Random pulse artificial neural network architecture.

Abstract

This thesis presents a modular hardware artificial neural network architecture using the random pulse data representation and processing. In random pulse machine, continuous variables can be represented as a probability of a pulse occurrence at a certain sampling time. Random pulse machines deal with analog variables while using digital technology to perform arithmetic and logic operations on binary pulses which are the information carries. Thus, large systems can be built to perform parallel analog computation on large amounts of input data using this technique. This is a good trade-off between the electronic circuit complexity and the computational accuracy and well suited for VLSI neural network. Simulations have been conducted to validate the performance of both 1-bit and 2-bit random pulse neural network. Perception application for linear classification and autoassociative memory for digit recognition are finally presented to illustrate the behavior of the developed artificial neural network.