Knowledge compilation has seen much progress in the last decade, especially as work in this area has been normalized into a systematic study of tractable languages, their relative succinctness, and their efficient support for various queries. What has been particularly exciting is the impact that knowledge compilation has had on several areas, such as probabilistic reasoning and probabilistic databases. In this talk, I will discuss a new area, machine learning, which is bound to be significantly impacted by knowledge compilation. In particular, I will discuss recent work in which knowledge compilation has been used to learn probabilistic models under massive logical constraints, and over combinatorial objects, such as rankings and game traces. I will further identify and discuss three specific roles for knowledge compilation in machine learning, which arise in defining (a) more structured probability spaces, (b) more expressive queries, and (c) new types of datasets that significantly generalize the standard datasets used in the machine learning literature.
Joint work with Arthur Choi and Guy Van den Broeck. |