Databases Reference
In-Depth Information
context of scalability must also be considered. Existing demand for large-scale
analyses can only be addressed using the enormous capacity of computational
networks. In addition, a scheme for scalable and fully distributed pattern
recognition is important because it will eliminate the implementation bottle-
neck of existing tightly coupled and highly iterative recognition algorithms.
Through these two mediums of change, pattern recognition can be used far
beyond its existing capabilities.
As previously discussed in Chapter 1, we believe that the means for ex-
isting artificial intelligence to emulate the functions of the human brain and
nervous system is through connectivity. Having a fully distributed approach
for information processing enables more information to be stored, and the ca-
pacity to process such information is significantly increased. With the advent
of seamless interconnectivity between smart devices, such as in the Internet-
of-Things (IoT), this capability for large-scale data processing can be further
extended. Furthermore, sensors attached to these large-scale computational
networks will provide an avenue for real-time information processing with a
life-long learning capability.
Imagine a fully interconnected sensory system composed of wireless sen-
sors and a distributed recognition algorithm that could learn how events are
happening and how it could adapt to any changes experienced throughout its
lifetime, i.e., as a pseudo-conscious system that acts as humans do, avoiding
hurtful situations using past experiences.
In the perspective of network evolution, we can see from the discussions in
later chapters of this topic that computational networks have evolved from a
simple local network to computing at tera- or peta-flop scale. Cloud computing
enables a highly scalable means for complex computations and promises enor-
mous resource availability. In addition, the granularity of computational net-
works has evolved from coarse-grained systems, such as a grid, to fine-grained
networks, such as WSNs. The question that remains to be answered is how
we can fully utilize such systems. This perhaps can only be answered if we are
able to change our perspective of computations from sequential Von Neumann
principles to a fully parallel and distributed computing approach. A paradigm
shift is required, from Von Neumann archetype of stored-program computer
to a purely in-network processing approach, in which computations can be
performed in parallel within the body of the network without experiencing
performance bottlenecks resulting from sequential instruction execution and
data operations.
In the following subsections, we will look at the contributions of this topic
from two perspectives: the fundamentals of the recognition process and the
idea of pattern recognition as a scalable commodity for information processing.
10.3.1 Changing the Fundamentals
For almost six decades, research on neural networks has focused on the
learning functions to improve the accuracy and e
ciency of neuron outputs.
