Pruning Convolutional Neural Network with Distinctiveness Approach

Abstract

Convolutional Neural Network (CNN) is one of the widely used deep learning frameworks in image classification, target detection and object recognition domains. Because of the complexity of the network structures, CNNs usually contains a large number of layers and channels, which makes algorithm time-consuming. Using existing CNN architectures to solve specific problems usually leads to many redundant parameters. This paper introduces an approach based on the distinctiveness rules to prune both fully connected layers and filters of the baseline network. We experiment with different pruning means to prune multiple layer neurons and filters by the distinctiveness rules and evaluate the performance of the approach. We also discuss the influence of threshold on the selection of redundant neurons in pattern space and trade-off network scale and accuracy. The result shows using the approach will not change the structure of the baseline network and 32 filters and 161 neurons are removed. The accuracy of the pruned network reaches 80.5% compared with 87% of the original model.