Deployment-Aligned Low-Precision Neural Architecture Search for Spaceborne Edge AI

A team led by Parampuneet Kaur Thind from the University of Toronto and the European Space Agency (ESA) has developed a novel neural architecture search (NAS) framework that explicitly addresses the deployment gap in low-precision edge AI systems. Traditionally, hardware-aware NAS optimizes models under full-precision assumptions and applies low-precision adaptation only after search, which often degrades on-device performance. The new approach, called deployment-aligned low-precision NAS, integrates FP16 numerical constraints directly into the search process, ensuring architectures are optimized from the ground up for low-precision deployment.

The researchers evaluated their method on vessel segmentation for spaceborne maritime monitoring using the Intel Movidius Myriad X VPU. While post-training quantization reduced on-device performance from 0.85 to 0.78 mIoU, their deployment-aligned approach achieved 0.826 mIoU on the same architecture (95,791 parameters). This recovery of approximately two-thirds of the deployment-induced accuracy gap was achieved without increasing model complexity or modifying the search space or evolutionary strategy.