Deep Neural Networks (DNNs) are now widely used in Cyber-Physical Systems (CPSs), both for sensor data or perception processing and also as neural network controllers. However, resource constraints often prevent a full local deployment of the DNNs, e.g., on edge devices. Implementing them on the cloud, on the other hand, is associated with large delays and large volumes of data transfers. This has resulted in the emergence of Split Computing (SC), where a part of the DNN is implemented on an edge device and the rest in the cloud. However, how to design control strategies with such a setup has not been sufficiently investigated in the past. In this paper, we study controller design strategies where state estimates from sensor data processed on an edge device are combined with estimates obtained from the cloud. While the former is associated with low delays, the state estimates have higher errors or uncertainties. The estimates from the cloud, obtained with larger DNNs, are, however, delayed. We show that the problem of sizing the DNNs on the edge and the cloud can be formulated as an optimization problem with the goal of maximizing system safety.
TBA.