# Efficient CNNs via Passive Filter Pruning

**Authors:** Arshdeep Singh, Mark D. Plumbley  
**arXiv:** [2304.02319](https://arxiv.org/abs/2304.02319)  
**Date:** 2023-04-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.02319v1/

## Key points

- Convolutional neural networks (CNNs) are resource-intensive due to their high computational complexity and memory storage requirements
- Recent research has focused on filter pruning methods to eliminate unnecessary filters from CNNs based on their importance
- Existing filter pruning methods can be categorized as "active" or "passive"
- Active methods use datasets and feature maps to quantify filter importance
- Passive methods compute filter importance using the entry-wise norm of the filters without involving data
- Entry-wise norm methods tend to eliminate smaller norm filters without considering their significance, leading to performance degradation when a high pruning ratio is applied
- The authors propose a passive filter pruning method that prunes filters based on their contribution to producing output by considering the operator norm of the filters
- The proposed method demonstrates better generalization across various CNN architectures compared to entry-wise norm-based pruning methods
- It is at least 4.5 times faster in computing filter importance compared to existing active filter pruning methods while achieving similar performance
- The efficacy of the proposed method is evaluated on audio scene classification and image classification tasks using different CNN architectures such as VGGish, DCASE21_Net, VGG-16, and ResNet-50
- The results show that the proposed method effectively prunes filters while maintaining or even improving performance in these tasks
- The paper presents an efficient passive filter pruning method for CNNs that considers the operator norm of filters to determine their importance
- The method outperforms existing entry-wise norm-based pruning methods and achieves comparable results to active filter pruning methods with significantly reduced computation time.

## Summary

Convolutional neural networks (CNNs) have achieved state-of-the-art performance in various applications, but they are resource-intensive due to their high computational complexity and memory storage requirements. To address this issue, recent research has focused on filter pruning methods that eliminate unnecessary filters from CNNs based on their importance. Existing filter pruning methods can be categorized as either "active" or "passive". Active methods use datasets and feature maps to quantify filter importance, while passive methods compute filter importance using the entry-wise norm of the filters without involving data. However, when a high pruning ratio is applied and a large number of filters need to be pruned, entry-wise norm methods tend to eliminate relatively smaller norm filters without considering their significance in producing node output. This leads to a degradation in performance. To overcome this limitation, the authors propose a passive filter pruning method that prunes filters based on their contribution to producing output by considering the operator norm of the filters. The proposed pruning method demonstrates better generalization across various CNN architectures compared to entry-wise norm-based pruning methods. Additionally, it is at least 4.5 times faster in computing filter importance compared to existing active filter pruning methods while achieving similar performance. The efficacy of the proposed method is evaluated on audio scene classification and image classification tasks using different CNN architectures such as VGGish, DCASE21_Net, VGG-16, and ResNet-50. The results show that the proposed method effectively prunes filters while maintaining or even improving performance in these tasks. In conclusion, this paper presents an efficient passive filter pruning method for CNNs that considers the operator norm of filters to determine their importance. The method outperforms existing entry-wise norm-based pruning methods and achieves comparable results to active filter pruning methods with significantly reduced computation time.