Efficient Network Inference via Hardware-Aware Architecture Search, Model Pruning & Quantization

As global navigation satellite system (GNSS) receivers become increasingly critical for infrastructure, there is a growing need for robust interference monitoring. While deep neural networks (DNNs) offer high accuracy for classifying and characterizing interference, they are often too computationally expensive and memory-intensive for deployment on embedded microcontrollers. This paper investigates how to bridge the gap between high-performance AI and the strict resource constraints of edge devices.

The authors employ a two-pronged strategy to optimize model efficiency. First, they apply iterative structured pruning and post-training static 8-bit integer quantization to a compact MCUNet baseline. Unlike element-wise pruning, structured pruning removes entire filters or channels, allowing for direct execution on standard hardware without needing specialized sparse-tensor support. Second, they utilize PrototypeNAS, a hardware-aware zero-shot neural architecture search (NAS) framework. This method optimizes the network architecture and pruning configuration simultaneously using analytical proxy metrics—such as feature map correlation and activation space separation—to estimate model quality without requiring full training. This allows the researchers to identify a Pareto-optimal set of models that balance computational cost (FLOPs) against memory (RAM/ROM) constraints.

The experiments show that the MCUNet baseline contains significant structural redundancy. Iterative structured pruning reduced the parameter count from approximately 593,000 to 18,500 (a 90% reduction) while maintaining competitive classification accuracy. The authors demonstrate that by using zero-shot NAS, they can further refine these models to be specifically tailored for diverse embedded platforms, including the iMXRT1062 MCU and various Raspberry Pi models. The resulting models maintain high performance in both interference classification and fine-grained characterization tasks, proving that automated hardware-aware design is a viable path for real-time edge AI in GNSS applications.