Structured pruning reduces AI fault risks and saves energy in space
Exploring the Trade-Off Between Structured Pruning and Fault Tolerance in Deep Neural Networks for Space Applications
Machine Learning
Summary
Deep neural networks used in space must handle random bit flips caused by cosmic rays. The authors studied how making these AI models narrower, by structured pruning, affects their ability to handle such faults. They found that while narrower models are more sensitive to faults during each operation, their shorter running time means they encounter fewer faults overall. This balance means it’s possible to make AI models more efficient without making them less reliable for space missions. Their work helps guide the design of AI systems that save energy and resist faults in space environments.
What this means in practice
- •For satellite system engineers: Optimize AI onboard satellites to reduce energy use and processing time while maintaining fault tolerance against cosmic ray effects.
- •For autonomous vehicle developers: Design AI models that balance robustness and efficiency when deployed in safety-critical systems requiring fault resilience.
Authors
Toon Vinck, Naïn Jonckers, Jaro De Roose, Jeffrey Prinzie, Peter Karsmakers
Abstract
Deep Neural Networks (DNNs) inherently exhibit a degree of robustness to bit-level faults due to their distributed representation of information. As a model increases in width, this information becomes more dispersed, theoretically reducing the impact of any single bit fault. In this paper, we empirically investigate the relationship between model width and robustness to Single Event Upsets (SEUs). We conduct a comprehensive experiment in which baseline models undergo iterative structured pruning to reduce their width while preserving task performance as much as possible. At each pruning stage, we run a targeted fault-injection campaign to evaluate the model's performance under simulated bit-flip scenarios. Our results show that, although structured pruning increases per-inference sensitivity to faults by reducing redundancy, this effect is effectively counterbalanced by shorter execution time, which lowers the probability of encountering an SEU. These findings suggest that structured pruning can yield significant energy and latency savings without compromising overall reliability, providing useful guidance for designing robust AI systems for space applications.