This paper examines the role of the architectural bias of CNNs, showing that the structural prior of locality and weight-sharing can improve performance of FCNs. The reviewers found the analysis novel and interesting and I agree with their consensus that this paper should be accepted.