Reviews: Relational recurrent neural networks

This paper proposes relational recurrent neural networks to incorporate relational reasoning into memory-based neural networks. More specifically, the authors design a new memory module, called relational memory core (RMC), to make memories to interact. Multi-head dot product attention is used for this purpose. Experiments are constructed on a few supervised learning and reinforcement learning tasks, which may profit from more capable relational reasoning across sequential information, to show the effectiveness of the proposed method. The main strengths of this paper are as follows: (1) It is a novel and well motivated idea to design a new deep learning architecture which can perform relational reasoning.