SEMINAR REPORT ON

Artificial Neural Network

SUBMITTED BY:SURAJ AGARWAL 107170 EC-3 B-TECH 3RD YEAR /5TH SEM

Index
1) Introduction...4 2) Definition..4 3) Structure of human brain..5 4) Neurons....5 5) Basics of ANN model..7 6) Artificial neural network..8 6.1) How ANN differs from Conventional Computer...9 6.2) ANN vs Von Neumann Computer..9 7) 8) Perceptron...10 Learning laws..11 8.1) Hebbs Rule.12 8.2) Hopefields Rule..12 8.3) Delta Rule12 8.4) Gradient Descent Rule.13 8.5) Kohonens Learning Rule....13 9) Basic structure of ANN13 10) Network architectures...14 10.1) Single Layer Feed Forward ANN14 10.2) Multi Layer Feed Forward ANN.15 10.3) Recurrent ANN16 11) Learning of ANN..17 11.1) Learning with a Teacher...17 11.2) Learning without a Teacher..18 11.3) Learning tasks...20
2

12) Control.22 13) Adaptation22 14) Generalization..23 15) Probabilistic ANN23 16) Advantages of ANN.24 17) Limitations of ANN.25 18) Applications of ANN...25 19) Conclusion...29 20) References...30

1.

Introduction <3><2>

Ever since eternity, one thing that has made human beings stand apart from the rest of the animal kingdom is, its brain .The most intelligent device on earth, the Human brain is the driving force that has given us the ever-progressive species diving into technology and development as each day progresses. Due to his inquisitive nature, man tried to make machines that could do intelligent job processing, and take decisions according to instructions fed to it. What resulted was the machine that revolutionized the whole world, the Computer (more technically speaking the Von Neumann Computer). Even though it could perform millions of calculations every second, display incredible graphics and 3-dimentional animations, play audio and video but it made the same mistake every time. Practice could not make it perfect. So the quest for making more intelligent device continued. These researches lead to birth of more powerful processors with high-tech equipments attached to it, super computers with capabilities to handle more than one task at a time and finally networks with resources sharing facilities. But still the problem of designing machines with intelligent self-learning, loomed large in front of mankind. Then the idea of initiating human brain stuck the designers who started their researches one of the technologies that will change the way computer work Artificial Neural Networks.
2.

Definition<3><6>

Neural Network is the specified branch of the Artificial Intelligence. In general, Neural Networks are simply mathematical techniques designed to accomplish a variety of tasks. Neural Networks uses a set of processing elements (or nodes) loosely analogues to neurons in the brain (hence the same, neural networks). These nodes are interconnected in a network that can then identify patterns in data as it is exposed to the data. In a sense, the network learns from the experience just as people do. Neural networks can be configured in various arrangements to perform a range of tasks including pattern recognition, data mining, classification, and process modeling.
4

3.

Structure of human brain <3><2>

biological structure of

As stated earlier, Neural Networks is very much similar to the human Brain. Following is the biological structure of brain is given. Sequential Functions: Rules Concepts Calculations Expert Systems Learn by Rules Functions of Brain Parallel Functions: Images Pictures Control

Neural Networks Learn by experience

As shown in table, left part of the brain consists of rules, concepts and calculations. It follows Rule Based Learning and hence solves the problem by passing them through rules. It has Sequential pairs of Neurons. Therefore, this part of brain is similar to the expert systems. Right part of the brain, as shown below in the figure; consist of functions, images, pictures, and controls. It follows parallel learning and hence learns through experience. It has parallel pairs of Neurons. Therefore, this brain is similar to the Neural Network.
4.

