Kernel norm-similarity based group pruning for compressing CNNs

Deep convolutional neural networks have demonstrated significant performance beyond traditional methods in more and more fields. However, with the increasing network size and the emergence of various complex structures, the increasing demand for computing and storage resources limits their deployment and application on resource-constrained devices. The mainstream structured pruning methods mostly focus on channel or filter pruning, and such coarse-grained pruning also leads to the loss of too many important parameters, which causes unnecessary performance reduction of the compressed network. In this paper, we propose a kernel norm-similarity based group pruning method to prune at a finer granularity while maintaining a regular structure of parameters. Specifically, we first adopt a K-PCA-based K-Means++ algorithm to realize the reasonable grouping of the filters. Afterward, we comprehensively evaluate the importance of the grouped kernels by combining their norm and similarity and then pruning those unimportant ones. Finally, the pruned convolutional layers are reconstructed into a dense group convolution structure. We conduct comprehensive experiments and evaluation of our method on various different models and datasets to prove its effectiveness. The experimental results indicate that our method enables CNNs to maintain better performance while achieving a higher compression ratio, which outperforms most current network pruning methods.