Recurrent Relational Networks for complex relational reasoning


Nov 03, 2017 (modified: Nov 03, 2017) ICLR 2018 Conference Blind Submission readers: everyone Show Bibtex
  • Abstract: A core component of human intelligence is the ability to reason about objects and their interactions which is something even state-of-the-art deep learning models struggle with. Santoro et al. (2017) introduced the relational network to add such relational reasoning capacity to deep neural networks but the proposed network is severely limited in the complexity of the reasoning it can perform. We introduce the recurrent relational network which can solve tasks requiring an order of magnitude more steps of reasoning. We apply it to solving Sudoku puzzles and achieve state-of-the-art results solving 96.6% of the hardest Sudoku puzzles. For comparison the relational network fails to solve any puzzles. We also apply our model to the BaBi textual QA dataset solving 19/20 tasks which is competitive with state- of-the-art sparse differentiable neural computers. The recurrent relational network is a general purpose module that can be added to any neural network model to add a powerful relational reasoning capacity.
  • TL;DR: We introduce Recurrent Relational Networks, a powerful and general neural network module for relational reasoning, and use it to solve 96.6% of the hardest Sudokus and 19/20 BaBi tasks.
  • Keywords: relational reasoning, graph neural networks