A Novel Lightweight Architecture of Deep Convolutional Neural Networks

Abstract

Deep convolutional neural networks have achieved much success in many computer vision tasks. However, a network has millions of parameters which limit its inference speed and usage for some situations with limited storage space. Low-rank based methods and pruning methods are verified effective to compress the number of parameters and accelerate inference speed of deep convolutional neural networks. As the price, the performance of the networks decreases. To overcome this problem, in this paper, we design a novel low-rank and sparse architecture of convolutional neural networks. Besides accelerating inference speed and reducing parameters, our approach achieves better performance than baseline networks.