(2017. 6) Relational Network
Submitted on 2017. 6
Adam Santoro, David Raposo, David G.T. Barrett, Mateusz Malinowski, Razvan Pascanu, Peter Battaglia and Timothy Lillicrap
Simple Summary
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
The difficulty of reasoning relations
Symbolic approaches - define relations between symbols using logic languages, and reason the relations using deduction and algebra.
Statistical learning - build representations from raw data and generalize across diverse and noisy conditions
Pros:
Learn to Infer relations: All-to-all
RNs are data efficient: Single function g to compute each relation
Operate on a set of objects: Order invariant
Tasks
CLVER : 95.5 % (super-human preformance)
Sort-of-CLEVR : > 94%
bAbi: Succeeded on 18/20 tasks
Dynamic physical systems: 93% conntion, 95% counting
Last updated
