Rocha, Paulo Valverde de Lacerda Paraiso; (1992) A fully integrated neural computing system. Doctoral thesis (Ph.D), UCL (University College London).

Text A_fully_integrated_neural_comp.pdf Download (7MB)

Abstract

This thesis investigates the Comprehensive Neural Computing (CNC) System - an integrated programming and execution environment for neural network models and their applications. At the core of the CNC is a neural computing architecture for integrating a range of hardware: sequential workstations, parallel computers, and specialised neural computers. This thesis comprises 4 major parts: the design of the CNC Architecture; the design of the CNC Neurocomputer, a fine-grain parallel machine; the CNC System simulator, implemented on a network of workstations; and the architecture assessment studies. The CNC Architecture was chosen to support the programming environment of the ESPRIT II GALATEA Project. GALATEA is the principal European neural computing project, which has CSF-Thomson (France), Philips/LEP (France), Siemens (Germany) and University College London as major partners. Its programming environment is a set of tools for programming and simulating neural networks that builds upon PYGMALION, the programming environment chosen for the CNC System. It comprises a graphic monitor, an algorithm library of common neural network models, a high-level neural network programming language based on C++, and an intermediate-level virtual machine language optimised for the mathematical operations involved in neural computations. The CNC centres on a communications architecture which supports the development of multi-neural network applications. It allows neural networks, or parts of networks, to be mapped and run on the most appropriate machine (e.g. workstations, Transputer-based parallel machines, and the CNC Neurocomputer). The communications architecture comprises a backplane bus specification (the IEEE Futurebus+), a high-level protocol (based on a message passing scheme), and a communication unit, which implements Futurebus+ and executes part of the high-level protocol. The CNC General-purpose Neurocomputer comprises an executive unit and an array of neural microprocessors. The executive unit performs the high-level message passing communication protocol and supervises the microprocessor array operation. The microprocessors are either general-purpose programmable RISC processors, or ASICs generated by a silicon compiler from a definition in PYGMALION'S neural network specification language. The CNC Architecture simulator was developed as the underlying tool for the investigation of applications using the CNC System. It uses workstations to simulate the CNC Neurocomputers, and a local area network for the emulation of the communications architecture hardware. A backplane bus simulator, also running on a workstation, is incorporated to allow monitoring of the system operation, and to investigate the suitability of the common bus approach to neural network applications. The assessment studies analyse the performance of the CNC Communications Architecture and CNC Neurocomputer. In particular, they present results of the simulation of test applications, and discuss the ideal system operating conditions and the suitability of the system for different neural applications. The studies demonstrate the versatility of the CNC Architecture and suggest that the CNC Architecture is suitable for integrating neural networks and other computing paradigms.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item