Geoscience Reference
In-Depth Information
13.1 INTRODUCTION
Neurocomputing is an emergent technology concerned with information processing systems that
autonomously develop operational capabilities in adaptive response to an information environment.
The principal information processing structures of interest in neurocomputing are computational
neural networks (CNNs). There are other classes of adaptive information systems such as genetic
learning systems, fuzzy learning systems and simulated annealing systems. Several features distin-
guish this approach to information processing from algorithmic and rule-based information systems
(see Fischer 1995):
1. Information processing is performed in parallel. Large-scale parallelism can produce a
significant increase in the speed of information processing (inherent parallelism).
2. Knowledge is not encoded in symbolic structures but rather in patterns of numerical
strength associated with the connections that exist between the processing elements (PEs)
of the system (connectionist type of knowledge representation) (Smolensky 1988).
3. Neural networks offer fault-tolerant solutions. These tools are able to learn from and make
decisions based on incomplete, noisy and fuzzy information.
4. Neurocomputing does not require algorithms or rule development and will often produce
a significant reduction in the quantities of software that need to be developed.
This alternative approach to information processing offers great potential for tackling difficult prob-
lems, especially in those areas of pattern recognition and exploratory data analysis for which the
algorithms and rules are not known, or where they might be known, but the software to implement
them would be too expensive or too time-consuming to develop. Indeed, with a neurocomputing
solution, the only bespoke software that would need to be developed will in most instances be for
relatively straightforward operations such as data pre-processing, data file input, data post-process-
ing and data file output. CASE (computer-aided software engineering) tools could be used to build
the appropriate routine software modules (Hecht-Nielsen 1990).
13.2 WHAT IS A COMPUTATIONAL NEURAL NETWORK?
Briefly stated, a CNN is a parallel distributed information structure consisting of a set of adaptive
processing (computational) elements and a set of unidirectional data connections. These networks
are
neural
in the sense that they have been inspired by neuroscience. No claim is made to them
being faithful models of biological or cognitive neural phenomena. In fact, the computational net-
works that are covered in this chapter have more in common with traditional mathematical and/
or statistical models, such as non-parametric pattern classifiers, statistical regression models and
clustering algorithms, than they do with neurobiological models.
The term CNN is used to emphasise rather than to ignore the difference between computational
and artificial intelligence (AI). Ignoring this difference might lead to confusion, misunderstanding
and misuse of neural network models in GeoComputation (GC). Computational intelligence (CI)
denotes the lowest
level of intelligence
which stems from the fact that computers are able to process
numerical (low-level) data without using knowledge in the AI sense. An AI system, in contrast, is
a CI system where added value comes from incorporating knowledge that humans possess - in the
form of non-numerical information, operational rules or constraints. Neural network implementa-
tions in the form of feedforward pattern classifiers and function approximators, which are consid-
ered at a later point, are therefore CI rather than AI systems.
Increased effort is now being made to investigate the potential benefits of neural network
analysis and modelling in various areas of GC. In particular, these computational devices would
appear to offer several important advantages that could be exploited, over and above those associ-
ated with what is now becoming known
as the traditional approach to geographical information
