🤖 machine learning

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

This paper proposes a hardware-aware neural architecture search (NAS) framework that integrates low-precision numerical constraints directly into the optimization process to bridge the accuracy gap between full-precision training and deployment on resource-constrained edge hardware, such as spaceborne VPUs.

Original authors: Parampuneet Kaur Thind, Vaibhav Katturu, Giacomo Zema, Roberto Del Prete

Published 2026-04-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Parampuneet Kaur Thind, Vaibhav Katturu, Giacomo Zema, Roberto Del Prete

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a professional chef training for a high-stakes cooking competition.

The Problem: The "Fancy Kitchen" Trap

Up until now, most AI researchers have been training their "AI chefs" in massive, state-of-the-art professional kitchens (these are like high-powered GPUs). These kitchens have every gadget imaginable, infinite counter space, and the most expensive ingredients. The chefs learn to make perfect, gourmet meals in this perfect environment.

However, the actual "competition" for this paper isn't in a luxury kitchen—it’s on a tiny, cramped food truck orbiting in space (this is the Spaceborne Edge AI). This food truck has very little power, tiny counters, and only basic, low-quality tools (this is the low-precision hardware).

The Mismatch: When the gourmet chef arrives at the food truck, they realize they can't use their fancy electric whisk or their precision scales. They try to make the same complex recipe with a manual hand-whisk and a cheap plastic spoon, and suddenly, the food tastes terrible. The "accuracy" of the meal has plummeted because the chef wasn't trained for the reality of the tools they’d actually be using.

The Solution: Training for the Food Truck

The researchers in this paper decided to change the training process. Instead of letting the AI chef train in the luxury kitchen and then sending them to the food truck, they introduced "simulated struggles" during the training phase.

They used a method called Deployment-Aligned NAS. Here is how it works:

The Architect (NAS): They use an automated system (Neural Architecture Search) that acts like a designer, trying out thousands of different "recipes" (network architectures) to see which one is the fastest and most accurate.
The Simulation (Low-Precision Training): While the designer is testing these recipes, the researchers don't let them use the fancy tools. They force the AI to practice using the "cheap plastic spoons" and "manual whisks" (the FP16 numerical constraints) right from the start.
The Real Test (Device-in-the-loop): They take the best recipes and actually cook them on the tiny "food truck" hardware (the Intel Movidius VPU) to see how they really perform in space.

The Results: A Robust Chef

By forcing the AI to deal with the limitations of low-precision math during its "schooling," the researchers found two major things:

No "Accuracy Crash": Usually, when you move a model from a big computer to a tiny space chip, the accuracy drops significantly (like a chef's meal becoming inedible). With this new method, the accuracy stayed much higher. They "recovered" about two-thirds of the quality that is usually lost.
Better "Recipes": The AI designer learned to pick architectures that weren't just "smart," but were "tough." It chose designs that were naturally resistant to the errors caused by using low-precision math.

Why does this matter for Space?

In space, we can't send massive amounts of data back to Earth because it's too slow and expensive. We need satellites to "think" for themselves—for example, spotting a ship in the ocean or detecting a forest fire—and make decisions instantly.

This paper provides a way to build "brains" for satellites that are small enough to fit on a tiny chip, fast enough to work in real-time, and smart enough to actually work when the "kitchen" gets tough.

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

1. Problem Statement

The deployment of Deep Neural Networks (DNNs) on spaceborne edge accelerators (e.g., CubeSats) is constrained by strict limits on power, memory, and latency. While Hardware-Aware Neural Architecture Search (NAS) has been developed to optimize architectures for specific devices, a critical "numerical mismatch" remains.

Current NAS pipelines typically optimize architectures using full-precision (FP32) training on GPUs and only apply low-precision adaptation (like quantization) after the search is complete. This decoupling causes a significant gap between optimization-time performance and deployment-time execution on low-precision hardware (like the Intel® Movidius™ Myriad™X VPU), where reduced mantissa precision and rounding errors can lead to substantial accuracy degradation.

2. Methodology

The authors propose integrating deployment-aligned low-precision training directly into the hardware-aware NAS evaluation loop. Instead of treating precision as a post-processing step, the search process itself is biased toward architectures that are numerically robust.

Search Strategy: The framework uses a Genetic Algorithm (GA) to navigate a discrete search space of single-path networks (up to 6 blocks). The search space includes various primitives like ConvAct, MBConv, CSPConv, and ResNet-style blocks.
Hardware-in-the-Loop (HITL): Candidate architectures are evaluated directly on the target Intel® Movidius™ Myriad™X VPU. The fitness function $F(a)$ jointly optimizes task accuracy (mIoU) and inference throughput (FPS).
Deployment-Aligned Training (The Core Innovation):
- Two-Stage Training: Candidates are first trained in FP32, then undergo FP16-aware fine-tuning.
- FP16 Projection: During fine-tuning, a projection operator $P(\cdot)$ is injected into the forward pass to simulate reduced mantissa precision and rounding.
- Straight-Through Estimator (STE): To maintain training stability, gradients are accumulated in FP32 while weights are rounded using STE.
- Activation Clipping: To prevent overflow and match the target device's dynamic range, activations are clipped to a symmetric bound ( $|x| \leq 12$ ).

3. Key Contributions

Deployment-Aligned NAS: A novel integration of low-precision training within the NAS loop, aligning optimization behavior with actual hardware execution.
Search-Level Numerical Alignment: Demonstrating that incorporating precision constraints during the search phase changes the selection dynamics, favoring "numerically robust" architectures.
Realistic Validation: Providing a comprehensive, device-in-the-loop evaluation on a VPU specifically for a maritime vessel segmentation task in Earth Observation (EO).

4. Results

The framework was evaluated on the HRSC2016 dataset for vessel segmentation. The comparison between Post-Training Quantization (PTQ) and the proposed Deployment-Aligned training yielded the following:

Quantitative Gains:
- PTQ Baseline: Converting an FP32 model to FP16 on-device caused the mIoU to drop from 0.85 to 0.78.
- Proposed Method: Deployment-aligned training achieved an on-device mIoU of 0.826 for the same architecture (95,791 parameters).
- Recovery: The method recovered approximately two-thirds of the accuracy gap induced by deployment without increasing model complexity.
Qualitative Improvements:
- PTQ models exhibited structural artifacts, such as fragmented vessel contours, internal holes in predicted regions, and false negatives for low-contrast targets.
- Deployment-aligned models preserved vessel morphology, connectivity, and hull contours much more reliably under low-precision execution.

5. Significance

This research shifts the paradigm of edge AI design from "optimize for accuracy, then compress" to "optimize for deployment-consistent execution." For spaceborne missions where autonomous, real-time decision-making is critical (e.g., disaster response or maritime monitoring), this approach ensures that the intelligence deployed on-board is both computationally efficient and numerically reliable, preventing the "brittle" performance often seen when high-precision models are forced into low-precision hardware.