Accelerated Learning of Probabilistic Data Models for Large-Scale Applications

I will first give an introduction to the task of data clustering, and will introduce two well-known algorithms: k-means and Expectation Maximization (EM) for Gaussian Mixture Models (GMMs). The properties, relations and computational complexities of the two algorithms will be explained and discussed. By using the example of a GMM as probabilistic model, I will then proceed by introducing truncated variational approximations for accelerated learning. Following their introduction, I will discuss the theoretical foundations of truncated approximations and their generalization to graphical models with discrete latents.

My presentation will close with applications of variational learning to standard and advanced data models. Highlights will include sublinear clustering of large datasets, `black-box' learning, large-scale deep models for semi-supervised learning, and unsupervised learning algorithms for visual, acoustic and medical data.