One of the elusive goals of modern AI research is to enable machines to learn concepts that are transferable from one task or domain to another. For example autonomous vehicles developed in one part of the world should perform adequately in the rest of the world as well. This has generally been very difficult to achieve as the theory regarding this form of learning has had conditions which could not be verified empirically. This project aims to remedy this by finding and testing new sets of conditions which are necessary and sufficient for successful domain adaptation while also being empirically verifiable. The ability to generalise between tasks and domains is often taken for granted in the context of modern AI. Therefore, it is of paramount importance to devise new theory and methodologies that build on it.