I. Introduction
Deep neural networks (DNNs) have been implemented in many cyber–physical systems (CPSs) for solving complex machine learning problems [1], [2]. However, deploying DNNs in a real-time environment is still subject to many limitations, due to their unpredictable execution times resulting from complex runtime behavior on appropriative hardware. Moreover, in many embedded applications, the timing constraints may change dynamically during runtime, e.g., smart autonomous cars [3] and drones [4], where the system must deal with changes in application requirements, such as the sampling period and resource availability.