Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.657487
Title: Implementing radial basis function neural networks in pulsed analogue VLSI
Author: Mayes, David J.
Awarding Body: University of Edinburgh
Current Institution: University of Edinburgh
Date of Award: 1997
Availability of Full Text:
Access through EThOS:
Full text unavailable from EThOS. Please try the link below.
Access through Institution:
Abstract:
The Radial Basis Function (RBF) neural network architecture is a powerful computing paradigm that can solve complex classification, recognition and prediction problems. Although the RBF is similar in structure to the ubiquitous Multilayer Perceptron (MLP) neural architecture, it operates in a different way. This thesis discusses the issues addressed, and the findings from, a project that involved implementing a Radial Basis Function neural network in analogue CMOS VLSI. The developed hardware exploits the pulse width modulation (PWM) neural method, which allows compact, low power hardware to be realised through a combination of analogue and digital VLSI techniques. Novel pulsed circuits were designed and developed, fabricated and tested in pursuit of a fully functioning RBF demonstrator chip. The theory underpinning the designs is discussed and measured hardware results from two test chips are presented along with an assessment of circuit performance. Although the circuits generally functioned as required, discrepancies between the actual and theoretical operation were observed. Thus suggested improvements to the original designs are made and the circuit and system level considerations for the final demonstrator chip are discussed. Measured results are presented from the final demonstrator chip, confirming the correct operation of its constituent circuits, along with results from experiments showing that, when modelled in software, the developed circuitry is capable of performing as well as an identically trained RBF with Gaussian non-linearities. However, further results indicated that the expected network performance would degrade when the neural parameters are quantised. Hardware experiments with the demonstrator chip indicated that it could be used as an RBF classifier, but its performance degraded for more complex problems. A summary of the probable reasons for the performance degradation is provided.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.657487  DOI: Not available
Share: