Use this URL to cite or link to this record in EThOS:
Title: Theoretical issues and practical considerations concerning confidence measures for multi-layer perceptrons
Author: Papadopoulos, Georgios
Awarding Body: University of Edinburgh
Current Institution: University of Edinburgh
Date of Award: 2000
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
The primary aim of this thesis is to study existing CM methods and assess their practicability and performance in harsh real-world environments. The motivation for this work was a real industrial application - the development of a paper curl prediction system. Curl is an important paper quality parameter that can only be measured after production. The available data were sparse and were known to be corrupted by gross errors. Moreover, it was suspected that data noise was not constant over input space. Three approaches were identified as suitable for use in real-world applications: maximum likelihood (ML), the approximate Bayesian approach and the bootstrap technique. These methods were initially compared using a standard CM performance evaluation method, based on estimating the prediction interval coverage probability (PI CP). It was found that the PI CP metric can only gauge CM performance as an average over the input space. However, local CM performance is crucial because a CM must associate low confidence with high data noise/low data density regions and high confidence with low noise/high data density regions. Moreover, evaluating local performance could be used to gauge the input-dependency of the noise in the data. For this reason, a new CM evaluation technique was developed to study local CM performance. The new approach, called classification of local uncertainty estimates (CLUES), was then used for a new comparison study, this time in the light of local performance. Three main conclusions were reached: the noise in the curl data was found to have input-dependent variance, the approximate Bayesian approach outperformed the other two in most cases, and the bootstrap technique was found to be inferior to both ML and Bayesian methods for data sets of input-dependent data noise variance.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available