Using ensemble learning to improve radiation tolerance of CNNs in space applications
Flavio Ponzina (UCSD)Earth 2
With the increasing complexity of space missions, improving the resilience of AI systems against radiation-induced errors is critical. This paper proposes the use of application-level optimizations of Convolutional Neural Network (CNN)-based systems to improve robustness in space applications without relying on expensive shielded electronics. In particular, our work demonstrates how a recently proposed methodology for zero-overhead ensembles of CNNs, named E2CNN, can be applied to two case studies involving 6D pose estimation for satellite navigation. Our results show that E2CNN achieves up to 5.48% higher accuracy compared to single-instance models under different error conditions, suggesting their effectiveness in improving resilience and output quality in low-earth orbit applications. The experimental evaluations indicate that even without memory protection, E2CNN offers superior performance, making them a promising solution for AI-based space systems.