Tractable Probabilistic Models

Current deep learning methods tend to be overconfident in out-of-distribution cases and cannot easily handle incomplete data, two common problems when applying machine learning to real-world tasks. Probabilistic Circuits (PCs) are tractable models which attempt to overcome these problems: They are well-calibrated (their posterior probability accurately reflects their chance of being correct) and have native abilities to handle incomplete data, making them well-suited to real-world data modeling tasks.

We are working on advancing the capabilities of PCs, such that they can be used to solve real-world data science problems. For example, we worked on PCs that exploit symmetries as common in relational domains, and on using PCs for explainable outlier detection.