MENDNet: Just-in-time Fault Detection and Mitigation in AI Systems with Uncertainty Quantification and Multi-Exit Networks

In this work, we design the MENDNet framework by modifying the codebase of Shallow Deep Networks by Y. Kaya et al (Link: https://github.com/yigitcankaya/Shallow-Deep-Networks) and incorporating appropriate additions for the purpose of fault detection and mitigation in Deep Neural Networks. The models are trained using train_networks.py. The experminets.ipynb notebook illustrates samples of all the experiments carried out in this work. Models used: VGG16, WideResNet32 and MobileNet. Datasets used: CIFAR-10 and CIFAR-100.