Dr Nicola L. C. Talbot CMath MIMA

About

I am an honorary lecturer in the School of Computing Sciences. My main line of work is with my publishing company Dickimaw Books.

Research Activities

Support Vector Machines

I sometimes involved in work on kernel learning methods in collaboration with Dr Gavin Cawley. Kernel learning methods, and in particular the Support Vector Machine, currently represent the state-of-the-art in statistical pattern recognition. Our work currently concentrates on the use of non-standard loss functionals, learning sparse and parsimonious representations and efficient metrics for model selection. See Dr Cawley's webpages for further details.

Top

A Bayesian Approach to Food Safety

Whilst working at the Institute of Food Research (Oct. 1996-Sept. 1999), I worked with Dr Gary Barker and Prof Mike Peck on Risk Assessment for the hazards associated with foodborne botulism and, in particular, hazards associated with the Sous Vide food manufacturing process, using Bayesian belief networks. Projects were funded by the former Ministry of Agriculture Fisheries and Food (now called the Department for Environment Food and Rural Affairs) and the European Union.

Related Links:

Top

Index Assignment for Vector Quantisation

I collaborated (Mar.-Sept. 1996) with Dr Gavin Cawley on research into index assignment for vector quantisation of speech and image signals over noisy image channels where the code book is reordered in order to minimise the effects of errors introduced by noise in the transmission channel.

Top

Comparison of Optimisation and Neural Network based Design Techniques

My Ph.D. research (University of Essex Oct. 1991-Sept. 1995) was concerned with an investigative comparison of neural and conventional optimization methods for use in computer-aided design of VLSI integrated circuits. The areas of placement and filter design were identified as suitable areas of research.

The placement problem was divided into three classes: quadratic assignment, force-directed placement and min-cut methods. For each class a neural and a conventional solution were implemented, and their performance analysed. It was shown that none of the neural methods used, with the possible exception of the Kohonen self-organising feature map, were better than a novel extension to a conventional algorithm developed during my research, where an optimal placement can be found from the solution to a system of linear simultaneous equations.

The investigation into filter design showed that neural networks are not well suited to continuous optimization problems, and a simple triangular interpolation algorithm generated better results.

My Ph.D. was funded by the EPSRC.

Top

School of Computing Sciences. Last Modified: 2012-10-18.