부산대학교 도서관

Deep Learning (DL) models are very efficient for many applications including, computer vision, natural language processing… Yet DL models require important computation resources making it particularly difficult to deploy these applications in constrained environments such as the Internet of Things (IoT). Offloading DL models to the cloud is one solution to this problem but has a number of drawbacks related to the trade-off between efficiency and latency, and other privacy issues. In this paper we try to solve this problem using two approaches, first by sharing the DL model between the cloud and the device and second by optimising the execution of the model using early exiting where inputs do not need to execute the model entirely. Both approaches are optimized automatically in order to choose the best sharing point and the best exiting point according to input. The solutions proposed could be easily generalized and are independent of applications and offer a good alternative in order to execute DL models locally.