W375 Westgate Building
10:00AM
Just as Newton observed a falling apple and uncovered the law of gravity, unveiling and leveraging the latent mechanisms behind observational data lies at the core of scientific discovery. This talk studies when and how latent variable models can be identified and exploited from purely observational data. I present a general theoretical framework showing how a structural view of the latent process yields nonparametric identifiability guarantees under realistic assumptions, and how these guarantees extend to flexible, task-driven objectives while leading to practical learning principles that benefit frontier models. I then demonstrate how leveraging identifiable latent mechanisms improves trustworthiness and effectiveness in both digital and embodied agents. These insights provide a principled path beyond the bottleneck of correlations fitting that cannot be resolved by scaling alone, toward next-generation models with causal understanding and actions.
Additional Information:
Yujia Zheng is a PhD candidate at Carnegie Mellon University, working in the causal learning and reasoning group advised by Prof. Kun Zhang. His research focuses on trustworthy machine learning, causal representation learning, and causality-based learning. His work has been supported by the Meta AIM PhD Fellowship and recognized with best paper honorable mention, orals, and spotlights at leading venues such as NeurIPS and ICML. He leads the development of causal-learn, a widely used open-source platform for causal learning. He has served as Publicity Co-Chair of UAI 2022, Local Arrangements Co-Chair of UAI 2023, and Workflow Co-Chair of CLeaR 2025 and 2026.
Details...
