Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

Partitioned Class Recognition by Computational Neural Network

IP.com Disclosure Number: IPCOM000114035D
Original Publication Date: 1994-Nov-01
Included in the Prior Art Database: 2005-Mar-27
Document File: 4 page(s) / 116K

Publishing Venue

IBM

Related People

Narasimha, MS: AUTHOR [+2]

Abstract

Disclosed is a method for constructing neural networks for recognizing partitioned classes and combining their results to get individual class recognition. Such a system can be trained more easily and provide more accurate recognition. A strategy based on the theory of finite projective planes gives the benefits of partitioned character class recognition while keeping the number of partitioned classes to a minimum.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 50% of the total text.

Partitioned Class Recognition by Computational Neural Network

      Disclosed is a method for constructing neural networks for
recognizing partitioned classes and combining their results to get
individual class recognition.  Such a system can be trained more
easily and provide more accurate recognition.  A strategy based on
the theory of finite projective planes gives the benefits of
partitioned character class recognition while keeping the number of
partitioned classes to a minimum.

      Disclosed is a method for constructing a neural network based
recognition system that, when compared to traditional neural network
based recognition systems using identical sets of feature
measurements, can be trained more easily and provides more accurate
recognition.  In addition, unlike the traditional approach, the
approach disclosed here is well suited for a parallel implementation
in a multiprocessor system in which the number of processors can be
as high as the number of classes to be recognized.  Training can
easily be performed in parallel on a network of workstations
significantly reducing the time required.

      When partitioned class recognition is used in an application,
the feature measurements of an instance of an unknown class are
processed (in parallel, if possible) by the perceptrons associated
with each class subset.  The outputs of the entire collection of
perceptrons are combined to produce a recognition result in a way
that is described below.

      A simple example of partitioned class recognition is
illustrated in the Figure.  In this example only three classes of
object are being recognized and four feature measurements are being
used.  The three classes, numerals {1, 2, 3} have been partitioned
into three subsets {1, 2}, {1, 3} and {2, 3}.  The blocks labeled N1,
N2, and N3 represent the entire hidden layer of a  neural network
that has been trained to recognize classes belonging to exactly one
of the subsets.  That is, N1 has been trained using instances of 1
and 2, N2 has been trained using instances of 1 and 3, and N3 has
been trained using instances of 2 and 3.  The Figure illustrates an
instance of a 2 being recognized correctly by having the smallest
output value for the summer corresponding to the number 2.

The following characteristics are desirable in defining the class
subsets:
  o  The union of the subsets in which any given class appears is the
      set of all classes.  (That is, each class appears at least once
      with every other class and at least one of the perceptrons is
      trained to distinguish each pair).
  o  Each class appears with every other class in no more than one
      subset.  (That is, duplications are minimized and the
      intersection of any two subsets is at most one class.)
  o  All classes make about the same number of total appearances in
      the complete collection of subsets.

      There are a large number of...