
gcottrell@ucsd.edu (BITNET)