Reviews: Recurrent Relational Networks

The authors propose a generalization of the Relational Network (RN) architecture proposed by Santoro 2017. Whereas RN does one step of processing on a fully-connected graph composed of nodes representing objects, the Recurrent Relational Network does multiple time steps of processing, maintaining a hidden state per node and parametrizing messages between nodes at each time step as a MLP. The new architecture is evaluated on several benchmarks. The first is bAbI, which is a popular text-based question-answering dataset consisting of 20 different type of tasks, each of which involves receiving several supporting facts and answering a question related to those facts. The proposed method solves all 20 tasks and seems to display less variability between different training runs in answering these questions compared to other published methods.