 Today, we are going to discuss on the topic classification algorithms based on neural networks. At the end of the session, the student will be able to demonstrate various classification algorithms based on neural networks and compare between them. An artificial neural network, what we have to assume now, often just called as a neural network is a mathematical model of the computational model based on biological neural networks. In other words, it is an emulation of the biological neural system that is mimicry of how our actual human beings neural system works. Why do we require to use this particular neural networks? They are useful for data mining and decision and support applications. People are good at generalizing from experience. However, computers excel at the following explicit instructions which can be done over and over again in a repetitive fashion. Neural network bridges this gap between the human beings and the computer and it does this with the help of modeling on the computer and the neural network behavior of the human brains is then experienced as a model. Neural networks are useful for pattern recognition or data classification through a learning process. As we all know, learning outputs a change and this change has to be observed in our neural network from the input. Neural networks simulate the biological system. When the learning involves, there is adjustments to the semantic connections between the neurons. When we look at the anatomy of such a sort of artificial neural network, we see that the neural network maps a set of input nodes to a set of output nodes. There is an input 0, 1 up to n and there are outputs 0, 1 up to m. So there are n inputs and m outputs in this particular network and the network itself is composed of an arbitrary number of nodes with an arbitrary topology. A neuron has many inputs but one output unit. The output can be excited or non-excited. The incoming signals from the neurons determine if a neuron is excited at a particular time. It is also called as firing of the particular neuron. The output is subject to attenuation in the synapse which are junction parts of a neuron. Now there are many topologies in which these neural networks are organized. There are two types basically. One is a feed-forward neural network. This is a unidirectional neural network. It has no feedback. It has no cycles. The error is detected right at the end and they are difficult to analyze such particular networks. Recurrent networks are biorectional and they have a feedback at every unit in your particular network. Now there is a concept which is called as training which has to be provided to your neural network. A neural network has to be configured in such that the application of a set of inputs produced a desired set of outputs. And this training of the neural network is done by feeding it teaching patterns and letting it change its weights according to some learning rule automatically. Now let us pause and think for a while on how neural networks are going to help to do a classification. Neural networks in data mining as I told you the first form is a feed-forward neural network. The simplified process for training a feed-forward neural network is as follows. Input data is present to the network and propagated through the network until it reaches an output layer. So the forward process produces a predicted output between the input and output layers there might be many layers which are called as hidden layers and the predicted output is subtracted from the actual output and an error value for the network is calculated. The neural network then uses a supervised learning which in most cases is back propagation to train the network. This back propagation is a basically a learning algorithm for adjusting the weights in the particular network. It starts with the weights between the output layer and the last hidden layer and works backward throughout the network till the input layer. Once the back propagation has finished the forward process continues. Back propagation is a common method of teaching the artificial neural network how to perform a given task. The basic propagation algorithm is used and it is a layered feed-forward artificial neural network where the output of one layer is given as the input of the other layer and so on we go on till the last layer is encountered. The basic propagation algorithm uses supervised learning which means that we provide the algorithm that it has to use. The techniques used are present training sample to the neural network compare the neural network's output to the desired output from a sample calculate the error in each output neuron and for each neuron calculate the output which has to be then scaled with a scaling factor and we have to go for finding out the local error. We adjust the weight at each neuron to lower the local error and we assign a blame to the local error to the neurons and giving this responsibility to the neurons to connect in stronger weights and we repeat this procedure on the neurons. When we look at the basis of a particular node which is used in our neural network we see many inputs weights are added to these inputs and then we have something called as bias node which adds the bias into our particular system. So every node is an element which performs a function which is the summated weight of the product of the weights assigned and the value of a particular input that is the weighted sum plus the biased weight and this function will be given as output from the particular node which will decide whether that neuron has to be fired or not. A simple perceptron is a classifier because it gives an output of one or zero. It has a binary logical application. It has a threshold linear value. It gives weights which may be minus one or one at random and it gives an output which is based on a weighted sum and the bias value and the output is generated which is in the binary form. This may be used to define a class if the output is zero as one class and the output is one is one class therefore a neural network can be used as a classifier. So initialize the weights in the network repeat the procedure for each of the particular node in the training set and find the output from our particular system using this algorithm which will be used to generate a zero or one finally as a classified output which determines a class. Advantages of neural network is high accuracy. They are able to approximate complex nonlinear mappings, noise tolerance, independence of prior assumptions about the distribution of data or the interaction between the factors, the ease of maintenance because neural networks can be updated by using fresh data making them useful for dynamic environments of usage. So we can apply such networks in finance, marketing, human resource and accounting. The design problems however are some sort of bottlenecks. There are no general methods to determine the optimal number of neurons necessary to solve a problem. It is difficult to select also the training data set which fully describes the problem to be solved. How do we improve this ANN? It is by using them with optimized algorithms like the genetic algorithms and neuro fuzzy systems. For our references we have used Dunham and Campbell. Thank you.