Neurons <3><2>

The conceptual constructs of a neural network stemmed from our early understanding of the human brain. The brain is comprised of billion and billions of interconnected neurons (some experts estimate upwards of 1011 neurons in the human brain). The fundamental building blocks of this massively parallel cellular structure are really quite simply when studied in isolation. A neuron receives incoming electrochemical signals from its dendrites and collects these signals at the neuron nucleus. The neuron nucleus has a internal threshold that determines if neuron itself fires in response to the incoming information. If the combined incoming signals exceeds this threshold then neuron fires and an electrochemical signal is sent to all neurons connected to the firing neuron on its output
5

connections or axons. Otherwise the incoming signals are ignored and the neuron remains dormant. There are many types of neurons or cells. From a neuron body (soma) many fine branching fibers, called dendrites, protrude. The dendrites conduct signals to the soma or cell body. Extending from a neurons soma, at a point called axon hillock (initial segment), is a long giber called an axon, which generally splits into the smaller branches of axonal arborization. The tips of these axon branches (also called nerve terminals, end bulbs, telondria) impinge either upon the dendrites, somas or axons of other neurons or upon effectors.

The axon-dendrite (axon-soma, axon-axon) contact between end bulbs and the cell it impinges upon is called a synapse. The signal flow in the neuron is (with some exceptions when the flow could be bi-directional) from the dendrites through the soma converging at the axon hillock and then down the axon the end bulbs. A neuron typically has many dendrites but only a single axon. Some neurons lack axons, such as the amacrine cells.

5.

Basics of Artificial Neural Models <3><2>

The human brain is made up of computing elements, called neurons, coupled with sensory receptors (affecters) and effectors. The average human brain, roughly three pounds in weight and 90 cubic inches in volume, is estimated to contain about 100 billion cells of various types.
7

A neuron is a special cell that conducts and electrical signal, and there are about 10 billion neurons in the human brain. The remaining 90 billion cells are called glial or glue cells, and these serve as support cells for the neurons. According to a simplified account, the human brain consists of about ten billion neurons -and a neuron is, on average, connected to several thousand other neurons. By way of these connections, neurons both send and receive varying quantities of energy. One very important feature of neurons is that they don't react immediately to the reception of energy. Instead, they sum their received energies, and they send their own quantities of energy to other neurons only when this sum has reached a certain critical threshold. The brain learns by adjusting the number and strength of these connections. Even though this picture is a simplification of the biological facts, it is sufficiently powerful to serve as a model for the neural net. The motivation for artificial neural network (ANN) researches is the belief that a humans capabilities, particularly in real-time visual perception, speech understanding, and sensory information processing and in adaptively as well as intelligent decision making in general, come from the organizational and computational principles exhibited in the highly complex neural network of the human brain. Expectations of faster and better solution provide us with the challenge to build machines using the same computational and organizational principles, simplified and abstracted from neurobiological of the brain.

Artificial Neural Network Model

8

6. Artificial Neural Network<3><2>

Artificial neural network (ANNs), also called parallel distributed processing systems (PDPs) and connectionist systems, are intended for modeling the organization principles of the central neurons system, with the hope that the biologically inspired computing capabilities of the ANN will allow the cognitive and sensory tasks to be performed more easily and more satisfactory than with conventional serial processors. Because of the limitation of serial computers, much effort has devoted to the development of the parallel processing architecture; the function of single processor is at a level comparable to that of a neuron. If the interconnections between the simplest fine-grained processors are made adaptive, a neural network results. ANN structures, broadly classified feedback), as recurrent (involving feedback) or nonnumerous processing

recurrent

(without

have

elements (also dubbed neurons, neurodes, units or cells) and connections (forward and backward interlayer connections between neurons in different layers, forward and backward interlayer connections or lateral connections between neurons in the same layer, and selfconnections between the input and output layer of the same neuron. Neural networks may not have differing structures or topology but are also distinguished from one another by the way they learn, the manner in which computations are performed (rule-based, fuzzy, even nonalorithmic), and the component characteristic of the neurons or the input/output description of the synaptic dynamics).These properties. Since the components and connections and their packaging under stringent spatial constraints make the system large-scale, the role of graph theory, algorithm, and neuroscience is pervasive. networks are required to perform significant processing tasks through collective local interaction that produces global

