What is it about?
Small satellites increasingly rely on artificial intelligence (AI) to perform complex tasks, such as controlling their positions and movements while flying in formation. However, satellites have limited power, and radiation in space can cause errors in their electronic circuits. Making AI reliable usually requires repeating calculations, which increases processing time and energy consumption. In this study, we developed a new approach that makes AI both faster and more resistant to errors. Our method identifies the calculations most vulnerable to errors and selectively repeats them to improve reliability. It also skips unnecessary calculations to save time and energy. Our results show that the proposed approach can reduce processing time by 16% compared with a conventional AI system without fault protection, while requiring only a small increase in circuit area. This technology could help enable faster, more reliable, and energy-efficient AI systems for future satellites and other devices operating in challenging environments.
Featured Image
Photo by Kevin Stadnyk on Unsplash
Why is it important?
As artificial intelligence (AI) becomes increasingly important for satellite operations, ensuring both reliability and energy efficiency is a major challenge. Radiation in space can cause errors in electronic circuits, while conventional fault-protection methods require additional computations, increasing processing time and energy consumption. Our research introduces a new approach that identifies which AI calculations are most vulnerable to errors and protects only those that need protection. Unlike conventional methods, our approach provides a mathematical guarantee on the maximum impact of certain errors, enabling more efficient and reliable AI processing. Importantly, our method achieves 16% faster processing than a conventional AI system without fault protection, while adding very little hardware overhead. This demonstrates that improving AI reliability does not necessarily require sacrificing speed or efficiency. These findings could contribute to the development of smaller, more energy-efficient, and more dependable AI systems for future satellite missions, particularly those involving coordinated satellite operations.
Perspectives
I believe that making AI more reliable should not come at the cost of significantly increased processing time or energy consumption. This is particularly important for small satellites, where computing resources and available power are extremely limited. One of the most exciting aspects of this research is that we demonstrated how AI can become both more reliable and faster by carefully selecting which computations need protection and eliminating unnecessary operations. I find it particularly encouraging that our approach can outperform even a system without fault protection in terms of processing speed. I hope this work will help change the perception that reliability and efficiency must always involve a trade-off. More broadly, I believe these ideas could contribute to the development of dependable, energy-efficient AI systems not only for space applications but also for robotics and other resource-constrained environments.
Yoichi Tomioka
The University of Aizu
Read the Original
This page is a summary of: Sparsity- and Tolerance-Aware Temporally Redundant Neural Networks, July 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3770743.3803961.
You can read the full text:
Contributors
The following have contributed to this page
