Fast, Locally Adaptive Inference for Machine Learning in Graphical Models

Graphical models are a powerful tool in machine learning with successful applications in diverse areas such as medical diagnosis, natural language processing, robotics, speech recognition and analysis of genetic data. Despite this success, modern data sets place new demands on the graphical modelling framework, because the models can be enormous, but exact inference in graphical models is intractable. Despite the extensive literature on approximate inference, there is still a huge gap between the largest data sets that we wish to analyse and the largest graphical models that we can handle.

In order to meet the challenges of these new applications, this project concerns new approximate inference algorithms for the large-scale graphical models that arise in practical applications of machine learning. Very few existing inference algorithms can handle extremely large models with continuous variables, and important classes of inference algorithms, such as Monte Carlo techniques, have not been scaled to such models at all. Computationally efficient inference would significantly expand the range of applications to which the graphical modelling framework can be applied.

The work in this proposal has the potential for substantial economic benefits, because better inference algorithms will enable the graphical modelling framework to be applied with less effort to a broader variety of problems. Graphical modelling provides a powerful, principled reasoning framework that has proven successful for problems in which there is noisy data that only indirectly indicates the true variables of interest. Clearly, this describes a broad variety of applications, both in academia and industry, and for this reason the range of successful applications of graphical modelling has been impressive---including mobile robotics, computer vision, social network analysis, managing online advertisements, and automatically extracting information from product reviews. Despite this success, we believe that graphical modelling has even broader potential for explosive success throughout industry, but this potential cannot be realised without fundamental research into scalable inference algorithms.

The problem is that graphical models are more difficult to apply than they should be. An external user canot really be expected to apply general graphical models without either collaborating with an expert in graphical models or becoming an expert themselves. Although the graphical modelling approach has high practical usefulness, unfortunately the expertise needed to apply it is also high. New, scalable inference inference algorithms have the potential to significant reduce the burden of applying the graphical modelling approach, enabling new industrial applications.