|
|
|
How neural networks make thinking machines
AN ARTIFICIAL Neural Network is an information processing technology derived from the concept that computers, like biological brains that learn from experience, can learn, identify, and process information to solve complex problems.
Neural Networks, also called connectionist architectures, parallel distributed processors and neuromorphic systems, are based on densely interconnected and parallel distributed architecture of computation. NN do not use traditional programming languages; these networks can programme and train themselves instead. Proponents think of NN technology as the next quantum leap in computing.
The neuron, basic unit of NN, receives a number of inputs, either from the original data or from other neurons. Each input has a corresponding weight attached to it that contributes to the total weightage of the neuron. A value is associated with a neuron. Classical NN usually have the input layer, the output and the hidden layer. Most NNs are either Feed Forward or Recurrent. In Feed Forward networks the signals progress from the input to the hidden and finally to the output layers. In Recurrent networks, signals can travel back and forth, hence creating very complex loops. The output of the network is calculated through an activation function and the output produced by the output layer is considered to be the output of the whole network.
It is possible to program NN by adjusting the weights and connections in the network if one knows the solution before hand, but this defeats the whole purpose of NN in the first place. One way to reach the goal is that while training the network, if it seems that it is getting further off the marks then the weights can be changed randomly and the network starts towards the solution afresh. However, most widely used algorithm for error minimization is back-propagation in which input data is presented to the network several times and the error is computed each time. The error is then back-propagated to the inputs which adjust the inputs in such a way that the error is less that before. The process is repeated again until the error is minimized.
NNs generally are used when the exact relationship between the inputs and the outputs is not known. The inputs and the outputs are gradually “learnt” by the NN accomplished either by supervised or unsupervised process. In supervised learning the network is given sample data that contains inputs and outputs and the network learns relationship between the two after many runs. In unsupervised learning the system is provided with only the inputs and not the outputs. The system then has to decide which features of the data it will use to model the data. Inputs are sometimes also called patterns and if the network recognizes the pattern correctly then it is said to have classified it correctly. The idea of fast thinking machines, able to make decisions based on what they identify, rather than working based on inbuilt programs, is not really new in the history. While NNs are gaining new hype and lot of resources are being pumped into labs and research centres around the world these days, the work to develop NNs had started before the proliferation of conventional computers that work on the Von Neumann architecture. The first artificial neuron was produced back in 1943 when a neurophysiologist Warren McCulloch jointly wrote a breakthrough paper on how a neuron might work. The concept was furthered by Donald Hebb who wrote a book entitled: The Organization of Behaviour in 1949 and explained how neural pathways strengthen each time they are used.
Research in the field of NNs came to an almost standstill after the publication of “Perceptrons” which discussed many limitations of NN. It was not until Japan initiated AI project in the 80s that interest was revived in 1990.
NNs with an ability to derive meaning from complicated data can be used to extract patterns and detect trends that are too complex to be noticed by either humans or conventional computers. A trained NN is as an expert in the category of information it has been given to analyse, for adaptive learning, self organization, and real time operations. NNS have revolutionized the way scientists solve many complex and real-world problems in science, engineering, economics, and business. Applications to solve problems are already developing given the promise neural chips hold for the future. Business is a major field of activity to adapt neural network technology. Since NN technology is best at identifying patterns or trends in data, it is well suited for prediction and forecasting business process including sales forecasting, industrial process control, customer research, data validation, risk management, target marketing — larger problems of larger businesses. Developers are working to create a NN technology for business applications using relational databases and Predictive Model Markup Language. It is safe to predict that NN applications would fit into these business areas and more. There is also a strong potential for using NN for database mining that is searching for patterns implicit within the explicitly stored information in databases.
NN technology is also contributing to other areas of research such as neurology and psychology. It is being used to model parts of living organisms and to investigate the internal mechanisms of the brain. American Defence Advanced Research Projects Agency is working on a project that may enable to recognize the habits and thinking styles of military commanders, human beings, in the battlefield so that some of their routine functions can be automated, freeing them to concentrate on more important tasks. Nasa is using NN technology for reading photographs taken from space and FBI for processing images taken at crime scenes. Another area where the scientists see the technology being used in near future is in interactive entertainment. Video games of the future could have characters with almost human intelligence, capable of not only understanding and acting on commands but also able to think and react. Scientists also argue that perhaps one day “conscious” networks might be produced.
NN are finding their application in unexpected fields. Efforts are also under way in combing this field with Bayesian Networks, for statistical reasoning. Another interesting idea is that of Neuromorphic computing, directly implementing the NN into hardware.
The computing world has a lot to gain from NN. Their ability to learn by experience makes them very flexible and powerful. They are also very well suited for real time systems because of their fast response and computational times which are due to their parallel architecture. Advancement beyond current commercial applications appears to be possible, and research is advancing the field on many fronts.
Raymond Kurzweil’s idea in which he had predicted about merging of mind and machines in future is no more fictional. The future is already upon us.
The writer
|
|
|
|
