NEURAL NETWORKS BASED NOISE CANCELLER

Abstract

This paper deals with the implementation of noise canceller using neural networks. Two types of noise canceller has been software implemented. Then the performance of them has been checked for different training algorithms.
At first, batch gradient descent algorithm is used with different noise levels. The effect of increasing noise levels is to degrade the performance (increasing the mean square error) for the two systems under consideration. Then for the same training algorithm, the effect of hidden layer neurons is investigated. It is clear that increasing hidden layer neurons degrades the performance of the two above systems.
The second training algorithm which is used is gradient descent with momentum. The effect of increasing momentum is to improve the performance of the two noise canceller systems.
The third training algorithm which is used is gradient descent with variable learning rate. As in the case of momentum, increasing learning rate improves the performance of the two above systems.
Then the effect of momentum with learning rate increase is investigated. It is clear that the effect of momentum with learning rate increase will improve the performance.
The last training algorithm which is used is resilient backpropagation and it gives the best performance at all.