Object-Oriented Learning (OOL): Perception, Representation, and Reasoning
International Conference on Machine Learning (ICML)
Friday July 17, 2020, Virtual Workshop
In recent years, several methods for learning disentangled representations (with respect to the underlying factors of variation) were proposed. One of them, Independently Controllable Factors (ICF), is learning a representation that is independently controllable. While it was successfully applied in stable grid-world environments, it is not able to account for the dynamics of factors of variation or other agents’ influence. We generalize this approach to situations with predictable changes in the environment that do not depend on the agent. In this situation, we want to learn policies that maximally change the dynamics of one component of the representation with minimal change of the dynamics of other components. We apply this approach to improve active, unsupervised representation learning of object-centric representations in environments with dynamics.