Time: July 7, 10-11 a.m.

Location: KEC 1007 

Zoom link.

Please attend in person if you can. Sriraam will be visiting here for that whole week. If you'd like to meet with him when he is here, please email him at sriraam.natarajan@utdallas.edu with your availability. 

10:00-10:45 a.m.: Review of probabilistic inference and introduction to tractable probabilistic models.

10:45-11 a.m.: Break.

11:00-11:45 a.m. Human-aligned learning of tractable probabilistic models and their applications to health care.

There will be coffee and cookies. 

Abstract: Probabilistic models deal with uncertainty in data-driven decision-making and modeling in a principled manner. Recent advances in GPU-accelerated computation have enabled probabilistic models to scale to large and complex data sets. However, as models grow in complexity, efficient exact inference (querying the model) becomes a challenge, hindering their feasibility in high-stakes domains like healthcare. This talk aims to introduce Deep Tractable Probabilistic Models (DTPMs), a special class of probabilistic models that balance expressiveness and tractability. First, I will introduce these models by explaining their semantics. Then, I will present our recent work on learning these models in the presence of rich, human-provided domain knowledge. Finally, I will conclude by closing-the-loop, i.e., explaining these models back to the expert. I will provide examples of healthcare problems, specifically those of adverse pregnancy outcomes and pediatric critical care to demonstrate the versatility of the learning algorithms.