Last updated: 19 December 2010

Applied Bayesian Inference

Natural Language Modelling and Visual Feature Tracking


The purpose of this dissertation is to show that the Bayesian approach to probabilistic modelling not only provides the correct framework for doing rational inference in principle, but that Bayesian inference can be used to construct accurate and computationally efficient solutions to complex problems in practice. The following two themes are most prominent in this work. Firstly, appropriate approximation is the key to developing mathematical models that retain sufficient accuracy about the parameters of interest in the model, while allowing the computational efficiency required of a real-time application. Secondly, any model is almost invariably used as part of a larger system, and the language of probabilities provides the correct framework within which to integrate simple models into more complex ones.

To support these themes, Bayesian solutions are presented to the following problems.

The main contributions of each of these solutions are in constructing a new Bayesian model, and in showing how approximate inference can be used to build computationally efficient solutions with good accuracy.

N-gram Language Modelling

Source code

Language-Aware Modelling

Source code

Visual Feature Tracking

Source code