You are on page 1of 18

INTRODUCTION Neural Networks or NN is the way the brain processes or understands or records and answers if there exists any

kind of query, thought, decision. So we have to understand how the brain processes or transmits the signals. We know the structure of the brain is quite complex and still the scientists or the researchers have not yet been able to understand it. Neural Networks helps us make or realize Artificial Neural Networks (ANN). Neural science is a study in a way itself. Here we will be relating neural networks or neural science to artificial networks or specifically artificial neural networks (ANN ). Now the question arises how does the brain computes or how it understands, responds, thinks. So to understand these questions we first have to know how a neuron works. Neuron is, the smallest individual unit of the nervous system. Our brain contains around 10^10 neurons and each neuron is further connected to 10^4 nerve cells. Here we will first understand how a neuron works, then this understanding is further applied to know our nervous system as a whole. Then, we relate it with the artificial neural networks. These ANN have different architectures, we use these kind of architectures for solving our problems. ANN have a widespread use; in the medical terms, signal processing and at every place where complexity arises. Our Nervous system can process parallel tasks. Neural Networks adopt various learning mechanisms. We learn through experience. So for the network to learn we have supervised learning and unsupervised learning. We the human beings have a very complex structure; whether it be brain, or our body as a whole. Till date researchers, scientists have not been able to find out how our brain computes or solves tasks. This study shows how complex our brain is. Neural Networks is related to Artificial Intelligence (AI), AI as the name suggests Intelligence is being given to a machine or to an artificial being. Information is received as well as computed by the machine or the artificial machine itself. Not only this the output is also given by the so called Artificial Intelligent system. Making a system alone by neural networks is difficult(not impossible) but for some more flexibility and advances we combine it with Fuzzy Systems or Fuzzy Logic's. Now this fuzzy logic or a fuzzy system responds to uncertainty. We can say to handle uncertain complexity we use fuzzy systems. Further we use Neural Networks for Face Recognition. Before opening the door the person has to go through FR test, once the face is recognized automatically the door is being opened. Neural Networks have a widespread use in medical applications. Using Electronics and medical as the base NN is further used to solve problems like blindness. Blindness is something which can be through eye diseases, etc. By using these networks researchers are able to solve problems. Neural Network architectures are classified as Single layer feed forward networks, Multilayer feed forward networks, and recurrent networks. Some well

known NN systems include back propagation network, ADALINE(Adaptive Linear Element), associative memory etc. We will discuss some of these in the next section.

HISTORICAL BACKGROUND

Neural network simulations appear to be a recent development. However, this field was established before the advent of computers, and has survived at least one major setback and several eras. Many important advances have been boosted by the use of inexpensive computer emulations. Following an initial period of enthusiasm, the field survived a period of frustration and disrepute. During this period when funding and professional support was minimal, important advances were made by relatively few researchers. These pioneers were able to develop convincing technology which surpassed the limitations identified by Minsky and Papert. Minsky and Papert, published a book (in 1969) in which they summed up a general feeling of frustration (against neural networks) among researchers, and was thus accepted by most without further analysis. Currently, the neural network field enjoys a resurgence of interest and a corresponding increase in funding. The first artificial neuron was produced in 1943 by the neurophysiologist Warren McCulloch and the logician Walter Pits. But the technology available at that time did not allow them to do too much. Why we use Neural Networks? Neural networks, with their remarkable ability to derive meaning from complicated or imprecise data, can be used to extract patterns and detect trends that are too complex to be noticed by either humans or other computer techniques. A trained neural network can be thought of as an "expert" in the category of information it has been given to analyze. This expert can then be used to provide projections given new situations of interest and answer "what if" questions. Other advantages include: 1. Adaptive learning: An ability to learn how to do tasks based on the data given for training or initial experience. 2. Self-Organization: An ANN can create its own organization or representation of the information it receives during learning time. 3. Real Time Operation: ANN computations may be carried out in parallel, and special hardware devices are being designed and manufactured which take advantage of this capability.

4. Fault Tolerance via Redundant Information Coding: Partial destruction of a network leads to the corresponding degradation of performance. However, some network capabilities may be retained even with major network damage.

HUMAN BRAIN

Fig.1: Shows Structure of Human Brain Human Brain is one of the most complicated things, which, on the whole, has been poorly understood. However the fundamental constituent of the brain that is, neuron, has made it possible. As I told you earlier brain contains about 10^10 neurons and each neuron is interconnected to other 10^4 neurons. A neuron is a small cell that receives electrochemical signals from its various sources and sends electrical impulses to other neurons. An average brain weighs about 1.5Kg and an average neuron about 1.5*10^-9 grams. Despite the different activities of the neurons all of them bear the same characteristics. Fig 1 shows the structure of the brain. The division of brain into four parts namely; Frontal lobe, Parietal lobe, Temporal lobe, occipital lobe. Now we talk about the neuron(smallest individual unit of the nervous system), neuron is composed of a nucleus cell body known as soma.

