The Statistical Structure of Identifiable Generative Models

Explore the statistical structure of identifiable generative models in this 49-minute talk by Johnny Xi from Valence Labs. Delve into the foundations of causal representation learning, examining how unstructured low-level data can be processed into high-level causal units. Investigate the concept of identifiability in generative models and its importance in making unique inferences based on observations and assumptions. Learn about recent progress in exploring specific assumptions and techniques for identifiability, and gain insights into the statistical structure of the identifiability problem itself. The talk covers statistical modeling and identifiability from the ground up, applying this framework to analyze generative models as a class of statistical models. Discover generic identification results that describe the properties of the identifiability problem without assuming any specific model. Additionally, explore historical notes from factor analysis and ICA, understand the differences in their definition of "strong" identifiability, and learn how non-linear generators can admit perfectly unique solutions by fixing multiple latent distributions. The presentation concludes with a Q&A session, allowing for further discussion and clarification of the concepts presented.

- Intro + Background
- Generative Models as Statistical Models
- Indeterminacy in Generative Models
- Strong Identifiability
- Conclusions
- Q&A