Browse Prior Art Database

Prespecifying Neural Network Weights

IP.com Disclosure Number: IPCOM000111287D
Original Publication Date: 1994-Feb-01
Included in the Prior Art Database: 2005-Mar-26
Document File: 2 page(s) / 80K

Publishing Venue

IBM

Related People

Camp Jr, WO: AUTHOR

Abstract

Disclosed is a method for prespecifying the weights of a neural network in order to decrease the likelihood of incorrect responses from the network due to a less than optimum training data set.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Prespecifying Neural Network Weights

      Disclosed is a method for prespecifying the weights of a neural
network in order to decrease the likelihood of incorrect responses
from the network due to a less than optimum training data set.

      This disclosure addresses a problem with the use of neural
networks to determine the category of an input pattern, particularly
when the input pattern consists of a large number of inputs in the
range of 1,000 to 1,000,000.  The networks often used for this task
are multilayer and would have interconnections numbering much greater
than 1,000,000.  It is also an observation of this field that the
number of unique training patterns must be substantially greater than
the number of weights to be trained, much like over determining a set
of equations.  Then there is the other problem that the number of
potential input patterns is 2 to the power of the number of inputs,
and one does not have any chance of exhaustively checking the
performance of the network after it is trained.  Both of these
problems are facets of the phenomenon called "combinatorial
explosion".

      Prior attempts to reduce the number of free variables in a
network include slaving the variables to one another among groups of
connections and prespecifying weights by copying the mammalian visual
system.  These work well for the early stages of image processing,
but do not work well beyond the point of finding edge oreintations in
an image.

      The premise of this invention is that it is better to construct
as much of the network by design to accomplish pattern recognition,
rather than rely on training with what will invariably be too few
patterns.  This invention is a method to construct the weights for
the portion of the network following the early stages where edge
orientations are found in the image.

      The input to the network to be described is an array of angle
orientations of edges found in the image.  Each input can be one of a
finite number of discrete angles representing the orientation of an
edge at that location to the image.  The network of weights is
constructed by forming connections from this array of inputs to an
output node for each pattern to be recognized.  Each point on the
array of inputs has connections for ea...