Databases Reference
In-Depth Information
network. The memorization can occur in the pre-execution stage or instanta-
neously during the recognition process. The former means that the GN algo-
rithm performs a supervised recognition; the latter represents an unsupervised
mechanism. Furthermore, GN algorithms perform recognition on patterns of
equivalent size. Therefore, the features of auto-AM have been fulfilled.
The scalability of the DHGN and other GN-based algorithms is owing to
the adoption of an associative memory approach. The DHGN is an associa-
tive memory system that is capable of recognizing patterns (either original or
noisy), and it is able to match multiple streams of input with historical data
in the network in real-time [61]. For a given pattern, the DHGN also performs
an internal association, in the sense that an association between elements of
a pattern is considered. For example, given a pattern, P , comprising five ele-
ments, {p
1
, p
2
, p
3
, p
4
, p
5
}, the DHGN also takes into account the associations
set {(p
1
, p
2
) , (p
2
, p
3
) , (p
3
, p
4
) , (p
4
, p
5
)}. The following subsection will fur-
ther discuss the architecture of the DHGN in line with its pattern recognition
process.
5.1.2 DHGN Computational Design
The DHGN formalizes the distributed HGN scheme described in Chapter 4.
By dividing and distributing subpatterns, the DHGN adds a clustering mech-
anism for pattern recognition. Each of the subpatterns undergoes a one-shot
recognition procedure. The results of the sub-recognition add cumulatively to
obtain the actual recognition result.
The DHGN network constitutes a number of DHGN subnets (HGN sub-
composition) and a Stimulator/Interpreter Module (SI module) node, as de-
scribed in Muhamad Amin and Khan [4]. Figure 5.1 shows the complete ar-
chitecture of the DHGN network. In this figure, the decomposition of binary
image pattern âKâ into subpatterns is illustrated. This decomposition is per-
formed by the SI module node. The input activates the GN nodes that cor-
respond to the bits of the input pattern. In doing so, each pattern element of
a subpattern is mapped to the relevant GNs in the respective subnet. Each
subnet integrates its responses and sends the results to the SI module to form
an overall response.
Communications within the DHGN network occur in a single-cycle environ-
ment, i.e., each pattern is passed through the network only once. Recognition
result is produced in the form of recall (pattern is known) or store (pattern
is memorized). Within each DHGN subnet, the communication between GNs
occurs once for each subpattern. By eliminating the need for an iterative
mechanism to recall or store patterns, the DHGN offers a fast recognition
procedure.
Each DHGN subnet is derived from a composition of inter-connected GNs.
The size of the subnet depends on the size of the subpattern and the number
of different elements in the subpattern. Therefore, to define the size of each
