System level solutions for analog array processors /

Neural network paradigms and image processing algorithms

Bibliographic Details
Main Author: Han, Gunhee, 1965-
Format: Thesis Book
Language:English
Published: [Place of publication not identified] : [publisher not identified] ; 1997.
Subjects:
Online Access:http://proxy.library.tamu.edu/login?url=http://proquest.umi.com/pqdweb?did=736580081&sid=1&Fmt=2&clientId=2945&RQT=309&VName=PQD
Description
Summary:Neural network paradigms and image processing algorithms
require intensive parallel computation. Due to its massive
parallelism, analog implementation is an attractive approach.
This is due, in part, to its low power consumption and small
silicon area. However, the analog system is subject to
process variations, especially when an array of processing
elements is large. More over, the analog system can not
physically implement a practical large array. The proposed
research is focused to find a system level solution
methodology to overcome the practical limits of conventional
analog parallel implementations. Since the major building
blocks of a neuro-image processor are the multiplier and the
integrator, a detailed study on multiplier is performed, and
a switched capacitor weighted integrator with offset
cancellation is developed. Two types of pseudo-parallel
mixed-signal neuro-image processors are developed. One is a
"General Purpose Neuro-Image Processor(GPNIP)" which can
process a host of neuro-image processing paradigms. This
processor adopts a Single Instruction Multiple Data (SIMD)
architecture, and multiplexing and pipelining schemes often
used in digital systems. Nevertheless, the computation is
performed in the discrete-time domain(sampled-data analog
signal). The other image processor is a "Pipelined Cellular
Neural Network(CNN) Processor". The computation of this
processor is performed in the continuous-time domain, i.e.,
as a continuous analog signal. The analog processors are
fabricated using conventional CMOS technology and
experimental results verifying theoretical results are
presented. Since the proposed processors are analog array
processors, a yield estimation methodology for analog array
processor is developed, and then applied to estimate the
yield of large CNN processors. These two processor are unique
by adopting system level concepts used in digital computation
and incorporating them into a mixed-signal pseudo-parallel
computation system. The GPNEP has been implemented to be
independent of process variation due to a built-in offset
cancellation scheme. The architecture allows the gain
correction as well. It has a tremendous architectural
flexibility, that allows a system designer to implement any
architecture based on this processor. Improved computing
efficiency and feasibility of large scale mixed-signal neuro
image processors are achieved in this research.
Item Description:Vita.
"Major Subject: Electrical Engineering".
Physical Description:xiii, 141 leaves : illustrations ; 28 cm.
Issued also on microfiche from University Microfilms Inc.
Bibliography:Includes bibliographical references: pages 123-140.