Databases Reference
In-Depth Information
distributed into different feature subspaces. These feature subspaces are pre-
processed and sent to a final module that conducts the classification process.
This approach alleviates the need for large training data sets for SVM.
Some pattern set distribution techniques also distribute patterns to process-
ing entities within a network. For instance, Lobo, Bandeira, and Moura-Pires
[43] proposed a distributed SOM for a ship recognition process using acous-
tic signatures. This type of technique requires that results are collected from
each processing entity and further processed by an intensive post-processing
mechanism. The pattern set distribution technique does not minimize the
computational complexity of the recognition algorithm. However, it reduces
the execution time and allows for parallel processing implementations. This
technique is suitable for recognition schemes that analyze a large number of
patterns. However, the technique does not fit well into systems that cater to
high-dimensional and large-scale data, such as Magnetic Resonance Imaging
(MRI) images.
Existing distributed pattern recognition approaches tend to employ the set
distribution technique. This technique alleviates the need for a large number
of training data sets, which leads to fast learning speeds. Nevertheless, the
complexity issue remains unsolved. Examples of DPR schemes that use the
set distribution technique include the works carried out in [42, 44, 45].
2.5 Current DPR Schemes
A number of purely distributed pattern recognition approaches have been
pursued. Several neural network schemes have been developed that have a dis-
tributed processing capability, such as the Hamming AM and Morphological
AM. Nevertheless, the algorithmic distribution capability has yet to be further
analyzed. In recent years, DPR methods based on the original Graph Neu-
ron (GN) algorithm have been developed. Established extensions include the
Hierarchical GN (HGN) [3] and Distributed Hierarchical GN (DHGN) [46]
algorithms. In this section, we will discuss briefly some of the fundamental
characteristics of these schemes.
2.5.1 Graph Neuron
Graph Neuron (GN) is a pattern recognition algorithm that implements a
simple associative memory (AM) architecture, which provides the capability
of pattern recall based on similar or incomplete patterns. In an associative
memory architecture, the store and recall operations are based on an asso-
ciation with the input rather than the address of the memory content as
is used in a conventional memory architecture. Therefore, pattern recogni-
