Reliable Identification of Redundant Kernels for Convolutional Neural Network Compression

Abstract To compress deep convolutional neural networks (CNNs) with large memory footprint and long inference time, this paper proposes a novel pruning criterion based on layer-wise L n -norms of feature maps to identify unimportant convolutional kernels. We calculate the L n -norm of the feature map outputted by each convolutional kernel to evaluate the importance of the kernel. Furthermore, we use different L n -norms for different layers, e.g., L 1 -norm for the first convolutional layer, L 2 -norm for middle convolutional layers and L ∞ -norm for the last convolutional layer. With the ability of accurately identifying unimportant convolutional kernels in each layer, the proposed method achieves a good balance between model size and inference accuracy. Experimental results on CIFAR, SVHN and ImageNet datasets and an application example in a railway intelligent surveillance system show that the proposed method outperforms existing kernel-norm-based methods and is generally applicable to any deep neural network with convolutional operations.