Implementation of Variable Step-Size LMS Filter in Neural Network

Abstract

This dissertation is an effort towards the implementation of an integrated ANN trained with backpropagation algorithm that can also function as a variable step size LMS filter. An artificial neural network is an emulation of biological neural system. An artificial neural network is an adaptive system. Learning rule is required to make the neural network adaptive. The implementation of the neural network suffers from various bottlenecks including massive consumption of computational resources and difficulty to determining the parameters of the network and training algorithm. This work discusses the effect of the step size on the training of neural network. A novel variable step size algorithm based on Principal Component Analysis (PCA) that derived from statistical analysis has been proposed. Furthermore, a novel approach is proposed to implement backpropagation algorithm in FPGA for effective resource utilization. Simulation and implementation results confirm the efficacy of the proposed techniques both in terms of generalization performance and hardware resource utilization.