Pattern Classification using Artificial Neural Networks

Abstract

Classification is a data mining (machine learning) technique used to predict group membership for data instances. Pattern Classification involves building a function that maps the input feature space to an output space of two or more than two classes.Neural Networks (NN) are an effective tool in the field of pattern classification, using training and testing data to build a model. However, the success of the networks is highly dependent on the performance of the training process and hence the training algorithm. Many training algorithms have been proposed so far to improve the performance of neural networks. In this project, we shall make a comparative study of training feedforward neural network using the three algorithms - Backpropagation Algorithm, Modified Backpropagation Algorithm and Optical Backpropagation Algorithm. These algorithms differ only on the basis of their error functions.We shall train the neural networks using these algorithms and taking 75 instances from the iris dataset (taken from the UCI repository and then normalised) ; 25 from each class. The total number of epochs required to reach the degree of accuracy is referred
to as the convergence rate. The basic criteria of comparison process are the convergence rate and the classification accuracy. To check the efficiency of the three training
algorithms, graphs are plotted between No. of Epochs vs. Mean Square Error(MSE). The training process continues till M.S.E falls to a value 0.01. The effect of using the momentum and learning rate on the performance of algorithm are also observed. The comparison is then extended to compare the performance of multilayer feedforward network with Probabilistic network.