The module builds on the core module “Probabilistic Modelling and Reasoning”, and will include selected models like restricted Boltzmann machines for modelling discrete data. It will cover approximate inference techniques for probabilistic graphical models: belief propagation and variational inference (convex relaxations). The second focus of the module is on Bayesian non-parametric models, of which Gaussian processes will be the running example. The module concludes with an overview of other Bayesian non parametric models, notably Dirichlet processes, with common tricks like collapsed Gibbs sampling.
Upon completion of the module the student will be able to:
- formulate and interpret problems fully as probabilistic (graphical) models;
- mathematically derive and practically implement a suite of inference techniques for probabilistic models, from established to state-of-the-art;
- explain the fundamental theory underlying Gaussian processes as a particular tool for tackling regression problems;
- demonstrate practical considerations necessary to implement these models effectively and efficiently.