NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:767
Title:CondConv: Conditionally Parameterized Convolutions for Efficient Inference


		
This work proposes a novel convolutional mechanism called CondConv which learns conditionally parameterized convolutions. While CondConv is novel enough, it also has a drawback: the number of parameters is increased significantly compared to ordinary convolutions (as argued by R2). The authors well address the concerns raised by the reviewers. In case the paper is accepted, the authors are encouraged to enhance the paper according to the rebuttal, especially adding discussions about the increased parameters issue.