Databases Reference
In-Depth Information
5.1.2.2.3 Top layer
These neurons are only responsible for communi-
cating the final index for each subpattern stored/recalled to the SI mod-
ule. The costs for communicating these indices were included in the macro-
communication evaluation.
This subsection has presented a detailed description of the DHGN architec-
ture for distributed pattern recognition. This architecture represents an ab-
stract formation of the network. In reality, this architecture can be deployed
in a coarsely distributed or finely distributed network environment.
5.1.3 Dual-Phase Recognition Procedure
The DHGN architecture that has been described in the previous subsection
comprises two important entities: the SI module and the DHGN subnets.
Recognition of patterns mainly occurs within each DHGN subnet. However, at
this instant, all each subnet knows is a sub-composition of the overall pattern.
The DHGN system must restructure the overall information of the pattern
and produce a result for the entire pattern, i.e., whether the input pattern
is known to the system or not. Another phase of recognition is required that
involves the results of the recognition processes executed by the subnets.
The recognition procedure for the DHGN implementation can be analogi-
cally represented as a distributed analysis procedure, as shown in Figure 5.2.
Imagine there is a large block of data that needs to be analyzed. Given a set
of analysts, this large block of data can be decomposed into sub-structures of
data, and each analyst would work on a sub-structure. In the end, the results
of the analysis must be recompiled to form an overall result for the analysis
of the large block of data.
The DHGN distributed pattern recognition performs pattern analysis in
two phases:
1. Subpattern recognition
2. Pattern reconstruction and recognition
Note that these two phases occur consequently and within a single-cycle recog-
nition mechanism.
5.1.3.1
Phase 1: Subpattern Recognition
In the DHGN implementation, the core recognition process is conducted at
the subpattern level. There are four stages involved in this process.
Stage 1.
After receiving an input from the SI module, each activated neu-
ron in the base layer will send a signal message to the nodes in
the adjacent columns containing the row number/address of the
activated neuron. The activated neurons on the edges of the layer
will only send the activation signal message to the neurons in the
penultimate columns. The activated neurons that receive the signal
