Fifth National Conference On RTICT2012 , Bannari Amman Institute Of Technology,Sathyamangalam-638401, 2-3 April 2012 STUDY OF NEURAL NETWORKS AND APPLICATION Sasikumar Gurumurthy, Tamilpriya.M and Renukadevi.R VIT University, Vellore. ABSTRACT An classification, through a learning Artificial Neural process. Learning in biological an systems involves adjustments to processing the synaptic connections that paradigm that is inspired by the exist between the neurons. This way biological nervous systems, is true of ANNs as well. Network (ANN) is information such as the brain, process Introduction: information. The key element of this paradigm is the 1.1 Why use neural networks? novel structure of the information processing system. It Neural networks, with their is remarkable ability to derive composed of a large number of meaning highly interconnected processing imprecise data, can be used to elements (neurons) working in extract patterns and detect trends unison that are too complex to be noticed to solve specific from complicated or problems. ANNs, like people, by learn by example. An ANN is computer techniques. A trained configured for a specific neural network can be thought of application, such as pattern as an "expert" in the category of data information it has been given to recognition or 1083 either humans or other Fifth National Conference On RTICT2012 , Bannari Amman Institute Of Technology,Sathyamangalam-638401, 2-3 April 2012 analyses. This expert can then be 1.3 Neural networks versus used to provide projections given conventional computers new situations of interest and Neural networks take a different approach to problem solving than that of conventional 1.2 Advantages: computers. 1. Adaptive learning: An computers Conventional use an algorithmic ability to learn how to do approach i.e. the computer follows tasks based on the data a set of instructions in order to given for training or initial solve experience. specific steps that the computer 2. Self-Organization: a problem. Unless the needs to follow are known the An ANN can create its own computer organization or problem. That restricts the problem the solving capability of conventional receives computers to problems that we representation information of it the to solve. But computers would be 3. Real Time Operation: ANN may solve already understand and know how during learning time. computations cannot so much more useful if they could be do things that we don't exactly carried out in parallel, and know how to do. special hardware devices are being designed and Neural manufactured which take information in a similar way the advantage human brain does. The network is of this capability. networks process composed of a large number of 1084 Fifth National Conference On RTICT2012 , Bannari Amman Institute Of Technology,Sathyamangalam-638401, 2-3 April 2012 highly interconnected processing collects signals from others elements (neurons) working in through a host of fine parallel structures called dendrites. to solve a specific problem. Neural networks learn by The example. be spikes of electrical activity programmed to perform a specific through a long, thin stand task. be known as an axon, which selected carefully otherwise useful splits into thousands of time is wasted or even worse the branches. At the end of network might be functioning each branch, a structure incorrectly. The disadvantage is called a synapse converts that because the network finds out the activity from the axon how to solve the problem by itself, into electrical effects that its operation can be unpredictable. inhibit or excite activity They The cannot examples must from 2. Human and Artificial Neurons neuron the electrical - investigating the similarities sends out axon into effects that inhibit or excite activity in 2.1 How the Human Brain the Learns? .When a neuron receives Much is sufficiently large compared unknown about how the with its inhibitory brain trains itself to process so theories abound. the human In neurons excitatory input that is still information, connected brain, a typical neuron 1085 Fifth National Conference On RTICT2012 , Bannari Amman Institute Of Technology,Sathyamangalam-638401, 2-3 April 2012 to deduce the essential features of neurons and their interconnections. We then typically program a computer to simulate these features. However because knowledge of our neurons is incomplete and our computing power is limited, our models are .A Components of a neuron necessarily gross idealizations of real networks of neurons. 3. Architecture of neural networks Figure 2.1.B The Synapse 3.1 Feed-forward networks 2.2 From Human Neurons to Feed-forward ANNs (figure 1) allow signals to travel one way Artificial Neurons only; from input to output. We conduct these There is no feedback (loops) i.e. neural networks by first trying the output of any layer does not 1086 Fifth National Conference On RTICT2012 , Bannari Amman Institute Of Technology,Sathyamangalam-638401, 2-3 April 2012 affect that same layer. Feed- is often used to denote feedback forward connections ANNs tend to be in single-layer organizations. straight forward networks that associate inputs with outputs. They are extensively used in pattern recognition. This type of organization is also referred to as bottom-up or top-down3.2 Feedback networks Figure 4.1 An example of a simple feed forward network Feedback networks (figure 1) can have signals traveling in both directions by introducing loops in the network. Feedback networks are very powerful and can get extremely complicated. Feedback networks are dynamic; their 'state' is changing continuously until they reach an equilibrium point. They remain at the equilibrium point until the Figure4.2 An example of a input complicated network. changes and a new equilibrium