In recent years, deep learning models have demonstrated excellent performance in complex tasks, but their large number of parameters and high computational costs have limited their application in resource-constrained scenarios. This paper proposes a structured pruning method based on reinforcement learning (TD3 algorithm), which performs structured pruning on a group-by-group basis to balance model compression efficiency and performance retention. The TD3 agent takes the parameter states of each group as observation inputs, dynamically adjusts the pruning rate of each group as actions, and designs a multi-objective reward function based on model accuracy,FLOPs, and the number of parameters to achieve autonomous optimization of pruning strategies. Experiments on ResNet56 and VGG19 with the CIFAR-100 dataset show that this method maintains high classification accuracy while significantly reducing parameters and computational complexity. Compared with traditional pruning methods, it is more adaptive and provides an effective solution for model deployment in resource-constrained environments.

Qianxi Li, Wenhui Zhang. Structured Pruning Based on Reinforcement Learning for CNN. Academic Journal of Computing & Information Science (2025), Vol. 8, Issue 7: 79-86. https://doi.org/10.25236/AJCIS.2025.080710.

[1] Simonyan K ,Zisserman A .Very Deep Convolutional Networks for Large-Scale Image Recognition. [J].CoRR,2014,abs/1409.1556

[2] Chowdhery A, Narang S, Devlin J, et al. Palm: Scaling language modeling with pathways[J]. Journal of Machine Learning Research, 2023, 24(240): 1-113.

[3] Hongbing Z .Application Research of Speech Signal Processing Technology Based on Cloud Computing Platform[J].International Journal of Information Technologies and Systems Approach (IJITSA),2021,14(2):20-37.

[4] You H, Li C, Xu P, et al. Drawing early-bird tickets: Towards more efficient training of deep networks.[J].CoRR,2019,abs/1909.11957

[5] Han S, Mao H ,Dally J W .Deep Compression: Compressing Deep Neural Network with Pruning, Trained Quantization and Huffman Coding.[J].CoRR,2015,abs/1510.00149

[6] Gao S, Zhang Y, Huang F, et al. BilevelPruning: unified dynamic and static channel pruning for convolutional neural networks[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2024: 16090-16100.

[7] Lin H, Xu H, Wu Y, et al. Duquant: Distributing outliers via dual transformation makes stronger quantized llms[J]. Advances in Neural Information Processing Systems, 2024, 37: 87766-87800.

[8] Lv J, Yang H, Li P. Wasserstein distance rivals kullback-leibler divergence for knowledge distillation[J]. Advances in Neural Information Processing Systems, 2024, 37: 65445-65475.

[9] Guo Z, Cheng X, Wu Y, et al. A wander through the multimodal landscape: Efficient transfer learning via low-rank sequence multimodal adapter[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2025, 39(16): 16996-17004.

[10] Fang G, Ma X, Mi M B, et al. Isomorphic pruning for vision models[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 232-250.

[11] Liu J, Tang D, Huang Y, et al. Updp: A unified progressive depth pruner for cnn and vision transformer[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2024, 38(12): 13891-13899.

[12] Wang B, Kindratenko V. Rl-pruner: Structured pruning using reinforcement learning for cnn compression and acceleration[J]. arXiv preprint arXiv:2411.06463, 2024.

[13] Wang Z, Li C, Wang X. Convolutional neural network pruning with structural redundancy reduction[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021: 14913-14922.

[14] Benbaki R, Chen W, Meng X, et al. Fast as chita: Neural network pruning with combinatorial optimization[C]//International Conference on Machine Learning. PMLR, 2023: 2031-2049.

[15] Diao E , Wang G , Zhan J ,et al.Pruning Deep Neural Networks from a Sparsity Perspective[J]. ArXiv, 2023, abs/2302.05601.DOI:10.48550/arXiv.2302.05601.

[16] Mnih V, Kavukcuoglu K, Silver D, et al. Human-level control through deep reinforcement learning [J]. nature, 2015, 518(7540): 529-533. 

[17] Mnih V, Badia A P, Mirza M, et al. Asynchronous methods for deep reinforcement learning[C]//International conference on machine learning. PmLR, 2016: 1928-1937.

[18] Lillicrap P T ,Hunt J J ,Pritzel A , et al.Continuous control with deep reinforcement learning. [J]. CoRR,2015,abs/1509.02971

[19] Fujimoto S, Hoof H, Meger D. Addressing function approximation error in actor-critic methods[C]//International conference on machine learning. PMLR, 2018: 1587-1596.

[20] Wang H , Qin C , Zhang Y ,et al.Neural Pruning via Growing Regularization[J]. 2020. DOI:10.48550/arXiv.2012.09243.