NeurIPS 2020

Self-Supervised Relational Reasoning for Representation Learning

Meta Review

This paper presents a relational reasoning based self-supervised learning framework. All the reviewers like the paper and believe it could be impactful. There are concerns about experiments (not extensive, on small datasets, smaller backbone and only on classification). The authors provided a strong rebuttal addressing some of the concerns on experiments. Some ablation experiment concerns still remain. The AC agrees with all the reviewers that this is indeed a good paper and with more experiments (specifically on ablation and ImageNet) can become a really strong submission. AC committee recommends acceptance.