needs to be found. Feedback architectures are also referred to as interactive or recurrent, although the latter term 1087 Fifth National Conference On RTICT2012 , Bannari Amman Institute Of Technology,Sathyamangalam-638401, 2-3 April 2012 weights on the connections between the input and the hidden units. The behavior of the output units depends on the activity of the hidden units and the weights between the hidden and output units This simple type of network is interesting because the hidden Figure 4.1 An example of a simple units are free to construct their own feed forward network representations of the input. The weights between the input and hidden units determine when each hidden unit is active, and so by modifying these weights, a hidden unit can choose what it represents .We also distinguish single-layer and multi-layer architectures. The single-layer organization, in which all units are connected to one another, constitutes the most general case and is of more potential computational power than hierarchically structured multi-layer Figure 4.2 An example of a organizations. complicated network. In multi-layer networks, units are often numbered 1088 Fifth National Conference On RTICT2012 , Bannari Amman Institute Of Technology,Sathyamangalam-638401, 2-3 April 2012 by layer, instead of following a global numbering. 3.4 Perceptrons The most influential work on neural nets in the 60's went Figure 4.4 under the heading of 'Perceptrons' a term coined by In 1969 Minsky and Papert wrote a Frank Rosenblatt. The book in which they described the Perceptrons (figure 4.4) turns limitations of single layer out to be an MCP model (neuron Perceptrons. The impact that the with weighted inputs) with some book had was tremendous and additional, fixed, pre-- caused a lot of neural network processing. Units labeled A1, A2, researchers to loose their interest. Aj , Ap are called association The book was very well written units and their task is to extract and showed mathematically that specific, localized featured from single layer Perceptrons could not do some basic pattern recognition the input images. operations like determining the Perceptrons mimic the basic idea parity of a shape or determining behind visual whether a shape is connected or system. They were mainly used in not. What they did not realized, pattern recognition even though until the 80's, is that given the the mammalian their capabilities extended a lot But to give you some more more. specific examples; ANN are also used in the following specific 1089 Fifth National Conference On RTICT2012 , Bannari Amman Institute Of Technology,Sathyamangalam-638401, 2-3 April 2012 paradigms: recognition of speakers 4. in communications; diagnosis of networks hepatitis; recovery of Applications of neural 4.1 Neural Networks in Practice telecommunications from faulty of Given this description of words; neural networks and how they undersea mine detection; texture work, what real world applications analysis; three-dimensional object are recognition; word networks have broad applicability recognition; and facial recognition. to real world business problems. In software; interpretation multimeaning Chinese hand-written they suited fact, they have 4.2 Neural networks in successfully medicine Neural already been applied in many industries. Since neural networks are best at identifying patterns or Artificial Neural Networks (ANN) for? are currently a trends in data, they are well suited 'hot' for prediction or forecasting needs research area in medicine and it is including: believed that they will receive extensive app appropriate training, multilevel Perceptrons can Sales forecasting do Industrial these operations. process control Customer research Data validation Risk management 1090 Fifth National Conference On RTICT2012 , Bannari Amman Institute Of Technology,Sathyamangalam-638401, 2-3 April 2012 Target marketing lication to 4.2.1 Modeling and Diagnosing biomedical systems in the next few the Cardiovascular System years. At the moment, the research Neural Networks are used is mostly on Modeling parts of the experimentally human body and recognizing human diseases from Diagnosis can be achieved by model cardiovascular building Neural networks are ideal in to a the system. model of recognizing diseases using scans cardiovascular since there is no need to provide a individual and comparing it with specific algorithm on how to the identify measurements the disease. Neural real system time of the an physiological taken from the networks learn by example so the patient. If this routine is carried out details of how to recognize the regularly, disease are not needed. What is medical conditions can be detected needed is a set of examples that are at an early stage and thus make the representative of all the variations process of combating the disease of the disease. The quantity of much easier. potential harmful examples is not as important as the A model of an individual's 'quantity'. The examples need to be cardiovascular system must mimic selected very carefully if the the system is to perform reliably and relationship among physiological variables (i.e., heart efficiently. rate, systolic and diastolic blood pressures, and breathing rate) at different physical activity levels. If a 1091 model is adapted to an Fifth National Conference On RTICT2012 , Bannari Amman Institute Of Technology,Sathyamangalam-638401, 2-3 April 2012 individual, then it becomes a conditions by fusing the data from model of the physical condition of the individual biomedical sensors. that individual. The simulator will 4.2.2 Electronic noses have to be able to adapt to the ANNs features of any individual without are to used the supervision of an expert. This experimentally implement calls for a neural network. electronic