Abstract
With the rise of modern neural network and deep learning techniques, rapid progress has been achieved on highly complex problems such as computer vision and natural language processing. The principal strengths of these methods lie in their ability to approximate complex functions from high-dimensional data without requiring carefully, hand-designed models of the underlying phenomena. However, if the methods are not constrained in ways that reflect the structural constraints underlying the data generating processes which brought about the observations, these methods will suffer from a number of limitations. These limitations include bias, poor model robustness, a dependence on expensive, annotated datasets, and an inability to reason about cause-effect relationships in the world. We demonstrate that each of these limitations can be addressed by integrating structural and causal constraints, thus highlighting that a hybrid between modern deep learning and a more traditional infusion of domain expertise represents an important approach to designing machine learning systems today. This is particularly true if researchers wish to use machine learning to develop trustworthy, interpretable, and/or causal decision processes