Recurrent Relational Networks

Authors:
Rasmus Palm Technical University Denmark
Ulrich Paquet DeepMind
Ole Winther Technical University of Denmark

Introduction:

This paper is concerned with learning to solve tasks that require a chain of interde-pendent steps of relational inference, like answering complex questions about therelationships between objects, or solving puzzles where the smaller elements of asolution mutually constrain each other.The authors introduce the recurrent relational net-work, a general purpose module that operates on a graph representation of objects.As a generalization of Santoro et al.

Abstract:

This paper is concerned with learning to solve tasks that require a chain of interde-pendent steps of relational inference, like answering complex questions about therelationships between objects, or solving puzzles where the smaller elements of asolution mutually constrain each other. We introduce the recurrent relational net-work, a general purpose module that operates on a graph representation of objects.As a generalization of Santoro et al. [2017]’s relational network, it can augmentany neural network model with the capacity to do many-step relational reasoning.We achieve state of the art results on the bAbI textual question-answering datasetwith the recurrent relational network, consistently solving 20/20 tasks. As bAbI isnot particularly challenging from a relational reasoning point of view, we introducePretty-CLEVR, a new diagnostic dataset for relational reasoning. In the Pretty-CLEVR set-up, we can vary the question to control for the number of relationalreasoning steps that are required to obtain the answer. Using Pretty-CLEVR, weprobe the limitations of multi-layer perceptrons, relational and recurrent relationalnetworks. Finally, we show how recurrent relational networks can learn to solveSudoku puzzles from supervised training data, a challenging task requiring upwardsof 64 steps of relational reasoning. We achieve state-of-the-art results amongstcomparable methods by solving 96.6% of the hardest Sudoku puzzles.

You may want to know: