Generative vector quantisation

Based on the assumption that a pattern is constructed out of features which are either fully present or absent, we propose a vector quantisation method which constructs patterns using binary combinations of features. For this model there exists an efficient EM-like learning algorithm which learns a set of representative codebook vectors. In terms of a generative model, the collection of allowed binary states 'generates' the set of codebook vectors. Thus, the method provides not only a compact description of the data in terms of clusters, but also an explanation of the individual clusters in terms of common elementary features. Preliminary results on image compression and handwritten digit analysis indicate that our approach is a computationally inexpensive alternative to more complex probabilistic generative graphical models.