Deep compression of convolutional neural networks with low-rank approximation

Abstract

The application of deep neural networks (DNNs) to connect the world with cyber physical systems (CPSs) has attracted much attention. However, DNNs require a large amount of memory and computational cost, which hinders their use in the relatively low-end smart devices that are widely used in CPSs. In this paper, we aim to determine whether DNNs can be efficiently deployed and operated in low-end smart devices. To do this, we develop a method to reduce the memory requirement of DNNs and increase the inference speed, while maintaining the performance (for example, accuracy) close to the original level. The parameters of DNNs are decomposed using a hybrid of canonical polyadic–singular value decomposition, approximated using a tensor power method, and fine-tuned by performing iterative one-shot hybrid fine-tuning to recover from a decreased accuracy. In this study, we evaluate our method on frequently used networks. We also present results from extensive experiments on the effects of several fine-tuning methods, the importance of iterative fine-tuning, and decomposition techniques. We demonstrate the effectiveness of the proposed method by deploying compressed networks in smartphones.