Databases Reference
In-Depth Information
TABLE 8.1: Comparison of Fine- and Coarse-Grained Network Specifications
Network Granularity
Specifications
Coarse-Grained
Fine-Grained
No. of Processing Nodes
Low to High
High
Processing Capacity
High
Low
Storage Capacity
High
Low
Energy Supply
High
Low
Example
Computational Grid
WSN
versely, fine-grained networks are defined as a network that comprises many
small processing nodes that perform simple and lightweight tasks, such as
the wireless sensor network (WSN). Table 8.1 compares the specifications of
coarse- and fine-grained computational networks.
The DHGN implementation for distributed pattern recognition takes into
account these two levels of granularity. This is essential in providing a scalable
and robust scheme that can be used in different network conditions. Further-
more, because of the network granularity considerations, the DHGN algorithm
is made aware of the resource availability of the computational network to be
used in the recognition process.
8.1.1 DHGN Configurations for Adaptive Granularity
Two configurations for the DHGN implementation will be presented: fully
distributed and clustered.
8.1.1.1
Fully Distributed Configuration
The original configuration of the DHGN algorithm described in Chapter 5
distributed all of the neurons in a DHGN subnet to the processing nodes. This
implies that each node is responsible for a single neuron in a DHGN subnet.
This configuration eliminates the requirement for high processing capability
and storage capacity because the computing node performs the recognition
process on a single atomic element of the input subpattern. However, the
communication costs for each node require considerable attention. Each node
is required to communicate frequently with other neighboring nodes to up-
date its bias array. Figure 8.1 shows the fully distributed configuration of the
DHGN algorithm for a WSN. Note that each neuron is mapped to a process-
ing node. Processing nodes that are close together are grouped into individual
DHGN subnets.
This fully distributed DHGN configuration can be deployed in a fine-grained
network that comprises sensor nodes with restricted computing resources, such
as WSN. A major challenge in this implementation is the rapid inter-node
communications required for message exchange during the recognition process.
