Neural networks applied to signal processing

Abstract

The relationship between the structure of a neural network and its ability to perform nonlinear mapping is analyzed. A new algorithm, called the conjugate gradient optimization method, for calculating the weights and thresholds of a neural network is presented. The performance of the conjugate gradient algorithm is then compared to the well known backpropagation method and shown to be more computationally efficient. A neural network using the conjugate gradient algorithm is then applied to three simple examples to demonstrate its signal processing capabilities. The first example illustrates the ability of the neural network to perform classification. The second compares the performance of a one-step linear predictor to a neural network for a nonlinear chaotic time series. The neural network predictor is shown to provide much greater accuracy than its linear counterpart. The final application presented demonstrates the ability of a neural network to perform channel equalization for a nonmininmum phase channel. Its performance is then compared to its linear equivalent.