noses. Electronic noses have several potential applications Another reason that justifies in telemedicine. Telemedicine is the use of ANN technology is the the practice of medicine over long ability of ANNs to provide sensor distances via a communication fusion which is the combining of link. The electronic nose would values different identify sensors. Sensor fusion enables the surgical ANNs identified odors would then be from to several learn complex odors in the remote environment. These relationships among the individual electronically transmitted sensor another where values, which would site a to door otherwise be lost if the values were generation system would recreate individually analyzed. In medical them. Because the sense of smell modeling this can be an important sense to the implies that even though each surgeon, telesmell would enhance sensor in a set may be sensitive telepresent surgery. and diagnosis, only to a specific physiological 4.2.3 Instant Physician variable, ANNs are capable of detecting complex medical An application developed in the mid-1980s called the "instant 1092 Fifth National Conference On RTICT2012 , Bannari Amman Institute Of Technology,Sathyamangalam-638401, 2-3 April 2012 physician" associative trained an memory auto using neural networks for database neural mining that is, searching for network to store a large number of patterns implicit within the medical records, each of which explicitly stored information in includes information on databases. Most of the funded diagnosis, and work in this area is classified as treatment for a particular case. proprietary. Thus, it is not possible After training, the net can be to report on the full extent of the presented with input consisting of work going on. Most work is a set of symptoms; it will then find applying neural networks, such as the the Hopfield-Tank network for symptoms, full stored pattern that optimization and scheduling. represents the "best" diagnosis and treatment. 4.3.1 Marketing 4.3 Neural Networks in business There application Business is a diverted field is a which marketing has been integrated with a neural network with several general areas of system. The Airline Marketing specialization such as accounting Tactician (a trademark abbreviated or financial analysis. Almost any as AMT) is a computer system neural network application would made fit into one business area or of technologies financial analysis. There is some various intelligent including expert systems. A feed forward neural potential for using neural networks network is integrated with the for business purposes, including AMT and was trained using back- resource allocation and scheduling. propagation to assist the marketing There is also a strong potential for 1093 Fifth National Conference On RTICT2012 , Bannari Amman Institute Of Technology,Sathyamangalam-638401, 2-3 April 2012 control of airline seat allocations. system was developed by the The adaptive neural approach was Nestor Company. This system was amenable expression. trained with 5048 applications of application's which 2597 were certified. The to Additionally, rule the environment changed rapidly and data constantly, borrower which required a related to property qualifications. and In a solution. conservative mode the system The system is used to monitor and agreed on the underwriters on 97% recommend booking advice for of the cases. In the liberal model each departure. S impact on the the system agreed 84% of the profitability of an airline and can cases. This is system run on an provide a technological advantage Apollo DN3000 and used 250K for users of the system. [Hutchison memory while processing a case & Stephens, 1987]4.3.2 Credit file in approximately 1 sec. continuously adaptive Evaluation 5. Conclusion The HNC Company, founded by Robert Hecht-Nielsen, has The computing world has a developed several neural network lot to gain from neural networks. applications. One of them is the Their ability to learn by example Credit Scoring systems which makes them very flexible and increase the profitability of the powerful. Furthermore there is no existing model up to 27%. The need to devise an algorithm in HNC neural systems were also order to perform a specific task; applied to mortgage screening. i.e. there is no need to understand Neural automated the internal mechanisms of that mortgage insurance under writing task. They are also very well suited network 1094 Fifth National Conference On RTICT2012 , Bannari Amman Institute Of Technology,Sathyamangalam-638401, 2-3 April 2012 REFERENCE for real time systems because of their fast response and 1. Kandel E.R. and Schwartz J.H., Principles of neural science, Elsevier/North Holland (1981) 2. Hungenahally S.K. and Jain L.C., Neuro intelligent systems, BPB Publication/India (1994) 3. Joey Rogers Analysis of artificial neural network, Academic press/USA (1997) 4. Abraham Silberschatz, Peter Baer Galvin and Greg Gagne., Operating system concepts John Wiley & sons/Asia, (2004) computational times which are due to their parallel architecture. Neural networks also contribute to other areas of research such as neurology and psychology. They are regularly used to model parts of living investigate organisms the and to internal mechanisms of the brain. Perhaps the most exciting aspect of neural networks is the possibility that some day 'conscious' networks might be produced. There is a number of scientists arguing that consciousness is a 'mechanical' property and that 'conscious' neural networks are a realistic possibility. Finally, I would like to state that even though neural networks have a huge potential we will only get the best of them when they are much information has a direct integrated with computing, AI, fuzzy logic and related subjects. 1095