Ph.D. Forum Abstract: Adapting Deep Learning-Based Sensing Systems to Cyber-Physical Dynamics

Deep neural networks are increasingly used in cyber-physical systems (CPS) to achieve better performance. The dynamic nature of a CPS resulting from the interactions between the physical processes and computational elements can affect the performance of the DNN model. The goal of my research is to explore and design adaptation approaches for a CPS that can achieve high and robust performance under dynamic conditions. The first approach aims to optimize resource allocation in CPS with concurrent sensors. The second approach focuses on quality inspection in production lines, aiming to optimize energy consumption and performance. The third approach is designed for autonomous driving and involves adapting the neural architecture in real-time to meet dynamic deadlines.