Web.
The web site itself may have changed. You can check the current page or check for previous versions at the Internet Archive. Yahoo! is not affiliated with the authors of this page or responsible for its content.
Artificial Neural Networks: The next intelligence
83




Artificial Neural Networks: The next intelligence

By Amit Khajanchi



The computing world has a lot to gain from neural networks. Their
ability to learn by example makes them very flexible and powerful.
..Perhaps the most exciting aspect of neural networks is the possibility
that some day 'conscious' networks might be produced.
1

-
Christos Stergiou and Dimitrios Siganos,
Department of Computing, Imperial College of London
Abstract
This paper is divided in two parts. Part one examines the relevance of Artificial
Neural Networks (ANNs) for various business applications. The first section sets the
stage for ANNs in the context of modern day business by discussing the evolution of
businesses from Industrial Revolution to current Information Age to outline why
business today are in critical need of technology that sifts through massive data.
Next section introduces Artificial Neural Network technology as a favorable
alternative to traditional analytics and informs the reader of the basic concept
underlying the technology. Finally, third section screens through four different
applications of ANNs to gain an insight into potential business opportunities that lie
abound.

84
Part Two focuses on a commercial venture that leverages ANN. For this purpose, a
case study of HNC Software, a technology company that patented fraud detection
application using Artificial Neural Networks , is examined. The case will present
companys profile, its competitive landscape, and various technology issues it
encountered through its growth phase. Finally, the paper will end by asking a critical
question that highlights the relevance of strategy in a technology venture.

Introduction
Modern day businesses face unique challenges that were nonexistent prior to the
Industrial and Internet Revolution. Industrial revolution brought about the concept
of economies of scale, mass production and standardization. Businesses competed
on the grounds of operational efficiency and scale of production. As a result,
successful organizations grew larger to accommodate these practices and faced an
increasing amount of coordination costs to fulfill their services. Advances in
communication technology such as telephone, television, fax and internet have
greatly enhanced the organizations ability to coordinate through chains of
geographically dispersed units, suppliers, and customers and enabled the large
corporations to minimize coordination costs. As a result of those advances and
improvements in transportation, businesses started competing on the grounds of
timely delivery of products/services and customer satisfaction.
Today, greater means available for distribution, communication, and production
facilities are not without their cost. What is at stake is the crucial element of
business--customer relationship. In the old days, a business manager knew the
customers inside and out and recommended products suited to their needs and
preference on a timely fashion. However, with the advent of mass marketing, mass
distribution, and mass production, business decision making (product planning, cross
selling, pricing decisions) has become detached from a unique customer to fit the
needs of an average potential customer. The truth is, the product that satisfies
an average customer has a low customer satisfaction rate because it is not optimized
to fit a unique customers needs and/or preferences. As a consequence, niche
players who can accurately define their customer segment have a greater advantage
over larger businesses that cannot.
Large enterprises strive to stay competitive by strategically utilizing modern day
analytics to understand and to be closely in tune with the changing needs and
preferences of their customers. Statistics (moving averages, ratio analysis, time
series analysis, regression analysis) and computational science (linear programming,
calculus, and simulation techniques) greatly enhance the businesses ability to
dissect the given data and organize it to create meaningful information to support
decision making. However, all these methods require man power (1) to organize
appropriate data; (2) to analyze meaningful information; and (3) to communicate
useful knowledge. This process is time consuming and worse yet, static (meaning
that the result of analysis is specific to the time frame of analysis). The Artificial
Neural Networks approach provides an attractive alternative that enables large
businesses to be adaptive to the changing needs and preferences for each customer
segment.
Overview of Artificial Neural Networks (ANNs)
The concept of neural networks is modeled after biological sensory mechanisms
where the neuron signals are transmitted to the brain and processed. This concept
85
moves away from traditional statistical models where data are analyzed based upon
holding everything else constant (ceteris paribus). The weakness in statistical
models lies in their inability to model the changing relationships between variables
(non-linear problem) and thus presents challenges in making a predictive analysis
where the underlying relationships are not constant. A neural network overcomes
this problem by being adaptive to real sets of data. Much like living organisms, a
neural network gets training and learns the tricks of the trade by observation and re-
adjusts its learning against new sets of data iteratively. Wen (2000)
2
states,
Artificial Neural Networks (ANNs) are an information processing technology
pertaining to the area of machine learning in artificial intelligence. A neural network
employs an adaptive structure that can be trained with application data to capture
complex relationships between input and out variables.
Neural Netw ork Architecture
The inherent power of neural networks lies in its ability to recognize the underlying
relationship between input and output data. According to Nasir (2001),
3
the
prototypical use of neural networks is in structural pattern recognition. Through a
preset learning algorithm and series of training iterations the network learns to
recognize patterns in the data sets and assigns weights to each variable (nodes).
Neural network architecture employs multiple layers of nodes. A node is where the
data is converted into values between 0 and 1 using sigmoid transfer function in a
network. Following figures illustrates this:
Figure 1: Processing algorithm of a node in a neural network (Wu 1994)
4


Figure 2: Graph of Sigmoid Transfer Function (Wu 1994)

Additional layer in neural network architecture models add complexity. For example,
two layer architecture (Figure 3) will contain an input layer and an output layer;
three layer architecture (Figure 4) will include a hidden layer in the middle
(unobservable variables); and a more complex network will have fourth threshold
(constraint) layer (Figure 5).
86
Figure 3: Two Layer Neural Network model (Chatterjee 2000)
5


Figure 4: Three layer Neural Network model (Wu 1994)

Figure 5: Four-layer feed-forward neural network model (Yue et al 2002)
6


Learning Algorith m
The crucial part of neural network alchemy is in its ability to learn from series of
iterations of input data (called the training period). The most basic algorithm that
enables this is known as the delta rule. In 1962, Widrow
7
provided this first learning
algorithm for 2 layer ANNs (a.k.a. perceptron). By 1986, with advances in
computational technology and further academic work, Rumelhart, et al.,
8
suggested
a more novel approach called the back propagation method that enabled learning for
multilayered feed forward networks. Essentially, delta rule was insufficient for
training networks with hidden layers that did not have direct inputs and outputs.
Bias
87
Back propagation provided a sounder training scheme for a network to incorporate
multilayers of nodes in its design according to Wen (2002).
In essence, the back propagation algorithm consists of three general steps (Wu
2000):
1. compute outputs
2. compare outputs with the desired targets, and
3. adjust connections weights and parameters of the activation function(s) to
remove as much output errors as possible.
Screening the Technology for Opportunity Recognition
Artificial Neural Networks have a multitude of real world applications. They are well
suited to dissect and analyze voluminous data and extract meaningful patterns and
relationships. There has been much academic research done in this area and the
results are promising. As Wray, et al., (1994)
9
mention, the advantages of neural
networks over statistical models are (1) ANNs requires no predefined knowledge of
underlying relationships between input and output variables; (2) ANNs associative
ability make them robu