Learning structured representations of data

Abstract

Bayesian networks have shown themselves to be useful tools for the analysis and modelling of large data sets. However, their complete generality leads to computational and modelling complexities that have limited their applicability. We propose an approach to simplify and constrain Bayesian networks that strikes a more useful compromise between generality and tractability. These constrained graphical will allow us to build computationally tractable models for large high-dimensional data sets. We also describe examples of data sets drawn from image and speech processing on which can (1) further explore this constrained set of graphical models, and (2) analyse their performance as a general-purpose statistical data analysis tool.