The Relationship between Classifier Factorisation and Performance in Stochastic Vector Quantisation

Abstract

We seek to address the issue of multiple classifier formation within Luttrell’s stochastic vector quantisation (SVQ) methodology. In particular, since (single layer) SVQs minimise a Euclidean distance cost function they tend to act as very faithful encoders of the input: however, for sparse data, or data with a large noise component, merely faithful encoding can give rise to a classifier with poor generalising abilities. We therefore seek to asses how the SVQs’ ability to spontaneously factorise into independent classifiers relates to overall classification performance. In doing so, we shall propose a statistic to directly measure the aggregate ’factoriality’ of code vector posterior probability distributions, which, we anticipate, will form the basis of a robust strategy for determining the capabilities of stochastic vector quantisers to act as unified classification/classifier-combination schemes.