David Dunson

Department of Statistical Science

Duke University


Bayesian nonparametric learning from high-dimensional data

In modern applications, it has become routine to collect high-dimensional and highly-structured data, with the number of "subjects", n, often substantially less than the number of variables, p, under study. In such settings, dimensionality reduction is crucial and one common strategy relies on "learning" of a lower dimensional subspace the data are concentrated near. After such learning, the effective number of parameters is hopefully substantially less than the sample size, so that one has a hope of obtaining reliable statistical inferences and predictions. In this talk, I will provide an overview and illustration of recent approaches for nonparametric Bayes learning of lower-dimensional structure in complex massive dimensional data. The common emphasis will be on defining a prior that fully accounts for uncertainty in the lower dimensional structure in a flexible manner, with tensor products providing a useful paradigm. The concepts will be illustrated through application to nonparametric regression in many continuous predictors, analysis of large contingency tables, functional data analysis and shape analysis. Theoretical properties will be mentioned briefly, but the emphasis will be on biomedical applications.


Back to Colloquium Series