Kedar Dhuru, Anand Gopalan, F.R.C.R, Contributing Editor                                                      Neural Networks
Feature Focus
"Cogito Ergo Sum (I think therefore I am)"

This Descartian aphorism has governed man's progress right from the Age of Enlightenment. The faith in the human intellect has always propelled man on the path of scientific achievement. Yet the human brain with all its mystery and complexity remains an elusive subject of study. Its mysteries wait to be unraveled fully. The "human information processing system" composed of neurons which switch at speeds about a million times slower than today's digital circuits, still efficiently performs a variety of complex tasks like speech recognition which remain elusive to most of today's advanced computer systems. Out of a need to imitate this complexity and functionality of the human brain rises the science of Neural Networks.

image3 The inputs from various connectors are assigned different weighted values. As a result a single neuron is capable of performing a myriad of functions depending on the values at its various connectors which effectively decide its function. In contrast to traditional computers, which have a fixed set of instructions, most NNs, have some sort of "training" rule whereby they learn from examples much in the same way that small children recognize animals from viewing examples of animals. A neural network is a massively parallel distributed processor that has a natural propensity for storing experiential knowledge and making it available for use. It resembles the brain in two respects:
· The network, through a learning process, acquires knowledge.
· Inter-neuron connections called "synaptic weights" are used to store the knowledge.
The most basic processing unit of a Neural Network is called an Artificial Neuron. The artificial neuron is simply a device, which fires when its total input exceeds a particular value.


Neural networks are a form of multiprocessor computer system, with simple processing elements, a high degree of interconnection, simple scalar messages, and adaptive interaction between elements. Neural Networks are of two kinds, Artificial and Biological Neural Networks. A large amount of work has gone into the task of developing artificial units which function like neurons and hence do away with their biological counterparts and the term Neural Networks is now used predominantly to describe Artificial Neural Networks. There is no universal definition of an NN. But most people in this field would say that NN is a network of many simple processors (units), each having a small amount of local memory. Communication channels (connections) which usually carry numeric data, encoded by various means connect the units. The units operate only on their local data and on the inputs they receive via the connections.
Although still a budding science Neural Networks already find applications in a variety of fields. Computer professionals use it to find the properties of non-symbolic information processing and learning systems in general. Engineers use it for signal processing and automatic control. Neuro-physiologists use it to describe and explore medium level brain controls. Biologists use it for interpreting nucleotide sequences.

Although theoretically, NNs can do anything that a normal digital computer can do, but practically, they are useful only for tasks like classification and function approximation which have a lot of training data but to which certain hard and fast rules cannot be applied. Unlike the brain, NN cannot be used to create data and information, which do not already exist in the training data. Thus, they cannot be applied to tasks involving the manipulation of symbols and memory. So although present day neural networks can increase the efficiency and flexibility of the modern day computer, the dream of simulating the human consciousness still belongs in the realms of science fiction.


Back main page Profile of the month: Sir C.V.Raman

The IEEE Bombay Section Student Newsletter


Web Edition of newsletter by Rohit Mordani, TSEC