How reliably can one deploy Deep Neural Networks (DNNs) to real-world applications? Answering this very question requires understanding all the relevant failure modes of DNNs, which have been largely investigated, however in small subsets and independently by different subgroups of researchers from multiple communities. Nevertheless, these failure modes might not be as dissimilar as we think they are, and to understand their interplay, similarities and dissimilarities, it is important to discuss them together. This is precisely the goal of our proposed tutorial where we will offer an overview of the efforts towards reliable DNNs (main challenges, evaluations, research directions, and trends) as well as in-depth coverage of the various paradigms for achieving it (uncertainty estimation, calibration, OOD detection, robustness to distribution shift).

A vast amount of literature in both machine learning and computer vision communities addressed recently one or more specific facets of reliability. This tutorial will focus on pragmatic and scalable approaches that can effectively improve reliability on complex computer vision tasks: from image classification trained on large repositories (e.g., ImageNet) to automatic perception for autonomous driving (object detection, segmentation, depth estimation, etc.). Specifically, the proposed tutorial would cover the following subjects: (1) Uncertainty estimation and different blindspots of modern DNNs, (2) Efficient Deep Ensembles, (3) Calibration of DNNs, (4) Out-of-distribution detection, (5) Robustness and generalization under distribution shift (adverse weather conditions).