Chung, K.T., Lee, C.K.M., Tsang, Y.P., Wu, C.H. and Asadipour, Ali ORCID: https://orcid.org/0000-0002-1449-2195, 2024, Journal Article, Multi-objective evolutionary architectural pruning of deep convolutional neural networks with weights inheritance Information Sciences, 685. pp. 1-18. ISSN 0020-0255
Abstract or Description: | Despite the ongoing success of artificial intelligence applications, the deployment of deep learning models on end devices remains challenging due to the limited onboard computational resources. A way to tackle this challenge is model compression through network pruning, which removes unnecessary parameters to reduce model size without significantly affecting performance. However, existing iterative methods require designated pruning rates and obtain a single pruned model, which lacks the flexibility to adapt to devices with heterogeneous computational capabilities. This paper considers network pruning in Deep Convolutional Neural Networks (DCNNs) and proposes a novel algorithm for structured filter pruning in DCNNs using a multi-objective evolutionary approach with a novel weights inheritance scheme and representation scheme to reduce the time cost of the optimization process. The proposed method provides solutions with multiple levels of tradeoff between performance and efficiency for various hardware specifications on edge devices. Experimental results demonstrate the effectiveness of the proposed framework in optimizing popular DCNN models in terms of model complexity and accuracy. Notably, the framework successfully made significant reductions in floating-point operations ranging from 40% to 90% of VGG-16/19 and ResNet-56/110 with negligible loss in accuracy on the CIFAR-10/100 dataset. |
---|---|
Official URL: | https://www.sciencedirect.com/science/article/pii/... |
Subjects: | Other > Mathematical and Computer Sciences > G400 Computer Science Other > Mathematical and Computer Sciences > G700 Artificial Intelligence |
School or Centre: | Research Centres > Computer Science Research Centre |
Funders: | This work was supported by the Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hong Kong, and the Research and Innovation Office of The Hong Kong Polytechnic University. |
Identification Number or DOI: | https://doi.org/10.1016/j.ins.2024.121265 |
Uncontrolled Keywords: | Network pruning; Multi-objective evolutionary algorithm; Deep convolutional neural networks; Deep learning |
Date Deposited: | 02 Sep 2024 13:27 |
Last Modified: | 02 Sep 2024 13:27 |
URI: | https://researchonline.rca.ac.uk/id/eprint/5945 |
Edit Item (login required) |