https://hgpu.org/?p=13924
PerforatedCNNs: Acceleration through Elimination of Redundant Convolutions