Nonparametric bayesian models for structured objects
Zoubin Ghahramani, Professor, University of Cambridge, Cambridge, U.K., Adjunct Faculty, University College London, London, U.K.
Bayesian nonparametrics provides an elegant framework for developing flexible models for machine learning and statistics. Much work has been done on nonparametric models of distributions (e.g. Dirichlet processes) and functions (e.g. Gaussian processes). However, many modelling tasks require latent variables with richer structures. I will describe our recent work on sparse matrices and graph structures (via the Indian Buffet Process), on hierarchies (via the Pitman-Yor Diffusion Tree), on models of covariance matrices (via the Wishart process), and on network structured regression.
Joint work with Ryan P. Adams, Tom Griffiths, David K. Knowles, Hanna M. Wallach, and Andrew G. Wilson.