Hierarchical Relational Inference

Common-sense physical reasoning in the real world requires learning about interactions of objects and their dynamics. The notion an abstract object, however, encompasses a wide variety that differ greatly terms complex behaviors they support. To address this, we propose novel approach to models as hierarchies parts may locally behave separately, but also act more globally single whole. Unlike prior approaches, our method learns unsupervised fashion directly from raw visual images discover objects, parts, relations. It explicitly distinguishes multiple levels abstraction improves over strong baseline at modeling synthetic real-world videos.