6.1 How Neural Networks differ from Conventional Computer?

Neural Networks perform computation in a very different way than conventional computers,
9

where a single central processing unit sequential dictates every piece of the action. Neural Networks are built from a large number of very simple processing elements that individually deal with pieces of a big problem. A processing element (PE) simply multiplies an output value (table lookup). The principles of neural computation come from the massive processing tasks, and from the adaptive nature of the that interconnected the PEs. parameters (weights)

6.2 Similarities and difference between neural net and von Neumann computer
Neural net Trained adjusting ( learning the by example) by Von Neumann computer Programmed with instruction (if-then analysis based on logic).

connection

strengths,

threshold and structure. Parallel(discrete asynchronous or continuous), digital, Sequential or serial

synchronous(with a clock)

May be fault-tolerant because of Distributed Not fault-tolerant. representation and Large -scale redundancy Self-organization during learning Memory & processing elements collocated Knowledge stored is adaptable Processing is anarchic and cycle time, which governs the processing speed, is in milliseconds range Software dependent Memory and processing elements separate Knowledge stored in address memory Processing is autocratic and cycle time, corresponds to processing one step of program in the CP during one clock cycle, is in nanosecond range
7.

Perceptron<2>

At the heart of every Neural Network is what is referred to as the perceptron (sometimes
10

called processing element or neural node) which is analogous to the neuron nucleus in the brain. The second layer that is very first hidden layer is known as perceptron. As was the case in the brain the operation of the perceptron is very simple; however also as is the case in the brain, when all connected neurons operate as a collective they can some very powerful learning capacity. Input signals are applied to the node via input connection (dendrites in the case of the brain.) The connections have strength which changes as the system learns. In neural networks the strength of the connections are referred to as weights. Weights can either excite or inhibit the signal. particular weights. At the perceptron all weighted input are summed. This sum value is than passed to a scaling function. The selection of scaling function is part of the neural network design. The structure of perceptron (Neuron Node) is as follow. transmission of the incoming Mathematically incoming signals values are multiplied by the value of those provide

Perceptron
11

8.

Learning Laws<2><3>

Many learning laws are in common use. Most of these are some sort of variation of the best known and oldest learning laws, hebbs rule. Research into different learning functions continues as new ideas routine show up in trade publication. Some researches have the modeling of biological learning as their main objective. Others are experimenting with adaptation of their perceptions of how nature handles learning. Either way, mans understanding of how neural processing actually works is very limited. Learning is certainly more complex than the simplification represented by the learning laws currently develop. A few of the major laws are presented as examples. 8.1 Hebbs Rule The first, and undoubtedly the best known, learning rule were introduced by Donald Hebb. The description appeared in his book the Organization of behavior in 1949. His basic rule is: if a neuron receives an input from strengthened. 8.2 Hopfield Law It is similar to Hebbs rule with the exception that it specifies the magnitude of the strengthening or weakening. It states, if the desired output and the input are both active and both inactive, increment the connection weight by the learning rate, otherwise decrement the weight by the learning rate. 8.3 The Delta Rule This rule is a further variation of Hebbs Rule. It is one of the most This rule is based on the simple idea of continuously commonly used. another neuron and if both are highly active (mathematically have the same sign), the weight between the neurons should be

modifying the strengths of the input

connections to reduce the difference (the delta) between the desired output value and the actual output of a processing element. Their rule changes the synaptic weights in the way that minimizes the mean squared error of the network. This rule is also referred to as
12

windrows-Hoff Learning rule and the least mean square (LMS) Learning Rule. The way that the Delta Rule works is that the delta rule error in the output layer is transformed by the derivative of the transfer function and is then used in the previous neural layer to adjust input connection weights. In other words, the back-propagated into previous layers one layer at a time. The process of back-propagating the network errors continues until the first layer is reached. The network typed called feed forward; back-propagation derives its name from this method of computing the error term. When using the delta rule, it is important to ensure that the input data set is well randomized. Well-ordered or structured presentation of the training set can lend to a network, which cannot converge to the desired accuracy. If that happens, then network is incapable of learning the problem. 8.4 The Gradient Descent Rule This rule is similar to the Delta rule in that the derivatives of the transfer used to modify the delta error before it is applied to the function is still

