A simple neural network module for relational reasoning

TL;DR

This paper introduces Relation Networks, a simple neural module that significantly improves relational reasoning in neural networks, achieving state-of-the-art results on visual and textual question-answering tasks and demonstrating the importance of explicit relational reasoning.

Contribution

The paper presents Relation Networks as a plug-and-play module that enhances neural networks' ability to perform relational reasoning across various tasks.

Findings

State-of-the-art performance on CLEVR visual question answering dataset

Improved reasoning ability in text-based question answering with bAbI tasks

Convolutional networks gain relational reasoning capacity when augmented with RNs

Abstract

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.