Warning: From recent investigations, it appears that all the SVM experiments in the above MPhil thesis were conducted with the SVM bias fixed at zero (i.e. if the SVM classifier is a linear discriminant (w,b) in a kernel-induced feature space, where `w' is the weight vector of the linear discriminant and `b' is its bias, then b=0). The thesis gives the impression that the biases for all SVM classifiers were free variables (this was in fact the original intention). However the fact that the biases for these classifiers were not free variables seriously limits the modelling ability of these SVM classifiers. It is important for the reader to take this into account when reading the thesis and making comparisons. Fuller investigations are required to check whether any experiments were performed where nonzero biases were allowed.
Warning: The experiments in the above report CUED/F-INFENG/TR.412 were conducted over a number of months during which the software was gradually made more robust. The experiments were therefore performed with slightly different versions of software. It is important for the reader to take this into account when reading the report and making comparisons.