* Over the few years I have been working (with Carl Rasmussen, David Barber and Francesco Vivarelli) on using Gaussian processes (GPs) for regression and classification problems. In a sense Gaussian processes are an alternative to neural networks for prediction, and a key question is how their performance compares with neural networks. Our results indicate a similar level of performance with networks when tested on a suite of real-world problems, and with more straightforward computational methods; GPs are particularly suitable when only a few hundred training examples are available. GPs have some properties which make them easier to analyze then neural nets, and there are also interesting links between networks and GPs. More information can be obtained on GPs from Carl Rasmussen's GP homepage.
* Image interpretation. Object recognition can be cast in a statistical framework (e.g. in work by Grenander). This approach argues for image understanding using generative models, i.e. explaining an image by instantiated objects. In previous work I used this framework to study the performance of deformable models for a character recognition task. Currently I am looking at identifying regions in outdoor scenes using this approach, and there is much work to be done in extending/applying the method.
* I am involved (along with Francesco Vivarelli) in the Validation and Verification of Neural Network Systems project supported by British Aerospace and EPSRC.
Other areas of interest include (i) applications of inference with Gaussian and non-Gaussian random fields (e.g. modelling wind fields) (ii) data visualization (iii) the role of machine learning in computer vision and (iv) trying to understand the processing and representations used in animal visual systems.