Fig2: Shows Structure of a Neuron Attached to the soma are long irregularly shaped filaments called dendrites. The dendrites behave as input channels that is, all inputs from other neurons arrive through dendrites. Dendrites look like branches of a tree during winter. Another type of link attached to the soma is the Axon. Unlike the Dendritic links, the axon is electrically active and serves as the output channel. Axons, which mainly appear on output cells are non linear threshold devices which produce voltage pulse called action potential or spike that lasts for about a millisecond. If the cumulative inputs received by the soma raise the internal electric potential, then neuron fires by propagating the action potential down the axon to excite or inhibit other neurons. The axon terminates in a specialized contact called synapse or synaptic junction that connects axon with dendritic links of another neuron. The synaptic junction, which is a very minute gap at the end of dendritic link contains a neuro - transmitter fluid. It is this fluid responsible for accelerating and retarding electric charges to soma. Each dendritic link can have many synapses acting on it thus bringing about massive interconnectivity. In general a single neuron can have many synaptic input and outputs. The size of the synapses are believed to be related to learning. It is the increased neuronal activity which is responsible for learning and memory.

MODEL OF AN ARTIFICIAL NEURON Human brain no doubt is a highly complex structure viewed as a massive interconnected network of simple processing elements called neurons. We saw how a neuron works, but now that we have to make an artificial network comprising of an

artificial neuron. Here we have a model that resembles a biological neuron. This is known as an artificial neuron. This is a basic model that forms artificial neural networks or ANN.

Fig3: Shows simple model of an artificial neuron( or the McCulloch and Pitts model) In Fig 3; x(1),(x2) till x(m) are the m inputs to the artificial neuron. w(k1), w(k2). w(km) are the weight attached to input links. We know that a biological neuron receives all inputs from dendrites, sums them and produces an output if the sum is greater than the threshold value. The input signal is passed on through the synapse which may accelerate or retard an arriving signal. It is this acceleration or retardation that is modeled by the weights. An effective synapse that transmits a stronger signal will have correspondingly larger weight while a weak synapse will have smaller weight. Thus weights are multiplicative factors of inputs to account for the strength for a synapse. Hence output received by the artificial neuron is : X = w(1)I(1) + w(2)I(2) + + w(m)I(m) 1.

To generate the final output the sum is passed on to a non linear filter T called Activation function or transfer function or squash function which releases the output. Y(output) = T(X) X comes from equation 1.

ACTIVATION FUNCTION OR THRESHOLDING FUNCTION There are some type of functions used as threshold function. The common functions used are shown below. 1. Step function The unit step function, u(t), is defined as

That is, u is a function of time t, and u has value zero when time is negative (before we flip the switch); and value one when time is positive (from when we flip the switch).

Fig4: Shows a step function 2. Signum function or Sign function :

Fig5: Shows Signum function

NEURAL NETWORK ARCHITECTURES

An Artificial Neural Network is defined as a data processing system consisting of a large number of simple highly interconnected processing elements (artificial neurons). We have a directed graph or digraph. NN are classified according to their learning mechanisms. The three fundamentally different classes of NN are given as follows. 1. Single layer Feed-forward Network 2. Multi layer Feed-forward Network 3. Recurrent Networks We will see the networks one by one as follows. 1. SINGLE LAYER FEED-FORWARD NETWORK This type of network consists of two layer, namely the input layer and the output layer. The input layer neurons receive the input signals and the output layer neurons send the signals out. The synaptic links carrying the weights connect every input neuron to the output neuron but not vice-versa. This kind of network is feed forward in type. Despite the two layers the network is termed as single layer because the output layer alone performs the computation.

Fig.6 : Shows a single layer Feed-forward Network 2. MULTILAYER FEED-FORWARD NETWORK This network as its name indicates is made up of multiple layers. Thus, architectures of this class besides processing an input and an output layer also have one or more intermediary layers known as hidden layers. The computational units of hidden layers are known as hidden neurons or hidden units. The hidden layer aids in performing useful intermediary computations before directing the input to the output layer. The input layer is connected to the output layer and the weights on these links are referred to as input-hidden layer weights. Again the hidden layer neurons are linked to the output neurons and these are referred to as hidden-output weights.

Fig.7: Shows three layer network, with four input nodes, five hidden nodes and one output node or neuron.

3. RECURRENT NETWORKS These networks differ from feed-forward network architectures in the sense that there is at least one feedback loop. Thus, in these networks, for example, there could exist one layer with feedback connections. There could also be neurons with self feedback links, that is, the output of a neuron is fed back to its input.

