ABSTRACT
Causal models follow an AI paradigm that is diametrically opposed to that of deep learning. Where deep learning is happy to trade off interpretability for increased predictive power, causal models are built on extreme interpretability. Rather than correlations, they capture causeand-effect relationships between variables, and use these relationships in conjunction with data in order to answer quantitative queries about the system being modeled. In this chapter, several ways in which causal and non-causal associations between variables arise are discussed, including from direct causation, through a confounder, or from a selection bias due to a common effect. The next section provides an introduction to causal graphs, and illustrates the effect of an intervention on the graph. Pearl’s do-calculus for converting a causal expression to a statistical expression is briefly described, and illustrated by a backdoor adjustment calculation to correct for a confounder. The importance of assumptions for causal inference is discussed, and the concept of causal discovery is described with examples. The final section covers several ways in which causal methods and equation discovery are combined with machine learning and deep learning techniques.
