Elastic Filter Prune in Deep Neural Networks Using Modified Weighted Hybrid Criterion

The deployment of Convolutional Neural Networks (CNNs) on edge devices has gradually become a hot topic in research and application. However, simply pursuing high-performance networks is no longer suitable for scenarios that require comprehensive optimization of limited resources, low power consumption, and real-time performance, due to the huge storage and computational overhead required by deep convolutional networks. Based on the Weighted Hybrid Criterion (WHC) method and inspired by Elastic Net, we propose an improved filter pruning method called Elastic Weighted Hybrid Criterion (E-WHC). This approach introduces a regularization parameter to balance the sparsity and importance of filters. Experimental results show that on the CIFAR-10 dataset, E-WHC on the ResNet-32 network achieves a 0.25% accuracy improvement while reducing more than 41.5% of floating-point operations compared to WHC. Similarly, on the ResNet-56 network, E-WHC achieves a 0.1% accuracy improvement while reducing over 28.4% of floating-point operations. We further analyze the reasons for this improvement and discuss the role of the regularization parameter in filter selection. The proposed research provides new insights for the improvement of filter selection methods in convolutional neural networks.