Fig.8: Shows Recurrent Networks with a self loop shown by D.

EARLY NEURAL NETWORK ARCHITECTURES Here we will discuss some of the early neural network architectures. These architectures were in between 1958 to 1988. Some of the architectures between this time period are being covered which are mostly studied and used as reference. Here we will discuss about Rosenblatts Perceptron, ADALINE, MADALINE. 1. Rosenblatts Perceptron: The Perceptron is a computational model of retina of the eye and hence, is named perceptron. The network comprises three units, the sensory unit S, Association unit A and Response unit R. The S unit comprising 400 photo detectors receives input images and provides a 0/1 electric signal as output. If the input signals exceed a threshold, then the photo detector outputs 1 else 0.The photo detectors are randomly connected to the Association unit A. The A unit comprises feature demons and predicates. The predicates examine the output of the S unit for specific features of the image. The third unir R comprises pattern recognizers or perceptrons, which receives the result in binary form. The weights of S and A are fixed while that of R is adjustable. The output of the R unit can be such that if the weighted sum of its inputs is less than or equal to zero than the output is zero, otherwise the output is the weighted sum itself. It could also be determined by a step function with binary values 0/1 or bipolar values -1/1. Thus, for a function yielding 0/1 output values is given as: Y= 1 for w(m)x(m)>0 0 otherwise

Here x is the input, w is the weight and y is the output.

Fig.9: Shows the Perceptron model; S is sensory unit, A is Activation unit, R is response unit.

The above figure shows the idea or the concept of this model. In the next figure, it shows a multi layered perceptron model.

Fig.10: Shows a Multilayered Perceptron model. There is one hidden layer, one input and output layer. 2. ADALINE Network The Adaptive Linear Neural Network framed b Bernard Widrow of Stanford university makes the use of supervised learning. Here there is only one output neuron and the output values are bipolar (-1 or +1). However, the inputs x(m) could be binary, bipolar, or real valued. The bias weight is w(0) with an input link of x(0)=+1. If the weighted sum of the inputs is greater than or equal to zero then the output is one, else 0. The supervised learning algorithm adopted by the network is similar to the perceptron learning algorithm. Devised by Widrow-Hoff (1960), the learning algorithm is also known as the Least Mean Square (LMS) or Delta rule.

Fig.11: Shows an ADALINE Network. Here the input weights are adjusted (not the output ones, as in perceptron) The rule of widrow is given as: W(new) = W(old) + Zz(t Y )x(input) Where Zz is the learning co-efficient, t is the target output, Y is computed output and x is input. ADALINE network has had the most successful applications because it is used virtually in all high speed modems and telephone switching systems to cancel the echo in long distance communication circuits. 3. MADALINE Network A MADALINE(Many ADALINE) network is created by combining a number of ADALINES. The use of multiple ADALINES helps counter the problem of nonlinear separability. For example, the MADALINE network with two units exhibits the capability to solve the XOR problem.

Fig.12: Shows MADALINE that is, Many ADALINE

LEARNING METHODS OF NEURAL NETWORKS


As mentioned before there are learning methods for the Neural Networks. We the human beings learn from experiences. Life is all about experiences, and if we are developing an intelligence network comparable to a human network then we need to train it accordingly. There are some methods shown, so as to how to train an artificial neural network. Broadly, learning of Neural Network is categorized as Supervised learning, Unsupervised learning, Reinforced learning(helping a network in its learning process). 1. SUPERVISED LEARNING In this, every input pattern that is used to train the network is associated with an output pattern, which is the target or the desired pattern. A teacher is assumed to be present during the learning process, when a comparison is made between the networks computed output and the correct output, to determine the error. The error can then be used to change network parameters, which finally results in an improved performance. Now this supervised learning is divided as Error correction Gradient descent and Stochastic methods.

Error correction Gradient descent is further divided as Least Mean Square and Back-propagation algorithms. 2. UNSUPERVISED LEARNING In this learning method, the target output is not presented to the network. It is as if there is no teacher to present the desired patterns and hence, the system learns of its own by discovering and adapting to structural features in the input patterns. 3. REINFORCED LEARNING In this method, a teacher though available, does not present the expected answer but only indicates if the computed output is correct or incorrect. The information provided helps the network in its learning process

