Exploiting Kernel Sparsity and Entropy for Interpretable CNN Compression

TLDR

CNN compression has become a major research focus, yet most methods fail to interpret inherent structures to identify implicit redundancy. This paper investigates CNN compression from a novel interpretable perspective. The authors link input feature maps to 2D kernels in a theoretical framework, introduce a kernel sparsity and entropy (KSE) indicator to quantify feature‑map importance, apply KSE‑based kernel clustering for precise compression, and evaluate the approach on CIFAR‑10, SVHN, and ImageNet 2012. The KSE‑based method achieves rapid, layer‑wise compression, outperforming prior approaches with 4.7× FLOPs reduction and 2.9× compression on ResNet‑50 while incurring only a 0.35% top‑5 accuracy drop on ImageNet 2012.

Abstract

Compressing convolutional neural networks (CNNs) has received ever-increasing research focus. However, most existing CNN compression methods do not interpret their inherent structures to distinguish the implicit redundancy. In this paper, we investigate the problem of CNN compression from a novel interpretable perspective. The relationship between the input feature maps and 2D kernels is revealed in a theoretical framework, based on which a kernel sparsity and entropy (KSE) indicator is proposed to quantitate the feature map importance in a feature-agnostic manner to guide model compression. Kernel clustering is further conducted based on the KSE indicator to accomplish high-precision CNN compression. KSE is capable of simultaneously compressing each layer in an efficient way, which is significantly faster compared to previous data-driven feature map pruning methods. We comprehensively evaluate the compression and speedup of the proposed method on CIFAR-10, SVHN and ImageNet 2012. Our method demonstrates superior performance gains over previous ones. In particular, it achieves 4.7× FLOPs reduction and 2.9× compression on ResNet-50 with only a top-5 accuracy drop of 0.35% on ImageNet 2012, which significantly outperforms state-of-the-art methods.