Dr. Rafal Kustra
University of Toronto

mR3: Flexible, computationally efficient model for very high-dimensional data         

In this talk I will introduce mR3: Multivariate, Reduced-Rank and Regularized model for exploratory analysis of high-dimensional data. The model was motivated by Neuroimaging data, where one routinely analyzes brain scans,with hundreds of thousands of elements each, with a goal of finding spatial associations with experimental covariates of interest. We have also applied this model to a time-course microarray data and I will mention other potential applications as well.

In the first part of this talk I will describe the Neuroimaging experiment that motivated it, then outline the model and underline its similarities with previous approaches, such as Penalized Discriminant of Hastie et al, and Curds and Whey procedure of Breiman and Friedman. I will also describe a computationally-efficient algorithm
for mR3 that enables us to do non-parametric inference despite huge data sizes. I will summarize the results of our simulation study and highlight some results from the analysis of neuroimaging data. I will conclude with some extensions we are working on currently, including a microarray data application.