Novel `topographic´ nonlinear feature extraction using radial basis functions for concentration coding in the `artificial nose´

A radial basis function technique is discussed which generates a nonlinear feature extraction mapping in which an ordering (possibly subjective) or similarity of the original data is also maintained. The technique is motivated by a real-data example and discussed in the context of an `artificial nose´-a chemical vapour analysis employing broad-band sensor arrays. The particular motivation is that of concentration coding in which the different classes (different levels of concentration) have a natural relative ordering (e.g. linear and uniformly monotonic) which is not reflected in the actual distribution of the data. Comparisons with generalised linear discriminant analysis are made