D-Pruner: Filter-Based Pruning Method for Deep Convolutional Neural Network

Abstract

The emergence of augmented reality devices such as Google Glass and Microsoft Hololens has opened up a new class of vision sensing applications. Those applications often require the ability to continuously capture and analyze contextual information from video streams. They often adopt various deep learning algorithms such as convolutional neural networks (CNN) to achieve high recognition accuracy while facing severe challenges to run computationally intensive deep learning algorithms on resource-constrained mobile devices. In this paper, we propose and explore a new class of compression technique called D-Pruner to efficiently prune redundant parameters within a CNN model to run the model efficiently on mobile devices. D-Pruner removes redundancy by embedding a small additional network. This network evaluates the importance of filters and removes them during the fine-tuning phase to efficiently reduce the size of the model while maintaining the accuracy of the original model. We evaluated D-Pruner on various datasets such as CIFAR-10 and CIFAR-100 and showed that D-Pruner could reduce a significant amount of parameters up to 4.4 times on many existing models while maintaining accuracy drop less than 1%.