Deployment-Aligned Low-Precision Neural Architecture Search for Spaceborne Edge AI
Training satellite AI models in low-precision arithmetic during architecture search — not after — recovers most of the accuracy lost when deploying to space-grade edge chips.
The Thesis
Most AI models designed for edge hardware are built and tested at full numerical precision, then converted to lower precision only at deployment. This mismatch causes accuracy to quietly degrade when the model finally runs on real hardware. This paper proposes baking low-precision arithmetic (specifically FP16 — a 16-bit floating point format that trades some numerical exactness for speed and power savings) directly into the architecture search process. The target application is vessel detection from satellite imagery, running on Intel's Myriad X chip — a power-sipping processor designed for vision tasks on spacecraft and drones. The result: a tiny 95,791-parameter model that recovers roughly two-thirds of the accuracy gap caused by post-hoc precision conversion, without adding any model complexity.
Catalyst
Small satellites (CubeSats and their successors) have matured to the point where operators want real-time onboard inference rather than downlinking raw imagery for ground processing — but the compute budgets remain brutal. Intel's Myriad X VPU has become a de facto edge inference chip for space applications precisely because it balances throughput with power draw. The gap between simulation-time training and on-chip deployment has become an urgent engineering problem as these missions move from prototype to production.
What's New
Prior hardware-aware neural architecture search (NAS) pipelines — including popular methods like Once-for-All and ProxylessNAS — optimize model structure using full-precision (FP32) math throughout, then apply quantization or precision reduction as a separate post-processing step. That late conversion often surprises the model with numerical behavior it was never trained to handle, causing accuracy to drop. This paper instead exposes candidate architectures to FP16 constraints during the search-and-fine-tuning loop itself, so the evolutionary search selects architectures that are inherently robust to the precision regime they will actually run in.
The Counter
The paper tests exactly one task (vessel segmentation), one chip (Myriad X), and one precision format (FP16). That is a very narrow base on which to claim a general principle about deployment-aligned NAS. The mIoU scores reported (0.826 vs 0.78) are meaningful, but there is no comparison against a state-of-the-art quantization-aware training baseline applied post-search — which is the real competition, not naive post-training conversion. FP16 is also not the most aggressive precision regime in actual space deployments; INT8 and INT4 quantization present harder problems that this approach does not address. The model is also extremely small (under 100K parameters), which limits how much the result generalizes to more capable architectures that might actually be needed for complex scene understanding. Finally, space-qualified hardware evolves slowly; by the time this methodology reaches production missions, the target chip may have changed.
Longs
- INTC — Myriad X VPU is the direct deployment target in this paper
- RKLB (Rocket Lab) — small satellite launch and bus provider feeding the spaceborne edge AI market
- KTOS (Kratos Defense) — satellite ground systems and space AI payloads
- UFO (Procure Space ETF) — broad exposure to commercial space segment
- PLTR — defense and intelligence satellite analytics pipelines
Shorts
- Vendors selling post-training quantization tools as standalone products — this paper argues the right fix is upstream, not downstream
- Cloud-based satellite imagery analytics firms whose moat depends on ground-side inference; onboard inference reduces their data pipeline advantage
Enablers (Picks & Shovels)
- Intel OpenVINO toolkit — the compiler and runtime that converts models to Myriad X inference format
- ONNX (Open Neural Network Exchange) — the interoperability format used to move models between training and deployment
- ESA (European Space Agency) open datasets for maritime vessel segmentation
- Evolutionary NAS frameworks such as NSGA-II used for multi-objective architectural search
Private Watchlist
- Latitude Connect (spaceborne edge AI payloads)
- Orbital Sidekick (satellite-based spectral intelligence, onboard inference)
- Xona Space Systems (edge compute for satellite navigation)
- SynSpace (satellite imagery analytics)
Resources
The Paper
Designing deep networks that meet strict latency and accuracy constraints on edge accelerators increasingly relies on hardware-aware optimization, including neural architecture search (NAS) guided by device-level metrics. Yet most hardware-aware NAS pipelines still optimize architectures under full-precision assumptions and apply low-precision adaptation only after the search, leading to a mismatch between optimization-time behavior and deployment-time execution on low-precision hardware that can substantially degrade accuracy. We address this limitation by integrating deployment-aligned low-precision training directly into hardware-aware NAS. Candidate architectures are exposed to FP16 numerical constraints during fine-tuning and evaluation, enabling joint optimization of architectural efficiency and numerical robustness without modifying the search space or evolutionary strategy. We evaluate the proposed framework on vessel segmentation for spaceborne maritime monitoring, targeting the Intel Movidius Myriad X Visual Processing Unit (VPU). While post-training precision conversion reduces on-device performance from 0.85 to 0.78 mIoU, deployment-aligned low-precision training achieves 0.826 mIoU on-device for the same architecture (95,791 parameters), recovering approximately two-thirds of deployment-induced accuracy gap without increasing model complexity. These results demonstrate that incorporating deployment-consistent numerical constraints into hardware-aware NAS substantially improves robustness and alignment between optimization and deployment for resource-constrained edge Artificial Intelligence (AI).