Databases Reference
In-Depth Information
TABLE 3.1: Store and Recall Responses of a GN Array
Input Sequence
Input Pattern
Output
First
XXYX
#### (Store)
Second
XXYX
XXYX (Recall)
the adjacent nodes that are activated in a particular pattern input phase. A
row of the bias array represents a part of the stored pattern. A bias entry is
defined if the set of adjacent neurons does not match any existing rows of the
bias array. A new pattern is found when at least one activated neuron cannot
find a matching entry in its bias array. In this stage, new patterns are stored,
and previously encountered patterns are recalled. Table 3.1 shows the process
when pattern “XXYX” is stored and then recalled. Note that when a pattern
is stored for the first time, the output from the GN network is a null entry,
represented by the “#” pattern in the table. A null response indicates that
no match was found, and the segments of the pattern were stored by the GN
array.
Stages 1 and 2 of the GN learning phase take place in a completely parallel
and decentralized manner. As shown in Figure 3.3, the maximum size of a bias
array is two and occurs in GN (Y,2) after the array has stored four patterns.
Scalability tests, using as many as 16,384 nodes, have shown that increases
in the size of the network result in nominal increases in the computational
complexity [33].
In the supervised GN approach, the size of the network depends on the size
of the patterns and the number of unique elements in the pattern used for
recognition or classification purposes. Given pattern P = a, the number of
GNs, N (a), required in a one-dimensional GN network analysis is given as
follows.
N (a) = s
a
•e
a
(3.1)
where s
a
represents the size, and e
a
is the number of unique elements of pat-
tern a. Eventually, an increase in the dimensions of the patterns will increase
the number of GNs in the GN network. Therefore, given the dimension of
pattern a as d
a
, the number of GNs can be determined as follows.
N (a) = s
a
•e
a
•d
a
(3.2)
The GN approach has been used in a number of applications involving pat-
tern recognition and classification. With lightweight and distributed features,
GN implementations have been applied in resource-constrained networks, such
as wireless sensor networks (WSNs). Khan and Mihailescu [2] proposed a GN
implementation for pattern recognition in a WSN. A simulation of sensory