connection weights. Here,

however, an additional proportional constant tied to the learning rate is appended to the final modifying factor acting upon the weights. This rule is commonly used, even though it converges to a point of stability very slowly. It has been shown that different learning rates for different layers of network help the learning process converge faster. In these tests, the learning rates for those layers close to the output were set lower than those layers near the input. This especially important for applications where the input data is not derived from a strong underlying model. 8.5 Kohonens Learning Law The procedure, developed by Teuvo Kohonen, was inspired by learning in biological systems. In this procedure, the processing elements complete for the opportunity to learn, or update their weights. The processing element with the largest output is declared the winner and has the capabilities of inhibiting its competitors as well as exciting its neighbors. Only the winner is permitted an output, and only the winner plus its neighbors are allowed to adjust their connection weights. Further, the size of the neighborhood
13

can vary during

the training period. The usual paradigm is to start with a

larger definition of the

neighborhood, and narrow in as the training process proceeds. Because the winning element is defined as the one that has the closest match to the input pattern, Kohonen networks model the distributed of the data and is sometimes referred to as selforganizing maps or self-organizing topologies.
9.

Basic Structure of artificial neural network<2>

Input layer: The bottom layer is known as input neuron network in this case x1 to x5 are input layer neurons. Hidden layer: The in-between input and output layer the layers are known as hidden layers where the knowledge of past experience / training is kept. Output Layer: The topmost layer which gives the final output. In this case z1 and z2 are output neurons.

Basic Structure Of Artificial Neural Network

10. Network architectures<1>

14

1). Single layer feed forward networks: In this layered neural network the neurons are organized in the form of layers. In this simplest form of a layered network, we have an input layer of source nodes those projects on to an output layer of neurons, but not vise-versa. In other words, this network is strictly a feed forward or acyclic type. It is as shown in figure:

Such a network is called single layered network, with designation single later referring to the o/p layer of neurons.

2). Multilayer feed forward networks: The second class of the feed forward neural network distinguishes itself by one or more hidden layers, whose computation nodes are correspondingly called neurons or units. The function of hidden neurons is intervened between the external i/p and the network o/p in some useful manner. The ability of hidden neurons is to extract higher order statistics is particularly valuable when the size of i/p layer is large. The i/p vectors are feed forward to 1 st hidden layer and this pass to 2 nd hidden layer and so on until the last layer i.e. output layer, which gives actual network response.

15

3). Recurrent networks: A recurrent network distinguishes itself from feed forward neural network, in that it has least one feed forward loop. As shown in figures output of the neurons is fed back into its own inputs is referred as self-feedback. A recurrent network may consist of a single layer of neurons with each neuron feeding its output signal back to the inputs of all the other neurons. Network may have hidden layers or not.

16

11.

Learning of ANNS <2/p-42><1/p-57>

ability of the

The property that is of primary significance for a neural network is the

network to learn from environment, and to improve its performance through learning. A neural network learns about its environment through an interactive process of adjustment applied to its synaptic weights and bias levels. Network becomes more knowledgeable about its environment after each iteration of the learning process. 11.1 Learning with a teacher: 1). Supervised learning: the learning process in which the teacher teaches the network by giving the network the knowledge of environment in the form of sets of the inputs outputs pre-calculated examples. As shown in figure

17

Neural network response to inputs is observed and compared with the

predefined output.

The difference is calculated refer as error signal and that is feed back to input layers neurons along with the inputs to reduce the error to get the perfect response of the network as per the predefined outputs. For example, when learning a language, we hear the sound of a word from a teacher. The sound is stored in the memory banks of our brain, and we try to reproduce the sound. When we hear our own sound, we mentally compare it (actual output) with the stored sound (target sound) and note the error. If error is large, we try again and again until it becomes significantly small; then we stop.

