Databases Reference
In-Depth Information
only respond to a pattern signal that has element “B” in the same position.
To generate the network that matches the criteria of the patterns that will be
used, a GN network must be able to obtain prior knowledge of the pattern.
This type of network is a supervised GN network.
3.1.2 Recognition Procedure
The GN recognition process involves the memorization of adjacency infor-
mation obtained from the edges of the graph. Adjacency information for each
GN is represented using the (left, right) formation. Each activated neuron
records the information retrieved from its adjacent left or right neuron. In
the GN terminology, this adjacency information is known as a bias entry,
and each neuron maintains an array of bias entries. The entries for the entire
stored pattern are collectively stored in the bias arrays. Each neuron holds a
single bias array, which contains all of the bias entries obtained in recogni-
tion processes. Because each neuron is only required to store a single array,
the storage complexity of the GN recognition process is low. Furthermore,
the bias array of each neuron stores only the unique adjacency information
derived from the input patterns.
In the graph-matching representation, pattern recognition based on a GN
network implements the graph comparison approach by treating each pattern
as a graph, each element of a pattern as a vertex, and the position between
elements as an edge. Consider the following example: Given two patterns, P
in
and P
st
, P
in
is said to match P
st
if the following conditions are satisfied:
1. The number of vertices in P
in
, V
in
, is equivalent to the number of vertices
in P
st
, V
st
, i.e., |V
in
| = |V
st
|.
2. The number of edges in P
in
, E
in
is equivalent to the number of edges in
P
st
, E
st
, i.e., |E
in
| = |E
st
|.
3. The bias entry, b ∈ B
in
for each vertex v ∈ V
in
is a subset of bias array,
B
st
, for each vertex v ∈ V
st
, i.e., b ∈ B
st
.
The pattern recognition process initially takes place in the following phases:
3.1.2.1
Pattern input phase
An input pattern, defined by p (value, position) pairs, is sequentially broad-
cast throughout the network. Each neuron responds only to the input pair that
corresponds to the pre-defined position and value settings of the neuron; it
disregards the remainder of the pattern. From Figure 3.3, GN X(1) has a
pre-defined value = “X” and position = 1 and will respond to the first letter
of pattern P1, i.e., “X”YXX, which is input as pair p1(X,1). This neuron will
ignore the rest of the message. Similarly, GN Y(2) will respond to the second
pair p2(Y,2); GN X(3) will respond to p3(X,3); and GN X(4) will respond
