This theme of research explores the relation between learning machines and reasoning frameworks. Instead of viewing machine learning systems as simple statistical models, I argue in (Bottou, 2011) that one should now study how they combine. The algebraic properties of these combinations can then be mapped to the algebraic properties of reasoning models. In (Bottou et al., 2012), we show how to leverage causal reasoning to understand the behavior of complex learning systems interacting with their environment and predict the consequences of changes to the system. Such predictions allow both humans and algorithms to select changes that improve both the short-term and long-term performance of such systems. This work was successfully applied to the Bing ad placement system around 2010-2011.