11.2 Learning without a teacher: Unlike supervised learning, in unsupervised learning, the learning process takes place without teacher that is there are no examples of the functions to be learned by the network. 1). Reinforcement learning / neurodynamic programming : In reinforcement learning, the learning of an input output mapping is performed through continued interaction with environment in order to minimize a scalar index of performance. As shown in figure.
18

In reinforcement learning, because no information on way the right output should be provided, the system must employ some random search strategy so that the space of plausible and rational choices is searched until a correct answer is found. Reinforcement learning is usually involved in exploring a new environment when some knowledge (or subjective feeling) about the right response to environmental inputs is available. The system receives an input from the environment and process an output as response. Subsequently, it receives a reward or a penalty from the environment. The system learns from a sequence of such interactions. In reinforced learning it uses a critic instead of a teacher which helps to indicate whether the actual input is same as target input or not. The critic does not present the target output to the network but presents only a pass/fail indication. Thus the error signal produced is binary: pass or fail. If the teacher indication is fail the network readjusts its parameter and tries again and again until it gets its output response right. During this process there is no indication if the output response is moving in the right direction or how close to the correct output it is. 2). Unsupervised learning: In unsupervised or self-organized learning there is no external
19

teacher or critic to over see the learning process. As indicated in figure.

Rather provision is made for a task independent measure of the quality of the representation that the network is required to learn and the free parameters of the network are optimized with respect to that measure. Once the network has become tuned to the statistical regularities of the input automatically. For example, show a person a set of different objects. Then ask him/her to separate them into groups, such that objects in a group have one or more common features that distinguish them from another group. When this is done, show the same person another object and ask him/her to place the object in one of the groups. If it does not belong to any of the existing groups, a new group may be formed. Even though unsupervised learning does not require a teacher, it requires guidelines to determine how it will form groups. 11.3 Learning tasks Pattern recognition: Humans are good at pattern recognition. We can recognize the familiar face of the person even though that person has aged since last encounter, identifying a familiar person by his voice on telephone, or by smelling the fragments comes to know the food etc. Pattern recognition is formally defined as the process where by a received pattern/signal is
20

data, it develops the ability to form internal

representation for encoding features of the input and there by to create the new class

assigned to one of a prescribed number of classes. A neural network performs pattern recognition by first undergoing a training session, during which the network is repeatedly present a set of input pattern along with the category to which each particular pattern belongs. Later, a new pattern is presented to the network that has not been seen before, but which belongs to the same pattern category used to train the network. The network is able to identify the class of that particular pattern because of the information it has extracted from the training data. Pattern recognition performed by neural network is statistical in nature, with the pattern being represented by points in a multidimensional decision space. The decision space is divided into regions, each one of which is associated with class. The decision boundaries are determined by the training process.

As shown in figure: in generic terms, pattern-recognition machines using neural network may take two forms. 1). To extract features through unsupervised network. 2). Features pass to supervised network for pattern classification to give final output.

21

12. Control<2><3>

The control of a plant is another learning task that can be done by a neural network; by a plant we mean a process or critical part of a system that is to be maintained in a controlled condition. The relevance of learning to control should not be surprising because, after all, the human brain is a computer, the output of which as a whole system are
22

actions. In the context of control, the brain is living proof that it is possible to build a generalized controller that takes full advantages of parallel distributed hardware, can control many thousands of processes as done by the brain to control the thousands of muscles.

`13. Adaptation<2><3>
The environment of the interest is no stationary, which means that the statistical parameters of the information bearing generated by the environment vary with the time. In situation of the kind, the traditional methods of supervised may learning may prove to be inadequate because the network is not equipped with the necessary means to track the statistical variation of the environment in which it operates. To overcome these shortcomings, it is desirable for a neural network to continually adapt its free parameters to variation in the incoming signals in a real time fashion. Thus an adaptive system responds to every distinct input as a novel one. In other words the learning process encountered in the adaptive system never stops, with learning going on while signal processing is being performed by the system. This form of learning is called continuous learning or learning on the fly.
23

14.

