Browse Prior Art Database

Neural Network with Winner-Take-All Behavior in the Hidden Layer

IP.com Disclosure Number: IPCOM000111096D
Original Publication Date: 1994-Feb-01
Included in the Prior Art Database: 2005-Mar-26
Document File: 2 page(s) / 43K

Publishing Venue

IBM

Related People

Camp Jr, WO: AUTHOR

Abstract

Disclosed is a method to enhance the ability of a pattern recognizing neural network to discriminate between patterns. If at each location in an image, the first layer of the network develops several analog values representing the correlation between any edge in the image and a discrete number of angles, and the second layer of the network multiplies each of these analog values by a vector with a preferred angle having a positive value and non-preferred angles having negative values, and this is done for all locations across an image, then one has a filter tuned to a particular pattern. However, if the first layer process results in a lot of ambiguous information about the angle of an edge in the image, outputs from that layer would exist for all or many of the possible angles.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 75% of the total text.

Neural Network with Winner-Take-All Behavior in the Hidden Layer

      Disclosed is a method to enhance the ability of a pattern
recognizing neural network to discriminate between patterns.  If at
each location in an image, the first layer of the network develops
several analog values representing the correlation between any edge
in the image and a discrete number of angles, and the second layer of
the network multiplies each of these analog values by a vector with a
preferred angle having a positive value and non-preferred angles
having negative values, and this is done for all locations across an
image, then one has a filter tuned to a particular pattern.  However,
if the first layer process results in a lot of ambiguous information
about the angle of an edge in the image, outputs from that layer
would exist for all or many of the possible angles.  When multiplied
by the following layer templates, this would result in a large
negative contribution to the sums for all the patterns.  Thus, what
is simply an ambiguous situation becomes a strongly negative
contribution to all the results of the second layer of the network.

      To avoid this situation, the network is made to have inhibitory
cross-connections between just the hidden nodes for each possible
angle representing one location in the image.  Similar connections
occur at all the similar groups of hidden nodes for each of the
locations in the input image.  These inhibitory connections force the
hidden...