A growing number of methods have been devised for learning probabilistic causal models from observational data alleviating the need for performing experiments to discover causal relations. Causal models represent relations such as quantity (variable) A causes B, meaning changing A will have an effect on the statistical distribution of B (e.g., starting to smoke increases the chances of getting lung cancer).

A fundamental and pervasive assumption in all of these methods is the Faithfulness assumption. Intuitively, Faithfulness states that if a A causes B, then A and B should be pair-wise correlated. However, it is possible that A and B are only multi-variately correlated, i.e., A correlates with B only when observed in the context of some other variables. There is growing theoretical and empirical evidence that Faithfulness may not hold in some systems (e.g., cells).

The talk will focus on new methods and theories for learning causal models in settings where the Faithfulness assumption may not hold. A basic understanding of statistical concepts such as multivariate and conditional correlations is assumed for the audience.