Explainable Network Pruning for Model Acceleration Based on Filter Similarity and Importance

Filter-level network pruning has effectively reduced computational cost, as well as energy and memory usage, for parameterized deep networks without damaging performance, particularly in computer vision applications. Most filter-level network pruning algorithms focus on minimizing the impact of pruning on network performance using either importance-based or similarity-based pruning approaches. However, no study has attempted to compare the effectiveness of the two approaches across different network configurations and datasets. To address these issues, this paper compares two explainable network pruning methods based on importance-based and similarity-based approaches to understand their key benefits and limitations. Based on the analysis findings, we propose an innovative hybrid pruning method and demonstrate its effectiveness using several models and datasets. The comparisons with other state-of-the-art filter pruning methods show the superiority of the new hybrid method.