Neural Foundations · 2017
A simple neural network module for relational reasoning
Proposed the Relation Network, a plug-in module that applies a shared learned function to every pair of objects and sums the results, and showed it lifts ordinary networks to superhuman accuracy on the CLEVR visual question answering benchmark.
Editorial record
Plain-language summary
Questions like 'is the ball the same colour as the cube?' require comparing objects, which standard convolutional and recurrent networks do poorly. A Relation Network takes all pairs of object representations, runs each pair through the same small network, and adds the outputs. This explicit all-pairs computation is a simple inductive bias for relational reasoning, and it is a close cousin of the pairwise interactions that self-attention later made general and scalable.
Source record
Provenance
- Record ID
- P-677
- Record created
- 2026-09-26
- Last reviewed
- 2026-09-26
- Record version
- 1
- https://arxiv.org/abs/1706.01427
- arXiv:1706.01427
Citation caveat: Citation metadata is approximate and marked unverified in the source dataset.