Reviews: Object-Oriented Dynamics Predictor

This paper addresses the problem of action-conditional video prediction via a deep neural network whose architecture specifically aims to represent object positions, relationships, and interactions. The learned models are shown empirically to generalize to novel object configurations and to be robust to minor changes in object appearance. Technical Quality As far as I can tell the paper is technically sound. The experiments are well-designed to support the main claims. I especially appreciated the attempts to study whether the network is truly capturing object-based knowledge as a human might expect (rather than simply being a really fancy pixel - pixel model).