Soft Error Induced System Errors in Image Inference Applications with Glow-Compiled CNNs on MCUs

Image recognition, using Glow™ compiled and quantized convolutional neural networks, was performed on members of the MNIST and FashionMNIST datasets on two different microcontroller cores while exposing them to a neutron beam at the Los Alamos Neutron Science Center. Four types of application-level effects were observed, with some or all effects in each category resulting in application hangs or misclassifications of the test images. Three effects manifest continuously until the core is reset, and the implications of these continuous application-level effects within the ASIL-D context are immense. The data also suggests that image complexity has a positive association with the number of application-level errors expected. Baseline memory tests were also performed on the two different cores as a validation of our methodology.