In many datasets, the driving cause for the behaviour we observe is tied to a hidden cause that would be measured using a scale that "wraps on itself" such as angles or a clock/calendar. More precisely, the topology of the latent space can be described as lying over a manifold comprised of combinations of circles, spheres or tori, which requires the tools of circular statistics. My research aims to study how to learn models using these variables in an efficient way through the framework of variational inference. I am also interested in the field of reinforcement learning and applications of machine learning for decision making.