Reviews: Runtime Neural Pruning

Update: I updated my score after RNN overhead clarification for more recent networks and their remark on training complexity. The authors propose a deep RL based method to choose a subset of convolutional kernels in runtime leading to faster evaluation speed for CNNs. I really like the idea of combining an RNN and using it to guide network structure. I have some doubts on the overhead of the decision network (see below) would like a comment on that before making my final decision. The experiments focus on VGG16 which is a very heavy network.