Generalization <2><3>

In back propagation learning we typically starts with a training sample and uses the back propagation algorithm to compute the synaptic weights of a multiplayer preceptor by loading (encoding) as many as of the training example as possible into the network. The hope is that the neural network so design will generalize. A network is said generalize well when the input output mapping computed by the network is correct or nearly so for the test data never used in creating or training the network; the term generalization is borrowed from psychology. A neural network that is design to generalize well will produced a correct input output mapping even when the input is slightly different from the examples used to train the network. When however a neural network learns too many input output examples the network may end up memorizing the training data. It may do so by finding a feature that is present in training data but not true for the underlining function that is to be modeled. Such a phenomena is referred to as an over fitting or over training. When the network is over trained it looses the ability to generalize between similar input output pattern.
15. The

probabilistic neural network<3>

network (PNN). In

Another multilayer feed forward network is the probabilistic neural addition to the input layer, the PNN has two hidden major difference from a feed forward

layers and an output layer. The

network trained by back propagation is that it can PNN is

be constructed after only a single pass of the training exemplars in its original form and two passes is a modified version. The activation function of a neural in the case of the patterns. statistically derived from estimating of probability density functions (PDFs) based on training

24

16. Advantages of Neural

Networks <3>

1) Networks start processing the data without any preconceived hypothesis. They start random with weight assignment to various input variables. Adjustments are made based on the difference between predicted and actual output. This allows for unbiased and better understanding of data. 2) Neural networks can be retained using additional input variables individuals. Once trained they can be called on to predict in a new patient. 3) There are several neural network models available to choose from in a particular problem. 4) Once trained, they are very fast. 5) Due to increased accuracy, results in cost saving. 6) Neural networks are able to represent any functions. Therefore they are called Universal Approximators. 7) Neural networks are able to learn representative examples by back propagating errors. and number of

25

17. Limitations of Neural

Network <3>

Low Learning Rate:- For problems requiring a large and complex network

architecture or having a large number of training examples, the time needed to train the network can become excessively long.
Forgetfulness : - The network tends to forget old training e xamples as it is

presented with new ones. A previously trained neural network that must be updated with new information must be trained using the old and new examples - there is currently no known way to incrementally train the network.
Imprecision: - Neural networks do not provide precise

numerical answer, but

rather relate an input pattern to the most p robable output state.

Black box approach: - Neural networks can be trained to transform an input pattern

to output but provide no insights to the physics behind the transformation.

Limited Flexibility: - The ANNS is designed and implemented for only one

particular system. It is not applicable to another system.

18.

Application of Artificial Neural Network <3>

neural networks the

In parallel with the development of theories and architectures for

scopes for applications are broadening at a rapid pace. Neural networks may develop intuitive concepts but are inherently ill suited for implementing rules precisely, as in the case of rule based computing. Some of the decision making tools of the human brain such as the seats of consciousness, thought, and intuition, do not seem to be within our capabilities for comprehension in the near future and are dubbed by some to be essentially no algorithmic. Following are a few applications where neural networks are employed presently:
1) Time Series Prediction: Predicting the future has always been one of humanitys

desires. Time series measurements are the means for us to characterize and understand a system and to predict in future behavior. Gershenfield and weighed defined three goals for time series analysis: forecasting, modeling, and
26

characterization.

