A Mixed-Signal Feed-Forward Neural Network Architecture Using A High-Resolution Multiplying D/A Conversion Method

Artificial Neural Networks (ANNs) are parallel processors capable of learning from a set of sample data using a specific learning rule. Such systems are commonly used in applications where human brain may surpass conventional computers such as image processing, speech/character recognition, intelligent control and robotics to name a few.;In this thesis, a mixed-signal neural network architecture is proposed employs a high resolution Multiplying Digital to Analog Converter (MDAC) designed using Delta Sigma Modulation (DSM). To reduce chip are, multiplexing is used in addition to analog implementation of arithmetic operations.;This work employs a new method for filtering the high bit-rate signals using neurons nonlinear transfer function already existing in the network. Therefore, a configuration of a few MOS transistors are replacing the large resistors required to implement the low-pass filter in the network. This configuration noticeably decreases the chip area and also makes multiplexing feasible for hardware implementation.