Artificial Neural Networks
Neelam Nehra, Pardeep Sangwan, Divya Kumar
Abstract
Artificial neural networks (ANNs) aim at learning discrete-valued, real-valued, and vector-valued representations in the machine learning (ML) field. Algorithms such as backpropagation based on gradient descent method for tuning network parameters in order to best fit a training dataset have already shown remarkable results in the field of ML. The ANN learning is robust to errors and has been effectively applied to different problems in the area of speech recognition, speaker recognition, and other domains of speech processing. The main emphasis of this chapter is on neural network (NN) representations and defining suitable problems for NN learning. The rest of the chapter covers numerous substitute designs for the primitive units making up an ANN such as perceptron units, sigmoid units, and linear units. This chapter also covers the learning algorithms for training single units. Later, the backpropagation algorithm for multilayer perceptron training is described in detail. Also, the general issues such as the representational capabilities of ANNs, overfitting problems, and substitutes to the backpropagation algorithm are also explained. Lastly, the concepts of convergence, local minima, generalization, overfitting, and stopping criterion are covered in detail.