Gastvortrag: Guy Van den Broeck
Mittwoch, 07.10.2020, 17.00 Uhr
Ort: Online session
Vortragender: Guy Van den Broeck
Probabilistic graphical models are a rich staple of probabilistic AI. However, they make a very specific choice of abstraction: probability distributions are represented by their variable-level (in)dependencies. In this talk I present some recent work on probabilistic models that go beyond classical PGMs, and make a radically different choice of abstraction; one that is computational. Concretely, I will discuss two classes of models: probabilistic circuits and probabilistic programs. Probabilistic circuits represent distributions through the computation graph of probabilistic inference. They move beyond PGMs by guaranteeing tractable inference for certain classes of queries. Probabilistic programs represent distributions through higher-level primitives of computation: iteration, branching, and procedural abstraction. They move beyond PGMs by looking "inside" of the dependencies. Finally, I will illustrate how these two computational abstractions are themselves closely related, by showing how the Dice probabilistic programming language compiles probabilistic programs into probabilistic circuits for inference.