A Parasitic Mechanism-Based Filter Pruning Method for Deep Convolutional Neural Networks

The convolutional neural network (CNN) shows excellent performances in various applications at the cost of huge storage and calculation overheads. To compress and accelerate the CNN, conventional methods usually prune the unimportant filters under the pre-selected standards, which should be artificially and carefully designed according to the specific application scenarios. In this paper, we propose a general CNN accelerating and compressing method, i.e., the parasitic mechanism (PAM)-based filter pruning algorithm. First, the parasitic layer is constructed on the basis of the convolutional structure and the gate module. Then, the parasitic layers are attached to the original CNN, which carry out the pruning process for the filters in each layer. The above attaching and pruning process is named as the parasitic mechanism. Compared with the traditional filter pruning methods which pre-select the pruning standards artificially, the PAM method intelligently learns the pruning standard in the training process, and prunes the unimportant filters under the constraint of maintaining the performance of the original CNN. As verified by simulations, the proposed PAM algorithm has a better pruning performance than the conventional filter pruning methods.