Abstract
Over the last decade, deep models have enjoyed wide empirical success. However, in practice, these models are not reliable due to their sensitivity against adversarial or natural input distributional shifts as well as a lack of meaningful reasoning behind their predictions. In this talk, I will show that a root cause of these issues is the heavy reliance of deep models on spurious features in their inferences. I will then explain our progress in understanding failure modes of deep learning and outline a roadmap towards developing trustworthy learning paradigms.