The idea of building an- intelligent machine out of artificial neurons (smallest sensor of brain) has been around for quite some time. Research in neural networks was virtually halted in the 19~0s because of its complexity in realizing it on hardware. In recent years neural network has become one of the most active areas in computer science due to introduction of faster digital computers, massive parallel computers, and discovery of powerful learning algorithms.

Human brain consists of about 10 interconnected neurons. Hence it is virtually impossible to duplicate the operations of brain with the present knowledge of neurons and computer technology. The new neural network architecture called connectionist architectures receives inspiration from known facts about how the brain works. The connectionist architecture is based on

• Large numbers of very simple neuron like processing elements.

• Large numbers of weighted connections between the elements (the weights represent the knowledge of the network,

• Highly parallel, distributed control and,

• Emphasis on learning internal representations automatically.

Computers are capable of amazing feats. They can effortlessly store vast quantities of information, they are very fast and they can perform large arithmetic calculations without error. But the current AI systems are not good at doing the simple tasks which humans routinely perform (such as seeing, talking and common reasoning). Perhaps the structure of the brain is somehow
suited to these tasks and not suited to tasks like high speed arithmetic calculation and vice versa.

Comparing human brain with present day computer, individual neurons is extremely slow device compared to its counterparts  in  digital computers. Neurons operate in milliseconds range, which is very far from present computers speed. Yet, humans can perform extremely complex tasks, like interpreting a visual scene or understanding a sentence, in just tenth of a second. In other words, we do in about ap hundred steps what current computers can not do in ten million steps. Though it sounds unrealistic, it is a fact, because unlike conventional computer the brain contains a huge number of processing elements that act in parallel. That means we are looking for massively parallel computers.

Another thing people seem to be able to do a better than computers is handle fuzzy situations. We have very large memories of visual,' auditory, and problem-solving episodes, and one key operation in solving new problems is finding closest matches to old situations. Inexact matching is something brain style models seem to be good at. Home appliances as washing machines, refrigerators based on neuro fuzzy logic are already in the market.

Neurons are failure prone devices. They are constantly dying (you have certainly lost a few since you began reading this article), and their operation is irregular .But a computer must operate perfectly, we certainly do not want it to reply in French when we ask a question in Nepali. Loss of information due to failure can be handled by having more than one copy of the information stored in the computer. This introduces redundancy into the system but it works perfectly even if some components of the computer fail. With present day VLSI (very large scale integration) circuit technology it is possible to build a billion component integrated circuit in which 95 percent of the components work correctly.

The artificial neurons are designed to mimic the structure of neurons. A set of inputs, each representing the output of another neuron are applied. Each input is multiplied by a corresponding weight, and all of the weighted inputs are summed to determine the output of the neuron.

A larger network of such neuron can be built and trained to perform some function. Training a neural network means changing the weights of the inputs so that it produces desired output. The most oRen used method of training known as 'supervised training' requires pairs of input and desired output. The actual output when all the inputs are applied is used to calculate the error (difference between desired and actual output). The input weights are adjust to until the desired output is obtained.

However, supervised training has been criticized for not being natural, because the human brain does not know the desired answer beforehand. A new training method known as 'unsupervised training' is proposed which does not requite the desired output to train. The training process extracts the statistical properties of the input set and learns to produce consistent output for the inputs which are similar.

The artificial neurons which may be implemented by software or hardware exhibit a surprising number of brain ' s characteristics.

Learning

They learn from experience. They can modify their behavior in response to their environment.

Generalization

Once trained a network's response can be, to a certain degree, insensitive to minor variation's in the input. This ability to see through noise and distortion to the pattern that lies within is vital to pattern recognition in a real world environment. Such a characteristic is not possible with conventional computer for which the inputs must be in a pre-specified pattern. Thus, it produces a system that can deal with the not-so-perfect world in which we live.

Some neural network are able to abstract essential characteristics from inputs containing irrelevant data. For example a network can be trained on a sequence of distorted versions of the letter A. After adequate training the network will be able to produce a perfectly formed letter, that means it has learned to produce something that it has never seen before.
 

Applicability and future

Artificial neural networks are not panacea. They are not well suited for tasks as calculating the payroll. Because of their capability to extract essential data from incomplete or incorrect data they are preferred over conventional computers for a large class of pattern recognition problem.

Potential applications are those where human intelligence functions effortlessly and conventional computation has proven cumbersome or inadequate. This application class is at least as large as that serviced by conventional computation, and the vision arises of artificial neural networks taking their place alongside of conventional computation. This will happen only if fundamental research yields results at a rapid rate, as today's theoretical foundations are inadequate to support such projection.