Novel methods for learning and inference essential to creating perceiving systems. Photo: Wolfram Scheible.
Our research combines modeling, learning and inference. When we can simply model the world (e.g. its physics) we do. When the complexity gets too great then we need learning. For example, object shapes are not governed by simple physics and to represent the statistical variation of object shapes we need learning. While much of our work in this area has focused on the human body, we see the body as just an interesting example of a deformable and articulated object.
Our goal is to apply our representations and learning methods much more widely. Our work on inference is applied across our work. It is typically grounded in probabilistic inference and exploits graphical models, belief propagation, and stochastic sampling to name just a few methods. In most problems we care about the variables of interest are continuous and high dimensional.