Forecasting

is

predicting

the

short-term evolution of the system. Modeling involves finding a description that accurately captures the features of the long-term behavior. The goal of characterization is to determined the fundamental properties of the system, such as the degrees of freedom or the amount of randomness. The traditional methods used for time series prediction are the moving average (ma), autoregressive (ar), or the combination of the two, the ARMA model. Neural network approaches produced some of the best short-term predictions. However, methods that reconstruct the state space by time delay embedding and develop a representation for the geometry in the systems state space yielded better longer-term predictions than neural networks in some cases.
2) Speech Generation : One of the earliest successful applications of the back

propagation algorithm for training multiplayer feed forward networks were in a speech generation system called NET talk, developed by Sejnowski and Rosenberg. Net talk is a fully connected layered feed forward network with only one hidden layer. It was trained to pronounce written English text. Turning a written English text into speech is a difficult task, because most phonological rules have exceptions that are context-sensitive. Net talk is a simplest network that learns the function in several hours using exemplars.
3) Speech Recognition : Kohonen used his self-organizing map for inverse

problem to that addressed by Net talk: speech recognition. He developed a phonetic typewriter for the Finnish language. The phonetic typewriter takes as input a speech as input speech and converts it into written text. Speech recognition in general is a much harder problem that turning text into speech. Current state-of-the-art English speech recognition systems are based on hidden Markov Model (HMM). The HMM, which is a Markov process; consist of a number of states, the transitions between which depend on the occurrence of some symbol.
4) Autonomous Vehicle Navigation: Vision-based autonomous vehicle and robot

guidance have proven difficult for algorithm-based computer vision methods, mainly because of the diversity of the unexpected cases that must be explicitly
27

dealt with in the algorithms and the real-time constraint. Pomerleau successfully demonstrated the potential of neural networks for overcoming these difficulties. His ALVINN (Autonomous Land Vehicle in Neural Networks) set a worked record for autonomous navigation distance. After training on a two-mile stretch of highway, it drove the CMU Navlab, equipped with video cameras and laser range sensors, for 21.2 miles with an average speed of 55 mph on a relatively old highway open to normal traffic. ALVINN was not distributed by passing cars while it was driven autonomously. ALVINN nearly doubled the previous distance world record for autonomous navigation. A network in ALVINN for each situation consists of a single hidden layer of only four units, an output layer of 30 units and a 30 X 32 retina for the 960 possible input variables. The retina is fully connected to the hidden layer, and the hidden layer is fully connected to the output layer. bipartite graphs.
5) Handwriting Recognition: Members of a group at AT&T Bell Laboratories have

The

graph

of the feed forward network is a node-coalesced cascade version of

been working in the area of neural networks for many years. One of their projects involves the development of a neural network recognizer for handwritten digits. A feed forward layered network with three hidden layers is used. One of the key features in this network that reduces the number of free parameters to enhance the probability of valid generalization by the network. Artificial neural network is also applied for image processing.
6) In Robotics Field: With the help of neural networks and artificial Intelligence.

Intelligent devices, which behave like human, are designed which are helpful to human in performing various tasks. Following are some of the application of Neural Networks in various fields: Business o Marketing o Real Estate Document and Form Processing
28

o Machine Printed Character Recognition o Graphics Recognition o Hand printed Character Recognition o Cursive Handwriting Character Recognition Food Industry o Odor/Aroma Analysis o Product Development o Quality Assurance Financial Industry o Market Trading o Fraud Detection o Credit Rating Energy Industry o Electrical Load Forecasting o Hydroelectric Dam Operation o Oil and Natural Gas Company Manufacturing o Process Control o Quality Control Medical and Health Care o Image Analysis o Drug Development

19. Conclusion

At last it can be said that after 20 or 30 years neural networks will be so developed that it can find the errors of even human beings and will be able to rectify that errors and make human being more intelligent .
29

20.References

Neural Network by: - Simon Haykin <1> Understanding neural networks and fuzzy logic by: - Stamatios V. Kartalopoulos <2>
30

Artificial Neural Network by:- Robert J. Schalkoff <3> Internet: en.wikipedia.org/wiki/Artificial_neural_network <4>

www.learnartificialneuralnetworks.com <5> http://www.ibm.com/developerworks/library/l-neural/ <6>

31