Traditional neural architecture pruning methods primarily focus on trade-offs between inference accuracy and computational cost ( e.g. , FLOPs), without accounting for system-level behaviors. In the context of cyber-physical systems (CPS), perception models and the downstream controllers that rely on their outputs are typically designed independently. In this paper, we study the joint design of neural architecture pruning strategies and the controllers that utilize the resulting inference outputs. We propose two key innovations: (i) incorporating system-level safety metrics into the pruning optimization loop, and (ii) a regression-based scaling law method that efficiently models the relationship between pruning ratios and system-level safety outcomes. This enables rapid pruning guided by system-level constraints, avoiding the inefficiency of exhaustive search or conventional neural architecture search approaches that focus on local accuracy or latency metrics. We demonstrate our approach on YOLO-based perception models deployed in control systems for distance estimation. To the best of our knowledge, this is the first work to introduce a scaling-law-guided pruning method for YOLO-based perception models, offering a principled design automation framework for safety-aware, machine learning–enabled CPS.
Chen et al. (Fri,) studied this question.