Abstract
Tractable circuits are located at the intersection of neural networks, logic, and probabilistic models, making them a promising methodology for reliable reasoning in AI. In this talk, I will showcase two recent advancements of tractable circuits. First, I will discuss their application to side-channel attacks in crypto systems, which requires a neuro-symbolic approach and combines logical knowledge, learning and probabilistic inference. Specifically, this approach enables for the first time side-channel attacks using fast and (almost) exact inference. Second, I will present recent endeavors to extend probabilistic circuits to continuous latent variable models, representing a new paradigm in circuit learning and inference and setting new state-of-the-art in various applications.