FUZZY LOGIC AND HYBRIDISATION


Fuzzy set theory proposed in 1965 by Lotfi A. Zadeh is a generalization of classical set theory. Fuzzy logic representations founded on Fuzzy set theory try to capture the way humans represent and reason with real world knowledge in the face of uncertainty. Uncertainty could arise due to generality, vagueness, ambiguity, chance, or incomplete knowledge. Alone Neural Networks can make a system but it would be difficult as well as non-flexible, so neural network combines with Fuzzy systems. Now we have to understand what a fuzzy set is; a fuzzy set can be defined mathematically by assigning to each possible individual in the universe of discourse, a value representing its grade of membership in the fuzzy set. This grade corresponds to the degree to which that individual is similar or compatible with the concept represented by the fuzzy set. In other words, we can say that fuzzy sets support a flexible sense of membership of elements to a set. In classical set theory, elements either belongs to or does not belong to a set and such sets are called crisp sets. But in fuzzy sets many degrees of membership(between 0 and 1 ) are allowed. Operations such as union, intersection, product, equality, difference, etc all are allowed on fuzzy sets. Just crisp set theory has influenced symbolic logic, fuzzy set theory has given rise to fuzzy logic. In crisp sets there is a true logic and a false logic. But here in fuzzy sets multi-valued truth tables such as absolutely true, fairly true, false, absolutely false, partly false, etc are supported. Fuzzy rules have found large advantages in real-world problems. Fuzzy rule also called as IF THEN rule have widespread applications nowa-days.

Now, if we combine Fuzzy Logic and Artificial Neural Networks we come across a new term called hybrid systems. We know alone Fuzzy Logic cannot implement large or complex systems and on the other hand if we see a system only of Neural Network is not flexible enough. Therefore if we combine Fuzzy and Neural systems we get a new system called Neuro-Fuzzy system. Hybridization of technologies can have its pitfalls and therefore needs to be done with care. If a technology can solve a problem then a hybrid technology has a better solution for it or it finds an alternative method to arrive at a solution of this. Though Hybrid systems have tremendous potential to solve problems, an inappropriate use of technology can back fire. For, it is improper to expect that if the individual technologies are good then hybridization of technologies should turn out to be even better. In fact, it is not likely for a hybrid technology to exhibit most of the weaknesses of the participating technologies and less of their strengths. Neural Networks, Fuzzy Logic, Genetic Algorithms are soft computing methods which have been inspired by biological computational processes and natures problem solving strategies. Genetic algorithms are inspired by the process of biological evolution, are adaptive search and optimization algorithms. Each of the technologies mentioned above, in their own right and merit, have provided solutions to a wide range of problems belonging to different domains.

APPLICATIONS OF NEURAL NETWORKS


Given this description of neural networks and how they work, what real world applications are they suited for? Neural networks have broad applicability to real world business problems. In fact, they have already been successfully applied in many industries. Since neural networks are best at identifying patterns or trends in data, they are well suited for prediction or forecasting needs including: industrial process control customer research data validation risk management

But to give you some more specific examples : ANN are also used in the following specific paradigms Recognition of speakers in communications Diagnosis of hepatitis Recovery of telecommunications from faulty software Interpretation of multi-meaning Chinese words

Undersea mine detection Texture analysis Three-dimensional object recognition Hand-written word recognition, and facial recognition Neural Networks in medicines (such as electronic nose, instant physician etc), Business activities etc.

Neural Networks in Control systems: Neural Networks are used for designing control systems. For example, when a truck driver reverses a truck and he has to quite accurate to come to the final point. There is no option of forwarding and then reversing, then again doing this. A network or a neural network is made for steering. Once the network can see the point where it needs to be parked then it is done easily. While doing this sought of a job manually needs much experience. Neural Networks in Medical Applications: As we know that Neural Network has widespread applications in the medical field, for example, The Bionic Eye for permanent blindness, Electronic Noses, Instant physician etc. Some of these are described as follows: Electronic noses

ANNs are used experimentally to implement electronic noses. Electronic noses have several potential applications in telemedicine. Telemedicine is the practice of medicine over long distances via a communication link. The electronic nose would identify odours in the remote surgical environment. These identified odours would then be electronically transmitted to another site where an door generation system would recreate them. Because the sense of smell can be an important sense to the surgeon, telesmell would enhance telepresent surgery. Instant Physician

An application developed in the mid-1980s called the "instant physician" trained an auto-associative memory neural network to store a large number of medical records, each of which includes information on symptoms, diagnosis, and treatment for a particular case. After training, the net can be presented with input consisting of a set of symptoms; it will then find the full stored pattern that represents the "best" diagnosis and treatment.

CONCLUSION
The computing world has a lot to gain from neural networks. Their ability to learn by example makes them very flexible and powerful. Furthermore there is no need to devise an algorithm in order to perform a specific task; i.e. there is no need to understand the internal mechanisms of that task. 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. Neural networks also contribute to other areas of research such as neurology and psychology. They are regularly used to model parts of living organisms and to investigate the 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 integrated with computing, AI, fuzzy logic and related subjects.

REFERENCES
1. Neural networks, Fuzzy Logic, and Genetic Algorithms by S.Rajasekaran and G.A. Vijayalakashmi Pai. 2.

You might also like