| The last decade has witnessed the revival and a new surge in the field of artificial neural network research. This is a thoroughly interdisciplinary area, covering neurosciences, physics, mathematics, economics, and electronics. Although artificial neural networks have found diverse applications in pattern recognition, signal processing, communications, control systems, optimization, among others, this is still a research field with many open problems in the areas of theory, applications, and implementations. Compared with the development in neural network theories, hardware implementation has lagged behind. In order to take full advantages of neural networks, dedicated hardware implementations are definitely needed. Today, harnessing VLSI technology to produce efficient implementations of neural networks may be the key to the future growth and ultimate success of neural network techniques.; This dissertation deals with the development of neural network models suitable for digital VLSI implementations. Since the state-of-the-art VLSI implementation technologies are basically a digital implementation medium, which offers many advantages over its analog counterpart, artificial neural networks must be adapted to an all-digital model in order to benefit from those advanced technologies. In this dissertation, new models of multilayer feedforward neural networks with single term powers-of-two weights, quantized neurons, and simplified activation functions are proposed to facilitate the hardware implementation in digital approach. Dedicated training algorithms and design procedures for these models are also developed. To demonstrate the feasibility of the presented models, performance analysis and simulation results are provided, and VHDL and FPGA designs are implemented. It has been shown that these proposed models can achieve almost the same performance as the original multilayer feedforward networks while obtaining significant improvement in digital hardware implementation in terms of silicon area and operation speed. By using the models developed in this dissertation, a digital implementation approach of multilayer feedforward neural networks becomes very attractive. |
