Reliability of Vision Transformers and CNNs on Edge AI Systems Under Neutron Radiation

The reliability of neural networks on Edge AI platforms is crucial for safety-critical applications, especially in environments exposed to radiation. This article presents an experimental evaluation of the reliability of Vision Transformers (ViTs) and convolutional neural networks (CNNs) deployed on low-power system-on-chip (SoC) devices, specifically the Jetson Orin Nano. The study explores the impact of neutron radiation on these models, focusing on the effect of optimization techniques such as tensor runtime (TensorRT), data precision reduction, and weight quantization. Results indicate that while optimization can improve inference performance, it may also increase the error rates in outputs. Additionally, including main memory in the radiation tests resulted in severe persistent errors, highlighting the need for thorough reliability assessments in real-world scenarios. A comparison between ViTs and CNNs revealed that ViTs, despite their complexity, show comparable or better reliability on modern SoCs, which is critical for deploying AI in safety-critical environments like aerospace and autonomous driving.