Probabilistic Modeling and Inference
A central challenge in machine learning is to quantify uncertainty via probabilistic models that capture statistical dependencies between many uncertainty quantities. We have recently initiated the study of learning and inference in Probabilistic Submodular Models
, a rich class of probabilistic models defined through submodular functions. This class contains and generalizes a number of extensively studied subclasses, such as determinantal point processes, and Ising models. A major benefit of such models is that they allow to capture complex, long-range interactions between many variables, which is useful, for example in computer vision and information retrieval. We develop novel algorithms for efficient approximate inference, using, e.g., variational
or sampling techniques, as well as learning such models from data.