The Robustness of Modern Deep Learning Architectures against Single Event Upset Errors

Abstract

Neural Networks have revolutionized computer vision and the field has advanced so rapidly that in less than a decade neural networks are on par with human performance at image classification tasks. Applying neural networks to robotics should drastically increase the capabilities of robots to navigate and interact with the world around them. However, robots are often used to investigate environments too dangerous for humans. Such environments include areas with high levels of radiation such as the Fukushima reactor or Jovian moons. Radiation causes single event upset (SEU) errors such as unexpected bit flips so robots must be robust to those kinds of errors. Before robots can fully utilize the advances of neural networks, the networks must be tested for their robustness against SEUs. SEUs are most likely to cause bit flips within the trained parameters of a neural network. This is because the memory that stores the trained parameters takes up the most surface area of a computer, and thus is more likely to be hit by high energy particles. Previous papers in this field have focused on older networks such as Multi-Layer Perceptrons, but Convolutional Neural Networks are the current state of the art when it comes to object classification tasks. This paper tests several modern neural network architectures for their robustness to bit flips in their weights and examines which aspects of each different architecture lead to greater robustness. The different architectures tested are VGG16, ResNet50, and InceptionV3. The experiments show that all three networks display bimodal distributions under memory errors, meaning that the networks either retain their trained classification accuracy, or drop to very low accuracies. Additionally, it only takes a small number of memory errors compared to the total size of the network for the neural network's performance to significantly degrade. When comparing the three architectures, VGG16 was the least robust against random bit flips in its trained weights while ResNet50 and InceptionV3 had similar levels of robustness against SEU-type memory errors. Some possible reasons for ResNet50 and Inception V3's robustness are their use of batch normalization or their use of shortcut connections.