Logo image
CNN Model Optimization Using a Hybrid Approach of Genetic Algorithm-based Pruning and Retraining with Knowledge Distillation
會議論文集

CNN Model Optimization Using a Hybrid Approach of Genetic Algorithm-based Pruning and Retraining with Knowledge Distillation

Kuan-Ling Chou, Cheng Lung Wang, Yung-Chih Chen, Wuqian Tang, Yi-Ting Li, Shih-Chieh Chang 和 Chun-Yao Wang
GLSVLSI 2025 - Proceedings of the Great Lakes Symposium on VLSI 2025, 頁碼.667-672
29/06/2025
Web of Science ID: WOS:001596652300103

摘要

Convolutional neural networks (CNN) have achieved significant success in various domains, e.g., image recognition, however, deploying CNN on resource-constrained devices remains challenging due to their high computation demands and large memory requirements. In this paper, we propose an approach for structured pruning of CNN using genetic algorithms, specifically focusing on channel pruning. The pruning process often results in a loss of accuracy, requiring retraining techniques to recover the network's accuracy to its original level. We integrate knowledge distillation into the retraining process to recover the model to a similar accuracy within fewer epochs. The experimental results show that our approach successfully reduces the amounts of parameters by 91.02% and computation demands by 90.96%, with a 1.82% decrease in accuracy only for a gesture recognition model with over 30 million parameters. The results demonstrate that our approach significantly reduces the computation of the model without dramatically decreasing its accuracy.

相關連結

指標

1 檢視次數

詳細資料

Logo image