Aktionsplan-Informatik: Development and Assessment of Complex Probabilistic Models in Machine Learning

• Applicant: Dr. Carl Edward Rasmussen
• Subject Area: Theoretical Computer Science
• Term: from 2003 to 2010
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 5401442

Modeling of large complex data sets requires complex models, as are currently being developed in Machine Learning. These models can be put on firm theoretical foundations of statistics and probability theory, eg in a Bayesian setting. The computation required for inference in these models include optimization or marginalisation over all free parameters in order to make predictions and evaluations of the model. Inference in all but the very simplest models is not analytically tractable, so approximate techniques such as variational approximations and Markov Chain Monte Carlo may be needed. Models include probabilistic kernel based models, such as Gaussian Processes and mixture models based on the Dirichlet Process. The project aims at understanding and design of these models, which requires simultaneously realistic assumptions about the data, and tractable (approximate) inference algorithms. Thorough empirical assessment of these models is necessary for their practical use. Careful empirical assessments are in themselves a non-trivial statistical inference problems.

• DFG Programme: Independent Junior Research Groups