University of Cambridge > > Machine Learning @ CUED > A marginal sampler for σ-Stable Poisson-Kingman mixture models

A marginal sampler for σ-Stable Poisson-Kingman mixture models

Add to your list(s) Download to your calendar using vCal

If you have a question about this talk, please contact Dr Jes Frellsen.

Infinite mixture models reposed on random probability measures like the Dirichlet process allow for flexible modelling of densities and for clustering applications where the number of clusters is not fixed a priori. This is because we can formulate the problem as a hierarchical model where the top level is a (discrete) random probability measure. In recent years, there has been a growing interest in using other random probability measures, beyond the classical Dirichlet process, for extending modelling flexibility . Examples include Pitman-Yor processes, normalized inverse Gaussian processes, and normalized random measures. Our understanding of these models has grown significantly over the last decade: there is an increasing realisation that while these models are nonparametric in nature and allow an arbitrary number of components to be used, they do impose significant prior assumptions regarding the clustering structure.

In this talk I will present a very wide class of random probability measures, called σ-stable Poisson-Kingman processes, and discuss its use for Bayesian nonparametric mixture modelling. This class of processes encompasses most known random probability measures proposed in the literature so far, and we argue that it forms a natural class to study. I will review certain characterisations which lead us to propose a tractable and exact posterior inference algorithm for the whole class.

This is joint work with Yee Whye Teh and Stefano Favaro. The talk is based on our preprint.

This talk is part of the Machine Learning @ CUED series.

Tell a friend about this talk:

This talk is included in these lists:

Note that ex-directory lists are not shown.


© 2006-2024, University of Cambridge. Contact Us | Help and Documentation | Privacy and Publicity