Paper available: "Massive Weight Sharing: A Cure for ..."

[Benny Lautrup]

FTP-host: connect.nbi.dk
FTP-file: neuroprose/lautrup.massive.ps.Z

----------------------------------------------

The following paper is now available:

Massive Weight Sharing: A Cure for Extremely Ill-posed Learning [12 pages]

B. Lautrup, L.K. Hansen, I. Law, N. Moerch, C. Svarer and S.C. Strother
CONNECT, The Niels Bohr Institute, University of Copenhagen, Denmark

Presented at the workshop on Supercomputing in Brain Research: From Tomography
to Neural Networks, HLRZ, Juelich, Germany, November 21-23, 1994

Abstract:
In most learning problems, adaptation to given examples is  well-posed  because
the number of examples far exceeds the number of  internal  parameters  in  the
learning machine. Extremely ill-posed learning problems are, however, common in
image and spectral analysis. They are characterized by a vast number of  highly
correlated inputs, e.g. pixel or pin values, and a modest number  of  patterns,
e.g. images or spectra. In this paper we show, for the case of  a  set  of  PET
images differing only in the values of  one  stimulus  parameter,  that  it  is
possible to train a neural network to learn the underlying rule  without  using
an excessive number of network weights or large amounts of computer  time.  The
method is based upon the observation that the standard learning rules  conserve
the subspace spanned by the input images.

Please do not reply directly to this message.

-----------------------------------------------

FTP-instructions:

unix> ftp connect.nbi.dk (or 130.225.212.30)
ftp> Name: anonymous
ftp> Password: your e-mail address
ftp> cd neuroprose
ftp> binary
ftp> get lautrup.massive.ps.Z
ftp> quit
unix> uncompress lautrup.massive.ps.Z

-----------------------------------------------

Benny Lautrup,
Computational Neural Network Center (CONNECT)
Niels Bohr Institute
Blegdamsvej 17
2100 Copenhagen
Denmark

Telephone: +45-3532-5200
Direct:    +45-3532-5358
Fax:       +45-3142-1016
e-mail:    lautrup at connect.nbi.dk