Databases Reference
In-Depth Information
Chapter 5
Recognition via
Divide-and-Distribute Approach
As discussed in the previous chapter, the effectiveness of one-shot learning
pattern recognition, such as Graph Neuron (GN)-based algorithms can be
improved by dividing patterns into subpatterns and distributing them across
multiple computational networks. This improvement has a two-fold effect.
First, the scalability of the recognition process improves. This approach al-
lows recognition to scale up with the size of patterns and the network capacity
to conduct the recognition. Second, the distribution of patterns into subpat-
terns of equal or different sizes allows for error encapsulation in a particular
subnet, and thus recognition is performed more accurately. Nevertheless, the
effects of error encapsulation can only be observed when the error is small and
concentrated.
Graph Neuron (GN)-based algorithms have been developed based on two
different concepts, graph-matching and associative memory. These two con-
cepts give GN-based algorithm implementation the added advantage of scal-
ability. The simple recognition procedure and lightweight algorithm of the
GN give it the ability to perform pattern recognition processes on distributed
systems. Furthermore, GN algorithms incur low computational and commu-
nication costs when deployed in a distributed system. Previous chapters have
analyzed both the GN and HGN approaches and introduced a distributed
version of the HGN, the Distributed Hierarchical Graph Neuron (DHGN).
5.1 Divide-and-Distribute Approach for One-Shot Learn-
ing IS-PR Scheme
An important aspect in the development of pattern recognition schemes is
the algorithmic design. A proper design is e
cient and has the ability to gen-
erate a more accurate classification strategy. In this chapter, the algorithmic
design and prospects of using the proposed DHGN algorithm for Internet-
scale pattern recognition schemes are extensively discussed. The proposed
algorithm extends the scalability of the existing Hierarchical Graph Neuron
(HGN) implementation by reducing the computational requirement incurred
73
