Title:  Score matching for new models with difficult normalising constants

Bio:  Kassel is a Postdoctoral Fellow at the Research School of Finance, Actuarial Studies, and Statistics. His current research focuses on developing methods for compositional data and other multivariate data with known mathematical constraints, with a particular emphasis on Riemannian manifolds. He has authored two R packages available on CRAN, scorematchingad and lacunaritycovariance, which implement some of his research findings.  Kassel's background includes work on bird species occupancy-detection models, spatial random sets, and remote sensing.

Abstract:  Score matching, introduced by Hyvärinen (2005), is an estimation technique that circumvents the challenge of intractable normalizing constants in model densities. Such situations can occur when probability densities are related to potential energies and/or involve constrained spaces. Recent work with my colleagues Andrew Wood and Janice Scealy applied score matching to compositional data generated by microarrays.

Score matching requires second-order derivatives of the density, which can be burdensome to calculate manually, and I'll show how automatic differentiation can significantly expedite the implementation process. Furthermore, we can draw on Windham's (1995) robustification method to enhance efficiency.

This presentation will discuss properties of score matching, application to compositional data, and our advancements in computation and robustness.

(References: Hyvärinen, 2005, Journal of Machine Learning Research; Windham, 1995, J R Stat Soc Series B Stat Methodol)