Using Feature Entropy to Guide Filter Pruning for Efficient Convolutional Networks

The rapid development of convolutional neural networks (CNNs) is usually accompanied by an increase in model volume and computational cost. In this paper, we propose an entropy-based filter pruning (EFP) method to learn more efficient CNNs. Different from many existing filter pruning approaches, our proposed method prunes unimportant filters based on the amount of information carried by their corresponding feature maps. We employ entropy to measure the information contained in the feature maps and design features selection module to formulate pruning strategies. Pruning and fine-tuning are iterated several times, yielding thin and more compact models with comparable accuracy. We empirically demonstrate the effectiveness of our method with many advanced CNNs on several benchmark datasets. Notably, for VGG-16 on CIFAR-10, our EFP method prunes 92.9% parameters and reduces 76% float-point-operations (FLOPs) without accuracy loss, which has advanced the state-of-the-art.