Three Modern Roles for Logic in AI | Adnan Darwiche | PODS 2020

Invited tutorial given at the database theory conference (PODS) on June 16, 2020. The associated PODS paper can be found at: The tutorial considers three modern roles for logic in artificial intelligence, which are based on the theory of tractable Boolean (and Arithmetic) circuits: (1) logic as a basis for computation, where compiling Boolean formulas into tractable circuits provides a systematic framework for solving NP and beyond NP problems (e.g., answering probabilistic queries on Bayesian networks); (2) logic for learning from a combination of data and knowledge in both supervised and unsupervised modes; and (3) logic for reasoning about the behavior of machine learning systems (e.g., explaining decisions, measuring their robustness and proving properties of machine learning classifiers). Among the tractable circuits considered in this tutorial are DNNF, d-DNNF, SDD and OBDD and PSDD circuits. A full course on the subjects discussed in this tuto