A simple neural network module for relational reasoning
This addresses the challenge of enabling neural networks to perform relational reasoning, which is crucial for general intelligence, by providing a simple module that enhances existing architectures.
The paper tackled the problem of relational reasoning in neural networks by introducing Relation Networks (RNs) as a plug-and-play module, achieving state-of-the-art, super-human performance on the CLEVR visual question answering dataset and improving performance on text-based and physical reasoning tasks.
Relational reasoning is a central component of generally intelligent behavior, but has proven difficult for neural networks to learn. In this paper we describe how to use Relation Networks (RNs) as a simple plug-and-play module to solve problems that fundamentally hinge on relational reasoning. We tested RN-augmented networks on three tasks: visual question answering using a challenging dataset called CLEVR, on which we achieve state-of-the-art, super-human performance; text-based question answering using the bAbI suite of tasks; and complex reasoning about dynamic physical systems. Then, using a curated dataset called Sort-of-CLEVR we show that powerful convolutional networks do not have a general capacity to solve relational questions, but can gain this capacity when augmented with RNs. Our work shows how a deep learning architecture equipped with an RN module can implicitly discover and learn to reason about entities and their relations.